content
stringlengths 1
15.9M
|
|---|
\section{Introduction} \label{sec:intro}
Conspicuous filamentary structures are ubiquitous in the interstellar medium (ISM). But how the filaments form is not clearly understood. Numerical simulations of the multiphase atomic ISM in Galactic disk show that cold, dense gas prefer to organize itself naturally in the filamentary network \citep{Tasker2009}. Widths derived from observed filaments support the argument that the filaments may form as a result of the dissipation of large-scale turbulence \citep{Arzoumanian2011}. Filaments are constituent of molecular clouds and may play a pivotal role in the process of star formation \citep[e.g.][]{Schneider1979,Arzoumanian2011,ZhangM2019}. According to simulation, massive star forming clumps are filamentary structures in their early stages \citep{Smith2009}. And \citet{Andre2010} observed filaments in both quiescent and star-forming clouds, where a large part of prestellar cores reside. \\
Filamentary structures have been observed by different tracers. For the reason that they are infrared dark, filaments can be found from extinction maps at optical \citep{Schneider1979} and infrared wavelengths \citep{Jackson2010,Kainulainen2013}. Far-infrared or submillimeter dust emission maps are also employed \citep{Andre2010,Hacar2013,Li2016}. Researches on filaments have been focused on nearby star-forming regions such as Taurus and Orion \citep{Palmeirim2013,Takahashi2013,Kainulainen2017}, or substructures such as fibers \citep{Hacar2018}. \\
A variety of filaments have been found with diverse lengths and morphology. For instance, infrared dark cloud ``Nessie'' \citep{Jackson2010}, is found to be as long as 430 pc in the CO map and runs along the Scutum-Centaurus Arm in position-position-velocity space \citep{Goodman2014}. Careful studies of the individual filaments provide important hints on their properties, the role they play in the star formation and so on. However, on one hand, individuality is inevitable. On the other hand, more distant filaments which also contain high-mass star formation have not been studied comprehensively. So an unbiased sample of filaments in the Milky Way is needed.\\
Large-scale ($\ge 10$\,pc) filaments may be connected to the large-scale, Galactic spiral structures. From a position-velocity analysis, \citet{Goodman2014} suggest that Nessie with length of hundreds of parsecs forms a bone-like feature that closely follows the center of the Scutum-Centaurus Arm of the Milky Way. The inspection of this linkage between large-scale filaments and spiral arms is extended to other Galactic positions in the inner Galactic plane \citep[e.g.][]{Ragan2014,Zucker2015,Wang2015,Abreu2016}. Unlike small-scale filaments \citep[e.g.][]{Schisano2014,Schisano2020,Koch2015,Li2016,Mattern2018} that nearby ones could be found, large-scale filaments in the Milky Way are typically far from us, so researches towards them have not come up until recent years, when modern multiwavelength survey that covers Galactic plane at high resolution and sensitivity. Several powerful algorithms have been developed to identify filaments through intensity or column density map such as DisPerSE, FilFinder, local Hessian matrix, getFilaments and related algorithms based on wavelet transforms \citep{Sousbie2011a,Sousbie2011b,Koch2015,Menshchikov2013,Molinari2010b,OssenkopfOkada2018,Schisano2014}. The algorithms above perform well in finding parsec-scale filaments. However, large-scale filaments are rather different. One important feature is that, they may not be continuous in mm or submm continuum images. Therefore, although several catalogues of large-scale filaments have been produced in the past few years \citep{Ragan2014, Wang2015,Wang2016, Abreu2016,Zucker2018} and different searching methods have different selection criteria to identify filaments, of which most are by inspection of dust characteristic artificially, except \citet{Wang2016}. \\
\citet{Wang2016} adopted minimum spanning tree (MST) algorithm to automatically identify filaments by connecting velocity-coherent clumps. The minimum spanning tree, a construct from graph theory, is the unique set of straight lines (``edges'') connecting a given set of points (``vortices'') without closed loops, such that the sum of the edge lengths is a minimum. Two algorithms have been developed independently by \citet{Kruskal1956,Prim1957}. MSTs have been associated for the fist time to cluster analysis by \citet{GowerRoss1969}. In astrophysics, minimum spanning trees have so far mainly been used to analyse the large-scale distribution of galaxy or galaxy clusters, and filamentary feathers have been found \citep[e.g.][]{Barrow1985,AdamiMazure1999,Doroshkevich2004,Colberg2007,ParkLee2009,Alpaslan2014,Naidoo2020,Pereyra2020}. MSTs have also been employed to identify star clusters \citep[e.g.][]{Cartwright2004,Schmeja2006,Gutermuth2009,Wu2017}. In high energy astrophysics, MSTs have been used for $\gamma$-ray source detection \citep[e.g.][]{DiGesuSacco1983,Campana2008,Campana2013}. Recently, MSTs have also been introduced to quantify core separations and mass segregation \citep[e.g.][]{Sanhueza2019,Dib2019}. However, MSTs have also been criticized for giving unreliable cluster catalogues due to ``chaining'' of unrelated structure when noise is present \citep{Getman2018}. Therefore, apart from kinematic coherence examination carryied by \citet{Wang2016}, we also did checks on the 2D (Galactic longitude and latitude) MSTs. The clumps \citet{Wang2016} employ are from Bolocam Galactric Plan Survey (BGPS) \citep{Rosolowsky2010}, which only cover half of the Galactic plane ($7.5^\circ\leq l\leq 195^\circ$). Now, ATLASGAL gives better estimation of distance, temperature and velocity towards a complete sample of Galactic clumps \citep{Urquhart2018}. High-resolution column density map derived from the Herschel Hi-GAL survey with PPMAP \citep{Marsh2017} also makes it possible to get a more accurate estimation of filament mass in the entire Galactic plane.\\
Physical properties of large-scale filaments associated with spiral structure have been inspected recently \citep[e.g.][]{Ragan2014,Abreu2016,Wang2015,Wang2016,Zucker2018}. However, most of the studies lack a large sample size, or use spiral arm models that do not have high-accuracy distance estimation. Now we are able to obtain a large sample of large-scale filaments and an updated spiral arm model fitted from trigonometric parallaxes of high-mass star-forming regions \citep{Reid2019} which gives us better assistance to study the relation between filaments and spiral arms. With the advance of the radial velocity measurements of ATLASGAL clumps \citep{Urquhart2014} and the updated spiral arm model \citep{Reid2019}, we extend the work of \citet{Wang2016} in the northern Galactic plane to the entire (except Galactic center) inner Galactic plane covered by ATLASGAL.
The paper is structured as follows. We describe the data and methods we use in Sect. \ref{sec:method}. Section \ref{sec:results} describes our results containing physical properties of our large-scale filaments and some statistics as well as their dense gas mass fraction. Then in Sect. \ref{sec:discussion}, we examine the robustness of MST method and investigate the Galactic distribution of the filaments. Fragmentation of large-scale filaments is also inspected. And finally, our conclusions are summarized in Sect. \ref{sec:conclusion}.
\section{Data and Method} \label{sec:method}
\subsection{ATLASGAL Galactic Clumps}
The APEX Telescope Large Area Survey of the Galaxy (ATLASGAL) is an unbiased 870 \textmu m submillimeter survey of the inner Galactic plane ($|l|<60^\circ$ with $|b|<1.5^\circ$). They identified 10163 compact sources with a resolution of $19^{\prime\prime}$ and generated a comprehensive and unbiased catalogue of massive, dense clumps, which are located across the inner Galaxy \citep{Urquhart2014}. Among the total 10163 sources, $\sim$ 8000 dense clumps located away from the Galactic center region ($300^\circ<l<355^\circ$ and $5^\circ<l<60^\circ$) were studied on the physical properties, such as velocity, by \citet{Urquhart2018}. The velocity information of clumps makes it possible to investigate their coherence in position-position-velocity (PPV) space.
\subsection{Herschel Column Density Map}\label{sec:Herschel}
The Herschel satellite \citep{Pilbratt2010} provides a powerful tool to study the structure of high-mass molecular clouds on different scales \citep[e.g.][]{Beuther2010, Schneider2012}. Hi-GAL, the Herschel infrared Galactic Plane Survey, makes an unbiased photometric survey of the inner Galactic plane by mapping a $2^\circ$ wide strip in the entire Galactic Plane in five wavebands between 70 \textmu m and 500 \textmu m \citep{Molinari2010}. High resolution temperature-differential column density maps with an angular resolution of 12 arcsec have been derived from Herschel Hi-GAL data using PPMAP tool, which represents improvement over those obtained with standard analysis techniques. \citep{Marsh2017}. We employ the temperature-integrated column density maps for the following mass calculation of our filaments.\\
\subsection{Filament Identification}\label{sec:ident}
\begin{figure*}
\begin{center}
\includegraphics[scale=0.3,angle=0]{figure1.jpg}
\caption{Two-color view of filaments. The color-coded circles denote clumps in filaments with various velocities. For backgrounds, cyan represents intermediate infrared 24 \textmu m emission on logarithmic scale from MIPSGAL \citep{Carey2009} and red shows submillimeter 870 \textmu m emission on linear scale from APEX + Planck combined image \citep{Csengeri2016}. Other two-color view of filaments and their column density maps are displayed in Appendix \ref{sec:maps}}
\label{twocolor}
\end{center}
\end{figure*}
The filaments are identified from ATLASGAL \citep{Urquhart2018} clumps usin minimum spanning tree (MST) method firstly used by \citet{Wang2016}\footnote{Code is available in \citet{Wang2021}}. The MST algorithm is adopted to isolate filaments in PPV sample, meaning clumps cluster as a filament only when they are close to each other with similar velocities. The clumps (circles in Fig. \ref{Mcal} (a) and filled circles in Fig. \ref{twocolor}) are connected with the cost of a minimum sum of edge lengths, where ``edges'' (rectangles in Fig. \ref{Mcal} (a) and white lines in Fig. \ref{twocolor}) mean straight lines linking clumps, and edge lengths are separation between each two clumps. The criteria for MST matching and filament selection follow \citet{Wang2016} :
\begin{itemize}
\item[(1)]The accepted MST must contain at least five ATLASGAL clumps: $N_{cl} \geq 5$.
\item[(2)]Only edges shorter than a maximum length (cut-off length) can be connected $(\Delta L < 0.1^\circ )$.
\item[(3)]For any two clumps to be connected, the difference in line-of-sight velocity (matching velocity $\Delta v$) must be less than 2 km s$^{-1}$.
\item[(4)]Linearity $f_L > 1.5$. Here linearity is defined to quantify the degree of similarity between the target with a linear shape.
\item[(5)]Projected length $L_{sum}\geq 10$ pc. We only focus on large-scale filaments in this work.
\end{itemize}
Initial values of cut-off length $\Delta L$ in (2), and matching velocity $\Delta v$ in (3), are chosen based on characteristics of previously known filaments. A variety of values around initial ones are tested and the best combination is chosen to identify known filaments as many as possible but not connect unrelated sources, see Appendix \ref{sec:parameter} for details. $N_{cl}$ in (1), minimum number of clumps refers to ``pruning'' level of MSTs in \citet{ParkLee2009} and \citet{Pereyra2020}. In their work, when a branch of an MST has less than 5 nodes, it is thought as a minor branch and should be removed from the tree. Linearity in (4) is the ratio between spread (standard deviation) of clumps along major axis and spread along minor axis. To define the major axis of a filament, we plot all the clumps in the projected sky (Galactic longitude as $x$ and Galactic latitude as $y$) and fitted a line as the major axis of this filament. But instead of ordinary least squares, we get this line with the help of principle component analysis (PCA), which is detailed in Appendix \ref{sec:PCA}. The minor axis is perpendicular to the major axis and pass through the mean position $(\overline{x},\overline{y})$. The larger $f_L$ means the shape is more likely to be linear. For a straight line, $f_L\rightarrow \infty $. The distances of clumps in a filament are thought to be the same. If the distances of clumps in one filament are different, a unified distance will be used (detailed in Sect. \ref{sec:properties}). The identified filaments are shown in Fig. \ref{twocolor} and Sect. \ref{sec:results}.\\
\begin{figure*}
\gridline{\fig{figure2a.png}{0.485\textwidth}{(a)}
\fig{figure2b.png}{0.5\textwidth}{(b)}
}
\gridline{\fig{figure2c.png}{0.5\textwidth}{(c)}
\fig{figure2d.png}{0.5\textwidth}{(d)}
}
\caption{Exemplification of how to define total mass and dense gas mass in a filament. In panel (a), background is Herschel column density map derived from PPMAP \citep{Marsh2017} in logarithmic scale. Circles are located in the same position with clumps in this filament and the diameters are four times the major axes of clumps. Rectangles are guided by ``edges'' of a filament. The length of a rectangle is the separation between two clumps in both ends and width is equal to the smaller diameter of the two circled in both ends. The boundary of a filament is a combination of these circles and rectangles shown in panel (c). And the mass of a filament is obtained by integrating column density within the boundary. Dense gas mass of a filament is the sum of mass within blue circles in panel (b). Blue circles are also located in the position of clumps. The diameters of circles are twice the major axes of clumps. Yellow masks in panel (d) show the boundary of dense gas and the green contour enclosing them is the sketch of the filament boundary.}
\label{Mcal}
\end{figure*}
\subsection{Mass Calculation}\label{sec:mass}
Mass of the filaments is calculated by summing mass within the boundary of each filament in the Hi-GAL based column density maps. The boundary of a filament is a combination of a series of circles and rectangles as shown in Fig. \ref{Mcal}. A circle represents the position of a clump with a diameter of four times the clump's major axis. The diameter is chosen referring to the aperture radius ATLASGAL used to measure clump flux density from Herschel, which is twice the major axis \citep{Urquhart2018}. And our mass is derived from Herschel column density map. This diameter is large enough that most of the source emission is within the boundary, while still being small enough to avoid including the background. Rectangles are consist of edges linking each two clumps in a filament. The length of a rectangle is the length of the edge (that is, separation between two clumps in both ends) and the width is the same as the smaller circle diameter of two clumps in both ends. Then the mass of filament is obtained by summing contributions from the pixels of the column-density map in the boundary. Owing to gaps in coverage on column density map, 29 filaments are affected (filaments with DGMF of ``-'' in Table \ref{t1}). For a filament with incomplete column density map, we can only obtain a total clump mass, which is the sum of ATLASGAL clump mass in this filament. So for these filaments, instead of integrating column density within the boundary, we firstly calculate the mean ratio of total clump mass to filament mass obtained from column density map of other 134 filaments with complete column density map. Then we scale total clump mass for these filaments by the mean ratio.\\
For dense gas mass calculation, several approaches have been employed in previous studies. They can be classified into two categories, one is calculating mass from a denser gas tracer within the same boundary as that of the whole filament mass \citep[e.g.][]{Ragan2014}. The other is to take contour with a higher level in the extinction map as the boundary to calculate dense gas, and contour with lower level as boundary of total mass \citep[e.g.][]{ZhangM2019}. To avoid systematic bias, we adopt the latter one. That is, our dense gas mass is also acquired from Herschel Hi-GAL column density map. Dense gas mass of a filament is derived from the sum of clump mass in this filament derived by an integral of column density in circle regions (shown as blue circles in Fig. \ref{Mcal} (b)) with diameter twice the major axis of clumps. We also tried to take clump mass directly from ATLASGAL catalogue \citep{Urquhart2018}, where mass is calculated with integrated 870 \textmu m flux density as dense gas. We compared clump mass from ATLASGAL and that from Herschel column density map with the same radius. And we find some clump masses from ATLASGAL are dramatically larger than that from Herschel column density map. Considering the consistency of calculation method between the dense mass and the total mass, we did not use mass from ATLASGAL catalogue. Another reason is that, column density map obtained with PPMAP considered temperature variance pixel-by-pixel while ATLASGAL takes a single temperature from spectral grey body fitting for each clump to calculate mass.\\
Unlike employing a fixed extinction or column density contour to define boundaries of filaments or dense gas region, our boundaries vary in different Galactic longitudes. As is known to us, column density in the inner Galactic plane is not the same in various Galactic longitudes, being larger in the center and decreasing outwards. A fixed extinction or column density contour, on one hand, in Galactic longitude where column density is high, might put extra mass that should not be in a filament in that filament. On the other hand, in less dense Galactic longitudes, the contour may contain nothing. Our polygon boundaries guided by filaments avoid this problem.
\section{Results} \label{sec:results}
We identify 163 large-scale filaments in the inner Galactic plane ($|l|<60^\circ$ with $|b|<1.5^\circ$) from ATLASGAL clumps and the filaments are shown in Fig. \ref{twocolor}. The color-coded circles in Fig. \ref{twocolor} denote clumps with different velocities. For backgrounds, cyan represents intermediate infrared 24 \textmu m emission on logarithmic scale from MIPSGAL \citep{Carey2009} and red shows submillimeter 870 \textmu m emission on linear scale from APEX + Planck combined image \citep{Csengeri2016}. Other two-color view of filaments and their column density maps are displayed in Appendix \ref{sec:maps}.
\subsection{Basic Physical Properties} \label{sec:properties}
Physical properties of the 163 filaments are shown in Table \ref{t1}. The meaning of columns are listed in Table \ref{t1_explain}. The determination of ATLASGAL clump distances can be roughly divided into four steps \citep{Urquhart2018}. Firstly, if maser parallax and spectroscopic distances are available, literature solution is adopted. Secondly, kinematic distances are determined for clumps with radial velocity. Thirdly, clumps located within the Solar circle face distance ambiguities, which are resolved using archival HI data. Fourthly, they group the clumps to assign or correct their distances. Typical uncertainty for kinematic distances of ATLASGAL clumps is 0.3 kpc. As mentioned at the end of Sect. \ref{sec:ident}, we consider clumps in a velocity-coherent filament have the same distance. This could be regarded as a correction for distance measurement of ATLASGAL clumps. Physical properties of clumps that we directly take from ATLASGAL catalogue relating with distance are radius and Galactocentric radius. Both are linear correlated with distance. So we could correct them by simply multiplying a factor $d_{fl}/d_{cl}$, where $d_{cl}$ is clump distance directly from ATLASGAL catalogue and $d_{fl}$ is the distance of the filament. In most of our filaments (95\%), clumps have the same distances. One probable reason for this is that most of velocity coherent structures found by MST are physically real structures, so clumps in the same structure naturally have the same distances. A small portion of MSTs may be a collection of clumps without physical connection. Another probable reason is that ATLASGAL applied friends-of-friends method to group the clumps \citep{Urquhart2018}. For some of clumps ($\sim10\%$) without well defined distances, distance of the cluster they belong is assigned to them. And for some well studied complexes with reliable distances, their associated clumps adopted the reliable distances.\\
We include standard deviations of Galactic longitude, Galactic latitude, velocity, velocity gradient, and height above Galactic mid-plane in Table \ref{t1}. For a fractal turbulent cloud with mass of 1000 M$_\odot$, the standard deviation of mass derived from PPMAP is 612 M$_\odot$ \citep{Marsh2015}. The uncertainties for $|\theta|$ are derived from bootstrap method. That is, we add Gaussian noise to Galactic longitude and latitude of clumps, and then get one $|\theta|$ with PCA. We repeat this process for 1000 times and take standard deviation of the 1000 values of $|\theta|$ as the uncertainty of $|\theta|$. \\
The statistics of the parameters are shown in the last few rows of Table \ref{t1}, including minimum, maximum, median, mean, standard deviation, skewness (S), and kurtosis (K). Skewness measures the asymmetry of probability distribution. For instance, normal distribution has $S=0$. Negative skewness means that the tail is on the left side of the distribution, and positive skewness indicates that the tail is on the right. Kurtosis characterizes how the distribution is compared to a normal distribution. The normal distribution has $K=0$. A distribution with $K>0$ is more centrally peaked than normal distribution while for $K<0$ is flatter.
{\setlength{\tabcolsep}{2pt}
\begin{longrotatetable}
\movetabledown=0.5in
\begin{deluxetable}{cccccccccccccccccccccccc}
\tablecaption{\label{t1}}
\tablecolumns{24}
\tabletypesize{\scriptsize}
\tablehead{
\colhead{(1)} & \colhead{(2)} & \colhead{(3)} & \colhead{(4)} & \colhead{(5)} & \colhead{(6)} & \colhead{(7)} & \colhead{(8)} & \colhead{(9)} & \colhead{(10)} & \colhead{(11)} & \colhead{(12)} & \colhead{(13)} & \colhead{(14)} & \colhead{(15)} & \colhead{(16)} & \colhead{(17)} & \colhead{(18)} & \colhead{(19)} & \colhead{(20)} & \colhead{(21)} & \colhead{(22)} & \colhead{(23)} & \colhead{(24)}\\
\colhead{ID} & \colhead{$l_{wt}$} & \colhead{$b_{wt}$} & \colhead{$v_{wt}$} & \colhead{$d$} & \colhead{N$_{cl}$} & \colhead{$L_{sum}$} & \colhead{$L_{end}$} & \colhead{$v_{grad}$} & \colhead{$T_{min}$} & \colhead{$T_{max}$} & \colhead{Mass} & \colhead{$M_{line}$} & \colhead{$N_{H_2}$} & \colhead{$n_{H_2}$} & \colhead{$\frac{L_{sum}}{w}$} & \colhead{$f_A$} & \colhead{$f_L$} & \colhead{$L_{wt}$} & \colhead{$R_{gc}$} & \colhead{$z$} & \colhead{$|\theta|$} & \colhead{Mor.} & \colhead{DGMF}
}
\startdata
1 & 5.37$\pm$0.02 & 0.13$\pm$0.06 & 11.0$\pm$1.1 & 2.9 & 6 & 12.8 & 9.3 & 0.41$\pm$0.06 & 10.8 & 18.0 & 4.6 & 364.1 & 1.8 & 5.8 & 10.0 & 7.03 & 3.02 & 11.1 & 5.4 & 25.1$\pm$3.3 & 82.9$\pm$2.7 & L,C & 0.20 \\
2 & 6.20$\pm$0.06 & -0.12$\pm$0.01 & 11.3$\pm$2.3 & 3.0 & 6 & 11.2 & 8.6 & 0.91$\pm$0.28 & 16.1 & 26.0 & 5.8 & 521.4 & 2.3 & 6.6 & 7.8 & 7.06 & 5.73 & 9.1 & 5.4 & 12.0$\pm$0.7 & 8.0$\pm$2.5 & X & 0.32 \\
3 & 6.16$\pm$0.05 & -0.61$\pm$0.02 & 17.4$\pm$1.2 & 3.0 & 8 & 11.8 & 7.2 & 1.02$\pm$0.10 & 14.6 & 25.0 & 9.7 & 824.8 & 2.2 & 3.7 & 5.0 & 4.43 & 3.11 & 8.5 & 5.4 & -13.7$\pm$1.3 & 18.8$\pm$2.7 & S & 0.38 \\
4 & 8.58$\pm$0.17 & -0.33$\pm$0.05 & 37.2$\pm$1.2 & 4.4 & 38 & 91.1 & 51.2 & 0.41$\pm$0.02 & 9.9 & 37.6 & 135.9 & 1491.3 & 3.6 & 5.6 & 34.9 & 8.07 & 3.96 & 57.3 & 4.0 & -10.2$\pm$3.7 & 7.6$\pm$0.4 & S,H & - \\
5 & 8.16$\pm$0.07 & 0.22$\pm$0.03 & 18.6$\pm$0.6 & 3.0 & 10 & 13.6 & 10.2 & 0.54$\pm$0.02 & 11.2 & 29.5 & 11.1 & 820.0 & 3.1 & 7.9 & 8.3 & 10.15 & 5.27 & 11.0 & 5.4 & 29.8$\pm$1.6 & 23.4$\pm$1.8 & L,H & 0.46 \\
6 & 8.51$\pm$0.04 & -0.64$\pm$0.04 & 17.0$\pm$0.6 & 3.0 & 9 & 12.8 & 8.4 & 0.49$\pm$0.02 & 11.9 & 26.8 & 2.6 & 200.4 & 1.1 & 4.3 & 11.7 & 5.21 & 1.86 & 12.3 & 5.4 & -15.1$\pm$2.0 & 41.2$\pm$3.6 & X & 0.26 \\
7 & 10.22$\pm$0.11 & -0.27$\pm$0.10 & 11.7$\pm$1.3 & 3.5 & 48 & 73.8 & 27.5 & 0.62$\pm$0.10 & 11.3 & 47.7 & 102.3 & 1386.5 & 3.0 & 4.2 & 25.3 & 6.61 & 2.42 & 35.7 & 4.9 & 0.9$\pm$6.1 & 41.2$\pm$0.5 & S,X & - \\
8 & 11.09$\pm$0.10 & -0.10$\pm$0.03 & 29.7$\pm$1.4 & 2.9 & 12 & 25.0 & 16.9 & 0.65$\pm$0.04 & 10.5 & 15.8 & 20.0 & 801.2 & 2.6 & 5.7 & 13.2 & 5.99 & 3.84 & 19.1 & 5.5 & 13.8$\pm$1.5 & 6.7$\pm$1.1 & S & 0.32 \\
9 & 12.49$\pm$0.04 & -0.21$\pm$0.09 & 34.8$\pm$0.9 & 2.6 & 7 & 16.8 & 14.1 & 0.24$\pm$0.01 & 10.0 & 18.5 & 8.3 & 496.1 & 2.2 & 6.7 & 12.2 & 9.75 & 5.04 & 15.2 & 5.9 & 9.9$\pm$4.1 & 65.3$\pm$1.5 & S & 0.20 \\
10 & 12.87$\pm$0.16 & -0.21$\pm$0.08 & 35.5$\pm$1.5 & 2.6 & 41 & 62.7 & 26.8 & 0.86$\pm$0.05 & 9.6 & 26.6 & 86.6 & 1382.3 & 4.5 & 9.8 & 32.8 & 6.21 & 2.36 & 35.2 & 5.9 & 9.9$\pm$3.8 & 17.8$\pm$0.5 & C,H & - \\
11 & 13.20$\pm$0.06 & 0.04$\pm$0.02 & 50.7$\pm$1.5 & 2.6 & 10 & 15.9 & 9.9 & 1.29$\pm$0.05 & 10.5 & 26.8 & 14.7 & 923.7 & 3.8 & 10.2 & 10.4 & 7.42 & 3.66 & 11.2 & 5.9 & 21.5$\pm$1.0 & 14.9$\pm$2.2 & S,H & 0.38 \\
12 & 12.69$\pm$0.02 & -0.16$\pm$0.03 & 54.9$\pm$1.4 & 4.8 & 6 & 12.9 & 8.1 & 1.21$\pm$0.04 & 13.7 & 26.9 & 36.2 & 2803.2 & 5.6 & 7.5 & 4.1 & 11.25 & 1.58 & 13.0 & 3.8 & 1.4$\pm$2.4 & 63.0$\pm$9.6 & X,H & 0.35 \\
13 & 13.21$\pm$0.10 & -0.34$\pm$0.03 & 38.9$\pm$1.2 & 2.6 & 14 & 23.9 & 12.7 & 0.91$\pm$0.04 & 9.3 & 26.0 & 15.3 & 637.5 & 3.1 & 9.9 & 18.6 & 2.81 & 3.29 & 14.9 & 5.9 & 4.0$\pm$1.6 & 8.0$\pm$1.0 & X & 0.26 \\
14 & 14.00$\pm$0.01 & -0.14$\pm$0.03 & 40.1$\pm$1.4 & 3.1 & 9 & 11.9 & 5.3 & 1.07$\pm$0.05 & 16.2 & 33.7 & 7.6 & 639.5 & 2.4 & 5.9 & 7.1 & 7.49 & 2.64 & 6.6 & 5.4 & 10.7$\pm$1.6 & 76.5$\pm$5.0 & C & 0.47 \\
15 & 14.36$\pm$0.11 & -0.14$\pm$0.05 & 38.8$\pm$1.7 & 3.1 & 30 & 51.2 & 19.4 & 0.76$\pm$0.05 & 10.1 & 29.2 & 42.5 & 830.3 & 2.8 & 6.1 & 27.3 & 4.63 & 2.69 & 24.2 & 5.4 & 10.8$\pm$2.9 & 16.7$\pm$0.7 & X & 0.43 \\
16 & 14.72$\pm$0.06 & -0.18$\pm$0.04 & 38.6$\pm$1.3 & 3.1 & 11 & 18.7 & 11.4 & 0.80$\pm$0.04 & 8.5 & 27.5 & 19.7 & 1058.4 & 3.3 & 6.7 & 9.2 & 4.11 & 2.29 & 15.1 & 5.4 & 8.5$\pm$2.0 & 26.0$\pm$2.3 & S,X & 0.33 \\
17 & 14.32$\pm$0.17 & -0.58$\pm$0.08 & 20.2$\pm$1.3 & 1.5 & 40 & 41.2 & 14.1 & 0.86$\pm$0.08 & 8.7 & 27.9 & 18.4 & 448.0 & 3.0 & 13.1 & 43.9 & 2.96 & 2.49 & 18.1 & 6.9 & 6.6$\pm$2.0 & 12.8$\pm$0.4 & L,X & - \\
18 & 15.05$\pm$0.09 & -0.66$\pm$0.07 & 19.8$\pm$1.6 & 2.0 & 38 & 40.3 & 13.5 & 1.04$\pm$0.12 & 16.5 & 53.0 & 72.4 & 1796.3 & 4.9 & 8.8 & 17.6 & 3.82 & 1.67 & 21.1 & 6.5 & -2.1$\pm$2.5 & 29.8$\pm$1.0 & H & - \\
19 & 15.65$\pm$0.03 & -0.22$\pm$0.05 & 57.0$\pm$0.2 & 11.6 & 6 & 41.2 & 39.0 & 0.07$\pm$0.01 & 16.6 & 20.9 & 44.8 & 1088.6 & 2.3 & 3.3 & 14.1 & 17.49 & 7.44 & 40.4 & 4.2 & -44.2$\pm$11.0 & 58.3$\pm$2.5 & C & 0.19 \\
20 & 13.87$\pm$0.05 & 0.25$\pm$0.03 & 48.1$\pm$1.2 & 3.9 & 6 & 19.3 & 9.8 & 0.15$\pm$0.01 & 15.6 & 33.8 & 10.9 & 565.4 & 2.1 & 4.9 & 11.2 & 2.98 & 1.51 & 16.2 & 4.6 & 34.1$\pm$2.3 & 21.2$\pm$6.5 & C,H & 0.28 \\
21 & 9.87$\pm$0.07 & -0.74$\pm$0.02 & 27.4$\pm$2.2 & 3.1 & 7 & 13.0 & 12.5 & 0.46$\pm$0.05 & 13.2 & 27.4 & 8.4 & 644.3 & 2.1 & 4.6 & 6.9 & 20.40 & 9.48 & 12.8 & 5.3 & -21.8$\pm$1.0 & 12.9$\pm$1.9 & L & - \\
22 & 16.38$\pm$0.06 & -0.60$\pm$0.06 & 40.1$\pm$2.2 & 3.3 & 17 & 26.6 & 13.1 & 0.61$\pm$0.05 & 10.7 & 16.5 & 17.3 & 651.2 & 2.8 & 7.8 & 18.1 & 4.13 & 1.97 & 18.7 & 5.2 & -16.6$\pm$3.4 & 40.5$\pm$1.6 & C,X & 0.34 \\
23 & 16.39$\pm$0.04 & -0.13$\pm$0.05 & 45.9$\pm$1.0 & 3.5 & 6 & 12.6 & 10.6 & 0.57$\pm$0.02 & 15.9 & 26.8 & 8.8 & 697.9 & 3.3 & 10.3 & 9.5 & 5.57 & 4.89 & 11.5 & 5.1 & 9.8$\pm$3.1 & 49.8$\pm$2.4 & S & 0.21 \\
24 & 18.28$\pm$0.06 & -0.26$\pm$0.02 & 68.4$\pm$1.0 & 4.9 & 6 & 15.5 & 14.9 & 0.58$\pm$0.01 & 11.1 & 19.8 & 35.2 & 2266.1 & 3.0 & 2.7 & 3.3 & 26.81 & 11.36 & 15.2 & 4.0 & -6.8$\pm$1.5 & 16.2$\pm$2.6 & L & 0.41 \\
25 & 18.61$\pm$0.05 & -0.09$\pm$0.05 & 45.3$\pm$0.6 & 3.4 & 6 & 13.7 & 12.6 & 0.27$\pm$0.01 & 12.0 & 23.7 & 13.6 & 990.5 & 3.2 & 6.8 & 7.1 & 9.72 & 5.45 & 13.5 & 5.3 & 13.0$\pm$3.1 & 44.0$\pm$2.2 & L,H & 0.29 \\
26 & 19.00$\pm$0.01 & -0.05$\pm$0.03 & 59.8$\pm$2.6 & 5.0 & 7 & 11.4 & 7.2 & 0.84$\pm$0.05 & 15.1 & 28.0 & 24.9 & 2185.5 & 4.0 & 4.9 & 3.3 & 3.39 & 3.16 & 8.6 & 4.0 & 10.9$\pm$2.6 & 76.8$\pm$5.5 & S & 0.39 \\
27 & 18.92$\pm$0.10 & -0.41$\pm$0.12 & 64.8$\pm$2.6 & 5.0 & 28 & 93.4 & 38.1 & 0.42$\pm$0.02 & 11.9 & 31.3 & 94.9 & 1015.8 & 1.9 & 2.3 & 27.8 & 8.25 & 2.83 & 46.6 & 4.0 & -20.8$\pm$10.6 & 51.2$\pm$0.6 & C & 0.45 \\
28 & 19.62$\pm$0.06 & -0.11$\pm$0.03 & 58.9$\pm$2.1 & 5.1 & 12 & 33.8 & 17.3 & 0.58$\pm$0.03 & 13.8 & 31.6 & 36.1 & 1069.5 & 2.9 & 5.2 & 14.6 & 4.49 & 2.18 & 23.4 & 3.9 & 5.0$\pm$2.9 & 20.0$\pm$2.5 & C & 0.27 \\
29 & 20.74$\pm$0.04 & -0.08$\pm$0.05 & 57.1$\pm$1.6 & 11.7 & 16 & 79.3 & 33.6 & 0.33$\pm$0.01 & 16.5 & 29.5 & 170.2 & 2146.9 & 2.0 & 1.3 & 12.0 & 2.09 & 2.04 & 47.0 & 4.9 & -14.3$\pm$10.3 & 57.0$\pm$2.2 & C & 0.43 \\
30 & 22.40$\pm$0.04 & 0.35$\pm$0.04 & 84.0$\pm$0.3 & 5.4 & 7 & 20.0 & 12.9 & 0.08$\pm$0.01 & 14.7 & 23.0 & 41.4 & 2073.4 & 3.1 & 3.0 & 4.7 & 4.57 & 1.57 & 21.0 & 3.9 & 47.2$\pm$3.4 & 37.7$\pm$5.3 & L,X & 0.39 \\
31 & 22.57$\pm$0.05 & -0.20$\pm$0.01 & 76.6$\pm$0.8 & 4.2 & 6 & 12.6 & 11.6 & 0.49$\pm$0.01 & 11.6 & 17.1 & 12.4 & 983.2 & 4.9 & 16.5 & 10.2 & 13.38 & 7.00 & 12.1 & 4.8 & 2.2$\pm$0.8 & 9.0$\pm$2.9 & S & 0.23 \\
32 & 22.94$\pm$0.01 & -0.28$\pm$0.09 & 62.2$\pm$1.8 & 4.2 & 8 & 22.7 & 17.6 & 0.62$\pm$0.02 & 13.3 & 25.0 & 19.5 & 861.5 & 2.5 & 4.7 & 10.5 & 9.78 & 10.09 & 17.9 & 4.8 & -3.5$\pm$7.0 & 87.6$\pm$1.3 & L & 0.27 \\
33 & 23.08$\pm$0.08 & -0.40$\pm$0.02 & 77.4$\pm$0.9 & 4.6 & 6 & 25.0 & 21.3 & 0.24$\pm$0.01 & 13.4 & 20.9 & 50.2 & 2009.7 & 3.0 & 2.9 & 5.9 & 13.65 & 5.54 & 22.6 & 4.5 & -15.9$\pm$2.0 & 12.8$\pm$1.8 & C & 0.31 \\
34 & 22.75$\pm$0.10 & -0.46$\pm$0.03 & 76.4$\pm$1.1 & 4.6 & 13 & 39.7 & 29.0 & 0.42$\pm$0.01 & 11.5 & 19.8 & 33.6 & 845.2 & 2.9 & 6.6 & 21.9 & 12.41 & 4.66 & 31.6 & 4.5 & -20.6$\pm$2.6 & 12.8$\pm$1.1 & S & 0.21 \\
35 & 23.40$\pm$0.05 & -0.21$\pm$0.07 & 101.5$\pm$2.7 & 5.9 & 31 & 77.7 & 30.3 & 0.47$\pm$0.02 & 12.0 & 44.1 & 322.9 & 4156.3 & 5.5 & 4.9 & 16.6 & 5.63 & 1.63 & 48.0 & 3.8 & -7.8$\pm$7.4 & 65.4$\pm$1.6 & X & 0.45 \\
36 & 23.28$\pm$0.04 & -0.27$\pm$0.04 & 61.0$\pm$1.6 & 4.6 & 6 & 16.8 & 12.3 & 0.90$\pm$0.03 & 11.9 & 26.2 & 22.4 & 1336.4 & 3.9 & 7.4 & 7.8 & 9.10 & 4.29 & 13.6 & 4.5 & -6.0$\pm$3.0 & 43.0$\pm$3.0 & C & 0.25 \\
37 & 23.22$\pm$0.03 & 0.03$\pm$0.04 & 76.0$\pm$2.1 & 5.9 & 7 & 18.2 & 11.9 & 0.77$\pm$0.03 & 13.0 & 25.6 & 29.2 & 1603.6 & 4.0 & 6.6 & 7.3 & 3.68 & 1.97 & 16.9 & 3.8 & 16.3$\pm$3.7 & 57.1$\pm$5.0 & C & 0.33 \\
38 & 23.97$\pm$0.03 & 0.15$\pm$0.02 & 79.6$\pm$1.9 & 4.8 & 10 & 18.5 & 8.7 & 0.62$\pm$0.02 & 13.5 & 38.7 & 38.4 & 2079.4 & 3.2 & 3.2 & 4.5 & 8.11 & 2.13 & 12.0 & 4.4 & 28.4$\pm$1.6 & 31.2$\pm$5.3 & L,H & 0.38 \\
39 & 24.04$\pm$0.03 & 0.18$\pm$0.07 & 109.8$\pm$4.0 & 7.8 & 11 & 45.4 & 31.7 & 0.97$\pm$0.05 & 10.9 & 27.8 & 101.1 & 2226.4 & 3.2 & 3.0 & 10.4 & 6.75 & 2.80 & 39.0 & 3.4 & 34.0$\pm$9.1 & 74.2$\pm$1.9 & L,X & 0.27 \\
40 & 23.92$\pm$0.18 & 0.51$\pm$0.06 & 95.4$\pm$1.1 & 5.8 & 22 & 84.1 & 66.7 & 0.20$\pm$0.01 & 10.3 & 24.0 & 82.4 & 979.2 & 2.1 & 2.9 & 28.3 & 15.88 & 10.22 & 68.0 & 3.8 & 65.0$\pm$5.9 & 17.2$\pm$0.4 & L & 0.27 \\
41 & 24.79$\pm$0.04 & 0.09$\pm$0.05 & 109.2$\pm$1.5 & 6.0 & 16 & 47.0 & 22.0 & 0.39$\pm$0.01 & 16.3 & 34.3 & 62.1 & 1320.5 & 2.1 & 2.2 & 11.9 & 3.27 & 2.03 & 30.9 & 3.8 & 22.7$\pm$5.3 & 53.2$\pm$2.2 & X & - \\
42 & 24.48$\pm$0.05 & -0.52$\pm$0.03 & 60.2$\pm$1.2 & 11.3 & 7 & 55.5 & 30.5 & 0.08$\pm$0.01 & 10.9 & 17.6 & 76.8 & 1382.1 & 2.3 & 2.5 & 14.7 & 4.30 & 2.31 & 40.3 & 5.1 & -99.0$\pm$5.0 & 14.6$\pm$3.7 & X & 0.19 \\
43 & 25.77$\pm$0.04 & 0.24$\pm$0.03 & 109.7$\pm$1.4 & 8.7 & 9 & 44.2 & 23.1 & 0.22$\pm$0.01 & 19.8 & 31.2 & 79.3 & 1794.8 & 1.8 & 1.2 & 7.1 & 7.54 & 3.22 & 27.2 & 3.8 & 45.0$\pm$4.3 & 34.2$\pm$3.1 & C & 0.43 \\
44 & 25.72$\pm$0.08 & -0.16$\pm$0.02 & 92.6$\pm$1.5 & 10.2 & 12 & 74.8 & 50.4 & 0.22$\pm$0.01 & 10.3 & 30.0 & 162.8 & 2176.6 & 3.2 & 3.0 & 17.3 & 8.49 & 4.47 & 55.2 & 4.5 & -21.9$\pm$4.1 & 9.0$\pm$1.4 & S & 0.29 \\
45 & 25.36$\pm$0.06 & -0.38$\pm$0.05 & 57.1$\pm$1.6 & 2.7 & 9 & 15.1 & 9.7 & 0.94$\pm$0.04 & 16.4 & 22.3 & 6.8 & 451.8 & 2.3 & 8.1 & 12.5 & 4.16 & 2.26 & 13.0 & 6.0 & 1.9$\pm$2.2 & 32.6$\pm$2.2 & L,X & 0.26 \\
46 & 28.18$\pm$0.05 & -0.05$\pm$0.03 & 97.1$\pm$1.1 & 6.1 & 7 & 26.1 & 17.1 & 0.29$\pm$0.01 & 15.3 & 29.9 & 29.7 & 1138.7 & 2.2 & 2.8 & 8.0 & 7.14 & 2.64 & 21.4 & 4.2 & 9.0$\pm$3.3 & 30.0$\pm$3.2 & S,H & - \\
47 & 28.24$\pm$0.02 & 0.04$\pm$0.02 & 107.1$\pm$1.6 & 6.1 & 9 & 18.0 & 8.5 & 0.64$\pm$0.01 & 15.5 & 29.9 & 61.3 & 3404.2 & 5.7 & 6.2 & 4.8 & 3.09 & 2.03 & 12.0 & 4.2 & 17.9$\pm$2.6 & 64.3$\pm$6.8 & S,H & 0.41 \\
48 & 28.66$\pm$0.05 & 0.03$\pm$0.03 & 98.3$\pm$2.5 & 7.4 & 6 & 30.3 & 22.4 & 0.76$\pm$0.03 & 18.3 & 27.0 & 67.9 & 2244.3 & 2.9 & 2.5 & 6.2 & 12.38 & 3.70 & 25.4 & 4.0 & 15.9$\pm$3.6 & 29.4$\pm$3.4 & C & 0.27 \\
49 & 28.79$\pm$0.02 & 0.20$\pm$0.03 & 105.3$\pm$1.4 & 7.4 & 7 & 21.7 & 13.6 & 0.72$\pm$0.01 & 18.5 & 26.5 & 41.3 & 1905.2 & 2.6 & 2.3 & 4.7 & 4.46 & 2.82 & 16.7 & 4.0 & 36.9$\pm$4.3 & 66.4$\pm$5.0 & S,H & 0.43 \\
50 & 28.57$\pm$0.03 & -0.28$\pm$0.05 & 88.0$\pm$1.8 & 4.7 & 8 & 17.3 & 12.5 & 0.43$\pm$0.01 & 10.3 & 29.7 & 40.6 & 2352.5 & 4.2 & 4.9 & 4.9 & 11.08 & 4.49 & 13.7 & 4.8 & -6.5$\pm$4.4 & 65.2$\pm$2.4 & L,C & 0.42 \\
51 & 26.55$\pm$0.03 & -0.29$\pm$0.04 & 107.7$\pm$0.8 & 7.6 & 8 & 28.5 & 16.0 & 0.31$\pm$0.01 & 16.7 & 26.8 & 25.3 & 889.1 & 2.0 & 3.1 & 10.4 & 4.35 & 1.79 & 23.9 & 3.7 & -27.3$\pm$5.1 & 54.7$\pm$4.5 & C & 0.40 \\
52 & 30.03$\pm$0.05 & 0.10$\pm$0.03 & 105.6$\pm$0.9 & 5.2 & 8 & 25.2 & 15.5 & 0.53$\pm$0.01 & 12.9 & 31.9 & 23.3 & 925.3 & 3.2 & 7.4 & 14.1 & 3.49 & 1.65 & 24.4 & 4.7 & 24.6$\pm$2.7 & 2.5$\pm$3.5 & X & 0.22 \\
53 & 30.60$\pm$0.04 & -0.11$\pm$0.02 & 113.1$\pm$1.3 & 5.2 & 9 & 15.3 & 12.2 & 0.64$\pm$0.01 & 15.0 & 23.6 & 33.8 & 2206.7 & 5.0 & 7.6 & 5.6 & 7.27 & 5.24 & 13.0 & 4.7 & 5.8$\pm$1.5 & 20.2$\pm$3.1 & S & 0.32 \\
54 & 30.34$\pm$0.10 & -0.22$\pm$0.07 & 104.0$\pm$1.5 & 5.2 & 22 & 87.6 & 33.9 & 0.33$\pm$0.01 & 12.0 & 30.0 & 96.5 & 1102.1 & 2.5 & 3.8 & 31.9 & 2.87 & 1.96 & 48.4 & 4.7 & -4.1$\pm$6.6 & 31.1$\pm$1.0 & X & 0.28 \\
55 & 30.86$\pm$0.05 & 0.10$\pm$0.03 & 38.9$\pm$1.5 & 2.7 & 6 & 13.1 & 7.8 & 0.52$\pm$0.03 & 17.1 & 31.3 & 13.7 & 1043.6 & 3.0 & 5.8 & 6.1 & 6.02 & 1.96 & 11.1 & 6.2 & 25.1$\pm$1.3 & 1.8$\pm$2.9 & C & 0.24 \\
56 & 32.03$\pm$0.05 & 0.07$\pm$0.01 & 95.6$\pm$1.0 & 5.2 & 7 & 16.6 & 15.6 & 0.86$\pm$0.01 & 10.4 & 27.2 & 30.9 & 1859.0 & 2.6 & 2.4 & 3.8 & 11.45 & 5.80 & 16.5 & 4.8 & 21.9$\pm$1.0 & 6.3$\pm$2.8 & S & 0.45 \\
57 & 33.22$\pm$0.05 & 0.01$\pm$0.03 & 100.2$\pm$1.3 & 6.5 & 6 & 32.0 & 17.0 & 0.18$\pm$0.01 & 18.8 & 27.8 & 42.0 & 1314.1 & 2.1 & 2.2 & 8.2 & 4.56 & 1.68 & 26.5 & 4.6 & 14.6$\pm$3.6 & 16.9$\pm$5.2 & S & 0.31 \\
58 & 33.63$\pm$0.10 & -0.01$\pm$0.02 & 104.0$\pm$1.0 & 6.5 & 19 & 62.5 & 41.1 & 0.32$\pm$0.01 & 10.4 & 27.3 & 92.8 & 1485.7 & 2.4 & 2.5 & 16.0 & 13.90 & 6.21 & 43.2 & 4.6 & 12.5$\pm$2.2 & 5.8$\pm$0.9 & L & 0.36 \\
59 & 34.28$\pm$0.10 & 0.16$\pm$0.07 & 57.2$\pm$1.7 & 1.6 & 25 & 24.2 & 12.3 & 1.33$\pm$0.06 & 11.4 & 33.3 & 17.9 & 739.4 & 3.8 & 12.9 & 19.9 & 8.03 & 4.01 & 13.7 & 7.1 & 26.9$\pm$1.9 & 34.3$\pm$0.7 & S,H & - \\
60 & 35.56$\pm$0.04 & 0.02$\pm$0.06 & 52.6$\pm$3.3 & 10.4 & 12 & 77.6 & 42.9 & 0.60$\pm$0.05 & 14.7 & 28.0 & 115.8 & 1491.5 & 1.3 & 0.8 & 11.1 & 4.68 & 2.48 & 55.1 & 6.1 & 11.1$\pm$11.2 & 62.6$\pm$2.0 & L,X & 0.41 \\
61 & 35.18$\pm$0.07 & -0.75$\pm$0.05 & 34.2$\pm$1.6 & 2.2 & 10 & 16.0 & 9.0 & 0.90$\pm$0.06 & 11.8 & 21.7 & 16.8 & 1050.8 & 4.2 & 10.9 & 10.2 & 4.65 & 1.91 & 13.1 & 6.7 & -7.5$\pm$1.7 & 24.2$\pm$2.4 & X,H & - \\
62 & 41.16$\pm$0.11 & -0.21$\pm$0.03 & 60.0$\pm$2.4 & 8.9 & 11 & 56.4 & 46.4 & 0.47$\pm$0.03 & 16.6 & 27.9 & 64.1 & 1138.0 & 2.0 & 2.3 & 15.9 & 9.16 & 8.32 & 47.7 & 6.1 & -20.1$\pm$4.0 & 11.8$\pm$1.0 & S & 0.29 \\
63 & 43.16$\pm$0.05 & -0.01$\pm$0.04 & 11.4$\pm$2.9 & 11.1 & 12 & 107.0 & 35.0 & 0.25$\pm$0.10 & 19.7 & 31.9 & 291.8 & 2727.4 & 1.9 & 0.9 & 12.1 & 2.70 & 1.67 & 54.6 & 7.6 & 7.3$\pm$6.9 & 16.4$\pm$3.4 & X,H & - \\
64 & 48.60$\pm$0.03 & 0.03$\pm$0.04 & 17.1$\pm$1.0 & 10.8 & 7 & 36.3 & 21.8 & 0.19$\pm$0.02 & 17.3 & 34.3 & 75.1 & 2069.3 & 1.6 & 0.8 & 4.4 & 4.12 & 1.90 & 31.7 & 8.2 & 17.6$\pm$6.8 & 62.4$\pm$5.1 & C & 0.42 \\
65 & 49.02$\pm$0.09 & -0.29$\pm$0.04 & 66.0$\pm$2.7 & 5.3 & 28 & 82.9 & 30.5 & 0.41$\pm$0.02 & 14.6 & 29.0 & 69.9 & 844.2 & 1.4 & 1.6 & 22.7 & 5.16 & 2.47 & 39.2 & 6.3 & -7.6$\pm$3.4 & 3.4$\pm$1.0 & X & 0.49 \\
66 & 49.45$\pm$0.12 & -0.38$\pm$0.05 & 60.9$\pm$5.0 & 5.3 & 34 & 102.0 & 42.9 & 0.42$\pm$0.05 & 17.3 & 35.2 & 375.2 & 3676.4 & 5.4 & 5.2 & 23.8 & 4.31 & 2.52 & 54.8 & 6.3 & -15.8$\pm$4.3 & 2.7$\pm$0.7 & L,X & - \\
67 & 49.38$\pm$0.02 & -0.27$\pm$0.05 & 50.4$\pm$1.5 & 5.3 & 12 & 25.5 & 14.6 & 0.42$\pm$0.02 & 15.8 & 30.0 & 47.0 & 1841.2 & 2.8 & 2.8 & 6.2 & 3.93 & 2.34 & 19.2 & 6.3 & -6.3$\pm$4.8 & 88.3$\pm$2.0 & S & 0.46 \\
68 & 53.18$\pm$0.07 & 0.06$\pm$0.03 & 22.8$\pm$1.4 & 4.0 & 7 & 16.5 & 15.5 & 0.36$\pm$0.03 & 12.8 & 21.8 & 6.1 & 369.2 & 1.4 & 3.5 & 10.1 & 11.06 & 6.34 & 16.3 & 6.8 & 25.5$\pm$1.7 & 18.5$\pm$2.1 & L & 0.36 \\
69 & 54.11$\pm$0.03 & -0.07$\pm$0.02 & 38.8$\pm$0.9 & 4.0 & 9 & 15.7 & 10.0 & 0.43$\pm$0.01 & 15.2 & 33.4 & 7.0 & 445.7 & 1.6 & 3.6 & 8.8 & 6.28 & 2.32 & 13.2 & 6.8 & 16.4$\pm$1.6 & 27.4$\pm$4.2 & X & 0.41 \\
70 & 59.59$\pm$0.06 & -0.21$\pm$0.02 & 27.1$\pm$0.8 & 2.2 & 6 & 10.4 & 6.4 & 0.39$\pm$0.02 & 15.3 & 20.8 & 1.7 & 162.0 & 0.9 & 3.3 & 9.2 & 3.13 & 2.67 & 7.9 & 7.5 & 15.6$\pm$0.9 & 5.5$\pm$2.9 & S & 0.37 \\
71 & 354.64$\pm$0.04 & 0.49$\pm$0.07 & -20.8$\pm$0.4 & 4.4 & 13 & 33.8 & 20.0 & 0.31$\pm$0.01 & 11.8 & 25.6 & 31.8 & 941.8 & 1.8 & 2.4 & 10.6 & 3.62 & 2.65 & 25.1 & 4.0 & 53.1$\pm$5.4 & 62.5$\pm$1.6 & S & 0.43 \\
72 & 354.18$\pm$0.04 & -0.05$\pm$0.02 & -31.3$\pm$1.9 & 11.4 & 6 & 45.2 & 25.1 & 0.19$\pm$0.02 & 13.6 & 40.2 & 63.0 & 1392.5 & 1.6 & 1.3 & 8.5 & 5.87 & 2.51 & 32.0 & 3.3 & -10.2$\pm$4.1 & 19.4$\pm$4.9 & S & 0.29 \\
73 & 353.11$\pm$0.12 & 0.72$\pm$0.20 & -3.6$\pm$2.0 & 1.4 & 124 & 98.5 & 18.8 & 1.75$\pm$1.40 & 15.4 & 49.3 & 40.1 & 407.2 & 2.1 & 7.1 & 81.1 & 2.45 & 1.71 & 29.0 & 7.0 & 39.5$\pm$4.8 & 85.8$\pm$0.3 & C,X & - \\
74 & 352.61$\pm$0.03 & -0.20$\pm$0.02 & -85.1$\pm$0.8 & 7.6 & 9 & 21.8 & 12.8 & 0.17$\pm$0.01 & 13.6 & 31.8 & 44.9 & 2063.2 & 2.4 & 1.8 & 4.0 & 2.87 & 2.43 & 16.5 & 1.3 & -18.6$\pm$3.2 & 40.1$\pm$4.5 & X,H & 0.47 \\
75 & 353.38$\pm$0.08 & -0.34$\pm$0.05 & -17.6$\pm$1.0 & 3.1 & 15 & 26.8 & 14.3 & 0.46$\pm$0.04 & 12.4 & 28.4 & 62.3 & 2326.9 & 5.4 & 8.4 & 10.0 & 4.87 & 3.92 & 16.1 & 5.3 & -0.1$\pm$2.9 & 33.4$\pm$1.2 & S,H & - \\
76 & 353.40$\pm$0.02 & -0.09$\pm$0.01 & -52.3$\pm$2.1 & 10.2 & 7 & 25.1 & 10.4 & 0.56$\pm$0.03 & 15.7 & 37.4 & 131.8 & 5257.9 & 3.8 & 1.8 & 2.9 & 4.57 & 2.27 & 13.8 & 2.2 & -12.4$\pm$1.5 & 2.5$\pm$7.1 & X & 0.51 \\
77 & 353.56$\pm$0.03 & -0.05$\pm$0.04 & -57.6$\pm$1.0 & 10.2 & 10 & 46.4 & 21.4 & 0.22$\pm$0.01 & 16.6 & 32.6 & 108.9 & 2348.5 & 2.7 & 2.0 & 8.4 & 3.11 & 2.06 & 29.8 & 2.2 & -5.4$\pm$6.9 & 61.7$\pm$3.9 & C,X & 0.35 \\
78 & 351.59$\pm$0.06 & 0.17$\pm$0.01 & -41.4$\pm$0.9 & 11.5 & 13 & 48.7 & 38.6 & 0.22$\pm$0.01 & 20.9 & 36.5 & 154.5 & 3173.0 & 2.4 & 1.2 & 5.9 & 9.20 & 4.26 & 42.6 & 3.5 & 35.1$\pm$2.7 & 3.3$\pm$1.9 & S & 0.47 \\
79 & 351.94$\pm$0.13 & 0.68$\pm$0.05 & -0.4$\pm$1.4 & 1.3 & 17 & 17.6 & 9.7 & 1.03$\pm$2.71 & 12.3 & 27.4 & 3.7 & 207.9 & 2.0 & 12.4 & 26.8 & 4.53 & 3.06 & 11.6 & 7.0 & 37.6$\pm$1.1 & 8.4$\pm$0.8 & S & 0.29 \\
80 & 351.27$\pm$0.31 & 0.69$\pm$0.11 & -4.2$\pm$2.5 & 1.3 & 106 & 78.5 & 29.1 & 1.44$\pm$1.03 & 10.2 & 46.6 & 63.3 & 806.4 & 5.0 & 20.2 & 77.5 & 12.21 & 4.89 & 31.4 & 7.0 & 37.8$\pm$2.5 & 16.3$\pm$0.1 & S,X & - \\
81 & 350.19$\pm$0.08 & 0.08$\pm$0.05 & -66.3$\pm$4.4 & 10.5 & 32 & 147.8 & 53.5 & 0.26$\pm$0.02 & 13.2 & 28.7 & 525.0 & 3552.9 & 2.6 & 1.3 & 17.4 & 5.65 & 2.57 & 67.8 & 2.7 & 16.4$\pm$8.7 & 24.2$\pm$0.9 & C,X & 0.52 \\
82 & 349.87$\pm$0.08 & 0.10$\pm$0.05 & -62.0$\pm$0.9 & 10.7 & 6 & 52.8 & 44.6 & 0.09$\pm$0.01 & 14.3 & 22.3 & 174.2 & 3300.5 & 1.8 & 0.6 & 4.6 & 5.42 & 3.11 & 53.0 & 2.9 & 21.4$\pm$9.0 & 29.8$\pm$2.1 & C & 0.29 \\
83 & 349.30$\pm$0.05 & 0.16$\pm$0.01 & -63.6$\pm$0.5 & 10.5 & 8 & 36.1 & 29.4 & 0.12$\pm$0.01 & 13.6 & 20.0 & 118.2 & 3277.0 & 3.1 & 1.9 & 5.4 & 6.30 & 4.11 & 32.7 & 2.8 & 31.5$\pm$2.6 & 8.1$\pm$2.9 & S & 0.33 \\
84 & 349.12$\pm$0.06 & 0.10$\pm$0.03 & -75.9$\pm$2.1 & 10.5 & 11 & 61.8 & 34.6 & 0.25$\pm$0.01 & 16.0 & 26.7 & 301.3 & 4878.1 & 3.0 & 1.2 & 6.0 & 3.61 & 2.68 & 43.2 & 2.8 & 20.2$\pm$4.8 & 12.6$\pm$2.3 & X & 0.38 \\
85 & 349.14$\pm$0.03 & 0.01$\pm$0.02 & 16.4$\pm$1.3 & 19.7 & 6 & 58.4 & 39.8 & 0.19$\pm$0.02 & 18.6 & 26.4 & 427.8 & 7324.3 & 2.9 & 0.7 & 3.7 & 4.56 & 2.44 & 51.4 & 11.6 & -14.4$\pm$7.3 & 27.6$\pm$5.5 & S & 0.38 \\
86 & 348.94$\pm$0.06 & 0.11$\pm$0.02 & -69.3$\pm$1.7 & 10.5 & 7 & 41.1 & 24.8 & 0.17$\pm$0.01 & 16.1 & 19.9 & 88.5 & 2152.1 & 2.7 & 2.2 & 8.2 & 3.49 & 3.08 & 29.5 & 2.8 & 23.3$\pm$3.4 & 1.8$\pm$2.0 & C & 0.26 \\
87 & 347.89$\pm$0.06 & 0.03$\pm$0.03 & -29.8$\pm$1.4 & 12.9 & 6 & 62.1 & 44.1 & 0.06$\pm$0.01 & 16.3 & 30.2 & 118.4 & 1906.5 & 1.5 & 0.8 & 8.0 & 9.12 & 4.44 & 48.4 & 5.0 & 3.3$\pm$6.0 & 22.0$\pm$2.6 & L,C & 0.31 \\
88 & 347.69$\pm$0.05 & 0.24$\pm$0.03 & -72.4$\pm$1.2 & 9.8 & 6 & 35.9 & 24.9 & 0.22$\pm$0.01 & 10.3 & 21.6 & 106.6 & 2972.1 & 3.1 & 2.1 & 5.9 & 3.41 & 2.10 & 34.4 & 2.4 & 44.8$\pm$4.7 & 16.9$\pm$4.2 & C & 0.27 \\
89 & 347.62$\pm$0.02 & 0.20$\pm$0.05 & -94.0$\pm$1.5 & 9.8 & 13 & 52.1 & 29.1 & 0.25$\pm$0.01 & 17.9 & 39.3 & 135.7 & 2604.9 & 1.9 & 0.9 & 6.1 & 4.77 & 2.90 & 35.4 & 2.4 & 38.5$\pm$8.8 & 78.4$\pm$2.4 & C & 0.54 \\
90 & 347.91$\pm$0.05 & -0.37$\pm$0.05 & -94.5$\pm$1.7 & 6.9 & 10 & 36.7 & 21.4 & 0.27$\pm$0.01 & 10.9 & 23.8 & 43.6 & 1187.3 & 1.6 & 1.4 & 7.9 & 3.92 & 1.76 & 32.4 & 2.2 & -34.4$\pm$6.3 & 47.4$\pm$2.6 & C & 0.48 \\
91 & 350.51$\pm$0.04 & -0.37$\pm$0.04 & -22.8$\pm$0.8 & 3.3 & 8 & 14.1 & 9.7 & 0.30$\pm$0.02 & 12.3 & 20.4 & 8.3 & 586.1 & 2.4 & 6.8 & 9.4 & 4.36 & 1.86 & 14.3 & 5.1 & -3.3$\pm$2.5 & 53.3$\pm$3.9 & C & 0.36 \\
92 & 349.92$\pm$0.08 & -0.53$\pm$0.02 & -24.8$\pm$2.0 & 3.3 & 13 & 25.7 & 12.4 & 0.96$\pm$0.11 & 12.9 & 31.3 & 14.5 & 563.5 & 1.3 & 1.8 & 9.2 & 5.26 & 3.25 & 14.5 & 5.1 & -12.5$\pm$1.3 & 1.9$\pm$1.3 & S & 0.43 \\
93 & 345.30$\pm$0.13 & -0.17$\pm$0.03 & -16.9$\pm$0.8 & 1.4 & 13 & 12.8 & 10.0 & 0.68$\pm$0.05 & 11.2 & 24.1 & 2.6 & 200.3 & 2.2 & 16.0 & 22.6 & 7.96 & 6.02 & 10.5 & 7.0 & 18.0$\pm$0.7 & 9.2$\pm$0.8 & L,X & 0.20 \\
94 & 345.41$\pm$0.06 & -0.94$\pm$0.04 & -20.4$\pm$1.1 & 1.4 & 19 & 14.6 & 6.0 & 1.15$\pm$0.09 & 11.7 & 38.7 & 4.9 & 333.4 & 2.0 & 7.8 & 13.9 & 4.48 & 1.77 & 9.1 & 7.0 & -1.0$\pm$1.0 & 26.9$\pm$2.0 & X,H & - \\
95 & 343.79$\pm$0.08 & -0.15$\pm$0.05 & -25.6$\pm$1.9 & 2.6 & 16 & 23.3 & 12.2 & 0.65$\pm$0.07 & 10.0 & 34.8 & 10.6 & 456.3 & 1.6 & 3.7 & 13.0 & 4.21 & 2.39 & 15.9 & 5.9 & 13.0$\pm$2.2 & 25.0$\pm$1.4 & C,S & 0.38 \\
96 & 343.49$\pm$0.03 & -0.03$\pm$0.05 & -28.9$\pm$1.1 & 2.6 & 14 & 17.8 & 9.5 & 0.81$\pm$0.04 & 17.4 & 36.9 & 8.1 & 456.6 & 1.8 & 4.8 & 11.3 & 3.84 & 2.16 & 13.0 & 5.9 & 18.2$\pm$2.3 & 67.5$\pm$2.5 & C,X & 0.49 \\
97 & 343.12$\pm$0.04 & -0.08$\pm$0.03 & -30.8$\pm$1.6 & 2.7 & 7 & 10.3 & 7.8 & 0.59$\pm$0.04 & 12.9 & 28.9 & 8.8 & 855.3 & 3.8 & 11.1 & 7.3 & 6.85 & 3.20 & 9.2 & 5.8 & 15.8$\pm$1.3 & 32.2$\pm$3.4 & L,H & - \\
98 & 344.16$\pm$0.10 & -0.61$\pm$0.05 & -24.3$\pm$1.3 & 2.2 & 34 & 35.7 & 13.7 & 1.22$\pm$0.09 & 10.8 & 27.7 & 45.3 & 1270.4 & 4.3 & 9.6 & 19.3 & 4.26 & 2.20 & 18.5 & 6.2 & -2.9$\pm$1.9 & 12.4$\pm$0.9 & X & - \\
99 & 342.54$\pm$0.05 & 0.18$\pm$0.01 & -41.7$\pm$0.5 & 12.5 & 9 & 43.0 & 36.1 & 0.10$\pm$0.01 & 11.5 & 23.2 & 122.1 & 2841.1 & 2.7 & 1.7 & 6.6 & 19.11 & 6.89 & 37.6 & 5.2 & 38.2$\pm$1.8 & 2.8$\pm$2.0 & L & 0.35 \\
100 & 342.06$\pm$0.05 & 0.42$\pm$0.01 & -70.8$\pm$0.9 & 10.9 & 7 & 29.1 & 28.1 & 0.27$\pm$0.01 & 13.1 & 22.3 & 146.6 & 5040.5 & 3.1 & 1.3 & 2.9 & 12.46 & 8.08 & 28.9 & 3.9 & 81.8$\pm$2.7 & 13.7$\pm$2.8 & L & 0.42 \\
101 & 340.93$\pm$0.04 & -0.32$\pm$0.07 & -44.8$\pm$1.0 & 3.3 & 10 & 21.5 & 11.9 & 0.37$\pm$0.01 & 10.2 & 21.5 & 19.2 & 891.4 & 2.8 & 5.8 & 10.8 & 2.85 & 1.86 & 17.5 & 5.4 & -0.1$\pm$3.9 & 86.8$\pm$2.4 & X & 0.29 \\
102 & 341.24$\pm$0.05 & -0.26$\pm$0.03 & -43.6$\pm$0.8 & 3.3 & 16 & 20.9 & 11.0 & 0.61$\pm$0.02 & 12.5 & 25.9 & 20.0 & 955.8 & 2.8 & 5.6 & 10.0 & 5.06 & 2.33 & 14.5 & 5.4 & 3.2$\pm$1.6 & 14.8$\pm$2.2 & C,X & 0.51 \\
103 & 339.95$\pm$0.03 & -0.55$\pm$0.02 & -92.3$\pm$0.9 & 10.1 & 6 & 25.9 & 18.0 & 0.25$\pm$0.01 & 12.8 & 31.0 & 40.8 & 1575.7 & 1.9 & 1.4 & 4.9 & 4.91 & 3.25 & 21.2 & 3.6 & -91.8$\pm$3.9 & 33.0$\pm$5.0 & X & 0.42 \\
104 & 338.59$\pm$0.02 & 0.03$\pm$0.05 & -22.7$\pm$2.2 & 13.6 & 11 & 80.7 & 39.6 & 0.15$\pm$0.02 & 13.1 & 27.4 & 369.9 & 4582.2 & 2.5 & 0.9 & 7.2 & 5.84 & 2.44 & 51.2 & 6.5 & 5.4$\pm$11.7 & 73.7$\pm$3.0 & C & 0.43 \\
105 & 338.41$\pm$0.05 & 0.00$\pm$0.02 & -47.6$\pm$2.0 & 2.7 & 6 & 10.5 & 6.6 & 0.95$\pm$0.06 & 22.1 & 34.3 & 8.1 & 777.7 & 2.5 & 5.2 & 5.3 & 7.01 & 3.11 & 7.8 & 5.9 & 19.7$\pm$0.7 & 1.5$\pm$2.5 & C & 0.31 \\
106 & 338.06$\pm$0.06 & -0.00$\pm$0.10 & -40.9$\pm$1.9 & 3.0 & 22 & 50.1 & 17.8 & 0.82$\pm$1.42 & 11.7 & 26.3 & 55.5 & 1109.5 & 3.6 & 7.9 & 26.3 & 3.51 & 1.70 & 27.5 & 5.7 & 18.8$\pm$5.4 & 78.9$\pm$1.3 & C,X & 0.29 \\
107 & 337.87$\pm$0.06 & -0.01$\pm$0.04 & -54.5$\pm$3.5 & 3.6 & 14 & 35.6 & 16.2 & 0.82$\pm$0.68 & 10.2 & 26.2 & 29.2 & 818.5 & 5.1 & 21.4 & 35.9 & 2.95 & 1.52 & 26.8 & 5.2 & 17.1$\pm$2.7 & 11.0$\pm$3.1 & C,X & 0.16 \\
108 & 337.67$\pm$0.04 & -0.05$\pm$0.04 & -48.1$\pm$1.7 & 12.1 & 11 & 59.4 & 33.1 & 0.29$\pm$0.11 & 15.4 & 25.5 & 308.4 & 5192.9 & 3.5 & 1.5 & 6.4 & 3.28 & 1.71 & 50.9 & 5.4 & -8.5$\pm$7.4 & 36.4$\pm$3.7 & C & 0.44 \\
109 & 337.70$\pm$0.04 & 0.10$\pm$0.04 & -75.2$\pm$1.0 & 10.8 & 6 & 31.1 & 27.7 & 0.15$\pm$0.01 & 17.0 & 22.9 & 183.4 & 5903.6 & 4.7 & 2.5 & 4.0 & 11.41 & 4.58 & 30.3 & 4.4 & 21.8$\pm$7.7 & 49.2$\pm$3.0 & C & 0.36 \\
110 & 337.41$\pm$0.03 & -0.39$\pm$0.05 & -41.5$\pm$0.6 & 3.0 & 8 & 10.7 & 7.0 & 0.43$\pm$0.01 & 15.9 & 30.4 & 11.7 & 1092.9 & 3.9 & 9.2 & 6.1 & 5.74 & 2.02 & 9.9 & 5.7 & -1.2$\pm$2.6 & 78.6$\pm$3.9 & C & - \\
111 & 336.88$\pm$0.10 & -0.02$\pm$0.15 & -75.7$\pm$3.2 & 4.7 & 64 & 212.8 & 48.9 & 0.46$\pm$0.26 & 13.3 & 47.1 & 171.1 & 804.0 & 1.6 & 2.0 & 66.2 & 3.69 & 2.39 & 63.7 & 4.4 & 14.3$\pm$12.5 & 60.0$\pm$0.3 & C,X & - \\
112 & 336.95$\pm$0.06 & -0.00$\pm$0.07 & -120.6$\pm$1.2 & 7.7 & 22 & 96.8 & 31.4 & 0.26$\pm$0.01 & 13.5 & 32.7 & 133.5 & 1379.6 & 2.1 & 2.1 & 23.3 & 3.51 & 1.56 & 51.1 & 3.3 & 9.3$\pm$9.0 & 49.8$\pm$1.7 & S,X & - \\
113 & 337.77$\pm$0.07 & -0.34$\pm$0.03 & -41.2$\pm$0.5 & 3.0 & 7 & 14.5 & 10.8 & 0.23$\pm$0.01 & 13.4 & 34.4 & 10.6 & 727.9 & 2.7 & 6.4 & 8.5 & 5.28 & 2.88 & 13.2 & 5.7 & 1.1$\pm$1.7 & 16.5$\pm$2.4 & X & 0.30 \\
114 & 337.93$\pm$0.09 & -0.48$\pm$0.03 & -39.0$\pm$1.4 & 2.9 & 15 & 25.7 & 16.0 & 0.72$\pm$0.04 & 10.8 & 34.6 & 21.8 & 846.7 & 2.2 & 4.0 & 10.9 & 7.51 & 2.90 & 19.4 & 5.8 & -4.9$\pm$1.7 & 10.1$\pm$1.3 & C & - \\
115 & 336.48$\pm$0.05 & -0.23$\pm$0.06 & -86.8$\pm$2.2 & 5.0 & 19 & 34.5 & 19.3 & 0.67$\pm$0.02 & 13.4 & 29.0 & 47.7 & 1385.2 & 2.6 & 3.3 & 10.4 & 3.37 & 2.30 & 25.5 & 4.2 & -4.7$\pm$5.1 & 49.6$\pm$1.5 & S & 0.44 \\
116 & 336.36$\pm$0.04 & -0.14$\pm$0.05 & -79.5$\pm$0.4 & 5.0 & 6 & 22.2 & 16.5 & 0.06$\pm$0.01 & 16.9 & 27.1 & 19.0 & 855.2 & 1.6 & 2.1 & 6.8 & 5.16 & 1.69 & 25.5 & 4.2 & 2.7$\pm$4.6 & 66.3$\pm$4.6 & C & 0.33 \\
117 & 335.19$\pm$0.08 & -0.34$\pm$0.07 & -41.0$\pm$2.8 & 2.9 & 25 & 45.3 & 15.5 & 0.84$\pm$0.08 & 10.3 & 29.3 & 22.1 & 487.4 & 2.0 & 5.5 & 30.1 & 3.91 & 1.61 & 24.8 & 5.8 & 2.3$\pm$3.4 & 37.7$\pm$1.4 & S,X & 0.34 \\
118 & 335.28$\pm$0.03 & -0.13$\pm$0.05 & -44.9$\pm$0.7 & 2.9 & 8 & 12.2 & 6.8 & 0.66$\pm$0.01 & 8.6 & 20.0 & 7.0 & 573.8 & 2.1 & 5.1 & 7.2 & 3.52 & 2.12 & 9.4 & 5.8 & 13.0$\pm$2.3 & 69.4$\pm$3.9 & C,X & 0.30 \\
119 & 334.04$\pm$0.02 & -0.00$\pm$0.06 & -86.1$\pm$1.3 & 5.2 & 9 & 27.3 & 17.1 & 0.47$\pm$0.01 & 11.4 & 23.1 & 20.9 & 766.0 & 2.7 & 6.2 & 15.3 & 3.39 & 2.45 & 22.0 & 4.3 & 15.2$\pm$5.3 & 87.3$\pm$2.2 & S,X & 0.22 \\
120 & 333.17$\pm$0.39 & -0.40$\pm$0.16 & -51.3$\pm$3.7 & 3.6 & 169 & 370.2 & 109.4 & 0.54$\pm$0.06 & 11.6 & 44.8 & 436.1 & 1178.1 & 2.7 & 3.9 & 133.4 & 10.46 & 4.24 & 121.0 & 5.4 & -7.0$\pm$10.0 & 18.4$\pm$0.1 & L,X & - \\
121 & 333.61$\pm$0.03 & 0.05$\pm$0.02 & -83.5$\pm$2.3 & 4.9 & 7 & 15.1 & 7.7 & 1.07$\pm$0.04 & 14.8 & 21.6 & 37.3 & 2466.5 & 3.1 & 2.6 & 3.1 & 4.77 & 1.60 & 12.3 & 4.5 & 20.2$\pm$1.7 & 3.3$\pm$6.3 & X & 0.40 \\
122 & 333.27$\pm$0.06 & 0.06$\pm$0.13 & -47.7$\pm$1.0 & 3.6 & 20 & 58.9 & 29.3 & 0.24$\pm$0.01 & 10.4 & 27.2 & 43.7 & 743.1 & 2.3 & 4.9 & 29.8 & 3.96 & 2.04 & 41.1 & 5.4 & 21.8$\pm$8.2 & 87.7$\pm$0.9 & X & 0.23 \\
123 & 333.19$\pm$0.03 & -0.07$\pm$0.05 & -87.8$\pm$1.4 & 5.2 & 12 & 41.3 & 15.3 & 0.31$\pm$0.01 & 16.1 & 27.1 & 61.0 & 1476.0 & 2.8 & 3.4 & 12.4 & 2.87 & 1.51 & 25.4 & 4.4 & 8.5$\pm$4.7 & 80.5$\pm$4.3 & C & 0.33 \\
124 & 333.06$\pm$0.04 & 0.01$\pm$0.08 & -43.9$\pm$1.6 & 3.6 & 10 & 25.3 & 13.7 & 0.66$\pm$0.03 & 17.0 & 30.8 & 28.0 & 1107.4 & 3.0 & 5.3 & 10.8 & 3.48 & 2.29 & 18.2 & 5.4 & 18.7$\pm$4.7 & 80.4$\pm$2.1 & C & 0.31 \\
125 & 332.76$\pm$0.06 & -0.00$\pm$0.02 & -95.3$\pm$1.0 & 5.6 & 9 & 24.0 & 18.5 & 0.29$\pm$0.01 & 9.6 & 31.3 & 24.6 & 1028.1 & 2.2 & 3.1 & 8.2 & 9.50 & 5.85 & 19.5 & 4.2 & 14.3$\pm$2.0 & 15.7$\pm$1.9 & S & 0.28 \\
126 & 332.36$\pm$0.13 & -0.06$\pm$0.12 & -48.4$\pm$1.6 & 3.1 & 40 & 69.4 & 28.4 & 0.54$\pm$0.02 & 8.8 & 35.5 & 57.9 & 834.2 & 2.5 & 5.0 & 33.6 & 6.40 & 1.51 & 47.0 & 5.8 & 15.9$\pm$6.3 & 39.3$\pm$0.7 & C & 0.36 \\
127 & 331.98$\pm$0.13 & -0.12$\pm$0.02 & -51.2$\pm$0.5 & 3.1 & 16 & 32.5 & 23.5 & 0.31$\pm$0.01 & 10.2 & 24.5 & 25.4 & 780.4 & 2.3 & 4.6 & 15.5 & 6.86 & 5.21 & 25.1 & 5.8 & 12.8$\pm$1.4 & 2.3$\pm$0.8 & S & 0.28 \\
128 & 331.64$\pm$0.04 & -0.07$\pm$0.04 & -85.0$\pm$0.7 & 5.3 & 8 & 22.2 & 13.8 & 0.23$\pm$0.01 & 16.8 & 26.4 & 31.8 & 1435.7 & 3.5 & 5.7 & 8.7 & 4.76 & 2.44 & 17.8 & 4.5 & 8.6$\pm$3.2 & 38.0$\pm$3.2 & S,X & 0.28 \\
129 & 331.69$\pm$0.03 & -0.22$\pm$0.05 & -47.6$\pm$0.9 & 3.1 & 7 & 13.4 & 8.7 & 0.37$\pm$0.01 & 13.6 & 21.3 & 10.5 & 778.5 & 2.1 & 3.8 & 5.9 & 4.05 & 2.13 & 11.9 & 5.8 & 7.2$\pm$2.9 & 84.1$\pm$3.5 & C & 0.25 \\
130 & 331.32$\pm$0.20 & -0.14$\pm$0.08 & -88.9$\pm$2.5 & 5.3 & 57 & 192.2 & 62.0 & 0.52$\pm$0.02 & 10.8 & 37.5 & 265.2 & 1379.7 & 2.3 & 2.6 & 51.7 & 5.22 & 3.03 & 74.3 & 4.5 & 2.6$\pm$7.0 & 9.8$\pm$0.3 & S,X & - \\
131 & 331.07$\pm$0.21 & -0.41$\pm$0.06 & -65.6$\pm$2.2 & 4.0 & 53 & 132.1 & 54.9 & 0.38$\pm$0.02 & 10.7 & 33.6 & 162.0 & 1226.2 & 2.5 & 3.5 & 43.9 & 7.73 & 3.50 & 63.2 & 5.2 & -11.1$\pm$4.3 & 1.6$\pm$0.3 & S,X & - \\
132 & 338.91$\pm$0.07 & 0.56$\pm$0.04 & -62.7$\pm$2.3 & 4.2 & 16 & 38.1 & 19.4 & 0.47$\pm$0.02 & 11.7 & 26.0 & 64.5 & 1691.0 & 3.1 & 3.8 & 11.3 & 3.93 & 2.13 & 26.7 & 4.7 & 57.4$\pm$3.2 & 24.1$\pm$1.7 & X & - \\
133 & 340.82$\pm$0.10 & -1.01$\pm$0.03 & -27.0$\pm$1.7 & 2.4 & 20 & 21.8 & 12.0 & 0.87$\pm$0.07 & 13.4 & 32.4 & 8.5 & 390.9 & 2.0 & 6.5 & 17.5 & 6.65 & 3.61 & 13.8 & 6.1 & -22.3$\pm$1.3 & 9.2$\pm$0.9 & S,X & 0.54 \\
134 & 334.64$\pm$0.03 & 0.43$\pm$0.05 & -65.2$\pm$0.4 & 4.0 & 13 & 19.3 & 13.4 & 0.30$\pm$0.01 & 10.1 & 21.1 & 8.9 & 462.2 & 1.5 & 3.1 & 9.8 & 6.92 & 3.92 & 15.1 & 5.0 & 47.7$\pm$3.7 & 66.0$\pm$1.9 & L,X & 0.45 \\
135 & 345.50$\pm$0.09 & 0.34$\pm$0.04 & -17.1$\pm$0.5 & 2.4 & 9 & 17.0 & 11.8 & 0.46$\pm$0.02 & 13.5 & 27.3 & 22.4 & 1317.6 & 4.4 & 9.8 & 9.1 & 4.75 & 2.77 & 14.6 & 6.1 & 34.1$\pm$1.7 & 18.0$\pm$1.7 & S,X & - \\
136 & 335.34$\pm$0.08 & 0.42$\pm$0.02 & -59.4$\pm$0.5 & 3.8 & 6 & 18.8 & 17.1 & 0.15$\pm$0.01 & 9.7 & 14.5 & 4.4 & 235.3 & 1.2 & 3.9 & 15.1 & 7.98 & 4.81 & 18.5 & 5.1 & 45.5$\pm$1.3 & 4.7$\pm$1.9 & S & 0.15 \\
137 & 327.99$\pm$0.04 & -0.07$\pm$0.04 & -49.7$\pm$1.5 & 3.1 & 9 & 15.3 & 10.2 & 1.44$\pm$0.06 & 12.7 & 35.4 & 9.0 & 591.0 & 3.1 & 10.7 & 12.8 & 3.56 & 1.61 & 16.3 & 6.0 & 15.7$\pm$2.3 & 42.0$\pm$3.5 & X & 0.25 \\
138 & 327.73$\pm$0.07 & -0.38$\pm$0.04 & -74.5$\pm$3.2 & 4.2 & 10 & 27.9 & 17.0 & 1.07$\pm$0.07 & 16.9 & 30.6 & 22.7 & 814.1 & 1.9 & 3.0 & 10.6 & 4.92 & 2.07 & 23.7 & 5.3 & -9.8$\pm$2.6 & 12.6$\pm$2.6 & C & 0.42 \\
139 & 327.29$\pm$0.04 & -0.56$\pm$0.03 & -46.3$\pm$1.9 & 3.0 & 11 & 14.6 & 6.5 & 0.83$\pm$0.05 & 16.1 & 38.4 & 34.8 & 2387.7 & 5.1 & 7.2 & 5.0 & 3.28 & 2.08 & 9.0 & 6.1 & -9.6$\pm$1.8 & 42.4$\pm$3.3 & C,H & - \\
140 & 327.15$\pm$0.07 & -0.29$\pm$0.07 & -61.2$\pm$1.5 & 3.5 & 19 & 41.3 & 21.4 & 0.42$\pm$0.01 & 9.6 & 27.2 & 26.6 & 643.1 & 2.4 & 5.8 & 24.5 & 4.77 & 1.93 & 30.9 & 5.8 & 1.5$\pm$4.0 & 43.0$\pm$1.4 & C,X & 0.31 \\
141 & 326.98$\pm$0.04 & -0.05$\pm$0.07 & -58.5$\pm$3.0 & 3.5 & 15 & 31.7 & 15.7 & 0.63$\pm$0.05 & 10.1 & 23.1 & 19.8 & 626.5 & 1.8 & 3.6 & 14.9 & 3.24 & 2.26 & 20.9 & 5.8 & 16.1$\pm$4.3 & 75.9$\pm$1.8 & C,X & 0.41 \\
142 & 326.80$\pm$0.04 & -0.11$\pm$0.03 & -56.3$\pm$0.7 & 3.5 & 6 & 10.8 & 6.5 & 0.74$\pm$0.01 & 12.9 & 26.1 & 7.8 & 719.2 & 2.4 & 5.3 & 5.8 & 3.20 & 1.94 & 9.4 & 5.8 & 12.5$\pm$1.9 & 37.2$\pm$5.0 & C & 0.34 \\
143 & 326.83$\pm$0.07 & -0.55$\pm$0.04 & -58.0$\pm$1.5 & 3.5 & 7 & 23.1 & 12.4 & 0.33$\pm$0.01 & 17.2 & 29.5 & 5.5 & 238.8 & 1.4 & 5.3 & 21.5 & 2.54 & 1.97 & 17.6 & 5.8 & -14.3$\pm$2.2 & 4.4$\pm$2.7 & X & 0.15 \\
144 & 326.66$\pm$0.05 & 0.60$\pm$0.11 & -39.8$\pm$1.7 & 1.8 & 40 & 37.7 & 15.4 & 0.92$\pm$0.05 & 13.7 & 34.0 & 15.4 & 409.7 & 1.8 & 5.3 & 26.7 & 5.52 & 2.31 & 20.4 & 6.9 & 40.7$\pm$3.4 & 80.5$\pm$0.7 & S,X & 0.55 \\
145 & 326.46$\pm$0.06 & 0.85$\pm$0.13 & -40.7$\pm$1.1 & 1.8 & 23 & 26.0 & 14.1 & 0.62$\pm$0.02 & 13.3 & 36.7 & 5.8 & 221.1 & 1.3 & 5.0 & 24.3 & 7.02 & 2.66 & 17.6 & 6.9 & 48.7$\pm$4.0 & 75.5$\pm$0.7 & S,X & 0.49 \\
146 & 345.24$\pm$0.06 & 1.04$\pm$0.02 & -15.2$\pm$0.9 & 2.4 & 13 & 14.3 & 10.0 & 0.88$\pm$0.07 & 16.3 & 41.5 & 7.8 & 548.7 & 1.8 & 3.8 & 7.4 & 7.16 & 4.08 & 11.2 & 6.1 & 63.4$\pm$0.7 & 5.8$\pm$1.8 & S,X & 0.52 \\
147 & 345.42$\pm$0.06 & 1.43$\pm$0.04 & -12.4$\pm$0.8 & 2.4 & 11 & 15.8 & 11.1 & 0.56$\pm$0.05 & 15.7 & 48.7 & 28.1 & 1782.2 & 5.1 & 9.5 & 7.2 & 4.72 & 1.84 & 16.4 & 6.1 & 79.9$\pm$1.6 & 17.1$\pm$2.8 & C & - \\
148 & 351.65$\pm$0.05 & -1.23$\pm$0.08 & -10.5$\pm$1.3 & 2.6 & 14 & 22.2 & 12.0 & 0.45$\pm$0.08 & 16.0 & 45.5 & 19.0 & 857.1 & 3.3 & 8.5 & 13.8 & 4.09 & 3.92 & 13.5 & 5.8 & -36.5$\pm$3.7 & 63.8$\pm$1.2 & C,X & - \\
149 & 323.88$\pm$0.06 & 0.03$\pm$0.03 & -57.8$\pm$1.1 & 10.0 & 14 & 61.6 & 42.3 & 0.17$\pm$0.01 & 12.1 & 25.9 & 62.5 & 1015.2 & 1.5 & 1.5 & 14.7 & 5.23 & 2.69 & 52.7 & 5.9 & 14.0$\pm$5.9 & 21.5$\pm$1.9 & S & 0.36 \\
150 & 322.16$\pm$0.09 & 0.63$\pm$0.04 & -54.9$\pm$1.2 & 3.3 & 13 & 25.6 & 16.3 & 0.82$\pm$0.03 & 13.3 & 30.4 & 27.4 & 1070.3 & 3.7 & 8.6 & 14.3 & 4.37 & 3.31 & 19.1 & 6.1 & 56.0$\pm$2.5 & 20.4$\pm$1.2 & L,H & - \\
151 & 321.07$\pm$0.05 & -0.52$\pm$0.02 & -60.6$\pm$0.7 & 9.3 & 12 & 44.6 & 26.2 & 0.24$\pm$0.01 & 18.5 & 30.2 & 53.8 & 1207.8 & 1.2 & 0.8 & 6.9 & 3.76 & 3.20 & 30.9 & 5.9 & -73.1$\pm$3.9 & 20.8$\pm$2.4 & C,X & 0.46 \\
152 & 320.28$\pm$0.07 & -0.30$\pm$0.02 & -65.8$\pm$1.2 & 8.6 & 11 & 46.6 & 36.4 & 0.22$\pm$0.01 & 16.6 & 31.8 & 61.0 & 1309.4 & 1.5 & 1.1 & 8.5 & 10.21 & 3.76 & 41.2 & 5.8 & -32.3$\pm$3.3 & 11.2$\pm$1.9 & S & 0.45 \\
153 & 320.40$\pm$0.02 & 0.13$\pm$0.03 & -4.9$\pm$1.5 & 12.5 & 10 & 42.0 & 20.5 & 0.27$\pm$0.11 & 12.1 & 30.8 & 81.6 & 1944.9 & 1.4 & 0.6 & 4.7 & 3.33 & 2.63 & 25.7 & 8.1 & 35.1$\pm$5.8 & 49.0$\pm$4.2 & C & 0.55 \\
154 & 320.17$\pm$0.03 & 0.83$\pm$0.04 & -39.8$\pm$2.1 & 2.4 & 14 & 13.4 & 5.4 & 0.84$\pm$0.06 & 14.1 & 33.5 & 4.7 & 350.8 & 1.4 & 3.5 & 8.3 & 2.67 & 1.75 & 8.2 & 6.7 & 56.0$\pm$1.8 & 56.6$\pm$3.3 & C,X & 0.62 \\
155 & 317.66$\pm$0.07 & 0.10$\pm$0.03 & -44.2$\pm$1.0 & 2.8 & 8 & 12.4 & 8.4 & 0.38$\pm$0.01 & 11.6 & 23.5 & 17.6 & 1416.6 & 5.7 & 15.0 & 7.9 & 3.59 & 3.08 & 10.0 & 6.6 & 26.1$\pm$1.3 & 12.9$\pm$2.2 & C & 0.40 \\
156 & 316.77$\pm$0.07 & -0.02$\pm$0.05 & -38.6$\pm$1.2 & 2.5 & 20 & 25.4 & 14.1 & 0.80$\pm$0.04 & 14.6 & 29.4 & 48.3 & 1898.4 & 6.5 & 14.7 & 13.9 & 5.50 & 2.55 & 17.9 & 6.7 & 20.6$\pm$2.0 & 32.7$\pm$1.4 & X,H & 0.51 \\
157 & 314.23$\pm$0.03 & 0.30$\pm$0.08 & -61.2$\pm$2.0 & 4.2 & 10 & 25.8 & 18.1 & 0.53$\pm$0.02 & 14.0 & 32.8 & 14.1 & 545.2 & 1.8 & 4.1 & 13.9 & 5.09 & 3.70 & 20.5 & 6.2 & 41.9$\pm$5.8 & 77.8$\pm$1.6 & L & 0.35 \\
158 & 311.60$\pm$0.04 & 0.30$\pm$0.01 & -48.1$\pm$0.9 & 4.1 & 6 & 10.2 & 7.6 & 0.33$\pm$0.01 & 12.1 & 36.5 & 11.5 & 1127.7 & 3.5 & 7.4 & 5.2 & 6.45 & 5.48 & 8.1 & 6.4 & 41.7$\pm$1.0 & 18.3$\pm$4.2 & S & - \\
159 & 309.92$\pm$0.03 & 0.39$\pm$0.08 & -57.9$\pm$1.4 & 5.5 & 6 & 34.4 & 23.6 & 0.13$\pm$0.01 & 16.7 & 35.1 & 27.3 & 793.4 & 1.4 & 1.7 & 9.8 & 4.58 & 3.04 & 28.2 & 6.4 & 56.4$\pm$7.6 & 89.0$\pm$1.5 & L,X & 0.31 \\
160 & 309.15$\pm$0.03 & -0.30$\pm$0.08 & -42.8$\pm$1.3 & 3.5 & 8 & 21.9 & 14.9 & 0.39$\pm$0.02 & 10.4 & 19.4 & 16.2 & 740.9 & 3.2 & 9.2 & 15.3 & 4.08 & 2.56 & 18.9 & 6.7 & 2.6$\pm$4.6 & 79.5$\pm$2.3 & C,X & 0.24 \\
161 & 305.55$\pm$0.06 & -0.01$\pm$0.04 & -38.2$\pm$1.4 & 3.8 & 11 & 24.3 & 17.3 & 0.52$\pm$0.03 & 17.7 & 30.0 & 15.4 & 631.3 & 1.6 & 2.7 & 9.8 & 7.14 & 2.80 & 21.3 & 6.9 & 20.8$\pm$2.5 & 30.5$\pm$2.1 & C & 0.45 \\
162 & 305.22$\pm$0.06 & -0.00$\pm$0.04 & -32.5$\pm$2.9 & 3.8 & 21 & 40.6 & 17.7 & 0.69$\pm$0.08 & 12.3 & 42.1 & 29.3 & 722.2 & 1.6 & 2.3 & 14.4 & 4.94 & 2.70 & 22.1 & 6.9 & 21.3$\pm$2.6 & 32.2$\pm$1.6 & S,X & 0.55 \\
163 & 305.30$\pm$0.09 & 0.23$\pm$0.06 & -38.8$\pm$2.1 & 3.8 & 29 & 75.5 & 22.2 & 0.66$\pm$0.05 & 16.1 & 51.2 & 100.6 & 1332.9 & 2.3 & 2.6 & 20.9 & 3.83 & 1.59 & 35.7 & 6.9 & 36.5$\pm$3.9 & 3.9$\pm$1.5 & C,X & - \\
\hline
min & 5.37 & -1.23 & -120.6 & 1.3 & 6 & 10.2 & 5.3 & 0.06 & 8.5 & 14.5 & 1.7 & 162.0 & 0.9 & 0.6 & 2.9 & 2.09 & 1.51 & 6.6 & 1.3 & -99.0 & 1.5 & - & 0.14 \\
max & 354.64 & 1.43 & 113.1 & 19.7 & 169 & 370.2 & 109.4 & 1.75 & 22.1 & 53.0 & 525.0 & 7324.3 & 6.5 & 21.4 & 133.4 & 26.81 & 11.36 & 121.0 & 11.6 & 81.8 & 89.0 & - & 0.62 \\
med & 320.40 & -0.07 & -20.4 & 4.2 & 10 & 25.9 & 16.0 & 0.47 & 13.3 & 28.0 & 31.8 & 1070.3 & 2.5 & 4.0 & 10.2 & 4.91 & 2.63 & 20.5 & 5.4 & 11.1 & 29.8 & - & 0.31 \\
mean & 202.39 & -0.07 & -2.5 & 5.3 & 16 & 39.7 & 20.6 & 0.53 & 13.5 & 29.4 & 65.0 & 1443.8 & 2.7 & 5.0 & 14.6 & 6.18 & 3.21 & 25.9 & 5.2 & 10.3 & 35.3 & - & 0.35 \\
Std & 155.33 & 0.36 & 60.9 & 3.3 & 20 & 40.8 & 14.3 & 0.32 & 2.8 & 7.2 & 89.2 & 1174.6 & 1.1 & 3.7 & 15.3 & 3.65 & 1.80 & 16.7 & 1.4 & 26.1 & 26.3 & - & 0.09 \\
S & -0.28 & 0.49 & 0.3 & 1.3 & 5 & 4.4 & 2.3 & 0.89 & 0.4 & 0.8 & 2.8 & 2.1 & 1.0 & 1.6 & 4.3 & 2.31 & 2.07 & 1.8 & 0.2 & -0.6 & 0.5 & - & 0.15 \\
K & -1.90 & 2.20 & -1.1 & 1.6 & 30 & 27.6 & 8.8 & 0.67 & -0.5 & 1.0 & 8.4 & 5.6 & 0.7 & 3.4 & 24.8 & 7.45 & 4.85 & 5.6 & 2.3 & 2.9 & -1.0 & - & -0.5
\enddata
\tablecomments {meaning of columns are detailed in Table \ref{t1_explain}. Column (1) ID of large-scale filaments; Cols. (2)-(4) flux-weighted longitude, latitude (degree), and LSR velocity (km s$^{-1}$); Column (5) distance (kpc); Column (6) Number of clumps in filaments; Column (7) sum of edge lengths (pc); Column (8) end-to-end length (pc); Column (9) velocity gradient (km s$^{-1}$ pc$^{-1}$); Cols. (10) and (11) minimum and maximum temperatures (K) of clumps; Column (12) mass of the filament ($10^3 M_\odot$); Column (13) line mass ($M_\odot$ pc$^{-1}$); Cols. (14) and (15) $H_2$ column density ($10^{22}$ cm$^{-2}$) and volume density ($10^3$ cm$^{-3}$); Column (16) sum of edge lengths over filament width; Column (17) aspect ratio; Column (18) linearity; Column (19) linearity-weighted length; Column (20) Galactocentric radius (kpc); Column (21) height above Galactic mid-plane (pc); Column (22) orientation angle between filaments major axes and Galactic mid-plane in the projected sky (degree); Column (23) morphology class; Column (24) dense gas mass fraction, ``-'' means no DGMF measurement.}
\end{deluxetable}
\end{longrotatetable}
}
\begin{table}[b]
{\begin{tabular}{llp{40.5em}}
\toprule
\tablecolumns{3}
Column & & \multicolumn{1}{c}{Explanation} \\
\colrule
(1) ID & & Assigned filament ID \\
(2) $l_{wt}$ &(deg) & Flux-weighted Galactic longitude \\
(3) $b_{wt}$ &(deg) & Flux-weighted Galactic latitude \\
(4) $v_{wt}$ &(km s$^{-1}$) & Flux weighted local standard of rest (LSR) velocity \\
(5) $d$ &(kpc) & Distance, the median of all clump distances in this filament \\
(6) N$_{cl}$ & & The number of clumps in the filament \\
(7) $L_{sum}$ &(pc) & The sum of ``edges'' (separation of each two connected clumps) of this filament \\
(8) $L_{end}$ &(pc) & End-to-end length along the major axis \\
(9) $v_{grad}$ &(km s$^{-1}$ pc$^{-1}$) & Mean velocity gradient on edges \\
(10) $T_{min}$ &(K) & Minimum temperature of clumps \\
(11) $T_{max}$ &(K) & Maximum temperature of clumps \\
(12) Mass &($10^3 M_\odot$) & Mass of the filament, derived from Herschel Hi-GAL column density map (detailed in Sect. \ref{sec:mass}) \\
(13) $M_{line}$ &($M_\odot$ pc$^{-1}$) & Line mass, mass per unit length \\
(14) $N_{H_2}$ &($10^{22}$ cm$^{-2}$) & $H_2$ column density \\
(15) $n_{H_2}$ &($10^3$ cm$^{-3}$) & $H_2$ volume density \\
(16) $\frac{L_{sum}}{w}$ & & The ratio between sum of edge lengths and width. The width is the mean of clump diameters \\
(17) $f_{A}$ & & Aspect ratio, the ratio of area between the circle enclosing the filament and the concave hull (detailed in Appendix \ref{sec:concavehull}). For an approximately elliptical distribution, $f_{A}$ is very similar with the aspect ratio of the ellipse \\
(18) $f_{L}$ & & Linearity, the ratio between spread (standard deviation) of clumps along major axis and spread along minor axis \\
(19) $L_{wt}$ & (pc) & Linearity-weighted length, $L_{wt} = L_{end}\cdot \sqrt{1+\dfrac{4}{f_L^2}}$ \\
(20) $R_{gc}$ &(kpc) & Galactocentric radius, median of clump Galactocentric radius \\
(21) z &(pc) & Height above Galactic mid-plane \\
(22) $|\theta|$ & (deg) & Orientation angle between filaments major axes and Galactic mid-plane in the projected sky (calculation of $|\theta|$ is from PCA, detailed in Appendix \ref{sec:PCA}) \\
(23) Mor. & & Morphology class following \citet{Wang2015}, ``L'' represents linear straight L-shape, ``C'' bent C-shape, ``S'' quasi-sinusoidal shape, ``X'' crossing of multiple filaments, ``H'' head-tail or hub-filament system \\
(24) DGMF & & Dense gas mass fraction, the ratio between dense gas mass and total mass of the filament. It is illustrated in Sect. \ref{sec:DGMF}\\
\botrule
\end{tabular}}
\caption{Explanation of physical properties listed in table \ref{t1} }
\label{t1_explain}
\end{table}
\subsection{Statistics}
\begin{figure*}
\gridline{\fig{figure3a.pdf}{0.3\textwidth}{(a)}
\fig{figure3b.pdf}{0.3\textwidth}{(b)}
\fig{figure3c.pdf}{0.3\textwidth}{(c)}
}
\gridline{\fig{figure3d.pdf}{0.3\textwidth}{(d)}
\fig{figure3e.pdf}{0.3\textwidth}{(e)}
\fig{figure3f.pdf}{0.3\textwidth}{(f)}
}
\gridline{\fig{figure3g.pdf}{0.3\textwidth}{(g)}
\fig{figure3h.pdf}{0.3\textwidth}{(h)}
\fig{figure3i.pdf}{0.3\textwidth}{(i)}
}
\gridline{\fig{figure3j.pdf}{0.3\textwidth}{(j)}
}
\caption{Physical properties of the large-scale filaments. Blue bars show distributions of filament properties. (a)-(f) are normalized distribution of Galactic longitude, Galactic latitude, height to Galactic mid-plane, Galactocentric radius, distance, and radial velocity for filaments. Red steps show distribution of ATLASGAL clumps with velocity information from \citet{Urquhart2018}. Panels (g)-(j) are mass, linearity-weighted length, line mass, and orientation angle of filaments. Red bars denote ``bones'' with the same longitude-velocity criteria with \citet{Wang2016}, those are criteria (6)$\sim$(8) in Sect. \ref{sec:bone}. While black bars show the new definition of bones in this work. We require that a new bone not only has similar velocity with an arm in the same Galactic longitude, but also has the similar distance to us with that arm (criterion (9) in Sect. \ref{sec:bone}).}
\label{sta}
\end{figure*}
Some representative physical properties of the filaments are shown in Fig. \ref{sta}. Panels (a)-(h) are normalized distribution of Galactic longitude, Galactic latitude, height to Galactic mid-plane, Galactocentric radius, distance, and radial velocity for filaments. Panels (g)-(j) are mass, linearity-weighted length, line mass, and orientation angle of filaments.\\
No filaments are found within $\left| l\right|<5^\circ$, due to the limitation of catalogue we use. The distribution of filaments with Galactic longitude has two peaks around $l=-22^\circ$ and $l=23^\circ$ shown in Fig. \ref{sta} (a). \citet{Mattern2018} also find the former peak but not the later one in their small-scale ATLASGAL filaments, owning to their Galactic longitude ranging from $l=-60^\circ$ to $l=18^\circ$. They attribute the decrease of filaments count towards center to difficulty in identifying filaments in confused structures. They think the suppression outwards means it is unlikely to find filaments towards outer Galaxy. In our sample, the position of filaments shows a similarity with those of ATLASGAL clumps, implying filaments tend to be found in the region where clumps are crowded, and this can be directly perceived in Fig. \ref{faceon} (c). So the peaks in distribution of filaments with Galactic longitude are merely due to peaks in distribution of ATLASGAL clumps. This result also implies that large-scale filaments are ubiquitous in the inner Galactic plane, and therefore, an unavoidable issue in the study of star formation. But we also note that this similarity in distribution of filaments and clumps does not always exist (for example, Galactocentric radius in Fig. \ref{sta} (d)). The distribution of Galactic latitude of filaments shown in Fig. \ref{sta} (b) has a peak and mean $b=-0.01^\circ$ close to Galactic mid-plane, which is consistent with \citet{Mattern2018}.\\
Of our filaments, F120 stands out, whose linearity-weighted length is 121 pc and the mass is $4.36\times10^5$ M$_\odot$. Except this extreme filament, the linearity-weighted lengths of filaments range from 7 to 74 pc, with mass of order of 10$^3$ M$_\odot$ to 10$^5$ M$_\odot$. We also find no obvious correlation between length and distance. The vertical distances from filaments to Galactic mid-plane (Fig. \ref{sta} (c)) are not symmetrical about the Galactic mid-plane (S=-0.6), which agrees with the result of northern sky \citep{Wang2016}.\\
The relation between mass and lengths (sum of edge lengths) of large-scale filaments is shown in Fig. \ref{ML} (a). The black curve is the best-fitted line using the least square method, whose expression is $lgL=0.46\,lgM-0.40$. This implies a tight power-law relation between length and mass, $M\sim L^{2.17}$. The power-law index is similar with that of filaments from BGPS sources \citep{Wang2016}, indicating a fractal dimension of 2.17. For comparison, the relation between mean cylinder radii and mass of large-scale filaments is also examined (Fig. \ref{ML} (b)), which gives a relation of $lgL=0.56\,lgM-2.14$. The corresponding dimension is 1.77, out of range 2 to 3 from \citet{Roman2010}, implying the mean cylinder radii is a less essential property for large-scale filaments.\\
Do filaments in the proximity of the Galactic mid-plane prefer to align with the Galactic mid-plane? The relation between orientation angle and height to the Galactic mid-plane of large-scale filaments is plotted in Fig. \ref{orivsz}. There is no obvious correlation between the two, which is identical to the result from northern sky \citep{Wang2016}.
\begin{figure*}
\gridline{\fig{figure4a.pdf}{0.5\textwidth}{(a)}
\fig{figure4b.pdf}{0.5\textwidth}{(b)}
}
\caption{Relation between the mass and lengths (a), cylinder radii (b) of large-scale filaments. Blue circles denote filaments and black lines are fitted results. }
\label{ML}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[scale=0.85,angle=0]{figure5.pdf}
\caption{Orientation angle vs. vertical height to the physical Galactic mid-plane of large-scale filaments. Blue filled circles show filaments while red crosses and green pluses mark bones with the same criteria with \citet{Wang2016} and bones in this work respectively.}
\label{orivsz}
\end{center}
\end{figure*}
\subsection{Dense Gas Mass Fraction} \label{sec:DGMF}
Dense gas mass fraction (DGMF) of a filament is the ratio of dense gas mass to total mass of the whole filament, which is an important quantity related to star formation rate (SFR) and star formation efficiency (SFE) of molecular clouds \citep{Heiderman2010, Lada2010, Lada2012}, although \citet{Kainulainen2013} suggest that DGMF and SFE may not be closely linked. To avoid deviation engendered by distinct methods to get mass, both dense gas mass and filament mass are derived from Herschel Hi-GAL column density map, which is illustrated in Sect. \ref{sec:Herschel}. After discarding filaments with incomplete column density information, 130 of the 163 large-scale filaments have DGMF measurement. Their DGMFs range from 14.7\% to 62.4\%, with a mean value of 35.6\%. The result is larger than that from \citet{Ragan2014} and \citet{Abreu2016}, who take the ratio between mass from ATLASGAL 870 \textmu m dust emission and mass from $^{13}$CO emission as DGMF. But our result is consistent with a value of 50\% from ``Nessie'' \citep{Goodman2014}, where DGMF is from the ratio of mass in an envelope traced by HNC observation to mass in cylinders with a fixed diameter and above a column density threshold. We conjecture that diverse definitions of dense gas mass account for the difference. We also employ similar dense gas mass calculation with \citet{Abreu2016}, where dense gas mass is acquired from ATLASGAL flux. The DGMFs become smaller.\\
\section{Discussion}
\label{sec:discussion}
\subsection{Comparison of our MST Filaments to MST Filaments from BGPS Sources and Other Previously Known Filaments}
To see whether the MST method is robust on sources from different catalogues, we compare our filaments from ATLASGAL Galactic clumps to those identified from BGPS by \citet{Wang2016} in a common longitude range the two survey reside ($7.5^\circ<l<60^\circ$). In the common region, 42 filaments are identified from BGPS sources while 67 are from ATLASGAL. The detailed comparison is in Appendix \ref{sec:BGPS&ATLASGAL}. On the whole, in the common Galactic region of the two catalogues ($7.5^\circ<l<60^\circ$), most (70\%) of the filaments identified from BGPS sources are also found from ATLASGAL clumps despite different methods they use to extract sources, distinct lower limit of source luminosities, and various surveys referred to obtain radial velocities of sources \citep{Shirley2013,Urquhart2018}. \\
The distinction between filaments from the two catalogues is mainly caused by the number of sources used to identify filaments. BGPS sources contain over 8400 continuum sources, but only 3126 of them have velocity measurement from HCO$^+(3-2)$ and/or N$_2$H$^+(3-2)$ spectral. Since we aim to identify velocity coherent filaments, velocity is indispensable for the MST method. For ATLASGAL sources, 7809 of them have radial velocities obtained from 21 archival molecular line surveys \citep{Urquhart2018}. In the overlaid longitude we compare ($7.5^\circ<l<60^\circ$), 2201 BGPS sources have velocity measurement while that for ATLASGAL sources is 3379. Therefore, it is natural that the number of filaments identified from ATLASGAL sources is larger than that from BGPS sources. More sources with velocity information also inform us that some of filaments identified from MST may be actually a part of a larger structure (for instance, BGPS filaments F2 and F3 are eastern and western part of ATLASGAL filament F4, respectively). While missing velocity information of a portion of sources may lead to failure to identify some filaments. \\
Our MST method identified some previously known large-scale filaments in the southern sky. Among three ``bone'' candidates in southern sky from \citet{Zucker2015}, filament BC\_355.31-0.29 and BC\_332.21-0.04 are found by MST. Filament BC\_355.31-0.29 is our F117. Filament BC\_332.21-0.04 is a complex consisting of F128, F127 and the southern part of F126. Filament G350.54+0.69 found by \citet{Hong-Li2018} is a part of our F80. Filament G316.75-0.1 \citep{Watkins2019} share a portion of spatial overlap with our F156.
\subsection{Large-scale Filaments in the Milky Way}
\label{sec:bone}
\begin{figure*}
\gridline{\fig{figure6a.pdf}{0.8\textwidth}{(a)}
}
\gridline{\fig{figure6b.pdf}{0.8\textwidth}{(b)}
}
\caption{(a) Longitude-velocity view of filaments and spiral arms. Spiral arms from \citet{Reid2019} are plotted as belts in a variety of colors. Our data do not support ``grand design'' spiral pattern, neither are they opposed to this pattern. We just simply show this optimistic presumption of spiral structure. Grey ``arms'' are arm segments which may join together to form arms. Filaments are shown as black circles and if it is a bone, a red cross is overlaid. Grey rectangles at the bottom mark the longitude range of our filaments. (b) Longitude-distance view of filaments and spiral arms. If a bone is also within an arm in this longitude-distance diagram, a blue plus is overlaid. (c) (In next page) Face-on view of filaments and spiral arms. Spiral arms are shown with their ``kink'' widths (taken from Table 2 of \citet{Reid2019}) which can be regarded as characteristic widths of arms. Red dots in the background mark clumps used to identify filaments from \citet{Urquhart2018}} \label{faceon}
\end{figure*}
\setcounter{figure}{5}
\begin{figure*}
\gridline{\fig{figure6c.pdf}{0.8\textwidth}{(c)}
}
\caption{(Continued) }
\end{figure*}
To examine the Galactic distribution of large-scale filaments, we exhibit locations of the filaments in the longitude-velocity (PV) space and draw Galactic spiral arms as reference (Fig. \ref{faceon} (a)). The spiral arm model is from a fitting result of Galactic high-mass star forming regions with trigonometric parallaxes \citep{Reid2019}. According to previous criterion to judge whether a large-scale filament is associated with spiral arms \citep{Ragan2014,Abreu2016}, if a filament has LSR velocity within $\pm$5 km s$^{-1}$ \citep{Wang2016} of a spiral arm in the same Galactic longitude, it is thought to be an arm filament. This criterion is illustrated in Fig. \ref{faceon} (a). Belts show spiral arms in PV space with widths of 10 km s$^{-1}$. If a filament locates in any of the belts, it is thought as an arm filament. Of our 163 large-scale filaments, 87 (53\%) are in spiral arms or spurs under this criterion. But this value is only for comparison to previous work. We will update the criteria to judge which filaments could be thought as arm filaments later. \citet{Ragan2014} finds most of their filaments are inter-arm filaments while \citet{Abreu2016} and \citet{Wang2016} find arm filaments percentages are 67\% and 80\%, respectively.\\
There exists some filaments lying in the center of spiral arms and so sketch out bones of the Milky Way \citep{Goodman2014,Zucker2015}. ``Bones'' are also found in our large-scale filaments if we add the following three additional criteria \citep{Wang2016}:
\begin{itemize}
\item[(6)]Lie very close to Galactic mid-plane, $|z|\leq 20$ pc;
\item[(7)] Run roughly parallel to arms in the projected sky, $\theta \leq 30^\circ $;
\item[(8)] Flux-weighted LSR velocity is within $\pm$5 km s$^{-1}$ of a spiral arm in the same Galactic longitude.
\end{itemize}
Their Galactic distribution is shown as ``crosses'' in Fig. \ref{faceon} and physical properties are denoted as red bars in Fig. \ref{sta}. Fig. \ref{faceon} (c) shows filaments and clumps (red filled circles) overlaid to Galactic spiral arms as viewed from the Northern Galactic Pole. The exhibited widths of spiral arms are intrinsic (Gaussian $1\sigma $) arm widths at Galactic radius of ``kinks'' \citep{Reid2019}. There are 23, 12, 20, and 6 large-scale filaments in Norma-Outer, Scutum-Centaurus-OSC, Sagittarius-Carina, and Perseus arm, respectively while for bones these values are 10, 8, 12, 5, respectively (summarized in Table \ref{t_arm}). Here, to judge whether a filament is associated with an arm, we still follow the criterion from previous work. The differences between arm filaments and bones are that arm filament need only to satisfy criterion (8) while bones should fulfill not only criterion (8) but also (6) and (7). Surprisingly, most filaments in Scutum-Centaurus-OSC arm and Perseus arm are bones (8/12 and 5/6) and these two arms have been thought as two dominant spiral arms \citep{Drimmel2000, Churchwell2009}. However, we note that bone fractions in the four arms have no significant differences and we lack a statistic sample.\\
From the face-on view of filaments in spiral arms, we notice that some bones thought as arm filaments are obviously not in any spiral arms or spurs (see red crosses between arms in Fig. \ref{faceon} (c)). That is because in the previous criteria used to judge whether a filament is in a spiral arm, we only require that the filament is velocity coherent with a spiral arm in the same Galactic longitude. However, the distance of the filament to that arm was not in consideration. So this is a 2D position-velocity (PV) match. Researches on Galactic location of filaments have also been confined to two-dimension match, PV \citep{Ragan2014, Abreu2016, Wang2016} or position-position (PP) \citep{Wang2015,Mattern2018}. Now we obtain better distance estimation of both filaments and spiral arms, position-position-velocity (PPV) match becomes possible.\\
Our new definition of ``bone'' adds a distance constraint compared with the previous one when judging whether a filament is in a spiral arm. So it becomes a PPV match. This constraint is another criterion:
\begin{itemize}
\item[(9)]The difference between its distance to us and the distance of its PV-related spiral arm to us is less than 1 kpc in the same Galactic longitude.
\end{itemize}
That is, for a bone with previous definition meeting criteria (6), (7), and (8), we further examine whether it is close enough to this arm. If a previous bone also meets criterion (9), it will pass our distance constraint and be thought as a bone by new definition. Otherwise it will be excluded from bones, because in this situation the filament may just look near an arm in line of sight. The 1 kpc tolerance is from a combination of uncertainties of filament distances and $1\sigma$ widths of arms. Distances from the Sun versus Galactic longitudes of filaments and spiral arms are plotted in Fig. \ref{faceon} (b). Bones satisfying our new definition are denoted with blue pluses. As we can see, previous bones (denoted in red crosses) far from spiral arms are eliminated in our new definition. And this is also clear in the face-on map. However, some of previous bones seem to be in one arm in Fig. \ref{faceon} (b) are also excluded. This situation occurs when a filament has similar velocity with a spiral arm at a Galactic longitude, but far from this arm and near another arm in distance.\\
\setlength{\tabcolsep}{3mm}{
{
\doublerulesep=2pt
\begin{table}[b]
{\begin{tabular}{@{}cccccc}
\toprule
Arm & Filaments & Bones & Bone fraction & Bones with distance constraint & Mean DGMF \\
\colrule
Norma-Outer & 23 (26.4$\%$) & 10 (20.4$\%$) & 43.5$\%$ & 5 (33.3$\%$) & 37.2$\%\pm2.9\%$ \\
Scutum-Centaurus-OSC & 12 (13.8$\%$) & 8 (16.3$\%$) & 66.7$\%$ & 1 (6.7$\%$) & 39.7$\%\pm2.6\%$ \\
Sagittarius-Carina & 20 (23.0$\%$) & 12 (24.5$\%$) & 60.0$\%$ & 3 (20.0$\%$) & 30.8$\%\pm2.2\%$ \\
Perseus & 6 (6.9$\%$) & 5 (10.2$\%$) & 83.3$\%$ & 1 (6.7$\%$) & 33.3$\%$ \\
Other Arm Segments & 26 (29.9$\%$) & 14 (33.3$\%$) & 53.8$\%$ & 5 (10.0$\%$) & 34.6$\%\pm1.7\%$\\
Total & 87 (100$\%$) & 49 (100$\%$) & 56.3$\%$ & 15 (100$\%$) & 35.3$\%\pm1.2\%$ \\
\botrule
\end{tabular}}
\caption{Number of filaments, bones with the definition of \citet{Wang2016} and our new definition of bones with distance constraint in each arm. Column ``Filaments'' means filaments meeting criteria (1)$\sim$(5) in each arm. Column ``Bones'' shows number of bones satisfying criteria (1)$\sim$(8). Column ``Bone fraction'' indicates that of filaments in each arm, how many are bones. Column ``Bones with distance constraint'' shows number of bones satisfying criteria (1)$\sim$(9). The last Column lists mean DGMF of filaments in each arm with standard error. Only one filament in Perseus has DGMF measurement, so the standard error of DGMF for Perseus is none. Apart from four main spiral arms, filaments associated with other arm segments or spurs are also listed in the 6th row. They consist of Local arm, 3 kpc and other segments and spurs, which may not be true spiral arms \citep{Reid2019}.}
\label{t_arm}
\end{table}
}
}
The amount of filaments in spiral arms decreases significantly after adding the distance constraint. Unlike bones, when a filament match a spiral arm in PPV space (at a certain Galactic longitude, they have similar distance from the Sun and similar velocity), it is thought as an arm filament. Of the 163 large-scale filaments, 138 are not in any spiral arms or spurs. While 8, 1, 4, and 1 are in Norma-Outer, Scutum-Centaurus-OSC, Sagittarius-Carina and Perseus arm (satisfy criteria (8) and (9)), respectively. The values for bones (satisfying criteria (6) $\sim$ (9)) are 5, 1, 3, and 1. The fact that 85\% of large-scale filaments are not in any spiral arms or structures seems to indicate that filament may not be associated so tightly with spiral arms as thought before. But we also note that due to the existence of distance ambiguity of filaments and uncertainty of spiral arm distance, especially in $4^{th}$ quadrant, some filaments that should be in spiral arms may be ruled out by the distance constraint, resulting in underestimation of arm filaments count. \\
If we do not care whether large-scale filaments are bones of the Milky Way and just want to know the influence of spiral arms to filaments, we could take filaments meeting criteria (1)$\sim$(5), (8), and (9) as arm filaments. And the others are inter-arm filaments. Between the two, we find no significant distinctions in some of their physical properties such as mean temperatures, non-thermal velocity dispersion, and column densities. But they also have differences in properties such as Galactocentric radius, height above Galactic mid-plane, mass, length. The detailed discussion for arm filaments and inter-arm filaments are in Appendix \ref{sec:env}. \\
\subsection{Dense Gas Mass Fraction}
Dense gas mass fraction is found to be associated with the Galactic position of filaments. From analysis of seven filaments identified from Galactic Ring Survey \citep[GRS,][]{Jackson2006}, \citet{Ragan2014} suggested that DGMF decreases as Galactocentric radius increase. They also found filaments near the Galactic mid-plane tend to have larger DGMF. In contrast, \citet{Abreu2016} find no obvious correlation between height above Galactic mid-plane and DGMF from inspection of their ten filaments identified in the fourth Galactic quadrant.\\
To investigate this issue in a larger statistical sample, we exam the DGMF and Galactic position of our 130 filaments with DGMF measurement. Results are shown in Fig. \ref{DGMF}. Surprisingly, DGMF is neither related to Galactocentric radius nor to height above Galactic mid-plane. We change the size of boundaries to get filament mass and dense gas mass, and get the similar result. We also try to use clump mass from ATLASGAL as dense gas mass, and the DGMF is still not correlated with Galactocentric radius or height above Galactic mid-plane. \\
\begin{figure*}
\gridline{\fig{figure7a.pdf}{0.45\textwidth}{(a)}
\fig{figure7b.pdf}{0.45\textwidth}{(b)}
}
\caption{Dense gas mass fraction and Galactic position of filaments}
\label{DGMF}
\end{figure*}
\subsection{Dense Clumps in Filaments}
\begin{figure*}
\gridline{\fig{figure8a.pdf}{0.5\textwidth}{(a)}
\fig{figure8b.pdf}{0.5\textwidth}{(b)}
}
\gridline{\fig{figure8c.pdf}{0.5\textwidth}{(c)}
\fig{figure8d.pdf}{0.5\textwidth}{(d)}
}
\caption{Comparison between clumps on- or off-filaments. Panel (a): clump mass vs. radius for ATLASGAL clumps. Brown filled circles and blue pluses denote clumps in filaments and not in, respectively. Solid and dotted lines are fitted results. In panel (b), column density distribution of clumps in filaments or not are shown in brown bars and blue steps, respectively. We bin the logarithmic radius with bin size of 0.1 and the mean logarithmic mass in each bin is shown in panel (c). Black curve denote clumps in filaments while red curve show those not in. Error bars are standard deviations of mass in each bin. In panel (d), brown triangles are ATLASGAL clumps in the northern sky we use. Red filled circles are clumps in BGPS filaments in \citet{Wang2016}. Blue crosses are BGPS clumps from \citet{EB15}. Lines are fitted results.}
\label{infl}
\end{figure*}
Dense clumps are thought to be birthplaces of massive stars \citep[e.g.][]{Motte2018}. Their typical mass is $10^3M_\odot$, and typical molecular hydrogen column density is $10^{22}\cdot$ cm$^{-2}$ with typical radius of 1 pc. According to \citet{Wang2016}, dense clumps in large-scale filaments are slightly denser than those not in, indicating that filaments prefer to gather material and assist the formation of massive stars. To examine whether this preference can be extended to the whole inner Galactic plane, we compare mass of clumps with various effective radius between in filaments or not. Differences between clumps in filaments or not are shown in Fig. \ref{infl} (a), where lines are fitted results. Surprisingly, the two have no remarkable distinctions. We also bin the logarithmic radius to see more clear (Fig. \ref{infl} (c)). The bins range from about -1.6 to 1.2 with a step of 0.1. In each bin, we calculate the mean logarithmic mass of clumps in filaments or not, and then plot them versus median logarithmic radius of each bin. The black curve is clumps in filaments and the red curve shows clumps not in. Error bars denote standard deviations of logarithmic clump mass in each bin. As we can see, the two curves almost coincide except for both ends, where the number of clumps is small. But even at both ends, the differences are within the error bars. Therefore, mass of clumps in filaments or not have no significant distinctions on the same scale.\\
We then make a histogram to show molecular hydrogen column density of clumps in filaments or not (Fig. \ref{infl} (b)). The densest clump has column density of about $25\times 10^{22}\cdot$ cm$^{-2}$, but the number of extremely dense clumps is small. So we only show density ranging from 0 to $6\times 10^{22}\cdot$ cm$^{-2}$. The median of molecular hydrogen column density for 2628 clumps in large-scale filaments and 4213 not in are $1.03\times 10^{22}\cdot$ cm$^{-2}$ and $1.28\times 10^{22}\cdot$ cm$^{-2}$, respectively. Therefore, for our filaments, clumps not in filaments are on contrary slightly denser than those in filaments. This result may account for why star formation efficiency and star formation rate surface density in filaments are similar to that in other star-forming regions discovered by \citet{ZhangM2019}. To figure out whether this result is caused by not demanding enough for filaments, we increase the critical linearity for filaments to do the same experiment. And we also use bone to examine this issue. However, the results do not change much. In some cases, filamentary structures have been thought to fragment through cylinder fragmentation. Clumps from cylinder fragmentation will be denser than that from Jeans fragmentation \citep{Wang2014}. Therefore, our results give us a hint that the fragmentation process of filamentary structures may not be cylinder fragmentation.\\
To figure out why we find no distinction between clumps on- and off-filaments, while \citet{Wang2016} got different results, we restore Fig. 5 (b) of \citet{Wang2016} and overlay ATLASGAL clumps in northern sky in our Fig. \ref{infl} (d). We only plot ATLASGAL clumps in northern sky in this panel because BGPS clumps we compare are almost in northern sky. Brown triangles are ATLASGAL clumps in the northern sky. Red filled circles are 496 clumps in BGPS filaments in \citet{Wang2016}. Blue crosses are 1710 BGPS clumps with well-constrained distance estimates from \citet{EB15}, hereafter EB15 clumps. Lines are fitted results. On the whole, masses of clumps in BGPS filaments (red filled circles in Fig. \ref{infl} (d)) are twice as much as masses of EB15 clumps (blue crosses in Fig. \ref{infl} (d)) on the same scale. That is due to their difference in temperatures, although these two categories of clumps are both BGPS clumps and mass is calculated with the same formula. In \cite{Wang2016}, temperature of clumps in BGPS filaments are searched from three catalogues, and clumps with no reported temperature are assumed to be the average of measured temperature, 15 K (see Sect. 4.2 in \citet{Wang2016} for details). While in \citet{EB15}, temperatures of BGPS clumps are randomly assigned such that the temperature distribution is a lognormal function with a mean of 20 K (Sect. 4.1.1 in \citet{EB15}). This difference in temperature causes a systematic deviation in mass calculation, leading to an underestimation of clump mass in \citet{EB15} by a factor of about 1.5.\\
Mass of ATLASGAL clumps is also derived with the same formula as that of clumps in BGPS filaments and EB15 clumps. But their temperatures are directly from greybody fit to the submillimeter dust emission (see Sect. 4 in \citet{Urquhart2018}), which are reliable. ATLASGAL clumps are consistent better with clumps in BGPS filaments in Fig. \ref{infl} (d), while systematically above EB15 clumps. Therefore, the differences of BGPS clumps in filaments or not found by \citet{Wang2016} is caused by systematic deviation of mass calculation rather than their intrinsic differences.\\
\subsection{Fragmentation of Large-scale Filaments}
\begin{figure*}
\gridline{\fig{figure9a.pdf}{0.5\textwidth}{(a)}
\fig{figure9b.pdf}{0.5\textwidth}{(b)}
}
\caption{(a) mean clump mass vs. mean edge length for the 163 filaments. The orange and blue lines are theoretical results for thermal Jeans and Turbulent Jeans fragmentation. The purple line is for cylinder fragmentation. The black dashed line is fitted result. (b) slope of the relation between mean fragment mass and mean separation if we only consider filaments with linearity larger than a set of values. Error bars denote statistical errors.}
\label{frag}
\end{figure*}
Linear filaments were approximately regarded as gas cylinders when investigating their fragmentation \citep[e.g.][]{Fischera2012,Wang2014,Wang2016}. According to \citet{Fermi1953} and \citet{Ostriker1964}, isothermal gas cylinder will fragment due to gravitational instability when mass per unit length exceeds a critical line mass. Under this condition, the fragments will equally space with separation $\lambda _{cl}$ and the fragment mass will be
\begin{equation}
M_{cl}=301.7M_\odot\left( \dfrac{\lambda _{cl}}{{\rm pc}}\right) ^3
\end{equation}
Alternatively, if the fragmentation is governed by Jeans instability, the relation between fragment mass and separation is different. For thermal Jeans fragmentation, the fragment mass is
\begin{equation}
M_{thJ}=13.29M_\odot\left( \dfrac{T}{{\rm 10K}}\right) \left( \dfrac{\lambda _{thJ}}{{\rm pc}}\right)
\end{equation}
While for turbulent fragmentation, the fragment mass is
\begin{equation}
M_{turbJ}=381.7M_\odot\left( \dfrac{\sigma}{{\rm km\cdot s^{-1}}}\right)^2\left( \dfrac{\lambda _{turbJ}}{{\rm pc}}\right)
\end{equation}
where $\sigma$ is turbulent linewidth which is well approximated by the velocity dispersion measured from dense gas tracers such as $\rm NH_3$. The above equations are deduced from \citet{Wang2014}.
As can be seen, the power law index for the relation between fragment mass and separation is 3 for cylinder fragmentation while 1 for Jeans fragmentation. So we show mean clump mass versus mean clump separation in each filament in Fig. \ref{frag} (a). The slope of the fitted line is 1.46, representing the power law index between mean clump mass and mean separation. If we take the projection effect into consideration, the slope will be even shallower. This power law index indicates that large-scale filaments are more likely to have Jeans fragmentation rather than cylinder fragmentation, which is in contrast to \citet{Wang2016} who suggest a cylinder fragmentation. We note that our sample is larger (163 filaments) compared with \citet{Wang2016} (54 filaments). However, it is too early to draw a conclusion because Eq. (1) for cylinder fragmentation is based on isothermal linear gas cylinder with infinite length. In practice, a filament could not be strictly isothermal, and it also has finite length with some bents.\\
To inspect whether straight filaments would prefer to have cylinder fragmentation, we get the power law index for filaments with linearity larger than a set of values. For instance, mean clump mass versus mean separation of filaments with linearity larger than 3 are plotted. We fit a line for mass and separation, and take the slope of this line as a power law index for this linearity. For filaments with linearity larger than 5, we repeat the procedure and get another power law index. Similarly, we choose critical linearity from 1.5 to 8.5 and get a series of slopes (power law index). Then we plot power law index versus linearity in Fig. \ref{frag} (b). As can be seen, the power law index has a rising trend in linearity between 4 and 6. This result gives us a hint that linearity might play an important role in fragmentation mechanism of filaments. And the larger the linearity is, the more likely to be cylinder fragmentation. For linearity larger than 6, the trend of the curve is uncertain due to large error resulted from small sample size.
\section{Conclusion}
\label{sec:conclusion}
We employ MST method \citep{Wang2016} to search for large-scale filaments in PPV space in the inner Galactic plane. The algorithm is applied to 7809 clumps with velocity measurement in ATLASGAL Galactic clumps catalogue in the range $|l|<60^\circ$ with $|b|<1.5^\circ$ \citep{Urquhart2018}. We produce a sample of large-scale filaments in the inner Galactic plane consisting of 163 filaments. We derive their physical properties and examine their DGMFs with the help of Herschel Hi-GAL column density map derived from PPMAP \citep{Marsh2017}. We inspect the Galactic distribution of the filaments and compare them with the latest spiral arm model \citep{Reid2019}. Fragmentation of large-scale filaments is also investigated. The main results are:
\begin{itemize}
\item[1.] Robustness of MST method is verified by a comparison between the results from ATLASGAL Galactic clumps and BGPS dust clumps. Most (70\%) of the filaments identified from BGPS dust clumps are found from ATLASGAL clumps in the common Galactic region of the two catalogues despite different methods used to extract them, distinct lower limit of clump luminosities and various surveys referred to obtain radial velocities of clumps.
\item[2.] The Galactic position of filaments are asymmetric about Galactic mid-plane, which may be an observational bias.
\item[3.] Dense gas mass fractions of filaments have no significant distinctions in different Galactic radii and vertical height above Galactic mid-plane, in contrast to previous studies.
\item[4.] Filaments are compared with updated spiral arm model and a new PPV match is employed to judge whether a filament is in a spiral arm. Under this matching method, a number of filaments that are thought to be in arm from PP or PV match are eliminated. So the number of filaments actually associated with arms decreases a lot.
\item[5.] Bone fraction and DGMF do not vary too much in different spiral arms.
\item[6.] Dense clumps in filaments have no obvious distinction in mass compared with those not in on the same scale.
\end{itemize}
\acknowledgments
We are grateful to Chao Wang, Wenyu Jiao, K. A. Marsh, Nannan Yue, Siju Zhang, Fengwei Xu, and an anonymous referee for helpful discussion that improved the scientific content as well as clarity of the manuscript. We acknowledge support by the National Key Research and Development Program of China (2017YFA0402702, 2019YFA0405100), the National Science Foundation of China (12041305, 11973013, 11721303), and the High-performance Computing Platform of Peking University through the instrumental analysis fund of Peking University (0000057511). This research has made use of SAOImageDS9, developed by Smithsonian Astrophysical Observatory. This research made use of Montage. It is funded by the National Science Foundation under Grant Number ACI-1440620, and was previously funded by the National Aeronautics and Space Administration's Earth Science Technology Office, Computation Technologies Project, under Cooperative Agreement Number NCC5-626 between NASA and the California Institute of Technology.
|
\section{Introduction}
Numerical simulation has been an important tool for plasma physics and fusion research since the inception of particle-in-cell methods in the 1950s~\citep{BirdsallLangdon}. Subsequent developments, such as the formulation of the gyrokinetic system of equations~\citep{Catto_lin,FriemanChen,Dubin} and flux-tube simulations that employ magnetic-field-following coordinates~\citep{Beer95}, have enabled computationally affordable numerical studies of otherwise enormously complex physical systems. These studies have been instrumental in various discoveries, such as the importance of zonal flows in reducing turbulent transport in axisymmetric systems~\citep{Dimits} and novel plasma modes~\citep{Hallatschek_tails,Parisi_modes}. Gyrokinetic simulations have also successfully matched turbulent heat fluxes from experimental data~\citep{Candy_2003}, have shown the importance of interactions between the ion and electron scales~\citep{Howard_2016}, and have been used to predict new phenomena that have been borne out in experiment~\citep{DiSiena_prediction}. For a review on the history and state of the art of plasma simulation for fusion science, see~\citet{White_2017} and references therein.
Gyrokinetic simulation of tokamak and stellarator plasmas can be divided into two distinct but related paradigms, that of `global' and `local' simulation. In the global paradigm, which was the original approach used in early gyrokinetic studies~\citep{Parker93}, the entire physical device (or a finite portion thereof) is modelled, which includes the spatial variation of the magnetic geometry and pressure profiles which make up the equilibrium state. This approach has the advantage of being able to capture effects arising from radial profile variation, such as turbulence spreading~\citep{Garbet_1994}.
However, due to radial variation in the equilibrium temperature and magnetic geometry, the gyro-averaging operator becomes spatially dependent, and thus difficult to calculate; early studies were forced to adopt simplifications that made calculation of the gyroaverage tractable. More sophisticated, non-spectral gyroaveraging schemes have also been developed~\citep{Jolliet_2007,Steiner_2015}, but these schemes either suffer from low accuracy or high computational cost~\citep{guadagni_cerfon_2017}. Additionally, global simulations typically use Dirichlet boundary conditions which nullify perturbations at the radial boundaries. With sufficiently large `buffer' regions at the radial boundaries, such boundary conditions are hoped to be `benign', though it is unclear what effects they have on physical observables without detailed studies for each simulation~\citep{Candy_globloc}. The `local' paradigm stems from the development of the flux-tube domain~\citep{Beer95}, which considers an infinitesimally thin box elongated along a magnetic field line. As the flux-tube samples the equilibrium profiles at a specific radial location, the gyrokinetic equation becomes homogeneous in the direction perpendicular to the magnetic field, and so statistical periodicity can be assumed, allowing the use of Fourier spectral methods and thereby permitting an analytical treatment of the gyroaveraging operator. This development led to a number of fixed-grid gyrokinetic codes~(e.g.,~\citep{gyro,gs2,gene}) that can probe the local characteristics of turbulence with high fidelity at only moderate computational costs. This, however, came at the expense of neglecting any influence from the effects of radial profile variation.
Formally, local gyrokinetics is the $(k_\perp L)^{-1} \rightarrow 0$ limit of global gyrokinetics, and numerical studies have indeed shown this to be the case~\citep{Candy_limit}. Here, $k_\perp$ is the inverse scale length of the fluctuations across the magnetic field and $L$ is a characteristic large-scale size (such as the device minor radius $a$). Recently, it has been shown by~\citet{Parra_globallocal} that the local and global approaches to gyrokinetic simulation can be bridged using an appropriate expansion of the gyrokinetic equation, and that an approach that sits between the local and global paradigms can be formulated. Indeed, radial profile variation of the pressure has been recently implemented in the flux-tube code CGYRO using periodic triangle waves~\citep{Candy_globshear,Candy_globloc}. In this manuscript, we formulate a new approach to hybrid global-local gyrokinetics that is related to the one laid out in~\citep{Parra_globallocal}, which includes novel boundary conditions, equilibrium flow shear algorithms, and sources and sinks. We then implement our approach using an extended version of the gyrokinetic flux-tube code~\texttt{stella}~\citep{Barnes_stella}, benchmark it against a number of standard linear and nonlinear test cases, and discuss some limitations with regards to the novel radial boundary conditions.
This paper is organized as follows: in \S \ref{sec:theory} we describe the gyrokinetic equation and lay out the theoretical framework that underlies our approach to global gyrokinetics. The numerical implementation of our approach in the flux-tube code \texttt{stella} is explained in~\S \ref{sec:numerical_imp}, with additional details on the profile variation of magnetic geometry, the moments of the distribution function and the fluxes given in~\ref{app:mag_geo} and \ref{app:fluxes}. In \S \ref{sec:tests} we perform benchmarks of our global approach and compare it to \texttt{stella} in local operation. We offer concluding remarks in~\S \ref{sec:conclusion}.
\section{Theoretical Model}\label{sec:theory}
\subsection{Gyrokinetic equation}\label{sec:GKE_delta}
The \texttt{stella} code numerically solves the first-order $\delta f$ gyrokinetic equation, which is derived by taking the Vlasov equation and then imposing the ordering
\begin{equation}\label{eqn:orderings}
\epsilon \sim \rho_\ast \sim \frac{\delta f_s}{f_s } \sim \frac{\omega}{\Omega_\mr{i}} \sim \frac{k_\parallel}{k_\perp} \sim \frac{Z_s e\varphi}{T_s}\ll 1,
\end{equation}
where $\epsilon$ is the ordering parameter, $\rho_\ast \doteq \rho_\mr{ref} / L$, $\rho_\mr{ref}$ is a reference thermal gyroradius where $\rho_\mr{ref}\sim \rho_\mr{i}$ and $\rho_\mr{i} = v_\perp / \Omega_\mr{i}$ is the ion gyroradius, $v_\perp$ is the velocity perpendicular to the magnetic field,
$f_s$ and $\delta f_s$ are respectively the total and perturbed distribution function for species $s$, $\omega$ is a characteristic frequency of the perturbations, $\Omega_\mr{i}$ is the ion gyrofrequency,
$k_\parallel$ is the characteristic inverse scale length of the perturbations along the magnetic field, $Z_s$ is the charge number, $e$ is the (positive) unit charge, $\varphi$ is the electrostatic potential and $T_s$ is the temperature for species $s$.
This ordering allows one to average over the fast gyromotion of charged particles in a strong magnetic field, which removes the dependence of the particle gyrophase. Thus, the dimensionality of phase space is reduced to three spatial dimensions and two velocity dimensions. Additionally, the fast gyrofrequency timescale is removed from the system. Finally, it is assumed that $f_s = F_s + \delta f_s$, where $F_s$
is a stationary background Maxwellian distribution given by
\begin{equation}
F_s = n_s\left(\frac{m_s}{2\upi T_s}\right)^{3/2}\exp\left( - \frac{m_s v_\parallel^2}{2T_s} - \frac{\mu_s B}{T_s}\right).
\end{equation}
Here, $s$ is the species index; $m_s$ and $n_s$ are the particle mass and density for species $s$, respectively; $v_\parallel = \eb \bcdot \bb{v}$; $\eb$ is the unit vector along the magnetic field; $\bb{v}$ is the peculiar velocity in the frame rotating with the plasma at the center of the radial domain, $\mu_s = m_s v_\perp^2 / 2B $ is the magnetic moment; and $B$ is the magnitude of the magnetic field. Neglecting electromagnetic effects and collisions, and considering only toroidal flow~\citep{Catto_flow}, the gyrokinetic equation is given by
\begin{align}\label{eqn:GKE}
\frac{\partial g_s}{\partial t} & +R\Omega_\zeta \bb{\hat{\zeta}}\bcdot \grad g_s + v_\parallel \left(\underline{\eb \bcdot \grad } g_s + \underline{\frac{Z_se}{T_s} F_s \eb \bcdot \grad }\langle \varphi \rangle_\bb{R} \right)- \frac{\mu_s}{m_s} \underline{\eb \bcdot \grad B} \frac{\partial g_s}{\partial v_\parallel} \nonumber \\ &+ \underline{\bb{v}_{\mr{M}s}}\bcdot \left(\grad_\perp g_s +\underline{\frac{Z_se}{T_s}F_s}\grad_\perp \langle \varphi\rangle_\bb{R}\right) + \underline{\langle\bb{v}_\bb{E} \rangle_\bb{R} \bcdot \grad_\perp} g_s + \underline{\langle \bb{v}_\bb{E} \rangle_\bb{R}\bcdot \grad\big|_E F_s} + \underline{\frac{m_s }{T_s}F_s\frac{Iv_\parallel}{B}\langle\bb{v}_\bb{E} \rangle_\bb{R} \bcdot \grad_\perp} \Omega_\zeta = 0,
\end{align}
where $g_s(\bb{R},v_\parallel, \mu_s, t) \doteq \ba{\delta f}_\bb{R}$ is the perturbed gyrocenter distribution function for species $s$, underlines indicate coefficients that have spatial dependence, $\langle \cdots \rangle_\bb{R}$ denotes gyroaveraging at constant gyrocenter position $\bb{R} = \bb{r} - \bb{\rho}$, $\bb{\rho} = \eb \btimes \bb{v}/\Omega_s$ is the gyroradius vector, $\Omega_s$ is the gyrofrequency for species $s$, $R$ is the device major radius, $\bb{\hat{\zeta}}$ is the toroidal unit vector, $I = RB_\mr{T}$ where $B_\mr{T}$ is the toroidal magnetic field strength, $\Omega_\zeta$ is the toroidal rotation which varies across flux surfaces,
$\grad|_\mr{E}$ is the gradient taken at constant kinetic energy $E=m_s v^2 / 2$, the magnetic and $\bb{E}\btimes \bb{B}$ drifts are given by
\begin{subequations}
\begin{align}
\bb{v}_{\mr{M}s} &= \frac{\eb}{\Omega_s}\btimes \left(\frac{\mu_s}{m_s} \grad B + v_\parallel^2 \bb{\kappa}\right), \\
\bb{v}_\bb{E} &= \frac{c}{B}\eb \btimes \grad_\perp \varphi,
\end{align}
\end{subequations}
respectively, $c$ is the speed of light, $\bb{\kappa} = \eb \bcdot \grad \eb$ is the magnetic field curvature and $\bb{E}$ and $\bb{B}$ are the electric and magnetic fields, respectively. The gyrokinetic equation~\eqref{eqn:GKE} is boosted to the frame moving along the mean toroidal flow $R \Omega_\zeta$. This flow is assumed to be subsonic, i.e., $R\Omega_\zeta \sim \epsilon v_\mr{thi}$, and is allowed to have a strong flow gradient, i.e. $R^2 \Omega'_\zeta \sim v_\mr{thi}$, where prime denotes differentiation with respect to $\ensuremath{r}$.
This ordering implies that the Coriolis and centrifrugal drifts, given by
\begin{subequations}
\begin{align}
\bb{v}_{\mr{Coriolis},s} &= \frac{2 v_\parallel \Omega_\zeta}{\Omega_{s}} \eb \btimes [(\grad_\bb{R} R \btimes \bb{\hat{\zeta}}) \btimes \eb]\\
\bb{v}_{\mr{centrifugal},s} &= - \frac{R\Omega^2_\zeta}{\Omega_s}\eb\btimes \grad_\bb{R}R
\end{align}
\end{subequations}
are respectively ordered $\rho_\ast$ and $\rho_\ast^2$ compared to the magnetic drifts, and so are not included in equation~\eqref{eqn:GKE}, while the terms dealing with perpendicular and parallel flow shear are retained.
In order to close equation~\eqref{eqn:GKE}, $\varphi$ must be determined. This is done using the quasineutrality equation
\begin{equation}\label{eqn:quasi}
\sum_s Z_s \delta n_s \doteq \sum_s Z_s \int \od^3 v\left(\langle g_s\rangle_\bb{r} + \frac{Z_s e}{T_s} F_s \left(\langle\langle \varphi\rangle_\bb{R}\rangle_\bb{r} - \varphi\right)\right) =0,
\end{equation}
where $\delta n_s$ is the perturbed gyrocenter density for species $s$, $\langle \cdots \rangle_\bb{r}$ is the gyroaverage taken at constant particle position $\bb{r}$ and the Debye length is taken to be much smaller than the electron gyroradius.
The first order $\delta f$ gyrokinetic equation~\eqref{eqn:GKE} forms the basis of both conventional global gyrokinetics and local flux-tube gyrokinetics, the latter being the $(k_\perp L)^{-1} \rightarrow 0$ limit of the former~\citep{Parra_globallocal}. To be more precise, for global gyrokinetics the geometrical coefficients and background pressure profiles \dblbrck{the underlined terms in~\eqref{eqn:GKE}} are allowed to vary in all spatial directions. In addition, the gyroaveraging operator may also have spatial dependence. As $(k_\perp L)^{-1} \rightarrow 0$, these terms no longer vary across turbulent eddies and can thus be treated as constants in the directions perpendicular to the magnetic field; the only variation of these quantities that persists is along the magnetic field. Local gyrokinetics typically assumes periodicity in a statistical sense in the perpendicular directions, and thus local flux-tube gyrokinetics is routinely performed using a pseudo-spectral approach on a Fourier basis and field-line-following Clebsch coordinates $\alpha$ and $\psi$, where $\psi$ is a flux label, $\alpha$ labels a fieldline, and $\bb{B} = \grad \alpha \btimes \grad \psi$. The coordinate along the direction of the magnetic field, denoted by $z$, is typically chosen in a way to simplify the underlying equations; some examples are the arc length along the magnetic field line, the toroidal angle $\zeta$ or the poloidal angle $\theta$.
\subsection{Subsidiary expansion of the gyrokinetic equation \label{sec:subsidiary}}
In practice, while local simulations employ the $(k_\perp L)^{-1} \rightarrow 0$ limit, the extent of their radial domains are typically the size of hundreds of gyroradii and thus cannot be considered `small' when compared to typical device parameters; it is this largeness of the radial box size that we exploit in order to incorporate profile variation in local flux-tube gyrokinetics. Defining the parameter $\Delta \doteq \ell_x/L$, where $\ell_x$ is the radial extent of a simulation, we now impose a subsidiary expansion on the gyrokinetic equation~\eqref{eqn:GKE} using the ordering
\begin{equation}\label{eqn:subsidiary}
\epsilon \sim \frac{1}{k_\perp L} \ll \Delta \ll 1.
\end{equation}
This ordering allows us to include the effects of profile variation without needing to go to higher order in $\rho_\ast$.\footnote{\label{footnote:rhostar}It is worthwhile to point out that most global gyrokinetic codes do not retain all higher order terms in $\rho_\ast$, e.g., diamagnetic flows~\citep{Lee_NF, Lee_POP, Lee_PPCF, Barnes_diamagnetic} and the parallel nonlinearity~\citep{Kniep_2004, Lin_parallel,McDevitt_2009, Candy_parallel,Barnes_parallel}}
In this manuscript, we focus on the radial profile variation found in axisymmetric geometry which is relevant to tokamaks. We defer the treatment of the `full flux surface' version of \texttt{stella}, which includes azimuthal variation found in three-dimensional systems relevant to stellarators, to a later publication. For tokamak geometry and equilibria, axisymmetry implies that, while we only consider profile variation along the magnetic field and across flux surfaces, radially global \texttt{stella} is also toroidally global, without needing any further consideration of the binormal direction $\eb \btimes \grad \psi$ (the direction perpendicular to the magnetic field lying within a flux surface). In order to make \texttt{stella} fully global, one must also include higher order effects in the parallel physics, such as the parallel nonlinearity (see references in footnote~\ref{footnote:rhostar}) and higher order parallel derivatives in the particle drifts~\citep{Sung_2013}; these extra terms are not considered in this manuscript.
To perform the subsidiary expansion given in~\eqref{eqn:subsidiary}, we first normalize the gyrokinetic equation to render it dimensionless. In general there are two ways to do this: have the normalizing quantities vary across flux surfaces or have them fixed to a reference surface at some radial position $r_0$. While the former is the most economical in terms of the velocity space resolution required, it introduces additional derivatives to the gyrokinetic equation (including with respect to $\mu_s$) as well as complicating the multiple flux-tube boundary conditions given in \S \ref{sec:radBC}. The \texttt{stella} code is efficiently parallelized over velocity space, allowing for large velocity-space grids, and so we opt for normalization at a fixed reference position.
Denoting reference values using a subscript `ref' and normalized quantities with a tilde, the normalized quantities are $\tilde{\bb{v}} = \bb{v} / v_{\mr{th}s}(\ensuremath{r}_0)$, $\tilde{v}_{\parallel} = v_\parallel/v_{\mr{th}s}(\ensuremath{r}_0)$, $\tilde{\mu}_s = \mu_s B_\mr{ref}/m_s v_{\mr{th}s}^2(\ensuremath{r}_0)$, $\tilde{B}= B/B_\mr{ref}$, $\tilde{m}_s = m_s/m_\mr{ref}$, $\tilde{n}_s = n_s/n_\mr{ref}$, $\tilde{T}_s = T_s/T_\mr{ref}$, $\tilde{I} = RB_\mr{T}/aB_\mr{ref}$, and $\tilde{\Omega}_\zeta = a \Omega_\zeta/v_\mr{th,ref}$. Here, $v_{\mr{th}s} = \sqrt{2 T_s(\ensuremath{r}) / m_s}$.
We also normalize time to $a/ v_{\mathrm{th,ref}}$, parallel lengths to $a$ and perpendicular lengths to $\rho_\mr{ref}$, where $a$ is the minor radius and $v_{\mathrm{th,ref}}=\sqrt{2T_\mr{ref}/m_\mr{ref}}$, $\rho_\mr{ref}= v_\mr{th,ref}/\Omega_\mr{ref}$ and $\Omega_\mr{ref} = eB_\mr{ref}/m_\mr{ref}c$.
Finally, we introduce the normalized distribution function and electrostatic potential $\tilde{g}_s = (a/\rho_\mr{ref})[g_s/F_s(\ensuremath{r}_0)]\exp[-v^2(\ensuremath{r}_0)/v_{\mr{th}s}^2(\ensuremath{r}_0)]$ and $\tilde{\varphi} = (e\varphi/T_\mr{ref})(a/\rho_\mr{ref})$, where $v^2(\ensuremath{r}) = v_\parallel^2 + 2B(\ensuremath{r}) \mu_s/m_s$. The normalized gyrokinetic equation is then
\begin{align}\label{eqn:gk_norm}
\frac{\partial \tilde{g}_s}{\partial \tilde{t}} &+ \frac{R}{a}\tilde{\Omega}_\zeta \bb{\hat{\zeta}}\bcdot \tilde{\grad} \tilde{g}_s + \frac{v_{\mr{th}s}(\ensuremath{r}_0)}{v_{\mr{th,ref}}}\tilde{v}_\parallel \eb \bcdot \tilde{\grad} \tilde{z} \left(\frac{\partial \tilde{g}_s}{\partial \tilde{z}} + \frac{\partial \langle\tilde{\varphi} \rangle_\bb{R} }{\partial \tilde{z}} \frac{Z_s}{\tilde{T}_s(\ensuremath{r})}\frac{F_s(\ensuremath{r})}{F_s(\ensuremath{r}_0)} \rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)} \right) - \frac{v_{\mr{th}s}(\ensuremath{r}_0)}{v_\mr{th,ref}}\tilde{\mu}_s \eb \bcdot \tilde{\grad} \tilde{B} \frac{\partial \tilde{g}_s}{\partial \tilde{v}_\parallel} \nonumber
\\ & + \tilde{\bb{v}}_{\mr{M}s}\bcdot \left(\tilde{\grad}_\perp \tilde{g}_s +\tilde{\grad}_\perp \langle\tilde{\varphi} \rangle_\bb{R}\frac{Z_s}{\tilde{T}_s(\ensuremath{r})}\frac{F_s(\ensuremath{r})}{F_s(\ensuremath{r}_0)}\rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)} \right)+ \langle\tilde{\bb{v}}_\bb{E} \rangle_\bb{R} \bcdot \tilde{\grad}_\perp \tilde{g}_s +\frac{\rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)} }{F_s(\ensuremath{r}_0)} \langle\tilde{\bb{v}}_\bb{E} \rangle_\bb{R} \bcdot a\grad\big|_E F_s\nonumber \\
& + {\frac{\tilde{m}_s }{\tilde{T}_s(\ensuremath{r})} \frac{v_{\mr{th}s}(\ensuremath{r}_0)}{v_\mr{th,ref}}\frac{F_s(\ensuremath{r})}{F_s(\ensuremath{r}_0)}\rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)}\frac{\tilde{I}\tilde{v}_\parallel}{\tilde{B}}\langle\tilde{\bb{v}}_\bb{E} \rangle_\bb{R} \bcdot \tilde{\grad}_\perp} \tilde{\Omega}_\zeta = 0,
\end{align}
where
\begin{subequations}
\begin{align}
\tilde{\bb{v}}_{\mr{M}s} &= \frac{T_s(\ensuremath{r}_0)}{Z_sT_\mr{r}}\frac{1}{\tilde{B}(\ensuremath{r})} \eb\btimes \left(\tilde{\mu}_s\tilde{\grad} \tilde{B} + \tilde{v}_\parallel^2 \tilde{\bb{\kappa}}\right), \\
\tilde{\bb{v}}_{\bb{E}} &= \frac{1}{2} \frac{1}{\tilde{B}(\ensuremath{r})}\eb \btimes \tilde{\grad}_\perp \tilde{\varphi}.
\end{align}
\end{subequations}
The quasineutrality equation is also normalized,
\begin{equation}\label{eqn:quasi_norm}
\sum_s Z_s \tilde{n}_s(\ensuremath{r}_0) \int \od^3 \tilde{v}\frac{1}{\upi^{3/2}}\left(\langle \tilde{g}_s\rangle_\bb{r} + \frac{Z_s }{\tilde{T}_s(\ensuremath{r})} \frac{F_s(\ensuremath{r})}{F_s(\ensuremath{r}_0)} \rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)} \left(\langle\langle \tilde{\varphi}\rangle_\bb{R}\rangle_\bb{r} - \tilde{\varphi}\right)\right) =0.
\end{equation}
\subsubsection{Radial variation of pressure and magnetic geometry \label{sec:var_outline}}
The expansion in $\Delta$ outlined in \S \ref{sec:GKE_delta} allows us to incorporate variation of the equilibrium pressure profiles into global \texttt{stella} by considering their Taylor expansions around a central radial location $r_0$:
\begin{subequations}
\begin{align}
n_s(r) = n_s(r_0) + \left. \frac{\od n_s}{\od r}\right|_{r=r_0}(r-r_0) +
\frac{1}{2}\left. \frac{\od^2 n_s}{\od r^2}\right|_{r=r_0}(r-r_0)^2 + \ldots , \\
T_s(r) = n_s(r_0) + \left. \frac{\od T_s}{\od r}\right|_{r=r_0}(r-r_0) +
\frac{1}{2}\left. \frac{\od^2 T_s}{\od r^2}\right|_{r=r_0}(r-r_0)^2 + \ldots .
\end{align}
\end{subequations}
The inclusion of the new terms $n_s''(r)$ and $T''_s(r)$, where primes denote differentiation with respect to $\ensuremath{r}$, modifies the source of free energy resulting from the gradients of the background distribution function $\left.\grad\right|_EF_s$. Additionally, the expansion in $\Delta$ also results in $n'$ and $T'$ modifying both the magnetic drift and parallel streaming terms through the adiabatic contribution $(Z_s/T_s)F_s \varphi$ of the perturbed distribution function. These first derivatives also introduce profile variation of the quasineutrality equation~\eqref{eqn:quasi}.
While the density profile is specified by $n_s(\ensuremath{r}_0)$,
$n'_s(\ensuremath{r}_0)$, $n''_s(\ensuremath{r}_0)$ (similarly for the temperature profiles), more relevant in determining growth rates and turbulent amplitudes are the gradient scale lengths $L_{n_s}^{-1} = - n'_s/n_s$ and $L_{T_s}^{-1} = - T'_s/T_s$. While global \texttt{stella} utilizes the second derivatives of the density and temperature, these are not required for profile variation of the gradient scale lengths:
\begin{equation}
\left(L_n^{-1}\right)' =\frac{n'^2_s}{n^2_s} -\frac{n''_s}{n_s},
\end{equation}
likewise for temperature, and so profile variation of the growth rates and turbulent amplitudes is expected, even with $n_s'' = T_s'' = 0$.
Along with radial variation of the background pressure profiles, global \texttt{stella} also includes variation of the magnetic geometry through the use of an extended Miller equilibrium model~\citep{miller}. The original Miller equilibrium parameterises a single axisymmetric flux surface using the model equations
\begin{subequations}\label{eqn:miller}
\begin{align}
R(r,\theta) &= R_0(r) + r \cos (\theta + \sin \theta \arcsin \delta (r)), \\
Z(r,\theta) &= \kappa(r) r \sin (\theta),
\end{align}
\end{subequations}
where $R$ and $Z$ are the radial and vertical position in cylindrical coordinates, $r$ is the flux surface label, $\theta$ is a poloidal coordinate, $R_0$ may include the radial displacement of a given flux surface (Shafranov shift), and $\delta(r)$ and $\kappa(r)$ set the triangularity and elongation of the flux surface, respectively. In order to calculate the geometrical coefficients of~\eqref{eqn:GKE}, one must first locally solve the Grad-Shafranov equation. This requires specification of the derivatives $R'_0(r)$, $\delta'(r)$, and $\kappa'(r)$, as well as specification of the safety factor $q$ \dblbrck{defined in \eqref{eqn:app_q}}, the normalized plasma pressure $\beta= 4 \upi p/B_\mr{ref}^2$, and their derivatives. Here, $p$ is the species-summed plasma pressure.
To extend the Miller equilibrium model to a global profile, one must ensure that the Grad-Shafranov equation is consistently satisfied over the entire profile; it is not sufficient to evaluate~\eqref{eqn:miller} at different radial locations $r$ and then to use these coefficients to individually solve the Grad-Shafranov equation at these locations. Global \texttt{stella} instead solves the Grad-Shafranov equation localized at some specified central radius $r=r_0$, and then also its first radial derivative centered at the same location. This ensures that the geometrical coefficients evolve across the radial domain in a way consistent with MHD equilibria and with the subsidiary expansion in $\Delta$ outlined in \S \ref{sec:GKE_delta}. This approach requires the specification of the second radial derivatives $q''(r)$, $\psi''(r)$ and $\beta''(r)$. Like with the density and the temperature, a non-zero value $q''(r)$ is not required for profile variation in the shearing parameter $\hat{s} \doteq (r/q) q'(r)$:
\begin{equation}
\hat{s}' = \frac{r q''}{q} + \frac{q'}{q} - \frac{r q'^2}{q^2}.
\end{equation}
The coefficients of~\eqref{eqn:gk_norm} that depend on $\ensuremath{r}$ are now Taylor expanded up to first order, and the resulting equations are given in \S \ref{sec:sim_equations}. Detailed calculations of the geometrical coefficients and their derivatives are given in \ref{app:mag_geo}.
\subsection{Coordinates and the parallel boundary condition}\label{sec:coords}
The coordinates used in \texttt{stella} for axisymmetric systems are $(x,y,\theta,v_\parallel,\mu_s)$, where $\theta$ is a poloidal coordinate which signifies the position along a magnetic field line (as used in the Miller equilibrium) and is zero at the outboard midplane.\footnote{The reader can refer to~\citet{Barnes_stella} for the use of \texttt{stella} in three-dimensional systems.} The $(x,y)$ coordinates are in the plane perpendicular to $\bb{B}$, with the binormal coordinate $y$ being defined in the same way for both the local and global versions of \texttt{stella}:
\begin{equation}\label{eqn:ycoord}
y = \frac{\psi'(\rcoord_0)}{B_\mr{ref}}(\alpha-\alpha_0),
\end{equation}
where $\psi$ is the poloidal flux function, $B_\mr{ref}$ is a reference magnetic field, $\alpha = \zeta - q \vartheta$, $\zeta$ is the toroidal angle and $\vartheta$ is the straight-field-line poloidal angle, given by~\eqref{eqn:vartheta_app}.
Currently, \texttt{stella} offers two options for the choice of the radial coordinate $x$, the first being the poloidal flux function $\psi$,
\begin{equation}\label{eqn:xcoord_loc}
x = \frac{q_0}{r_0 B_\mr{ref}}(\psi-\psi_0),
\end{equation}
while the second is the safety factor $q$,
\begin{equation}\label{eqn:xcoord_glob}
x = \frac{\psi'(\rcoord_0)}{B_\mr{ref}} (q-q_0).
\end{equation}
For local \texttt{stella}, $\psi$ is used as the radial coordinate by default. Conversely, global \texttt{stella} is made to use $q$ as the radial coordinate in order to simplify the implementation of the parallel boundary condition typically used in flux-tube simulations~\cite{Beer95}. This boundary condition, which is for the $\theta$ dimension,\footnote{While $\theta$ is used for the parallel coordinate, $\vartheta$ is the variable that appears in the definition of $\alpha$. The angles $\theta$ and $\vartheta$ can be made to match at the inboard midplane ($\vartheta(\theta = \uppi) = \upi$ and $\vartheta(\theta = -\upi) =-\upi$ ), the location at which the boundary condition is applied in simulation. Therefore, either variable can be used in the parallel boundary condition.} asserts $2\upi N$ periodicity of a quantity $A$ at fixed toroidal angle $\zeta$, rather than for fixed $\alpha$,
\begin{equation}\label{eqn:parBC}
A(\psi, \alpha(\psi,\zeta, \theta = 0), \theta = 0) = A(\psi,\alpha(\psi,\zeta, \theta =2 \upi N ), \theta = 2 \upi N).
\end{equation}
Here, $N$ is an integer indicating the number of poloidal turns that is chosen as an input parameter for simulation.
To illustrate the difficulty of applying this boundary condition in the general case of arbitrary $q$ profile, we first consider the local case where $\psi$ is used as the radial coordinate. Then, with $\alpha = \zeta - q(\psi) \vartheta$,~\eqref{eqn:parBC} can be expressed in Fourier space as
\begin{equation}
\sum_{k_\psi, k_\alpha} A_\bb{k}(\theta = 0) \rme^{\imag k_\psi (\psi-\psi_0) + \imag k_\alpha (\alpha - \alpha_0)} = \sum_{k_\psi, k_\alpha} A_\bb{k}(\theta= 2\upi)\rme^{\imag k_\psi (\psi-\psi_0) + \imag k_\alpha (\alpha - \alpha_0) - 2\upi N \imag k_\alpha [q_0 + (\od q / \od \psi)(\psi-\psi_0) + (\od^2 q/ \od \psi^2) (\psi-\psi_0)^2/2 + \ldots]}.
\end{equation}
In the flux-tube limit, only the first two terms in the Taylor expansion of $q(\psi) \approx q_0 + \od q/ \od \psi (\psi-\psi_0)$ are kept, and so the parallel boundary condition amounts to matching a quantity at either end of the $\vartheta$ domain at different radial wavenumber $k_\psi$,
\begin{equation}\label{eqn:twistshift_psi}
A_{k_\psi, k_\alpha}(\theta = 0) = C_k A_{k_\psi + \upDelta k_\psi, k_\alpha}(\theta= 2\upi),
\end{equation}
where $C_k = \exp(-2\upi N \imag k_\alpha q_0 ) $ and $\upDelta k_\psi = 2\upi N k_\alpha (\od q / \od \psi)$. Equation~\eqref{eqn:twistshift_psi} allows for an efficient and straightforward implicit treatment of the parallel streaming term, which otherwise can set a stringent constraint on the simulation time step. In practice, as the phase-shift due to $C_k$ is of order $\rho_\ast^{-1}$, a $\rho_\ast$-small adjustment in the radial position of the simulation domain allows for $C_k = 1$; this is equivalent to adjusting the positions of the mode rational surfaces that are introduced by the parallel boundary condition~\cite{Ball_prl,Ajay2020}.
One must keep more terms in the Taylor expansion of $q(\psi)$ for global simulations, and as a result the parallel boundary condition cannot be expressed using a simple wavenumber shift. In general, the quantity would have to be inverse-Fourier transformed to $(x,k_y)$ space, have an $x$-dependent phase shift in the $y$ direction applied, and then Fourier transformed back into $(k_x,k_y)$ space, thus coupling all the $k_x$ modes at the $\theta$ boundaries. This complication can be avoided altogether if $q$ is chosen as the radial coordinate, rather than $\psi$. In this case,~\eqref{eqn:parBC} can be expressed in Fourier space as
\begin{equation}
\sum_{k_q, k_\alpha} A_\bb{k}(\theta = 0) \rme^{\imag k_q (q-q_0) + \imag k_\alpha (\alpha - \alpha_0)} = \sum_{k_q, k_\alpha} A_\bb{k}(\theta= 2\upi)\rme^{\imag (k_q - 2 \upi N k_\alpha)(q-q_0) + \imag k_\alpha (\alpha - \alpha_0) - 2\upi N \imag k_\alpha q_0 },
\end{equation}
and so now
\begin{equation}\label{eqn:twistshift_q}
A_{k_q, k_\alpha}(\theta = 0) = C_k A_{k_q + \upDelta k_q, k_\alpha}(\theta= 2\upi),
\end{equation}
where $C_k = \exp(-2\upi N \imag k_\alpha q_0 ) $ and $\upDelta k_q = 2\upi N k_\alpha $. This, however, comes at the cost of only allowing global simulations of systems with a monotonic $q$ profile. We thus choose $q$ as the radial coordinate for global \texttt{stella}, though alternative formulations of the parallel boundary condition will be investigated in future work.
For global simulations, the radial coordinate $x$ (equivalently, $q$) can be related to the physical location $r$ using
\begin{align}
q-q_0 =\tilde{x}\rho_\ast\frac{aB_\mr{ref}}{\psi'} = q'(\rcoord_0) (\ensuremath{r} - \ensuremath{r}_0) + \frac{q''(\rcoord_0)}{2} (\ensuremath{r}-\ensuremath{r}_0)^2 + \cdots
\end{align}
where $\tilde{x}\doteq x/\rho_\mr{ref}$ and~\eqref{eqn:xcoord_glob} has been used for the first equality.
Defining $\upDelta \ensuremath{r} = \ensuremath{r} - \ensuremath{r}_0$ and keeping terms up to order $\upDelta \ensuremath{r}^2$, the physical location in terms of $x$ is given by the positive root of the quadratic equation,
\begin{align}\label{eqn:grid_quad}
\upDelta \ensuremath{r} = \frac{q'}{q''}\left[-1 + \left(1 + 2 \tilde{x} \rho_\ast \frac{q''}{q'^2}\frac{aB_\mr{ref}}{ \psi'} \right)^{1/2}\right].
\end{align}
These results are readily generalized if $\psi$ is used as the radial coordinate instead.
The extent of the radial domain in $q$ must be chosen to ensure $\ensuremath{r}/a \in (0,1)$. Note that for $q'' \ne 0$, a grid centered in $q$ around $q_0$ will not be centered in $\ensuremath{r}$ around $\ensuremath{r}_0$. By default, global \texttt{stella} constructs the radial grid to be centered in $\ensuremath{r}$, which leads to the two constraint equations
\begin{align*}
\frac{q''}{2q'} \upDelta \ensuremath{r}_+^2 - \upDelta \ensuremath{r}_+ - \tilde{x}_-\rho_\ast \frac{aB_\mr{ref}}{q'\psi'} &= 0, \\
\frac{q''}{2q'} \upDelta \ensuremath{r}_+^2 + \upDelta \ensuremath{r}_+ - \tilde{x}_+\rho_\ast \frac{aB_\mr{ref}}{q'\psi'} &= 0,
\end{align*}
where $\tilde{x}_\pm$ are the left and rightmost values of the $\tilde{x}$ grid, $\tilde{x}_+ = \tilde{x}_- + \tilde{\ell}_x $, $\tilde{\ell}_x\doteq \ell_x/\rho_\mr{ref}$, and $\upDelta r_+ = \ensuremath{r}_+ - \ensuremath{r}_0$ where $\ensuremath{r}_+$ is the rightmost point of the radial grid. Solving this system for $\upDelta \ensuremath{r}$ and $\tilde{x}_-$ gives
\begin{subequations}
\begin{align}\label{drho}
\upDelta \ensuremath{r}_+ &= \frac{ \tilde{\ell}_x\rho_\ast}{2} \frac{aB_\mr{ref}}{q'\psi'}, \\
\tilde{x}_- &= -\frac{\tilde{\ell}_x}{2} \left( 1- \frac{ \tilde{\ell}_x\rho_\ast}{4} \frac{q''}{q'^2}\frac{aB_\mr{ref}}{\psi'}\right).
\end{align}
\end{subequations}
Thus, the physical radial box size $\ell_\ensuremath{r} = 2\upDelta r_+ =(aB_\mr{ref}/q'\psi') \tilde{\ell}_x\rho_\ast$ is determined by the input parameters $\rho_\ast$, $\tilde{\ell}_x$ and $q'$, along with $\psi'$ calculated from the Miller geometry (\ref{app:mag_geo}).
If $q'' = 0$, then $\tilde{x}_- =- \tilde{\ell}_x/2$ and the grid becomes centered in $x$ (or $q$) as well.
\subsection{Radial boundary conditions}\label{sec:radBC}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{plots/method4}
\caption{Illustration of the novel radial boundary condition which employs multiple flux-tube simulations. The simulations on the left and right, which are local simulations, are performed concurrently to and indepedent of the central `global' simulation. At every timestep, information of the distribution function from the side domains is directly copied into analogous regions in the central domain. This transfer of information is done in specified boundary regions. }
\label{fig:radial_BC}
\end{figure}
The global version of \texttt{stella} employs a novel set of radial boundary conditions which feeds information generated from independent local flux-tube simulations into a central `global' simulation. The aim of these boundary conditions is two-fold: firstly, we replace the Dirichlet boundary condition typically used in global simulations with a more physically motivated forcing. By doing so, particle, momentum, and heat fluxes, which are self-consistently generated by local simulation, can be specified at the boundaries in a statistical sense. The second aim of our novel boundary conditions is that by feeding in `good' information at the radial boundaries, the Gibbs phenomena that typically arise near a discontinuity when using a Fourier spectral approach is largely mitigated.
The novel radial boundary condition, which is illustrated in figure~\ref{fig:radial_BC}, works as follows: a boundary region is specified for each flux-tube domain with a given number of radial collocation points $N_\mr{boundary}$ and radial width $\tilde{L}_\mr{boundary}$; the domains of the local simulations each contain a single instance of this boundary, while the central domain contains a boundary on each radial end. Then, at every timestep (or, alternatively, every Runge-Kutta explicit and implicit substep), information from the distribution functions in the left and right domains ($g_{s, \mr{L}}$ and $g_{s,\mr{R}}$, respectively) is copied directly into the boundary region of the distribution function of the central domain:
\begin{subequations}\label{eqn:radial_BC}
\begin{align}
\tilde{g}_{s}(\tilde{x} + \tilde{x}_-) &= g_{s,\mr{L}}(\tilde{x} + \tilde{x}_{-,\mr{L}}), \\
\tilde{g}_{s}(\tilde{x}+ \tilde{x}_+ - \tilde{L}_\mr{boundary}) &= g_{s,\mr{R}}(\tilde{x}+ \tilde{x}_{+,\mr{R}} - \tilde{L}_\mr{boundary}),
\end{align}
\end{subequations}
for $0 < \tilde{x} < \tilde{L}_\mr{boundary}$. The locations of the boundary regions here are chosen to be consistent with the parallel boundary condition: while local simulations employing the parallel boundary condition are periodic in the radial direction, they are not homogeneous due to the appearance of mode rational surfaces at distinct radial locations. We stress here that these radial locations remain equally spaced due to our choice of $q$ as the radial coordinate, regardless of the profile of the magnetic safety factor. The placement of the boundary regions outlined in~\eqref{eqn:radial_BC} and illustrated in figure~\ref{fig:radial_BC} ensures that these mode rational surfaces line up correctly when the parallel-boundary phase shift $C_k$ is chosen to be consistent in every domain (see \S \ref{sec:coords}). Intuitively, the size of the $\tilde{L}_\mr{boundary}$ must be larger than the largest Larmor radius contained in the simulation in order for information not to leak through the edge of the radial simulation domain. Numerical tests show that physical observables such as the heat flux converge once the boundary layer size meets this criteria, and that numerical instabilities may appear if the region is too small.
The two auxiliary local simulations are seeded with different initial conditions than the central domain and all three are run concurrently with the same timestep (which may be adjusted during the simulation based on the CFL constraint). Currently, in order to avoid interpolation when mapping onto the central domain, the left and right domains use the same grid resolution and spacing in the parallel and binormal coordinates, and also share the same velocity space grid. While the grid spacing in the radial direction must also match, the left and right simulation domain are free to have different radial extents. While this set-up effectively triples the computational cost of the equivalent local simulation, it does so in an embarrassingly parallel way; a fact of which the parallelism of~\texttt{stella} readily takes advantage.
\section{Numerical Implementation \label{sec:numerical_imp}}
\subsection{Simulation equations \label{sec:sim_equations}}
Global \texttt{stella} uses a pseudospectral approach to integrate the gyrokinetic equation. Here, the linear terms with constant coefficients are evaluated in Fourier space, while nonlinear terms and terms with radial profile variation are evaluated in real space. The Fourier transformed gyrokinetic equation, including the radial variation terms resulting from the Taylor expansion outlined in \S \ref{sec:subsidiary}, is
\begin{align}\label{eqn:GKE_FT}
\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t} &+ \imag k_y\tilde{\Omega}_\zeta \tilde{g}_{\bb{k},s} + \frac{v_{\mr{th}s}(\rcoord_0)}{v_{\mr{th,ref}}} \tilde{v}_\parallel {\eb \bcdot \tilde{\grad} \tilde{z}} \Bigg(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial \tilde{z}} + {\frac{Z_s}{\tilde{T}_s(\rcoord_0)} } \frac{\partial J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{\varphi}_{\bb{k}} }{\partial \tilde{z}}\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \Bigg) - \frac{v_{\mr{th}s}(\rcoord_0)}{v_{\mr{th,ref}}} \tilde{\mu}_s{\eb \bcdot \tilde{\grad} \tilde{B}} \frac{\partial \tilde{g}_{\bb{k},s}}{\partial \tilde{v}_\parallel} \nonumber\\
&+ \imag \tilde{\omega}_{\mr{D},\bb{k},s}\left( \tilde{g}_{\bb{k},s} +{\frac{Z_s}{\tilde{T}_s(\rcoord_0)}} J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{\varphi}_{\bb{k}}\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \right)
+ (\imag\tilde{\omega}_{\ast,\bb{k},s} + \imag\tilde{\omega}_{\zeta,\bb{k},s})J_0(a_{\bb{k},s})\tilde{\varphi}_\bb{k} + \mathcal{N}_{\bb{k},s} \nonumber\\
= &- \frac{v_{\mr{th}s}(\rcoord_0)}{v_{\mr{th,ref}}} \tilde{v}_\parallel \left({\eb \bcdot \tilde{\grad} \tilde{z}}\right)' \ensuremath{\hat{\mathcal{X}}} \Bigg(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial \tilde{z}} + {\frac{Z_s}{\tilde{T}_s(\rcoord_0)} } \frac{\partial J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{\varphi}_{\bb{k}} }{\partial \tilde{z}}\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \Bigg) + \frac{v_{\mr{th}s}(\rcoord_0)}{v_{\mr{th,ref}}} \tilde{\mu}_s\left(\eb \bcdot \tilde{\grad} \tilde{B}\right)' \ensuremath{\hat{\mathcal{X}}} \frac{\partial \tilde{g}_{\bb{k},s}}{\partial \tilde{v}_\parallel} \nonumber \\
& - \frac{v_{\mr{th}s}(\rcoord_0)}{v_{\mr{th,ref}}} \tilde{v}_\parallel {\eb \bcdot \tilde{\grad} \tilde{z}} \ensuremath{\hat{\mathcal{X}}} \Bigg[ \left(\frac{F_{s}(\ensuremath{r})}{\tilde{T}_s}\right)'\frac{\partial J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{\varphi}_{\bb{k}} }{\partial \tilde{z}} \frac{Z_s\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)}}{F_s(\ensuremath{r}_0)} + {\frac{Z_s}{\tilde{T}_s(\rcoord_0)} } \frac{\partial J'_0(a_{\bb{k},s}) \tilde{\varphi}_{\bb{k}} }{\partial \tilde{z}}\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \Bigg] \nonumber\\
&- \ensuremath{\hat{\mathcal{X}}} \left[\left( \imag\tilde{\omega}_{\mr{D},\bb{k},s}\right)'\left( \tilde{g}_{\bb{k},s} +{\frac{Z_s}{\tilde{T}_s(\rcoord_0)}} J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{\varphi}_{\bb{k}}\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \right)\right]
- \ensuremath{\hat{\mathcal{X}}}\left[\left( ( \imag \tilde{\omega}_{\ast,\bb{k},s} + \imag \tilde{\omega}_{\zeta,\bb{k},s})J_0(a_{\bb{k},s})\right)'\tilde{\varphi}_\bb{k} \right]\nonumber \\
&- \ensuremath{\hat{\mathcal{X}}}\Bigg[ \left(\frac{F_{s}(\ensuremath{r})}{\tilde{T}_s}\right)' \imag \tilde{\omega}_{\mr{D},\bb{k},s} J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{\varphi}_{\bb{k}} \frac{Z_s\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)}}{F_s(\ensuremath{r}_0)} \Bigg]
- \imag \tilde{\omega}_{\mr{D},\bb{k},s}\ensuremath{\hat{\mathcal{X}}}\Bigg( {\frac{Z_s}{\tilde{T}_s(\rcoord_0)} } J'_0(a_{\bb{k},s}) \tilde{\varphi}_{\bb{k}} \rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \Bigg) + \left(\mathcal{N}_{\bb{k},s}\right)',
\end{align}
where all quantities that depend on $\ensuremath{r}$ have been evaluated at $\ensuremath{r}_0$.
Here, $a_{\bb{k},s} =(\tilde{m}_s \tilde{T}_s/Z_s^2)^{1/2} \tilde{k}_\perp \tilde{v}_\perp / \tilde{B}$, $ \tilde{k}_\perp^2 = k_x^2\rho_\mr{ref}^2 |\grad x|^2 + 2 k_x k_y \rho_\mr{ref}^2(\grad x \bcdot \grad y) + k^2_y\rho_\mr{ref}^2 |\tilde{\grad} y|^2$,
\begin{subequations}
\begin{align}
\tilde{\omega}_{\mr{D},\bb{k},s} &= \frac{a\tilde{T}_s(\ensuremath{r}_0)}{Z_s \tilde{B}(\ensuremath{r})}\left(\tilde{v}_\parallel^2 \eb \btimes \bb{\kappa} + \tilde{\mu}_s \eb \btimes \grad \tilde{B}\right)\bcdot\left(k_y \rho_\mr{ref} \grad y + k_x \rho_\mr{ref}\grad x\right),\\
\tilde{\omega}_{\ast,\bb{k},s} &= \frac{k_y \rho_\mr{ref}}{2} \frac{aB_\mr{ref}}{\psi'} \frac{\od y}{\od \alpha}\frac{\rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)}}{F_s(\ensuremath{r}_0)}\left.\frac{\od F_s(\ensuremath{r})}{\od \ensuremath{r}}\right|_E,\\
\tilde{\omega}_{\zeta,\bb{k},s} &= -k_y\rho_\mr{ref}\sqrt{\frac{\tilde{m}_s}{\tilde{T}_s}} \frac{q a}{r}\frac{\tilde{I}}{\tilde{B}} \tilde{v}_\parallel \rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \tilde{\gamma}_E,
\end{align}
\end{subequations}
$\tilde{\gamma}_E = (\ensuremath{r}/q)(\od \Omega_\zeta /\od \ensuremath{r}) (a/v_\mr{th,ref})$,
and the nonlinear term for when $q$ is the radial coordinate is given by
\begin{equation}
\mathcal{N}_{\bb{k},s} = \frac{B_\mr{ref}}{2}\frac{\od y }{\od \alpha}\frac{\od x}{\od q}\frac{q'}{\psi'}\ensuremath{\mathcal{F}_\bb{k}}\left[\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_y \rho_\mr{ref}J_0(a_{\bb{k},s})\tilde{\varphi}_\bb{k}\right)\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_x \rho_\mr{ref}\tilde{g}_{\bb{k},s}\right) - \ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_x \rho_\mr{ref}J_0(a_{\bb{k},s})\tilde{\varphi}_\bb{k}\right)\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_y \rho_\mr{ref}\tilde{g}_{\bb{k},s}\right)\right],
\end{equation}
where $\ensuremath{\mathcal{F}_\bb{k}}$ and $\ensuremath{\mathcal{F}_\bb{k}}^{-1}$ denote the forward and inverse Fourier transforms, respectively.
The terms dealing with radial profile variation are denoted using a prime; here,
\begin{subequations}
\begin{align}
\left(\frac{F_{s}(\ensuremath{r})}{\tilde{T}_s}\right)' &= \frac{F_{s}}{\tilde{T}_s} \left[ \frac{n'_s}{n_s} + \frac{T'_s}{T_s}\left(\frac{E}{T_s} - \frac{5}{2}\right)- \frac{\mu_s B'}{T_s}\right], \\
\left(\frac{1}{\psi' } \left. \frac{\od F_{s}(\ensuremath{r})}{\od \ensuremath{r}} \right|_E \right)' &= \frac{1}{\psi'} \bigg\{ \frac{n''_s}{n_s} - \frac{n'^2_s}{n_s^2} + \left(\frac{T''_s}{T_s} - \frac{T'^2_s}{T_s^2}\right)\left(\frac{ E}{T_s}-\frac{3}{2}\right)- \frac{T'^2_s}{T_s^2}\frac{E}{T_s} + \frac{T'_s}{T_s} \frac{\mu_s B'}{T_s} \nonumber \\
&\quad+ \left[\frac{n'_s}{n_s} + \frac{T'_s }{T_s}\left(\frac{ E}{T_s}-\frac{3}{2}\right) \right]\left[\frac{n'_s}{n_s}+ \frac{T'_s}{T_s}\left(\frac{ E}{T_s}-\frac{3}{2}\right) - \frac{\mu_s B'}{T_s}- \frac{\psi''}{\psi'}\right]\bigg\}
F_{s}, \\
\left(\mathcal{N}_{\bb{k},s}\right)' &=\ensuremath{\hat{\mathcal{X}}} \left( \frac{q''}{q'} - \frac{\psi''}{\psi'}\right) \mathcal{N}_{\bb{k},s} + \frac{B_\mr{ref}}{2}\frac{\od y }{\od \alpha}\frac{\od x}{\od q}\frac{q'}{\psi'}\ensuremath{\mathcal{F}_\bb{k}}\bigg[\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_y \rho_\mr{ref} \ensuremath{\hat{\mathcal{X}}} J'_0(a_{\bb{k},s})\tilde{\varphi}_\bb{k}\right)\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_x \rho_\mr{ref}\tilde{g}_{\bb{k},s}\right) \nonumber\\
&\qquad - \ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_x \rho_\mr{ref} \ensuremath{\hat{\mathcal{X}}} J'_0(a_{\bb{k},s})\tilde{\varphi}_\bb{k}\right)\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(\imag k_y \rho_\mr{ref}\tilde{g}_{\bb{k},s}\right)\bigg],\\
J'_0(a_{\bb{k},s}) &= - J_1(a_{\bb{k},s})a_{\bb{k},s}\left(\frac{({k}_\perp^2)'}{{k}_\perp^2} - \frac{{B}'}{{B}}\right).
\end{align}
\end{subequations}
In these equations, tildes have been suppressed for fractional terms that are dimensionless.
The other radial profile terms dealing with the magnetic geometry (see \S \ref{sec:var_outline}) are computed in \ref{app:mag_geo}. The radial operator $\ensuremath{\hat{\mathcal{X}}}$ applies the $\ensuremath{r}$ profile to all the terms appearing to the right of it, and is defined as
\begin{equation}\label{eqn:spaceOperator}
\ensuremath{\hat{\mathcal{X}}} (\cdots) \doteq \ensuremath{\mathcal{F}_\bb{k}} \left\{ (\ensuremath{r}_\mr{clamped} - \ensuremath{r}_0) \ensuremath{\mathcal{F}_\bb{k}}^{-1}\left\{ \cdots\right\}\right\},
\end{equation}
where
\begin{equation}
\ensuremath{r}_\mr{clamped} \doteq \left\{ \begin{matrix}
\ensuremath{r}(\tilde{x}_- + \tilde{L}_\mr{boundary}) & \tilde{x}_- \le \tilde{x} \le \tilde{x}_- + \tilde{L}_\mr{boundary},\\
\ensuremath{r}(\tilde{x}) & \tilde{x}_- + \tilde{L}_\mr{boundary} \le \tilde{x} \le \tilde{x}_+ - \tilde{L}_\mr{boundary},\\
\ensuremath{r}(\tilde{x}_+ - \tilde{L}_\mr{boundary}) & \tilde{x}_+ - \tilde{L}_\mr{boundary} \le \tilde{x} \le \tilde{x}_+.
\end{matrix} \right.
\end{equation}
We use this definition of $\ensuremath{r}_\mr{clamped}$ in order to render the evolution of $g_s$ within the radial boundary region as consistent as possible between the three simulation domains. Global \texttt{stella} also includes an option to employ the radial boundary conditions suggested by Candy et al.~\citep{Candy_globshear,Candy_globloc} by using a triangle waveform in the definition of~\eqref{eqn:spaceOperator}. Additionally, both approaches can be combined in order to further mitigate Gibbs phenomena in the higher radial derivatives of the distribution function and electrostatic potential. This is done by placing one of the boundary regions of the central domain into the middle of the box.
Adding global effects as a next-order correction in a Taylor expansion provides two advantages over using arbitrary profile variation: Firstly, as all the global terms utilize the same spatial operator $\ensuremath{\hat{\mathcal{X}}}$, it greatly reduces the number of Fourier transforms needed to implement the terms numerically. Secondly, as only quadratic variations of the kinetic and magnetic geometry profiles are added, there are fewer free parameters needed to specify the physical system, and so the problem of 'flux-matching' global simulations to experimental results is somewhat simplified. Future versions of global~\texttt{stella} will aim to allow arbitrary profile variation in both the kinetic and magnetic profiles, though the planned inclusion of electromagnetic effects will make this a more challenging task.
The Fourier-transformed quasineutrality equation is
\begin{align}
\sum_s Z_s & \tilde{n}_s(\ensuremath{r}_0) \frac{2}{\uppi^{1/2}} \int \od \tilde{v}_\parallel \int \od \tilde{\mu}_s \,\tilde{B}\left(J_0(a_{\bb{k},s}(\rcoord_0)) \tilde{g}_{\bb{k},s} + \frac{Z_s }{\tilde{T}_s(\ensuremath{r}_0)} \rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)}\left(J^2_0(a_{\bb{k},s}) -1\right)\tilde{\varphi}_{\bb{k}}\right) \nonumber \\
= &-\ensuremath{\hat{\mathcal{X}}}\sum_s Z_s \tilde{n}_s(\ensuremath{r}_0) \frac{2}{\uppi^{1/2}} \int \od \tilde{v}_\parallel \int \od \tilde{\mu}_s \tilde{B}\Bigg\{\left(J'_0(a_{\bb{k},s}) + \frac{B'}{B}J_0(a_{\bb{k},s}) \right) \tilde{g}_{\bb{k},s} \nonumber
\\ & + \frac{Z_s }{\tilde{T}_s(\ensuremath{r}_0)} \rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)}\left(J^2_0(a_{\bb{k},s}) -1\right) \left[ \frac{n'_s}{n_s} + \frac{T'_s}{T_s}\left(\frac{E}{T_s} - \frac{5}{2}\right) + \frac{B'}{B}\left(1 - 2 \tilde{\mu}_s \tilde{B}\right) - \frac{2J_0(a_{\bb{k},s})J'_0(a_{\bb{k},s})}{1-J^2_0(a_{\bb{k},s})}\right]\tilde{\varphi}_{\bb{k}}\Bigg\}.
\end{align}
Solving quasineutrality requires special considerations, which are given in \S \ref{sec:quasi_solve}.
Some comments on the Fourier transforms in the spatial operator $\ensuremath{\hat{\mathcal{X}}}$ are in order. The inclusion of terms with a linear profile $\ensuremath{r} - \ensuremath{r}_0$ introduces the effect of profile shearing, which results in the advection of the radial wavenumber of a mode in Fourier space~\citep{Candy_globloc}. In the case of a positive advection speed in $k_x$, a mode's radial wavenumber may reach the maximal radial wavenumber resolved in the simulation, $k_{x,\rm{max}}$. If no other steps are taken, it will then unphysically wrap around to the opposite extreme at $k_{x,\rm{min}} = -k_{x,\mr{max}}$. (For negative advection speed in $k_x$, wrap-around from $k_{x,\rm{min}}$ to $k_{x,\rm{max}}$ will instead occur.) This is related to the phenomena of aliasing which could occur when the distribution function $g_{\bb{k},s}$ and electrostatic potential $\varphi_\bb{k}$ are inverse Fourier transformed to real space in order to calculate the nonlinear term $\mathcal{N}_{\bb{k},s}$. To avoid aliasing by the nonlinear term, some form of dealiasing is performed, the conventional method being the `2/3rds approach' ~\citep{Orszag_aliasing}, where $g_{\bb{k},s}$ and $\varphi_\bb{k}$ are inverse Fourier transformed onto a grid with 50\% more collocation points in each Fourier transformed dimension than the number of corresponding Fourier modes. This approach to dealiasing for the nonlinear term can be justified as a spectral cut-off of the distribution function and electrostatic potential that can be enforced with sufficient (hyper)viscosity, and thus their truncation when transforming back to Fourier space should not have any significant effect. However, this approach is unsuitable for the operator $\ensuremath{\hat{\mathcal{X}}}$, as truncation of the discontinuous function $\ensuremath{r} - \ensuremath{r}_0$ introduces Gibbs phenomena which results in numerical artefacts when convolved with another term. Additionally, in our experience dealiasing by spectral cut-off does not prevent wrap-around.
The Fourier transforms that appear in the spatial operator $\ensuremath{\hat{\mathcal{X}}}$ are thus performed without any additional padding, relying only on hyper-dissipation (\S \ref{sec:hyperdissipation}) to prevent radial wavenumber wrap-around due to profile shearing. This approach can be justified by the smallness of profile shearing due to $\ensuremath{\hat{\mathcal{X}}}$: while $(\ensuremath{r} - \ensuremath{r}_0)/a$ is of order $\Delta$, profile shearing depends on its derivative $\od\, (\ensuremath{r} - \ensuremath{r}_0)/\od x$, which is formally of order $\rho_\ast \ll \Delta$. The hyper-dissipative damping rate at the radial wavenumber grid boundaries can then be made to be faster than the rate at which a radial mode advects across of radial wavenumber cell of width $\upDelta k_x = 2 \upi/\ell_x$.
\subsection{Integration scheme \label{sec:int_scheme}}
The \texttt{stella} code, either in local flux-tube or global operation, employs an operator-split time integration scheme that alleviates the timestep constraint caused by the fast parallel electron dynamics. In the local version, this scheme comprises three main steps~\citep{Barnes_stella},
\begin{equation}
\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t} =\left(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t}\right)_1 +\left(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t}\right)_2
+ \left(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t}\right)_3,
\end{equation}
where
\begin{subequations}\label{eqn:timestep}
\begin{align}
\left(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t}\right)_1 &= -\imag \omega_{\mr{D},\bb{k},s}\left(\tilde{g}_{\bb{k},s} + J_0(a_{\bb{k},s}) \tilde{\varphi}_\bb{k}\frac{Z_s}{\tilde{T}_s} \frac{F_s(\ensuremath{r})}{F_s(\ensuremath{r}_0)}\rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)} \right)- \imag \omega_{\ast,\bb{k},s}J_0(a_{\bb{k},s})\tilde{\varphi}_\bb{k} - \mathcal{N}_\bb{k}, \label{eqn:timestep_exp}\\
\left(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t}\right)_2 &= \frac{v_{\mr{th},s}(\ensuremath{r}_0)}{v_{\mr{th,ref}}}\tilde{\mu}_s \eb\bcdot \tilde{\grad} \tilde{B} \frac{\partial \tilde{g}_{\bb{k},s}}{\partial \tilde{v}_\parallel},\label{eqn:timestep_mir}\\
\left(\frac{\partial \tilde{g}_{\bb{k},s}}{\partial t}\right)_3 &= -\frac{v_{\mr{th},s}(\ensuremath{r}_0)}{v_{\mr{th,ref}}}v_\parallel \eb\bcdot \tilde{\grad} z \left( \frac{\partial \tilde{g}_{\bb{k},s}}{\partial z} + \frac{\partial J_0(a_{\bb{k},s}) \tilde{\varphi}_\bb{k}}{\partial z}\frac{Z_s}{\tilde{T}_s} \frac{F_s(\ensuremath{r})}{F_s(\ensuremath{r}_0)}\rme^{-{v^2(\ensuremath{r}_0)}/v_{\mr{th}s}^2(\ensuremath{r}_0)} \right).\label{eqn:timestep_prl}
\end{align}
\end{subequations}
The first step, which includes the magnetic and $\bb{E} \btimes \bb{B}$ drifts, is an explicit step which utilizes a strong stability preserving, third-order Runge-Kutta method. The next two steps, which respectively perform the parallel acceleration and streaming, are done implicitly. Additional implicit steps, such as those performing equilibrium flow shear or applying dissipative operators, can also be included in the time integration scheme.
Global \texttt{stella} incorporates global effects by computing the terms on the right-hand-side of \eqref{eqn:GKE_FT} that account for the radial variation of pressure and magnetic geometry, and including them in the explicit step given by~\eqref{eqn:timestep_exp}. The explicit treatment of the radial corrections is not expected to place a limit on the time step, provided $\Delta$ is not too large; indeed, we have found this to be the case for the ion scale simulations that we have performed. { Finally, while the parallel streaming term~\eqref{eqn:timestep_prl} can in principle be calculated implicitly for global~\texttt{stella}, for reasons detailed in \S \ref{sec:quasi_solve} this term is calculated explicitly by default, as well as for all the numerical benchmarks performed in this paper; future versions of \texttt{stella} will include implicit global algorithms for the parallel dynamics. }
\subsection{Quasineutrality \label{sec:quasi_solve}}
At every integration sub-step that involves the electrostatic potential, the quasineutrality equation~\eqref{eqn:quasi_norm} must be used to recompute $\tilde{\varphi}_\bb{k}$ using the updated value of $\tilde{g}_{\bb{k},s}$. This involves performing a velocity-space integration of the distribution function, and then an inversion of the spectral double-gyro-average operator
\begin{equation}
Q= \sum_s \int \od^3 v \, (Z_s^2 e/T_s) F_s \left(1-J_0^2(a_{\bb{k},s})\right).
\end{equation}
The latter step is trivial in local flux-tube simulations as the operator is a diagonal operator in $\bb{k}$-$\theta$ space. In global operation, the quasineutrality equation including global profile variation can be expressed as
\begin{equation}\label{eqn:quasi_rad}
\Theta \tilde{\varphi}_\bb{k} +\ensuremath{\hat{\mathcal{X}}}(\Theta'\tilde{\varphi}_\bb{k}) = \frac{2}{\upi^{1/2}}\sum_s Z_s \int \od \tilde{v}_\parallel \int \od \tilde{\mu}_s \left\{ \tilde{B}(\ensuremath{r}_0) J_0(a_{\bb{k},s}(\rcoord_0))\tilde{g}_{\bb{k},s} + \ensuremath{\hat{\mathcal{X}}} \left[ \left(\tilde{B}' J_0(a_{\bb{k},s}) + \tilde{B} J'_0(a_{\bb{k},s})\right)_{\ensuremath{r} = \ensuremath{r}_0}\tilde{g}_{\bb{k},s}\right]\right\},
\end{equation}
where
\begin{subequations}
\begin{align}
\Theta &=\sum_s \frac{2}{\pi^{1/2}}\int \od \tilde{v}_\parallel \od \tilde{\mu}_s \, \tilde{B}(\ensuremath{r}_0) Z_s^2 \frac{\tilde{n}_s(\ensuremath{r}_0)}{\tilde{T}_s(\ensuremath{r}_0)}\left(1-J_0^2(a_{\bb{k},s}(\rcoord_0)) \right)\, \rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)},\\
\Theta' &= \sum_s \frac{2}{\pi^{1/2}}\int \od \tilde{v}_\parallel \od \tilde{\mu}_s \, \tilde{B}(\ensuremath{r}_0) Z_s^2 \frac{\tilde{n}_s(\ensuremath{r}_0)}{\tilde{T}_s(\ensuremath{r}_0)}\left(1-J_0^2(a_{\bb{k},s}(\rcoord_0)) \right)\, \rme^{-v^2(\rcoord_0)/v^2_{\mr{th}s}(\rcoord_0)} \times \nonumber
\\ & \hspace{6cm} \left[ \frac{n'_s}{n_s} + \frac{T'_s}{T_s} \left(\frac{E}{T_s} - \frac{5}{2}\right) + \frac{B'}{B}\left(1 - 2 \tilde{\mu}_s \tilde{B}\right) - \frac{2J_0(a_{\bb{k},s})J'_0(a_{\bb{k},s})}{1-J^2_0(a_{\bb{k},s})}\right]_{\ensuremath{r}=\ensuremath{r}_0}.
\end{align}
\end{subequations}
Solving the quasineutrality equation~\eqref{eqn:quasi_rad} requires two steps: first the velocity integration of $\tilde{g}_{\bb{k},s}$ is carried out (which includes taking the radial variation of $B$, $k_\perp$ and $v_\perp$ into account). The second step is then solving for $\varphi$. However, the Fourier transforms that appear on the left-hand-side of~\eqref{eqn:quasi_rad} add the complication that every radial mode couples together, and so global \texttt{stella} offers the user two options in solving for $\varphi$: a perturbative approach and an exact approach.
In the former, the electrostatic potential is decomposed into two pieces $\tilde{\varphi}_\bb{k} = \varphi_0 + \varphi_1$, where
\begin{align}
\varphi_0 &= \frac{1}{\Theta}\frac{2}{\upi^{1/2}}\sum_s Z_s \int \od \tilde{v}_\parallel \int \od \tilde{\mu}_s \tilde{B}(\ensuremath{r}_0) J_0(k_\perp(\ensuremath{r}_0)v_\perp(\ensuremath{r}_0)/\Omega_s(\ensuremath{r}_0))\tilde{g}_{\bb{k},s}, \\
\varphi_1 &= \frac{1}{\Theta}\ensuremath{\hat{\mathcal{X}}}\left[(2/\pi^{1/2})\sum_s Z_s \tilde{n}_s(\psi_0)\int \od \tilde{v}_\parallel \int \od \tilde{\mu}_s\, \tilde{g}_{\bb{k},s}\tilde{B}J_0(a_{s0})\left(\frac{J'_0}{J_0} + \frac{\tilde{B}'}{B}- \frac{\Theta'}{\Theta }\right)_{\ensuremath{r}=\ensuremath{r}_0}\right].
\end{align}
Here, $\varphi_0$ is the electrostatic potential obtained in the $\rho_\ast$, $\Delta \rightarrow 0$ limit, while $\varphi_1$ is obtained by grouping the perturbative correction $\ensuremath{\hat{\mathcal{X}}}(\Theta'\varphi_0)$ together with the radial corrections of the gyroaverage and velocity integration of $\tilde{g}_{\bb{k},s}$. This new term, $\varphi_1$, is then included as an additional set of terms in the explicit step ~\eqref{eqn:timestep_exp}. As the variation of $\varphi_1$ is now included as a correction to the gyrokinetic equation, the zeroth-order portion of the parallel streaming term can remain implicit, and no further change in the response matrix approach is needed. The main drawback of the perturbative approach is sensitivity to any near-cancellations of $1- J_0^2(a_{\bb{k},s})$ at low wavenumbers, which is particularly problematic for $k_y=0$ zonal modes as information in higher wavenumber modes (for which $J_0^2(a_{\bb{k},s})-1$ is not small) may transfer to larger scales due to Fourier convolution with $\ensuremath{r}-\ensuremath{r}_0$, and so the $\Delta \ll 1$ expansion may break down even for moderately small $\Delta$.
Solving~\eqref{eqn:quasi_rad} exactly entails finding a solution to the linear equation
\begin{equation}\label{eqn:qn_full}
\mathsfbi{Q}\bcdot \bb{\tilde{\varphi}}_{\bb{k}} = \bb{G},
\end{equation}
where $\mathsfbi{Q}\doteq \Theta + \ensuremath{\hat{\mathcal{X}}}\Theta'$ is the complete quasineutrality operator, $\tilde{\bb{\varphi}}_{\bb{k}}$ is the column vector comprised of all $k_x$ modes of $\tilde{\varphi}_\bb{k}$ at a given $z$ location and binormal mode-number $k_y$, and $\bb{G}$ is the right-hand-side of~\eqref{eqn:quasi_rad} expressed as a column vector. Solving~\eqref{eqn:qn_full} can be done efficiently through the use of an LU decomposition and back substitution~\citep{numerical_recipes}. Thus, $\varphi$ encompasses all the radial variation in the quasineutrality equation. (Radial variation of the gyroaveraging of $\varphi$ appearing in the gyrokinetic equation is still treated separately.) This approach unfortunately has the drawback of complicating the implicit solve of the parallel streaming term: the tridiagonal solve of~\eqref{eqn:timestep_prl} couples all modes connected by the parallel boundary condition, while the quasineutrality equation couples all modes radially, and so for a given binormal mode-number $k_y$, all radial and parallel grid points are coupled, resulting in a linear equation with a matrix of size $(N_xN_z)^2$, where $N_x$ is the number of radial modes and $N_z$ is the number of grid points in the parallel direction. Currently, when the full solve of quasineutrality is employed, global \texttt{stella} solves the parallel streaming term explicitly, though future versions of the code will include a modified response matrix algorithm that incorporates the exact quasineutrality equation.
Another complication that is introduced in the $\Delta$ expansion is the coupling of the $\varphi_{\bb{k} = \bb{0}}$ mode---which determines the parallel electric field $E_\parallel$---to other $\bb{k} \ne \bb{0}$ modes in the gyrokinetic equation. In the local limit, this mode does not enter the $\bb{E}\btimes \bb{B}$ nonlinearity nor does the non-linearity act on it; rather, it enters only in the parallel streaming term $v_\parallel\eb\bcdot \grad$, which is one-point in Fourier space in the local limit, and so the $\bb{k} =\bb{0}$ mode can be neglected.\footnote{Formally, this mode cannot be described in the flux-tube limit, as it does not obey the ordering $(k_\perp L)^{-1} \ll 1 $.} In the global limit, the $\bb{k} = \bb{0}$ mode of both the gyrokinetic and quasineutrality equations involve many $\bb{k}\ne \bb{0}$ modes, and so $\varphi_{\bb{k} = \bb{0}}$ should be solved self-consistently.
The ease of determining $\varphi_{\bb{k}=\bb{0}}$ relies on how the species are treated; for cases with both kinetic ions and electrons, $\varphi_{\bb{k}=\bb{0}}$ is entirely absent in the quasineutality equation as $J_1 = J_0^2 - 1 = 0$ for $\bb{k} = \bb{0}$, and so $\varphi_{\bb{k}=0}$ must be determined from the parallel streaming term, using the $\bb{k} = \bb{0}$ quasineutrality equation at every $z$ location as its own solvability condition. This is most conveniently done when the parallel streaming is handled implicitly, and so the current version of global \texttt{stella}, which calculates this term explicitly, zeros out $\varphi_{\bb{k}=0}$. Future versions with the implicit parallel solve will aim to self-consistently evolve this mode.
In the case of adiabatic ions or electrons, $\varphi_{\bb{k}=\bb{0}}$ enters the quasineutrality equation through the Boltzmann response of the adiabatic species. For adiabatic ions with Boltzmann response $\delta n_\mr{i}/n_\mr{i} =(Z_\mr{i}e/T_\mr{i})\varphi$, including this mode is straightforward and equation~\eqref{eqn:qn_full} for $k_y = 0$ can be used as-is.
For adiabatic electrons, a modified Boltzmann response is used for the electron density,
\begin{equation}\label{eqn:elec_boltz}
\frac{\delta n_\mr{e}}{n_\mr{e}} = \frac{q}{T_\mr{e}}\left(\varphi - \ba{\varphi}_\psi\right),
\end{equation}
where the flux surface average $\ba{A}_\psi$ of a quantity $A$ is given by
\begin{equation}
\ba{A}_\psi = \frac{\int \od y \int
\od z \, \mathcal{J} A }{\int \od y \int \od z \,\mathcal{J} } \approx \frac{\int \od y \int
\od z \, \mathcal{J}(\rcoord_0) A }{\int \od y \int \od z \,\mathcal{J}(\rcoord_0) } + \frac{\ensuremath{r}_\mr{clamped} -
\ensuremath{r}_0}{\int \od y \int \od z \,\mathcal{J}(\rcoord_0)}\int \od y \int \od z \, A \left(\mathcal{J}'(\rcoord_0) - \mathcal{J}(\rcoord_0)\frac{\int \od z \, \mathcal{J}'(\rcoord_0)}{\int \od z \, \mathcal{J}(\rcoord_0)}\right),
\end{equation}
where $\mathcal{J}$ and $\mathcal{J}'$ are given in~\ref{app:mag_geo}.
Quasineutrality for the $k_y= 0$ mode then takes the form
\begin{equation}\label{eqn:quasi_rad_matrix}
\mathsfbi{Q}\bcdot \bb{\tilde{\varphi}}_\bb{k} - \bb{G} = -\mathsfbi{C}\bcdot ({\bb{\tilde{\varphi}}_\bb{k}}-\ba{\bb{\tilde{\varphi}}_\bb{k}}_\psi),
\end{equation}
where $\mathsfbi{Q}$ now excludes the electron response and $\mathsfbi{C} \doteq \tilde{n}_\mr{e}/\tilde{T}_\mr{e} + \ensuremath{\hat{\mathcal{X}}} (\tilde{n}_\mr{e}/\tilde{T}_\mr{e} )'$. The electron Boltzmann response~\eqref{eqn:elec_boltz} now imposes a solvability constraint on the quasineutrality equation
\begin{equation}\label{eqn:solvability}
\langle \mathsfbi{Q}\bcdot \bb{\tilde{\varphi}}_\bb{k} \rangle_\psi - \langle\bb{G}\rangle_\psi = 0.
\end{equation}
Formally, this equation is over-determined: \eqref{eqn:solvability} provides $N_x$ equations for $N_x-1$ unknowns ($\tilde{\varphi}_{k_x\neq 0,k_y = 0}(z)$). While \eqref{eqn:solvability} can be made to be satisfied by the initial conditions, errors are introduced by the coupling of multiple flux-tube simulations needed for the novel radial boundary conditions;
the gyroaveraging of $\tilde{g}_{\bb{k},s}$ and $\tilde{\varphi}_\bb{k}$ that appears in~\eqref{eqn:quasi_rad} will sample points within the physical and boundary regions of the central domain when performed near the boundary. This results in discrepancies of $\varphi_\bb{k}$ within the boundary regions between the central and auxiliary flux-tube domains, which then leads to errors in~\eqref{eqn:solvability}. To circumvent this issue, we include a correction term
in the quasineutrality equation as
\begin{equation}\label{eqn:quasi_rad_matrix_tweak}
\mathsfbi{Q}\bcdot \bb{\tilde{\varphi}}_\bb{k} - \bb{G} -
\langle\langle\mathsfbi{Q}\bcdot \bb{\tilde{\varphi}}_\bb{k} - \bb{G} \rangle_\psi\rangle_{\bb{x}_\perp} = -\mathsfbi{C}\bcdot ({\bb{\tilde{\varphi}}_\bb{k}}-\ba{\bb{\tilde{\varphi}}_\bb{k}}_\psi),
\end{equation}
where
\begin{equation}\label{eqn:perp_areal_average}
\ba{A_\bb{k}}_{\bb{x}_\perp} = \ensuremath{\mathcal{F}_\bb{k}}\left( \frac{\mr{TH}(\tilde{x})}{\ell_y(\ell_x - 2\tilde{L}_\mr{boundary})}\int_0^{\ell_y}\od \tilde{y} \int_{\tilde{x}_- + \tilde{L}_\mr{boundary}}^{\tilde{x}_+ - \tilde{L}_\mr{boundary}}\od \tilde{x}\,\ensuremath{\mathcal{F}_\bb{k}}^{-1}\left(A_\bb{k}\right)\right)
\end{equation}
is the perpendicular areal average within the physical region, and
\begin{equation}
\mr{TH}(\tilde{x}) = \left\{\begin{matrix}
1 & \tilde{x}_- + \tilde{L}_\mr{boundary}< \tilde{x} < \tilde{x}_+ - \tilde{L}_\mr{boundary}, \\ 0 & \textrm{otherwise},
\end{matrix}\right.
\end{equation}
is a top-hat function that excludes the boundary region.
Equation~\eqref{eqn:quasi_rad_matrix_tweak} can then be solved after imposing a gauge potential, which we choose to be $\bb{\hat{k}}_0\bcdot \langle\mathsfbi{C}\bcdot\bb{\hat{k}}_0\bb{\hat{k}}_0\bcdot \bb{\tilde{\varphi}}_\bb{k}\rangle_\psi = 0$, where $\bb{\hat{k}}_0$ is the unit column vector for the $\bb{k}=\bb{0}$ modes. The coupling of the $\bb{k} = \bb{0}$ modes to and from modes with $\bb{k}\ne \bb{0}$ is $\mathcal{O}(\Delta)$, and so the effect of the $\bb{k} = \bb{0}$ modes on physical observables is $\mathcal{O}(\Delta^2)$. Numerical tests confirm the scaling of $|\tilde{\varphi}_{\bb{k} =\bb{0}}| \sim \Delta$, and so we expect the impact of the correction introduced in~\eqref{eqn:quasi_rad_matrix_tweak} to be small.
\subsection{Sources and sinks}\label{sec:sources}
When the novel radial boundary condition is used for global simulations in \texttt{stella}, a mismatch of flux between the left and right simulation domains results in a pile-up or deficit of particles and heat in the central domain (the net flux of toroidal angular momentum in an up-down symmetric local simulation with no mean flow shear is zero, see~\citet{Parra_symmetry}). This necessitates the inclusion of a source or sink in the central domain. Global \texttt{stella} is equipped with two types of sinks: one based on a Krook-type operator and a new projection-based operator that exploits the scale-separated nature of our $\delta f$ approach.
The Krook type operator, commonly employed in global simulations, is of the form
\begin{equation}\label{eqn:krook_operator}
D_\mr{K}(\tilde{g}_{\bb{k},s}) = - \nu_\mr{S}
\ba{\ba{\tilde{g}^\mr{even}_{\bb{k},s} - \frac{\tilde{F}_s (\ensuremath{r})}{\tilde{n}_s(\ensuremath{r})} \ba{\int \od^3 \tilde{v} \, J_{0s}\tilde{g}^\mr{even}_{\bb{k},s}}_\psi}_{\bb{x}_\perp}}_t,
\end{equation}
where $\nu_\mr{S}$ sets the strength of the source operator, $\tilde{F}_s = (v^3_\mr{{th}s}/n_\mr{ref})F_s$,
\begin{equation}
\tilde{g}^\mr{even}_{\bb{k},s} = \frac{1}{2} \left[\tilde{g}_{\bb{k},s}(v_\parallel) + \tilde{g}_{\bb{k},s}(-v_\parallel)\right]
\end{equation}
is the even-in-$v_\parallel$ component of the distribution function, $\langle \cdots \rangle_{\bb{x}_\perp}$ is given by~\eqref{eqn:perp_areal_average}, and
\begin{equation}\label{eqn:time_average}
\ba{\cdots }_t \doteq \frac{\int_0^t \od t' \exp(t'/\tau_\mr{S})(\cdots)}{\int_0^t \od t' \exp(t'/\tau_\mr{S})}
\end{equation}
is an exponentially-weighted time average, where $\tau_\mr{S}$ sets the averaging window~\citep{Candy_globloc}. The Krook operator given by~\eqref{eqn:krook_operator} is similar to that advocated by~\citet{McMillan_sources}, where the even-in-$v_\parallel$ portion of $\tilde{g}_{\bb{k},s}$ is used in order to conserve momentum, and the zeroth moment of $\tilde{g}_{\bb{k},s}$ is subtracted in order to avoid introducing density perturbations on a flux surface.
This operator is then included as an additional term on the right-hand-side of~\eqref{eqn:GKE_FT}, and operates only on the $k_y = 0$ modes. This, along with the time-averaging, ensures that this operator only affects the large-scale, long-time build up of particles and energy. The time-averaging procedure~\eqref{eqn:time_average} can be implemented numerically in a way that minimizes storage requirements, necessitating only an additional array the size of $\tilde{g}_{\bb{k}_y =0,s}$ that conglomerates all the information needed from the previous timesteps.
The projection-operator-based sink, provided as a more physically-motivated alternative to the standard Krook operator, is rooted in the $\delta f$ gyrokinetic formalism. Here, the full distribution function $f$ is decomposed into a large-scale, long-time component $F$ and a small-scale, short-time component $\delta f$. There then exists a suitable large-scale, long-time transport average $\ba{\cdots}_\mr{T}$ such that $\ba{f}_\mr{T} = F$ and $\ba{\delta f}_\mr{T} = 0$. The $\delta f$ gyrokinetic equation can then be obtained from
\begin{equation}
\frac{\partial \delta f}{\partial t} = \frac{\partial f}{\partial t} - \ba{\frac{\partial f}{\partial t}}_\mr{T}.
\end{equation}
It is this concept on which we base the projection-operator sink. Like the Krook-operator above, we let $\ba{\cdots}_\mr{T} =\langle \langle \cdots \rangle_{\bb{x}_\perp}\rangle_t$. This operator is then applied to the entire right-hand-side of the gyrokinetic equation, resulting in
\begin{align}\label{eqn:GKE_proj}
\frac{\partial g_s}{\partial t} = &- v_\parallel {\eb \bcdot \grad z} \left(\frac{\partial g_s}{\partial z} + {\frac{Z_se}{T_s} F_s} \frac{\partial \langle \varphi \rangle_\bb{R}}{\partial z} \right) + \frac{\mu_s}{m_s} {\eb \bcdot \grad B} \frac{\partial g_s}{\partial v_\parallel} \nonumber - {\bb{v}_{\mr{M}s}}\bcdot \left(\grad_\perp g_s +{\frac{Z_se}{T_s}F_s}\grad_\perp \langle \varphi\rangle_\bb{R}\right) \\
&- {\langle\bb{v}_\bb{E} \rangle_\bb{R} \bcdot \grad_\perp} g_s - {\langle \bb{v}_\bb{E} \rangle_\bb{R}\bcdot \grad\big|_E F_s} - \Bigg\langle- v_\parallel {\eb \bcdot \grad z} \left(\frac{\partial g_s}{\partial z} + {\frac{Z_se}{T_s} F_s} \frac{\partial \langle \varphi \rangle_\bb{R}}{\partial z} \right) + \frac{\mu_s}{m_s} {\eb \bcdot \grad B} \frac{\partial g_s}{\partial v_\parallel} \nonumber \\
& - {\bb{v}_{\mr{M}s}}\bcdot \left(\grad_\perp g_s +{\frac{Z_se}{T_s}F_s}\grad_\perp \langle \varphi\rangle_\bb{R}\right) - {\langle\bb{v}_\bb{E} \rangle_\bb{R} \bcdot \grad_\perp} g_s - {\langle \bb{v}_\bb{E} \rangle_\bb{R}\bcdot \grad\big|_E F_s} \Bigg\rangle_\mr{T}.
\end{align}
This operator is simple to implement numerically: one first calculates the distribution function for the next timestep $\tilde{g}^{n+1}_{\bb{k},s, \mr{int}}$ as one normally would without any sinks, and then the actual distribution function at the next time step $\tilde{g}^{n+1}_{\bb{k},s}$ is given by
\begin{equation}
\tilde{g}^{n+1}_{\bb{k},s} = \tilde{g}^{n+1}_{\bb{k},s, \mr{int}} - \upDelta t^n\ba{(\tilde{g}^{i+1}_{\bb{k},s, \mr{int}}-\tilde{g}^{i}_{\bb{k},s})/\upDelta t^i}_{\mr{T},\, i\,\in\, [0, n]},
\end{equation}
where $\upDelta t^i$ is the (possibly time dependent) simulation time step at step $i$ separating $\tilde{g}^i_{\bb{k},s}$ and $\tilde{g}^{i+1}_{\bb{k},s}$.
Apart from being a more physically motivated sink, this method also has the advantage of only requiring one parameter to be specified, $\tau_\mr{S}$, which must be chosen to be longer than any microscopic time-scale of interest; in practice, we find for global simulations that $\Delta^{-1}(a/v_\mr{thi}) \lesssim \tau_\mr{S} \lesssim \rho_\ast^{-1}(a/v_\mr{thi})$ is sufficient for convergence.
When employing the multiple-flux-tube radial boundary condition, discrepancies between the turbulence supplied by the auxiliary simulations and that of the central domain may sometimes result in strong shear layers just outside the boundary region. In order to mitigate these shear layers, global~\texttt{stella} supplies localized Krook-type sinks which can operate inside a subregion of length $L_\mr{K}$ of the boundary region in order to smooth the transition between the supplied turbulence and the modelled turbulence in the central region. These operators take the form
\begin{equation}
D_\mr{BC}[\tilde{g}_{k_ys}(\tilde{x})] = \left\{\begin{matrix}
\eta(\tilde{x}-\tilde{x}_- - \tilde{L}_\mr{boundary})(g_{s,\mr{L}}(\tilde{x}) - \tilde{g}_s(\tilde{x})) & \quad & \tilde{x}_- + (\tilde{L}_\mr{boundary} -L_\mr{K})< \tilde{x} < \tilde{x}_- + \tilde{L}_\mr{boundary} \\
\eta(\tilde{x}-\tilde{x}_++\tilde{L}_\mr{boundary})(g_{s,\mr{R}}(\tilde{x}) - \tilde{g}_s(\tilde{x})) & \quad & \tilde{x}_+ - \tilde{L}_\mr{boundary} < \tilde{x} < \tilde{x}_+ - (\tilde{L}_\mr{boundary}-L_\mr{K}) \\
0 & \quad & \textrm{elsewhere}\end{matrix} \right.
\end{equation}
where $\eta(x) = \nu_\mr{BC}[ 1-(1-\rme^{-3(L_\mr{K} -|x|)/L_\mr{K}})/(1-\rme^{-3})]$ and $\nu_\mr{BC}$ sets the strength of the operator. In practice, $\nu_\mr{BC}$ is chosen to be comparable to the growth rate of the fastest growing mode over all three domains. When these Krook operators are used in simulations with the novel boundary condition, $g_\mr{s,\mr{L}}$ and $g_\mr{s,\mr{R}}$ are only directly copied into the Krook-free portion of the boundary region. Like the case with $\tilde{L}_\mr{boundary}$, the size of the Krook damping region should be similar to the size of the largest Larmor orbit contained in the simulation.
\subsection{Equilibrium flow shear}\label{sec:flowshear}
The local version of \texttt{stella} currently implements equilibrium flow shear using the discrete wavenumber-shift method~ formulated by \citet{Hammett_FlowShear} and includes the nonlinear corrections advocated by~\citet{McMillan_FlowShear} which renders the scheme semi-continuous.
The wavenumber advection method is based on the fact that for a linear background velocity profile $v_{\bb{E},0,y} = \gamma_E\tilde{x}$, where $\gamma_E$ is the shear rate, the advection equation
\begin{equation}
\frac{\partial g}{\partial t} + v_{\bb{E},0,y}(x) \frac{\partial g}{\partial y} + \ldots
\end{equation}
is equivalent to solving the system with a time-dependent $k_x$,
\begin{equation}\label{flowshear}
\frac{\partial }{\partial t}g(k_x - k_y \gamma_E t,k_y) + \ldots.
\end{equation}
For gyrokinetic codes employing an implicit integration scheme, having a time-dependent $k_x$ would require a recalculation of the response matrices at every simulation time step, and thus would be prohibitively expensive computationally. The wavenumber shift approach instead remaps the distribution function discretely.
To determine when and how this remapping occurs, we define a $k_y$-dependent marker wavenumber $k_x^\ast(t) \doteq - k_y \gamma_E t$ that evolves in time. Whenever $|k_x^\ast(t) - k_x^\ast(t^*_\mr{prev})| > \upDelta k_x /2$, where
\begin{equation}
t^*_\mr{prev} = \upDelta t_E \left[\operatorname{floor}\left(\frac{t_\mr{prev}+\upDelta t_E/2}{\upDelta t_E}\right) - \frac{1}{2}\right],
\end{equation}
$t_\mr{prev}(k_y)$ is the time of the previous remapping for binormal wavenumber $k_y$, $\upDelta t_E = |\upDelta k_x/ \gamma_E k_y|$ is the nominal time between remappings for binormal wavenumber $k_y$, and $\upDelta k_x$ is the radial wavenumber grid spacing, then information of the distribution function $g_{\bb{k},s}$ at radial wavenumber $k_x$ is transferred to radial wavenumber $k_x - \operatorname{sign}(\gamma_E)n_\mr{shift}\upDelta k_x$, i.e.,
\begin{equation}
g^{n+1}_{k_x - \operatorname{sign}(\gamma_E)n_\mr{shift}\upDelta k_x, k_y} = g^{n}_{k_x, k_y},
\end{equation}
where $n_\mr{shift} = \mr{round}[|k_x^\ast(t) - k_x^\ast(t^*_\mr{prev})|/\upDelta k_x]$ is the number of radial wavenumbers by which the distribution function is to be shifted. This remapping, if it is to occur, is done at the beginning of a simulation time step before any other term in the gyrokinetic equation is calculated. This scheme, which leaves the radial wavenumber grid unchanged, ensures that $k_x$ on the fixed grid is as close to the continuous time $k_x(t) = k_x(t=0) + k_x^\ast (t)$ as possible.
Note that the $k_y=0$ zonal modes are unaffected by this shearing, and so do not require remapping.
It has been recently shown that the accuracy of the wavenumber advection scheme can be improved by incorporating a continuously varying $k_x$, as well as a continuous flow shear, into the numerical integration of the nonlinear terms that appear in the gyrokinetic equation~\citep{McMillan_FlowShear,christen_2021}. These modifications are done to avoid the problem where modes at different $k_y$ that are initially uncoupled nonlinearly become coupled due to staggered remapping. As the nonlinear terms are typically integrated using an explicit scheme, these modifications can be made at little extra computational cost, and are implemented in a straightforward manner: first, the substitution $k_x \rightarrow k_x - k_y\gamma_E(t - t_\mr{prev})$ is made for every $k_x$ appearing in the nonlinear term.\footnote{This substitution is not made in the $J_0$ Bessel functions that appear in the nonlinear term. This is not needed, however, to avoid spurious cross-coupling between uncorrelated modes.} Second, a phase-shift to $g_{\bb{k},s}$ and $\varphi_\bb{k}$ is applied by multiplying an appropriate prefactor,
\begin{equation}\label{eqn:prefactor}
\varphi_{k_y}(x) \rightarrow \varphi_{k_y}(x) \exp[-\imag k_y \gamma_E x(t-t^*_\mr{prev})].
\end{equation}
This renders the effects of shearing semi-continuous. Once the nonlinear term is calculated, the result is divided by this prefactor, and the numerical integration then proceeds as usual.
The global version of \texttt{stella} employs a continuous form of equilibrium flow shear, which simply adds flow shear as a background linear profile. While this approach has been used in gyrofluid simulation in the past~\citep{Waltz_1994}, it has the disadvantage of introducing a region of very strong shear at the radial boundaries of the simulation where the shear profile experiences a discontinuity, and thus results in excessive levels of turbulence stabilization. This drawback can be avoided entirely by utilizing the novel boundary conditions detailed in \S \ref{sec:radBC}; as the boundary regions are replenished with `good' information at every time step, the effects of the large amounts of shear at the profile discontinuities is reduced. When this approach is used, the shear rate is made continuous across all three domains by using the wavenumber shift approach, along with nonlinear corrections, in the left and right domains. Prior to communicating information of the left and right distribution functions, $g_{s,\mr{L}}$ and $g_{s,\mr{R}}$ are multiplied by the phase factor given in~\eqref{eqn:prefactor} in order to render the shearing semi-continuous. These distribution functions are then Doppler shifted in order to match the velocities at the boundaries of the central domain. After these steps are performed, the information can be copied as usual.
{
Like the spatial operator $\ensuremath{\hat{\mathcal{X}}}$ detailed in \S \ref{sec:sim_equations}, the continuous approach to flow-shear space suffers the drawback of wrap-around, where a mode at one extreme of the radial wavenumber grid gets transferred to the opposite extreme, rather than moving off-grid.
In the discrete wavenumber shift approach, which is done in $k_x$-$k_y$ space, the wrap-around problem is easily avoided: information of the distribution function that gets sheared off the radial wavenumber grid is simply zeroed out. In the global version of equilibrium shear, which is done in $x$-$k_y$ space, a mode at $k_{x,\mr{max}}$ would unphysically wrap around to $-k_{x,\mr{max}}$ if no additional steps were taken. Unlike the shearing resulting from the $\rho_\ast$-small terms due to radial profile variation, the shearing from the equilibrium flow is of order unity, and so one cannot rely on hyper-dissipation alone to prevent wrap-around. Fortunately, the advection speed $v_{k_x} = - \gamma_E k_y$ due to equilibrium flow shear is well defined for every $k_y$, and is uniform across the entire $z$ domain. We thus prevent wrap-around with global shear by zeroing out the distribution function at $k_x = \pm k_{x,\mr{max}}$ over the same $k_y$-dependent frequency that a sheared mode takes to cross a single cell in the radial wavenumber grid, i.e., whenever $|k_x^\ast(t) - k_x^\ast(t^{**}_\mr{prev})| > \upDelta k_x$, where $t^{**}_\mr{prev} = \upDelta t_E \operatorname{floor}(t_\mr{prev}/\upDelta t_E)$ and $t_\mr{prev}$ here refers to the last time a mode at $k_y$ was zeroed out. We have found that this largely eliminates any wrap-around of information across the boundaries of the $k_x$ grid.
}
\subsection{Hyper dissipation}\label{sec:hyperdissipation}
A fourth-order hyper-dissipation operator is used in \texttt{stella} in order to prevent a pile-up of energy at small perpendicular scales due to turbulent cascade. This operator, which is applied to the distribution function as an additional implicit step in the operator splitting scheme, has the form
\begin{equation}
g_{n+1} = \frac{1}{1+ \upDelta t D_\mr{hyper}[k^2_{\perp,\textrm{hyper}}/\operatorname{max}(k^2_{\perp,\textrm{hyper}})]^2}g_n,
\end{equation}
where $D_\mr{hyper}$ is an adjustable coefficient which sets the strength of the dissipation and $k^2_{\perp,\rm{hyper}}(k_x,k_y)$ is some measure of the squared perpendicular wavenumber. For local \texttt{stella},
\begin{equation}
k^2_{\perp,\mr{hyper}} = k_\perp^2 = k_x^2 |\grad x|^2 + 2 k_xk_y(\grad x\bcdot \grad y) + k^2_y |\grad y|^2
\end{equation}
is the physical squared perpendicular wavenumber. The geometrical coefficients $|\grad x|^2$, $\grad x\bcdot \grad y$, and $|\grad y|^2$ vary in $\vartheta$, and ensure continuity across the parallel boundary for modes connected by the parallel boundary condition (\S \ref{sec:coords}). These coefficients, however, also vary radially along with the geometric profile variation, thus complicating the application of hyper-dissipation for global simulations. This difficulty can be avoided by using an alternative definition for $k_{\perp,\mr{hyper}}$:
\begin{equation}
k^2_{\perp,\mr{hyper}} = \left\{ \begin{matrix}k_x^2 & \quad & k_y = 0, \\ k_y^2[1+\lambda(\theta - \theta_0)^2]& \quad & k_y \ne 0,\end{matrix}\right.
\end{equation}
where $\theta_0 \doteq k_x/k_y$ is the ballooning angle and $\lambda = 1$ when $q$ is used as the $x$ coordinate.\footnote{Our definition of the ballooning angle differs from the standard one, $\theta_0 = k_x/k_y \hat{s}$, as we use $q$ as the radial coordinate instead of $\psi$.} This form of $ k^2_{\perp,\mr{hyper}}$ retains the $\theta$ dependence that ensures continuity across the parallel boundary, while doing away with the radial variation of the geometrical coefficients. In practice, the results of a nonlinear simulation should be insensitive to the exact form of the hyper-dissipation operator, provided the small perpendicular scales are well resolved.
\section{Numerical benchmarks}\label{sec:tests}
In this section we perform a variety of tests in order to demonstrate the efficacy of our new approach to global gyrokinetics; namely, that of the novel radial boundary conditions and the inclusion of next-order corrections in the underlying gyrokinetic equations. Unless otherwise noted, all simulations used standard Cyclone Base Case (CBC) parameters which are tabulated in Table~\ref{tab:CBC_param}, and we define $\rho_\ast = \rho_\mr{ref}/a$. Additionally, the values $N_x$ and $N_y$ refer to the number of modes resolved in the radial and binormal direction after $2/3$ dealiasing, and so the number of collocation points when calculating the nonlinear terms is effectively larger by a factor of $3/2$. This dealiasing is not performed during the communication of the novel radial boundary condition, and so the boundary size $N_\mr{boundary}$ should be compared with $N_x$, rather than $3N_x/2$.
\begin{table}
\centering
\begin{tabular}{c|c}
input parameter & value \\
\hline
$m_\mr{i}/m_\mr{ref}$ & 1 \\
$n_\mr{i}(\ensuremath{r}_0)/n_\mr{ref}$ & 1 \\ $T_\mr{i}(\ensuremath{r}_0)/T_\mr{ref}$ & 1 \\
$n_\mr{i}/ n_\mr{e}$ & 1 \\
$T_\mr{i}/T_\mr{e}$ & 1 \\
$R_0/L_{n_\mr{i}} = -(R_0/n_\mr{i}) n'_\mr{i}$ & 2.2 \\
$R_0/L_{T_\mr{i}} = -(R_0/T_\mr{i}) T'_\mr{i}$ & 6.9 \\
$\epsilon = \ensuremath{r}_0/a $ & 0.5\\
$R_0/a$ & 2.77778 \\
\end{tabular}
\hspace{1cm}
\begin{tabular}{c|c}
input parameter & value \\
\hline
$q$ & 1.4 \\
$\hat{s}$ & 0.796 \\
$\kappa$ & 1 \\
$\kappa'$ & 0 \\
$\delta$ & 0 \\
$\delta'$ & 0 \\
$\beta$ & 0 \\
$\beta'$ & 0 \\
\\
\end{tabular}
\caption{Parameters of the Cyclone Base Case (CBC).}
\label{tab:CBC_param}
\end{table}
\subsection{Continuity of the radial boundary conditions}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/phiout_010200}
\caption{Snapshot of the electrostatic potential in the perpendicular plane situated at the outboard midplane for a simulation using CBC parameters with no radial profile variation. Dashed lines denote the simulation domain boundaries.}
\label{fig:CBC_omp_section}
\end{figure}
\begin{figure}
\centering
\includegraphics{plots/CBC_fluxes}
\caption{Ion heat fluxes in each domain of a multiple-domain nonlinear simulation without radial profile variation ($\rho_\ast =0$).}
\label{fig:CBC_fluxes}
\end{figure}
To test the performance of the new radial boundary conditions, we perform a simulation using CBC parameters without any profile variation ($\rho_\ast = \Delta = 0$) in order to demonstrate continuity between the physical domains. This simulation uses adiabatic electrons, 85 dealiased modes in each of the the radial and binormal directions, 32 points along the magnetic field, 24 points in the parallel velocity space, 6 points in $\mu$ space, and $D_\mr{hyper} = 0.05$. The boundary regions in the center domain each use ten co-location points, and the boundary Krook operator is applied in the inner eight points with $\nu_\mr{K}(a/v_\mr{th,ref}) = 0.5$. Each flux-tube domain is seeded with small amplitude random noise, and after a period of exponential growth in the initial linear regime, the nonlinearity becomes dynamically important and the domains reach a saturated turbulent state. No other sources or sinks are used in these simulations.
Figure~\ref{fig:CBC_omp_section} shows a snapshot of the electrostatic potential for each simulation domain in the perpendicular plane situated at the outboard midplane. As no radial profile variation is included in these simulations, the statistical characteristics of the turbulence in each simulation domain should be identical to one-another. Indeed, the cross-section of $\varphi$ shows similar turbulence across all the domains, without any spurious behaviour occurring near the domain boundaries. The heat fluxes in each domain are also shown in figure~\ref{fig:CBC_fluxes} to be statistically identical, and the linear growth rates (not shown) are also consistent. The novel radial boundary condition is thus shown to be well behaved. An expression for the heat flux is given in~\ref{app:fluxes}.
\subsection{Equilibrium flow shear}
Our global implementation of equilibrium flow shear is tested by comparing the resulting fluxes to those obtained by local flux-tube simulations using the standard wavenumber shift method~\citep{Hammett_FlowShear} along with non-linear corrections~\citep{McMillan_FlowShear}. Unlike radial profile variation, equilibrium flow shear is not a $\rho_\ast$-small effect, and so these tests are performed without other global effects (i.e. $\rho_\ast = \Delta = 0$).
In these tests, our global simulations use the CBC parameters and a resolution of $(N_y,N_z,N_{v_\parallel}, N_\mu) = (21,16,24,6)$. The side domains use $N_x = 43$ with a radial extent of $\ell_x = 62.8\rho_\mr{i}$ while the central domain uses $N_x = 85$ dealiased modes with a radial extent of $\ell_x = 125 \rho_\mr{i}$. The boundary regions use 16 cells each, with a Krook operator applied on the outer 10 cells with a damping rate of $0.2$. If the Krook operator is not used, then shear layers can develop at the interfaces of the boundary region which may counteract the effect of the background flow shear. The local simulations use the same parameters as the central domain in the global simulation, and adiabatic electrons are employed throughout. No other sources or sinks are used in the simulations.
Figure~\ref{fig:shear_both_LR} shows the total heat flux (left) and parallel momentum flux (right) for the global implementation of flow shear (yellow line) along with the result from local flux-tube simulations (blue line). Expressions for these fluxes are given in \ref{app:fluxes}. These results reveal that our global implementation of flow shear compares well to the more standard method. { Also included in this plot as black squares are the results from simulations of using global flow shear without the periodical zeroing-out of information at the boundary. As argued in \S \ref{sec:sim_equations} and \S \ref{sec:flowshear}, the results are insensitive to this zeroing for small values of shearing due to the presence of finite hyper-dissipation. However, for large flow shear the information can escape this hyper-dissipation, thus causing an over-prediction of both heat and momentum flux when the zeroing-out dealiasing routine is not performed; this over-prediction worsens as shearing is increased.}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/shear_both_LR2.eps}
\caption{Comparison of the total heat flux (left) and momentum flux (right) between a local flux-tube simulation (yellow line) and a global simulation (blue line) with no profile variation ($\rho_\ast = 0$) across a range of shear rates $\gamma_E$. Each simulation employs both parallel and perpendicular flow shear. Black squares denote results from the global simulation without the zeroing out described in the last paragraph of \S \ref{sec:flowshear}.}
\label{fig:shear_both_LR}
\end{figure}
\subsection{Consistency of the first derivatives}
Our approach to global gyrokinetics involves the Taylor expansion of the geometrical coefficients (given in~\ref{app:mag_geo}), as well as the density and temperature profiles. This introduces various additional terms in the gyrokinetic and quasineutrality equations, as well as the gyro-averaging operator. In order to verify the implementation of these terms, we compare the geometrical coefficients computed in two separate local flux-tube simulations; one simulation is performed at a displaced radial location $\ensuremath{r} = \ensuremath{r}_0 + \upDelta \ensuremath{r}_+$ that does not employ the additional terms in the Taylor expansion, and a simulation located at $\ensuremath{r} = \ensuremath{r}_0$ which then employs these additional terms to evaluate the coefficients at $\ensuremath{r}_0 + \upDelta \ensuremath{r}_+$ (thus placing the flux tube at $\ensuremath{r}=\ensuremath{r}_0 + \upDelta r_+$). We then compute the absolute error between the geometrical coefficients of the two simulations and plot this as a function of $\rho_\ast$ (equivalently, $\Delta$); proper implementation of the terms resulting from the first-order Taylor expansion results in a scaling of the absolute error with $\rho_\ast^2$. Plotted in figure~\ref{fig:milerr} is the error scaling of various geometrical coefficients with $\rho_\ast$ which do indeed exhibit a $\rho_\ast^2$ scaling for a broad range of $\rho_\ast$. While this figure only displays a few geometrical coefficients, these scalings have also been verified for all other geometrical coefficients, as well as for terms resulting from radial variation in density and pressure. As a final note, the error scalings for some quantities in figure~\ref{fig:milerr} do break from the $\rho_\ast^2$ scaling for $\Delta \gtrsim 0.3$; however, for these order-unity values of $\Delta$ the Taylor expansion is expected to breakdown.
\begin{figure}
\centering
\includegraphics{plots/milerr}
\caption{Convergence of several geometric factors as a function of $\rho_\ast$. Solid lines denote the line-averaged absolute error, e.g. $\delta R = L_z^{-1}\int \od z |R(r + \upDelta r, z) - R(r,z) - \upDelta r R'(r,z)|$. Dotted line denotes the ideal scaling $\rho_\ast^2$. Here, $\rho_\ast = \rho_\mr{ref} / a$.
}
\label{fig:milerr}
\end{figure}
\subsection{Convergence to the $(k_\perp L)^{-1} \rightarrow 0$ limit}
Local flux-tube gyrokinetics is obtained by formally taking the $(k_\perp L)^{-1} \rightarrow 0$ limit of the gyrokinetic equations~\citep{Parra_globallocal}. Thus, a stringent test that all global gyrokinetic codes must pass is to prove convergence to local gyrokinetic simulations in the $(k_\perp L)^{-1} \rightarrow 0$ limit, which can be done by letting $\Delta$ tend towards zero. We test this convergence in $\texttt{stella}$ using the Cyclone Base Case parameters with adiabatic electrons, as well as the additional second derivatives $ q''a^2 = 5$, $ n''(a^2/n) =-1 $, $ T''_\mr{i}(a^2/T_\mr{i}) =-4$. These second derivatives have been chosen in order to provide appreciable modification to the gradient scale lengths as $\Delta$ approaches order unity.
The multiple flux-tube boundary conditions are employed, and each simulation domain has a resolution of $(N_x, N_z,N_\mu, N_{v_\parallel}) = (43,24,12,48)$, a single poloidal turn and a radial extent in terms of gyroradii of $\ell_x = 60 \rho_\mr{i}$. The size of the radial boundary is 4 collocation points. The spatial resolution is held fixed between the simulations, and so as $\rho_\ast$ is increased, the physical region of the device that is sampled (as a portion of the minor radius $a$) is also increased. Sources or sinks are not needed in linear simulations since profile relaxation is a nonlinear effect, and so sources or sinks are not used in the linear simulations reported in this section. Each domain is seeded with a single $k_y\rho_\mr{i} = 0.2$ mode which, in a linear simulation, decouples from other modes with different binormal wavenumber. As the system evolves, the growth rate of the volume-averaged heat flux (\ref{app:fluxes}) eventually reaches a constant value, which is plotted in the left panel of figure~\ref{fig:rhostar_scan} as a function of $\rho_\ast$. The global simulations are indeed shown to converge to the local result, which is denoted in the figure with a dotted line. Apart from the largest value of $\rho_\ast$ at which the $\Delta$ expansion begins to breakdown, this growth rate is nearly a monotonically increasing function of $\rho_\ast$. This is due to the radial domain sampling regions of larger $R/L_T$ near the outer edge of the torus, an effect that disappears as $\rho_\ast$, and thus the radial extent in terms of $r$, is decreased. Additional simulations for $\rho_\ast = 0.002$ where $N_z$, $N_\mu$, and $N_{v_\parallel}$ are separately halved have been performed to ensure convergence of the growth rates.
We also perform a $(k_\perp L)^{-1} \rightarrow 0$ convergence test for nonlinear simulations by comparing the time- and volume-averaged ion heat flux of the saturated turbulent state. This is performed using the same parameters as in the linear simulations of the previous paragraph, but with the resolution of $(N_x, N_y, N_z,N_\mu, N_{v_\parallel}) = (171,85,16,8,48)$. Additionally, the simulations employ the projection-based source with $\tau_\mr{S}(v_\mr{th,ref}/a) = 50$, and no boundary-region Krook operator.
The results of these simulations are shown on the right panel of figure~\ref{fig:rhostar_scan}, along with the result of a nonlinear flux-tube simulation performed at $\ensuremath{r}/a = 0.5$. While the heat flux signals are much noisier than the growth rates of the linear simulations, convergence in $\rho_\ast$ is indeed demonstrated for small enough $\rho_\ast$.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/rhostar_scan2}
\caption{(\emph{left}) Time-averaged growth rate of the volume-average ion heat flux as a function of $\rho_\ast$ for radially global linear simulations. Dotted line indicates the growth rate of a local linear simulation centered at $r/a = 0.5$. (\emph{right}) Time- and volume-averaged ion heat flux in a saturated turbulent state for radially global nonlinear simulations. Error bars denote the standard error of the mean. Solid line indicates the time- and volume-averaged heat flux of a local nonlinear simulation centered at $r/a = 0.5$, while the dotted lines denote the standard error of the mean for that simulation. Here, $\rho_\ast = \rho_\mr{ref} / a$.}
\label{fig:rhostar_scan}
\end{figure}
Finally, in order to showcase the utility of the hybrid global-local approach, we perform a $(k_\perp L)^{-1} \rightarrow 0$ convergence test where the physical region of the device that is sampled is held fixed, rather than the box size in terms of the number of gyroradii. Three sets of simulations are used, employing the multiple-flux-tube boundary condition, the periodic-triangle-wave boundary condition detailed in~\citet{Candy_globloc}, as well are Dirichlet boundary condition. The numerical resolution for these simulations are $(N_y, N_z, N_\mu, N_{v_\parallel}) = (43,16,8,36)$. The radial resolution for the simulations employing the multiple-flux-tube and Dirichlet boundary conditions are $N_x = 43$, $171$, $341$ for $\rho_\ast^{-1} = 250$, $500$, $1000$, $2000$, respectively, while using a boundary size of four collocation points for $\rho_\ast^{-1} = 250$ and eight for the remaining three values of $\rho_\ast$. Additionally, the boundary Krook operator is applied to half of those points for these simulations. Dirichlet boundaries are applied similarly to the multiple-flux-tube approach, but with zeroes instead of information sourced from auxiliary local simulations. The simulations employing the periodic triangle wave boundary condition have radial box sizes and resolutions that are twice as large as the other two analogous cases, but do not need to use auxiliary flux-tube simulations or a boundary-region Krook operator. All simulations employ the Krook-based source with $\tau_\mr{S}(v_\mr{th,ref}/a) = 50$ and $\nu_\mr{K}(a/v_\mr{th,ref}) = 0.15$. We note that the periodic triangle simulations require twice the computational resources to perform compared to analogous simulations employing Dirichlet boundary conditions, while the multiple flux-tube simulations requires three times as much, though this can be mitigated in the future by using smaller radial domains in the auxiliary simulations and appropriate load balancing.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{plots/rho_scale}
\caption{Ion heat fluxes as a function of $\rho_\ast$ for simulations where the sampled physical region is held fixed. Included are simulations which use a multiple-flux-tube radial boundary condition (purple line, \S \ref{sec:radBC}), the periodic triangle wave boundary condition (green line, \citep{Candy_globloc}), as well as Dirichlet boundary conditions (blue line). Fluxes are time averaged over the saturated states and volume averaged within the center half of the simulation domain. The dotted line is the result of a local simulation located at $\ensuremath{r}/a = 0.5$, while the shaded region denotes the standard error of that mean. }
\label{fig:rhostar_scan_fixbox}
\end{figure}
The results from these simulations are given in figure~\ref{fig:rhostar_scan_fixbox}, which displays the ion heat flux as a function of $\rho_\ast$. Here, heat fluxes are time-averaged over the saturated state, while volume-averaged over the half-domain centered around $\ensuremath{r}/a = 0.5$; for the simulations employing periodic triangle wave boundary conditions, this is the half-domain of the physical region of the simulation, which is the left half of the simulation domain. We see from figure~\ref{fig:rhostar_scan_fixbox} that
the simulations using multiple flux-tube and triangle wave boundary conditions agree with local result for $\rho_\ast$ at least as large as $\rho_\ast \approx 1/250$, while the simulations employing Dirichlet boundary conditions significantly underpredict the fluxes for $\rho_\ast^{-1} \lesssim 500$. This archetypal under-prediction of the fluxes relative to the local result is often attributed to global effects, such as profile shearing, that are absent in local codes: The results shown here instead indicate that the artificial damping of fluctuations associated with Dirichlet boundary conditions is responsible for the reduction in flux with $\rho_\ast$. Such a concern has already been raised in~\citet{Candy_globloc}, who found a similar insensitivity of the heat flux to variation of $\rho_\ast$ when using the periodic-triangle-wave boundary condition.
\subsection{Rosenbluth-Hinton tests}
A standard benchmark for gyrokinetic codes is the Rosenbluth-Hinton test~\citep{Rosenbluth}, which verifies the implementation of the parallel streaming terms, the radial magnetic drifts and the $k_y = 0$ mode in the $k_\perp \rho_\mr{i}\rightarrow 0$ limit. In this test, the residual level of a zonal electrostatic perturbation that has been damped by collisionless processes is measured and compared to the theoretical prediction for large aspect ratio circular geometry:
\begin{equation}\label{eqn:RH_res}
\mr{RH}_\mr{XC} \doteq \frac{\varphi(t)}{\varphi(0)} = \left[1 + q^2(1.64 + 0.5\sqrt{\epsilon} + 0.361 \epsilon)/\sqrt{\epsilon}\right]^{-1}.
\end{equation}
This residual level, originally calculated by~\citet{Rosenbluth}, was given to higher accuracy in $\epsilon$ and $q$ by~\citet{Xiao2006}.
To perform the Rosenbluth-Hinton test using global \texttt{stella}, we run a multiple flux-tube linear simulation with velocity space resolution $(N_{v_\parallel}, N_\mu) = (192,6)$, $N_z = 128$, $N_x = 43$ for the side domains and $N_x = 85$ for the central domain. The radial extent of the side domains is $\ell_x= 300\rho_\mr{i}$ while for the center it is $\ell_x = 600\rho_\mr{i}$.
Flat density and temperature profiles are used for all simulations, while the magnetic geometry is taken from the CBC, along with $q''a^2 = 0.5$ and $\rho_\ast = 0.0002$ ($\Delta = 0.12$). Each boundary region uses eight collocation points and no dissipation is used, and adiabatic electrons are employed. Simulations are initialized with a single large-wavelength $k_x\rho_\mr{i} = 2\uppi / 300$ mode, and after some time a residual level is eventually maintained. The left panel of figure~\ref{fig:RH_test} shows the residual level of the zonal flow at the center of the middle domain (at $\rho = 0.5$) along with the analytical prediction given by~\eqref{eqn:RH_res} (this denoted with a dotted line). It is clear that global \texttt{stella} captures the correct residual level at the center of the domain.
While the original Rosenbluth-Hinton calculation is performed in the local limit, we can let~\eqref{fig:RH_test} be a function of $r$ through $\epsilon$ and $q$ and ascertain how well \texttt{stella} attains this `global' version of the residual level. The right panel of figure~\ref{fig:RH_test} displays the final zonal potential at the end of the simulation (purple line), as well as the initial potential profile scaled by the predicted residual fraction given by~\eqref{eqn:RH_res} (green line). At least at this modest $\rho_\ast$, the global simulation captures the correct residual level over the entire domain.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/RH}
\caption{(\emph{left}) Rosenbluth-Hinton residual of the electrostatic potential $\varphi$ (solid line) at the center of the radial domain for a global simulation. Dotted line denotes expected residual level. (\emph{right}) Comparison of the (time-averaged) electrostatic potential in its final state (solid line) to the initial electrostatic potential scaled by the expected radially dependent residual fraction (dotted line) as given by~\citet{Xiao2006}. Shaded area denotes the boundary region of the central flux tube.}
\label{fig:RH_test}
\end{figure}
\subsection{Sources and sinks}
\begin{figure}
\centering
\includegraphics{plots/sources_both}
\caption{Evolution of the box-averaged temperature as a function of time from simulations using a variety of sources, as well as a simulation with no source. Simulations with sources where time averaging is performed using a short-time (long-time) window are denoted with solid (dash-dotted) lines. Simulations in the top panel are performed with the Krook operator in the boundary region, while simulations in the bottom panel are performed without this operator. }
\label{fig:sources}
\end{figure}
In general, sources and sinks are required for global nonlinear simulations because a mismatch of flux between the two boundaries can lead to an accumulation or deficit of particles, momentum and heat in the central domain. Our scale-separated approach to global gyrokinetics involves a novel projection operator source that is detailed in \S \ref{sec:sources}. Here, we demonstrate the efficacy of this new source by comparing it to global simulations using a more conventional Krook-based source, as well as comparing it to global simulations that do not employ sources. These simulations use CBC parameters with adiabatic electrons along with $ q'' a^2 = 5$, $n'' (a^2/n) = -1$ and $T''_\mr{i} (a^2/T_\mr{i}) = -4$. The spatial and velocity resolution for each domain is $(N_x, N_y,N_z,N_\mu, N_{v_\parallel}) = (85,85,16,8,48)$, $D_\mr{hyper} = 0.2$, and $\rho_\ast = 10^{-3}$ for the central domain. The binormal extent for all domains is $\ell_y = 20\upi \rho_\mr{i}(\ensuremath{r}_0)$ while the radial extent is $\ell_x = (6/2\upi)\ell_y$. Each boundary region is comprised of six collocation points.
The box-averaged ion temperatures fluctuations from these simulations are plotted in figure~\ref{fig:sources} as a function of time. The black solid line denotes the simulation with no sources, while the orange and blue lines denote simulations with the projection-type and Krook-type sinks, respectively. Simulations in the top panel apply the Krook operator to the three points facing the physical region with $\nu_\mr{BC} = 0.1$, while the simulations in the bottom do not. For the given simulation parameters, the turbulence is strongest in the rightmost domain, resulting in a strong right-going radial flux of heat. As a result, the simulation with no sources and sinks results in an unbounded secular decrease of temperature. On the other hand, simulations employing sources and sinks with long-windowed time averaging eventually asymptote to finite values.
These values for the projection-type operator converge as the averaging window size is increased, and are sufficiently small to prevent profile relaxation.
In the case of the Krook-operator source, the box-averaged ion temperature eventually decays to zero, though the details on how this happens is sensitive to the precise values of $\nu_\mr{S}$ and $\tau_\mr{S}$.
In practice, one should choose a averaging window size that is longer than any of the turbulent timescales of interests; for our global simulations, this is roughly $\Delta^{-1}(a/v_\mr{thi}) \lesssim \tau_\mr{S} \lesssim \rho_\ast^{-1}(a/v_\mr{thi})$. Thus, figure~\ref{fig:sources} indicates that both the Krook-based and projection-operator-based sources are equally effective at controlling the box-averaged temperature in the simulation domain.
\subsection{Radial heat flux profiles}
\sloppy As a final test of our global approach to gyrokinetics, we compare the radial profile of the ion-temperature-gradient-driven nonlinear heat flux (\ref{app:fluxes}) resulting from a global simulation to the heat fluxes obtained at multiple radial positions with local simulations. These simulations use CBC parameters with adiabatic electrons along with $ q'' a^2 = 5$, $n'' (a^2/n) = -1$ and $T''_\mr{i}(a^2/T_\mr{i}) = -4$. These parameters result in the radial profiles displayed in the left panel of figure~\ref{fig:flux_profile}. The spatial and velocity resolution for each domain is $(N_x, N_y,N_z,N_\mu, N_{v_\parallel}) = (171,85,16,8,48)$, $D_\mr{hyper} = 0.2$, and $\rho_\ast = 10^{-3}$ for the central domain. The binormal extent for all domains is $\ell_y = 20\upi \rho_\mr{i}(\ensuremath{r}_0)$ while the radial extent is $\ell_x = (6/2\upi)\ell_y$. Each boundary region is comprised of ten collocation points, with the Krook operator being applied to the six points facing the physical region with $\nu_\mr{BC} = 0.1$. The projection operator source is employed in the centre domain within the physical region with an averaging window time of $\tau_\mr{S}(v_\mr{th,ref}/a) = 50$.
The resulting radial profile of the heat flux is plotted in the right panel of figure~\ref{fig:flux_profile}, which displays the long-time average after the turbulence has reached a saturated state. Additionally, the results of five separate local simulations are plotted using crosses. The shaded region and error bars of the local results are calculated from the standard deviation of the mean for a continuous signal, i.e., the standard deviation divided by $\sqrt{t_\mr{ave}/t_\mr{auto}}$, where $t_\mr{ave}$ is the time window of the averaging and $t_\mr{auto}$ is the autocorrelation time of the signal. Boundary regions are denoted by the shaded areas, and the profile is given as a function of the minor radius $\ensuremath{r}$. Like the simulations from the previous section, the resulting turbulence is stronger on the right side of the simulation domain, which results in a much larger heat flux.
Similar to the results given by~\citet{Candy_globloc},
the heat flux of the global simulation across the radial domain captures the radial variation of the heat fluxes resulting from by local simulations, with some minor mismatch near the boundaries. These discrepancies can be the result of a number of effects, such as profile shearing~\citep{Waltz_1994} and turbulence spreading~\cite{Garbet_1994}. With regards to the latter, the separation between any two local simulations in figure~\ref{fig:flux_profile} is approximately a single poloidal gyroradius $\rho_\mr{pol} = (B/B_\mr{P})\rho_\mr{ref}$, where $B_\mr{P}$ is the poloidal magnetic field strength, and so figure~\ref{fig:flux_profile} is consistent with turbulence spreading across eddies that have a radial extent on the order of $\rho_\mr{pol}$~\citep{Rotation_theory}. In-depth studies of both profile shearing and turbulence spreading our global approach to gyrokinetics are ongoing, and will be reported in a future publication.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{plots/flux_profile_both}
\caption{(\emph{left}) Radial profile of the equilibrium density and temperature, and magnetic safety factor $q$. (\emph{right}) Radial profile of the total flux-surface-averaged heat flux $Q$ (solid line) from a global $\texttt{stella}$ simulation using CBC parameters with $\rho_\ast = 10^{-3}$. Shaded areas on the left and right denote the boundary regions of the central flux tube. Yellow crosses denote local simulations performed at the specified radial location. Vertical error bars and blue shaded area encompassing the solid line are calculated using the standard error of the mean appropriate for a continuous time signal. The asymmetry of the size of the boundary regions in the physical $r/a$ space is due to $q'' \ne 0$. }
\label{fig:flux_profile}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
The gyrokinetic flux-tube code \texttt{stella} has been extended to allow for radially global simulations using a novel hybrid approach that retains the use of a spectral basis. While this is not the first code to take a hybrid approach to global gyrokinetics~\citep{Candy_globloc}, what makes \texttt{stella} unique are the novel radial boundary conditions that couple multiple flux-tube simulations; a subsidiary expansion of the gyrokinetic equation~\citep{Parra_globallocal} that incorporates the effects of radial profile variation in every term, rather than just $\omega_\ast$; and next-order corrections in the magnetic geometry using an extended form of the Miller equilibrium equations.
These new features have been benchmarked using a number of standard test cases, and are readily applicable to a wide variety of physical problems, such as turbulence spreading~\citep{Garbet_1994,Lin_2004}, and the generation of intrinsic rotation due to the breaking of symmetries in the gyrokinetic equation~\citep{Rotation_theory,Wang_2010,Waltz_2011,Hornsby_2018}. We stress that while our novel radial boundary condition provides a more physically motivated alternative to the standard Dirichlet boundary condition, it does not obviate the need for Krook-type operators which smoothly connect the buffer regions to the physical one.
Global \texttt{stella} could be usefully extended to include electromagnetic effects and arbitrary profile variation using an equilibrium solver such as EFIT~\citep{EFIT_1985}. These additions would respectively enable global simulation of high-$\beta$ tokamaks and regions of high pressure gradients such as the tokamak pedestal, the latter of which may require the resolution of $\rho_\ast$-small physical effects with high fidelity. Finally, integrating the non-perturbative solution of quasineutrality with \texttt{stella}'s implicit solver of the parallel streaming term will lead to significant numerical savings when performing global simulation with kinetic electron effects, and may enable global multiscale simulations with relatively modest computational costs~\citep{Howard_2016}.
As a concluding remark, while we offer an approach to global gyrokinetics using flux-tubes, it is not the only one. Spectral methods for non-periodic domains has long been an important subject of research~\citep{Wang_1998, Agnon_1999,Boyd_2007,Huybrechs_2010}, and spectrally accurate gyroaveraging schemes for non-periodic domains have also been recently developed~\citep{guadagni_cerfon_2017}. While still in its infancy, hybrid global-local gyrokinetics will offer a physically robust and computationally efficient complement to the already vast array of global gyrokinetic codes available to the plasma physics community.
\section*{Source code}
The \texttt{stella} source code is available at the GitHub repository~\url{https://github.com/stellaGK/stella}, while documentation can be found at ~\url{https://stellagk.github.io/stella/}.
\section*{Data availability statement}
The input files that are used for the numerical tests featured in this manuscript, as well as instructions for their use, are
available at the following URL: \url{https://ora.ox.ac.uk/objects/uuid:68e46124-82e5-4a0b-ac78-d8f20bacdc59}.
\section*{Acknowledgements}
The authors would like to thank the reviewers for their careful reading of the manuscript, as well as their helpful comments and suggestions. This work has been carried out within the framework of the EUROfusion Consortium and has received funding from the Euratom research and training programme 2014-2018 and 2019-2020 under grant agreement No 633053. The views and opinions expressed herein do not necessarily reflect those of the European Commission. This work was supported in part by the Engineering and Physical Sciences Research Council (EPSRC) [Grant Number EP/R034737/1]. The authors acknowledge EUROfusion, the EUROfusion High Performance Computer (Marconi-Fusion) under the project MULTEI and OXGK, and the use of ARCHER through the Plasma HEC Consortium [EPSRC Grant Number EP/R029148/1] under the project e607. This work was also supported by the U.S. Department of Energy under contract number DE-AC02-09CH11466. The United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes.
|
\section{Introduction}
In this work we consider the solution $u=u(t,x)$ of the one-dimensional integro-differential equation
\begin{equation}\label{nonlocal}
\partial_t u (t,x)=\int_\mathbb{R} J(y)\left[u(t,x-y)-u(t,x)\right]d y+f\left(u(t,x)\right),\quad t>0,\, x\in\mathbb{R},
\end{equation}
supplemented with some compactly supported initial data $u(0,x)=u_0(x)$. The function $J:\mathbb{R}\to \mathbb{R}$ is a nonnegative dispersal kernel of total mass 1 allowing, in population dynamics models, to take into account long distance dispersal events. The nonlinearity $f=f(u)$ encodes the demographic assumptions and, in this manuscript, we focus on the case where it is of the {\it bistable} or of the {\it ignition} type. Precise assumptions will be given later on. Our concern is to understand the {\it threshold phenomenon} for propagation. As far as we know, we provide the first quantitative estimates for such phenomenon in \eqref{nonlocal}.
\medskip
The solutions of \eqref{nonlocal} share some properties with the ones of the local diffusion equation
\begin{equation}\label{local}
\partial_t u (t,x)=\partial_{xx}u(t,x)+f\left(u(t,x)\right),\quad t>0,\,x\in\mathbb{R}.
\end{equation}
In particular both problems exhibit a \textit{threshold phenomenon}, meaning that \lq\lq small" initial data lead to extinction, whereas \lq\lq large" initial data lead to propagation. In the local diffusion case \eqref{local}, we refer to \cite{Aro-Wei-78} for such a property, while the {\it sharp} threshold phenomenon was more recently investigated \cite{zlatovs2006sharp}, \cite{matano2010}, \cite{Pol-11}, \cite{Mur-Zho-13, Mur-Zho-17}, through different technics. On the other hand, the nonlocal diffusion case \eqref{nonlocal} is more delicate, see subsection \ref{ss:towards} below, but some progresses were recently achieved \cite{alfaro2017fujita}, \cite{berestycki2017non}, \cite{limlong}.
Very recently, the authors (including the first two authors of the present work) of \cite{alfaro2020quantitative} have provided a sharp {\it quantitative} estimate of the threshold phenomenon in \eqref{local}. To be more precise, let $f=f(u)$ be of the ignition or bistable type between $0$ and $1$, with threshold $\theta\in (0,1)$, and \lq\lq $1$ more stable than $0$'' (see Assumption \ref{ass:f} below). Then, for any $\epsilon \in (0,1-\theta)$ there are two lengths $0<L_\epsilon^{\rm ext}<L_\epsilon^{\rm prop}<+\infty$ such that the solution $u_L^\epsilon=u_L^\epsilon(t,x)$ of \eqref{local} starting from
$$
\phi_L^\epsilon(x):=(\theta+\epsilon) \mathds{1} _{(-L,L)}(x),
$$
satisfies
\begin{eqnarray*}
\lim_{t\to+\infty} u_L^\epsilon (t,x)=\begin{cases} 0 & \text{ uniformly for $x\in\mathbb{R}$, if $0<L<L_\epsilon^{\rm ext}$},\\
1 & \text{ locally uniformly for $x\in\mathbb{R}$, if $L>L_\epsilon^{\rm prop}$}.
\end{cases}
\end{eqnarray*}
Among others, the study \cite{alfaro2020quantitative} provides some estimates of the threshold values $L_\epsilon^{\rm ext}$ and $L_\epsilon^{\rm prop}$ as $\epsilon \to 0^+$, namely
$$
0<\liminf_{\epsilon \to 0+} \frac{L_\epsilon^{\rm ext}}{-\ln \epsilon }\leq \limsup_{\epsilon\to 0^+} \frac{L_\epsilon^{\rm prop}}{-\ln \epsilon}<+\infty.
$$
As far as we know, for integro-differential equation such as \eqref{nonlocal} there is no such quantitative estimates of this threshold phenomenon in the literature. The goal of the present work is to fill this gap by first deriving sufficient conditions for the existence of similar critical lengths $L_\epsilon^{\rm ext}$ ad $L_\epsilon^{\rm prop}$ for suitable levels $\epsilon>0$, and secondly to provide estimates of these quantities in some asymptotic regime, namely when the height $\theta+\epsilon$ of the initial data tends to $\theta$.
Note that the analysis in \cite{alfaro2020quantitative} is based on the explicit formula of the heat kernel. For the nonlocal diffusion case, the expression of the \lq\lq corresponding heat kernel'', see \eqref{def-K}--\eqref{def-psi}, is much more complicated and estimating its behaviour is far from obvious. However, refined estimates for the tails of the kernel $J$, as well as for its $n-$folds convolution $J^{*(n)}$, can be obtained for a large variety of kernels. Such information, mainly coming from the probability theory, combined with the construction of accurate sub- and super-solutions will allow us to derive some precise estimates of the threshold phenomenon for the nonlocal diffusion equation \eqref{nonlocal}.
It turns outs that these estimates of the threshold phenomenon strongly depend on the decay rate at infinity for the kernel $J$, or {\it tails} of $J$. For instance, when the kernel $J$ has an exponential decay then the critical lengths (when exist) both behave like for the local problem \eqref{local}. On the other hand, for kernels with {\it heavy} tails (algebraic decay, Weibull-like tails, etc.), such estimates become more complicated and can take a large variety of different forms.
\medskip
Let us first present the assumptions on the dispersal kernel $J$ and the nonlinearity $f$.
\begin{assumption}[Dispersal kernel]\label{ass:kernel} The kernel $J:\mathbb{R}\to \mathbb{R}$ satisfies the following.
\begin{itemize}
\item [(i)] $J$ is nonnegative, even and satisfies $\int_\mathbb{R} J(x)dx=1$;
\item [(ii)] $J$ is nonincreasing in $(0,+\infty)$.
\end{itemize}
\end{assumption}
\begin{assumption}[Expansion form]\label{ass:kernel-bis}
The Fourier transform of $J$ has an expansion
\begin{equation}
\label{J-Fourier-ass}
\widehat J(\xi)=1-a\vert \xi\vert ^{\beta}+o(\vert \xi \vert ^{\beta}), \quad \text{ as } \xi \to 0,
\end{equation}
for some $0<\beta\leq 2$ and $a>0$.
\end{assumption}
As observed and proved in \cite{chasseigne2006asymptotic}, expansion \eqref{J-Fourier-ass} plays a crucial role in the behavior of the linear equation $\partial _t u=J*u-u$, but also in some nonlinear phenomena, such as the Fujita blow up phenomenon and the hair trigger effect \cite{alfaro2017fujita}.
Notice that expansion \eqref{J-Fourier-ass} contains some information on the tails of $J$. Indeed, for kernels which have a finite second momentum, namely $m_2(J):=\int_{\mathbb{R}}x^2J(x)d x<+\infty$, expansion \eqref{J-Fourier-ass} holds true with $\beta =2$, as can be seen in \cite[Chapter 2]{durrett2019probability} among others. In particular, this is the case for kernels which are compactly supported, exponentially bounded, or which decrease like $\mathcal O\left(\frac{1}{\vert x\vert^{3+\epsilon}}\right)$ with $\epsilon>0$. On the other hand, when $m_2(J)=+\infty$ then more general expansions are possible. For example, for algebraic tails satisfying
\begin{equation}
\label{heavytails}
J(x)\sim \frac{C}{\vert x\vert ^{\alpha}} \quad \text{ as }\vert x\vert \to \infty, \quad\text{ with } 1<\alpha< 3,
\end{equation}
then \eqref{J-Fourier-ass} holds true with $\beta= \alpha- 1\in(0,2)$. This fact is related to the stable laws of index $\beta\in(0,2)$ in probability theory, and a proof can be found in \cite[Chapter 2, subsection 2.7]{durrett2019probability}. In particular it contains the case of the Cauchy law $J(x)=\frac{1/\pi}{1+x^2}$, for which
$$
\widehat J(\xi)=e^{-\vert\xi\vert}=1-\vert\xi\vert +o(\vert \xi\vert), \quad \text{ as } \xi \to 0,
$$
and $\beta =1$, despite the nonexistence of the first moment $m_1(J):=\int \vert x\vert J(x)d x$.
\medskip
Throughout this note the reaction term arising in \eqref{nonlocal} will satisfy the following set of hypothesis.
\begin{assumption}[Nonlinearity]\label{ass:f}
The function $f: \mathbb{R}\to\mathbb{R}$ is Lipschitz continuous. There is a threshold $\theta\in (0, 1)$ such that
\begin{equation}\label{0-theta-1}
f(u)=0,\;\forall u\in\{0, \theta, 1\},
\end{equation}
and
\begin{equation}\label{bist-ignition}
f(u)>0,\;\forall u\in(\theta, 1), \quad\text{and}\quad
\begin{cases}
f(u)<0,\;\forall u\in (0, \theta), \quad(\text{Bistable Case}),\\
\text{or}\\
f(u)=0,\;\forall u\in (0, \theta), \quad(\text{Ignition Case}).
\end{cases}
\end{equation}
In the bistable case, we further require
\begin{equation}\label{positive}
\int_{0}^{1}f(s) ds>0,
\end{equation}
Moreover, in both cases, we require that there are $r^->0$ and $\delta\in (\theta,1)$ such that
\begin{equation}\label{r-moins}
f(u)\geq r^-\left(u-\theta\right),\;\forall u\in [0,\delta].
\end{equation}
\end{assumption}
Notice that for $r>0$, the usual cubic bistable nonlinearity
$$
f(u)=ru(u-\theta)(1-u)\mathds{1}_{(0, 1)}(u)
$$
satisfies the above assumptions as soon as $\theta<\frac{1}{2}$, and so does the ignition nonlinearity
$$
f(u)=r(u-\theta)(1-u)\mathds{1}_{(\theta, 1)}(u).
$$
\medskip
The organization of this work is as follows. In section \ref{s:basic}, we list some basic facts and key lemmas, in particular some estimates for the tails of $n-$folds convolution kernel (coming from sum of i.i.d. random variables in probability theory).
In section \ref{main results} we state our main results including our quantitative estimates on the threshold phenomenon \lq\lq extinction vs. propagation''. In sections \ref{extinction} and \ref{non-extinction} we inquire on extinction and non-extinction phenomena, respectively, in some related toy models.
Next, in section \ref{quantitative}, we build on these preliminary results to prove our main results. Finally, in section \ref{propagation} we establish a sufficient condition for \lq\lq propagation to occur'', which is another one of our main results.
\section{Preliminaries}\label{s:basic}
In this section, after presenting some notations and basic facts, we collect some estimates for the tails of $J^{*(i)}$ ($i\geq 1$) coming from the probability theory.
\subsection{Notations and basic linear facts}\label{ss:linear}
If $f\in L^1(\mathbb{R})$, we define its Fourier transform $\mathcal{F}(f) = \widehat{f}$ and its inverse Fourier transform $\mathcal{F}^{-1} (f)$ by
\begin{equation*}
\widehat{f}(\xi):=
\int_{\mathbb{R}} e^{-i\xi x} f(x) dx, \quad \mathcal{F}^{-1}(f)(x)
:=\int_{\mathbb{R}} e^{ix\xi} f(\xi) d\xi.
\end{equation*}
With this definition, we have, for $f$, $g\in L^1(\mathbb{R})$,
$$
\widehat{f*g}=\widehat f \, \widehat g,
$$
and $f=\frac{1}{2\pi}\mathcal F ^{-1}(\mathcal F(f))$ if $f$, $\mathcal F (f) \in L ^1(\mathbb{R})$.
\bigskip
Formally applying the Fourier transform to equation
\begin{equation}\label{nonlocal-heat}
\partial _t u=J*u-u
\end{equation}
yields
$$
\frac{d}{dt}\widehat{u}(t, \xi) = \widehat{u}(t, \xi) (\widehat{J}(\xi)-1),
$$
where $\xi$ is seen as a parameter and which is solved as
\begin{equation*}
\widehat{u}(t, \xi) = e^{t(\widehat{J}(\xi)-1)} \widehat{u}_0(\xi).
\end{equation*}
Applying the inverse Fourier transform, we see that the fundamental solution writes as
\begin{equation}\label{def-K}
K(t, \cdot)=e^{-t}\delta_0+\psi(t, \cdot),
\end{equation}
where
\begin{equation}\label{def-psi}
\psi(t,\cdot) := e^{-t} \sum_{i=1}^{\infty} \frac{t^i}{i!} J^{\ast(i)},
\end{equation}
where $\delta_0$ is the Dirac mass at $0$ and $J^{\ast(i)}:=J\ast\cdots\ast J$ is the convolution of $J$ with itself $i-1$ times.
Hence the unique bounded solution to $\partial _tu=J*u-u$ with initial data $u_0\in L^\infty(\mathbb{R})$ is given by
$$
u(t, x) = K(t, \cdot)\ast u_0(x)=e^{-t}u_0(x) + \psi(t, \cdot) \ast u_0(x).
$$
Obviously, though the convolution of a Dirac mass by a $L^\infty$ function is not pointwise well defined, we let $\delta_0\ast u_0=u_0$. Moreover, from the normal convergence of the series
$\sum_{i=1}^{\infty} \frac{t^i}{i!} J^{\ast(i)}$
and $\sum_{i=1}^{\infty} \frac{t^{i-1}}{(i-1)!} J^{\ast(i)}$ in $C([0, T], L^1(\mathbb{R}))$, we deduce that the function $t\in[0, \infty)\to\psi(t, \cdot)\in L^1(\mathbb{R})$ is of class $C^1$ and that
\begin{equation}\label{partial}
\partial_t \psi(t, x) = J\ast \psi (t, \cdot)(x) - \psi (t, x) + e^{-t} J(x).
\end{equation}
Notice also that
\begin{equation}\label{1-e^t}
\int_{\mathbb{R}} \psi (t, x) dx = 1-e^{-t}, \quad \forall t\ge0.
\end{equation}
\subsection{Some key estimates on the tails of $J^{*(i)}$ ($i\geq 1$)}\label{ss:linear2}
Observe that if $J\in L^1(\mathbb{R})$ satisfies Assumption \ref{ass:kernel-bis} then so does $J^{*(i)}$, with $i\geq 1$, but with the expansion
\begin{equation}
\label{J-Fourier-ass-i}
\widehat {J^{*(i)}} (\xi)=1-i\,a\vert \xi\vert ^{\beta}+o(\vert \xi \vert ^{\beta}), \quad \text{ as } \xi \to 0.
\end{equation}
Notice also that $\widehat{J}$ is a real function since $J$ is even. In the sequel, for $i\geq 1$ and $L>0$, we denote
\begin{equation}
\label{def-R-i-L}
R_i(L):=\int _{\vert x\vert \geq L} J^{*(i)}(x)dx
\end{equation}
the tails of $J^{*(i)}$. The expansion \eqref{J-Fourier-ass-i} on the low frequencies of the Fourier transform $\widehat {J^{*(i)}}$ is very related to the tails of $J^{*(i)}$. The following first bound is provided
by \cite[Chapter 2 subsection 2.3.c (3.5)]{durrett2019probability}.
\begin{lemma}\label{lem:Durret}
Let Assumption \ref{ass:kernel}-(i) be satisfied. For any $i\geq 1$ and $L>0$, one has
\begin{equation}
R_i(L)\leq \frac{L}{2}\int_{-\frac{2}{L}}^{\frac{2}{L}} \left[1-\left(\widehat J(\xi)\right)^i\right] d\xi.
\end{equation}
\end{lemma}
As $L\to+\infty$, we have the following result quoted from \cite[Theorem 5]{pitman1968}.
\begin{lemma}\label{lem:tails}
Let Assumptions \ref{ass:kernel}-(i) and \ref{ass:kernel-bis} hold.
\begin{itemize}
\item [(i)]
If $\beta=2$ then, for any $i\geq 1$, one has
\begin{equation}\label{tails-beta-equal-2}
\int _0^L uR_i(u) du \sim i\,a, \quad \text{ as } L\to +\infty.
\end{equation}
\item [(ii)]
If $0<\beta <2$ then there is $C=C(\beta)>0$ such that, for all $i\geq 1$,
\begin{equation}\label{tails-beta-lower-2}
R_i(L) \sim i \,\frac{aC}{L ^{\beta}}, \quad \text{ as } L\to +\infty.
\end{equation}
\end{itemize}
\end{lemma}
When $0<\beta <2$, the asymptotic behavior of $R_i(L)$ is thus very precisely described by Assumption \ref{ass:kernel-bis}. On the other hand, when $\beta=2$, Assumption \ref{ass:kernel-bis} is not enough to capture the asymptotic behavior of $R_i(L)$: precise estimates, as needed in this work, strongly depend on the decay rate of $J$ at infinity. Below we discuss such estimates for different types of kernels.
The following result deals with exponentially bounded kernels and is known, in probability theory, as the large deviations Cram\'er theorem (adapted to our situation). We refer the reader to \cite[section 2.2]{Dembo-Zeitouni} for more details and proofs. The uniformity of the limit \eqref{cramer} as stated below is a consequence of \cite{hoglund}, see Appendix \ref{s:appendix}.
\begin{lemma}\label{lem:expo}
Let Assumption \ref{ass:kernel}-(i) be satisfied and assume further
\begin{equation}\label{expo-decay}
\exists \lambda _0>0,\;\;\int_\mathbb{R} e^{\lambda _0 x} J(x)d x<+\infty.
\end{equation}
Define the logarithmic moment generating function $\Lambda:\mathbb{R}\to (-\infty,+\infty]$ as
$$
\Lambda(\lambda):=\ln \left(\int_\mathbb{R} e^{\lambda x}J(x)d x\right),
$$
and its Fenchel-Legendre transform as
$$
\Lambda ^*(x):=\sup_{\lambda \in \mathbb{R}} ( \lambda x -\Lambda(\lambda)).
$$
Then, $\Lambda ^*$ is nondecreasing on $(0,+\infty)$. Also, there exists $0<L_1$ such that $\Lambda^*(L)<+\infty$ for all $L\in [0,L_1]$ and, for all $0<L_0<L_1$, one has
\begin{equation}\label{cramer}
\lim_{i\to+\infty}\frac{1}{i}\ln R_i(iL) =-\Lambda ^*(L), \quad \text{ uniformly for $L\in [L_0,L_1]$}.
\end{equation}
Moreover we also have for any $L>0$, $i\geq 1$ and $\lambda \in \mathbb{R}$,
\begin{equation}\label{esti-expo}
R_i(iL)\leq e^{-i \Lambda^*(L)}\leq e^{-i(\lambda L- \Lambda(\lambda))}.
\end{equation}
\end{lemma}
\medskip
When the kernel $J$ is not exponentially bounded (but still $\beta=2$), there is a large variety of possible tails. The following examples are typical, see e.g. \cite{mikosch1998large}, and will be considered in the following.
\medskip
\noindent {\bf Regularly varying tails $RV(\alpha)$}:
\begin{equation}\label{tail-RV}
R_1(L)=L^{-\alpha} S(L),\quad \forall L>0,
\end{equation}
for $\alpha>2$ and where $S:(0,+\infty)\to(0,+\infty)$ is a slowly varying function (that is, for all $\nu>0$, $S(\nu x)\sim S(x)$ as $x\to+\infty$).
\medskip
\noindent {\bf Lognormal-type tails $LN(\gamma, \lambda, \rho)$:}
\begin{equation}\label{tail-LN}
R_1(L) \sim cL^{\rho}e^{-\lambda \ln^\gamma L},\quad \text{ as } L\to+\infty,
\end{equation}
for some $\gamma>1$, $\lambda>0$, $\rho \in \mathbb{R}$ and appropriate constant $c=c(\rho, \gamma)$.
\medskip
\noindent {\bf Weibull-like tails $WE(\alpha,\lambda,\rho)$:}
\begin{equation}\label{tail-WE}
R_1(L)\sim cL^{\rho}e^{-\lambda L^\alpha},\quad \text{ as } L\to+\infty,
\end{equation}
for some $0<\alpha<1$, $\lambda>0$, $\rho \in \mathbb{R}$ and appropriate constant $c=c(\rho, \alpha)$.
\medskip
For these three type classes of kernels, refined estimates of $R_i(L)$ when $i\gg 1$ and $L\gg 1$ are known in the literature: the next lemma is taken from \cite[Proposition 3.1]{mikosch1998large}.
\begin{lemma}\label{LDH}
For the three classes of kernels above, assume $J$ is normalized by $m_2(J)=1$. Then define the threshold sequence $d_n$ accordingly to
\begin{center}
\begin{tabular}{|l|l|
\hline
\qquad Distribution & \qquad $d_n$\\
\hline
$RV(\alpha)$, \quad\qquad $\alpha>2$ & $n^{1/2}\ln^{1/2} n $\\
$LN(\gamma,\lambda,\rho)$, \quad $1<\gamma\leq 2$ & $n^{1/2}\ln^{\gamma/2}n$\\
$LN(\gamma,\lambda,\rho)$, \quad $2<\gamma$ & $n^{1/2}\ln^{\gamma-1}n$\\
$WE(\alpha,\lambda,\rho)$, \; $0<\alpha<1$ & $n^{1/(2-2\alpha)}$\\
\hline
\end{tabular}\\
\medskip
Table 1: The threshold sequences $d_n$.
\end{center}
Then, for any sequence $\gamma_n\gg d_n$, one has
\begin{equation}\label{conclusion-table1}
\lim_{i\to+\infty} \sup_{L\geq \gamma_i}\left\vert \frac{R_i(L)}{iR_1(L)}-1\right \vert=0.
\end{equation}
\end{lemma}
\medskip
Last, for some specific forms of Weibull-like tails satisfying ($0<\alpha <1$)
\begin{equation}
\label{tail-weibull-zero}
R_1(L)\sim c e^{-L^\alpha}, \quad \text{ as }L\to+\infty,
\end{equation}
more refined estimates are known: according to \cite[(2.32)]{Nagaev1979} there exists some constant $C>0$ such that, for all $L>0$ and $i\geq 1$,
\begin{equation}\label{WEbe=0}
R_i(L)\leq C\left[\exp\left(-\frac{L^2}{20 i}\right)+iR_1\left(\frac L 2\right)\right].
\end{equation}
\section{Main results}\label{main results}
For $\epsilon\in(0, 1-\theta]$ and $L>0$, we consider the family of initial data $\phi_L^\epsilon$ given by
\begin{equation}\label{initial}
\phi_L^\epsilon(x):=(\theta+\epsilon)\mathds{1}_{(-L,L)}(x),\quad x\in\mathbb{R},
\end{equation}
wherein $\mathds{1}_A$ denotes the characteristic function of the set $A$. We denote by $u_L^\epsilon=u_L^\epsilon(t,x)$ the solution of \eqref{nonlocal} starting from the initial datum $\phi_L^\epsilon$.
The results presented in this section are concerned with the derivation of asymptotic estimates ($\epsilon\ll 1$) of the size $L$ such that the solution $u_L^\epsilon(t,x)$ goes extinct or propagates at large times. We split our main results into two parts. The first subsection is related to extinction while the second is concerned with propagation.
\subsection{Extinction results}\label{ss:ext}
Let us recall that, for $i\geq 1$ and $L>0$, $R_i(L)$ denotes the tails of $J^{*(i)}$ as defined in \eqref{def-R-i-L}. Our general extinction criterion reads as follows.
\begin{theorem}[Extinction]\label{ext}
Let Assumptions \ref{ass:kernel} and \ref{ass:f} be satisfied. Let $\epsilon>0$ be fixed and set
\begin{equation}\label{r^+}
r^+:=\sup_{u\in (\theta,1]} \frac{f(u)}{u-\theta}>0.
\end{equation}
Assume $T>0$ and $L>0$ are such that
\begin{equation}\label{criterion}
\theta e^{-(r^++1)T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}R_i(L)\geq \epsilon.
\end{equation}
Then the function $u_L^\epsilon$ is going to extinction at large times, namely
\begin{equation*}
\lim_{t\to+\infty}\sup_{x\in\mathbb{R}}u_L^\epsilon(t, x)=0.
\end{equation*}
\end{theorem}
The above result has various implications exploring the asymptotic size ensuring the extinction of the solution. More precisely, for each $\epsilon\in (0,1-\theta]$, define the interval
$$
\mathcal L_\epsilon^{\rm ext}:=\left\{L>0:\;\lim_{t\to+\infty}\sup_{x\in\mathbb{R}}u_L^\epsilon(t, x)=0\right\},
$$
as well as
\begin{equation}\label{DEF-Lext}
L_\epsilon^{\rm ext}:=\begin{cases}\sup\mathcal L_\epsilon^{\rm ext}\in (0,\infty] & \text{ if }\; \mathcal L_\epsilon^{\rm ext}\neq \emptyset,\\
0 & \text{ if } \;\mathcal L_\epsilon^{\rm ext}=\emptyset.
\end{cases}
\end{equation}
The next corollaries are concerned with a lower bounded of $L_\epsilon^{\rm ext}$ as $\epsilon\to 0^+ $.
We start with the case where the kernel $J$ has a slow decay at infinity, meaning that $0<\beta<2$ in the expansion \eqref{J-Fourier-ass}.
\begin{corollary}[Asymptotic extinction criterion, $0<\beta<2$]\label{asmptotic1}
Let Assumptions \ref{ass:kernel}, \ref{ass:kernel-bis} with $0<\beta <2$, and \ref{ass:f} be satisfied. Then
\begin{equation*}
\liminf_{\epsilon \to 0^+} \; \epsilon^{\frac{1}{\beta}}L^{\rm ext}_\epsilon\geq C^-:= \left(\theta aC e^{-1}/r^+ \right)^{1/\beta}>0,
\end{equation*}
where $a>0$ and $C=C(\beta)>0$ come from the asymptotic expansion \eqref{J-Fourier-ass} and Lemma \ref{lem:tails}-$(ii)$, respectively.
\end{corollary}
Next, we explore a similar question when $\beta=2$. We start with the case of exponentially bounded kernel.
\begin{corollary}[Asymptotic extinction criterion for exponentially bounded kernels]\label{asmptotic2} Let Assumptions \ref{ass:kernel} and \ref{ass:f} be satisfied and assume further the exponential decay \eqref{expo-decay}. Then
\begin{equation*}
\liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{\ln \frac{1}{\epsilon}}\ge C^->0,
\end{equation*}
for some $C^-=C^-(r^+,J)>0$.
\end{corollary}
We pursue the $\beta=2$ case by presenting a series of classical kernels presented in subsection \ref{ss:linear2}.
\begin{corollary}[Asymptotic extinction criterion, some examples with $\beta=2$]\label{ext:examples}
Let Assumptions \ref{ass:kernel} and \ref{ass:f} be satisfied. Then the following estimates hold.
\begin{itemize}
\item [(i)] Assume that there are $c_2>c_1>0$, $\alpha>2$ and $x_0\in \mathbb{R}$ such that $ \frac{c_1}{\vert x \vert^{1+\alpha}}\leq J(x) \leq \frac{c_2}{\vert x \vert^{1+\alpha}}$ for all $\vert x\vert \geq \vert x_0\vert $. Then
$$
\liminf_{\epsilon \to 0^+}\; \epsilon^{\frac {1} {\alpha}}L_\epsilon^{\rm ext}\ge C^->0,
$$
for some $C^-=C^-(\theta,r^+,T,c_1,\alpha)>0$.
\item[(ii)] Assume that $J$ has regularly varying tails $RV(\alpha)$ with $\alpha>2$. Then
$$
\liminf_{\epsilon \to 0^+}\; \epsilon^{\frac {1} {\alpha'}}L_\epsilon^{\rm ext}\ge 1,
$$
for any given $\alpha'>\alpha$.
\item [(iii)] Assume that $J$ has lognormal-type tails $LN(\gamma,\lambda,\rho)$ with $\gamma>1$, $\lambda>0$ and $\rho \in \mathbb{R}$. Then
$$
\liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{e^{\left(\frac{1}{\lambda '}\ln \frac 1 \epsilon\right)^{\frac 1 \gamma}}}\ge 1,
$$
for any given $\lambda '>\lambda$.
\item [(iv)] Assume that $J$ has Weibull-like tails $WE(\alpha,\lambda,\rho)$ with $0<\alpha <1$, $\lambda>0$ and $\rho \in \mathbb{R}$. Then
$$
\liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{\left(\frac 1{\lambda '}\ln \frac 1 \epsilon\right)^{\frac1\alpha}}\ge 1,
$$
for any given $\lambda '>\lambda$. Furthermore, if $\rho=0$,
$$
\liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{\left(\frac 1{\lambda }\ln \frac 1 \epsilon\right)^{\frac1\alpha}}\ge C^-,
$$
for some $C^-=C^-(\theta,r^+,T,J)>0$.
\end{itemize}
\end{corollary}
\subsection{Propagation results}\label{ss:prop}
To state our propagation results, we need to assume that for any level $\theta+\epsilon>\theta$, there exists a size $L$ such that the solution $u_L^\epsilon$ of \eqref{nonlocal} with initial data \eqref{initial} propagates. This assumption will be largely discussed below.
\begin{assumption}[Propagation assumption]\label{ass:prop} Let Assumptions \ref{ass:kernel}-(i) and \ref{ass:f} hold. We assume that for each $\epsilon\in (0,1-\theta]$ there exists $L>0$ large enough such that $u_L^\epsilon(t, x)$ propagates, in the sense that
$$
\lim_{t\to+\infty} \; u_L^\epsilon(t,x)=1\; \text{ locally uniformly for on $\mathbb{R}$}.
$$
\end{assumption}
Now using this and similarly as above, for each $\epsilon\in (0,1-\theta]$, we define the non-empty interval
$$
\mathcal L_\epsilon^{\rm prop}:=\left\{L>0:\;\lim_{t\to+\infty}u_L^\epsilon(t, x)=1\text{ locally uniformly on $\mathbb{R}$}\right\},
$$
as well as the quantity
\begin{equation}\label{DEF-Lprop}
L_\epsilon^{\rm prop}:=\inf\mathcal L_\epsilon^{\rm prop}\in [0,\infty).
\end{equation}
We now state our general propagation criterion (let us recall that $\delta\in(\theta,1)$ is coming from \eqref{r-moins}).
\begin{theorem}[Propagation]\label{theo-propagation}
Let Assumptions \ref{ass:kernel}-(i), \ref{ass:f} and \ref{ass:prop} holds. For any $\alpha\in (0,\delta-\theta)$ and $m\in (0,1)$
there exists $\epsilon_0>0$ small enough such that, for all $\epsilon\in (0,\epsilon_0)$, one has $L_\epsilon^{\rm prop}\le L$ where $L$ satisfies the following condition
\begin{equation}\label{Condition}
\frac{\epsilon}{2(\theta+\epsilon)}\geq e^{-T_\epsilon }\sum_{i=1}^{+\infty} \frac{T_\epsilon^i}{i!}R_i\left((1-m)L\right),\quad\text{ with } T_\epsilon=\frac{1}{r^-}\ln\frac{2\alpha}{\epsilon},
\end{equation}
where we recall that $R_i(L)$ denotes the tails of $J^{*(i)}$ as defined in \eqref{def-R-i-L},
\end{theorem}
In the following we derive an upper bound of $L_\epsilon^{\rm prop}$ as $\epsilon\to 0^+$. As before, this asymptotic strongly depends on the tails of the kernels $J^{*(i)}$.
\begin{corollary}[Asymptotic propagation criterion, $0<\beta<2$]\label{prop:corollary1}
Let Assumptions \ref{ass:kernel}-(i), \ref{ass:kernel-bis} with $0<\beta<2$, \ref{ass:f} and \ref{ass:prop} be satisfied. Then
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac 1 \beta}}\le C^+,
$$
for some $C^+=C^+(\theta,r^-,J)>0$.
\end{corollary}
Next, we explore a similar question when $\beta=2$. For such cases with $\beta=2$, we assume $r^-\in(0, 1)$, where $r^-$ is from \eqref{r-moins}. We start with the case of exponentially bounded kernel again.
\begin{corollary}[Asymptotic propagation criterion for exponentially bounded kernels]\label{prop:corollary2}
Let Assumptions \ref{ass:kernel}-(i), \ref{ass:f} and \ref{ass:prop} be satisfied. Assume further the exponential decay \eqref{expo-decay} and $r^-\in(0, 1)$. Then
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\ln\frac 1 \epsilon}\le C^+,
$$
for some $C^+=C^+(r^-,J)>0$.
\end{corollary}
We pursue the $\beta=2$ case by presenting a series of classical kernels given in subsection \ref{ss:linear2}.
\begin{corollary}[Asymptotic propagation criterion, some examples with $\beta=2$]\label{prop:corollary3}
Let Assumptions \ref{ass:kernel}-(i), \ref{ass:f} and \ref{ass:prop} be satisfied. In addition, assume $r^-\in(0, 1)$ and $m_2(J)=\int_{\mathbb{R}}x^2J(x)dx=1$. Then we have the following results.
\begin{itemize}
\item [(i)] Assume that there are $c_2>c_1>0$, $\alpha>2$ and $x_0\in \mathbb{R}$ such that $\frac{c_1}{\vert x\vert^{1+\alpha}}\leq J(x)\leq \frac{c_2}{\vert x\vert ^{1+\alpha}}$ for all $\vert x\vert \geq \vert x_0\vert$ then
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac 1 \alpha}}\le C^+,
$$
for some $C^+=C^+(\theta, r^-, c_2, \alpha)>0$.
\item [(ii)] Assume that $J$ has regularly varying tails $RV(\alpha)$ with $\alpha>2$. Then
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac 1 {\tilde \alpha}}}\le 1,
$$
for any given $0<\tilde \alpha<\alpha$.
\item [(iii)] Assume that $J$ has lognormal-type tails $LN(\gamma,\lambda,\rho)$ with $\gamma>1$, $\lambda>0$, $\rho \in \mathbb{R}$. Then
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\exp \left[\left(\frac{1}{\tilde \lambda} \ln \frac 1\epsilon\right)^\frac{1}{\gamma}\right]}\le 1,
$$
for any given $0<\tilde \lambda<\lambda$.
\item [(iv)] Assume that $J$ has Weibull-like tails $WE(\alpha,\lambda,\rho)$ with $0<\alpha<1$, $\lambda>0$, $\rho \in \mathbb{R}$. Then
$$
\limsup_{\epsilon\to 0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac 1{\tilde \lambda}\ln \frac 1\epsilon\right)^{\frac 1\alpha}}\leq 1, \quad \text{ if } \quad 0<\alpha<\frac 2 3,
$$
for any given $0<\tilde \lambda<\lambda$, while
$$
\limsup_{\epsilon\to 0^+}\frac{L_\epsilon^{\rm prop}}{\left(\ln \frac 1 \epsilon \right)^{\frac{1}{2(1-\tilde \alpha)}}}\leq 1, \quad \text{ if } \quad \frac 2 3\leq \alpha <1,
$$
for any given $0<\alpha<\tilde \alpha<1$. Furthermore, if $\rho=0$,
$$
\limsup_{\epsilon\to 0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac 1 \lambda \ln \frac 1 \epsilon \right)^{\frac{1}{\alpha}}}\leq C^+,
$$
for some $C^+=C^+(J)>0$.
\end{itemize}
\end{corollary}
Note that when $\rho=0$ in $(iv)$, the assumption $m_2(J)=1$ is not necessary since we directly rely on estimate \eqref{WEbe=0}.
\subsection{Combining, summarizing and commenting the above results}
Let the dispersal kernel satisfy Assumptions \ref{ass:kernel} and \ref{ass:kernel-bis}. Let the nonlinearity $f$ be of the bistable or ignition type in the sense of Assumption \ref{ass:f}. We considered the solution $u_L^\epsilon=u_L^\epsilon(t,x)$ to the one-dimensional reaction integro-differential equation \eqref{nonlocal} starting from the step function $\phi_L^\epsilon(x):=(\theta+\epsilon)\mathds{1}_{(-L,L)}(x)$. Under the propagation Assumption \ref{ass:prop}, to be discussed in the next subsection, we investigated the values $0<L_\epsilon^{ext}\leq L_\epsilon^{prop}<+\infty$ in the asymptotic regime $\epsilon\to 0^+$, meaning that the height of the initial step function tends to the threshold value $\theta$ of nonlinearity $f$. For a large variety of dispersal kernels, we proved some estimates stated in subsection \ref{ss:ext} and \ref{ss:prop} that we now combine and comment.
\medskip
If $0<\beta <2$ in Assumption \ref{ass:kernel-bis}, meaning in particular that $m_2(J)=\int_{\mathbb{R}}x^2J(x)dx=+\infty$, then
$$
0<\liminf_{\epsilon \to 0^+}\frac{L_\epsilon^{\rm ext}}{\left(\frac 1 \epsilon\right)^{\frac 1 \beta}}\le \limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac 1 \beta}}<+\infty, \quad \text{ (heavy tails)}.
$$
In this situation, the \lq\lq main term'' is of magnitude $\left(\frac 1\epsilon\right)^{\frac1\beta}$, but the lower and upper bounds coincide only up to a logarithmic term.
\medskip
If $m_2(J)=\int_{\mathbb{R}}x^2J(x)dx<+\infty$, meaning $\beta=2$ in Assumption \ref{ass:kernel-bis}, we distinguish the following prototype situations under the additional assumption $r^-\in(0,1)$, that we believe to be only technical.
If $J$ is exponentially bounded as in \eqref{expo-decay} then
$$
0<\liminf_{\epsilon \to 0^+}\frac{L_\epsilon^{\rm ext}}{\ln \frac 1 \epsilon}\le \limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\ln\frac 1 \epsilon}<+\infty, \quad \text{ (exponentially bounded tails)}.
$$
In this situation, the lower and upper bounds are {\it sharp} in the sense that they coincide up to a multiplicative constant.
If $J$ has \lq\lq not heavy algebraic tails'', in the sense that there are
$c_2>c_1>0$, $\alpha>2$ and $x_0\in \mathbb{R}$ such that
$$
\frac{c_1}{\vert x \vert^{1+\alpha}}\leq J(x) \leq \frac{c_2}{\vert x \vert^{1+\alpha}}, \quad \text{ for all } \vert x\vert \geq \vert x_0\vert,
$$
and moreover $m_2(J)=1$, then
$$
0<\liminf_{\epsilon \to 0^+}\frac{L_\epsilon^{\rm ext}}{\left(\frac 1 \epsilon\right)^{\frac 1 \alpha}}\le \limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac 1 \alpha}}<+\infty, \quad \text{ (not heavy algebraic tails)}.
$$
In this situation, the main term is of magnitude $\left(\frac 1\epsilon\right)^{\frac1\alpha}$, but the lower and upper bounds coincide only up to a logarithmic term.
If $J$ has regularly varying tails $RV(\alpha)$ with $\alpha>2$ and $m_2(J)=1$, then for any $0<\tilde \alpha<\alpha<\alpha'$,
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac 1 {\tilde \alpha}}}\leq 1\leq \liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{\left(\frac 1 \epsilon\right)^{\frac {1} {\alpha'}}} \quad \text{ (regularly varying tails)}.
$$
The situation is here less clear than the not heavy algebraic tails since, from the above, it is not clear that the main term is of magnitude $\left(\frac 1\epsilon\right)^{\frac1\alpha}$.
If $J$ has lognormal-type tails $LN(\gamma,\lambda,\rho)$ with $\gamma>1$, $\lambda>0$, $\rho \in \mathbb{R}$ and $m_2(J)=1$, then for any $0<\tilde \lambda<\lambda<\lambda'$,
$$
\limsup_{\epsilon\to0^+}\frac{L_\epsilon^{\rm prop}}{\exp \left[\left(\frac{1}{\tilde \lambda} \ln \frac 1\epsilon\right)^\frac{1}{\gamma}\right]}\le 1 \leq
\liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{\exp\left[\left(\frac{1}{\lambda '} \ln \frac 1 \epsilon\right)^{\frac 1 \gamma}\right]}, \quad \text{ (lognormal-type tails)}.
$$
If $J$ has Weibull-like tails $WE(\alpha,\lambda,\rho=0)$ with $0<\alpha<1$, $\lambda>0$ (see above for the $\rho\neq 0$ case) then
$$
0< \liminf_{\epsilon \to 0^+}\; \frac{L_\epsilon^{\rm ext}}{\left(\frac 1 \lambda \ln \frac 1 \epsilon\right)^{\frac 1\alpha}}\leq \limsup_{\epsilon\to 0^+}\frac{L_\epsilon^{\rm prop}}{\left(\frac 1 \lambda \ln \frac 1 \epsilon \right)^{\frac{1}{\alpha}}}<+\infty, \quad \text{ (Weibull-like tails $WE(\alpha, \lambda, \rho=0)$)}.
$$
In this situation, the main term is of magnitude $\frac{1}{\left(\ln \frac 1 \epsilon\right)^{\frac 1 \alpha}}$ and the lower and upper bounds are sharp.
\medskip
As a conclusion, our analysis reveals that the main term (or, at least, the approximate main term) driving the asymptotics of $L_\epsilon^{ext}$ and $L_\epsilon^{prop}$ is, in some sense, selected by the tails of $J$ and can take a large variety of different forms.
\subsection{Towards Assumption \ref{ass:prop}}\label{ss:towards}
The propagation threshold in nonlocal diffusion problems is of fundamental importance and independent interest. In this subsection, we derive some sufficient conditions for the propagation Assumption \ref{ass:prop} to hold. We place ourselves in the following {\it bistable} situation.
\begin{assumption}[Dispersal kernel and bistable nonlinearity]\label{TWA} The dispersal kernel $J\in C^1(\mathbb{R})$ is nonnegative, even, and satisfies $\int_\mathbb{R} J(x)dx=1$. Assume further $J'\in L^1(\mathbb{R})$ and the existence of a finite first moment, namely
\begin{equation}\label{first_moment}
\int_\mathbb{R} |y|J(y)dy<+\infty.
\end{equation}
The nonlinearity $f\in C^1(\mathbb{R})$ is bistable in the sense of \eqref{0-theta-1}---\eqref{bist-ignition} and further satisfies $f'(0)<0$, $f'(1)<0$.
\end{assumption}
A traveling wave solution $(c,U)$ for \eqref{nonlocal} is a speed $c\in \mathbb{R}$ and a profile $U=U(z)$ solving
\begin{equation}\label{TW}
\begin{cases}
J\ast U-U-cU'+f(U)=0,\quad \text{on }\,\mathbb{R},\\
U(-\infty)=0, \quad U(+\infty)=1,
\end{cases}
\end{equation}
meaning in particular that $u(t,x):=U(x+ct)$ solves \eqref{nonlocal}.
The existence, uniqueness and properties of traveling wave solutions in nonlocal diffusion equations with monostable, ignition or bistable nonlinearities has attracted a lot of attention, see e.g. \cite{Erm-Mac-93}, \cite{bates1997traveling}, \cite{Che-97}, \cite{coville2006uniqueness,coville2007travelling}, \cite{Cov-Dup-07}, \cite{Mel-Roq-Sir-14}, \cite{Gui-Hua-15}, \cite{alfaro2017propagation} and the references therein.
Under Assumption \ref{TWA}, the existence of traveling waves is studied in the important work of Bates et al. \cite{bates1997traveling} (to which we refer for more precise statements): there exists a monotone solution, in the {\it weak} sense, to \eqref{TW}. The smoothness of the profile $U$ is a delicate issue. More precisely, assuming further a structure condition on $g(u):=u-f(u)$, see hypothesis (H3) in \cite{bates1997traveling}, some of the conclusions of \cite[Theorem 3.1]{bates1997traveling} are as follows: if $c\neq 0$ then the profile $U$ is smooth and the sign of the speed is that of $\int_0^1f(u)du$; on the other hand if $c=0$ there are some situations, see \cite[end of Section 3]{bates1997traveling}, where the profile $U$ is discontinuous, and this can happen even if $\int_0^1f(u)du > 0$. Moreover,
the $c=0$ case is ruled out by the non-existence of a {\it null-truncation} of $g(u)=u-f(u)$ in the sense of \cite[Theorem 1]{berestycki2017non}. In particular, a sufficient condition to enforce $c>0$ is to assume
\begin{equation}
\label{hyp:g-monotone-prop}
u \mapsto u-f(u) \text{ is monotone and } \int_0^1f(u)du > 0.
\end{equation}
This shows the relevance of our sufficient conditions for propagation to occur, which we now state.
\begin{theorem}[Propagation threshold]\label{theo-ass-prop}
Let Assumption \ref{TWA} be satisfied and assume $\int_0^1f(u)du>0$. Assume the existence of $(c,U)$ an increasing traveling wave solution to \eqref{nonlocal} with $c\neq 0$ (and thus $c>0$ from the above discussion). Then the conclusions of the propagation Assumption \ref{ass:prop} hold: for each $\epsilon\in (0,1-\theta]$ there exists $L>0$ large enough such that $u_L^\epsilon(t, x)$ propagates, in the sense that
\begin{equation}\label{qqch}
\lim_{t\to+\infty} \; u_L^\epsilon(t,x)=1\; \text{ locally uniformly for on $\mathbb{R}$}.
\end{equation}
\end{theorem}
Theorem \ref{theo-ass-prop} thus provides sufficient conditions for the existence of $L^{\rm prop}_\epsilon<+\infty$, see \eqref{DEF-Lprop}, in the bistable case, and thus in the ignition case due to comparison arguments.
For classical diffusion bistable equations, the fact that \lq\lq sufficiently large'' step initial data lead to propagation is well-known, see \cite{Aro-Wei-78} for \lq\lq sufficiently high'' step initial data, and \cite{fife1977approach} for step initial data with any height larger than $\theta$. In the nonlocal diffusion setting, let us mention two propagation threshold properties proved by Berestycki and Rodr\' iguez \cite[Theorem 10]{berestycki2017non} and Lim
\cite[Theorem 1.2]{limlong} for \lq\lq sufficiently high'' step initial data and under the assumption that the kernel is exponentially bounded in the sense of \eqref{expo-decay}. These results also require a structure assumption as
\eqref{hyp:g-monotone-prop} or a \lq\lq $c>0$ traveling wave'' assumption. They
mainly rely on the construction of a compactly supported sub-solution using energy methods \cite{berestycki2017non} or traveling wave solutions \cite{limlong}. Theorem \ref{theo-ass-prop} improves these results into two directions. First, we cover more dispersal kernels by only requiring the existence of a finite first moment, see \eqref{first_moment}. Second, and even more important, we allow the height of the step initial data which we consider to be $\theta+\epsilon$ for any small $\epsilon>0$. This is achieved by performing a nonlocal genralization of the arguments in \cite{fife1977approach}. This requires a finite first moment of the dispersal kernel to guarantee the existence of a traveling wave solution $(c,U)$ and to collect some useful integrability properties of the wave $U$ at $\pm \infty$. It would be interesting to investigate on the existence (or not) of a propagation threshold for very heavy kernels not admitting a finite first moment, and for step initial data with height $\theta+\epsilon$. We believe that this is a delicate issue.
\section{Extinction criterion}\label{extinction}
In this section we consider the following semilinear problem
\begin{equation}\label{super}
\begin{cases}
\partial _t w=J\ast w-w+g(w), & t>0, \, x\in\mathbb{R},\\
w(0, x)=(\theta+\epsilon)\mathds{1}_{(-L, L)}(x), & x\in\mathbb{R},
\end{cases}
\end{equation}
where $\epsilon>0$, $L>0$ and
$$
g(u):=r^+(u-\theta)_+.
$$
Here $r^+>0$ and $\theta>0$ are given and fixed parameter and the subscript $+$ is used to denote the positive part of a real number. We start with a criterion for extinction which does not require $\epsilon$ to be small.
\begin{proposition}[Extinction criterion]\label{prop:extinction} Let Assumption \ref{ass:kernel} be satisfied. Let $\epsilon>0$ be fixed. Assume that $T>0$ and $L>0$ are such that
\begin{equation}\label{ext-criterion}
\theta e^{-(r^++1)T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}R_i(L)\geq \epsilon.
\end{equation}
Then the solution $w=w(t, x)$ to \eqref{super} satisfies
\begin{eqnarray}
\sup_{x\in\mathbb{R}}w(T, x)\leq\theta,
\end{eqnarray}
and is thus going to extinction at large times.
\end{proposition}
\begin{proof}
Consider $v=v(t, x)$ the solution to the linear Cauchy problem
\begin{equation*}
\begin{cases}
\partial _t v=J\ast v-v, & t>0, \, x\in\mathbb{R},\\
v(0, x)=w(0, x), & x\in\mathbb{R}.
\end{cases}
\end{equation*}
Recalling the definitions of $K$ and $\psi$ in \eqref{def-K} and \eqref{def-psi}, one gets $v(t, x)=(\theta+\epsilon)K(t, \cdot)\ast \mathds{1}_{(-L, L)}(x)$. Let us first prove, through a regularization argument and \cite{xu2021spatial}, that
\begin{equation}\label{def-V}
V(t):=\|v(t, \cdot)\|_{L^\infty(\mathbb{R})}=(\theta+\epsilon)\int_{|x|< L}K(t, x)dx=(\theta+\epsilon)\left[\int_{|x|< L}\psi(t, x)dx+e^{-t}\right].
\end{equation}
Take two sequences $(u_n^-)_{n\geq 0}$, $(u_n^+)_{n\geq 0}$ of continuous, symmetric and nonincreasing on $\mathbb{R}^+$ functions such that, for all $n\geq 0$ and $x\in \mathbb{R}$,
$$
0\leq u_n^-(x)\leq w(0,x)\leq u_n^+(x)\leq (\theta+\epsilon)\mathds{1}_{(-2L, 2L)}(x),
$$
$u_n^-(0)=\theta+\epsilon$, and $u_n^\pm (x)\to w(0,x)$, as $n\to+\infty$, for almost all $x\in\mathbb{R}$. Denoting by $v_n^\pm=v_n^\pm(t,x)$ the solutions to $\partial _t v=J*v-v$ starting from $u_n^\pm$, the comparison principle yields
\begin{equation}\label{cp-yields}
0\leq v_n^-(t,x)\leq v(t,x)\leq v_n^+(t,x)\leq \theta+\epsilon.
\end{equation}
Now, since the initial data $u_n^\pm$ are continuous, symmetric and decreasing on $\mathbb{R}^+$ and since the kernel $J$ is symmetric, Xu et al. \cite[Theorem 2.5]{xu2021spatial} proved that, for each $t>0$ and $n\geq 0$, both functions $x\mapsto v_n^-(t,x)$ and $x\mapsto v_n^+(t,x)$ are also symmetric and nonincreasing on $\mathbb{R}^+$. Hence one obtains, for $t>0$ and $n\geq 0$ ,
$$
\|v_n^\pm(t, \cdot)\|_{L^\infty(\mathbb{R})}=v_n^\pm(t,0)=\int_{\mathbb{R}}\psi(t, x)u_n^\pm(x) dx+e^{-t}u_n^\pm(0).
$$
From \eqref{cp-yields} we deduce that, for any $t>0$ and $n\geq 0$,
$$
\int_{\mathbb{R}}\psi(t, x)u_n^-(x) dx+e^{-t}u_n^-(0)\leq V(t)\leq
\int_{\mathbb{R}}\psi(t, x)u_n^+(x) dx+e^{-t}u_n^+(0).
$$
Recalling that $u_n^\pm(0)=\theta+\epsilon$, we let $n\to+\infty$ which, using Lebesgue convergence theorem, yields \eqref{def-V}.
Following Alfaro et al. \cite{alfaro2020quantitative}, we aim at constructing a super-solution to \eqref{super} in the form $W(t, x):=v(t, x)\varphi(t)$, with $\varphi(0)=1$ and $\varphi(t)>0$. As in \cite[Proof of Proposition 2.1]{alfaro2020quantitative}, we choose
$$ \varphi(t)=e^{r^+t}\left(1-\int_{0}^{t}r^+e^{-r^+s}\frac{\theta}{V(s)}ds\right). $$
Observe that $V(0)\varphi(0)>\theta$ and denote by $T>0$ the first time where $V(T)\varphi(T)=\theta$ (obviously we let $T=+\infty$ if such a time does not exist). Then $\left(\varphi(t)-\frac{\theta}{V(t)}\right)_+= \varphi(t)-\frac{\theta}{V(t)}$ for all $t\in [0,T)$, and thus one can check that $W(t,x)=v(t,x)\varphi(t)$ is a super-solution to \eqref{super} on the time interval $(0,T)$. In particular, if $T<+\infty$, it follows from the comparison principle that $w(T,\cdot)\leq W(T,\cdot)\leq \theta$, and we are done.
The condition $T<+\infty$ rewrites as: there exists $T>0$ such that $F_L(T)=0$ where
$$
F_L(t):=\theta\left(1-\frac{e^{-r^+t}}{A_L(t)}\right)+\epsilon-\int_{0}^{t}r^+e^{-r^+s}\frac{\theta}{A_L(s)}ds=0,
$$
wherein
$$
A_L(t):=\int_{|x|< L}\psi (t, x)dx+e^{-t}.
$$
We claim (see below for a proof) that $A_L'(t)\leq 0$ for all $t>0$. As a result, since $F_L(0)=\epsilon$ and $F'_L(t)=\theta e^{-r^+t}\frac{A'_L(t)}{A^2_L(t)}\leq0$, we are left to find $T>0$ such that $F_L(T)\leq 0$, that is
\begin{equation*}
1+\frac{\epsilon}{\theta}\leq r^+\int_{0}^{T}\frac{e^{-r^+s}}{A_L(s)}ds+\frac{e^{-r^+T}}{A_L(T)}.
\end{equation*}
Integrating by parts this is equivalent to
\begin{equation}\label{to-be-reached}
\frac{\epsilon}{\theta}\leq -\int_{0}^{T}e^{-r^+s}\frac{A'_L(s)}{A_L^{2}(s)}ds.
\end{equation}
But, since $0<A_L(s)\leq1$ (recall \eqref{1-e^t}) and $A'_L(s)\leq 0$ for all $s>0$,
\begin{eqnarray}
-\int_{0}^{T}e^{-r^+s}\frac{A'_L(s)}{A_L^{2}(s)}ds
&\geq & -\int_{0}^{T}e^{-r^+s}A'_L(s)ds\nonumber\\
&\geq & e^{-r^+T}(1-A_L(T))\nonumber\\
& = & e^{-r^+T}\left(1-e^{-T}-\int_{|x|< L}\psi (T, x)dx\right)\nonumber\\
&= & e^{-r^+T}\left(1-e^{-T}-e^{-T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}\int_{|x|< L}J^{*(i)}(x)dx\right)\nonumber\\
&= & e^{-r^+T}\left(1-e^{-T}-e^{-T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}\left(1-R_i(L)\right)\right)\nonumber\\
&=& e^{-r^+T}\left(1-e^{-T}-e^{-T}(e^T-1)+e^{-T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}R_i(L)\right)\nonumber\\
&=& e^{-(r^++1)T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}R_i(L).\nonumber
\end{eqnarray}
From this and \eqref{to-be-reached}, we conclude that \eqref{ext-criterion} enforces $T<+\infty$ and is thus a sufficient condition for extinction.
\end{proof}
\medskip
It remains to prove the claim that $A'_L(t)\leq0$ for all $t>0$.
\medskip
\begin{proof}
Observe that $A_L(t)=e^{-t}+\int_{\mathbb{R}}\psi(t, x)u_0(x)dx$, where $u_0=\mathds{1}_{(-L, L)}$, and thus
\begin{eqnarray}
A'_L(t)
&=&-e^{-t}+\int_{\mathbb{R}}\partial_t\psi(t, x)u_0(x)dx\nonumber\\
&=&-e^{-t}u_0(0)+\int_{\mathbb{R}}\left(J\ast \psi(t, \cdot)(x)-\psi(t,x)+e^{-t}J(x)\right)u_0(x)dx\nonumber\\
&=&\int_{\mathbb{R}}J(y)\int_{\mathbb{R}}\left(\psi(t,x-y)u_0(x)-\psi(t, x)u_0(x)\right)dxdy+e^{-t}\int_{\mathbb{R}}J(x)\left(u_0(x)-u_0(0)\right)dx\nonumber\\
&=&\int_{\mathbb{R}}J(y)\left(\psi(t,\cdot)*u_0(y)-\psi(t, \cdot)*u_0(0)\right)dy+e^{-t}\int_{\mathbb{R}}J(x)\left(u_0(x)-u_0(0)\right)dx\nonumber\\
&=&\int _{\mathbb{R}} J(x)\left(u(t,x)-u(t,0)\right)dx,\nonumber
\end{eqnarray}
where we have used \eqref{partial} and the symmetry of $\psi$ and where $u=u(t,x)$ is the solution of $\partial _t u=J*u-u$ starting from $u_0$. Now, since both $J$ and $u_0$ are symmetric and nonincreasing on $\mathbb{R}^+$, it follows from Xu et al. \cite[Theorem 2.5] {xu2021spatial} (and a regularization argument as above) that $u(t,x)\leq u(t,0)$ for all $t>0$, $x\in \mathbb{R}$. Thus we have $A'_L(t)\leq0$ for all $t>0$.
\end{proof}
\medskip
We now take a kernel $J$ satisfying Assumptions \ref{ass:kernel} and \ref{ass:kernel-bis} with, in particular, the expansion \eqref{J-Fourier-ass} for some $0<\beta\leq 2$. We consider the above extinction criterion in the limit $\epsilon \to 0$, revealing the role of the dispersal kernel. Notice that the results below connect to the {\it control issue}: given a \lq\lq final time'' $T>0$, how large can we choose the size of the initial data so that the solution is everywhere smaller than the threshold $\theta$ at time $T$?
We start with the case $0<\beta< 2$. Our analysis seems to reveal that the \lq\lq final time'' does not affect the order of magnitude $L\sim \epsilon^{-1/\beta}$ but plays a role in the value of the \lq\lq multiplicative constant''.
\begin{corollary}[Asymptotic extinction criterion, $0<\beta< 2$]\label{asmptotic}
Let Assumptions \ref{ass:kernel} and \ref{ass:kernel-bis} be satisfied with $0<\beta< 2$. Let $T>0$ be a fixed time. Then, for any $0<\gamma<1$, there exists $\epsilon_0>0$ small enough such that, for each $\epsilon\in(0, \epsilon_0)$ and for each
$$
0<L<\gamma \left(\theta aC Te^{-r^+T}\right)^{1/\beta}\epsilon ^{-1/\beta},
$$
the solution $w=w(t, x)$ to \eqref{super} is everywhere smaller than $\theta$ at time $T$ and is thus going to extinction at large times. Here constants $a>0$ and $C=C(\beta)>0$ come from the asymptotic expansion \eqref{J-Fourier-ass} and Lemma \ref{lem:tails}-(ii), respectively.
\end{corollary}
\begin{proof}
For a given $0<\gamma<1$, we select an integer $N$ large enough so that
$$
\sum_{i=0}^N \frac{T^{i}}{i!}\geq \gamma ^{\beta/2}e^T.
$$
Next, from \eqref{tails-beta-lower-2}, there is $L_0>0$ large enough such that, for all $L\geq L_0$ and $1\leq i \leq N+1$,
$$
R_i(L)\geq \gamma ^{\beta/2} i\, \frac{aC}{L^\beta}.
$$
Hence, if $L\geq L_0$ then
$$
\theta e^{-(r^++1)T}\sum_{i=1}^{+\infty} \frac{T^i}{i!}R_i(L)\geq \theta e^{-(r^++1)T}\sum_{i=1}^{N+1}\frac{T^i}{i!}\gamma^{\beta/2} i\,\frac{aC}{L^\beta}\geq \gamma ^{\beta} \theta T e^{-r^+T}\frac{aC}{L^\beta}.
$$
As a result, to check the extinction criterion \eqref{ext-criterion} it is enough to have
$$
L\geq L_0 \quad \text{ and }\quad\epsilon \leq \gamma ^{\beta} \theta T e^{-r^+T}\frac{aC}{L^\beta}.
$$
Hence, there is $\epsilon _0>0$ such that, for all $\epsilon \in (0,\epsilon_0)$, the conclusion of the corollary holds true for $L=\gamma \left(\theta aC Te^{-r^+T}\right)^{1/\beta}\epsilon ^{-1/\beta}$, and also for smaller $L$ from the comparison principle.
\end{proof}
\medskip
On the other hand, the case $\beta=2$ is more tricky as revealed by the following corollaries.
\begin{corollary}[Asymptotic extinction criterion for exponentially bounded kernels]\label{lighttail}
Let Assumption \ref{ass:kernel} be satisfied and assume further the exponential decay \eqref{expo-decay}. Then there exists $\epsilon_0>0$ small enough such that, for each $\epsilon\in(0, \epsilon_0)$ and for each $L$ satisfying
$$
0<L<C^-\ln\frac{1}{\epsilon},
$$
with $C^-=C^-(r^+,J)>0$, the solution $w=w(t, x)$ to \eqref{super} is everywhere smaller than $\theta$ at time $T=\tilde C^{-}\ln \frac 1 \epsilon$, $\tilde C^-=\tilde C^-(r^+,J)>0$, and is thus going to extinction at large times.
\end{corollary}
\begin{proof} We shall use the large deviations Cram\'er theorem, as stated in Lemma \ref{lem:expo}. We use the criterion \eqref{ext-criterion} with a single \lq\lq large'' term $i=i(\epsilon)$: to prove extinction it is sufficient to obtain $
\theta e^{-(r^++1)T}\frac{T^i}{i!} R_i(L) \ge \epsilon$ or, equivalently,
\begin{equation}\label{exp-super}
-(r^++1)\frac T i+\ln T-\frac 1 i \ln (i!)+\frac 1 i \ln R_{i}(L)\geq \frac 1 i (\ln\epsilon-\ln\theta).
\end{equation}
Recalling that $L_1>0$ is defined in Lemma \ref{lem:expo} let us fix $0<L_0<s_0<L_1$. Then since $\Lambda^*$ is nondecreasing on $(0,+\infty)$, one has
$0\leq \Lambda^*(L_0)\leq \Lambda^*(s_0)\leq \Lambda^*(L_1)<+\infty$, while according to Lemma \ref{lem:expo} the convergence in \eqref{cramer} holds uniformly in any compact subset of $(0,L_1]$.
We now choose
$$L=Cs_0\ln\frac{1}{\epsilon}, \quad T=i=[C\ln \frac 1 \epsilon],
$$
where $C>0$ is some constant to be determined so that \eqref{exp-super} is satisfied for all $\epsilon$ small enough.
From Stirling's formula, we see that there is $K>0$ such that $\frac 1 i \ln (i!)\leq \ln i +K$ for all $i\geq 1$. Hence the condition \eqref{exp-super} is reached as soon as
\begin{equation*}
-(r^++1)-K+\frac 1 i \ln R_{i}(L)\geq \frac 1 i(\ln\epsilon-\ln\theta).
\end{equation*}
As $\epsilon\to 0$, the right hand side tends to $-1/C$. On the other hand, we write
\begin{equation*}
\frac 1 i\ln R_{i}(L)=\left(\frac 1 i \ln R_i\left(i\frac L i\right)+\Lambda^*\left(\frac L i\right)\right)-\Lambda^*\left(\frac{L}{i}\right).
\end{equation*}
Since $s_0\leq \frac L i\leq \frac{Cs_0\ln \frac 1 \epsilon}{C\ln \frac 1 \epsilon -1}$,
the first term in the right-hand converges to $0$ as $\epsilon\to 0$ due to the uniformity of the convergence in \eqref{cramer}, while the monotonicity of $\Lambda^*$ yields, for all $\epsilon$ small enough,
$$
-\Lambda^*\left(\frac{L}{i}\right)\geq -\Lambda^*(L_1).
$$
As a consequence we get
\begin{equation*}
\liminf_{\epsilon\to 0}\frac 1 i\ln R_{i}(L)\geq -\Lambda^*(L_1).
\end{equation*}
As a result it suffices to have $-(r^++1)-K-\Lambda^*(L_1)\geq -1/C$, which is achieved by taking $C=C(r^+,s_0)=C(r^+,J)>0$ small enough. The extinction for smaller $L$ follows from the comparison principle.
\end{proof}
\begin{corollary}[Asymptotic extinction criterion, some examples with $\beta=2$]\label{remarkbeta=2}
Let Assumption \ref{ass:kernel} be satisfied. Let $T>0$ be a fixed time.
\begin{itemize}
\item [(i)] Assume that there are $c_2>c_1>0$, $\alpha>2$ and $x_0\in \mathbb{R}$ such that $ \frac{c_1}{\vert x \vert^{1+\alpha}}\leq J(x) \leq \frac{c_2}{\vert x \vert^{1+\alpha}}$ for all $\vert x\vert \geq \vert x_0\vert $. Then there is $C^-=C^-(\theta,r^+,T,c_1,\alpha)>0$ such that, for each $\epsilon>0$ and each $L$ satisfying
$$
0<L<C^-\frac 1 {\epsilon^{\frac{1}{\alpha}}},
$$
the solution $w=w(t, x)$ to \eqref{super} is everywhere smaller than $\theta$ at time $T$ and is thus going to extinction at large times.
\item[(ii)] Assume that $J$ has regularly varying tails $RV(\alpha)$ with $\alpha>2$. Let us fix $\alpha'>\alpha$. Then there exists $\epsilon_0>0$ small enough such that, for each $\epsilon\in(0, \epsilon_0)$ and for each $L$ satisfying
$$
0<L<\frac 1 {\epsilon^{\frac{1}{\alpha'}}} ,
$$
the solution $w=w(t, x)$ to \eqref{super} is everywhere smaller than $\theta$ at time $T$ and is thus going to extinction at large times.
\item [(iii)] Assume that $J$ has lognormal-type tails $LN(\gamma,\lambda,\rho)$ with $\gamma>1$, $\lambda>0$ and $\rho \in \mathbb{R}$. Let us fix $\lambda '>\lambda$. Then there exists $\epsilon_0>0$ small enough such that, for each $\epsilon\in(0, \epsilon_0)$ and for each $L$ satisfying
$$
0<L< e^{\left(\frac 1 {\lambda '} \ln\frac 1 \epsilon\right)^{\frac 1 \gamma}},
$$
the solution $w=w(t, x)$ to \eqref{super} is everywhere smaller than $\theta$ at time $T$ and is thus going to extinction at large times.
\item [(iv)] Assume that $J$ has Weibull-like tails $WE(\alpha,\lambda,\rho)$ with $0<\alpha <1$, $\lambda>0$ and $\rho \in \mathbb{R}$. Let us fix $\lambda '>\lambda$. Then there exists $\epsilon_0>0$ small enough such that, for each $\epsilon\in(0, \epsilon_0)$ and for each $L$ satisfying
\begin{equation}\label{bidule}
0< L < \left(\frac 1 {\lambda '} \ln \frac 1 \epsilon\right)^{\frac 1 \alpha},
\end{equation}
the solution $w=w(t, x)$ to \eqref{super} is everywhere smaller than $\theta$ at time $T$ and is thus going to extinction at large times.
\end{itemize}
\end{corollary}
\begin{proof} We use the criterion \eqref{ext-criterion} with only the term $i=1$: to prove extinction it is sufficient to obtain $\theta e^{-(r^++1)T}T R_1(L)\geq \epsilon$.
In case $(i)$ since $R_1(L)\geq \frac{2c_1}{\alpha}L^{-\alpha}$ the conclusion easily follows.
In case $(ii)$, from \eqref{tail-RV} the condition is recast
$$
CS(L)L^{-\alpha}\geq \epsilon, \quad C:=\theta e^{-(r^++1)T}T,
$$
where $S$ is a slowly varying function. Letting $L=\frac 1 {\epsilon^{\frac{1}{\alpha'}}}$, this is recast
$$
CL^{\alpha'-\alpha}S(L)\geq 1,
$$
which is true for $L\gg 1$ from known properties of slowly varying functions, see e.g. \cite[VIII, Lemma 2]{Feller}. As a result, extinction does occur for $0<\epsilon<\epsilon _0$ with $\epsilon_0>0$ small enough and $L=\frac 1 {\epsilon^{\frac{1}{\alpha'}}}$. The extinction for smaller $L$ follows from the comparison principle.
In case $(iii)$, from \eqref{tail-LN} the condition is recast
$$
C L^\rho e^{-\lambda \ln ^\gamma L}\geq \epsilon, \quad C:=\theta e^{-(r^++1)T}T c.
$$
Letting $L=e^{\left(\frac 1 {\lambda '} \ln\frac 1 \epsilon\right)^{\frac 1 \gamma}}$, this is recast
$$
CL^{\rho}e^{(\lambda '-\lambda) \ln ^\gamma L}\geq 1,
$$
which is true for $L\gg 1$. As a result, extinction does occur for $0<\epsilon<\epsilon _0$ with $\epsilon_0>0$ small enough and $L=e^{\left(\frac 1 {\lambda '} \ln\frac 1 \epsilon\right)^{\frac 1 \gamma}}$. The extinction for smaller $L$ follows from the comparison principle.
In case $(iv)$, we recast the condition thanks to \eqref{tail-WE} and use a similar argument.
\end{proof}
\begin{remark}\label{rem:as-easily} As easily seen from the proof, the estimate in case $(ii)$ can be slightly improved (namely $\alpha '$ can be taken equal to $\alpha$) when the slowly varying function $S$ satisfies $S(L)\to +\infty$ as $L\to +\infty$. Similarly the estimates in $(iii)$ and $(iv)$ can be slightly improved (namely $\lambda '$ can be taken equal to $\lambda$) when $\rho >0$. Last, in $(iv)$, when $\rho=0$, \eqref{bidule} can be replaced by
\begin{equation}\label{bidule2}
0<L<\left(\frac{1}{\lambda}\ln \frac C \epsilon\right)^{\frac 1 \alpha},\text{ where }C:=\theta e^{-(r^++1)T}Tc.
\end{equation}
\end{remark}
\section{Non-extinction criterion}\label{non-extinction}
In this section, we fix $r^->0$, $\theta\in(0, 1)$, and define the linear function
\begin{equation}\label{g}
\tilde{g}(w):=r^-(w-\theta).
\end{equation}
For $\epsilon>0$ and $L>0$, we consider $w_L^\epsilon=w_L^\epsilon(t, x)$ the solution of the Cauchy problem
\begin{equation}\label{sub}
\begin{cases}
\partial _t w=J\ast w-w+\tilde{g}(w), & t>0,\, x\in\mathbb{R},\\
w(0, x)=(\theta+\epsilon)\mathds{1}_{(-L, L)}(x), & x\in\mathbb{R}.\end{cases}
\end{equation}
We start with a criterion for non-extinction which does not require $\epsilon$ to be small.
\begin{proposition}[Non-extinction criterion]\label{nonextinction}
Let Assumption \ref{ass:kernel}-(i) be satisfied. Let $\epsilon>0$ and $L>0$ be given. Let $\eta\in(\theta, 1)$ and $m\in (0,1)$ be given. Define
\begin{equation}\label{def:Tep}
T_\epsilon:=\frac{1}{r^-}\ln \frac{2(\eta-\theta)}{\epsilon}.
\end{equation}
\begin{itemize}
\item [(i)] For all $0<t\leq T_\epsilon$, all $x\in \mathbb{R}$, $w_L^\epsilon(t, x)\leq \theta+2(\eta-\theta)$.
\item [(ii)] If
\begin{equation}\label{nonextinctioncondition}
\frac{\epsilon}{2(\theta+\epsilon)}\geq e^{-T_\epsilon }\sum_{i=1}^{+\infty} \frac{T_\epsilon^i}{i!}R_i\left((1-m)L)\right),
\end{equation}
then
\begin{equation}\label{conclusion}
\min_{|x|\leq mL} w_L^\epsilon(T_\epsilon, x)\geq \eta.
\end{equation}
\end{itemize}
\end{proposition}
\begin{proof}
Notice that the function $w^\epsilon_L(t, x)$ is given by
\begin{equation}\label{w^ep_L}
w^\epsilon_L(t, x)=\theta+e^{r^-t}(v(t, x)-\theta),
\end{equation}
where $v=v(t, x)$ denotes the solution of the linear equation
$$
\partial _t v =J\ast v-v, \; t>0, \, x\in\mathbb{R},
$$
starting from $w^\epsilon_L(0, x)=(\theta+\epsilon)\mathds{1}_{(-L, L)}(x)$. From the comparison principle $v(t,x)\leq \theta +\epsilon$ for all $t>0$ and $x\in \mathbb{R}$, and thus, for all $0<t\leq T_\epsilon$,
$$
w^\epsilon(t,x)\leq \theta+\frac{2(\eta-\theta)}{\epsilon}(\theta+\epsilon-\theta)=\theta+2(\eta-\theta),\; \forall x\in\mathbb{R},
$$
which proves $(i)$.
Next, we know from subsection \ref{ss:linear} that $v=v(t,x)$ is given by
$$
v(t,x)=(\theta+\epsilon) \left[e^{-t} \mathds{1}_{(-L, L)}(x)+\int_{-L}^L\psi(t,x-y)dy\right].
$$
From \eqref{1-e^t} we obtain
$$
\frac {v(t,x)}{\theta+\epsilon}=1+e^{-t} \left(\mathds{1}_{(-L, L)}(x)-1\right)-\int_{|y|\geq L}\psi(t,x-y)dy,
$$
and thus
$$
w^\epsilon_L(t,x)=\theta+(\theta+\epsilon)e^{r^-t}\left[e^{-t} \left(\mathds{1}_{(-L, L)}(x)-1\right)-\int_{|y|\geq L}\psi(t,x-y)dy+\frac{\epsilon}{\theta+\epsilon}\right].
$$
We now restrict to $x$ satisfying $|x|\leq mL$. Since $|y|\geq L$ ensures that $|x|\leq mL\leq m|y|$ and $|x-y|\geq(1-m)|y|$, we deduce that
\begin{eqnarray}
w^\epsilon_L(t, x)
&\geq& \theta+(\theta+\epsilon)e^{r^-t}\left[\frac{\epsilon}{\theta+\epsilon}-\int_{|z|\geq (1-m)L}\psi(t, z)dz\right]\nonumber\\
&=& \theta+(\theta+\epsilon)e^{r^-t}\left[\frac{\epsilon}{\theta+\epsilon}-e^{-t}\sum_{i=1}^{+\infty} \frac{t^i}{i!}\int_{\vert z\vert \geq (1-m)L}J^{*(i)}(z)dz\right].\nonumber
\end{eqnarray}
It follows from this and \eqref{nonextinctioncondition} that
$$
\min_{|x|\leq mL}w^\epsilon_L(T_\epsilon, x)\geq \theta+(\theta+\epsilon)\frac{2(\eta-\theta)}{\epsilon}\left[\frac{\epsilon}{\theta+\epsilon}-\frac{\epsilon}{2(\theta+\epsilon)}\right]=\eta,
$$
which concludes the proof of $(ii)$.
\end{proof}
\begin{remark}\label{Rem1}
Note that the right hand side of \eqref{nonextinctioncondition} is decreasing with respect to $L>0$ from $1-e^{-T_\epsilon}$ when $L=0$ to $0$ when $L\to +\infty$.
\end{remark}
We now take a kernel $J$ satisfying Assumptions \ref{ass:kernel}-$(i)$ and \ref{ass:kernel-bis} with, in particular, the expansion \eqref{J-Fourier-ass} for some $0<\beta\leq 2$. We consider the above non-extinction criterion in the limit $\epsilon \to 0$, revealing the role of the dispersal kernel.
\begin{corollary}[Asymptotic non-extinction criterion, $0<\beta\le2$]\label{prop:0<beta<2}
Let Assumptions \ref{ass:kernel}-(i) and \ref{ass:kernel-bis} with $0<\beta\le2$ be satisfied. Let $\eta\in(\theta, 1)$ and $m\in(0, 1)$ be given. Then for all $\epsilon>0$, there is $L_\epsilon>0$ such that the conclusion \eqref{conclusion} holds for all $L\geq L_\epsilon$. Furthermore, there is $\epsilon_0>0$ small enough such that, for all $0<\epsilon<\epsilon_0$, $L_\epsilon$ can be chosen as follows,
$$
L_\epsilon=\frac{C^+}{1-m}\left(\frac{1}{\epsilon}\ln\frac{1}{\epsilon}\right)^{\frac1\beta},
$$
for some constant $C^+=C^+(\theta, r^-, J)>0$.
\end{corollary}
\begin{proof}
Observe first that the existence of $L_\epsilon$ directly follows from Remark \ref{Rem1}. We now turn to the asymptotic of $L_\epsilon$ as $\epsilon\to 0$. From Lemma \ref{lem:Durret} we can write
\begin{equation*}
\begin{split}
e^{-T_\epsilon}\sum_{i=1}^{\infty} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right)&\leq e^{-T_\epsilon}\sum_{i=1}^{\infty} \frac{T_\epsilon^i}{i!}\frac{(1-m)}{2}L \int_{|\xi|\leq \frac{2}{(1-m)L}}\left[1-\left(\widehat J(\xi)\right)^i\right]d\xi\\
&\leq e^{-T_\epsilon}\sum_{i=1}^{\infty} \frac{T_\epsilon^i}{i!}\frac{(1-m)}{2}L \int_{|\xi|\leq \frac{2}{(1-m)L}}i (1-\widehat J(\xi))d\xi,
\end{split}
\end{equation*}
from the mean value theorem. Hence, from \eqref{J-Fourier-ass}, there is $C=C(\beta,a)>0$ such that, for all $\epsilon>0$ and all $L>0$ large enough,
\begin{equation*}
\begin{split}
e^{-T_\epsilon}\sum_{i=1}^{\infty} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right)
&\leq C e^{-T_\epsilon}\sum_{i=1}^{\infty} \frac{iT_\epsilon^i}{i!}\frac{1}{(1-m)^\beta L^\beta}=C \frac{T_\epsilon}{(1-m)^\beta L^\beta}.
\end{split}
\end{equation*}
From this and the definition of $T_\epsilon$ in \eqref{def:Tep}, we see that the non-extinction criterion \eqref{nonextinctioncondition} is satisfied, for all $\epsilon>0$ small enough, as soon as $L>L_\epsilon:=\frac{C^+}{1-m}(\frac 1 \epsilon \ln \frac 1 \epsilon)^{\frac1\beta}$ for some $C^+=C^+(\theta,r^-,\beta,a)>0$, which proves the desired result.
\end{proof}
\medskip
On the other hand, the case $\beta=2$ is more tricky as revealed by the following corollaries. Recall that $r^-$ is defined in \eqref{r-moins} and $m_2(J)=\int_{\mathbb{R}}x^2J(x)dx$.
\begin{corollary}[Asymptotic non-extinction criterion for exponentially bounded kernels]\label{prop:exp-decay}
Let Assumption \ref{ass:kernel}-(i) be satisfied. Assume further the exponential decay \eqref{expo-decay} and $r^-\in(0, 1)$. Let $\eta\in(\theta, 1)$ and $m\in(0, 1)$ be given. Then for all $\epsilon>0$, there is $L_\epsilon>0$ such that the conclusion \eqref{conclusion} holds for all $L\geq L_\epsilon$. Furthermore, there is $\epsilon_0>0$ small enough such that, for all $0<\epsilon<\epsilon_0$, $L_\epsilon$ can be chosen as follows,
$$
L_\epsilon=\frac{C^+}{1-m}\ln\frac{1}{\epsilon},
$$
for some constant $C^+=C^+(r^-, J)>0.
\end{corollary}
\begin{corollary}[Asymptotic non-extinction criterion, some examples with $\beta=2$]\label{prop:examples-be=2}
Let Assumption \ref{ass:kernel}-(i) be satisfied. Assume $r^-\in(0, 1)$ and $m_2(J)=1$. Let $\eta\in(\theta, 1)$ and $m\in(0, 1)$ be given. Then for all $\epsilon>0$, there is $L_\epsilon>0$ such that the conclusion \eqref{conclusion} holds for all $L\geq L_\epsilon$. Furthermore, there is $\epsilon_0>0$ small enough (possibly depending on $\tilde \alpha$ or $\tilde \lambda$ in cases $(ii), (iii)$ and $(iv)$, see below) such that, for all $0<\epsilon<\epsilon_0$, $L_\epsilon$ can be chosen as follows.
\begin{itemize}
\item [(i)] Assume that there are $c_2>c_1>0$, $\alpha>2$ and $x_0\in \mathbb{R}$ such that $\frac{c_1}{\vert x\vert^{1+\alpha}}\leq J(x)\leq \frac{c_2}{\vert x\vert ^{1+\alpha}}$ for all $\vert x\vert \geq \vert x_0\vert$, then
$$
L_\epsilon= \frac{C^+}{1-m}\left(\frac1\epsilon\ln\frac1\epsilon\right)^{\frac1\alpha},
$$
for some constant $C^+=C^+(\theta, r^-, c_2, \alpha)>0$.
\item [(ii)] Assume that $J$ has regularly varying tails $RV(\alpha)$ with $\alpha>2$, then
$$
L_\epsilon= \left(\frac1\epsilon\ln \frac1\epsilon\right)^{\frac{1}{\tilde \alpha}},
$$
for any given $0<\tilde \alpha<\alpha$.
\item[(iii)] Assume that $J$ has lognormal-type tails $LN(\gamma,\lambda,\rho)$ with $\gamma>1$, $\lambda>0$, $\rho \in \mathbb{R}$, then
$$
L_\epsilon= \exp \left[\left(\frac{1}{\tilde \lambda} \ln \frac 1\epsilon\right)^\frac{1}{\gamma}\right],
$$
for any given $0<\tilde \lambda<\lambda$.
\item[(iv)] Assume that $J$ has Weibull-like tails $WE(\alpha,\lambda,\rho)$ with $0<\alpha<1$, $\lambda>0$, $\rho \in \mathbb{R}$ then
$$
L_\epsilon= \frac{1}{1-m}\left(\frac{1}{\tilde \lambda}\ln \frac 1\epsilon\right)^{\frac 1\alpha}, \quad \text{ if } \quad 0<\alpha<\frac 2 3,
$$
for any given $0<\tilde \lambda<\lambda$, while
$$
L_\epsilon =\left(\ln \frac 1 \epsilon \right)^{\frac{1}{2(1-\tilde \alpha)}},\quad \text{ if } \quad \frac 23\leq \alpha<1,
$$
for any given $0<\alpha<\tilde \alpha<1$.
\end{itemize}
\end{corollary}
\begin{corollary}[Asymptotic non-extinction criterion, Weibull-like tails $WE(\alpha, \lambda, \rho=0)$]\label{prop:Weibull}
Let Assumption \ref{ass:kernel}-(i) be satisfied. Further assume that $J$ has Weibull-like tails $WE(\alpha, \lambda, \rho=0)$ and $r^-\in(0, 1)$. Let $\eta\in(\theta, 1)$ and $m\in(0, 1)$ be given. Then for all $\epsilon>0$, there is $L_\epsilon>0$ such that the conclusion \eqref{conclusion} holds for all $L\geq L_\epsilon$. Furthermore, there is $\epsilon_0>0$ small enough such that, for all $0<\epsilon<\epsilon_0$, $L_\epsilon$ can be chosen as follows,
$$
L_\epsilon=\frac{C^+}{1-m}\left(\frac 1 \lambda \ln \frac 1 \epsilon\right)^{\frac 1\alpha},
$$
for some $C^+=C^+(J)>0$.
\end{corollary}
Again the existence of $L_\epsilon$ in Corollaries \ref{prop:exp-decay}, \ref{prop:examples-be=2} and \ref{prop:Weibull} directly follows from Remark \ref{Rem1}. We now turn to the estimates in Corollaries \ref{prop:exp-decay}, \ref{prop:examples-be=2} and \ref{prop:Weibull}. We assume
\begin{equation}\label{hypr-}
0<r^-<1.
\end{equation}
As a preparation we consider the sum of the series in \eqref{nonextinctioncondition} and roughly show that, for $t$ large enough,
the main contribution corresponds to the indexes $i$ around $t$. To see this, observe that for any $t>0$ and any $M(t)\in\mathbb N$ such that $M(t)\leq t$ one has, using $i!\geq i^i e^{-i}$ and since $i\mapsto i+i\ln t-i\ln i$ is increasing on $(0,t)$,
\begin{equation}
\label{M}
e^{-t}\sum_{i=1}^{M(t)} \frac{t^i}{i!}\leq e^{-t}\sum_{i=1}^{M(t)} \exp\left(i+i\ln t-i\ln i\right)\leq e^{-t} M(t)\left(\frac{et}{M(t)}\right)^{M(t)}.
\end{equation}
On the other hand, for any $t>0$ and any $N(t)\in \mathbb{N}$ such that $N(t)>et$, one has
\begin{equation}
\label{N}
e^{-t}\sum_{i=N(t)}^{+\infty} \frac{t^i}{i!}\le e^{-t} \sum_{i=N(t)}^{+\infty} \left(\frac{et}{i}\right)^i \le e^{-t}\sum_{i=N(t)}^{+\infty} \left(\frac{et}{N(t)}\right)^i= e^{-t}\left(\frac{et}{N(t)}\right)^{N(t)}\frac{N(t)}{N(t)-et}.
\end{equation}
Let us choose $0<\gamma^-<1$ such that
$$
\gamma^-\ln (e/\gamma^-)<1-r^-,
$$
and $M(t)=[\gamma^- t]$. Then $\gamma^-t-1<M(t)\leq \gamma ^- t < t$ and thus \eqref{M} yields
$$
e^{-t}\sum_{i=1}^{M(t)} \frac{t^i}{i!}\leq \gamma^- t e^{-t} \exp\left(\gamma^- t\ln \frac{et}{\gamma^- t-1}\right),
$$
so that
$$
e^{-t}\sum_{i=1}^{M(t)} \frac{t^i}{i!}\leq \gamma^- t \exp\left[(\gamma^- \ln (e/\gamma^-)-1)t+\mathcal O(1)\right]=o(e^{-r^- t}), \quad \text { as } t \to +\infty.
$$
Let us next choose $N(t)=[3t]+1>et$, so that \eqref{N} yields
$$
e^{-t}\sum_{i=N(t)}^{\infty} \frac{t^i}{i!}\leq e^{-t}\frac{[3t]+1}{[3t]+1-et}=o(e^{-r^- t}), \quad \text { as } t \to +\infty.
$$
To summarize the above analysis, since $J^{*(i)}\geq 0$ and $\int_{\mathbb{R}} J^{*(i)}(x)dx=1$ for any $i\geq 1$, we have proved the following lemma.
\begin{lemma}\label{LE-E1}
Assume that $0<r^-<1$. Select $\gamma^->0$ small enough so that $\gamma^-\ln (e/\gamma^-)<1-r^-$. Defining
$$
M(t):=[\gamma^-t], \quad N(t):=[3t]+1,
$$
we have
\begin{equation}\label{E1}
e^{-t}\sum_{i=1}^{+\infty} \frac{t^i}{i!} R_i\left((1-m)L\right)=e^{-t}\sum_{i=M(t)}^{N(t)} \frac{t^i}{i!}R_i\left((1-m)L\right)+o\left(e^{-r^- t}\right), \quad \text { as } t \to +\infty,
\end{equation}
independently of $m\in(0,1)$ and $L>0$.
\end{lemma}
Let us recall that $T_\epsilon=\frac{1}{r^-}\ln \frac{2(\eta-\theta)}{\epsilon}$ and define $M_\epsilon:=M(T_\epsilon)$ and $N_\epsilon:=N(T_\epsilon)$. Then, from Lemma \ref{LE-E1}, we see that the non-extinction criterion \eqref{nonextinctioncondition} is satisfied as soon as
\begin{equation}\label{estrb2}
\frac{\epsilon}{4\theta}\geq e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right).
\end{equation}
holds for $\epsilon>0$ small enough. We now explore the new criterion \eqref{estrb2} with various assumptions.
\medskip
\begin{proof}[\bf Proof of Corollary \ref{prop:exp-decay}] Here we assume that $J$ is exponentially bounded in the sense of \eqref{expo-decay}. Let us fix $\lambda>0$ so that $\Lambda(\lambda)<+\infty$. Now choose $L=T_\epsilon\ell$ with $\ell>0$ to be determined later. It follows from \eqref{esti-expo} that
\begin{equation*}
\begin{split}
e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right)&\leq e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} \exp\left(-i\left(\lambda\frac{(1-m)\ell T_\epsilon}{i}-\Lambda(\lambda)\right)\right)\\
&\leq e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} \exp\left(-i\left(\lambda\frac{(1-m)\ell}{6}-\Lambda(\lambda)\right)\right),
\end{split}
\end{equation*}
for $\epsilon>0$ small enough since then $\frac{T_\epsilon}{i}\geq \frac{T_\epsilon}{N_\epsilon}=\frac{T_\epsilon}{[3T_\epsilon]+1}\geq \frac 1 6$. Hence we reach
\begin{equation*}
\begin{split}
&e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right)
\leq e^{-T_\epsilon}\exp\left(T_\epsilon e^{-\left(\lambda\frac{(1-m)\ell}{6}-\Lambda(\lambda)\right)}\right).
\end{split}
\end{equation*}
As a result the criterion \eqref{estrb2} is satisfied as soon
$$
e^{-\left(\lambda\frac{(1-m)\ell}{6}-\Lambda(\lambda)\right)}\leq \frac{-\ln (4\theta)+\ln \epsilon+T_\epsilon}{T_\epsilon}.
$$
Recalling $T_\epsilon=\frac{1}{r^-}\ln\frac{2(\eta-\theta)}{\epsilon}$ this rewrites as
$$
e^{-\left(\lambda\frac{(1-m)\ell}{6}-\Lambda(\lambda)\right)}\leq \frac{-\ln (4\theta)}{T_\epsilon}+1-r^-\frac{-\ln\epsilon}{\ln(2(\eta-\theta))-\ln\epsilon}.
$$
Since the right hand side tends to $1-r^->0$ as $\epsilon \to 0$, one can choose $\ell>0$ large enough so that the above inequality holds for $\epsilon>0$ small enough. This completes the proof of Corollary \ref{prop:exp-decay}.
\end{proof}
\medskip
Now, we denote $d_n$ the threshold sequence as given by Table 1 in Lemma \ref{LDH}, and $\gamma_n$ a sequence such that $\gamma _n \gg d_n$. From the uniform conclusion \eqref{conclusion-table1} of Lemma \ref{LDH}, there is $\epsilon_0>0$ such that, for all $0<\epsilon<\epsilon_0$, all $L\geq \gamma_{M_\epsilon}$,
$$
e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right)\leq
2e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!}iR_1\left((1-m)L\right)\leq
2 T_\epsilon R_1\left((1-m)L\right).
$$
As a consequence the criterion \eqref{estrb2} is asymptotically satisfied as soon as $L\geq L_\epsilon$ where
\begin{equation}\label{condition}
L_\epsilon \gg d_{M_\epsilon}\quad \text{ and }\quad
R_1\left((1-m)L_\epsilon\right)\leq C\frac{\epsilon}{\ln \frac{1}{\epsilon}},
\end{equation}
where $C=C(\theta, r^-)$ is a positive constant and where we recall that $M_\epsilon=[\gamma^-T_\epsilon]$, $T_\epsilon=\frac{1}{r^-}\ln\frac{2(\eta-\theta)}{\epsilon}$. We now rely on Table 1 to compute $L_\epsilon$ for different kernels in Corollary \ref{prop:examples-be=2}.
\medskip
\begin{proof}[\bf Proof of Corollary \ref{prop:examples-be=2}-$(i)$] In that case, since $$
R_1\left((1-m)L_\epsilon\right)\le \frac{2c_2}{\alpha(1-m)^\alpha}\frac{1}{L_\epsilon^\alpha},
$$
the second condition in \eqref{condition} is asymptotically satisfied if
$$(1-m)L_\epsilon=C^+\left(\frac1\epsilon\ln\frac 1\epsilon\right)^{\frac1\alpha}\text{ with appropriate } C^+=C^+(\theta, r^-, c_2, \alpha)>0.
$$
In view of the line concerning $RV(\alpha)$ in Table 1 of Lemma \ref{LDH}, the first condition in \eqref{condition} is also satisfied. This proves Corollary \ref{prop:examples-be=2}-$(i)$.
\end{proof}
\vspace{1ex}
\begin{proof}[\bf Proof of Corollary \ref{prop:examples-be=2}-$(ii)$] For $RV(\alpha)$ with $\alpha>2$ and any given $0<\tilde \alpha<\alpha$, it follows from \eqref{tail-RV} and known properties of slowly varying functions, see e.g. \cite[VIII, Lemma 2]{Feller}, that the second condition in \eqref{condition} is asymptotically satisfied if
$$
L_\epsilon=\left(\frac1\epsilon\ln\frac1\epsilon\right)^{\frac{1}{\tilde \alpha}}.
$$
In view of the line concerning $RV(\alpha)$ in Table 1 of Lemma \ref{LDH}, the first condition in \eqref{condition} is also satisfied. This proves Corollary \ref{prop:examples-be=2}-$(ii)$.
\end{proof}
\medskip
\begin{proof}[\bf Proof of Corollary \ref{prop:examples-be=2}-$(iii)$] For $LN(\gamma,\lambda,\rho)$ with $\gamma>1$, $\lambda>0$, $\rho \in \mathbb{R}$, and given $0<\tilde \lambda<\lambda$, it follows from \eqref{tail-LN} that the second condition in \eqref{condition} is asymptotically satisfied if
$$
L_\epsilon= \exp \left[\left(\frac{1}{\tilde \lambda} \ln \frac 1\epsilon\right)^\frac{1}{\gamma}\right].
$$
In view of the two lines concerning $LN(\gamma,\lambda,\rho)$ in Table 1 of Lemma \ref{LDH}, the first condition in \eqref{condition} is also satisfied. This proves Corollary \ref{prop:examples-be=2}-$(iii)$.
\end{proof}
\medskip
\begin{proof}[\bf Proof of Corollary \ref{prop:examples-be=2}-$(iv)$] For $WE(\alpha,\lambda,\rho)$ with $0<\alpha<1$, $\lambda>0$, $\rho \in \mathbb{R}$, in view of \eqref{tail-WE} and the line concerning $WE(\alpha,\lambda,\rho)$ in Table 1 of Lemma \ref{LDH}, to reach the condition \eqref{condition} it is sufficient to have
\begin{equation}\label{condition2}
L_\epsilon\gg \left(\ln \frac 1 \epsilon\right)^{\frac{1}{2(1-\alpha)}} \quad \text{ and } \quad L_\epsilon^\rho e^{-\lambda (1-m)^\alpha L_\epsilon^\alpha} \leq C' \frac{\epsilon}{\ln \frac 1 \epsilon},
\end{equation}
for some appropriate $C'=C'(\theta,r^-,m,\alpha)$.
When $0<\alpha<\frac 23$, the choice
$$
L_\epsilon=\frac{1}{1-m}\left(\frac{1}{\tilde \lambda}\ln \frac 1 \epsilon\right)^{\frac 1 \alpha},
$$
for any given $0<\tilde \lambda<\lambda$, ensures that both conditions in \eqref{condition2} are asymptotically satisfied.
On the other hand, when $\frac 23 \leq \alpha <1$, the first condition in \eqref{condition2} prevents such a choice and we are compelled to take
$$
L_\epsilon=\left(\ln \frac 1 \epsilon\right)^{\frac{1}{2(1-\tilde \alpha)}},
$$
for any given $\frac23\le\alpha<\tilde \alpha<1$, for the two conditions in \eqref{condition2} to be asymptotically satisfied. This proves Corollary \ref{prop:examples-be=2}-$(iv)$.
\end{proof}
\medskip
Observe that when $\frac 2 3 \leq \alpha<1$ the above estimate on $L_\epsilon$ is not so good in particular when $\alpha\to1$. However for some specific forms of Weibull-like tails $WE(\alpha,\lambda,\rho)$ with $\rho=0$, we can rely on \eqref{WEbe=0} to obtain a sharper estimate.
\medskip
\begin{proof}[\bf Proof of Corollary \ref{prop:Weibull}] By a change of variable, the estimate for $WE(\alpha, \lambda, \rho=0)$ can be transformed into one of \eqref{tail-weibull-zero}, namely for $WE(\alpha, 1, \rho=0)$. From \eqref{tail-weibull-zero} and \eqref{WEbe=0}, we have, up to a multiplicative constant,
\begin{eqnarray}
e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} R_i\left((1-m)L\right) & \leq & e^{-T_\epsilon}\sum_{i=M_\epsilon}^{N_\epsilon} \frac{T_\epsilon^i}{i!} \left[\exp\left(-\frac{(1-m)^2}{20i}L^2\right)+ i \exp\left(-\frac{(1-m)^\alpha}{2^\alpha}L^\alpha\right)\right]\nonumber\\
& \leq & \exp\left(-\frac{(1-m)^2}{20N_\epsilon}L^2\right)+ T_\epsilon \exp\left(-\frac{(1-m)^\alpha}{2^\alpha}L^\alpha\right).\nonumber
\end{eqnarray}
Recalling $N_\epsilon=N(T_\epsilon)=[3T_\epsilon]+1$ and $T_\epsilon=\frac{1}{r^-}\ln\frac{2(\eta-\theta)}{\epsilon}$, one can check that the above right hand side is asymptotically smaller than $\frac{\epsilon}{4\theta}$ if we choose
$$
L=L_\epsilon:=\frac{C^+}{1-m}\left(\ln \frac 1 \epsilon\right)^{\frac 1 \alpha}
$$
with $C^+=C^+(J)>0$ large enough. As a result, the non-extinction criterion \eqref{estrb2} is satisfied. We have thus reached the announced sharper estimate when $\rho=0$ (which, when $\alpha \to 1$, is consistent with the exponential case shown in Corollary \ref{prop:exp-decay}). This completes the proof of Corollary \ref{prop:Weibull}.
\end{proof}
\section{Quantitative estimates of the threshold phenomena}\label{quantitative}
In this section, relying on Sections \ref{extinction} and \ref{non-extinction}, we complete the proof of the main results of Section \ref{main results}. We denote $u_L^\epsilon=u^\epsilon_L(t, x)$ the solution to
$$
\partial _t u=J\ast u-u+f(u),
$$
starting from $\phi_L^\epsilon(x)=(\theta+\epsilon)\mathds{1}_{(-L, L)}(x)$. We start with the extinction results.
\medskip
\begin{proof}[Proof of Theorem \ref{ext}]
By \eqref{r^+}, $u_L^\epsilon$ is a sub-solution to problem \eqref{super}. As a result, Theorem \ref{ext} follows from Proposition \ref{prop:extinction} and the comparison principle.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary \ref{asmptotic1}] The proof is a rather straightforward combination of Corollary \ref{asmptotic} and the fact that $ \max_{T>0}T e^{-r^+T}= e^{-1}/r^+$.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary \ref{asmptotic2}] The proof is a straightforward consequence of Corollary \ref{lighttail}.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary \ref{ext:examples}] The proof is a straightforward consequence of Corollary \ref{remarkbeta=2} and Remark \ref{rem:as-easily}.
\end{proof}
\medskip
Let us now investigate the propagation results.
\medskip
\begin{proof}[Proof of Theorem \ref{theo-propagation}]
From \eqref{r-moins} in Assumption \ref{ass:f}, one has
$$
f(w)\geq \tilde{g}(w),\quad\forall w\in (-\infty, \delta],
$$
where $\tilde{g}(w)=r^-(w-\theta)$ was defined in \eqref{g}. Let $0<\alpha<\frac{\delta-\theta}2$ be given small enough so that we can define a Lipschitz continuous function $\tilde{f}: \mathbb{R}\to\mathbb{R}$ such that $\tilde f\leq f$,
\begin{eqnarray}
\tilde{f}(w)=\begin{cases}
f(w)\quad &\text{for } w\in (-\infty, \theta) \cup [\delta, \infty),\\
r^-(w-\theta)\quad &\text{for } w\in [\theta, \theta+2\alpha],
\end{cases}
\end{eqnarray}
and $\tilde f$ satisfies Assumption \ref{ass:f} (in particular $\int_0^1\tilde f(s)ds>0$ and \eqref{r-moins} holds on $[0,\theta+2\alpha]$). Denote $\tilde{w}=\tilde{w}(t, x)$ the solution to
\begin{equation}\label{tildew}
\partial _t \tilde{w}=J\ast \tilde{w}-\tilde{w}+\tilde{f}(\tilde{w}),
\end{equation}
starting from $ \tilde{w}(0, x)=\phi _L ^\epsilon(x)$, so that $\tilde{w}(t, x)\leq u_L^\epsilon(t, x)$ from the comparison principle. Consider the time $T_\epsilon=\frac{1}{r^-}\ln\frac{2\alpha}{\epsilon}$. For $0<\epsilon\leq 2 \alpha$ we know from Proposition \ref{nonextinction} $(i)$ (setting $\eta=\theta+\alpha$) that $\tilde{w}\leq\theta+2\alpha\le\delta$ on $[0, T_\epsilon]\times\mathbb{R}$. Since
$$
\tilde{f}(w)\geq \tilde g (w), \quad\forall w\in(-\infty, \theta+2\alpha],
$$
one obtains from Proposition \ref{nonextinction} $(ii)$ that, for any given $m\in(0, 1)$, there exists $\epsilon_0>0$ such that for all $\epsilon\in(0, \epsilon_0)$ and $L>L_\epsilon$, where $L_\epsilon$ satisfies \eqref{nonextinctioncondition}, one has
$$
u_L^\epsilon(T_\epsilon,x)\geq \tilde{w}(T_\epsilon, x)\geq(\theta+\alpha)\mathds{1}_{(-mL_\epsilon,mL_\epsilon)}.
$$
From the propagation Assumption \ref{ass:prop} for the nonlinearity $f$, we know that $L_\alpha^{prop}<+\infty$ exists, that is, for $\ell > L_\alpha^{prop}$ the solution to \eqref{nonlocal} starting from $(\theta+\alpha)\mathds{1}_{(-\ell,\ell)}$ propagates. As a result, for $\epsilon>0$ small enough so that $mL_\epsilon> L_\alpha^{prop}$, one has $u_L^\epsilon(T_\epsilon+t, x)\to 1$ as $t\to+\infty$ locally uniformly in space and therefore $u_L^\epsilon(t, x)\to1$ as $t\to+\infty$ locally uniformly in space. We have thus proved that, for $L_\epsilon$ satisfying \eqref{nonextinctioncondition}, we have $L_\epsilon^{prop}\leq L_\epsilon$ for $\epsilon>0$ small enough. This completes the proof.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary \ref{prop:corollary1}]
Combining the arguments in the proof of Theorem \ref{theo-propagation} with Corollary \ref{prop:0<beta<2}, one can obtain the desired results.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary \ref{prop:corollary2}]
Combining the arguments in the proof of Theorem \ref{theo-propagation} with Corollary \ref{prop:exp-decay}, one can obtain the desired results.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary \ref{prop:corollary3}]
Combining the arguments in the proof of Theorem \ref{theo-propagation} with Corollary \ref{prop:examples-be=2} and Corollary \ref{prop:Weibull}, one can obtain the desired results.
\end{proof}
\section{Propagation threshold}\label{propagation}
This section is devoted to the proof of the propagation threshold result, namely Theorem \ref{theo-ass-prop}. We shall rely on some ideas developed by
Fife and McLeod in \cite[Theorem 3.2]{fife1977approach} and crucially make use of the following integrability properties of the wave profile $U$.
\begin{lemma}\label{integrability}
Let Assumption \ref{TWA} be satisfied. Then any monotone traveling wave $(c,U)$ solving \eqref{TW}, and whose existence follows from \cite{bates1997traveling},
satisfies the integrability properties
$$
\int_{-\infty}^0 U(x)dx<+\infty \quad \text{ and } \quad \int_0^{+ \infty}(1-U(x))dx<+\infty.
$$
\end{lemma}
\begin{proof} This is nothing else than \cite[(5.9)]{bates1997traveling} but, to enlight the importance of the finite first moment hypothesis \eqref{first_moment} and for the convenience of the reader, we give a proof of the integrability of $1-U$ in $+\infty$ (the other one being similar).
From the equation satisfied by $U$, $U(+\infty)=1$, and the assumption $f'(1)<0$, there are $R>0$ and $C_1>0$ such that
\begin{equation}\label{ineqq}
-(J\ast U)(x)+U(x)+cU'(x)=f(U(x))\ge C_1(1-U(x))>0,\quad \forall x\geq R.
\end{equation}
Now we consider $\ell>0$. When $c\neq 0$ the regularity of the wave, see subsection \ref{ss:towards}, enables to write
\begin{eqnarray*}
\int_R^{R+\ell} (J\ast U(x)-U(x))dx&=&\int_R^{R+\ell} \int_{-\infty}^{+\infty} J(y)[U(x-y)-U(x)] \,dy dx\nonumber\\
&=&-\int_R^{R+\ell} \int_{-\infty}^{+\infty} J(y)\int_0^1 yU'(x-sy) \, ds dy dx\nonumber\\
&=&-\int_{-\infty}^{+\infty} yJ(y) \int_0^1\left( \int_{R}^{R+\ell} U'(x-sy) dx\right) ds dy\nonumber\\
&=&-\int_{-\infty}^{+\infty} yJ(y)\int_0^1\left(U(R+\ell-sy)-U(R-sy)\right) ds dy,\nonumber
\end{eqnarray*}
thanks to Fubini's theorem. Note that, when $c=0$, a mollifying argument as in \cite[Lemma 3.2]{alfaro2017propagation} shows that the above conclusion is still valid. As a result, since $0\leq U\leq 1$, we get, for any $\ell>0$,
\begin{equation*}
\left|\int_R^{R+\ell} (J\ast U(x)-U(x))dx\right| \leq 2\int_{-\infty}^{+\infty}|y|J(y)dy.
\end{equation*}
On the other hand, for any $\ell>0$,
$$
\left|\int_R^{R+\ell} cU'(x) dx \right|= \vert c\vert \left( U(R+\ell)-U(R) \right) \le \vert c\vert.
$$
We thus deduce from \eqref{ineqq} that, for any $\ell>0$,
$$
C_1\int_R^{R+\ell}(1-U(x))dx \le \vert c\vert +2\int_{-\infty}^{+\infty}|y|J(y)dy <+\infty,
$$
which implies that $1-U\in L^1(R,\infty)$ and completes the proof of the lemma.
\end{proof}
\medskip
We now turn to the proof of Theorem \ref{theo-ass-prop}.
\medskip
\begin{proof}[Proof of Theorem \ref{theo-ass-prop}] Let Assumption \ref{TWA} be satisfied and assume $\int_0^1f(u)du>0$. Consider $(c,U)$ an increasing traveling wave solution to \eqref{nonlocal} with $c>0$. We aim at showing that, for any $\epsilon\in(0,1-\theta]$, there exists $L>0$ large enough such that propagation occurs for $u_L^\epsilon$ the solution to \eqref{nonlocal} starting from \eqref{initial}. To do so and as mentioned above, we rely on the approach of \cite{fife1977approach} for the classical diffusion case.
Let us consider the function $\underline{u}$ given by
$$
\underline{u}(t, x):=U_+(t, x)+U_-(t, x)-1-q(t),
$$
with $U_\pm(t, x):=U(\zeta_\pm(t,x))$, where $\zeta_\pm(t,x)$ take the form
\begin{equation*}
\zeta_+(t,x)=x+ct-\xi(t),\;\; \zeta_-(t,x)=\zeta_+(t,-x)=-x+ct-\xi(t).
\end{equation*}
Here $q=q(t)$ and $\xi=\xi(t)$ are functions to be determined for $\underline{u}$ to be a sub-solution to \eqref{nonlocal}.
From the above and the $U$-equation, we straightforwardly compute, for $t>0$ and $x\in \mathbb{R}$,
\begin{eqnarray}\label{sub2}
N\underline{u}(t,x) &:=& \partial_t \underline{u}(t,x)-J\ast\underline{u}(t,x)+\underline{u}(t,x)-f\left(\underline{u}(t,x)\right)\nonumber \\
&{}=&-\xi'(t)\left[U'(\zeta_+(t,x))+U'(\zeta_-(t,x))\right]\nonumber\\
&&+f\big(U_+(t,x)\big)+f\big(U_-(t,x)\big)-f\Big(U_+(t,x)+U_-(t,x)-1-q(t)\Big)-q'(t).
\end{eqnarray}
Before going further, let us introduce some notations. Denote $\alpha:=\theta+\epsilon\in (\theta,1]$ the fixed height of the step initial data. Fix two constants $1-\alpha<q_0<q_1<1-\theta$ so that
$$
\theta<1-q_1<1-q_0<\alpha,
$$
and define the function $\Phi$, continuous on $\mathbb{R}\times [0,+\infty)$, as
$$
\Phi(u, s):=
\begin{cases}
\frac{f(u-s)-f(u)}s, \quad &\text{ if } s>0,\\
-f'(u), \quad &\text{ if } s=0.
\end{cases}
$$
Moreover, for $0<s\leq q_1$ we have $\theta<1-q_1\leq1-s<1$, so that $\Phi(1, s)>0$. Also $\Phi(1, 0)=-f'(1)>0$. Thus there exists $\mu>0$ such that $\Phi(1, s)\geq 2\mu$ for $0\leq s\leq q_1$. By continuity, there exists a $\delta>0$ such that
$\Phi(u, s)\geq\mu$ for $1-\delta\leq u\leq1$ and $0\leq s\leq q_1$. It then follows that
\begin{equation}\label{f'(1)<0}
f(u-s)-f(u)\geq \mu s, \quad \text{ for all } 1-\delta\leq u\leq1 \text{ and } 0\leq s\leq q_1.
\end{equation}
Last, we fix $b>0$ large enough so that
\begin{equation}\label{def-b}
f(u)\leq b(1-u), \quad \text{ for all } 0\leq u \leq 1.
\end{equation}
\begin{claim}\label{limitbehavior}
For any $s_0\geq 0$,
$$
g(t):= \int_{0}^{t}e^{-\mu(t-s)}(1-U(cs+s_0))ds
$$
tends to $0$ as $t\to+\infty$.
\end{claim}
\begin{proof}
Observe first that $0<g(t)\leq\int_0^t e^{-\mu(t-s)} ds=\frac{1}{\mu}\left(1-e^{-\mu t}\right)$, so that $g\in L^\infty(0, \infty)$. Next note that
$g'(t)=1-U(ct+s_0)-\mu g(t)$, so that $g'\in L^\infty(0, \infty)$. Next, by Fubini-Tonelli's theorem,
\begin{eqnarray}
\int_0^{+\infty} g(t) dt &=& \int_0^{+\infty} \int_0^t e^{-\mu(t-s)}(1-U(cs+s_0)) ds dt = \int_0^{+\infty} e^{\mu s} (1-U(cs+s_0)) \int_s^{+\infty} e^{-\mu t} dt ds\nonumber\\
&=&\frac{1}{\mu}\int_0^{+\infty}(1-U(cs+s_0)) ds=\frac{1}{c\mu}\int_{s_0}^{+\infty}(1-U(x))dx\nonumber\\
&\le&\frac{1}{c\mu}\int_{0}^{+\infty}(1-U(x))dx=\frac{1}{c\mu}\norm{1-U}_{L^1(0, \infty)}<+\infty,\label{g(t)}
\end{eqnarray}
from Lemma \ref{integrability}, so that $g\in L^1(0, \infty)$. Now, the combination of $g\in L^1(0, \infty)$ and $g'\in L^\infty(0, \infty)$ enforces $g(t)\to 0$ as $t\to+\infty$, which completes the proof of the claim.
\end{proof}
\medskip
From $g(0)=0$ and the above claim, $g$ attains its maximum at some $t_0>0$, and
\begin{equation}\label{g(t_0)}
g(t_0)=\max_{t\geq 0}g(t)=\frac{1}{\mu}(1-U(ct_0+s_0))\le \frac{1}{\mu}(1-U(s_0)).
\end{equation}
For constants $\xi_0<0$ and $\eta_0>0$ with $s_0:=-\eta_0-\xi_0>0$ to be be determined below, we select
\begin{equation}\label{q}
q(t):=q_0e^{-\mu t} + bg(t)=q_0e^{-\mu t} + b\int_{0}^{t}e^{-\mu(t-s)}(1-U(cs+s_0))ds,\quad t \geq 0.
\end{equation}
We also let $\xi(t)=\xi_0+\eta(t)$ where $\eta$ is to be selected below with the properties
\begin{equation}\label{eta-to-be-selected}
\eta(0)=0, \; \eta'(t)> 0,\; \eta(t)\leq \eta_0\leq -\xi_0.
\end{equation}
In the sequel, we aim at reaching $N\underline u(t,x)\leq 0$ for all $x\in \mathbb{R}$, $t> 0$. Since both $\underline u(t,\cdot)$ and $J$ are symmetric, it is sufficient to work with $x\geq 0$. Since $U'>0$ we have, for all $x\geq 0$ and $t> 0$,
\begin{eqnarray}
1-U_+(t,x)+q(t) &=& 1-U(x+ct-\xi(t))+q(t) \nonumber\\
&\le& 1-U(ct-\xi_0-\eta(t))+q_0+bg(t_0)\nonumber\\
&\leq& 1-U(-\xi_0-\eta_0)+q_0+\frac{b}{\mu}(1-U(-\xi_0-\eta_0))\nonumber\\
&=&(1+b/\mu)(1-U(s_0))+q_0.\nonumber
\end{eqnarray}
Choose $s_0>0$ large enough so that
$(1+b/\mu)(1-U(s_0))+q_0\leq q_1$.
As a consequence, for any such choice, one has, for all $t\geq 0$ and $x\geq 0$,
\begin{equation}\label{q_1}
0 \le 1-U_+(t,x)+q(t) \le q_1.
\end{equation}
Below we complete the construction of the sub-solution by investigating the sign of $N\underline u(t,x)$ for $x\geq 0$ and $t> 0$. To do so, recalling that $\delta>0$ was chosen above for \eqref{f'(1)<0} to hold, we split our analysis according to the value of $U_-(t,x)$.
\medskip
\noindent \textbf{First case: $1-\delta\le U_-(t,x) \le 1$.} Then, from \eqref{f'(1)<0} and \eqref{q_1},
\begin{equation}\label{truc}
f\big(U_-(t,x)\big)-f\Big(U_-(t,x)-(1-U_+(t,x)+q(t))\Big)
\leq-\mu(1-U_+(t,x)+q(t)).
\end{equation}
Plugging this into \eqref{sub2}, using $U'>0$, $\xi'(t)>0$ and \eqref{def-b}, we reach
\begin{eqnarray*}
N\underline{u}(t,x) &\le& -\mu(1-U_+(t,x)+q(t))+b(1-U_+(t,x))-q'(t)\nonumber\\
&=&(b-\mu)(1-U_+(t,x))-\mu q(t)-q'(t)\nonumber\\
&\le & b(1-U_+(t,x))-\mu q(t)-q'(t)\nonumber\\
&=& b(1-U(x+ct-\xi(t)))-\mu q(t)-q'(t)\nonumber\\
&=& b(1-U(x+ct-\xi(t)))-b(1-U(ct-\xi_0-\eta_0))
\end{eqnarray*}
from the definition of $q(t)$ in \eqref{q}. Since $U'>0$, $x\geq 0$ and $-\xi(t)\geq -\xi_0-\eta_0$, we end up with $N\underline u(t,x)\leq 0$.
\medskip
\noindent \textbf{Second case: $0\le U_-(t,x)\le \delta$.} Let us recall that
$f\in C^1(\mathbb{R})$ and $f'(0)<0$. Therefore, up to modify $f$ on $(-\infty,0)$ (which is harmless for the problem under consideration since solutions are nonnegative), we may assume that there are $\tilde{\mu}>0$ and $\tilde{\delta}>0$ such that
\begin{equation}\label{concave}
f'(u)\le -\tilde{\mu},\quad \forall u\in(-\infty, \tilde{\delta}].
\end{equation}
Also, up to reducing $\mu$ and $\delta$ appearing in \eqref{f'(1)<0} if necessary, we may assume $0<\delta\le\tilde{\delta}$ and $0<\mu\le\tilde{\mu}$. As a result,
\begin{equation*}
f(u)-f(u-s)=\int_{u-s}^u f'(\sigma)d\sigma \leq -\mu s, \quad \text{ for all } -\infty< u\leq \delta \text{ and } s\ge 0.
\end{equation*}
From this we, again, deduce \eqref{truc} and conclude as in the first case.
\medskip
\noindent \textbf{Third case: $\delta\leq U_-(t,x)\leq 1-\delta$.} If we denote $C>0$ the Lipschitz constant of $f$ on the interval $[\delta-q_1,1-\delta]$, we deduce from $\delta\leq U_-(t,x)\leq 1-\delta$ and \eqref{q_1} that
\begin{equation}\label{truc-bis}
f\big(U_-(t,x)\big)-f\Big(U_-(t,x)-(1-U_+(t,x)+q(t))\Big)
\leq C(1-U_+(t,x)+q(t)).
\end{equation}
From \eqref{def-b}, we have $f(U_+(t,x))\leq b(1-U_+(t,x))$. Moreover, in this third case, we have
\begin{equation}\label{middle}
U'(\zeta_+(t,x))+U'(\zeta_-(t,x))\geq U'(\zeta_-(t,x))\geq \min_{U^{-1}(\delta)\leq z\leq U^{-1}(1-\delta)} U'(z) := \vartheta>0.
\end{equation}
Plugging this into \eqref{sub2}, we get
\begin{eqnarray*}
N\underline{u}(t,x)&\le& -\vartheta\xi'(t)+(C+b)(1-U_+(t,x))+Cq(t)-q'(t)\nonumber\\
&=& -\vartheta\eta'(t) +C(1-U_+(t,x))+(C+\mu)q(t)+b\left(U(ct+s_0)-U(\zeta_+(t,x)\right)
\end{eqnarray*}
from computing $q'(t)$. Since $\zeta_+(t,x)=x+ct-\xi(t)\geq ct-\xi_0-\eta_0=ct+s_0$ and $U'>0$, we obtain
\begin{equation}
\label{sub3}
N\underline{u}(t,x)\leq -\vartheta\eta'(t) +C\left(1-U(x+ct-\xi(t))\right)+(C+\mu)q(t).
\end{equation}
We now select
\begin{eqnarray}
\eta(t) &:=& \frac{C}{\vartheta}\int_{0}^{t}(1-U(cs+s_0))ds+\frac{(C+\mu)q_0}{\vartheta}\int_{0}^{t}e^{-\mu s}ds\nonumber\\
&& +\frac{b(C+\mu)}{\vartheta}\int_{0}^{t}\int_{0}^{s}e^{-\mu(s-\tau)}(1-U(c\tau+s_0))d\tau ds.\nonumber
\end{eqnarray}
Obviously $\eta(0)=0$ and
\begin{equation}\label{vartheta-prime}
\vartheta \eta'(t)=C\left(1-U(ct+s_0)\right)+(C+\mu)q(t)>0,
\end{equation}
and thus $\eta(t)\leq \eta(+\infty)$ for all $t\geq 0$. We estimate $\eta(+\infty)$ as follows:
\begin{eqnarray*}
\eta(+\infty)&=& \frac{C}{\vartheta}\int_{0}^{+\infty}(1-U(cs+s_0))ds+\frac{(C+\mu)q_0}{\vartheta}\int_{0}^{+\infty}e^{-\mu s}ds\nonumber\\
&& +\frac{b(C+\mu)}{\vartheta}\int_{0}^{+\infty}\int_{0}^{s}e^{-\mu(s-\tau)}(1-U(c\tau+s_0))d\tau ds\\
&\le& \frac{C}{c\vartheta}\int_0^{+\infty}(1-U(x))dx+\frac{(C+\mu)q_0}{\vartheta\mu}+\frac{b(C+\mu)}{c\vartheta\mu}\int_0^{+\infty}(1-U(x))dx=:\eta_0,\nonumber
\end{eqnarray*}
from the same computation as in \eqref{g(t)}.
Plugging \eqref{vartheta-prime} into \eqref{sub3}, we reach $
N\underline{u}(t,x) \leq C\left(U(ct+s_0)-U(x+ct-\xi(t)\right)
$ which is nonpositive as already argued above.
\medskip
\noindent \textbf{Conclusion.} With the above choices, we have therefore verified that $N\underline{u}(t, x)\leq0$ for all $(t, x)\in(0, +\infty)\times\mathbb{R}$. For $|x|\le L$, one has
$$
\underline{u}(0, x)=U(x-\xi_0)+U(-x-\xi_0)-1-q_0< 1-q_0<\alpha=\alpha\mathds{1}_{(-L, L)}(x).
$$
For $\vert x\vert \geq L$, one has
$$
\underline{u}(0, x)=U(x-\xi_0)+U(-x-\xi_0)-1-q_0\leq U(-L-\xi_0)-q_0<0,
$$
if $L=L(\xi_0)>0$ is large enough. As a result, for such a large $L>0$,
$$
\underline{u}(0, x)\leq\alpha\mathds{1}_{(-L, L)}(x),\quad \forall x\in\mathbb{R}.
$$
It follows from the comparison principle that
$u(t,x)\geq \underline u(t,x)$ for all $t\geq 0$, $x\in \mathbb{R}$. Since $\underline u$ satisfies \eqref{qqch}, so does $u$ and the proof is complete.
\end{proof}
|
\section{Introduction}
This paper revisits the seminal work of Malkus (1956) which attempted to build a theory of shear turbulence. This theory was based upon maximising the momentum transport (or equivalently dissipation rate) achieved by the flow amongst all those with a marginally-stable mean profile. Malkus clearly had a statistical form of marginal stability in mind but, to make progress, had to resort to specifying marginality with respect to the then 50-year-old Orr-Sommerfeld equation \citep{Orr1907, Sommerfeld1908}. So posed, his marginal stability idea was quickly repudiated (Reynolds \& Tiederman 1967 and more recently Iyer {\it et al.} 2019) although further studies showed it could be made to work if {an anisotropic eddy viscosity model was used} \citep{ReynoldsHussain72, Sen00,Sen07,Malkus79}. The concept of statistical stability was, however, central to Malkus's thinking (remaining so throughout his career, e.g. Malkus 1996 and 2003) and is clearly different from stability as viewed within the context of the governing Navier-Stokes equations (epitomised by the celebrated Orr-Sommerfeld equation).
For example, it is fairly uncontentious to assert that the turbulent attractor in, say, pressure-driven channel flow {at high enough Reynolds number and large enough domain} has stationary statistics (defined by averaging over one or more homogeneous directions or in an ensemble sense) and so within the partial differential equations which govern how these statistics evolve, the realised turbulence is a stable fixed point - i.e. turbulence is {\em statistically} stable to infinitesimal perturbations of the statistics. This is in contrast to the time-dependent turbulent attractor as viewed in the Navier-Stokes equations where adding a small disturbance may well see that disturbance grow and never decay yet the original statistics still recover (an example is shown herein). The difference, of course, is that a flow disturbance can have a component along the turbulent trajectory and hence acts as a time-shift: this part of the disturbance never decays to zero but does not affect the statistics \cite[e.g.][]{Nikitin18}.
For his theory, Malkus wanted a statistical stability criterion based only on the lowest order statistic - the mean flow. Ideally more statistical information needs to be incorporated and it is our objective here to attempt this.
At the very least, doing so should improve the now-standard approach of carrying out linear stability analysis of the mean profile of time-dependent flows in an attempt to understand observed coherent structures \cite[e.g.][]{CrightonGaster76, Gaster85, Roshko93, Lesshafft06,
Barkley06, SippLebedev07,Akervik08,Sipp10, Mantic-Lugo14, Beneddine16,Lefauve18}.
The motivation for this work comes from two different directions. The first is the general approach of applying linear analysis around the mean profile of a time-varying, possibly turbulent flow to deduce information about the likely fluctuations seen. Initially, this took the form of linear stability analysis stimulated by Malkus's work which tends to work well in free-shear turbulent flows like jets where inviscid (inflectional) instabilities dominate \citep{CrightonGaster76,Gaster85,Roshko93} but less well in wall-bounded situations {such as channel flow} where viscosity can be important \citep[although there have been successes e.g.][]{Barkley06,SippLebedev07,Beneddine16}. Driven by the fact that shear flow mean profiles tend to be linearly stable, this approach subsequently diversified into non-modal analysis \citep{Butler93, DelAlamo06, Cossu09} and input-output or resolvent analyses \citep{Jovanovic05, Cherny05, HwangCossu10a, HwangCossu10b, McKeonSharma10, Moarref12, Blesbois13, SharmaMcKeon13, Schmidt18}. The resolvent approach has been particularly illuminating in showing exactly how far linear analysis can go given the mean flow profile \citep{McKeon17, Jovanovic21}. Predictions can be made of the dominant fluctuation response at a given frequency and wavevector which typically resonates with observations at least up to amplitude and phase. In some sense this solves `half' the problem of turbulence - given a mean flow, linear analysis around this can extract the dominant fluctuation structures - and refocuses attention on the `other' half - predicting the mean profile. Identifying the required amplitudes and phases of the all fluctuation fields to support the observed mean profile, however, may approach the difficulty of 'just' solving the Navier-Stokes equations. One alternative is to appeal to some simpler `organising' physical principle as epitomised in \cite{Malkus56} and then the idea of statistical stability seems a key concept.
The second motivation is the resurgence of interest in dealing with statistical quantities directly through a cumulant expansion of time-varying flows \citep{Hopf52, Orszag73, Frisch95}. Generating the evolution equations for cumulants immediately highlights the closure problem of turbulence in which the time-derivative of a $n$th-order cumulant requires knowledge of the $(n$+$1)$th cumulant and so forth. Present-day computational power has, however, renewed interest in the pursuit of simple cumulant-discard closures to see how well they do in modelling flows in a variety of contexts: e.g. in atmospheric dynamics \citep{FarrellIoannou03, Marston08, SrinivasanYoung12, Tobias13, ParkerKrommes14}; astrophysics \citep{Tobias11}; plasmas \citep{FarrellIoannou09,ParkerKrommes13} and wall-bounded shear flows \citep{FarrellIoannou12, Constantinou14, Farrell16}. The most popular closure ignores third and higher order cumulants - commonly referred to as CE2 - and has the nice property of being exact for a quasilinear (or historically a `mean field theory') version of the Navier-Stokes equations \citep[e.g. see the recent review by][]{MarstonTobias22}. Significantly for our purposes here, CE2 and higher closures (CE$n$ where $n \geq 3$ is the highest order cumulant retained) present the most natural framework in which to extend Malkus's idea of statistical stability. In what follows, the focus will be on using CE2 and its relationship to the quasilinear version of the Navier-Stokes equations to develop an improved version of the usual Orr-Sommerfeld analysis (or `OS analysis' for short) of the mean profile: see Figure \ref{fig:1}. {The overall objective is to develop a way to judge whether a state is statistically stable or not using a subset of its statistics.}
It is worth emphasizing that the approach taken here is perfectly general and not confined to shear flows or even fluid mechanics. The key idea given a physical system described by a PDE is to generate evolution equations - `statistical' equations - for the first few statistics of the solution. If the solution of the original PDE is spatiotemporally complicated but has steady statistics, the premise pursued here is that a better way to assess the stability of this solution is to look at the linear stability {\em within} the statistical equations rather than the original PDEs. The former is an approximation given a finite number of the system's statistics are considered but yields a spectral problem built upon a steady state in statistical space. The latter has to rely on a costly ensemble of simulations where a distribution of small perturbations are each added to the solution and their evolution monitored \citep{Iyer19,Duguet22}. A good complementary example of where this approach would be useful is convection \citep{Malkus54, Wen22}.
\begin{figure}
\centerline{ \includegraphics{fig1.png}}%
\caption{\label{fig:1} Comparison between standard and statistical considerations of linear stability of turbulent states. On the left are physical space equations (methods are ordered top to bottom by decreasing non-linearity): Full Navier-Stokes equations (NS), Quasilinear approximation (QL), Orr-Sommerfeld equation around a turbulent mean velocity profile (OS), Extended and minimal Extended Orr-Sommerfeld equations (EOS and mEOS). On the right are statistical space equations: CE2 equations which include up to second order statistics. Black arrows indicate the standard path of turbulent flow stability analysis leading to OS; blue arrows indicate the statistical approach to turbulent flow stability analysis leading to EOS and mEOS, emphasising how the steady statistical state $U,\bm{C^{mn}}$ can be used to obtain the steady state counterpart $U,\bm{\tilde{u}_0^{mn}}$ in the physical space.
}
\end{figure}
The plan of this paper is to illustrate the analysis within the context of channel flow described in section \ref{II}. The Reynolds number is assumed high enough for the computational box used that the flow can be taken as approximately statistically stationary.
\rk{An averaging procedure will be assumed below such that the mean flow can be assumed to only depend on the cross-stream variable $y$ i.e. ${\bf U}=U(y){\bf \hat{x}}$. This could be ensemble averaging or averaging over the streamwise ($x$) and spanwise ($z$) directions.}
\rk{Using the latter spatially-averaging approach}, Malkus pointedly only chose a spanwise average so that his mean profile ${\bf U}=U(x,y) {\bf \hat{x}}$ could depend on the cross-stream {\em and} streamwise variables. As a result, his Orr-Sommerfeld analysis targeted the stability of a streamwise-independent mean profile $U(y){\bf \hat{x}}$ to streamwise-dependent {\em mean flow} disturbances $\delta {\bf U}(y) \exp(ik(x-ct))$ rather than fluctuation fields defined as having non-vanishing spanwise dependence. Contrarily, there are growing arguments to only streamwise average to retain the spanwise structure of the mean flow, i.e. ${\bf U}=U(y,z){\bf \hat{x}}$ (e.g. see Table 1 of \cite{Lozano-Duran21} for a sample list of relevant works). Despite this, the focus here is on the simplest mean flow definition {for a statistically steady state, ${\bf U}=U(y){\bf \hat{x}}$,} given the central role this plays in resolvent flow analysis but, there's no doubt, extending the mean flow definition is clearly an important direction to extend the approach discussed here. Standard Orr-Sommerfeld analysis is recalled in section \ref{II.A}.
Section \ref{II.B} then introduces the cumulant expansion approach \citep{Hopf52,Orszag73,Frisch95} and the hierarchy of evolution equations for these statistical quantities. Solving these equations to (hopefully) reach a steady state is an appealingly direct way to estimate the properties of statistically steady turbulent flows since it avoids having to average across large DNS-generated datasets.
Here, however, the emphasis is on the concept of stability in this statistical framework and its relationship to (dynamic) stability within the Navier-Stokes equations (see section \ref{II.C}) {\em not} on the accuracy of suitable closures of the cumulant equations in capturing the reference flow state. {That is, our strategy is to apply a statistical stability criterion from a statistical closure to a solution of the Navier-Stokes equations to approximate its statistical stability there. }
Our particular focus will be on the simplest non-trivial cumulant-discard scheme CE2 given the rapidly increasing dimensionality of the approach: CE$n$ works with cumulants up to order $n$ which, before exploiting any symmetries, is typically a $3n$-$2$ rank tensor (3 spatial coordinates per field reduced by 2 averaging directions). CE2 is exact for the quasilinearized Navier-Stokes equations - or QL equations - and translating what a statistically-steady state in CE2 means for the QL equations is a crucial step discussed in \S\ref{II.D}. Appendix \ref{appendixA} presents an equally important discussion on how the stability predictions within the CE2 and QL systems are related. Quasilinearization \citep{Vedenov61, Herring63, Herring64} has enjoyed a resurgence of interest recently \cite[e.g.][]{Hernandez20, Skitka20, OConnor21, MarstonTobias22} given its accessibility, and together with its `generalised' elaboration \citep{Marston16} in which the definition of what constitutes a mean is extended, has the ability to focus on different parts of the nonlinearity in the Navier-Stokes equations \cite[e.g.][]{Hwang21}.
Section \ref{III.A} then introduces an extended version of Orr-Sommerfeld analysis - or `EOS' analysis - based on translating the statistical stability problem in CE2 back to the dynamical equations. Intriguingly, this is {\em not} the same as just working within the QL approximation (as Appendix \ref{appendixA} makes clear). Applying EOS analysis carries a substantial overhead so we consider a reduced (practical) version referred to as `minimally extended Orr-Sommerfeld analysis' - or `mEOS' analysis - in section \ref{III.B} which is almost as cheap to apply as OS analysis.
Interestingly, applying the same strategy of mapping a cumulant-based system back to the underlying dynamical equations can only go one level higher in sophistication and requires a jump directly to CE$\infty$. This transforms EOS into the familiar linearised Navier-Stokes equations albeit based around the steady base state derived by assuming stationary 2nd rank cumulants. This `infinitely extended Orr-Sommerfeld' analysis is described in \S\ref{III.B}. Section \ref{IV} then explores the performance of OS, EOS and mEOS analyses on 4 different turbulent states realised in 2D channel flow which {is used as an approximation of a statistically steady flow}. The limitations of the analysis are discussed in \S\ref{V} followed by a summary and final thoughts in section \S\ref{VI}.
\section{\label{II}Formulation: channel flow}
For context in this work we consider channel flow ${\bf u}^*({\bf x}^*,t^*)$ of a fluid with density $\rho^*$ and kinematic viscosity $\nu^*$ between two parallel plates at $y^*=\pm h^*$ across which a \rk{time-dependent} pressure gradient $9\rho^* U^{*2}G(t^*)/4h^*\,{\bf \hat{x}}$ is imposed such that the bulk flow
\begin{equation}
U^*:= \frac{1}{4h^{*2}L_z} \int^{h^* L_z}_{-h^*L_z}\int^{h^*}_{-h^*} \,{\bf u}^*\, dy^* dz^*
\end{equation}
is kept constant (unstarred/starred quantities are dimensionless/dimensional and periodicity is imposed across the spanwise domain $z^* \in h^*[-L_z,L_z]$). Non-dimensionalizing the Navier-Stokes equations using $h^*$, $3U^*/2$ (so Reynolds numbers based on the bulk speed and the laminar centreline speed $U^{*c}=3U^*/2$ correspond for the laminar parabolic flow) and $\rho^*$ leads to
\begin{equation}
\begin{aligned}
{\bf u}_t+ {\bf u} \cdot \bm{\nabla} {\bf u} &= G(t){\bf \hat{x}}-\bm{\nabla} p +\frac{1}{Re} \nabla^2 {\bf u},\\
\bm{\nabla} \cdot {\bf u} &= 0
\end{aligned}
\label{eq:NSfull}
\end{equation}
with $\frac{1}{4L_z}\int^{L_z}_{-L_z}\int^1_{-1} u\, dydz=2/3$ where ${\bf u}={\bf u}^*/U^*=u {\bf \hat{x}}+v {\bf \hat{y}}+w {\bf \hat{z}}$, $t=3U^*/(2h^*) t^*$ and $Re:=3U^* h^*/2\nu^*$.
We also impose streamwise periodicity of the flow over $x^* \in h^*[-L_x,L_x]$ so the non-dimensionalised flow domain is $(x,y,z) \in [-L_x,L_x] \times [-1,1] \times [-L_z,L_z]$ with non-slip boundary conditions on the plates at $y=\pm 1$ (fundamental wavenumbers in $x$ and $z$ are labelled $\alpha:=\pi/L_x$ and $\beta:=\pi/L_z$ respectively).
\rk{In terms of an averaging procedure, a number of choices present themselves: ensemble averaging, spatial averaging and time averaging (or even a combination thereof) which all should be equivalent for a statistically stationary system in a large enough domain. However, in what follows we want to treat a numerical experiment in a finite domain with finite $Re$ and then ensemble averaging is the most natural choice as will be come clear below. } This averaging process is denoted by an overbar, $\overline{(\,\cdot\,)}$,
and then the flow can be decomposed into a mean $U(y,t) {\bf \hat{x}}:=\overline{{\bf u}({\bf x},t)}$ and fluctuation part ${\bf \tilde{u}}:={\bf u}-\overline{{\bf u}}=\tilde{u} {\bf \hat{x}}+\tilde{v} {\bf \hat{y}} + \tilde{w} {\bf \hat{z}}$ (due to symmetry, a vanishing mean spanwise component is assumed so $( \,\overline{\tilde{v} \tilde{w}} \,)_y=0$).
The Navier-Stokes equations can be similarly decomposed into a mean part,
\begin{equation}
U_t -\frac{1}{Re}U_{yy} = G - ( \,\overline{\tilde{u} \tilde{v}} \,)_y
\label{U}
\end{equation}
and a fluctuation part,
\begin{equation}
{\bf \tilde{u}}_t = \frac{1}{Re} \nabla^2 {\bf \tilde{u}} -\bm{\nabla} \tilde{p} -{U} {\bf \tilde{u}}_x - \tilde{v} U_y {\bf \hat{x}} -({\bf \tilde{u}} \cdot \bm{\nabla} {\bf \tilde{u}} -\overline{{\bf \tilde{u}} \cdot \bm{\nabla} {\bf \tilde{u}}}),
\label{fu}
\end{equation}
which is incompressible
\begin{equation}
\bm{\nabla}\cdot {\bf \tilde{u}} = 0 \label{divu}
\end{equation}
where subscripts denote derivatives (e.g. $U_y:=dU/dy$).
\subsection{\label{II.A}Orr-Sommerfeld stability analysis}
Given a possibly turbulent flow $(U,{\bf \tilde{u}})$, the `standard' linear stability analysis is to consider small (a.k.a. infinitesimal) perturbations $(0, \delta {\bf \tilde{u}})$ to a base state $(U,{\bf 0})$ where the fluctuation field is ignored and the mean flow $U$ is assumed steady. As a result only the equations (\ref{fu}) and (\ref{divu}) need be perturbed (and hence linearised) and since $U=U(y)$, the ensuing eigenvalue calculation is parameterised by a streamwise and spanwise wavenumber. Squire's theorem \cite{Squire33} is usually invoked to focus the search for instability to spanwise-independent perturbations and leads to the celebrated Orr-Sommerfeld equation \citep{Orr1907,Sommerfeld1908}. In primitive variables, $\tilde{w}$ then decouples from $\tilde{u}$ and $\tilde{v}$ and can be ignored, leaving the reduced eigenproblem
\begin{align}
-im \alpha c\, \tilde{u} & = \frac{1}{Re}(\tilde{u}_{yy}-m^2\alpha^2 \tilde{u})-im \alpha \tilde{p} -im \alpha U\tilde{u}-U_y \tilde{v}, \label{(7)}\\
-im \alpha c \,\tilde{v} & = \frac{1}{Re}(\tilde{v}_{yy}-m^2 \alpha^2\tilde{v})-\tilde{p}_y -im \alpha U\tilde{v}, \label{(8)}\\
0 & = im \alpha \, \tilde{u}+\tilde{v}_y \label{(9)}
\end{align}
where $(\tilde{u},\tilde{v},\tilde{p}) \propto e^{im \alpha (x-ct)}$ and $c:=c_r+ic_i$ is the (complex) eigenvalue. The Orr-Sommerfeld equation is then reached by defining a streamfunction $\psi$ (so $\tilde{u}=\psi_y$ and $\tilde{v}=-im \alpha \psi$) and eliminating the pressure so that
\begin{equation}
(U-c)(\partial_y^2-m^2 \alpha^2)\psi-U_{yy} \psi=\frac{1}{i m \alpha Re} (\partial_y^2-m^2 \alpha^2)^2 \psi.
\end{equation}
In what follows, we actually work with the primitive variable formulation (\ref{(7)})-(\ref{(9)}) as it is clearer to interpret the origin of new terms added below, the eigenvalue problem is better conditioned (2 equations with second order operators as opposed to one with a fourth order operator) {\em and} it is easily extended to 3D if needed (for 3D disturbances, the Orr-Sommerfeld equation must be augmented with Squire's equation for the cross-stream vorticity).
Nevertheless, we refer to this general approach of doing linear stability around a {turbulent} mean as `Orr-Sommerfeld analysis' - or `OS analysis' for short - in recognition of its conception in Malkus's work.
The (matrix) size of the eigenproblem (\ref{(7)})-(\ref{(9)}) is only $ 3N_y \times 3N_y$ for each streamwise wavenumber $m$ and so needs to be repeated $N_x$ times ($N_x,N_y$ represent the streamwise and wall-normal truncations). When the mean profile is known from simulations or experiments, for typical resolutions, this is an easily accessible procedure which, through the structure of the most unstable eigenvectors, can shed some light on the dominant structures of the turbulent flow.
A popular extension to the standard OS analysis involves using an eddy viscosity $E(y)$ instead of the molecular viscosity. An eddy viscosity can be determined self-consistently from the Reynolds stress needed to sustain the turbulent mean profile (e.g. \cite{ReynoldsTiederman67}) using the steady version of (\ref{U}):
\begin{equation}
\frac{1}{Re} [(\,1+E(y)\,)U_y]_y:=\frac{1}{Re}U_{yy}- ( \,\overline{\tilde{u} \tilde{v}} \,)_y=-G
\, \Rightarrow \,
E(y) := -Re \frac{ \overline{ \tilde{u} \tilde{v}} }{U_y}
=
-\frac{Re}{U_y} \int^y_0 G(\bar{y})d\bar{y}-1
\label{eq:eddy_viscosity}
\end{equation}
Then, the eigenproblem can be modified to:
\begin{align}
-im \alpha c\, \tilde{u} & = \frac{1}{Re} \left( \left[(1+E)\tilde{u}_{y} \right]_y-m^2\alpha^2 (1+E) \tilde{u} \right)-im \alpha \tilde{p} -im \alpha U\tilde{u}-U_y \tilde{v}, \\
-im \alpha c \,\tilde{v} & = \frac{1}{Re}\left( \left[(1+E)\tilde{v}_{y} \right]_y-m^2\alpha^2 (1+E) \tilde{v} \right)-\tilde{p}_y -im \alpha U\tilde{v}, \\
0 & = im \alpha \, \tilde{u}+\tilde{v}_y.
\end{align}
We will use this to assess the performance of our extended Orr-Sommerfeld analysis in \S\ref{IV.C}.
\subsection{\label{II.B}Statistics: cumulants}
In this section we consider a statistical framework for the flow by working with the equal-time cumulants of the flow \citep{Hopf52,Orszag73, Frisch95}. Even within this framework, we specialise further to exclusively consider equal-$x$-and-$z$ cumulants which are the subset of cumulants which influence the mean flow. (In fact, only equal-$y$ cumulants are needed in the mean flow equation - see (\ref{C-U}) below. However, the evolution equations for these are not available without also solving for `non-equal' $y$ cumulants.) The first cumulant is the mean $U(y,t)$. The second cumulant is the symmetric matrix
\begin{equation}
{\bf C}(y_1,y_2,t):= \overline{{\bf \tilde{u}}(x,y_1,z,t) \otimes {\bf \tilde{u}}(x,y_2,z,t)}
=\left(\begin{array}{ccc}
C_{11} & C_{12} & C_{13} \\
C_{21} & C_{22} & C_{23} \\
C_{31} & C_{32} & C_{33}
\end{array} \right),
\end{equation}
where we introduce the notation $C_{ij}(1,2):= \overline{[{\bf \tilde{u}}(x,y_1,z,t)]_i [{\bf \tilde{u}}(x,y_2,z,t)]_j}\quad$ (here $[{\bf \tilde{u}}]_1=\tilde{u}$, $[{\bf \tilde{u}}]_2=\tilde{v}$ and $[{\bf \tilde{u}}]_3=\tilde{w}$) to emphasize the $y$ arguments and de-emphasize the implicit time dependence, and the third cumulant is the third order tensor
\begin{equation}
C_{ijk}^{(3)}(1,2,3):=\overline{[{\bf \tilde{u}}(x,y_1,z,t)]_i [{\bf \tilde{u}}(x,y_2,z,t)]_j [{\bf \tilde{u}}(x,y_3,z,t)]_k}.
\end{equation}
These correspond to the second and third central moments of the flow respectively. Cumulants and central moments, however, diverge at fourth order and beyond, e.g.
\begin{align}
C_{ijkl}^{(4)}(1,2,3,4)&:=
\overline{ [{\bf \tilde{u}}(x,y_1,z,t)]_i [{\bf \tilde{u}}(x,y_2,z,t)]_j [{\bf \tilde{u}}(x,y_3,z,t)]_k [{\bf \tilde{u}}(x,y_4,z,t)]_l
}
\\
& \qquad
-C_{ij}(1,2)C_{kl}(3,4)
-C_{ik}(1,3)C_{jl}(2,4)-C_{il}(1,4)C_{jk}(2,3).
\end{align}
We will not go this high in the cumulant expansion used here but just note that the $n^{th}$ order cumulant is $n$-dimensional in space so that storage when doing computations becomes prohibitive very quickly. Hence the onus is on applying some sort of closure as soon as possible.
To derive evolution equations for the cumulants, we introduce a double Fourier series representation of the flow
\begin{equation}
{\bf \tilde{u}}(x,y,z,t):= \sum_m \sum_n {\bf \tilde{u}}^{mn}(y,t) e^{im \alpha x+in \beta z}
\end{equation}
where $m,n \in {\mathbb Z}$. Clearly ${\bf \tilde{u}}^{-m-n}={\bf \tilde{u}}^{*mn}$ (the complex conjugate of ${\bf \tilde{u}}^{mn}$) for a real flow but it will be clearer not to build this into the notation in anticipation of deriving perturbation equations later. Hence, we write
\begin{align}
C_{ij}(1,2) &= \sum_m \sum_n \biggl\{ C_{ij}^{mn}(1,2):= [{\bf \tilde{u}}^{mn} (y_1,t)]_i [{\bf \tilde{u}}^{-m-n} (y_2,t)]_j \biggr\} \label{C2}\\
C^{(3)}_{ijk}(1,2,3) &= \sum_m \sum_n \biggl\{ C_{ijk}^{(3)mn}(1,2,3)
:=\sum_p \sum_q [{\bf \tilde{u}}^{mn}(y_1,t)]_i [{\bf \tilde{u}}^{pq}(y_2,t)]_j \nonumber\\ & \hspace{6cm} \times [{\bf \tilde{u}}^{-(m+p)-(n+q)}(y_3,t)]_k \biggr\} \label{C3}
\end{align}
(note e.g. $C_{ij}^{mn}(1,2)=C_{ji}^{-m-n}(2,1)$\,).
Equations to evolve the cumulants are obtained by temporally differentiating their definitions in (\ref{C2}) and (\ref{C3}) and using (\ref{fu}). For example, for the second order cumulant,
\begin{align}
\partial_t C_{ij}^{mn}(1,2) & = [{\bf \tilde{u}}^{mn}(y_1,t)]_i \partial_t [{\bf \tilde{u}}^{-m-n}(y_2,t)]_j
+\partial_t [{\bf \tilde{u}}^{mn}(y_1,t)]_i [{\bf \tilde{u}}^{-m-n}(y_2,t)]_j , \nonumber \\
& = \frac{1}{Re} \left(\partial^2_1+\partial^2_2 -2m^2\alpha^2-2n^2\beta^2\right) C_{ij}^{mn}(1,2) \nonumber \\
& \hspace{2cm} -\left[\begin{array}{c}
-im\alpha \\ \partial_2 \\ -in\beta \end{array}\right]_j C_{i4}^{mn}(1,2)
-\left[\begin{array}{c}
im\alpha \\ \partial_1 \\ in\beta \end{array}\right]_i C^{mn}_{4j}(1,2)
\nonumber \\
& \hspace{2cm} +im\alpha [\,U(2)-U(1)\,]C_{ij}^{mn}(1,2)
-U_y(2)C_{i2}^{mn}(1,2) \delta_{1j}
\nonumber \\
& -U_y(1)C_{2j}^{mn}(1,2) \delta_{i1}
-\left[\begin{array}{c}
-im\alpha \\ \partial_2 \\ -in\beta \end{array}\right]_k C_{ijk}^{(3)mn}(1,2,2)
-\left[\begin{array}{c}
im\alpha \\ \partial_1 \\ in\beta \end{array}\right]_k C_{ikj}^{(3)-m-n}(2,1,1)
\label{C2a}
\end{align}
where $\partial_i:= \partial_{y_i}$, $U(i)=U(y_i)$ for $i=1,2$ and $C_{i4}^{mn}(1,2):= [{\bf \tilde{u}}^{mn} (y_1,t)]_i \tilde{p}^{-m-n}(y_2,t)=:C_{4i}^{-m-n}(2,1)$ are 3 extra `velocity-pressure' cumulants that get generated. Incompressibility conditions give the required 3 extra matrix constraints
\begin{equation}
\left[\begin{array}{c}
im \alpha\\ \partial_1 \\ in \beta \end{array}\right]_i C_{ij}^{mn}(1,2)=0 \qquad j \in \{1,2,3\}
\label{C-incompressible}
\end{equation}
along with the equation for the mean equation (\ref{U})
\begin{equation}
U_t -\frac{1}{Re}U_{yy} = G - \sum_m \sum_n \partial_y C_{12}^{mn}(y,y,t)
\label{C-U}
\end{equation}
to close the system. The infamous closure problem of the Navier-Stokes equation is immediately evident here in that the evolution equation for the 2nd order cumulant depends on the 3rd order cumulant, a pattern which continues for higher order cumulants so the system never closes. A popular (lowest) closure - commonly called CE2 - is to simply ignore the 3rd order cumulant which is equivalent to ignoring the fluctuation-fluctuation term (last bracketed term on the rhs of (\ref{fu})\,). This is the quasilinear approximation or sometimes referred to as mean field theory \cite[e.g.][]{Vedenov61,Herring63, Herring64},
\begin{align}
U_t &= \frac{1}{Re}U_{yy}+G - ( \,\overline{\tilde{u} \tilde{v}} \,)_y
\label{QL-U}\\
{\bf \tilde{u}}_t &= \frac{1}{Re} \nabla^2 {\bf \tilde{u}} -\bm{\nabla} \tilde{p} -{U} {\bf \tilde{u}}_x - \tilde{v} U_y {\bf \hat{x}}, \label{QL_fu}\\
0 &=\bm{\nabla}\cdot {\bf \tilde{u}} . \label{QL}
\end{align}
The defining feature of this approximation is that (\ref{QL_fu}) is linear in ${\bf \tilde{u}}$ so that fluctuations with different wavenumbers are only coupled in the mean flow equation (\ref{QL-U}). This linearity also means that any fluctuation field (parametrised by streamwise and spanwise wavenumber) can not be consistently in the stable manifold of $U$ as it varies with time. In particular, if $U$ is steady, only marginally-stable fluctuation fields \rk{(typically with a temporal frequency) can be non-vanishing}. Malkus argued for this model (and its marginal stability implications) on the basis that the fluctuation-fluctuation nonlinear term was only stabilising. This would be reasonable if bifurcations from unidirectional shear flows were always supercritical but, some decades later, subcriticality has been realised the more generic situation \citep{Kerswell05, Eckhardt07,Kawahara12, Graham21}.
Applying a closure at next order so $C^{(4)}$ is some assumed function of the lower cumulants or simply ignored (termed CE3) is less straightforward as the ensuing positive definiteness of the second cumulant is not automatic \citep{MarstonQiTobias14}. This difficulty explains the popularity of the lower-order CE2 approximation \rk{where, for example, the existence and stability of steady solutions has recently been investigated for ODE systems \citep{Li21,Li22}.}
\subsection{\label{II.C}Approximations to statistical stability}
The approach here is to consider the stability within the cumulant framework as this presents a natural way to assess stability of the flow statistics. Ideally, this should be attempted for a cumulant expansion which is high enough order to show a robustness against including even higher order cumulants. However, the rate at which the dimensionality of this procedure explodes means that only second and perhaps third order closures are currently practical. As a result, we focus on CE2 here and identify a clear way to progress to higher order (see \S\ref{III.B})
CE2 is the statistical equations (\ref{C2a})-(\ref{C-U}) with the third order cumulants in (\ref{C2a}) set to zero. The corresponding equations for perturbations $(\delta U, \delta C_{ij}^{mn} )$ upon a base statistical state $(U,C_{ij}^{mn})$ - hereafter referred to as the $\delta$CE2 problem - are
\begin{align}
\partial_t \,\delta C_{ij}^{mn}(1,2)
& = \frac{1}{Re} \left(\partial^2_1+\partial^2_2 -2m^2\alpha^2-2n^2\beta^2\right) \delta C_{ij}^{mn}(1,2) \nonumber \\
&-\left[\begin{array}{c}
-im \alpha\\ \partial_2 \\ -in \beta \end{array}\right]_j \delta C_{i4}^{mn}(1,2)
-\left[\begin{array}{c}
im\alpha \\ \partial_1 \\ in\beta \end{array}\right]_i \delta C^{mn}_{4j}(1,2)
\nonumber \\
&+im \alpha [\,U(2)-U(1)\,] \delta C_{ij}^{mn}(1,2)
-U_y(2) \delta C_{i2}^{mn}(1,2) \delta_{1j}
-U_y(1) \delta C_{2j}^{mn}(1,2) \delta_{i1} \nonumber\\
& +im \alpha [\,\delta U(2)- \delta U(1)\,] C_{ij}^{mn}(1,2)
-\delta U_y(2) C_{i2}^{mn}(1,2) \delta_{1j}
-\delta U_y(1) C_{2j}^{mn}(1,2) \delta_{i1} \label{dC_1}\\
0 &= \left[\begin{array}{c}
im\alpha\\ \partial_1 \\ in \beta \end{array}\right]_i \delta C_{ij}^{mn}(1,2),\label{dC_incompress}\\
\partial_t \delta U &= \frac{1}{Re}\delta U_{yy} +\delta G - \sum_m \sum_n \partial_y \delta C_{12}^{mn}(y,y,t).
\label{C-U1}
\end{align}
An equivalent equation arises in non-modal stability theory \citep[e.g.][]{FarrellIoannou93, Jovanovic05}.
If the pressure gradient is kept fixed $\delta G=0$, otherwise the constant volume flux condition $\int^1_{-1} \delta U dy=0$ is the extra constraint required.
Crucially the ansatz $(\delta U, \delta C_{ij}^{mn}) \propto e^{ \lambda t}$ is possible if the base state is independent of time - in other words the base flow is {\em statistically} steady.
Even in the cheapest 2D situation, the CE2 stability problem requires handling $5 N_x$ correlation matrices of size $N_y^2$ all linked through the mean equation. This leads to a matrix of size $ \left( N_y+5 N_x N_y^2\right)^2$
or $\approx 25N_x^2 N_y^4$ elements which is out of reach even for modest resolutions. Given this, we explore a route to potentially still capture the essence of the CE2 statistical stability approximation without the considerable cost. To do this, we discuss a connection back to the equations of motion which plausibly retains the steadiness of the stability problem.
\subsection{\label{II.D}Steady statistics and simplications}
\rk{A statistically steady base flow has steady mean $U$ and cumulants ${\bf C}$ with their Fourier components ${\bf C}^{mn}(1,2)$ also steady under ensemble averaging. This ensures that the associated stability problem (\ref{dC_1})-(\ref{C-U1}) has temporally-constant coefficients and is therefore a (conceptually at least) simple eigenvalue problem.}
\rk{In what follows below, we will not attempt to solve this directly as it is too unwieldy. Instead, a smaller, more practical QL stability problem is sought as a good proxy for it. The key in doing this is identifying a suitable base velocity field around which to develop a QL-type stability problem. A straightforward approach is to find the `best' rank-1 approximation of each ${\bf C}^{mn}$ and use the associated velocity field, ${\bf \tilde{u}}^{mn}_0(y)$, as representative of the base flow. This can be accomplished by minimising the Frobenius matrix norm of the difference},
%
\begin{align}
||{\bf C}^{mn}-{\bf \tilde{u}}^{mn}_0 \otimes {\bf \tilde{u}}^{*mn}_0||^2_F &:=
\sum_{i,j=1}^3 \sum_{p,q=1}^N
(\,C^{mn}_{i,j}(p,q)-[{\bf \tilde{u}}^{mn}_0(y_p)]_i[{\bf \tilde{u}}^{*mn}_0(y_q)]_j\,) \nonumber \\
&\hspace{1.75cm}\times (\,C^{*mn}_{ij}(p,q)-[{\bf \tilde{u}}^{*mn}_0(y_p)]_i[{\bf \tilde{u}}^{mn}_0(y_q)]_j\,)
\nonumber\\
& \hspace{-0.5cm}=\left\|\left(\begin{array}{ccc}
C^{mn}_{11} & C^{mn}_{12} & C^{mn}_{13} \\
C^{mn}_{21} & C^{mn}_{22} & C^{mn}_{23} \\
C^{mn}_{31} & C^{mn}_{32} & C^{mn}_{33}
\end{array} \right) -
\left[\begin{array}{c} \tilde{u}^{mn}_{(1:N)}\\ \tilde{v}^{mn}_{(1:N)} \\ \tilde{w}^{mn}_{(1:N)}
\end{array} \right]
\left[\begin{array}{ccc} \tilde{u}^{*mn}_{(1:N)} &\tilde{v}^{*mn}_{(1:N)} & \tilde{w}^{*mn}_{(1:N)}
\end{array} \right]
\right\|^2_F
\end{align}
where e.g. $\tilde{u}^{mn}_{(1:N)}:=[ \tilde{u}^{mn}(y_1)\, \,\, \tilde{u}^{mn}(y_2)\,\, \cdots\,\, \tilde{u}^{mn}(y_N)]^T$. Since ${\bf C}^{mn}=[{\bf C}^{mn}]^H$ (the Hermitian conjugate) and positive definite, the required ${\bf \tilde{u}}^{mn}_0$ is the leading right eigenvector of ${\bf C}^{mn}$ associated with the largest eigenvalue $\sigma_1$ scaled so that $|{\bf \tilde{u}}^{mn}_0|=\sqrt{\sigma_1}$.
\rk{We now discuss how the QL stability problem based upon the ${\bf \tilde{u}}^{mn}_0$ relates to the CE2 stability problem based on ${\bf C}^{mn}={\bf \tilde{u}}^{mn}_0 \otimes {\bf \tilde{u}}^{*mn}_0$. Possible disturbances partition into two types: Type A where the disturbance has energy in Fourier pairings not excited in the base flow and Type B where the disturbance has energy in Fourier pairings which are a subset of those present in the base flow i.e. $\bm{\delta} {\bf \tilde{u}}^{mn}$ is only non-zero if ${\bf \tilde{u}}_0^{mn}$ is (see {Appendix \ref{appendixA}} for more detail). The former (Type A) case is straightforward (since $\bm{\delta} {\bf C}={\bf 0}$) so we focus here on the Type B situation. The QL stability problem - hereafter the $\delta$QL problem - is}
\begin{align}
\partial_t \bm{\delta} {\bf \tilde{u}}^{mn} &= {\mathbb L}^{mn}(U) \bm{\delta} {\bf \tilde{u}}^{mn} +{\mathbb J}^{mn}(\delta U) {\bf \tilde{u}}_0^{mn} \qquad \qquad \forall (m,n) \neq (0,0) \label{dQL}\\
\partial_t \delta U &= \frac{1}{Re}\delta U_{yy} +\delta G - \sum_m \sum_n \partial_y
\biggl\{
\tilde{u}_0^{mn} \delta \tilde{v}^{-m-n}
+\delta \tilde{u}^{mn} \tilde{v}_0^{-m-n}
\biggr\}
\end{align}
where
\begin{align}
{\mathbb L}^{mn}(U)\, {\bf \tilde{u}}^{mn} &:= \frac{1}{Re} [\,\partial_y^2-m^2\alpha^2 -n^2 \beta^2\,]{\bf \tilde{u}}^{mn}
-\left[\begin{array}{c} im\alpha \\ \partial_y \\ in\beta \end{array} \right] \tilde{p}^{mn} -im \alpha{U} {\bf \tilde{u}}^{mn} - \tilde{v}^{mn} U_y {\bf \hat{x}}, \label{L}\\
{\mathbb J}^{mn}(\delta U) &:=\lim_{\epsilon \rightarrow 0} \, \frac{1}{\epsilon}[\,{\mathbb L}^{mn}(U+ \epsilon\delta U)-{\mathbb L}^{mn}(U)\,]={\mathbb L}^{mn}(U+\delta U)-{\mathbb L}^{mn}(U) \label{J}
\end{align}
(as ${\mathbb L}^{mn}$ is affine in $U$). This has temporally-constant coefficients and therefore admits an eigenfunction of the form
\begin{equation}
(\delta U(y,t), \, \bm{\delta} {\bf \tilde{u}}(x,y,z,t))
=\biggl( \delta \hat{U}(y) e^{\lambda t}, \,\sum_m \sum_n \biggl[
\bm{\delta} {\bf \tilde{u}}^{mn}(y,t):=\bm{\delta} \hat{{\bf u}}^{mn}(y) e^{\lambda t}
\biggr]e^{i(m \alpha x+ n \beta z)}
\biggr)
\end{equation}
\rk{The key point is that this has a direct equivalent in the CE2 problem (\ref{dC_1})-(\ref{C-U1}) where the same eigenvalue $\lambda$ has the corresponding eigenfunction ${\bf C}^{mn}:= {\bf \tilde{u}}_0^{mn} (y_1,t)\otimes \bm{\delta} {\bf \tilde{u}}^{-m-n}(y_2,t)+ \bm{\delta} {\bf \tilde{u}}^{mn} (y_1,t)\otimes {\bf \tilde{u}}_0^{-m-n}(y_2,t) =\hat{{\bf C}}^{mn}(y_1,y_2)e^{\lambda t}$. Given this, instability within the much-more-tractable $\delta$QL problem is therefore sufficient to conclude statistical instability within $\delta$CE2}. Going further to claim that the stability within the $\delta$QL problem also implies stability within the $\delta$CE2 problem seems very likely but is not assured (see Appendix \ref{appendixA}). With this one caveat, we nevertheless assume that the stability of the (much) smaller QL system is a proxy for the statistical stability of the CE2 system: in particular, stability in the $\delta$QL system is taken to imply statistical stability within $\delta$CE2. This allows us to generate an extended version of Orr-Sommerfeld analysis but, before pursuing this in the next section, we make a remark and an observation.
The remark is that certain notational liberties have been taken here to keep the discussion as clear as possible. For example, the operator ${\mathbb L}^{mn}$ strictly maps an incompressible flow field to another which involves a supplementary scalar field (the pressure) entering into the definition. As is well known, this is determined by imposing incompressibility. An implicitly-incompressible representation for the velocity field could be used (e.g. reducing the problem down to just using wall-normal velocity and vorticity) to avoid this wrinkle but staying with primitive variables makes the various manipulations as clear as possible.
The observation is that the special form of the nonlinearity in the QL and CE2 formulations means that different wavenumber pairings only interact through the mean flow equation. As a result eigenvalues can be sought within any subset of the wavenumbers possible and these are still valid within the full system of wavenumbers i.e. this is not a truncation merely a subclass of disturbances. In particular, for CE2 the simplest perturbation of the base state $(U,{\bf C})$ just consists of perturbations in one wavenumber pairing and the mean flow,
\begin{equation}
(\,\delta \hat{U}, \,\,\bm{\delta} \hat{{\bf C}}^{mn}\,) \,e^{\lambda t}.
\end{equation}
Even this leads to an eigenvalue matrix calculation of size $(N_y+9N_y^2) \times (N_y+9N_y^2)$ in 3D. The equivalent QL perturbation
has to consider
\begin{equation}
(\,\delta \hat{U}, \,\,\bm{\delta} {\bf \hat{u}}^{mn},\,\,\bm{\delta} {\bf \hat{u}}^{-m-n}\,) \,e^{\lambda t}
\end{equation}
and requires an eigenvalue matrix calculation of size $(N_y+8N_y) \times (N_y+8N_y)$ which is {\it much} smaller (or $(N_y+6N_y)\times (N_y+6N_y)$ in 2D - see Appendix \ref{appendixB}).
\section{\label{III}Extended Orr-Sommerfeld analysis}
We refer to the $\delta$QL problem which includes some information on the second-order-flow statistics of the base flow as `Extended Orr-Sommerfeld' stability analysis. The approach is as follows.
\begin{enumerate}
\item {\em Estimate the base flow statistics} $(U, {\bf C})$
\item {\em Approximate each base cumulant component tensor ${\bf C}^{mn}$ as rank 1 to obtain a representative physical base field ${\bf \tilde{u}}^{mn}_0(y)$.}
\item {\em Solve the Extended Orr-Sommerfeld eigenvalue problem for small perturbations $(\delta U, \,\bm{\delta} {\bf \tilde{u}}, \,\delta \tilde{p})$ which is}
\begin{equation}
\begin{aligned}
\partial_t\delta U&=\frac{1}{Re}\partial^2_y \delta U+\delta G - \sum_{m,n}
\partial_y \left(\delta \tilde{u}^{mn} \tilde{v}^{-m-n}_0 +\tilde{u}_0^{mn}\delta \tilde{v}^{-m-n} \right)
\end{aligned}
\label{eq:EOS_mean}
\end{equation}
coupled with the fluctuation equation for every Fourier mode $(m,n)$
\begin{gather}
\partial_t{\bm{\delta} {\bf \tilde{u}}}^{mn} = \frac{1}{Re} \nabla^2 \bm{\delta} {\bf \tilde{u}}^{mn}-\bm{\nabla} \delta \tilde{p}^{mn} - im U \bm{\delta} {\bf \tilde{u}}^{mn} - \delta \tilde{v}^{mn}U_y\bm{\hat{x}}
\textcolor{blue}{-im{\delta U} {\bf \tilde{u}}_0^{mn} }
\textcolor{red}{-\tilde{v}_0^{mn} \delta U_y \bm{\hat{x}}}, \label{eq:EOS_fluct} \\
im \alpha\, \delta\tilde{u}^{mn}+\partial_y\, \delta \tilde{v}^{mn}+in\beta\, \delta \tilde{w}^{mn} = 0
\label{eq:EOS_incomp}
\end{gather}
(coloured terms depend on $\delta U$ to be discussed later).
\end{enumerate}
This eigenvalue problem has size $ (N_y+4 N_x N_y N_z)^2 \approx 16N_x^2N_y^2N_z^2$ which is probably impractical for all but the smallest systems since all the wavenumber pairings are coupled through the mean equation. A natural way to simplify the calculation is to only include a targeted subset of the wavenumber pairings and, going further, to only consider one wavenumber pairing. This latter approximation is obviously the most extreme but is also closest in spirit to the original Orr-Sommerfeld analysis - we call this {\em `minimally Extended Orr-Sommerfeld'} (mEOS) analysis.
\subsection{\label{III.A}minimally Extended Orr-Sommerfeld equations (mEOS)}
Here, the sum over wavenumber pairings in (\ref{eq:EOS_mean}) is removed leaving the mean flow equation (\ref{eq:EOS_mean})
coupled with just one Fourier mode equation (\ref{eq:EOS_fluct}).
This eigenvalue problem focusses on the coupling between an individual Fourier mode and the mean flow and its size is now $ (5 \times N_y)^2 = 25 N_y^2$ so comparable to OS analysis. Just as for OS analysis, it needs to be repeated for each Fourier wavenumber of interest.
\subsection{\label{III.B}infinitely-Extended Orr-Sommerfeld equations (iEOS)}
Before going on to test EOS and mEOS on some statistically steady flows, it's worth briefly discussing how far this approach of using the cumulant framework to further motivate more sophisticated versions of Orr-Sommerfeld analysis can go. In fact, it turns out that only one further enhancement is possible and then only for CE$\infty$. This is because any intermediate closure will lead to an equation which when `unwrapped' at the highest level contradicts those obtained at lower cumulant orders.
To see this, consider CE3, the next closure after CE2 in which ${\bf C}^{(4)}$ is ignored. Unwrapping the statistical equation for ${\bf C}$ will recover the Navier-Stokes equations whereas unravelling the ${\bf C}^{(3)}$ equations will not as the nonlinear term leading to ${\bf C}^{(4)}$ has been dropped (any other closure will also suffer this inconsistency). The only way to avoid this is: 1) to only unwrap one cumulant equation - so CE2 which leads to EOS, or 2) ensure that all unwrapped equations are the same which means no closure and CE$\infty$. In this latter case, the Navier-Stokes equations re-emerge which, under perturbation around a steady base state, lead to the linearized Navier-Stokes equations as the ultimate extension of OS analysis, i.e.
\begin{equation}
\begin{aligned}
\partial_t\delta U&=\frac{1}{Re}\partial^2_y \delta U+\delta G - \sum_{m,n}
\partial_y \left(\delta \tilde{u}^{mn} \tilde{v}^{-m-n}_0 +\tilde{u}_0^{mn}\delta \tilde{v}^{-m-n} \right)
\end{aligned}
\label{eq:iEOS_mean}
\end{equation}
and for each wavenumber pair $(m,n)$
\begin{gather}
\partial_t{\bm{\delta} {\bf \tilde{u}}}^{mn} = \frac{1}{Re} \nabla^2 \bm{\delta} {\bf \tilde{u}}^{mn}-\bm{\nabla} \delta \tilde{p}^{mn} - im \alpha U \bm{\delta} {\bf \tilde{u}}^{mn} - \delta \tilde{v}^{mn}U_y\bm{\hat{x}}
-im \alpha {\delta U} {\bf \tilde{u}}_0^{mn}
-\tilde{v}_0^{mn} \delta U_y \bm{\hat{x}} \nonumber \\
\hspace{8cm}-[\bm{\delta} {\bf \tilde{u}} \cdot \bm{\nabla} {\bf \tilde{u}}_0+{\bf \tilde{u}}_0 \cdot \bm{\nabla} \bm{\delta} {\bf \tilde{u}}]^{mn} \nonumber\\
im \alpha \, \delta \tilde{u}^{mn}+\partial_y \delta \tilde{v}^{mn}+in \beta \, \delta \tilde{w}^{mn} = 0
\label{eq:iEOS_fluct}
\end{gather}
This could be called an {\em `infinitely-Extended Orr-Sommerfeld'} problem (or iEOS for short) as there is no further extension possible with the cumulant expansion framework considered here. It's worth remarking that the extra term added in iEOS would get dropped in any minimal version where the individual wavenumber pairings are considered separately: i.e. minimalizing iEOS is just mEOS. There is, however, considerable potential in avoiding this drastic reduction in favour of retaining small subsets of interacting wavenumber pairings such as triads which interact through the newly present term. This or iEOS will not be pursued further here to avoid overcomplicating the discussion.
\begin{figure}
\centerline{
\includegraphics[width=0.5\textwidth]{fig2a.png}%
\includegraphics[width=0.5\textwidth]{fig2b.png}}%
\caption{\label{fig:meanprofs} Mean velocity profiles (solid lines) and streamwise root-mean-squared velocity profiles (dashed lines) for the four 2D channel test states at $Re=36,300$. Left: states with an applied pressure gradient: symmetric S (light blue) and asymmetric A (pink). Right: body forced states F1 (blue) and F2 (red). Black solid line shows the laminar parabolic profile in both plots for reference.}
\end{figure}
\section{\label{IV}Application}
We now test the extended stability approaches, EOS and mEOS, upon turbulent states obtained by direct numerical simulations of 2D channel flow.
At high enough $Re$, the expectation is that the flow will be statistically stationary \citep{Malkus56}.
The simulations are performed using the open-source partial differential equation solver Dedalus \citep{dedalus}. The 2D flow is simulated using a vorticity-streamfunction formulation so that $(u,v):=(\psi_y,-\psi_x)$ and
\begin{equation}
\omega_t + \psi_y \omega_x - \psi_x \omega_y = \frac{1}{Re} \nabla^2 \omega - \partial_y f(y,t), \quad
\omega:=- \nabla^2 \psi
\label{eq:vorticity2D}
\end{equation}
where the flow is driven by a streamwise body force $f(y,t){\bf \hat{x}}$ so that the volume flux is fixed. $Re$ is defined as in \S \ref{II.A} so the following conditions are imposed
\begin{equation}
\psi\left(-1\right)=0, \quad \psi\left(1\right)=\tfrac{4}{3}, \quad
\psi_y\left(-1\right)=0, \quad \psi_y\left(1\right)=0.
\label{flux}
\end{equation}
with the latter two reflecting the presence of non-slip walls.
Following earlier work \citep{Falkovich_2018}, the length of the channel $2L_x$ is set to 4 times its height ($L_x=4$), $Re=36,300$ and computationally, $1024$ Fourier modes are used to discretize in the $x$ direction and $256$ Chebyshev modes in $y$ (see \cite{MK21} for details).
The streamwise body force is defined in terms of a profile function $g(y)$ as follows
\begin{equation}
f(y,t):=G(t)\left(g(y)-1\right)
\end{equation}
where setting $g=0$ recovers the usual $y$-independent applied pressure gradient $-G(t)$. For this situation it is already known that there is bistability with two statistically steady states possible in 2D channel flow at $Re=36,300$ \citep{MK21}: a state which is statistically symmetric about the channel midplane - the symmetric state S - and another which is statistically asymmetric - the asymmetric state A: see Fig. \ref{fig:meanprofs}(left). Choosing non-zero $g(y)$ is a device to diversify the test states available with two extra examples generated by using profiles $g_1:=(1-y^2)^6$ (state F1) and $g_2:=\cos \tfrac{3}{2} \pi y^2$ (state F2).
\begin{figure}
\centerline{
\includegraphics[width=0.5\textwidth]{fig3a.png}%
\includegraphics[width=0.5\textwidth]{fig3b.png}}%
\caption{\label{fig:energydistr} Fluctuation energy as a ratio to the total flow energy for the four 2D channel test states at $Re=36,300$: symmetric S (light blue), asymmetric A (pink), F1 (blue) and F2 (red). Left: decomposition by streamwise wavenumber $k$, right: temporal fluctuations over a typical time-averaging window. }
\end{figure}
\subsection{Statistically stable test states}
\rk{A spatial average over the streamwise direction and temporal averaging was used as a proxy for the ensemble-averaged statistics discussed in deriving EOS (these are equivalent for a statistically steady flow over a large enough domain and long enough time).}
The mean profiles for states S, A, F1 and F2 are shown in Fig. \ref{fig:meanprofs} along with their streamwise root-mean-squared velocity profiles, all obtained by time-averaging over $10^4$ time units.
Their respective power spectra are shown in Fig. \ref{fig:energydistr} (left). While the dominant wavenumber is the same for all four states ($k_d=2\alpha=\pi/2$), significant differences are seen at neighbouring streamwise wavenumbers $2k_d$ and $3k_d$ for example. Typical temporal variations of the fluctuation field energy $E_{fluct}$ compared to the total energy of the flow $E$ - Fig. \ref{fig:energydistr} (right) - show the desired wide variety of mean fluctuation energy and fluctuation amplitudes. In particular, the amplitude of the fluctuations in state F2 are $\approx 60\%$ of the mean and only $\approx 30\%$ for state A.
For each state the second-order cumulant matrices were computed every unit of time and then averaged over a period of $10^3$. During the process of calculating EOS and mEOS eigenvalues, correlation matrices time-averaged over different time windows t = 500, 1000, 1500 were used as well as correlation matrices time-averaged with different time-steps (0.5 and 1) for one time window t=500. The qualitative results were all the same, with only minor quantitative differences which did not affect the relative positions of eigenvalues obtained by different methods. The time-averaged cumulant matrix was then diagonalised and the flow field corresponding to the leading (real) eigenvalue used to typify the base fluctuation field. The implications of the time-averaging to the rank of the matrix are discussed in \S\ref{V.A}.
\begin{figure}
\centerline{\includegraphics[width=0.5\textwidth]{fig4a.png}
\includegraphics[width=0.5\textwidth]{fig4b.png}}%
\caption{\label{fig:eval_convergence} Typical eigenvalue convergence with respect to the total number of the streamwise wavenumbers $N_x$ included in the EOS model, shown for $k=2$ (left) and $k=3$ (right) for the F2 state.
}
\end{figure}
The EOS problem - equations (\ref{eq:EOS_mean})-(\ref{eq:EOS_fluct}) - and the mEOS problem were posed as generalised matrix eigenvalue problems of matrix size $\approx \left( 4 N_x N_y\right)^2$ and $(5N_y)^2$ respectively.
For EOS, $N_x=20$ streamwise wavenumbers were used as a balance between accurate eigenvalue convergence and computational accessibility: e.g. see Fig. \ref{fig:eval_convergence}. Wall-normal resolution of the eigenvalue problem was also varied to ensure eigenvalue convergence in the mEOS case. Doubling the wall-normal resolution from the DNS resolution of $256$ Chebyshev modes to $512$ produced less that a $10^{-8}$ relative error in the eigenvalues so all stability calculations were done using the DNS wall-normal resolution.
\subsection{\label{IV.C}Comparisons}
Here we compare the eigenvalues obtained by standard (OS), extended (EOS) and minimally extended Orr-Sommerfeld (mEOS) stability analyses on the four different turbulent states. Since all our test cases are {presumed} statistically steady, for a model to be good at predicting statistical stability of the turbulent state, we expect all the eigenvalues to be stable, i.e. to have a negative real part. For the symmetric (top left, figure \ref{fig:EOSvsNOS}) and asymmetric (top right, figure \ref{fig:EOSvsNOS}) states, we observe no significant difference in the leading eigenvalues between the three stability models. While all three models predict the asymmetric state to be statistically stable, they also all predict statistical instability for the symmetric state. For these test states, the leading eigenvalue is associated with the first streamwise wavenumber (marked blue in the plot). We observe some changes for the other eigenvalues which we do not examine further as they are all stable and do not affect characterization of statistical stability of the turbulent states.
Since all streamwise wavenumbers are coupled in the extended model, the leading wavenumber for the eigenvalue is assigned by examining the power spectrum of the corresponding eigenvector. Example power spectra for leading eigenvalues are shown in figure \ref{fig:evect_spectrum} where it is apparent that the mean velocity component of an eigenvector is at least an order of magnitude smaller than the leading wavenumber contribution.
\begin{figure}
\centerline{
\includegraphics[width=0.5\textwidth]{fig5a.png}%
\includegraphics[width=0.5\textwidth]{fig5b.png}}
\centerline{
\includegraphics[width=0.5\textwidth]{fig5c.png}%
\includegraphics[width=0.5\textwidth]{fig5d.png}}%
\caption{\label{fig:EOSvsNOS}
Comparison of the eigenvalues $\sigma=-im\alpha(c_r+i c_i)$ - so growth rate is $m \alpha c_i$ - for OS ($\bullet$), OS with eddy viscosity ($\square$), EOS ($\Diamond$) and mEOS analysis ($\medcirc$). Only the leading eigenvalues for {\color{blue}$m=1$ (blue)}, {\color{BrickRed}$m=2$ (red)} and {\color{OliveGreen}$m=3$ (green)} are shown.
Top left: symmetric S, top right: asymmetric A, bottom left: F1, bottom right: F2 states.
}
\end{figure}
The leading eigenvalue for the body-forced states states F1 (bottom left, figure \ref{fig:EOSvsNOS}) and F2 (bottom right, figure \ref{fig:EOSvsNOS}) is associated with the second streamwise wavenumber (marked red in the plot). For both F1 and F2 states, this eigenvalue is unstable in the standard stability analysis but becomes stable in EOS. The stabilisation effect is also captured by mEOS but isn't so strong. While the statistical stability of the F1 state is predicted correctly by the extended model, the improvement is only partial for the F2 state. In addition to the unstable leading eigenvalue which is stabilised by the extended model, there are two subsequent unstable eigenvalues corresponding to the first and third streamwise wavenumbers (marked blue and green on the plot). Similar to the symmetric state case, these two unstable eigenvalues see no improvement towards stability when the extended stability models are applied.
We also repeat the standard OS analysis with eddy viscosity as described in section \ref{II.A} (marked using squares in figure \ref{fig:EOSvsNOS}). While eddy viscosity has a general stabilising effect on the OS eigenvalues, this effect is rarely enough to stabilise the unstable OS eigenvalues. Significant improvement is only seen for the symmetric S state (top left plot) where the unstable eigenvalue corresponding to $m=1$ is stabilised, while the unstable eigenvalue corresponding to $m=3$ for the F2 state (bottom right plot) moves to just $\mathcal{O} (10^{-4})$ under the instability line. However, the most unstable eigenvalues corresponding to $m=2$ for the states F1 and F2 where significant improvement is seen using the extended OS analyses, are not notably affected by the eddy viscosity.
Note, to avoid the singularity in the expression for eddy viscosity (\ref{eq:eddy_viscosity}) for the asymmetric state, the mean velocity profile $U$ was symmetrised by $\frac{1}{2}\left(U(y)+U(-y)\right)$ and the Reynolds stress $\overline{ \tilde{u} \tilde{v}}$ was anti-symmetrised by $\frac{1}{2}\left( \overline{ \tilde{u} \tilde{v}}(y) - \overline{ \tilde{u} \tilde{v}}(-y) \right)$.
\begin{figure}
\centerline{\includegraphics[width=0.5\textwidth]{fig6a.png}
\includegraphics[width=0.5\textwidth]{fig6b.png}}
\caption{
\label{fig:evect_spectrum}
Power spectra for the EOS eigenvectors for the leading eigenvalues. Left: symmetric S state (\mbox{\boldmath $\times$}) and asymmetric A state ({\color{red}$\bullet$}) with $m=1$;
Right: states F1 ({\color{cyan}\mbox{$\blacksquare$}}) and F2 ({\color{blue}\mbox{\boldmath $+$}}) with $m=2$. The mean velocity perturbation ($m=0$) is at least an order of magnitude smaller than the leading wavenumber contribution to the power spectrum in all 4 cases.
}
\end{figure}
So, in summary, for each unstable eigenvalue observed in our test cases, the standard and extended models either agree on their statistical stability prediction, or the extended model shifts the prediction towards statistical stability. The stabilising effect is qualitatively captured by both EOS and mEOS with the mean velocity components making a non-zero power contribution to the eigenvectors of the extended model. We now examine why the extended stability models sometimes perform better than OS and sometimes make very little improvement.
\begin{table}
\begin{center}
\begin{tabular}{@{}crccr@{}}
\quad & \hspace{1.5cm} &\hspace{1cm} & \quad & \hspace{1.5cm} \\ \hline \hline
& & & & \\
\multicolumn{2}{c}{F2 state: leading eigenvalue for $m=2$} & &
\multicolumn{2}{c}{S state: leading eigenvalue for $m=1$} \\
& & & &\\ \cline{1-2} \cline{4-5}
& & & &\\
$\sigma_{OS}$ & 0.06058 + 0.8888i & & $\sigma_{OS}$ & 0.0024849 + 0.25601i\\
$\sigma_{OS}+\textcolor{blue}{\Delta \sigma_A}+\textcolor{red}{\Delta \sigma_S}$ & -0.16937+ 0.80087i
& &
$\sigma_{OS}+\textcolor{blue}{\Delta \sigma_A}+\textcolor{red}{\Delta \sigma_S}$ & 0.0029878 + 0.25637i \\
$\sigma_{mEOS}$ & -0.02294 + 0.7043i
& &
$\sigma_{mEOS}$ & 0.0028772 + 0.25653i \\
& & & & \\
$\textcolor{blue}{\Delta \sigma_A}$ & \textcolor{blue}{0.02738+0.05860i}
& &
$\textcolor{blue}{\Delta \sigma_A}$ & \textcolor{blue}{0.00011499-0.0000561i}\\
$\textcolor{red}{\Delta \sigma_S}$ & \textcolor{red}{-0.25733-0.02933i} & &
$\textcolor{red}{\Delta \sigma_S}$ & \textcolor{red}{0.00038787+0.0004194i}\\
& & & \\ \hline \hline
\end{tabular}
\end{center}
\caption{Perturbation analysis results for the leading eigenvalue in the F2 state (left) and the symmetric S state (right). The first order change in the eigenvalue caused by the advection ({\color{blue}$\Delta \sigma_A$}) and shear ({\color{red}$\Delta \sigma_S$}) are shown in the minimal extended Orr-Sommerfeld model (mEOS). Unperturbed (standard Orr-Sommerfeld) eigenvalues are given by $\sigma_{OS}$ and mEOS eigenvalues are given by $\sigma_{mEOS}$. The Table shows that F2 changes to being stable when using mEOS whereas S remains unstable. }
\label{tb:perturbationanalysis}
\end{table}
\begin{figure}
\centerline{
\includegraphics[width=0.8\textwidth]{fig7.png}}
\caption{\label{fig:perturbation_analysis_F2} \vm{Perturbation analysis results for F2 state. Top row: streamwise component of the unperturbed left eigenvector $\delta u_{sL}$, streamwise component of the perturbed advection term $-im\alpha u_{0s}^{m0} \delta U_{sR}$ and streamwise component of the perturbed shear term $-v_{0s}^{m0} \partial_y \delta U_{sR}$.
Bottom row: wall-normal component of the unperturbed left eigenvector $\delta v_{sL}$, wall-normal component of the perturbed advection term $-im\alpha v_{0s}^{m0} \delta U_R$ and mean profile component of the unperturbed right eigenvector $-\delta U_{sR}$. Solid (dashed) lines indicate real (imaginary) parts of the vectors. Subscript $(\cdot)_s$ refers to the sine components of the eigenfunctions as explained in {Appendix \ref{appendixB}}. }}
\end{figure}
%
%
\begin{figure}
\centerline{
\includegraphics[width=0.8\textwidth]{fig8.png} }
\caption{\label{fig:perturbation_analysis_S} \vm{Perturbation analysis results for S state. Top row: streamwise component of the unperturbed left eigenvector $\delta u_{cL}$, streamwise component of the perturbed advection term $-im\alpha u_{0c}^{m0} \delta U_{cR}$ and streamwise component of the perturbed shear term $-v_{0c}^{m0} \partial_y \delta U_{cR}$.
Bottom row: wall-normal component of the unperturbed left eigenvector $\delta v_{cL}$, wall-normal component of the perturbed advection term $-im\alpha v_{0c}^{m0} \delta U_R$ and mean profile component of the unperturbed right eigenvector $-\delta U_{cR}$. Solid (dashed) lines indicate real (imaginary) parts of the vectors. Subscript $(\cdot)_c$ refers to the cosine components of the eigenfunctions as explained in {Appendix \ref{appendixB}}. }}
\end{figure}
\subsection{Eigenvalue perturbation analysis}
{In an attempt to rationalise} the improvement (stabilisation) or not in mEOS of the leading eigenvalues, {we carry out} an eigenvalue perturbation analysis. While this neglects the interactions between different streamwise wavenumbers, it includes the minimal improvement that our models offer: adding the mean velocity perturbation into the stability consideration. This leads to an additional equation which governs the evolution of the mean velocity perturbation, and two extra terms - advection and shear - which affect the fluctuation perturbations. While formally limited to only small effects, the perturbation analysis provides a framework to study the structures of the mean velocity perturbations, base fluctuations and fluctuation perturbations and how their interactions affect the eigenvalues.
The unperturbed mEOS eigenvalue problem is the OS problem (equation (\ref{eq:EOS_fluct}) for one Fourier pairing without the $\delta U$ coloured terms) {\em and} (\ref{eq:EOS_mean}) (we are imagining a small number $\varepsilon$ inserted in front of the coloured terms and developing a perturbation expansion in this but actually $\varepsilon=1$). The latter is passive: it sets $\delta U$ but there is no feedback to the OS equation. Discretized, it reads
\begin{equation}
A \bm{y_R} = \sigma B \bm{y_R}
\end{equation}
with matrices $A,B$, eigenvalue $\sigma$ and right eigenvector $\bm{y_R}:=[\delta U\,\, \bm{\delta} {\bf \tilde{u}}^{m0}]^T$. The coloured $\delta U$ terms in (\ref{eq:EOS_fluct}) are then treated as perturbations of $A$ which causes $(\sigma,\bm{y_R}) \rightarrow (\sigma+\Delta \sigma, \,\bm{y_R+\Delta y_R})$ so that
\begin{equation}
\left(A + {\color{blue}\delta A_A} +{\color{red}\delta A_S} \right)\left(\bm{y_R} + \bm{\delta y_R } \right) = \left( \sigma +\delta \sigma \right) B \left( \bm{y_R} +\bm{\delta y_R} \right)
\end{equation}
where ${\color{blue}\delta A_A\bm{y_R}}$ corresponds to the advection term ${\textcolor{blue}{-im\alpha {\bf \tilde{u}}_0^{m0}\delta U_R }}$ and ${\color{red}\delta A_S \bm{y_R}}$
to the shear term $\textcolor{red}{-\tilde{v}_0^{m0} \partial_y \delta U_R \bm{\hat{x}}}$.
By taking the inner product using the corresponding left eigenvector of the unperturbed system $\bm{y_L}$ (${\bm y_L}^{\dag} A=\sigma {\bm y_L}^{\dag} B$), and considering first order terms only, an expression then follows for the first order perturbation to the eigenvalue $\delta \sigma$:
\begin{equation}
{\delta \sigma} =
\textcolor{blue}{\delta \sigma_A} +
\textcolor{red}{\delta \sigma_S}:=
\frac{\bm{y_L}^{\dagger}\textcolor{blue}{\delta A_A\bm{y_R}}}{\bm{y_L}^{\dagger}B\bm{y_R}}+\frac{\bm{y_L}^{\dagger}\textcolor{red}{\delta A_S\bm{y_R}}}{\bm{y_L}^{\dagger}B\bm{y_R}}.
\end{equation}
We present the two most interesting cases of the eigenvalue perturbation analysis. In the first case, we perform eigenvalue perturbation analysis on the leading eigenvalue of the standard Orr-Sommerfeld analysis for the F2 state, which is seen to stabilise in mEOS and EOS. The second case corresponds to the leading eigenvalue of the symmetric state, which shows no significant difference between the standard and extended approaches. The eigenvalues as well as perturbation analysis predictions are summarised in Table \ref{tb:perturbationanalysis}. The left and right eigenvectors and perturbed terms are shown in figures \ref{fig:perturbation_analysis_F2} and \ref{fig:perturbation_analysis_S}.
For the body-force-driven F2 case, the advection term has a small and positive contribution to the real part of $\delta \sigma$ while the shear term has a large negative contribution to the real part of $\delta \sigma$, as summarised in Table 1.
This latter prediction is only qualitatively correct as the shear term is not a small perturbation giving a much larger predicted decay rate of $-0.169$ than the actual value of $-0.023$ obtained from mEOS. The perturbation analysis, however, indicates that the shear in the perturbed mean field is causing the stabilisation.
Looking in more detail at the fields in figure \ref{fig:perturbation_analysis_F2}, it is clear that the active regions of the left and right eigenvectors coincide. In particular, the shear in $\delta U_R$ coincides with where the left eigenvector is significant and is of opposite phase giving the large stabilising effect. In contrast, the resulting advection term is not only an order of magnitude smaller than the shear term, but also has the same sign contribution as the left eigenvector, yielding a small destabilizing contribution to the eigenvalue.
For the symmetric S case, the perturbation analysis shows both advection and shear terms make small destabilizing contributions to the eigenvalue commensurate with mEOS although there is still an error due to the finiteness of the perturbation. Examining the structure of the left and right eigenvectors shown in figure \ref{fig:perturbation_analysis_S} indicates that they are ill-matched. The left eigenvector is concentrated much closer to the channel walls than the right eigenvector. Moreover, even though the right eigenvector component $\delta U$ has some structure near the wall, the interaction of this component with the perturbed operators moves this structure towards the centre of the wall for the shear term, avoiding any significant interaction with the left eigenvector. The advection term has some overlap with the left eigenvector, but it is also an order of magnitude smaller than the shear term, overall resulting in small destabilizing contributions. It's also worth remarking that even at resolution of $2048$ wall-normal Chebyshev modes, the eigenvectors possess the same small scale structure seen in the figures.
The `take home' message from this section is that the mean velocity perturbation, even if an order magnitude smaller than the other components of the eigenvector - see Figure \ref{fig:evect_spectrum}, can interact with the base flow fluctuation fields in a way that produces a strong stabilising influence.
\section{\label{V}Model limitations}
In this section, we discuss the various approximations made in deriving EOS and mEOS which include time-averaging the statistics, approximating correlation matrices as rank 1 and neglecting higher order statistics. {As already mentioned, the channel flow states considered above were assumed statistically steady. Making this assumption is required to apply Malkus's OS analysis in the first place and so subsequently EOS and mEOS, but is only an approximation. Time-averaging the statistics in a small computational box is a commonly-used approach to get a better estimate of the steady statistics hypothetically generated in an infinite computational box. }
\subsection{\label{V.A}Rank 1 approximation of correlation matrices}
Time-averaging the correlation matrix increases its rank from 1 up to potentially its full dimension - see Fig. \ref{fig:SVD} for singular values $\sigma_i$ of correlation matrices time-averaged over 1000 time units. {This is not necessarily a reflection that the statistics are non-stationary as a fluctuation field with two frequencies will give a rank-2 correlation matrix under averaging.}
For $m=2$ which is the dominant streamwise wave-number in all of the test states, a gap between the leading and the second singular values is larger than an order of magnitude indicating that a rank-1 approximation of the time-averaged correlation matrix may be reasonable. On the other hand, time-averaged correlation matrices for $m \in\{1,3\}$ do not show a significant gap between the leading and second singular values. In this case, it is much more likely that some important fluctuation field features might not be accurately represented by the rank 1 representation of the correlation matrix and, as a result, the approach is less justified.
In an attempt to explore a more accurate representation of
${\bf C}^{mn}=\sum_{i=1} \sigma_i \,{\bf \hat{e}_i} \otimes \, {\bf \hat{e}_i}$ (where
${\bf \hat{e}_i}$ is the $i^{th}$ normalised eigenvector and $\sigma_i$ the largest eigenvalue or singular value of ${\bf C}^{mn}$ as it is Hermitian) beyond just
the rank-1 approximation ${\bf C}^{mn} \approx \sigma_1 \,{\bf \hat{e}_1} \otimes \, {\bf \hat{e}_1}$ (${\bf \tilde{u}}^{mn}_0=\sqrt{\sigma_1} {\bf \hat{e}_1}$), a weighted average of the leading $N$ fields,
\begin{equation}
{\bf \tilde{u}}^{mn}_0=\sum_{i=1}^{N} \frac{\sigma_i}{\sum_{i=1}^{N} \sigma_i} \,\sqrt{\sigma_i} {\bf \hat{e}_i},
\end{equation}
was also considered for the F2 state where both $m=1$ and $m=2$ wavenumbers have a non-rank 1 time-averaged correlation matrix $C^{m0}$. However, including $N=1,2,5$ or even $10$ eigenvectors into the expansion showed no significant difference in the leading eigenvalues (less than $1\%$ difference in the absolute value of the eigenvalue).
Understanding the effect of a rank-1 approximation really requires time-stepping the full statistical equations which is beyond the scope of this work.
\begin{figure}
\centerline{
\includegraphics[width=0.5\textwidth]{fig9a.png}%
\includegraphics[width=0.5\textwidth]{fig9b.png}}
\centerline{\includegraphics[width=0.5\textwidth]{fig9c.png}%
\includegraphics[width=0.5\textwidth]{fig9d.png}}%
\caption{\label{fig:SVD} Leading singular values $\sigma_i$ of the correlation matrices ${\bf C}^{m0}$ time-averaged over 1000 time units. Shown for {\color{blue}$m=1$ (blue)}, {\color{BrickRed}$m=2$ (red)} and {\color{OliveGreen}$m=3$ (green)}.
Top left: symmetric S, top right: asymmetric A, bottom left: F1, bottom right: F2 states. Note the gap after the leading singular value for {\color{BrickRed}$m=2$ (red)} across all states confirming rank 1 approximation of the correlation matrix for the cases when mEOS and EOS analyses show stabilisation.}
\end{figure}
\subsection{Time-averaged statistics}
To understand the effect of time-averaging the statistics in the Extended Orr-Sommerfeld stability models, numerical experiments were performed in which the perturbations to the turbulent base flow were exposed to a time-varying base flow.
This experiment was first performed for the symmetric (S) state where the unstable Orr-Sommerfeld eigenvalue remains unstable in EOS and mEOS.
From the DNS data, 10 turbulent flow histories were taken of length $\Delta t=500$ separated by at least 1000 time units to avoid statistical dependence. Each was taken as a turbulent, time-dependent base flow.
Two random perturbations with energy $E_{pert} = 10^{-5} E_{base}$ of the base flow were then added to each of the 10 base flow histories (making a total of 20 cases)
and then their evolution computed over a $\Delta t=500$ time window using the QL equations. For $(\delta U, \bm{\delta}{\bf \tilde{u}}^{m0})$, these read
\begin{align}
\partial_t\delta U&=\frac{1}{Re}\partial^2_y \delta U+\delta G -[\,{\bf \tilde{u}}_0(t).\bm{\nabla} \bm{\delta} {\bf \tilde{u}}+ \bm{\delta}{\bf \tilde{u}}.\bm{\nabla} {\bf \tilde{u}}_0(t)\,]^{00} \\
\partial_t{\bm{\delta} {\bf \tilde{u}}}^{m0} &= \frac{1}{Re} \nabla^2 \bm{\delta} {\bf \tilde{u}}^{m0}-\bm{\nabla} \delta \tilde{p}^{m0} - im U \bm{\delta} {\bf \tilde{u}}^{m0} - \delta \tilde{v}^{m0}U_y\bm{\hat{x}}
-im{\delta U} {\bf \tilde{u}}_0^{m0} -\tilde{v}_0^{m0} \delta U_y \bm{\hat{x}}, \quad m>0
\label{QL}
\end{align}
so that the perturbation-base flow fluctuation interactions are not included in the perturbation fluctuation equation (compare with (\ref{linearNS}) below where they are retained).
Averaging the perturbation energy growth over the 20 different simulations revealed an unstable mode with exponential growth after initial transients decayed. The growth rate of this unstable mode was found to be $\sigma=0.00263$ which is within $5\%$ of the growth rate predicted by the Extended Orr-Sommerfeld analysis - see Fig. \ref{fig:QLvsDNS}. The most unstable eigenvectors of the EOS and mEOS analyses agree and are observed as the growing structure in this QL experiment - see Fig. \ref{fig:11}. We conclude that time dependence of the base flow is not enough to recover the statistical stability of the symmetric state S.
\begin{figure}
\centerline{
\includegraphics[width=0.8\textwidth]{fig10a.png} }
\centerline{
\includegraphics[width=0.8\textwidth]{fig10b.png}}
\caption{\label{fig:QLvsDNS} Results of various numerical experiments: time evolution of the perturbation energy $E_{pert}$ as a ratio to the base flow energy $E_{base}$ averaged over multiple DNS runs. Perturbations were time-stepped using linearised quasilinear (blue), linearised Navier-Stokes (orange) or full non-linear Navier-Stokes (green) equations. For each case, base flow was time-dependant and generated using full Navier-Stokes direct numerical simulations. Results shown for the symmetric S (top) and body-force driven BF2 (bottom) states. Black dashed lines are provided as guides to emphasise the zero gradient of the saturated states and to indicate the growth rate $re(\sigma)$ of the unstable states. For comparison, the real part of the leading mEOS eigenvalue $re(\sigma_{mEOS})$ is given for each of the states.}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[width=0.33\textwidth]{fig11a.png}
\includegraphics[width=0.33\textwidth]{fig11b.png}
\includegraphics[width=0.33\textwidth]{fig11c.png} }
\caption{\label{fig:11} Comparison of the leading eigenvectors from the EOS (left), mEOS $m=1$ (middle) analyses and an instantaneous snapshot of the velocity perturbation field in the QL experiment at $t=100$ (right) for the symmetric (S) state. Only the streamwise velocity component is shown. There is a good correspondence in the leading eigenvector across EOS and mEOS analyses and the QL experiment using a time-dependent base flow. }
\end{figure}
\subsection{Importance of higher order statistics}
To evaluate the need of flow statistics beyond the second order to recover statistical stability of the symmetric (S) state, we repeated the numerical experiment described above but this time the perturbation fields were time-stepped using the linearised Navier-Stokes equations where perturbation-base flow fluctuation interactions are now included and the base flow is time varying:
\begin{align}
\partial_t{\bm{\delta} {\bf \tilde{u}}}^{m0} & = \frac{1}{Re} \nabla^2 \bm{\delta} {\bf \tilde{u}}^{m0}-\bm{\nabla} \delta \tilde{p}^{m0} - im U(t) \bm{\delta} {\bf \tilde{u}}^{m0} - \delta \tilde{v}^{m0}U_y(t)\bm{\hat{x}}
-im{\delta U} {\bf \tilde{u}}_0^{m0}(t)\nonumber \\
& \hspace{2cm}-\tilde{v}_0^{m0}(t) \delta U_y \bm{\hat{x}}
-[\,{\bf \tilde{u}}_0(t).\bm{\nabla} \bm{\delta} {\bf \tilde{u}}+ \bm{\delta}{\bf \tilde{u}}.\bm{\nabla} {\bf \tilde{u}}_0(t)\,]^{m0}.
\label{linearNS}
\end{align}
We repeat the experiment with initial $E_{pert} = 10^{-5} E_{base}$ and $E_{pert} = 10^{-7} E_{base}$ obtaining the same qualitative results. When the energy growth of the perturbation field is time-averaged over 20 different runs, the perturbation growth is seen to saturate after an initial transient growth - see Fig. \ref{fig:QLvsDNS} (top). This suggests that the higher order statistics ignored in QL are important here for recovering statistical stability of the symmetric (S) turbulent state.
\begin{figure}
\centerline{
\includegraphics[width=0.5\textwidth]{fig12a.png}
\includegraphics[width=0.5\textwidth]{fig12b.png}}
\centerline{ \includegraphics[width=0.5\textwidth]{fig12c.png} \includegraphics[width=0.5\textwidth]{fig12d.png} }
\caption{\label{fig:12} Comparison of the leading eigenvectors from the mEOS $m=1$ (top left) and mEOS $m=3$ (top right) analysis with the instantaneous snapshots of the perturbation fields obtained from the QL experiment (bottom left) and linear NS experiment (bottom right) for the F2 state. The structures of the leading mEOS eigenvectors seem unrelated to the growing perturbations in the QL or LNS regimes. }
\end{figure}
\subsection{Non-linear effects}
Extended Orr-Sommerfeld stability analysis applied to the body force F2 state shows a significant stabilisation of the eigenvalue corresponding to the second streamwise wavenumber, but shows no significant difference for the other two unstable eigenvalues of the standard Orr-Sommerfeld approach. Repeating the numerical experiments described above applied to the F2 state shows that both the quasi-linear (QL) equations {\em and} the linearised Navier-Stokes (LNS) equations yield rapid exponential growth of the perturbations applied to the time-dependent base flow: see Figure \ref{fig:QLvsDNS} (bottom). Periodically renormalising the growing perturbation in the latter LNS case and continuing the simulation shows sustained growth over a period of O(1000) time units (not shown). This behaviour is in contrast to what was found for state S where the perturbation saturated (see Figure \ref{fig:QLvsDNS} (top)\,). On the face of it, this experiment seems to show that state F2 is linearly unstable but this does not necessarily follow. Time-stepping a perturbation using the linearised Navier-Stokes equations (or in the tangent space of a turbulent attractor) can continually sample the stretching or contracting dynamics {\em along} the attractor which can dominate the anticipated decaying behaviour perpendicular to the attractor. It seems stretching along the attractor dominates for state F2 but not for state S.
A comparison of the leading growing structures in Figure \ref{fig:12} is consistent with this: the QL and LNS flows bear no similarity with the predictions of mEOS. These LNS calculations for the state F2 highlight perfectly the difficulty in assessing the (linear) statistical stability using the Navier-Stokes equations touched on in the introduction.
\begin{figure}
\centerline{
\includegraphics[width=0.8\textwidth]{fig13.png}}
\caption{\label{fig:13} Time evolution of the relative difference between correlation matrices of the full flow $C^{m0}$ and base flow $C^{m0}_0$ shown for {\color{blue}$m=1$ (blue)}, {\color{BrickRed}$m=2$ (red)} and {\color{OliveGreen}$m=3$ (green)}. The correlation matrices were time-averaged over a rolling time window to display the statistical behaviour. Shown for
the full non-linear NS experiment
where the difference slowly approaches $0$ signifying the recovery of the same statistical state. }
\end{figure}
Repeating the experiment (with $E_{pert} = 10^{-5} E_{base}$) but now time-stepping the perturbation field with the full Navier-Stokes equations, we see the perturbation energy saturate after an initial transient: see Fig. \ref{fig:QLvsDNS} (bottom). Importantly, and as anticipated, the statistics of the turbulent state F2 recover albeit slowly. Figure \ref{fig:13} shows the relative difference between the correlation matrices averaged over a rolling time window between the perturbed and unperturbed F2 state. The statistical differences ebb away after an initial transient of growth. Since F2 is statistically linearly stable, this initial growth indicates that the statistical evolutionary operator linearised around the state F2 fixed point has to be non-normal. Finally, Fig. \ref{fig:14} shows the evolution of the streamwise perturbation velocity field. We see that the highly-localised structure visible at $t=10$ spreads in the channel at $t=20$ and reaches its saturated state at $t=30$ which does not change qualitatively even at much longer times $t=200$. It takes much longer for the statistics to recover suggesting the adjustment along the attractor happens much quicker than normal to it: see Figure \ref{fig:13}.
\begin{figure}
\centerline{
\includegraphics[width=0.5\textwidth]{fig14a.png}
\includegraphics[width=0.5\textwidth]{fig14b.png}}
\centerline{
\includegraphics[width=0.5\textwidth]{fig14c.png}
\includegraphics[width=0.5\textwidth]{fig14d.png}}
\caption{\label{fig:14} Snapshots of the streamwise velocity component of the perturbation field taken from the full non-linear Navier-Stokes experiment with the F2 state. At time $t = 10$ the unstable structure is still growing until time $t = 20$ when it reaches its maximum size and starts to spread out through the channel, at time $t = 30$ the perturbation saturates and shows no qualitative change at a much later time $t = 200$.}
\end{figure}
\section{\label{VI}Discussion}
\subsection{Summary}
We first summarise what has been done in the paper. The motivating objective has been to develop a theoretical approach which can assess the statistical stability of a turbulent state and goes beyond just examining the mean flow as Malkus did in the 1950s \citep{Malkus56}. That such an approach should exist in some form is predicated on the fact that the statistics of a disturbed turbulent flow always return to their undisturbed values if the disturbance is not too large regardless of whether it decays or not. That is, a turbulent flow is assumed {\em linearly} statistically stable (note this doesn't preclude the possibility of the statistical disturbance initially increasing before ultimately decaying away as the linear statistical operator may be non-normal). The way forward is to include more statistical information than just the mean of the flow in the analysis and the simplest such is the 2nd rank cumulant or 2-point equal-time correlations of the flow fluctuations. To do this, we have proposed to examine the statistical stability of the turbulent flow from within a cumulant equation framework truncated to just consider these statistical quantities (called CE2). Even in CE2, however, the ensuing statistical spectral problem is so unwieldy that a substitute has had to be sought. Here the well-known connection between CE2 and the Quasilinear (QL) equations can be exploited and we have argued that a suitably-designed QL spectral problem defined in \S\ref{III} can act as a good proxy to detect statistical instability. The key in designing this QL problem is to translate a statistically-steady base flow in CE2 into a representative {\em steady} base flow in the QL problem with no pretence that this base flow satisfies the QL equations itself {(the QL equations offer an approximation of uncertain accuracy to the Navier-Stokes equations after all).}
The issue of how the statistical spectral problem in CE2 is related to the QL spectral problem is an intriguing one which seems not to have been discussed before. It is important because it underpins the tacit assumption that a CE2 simulation starting from given initial conditions should exactly shadow (statistically) the equivalent QL simulation given one is derivable from the other \citep[e.g.][]{Marston08,Tobias11, Tobias13}. However, this ignores the difference in dimensionality between the two formulations and the possibility of triggering unstable `non-physical' perturbations in the larger CE2 problem (meaning perturbations which cannot exist in the QL setting). In Appendix \ref{appendixA} we have tried to clarify the different types of instability which exist in these formulations - Type A and B - and how they are related. In particular, for Type A instabilities, the spectral QL problem captures the most unstable disturbance possible in the CE2 system and so is an accurate proxy for CE2: CE2 can only be Type A unstable if QL is. For Type B, the situation is less certain. While it remains the case that any Type B instability in QL is mirrored in CE2 and CE2 will likely have further Type B instabilities, it is not clear whether CE2 can have a Type B instability and QL not. Given the as-yet-unreported consequence of this - CE2 could suffer a bifurcation away from QL - it is tempting to assume that this is not possible and then that QL also is a good proxy for Type B perturbations. With this one caveat, probing the QL spectral problem - named here Extended Orr-Sommerfeld analysis (EOS) - then appears the best way to examine the statistical stability of a flow where the 2-pt correlations as well as the mean profile of the flow are incorporated.
Despite improving upon the CE2 spectral problem, EOS is still unfortunately quite costly to implement so a reduced version was proposed in \S\ref{III.A} - minimally Extended Orr-Sommerfeld analysis (mEOS) - which has a similar cost to the usual Orr-Sommerfeld analysis of the flow mean and is the most straightforward extension of doing linear stability analysis only around the mean flow. Just as in OS-analysis one chooses a particular Fourier mode $(m,n)$ to analyse but now extra information about the corresponding base Fourier mode is included. mEOS was found to capture the qualitative effect of EOS with either both producing an improved estimate of statistical stability (e.g. the $m=2$ mode for state F1: see figure \ref{fig:EOSvsNOS} bottom left) or offering no correction of the OS eigenvalue (e.g. the $m=1$ mode for state F1).
Reassuringly, both approaches never made a poorer statistical stability prediction than OS analysis at least over the turbulent channel flow states tested.
\rk{mEOS works comparatively well because the structure of the key eigenmodes in the EOS problem tend to consist of one dominant (primary) fluctuation field interacting with a mean flow perturbation (see figure \ref{fig:evect_spectrum}). This is presumably because the comparatively small size of the cumulants or base fluctuation fields relative to the mean profile (e.g. indicated by the singular values in figure \ref{fig:SVD}) means that only the few fluctuation fields close to marginality on the mean flow play a role. This may well change with increasing $Re$ but the relative size of the cumulants should also increase possibly acting to deprioritise this near-marginal stability property. Clearly higher $Re$ flows warrant further study. }
\noindent
\subsection{Implications}
By incorporating 2-pt fluctuation correlations, EOS and its slimmed down version, mEOS, represent an attempt to generate a better way to assess the statistical stability of a turbulent flow than just considering the mean profile. Malkus (1956) wanted to use this to constrain the set of realisable turbulent mean profiles and it still remains an interesting requirement given the paucity of predictive alternatives. The concept of marginal stability still periodically rears its head in the literature to post-rationalise turbulent mean profiles (e.g. see \cite{Brauckmann17} and references therein) and EOS represents a finer tool for this. As noted in the introduction, the mean flow is increasingly taken as the starting point for resolvent flow analysis. To move this towards a full wall-bounded turbulence theory, a way has to be found to close the cycle between the fluctuation field and the mean. Just maybe, requiring statistical stability with respect to the dominant fluctuation fields could help constrain this.
A byproduct of this work is that it has also suggested an improved approach for predicting coherent structures in turbulent flows. {Coherence here means weakly damped or growing fluctuation {\em and} mean disturbances which emerge from the linear operator built around the base flow rather than, say, POD modes derived from the 2-pt correlations which have no direct relation to the underlying equations}. Initial work \citep{CrightonGaster76,Gaster85,Roshko93} linearising around the turbulent mean proved successful in free shear flows but less so when viscosity plays a leading role. Admittedly, EOS or mEOS needs extra (second cumulant) information about the base flow which adds cost to the analysis but the potential for better predictions is clear. A popular (cheap) alternative approach has been to apply OS analysis around the mean profile which incorporates an eddy viscosity to acknowledge the presence of the base fluctuations \citep{ReynoldsHussain72, Sen00, Sen07,DelAlamo06, Cossu09,HwangCossu10a, HwangCossu10b}. This eddy viscosity is taken to be that which is needed to sustain the turbulent mean profile and so makes no attempt to incorporate fluctuation information for each excited wavenumber of the base flow as in EOS. \rk{In the well-known triple expansion of Hussain \& Reynolds \citep{Hussain70}, the coherent wave part ($\tilde{f}$ in their equation (1.1)) is the equivalent of $\bm{\delta} {\bf \tilde{u}}^{mn}$ which there is excited by a wavemaker and then studied as it travels downstream in a turbulent flow. Orr-Sommerfeld analysis was found unable to reconcile what was seen experimentally but EOS should do better. }
The cumulant framework examined here to motivate EOS was used only to the first non-trivial level where only the first and second cumulants are retained. Whether this represents a good approximation or not to modelling the turbulent flow itself is a question not considered here. Rather the focus is on the associated spectral problem characterising the statistical stability {\em within} this framework {\em taking} the turbulent base flow as a given {from DNS or experiment}. Clearly considering higher cumulants beyond the the second would be better but the computational cost of handling the ensuing spectral problem becomes prohibitive. Interestingly, `projecting down' to an underlying Navier-Stokes-type problem fails immediately unless no closure at all is performed which means CE$\infty$ is considered. This gives what we have called the infinitely Extended Orr-Sommerfeld analysis (iEOS) which is as costly as EOS and unfortunately does not have a `minimal' version. Again, only targeting certain subsets of fluctuation fields - e.g. key triads suggested by resolvent flow analysis - may yet make this tractable.
\subsection{Future perspectives}
In terms of future work, an obvious focus is to improve the main approximation in EOS which is deriving a representative base flow. Here, we have used the leading singular vector of the second cumulant for each excited (vector) wavenumber but other strategies involving information from higher order cumulants are certainly possible (leaving aside the cost of computing these...). This would seem sensible in the case of iEOS which is a proxy for CE$\infty$ where {\em all} the cumulants are retained and the only other cumulant system beyond CE2 which allows a Navier-Stokes-type proxy. Pragmatically, a steady base flow has been sought which leads to a (conceptually) simple eigenvalue problem. Trying to go beyond this profoundly complicates matters.
One interesting future direction hinted at in the introduction is to relax the averaging procedure to just being over the streamwise direction or, more generally, redefine what constitutes the `mean'. A key feature we have been able to exploit is the connection between CE2 and the QL equations which still holds if the mean is more flexibly defined
to include some wavenumber subset of the fluctuation field. The then `generalised' QL equations \citep{Marston16} have the advantage of including more of the nonlinearity in the Navier-Stokes equations and can interpolate between the QL equations and the full Navier Stokes equations depending on how the mean is defined. A minimal mean extension would be to include one spanwise wavenumber field whereas including {\em all} fluctuation fields with no streamwise variation is equivalent to a mean defined by only streamwise averaging. There are many theoretical reasons for considering the latter, where the mean is $U(y,z){\bf \hat{x}}$ rather just $U(y){\bf \hat{x}}$, in wall-bounded flows (e.g. the SSP/VWI mechanism \cite{Waleffe97, Hall10}) but capturing it experimentally is very time-consuming task and using it reduces the predictive power of resolvent flow analysis. Even in computations, the spanwise homogeneity needs to be broken otherwise there is no reason to suppose the mean flow should not be spanwise-invariant (and then the spanwise structure of the mean profile needs to be divorced from the source of spanwise inhomogeneity). Nevertheless, this choice is starting to receive attention at least at the level of doing OS analysis on $U(y,z){\bf \hat{x}}$ (see \cite{Lozano-Duran21} and references herein) and this analysis can be extended as envisaged here.
Many challenges remain. Statistical stability is an important fundamental feature of turbulent flows yet is difficult to assess and hence to exploit. Hopefully, the work reported here represents a useful but admittedly small step forward.
|
\section{Introduction}
With the recent breakthroughs in the field of asteroseismology, as a result of the so-called `space-based photometry revolution' that was started by the CoRoT \citep{Baglin} and \textit{Kepler} \citep{Borucki} satellites and is still in full swing with TESS \citep{Ricker2015}, this young field of stellar physics has become a natural complement to the field of exoplanetology. Indeed, the need for precise and accurate stellar parameters remains a key issue for stellar modellers and a central aspect for the preparation of future missions such as PLATO \citep{Rauer}. This led to numerous synergies between exoplanetology and asteroseismology, illustrating the common interests of both fields in providing accurate depictions of exoplanet-host stars \citep[see e.g.][amongst others]{JCDKOI2010, Batalha2011, Huber2013,Huber2013Science,Davies2015,Silva2015,Lundkvist2016,Davies2016,Huber2018,Campante2018}. Beyond the need for accurate and precise planetary masses and radii, a good depiction of the host star is also required to properly understand the habitability, formation history, and dynamical evolution of planetary systems \citep{Huber2013Science,Farr2018}. Indeed, having access to temporal information by using stellar models also allows to put the current state of observed planetary systems in a `historical' perspective, permitting one to constrain additional physical phenomena such as the effects of tides and atmospheric evaporation on the architecture of exoplanetary systems.
In this paper, we demonstrate the importance of thorough seismic modelling for exoplanet-host red giant branch (RGB) stars. We show that going beyond the use of the standard seismic scaling relations is of primary importance to determine robust and accurate stellar parameters. We applied dedicated seismic modelling techniques to three red giants observed by TESS, for which long-period massive planets have been detected by the CASCADES survey (Ottoni et al. submitted), namely $\rm{HD22532}/\rm{TIC200841704}$, $\rm{HD64121}/\rm{TIC264770836}$, and $\rm{HD69123}/\rm{TIC146264536}$ that have been detected using radial velocity measurements with the CORALIE echelle spectrograph. All of these planets are long period Jupiter-like planets with periods of $872$, $623$, and $1193$ days, for which we could determine their masses to be between $2$ and $3$ Jupiter masses (Ottoni et al. submitted).
We start in Sect. \ref{AsteroObs} by describing the peakbagging procedure used to determine the individual pulsation frequencies for all targets. We focus on the analysis of radial $(\ell=0)$ and quadrupole $(\ell=2)$ modes which are used afterwards in the modelling. In Section \ref{SecModelling}, we detail the various procedures used to determine the fundamental properties of each star. We first illustrate in Sect. \ref{SecGlobalIndex} the use of the seismic scaling relations, discussing the impact of the various corrections proposed in the literature and the potential issues related to the robustness and accuracy of these techniques.
In Sects. \ref{SecIndFreq} and \ref{SecFreqInv}, we present the results of seismic modelling using individual frequencies and frequency differences, coupled to stellar evolutionary models. We compare the results of this dedicated modelling of each target to the various results of the scaling relations and discuss the implications of our tests for the requirements of precise stellar parameters for the purpose of exoplanetology. Finally, in Sect. \ref{SecOrbital}, we discuss the dynamical evolution of the planetary systems for each target under the effects of both dynamical and equilibrium tides in a fully coupled way with the evolution of the host star.
\section{Determination of oscillation parameters and frequencies}\label{AsteroObs}
We measured the observed asteroseismic oscillation mode frequencies, $\nu_{n, \ell}$, for the three targets using TESS photometric time series. We first constructed power spectra from the observed photometric flux. Using the peakbagging package \texttt{PBjam}\footnote{See \url{https://github.com/grd349/PBjam}} \citep{Nielsen2020}, we then determined the locations of radial and quadrupolar oscillation modes. Some evidence of dipolar modes could be seen in the spectra; however, as a result of the short duration of the observations, the frequency resolution could not allow us to provide a clear and unambiguous identification of these modes.
\begin{table}
\begin{center}
\caption{TESS sectors available for each target.}
\label{tab:sectors}
\begin{tabular}{c|c|c}
\hline\hline
\textbf{HD ID} & \textbf{TIC ID} & TESS Sectors \\
\hline
HD69123 & TIC146264536 & 7, 8 \\
HD22532 & TIC200841704 & 3, 4 \\
HD64121 & TIC264770836 & 8, 9 \\
\hline
\end{tabular}
\end{center}
\end{table}
Using the \texttt{lightkurve} package \citep{lightkurvecollaborationLightkurveKeplerTESS-2018} with \texttt{astropy} \citep{AstropyCollaboration.Robitaille.ea2013,AstropyCollaboration.Price-Whelan.ea2018} and \texttt{astroquery} \citep{Ginsburg.Sipocz.ea2019}, we constructed the power spectra as follows. We downloaded TESS light curves from the Mikulski Archive for Space Telescopes (MAST) for the available sectors given in Table~\ref{tab:sectors}. For each star, we stitched the pre-search data conditioning simple aperture photometry \citep[PDCSAP,][]{Stumpe.Smith.ea2012,Smith.Stumpe.ea2012} flux from each sector together. We discarded $5$-$\sigma$ outliers and removed low-frequency trends using a Savitzky-Golay filter. We then determined the power spectrum for each star using the Lomb-Scargle method \citep{Lomb1976,Scargle1982} and divided the power by an estimate of the background to determine the signal-to-noise ratio (S/N; where an S/N of 1 indicates the absence of any signal).
\begin{table*}
\begin{center}
\caption{Global stellar properties used as inputs for the PBjam peakbagging pipeline.}
\label{tab:seismo_input}
\begin{tabular}{c|c|c|c}
\hline \hline
\textbf{Input} & HD22532 & HD64121 & HD69123 \\
\hline
$\Delta\nu\,(\mu\mathrm{Hz})$ & $10.68\pm0.07$ & $11.33\pm0.09$ & $7.32\pm0.07$ \\
$\nu_{\mathrm{max}}\,(\mu\mathrm{Hz})$ & $129.38\pm2.12$ & $138.51\pm1.32$ & $88.58\pm1.54$\\
$T_{\mathrm{eff}}\,(\mathrm{K})$ & $5067\pm70$ & $4980\pm70$ & $4787\pm70$ \\
$G_{\mathrm{BP}} - G_{\mathrm{RP}}\,(\mathrm{dex})$ & $1.09\pm0.01$ & $1.08\pm0.01$ & $1.18\pm0.01$ \\
\hline
\end{tabular}
\end{center}
\end{table*}
We used the \texttt{PBjam} package to measure the observed radial, $\nu_{n,\ell=0}$, and quadrupolar, $\nu_{n,\ell=2}$, oscillation modes for each target. Initially, we performed the mode identification based on the methods of \citet{daviesAsteroseismologyRedGiants-2016} and a prior probability distribution constructed from thousands of stars already analysed using \texttt{PBjam}. The values and uncertainties of input parameters used to select stars from the prior are given in Table~\ref{tab:seismo_input}. The input frequency at maximum power, $\nu_{\mathrm{max}}$, and the large frequency separation, $\Delta\nu$ were determined using the layered approach of \citet{Elsworth_2020}, combining for $\Delta\nu$ the power spectrum of the power spectrum approach \citep{Hekker2010}, the universal pattern method \citep{Mosser2011}, and two other criteria for the mode properties \citep[See][ for additional details]{Elsworth_2020}. We also adopted the input effective temperatures, $T_{\mathrm{eff}}$, and colours, $G_{\mathrm{BP}} - G_{\mathrm{RP}}$, from \textit{Gaia} DR2 \citep{GaiaCollaboration.Prusti.ea2016,gaiacollaborationGaiaDataRelease-2018}. The inputs primarily determined the window in which we selected stars from the prior for subsequent mode identification. As a result, the inputs have very little effect on the final peakbagging step.
We identified initial mode locations, $\nu_{n,\ell}^{\,\prime}$, by fitting the asymptotic relation \citep{2013A&A...550A.126M} to 9 radial orders using Bayes' Theorem, $P(\boldsymbol{\theta}|D) \propto P(\boldsymbol{\theta}) P(D|\boldsymbol{\theta})$. The likelihood of the S/N data, $D$, given the model parameters, $\boldsymbol{\theta}$, is given by the term $P(D|\boldsymbol{\theta})$. The prior distribution of the model, $P(\boldsymbol{\theta})$, was obtained using a kernel density estimate for the population selected in the previous step. We sampled the posterior, $P(\boldsymbol{\theta}|D)$, using the \texttt{emcee} package \citep{foreman-mackeyEmceeMCMCHammer-2013}.
\begin{table*}
\begin{center}
\caption{Individual observed radial $(\ell=0)$ and quadrupolar $(\ell=2)$ oscillation modes, with their statistical uncertainties for each star.}
\label{tab:seismo_output}
\begin{tabular}{cc|cc|cc}
\hline \hline
\multicolumn{2}{c|}{HD69123} & \multicolumn{2}{c|}{HD22532} & \multicolumn{2}{c}{HD64121}\\
\hline
$\nu_{n,0}\,(\mu\mathrm{Hz})$ & $\nu_{n,2}\,(\mu\mathrm{Hz})$ & $\nu_{n,0}\,(\mu\mathrm{Hz})$ & $\nu_{n,2}\,(\mu\mathrm{Hz})$ & $\nu_{n,0}\,(\mu\mathrm{Hz})$ & $\nu_{n,2}\,(\mu\mathrm{Hz})$ \\
\hline
$66.90\pm0.12$ & --- & --- & $96.53\pm0.25$ & $104.64\pm0.30$ & --- \\
$74.00\pm0.15$ & --- & $108.23\pm0.14$ & $107.07\pm0.24$ & $115.99\pm0.27$ & $114.22\pm0.29$ \\
$81.29\pm0.15$ & $80.41\pm0.18$ & $118.95\pm0.10$ & --- & $127.02\pm0.10$ & $125.74\pm0.29$ \\
$88.43\pm0.20$ & $87.51\pm0.13$ & $129.45\pm0.10$ & $128.15\pm0.24$ & $138.44\pm0.23$ & $136.93\pm0.24$ \\
$95.82\pm0.05$ & $94.97\pm0.09$ & $140.07\pm0.11$ & --- & $149.67\pm0.11$ & --- \\
$103.36\pm0.10$ & $102.48\pm0.20$ & $150.73\pm0.18$ & $149.46\pm0.15$ & $161.06\pm0.14$ & --- \\
$110.80\pm0.16$ & $109.83\pm0.17$ & --- & --- & $172.53\pm0.32$ & $171.12\pm0.30$ \\
$118.43\pm0.18$ & --- & $172.31\pm0.30$ & --- & --- & --- \\
\hline
\end{tabular}
\end{center}
\end{table*}
Finally, we fitted a Lorentzian profile to each mode individually by sampling its posterior distribution using the Bayesian package \texttt{PyMC3} \citep{Salvatier.Wiecki.ea2016}. We used the initial mode identification as a prior on the Lorentzian centers, such that $\nu_{n,\ell} \sim \mathcal{N}(\nu_{n,\ell}^{\,\prime}, 0.03\Delta\nu)$, where $\mathcal{N}(\mu, \sigma)$ is a normal distribution with mean $\mu$ and standard deviation $\sigma$. All of the other parameters from the previous steps were relaxed. The resulting radial and quadrupolar mode locations are given in Table~ \ref{tab:seismo_output} and marked in Fig.~\ref{fig:seismo_model}. We also provide the power spectrum for each star in Appendix \ref{SecAddPlots}. We discarded results where the uncertainty over the prior standard deviation, $0.03\Delta\nu$, was greater than $95\%$, indicating uninformative data.
\begin{figure}
\centering
\begin{subfigure}[t]{0.47\textwidth}
\centering
\includegraphics[width=1.0\textwidth]{echelle_TIC146264536.png}
\end{subfigure}\\
~
\begin{subfigure}[t]{0.47\textwidth}
\centering
\includegraphics[width=1.0\textwidth]{echelle_TIC200851704.png}
\end{subfigure}\\
~
\begin{subfigure}[t]{0.47\textwidth}
\centering
\includegraphics[width=1.0\textwidth]{echelle_TIC264770836.png}
\end{subfigure}
\caption{Observational Echelle diagram for each star. Locations of the radial ($\ell=0$) and quadrupolar ($\ell=2$) oscillation modes are given by blue and orange circles, respectively. The S/N scale is indicated in greyscale on the right-hand side of the plots.}
\label{fig:seismo_model}
\end{figure}
\section{Seismic modelling}\label{SecModelling}
In this section, we carry out a detailed modelling of all three targets adding at each step a level of refinement in the procedure, discussing the consistency between the different approaches and their intrinsic limitations. First, we start with seismic scaling relations of global seismic indexes in Sect. \ref{SecGlobalIndex}. In Sect. \ref{SecIndFreq}, we carry out the modelling of each target using individual radial frequencies, we then discuss the precision, consistency and limitations of these results with previous approaches. Finally, we add one additional refinement in Sect. \ref{SecFreqInv} to the modelling procedure by carrying out an inversion of the mean density combining it in a fit of the small frequency separations in a Levenberg-Marquardt algorithm, computing the evolutionary models individually.
We start by presenting the classical constraints available for each target in Table~\ref{tabClassicConst}. We note that the evolutionary status of our targets is unambiguously determined by their global seismic parameters. Such ambiguity could have been an issue for the mean density inversions. Indeed, as demonstrated in \citet{Buldgen2019}, the mean-density inversion can be strongly biased if the evolutionary status of an RGB target is unknown, especially in the case of low-mass clump stars that can be mistaken for high-mass first-ascent RGB stars. We also mention that in the seismic modelling, we use an uncertainty of $70$K on the effective temperature to avoid overfitting, although this has a negligible impact on the final solution as the seismic data are, in comparison, much more precise.
\begin{table*}[t]
\caption{Classical constraints for the targets.}
\label{tabClassicConst}
\centering
\begin{tabular}{r | c | c | c | c }
\hline \hline
\textbf{Identifiers} & $\rm{HD22532}$ & $\rm{HD64121}$ & $\rm{HD69123}$ \\ \hline
$T_{\rm{eff}}$ $\rm{(K)}$ &$5038\pm24$&$5078\pm22$&$4842\pm41$\\
$\rm{L}$ $\rm{(L_{\odot})}$ &$18.80\pm0.33$&$17.70\pm0.30$&$29.51\pm0.57$\\
$\left[ \rm{Fe}/\rm{H} \right]$ $\left( \rm{dex}\right)$ &$-0.21\pm0.02$&$-0.19\pm0.02$&$0.05\pm0.03$\\
$\rm{\log g}$ $\left( \rm{dex}\right)$ &$3.09\pm0.07$& $2.91\pm0.16$&$2.86\pm0.11$\\
\hline
\end{tabular} \\
\small{\textit{Note:} See Ottoni et al. (submitted) for the discussion on the determination of the non-seismic parameters.}
\end{table*}
\subsection{Scaling relations and global indexes}\label{SecGlobalIndex}
Global seismic indexes have the advantage of providing a quick estimate of the global parameters for solar-like oscillators at a very low numerical cost, making them very useful for the studies of large samples of stars, or studies where the data quality does not allow for a more detailed modelling procedure. The seismic relations were originally presented by \citet{Brown1991} and \citet{Kjeldsen1995}, who originally aimed at predicting the pulsations properties of observed stars, although the link between the large frequency separation and the mean density has been known since \citet{Vandakurov1967}. The original scaling relations were thus written as functions of the observed large frequency separation, $\Delta \nu$, and the frequency of maximum power, $\nu_{\rm{max}}$, from a simple scaling law with the solar reference values.
More recently, the relations were re-arranged and used with the aim of estimating stellar fundamental parameters, such as the mass and radius \citep{Stello2008, Kallinger2010}, and they have since been widely used. However, while their success and usefulness have been recognised in the community, the direct use of the scaling relations to determine mass and radii estimates have come under some criticism, especially the scaling relation between the large frequency separation and the mean density. Indeed, comparisons between seismic and `dynamical' masses for red giant eclipsing binaries have shown the limitations of the use of the scaling relations \citep{Gaulme2016, Brogaard2018,Benbakoura2021}.
A key point mentioned by \citet{Brogaard2018} is the sensitivity of the results for the mass and radius with the applied corrections to relations, especially the $\Delta \nu$ relation. Indeed, some calibrated corrections have been proposed in the literature and applied in practical cases \citep[see e.g.][for some examples]{White2011, Sharma2016, Rodrigues2017, Yu2018, Kallinger2018} while other authors have even discussed the actual value of the large frequency separation to be used when applying them \citep{Mosser2013}. This implies that while formally simple, the scaling relations are also tweaked by correction factors to make them more robust, but these corrections significantly influence the results at the level of precision expected from seismic analyses. Similar conclusions on the sensitivity of the relations to their corrections are reached when comparing seismic parallaxes to those of \textit{Gaia} \citep{Hall2019, Khan2019, Zinn2019}. Indeed, the corrections for the $\Delta \nu$ and $\nu_{\rm{max}}$ relations defined in the literature can actually lead to variations in the determined masses and radii by more than $1\sigma$.
In this section, we illustrate and discuss this lack of robustness of the scaling relations in the context of the requirements of precise mass and radii determinations for the purpose of in-depth studies of exoplanetary systems. The seismic scaling relations for the mass and radius determinations are defined as
\begin{align}
\frac{\rm{M}}{\rm{M_{\odot}}} & \approx \left( f_{\nu_{\rm{max}}}\frac{\nu_{\rm{max}}}{\nu_{\rm{max},\odot}}\right)^{3}\left(f_{\Delta \nu}\frac{\Delta \nu}{\Delta \nu_{\odot}}\right)^{-4} \left(\frac{T_{\rm{eff}}}{T_{\rm{eff},\odot}}\right)^{3/2}, \label{eq:ScalingrelMass}\\
\frac{\rm{R}}{\rm{R_{\odot}}} & \approx \left( f_{\nu_{\rm{max}}}\frac{\nu_{\rm{max}}}{\nu_{\rm{max},\odot}}\right)\left(f_{\Delta \nu}\frac{\Delta \nu}{\Delta \nu_{\odot}}\right)^{-2} \left(\frac{T_{\rm{eff}}}{T_{\rm{eff},\odot}}\right)^{1/2}, \label{eq:ScalingrelRad}
\end{align}
where we have included correction factors $f_{\nu_{\rm{max}}}$ and $f_{\Delta \nu}$ that are commonly used and $\rm{M_{\odot}}$, $\rm{R_{\odot}}$, $\Delta \nu_{\odot}$ and $\nu_{\rm{max},\odot}$ are the solar mass, radius, average large frequency separation and frequency of maximum power, respectively. The solar reference values taken here are those of \citet{Huber2011}, namely $\Delta \nu_{\odot}=135.1 \mu\rm{Hz}$ and $\nu_{\rm{max},\odot}=3090\mu\rm{Hz}$. In practice, these corrections are often derived from grids of stellar models. Some authors have also advocated for additional terms in the scaling relations \citep{Viani2017} or even non-linear generalisations \citep{Kallinger2018}. Here we focus on the usual form of the scaling relations and compare them to more sophisticated modelling techniques. Recently, \citet{Li2021} have also investigated the intrinsic scatter of the scaling relations. They make the hypothesis that they provide reference values for stellar parameters and they used the red giant branch bump and the zero age core helium buring stage to measure the scatter in their stellar population. They find a limited scatter around these features when comparing them to their \textsl{Galaxia} synthetic population model \cite{Sharma2011} and point towards the difficulty to reduce the scatter by correcting the $\delta \nu$ scaling relation.
Equations \ref{eq:ScalingrelMass} and \ref{eq:ScalingrelRad} have been widely used in the field of asteroseismology to determine masses and radii of thousands of stars simultaneously. However, a strong weakness resides in the hypothesis of homology between the Sun and the observed star. Local approaches of the problem can also be applied, such as illustrated in \citet{ReeseDens} where the scaling relation between $\Delta \nu$ and the mean density is treated as a variational formula that is applied between a given reference model and the observed star. This local method was, however, found to be suboptimal when compared to the SOLA method, as shown by \citet{ReeseDens}, \citet{Buldgen}, and \citet{Buldgen2019}. Another way of using global seismic parameters consists in fitting the observed values to those given by theoretical models, as is done in the PARAM software \citep{Rodrigues2017}. This approach gets rid of the scaling with respect to the Sun, and it is likely preferable for the direct use of the scaling relations when only global seismic indexes are determined.
In Fig. \ref{FigScaling}, we illustrate the results of the application of the `raw' scaling relations, namely fixing $f_{\Delta \nu}$ and $f_{\nu_{\rm{max}}}$ to $1$, as well as the results obtained with two corrections found in the literature\footnote{We note that for HD22532, the correction from \citet{White2011} is almost zero and thus the results are similar to the `raw' scaling relations.}. The first issue we note is that depending on the applied corrections, the results can be significantly different by $1\sigma$ or more. For example, the correction derived from model grids by \citet{Rodrigues2017} leads to a systematic decrease in the estimated mass and radius. While this leads to mass determinations in better agreement with those from detailed modelling (see Sec. \ref{SecIndFreq}) for all stars, the fact that it may lead to variations of more than $1\sigma$ is an indicator that one should use the scaling relations with care. As for the correction based on $\rm{T_{eff}}$ of \citet{White2011}, the changes are of much smaller amplitude for our sample in this case. Overall, the significant differences between the various corrections confirms the results of \citet{Gaulme2016} and \citet{Brogaard2018}. Indeed, in comparing the values for the masses and radii to those determined from seismic inversions with \textit{Gaia} and spectroscopy in table \ref{tabMassModInd}, we can see that despite the corrections, the accuracy is still insufficient for HD69123. Similar conclusions were reached by \citet{Brogaard2018} when analysing the behaviour of the scaling relations for eclipsing binaries in detail.
\begin{figure}
\centering
\includegraphics[width=8cm]{ScalingResultsNew3}
\caption{Mass and radii determinations from the scaling relations using various corrections found in the literature as compared to the values determined from the inversion procedure and the radii determined using \textit{Gaia} and spectroscopic constraints shown in Table \ref{tabMassModInd}.}
\label{FigScaling}
\end{figure}
From this test, it appears that final values of the seismic masses and radii determined from the scaling relations significantly depends on the underlying correction factor $f_{\nu_{\rm{max}}}$ and $f_{\Delta \nu}$ introduced in the scaling relations for the mass and radius, which is unsuitable for the detailed modelling we wish to do here to be able to follow the evolution of the planetary system. Indeed, an accurate and reliable mass determination is required to properly follow the various evolutionary phases that will influence the dynamical evolution of the system as well as the evaporation of the planetary atmospheres. To that end, a detailed modelling seems more suitable as it takes more seismic and non-seismic constraints into account, providing a more reliable solution. Such an example of a more thorough analysis is carried out in the following sections.
\subsection{Individual frequencies modelling}\label{SecIndFreq}
The modelling using individual frequencies can be carried out using various approaches. In what follows, we make use of the Asteroseismic Inference on a Massive Scale software \citep[AIMS,][]{Rendle2019, Montalban2020} to model the targets and compare the results with the ones from the scaling relations.
For the purpose of these computations, two separate grids of models were computed. One grid is dedicated to $\rm{HD69123}$ and another grid to $\rm{HD22532}$ and $\rm{HD64121}$. The grid parameters are given in Table~\ref{tabGridProp}\footnote{Fits were also computed with the original grid of \citet{Rendle2019} for the sake of completeness but are not presented here, the results were consistent within $2\sigma$ but differences were seen in $T_{\rm{eff}}$ and $\rm{L}$, as a result of the different boundary conditions of the models of \citet{Rendle2019} that assumed an Eddington $T({\tau})$ relation.}. The grids were computed with the Liège stellar evolution code \citep[CLES,][]{ScuflaireCles} and the adiabatic frequencies were computed with the Liège stellar oscillation code \citep[LOSC,][]{ScuflaireOsc}
\begin{table*}[t]
\caption{Properties of the AIMS stellar evolution models grids.}
\label{tabGridProp}
\centering
\begin{tabular}{r | c | c }
\hline \hline
\textbf{Parameters}&\textbf{Solar $Z$ grid}& \textbf{Low $Z$ grid} \\ \hline
Mass $\left( M_{\odot} \right)$&$\left[1.00-2.2\right]$ $(0.02\; \rm{step})$ &$\left[1.10-1.90\right]$ $(0.02\; \rm{step})$ $(0.02\; \rm{step})$ \\
$X_{0}$ &$\left[0.68, 0.72 \right]$ $(0.01\; \rm{step})$&$\left[0.71, 0.75 \right]$ $(0.01\; \rm{step})$\\
$Z_{0}$&$\left[ 0.010, 0.040 \right]$ $(0.001\; \rm{step})$&$\left[ 0.006, 0.010 \right]$ $(0.001\; \rm{step})$\\
$\alpha_{\rm{MLT}}$&$2.03$&$2.03$\\
$\nu_{\rm{max}}$ cutoff $\left( \mu \rm{Hz}\right)$&$40$&$60$\\
\hline
\end{tabular}
\end{table*}
The solar mixture used for the grid is that of \citet{AGSS09}, the OPAL opacities \citep{OPAL} and the FreeEOS equation of state were used \citep{Irwin}. In all grids, a solar-calibrated value of the mixing-length parameter fixed at $2.03$ has been used. The atmosphere model used is Model C of \citet{Vernazza}. Microscopic diffusion was not considered in the models and a core overshooting value of $0.15\rm{H_{P}}$, with $\rm{H_{P}}=\left(\frac{dr}{d \ln P}\right)$ being the pressure scale height, was used as well as an envelope overshooting value of $0.15\rm{H_{P}}$. The overshooting regions are considered fully mixed in all cases and the temperature gradient was fixed to the adiabatic gradient in the case of core overshooting and to the radiative gradient in the case of envelope overshooting. The inclusion of core overshooting, although here quite crude, was motivated by its requirement to reproduce seismic observations by \citet{Deheuvels2016}, \citet{Bossini2015}, and \citet{Bossini2017} as well as for eclipsing binaries \citep{Claret2016,Claret2018,Claret2019}, whereas envelope overshooting has been found to be required to reproduce the position of the RGB-Bump in Kepler data \citep{Khan2018}. We also mention here that neglecting microscopic diffusion in the models leads to biases in our determinations of fundamental parameters, especially the age determination that is dominated by the duration of the MS phase. In addition, uncertainties regarding the properties of core overshooting dominate the uncertainties on the ages we report\footnote{This remains true at all stages of our modelling procedure, whatever the seismic analysis performed.}, as all stars exhibit a convective core on the main sequence.
The modelling was carried out by fitting the individual radial modes of the stars, using the two-terms surface correction of \citet{Ball1}, in conjunction with $\rm{T_{eff}}$, $\left[ \rm{Fe/H}\right]$, and $\rm{L}$. As seen from Figs \ref{FigEchAIMS}, \ref{FigDistribMass} and \ref{FigDistribRadius}, the MCMC modelling has found a well-defined solution in all cases. The observed individual frequencies were corrected from the line-of-sight Doppler velocity shifts following the recommendations of \citet{Davies2014}.
However, we can see that the models have slightly different values from the radii from Ottoni et al. (submitted), which were not included in the dataset for the MCMC modelling. This can be regarded as a word of caution against the systematic use of individual frequencies in seismic modelling (as mentioned in \citet{Rendle2019} and in \citet{Buldgen2019}) as they heavily relied on the surface corrections used to determine the optimal solution \citep[see e.g.][for a discussion in the context of eclipsing binaries and stellar clusters]{Jorgensen2020}. If the individual modes are determined with very high precision, this can lead to extremely precise, although not fully accurate, solutions. In our specific cases, the seismic radius from the individual frequency modelling is only slightly different from the one determined from the parallaxes and spectroscopic observations, and within $1.5\sigma$ error bars.
\begin{figure*}
\centering
\includegraphics[width=15cm]{EchelleAlNew2}
\caption{Echelle diagram illustrating the agreement between theoretical and observed radial frequencies for the AIMS solutions using the grids of table \ref{tabGridProp}.}
\label{FigEchAIMS}
\end{figure*}
\begin{figure*}
\begin{flushleft}
\begin{minipage}{\textwidth}
\includegraphics[trim=70 25 67 95, clip, width=0.325\linewidth]{histogramMassTIC264770836} \includegraphics[trim=70 25 70 95, clip, width=0.325\linewidth]{histogramMassTIC146264536}
\includegraphics[trim=70 25 70 95, clip, width=0.325\linewidth]{histogramMassTIC200841704}
\end{minipage}
\end{flushleft}
\caption{Probability distribution functions for the mass of $\rm{HD64121}$ (top left), $\rm{HD22532}$ (bottom), and $\rm{HD69123}$ (top right).The vertical red line in the plots indicates the position of the best model in the grid (without interpolation by AIMS).}
\label{FigDistribMass}
\end{figure*}
\begin{figure*}
\begin{flushleft}
\begin{minipage}{\textwidth}
\includegraphics[trim=70 25 70 95, clip, width=0.325\linewidth]{histogramRadiusTIC264770836}
\includegraphics[trim=70 25 70 95, clip, width=0.325\linewidth]{histogramRadiusTIC146264536}
\includegraphics[trim=70 25 70 95, clip, width=0.325\linewidth]{histogramRadiusTIC200841704}
\end{minipage}
\end{flushleft}
\caption{Probability distribution functions for the radius of $\rm{HD64121}$ (top left), $\rm{HD22532}$ (bottom), and $\rm{HD69123}$ (top right). The vertical red line in the plots indicates the position of the best model in the grid (without interpolation by AIMS).}
\label{FigDistribRadius}
\end{figure*}
Overall, the solution found by AIMS is of sufficient quality to serve as an initial condition for the next modelling step. However, since we did not exploit the information of the quadrupolar modes, we can assume that the modelling can be further refined. In addition, having clearly identified dipolar modes at our disposal would allow for an even more refined modelling, especially regarding core conditions (e.g. size of the helium core and boundary of the hydrogen burning shell). Unfortunately, as mentioned above, the dataset at our disposal was insufficient to fully identify them and thus exploit them in our theoretical modelling.
Indeed, the individual radial frequencies constrain the mean density almost solely, while their relatively high precision implies that the modelling focuses on reproducing them specifically\footnote{The issue would be even worse when directly fitting the individual frequencies of a main-sequence solar-like oscillator, due to higher relative precision and the larger number of constraints if non-radial modes are also included.}. Multiple strategies can be used to mitigate this aspect, one is to decrease the weight of the seismic constraints in the modelling or to increase the weight of the non-seismic constraints. None of these approaches are optimal in practice and the modelling strategies should be adapted to avoid such issues, especially if one wishes to study large samples automatically.
From a physical point of view, it should be noted that the apparent contradiction between seismic radii values and the radii determined from \textit{Gaia} parallaxes and spectroscopy may stem from the surface correction, or from the use of a fixed solar mixing-length parameter value for the grid\footnote{Other physical ingredients of the models could be at play but a detailed analysis of all degeneracies is beyond the scope of our study and would require a larger sample with higher-quality data.}.
\subsection{Seismic inversions and frequency differences}\label{SecFreqInv}
The inversions of the mean density were carried out using the reference models computed with AIMS, following the approach of \citet{ReeseDens}. As shown by \citet{Buldgen2019}, this approach is suitable to exploit the information of radial oscillations of first-ascent RGB and core Helium-burning stars, given that their evolutionary status is known. In the current study, this was the case for all our targets.
The inversion was computed following the variational formalism of \citet{Dziemboswki90}, using the SOLA method \citep{Pijpers} to carry out the inversion. The cost function minimised by the inversion is formally written as follows:
\begin{align}
\mathcal{J}_{\bar{\rho}}(c_{i})=&\int_{0}^{1}\left[K_{\mathrm{Avg}} - \mathcal{T}_{\bar{\rho}}\right]^{2}dx + \beta \int_{0}^{1}\left( K_{\mathrm{Cross}}\right)^{2}dx \nonumber \\ &+ \lambda \left[ 2 -\sum_{i}c_{i} \right] + \tan \theta \frac{\sum_{i}\left(c_{i}\sigma_{i}\right)^{2}}{<\sigma^{2}>}, \label{eq:CostSOLA}
\end{align}
where we have defined the target function of the inversion, the averaging and cross-term kernels,
\begin{align}
\mathcal{T}_{\bar{\rho}}&=4\pi x^{2} \frac{\rho}{\rho_{R}}, \\
K_{\mathrm{Avg}}&=\sum_{i}c_{i}K^{i}_{\rho,\Gamma_{1}}, \\
K_{\mathrm{Cross}}&=\sum_{i}c_{i}K^{i}_{\Gamma_{1},\rho},
\end{align}
with the radial position of an element of stellar plasma divided by the photospheric stellar radius $x=r/R$, being $\rho$ the local density, and $\rho_{R}=M/R^{3}$, with $M$ being the stellar mass and $R$ being the photospheric stellar radius. We have also introduced the parameters $\beta$ and $\theta$, defining the trade-off problem between the fit of the target, the contribution of the cross term and the amplification of observational error bars of the individual frequencies, denoted $\sigma_{i}$. The $K^{i}_{\rho,\Gamma_{1}}$ and $K^{i}_{\Gamma_{1},\rho}$ are the structural kernel functions, derived from the variational analysis of the pulsation equations and $<\sigma^{2}>=\frac{1}{N}\sum_{i=1}^{N}\sigma^{2}_{i}$ with $N$ being the number of observed oscillation modes. In Eq. \ref{eq:CostSOLA}, we have also defined the inversion coefficients $c_{i}$ and $\lambda$, a Lagrange multiplier. The third term is based on homologous reasoning described in \citet{ReeseDens}, which also serves to derive a non-linear generalisation of the method where the mean density, denoted $\bar{\rho}_{\mathrm{Inv}}$, was computed from
\begin{align}
\bar{\rho}_{\mathrm{Inv}}=\left(1+\frac{1}{2}\sum_{i}c_{i}\frac{\delta \nu_{i}}{\nu_{i}}\right)^{2}\bar{\rho}_{\mathrm{Ref}},
\end{align}
with $\bar{\rho}_{\mathrm{Ref}}$ being the mean density of the reference model and $\frac{\delta \nu_{i}}{\nu_{i}}$ being the relative differences between the observed and theoretical frequencies defined as $\frac{\nu_{\mathrm{Obs}}-\nu_{\mathrm{Ref}}}{\nu_{\mathrm{Ref}}}$. If we use the non-linear generalisation, the errors on the mean density value from the inversion are given by
\begin{align}
\sigma_{\bar{\rho}_{\mathrm{Inv}}}=\bar{\rho}_{\mathrm{Ref}}\left(1+\frac{1}{2}\sum_{i}c_{i}\frac{\delta \nu_{i}}{\nu_{i}} \right)\sqrt{\sum_{i}c^{2}_{i}\sigma^{2}_{i}}.
\end{align}
The results of the inversion for the mean density are the following: $\bar{\rho}^{\rm{HD69123}}_{\rm{Inv}}=0.00436 \pm 0.00004$ $\rm{g/cm^{3}}$, $\bar{\rho}^{\rm{HD22532}}_{\rm{Inv}}=0.00907 \pm 0.00007$ $\rm{g/cm^{3}}$, and $\bar{\rho}^{\rm{HD64121}}_{\rm{Inv}}=0.01025 \pm 0.00025$ $\rm{g/cm^{3}}$. The uncertainties were determined from both the analysis of model-dependencies as well as the use of various surface corrections, as was done in \citet{Buldgen2019}.
The mean density determined from the inversion was then used as a constraint for an additional forward modelling step combined with the small frequency separations. This additional step was carried out with a Levenberg-Marquardt minimisation technique, which allowed us to consider additional free parameters and to compute the evolutionary models directly with CLES.
The constraints used for this local minimisation were as follows: $\left[\rm{Fe}/\rm{H} \right]$, $\rm{T_{eff}}$, $\rm{L}$, $\bar{\rho}$, and the individual $d_{0,2}$. The free parameters were the mass, the age, the initial chemical composition, and the mixing-length parameter of convection. The main motivation for the use of the individual $d_{0,2}$ as constraints for the modelling is that \citet{Montalban2010, Montalban2012} have shown that these quantities are sensitive to the stellar mass for a given value of the large frequency separation due to the radius sensitivity of $d_{0,2}$ for a given mean density.
The results of this third modelling step are given in table~\ref{tabFinMod} and the agreement in terms of an individual small separation are illustrated in Fig. \ref{Figd02} for one target of our sample. In Fig. \ref{FigHRAll}, we illustrate the evolutionary tracks computed with CLES for each target, using the optimal parameters derived from the seismic modelling.
\begin{figure}
\centering
\includegraphics[width=8cm]{d02New2}
\caption{Comparisons between the theoretical and the observed $d_{0,2}$ for TIC$146264536$/HD$69123$.}
\label{Figd02}
\end{figure}
Given this third modelling step, we can see that a better agreement in terms of the radius was found for the modelling, implying a better mass determination. However, given the degenerate interplay between the initial chemical composition and the mixing-length parameter that impacts our final results, we can safely state that the estimation of the age of the targets could still be refined by exploiting the information of individual dipolar mixed modes, following for example the approach of \citet{Deheuvels2011} and \citet{Noll2021}. In our case, however, these modes were not identified in the power spectrum and were thus absent from the modelling.
\begin{table*}[t]
\caption{Parameters of the optimal stellar models found from the Levenberg-Marquardt minimisation.}
\label{tabFinMod}
\centering
\begin{tabular}{r | c | c | c | c }
\hline \hline
\textbf{Identifiers}& $\rm{HD22532}$ & $\rm{HD64121}$ & $\rm{HD69123}$\\ \hline
$\rm{M}$ $\rm{(M_{\odot})}$ &$1.23\pm0.07$&$1.19\pm0.07$&$1.43\pm0.06$\\
$\rm{R}$ $\rm{(R_{\odot})}$ &$5.77\pm0.04$&$5.49\pm0.04$&$7.78\pm0.05$\\
$\rm{X_{0}}$ &$0.71\pm0.02$&$0.72\pm0.02$&$0.72\pm0.01$\\
$\rm{Z_{0}}$ &$0.0081\pm0.001$&$0.0078\pm0.001$&$0.0137\pm0.001$\\
$\rm{\alpha_{MLT}}$ &$2.06\pm0.1$& $2.15\pm0.1$&$2.08\pm0.1$\\
$\rm{Age}$ $\rm{(Gyr)}$ &$4.03\pm1.1$&$4.51\pm1.0$&$3.31\pm1.5$\\
\hline
\end{tabular}
\end{table*}
It is also clear that the values in tab. \ref{tabFinMod} are subject to the intrinsic limitations of the stellar evolution code and could substantially vary if we changed the physics of the models on both the micro- and macroscopical level. However, for each star, we could also provide a model independent mass interval from the combination of the mean density determined by the inversion and the radius determined from the \textit{Gaia} parallaxes and the spectroscopic data. These intervals are provided in table~\ref{tabMassModInd}. While these values were computed independently from any stellar model, they still suffer from some limitations, namely the intrinsic accuracy of the inversion techniques discussed in \citet{Buldgen2019} as well as that of the radius determination. In other words, these mass intervals are sensitive to the stellar spectroscopic and astrometric observations, the bolometric correction and the extinction laws used in conjunction with the \textit{Gaia} data.
\begin{table*}[t]
\caption{Model-independent mass intervals from seismic inversions and radii from \textit{Gaia} parallaxes and spectroscopy.}
\label{tabMassModInd}
\centering
\begin{tabular}{r | c | c | c | c | c}
\hline \hline
\textbf{Identifiers}& $\rm{HD22532}$ & $\rm{HD64121}$ & $\rm{HD69123}$\\ \hline
$\rm{M}$ $\rm{(M_{\odot})}$ &$1.21\pm0.10$&$1.18\pm0.11$&$1.43\pm0.12$\\
$\rm{R}$ $\rm{(R_{\odot})}$ &$5.70\pm0.12$&$5.44\pm0.11$&$7.72\pm0.17$\\
\hline
\end{tabular}
\end{table*}
When compared with the values determined by both AIMS and the seismic scaling relations, we can see that the match is quite good for all three targets. However, at the desired level of precision, it appears that the disagreements can be of more than $1\sigma$ for some targets, as a result of the correction factors. Again this illustrates the lack of robustness of the scaling relations at this level of precision and the need for a detailed stellar modelling procedure to study planetary systems in detail.
\begin{figure*}
\centering
\includegraphics[width=18cm]{HRAllNew2}
\caption{Hertzsprung-Russel diagram showing the tracks of all targets, the location of the current position of the system being given by the red star.}
\label{FigHRAll}
\end{figure*}
\section{Orbital evolution}\label{SecOrbital}
In this section, we aim to study the orbital evolution of the three planetary systems HD22532 (TIC200841704), HD64121 (TIC264770836), and HD69123 (TIC146264536). The orbital parameters used in this study are presented in Table~\ref{param_orb}, in which we report the orbital period in units of days, the values of the semi-major axis in AU, and finally the upper limits on the planetary masses expressed in units of Jupiter masses. We note that the inclinations of the systems are unknown.
First, we computed stellar models representative of the host stars of the systems with the CLES stellar evolution code. We used the stellar parameters derived from the asteroseismic modelling as input and constraints. Second, we coupled the stellar models to our orbital evolution code \citep{Privitera2016A, Privitera2016B, Privitera2016c, Meynet2017, Rao2018}, taking the exchange of angular momentum between the planetary orbit and the star into account, occurring through the dissipation of tides in the stellar convective envelope. Thanks to this approach, we can investigate whether dynamical and/or equilibrium tides had a significant impact during the past history of the systems.
The physics adopted in the orbital evolution code is described in more detail in the work of \citet{Rao2018}. Here, we briefly recall the fundamental equations. The equation for the total change of the orbital distance reads as follows:
\begin{align}
\rm \left(\dot{a}/a \right) = - \dfrac{\dot{M}_{\star} +\dot{M}_{pl}}{ M_{\star} + M_{pl}} - \dfrac{2}{M_{pl} v_{pl}} \left[ F_{fri} + F_{gra} \right] + \left( \dot{a}/a\right)_{t}, \label{eq:total_a}
\end{align}
where the term $\rm \dot{M}_{\star} = - \dot{M}_{loss}$ indicates the stellar mass loss rate, $\rm M_{pl}$ is the mass of the planet, $\rm \dot{M}_{pl}$ is the rate of change of the planetary mass and $\rm v_{pl}$ is the planetary orbital velocity. Futhermore, $\rm F_{fri}$ and $\rm F_{gra}$ represent the contribution due to the frictional and gravitational drag forces respectively, whose expressions are taken as in \citet{Villaver2009, Mustill2012, Villaver2014}. The term $\rm \left( \dot{a}/a \right)_{t} $ includes the impacts of dynamical and equilibrium tides. In particular, the expression for equilibrium tides \citep{Zahn1966, Alexander1976, Zahn1977, Zahn1989, LS1984b, Villaver2009, MV2012, Villaver2014} is taken as in \citet{Privitera2016B} and their contribution is accounted for only when a stellar convective envelope is present. The equation for equilibrium tides is
\begin{align}
\rm \left( \dot{a}/a \right)_{eq} = \frac{f}{\tau_{cz}} \frac{M_{env}}{M_{\star}} q(1+q) \left( \frac{R_{\star}}{a} \right)^8 \left( \frac{\Omega_{\star}}{\omega_{pl}} - 1\right).\label{eq:equi_tides}
\end{align}
The term $\rm f$ is a numerical factor obtained from integrating the viscous dissipation of the tidal energy across the convective zone \citep{Villaver2009}, $\rm M_{env}$ is the mass of the convective envelope, $\rm q$ is the ratio between the mass of the planet and the one of the star ($\rm q = M_{pl}/M_{\star}$), $\rm \Omega_{\star}$ is the stellar surface rotation rate, $\rm \omega_{pl} = 2\pi/ P_{orb}$ is the orbital frequency of the planet, and $\rm \tau_{cz} $ is the convective turnover timescale.
As mentioned above, we also accounted for the impact of dynamical tides, in the form of a frequency-averaged tidal dissipation of inertial waves excited in the convective envelope of a rotating star, for which the Coriolis force is the restoring force \citep{BolmontMathis2016,Gallet2017,Bolmont2017,Benbakoura2019}. In our study, we assume the planet is on a circular-coplanar orbit around its host star. In this context, the effect of dynamical tides is accounted for whenever the condition $\rm \omega_{pl} < 2~\Omega_{\star}$ is satisfied since a planet orbiting a star on a circular-coplanar orbit is able to excite inertial waves when the orbital frequency is lower than twice the stellar rotation rate \citep{Ogilvie2007}. The expression of the change in orbital distance under the effects of dynamical tides is taken from \citet{Ogilvie2013} and \citet{Mathis2015} and is written as
\begin{align}
\rm \left( \dot{a}/a \right)_{dyn} = \left( \dfrac{9}{2Q^{\prime}_{d}} \right)q \omega_{pl} \left( \frac{R_{\star}}{a} \right)^5 \dfrac{(\Omega_{\star} - \omega_{pl})}{\mid \Omega_{\star} - \omega_{pl}\mid},
\label{eq:dyn_tides}
\end{align}
with $\rm Q^{\prime}_{d} = 3/(2D_{\omega})$ and $\rm D_{\omega} = D_{0\omega}D_{1\omega}D_{2\omega}^{-2}$. The `D' terms are defined as
\begin{align}
\begin{cases}
\rm D_{0\omega} = \dfrac{100\pi}{63} \epsilon^{2} \dfrac{\alpha^5}{1 - \alpha^5} (1 - \gamma)^2,\\
\rm D_{1\omega} = (1 - \alpha)^4 \left( 1 + 2\alpha + 3\alpha^2 + \frac{3}{2} \alpha^3 \right)^2 ,\\
\rm D_{2\omega} = 1 + \frac{3}{2}\gamma + \frac{5}{2 \gamma}\left( 1 + \frac{\gamma}{2} - \frac{3 \gamma^2}{2} \right) \alpha^3 - \frac{9}{4}\left(1 - \gamma\right)\alpha^5 ,
\end{cases}
\end{align}
where $\rm \alpha = R_{c}/R_{\star}$, $\rm \beta = M_{c}/M_{\star}$, $\rm \gamma = \dfrac{\alpha^3 (1 - \beta)}{\beta (1 - \alpha^3)}$, $\rm \epsilon = \dfrac{\Omega_{\star}}{\sqrt{\dfrac{\rm GM_{\star}}{\rm R_{\star}^3}}}$. Furthermore, $\rm M_{c}$ and $\rm R_{c}$ represent the mass and the radius of the radiative core. The term $\rm D_{\omega}$ was computed in \citet{Ogilvie2013} as the frequency-averaged tidal dissipation.
Whenever the orbital distance of the planet becomes equal to the corotation radius, defined as the distance at which $\rm \omega_{pl} = \Omega_{\star}$, tides become inefficient. When this condition is not satisfied, the tides widen or shrink the orbit, when the planet is beyond or inside the corotation radius, respectively.
Magnetic braking at the stellar surface is taken into account by using the formalism of \citet{Matt2015,Matt2019}, for which the magnetic torque writes
\begin{align}
\rm \dfrac{dJ}{dt} =
\begin{cases}
\rm -T_{\odot} \left(\dfrac{R}{R_{\odot}} \right)^{3.1} \left( \dfrac{M}{M_{\odot}} \right)^{0.5} \left(\dfrac{\tau_{cz}}{\tau_{cz \odot}} \right)^{p} \left( \rm \dfrac{\Omega}{\Omega_{\odot}} \right)^{p+1} &, \rm \text{if} ~ \left( Ro > Ro_{\odot}/\chi \right),\\
\rm -T_{\odot} \left(\dfrac{R}{R_{\odot}} \right)^{3.1} \left( \dfrac{M}{M_{\odot}} \right)^{0.5} \chi^{p} \left( \dfrac{\Omega}{\Omega_{\odot}} \right) & , \rm \text{if} ~ \left( Ro \leq Ro_{\odot}/\chi \right) ,
\end{cases}
\end{align}
where $\rm R_{\odot}$ and $\rm M_{\odot}$ are the radius and the mass of the Sun, and $\rm R$ and $\rm M$ are the radius and the mass of the stellar model. Furthermore, $\rm \tau_{cz}$ is the convective turnover timescale and $\rm Ro$ is the Rossby number, defined as the ratio between the stellar rotational period and the convective turnover timescale ($\rm Ro = P_{\star}/\tau_{cz}$). The quantity $\rm \chi \equiv Ro_{\odot} / Ro_{sat}$ indicates the critical rotation rate for stars with a given $\rm \tau_{cz}/\tau_{cz_{\odot}}$, defining the transition from saturated to unsaturated regime. We considered $\rm \chi$ to be equal to $\rm 10$ as in \citet{Matt2015} and \citet{Eggenberger2019a}. The exponent $\rm p$ was considered equal to $\rm 2.3$ and the constant $\rm T_{\odot}$ was calibrated in order to reproduce the solar surface rotation rate \citep{Eggenberger2019a}.
We accounted for the evaporation of the planetary atmosphere following the formalism of the Jeans escape regime \citep{Villaver2007} or the one of the hydrodynamic escape regime, depending on the properties of the planetary system under study. When using the hydrodynamic escape regime, we computed the planetary mass loss by using the energy limited formula as in the work of \citep{Lecavelier2007,Erkaev2007}.
For each of the systems considered, we initially studied the evolution of the orbits under the impact of dynamical and equilibrium tides at a fixed planetary mass. Therefore, we did not account for the atmospheric mass loss at this stage. We took the values of the semi-major axis and the masses reported in Table~\ref{param_orb} as initial input and tested whether we were able to reproduce the position of the planets at the ages of the systems.
The rotational history of the star being unknown, we shall considered a range of initial surface rotation rates representative of different kinds of rotators. We considered three different values for the initial surface rotation rate (namely $\rm \Omega_{in} = 3.2, 5$, and $\rm 18~ \Omega_{\odot}$), covering the range for slow, medium and fast rotators as deduced from surface rotation rates of solar-type stars observed in open clusters at different ages \citep{Eggenberger2019a}. A disk lifetime of 2 Myr and 6 Myr was used for the fast and medium-slow rotating case, respectively. During the disk-locking phase, we assumed the surface angular rotation remained constant. We also assumed the star rotating as a solid body. In Fig.~\ref{Om_Surf}, we report the evolution of the surface rotation rates computed for HD64121. The tracks show that at the beginning of the evolution the star experiences a spin-up due to the rapid contraction of the structure occurring during the PMS phase, reaching a maximum at the age of $\rm \sim 16$ Myr. After the peak, the braking due to magnetised stellar winds started to be efficient. Nevertheless, we notice that for a star such as HD64121, the efficiency of this process is reduced by the presence of a very shallow external convective envelope. The rotation rate evolutionary tracks therefore remain quite flat for the duration of the MS phase, until the star reaches the RGB, expanding and significantly reducing its rotation rate.
It is worth noting that when considering a value of $\rm \Omega_{in}$ as large as $\rm 18~\Omega_{\odot}$ and a disk-locking timescale of 2 Myr, the surface rotation rate of the star rapidly reaches overcritical values. We attempted to mitigate this effect by using a longer disk-locking timescale of 6 Myr as in the case of the slow and medium rotators, and we managed to maintain the track at subcritical values (as shown in Fig.~\ref{Om_Surf}). This method did not work for the systems HD22532 and HD69123, for which we excluded the fast rotating track as a viable rotational history of the host star. This result emphasizes how such an analysis of planetary systems can help to lift the degeneracy on the possible rotational histories of the stars.
Using this initial set up, we did not find any appreciable evolution of the orbital distances for our systems. This is an expected outcome, which is in agreement with previous results reported in \citet{Privitera2016B}, given the relatively large value for the orbital distances together with the moderate planetary masses. Even when considering the highest value for the initial surface rotation for HD64121, we did not obtain significant impacts on the planetary orbit. We explored a range of lower initial orbital distances for HD64121b, using an intermediate value for the surface rotation rate of the host star ($\rm \Omega_{in} = 5~\Omega_{\odot}$), in order to determine the maximal value below which the orbit of the planet would be significantly affected by tides, eventually leading to an engulfment at earlier times with respect to the age of the system. As shown in the top panel of Fig.~\ref{Orbit_Evol}, we find that the maximal value of the initial orbital distance below which the orbit would be impacted and the planet subsequently engulfed is $\rm a_{max} \approx 0.5$ AU. We would expect $\rm a_{max}$ to shift at higher values when considering larger surface rotation rates ($\rm \Omega_{in} = 18~\Omega_{\odot}$).
Subsequently, we computed the evolution of the systems by also including the impact of atmospheric evaporation. For the three systems considered, we followed the evolution of the escape parameter, assuming a value of the Bond albedo of $\rm A = 0.5$ \citep{Villaver2007}. We find that the escape parameters always have values above the threshold representing the switch from Jeans escape to hydrodynamic escape regime, here considered as $\rm E \gtrsim 20$ \citep{Villaver2007, Villaver2009}. In this context, the evaporation rates computed within the Jeans escape regime conditions resulted in a negligible impact on the evolution of the planetary masses.\\
\begin{table*}
\caption{Orbital parameters of HD22532b, HD64121b, and HD69123b.}
\label{param_orb}
\centering
\begin{tabular}{r | c | c | c }
\hline \hline
\textbf{Identifiers}& $\rm{HD22532b}$ & $\rm{HD64121b}$ & $\rm{HD69123b}$ \\ \hline
$\rm P ~(days)$ & 872.6~$\pm$~2.8 & 623.0~$\pm$~3.4 & 1193.3~$\pm$~7.0\\
$\rm a ~(AU)$ &$1.900 \pm 0.004$ & $1.510 \pm 0.006$ & $2.482 \pm 0.010$\\
$\rm M_{pl} sin(i) ~ (M_{J})$ & $2.12 \pm 0.09$ & $ 2.56 \pm 0.19$ & $ 3.04 \pm 0.16$\\
\hline
\end{tabular} \\
\end{table*}
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{HD64121_Omega_evolution.png}
\caption{Surface angular velocity evolution for HD64121 (TIC 264770836) in the case of fast rotator ($\rm \Omega_{in} = 18 \times \Omega_{\odot}$, black solid line), a medium rotator ($\rm \Omega_{in} = 5 \times \Omega_{\odot}$, blue solid line) and a slow rotator ($\rm \Omega_{in} = 3.2 \times \Omega_{\odot}$). The red dashed line represents the critical velocity limit. The black dotted vertical line indicates the age of the system.}
\label{Om_Surf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{HD64121b_Orbits_amin.png}
\caption{Impact of a change in orbital distance (blue solid lines) for HD64121b, for a host-star initial surface rotation of $\rm \Omega_{in} = 5 \times \Omega_{\odot}$. The magenta area shows extension of the stellar convective envelope, while the cyan area represents the extension of the stellar core.}
\label{Orbit_Evol}
\end{figure}
\section{Conclusion}\label{SecConc}
In this study, we have carried out a detailed comparative modelling of three exoplanet-host red giant stars. Each star is orbited by a long-period Jupiter-like planet that has been detected in radial velocities within the CASCADES survey (Ottoni et al. submitted) using the CORALIE spectrograph. We used the TESS photometric time series of each target to obtain seismic constraints. We extracted the individual radial and quadrupolar oscillation frequencies using the \texttt{PBjam} peakbagging package \citep{Nielsen2020} as well as the global seismic parameters $\left\langle \Delta \nu \right\rangle $ and $\nu_{\rm{max}}$, following the approach of \citet{Elsworth_2020}. Then we have analysed various approaches to determine the stellar fundamental parameters and computed the dynamical evolution of the system as well as the evaporation of the planetary atmospheres for each target, using the optimal seismic solution for the stellar properties.
The first step was to determine the stellar properties from the seismic scaling relations. As expected, a correct order of magnitude guess is obtained for this case, but the main issue is the strong dependency on the underlying correction of the scaling relations, which can lead to inaccurate results. We conclude that for the purpose of the detailed analysis carried out here, the seismic scaling relations do not offer a sufficient degree of accuracy and robustness. However, this does not imply that they cannot be useful for the analysis of large samples, for which extensive computations are either not justified or not accessible due to limitations of the seismic data. In that sense, using global seismic parameters such as $\left\langle \Delta \nu \right\rangle $ and $\nu_{\rm{max}}$, either in conjunction with model grids or with the scaling relations, still offers a uniform approach to do comparisons on large samples \citep[such as done for example by][for the case of the \textit{Gaia} offset analysis]{Khan2019}.
In a second modelling step, we carried out a fit of the individual radial modes in conjunction with non-seismic constraints using the AIMS software and a predetermined grid of stellar models. We have shown that these results are already more accurate. From this second step, we could determine a value for the mean density of each star by means of the SOLA inversion \citep{ReeseDens}, which could then be used both as an additional constraint and to derive a model-independent mass value when combined with \textit{Gaia} parallaxes and spectroscopic constraints. Such a constraint is of great interest for both the purpose of exoplanetology and also for even more advanced modelling aiming to exploit the information of dipolar oscillation modes.
Finally, the last step of our seismic modelling was carried out using a local minimisation algorithm and computing the stellar evolution models on the fly, fitting a series of seismic and non-seismic constraints. This allowed us to refine the fundamental properties of the star and to eliminate any remaining inconsistencies regarding the seismic and parallax-estimated radii for all targets. All the masses were consistent with the model-independent interval derived above, as a result of the constraints used in the minimisation. The evolutionary tracks found from this procedure were then used to study the dynamical evolution and the evaporation of the planetary atmospheres for each planet around the stars of our sample.
From this analysis, we have shown that no significant migration of the planets is expected, given their long periods. We have also shown that two out of three stars could not have had initial rotation velocities of $\rm 18~ \Omega_{\odot}$, as they would have then reached the critical velocity.
To summarise, we have demonstrated the importance of a thorough seismic modelling for the purpose of exoplanetology and illustrated its application in studying the fate of three long period planets around red giant stars. It is, however, evident that the modelling technique can be even more thorough than what is shown here should the data quality improve or additional strategies be developed in the future. In both cases, the conclusion that a refined approach is required to study the properties of exoplanetary systems in depth would still hold.
\section*{Acknowledgements}
We thank the referees for their suggestions and careful reading of the manuscript. G.B. acknowledges fundings from the SNF AMBIZIONE grant No. 185805 (Seismic inversions and modelling of transport processes in stars). C.P. acknowledges fundings from the Swiss National Science Foundation (project Interacting Stars, number 200020-172505). P.E. has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No 833925, project STAREX). WHB, GD and AL thanks the UK Science and Technology Facilities Council (STFC) for support under grant ST/R0023297/1. This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (CartographY GA. 804752). The research leading to this paper has received funding from the European Research Council (ERC grant agreement No. 772293 for the project ASTEROCHRONOMETRY). This article used an adapted version of InversionKit, a software developed within the HELAS and SPACEINN networks, funded by the European Commissions's Sixth and Seventh Framework Programmes.
|
\section{Introduction}
The theory of the strong interaction, Quantum Chromodynamics~(QCD), has a plethora of unknown properties and may offer a rich phase diagram that can only be revealed in the study of physical systems of very high baryon density and temperature. The details of the possible chiral and deconfinement QCD transitions are known, from first principle calculations, only in a narrow region close to vanishing baryon density~\cite{Borsanyi:2013bia,HotQCD:2014kol,Bazavov:2017dus} where both transitions appear as a smooth crossover.
To pin down the properties of hot and dense QCD matter has become the focus of various experimental programs ranging from laboratory experiments on earth to astrophysical data obtained from neutron star radii and binary neutron star mergers. Heavy ion collider programs at the BES at RHIC, NA61/Shine at CERN, CBM at GSI/FAIR, NICA in Russia and HIAF in China, and J-PARC-HI in Japan aim at studying the phase diagram of nearly isospin symmetric QCD matter at baryon densities of several times the nuclear matter ground state density and temperatures between 50 and more than 250 MeV.
Here, the emphasis is mostly on the exploration of the phase transition of QCD matter from a
confined hadronic phase to a phase where chiral symmetry is restored and quarks are eventually deconfined.
At the same time, complementary astrophysical observations of binary neutron star mergers and supernova explosions can also create (iso-spin asymmetric) matter of comparable density and temperatures up to $50$ MeV \cite{ourletter}.
Similarly, to the heavy ion collisions, cold neutron star matter is dynamically compressed and heated in binary neutron star mergers~(BNSM) which were experimentally detected by measuring gravitational waves~\cite{LIGOScientific:2017vwq,LIGOScientific:2020aai,Bauswein:2012ya,Most:2018eaw}.
While the systems created in such different scenarios vary in size over many orders of magnitude, they share a common unknown and defining property, the equation of state (EoS) of dense nuclear matter. Extracting the equation of state and its properties, like phase transitions or softest points, has been a defining challenge to many experimental and theoretical programs over the last decade.
To do so, model simulations that can incorporate different possible equations of state are compared to experimental observables. The challenge in such an approach is that many features not related to the EoS, such as microscopic properties, unknown transport parameters or boundary conditions, are not well constrained. In the current 'state-of-the-art' of ultra-relativistic heavy ion collisions, the dynamical evolution of the collisions is divided into roughly three phases \cite{Paiva:1996nv,Hama:2004rr,Bass:2000ib,Hirano:2001eu,Kolb:2003dz,Hirano:2004en,Hirano:2012kj,Nonaka:2006yn,Petersen:2008dd,Werner:2010aa,Gale:2013da,Shen:2014vra}. In this so-called 'standard model of ultra-relativistic heavy ion collisions', the reaction starts with an initial non-equilibrium phase where the kinetic energy of the two incoming nuclei lose a fraction of their longitudinal momentum and create a pre-equilibrium fireball. This phase is usually described by string models or QCD inspired non-equilibrium approaches, e.g. via a color glass condensate model or quantum kinetic theory~\cite{Iancu:2003xm,Gelis:2010nm,Albacete:2014fwa,Drescher:2006pi,Drescher:2006ca,Schenke:2012wb,Rybczynski:2013yba,Moreland:2014oya}.
Due to its violent non-equilibrium nature this phase of the reaction does generally not depend on the equation of state. After this energy deposition and a sufficient equilibration time, the near-equilibrium evolution can be described by (viscous) hydrodynamic or transport theoretical approaches. Here, an equation of state and transport properties of the medium can be included in the simulations. Finally once the system has hadronized, hadronic rescattering and the final freeze-out phase occurs \cite{Bass:2000ib,Steinheimer:2017vju}. As described above in this approach the EoS enters in the well defined equilibrium phase. Of course the applicable degrees of freedom vary with the collision energy.
However, this 'standard picture' is only well justified at very high beam energies, i.e. when the initial interpenetration time of the incoming nuclei is very short and can be well separated from the subsequent expansion. This is generally found to be the case of heavy ion collisions above $\sqrt{s_\mathrm{NN}}\approx 15$ GeV (corresponding to ${\mathrm{E_{lab}} \gtrapprox 100 A \mathrm{GeV}}$), where mainly the energy of the incoming nuclei is stopped while the baryon number of the participant nuclei is observed far from mid-rapidity. At significantly lower beam energies, the interpenetration time can last as long or even longer than the expansion phase. In such a scenario a large amount of the baryon number is stopped in the central collision region and a system of high baryon density is created around mid-rapidity.
The lower beam energies are exactly what is needed to study the EoS at the highest baryon densities. This also means that here the initial compression phase can not be separated from the expansion stage and the observables will therefore also be dependent on the equation of state in the initial compression phase. In particular this will be true in the presence of a phase transition.
It is therefore necessary to study the effects of the EoS on the initial compression at lower beam energies and also to devise new methods on how the dynamical evolution of such a system can be described. To achieve both of the above, a consistent treatment of the equation of state throughout the entire collision is necessary.
The challenges for the present paper are twofold:
\begin{itemize}
\item First, a new method is introduced on how a realistic chiral mean field equation of state (CMF-EoS) can be incorporated in a non-equilibrium Quantum-Molecular-Dynamics transport approach (UrQMD).
\item Second, the densities and temperatures achieved in this new approach (UrQMD-CMF) are compared with the evolution modeled with a relativistic (3+1) dimensional ideal fluid dynamical approach, where both approaches incorporate the same equation of state.
\end{itemize}
Then we ask I) how sensitive is the initial compression on the equation of state to the different assumptions made in both approaches (i.e. full local thermalization in contrast to (non-)equilibrium transport dynamics) and II) up to which beam energy is a simple modeling of the systems evolution within a one-fluid hydrodynamic model equivalent to that of a full microscopic transport simulation.
\section{Methods}
In the following section the models used in the paper are described. These include the hydrodynamic model with its initialization routine, and the microscopic transport model UrQMD. Finally, the equations of state employed in these models are introduced alongside with the formalism for their consistent implementation in the dynamical models.
\subsection{Hydrodynamic approach}\label{ref:hydro}
The full 3+1D evolution of a heavy ion collision can be simulated by (ideal) relativistic hydrodynamics \footnote{Currently we neglect viscous and dissipative effects since we are mainly interested in the bulk evolution of the system}. These equations describe the conservation of energy and momentum given by
\begin{equation}
\partial_{\mu} T^{\mu \nu} =0\ ,
\end{equation}
as well as the conservation of the baryon four current
\begin{equation}\label{consr}
\partial_{\mu} j^{\mu}=0 \ .
\end{equation}
In the following, the \texttt{SHASTA} algorithm~\cite{Boris:1973tjt,Rischke:1995ir} is used for the flux-corrected relativistic numerical solution of the above equations.
The equations are solved on a Cartesian $200\times200\times200$ grid with a cell size ${0.2\times0.2\times0.2~{\rm fm^{-3}}}$ and the time-step is fixed to ${\delta t=0.4\times0.2=0.08}$ fm/c. To close this set of hydrodynamical equations an equation of state is necessary. The EoS can be treated as a free input to the equations, which is provided by a table, using only the constraints of strangeness neutral $n_S=0$ matter with a charge to baryon fraction of $n_Q/n_B=0.4$.
The hydrodynamic initial state is given by two counter streaming Lorenz-contracted Wood-Saxon distributions of baryon charge $n_{WS}$ (and corresponding energy density) representing the two colliding (cold) nuclei:
\begin{equation}
n_{WS} = \gamma_{CM} \frac{n_0}{1+\exp{\left(\frac{\Delta r-R}{a}\right)}}\,.
\end{equation}
Here, $\gamma_{CM}$ is defined by the collision energy in the center of mass ($CM$) frame. $\Delta r$ is the distance from the nucleus center and is Lorentz contracted along the z-axis. The parameters of the WS distribution are the nuclear saturation density $n_0$ as well as the nuclei radius $R$ and the surface thickness $a$. The values of $R=6.6$~fm and $a=0.5$~fm correspond to the known properties of $Au$ nuclei. This initialization procedure, contrary to the 'standard' hybrid model, allows for a hydrodynamic treatment, with inclusion of the EoS, of both the early entropy production and of the expansion stages, which is important for low collision energies where the interpenetration times are of the same magnitude as the systems lifetimes.
\subsection{Microscopic Transport approach}\label{ref:transport}
The non-equilibrium microscopic description of the heavy ion collisions is done with the \texttt{UrQMD} transport approach~\cite{Bass:1998ca,Bleicher:1999xi}. \texttt{UrQMD} is based on the covariant propagation of hadrons on classical trajectories in combination with stochastic binary scatterings, color string formation and resonance excitation and decays. The hadrons interaction criteria are based on a geometric interpretation of their scattering cross section. The cross sections for these scatterings are taken either from experimental measurements where available \cite{ParticleDataGroup:2020ssz}, or are calculated e.g. from the principle of detailed balance.
In its default setup the model corresponds to a hadronic cascade and can be readily used to describe the final state spectra of hadrons over a wide range of beam energies. It was shown that the effective equation of state of the \texttt{UrQMD} 'cascade model' corresponds to a Hadron Resonance Gas (HRG) with the respective degrees of freedom \cite{Bravina:2008ra}.
Extending the equations of motion to non-trivial hadronic interactions, and consequently to any possible equation of state, is not straightforward.
Early, a non-relativistic QMD approach~\cite{Aichelin:1991xy} was developed to incorporate a density dependent Skyrme interaction~\cite{konopka}.
In this QMD part of the \texttt{UrQMD} model, the change of momenta of the baryons, due to a density dependent potential, is calculated using the non-relativistic equations of motion:
\begin{eqnarray}\label{motion}
\dot{\textbf{r}}_{i}=\frac{\partial \langle H \rangle}{\partial\textbf{p}_{i}},
\quad \dot{\textbf{p}}_{i}=-\frac{\partial \langle H \rangle}{\partial \textbf{r}_{i}},
\end{eqnarray}
where {\it $\langle H \rangle$} is the total Hamiltonian function of the system. The Hamiltonian of each baryon, $H_i=E^{\mathrm{kin}}_i + V_i$, comprises the kinetic energy and the mean field potential energy $V_i=E_{\mathrm{field}}/A$ of the baryon $i$. The mean field potential energy per baryon can be related to a density dependent single particle energy:
\begin{equation}
U_i(n_B)=\frac{\partial (n_B \cdot V_i(n_B))}{\partial n_B}\,.
\end{equation}
In the Skyrme \texttt{UrQMD} approach \cite{Aichelin:1991xy,konopka,Bass:1998ca} the density dependence of the single particle energy for all baryons is given by a simple form:
\begin{equation}
\label{eq:usky}
U_{\mathrm{Skyrme}}(n_B)= \alpha (n_B/n_0) + \beta (n_B/n_0)^{\gamma}\,.
\end{equation}
Two out of the three parameters ($\alpha$, $\beta$ and $\gamma$) are usually constraint by the nuclear matter saturation density and binding energy, while the remaining unconstrained property is the nuclear incompressibility, defining the so-called stiffness of the EoS.
Such a single free parameter approach to describe the equation of state of dense QCD matter has a significant shortcoming: the equation of state for densities above nuclear saturation is fixed by parameters which are defined solely at saturation density.
A similar problem occurs when a purely nuclear relativistic mean field model is implemented in QMD \cite{Nara:2019qfd,Nara:2020ztb}, although such an approach does also allow for the inclusion of additional degrees of freedom and thus a more complex phase structure.
Recently, another idea has been put forward where additional terms are added in equation eq.(\ref{eq:usky}) which allow for describing non-trivial features like a phase transition in the potential \cite{Sorensen:2020ygf}. However, this density functional approach suffers from a serious problem characteristic to the Skyrme potential: the speed of sound of this EoS eventually becomes superluminal at large baryon densities, even at $T=0$. Below we introduce a different way to replace the limited Skyrme potential by a, more realistic, density dependent equation of state.
Once the mean field potential is known, the change of momentum of each baryon in accord with Hamiltons equations of motion can be calculated as:
\begin{eqnarray}
\dot{\textbf{p}}_{i}=-\frac{\partial H }{\partial \textbf{r}_{i}} = -\frac{ \partial{ V(n_B) }}{\partial n_B}\cdot\frac{\partial n_B({\bf r}_i)}{\partial \textbf{r}_{i}}\,.
\end{eqnarray}
Besides the derivative of the mean field potential energy, only the local density and its gradient is required for each baryon.\footnote{Ignoring a possible momentum dependence of the potential.} This is calculated by assuming that each particle can be treated as a Gaussian wave packet~\cite{Aichelin:1991xy,Bass:1998ca}. With such an assumption, the local interaction baryon density $n_B(r_i)$ at location ${\bf r}_i$ of the $i$-th particle in the computational frame is:
\begin{equation}
n_B(r_i) = \left(\frac{\alpha}{\pi}\right)^{3/2}\sum_{j,\,j\neq i} B_j \exp{\left(-\alpha({\bf r}_i-{\bf r}_j)^2\right)} \, ,
\end{equation}
where $\alpha=\frac{1}{2 L^2}$, with $L=\sqrt{2}$ fm, is the effective range of the interaction. The summation runs over all baryons, $B_j$ is the baryon charge of the $j$-th baryon.
In the following, the QMD implementation will assume, for simplicity, that the mean field potential for all baryon types is the same as for the nucleons.
\subsection{Equations of state}
The present paper aims at estimating the role of a realistic and cosnsitent equation of state for the compression in heavy ion collisions. The two approaches to simulating heavy ion dynamics, introduced in sections \ref{ref:hydro} and \ref{ref:transport}, allow to incorporate the interactions via an EoS. In the following we concentrate on two EoSs, a simple ideal HRG, and a realistic Chiral Mean Field~(CMF) model, which incorporates all interactions essential for a realistic description of nuclear matter, neutron stars, and hot QCD matter.
{\it The HRG model.} The Hadron Resonance Gas model is an approximation to confined hadronic QCD matter~\cite{Dashen:1969ep}. It is based on the assumption that a gas of interacting hadrons
can be described (if the width of the resonances is smaller than the temperature) by inclusion of all hadron species and their resonances as explicit degrees of freedom in the partition function. This partition function then mimics the basic thermodynamic properties of QCD at low temperatures and small densities. The HRG was shown to successfully describe the properties of lattice QCD thermodynamics below the chiral transition \cite{Karsch:2003zq,Huovinen:2009yb,Borsanyi:2010bp,Ratti:2010kj,Alba:2015iva,Bazavov:2012jq,Huovinen:2017ogf,Vovchenko:2014pka}. However, due to a lack of many-body and long-range interactions the model is not able to describe basic features of QCD phenomenology such as a bound nuclear ground state or deconfinement.
Multiple extensions of the model have been developed over the years (for a survey, see~\cite{Walecka:1974qa,Dutra:2012mb,Dutra:2014qga,Vovchenko:2015vxa,Vovchenko:2016rkn,Vovchenko:2017cbu,Poberezhnyuk:2017yhx}). However, all modifications have similar shortcomings as the few-parameter description of the EoS in the Skyrme model.
As discussed earlier the \texttt{UrQMD} model in the cascade mode will have an equilibrium state that is equivalent to the HRG model description of QCD matter~\cite{Bravina:2008ra}. As in the cascade mode of \texttt{UrQMD} only elastic scatterings and resonance excitations occur\footnote{Here string formation is omitted.}, the HRG is a good approximation for the effective EoS of the model.
A comparison of the hydro simulations with a HRG equation of state and \texttt{UrQMD} in cascade mode has been used as a reference to study the effects of instant equilibration on the dynamics \cite{Petersen:2008dd}. Note, it is the cascade mode that is commonly used to calculate the initial compression phase in the prevalent hybrid models of heavy ion collisions \cite{Steinheimer:2007iy}.
{\it CMF model.} The Chiral Mean Field model~\cite{Papazoglou:1998vr,Steinheimer:2010ib,Motornenko:2019arp} is an approach for the description of QCD thermodynamics for a wide range of temperatures and densities. The effective degrees of freedom of the CMF model include a complete list of known hadrons as well as the three light quark flavors plus a gluon contribution. The CMF contains the transition between quarks and hadronic degrees of freedom, the liquid vapor transition in nuclear matter, as well as chiral symmetry restoration are driven by mean fields. Parity doubling introduces heavy parity partners to the baryons of the lowest octet~\cite{Steinheimer:2011ea,Aarts:2018glk}. The baryons and their parity partners interact via mesonic mean fields (attractive scalar $\sigma,~\zeta$ and repulsive $\omega,~\rho,~\phi$ meson exchanges). The effective masses of the parity partners depend on the chiral fields, therefore the partners become mass-degenerate as the chiral symmetry is restored. A detailed description of the CMF model with its parameters can be found in \cite{Motornenko:2020yme}.
The CMF model describes many aspects of QCD phenomenology. It has been successfully applied in an analysis of lattice QCD data~\cite{Motornenko:2020yme}, the description of cold neutron stars~\cite{Motornenko:2019arp}, and has been employed as EoS in the hydrodynamic simulations of both heavy ion collisions and binary neutron star mergers~\cite{Seck:2020qbx,ourletter}.
The effective masses of the ground-state octet baryons and their parity partners (assuming isospin symmetry) read~\cite{Steinheimer:2011ea}:
\begin{eqnarray}
m^*_{b\pm} &=& \sqrt{ \left[ (g^{(1)}_{\sigma b} \sigma + g^{(1)}_{\zeta b} \zeta )^2 + (m_0+n_s m_s)^2 \right]} \nonumber \\
& \pm & g^{(2)}_{\sigma b} \sigma \ ,
\label{eq:mass}
\end{eqnarray}
where the various coupling constants $g^{(*)}_{*b}$ are determined by vacuum masses and by nuclear matter properties. $m_0$ refers to a bare mass term of the baryons which is not generated by the breaking of chiral symmetry, and $n_s m_s$ is the ${\rm SU}(3)_f$-breaking mass term that generates an explicit mass corresponding to the strangeness $n_s$ of the baryon. The single-particle energy of the baryons, therefore, becomes a function of their momentum $k$ and effective masses: $E^*=\sqrt{k^2+m_b^{* 2}}$.
Similar to the effective mass $m_b*$ which is modified by the scalar interactions, the vector interactions lead to a modification of the effective chemical potentials for the baryons and their parity partners:
\begin{equation}
\mu^*_b=\mu_b-g_{\omega b} \omega-g_{\phi b} \phi-g_{\rho b} \rho\,.
\label{eq:chem}
\end{equation}
Note that the couplings of nucleons and hyperons to the mean fields were fixed to reproduce nuclear binding energies $E_0/B\approx-15.2$~MeV as well as the asymmetry energy $S_0\approx31.9$~MeV, and incompressibility $K_0\approx267$~MeV.
The phase diagram of the CMF model includes three critical regions: the nuclear liquid-vapor phase transition, chiral symmetry restoration, and the transition to quark matter~\cite{Motornenko:2019arp}. The model predicts two first-order phase transitions. The first one is associated with the nuclear liquid-vapor phase transition at $n_B\sim n_0$. The second one appears at about four times the normal nuclear density $4 n_0$ due to the chiral symmetry restoration. This chiral transition however shows only a small latent heat and the critical endpoint of this transition occurs already at $T_{\rm CP}\approx 17$ MeV.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{hydro_eos.pdf}
\caption{(Color online) The CMF equation of state, represented as pressure as a function of the baryon density, for different values of constant entropy per baryon (S/A). The black solid line corresponds to the CMF-EoS at zero temperature, where around 4$n_0$ a small kink in the pressure due to the phase transition can be observed. The same relation for the HRG-EoS at vanishing temperature is also shown as a grey line. The pressure in the HRG is substantially lower than in the CMF model.}
\label{fig:eos_comp}
\end{figure}
The CMF-EoS along different trajectories of fixed entropy per baryon is shown in Fig. \ref{fig:eos_comp}. This depiction is useful since one can see several relevant features in the CMF EoS. First, along the curve at zero entropy per baryon ($T=0$) a small kink in the pressure is observed which signals a very weak phase transition around four times saturation density. This kink disappears at higher entropies per baryon. Secondly for values of $S/A$ up to 10, the pressure only very mildly depends on the finite temperature and is dominated by the density dependence \footnote{Which supports our approach of assuming a mainly density dependent EoS in the implementation in UrQMD later.}. Finally, we also show the $T=0$ EoS in the HRG model as a grey line compared to the corresponding black line of the CMF. The CMF shows clearly a much larger pressure due to the mean field interactions, which will lead to observable effects in the dynamic simulations.
In the CMF model the single nucleon potential is given by the interactions with the chiral and repulsive mean fields. At $T=0$, in the CMF model, it can be calculated from the self energy of the nucleons as:
\begin{equation}
\label{eq:cmfnp}
\mathrm{U}_{CMF}= m_{N}^{*} - m_{N}^{vac} - \mu_{N}^{*} + \mu_{N} \,,
\end{equation}
where $m_{N}^{vac}$ and $\mu_{N}$ are the vacuum mass and chemical potential of the nucleon calculated only from the charge constraints and $m_{N}^{*}$ and $\mu_{N}^{*}$ are the corresponding effective nucleon mass (eq.(\ref{eq:mass})) and effective chemical potential (eq.(\ref{eq:chem})) generated through the interactions with the scalar and vector mean fields.
To set the stage, the CMF potential $\mathrm{U}_{CMF}$ is shown in panel (a) of Fig. \ref{fig:skyrme-pot} where we contrast the CMF single particle potential $\mathrm{U}_{CMF}$, as a function of baryon density $n_B$ in units of the ground-state baryon density, with two different Skyrme potentials $\mathrm{U}_{Sky}$ (resulting in different equations of state). We show the Chiral mean field EoS (full orange line) in comparison to the well known hard Skyrme potential (dotted red line) and the soft Skyrme potential (dotted green line).
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{nucl_opot.pdf}
\caption{(Color online) (a) Nuclear potential $V$ as a function of baryon density $n_B$ in units of the ground-state baryon density for three different potentials (resulting in a different equations of state).
(b) The resulting field energy per baryon $E_{\mathrm{field}}/A$ for the three different equations of state.
(c) Derivative of the field energy per nucleon with respect to the baryon density as a function of baryon density $\rho_B$ in units of the ground-state baryons density for three different potentials
We show the Chiral mean field EoS (full orange line) in comparison to the well known hard Skyrme potential (dotted red line) and the soft Skyrme potential ( dotted green line).
}
\label{fig:skyrme-pot}
\end{figure}
\subsubsection{The CMF EoS in UrQMD}
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{rho_time.pdf}
\caption{(Color online) Time evolution of the baryon density in the central volume of the reaction for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of coarse grained \texttt{UrQMD} simulations and the dashed line shows the results for one-fluid (3+1) dimensional hydrodynamic calculations for the same systems and energies. The green lines in the upper row are calculated using the hadron resonance gas EoS in hydro and for the conversion from $(\varepsilon,\rho_B)$ to the thermodynamic quantities, while the red lines in the lower row show the results for the CMF-EoS.
}
\label{fig:rho-time}
\end{figure*}
To implement the CMF-EoS in the QMD part of the \texttt{UrQMD} model we essentially need to calculate the effective field energy per baryon of any particular model which can then be used in the QMD equations of motion.
In the CMF model the nucleons interaction is described relativistically via scalar and vector mean fields which are not present in UrQMD. In addition, the CMF model is not only restricted to nucleons, thus, the single nucleon potential $\mathrm{U}_{CMF}$ as defined in eq. (\ref{eq:cmfnp}) is not suitable to calculate the relevant mean field potential that is required for the equations of motion.
Fortunately, the effective field energy per baryon $E_{\mathrm{field}}/A$ calculated from the CMF model can be used, i.e. the relevant quantity which enters the equations of motion:
\begin{equation}
V_{CMF}=E_{\mathrm{field}}/A=E_{\mathrm{CMF}}/A - E_{\mathrm{FFG}}/A\,,
\end{equation}
where $E_{\mathrm{CMF}}/A$ is the total energy per baryon at $T=0$ from the CMF model and $E_{\mathrm{FFG}}/A$ is the energy per baryon from a free non-interacting Fermi-gas.
The resulting effective field energy per baryon, as a function of the baryon density, from the CMF model is shown as a solid line in panel (b) of figure \ref{fig:skyrme-pot}, again compared to the known curves from the hard and soft Skyrme EoS.
Finally, panel (c) of Fig. \ref{fig:skyrme-pot} shows the derivative of the field energy per nucleon with respect to the baryon density as a function of baryon density $n_B$ in units of the ground-state baryon density for the three different potentials.
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{T_time.pdf}
\caption{(Color online) Time evolution of the average Temperature in the central volume of the reaction for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of coarse grained \texttt{UrQMD} simulations and the dashed line shows the results for one-fluid (3+1) dimensional hydrodynamic calculations for the same systems and energies. The green lines in the upper row are calculated using the hadron resonance gas EoS in hydro and for the conversion from $(\varepsilon,\rho_B)$ to the thermodynamic quantities, while the red lines in the lower row show the results for the CMF EoS.
In the coarse-graining procedure, only participants are used for the averaging, so the Temperature appears to 'jump' to a finite value.}
\label{fig:T-time}
\end{figure*}
What can be observed is that the CMF-EoS shows a behavior similar to that of the soft Skyrme potential for sub-saturation (up to saturation) density, then becomes even stiffer than the hard Skyrme potential and finally shows a significant softening compared to the hard Skyrme which essentially becomes superluminal at large densities. Around four times nuclear saturation density the CMF-EoS shows a small kink in the derivative of the field energy per baryon which is due to the weak chiral phase transition at $T=0$. Since this transition is only very weak we expect no significant effects of this transition on the dynamic evolution, in particular at finite temperatures where the kink will be smeared out by the thermal energy.
Regarding its phase structure, the CMF model has several appealing features:
\begin{enumerate}
\item A nuclear incompressibility compatible with experimental observations.
\item A stiff super-saturation nuclear equation of state which is required to explain astrophysical observations.
\item A "softening" of the equation of state at even higher densities due to the slow approach to the high density limit of a free gas of three quark flavors.
\end{enumerate}
Having now established a method in which any equation of state can be easily introduced in the QMD part of UrQMD, we will first study the dynamic evolution of bulk properties and their dependence on the EoS in the two dynamical approaches, hydrodynamics and microscopic transport.
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{t_rho_pd.pdf}
\caption{(Color online) Expansion trajectories (along the time evolution in the central cell) in the temperature-baryons density plane for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of coarse grained \texttt{UrQMD} simulations and the dashed line shows the results for one-fluid (3+1) dimensional hydrodynamic calculations for the same systems and energies. The green lines show the results using the hadron resonance gas EoS for the conversion from $(\varepsilon,n_B)$ to the thermodynamic quantities, while the red lines show the results for the CMF-EoS.
}
\label{fig:trho_isen}
\end{figure*}
\section{Results on the bulk evolution properties}
In this first work we will focus solely on bulk properties of the fireball created in central heavy ion collisions to establish the gross features of the CMF-EoS in the transport and hydrodynamic simulations. While it is known that other observables like radial flow and its higher moments can be very sensitive to the equation of state, the main focus of the present work is to establish the effect of the equation of state on the system created in different dynamical implementations of the same EoS. Studies of flow, cluster production as well as correlations and fluctuations will be explored in detail in future investigations.
In the hydrodynamic model the expected properties of matter at different beam energies can be extracted in a straightforward way. Here, only one single event per beam energy, with impact parameter $b=2$ fm, is sufficient to average the thermodynamic properties in the central volume (a cubic volume of length $l=2$fm) of central Au+Au collisions at various beam energies. The local energy density and net baryon density are explicitly propagated in the hydrodynamic framework and quantities like the Temperature, pressure as well as entropy density can be directly and unambiguously related to these volume averaged densities via the equation of state.
In the microscopic transport treatment, the equivalent expectation values for the local energy and baryon number densities can also be calculated by a coarse graining procedure \cite{Endres:2015fna}. In this procedure, a large number of events of a given beam energy and centrality are calculated and the total energy and baryon density in the central volume, a cube of length 2~fm, can be calculated as sum of the energy and net baryon charge of participants in that volume. In this study, for a given beam energy, we use 1000 events with impact parameter less than 3.4 fm to perform the coarse graining. To extract the thermodynamic quantities like Temperature, pressure and entropy density a mapping to the effective equation of state, that is used, is necessary. In our simulations this mapping is done by using either the HRG-EoS (for the \texttt{UrQMD} cascade simulations) or the CMF-EoS (for the corresponding CMF-UrQMD simulation). Note that this procedure assumes that the system is close to local equilibrium which is not necessarily the case in the \texttt{UrQMD} transport model, especially at very early and late times. Thus the extracted values for the temperature, pressure and entropy density may (and as we later see will) vary due to deviations from equilibrium.
We begin in Fig. \ref{fig:rho-time} with a comparison of the time evolution of the baryon density in the central volume of the reaction for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of the coarse grained \texttt{UrQMD} simulations, the dashed lines show the results for the one-fluid 3+1 dimensional hydrodynamic calculations for the same systems and energies. The green lines show the results using the HRG-EoS in hydro and for the coarse-graining conversion from $(\varepsilon,\rho_B)$ to the thermodynamic quantities, while the red lines show the results for the CMF-EoS.
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{p_rho.pdf}
\caption{(Color online) Time evolution of the pressure in units of the ground-state energy density in the central cell of the reaction for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of coarse grained \texttt{UrQMD} simulations and the dashed line shows the results for one-fluid (3+1) dimensional hydrodynamic calculations for the same systems and energies. The green lines show the results using the hadron resonance gas EoS for the conversion from $(\varepsilon,\rho_B)$ to the thermodynamic quantities, while the red lines show the results for the CMF-EoS.
}
\label{fig:prho}
\end{figure*}
In the time evolution of the baryon density in figure \ref{fig:rho-time} one can clearly observe that the full hydrodynamic simulation and the transport simulation with the CMF-EoS give almost identical results up to the highest beam energies. Only for the beam energy of $E_\mathrm{lab}= 10.0 A$ GeV the transport simulation yields a smaller compression due to the effect of transparency which cannot be described in a 1-fluid simulation (although it is known that 3-fluid models can reproduce this effect). In the case of the HRG-EoS compared to the cascade model, the agreement of the density evolution is not as exact, yet still comparable. This is somewhat expected since the EoS in the cascade mode is not explicitly introduced and enters only implicitly by the set of degrees of freedoms. The fact that the transport model and the hydrodynamic model agree so well in the compression is not a trivial result but shows clearly, that the maximally reached compression, for low beam energies where transparency can be neglected, does to first order depend on the work that needs to be done against the pressure of the compressed system. The compression reached varies drastically, by almost a factor of 2, between the two equations of state used. This finding has important consequences since it means any observable that is sensitive to the maximal compression reached in these collisions would be a very good and almost model independent messenger for the equation of state of dense QCD matter.
Figure \ref{fig:T-time} shows the time evolution of the equilibrium temperature in the central volume of the same reactions. The colors and line styles are the same as in figure \ref{fig:rho-time}.
In the comparison of the (equilibrium) temperature, the differences between the hydro and transport approaches are more obvious. In particular at very early times, the temperature in the coarse grained transport simulations is significantly larger than in the fluid simulation. This can be understood as an effect of the non-equilibrium state of the microscopic transport which is then mapped on an equilibrium temperature. The effect of the non-equilibrium is not observed in the baryon density which is explicitly propagated and conserved in both approaches but in the temperature which is inferred implicitly. Nevertheless, after a few fm/c, even in the non-equilibrium approach the temperatures reached agree within 5-10 MeV. Notably the HRG simulations give also a systematically larger temperature, however the increase as compared to the CMF is only on the order of $~10$ MeV.
Having extracted the time evolution of the baryon density as well as the effective temperature, the expansion dynamics of the systems studied can also be depicted in the T-$n_B$ phase diagram.
Therefore, we explore the expansion trajectories along the time evolution in the central cell in Fig. \ref{fig:trho_isen} in the temperature-baryon density plane for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of coarse grained \texttt{UrQMD} simulations and the dashed line shows the results for one-fluid (3+1) dimensional hydrodynamic calculations for the same systems and energies. The green lines show the results using the HRG-EoS for the conversation from $(\varepsilon,n_B)$ to the thermodynamic quantities, while the red lines show the results for the CMF-EoS. Note that for this comparison, we start the trajectories at the point of largest compression after which, in the case of the ideal hydrodynamics, they follow lines of constant entropy per baryon.
The most significant difference is that the HRG curves start at a much larger density. On the other hand, the trajectories become very close at lower densities. This means that at the time that the systems reach freeze out, at $n_B \leq n_0$, the thermodynamic conditions are very similar for the different models and equations of state.
Much of the compression as well as consecutive expansion of the system strongly depends on the pressure reached during the initial phase. As we have seen a higher pressure in the EoS (harder EoS) leads to smaller densities. On the other hand the amount of radial as well as directed and elliptic flow produced will depend on the pressure which drives the expansion stage.
Finally, we show in Fig. \ref{fig:prho} the time evolution of the pressure in units of the ground-state energy density in the central cell of the reaction for central Au+Au reactions at $E_\mathrm{lab}=1.23, 2.0, 4.0, 6.0, 10.0 A$ GeV (from left to right). The full lines show the results of coarse grained \texttt{UrQMD} simulations and the dashed line show the results for one-fluid (3+1) dimensional hydrodynamic calculations for the same systems and energies. The green lines show the results using the HRG-EoS, while the red lines show the results for the CMF-EoS. Again, the extraction of the pressure from the local densities $(\varepsilon,n_B)$ is straightforward in the hydro model while for the coarse grained approach we assume local equilibrium and isotropic pressure which allows us to read of the effective pressure from the EoS table as described above.
Most notably is that the maximum pressure is reached at different times, depending on the equation of state used. This is due to the maximum density also being reached at different times as shown in figure \ref{fig:rho-time}. This different time dependence of the pressure evolution is likely to have significant consequences on the generated flow, which we will study in detail in a forthcoming publication.
\subsection{Entropy production}
As shown in figure \ref{fig:trho_isen} the expansion in both the hydro and transport models follows approximately the same isentropic trajectories. However, we expect that the final entropy per baryon will be different in the two approaches since the microscopic transport has a finite viscosity (shear and bulk) and the system will be only in partial chemical equilibrium at best.
To complete the comparison, Fig. \ref{fig:subera} depicts the entropy production per baryon as a function of beam energy for central Au+Au reactions in the energy range from $E_\mathrm{lab}=1.23A$ GeV to $10 A$ GeV. The lines denote calculations using the coarse grained \texttt{UrQMD} model with CMF-EoS (full red line), the 3+1D one-fluid hydrodynamics calculation (dotted red line) and the one-dimensional relativistic shock model, i.e. the Taub adiabate (dashed grey line). For the hydrodynamics as well as the \texttt{UrQMD} coarse grained simulations, the entropy was extracted from the CMF model implicitly, as described for the temperature above, knowing the local energy and baryon densities. In the case of the hydro simulation, S/A as a function of time is essentially a constant throughout the expansion stage. In the transport simulation it only shows a slight increase. Here, we compare the values of S/A at the end of the expansion i.e. when the density drops below $n_B=n_0$.
The full 3+1D ideal hydrodynamic simulation produces almost exactly the same entropy per baryon as the analytic 1-D shock solution (Taub adiabat). In general the entropy per baryon in the hydrodynamic case is smaller than in the non-equilibrium transport which is expected. The difference between these two scenarios grows with increasing energy which is also expected from the increasing transparency which leaves a smaller baryon number in the center of the collision zone. Furthermore, it is known that the system at late times can only be describes as being in partial chemical equilibrium. Mapping such a system onto an equilibrium EoS to calculate the entropy per baryon will yield larger values of the effective (equilibrium) S/A.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{sovera.pdf}
\caption{(Color online) Entropy production per baryon as a function of beam energy for central Au+Au reactions in the energy range from $E_\mathrm{lab}=1.23A$ GeV to $10 A$ GeV. The lines denote calculations using the coarse grained \texttt{UrQMD} model with CMF EoS (full red line), a one-fluid hydrodynamics calculation (dotted red line) and the one-dimensional shock model, i.e. the Taub adiabate (dashed grey line).
}
\label{fig:subera}
\end{figure}
\section{Conclusion}
A method was introduced that enables us to implement any density dependent equation of state in the QMD part of the \texttt{UrQMD} model. It was shown that for low beam energies $E_\mathrm{lab} \leq 6A$ GeV, the bulk evolution of the density in this new description agrees very well with a relativistic 1-fluid simulation with the same equation of state. The effective temperature from the \texttt{UrQMD} simulation is slightly increased compared to the ideal hydrodynamic model due to non-equilibrium effects.
Our results highlight the importance of the equation of state for the initial compression phase in nuclear collisions at low beam energies and, at the same time, provide a method on how it can be introduced in a consistent manner.
It was also shown that the expansion in both models follows closely an almost isentropic expansion as expected for the corresponding EoS.
In the present study the CMF equation of state was used for both the hydro and \texttt{UrQMD} evolution. It describes properties of dense nuclear matter, astrophysical observables as well lattice QCD thermodynamics and includes a transition from hadronic to quark degrees of freedom.
The total entropy per baryon produced in both scenarios was compared and it was found that the \texttt{UrQMD} model shows a slightly increased entropy production at low beam energies due to the non-equilibrium nature of the transport simulation. The entropy per baryon at the highest beam energies under investigation are significantly higher in the transport model due to the expected baryon transparency which transports the net baryon number away from mid-rapidity. However, the application of the presented approach is questionable for higher beam energies, where the stopping and energy deposition is dominated by partonic interactions (e.g. strings) for which the baryonic mean field QMD approach is not a suitable description.
Having established this new method enables us to now study the effect of different possible equations of state within the microscopic transport approach. The only assumptions are that the EoS is dominated by its density dependence i.e. fermions, and the effective potentials which govern the interactions have only a mild explicit temperature dependence. One should note that the density dependent forces that are assumed, are independent of the degree of equilibration reached throughout the collision.
In the future this method can readily be extended to include also a strong first order phase transition. This consistent description of the whole collision that does not require any ad-hoc matching of different phases will allow us to study possible observable signals of this transition in heavy ion collisions.
\begin{acknowledgement}
The authors thank Volker Koch for insightful discussions and
comments related to this work.
AM acknowledges the Stern-Gerlach Postdoctoral fellowship of the Stiftung
Polytechnische Gesellschaft. J.S. and MOK thanks the Samson AG and J.S. the BMBF through the ErUM-Data project for funding. MOK thanks GSI for funding. This work was supported by a PPP program of the DAAD.
MB acknowledges support by the EU-STRONG 2020 network.
Y.N. acknowledges support from JSPS KAKENHI Grant Number JP21K03577.
The computational resources for this project were provided by the Center for Scientific Computing of the GU Frankfurt and the Goethe-HLR.
\end{acknowledgement}
|
\section{Introduction}
A possibility to extract the top-quark mass $m_t$ in a well defined mass renormalization scheme is given by the analysis of the normalized $\rho_{\mathrm{s}}$ distribution, as first discussed in~\cite{Alioli:2013}, which is defined as $(1/{\sigma_{t\bar{t}j}}) \, {d \sigma_{t\bar{t}j}}/{d \rho_\mathrm{s}}$, with $\rho_s = {2 m_0}/{m_{t\bar{t}j}}$ and $m_0 = 170\,\mathrm{GeV}.$ \par
A recent extraction of the ATLAS collaboration \cite{ATLAS:2019JHEP} of $m_t$ in $t\bar{t}j$ hadroproduction using the $\rho_{\mathrm{s}}$ distribution led to an on-shell value of $m_t^{\mathrm{pole}} = 171.1 \pm 0.4 (\mathrm{st}) \pm 0.9 (\mathrm{sy})\, ^{+0.7}_{-0.3}(\mathrm{th})\,\mathrm{GeV},$
in which the theory uncertainty is sizeable. Thereby, the scale uncertainty dominates the theory uncertainty and contributes to an uncertainty on $m_t^{\mathrm{pole}}$ with $^{+0.6}_{-0.2}\,$GeV, while the PDF and $\alpha_S$ uncertainty amounts to $\pm 0.2\,$GeV.\par
In \cite{Bevilacqua:2016} a better description of the high energy tails of differential NLO distributions of the process $pp \rightarrow t\bar{t}j$ with fully leptonic top-quark decay was found when using dynamical renormalization and factorization scales ($\mu_R,\mu_F$) instead of a static scale $\mu_0=m_t$. We therefore present a study of the scale and parton distribution function (PDF) uncertainty using a static and two dynamical scales, with more inclusive cuts in contrast to \cite{Bevilacqua:2016}, similar to those recently adopted by the experimental collaborations, and a special focus on the discussed $\rho_{\mathrm{s}}$ distribution.
\section{Theory uncertainty in the $\rho_{\mathrm{s}}$ distribution}
The NLO differential cross sections of the $pp \rightarrow t\bar{t}j$ process at $\sqrt{s}=13\,$TeV for $m_t^{\mathrm{pole}}=172\,$GeV presented in the following were obtained with the \texttt{ttbarj V2} implementation in the POWHEG-BOX \cite{Alioli:2010}. In contrast to the previous \texttt{ttbarj V1} implementation \cite{Alioli:2012}, in \texttt{ttbarj V2} all hard scattering amplitudes are calculated with \texttt{OpenLoops2} and the calculation can be parallelized. During the analysis at least one jet satisfying the kinematic cuts $p_T^j > 30\,$GeV and $|\eta_j|<2.4$ was required, with jets reconstructed using the anti-$k_T$ jet clustering algorithm with $R=0.4$. The default PDFs and $\alpha_S(M_Z)$ value were taken from the CT18NLO PDF set. Besides the static scale $\mu_0=m_t^{\mathrm{pole}}$ also the two dynamical scales $\mu_0=H_T^B/2$ and $H_T^B/4$ were investigated defined through $H_T^B = \sqrt{{p_{T,t}^B}^2 + {(m_t^{\mathrm{pole}})}^2} + \sqrt{{p_{T,\bar{t}}^B}^2 + {(m_t^{\mathrm{pole}})}^2} + p_{T,j}^B$. The superscript $B$ means that the kinematic variables are evaluated at the underlying Born level in the POWHEG-BOX.\\
In Fig.~\ref{fig:scalevar_3panels} the $\rho_{\mathrm{s}}$ distribution and the corresponding seven point scale variation graphs, using the three scale definitions described above, are shown explicitly. The graphs are obtained from $(\mu_R,\mu_F) = (K_R,K_F)\mu_0$ by varying $K_R, K_F \in \{ 0.5, 1, 2\}$, leaving out the extreme combinations $(0.5,2)$ and $(2,0.5)$. In fact, in case of the static scale, for low values of $\rho_{\mathrm{s}}$, which correspond to large values of $m_{t\bar{t}j}$ and as such to the high-energy region, the description of the $\rho_s$ distribution seems to be unreliable and the scale uncertainty increases rapidly. Further, a crossing of the graphs obtained with different values of $K_R, K_F$ is seen in the interval $0.1 \lesssim \rho_{\mathrm{s}} \lesssim 0.3$. These distinctive features are not observed using either dynamical scale and the scale variation does not induce such a large shape variation as seen in the static scale case.
\begin{figure}[!h!tbp]
\centering
\includegraphics[width=\textwidth]{plots/rho_scalevar_3panels.pdf}
\caption{NLO prediction of the $\rho_\mathrm{s}$ distribution and the seven point scale variation using as nominal $\mu_R,\mu_F$ the scales $\mu_0=m_t,\, H_T^B/2$ and $H_T^B/4$ (from left to right).}
\label{fig:scalevar_3panels}
\end{figure}
This leads to the observation that the scale uncertainty in the normalized $\rho_{\mathrm{s}}$ distribution is clearly reduced using a dynamical instead of the static scale, as shown in Fig.~\ref{fig:scalevar_3panels_norm}.
\begin{figure}[!h!tbp]
\centering
\includegraphics[width=\textwidth]{plots/rho_scalevar_3panels_norm.pdf}
\caption{Same as in Fig.~\ref{fig:scalevar_3panels}, but for the normalized $\rho_\mathrm{s}$ distribution.}
\label{fig:scalevar_3panels_norm}
\end{figure}
Further investigation of the scale uncertainty showed that using a dynamical scale with either $R=0.4$ or $R=0.8$ leads to a similar scale uncertainty, while in the static scale case the scale uncertainty can be reduced in the region of low $\rho_{\mathrm{s}}$ using $R=0.8$. Comparing the NLO and LO scale variation uncertainty bands, these overlap over the whole region of the $\rho_{\mathrm{s}}$ distribution using either dynamical scale, while in the predictions obtained with the static scale, the NLO and LO scale variation bands start to depart from each other in the high-energy region.\\
Furthermore, we investigated the approximate NLO PDF uncertainty in the $\rho_\mathrm{s}$ distribution using four different modern NLO PDF sets CT18NLO, ABMP16, MSHT20 and NNPDF3.1 and the dynamical scale $\mu_0=H_T^B/4$.
In the bulk of the $\rho_\mathrm{s}$ distribution the predictions obtained with the different PDF sets agree well, while differences are observed in the high-energy tail, which are not covered by the PDF uncertainty. This was found to stem from differences in the large $x$-gluon distributions of the corresponding PDF fits.
\section{Conclusions}
The $\rho_\mathrm{s}$ distribution seems to be better described, especially at low $\rho_{\mathrm{s}}$, when using the dynamical scale $H_T^B/4$ instead of the static scale $\mu_0=m_t$. This is concluded from the observation that the graphs obtained with different $(K_R, K_F)$ values do not cross using either described dynamical scale and the NLO and LO scale variation bands overlap over the whole $\rho_{\mathrm{s}}$ distribution, while using the static scale these only barely overlap in the high-energy region. The shape variation of the $\rho_\mathrm{s}$ distribution induced by the scale variation is smaller when adopting the dynamical scales instead of the static scale. This leads to a strongly reduced scale variation uncertainty in the normalized $\rho_\mathrm{s}$ distribution, that can be used for the experimental extraction of the top quark mass. In fact, using the dynamical scales the PDF uncertainty is comparable in size to the scale variation uncertainty in the normalized $\rho_s$ distribution. It was additionally found that while the size of the scale uncertainty does not show dependence on the $R$-parameter in the anti-$k_T$ jet clustering algorithm using either of the two dynamical scales, the statistics can be increased by using a larger $R$-value.
|
\section{Appendix}
\subsection{Proof of Theorem \ref{thm: sufficiency of BSS}}
\label{sec: sufficiency of BSS}
Recall that $\mathscr{A}_{s,k}:= \{\mathcal{D} \subseteq \mathscr{A}_s : \abs{\mathcal{S}\setminus \mathcal{D}} = k\}$ and
\begin{equation}
\label{eq: Schur comp}
\Delta(\mathcal{D}) = \widehat{\Sigma}_{ \mathcal{S} \setminus \mathcal{D}, \mathcal{S} \setminus \mathcal{D}} - \widehat{\Sigma}_{ \mathcal{S} \setminus \mathcal{D}, \mathcal{D}}\widehat{\Sigma}_{\mathcal{D}, \mathcal{D}}^{-1} \widehat{\Sigma}_{\mathcal{D}, \mathcal{S} \setminus \mathcal{D}}
\end{equation}
Note that for a constant $\eta>0$ and $\mathcal{D} \in \mathscr{A}_{s,k}$ we have the following:
\begin{equation}
\label{eq: RSS diff}
\begin{aligned}
& n^{-1} (R_\mathcal{D} - R_\mathcal{S}) = n^{-1}\{ Y^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_\mathcal{D} ) Y - Y^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_\mathcal{S}) Y\}\\
&= n^{-1} \left\{ (X_{\mathcal{S} \setminus \mathcal{D}} \beta_{\mathcal{S} \setminus \mathcal{D}} + E)^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_{\mathcal{D}}) (X_{\mathcal{S} \setminus \mathcal{D}} \beta_{\mathcal{S} \setminus \mathcal{D}} + E) - E^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_\mathcal{S}) E \right\}\\
& = \eta \beta_{\mathcal{D}\setminus \mathcal{S}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}} + 2^{-1}(1-\eta) \beta_{\mathcal{S}\setminus \mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}} - 2 \left\{n^{-1} ({\sf I}_n - P_{\mathcal{D}})X_{\mathcal{S}\setminus \mathcal{D}}\beta_{\mathcal{S}\setminus \mathcal{D}}\right\}^{ \mathrm{\scriptscriptstyle T} } (-E)\\
&\quad + 2^{-1}(1-\eta) \beta_{\mathcal{S}\setminus \mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}} - n^{-1} E^{ \mathrm{\scriptscriptstyle T} }(P_\mathcal{D} - P_\mathcal{S}) E.
\end{aligned}
\end{equation}
Also, let $\widetilde{E}:= (-E)$.
Now, note that BSS chooses the correct model if
$$
\min_{\mathcal{D} \in \mathscr{A}_s} n^{-1}(R_\mathcal{D} - R_\mathcal{S}) > 0.
$$
Hence, by union bound we have the following:
\begin{equation}
\label{eq: error prob BSS}
\pr (\widehat{\mathcal{S}}_{\rm best} \neq \mathcal{S}) \leq \sum_{k =1}^s \pr \left( \min_{\mathcal{D} \in \mathscr{A}_{s,k} } n^{-1} (R_\mathcal{D} - R_\mathcal{S}) <0\right).
\end{equation}
Thus, under the light of equation \eqref{eq: RSS diff} it is sufficient to show the followings with high probability:
\begin{equation}
\label{eq: quant 1}
\max_{\mathcal{D} \in \mathscr{A}_{s,k}} \widehat{\gamma}_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} < \frac{(1-\eta)n^{1/2}}{4} \min_{\mathcal{D}\in \mathscr{A}_{s,k}} \norm{\gamma_\mathcal{D}}_2,
\end{equation}
\begin{equation}
\label{eq: quant 2}
\max_{\mathcal{D}\in \mathscr{A}_{s,k}} n^{-1} \left\{ E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S}) E \right\} < \frac{1 - \eta}{2} \min_{\mathcal{D}\in \mathscr{A}_{s,k}}\norm{\gamma_\mathcal{D}}_2^2,
\end{equation}
for every $k \in [s]$.
We will analyze the above events separately. We recall the two important spaces below:
\[
\mathcal{T}_k:= \{ \widehat{\gamma}_\mathcal{D} : \mathcal{D} \in \mathscr{A}_{s,k} \},
\]
\[
\mathcal{G}_k := \{ P_\mathcal{D} : \mathcal{D} \in \mathscr{A}_{s,k} \}.
\]
\textbf{Linear term:} We begin with analyzing the supremum process in Equation \eqref{eq: quant 1}. First note the following:
\[
\max_{\mathcal{D}\in \mathscr{A}_{s,k}} \widehat{\gamma}_{\mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} = \max_{\mathcal{D}\in \mathscr{A}_{s,k}} (\widehat{\gamma}_{\mathcal{D}} - \widehat{\gamma}_{\mathcal{D}_{0,k}})^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} + \widehat{\gamma}_{\mathcal{D}_{0,k}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E},
\]
where $\mathcal{D}_{0,k} \in \mathscr{A}_{s,k} $ is a fixed subset such that $\norm{P_{\mathcal{D}_{0,k}} - P_\mathcal{S}}_{{\rm op}} = {\sf d}_{\mathcal{G}_k}^\circ$.
Let $f_\mathcal{D} : = (\widehat{\gamma}_{\mathcal{D}} - \widehat{\gamma}_{\mathcal{D}_{0,k}})^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} $ and $\norm{f} := \max_{\mathcal{D}\in \mathscr{A}_{s,k}} f_\mathcal{D}$.
By Borell-TIS inequality \citep[Theorem 2.1.1]{adler2007random}, we have
\begin{equation}
\label{eq: borell-TIS}
\pr \left\{ \norm{f} - \mathbb{E}(\norm{f}) \geq \sigma u \right\} \leq {\sf exp} \left(- \frac{u^2}{2 {\sf D}^2_{\mathcal{T}_k}}\right),
\end{equation}
for all $u>0$. Also, by Theorem 1 of Section 14 in \cite{lifshits1995gaussian} and using the fact that $\abs{\mathcal{T}_k} = \binom{p-s}{k} \binom{s}{k}\leq (e p)^{2k}$, we have
\[
\mathbb{E}(\norm{f}) \leq 4\sqrt{2} \sigma \int_{0}^\infty \sqrt{\log \mathcal{N}(\varepsilon, \mathcal{T}_k, \norm{\cdot}_2)} \; {\rm d}\varepsilon \leq 8 {\sf D}_{\mathcal{T}_k} \sigma \sqrt{k \log(ep)}.
\]
Setting $u = 2{\sf D}_{\mathcal{T}_k}\sqrt{ k \log(ep)}$ in Equation \eqref{eq: borell-TIS} we get
\begin{equation}
\label{eq: linear deviation}
\pr (\norm{f} \geq 10 {\sf D}_{\mathcal{T}_k} \sigma \sqrt{ k \log(ep)}) \leq (ep)^{-2k}.
\end{equation}
Finally note that $\widehat{\gamma}_{\mathcal{D}_{0,k}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} \sim N(0,1)$. Hence, $\pr(\widehat{\gamma}_{\mathcal{D}_0}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} \geq c_\mathcal{T} \sigma \sqrt{ k \log(ep)}) \leq (ep)^{- c_\mathcal{T}^2 k/2}$, where $c_\mathcal{T}>0$ is an arbitrary constant. Thus, for $s\geq 1$ we have
\begin{equation}
\label{eq: linear deviation 2}
\pr \left\{\max_{\mathcal{D} \in \mathscr{A}_{s,k}} \widehat{\gamma}_{\mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} \geq (10 {\sf D}_{\mathcal{T}_k} + c_\mathcal{T}) \sigma \sqrt{k \log(ep) } \right\} \leq (ep)^{-2k} + (ep)^{-c_\mathcal{T}^2 k/2}.
\end{equation}
\textbf{Quadratic term:}
Here we study the quadratic supremum process in Equation \eqref{eq: quant 2}. For a fixed constant $c_\mathcal{G}>0$, by union bound we have,
\begin{equation}
\label{eq: quad decomposition}
\begin{aligned}
&\pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{s,k}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S})E > \sigma^2 u + \sigma^2 c_\mathcal{G} u_0\right\}\\
& \leq \pr \left\{ n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_{\mathcal{D}_{0,k}} - P_\mathcal{S})E > \sigma^2 u_0 c_\mathcal{G} \right\} + \sum_{\mathcal{D} \in \mathscr{A}_{s,k}} \pr \left\{ n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{D}_{0,k}})E > \sigma^2 u \right\}.
\end{aligned}
\end{equation}
Now we will use Theorem 1.1 of \cite{rudelson2013hanson}. In particular, working out the explicit constants in the proof of the theorem we get,
\[
\pr \left\{ n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{D}_{0,k}})E > \sigma^2 u \right\} \leq 2 {\sf exp} \left\{- \min \left( \frac{n^2 u^2}{256 \norm{P_\mathcal{D} - P_{\mathcal{D}_{0,k}}}^2_F}, \frac{n u}{16\sqrt{2} \norm{P_\mathcal{D} - P_{\mathcal{D}_{0,k}}}_{{\rm op}}}\right) \right\}, \; \forall u>0.
\]
Also note that for $\mathcal{D} \in \mathscr{A}_{s,k}$
$$
\norm{P_\mathcal{D} - P_{\mathcal{D}_{0,k}}}_F^2 \leq {\sf rank}(P_\mathcal{D} - P_{\mathcal{D}_{0,k}}) \norm{P_\mathcal{D} - P_{\mathcal{D}_{0,k}}}_{{\rm op}}^2 \leq 4k {\sf D}^2_{\mathcal{G}_k}.
$$
Thus, we have
\begin{equation}
\label{eq: quad deviation}
\begin{aligned}
& \leq \pr \left\{ n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{D}_{0,k}})E > \sigma^2 u \right\}\\
&\leq 2 {\sf exp} \left\{- \min \left( \frac{n^2 u^2}{1024 k {\sf D}^2_{\mathcal{G}_k}}, \frac{n u}{16\sqrt{2} {\sf D}_{\mathcal{G}_k}}\right) \right\}.
\end{aligned}
\end{equation}
Setting $u = 4 \times \{16\sqrt{2} {\sf D}_{\mathcal{G}_k} k \log (ep)/n\}$ we get
\begin{equation}
\label{eq: quad deviation 2}
\pr \left\{ n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{D}_{0,k}})E > \frac{64 \sqrt{2} {\sf D}_{\mathcal{G}_k} \sigma^2 k \log(ep)}{n} (1 - c_\mathcal{G}) \right\} \leq 2 {\sf exp} \left\{- 4 k \log (ep) \right\}.
\end{equation}
This shows that
\begin{equation}
\label{eq: quad deviation 3}
\begin{aligned}
&\pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{s,k}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{D}_{0,k}})E > \frac{64 \sqrt{2} {\sf D}_{\mathcal{G}_k} \sigma^2 k \log(ep)}{n}\right\}\\
&\leq 2 \binom{p-s}{k} \binom{s}{k} {\sf exp}(- 4k \log(ep))\\
&\leq 2(ep)^{-2k}.
\end{aligned}
\end{equation}
By a similar argument, setting $u_0 = 4 \times \{16\sqrt{2} {\sf d}_{\mathcal{G}_k}^\circ k \log (ep)/n\}$ we can show
\begin{equation}
\label{eq: quad deviation 4}
\begin{aligned}
& \pr \left\{ n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_{\mathcal{D}_{0,k}} - P_\mathcal{S})E > \frac{64 \sqrt{2} c_\mathcal{G} {\sf d}_{\mathcal{G}_k}^\circ \sigma^2 k \log(ep)}{n}\right\}\\
& \leq 2 {\sf exp} \left\{- 4c_\mathcal{G} k \log(ep) \right\}\\
& = 2 (ep)^{- 4 c_\mathcal{G} k},
\end{aligned}
\end{equation}
Combining Equation \eqref{eq: quad deviation 3} and Equation \eqref{eq: quad deviation 4} yields
\begin{equation}
\label{eq: quad deviation 5}
\pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{s,k}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{S}})E > 64 \sqrt{2} ({\sf D}_{\mathcal{G}_k} + c_\mathcal{G} {\sf d}^\circ_{\mathcal{G}_k})\sigma^2\frac{ k \log(ep)}{n}\right\} \leq 2 (ep)^{-2k} + 2 (ep)^{-4 c_\mathcal{G} k}.
\end{equation}
Setting $\eta = 0$ in Equation \eqref{eq: quant 1} and \eqref{eq: quant 2}, we have following requirements for vanishing error probability in Equation \eqref{eq: error prob BSS}:
\begin{itemize}
\item
\[
\min_{\mathcal{D} \in \mathscr{A}_{s,k}}\frac{\norm{\gamma_\mathcal{D}}_2}{\sqrt{k}} \geq 4(10 {\sf D}_{\mathcal{T}_k} + c_\mathcal{T}) \sigma \left\{ \frac{\log(ep)}{n}\right\}^{1/2}.
\]
\item $$\min_{\mathcal{D} \in \mathscr{A}_{s,k}}\frac{\norm{\gamma_\mathcal{D}}_2^2}{k} \geq 64 \sqrt{2} ({\sf D}_{\mathcal{G}_k} + c_\mathcal{G} {\sf d}^\circ_{\mathcal{G}_k})\sigma^2 \left\{ \frac{\log(ep)}{n}\right\}.$$
\end{itemize}
Thus one sufficient condition for above condition to hold for all $k \in [s]$ is
\begin{equation}
\label{eq: margin cond 1}
\widehat{\tau}(s) = \min_{\mathcal{D}\neq \mathcal{S}}\frac{\beta_{\mathcal{S}\setminus \mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}}}{\abs{\mathcal{S}\setminus \mathcal{D}}} \geq \max_{k \in [s]}\left\{16( 10 {\sf D}_{\mathcal{T}_k} + c_\mathcal{T})^2, 64\sqrt{2} ({\sf D}_{\mathcal{G}_k} + c_\mathcal{G} {\sf d}^\circ_{\mathcal{G}_k}) \right\} \sigma^2 \frac{\log(ep)}{n}.
\end{equation}
Thus, under \eqref{eq: margin cond 1}, Equation \eqref{eq: error prob BSS} yields
\begin{align*}
\pr(\widehat{\mathcal{S}}_{\rm best} \neq \mathcal{S}) &\leq \sum_{k=1}^s (ep)^{-2k} + (ep)^{-c_\mathcal{T}^2 k/2} + 2 (ep)^{-2k} + 2 (ep)^{-4 c_\mathcal{G} k}\\
& \leq 6 (ep)^{-2} + 2(ep)^{-c_\mathcal{T}^2 /2} + 4 (ep)^{-4 c_\mathcal{G} }.
\end{align*}
\subsection*{Retaining complexitites}
Recall that $\mu = X_\mathcal{S} \beta_\mathcal{S}$ and $\gamma_\mathcal{D} = n^{-1/2}({\sf I}_n - P_\mathcal{D}) \mu$.
\[
\mathcal{T}_{\mathcal{I}_0}:= \{ \widehat{\gamma}_\mathcal{D} : \mathcal{D} \in \mathscr{A}_{s} , \mathcal{D} \cap \mathcal{S} = \mathcal{I}_0 \},
\]
\[
\mathcal{G}_{\mathcal{I}_0} := \{ P_\mathcal{D} - P_{\mathcal{I}_0} : \mathcal{D} \in \mathscr{A}_{s}, \mathcal{D} \cap \mathcal{S} = \mathcal{I}_0 \}.
\]
Define the following bracketing integrals:
\begin{equation*}
\label{eq: proj-entropy-integrals}
\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}}:= \int_0^\infty \sqrt{\log \mathcal{N}(\mathcal{G}_{\mathcal{I}_0}, \norm{\cdot}_{{\rm op}}, \varepsilon)} \;{\rm d}\varepsilon /\{\sqrt{(s - \abs{\mathcal{I}_0}) \log(ep)}\}
\end{equation*}
\begin{equation*}
\label{eq: cross-entropy-intergral}
\mathscr{J}_{\mathcal{T}_{\mathcal{I}_0}}:= \int_0^\infty \sqrt{\log \mathcal{N}(\mathcal{T}_{\mathcal{I}_0}, \norm{\cdot}_{2}, \varepsilon)} \;{\rm d}\varepsilon / \{\sqrt{(s - \abs{\mathcal{I}_0}) \log(ep)}\}
\end{equation*}
\[
\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} > \{\log(ep)\}^{-1/2}\; \forall\; \mathcal{I}_0 \subset \mathcal{S}
\sapta{Assumption}
\]
Finally define the set $\mathscr{A}_{\mathcal{I}_0}: = \{\mathcal{D} \in \mathscr{A}_s: \mathcal{S} \cap \mathcal{D} = \mathcal{I}_0\}$. We also set $\abs{\mathcal{I}_0} = s - k$ for $k\in [s]$. Thus, $\mathscr{A}_{\mathcal{I}_0} \subset \mathscr{A}_{s,k}$.
\textbf{Linear term:}
Let $f_\mathcal{D} : = \widehat{\gamma}_{\mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} $ and $\norm{f} := \max_{\mathcal{D}\in \mathscr{A}_{\mathcal{I}_0}} f_\mathcal{D}$.
By Borell-TIS inequality \citep[Theorem 2.1.1]{adler2007random}, we have
\begin{equation}
\label{eq: borell-TIS.com}
\pr \left\{ \norm{f} - \mathbb{E}(\norm{f}) \geq \sigma u \right\} \leq {\sf exp} \left(- \frac{u^2}{2 }\right),
\end{equation}
for all $u>0$. Also, by Theorem 1 of Section 14 in \cite{lifshits1995gaussian} we have
\[
\mathbb{E}(\norm{f}) \leq 4\sqrt{2} \sigma \int_{0}^\infty \sqrt{\log \mathcal{N}(\varepsilon, \mathcal{T}_{\mathcal{I}_0}, \norm{\cdot}_2)} \; {\rm d}\varepsilon = 4\sqrt{2} \sigma \mathscr{J}_{\mathcal{T}_{\mathcal{I}_0}} \sqrt{k\log(ep)}.
\]
Setting $u = 2c_\mathcal{T}\sqrt{ k \log(ep)}$ in Equation \eqref{eq: borell-TIS.com} we get
\begin{equation}
\label{eq: linear deviation.com}
\pr (\norm{f} \geq 4\sqrt{2} \sigma \mathscr{J}_{\mathcal{T}_{\mathcal{I}_0}} \sqrt{k\log(ep)}+ 2 c_\mathcal{T} \sigma \sqrt{ k \log(ep)}) \leq (ep)^{-2c_\mathcal{T}^2 k}.
\end{equation}
Thus, for $s\geq 1$ we have
\begin{equation}
\label{eq: linear deviation 2.com}
\pr \left\{\max_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}} \widehat{\gamma}_{\mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} \geq (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{I}_0}} +
2 c_\mathcal{T}) \sigma \sqrt{k \log(ep) } \right\} \leq (ep)^{-2 c_\mathcal{T}^2 k}.
\end{equation}
\textbf{Quadratic term:}
Here we study the quadratic supremum process in Equation \eqref{eq: quant 2}. First, define the two projection operators $P_{\mathcal{D} \mid \mathcal{I}_0} = P_\mathcal{D} - P_{\mathcal{I}_0}$ and $P_{\mathcal{S} \mid \mathcal{I}_0}: = P_\mathcal{S} - P_{\mathcal{I}_0}$. For a fixed constant $c_\mathcal{G}>0$, by union bound we have,
\begin{equation}
\label{eq: quad decomposition}
\begin{aligned}
&\pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S})E > \sigma^2 u + \sigma^2 c_\mathcal{G} u_0\right\}\\
& \pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}} E^{ \mathrm{\scriptscriptstyle T} } (P_{\mathcal{D} \mid \mathcal{I}_0} - P_{\mathcal{S}\mid \mathcal{I}_0})E > \sigma^2 u + \sigma^2 c_\mathcal{G} u_0\right\}\\
& \leq \pr \left\{ n^{-1} (k - E^{ \mathrm{\scriptscriptstyle T} } P_{\mathcal{S}\mid \mathcal{I}_0} E) > \sigma^2 u_0 c_\mathcal{G} \right\} + \pr \left\{ n^{-1} \max_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}}(E^{ \mathrm{\scriptscriptstyle T} } P_{\mathcal{D}\mid \mathcal{I}_0} E - k) > \sigma^2 u \right\}.
\end{aligned}
\end{equation}
By Theorem 3.1 of \cite{krahmer2014suprema} we have the following for any $t\geq 0$:
\begin{equation}
\label{eq: quad-deviation.com}
\begin{aligned}
&\pr \left\{\max_{\mathcal{D} \in \mathscr{A}_{s} } \abs{E^{ \mathrm{\scriptscriptstyle T} } P_{\mathcal{D}\mid \mathcal{I}_0} E - k}/\sigma^2\geq c_1 c_2 \mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}}(c_2 \mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} \sqrt{k \log(ep)} + \sqrt{k}) \sqrt{k \log(ep)} + c_1 \sqrt{k} + t
\right\}\\
&\leq {\sf exp} \left( - \min \left\{ \frac{t^2}{V^2}, t \right\} \right),
\end{aligned}
\end{equation}
where $V = c_2 \mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} \sqrt{k \log (ep)} + \sqrt{k}$. Recall that ${\sf diam} (\mathcal{G}_{\mathcal{I}_0}, \norm{\cdot}_{{\rm op}})={\sf D}_{\mathcal{G}_{\mathcal{I}_0}} \leq 1$, then we have $\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} \leq {\sf D}_{\mathcal{G}_{\mathcal{I}_0}} $. Also, recall that $\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} > \{\log(ep)\}^{-1/2}$ for all $\mathcal{I}_0 \subset \mathcal{S}$. Setting $t = c_2 \mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}}\sqrt{k \log (ep)}( c_2 \mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}}\sqrt{k \log (ep)} + \sqrt{k})$ in \eqref{eq: quad-deviation.com} we get
\[
\pr \left\{ \max_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}} E^{ \mathrm{\scriptscriptstyle T} } P_{\mathcal{D}\mid \mathcal{I}_0} E - k \geq c_3 \sigma^2\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} k \log(ep) \right\} \leq (ep)^{- c_4 k}
\]
By a similar argument, setting $u_0 = k\log (ep)/n$ we can show
\begin{equation}
\label{eq: quad deviation 4.com}
\begin{aligned}
& \pr \left\{ n^{-1} \abs{E^{ \mathrm{\scriptscriptstyle T} } P_{\mathcal{S}\mid \mathcal{I}_0} E - s} > \frac{ c_\mathcal{G} \sigma^2 k \log(ep)}{n}\right\}\\
& \leq 2 {\sf exp} \left\{- c_5 c_\mathcal{G} k \log(ep) \right\}\\
& = 2 (ep)^{- c_5 c_\mathcal{G} k},
\end{aligned}
\end{equation}
Combining Equation \eqref{eq: quad deviation 3} and Equation \eqref{eq: quad deviation 4} yields
\begin{equation}
\label{eq: quad deviation 5}
\pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{S}})E > c_6 (\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} + c_\mathcal{G} )\sigma^2\frac{ k \log(ep)}{n}\right\} \leq (ep)^{- c_4 k} + 2 (ep)^{- c_5 c_\mathcal{G} k}.
\end{equation}
Now recall that
\begin{equation}
\widehat{\tau}(s) \triangleq \min_{\mathcal{D}\neq \mathcal{S}}\frac{\beta_{\mathcal{S}\setminus \mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}}}{\abs{\mathcal{S} \setminus \mathcal{D}}} \geq \max\left\{ \max_{ \mathcal{I}_0 \subset \mathcal{S}} (c_6\mathscr{J}_{\mathcal{T}_{\mathcal{I}_0}} + 2 c_{\mathcal{T}})^2, \max_{ \mathcal{I}_0 \subset \mathcal{S}} c_5 (\mathscr{J}_{\mathcal{G}_{\mathcal{I}_0}} + c_\mathcal{G} ) \right\} \frac{\sigma^2 \log(ep)}{n}.
\label{eq: sufficiency condition for BSS}
\end{equation}
Thus we have
\begin{align*}
&\pr (\widehat{\mathcal{S}}_{\rm best} \neq \mathcal{S})\\
& \leq \sum_{k =1}^s \sum_{\mathcal{I}_0 \subset \mathcal{S} : \abs{\mathcal{I}_0} = s-k} \pr \left\{\min_{\mathcal{D} \in \mathscr{A}_{\mathcal{I}_0}} n^{-1} (R_\mathcal{D} - R_\mathcal{S}) <0 \right\}\\
& \leq \sum_{k = 1}^s \sum_{\mathcal{I}_0 \subset \mathcal{S}: \abs{\mathcal{I}_0} = s-k} \left\{(ep)^{-2 c_\mathcal{T}^2 k} + (ep)^{- c_4 k} + (ep)^{-c_5c_\mathcal{G} k}\right\}\\
& \leq \sum_{k = 1}^s \binom{s}{k} \left\{(ep)^{-2 c_\mathcal{T}^2 k} + (ep)^{- c_4 k} + (ep)^{-c_5c_\mathcal{G} k}\right\}\\
& \leq \sum_{k=1}^s (es)^k\left\{(ep)^{-2 c_\mathcal{T}^2 k} + (ep)^{- c_4 k} + (ep)^{-c_5c_\mathcal{G} k}\right\}\\
& \leq \sum_{k=1}^s {\sf exp} \left[-k \{2 c_\mathcal{T}^2 \log(ep) - \log(es)\} \right] + {\sf exp}\left[-k \{c_4 \log(ep) - \log(es)\}\right]\\
& \quad + {\sf exp}\left[-k\{c_5 c_\mathcal{G} \log(ep) - \log(es)\}\right].
\end{align*}
Hence, using the fact that $ 1 + \log s < 0.5 \min\{2 c_\mathcal{T}^2, c_4, c_5 c_\mathcal{G}\} \log (ep) $, the error probability goes to 0 as $p \to \infty$.
\subsection{Correlated random feature model example}
\label{sec: random feature model example}
Consider the model \eqref{eq: base model}. We assume that the rows of $X$ are independently generated from $p$-dimensional multivariate Gaussian distribution with mean-zero and variance-covariance matrix
\[
\Sigma = \begin{pmatrix}
1 & \b0_{(p-1) \times 1}\\
\b0_{1 \times (p-1)} & (1- \mu){\sf I}_{p-1} + \mu {\sf ind}_{p-1} {\sf ind}_{p-1}^{ \mathrm{\scriptscriptstyle T} }
\end{pmatrix},
\]
and true model is $\mathcal{S} = \{1\}$. In this case
\[
\widehat{\tau} = \beta_{1}^2 \min_{j \neq 1} \left\{ \frac{\norm{X_{1}}_2^2}{n} - \frac{(X_{1}^{ \mathrm{\scriptscriptstyle T} } X_{j}/n)^2}{\norm{X_{j}}^2/n} \right\} = \frac{\beta_1^2 \norm{X_1}_2^2}{n} - \beta_1^2 \min_{j \neq 1} \frac{(X_1^{ \mathrm{\scriptscriptstyle T} } X_j/n)^2}{\norm{X_j}_2^2/n}.
\]
We start with providing an upper bound on teh margin quantity $\widehat{\tau}$. Setting, $\delta = 1/2$ in Equation \eqref{eq: LM1-simple} and \eqref{eq: LM2-simple} we get
\begin{equation}
\label{eq: norm concentration 1}
\pr \left( \frac{\norm{X_j}_2^2}{n} \geq 1 + \delta_1\right) \leq {\sf exp}(-n \delta_1^2/16), \quad \forall j \in [p].
\end{equation}
\begin{equation}
\label{eq: norm concentration 2}
\pr \left( \frac{\norm{X_j}_2^2}{n} \leq 1 - \delta_1\right) \leq {\sf exp}(-n \delta_1^2/4), \quad \forall j \in [p].
\end{equation}
Using Equation \eqref{eq: norm concentration 2}, we also get the following:
\begin{equation}
\label{eq: norm concentration 3}
\begin{aligned}
\pr \left(\min_{j \neq 1} \frac{\norm{X_j}_2^2}{n} \leq 1 - \delta_1\right) \leq p {\sf exp}(-n \delta_1^2/4).
\end{aligned}
\end{equation}
Let $X_j = (X_{j,1}, \ldots, X_{j,n})^{ \mathrm{\scriptscriptstyle T} }$ for all $j \in [p]$.
Now note that $\norm{X_{1,u} X_{j,u}}_{\psi_1} \leq \norm{X_{1,u}}_{\psi_2} \norm{X_{2,u}}_{\psi_2} \leq 4. $
Thus, by Berstein's inequality we have
\[
\pr \left( \abs{\frac{X_1^{ \mathrm{\scriptscriptstyle T} } X_j}{n}} > \varepsilon \right) \leq {\sf exp} \left( - C n \varepsilon^2 \right),
\]
where $\varepsilon$ is a positive constant smaller than 1 and $C>0$ is a universal constant. Thus, we have
\begin{equation}
\label{eq: cross term 1}
\pr \left( \max_{j \neq 1}\abs{\frac{X_1^{ \mathrm{\scriptscriptstyle T} } X_j}{n}} > \varepsilon\right) \leq p {\sf exp}(- C n \varepsilon^2).
\end{equation}
Combining \eqref{eq: norm concentration 2}, \eqref{eq: norm concentration 3} and \eqref{eq: cross term 1} we have
\begin{equation}
\label{eq: prob margin}
\pr \left\{ \widehat{\tau} \geq \beta_1^2 \left(1 - \delta_1 - \frac{\varepsilon^2}{1 - \delta_1} \right) \right\} \geq 1 - {\sf exp}(-n \delta_1^2/16) - p{\sf exp}(-n \delta_1^2/4) - p {\sf exp}(-5C \varepsilon^2 n) = 1 + o(1/p),
\end{equation}
if $n \gtrsim \max \{\delta_1^{-2}, \varepsilon^{-2}\}\log p$. Again by Bernstein's inequality,
it can also be shown that
\begin{equation}
\label{eq: cross term 2}
\pr \left( \max_{j,k \neq 1}\abs{\frac{X_k^{ \mathrm{\scriptscriptstyle T} } X_j}{n} - \mu} \leq \mu\phi\right) \geq 1 - p^2 {\sf exp}(- C_1 n \phi^2) = 1 + o(1/p),
\end{equation}
when $\mu \in (0.75,1)$ and $n \gtrsim \phi^{-2} \log p$.
Next we will analyze the geometric quantities.
In this case we have
\[
\widehat{\gamma}_j = \frac{X_{1} - \frac{X_j^{ \mathrm{\scriptscriptstyle T} } X_{1}}{\norm{X_j}^2}.X_{j}}{\sqrt{\norm{X_{1}}^2 - \frac{(X_{1}^{ \mathrm{\scriptscriptstyle T} } X_{j})^2}{\norm{X_{j}}^2}}}.
\]
Note that \begin{align*}
\norm{\widehat{\gamma}_j - \widehat{\gamma}_k}_2^2 & = 2 (1 - \widehat{\gamma}_j^{ \mathrm{\scriptscriptstyle T} } \widehat{\gamma}_k)
\end{align*}
and
\begin{align*}
\widehat{\gamma}_j^{ \mathrm{\scriptscriptstyle T} } \widehat{\gamma}_k &= \dfrac{\norm{X_{1}}^2/n - \frac{
(X_j^{ \mathrm{\scriptscriptstyle T} } X_{1}/n)^2 }{\norm{X_j}^2/n} - \frac{
(X_k^{ \mathrm{\scriptscriptstyle T} } X_{1}/n)^2 }{\norm{X_k}^2/n} +\frac{
(X_j^{ \mathrm{\scriptscriptstyle T} } X_{1}/n) (X_k^{ \mathrm{\scriptscriptstyle T} } X_{1}/n) (X_j^{ \mathrm{\scriptscriptstyle T} } X_{k}/n)}{(\norm{X_j}^2/n) (\norm{X_k}^2/n)}}{\sqrt{\norm{X_{1}}^2/n - \frac{(X_{1}^{ \mathrm{\scriptscriptstyle T} } X_{j}/n)^2}{\norm{X_{j}}^2/n}} \sqrt{\norm{X_{1}}^2/n - \frac{(X_{1}^{ \mathrm{\scriptscriptstyle T} } X_{k}/n)^2}{\norm{X_{k}}^2/n}}}\\
& \approx \frac{1 - \varepsilon_1^2 - \varepsilon_2^2 - \mu \varepsilon_1 \varepsilon_2}{\sqrt{1 - \varepsilon_1^2} \sqrt{1 - \varepsilon_2^2}}.
\end{align*}
Hence by similar argument, for $n \gtrsim \log p$ we get
\[
\pr \left( \widehat{\gamma}_j^{ \mathrm{\scriptscriptstyle T} } \widehat{\gamma}_k \geq \frac{1-\delta_1 - \frac{\varepsilon_1^2}{1 - \delta_1}- \frac{\varepsilon_2^2}{1 - \delta_1} + 4 \mu (1 - \phi) \frac{\varepsilon_1 \varepsilon_2}{(1 + \delta_1)(1 + \delta_2)}}{\sqrt{1 + \delta_1 - \frac{\varepsilon_1^2}{1 + \delta_1}} \sqrt{1 + \delta_1 - \frac{\varepsilon_2^2}{1 + \delta_1}}}\right) = 1 + o(1/p^2).
\]
Hence, an union bound argument yields
\begin{equation}
\label{eq: diam T}
\pr \left\{ {\sf D}_{\mathcal{T}_1} \leq 2 \left(1 - \frac{1-\delta_1 - \frac{\varepsilon_1^2}{1 - \delta_1}- \frac{\varepsilon_2^2}{1 - \delta_1} + 4 \mu (1 - \phi) \frac{\varepsilon_1 \varepsilon_2}{(1 + \delta_1)(1 + \delta_2)}}{\sqrt{1 + \delta_1 - \frac{\varepsilon_1^2}{1 + \delta_1}} \sqrt{1 + \delta_1 - \frac{\varepsilon_2^2}{1 + \delta_1}}}\right) \right\} = 1 + o(1/p),
\end{equation}
provided that $n \gtrsim \max\{\delta_1^{-2}, \varepsilon^{-2}, \varepsilon_1^{-2}, \varepsilon_2^{-2}, \phi^{-2}\} \log p$.
For $j, k \neq 1$, let $\theta_{j,k}$ denote the
angle between $X_j$ and $X_k$.
\[
\norm{P_j- P_k}_{{\rm op}} = \sin(\theta_{j,k}) = \sqrt{1 - \cos^2(\theta_{j,k})} = \sqrt{1 - \left( \frac{X_j^{ \mathrm{\scriptscriptstyle T} } X_k}{\norm{X_j} \norm{X_k}}\right)^2}.
\]
By similar arguments it follows that
\begin{equation}
\label{eq: dima G}
\pr \left( {\sf D}_{\mathcal{G}_1} \leq \sqrt{1 - \left(\frac{\mu (1 - \phi)}{(1 + \delta_1)}\right)^2}\right) = 1 + o(1/p).
\end{equation}
Finally, for ${\sf D}_{\mathcal{G}_1}^\circ$ we have the following for large $n$:
\[
\pr \left( {\sf D}_{\mathcal{G}_1}^\circ \geq \sqrt{1 - \frac{\varepsilon^2}{1 - \delta_1}}\right) = 1+ o(1/p)
\]
To keep things simple, we set a common value $\delta_0$ for all the parameters $\varepsilon , \varepsilon_1 , \varepsilon_2 , \delta_1 , \phi $. Thus, for $n \gtrsim \delta_0^{-2} \log p$, the following events are true with probability $1 + o(1/p)$:
\[
\widehat{\tau} \geq \beta_1^2 \left( 1 - \frac{1 - 2\delta_0}{1 - \delta_0}\right),
\]
\[
{\sf D}_{\mathcal{T}_1} \leq 2 \left(
6 \delta_0 - 4 \mu \delta_0^2
\right) =: f_1(\delta_0),
\]
\[
{\sf D}_{\mathcal{G}_1} \leq \sqrt{1 - \mu^2 \left(\frac{1 - \delta_0}{1 + \delta_0}\right)^
2} =: f_2(\mu, \delta_0),
\]
\[
{\sf D}_{\mathcal{G}_1}^\circ \geq \sqrt{1 - \frac{\delta_0^2}{1 - \delta_0}}=: f_3(\delta_0),
\]
provided that $\delta$ is sufficiently small. Thus, to hold the margin condition in Theorem \ref{thm: sufficiency of BSS}, it is sufficient to have the following:
\[
\beta_1^2 \left(1 - \frac{1-2\delta_0}{1-\delta_0} \right) \geq
\max\left\{ 16 (10 f_1(\delta_0) + c_\mathcal{T} )^2, 64\sqrt{2}(f_2(\mu, \delta_0) + c_\mathcal{G})\right\} \sigma^2 \frac{\log p}{n},
\]
which is is easily achieved for small $\delta$ and $\mu \approx 1$, if $\beta_1^2 \gtrsim \sigma^2 \log(p)/n$.
\subsection{Proof of Theorem \ref{thm: necessary condition BSS}}
\label{sec: ncessary condition BSS}
We again start with the difference of RSS between a candidate model $\mathcal{D} \in \mathscr{A}_{s,k}$ and the true model $\mathcal{S}$:
\begin{equation}
\label{eq: RSS diff 2}
\begin{aligned}
& n^{-1} (R_\mathcal{D} - R_\mathcal{S}) = n^{-1}\{ Y^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_\mathcal{D} ) Y - Y^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_\mathcal{S}) Y\}\\
&= n^{-1} \left\{ (X_{\mathcal{S} \setminus \mathcal{D}} \beta_{\mathcal{S} \setminus \mathcal{D}} + E)^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_{\mathcal{D}}) (X_{\mathcal{S} \setminus \mathcal{D}} \beta_{\mathcal{S} \setminus \mathcal{D}} + E) - E^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_\mathcal{S}) E \right\}\\
& = \beta_{\mathcal{S}\setminus \mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}} - 2 \left\{n^{-1} ({\sf I}_n - P_{\mathcal{D}})X_{\mathcal{D}\setminus \mathcal{S}}\beta_{\mathcal{D}\setminus \mathcal{S}}\right\}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} - n^{-1} E^{ \mathrm{\scriptscriptstyle T} }(P_\mathcal{D} - P_\mathcal{S}) E.
\end{aligned}
\end{equation}
First of all, in order achieve model consistency, the following is necessary for any $k \in [s]$:
\begin{equation}
\label{eq: RSS diff necessary}
\min_{\mathcal{D}\in \mathscr{C}_{j_0}} n^{-1}(R_\mathcal{D} - R_{\mathcal{S}}) >0.
\end{equation}
Recall that $$\widehat{\tau}_{j_0} = \max_{\mathcal{D} \in \mathscr{C}_{j_0}} \Delta(\mathcal{D}) \beta_{j_0}^2.$$
Next we note that
\begin{equation}
\label{eq: RSS diff 3}
\begin{aligned}
\min_{\mathcal{D} \in \mathcal{D}_{j_0}} n^{-1}(R_\mathcal{D} - R_\mathcal{S}) &\leq \min_{\mathcal{D} \in \mathscr{C}_{j_0}} \{ \beta_{\mathcal{S}\setminus \mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \Delta(\mathcal{D}) \beta_{\mathcal{S}\setminus \mathcal{D}} - 2 n^{-1/2} \gamma_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} - n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S}) E\}\\
& \leq \widehat{\tau}_{j_0} +2 n^{-1/2}\max_{\mathcal{D} \in \mathscr{C}_{j_0}} \abs{\gamma_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E}} - n^{-1}\max_{\mathcal{D} \in \mathscr{C}_{j_0}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S}) E\\
& \leq \widehat{\tau}_{j_0} + 2 (\widehat{\tau}_{j_0}/n)^{1/2} \max_{\mathcal{D} \in \mathscr{C}_{j_0}} \abs{\widehat{\gamma}_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E}} - n^{-1}\max_{\mathcal{D} \in \mathscr{C}_{j_0}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S}) E.
\end{aligned}
\end{equation}
Similar to the proof of Theorem \ref{thm: sufficiency of BSS}, we define $f_\mathcal{D} : = \widehat{\gamma}_{\mathcal{D}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} $ and $\norm{f} := \max_{\mathcal{D}\in \mathscr{C}_{j_0}} f_\mathcal{D}$. Hence, we have
\begin{equation}
\label{eq: max T}
\max_{\mathcal{D}\in \mathscr{C}_{j_0}} \abs{\widehat{\gamma}_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E}} = \max_{\mathcal{D} \in \mathscr{C}_{j_0}} f_\mathcal{D} \vee (-f_\mathcal{D})
\end{equation}
By Borell-TIS inequality \citep[Theorem 2.1.1]{adler2007random}, we have
\begin{equation*}
\label{eq: borell-TIS 2}
\pr \left\{ \norm{f} - \mathbb{E}(\norm{f}) \geq \sigma u \right\} \leq {\sf exp} \left(- \frac{u^2}{2 }\right),
\end{equation*}
for all $u>0$. Setting $u = c_\mathcal{T}\sqrt{ \log (ep)}$ we get
\[
\pr \left\{ \norm{f} - \mathbb{E}(\norm{f}) \geq c_\mathcal{T}\sigma \sqrt{ \log (ep)}\right\} \leq (ep)^{-c_\mathcal{T}^2 /2}.
\]
\[
\mathbb{E}(\norm{f}) \leq 4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} \sigma \sqrt{ \log(ep)},
\]
which ultimately yields
\[
\pr \left\{ \norm{f} \geq (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T}) \sigma \sqrt{ \log(ep)}\right\} \leq (ep)^{-c_\mathcal{T}^2 /2}.
\]
Finally. using \eqref{eq: max T} we have
\begin{equation}
\label{eq: lin deviation}
\pr \left\{ \max_{\mathcal{D}\in \mathscr{C}_{j_0}} \abs{\widehat{\gamma}_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E}} \geq (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T}) \sigma \sqrt{ \log(ep)}\right\} \leq 2 (ep)^{-c_\mathcal{T}^2 /2}.
\end{equation}
Next, we will lower bound the quadratic term in Equation \eqref{eq: RSS diff 3} with high probability.
similar to the proof of Theorem \ref{thm: sufficiency of BSS}, we consider the decomposition
\[
\max_{\mathcal{D} \in \mathscr{C}_{j_0}}n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_\mathcal{S})E = \max_{j \notin \mathcal{S}} n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (\widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_{j_0}\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} })E.
\]
For the maximal process we will use Theorem 2.10 of \cite{adamczak2015note}. We begin with the definition of concentration property.
\begin{definition}
\label{def: concentration property}
Let $Z$ be random vector in $\mathbb{R}^n$. We say that $Z$ has concentration property with constant $K$ if for every 1-Lipschitz function $\varphi : \mathbb{R}^n \to \mathbb{R}$, we have $\mathbb{E} \abs{\varphi(Z)} < \infty$ and for every $u>0$,
\[
\pr \left( \abs{\varphi(Z) - \mathbb{E}(\varphi(Z))} \geq u\right) \leq 2 {\sf exp}(- u^2/K^2).
\]
\end{definition}
Note that the Gaussian vector $E$ enjoys concentration property with $K = \sqrt{2}\sigma$ \citep[Theorem 5.6]{boucheron2013concentration}. Let $Q_1 : = \max_{j \notin \mathcal{S}} E^{ \mathrm{\scriptscriptstyle T} } (\widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_{j_0}\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} })E$.
By Theorem 2.10 of \cite{adamczak2015note} we conclude that
\[
\pr \left\{ n^{-1}\abs{Q_1 - \mathbb{E}(Q_1)} \geq t\right\} \leq 2 {\sf exp} \left\{ - \frac{1}{C_0 \sigma^2} \min \left( \frac{n^2 t^2}{16}, \frac{n t}{2}\right)\right\},
\]
where $C_0$ is universal positive constant.
Setting $t = C_0 \delta \sigma^2 \log(ep)/n$ in the above equation we get
\begin{equation}
\label{eq: thm2 : quad deviation 2}
\pr \left\{ n^{-1}\abs{Q_1 - \mathbb{E}(Q_1)} \geq C_0 \delta \sigma^2 \log(ep)/n \right\} \leq 2 (ep)^{-\frac{ \delta}{2 } }.
\end{equation}
Next, we will lower bound the expected value of $Q_1$. First, note the following:
\begin{equation}
\label{eq: expected quadratic}
\begin{aligned}
\mathbb{E}(Q_1) & = \mathbb{E} \left\{\max_{j \notin \mathcal{S}} E^{ \mathrm{\scriptscriptstyle T} } (\widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_{j_0} \widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} }) E\right\} \\
& = \mathbb{E} \left\{\max_{j \notin \mathcal{S}} (\widehat{u}_{j}^{ \mathrm{\scriptscriptstyle T} } E)^2 \right\} - \sigma^2\\
& \geq \mathbb{E}\left\{\max_{j \in \mathcal{J}_0} (\widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } E)^2\right\} - \sigma^2\\
& \geq \left\{\mathbb{E} \max_{j \in \mathcal{J}_0} (\widehat{u}_{j}^{ \mathrm{\scriptscriptstyle T} } E) \right\}^2 - \sigma^2
\end{aligned}
\end{equation}
By Sudakov lower bound we have
\[
\mathbb{E} \max_{j \in \mathcal{J}_0} \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } E \geq \sup_{\delta>0} \frac{\sigma \delta}{2} \sqrt{\log \mathcal{M}(\delta, \{\widehat{u}_j\}_{j \in \mathcal{J}_0}, \norm{\cdot}_2)} \geq \sigma \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}} \sqrt{ \log(ep)}.
\]
Thus, \eqref{eq: expected quadratic} yields
\[
\mathbb{E}(Q_1) \geq \sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2 \log (ep) - \sigma^2.
\]
Thus, combined with \eqref{eq: thm2 : quad deviation 2} we finally get
\[
\pr \left[ Q_1 \geq \sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2 \log (ep) - \sigma^2 - C_0 \delta \sigma^2 \log(ep) \right] \geq 1 - 2 (ep)^{-\delta/2 }.
\]
Setting $\delta = \frac{\mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{2 C_0}$, we get
\[
\pr \left[ Q_1 \geq \frac{\sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{2} \log (ep) - \sigma^2 \right] \geq 1 - 2 (ep)^{-\frac{\mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{4 C_0} }.
\]
Thus, finally combining the above with \eqref{eq: RSS diff 3} and \eqref{eq: lin deviation} we get
\begin{equation}
\label{eq: RSS diff 4}
\begin{aligned}
&\min_{\mathcal{D} \in \mathscr{C}_{j_0}} n^{-1}(R_\mathcal{D} - R_\mathcal{S})\\
& \leq \widehat{\tau}_{j_0} + 2 \widehat{\tau}_{j_0}^{1/2} (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T}) \sigma \sqrt{\frac{\log(ep)}{n}} - n^{-1}\left( \frac{\sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{2} \log (ep) - \sigma^2 \right)
\end{aligned}
\end{equation}
with probability at least $1 - 2 (ep)^{-c_{\mathcal{T}}^2/2} - 2(ep)^{- \frac{\mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{4 C_0}}$. Thus, for large $p$ we have
\begin{equation*}
\begin{aligned}
&\min_{\mathcal{D} \in \mathscr{C}_{j_0}} n^{-1}(R_\mathcal{D} - R_\mathcal{D})\\
& \leq \widehat{\tau}_{j_0} + 2 \widehat{\tau}_{j_0}^{1/2} (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T}) \sigma \sqrt{\frac{\log(ep)}{n}} - \frac{\sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{4} \frac{\log (ep)}{n}
\end{aligned}
\end{equation*}
with probability at least $1 - 2 (ep)^{c_{\mathcal{T}}^2/2} - 2(ep)^{- \frac{\mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{4 C_0}}$.
Thus, in light of \eqref{eq: RSS diff necessary}, the following is necessary:
\[
\widehat{\tau}_{j_0}\geq \left\{\frac{\sqrt{b_1^2 + 4 a_1} - b_1}{2}\right\}^2,
\]
where
\[
b_1 = 2 (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T}) \sigma \sqrt{\frac{\log(ep)}{n}},
\]
\[
a_1 = \frac{\sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{4} \frac{\log (ep)}{n}.
\]
This finishes the proof.
\subsubsection*{Another decomposition}
Let $\mu = X_\mathcal{S} \beta_\mathcal{S}$. Also, by definition it follows that $\gamma_\mathcal{S} = 0$. Denote the only element of $\mathcal{D} \setminus \mathcal{S}$ by $j$. We then have that
\begin{equation}
\label{eq: lin diff gamma}
\gamma_\mathcal{D} = \gamma_\mathcal{D} - \gamma_\mathcal{S} = n^{-1/2}(P_\mathcal{S} - P_\mathcal{D}) \mu = n^{-1/2} (\widehat{u}_{j_0} \widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } )\mu ,
\end{equation}
and that
\begin{equation}
\label{eq: squre diff gamma}
\begin{aligned}
\norm{\gamma_\mathcal{D}}_2^2 - \norm{\gamma_\mathcal{S}}_2^2 &= n^{-1} \left\{(\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu)^2 - (\widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } \mu)^2 \right\}\\
& = n^{-1} \left( \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} - \abs{\widehat{u}_{j}^{ \mathrm{\scriptscriptstyle T} } \mu}\right) \left( \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} + \abs{\widehat{u}_{j}^{ \mathrm{\scriptscriptstyle T} } \mu}\right)\\
& \leq \frac{2}{n} \lambda_0 \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu},
\end{aligned}
\end{equation}
Also, due to Assumption \ref{assumption: anchor point}\ref{item: projected margin separation}, we have $\abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} > 4\lambda_0$. Therefore, we have that
\[
(\widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } \mu )^2 \geq (\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu )^2 - 2 \lambda_0 \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} \geq \frac{(\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu )^2}{2}.
\]
Then, due to Lemma 6.1 of \cite{fan2020best} and \eqref{eq: lin diff gamma} we have
\[
\norm{\gamma_{\mathcal{S}_0 \cup \{j\}}}_2 \geq (2n)^{-1/2} \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} {\sf d}^\circ_\mathcal{U}, \quad \forall\; j \in \mathcal{J}_0,
\]
where ${\sf d}^\circ_\mathcal{U}:= \min_{j\in \mathcal{J}_0} \norm{\widehat{u}_j - \widehat{u}_{j_0}}_2$.
This shows that for any model $j_1, j_2 \in \mathcal{J}_0$, the following holds:
\[
\norm{\gamma_{\mathcal{S}_0 \cup \{j_1\}} - \gamma_{\mathcal{S}_0 \cup \{j_2\}}}_2\geq (2n)^{-1/2} \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} {\sf d}^\circ_\mathcal{U} \norm{\widehat{\gamma}_{\mathcal{S}_0 \cup \{j_1\}} - \widehat{\gamma}_{\mathcal{S}_0 \cup \{j_2\}}}_2.
\]
Setting $\delta^\prime = \delta (2n)^{-1/2} \abs{\widehat{u}^{ \mathrm{\scriptscriptstyle T} }_{j_0} \mu }{\sf d}^\circ_\mathcal{U}$, the above inequality shows that
\begin{align*}
\sup_{\delta^\prime >0} \frac{\delta^\prime}{2} \sqrt{\log \mathcal{M}(\delta^\prime, \{\gamma_{\mathcal{S}_0 \cup \{j\}}\}_{j \in \mathcal{J}_0}, \norm{\cdot}_2)} &\geq (2n)^{-1/2} \abs{\widehat{u}^{ \mathrm{\scriptscriptstyle T} }_{j_0} \mu }{\sf d}^\circ_\mathcal{U} \sup_{\delta>0} \frac{\delta}{2} \sqrt{\log \mathcal{M}(\delta, \{\widehat{\gamma}_{\mathcal{S}_0 \cup \{j\}}\}_{j \in \mathcal{J}_0}, \norm{\cdot}_2)}\\
& = (2n)^{-1/2} \abs{\widehat{u}^{ \mathrm{\scriptscriptstyle T} }_{j_0} \mu }{\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}\sqrt{ \log(ep)}
\end{align*}
\begin{equation}
\label{eq: RSS diff 5}
\begin{aligned}
& \min_{\mathcal{D} \in \mathscr{C}_{j_0}} n^{-1}(R_\mathcal{D} - R_\mathcal{S})\\
&\leq \min_{\mathcal{D} \in \mathscr{C}_{j_0}} \{ n^{-1} \mu^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{S} - P_\mathcal{D}) \mu - 2 n^{-1/2} \gamma_\mathcal{D}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} + n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{S} - P_\mathcal{D}) E\}\\
& \leq \min_{j \in \mathcal{J}_0} \left[ n^{-1} \{(\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu)^2 - (\widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } \mu)^2\} - 2 n^{-1/2}\mu^{ \mathrm{\scriptscriptstyle T} }(\widehat{u}_{j_0} \widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } )^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} + n^{-1} E^{ \mathrm{\scriptscriptstyle T} } (\widehat{u}_{j_0} \widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} } )E \right]\\
& \leq \frac{2}{n} \lambda_0 \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} -2 n^{-1/2}\max_{j \in \mathcal{J}_0} \gamma_{\mathcal{S}_0 \cup \{j\}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} + n^{-1}\max_{j \in \mathcal{J}_0} E^{ \mathrm{\scriptscriptstyle T} } (\widehat{u}_{j_0} \widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } - \widehat{u}_j \widehat{u}_j^{ \mathrm{\scriptscriptstyle T} }) E.
\end{aligned}
\end{equation}
Note that $\gamma_{\mathcal{S}_0 \cup \{j\}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} = (\gamma_{\mathcal{S}_0 \cup \{j\}} - \gamma_\mathcal{S})^{ \mathrm{\scriptscriptstyle T} } \widetilde{E}$ and that
\[
\norm{\gamma_{\mathcal{S}_0 \cup \{j\}} - \gamma_\mathcal{S}}_2 \leq n^{-1/2} \left(\abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu} + \abs{\widehat{u}_{j}^{ \mathrm{\scriptscriptstyle T} } \mu}\right) \leq 2 n^{-1/2} \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu}.
\]
Again, an application of Borell-TIS inequality yields
\begin{equation}
\label{eq: linear term lower bound}
\pr \left\{ \max_{j \in \mathcal{J}_0} \gamma_{\mathcal{S}_0 \cup \{j\}}^{ \mathrm{\scriptscriptstyle T} } \widetilde{E} \geq \frac{{\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}}{\sqrt{2}} \sigma \abs{\widehat{u}_{j_0} \mu} \sqrt{\frac{\log (ep)}{n}} - 2 c_\mathcal{T} \sigma \abs{\widehat{u}_{j_0} \mu} \sqrt{\frac{\log (ep)}{n}}\right\} \geq 1- (e p)^{- c_\mathcal{T}^2 /2},
\end{equation}
for any $c_\mathcal{T}>0$. Finally, by Assumption \ref{assumption: anchor point}\ref{item: anchor point}, we have $\abs{\widehat{u}_{j_0} \mu}\geq \xi \sigma\sqrt{\log (ep)}$. Thus, by a similar argument, it can be shown that
\begin{equation*}
\pr \left\{ n^{-1}\max_{\mathcal{D} \in \mathscr{A}_{\mathcal{S}_0}} E^{ \mathrm{\scriptscriptstyle T} } (P_\mathcal{D} - P_{\mathcal{S}})E > c_6 (\mathscr{J}_{\mathcal{G}_{\mathcal{S}_0}} + c_\mathcal{G} )\sigma \abs{\widehat{u}_{j_0}^{ \mathrm{\scriptscriptstyle T} } \mu}\frac{ \sqrt{\log(ep)}}{n}\right\} \leq (ep)^{- c_4 \xi} + 2 (ep)^{- c_5 c_\mathcal{G} \xi}.
\end{equation*}
Thus setting $c_\mathcal{T} = \frac{{\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}}
{4\sqrt{2}} $ we get
\[
\min_{\mathcal{D} \in \mathscr{C}_{j_0}} n^{-1}(R_\mathcal{D} - R_\mathcal{S}) \leq \frac{2}{n} \lambda_0 \abs{\widehat{u}_{j_0} \mu} - \frac{{\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}
}{2\sqrt{2}}\sigma \abs{\widehat{u}_{j_0} \mu} \frac{\sqrt{\log (ep)}}{n} + c_6 (\mathscr{J}_{\mathcal{G}_{\mathcal{S}_0}} + c_\mathcal{G} )\sigma \abs{\widehat{u}_{j_0} \mu} \frac{\sqrt{\log (ep)}}{n}
\]
with probability at least $1 - (ep)^{-({\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}})^2 /32} - (ep)^{-
C^\prime \xi} - (ep)^{- A c_\mathcal{G} \xi}$. Thus one necessary condition is
\[
\lambda_0 \gtrsim \left\{\frac{{\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}
}{2\sqrt{2}} - C (\mathscr{J}_{\mathcal{G}_{\mathcal{S}_0}} + c_\mathcal{G} )\right\} \sigma\sqrt{\log (ep)},
\]
which yields,
\begin{equation}
\widehat{\tau}_{j_0} \gtrsim \left\{\frac{{\sf d}^\circ_\mathcal{U} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}
}{2\sqrt{2}} - C (\mathscr{J}_{\mathcal{G}_{\mathcal{S}_0}} + c_\mathcal{G} )\right\}^2 \frac{\sigma^2 \log (ep)}{n}.
\label{eq: second necessary condition}
\end{equation}
{
\color{blue}
The other condition was:
\begin{equation}
\widehat{\tau}_{j_0}\geq \left\{\frac{\sqrt{b_1^2 + 4 a_1} - b_1}{2}\right\}^2,
\label{eq: first necessary condition}
\end{equation}
where
\[
b_1 = 2 (4\sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T}) \sigma \sqrt{\frac{\log(ep)}{n}},
\]
\[
a_1 = \frac{\sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2}{4} \frac{\log (ep)}{n}.\]
}
\paragraph{Analyzing cases:}
\begin{itemize}
\item if $\mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2 \gtrsim 2(4 \sqrt{2} \mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}} + c_\mathcal{T})^2$ then condition \eqref{eq: first necessary condition} can be written as
\[
\widehat{\tau}_{j_0} \gtrsim \sigma^2 \mathscr{H}_{\mathcal{G}_{\mathcal{S}_0}}^2 \frac{\log(ep)}{n}.
\]
\item If ${\sf d}_\mathcal{U}^\circ \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}\gtrsim (\mathscr{J}_{\mathcal{G}_{\mathcal{S}_0}} + c_\mathcal{G})$ then condition \eqref{eq: second necessary condition} can be written as
\[
\widehat{\tau}_{j_0} \gtrsim \sigma^2 {\sf d}_\mathcal{U}^{\circ 2} \mathscr{H}_{\mathcal{T}_{\mathcal{S}_0}}^2 \frac{\log(ep)}{n}.
\]
\end{itemize}
Thus in certain regimes, the complexity of $\mathcal{T}_{\mathcal{S}_0}$ governs the margin boundary, whereas in other regimes complexity $\mathcal{G}_{\mathcal{S}_0}$ controls the margin condition. This is almost similar to the sufficient condition in \eqref{eq: sufficiency condition for BSS}.
\sapta{
\begin{itemize}
\item Probably $c_\mathcal{T}$ can be replaced by $\mathscr{J}_{\mathcal{T}_{\mathcal{S}_0}}$ and $c_\mathcal{G}$ can be replaced by $\mathscr{J}_{\mathcal{G}_{\mathcal{S}_0}}$ at the cost of a vanishing probability with worse rate. (No problem with that)
\item Here the quantities $\mathscr{H}$'s can be thought of as minimum pairwise separability of the corresponding spaces.
\item Finally, incorporate diameter and minimum pairwise distance to sort of establish a smooth bridge between the complexities.
\end{itemize}
}
\section{Technical lemmas}
\begin{lemma}[\cite{laurent2000adaptive}]\label{lemma: chi_tail_bound}
Let $W$ be chi-squared random variable with degrees of freedom $m$. Then, we have the following large-deviation inequalities for all $x>0$
\begin{equation}
\mathbb{P}(W-m>2\sqrt{mx}+2x)\leq {\sf exp}(-x),\quad and
\label{eq: LM1}
\end{equation}
\begin{equation}
\mathbb{P}(W-m<-2\sqrt{mx})\leq {\sf exp}(-x).
\label{eq: LM2}
\end{equation}
\end{lemma}
If we set $x = mu$ in Equation \eqref{eq: LM1} for $u>0$, then we get
\[
\pr \left(\frac{W}{m} -1 \geq 2 \sqrt{u} + 2 u \right) \leq {\sf exp}(-mu).
\]
Note that for $u <1$, we have $2\sqrt{u} + 2u< 4 \sqrt{u}$. Thus, setting $u = \delta/16$ for any $\delta <1$, we get
\begin{equation}
\label{eq: LM1-simple}
\pr \left( \frac{W}{m} - 1 \geq \delta\right) \leq {\sf exp}(- m \delta^2/16).
\end{equation}
Similarly, setting $x = mu$ and $u = \delta/4$ in Equation \eqref{eq: LM2}, we get
\begin{equation}
\label{eq: LM2-simple}
\pr \left( \frac{W}{m} - 1 \leq -\delta\right) \leq {\sf exp}(- m \delta^2/4).
\end{equation}
\end{appendices}
\section{Introduction}
\label{sec: Introduction}
Consider $n$ independent observations $(x_i, y_i)_{i \in [n]}$ of a random pair $(x, y)$ drawn from the following linear regression model:
\begin{equation}
\begin{split}
(x, \varepsilon)\sim \calP_x\times \calP_\varepsilon,\\
y = x^{ \mathrm{\scriptscriptstyle T} } \beta + \varepsilon,
\end{split}
\label{eq: base_model}
\end{equation}
where $\calP_x$ is the $p$-dimensional isotropic Gaussian distribution ${\sf N}_p(0, {\sf I}_p)$, and $\calP_\varepsilon$ is the standard Gaussian distribution on ${\mathbb{R}}$. In matrix notation, the observations can be represented as
\[
Y = X\beta + E,
\]
where $Y = (y_1, y_2, \ldots, y_n)^{ \mathrm{\scriptscriptstyle T} }, X = (x_1,x_2,\ldots, x_n)^{ \mathrm{\scriptscriptstyle T} }$ and $E = (\varepsilon_1, \varepsilon_2, \ldots, \varepsilon_n)^{ \mathrm{\scriptscriptstyle T} }$. The vector $\beta$ is unknown but sparse in the sense that $\norm{\beta}_0 := \sum_{j =1}^p {\sf ind}(\beta_j \neq 0) = s$, which is much smaller than $p$. Denote by $\calS(v)$ the set of non-zero coordinates of a vector $v\in {\mathbb{R}}^p$. Lastly we denote by $\pr_{\beta_0} (\cdot)$ and $\mathbb{E}_{\beta_0}(\cdot)$ the probability measure and the expectation with $\beta = \beta_0$ respectively. In this paper, we focus on the variable selection problem, i.e., identifying $\calS(\beta)$. We primarily use the 0-1 loss, i.e., $\pr_\beta(\calS(\hat{\beta})\neq \calS(\beta))$, to assess the quality of the selected model $\calS(\hat\beta)$.
{The isotropic Gaussian design has been widely used to conduct precise analysis of variable selection procedures \citep{fletcher2009necessary, genovese2012comparison, ndaoud2020optimal, su2017false}. Specifically, these works either derived the necessary and sufficient condition for exact model recovery \citep{fletcher2009necessary}, or established tight asymptotic bounds of model selection error \citep{genovese2012comparison, ndaoud2020optimal, su2017false}. The isotropic Gaussian design is also used in compressed sensing to generate a measurement matrix \citep{candes2006robust, candes2006near, donoho2006compressed}, so that one can sense the sparse signals with few measurements of the high-dimensional signal.}
Recently there has been growing interest in the variable selection problem in the presence of \textit{weak} and \textit{rare} signals \citep{genovese2012comparison, ji2012ups}.
This regime is ubiquitous in modern data analytics such as those in Genome-Wide Association Study (GWAS). There the
genes that exhibit detectable association with the trait of interest can be extremely few with weak effects \citep{wellcome2007genome, marttinen2013genome}.
Moreover, the number of subjects $n$ typically ranges in thousands, while the number of features $p$ can range from tens of thousands to hundreds of thousands. Such a high dimension further adds to the difficulty of identifying the weak signals. Weak and rare signals also arise in multi-user detection problems \citep{arias2011global} where one typically uses linear model of the form \eqref{eq: base_model}. There the $j$th column of $X$, denoted by $X_j$, is the channel impulse response for user $j$. The signal received from user $j$ is $\beta_j X_j$. Thus $\beta_j =0$ means that $j$th user is not sending any signal. It is a common practice to model the mixing matrix $X$ as random with i.i.d. entries. Under the presence of strong noise, one might be interested in knowing whether information is being transmitted or not. Typically, in some applications it is reasonable to assume that a very few numbers of users are sending signals. Also due to strong noise environment the signals become quite weak, making them harder to detect. Therefore, from an application point of view, understanding variable selection in \textit{weak} and \textit{rare} signal regimes is crucial. Despite its importance, the problem remains fairly underexplored. Typically most of the popular methods including LASSO \citep{tibshirani1996regression}, SCAD \citep{fan2001variable}, MC+ \citep{zhang2010nearly}, etc. have been extensively analyzed in terms of 0-1 loss when the signals are uniformly strong \citep{zhao2006model, guo2015model,zhang2010nearly} in the sense that
\[
\min_{j \in \calS(\beta)}\abs{\beta_j}\gg \left(\frac{\log p}{n}\right)^{1/2}.
\]
However, very little is known about the variable selection problem in the presence of weak and rare signals.
Besides the weakness and rarity of signals, heterogeneity in the signal {\em strength} is another important feature of modern data applications that has not yet received sufficient attention. One limitation of the existing literature on variable selection in the \textit{weak} and \textit{rare} signals regime is that it typically assumes that the true signals are homogeneous \citep{genovese2012comparison, ji2012ups, jin2014optimality}. \cite{ji2012ups} refer to this setup as the \textit{Asymptotically Rare and Weak} (ARW) signal regime.
Many popular approaches have been shown to enjoy satisfactory variable selection properties under the ARW regime.
For instance, \cite{genovese2012comparison} showed that both LASSO and marginal screening enjoy model consistency in terms of Hamming loss under independent random design. \cite{ji2012ups} and \cite{jin2014optimality} investigated the same problem under sparsely correlated design. They proposed two-stage screen and clean algorithms that also exhibit model consistency in terms of Hamming loss.
However, their theory heavily relies on homogeneous signals and does not extend to the heterogeneous case that is of interest to us.
In reality, the ARW setup seldom occurs: the signals almost always have different strengths \citep{li2019weak}. It is not yet fully understood whether one can achieve exact support recovery in the presence of weak, rare {\em and heterogeneous} signals.
To underscore the contrasting effects of homogeneous and heterogeneous signals in terms of exact model recovery, we study the variable selection property of marginal screening (see Section \ref{sec: Thresholding procedures}). We show that under the presence of strong heterogeneity in the signal, marginal screening fails to recover the exact model with probability converging to 1, whereas under homogeneous signal it can recover the exact model asymptotically \citep{genovese2012comparison}. It turns out that due to heterogeneity, the spurious correlations become large and create impediment to selecting the exact model. In correlated design, a different problem known as \textit{unfaithfulness} \citep{wasserman2009high, robins2003uniform} prevents marginal screening from achieving model consistency. Specifically, due to ``correlation cancellation'', marginal correlation between $Y$ and $X_j$ becomes negligible even when $\beta_j$ is large and this ultimately leads to \textit{false negatives}. In this paper, we study independent random design model in which correlation cancellation does not occur. Instead, we identify a different source of problem under the presence of signal heterogeneity that affects exact variable selection performance of marginal screening.
On the computational side, modern methods like LASSO, SCAD, MC+ were initially motivated as alternatives to Best Subset Selection (BSS). BSS is in general an NP-hard optimization problem and was believed to be practically intractable even for $p$ as small as $30$. Thanks to recent advancements in algorithms and hardware, the optimal solution to the BSS problem can now be computed, sometimes with approximations, for some practical settings. \cite{Jain2014iterative} showed that a wide family of iterative hard thresholding (IHT) algorithms can approximately solve the BSS problem, in the sense that they can achieve similar goodness of fit with the best subset with slight violation of the sparsity constraint. \cite{liu2020between} studied the optimal thresholding operator for such iterative thresholding algorithms, which manages to exploit fewer variables than IHT to achieve the same goodness fit as BSS.
\cite{bertsimas2016best} viewed the BSS problem through the lens of mixed integer optimization (MIO) and showed that for $n$ in 100s and $p$ in 1000s, the MIO algorithm can obtain a near optimal solution reasonably fast.
More recently, \cite{bertsimas2020sparse} developed a new cutting plane method that solves to provable optimality the Tikhonov-regularized \citep{Tikhonov1943OnTS} BSS problem with $n$ and $p$ in the 100s and 1000s respectively. A recent work \citep{zhu2020polynomial} proposed an iterative splicing method called Adaptive Best Subset Selection (ABESS) to solve the BSS problem. They also showed that ABESS enjoys both statistical accuracy and polynomial computational complexity when the design matrix satisfies sparse Reisz condition and minimum signal strength is of order $\Omega\{(s \log p \log \log n/n)^{1/2}\}$.
Given these recent advances in solving BSS, there has been growing acknowledgement that BSS enjoys significant statistical superiority over the aforementioned alternative methods. \cite{bertsimas2016best} and \cite{bertsimas2020sparse} numerically demonstrated higher predictive power and lower false discovery rate (FDR) respectively of the BSS solution compared to LASSO. \cite{fan2020best} and \cite{zhu2021early} reported that the approximate BSS solutions provided by IHT have much fewer false discoveries than LASSO, SCAD and SIS, especially in the presence of highly correlated design. They also theoretically showed that the model selection behavior of BSS does not explicitly depend on the restricted eigenvalue condition for the design \citep{bickel2009simultaneous, van2009conditions}, a condition which appears unavoidable (assuming a standard computational complexity conjecture) for any polynomial-time method \citep{zhang2014lower}. This suggests that BSS is robust against design collinearity in terms of model selection.
In this paper, we mainly focus on the \emph{precise} asymptotic bound, i.e., the bound with the optimal constant, for the minimum signal strength that allows BSS to achieve model consistency. Under a specific asymptotic setup, we show that BSS achieves asymptotic exact recovery of the true model once the minimum signal strength is above the information theoretic lower bound, meaning that BSS is optimal in terms of the requirement on the signal strength. In contrast, previous works such as \cite{wainwright2009information} and \cite{fan2020best} analyze BSS from a non-asymptotic perspective: they show that BSS can achieve model consistency under the optimal rate of the sample complexity. But their analyses are tight only up to a multiplicative constants.
The rest of the paper is organized as follows. Section \ref{sec: Ultra rare and weak minimum signal regime} introduces the Asymptotically Ultra-Rare and Weak Minimum signal (AURWM) regime that accommodates heterogeneous signal strengths.
Section \ref{sec: Thresholding procedures}
shows that in the presence of strong heterogeneity of the signal strength, marginal screening procedures fail to achieve model consistency under the AURWM regime with probability converging to 1. In Section \ref{sec: Analysis of ML decoder}, we derive the asymptotic minimax 0-1 loss under the AURWM regime and show that BSS is optimal in terms of the requirement on the minimum signal strength.
In Section \ref{sec: A novel algorithm}, we propose a computationally tractable two-stage algorithm that also enjoys model consistency under essentially the same condition as BSS. Finally, in Section \ref{sec: simulations}, we carry out simulation studies and numerically demonstrate the superiority of our method over other competing methods.
\textbf{Notation.} Let ${\mathbb{R}}$ and ${\mathbb{R}}_+$ denote the set of real numbers and the set of non-negative real numbers respectively. Denote by ${\mathbb{R}}^p$ the $p$-dimensional Euclidean space and by ${\mathbb{R}}^{p\times q}$ the space of real matrices of order $p\times q$. For a positive integer $K$, denote by $[K]$ the set $\{1, 2, \ldots, K\}$.
Regarding vectors and matrices, for a vector $v \in {\mathbb{R}}^p$, we denote by $\norm{v}_2$ the $\ell_2$-norm of $v$. We use ${\sf I}_{p}\in {\mathbb{R}}^{p \times p}$ to denote the $p$-dimensional identity matrix.
For a matrix $A \in {\mathbb{R}}^{p\times p}$, we denote by $A_j$ and $a_j$ the $j$th column and the transposed $j$th row of $A$ respectively. We let ${\sf tr}(A) = \sum_{j=1}^p A_{jj}$ denote the trace of $A$.
Throughout the paper, let $O(\cdot)$ (respectively $\Omega(\cdot)$) denote the standard big-O (respectively big-Omega) notation, i.e., we say $a_n = O(b_n)$ if there exists a universal constant $C>0$, such that $a_n \leq C b_n$ (respectively $a_n\geq C b_n$) for all $n \in \mathbb{N}$. Sometimes for notational convenience, we write $a_n \leq_c b_n$ in place of $a_n = O(b_n)$ and $a_n \gtrsim b_n$ in place of $a_n = \Omega(b_n)$. We write $a_n \sim b_n$ if $a_n = O(b_n)$ and $a_n = \Omega(b_n)$. We denote by $\Omega_{\pr}$ the big-Omega in probability: for a set of random variables $Z_n$ and a set of constants $a_n$, $X_n = \Omega_{\pr}(a_n)$ means that for any $\varepsilon_0>0$, there exist $C_{\varepsilon_0} >0$ and $n_{\varepsilon_0}\in \mathbb{N}$, both of which depend on $\varepsilon_0$, such that
\[
\pr(\abs{X_n/a_n} < C_{\varepsilon_0})\leq \varepsilon_0, \quad \forall n\geq n_{\varepsilon_0}.
\]
We use $\overset{\rm p}{\to}$ and $\overset{\rm d}{\to}$ to denote convergence in probability and distribution respectively. Also we say $X \overset{\rm d}{=} Y$ for two random variables $X,Y$ if their distributions are equal. We denote by ${\sf ind}(\cdot)$ the indicator function.
Finally, regarding probabilistic distributions, we use ${\sf N}(0,1)$ to denote the standard Gaussian distribution. We use ${\sf N}_p(0, \Sigma)$ to denote the $p$-dimensional Gaussian distribution with mean zero and variance-covariance matrix $\Sigma \in {\mathbb{R}}^{p \times p}$. We denote by ${\sf {Ber}}(\pi)$ the Bernoulli distribution with success probability $\pi\in [0,1]$.
\section{Ultra rare and weak minimum signal regime}
\label{sec: Ultra rare and weak minimum signal regime}
In this section, we focus on a specific asymptotic setup that allows \emph{heterogeneity} among the sparse signals in high dimension.
Throughout our paper, we consider the following signal class:
$$
{\mathcal{M}}_{s}^{a}:=\{ \beta \in {\mathbb{R}}^p: \Vert \beta \Vert_0 = s, \min_{j \in \calS(\beta)}\vert\beta_j\vert \geq a \}.
$$
Here $a$ denotes the minimum signal strength of $\beta$. Note that the signal class ${\mathcal{M}}_s^a$ only imposes a lower bound for the minimum signal strength and thus allows arbitrarily large magnitudes across the true signals. This implicitly accommodates heterogeneity in the signal, which is in sharp contrast with the homogeneous signal setup considered by \cite{genovese2012comparison}.
Now we are in a position to introduce the \textit{Asymptotically Ultra Rare and Weak Minimum signal} regime (AURWM), in which we mainly consider the signal class above with
\begin{equation}
a = \left( \frac{2r \log p}{n} \right)^{1/2} \quad
\text{and}\quad s = O(\log p).
\label{eq:aurwm}
\end{equation}
Besides, we set the sample size $n$ as
\[
n = \floor{p^k}, \quad 0 < k<1.
\]
The assumption that $s\leq_c \log p$ characterizes the ultra-rarity of the signals, which is common in genetic studies such as GWAS \citep{yang2020prioritizing}. Unless stated otherwise, from now on our statistical analysis follows the scalings of $n, p, s, a$ in this AURWM regime. We say a support estimator $\widehat{\calS}$ achieves \textit{asymptotic consistent recovery} in the AURWM regime if
\begin{equation}
\lim_{p \to \infty} \sup_{\beta \in {\mathcal{M}}_s^a} \pr_\beta(\widehat{\calS}\neq \calS(\beta) ) = 0.
\label{eq: consistent 0-1 loss recovery}
\end{equation}
This paper mainly focuses on the criterion \eqref{eq: consistent 0-1 loss recovery} to measure the quality of exact recovery performance for an estimator $\widehat{\calS}$.
It is also worth mentioning that a relevant but different asymptotic setup is studied by \cite{genovese2012comparison} and \cite{ji2012ups}. There the authors assumed a Bayesian model such that all the signals are independent and identicially distrbuted and that the sparsity $s\sim p^{1-\vartheta}$ for some $\vartheta \in (0,1)$. Under such a setup they obtained asymptotically tight phase transition boundaries with respect to Bayesian Hamming risk, which partitions the $r \mhyphen \vartheta$ plane into three regions: (a) Region of exact recovery, (b) Region of almost recovery, (c) Region of no recovery. We skip the details of these results for brevity. The major differences between their setup and ours are twofold: (1) They essentially assume homogeneous signals; (2) They assume $s$ to grow in a polynomial fashion with respect to $p$.
\section{Marginal screening under heterogeneous signal}
\label{sec: Thresholding procedures}
Marginal screening (MS) is one of the most widely used variable selection methods in practice. It selects the variables with top absolute marginal correlation with the response. Formally, for any $j \in [p]$, write $\mu_j := X_j^{ \mathrm{\scriptscriptstyle T} } Y/n$. Given any possibly data-driven threshold $\tau(X,Y)$,
define the marginal screening estimator as follows:
\begin{equation}
\widehat{\calS}_{\tau}:= \{ j\in [p]: \abs{\mu_j} \geq \tau(X,Y)\}.
\label{eq: thresholding estimator}
\end{equation}
Note that $\mu_j$ is essentially equivalent to the marginal correlation between $X_j$ and $Y$ because of isotropy of $X$.
Marginal screening has been applied in various fields for feature selection and dimension reduction, including biomedicine \citep{huang2019marginal, lu2005marginal, leisenring1997marginal}, survival data analysis \citep{hong2018conditional, li2016survival}, economics and econometrics \citep{wang2020asset, huang2014feature}.
Besides the broad applications, marginal screening has been shown to enjoy some desirable statistical properties. \cite{fan2008sure} established the sure screening property of marginal screening under an ultra-high dimensional setup, which serves as theoretical justification for MS to be used for dimension reduction in many applications. Later, \cite{genovese2012comparison} showed that MS enjoys the minimax optimal rate under Hamming loss with homogeneous signals. Nevertheless, as mentioned in Section \ref{sec: Introduction}, precise asymptotic characterization of the 0-1 loss of MS remains fairly underexplored under high dimension, especially in the presence of heterogeneity in signal strength.
\subsection{Failure of MS in the AURWM regime}
In this section, we study the 0-1 risk of the MS estimator. Define $\T := \{\widehat \calS_\tau \mid \tau: {\mathbb{R}} ^ {n \times p} \times {\mathbb{R}} ^ n \to {\mathbb{R}}_+ \}$, which is the class of all possible marginal screening estimators. Perhaps surprisingly, under the AURWM regime, we show that MS fails to achieve exact model recovery in the minimax sense.
\begin{theorem}
\label{thm: Thresholding fails in ARMW}
Under the AURWM regime, none of the MS estimators of form \eqref{eq: thresholding estimator} can achieve asymptotic exact recovery, i.e.,
\[
\lim_{p \to \infty}\inf_{\widehat{\calS}_\tau \in \T} \sup_{\beta \in {\mathcal{M}}_s^a} \pr_\beta(\widehat{\calS}_{\tau} \neq \calS(\beta)) = 1.
\]
\end{theorem}
To understand the main message of this theorem, it is instructive to compare it with the parallel result in \cite{genovese2012comparison} with homogeneous signal. Specifically, \cite{genovese2012comparison} consider a Bayesian setup where all the signal coefficients are independent and identically distributed Bernoulli random variables (up to a universal constant). Under the AURWM regime, $s = O(\log p)$, which implies that $\vartheta =1$ in Theorem 10 of \cite{genovese2012comparison}. Then Theorem 10 in \cite{genovese2012comparison} says that when $r > 1$, MS enjoys consistency in terms of Hamming risk and thus 0-1 risk too. In contrast, when we broaden the signal class to ${\mathcal{M}}_s^a$ that embraces possibly heterogeneous signals, the same model consistency fails to hold any more for MS as shown in Theorem \ref{thm: Thresholding fails in ARMW}. This comparison clearly reveals the curse of signal heterogeneity on MS.
To see how signal heterogeneity hurts MS, for any $j \in [p]$, write $\mu_j$ as
\begin{equation}
\mu_j = \frac{ \beta_j \norm{X_j}_2^2}{n} +\frac{ X_j^{ \mathrm{\scriptscriptstyle T} }(\sum_{\ell\neq j} X_\ell \beta_\ell + E)}{n} =: \mu_j^{(1)} + \mu_j ^ {(2)}.
\label{eq: alpha_mr_slector_simplified}
\end{equation}
Here $\mu^{(1)}_j = n ^ {-1}\beta_j \norm{X_j}_2 ^ 2$ represents the marginal contribution from $\beta_j$ to $\mu_j$, and $\mu_j^{(2)}$ represents the random error of $\mu_j$ due to the cross covariance between $X_j$ and the other signals and noise. Suppose there are spiky signals among $\{\beta_\ell\}_{\ell \neq j}$. Though $\E(\mu ^ {(2)}_j) = 0$ regardless of the magnitude of $\beta$, the spiky signals may incur large variance of $\mu_j ^ {(2)}$ and overwhelm the magnitude of $\mu ^ {(1)}_j$, which is the essential indicator of the significance of $\beta_j$. Consequently, one cannot tell if $\beta_j$ is a true variable based on only $\mu_j$ in the presence of spiky signals.
\section{Best subset selection}
\label{sec: Analysis of ML decoder}
Now we shift our focus to BSS, one of the most classical variable selection approaches. With the oracle knowledge of true sparsity $s$, BSS solves for
\[
\hat{\beta}_{\rm best}\in {\arg\min}_{\beta \in {\mathbb{R}}^p, \norm{\beta}_0= s} \norm{Y - X\beta}_2^2.
\]
Define $P_{\D}:= X_{\D}(X_{\D}^{ \mathrm{\scriptscriptstyle T} } X_{\D})^{-1} X_{\D}^{ \mathrm{\scriptscriptstyle T} }$, which is the orthogonal projection operator onto the column space of $X_{\D}$. The BSS above can be alternatively viewed as solving for
\begin{equation}
\widehat{\calS}_{\rm best} := {\arg\min}_{\D \subseteq [p]: \abs{\D}=s} \frac{Y^{ \mathrm{\scriptscriptstyle T} } ({\sf I}_n - P_{\D})Y}{n} = {\arg\max}_{\D \subseteq [p]: \abs{\D}=s} \frac{Y^{ \mathrm{\scriptscriptstyle T} } P_{\D} Y}{n}.
\label{eq: ML decoder}
\end{equation}
Using a union bound as in \cite{wainwright2009information} or \cite{fan2020best}, one can show that there exists a universal positive constant $\varphi$ (approximately equal to 0.618) such that whenever $r> 4/(1- \varphi)$, BSS achieves model consistency, i.e.,
\[
\lim_{p \to \infty}\sup_{\beta \in {\mathcal{M}}_s^a}\pr_\beta(\widehat{\calS}_{\rm best} \neq \calS(\beta)) =0.
\]
We emphasize that the requirement on $r$ here is more stringent than needed: we will show that BSS achieves model consistency whenever $r > 1$, which turns out to be necessary for any approach to obtain exact support recovery.
\subsection{Exact support recovery of BSS }
In the following theorem, we show that
$r>1$ is sufficient for BSS to achieve asymptotic exact recovery. Recall that $n = \floor{p^k}$ with $0 < k < 1$.
\begin{theorem}
\label{thm: tight phase transition boundary for ML}
Let $r>1$ and write $\delta = r-1$. Then there exists a universal positive constant $C_0$ such that whenever
$$ s< C_0 \min \bigg\{2k, \frac{4 \delta}{1+ \delta}, \frac{\delta^2}{\{(1+ 0.75\delta)^{1/2}+ (1+ 0.5\delta)^{1/2}\}^{2}} \bigg\}\log p,$$ we have
\[
\lim_{p\to \infty} \sup_{\beta\in {\mathcal{M}}_s^a} \pr_\beta (\widehat{\calS}_{\rm best} \neq \calS(\beta))=0.
\]
\end{theorem}
In order for BSS to achieve model consistency, we need to ensure that the maximum spurious correlation, i.e., correlation between the spurious variables and the response, is well controlled so that the best subset does not involve any false discovery.
One important ingredient of our analysis is the asymptotic distribution of the maximum spurious correlation due to \cite{fan2018discoveries}, based on which we can derive the sharp constant in the minimum signal strength for BSS to be model-consistent.
It is worth emphasis that pursuing the exact asymptotic distributions is crucial to obtain constant-sharp results; typically, standard non-asymptotic analysis can only yield optimal rates rather than optimal constants.
Note also that Theorem \ref{thm: tight phase transition boundary for ML} requires $s$ to grow slowly. Given that we have at least $\binom{p-s}{s}$ spurious models and that this number increases with respect to $s$ when $s$ is small, a larger $s$ implies higher maximum spurious correlation due to randomness and thus thinner chance for the best subset to remain the true model.
In the next section, we show that $r > 1$ is the weakest possible requirement on the minimum signal strength for any approach to achieve asymptotic model consistency.
\subsection{Necessary condition for exact recovery}
In this section, we show that under the AURWM regime, $r=1$ is the information theoretic boundary of the minimum signal strength for exact model recovery. In other words, if $r<1$, then no method can achieve model consistency. Towards this end, we study the minimax 0-1 loss $$\inf_{\widehat{\calS}}\sup_{\beta \in {\mathcal{M}}_s^a} \pr_\beta(\widehat{\calS}\neq \calS(\beta)),$$
where the infimum is taken over the class of all possible measurable functions $\widehat{\calS}: (X, Y)\to \{ \D\subseteq [p]: \abs{\D}=s\}$.
The next theorem establishes a lower bound of the above minimax 0-1 loss.
\begin{theorem}[Necessary Condition]
\label{thm: Information theoretic boundary}
Under the AURWM regime, when $r < 1$, there exists a universal positive constant $c$ such that
\[
\lim_{p\to \infty}\inf_{\widehat{\calS}} \sup_{\beta \in {\mathcal{M}}_a^s} \pr_\beta(\widehat{\calS} \neq \calS(\beta)) \geq c.
\]
\end{theorem}
Theorem \ref{thm: Information theoretic boundary} suggests that $r\geq 1$ is a necessary condition for consistent support recovery. Combining this with Theorem \ref{thm: tight phase transition boundary for ML}, we can see that BSS is optimal in terms of requirement on the minimum signal strength to achieve exact model recovery.
The proof of Theorem \ref{thm: Information theoretic boundary} leverages Theorem 1 in \cite{wang2010} and detailed proof can be found in Section 9.1 of the supplementary material.
\section{Achieving statistical optimality with computational efficiency}
\label{sec: A novel algorithm}
In spite of the optimality of BSS in terms of model selection, its NP-hardness seriously restricts its practical applicability. To address the computational issue, we propose a two-stage algorithm called ETS (Estimate then Screen) that combines the well-known IHT algorithm \citep{blumensath2009iterative, Jain2014iterative} with a follow-up coordinate screening step. We show that in the AURWM regime, ETS enjoys the same selection optimality as BSS in terms of the requirement on the minimum signal strength.
\subsection{The ETS algorithm}
\label{sec: IHT based ETS algorithm}
In this section we introduce our ETS algorithm (Algorithm \ref{alg: ETS}) in detail. Given a partition parameter $0 < \gamma < 1$, ETS first splits the full sample $(x_i, Y_i)_{i \in [n]}$ into two subsamples $\D_1, \D_2$ of respective sizes $n_1 = \floor{\gamma n}$ and $n_2 = n - n_1$. Then ETS performs two main steps on these two sub-samples respectively. It applies IHT to $\D_1$ to obtain an estimator $\hat{\beta}^{\rm iht}$ of the true signal vector $\beta$. Then ETS performs a coordinatewise screening based on $\D_2$ and $\hat{\beta}^{\rm iht}$ to select the true variables. To be self-contained, in Algorithm \ref{alg: two-stage IHT} we give the pseudo-code of the IHT algorithm. Let $f: {\mathbb{R}}^p \to {\mathbb{R}}, \theta \mapsto f(\theta)$ be a differentiable objective function and
$P_{\hat{s}}^0(\cdot)$ be the projection operator that projects a vector onto the space of $\hat{s}$-sparse vectors by selecting the $\hat{s}$ largest elements in magnitude.
\begin{algorithm}[h]
\SetAlgoLined
Input: Objective function $f$, sparsity level $\hat s$, step size $h$ \;
$\beta^{(0)} =0$\; $t =0$ \;
\While{not converged}{
$ \beta^{(t+1)} = P_{\hat{s}}^0(\beta^{(t)} - h \nabla_\theta f(\beta^t))$\;
$t\leftarrow t+1$
}
Output: $\hat{\beta}^{\rm iht} = \beta^{(t)}$.
\caption{IHT}
\label{alg: two-stage IHT}
\end{algorithm}
For $\ell \in \{1, 2\}$, let $X^{(\ell)} \in {\mathbb{R}}^{n_\ell \times p}$ and $Y^{(\ell)} \in {\mathbb{R}} ^ {n_{\ell}}$ denote the design matrix and the response vector of the $\ell$th sub-sample respectively. ETS computes $\hat{\beta}^{\rm iht}$ based on the first sub-sample $(X^{(1)},Y^{(1)})$ with the square loss as the objective function:
\begin{equation}
f_{n_1}(\theta; X^{(1)}, Y^{(1)}) := \frac{1}{n_1}\Vert Y^{(1)} - X^{(1)} \theta\Vert_2^2.
\label{eq: empirical_1_square_error_loss}
\end{equation}
Next comes the screening step of ETS. For each $i \in [p]$, define
\begin{equation}
\label{eq:delta}
\Delta_i := \frac{
X_i^{(2){ \mathrm{\scriptscriptstyle T} }} \left( Y^{(2)} - \sum_{j\neq i} X_{ j }^{(2)} \hat{\beta}^{\rm iht}_j \right) }{ \Vert X_i^{(2)} \Vert_2}
\end{equation}
and
\begin{equation}
\kappa_{\varsigma}(u) := \frac{a\norm{u}_2}{2} + \frac{\varsigma^2 \log p}{a \norm{u}_2}, \quad \forall u \in {\mathbb{R}}^{n_2},
\label{eq: iht_hat_eta_threshold}
\end{equation}
where $\varsigma > 0$ is specified later.
ETS selects the $i$th variable if and only if $\abs{\Delta_i}> \kappa_{\varsigma}(X_i^{(2)})$.
To see why we can screen variables based on $\{\Delta_i\}_{i \in [p]}$, note that
\begin{equation}
\Delta_i = \beta_i \Vert X_i^{(2)}\Vert_2 + \frac{X_i^{(2){ \mathrm{\scriptscriptstyle T} }}\bigl( \sum_{j \neq i} X_j^{(2)} (\beta_j - \hat{\beta}_j^{\rm iht}) + E \bigr)}{\Vert X_i^{(2)}\Vert_2}.
\label{eq: alpha_distribution}
\end{equation}
Some probability argument shows that conditioned on $\D_1$ and $X_i^{(2)}$, $\Delta_i$ is distributed as:
$$ \Delta_i \,\bigr\vert\, \big(\D_1, X_{i}^{(2)}\big) \overset{\rm d}{=} \beta_i \big\Vert X_i^{(2)}\big\Vert_2 + \bigg\{1 + \sum_{j \neq i}(\beta_j - \hat{\beta}_j^{\rm iht})^2 \bigg\}^{1/2} g_i, $$
where $g_i \sim {\sf N}(0,1)$ and is independent of $X_i^{(2)}$. If IHT performs well in the sense that $\Vert\hat{\beta}^{\rm iht} - \beta\Vert_2$ is small, then for all $i \in \calS(\beta)$, $\beta_i \Vert X_i^{(2)}\Vert_2$ becomes the dominant term in $\Delta_i$. In contrast, for all $i \notin \calS(\beta)$, $\beta_i \Vert X_i^{(2)}\Vert_2 = 0$ and we thus expect $\Delta_i$ to be small. This suggests the existence of a threshold $t(\cdot)$ on $(\Delta_i)_{i \in [p]}$ that distinguishes the true support $\calS(\beta)$ from the irrelevant variables. We follow \cite{ndaoud2020optimal} to choose the threshold function in \eqref{eq: iht_hat_eta_threshold}, which is shown to be a reasonable choice to identify the true variables.
For each $i \in [p]$, define $\hat{\eta}_i(X,Y) := {\sf ind} \{
\abs{\Delta_i}> \kappa_{\varsigma}(X_i^{(2)})
\}$
and write $$\hat{\eta}(X,Y) := (\hat{\eta}_1(X,Y), \ldots, \hat{\eta}_p(X,Y)) ^ { \mathrm{\scriptscriptstyle T} }.$$
The selector $\hat{\eta}(X,Y)$ is the final estimate of the support $\mathcal{S}(\beta)$ produced by the ETS algorithm. Algorithm \ref{alg: ETS} shows the detailed steps of the ETS algorithm.
\begin{algorithm}[h]
\SetAlgoLined
Input: Data $\mathcal{D} = \{(x_i, Y_i)\}_{i=1}^n$, sparsity level $\hat s$, step size $h$, partition parameter $\gamma$ , threshold parameter $\varsigma$\;
1. Randomly partition the whole dataset $\mathcal{D}$ into two disjoint subsets $\mathcal{D}_1= (X^{(1)}, Y^{(1)})$ and $\mathcal{D}_2 = (X^{(2)}, Y^{(2)})$ \;
2. Run IHT (Algorithm \ref{alg: two-stage IHT}) with the objective function $f_{n_1}(\theta; X^{(1)}, Y^{(1)})$ in \eqref{eq: empirical_1_square_error_loss} and compute $\hat{\beta}_{\rm iht}$\;
3. Construct the statistics $\{\Delta_i\}_{i=1}^p$ and thresholds $\{\kappa_\varsigma(X_i^{(2)})\}_{i=1}^p$ using \eqref{eq:delta}-\eqref{eq: iht_hat_eta_threshold}\;
4. Finally compute the selector $\hat{\eta}(X,Y)$\;
Output: The selector $\hat{\eta}(X.Y)$.
\caption{ETS}
\label{alg: ETS}
\end{algorithm}
Next, we define the binary decoder of the true support $\calS(\beta)$ as
$\eta_\beta := ({\sf ind}\{\beta_1 \neq 0\}, \ldots, {\sf ind}\{ \beta_p \neq 0\}) ^ { \mathrm{\scriptscriptstyle T} }$. The next theorem shows that ETS can achieve exact recovery under suitable choices of the tuning parameters.
\begin{theorem}
Assume $r> 1$ and write $\delta = r - 1$. Then there exist universal positive constants $A_1, A_2$ {and suitable choices of tuning parameters $\hat{s}$, $h$} such that with IHT iteration count $t \ge \log\{A_1 (1 + \|\beta\|_2 ^ 2)/ \delta\}$, $\gamma \in (0, \delta/(8+ 8\delta))$ and $\varsigma = (1 + A_2\delta)^{1/2} $, we have that $\lim_{p\to \infty}\sup_{\beta\in {\mathcal{M}}_{a}^s} \pr_\beta (\hat{\eta}\neq \eta_\beta) =0.$
\label{thm: IHT-scrrening exact recovery rate}
\end{theorem}
Note that as the signal strength parameter $r$ approaches the information-theoretic boundary, i.e., as $\delta$ approaches 0, ETS requires more iterations to achieve model consistency. This is not surprising: intuitively, weaker signals are harder to identify than strong ones.
Besides, ETS does not require the knowledge of the true sparsity $s$, but requires the knowledge of $a$ in the second stage for accurate screening. If the true sparsity $s$ is known, then we can enforce ETS to select exactly $s$ features as follows. Let $\abs{\Delta}_{(m)}$ denote the $m$th largest value of $\{\abs{\Delta_i}\}_{i \in [p]}$. For each $i \in [p]$, define
\begin{equation}
\hat{\eta}_i (X, Y; s) = {\sf ind}\{\abs{\Delta_i} \geq \abs{\Delta}_{(s)}\}.
\label{eq: selector_ETS_tilde}
\end{equation}
Hence $\hat{\eta}(s): = (\hat{\eta}_1(s), \ldots, \hat{\eta}_p(s))^{ \mathrm{\scriptscriptstyle T} }$ selects exactly $s$ features.
The following corollary shows that under the same conditions of Theorem \ref{thm: IHT-scrrening exact recovery rate}, $\hat\eta(s)$ achieves model consistency.
\begin{corollary}
Assume $r > 1$ and write $\delta = r - 1$. If $t \ge \log\{A_1 (1 + \|\beta\|_2 ^ 2)/ \delta\}$ with the same $A_1$ in Theorem \ref{thm: IHT-scrrening exact recovery rate} and $\gamma \in (0, \delta/(8+ 8\delta))$, we have that
\begin{equation}
\lim_{p \to \infty} \sup_{\beta \in {\mathcal{M}}_s^a} \pr_\beta (\hat{\eta}(s) \neq \eta_\beta) = 0.
\label{eq: model_consistecy_ETS_tilde}
\end{equation}
\label{cor: model_consistecy_ETS_tilde}
\end{corollary}
Detailed proofs of Theorem \ref{thm: IHT-scrrening exact recovery rate} and Corollary \ref{cor: model_consistecy_ETS_tilde} can be found in Section 9.2 and Section 9.3 of supplementary materials respectively. Besides, the sparsity level $\hat{s}$ and the step-size $h$ are chosen based on certain geometric properties of $f_{n_1}(\theta)$ and details can be found in Section 9.2 of the supplementary materials.
\section{Numerical experiments}
\label{sec: simulations}
In this section, we first numerically investigate the probability for MS to achieve exact recovery of the true model with growing ambient dimension $p$ under both homogeneous and heterogeneous signal setup. Our results show that while MS exhibits model consistency under the homogeneous signal regime, it completely fails to do so under the heterogeneous signal regime, which is consistent with Theorem \ref{thm: Thresholding fails in ARMW}.
We then conduct simulation experiments to demonstrate superiority of ETS over competing methods including LASSO and MS as signal strength grows or signal heterogeneity grows. To this end, we mention that we do not numerically compare {\em exact} BSS in this section mainly due to computational issues. In most of our simulation setups we consider $p$ in thousands and BSS generally suffers from high computational costs in such regimes, even with the help of modern optimization tools such as Gurobi \citep{hastie2020best}. Instead, we focus on LASSO and ETS, both of which are two different computational surrogates of the BSS problem.
\subsection{Exact recovery performance of MS}
In Figure \ref{fig: prop_recovery_path_plot_ms}, we demonstrate the asymptotics of MS under both homogeneous and heterogeneous signal patterns. We consider $p \in \{ 1000, 2000, \ldots, 8000\}$ and signal strength parameter $r \in \{2,3,4,5,6\}$. We set $s = \floor{2 \log p}$ and $n = \floor{p^{0.9}}$. We let $\tau(X, Y)$ in \eqref{eq: thresholding estimator} be equal to the $s$th largest value of $\{\abs{\mu_1}, \ldots, \abs{\mu_p}\}$, so that MS always chooses a model of size $s$.
For the homogeneous signal setup, we consider $\beta$ with $\norm{\beta}_0 =s$ and $\beta_j = a$ for all $j \in \calS(\beta)$, where $a$ is defined in \eqref{eq:aurwm}. This implies that the SNR varies between 0.19 and 2.15 across different choices of $(r,p)$. For the heterogeneous signal setup, we consider $\beta$ with $(s-1)$ active coordinates equal to $a$ and one ``spiky'' coordinate equal to $\{10 - (s-1)a^2\}^{1/2}$. This ensures that the SNR is fixed at 10 for all choices of $r,p$.
Figure \ref{fig: prop_recovery_path_plot_ms}(a) shows that under homogeneous signal MS is able to recover the exact model with probability converging to 1 as $p$ grows. In contrast, Figure \ref{fig: prop_recovery_path_plot_ms}(b) shows that under heterogeneous signal MS never achieves exact model recovery: plots for all values of $r$ are at level 0. Such a contrast corroborates Theorem \ref{thm: Thresholding fails in ARMW}: signal spikes can give rise to substantial spurious correlation and jeopardize the accuracy of MS.
\begin{figure}[h]
\centering
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/ms_asymp/prop_recovery_ms.pdf}
\caption{Proportion of exact recovery under weak homogeneous signal. ($0.19 \leq \text{SNR} \leq 2.15$)}
\end{subfigure}\hfill
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/ms_asymp/prop_recovery_ms_hetero.pdf}
\caption{Proportion of exact recovery under heterogeneous signal. (SNR = 10)}
\end{subfigure}
\caption{Asymptotics of MS with growing dimension $p$.}
\label{fig: prop_recovery_path_plot_ms}
\end{figure}
\subsection{Effect of growing signal strength}
Here we numerically compare the probability of exact support recovery of ETS with those of LASSO and MS as signal strength parameter $r$ grows. We investigate both homogeneous and heterogeneous signal patterns. We set $p =2000$, $s \in \{13, 52\}$ and $n = \floor{p^{0.9}} = 935$. We set signal strength parameter $r \in \{1.5, 2, 2.5, \ldots, 9\}$ in \eqref{eq:aurwm}. The support $\calS$ is chosen uniformly over all the size-$s$ subsets of $[p]$, and each support coordinate of $\beta$ is chosen as follows:
\[
\beta_j =(1 - b_j) (1 + Z_j^2/n)^{1/2} a + b_j r^{1/2} \quad \forall j\in \calS,
\]
where $(Z_j)_{j \in \calS} \overset{i.i.d.}{\sim} {\sf N}(0,1)$, and where $(b_j)_{j \in \calS} \overset{i.i.d.}{\sim} {\sf {Ber}}(\pi)$ with $\pi \in \{0,0.2\}$. $\pi =0$ corresponds to the homogeneous signal pattern, and $\pi =0.2$ corresponds to the heterogeneous signal pattern, where spiky signals are present with probability $0.2$. Each entry $x_{ij}$ of the design matrix $X$ is generated independently from ${\sf N}(0,1)$.
In this experiment, we grant all the approaches with the knowledge of $s$, so that the comparison is fair. Using this oracle knowledge, we only look at the solutions of the aforementioned three methods with sparsity exactly equal to $s$. Specifically, for LASSO, we look at the solution path and select the model of size exactly equal to $s$. For MS, we just select the top $s$ variables corresponding to the largest absolute values of $\mu$'s.
For ETS, we do not split data for estimation and screening separately; instead we use the full data in both steps. Specifically, we replace $X ^ {(1)}$ and $Y ^ {(1)}$ with $X$ and $Y$ respectively in \eqref{eq: empirical_1_square_error_loss} and replace $X ^ {(2)}$ and $Y ^ {(2)}$ with $X$ and $Y$ respectively in \eqref{eq:delta}. We set gradient step size $h=0.5$ in IHT. We choose projection size $\hat{s}$ by cross validation in terms of mean squared prediction error. Lastly, for selecting exactly $s$ features we use \eqref{eq: selector_ETS_tilde} in the screening stage of ETS. It is worthwhile to mention that from an application point of view, incorporating data splitting in ETS is not necessary as we are only interested in identifying the active signals, which is akin to point estimation. Also, given the fact that $n \ll p$ in high dimensional regime, using full sample in both the estimation and screening step delivers greater sample efficiency and provides better inference.
Next, for each choice of $ r$, we run LASSO, MS and ETS over 200 independent Monte Carlo experiments to compute the empirical probability of exact recovery. Figure \ref{fig: propo_recovery_path_plot} presents the results. We make the following important observations:
\begin{enumerate}
\item All the three methods enjoy higher chance of exact support recovery as the signal strength grows;
\item MS completely fails to achieve exact support recovery when $s$ becomes large (compare panels (a) and (c)) or the signal becomes heterogeneous (compare panels (a) and (b)).
\item LASSO and ETS are insensitive to heterogeneity of the signal. However, LASSO suffers from larger sparsity, while ETS is much more robust against it.
\item Overall, ETS is the best among all the three methods in terms of exact support recovery.
\end{enumerate}
\begin{figure}[h]
\centering
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/iht_nopart/prop_recovery_ms_homo_s13.pdf}
\caption{Homogeneous signal pattern $(\pi =0, s= 13)$}
\end{subfigure}\hfill
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/iht_nopart/prop_recovery_ms_hetero_s13.pdf}
\caption{Heterogeneous signal pattern $( \pi =0.2, s=13)$}
\end{subfigure}\hfill
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/iht_nopart/prop_recovery_ms_homo_s52.pdf}
\caption{Homogeneous signal pattern $(\pi =0, s= 52)$}
\end{subfigure}\hfill
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/iht_nopart/prop_recovery_ms_hetero_s52.pdf}
\caption{Heterogeneous signal pattern $(\pi =0.2, s= 52)$}
\end{subfigure}
\caption{Plot of proportion of exact recovery for varying $r$.}
\label{fig: propo_recovery_path_plot}
\end{figure}
\subsection{Effect of growing heterogeneity}
In this numerical experiment we study the effect of growing heterogeneity on ETS, LASSO and MS. We set $p = 2000, s = 13$, $n = \floor{p^{0.9}} = 935$ and $r\in \{2,6\}$ in \eqref{eq:aurwm}. Next, we introduce ${\mathsf{n}_{\rm spike}}$, the number of ``spiky'' signals in $\calS$. We vary ${\mathsf{n}_{\rm spike}}$ in $\{0\} \cup [6]$. The case ${\mathsf{n}_{\rm spike}} = 0$ corresponds to the homogeneous signal setup where the true signals are set as $a$ uniformly. For ${\mathsf{n}_{\rm spike}}>0$, we randomly set $(s- {\mathsf{n}_{\rm spike}})$ signals in $\calS$ to be equal to $a$ and the remaining signals to be equal to ${a_{\rm spike}}$, which is defined as
$$
{a_{\rm spike}} := \bigg\{ \frac{(2 - sa^2)}{{\mathsf{n}_{\rm spike}}} + a^2\bigg\}^{1/2}.
$$
Such a choice of ${a_{\rm spike}}$ ensures that the SNR always equals 2 whenever ${\mathsf{n}_{\rm spike}} > 0$. We again perform ETS, LASSO and MS over 200 Monte Carlo simulations for each choice of $r$ and ${\mathsf{n}_{\rm spike}}$ to obtain the empirical probability of exact support recovery. Similarly to the previous sections, we assume that the true sparsity $s$ is known and we apply the three methods in the same fashion as before.
Figure \ref{fig: prop_recovery_path_plot_varying_spike} shows again the detrimental effect of heterogeneity on MS in terms of exact recovery. In both panels we see a significant drop in the proportion of exact recovery for MS when ${\mathsf{n}_{\rm spike}}$ changes from 0 to 1. This is consistent with the theory in Section \ref{sec: Thresholding procedures}. However, in Figure \ref{fig: prop_recovery_path_plot_varying_spike}(b) we see that the proportion of exact recovery is slowly increasing as ${\mathsf{n}_{\rm spike}}$ grows from 1 to 6. This is because as ${\mathsf{n}_{\rm spike}}$ increases, ${a_{\rm spike}}$ monotonically decreases, so that the signals become more homogeneous. MS is then able to recover the exact model more frequently. We do not see a similar phenomenon in Figure \ref{fig: prop_recovery_path_plot_varying_spike}(a) because ${a_{\rm spike}}$ is too large. Another important observation is that while ETS and LASSO are both performing nearly perfectly when $r =6$, ETS significantly outperforms both LASSO and MS when $r = 2$. Therefore, ETS is again the overall winner.
\begin{figure}[h]
\centering
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/varying_spike/prop_recovery_varying_spike_r2_s13.pdf}
\caption{$r=2$}
\end{subfigure}\hfill
\begin{subfigure}{0.47\linewidth}
\includegraphics[width=\linewidth]{figures2/varying_spike/prop_recovery_varying_spike_r6_s13.pdf}
\caption{$r=6$}
\end{subfigure}\hfill
\caption{Plot of proportion of exact recovery with varying ${\mathsf{n}_{\rm spike}}$.}
\label{fig: prop_recovery_path_plot_varying_spike}
\end{figure}
\section{Conclusion}
In this paper, we study exact support recovery in high-dimensional sparse linear regression with independent Gaussian design. We focus on the AURWM regime that not only accommodates \textit{rare} and \textit{weak} signals as the ARW regime does, but also allows \textit{heterogeneity} in the signal strength.
Our first theoretical result (Theorem \ref{thm: Thresholding fails in ARMW}) shows that marginal screening fails to achieve exact support recovery under the AURWM regime.
The main reason is that the presence of ``spiky'' signals increases the maximum spurious marginal correlation, thereby blinding the marginal screening procedure to weak signals. Therefore, one needs to be cautious with usage of marginal screening for variable selection in practice.
In contrast, we show that BSS is robust to signal spikes and is able to achieve model consistency under the AURWM regime with the optimal requirement on signal strength (Theorem \ref{thm: tight phase transition boundary for ML}). The primary reason behind this is that unlike MS, BSS takes into account multiple features simultaneously and thus selects variables based on their capability of fitting the residualized responses given the other variables rather than the responses themselves. Therefore, spiky signals do not affect BSS: They are very likely to be in plausible candidate models in the first place and their effect on the response has been removed in the residualization procedure. Given the recent computational advancements in solving BSS, our positive result on BSS makes it more appealing from an application point of view.
However, it is worth mentioning that even with modern advances in optimization, BSS suffers from high computational costs when the ambient dimension is high.
To address this issue, we propose a computationally tractable two-stage method ETS that delivers essentially the same optimal exact recovery performance as BSS (Theorem \ref{thm: IHT-scrrening exact recovery rate}). Similar to BSS, ETS seeks for the features that exhibit high explanation power for the residuals from the model that excludes these features themselves (see \eqref{eq:delta}). Therefore, ETS is robust to spiky signals. This fact together with the slowly growing sparsity condition in \eqref{eq:aurwm} yields the optimal exact recovery accuracy of ETS.
Our work naturally raises several important questions for future research. One question is whether similar optimality results hold for BSS and ETS when the sparsity $s$ grows faster than $\log p$. The same question can also be asked for \textit{correlated} random design. Another direction of our interest is studying the problem of exact recovery in a \textit{distributed} setting where data are stored at different places and communication between them is restricted.
\section{Proofs}
\label{sec: some important proofs}
In this section we collect the proofs of some important results of this paper. Specifically, we present the proofs of Theorem \ref{thm: Thresholding fails in ARMW} and Theorem \ref{thm: tight phase transition boundary for ML}. The first one shows that marginal screening fails to achieve exact recovery in the AURWM regime, whereas BSS achieves model consistency under optimal requiremnt on the signal strength.
\subsection{Proof of Theorem \ref{thm: Thresholding fails in ARMW}}
\label{sec: Proof of thm: Thresholding fails in ARMW }
Consider a MS procedure $\widehat{\calS}_\tau \in \T$. To start with note that
\[
\pr_\beta(\widehat{\calS}_\tau = \calS(\beta) ) = \pr_\beta\left(\max_{j \in S^c} \abs{\mu_j} < \tau(X,Y) \leq \min_{j \in S} \abs{\mu_j} \right)\leq \pr \left(\max_{j \in S^c} \abs{\mu_j} < \min_{j \in S} \abs{\mu_j} \right).
\]
Recall that for for $j \in [p]$ we have
$$
\mu_j =
\begin{cases}
\beta_j \norm{X_j}^2_2/n + \omega_j \norm{X_j}_2 g_j/n & \text{if $j \in \calS(\beta)$}\\
\omega_j \norm{X_j}_2 g_j/n & \text{if $j \notin \calS(\beta)$},
\end{cases}
$$
where $\omega_j^2 = 1 + \sum_{k \neq j}\beta_k^2 $ and $g_j= X_j^{ \mathrm{\scriptscriptstyle T} }(\sum_{k\neq j} X_k \beta_k + E
) /(\omega_j \norm{X_j}_2)\sim {\sf N}(0,1) $. Thus we have
\begin{equation}
\pr_\beta(\widehat{\calS}_\tau = S) \leq \pr_\beta\left(\max_{j \in S^c}{\omega_j \norm{X_j}_2 \abs{g_j}}/(2n \log p )^{1/2} < \min_{j\in S} \vert\beta_j \norm{X_j}_2^2 + \omega_j \norm{X_j}_2 g_j\vert/( 2 n \log p )^{1/2} \right).
\label{eq: upper bound accuracy}
\end{equation}
Right hand side of Equation \eqref{eq: upper bound accuracy} does not depend on $\widehat{\calS}_\tau$ hence the above inequality is valid uniformly over the class $\T$. Now choose a sequence $\{c_p\}_{p=1}^\infty$ such that $\lim_{p \to \infty}c_p^2/r \geq 1$ .
Next construct a sequence of $\beta^{(p)}$ in the following manner,
\begin{itemize}
\item Consider the set $\calS_0 = \{1,\ldots, s\} \subseteq [p]$ with $s=O(\log p)$.
\item Set $\beta_{1}^{(p)} = c_p$. For all other $i\in \calS_0\setminus \{1 \}$ set $\beta_i^{(p)} = a = (2r (\log p)/n)^{1/2}$.
\item Set $\beta_i^{(p)} =0$ if $i \notin \calS_0$.
\end{itemize}
In this setup we have $\omega_j \sim (1 +c_p^2)^{1/2}$ for all $j \neq 1$.
Now fix $k_0\in \calS_0\setminus\{1\} $ (say $k_0 = 2$). From Equation \eqref{eq: upper bound accuracy} it can be concluded that
\begin{align*}
&\sup_{\widehat{\calS}_\tau \in \T}\pr_{\beta^{(p)}}(\widehat{\calS}_\tau = \calS_0) \\
& \leq \pr_{\beta^{(p)}}\left(\max_{j \in S^c}{\omega_j \norm{X_j}_2 \abs{g_j}}/(2n \log p )^{1/2} < \vert \beta_{k_0} \norm{X_{k_0}}_2^2 + \omega_{k_0} \norm{X_{k_0}}_2 g_{k_0} \vert/(2 n \log p )^{1/2} \right).
\end{align*}
An elementary calculation shows that $\{g_j\}_{j\in \calS_0^c}$ are i.i.d. standard Gaussian. Also note that $\omega_j > (1+c_p^2)^{1/2}$ for all $j \in \calS_0^c$. Using these facts and lemma 3 from \cite{fletcher2009necessary} we get
\[
\frac{1}{(1 + c_p^2)^{1/2}}\max_{j \in \calS_0^c} \omega_j \frac{\norm{X_j}_2 \abs{g_j}}{(2n \log p)^{1/2}}\geq \min_{j \in \calS_0^c} \frac{\norm{X_j}_2}{n^{1/2}} \max_{j \in \calS_0^c} \frac{\abs{g_j}}{(2 \log p)^{1/2}} \overset{\rm p}{\to} 1.
\]
But recall that $\beta_{k_0} = \{2r (\log p)/n\}^{1/2}$ and thus we have the following:
$$ \pr( \vert \beta_{k_0} \norm{X_{k_0}}_2^2 + \omega_{k_0} \norm{X_{k_0}}_2 g_{k_0}\vert / \{(1 + c_p^2) (2 n \log p )\}^{1/2} <1) \to 1 .$$ This tells that,
\[
\lim_{p \to \infty} \sup_{\widehat{\calS}_\tau \in \T} \inf_{\beta \in {\mathcal{M}}_{s}^a} \pr_{\beta} (\widehat{\calS}_\tau = \calS(\beta))\leq \lim_{p\to \infty}\sup_{\widehat{\calS}_\tau \in \T}\pr_{\beta^{(p)}}(\widehat{\calS}_\tau = \calS_0) = 0.
\]
This finishes the proof.
\subsection{Proof of Theorem \ref{thm: tight phase transition boundary for ML}}
\label{sec: proof of ind-design}
In this proof we repararmetrize $\delta$ by $8\delta_0$ for algebraic convenience. The main result can be salvaged by back substituting $8\delta_0$ by $\delta$ in all the main equations in this section. Also, for brevity of notation, in this proof we use $\widehat{\calS}$ and $\calS$ to denote that oracle BSS estimator and $\calS(\beta)$ respectively. Recall that BSS is defined as
\[
\widehat{\calS} = {\arg\max}_{\D, \abs{\D}=s} \norm{P_{\D} Y}_2^2 = {\arg\min}_{\D:\abs{\D}=s} \norm{({\sf I}_n-P_{\D})Y}_2^2.
\]
Thus from the above definition we have the following equality:
\begin{align*}
\pr(\widehat{\calS}\neq \calS) & = \pr \left(\norm{P_{\calS} Y }_2^2 < \max_{\D\neq \calS} \norm{P_{\D} Y}_2^2 \right).
\end{align*}
Now we will try to understand how the quantity $\norm{P_{\calS} Y}_2^2$ behaves asymptotically. First it is easy to see that $P_{\calS} Y = \sum_{j \in \calS}X_j \beta_j + P_{\calS} E$. Note that $\sum_{j \in \calS} X_j \beta_j = \norm{\beta}_2 \tilde{\varepsilon} $ where $\tilde{\varepsilon} \sim {\sf N}_n(0, {\sf I}_{ n})$ and independent of the noise $z$. Hence we up with the following:
\begin{align*}
\norm{P_{\calS} Y }_2^2& = \Big\Vert\sum_{j \in S}X_j \beta_j \Big\Vert_2^2 + 2 E^{ \mathrm{\scriptscriptstyle T} } P_{\calS} \left( \sum_{j \in S} X_j \beta_j\right) + E^{ \mathrm{\scriptscriptstyle T} } P_{\calS} E\\
& = \Big\Vert\sum_{j \in S}X_j \beta_j \Big\Vert_2^2 + 2 E^{ \mathrm{\scriptscriptstyle T} } \left(\sum_{j \in S} X_j\beta_j \right) + E^{ \mathrm{\scriptscriptstyle T} } P_{\calS} E\\
& = \norm{\beta}_2^2 \norm{\tilde{\varepsilon}}_2^2 + 2 \norm{\beta}_2 E^{ \mathrm{\scriptscriptstyle T} } \tilde{\varepsilon} + E^{ \mathrm{\scriptscriptstyle T} } P_{\calS} E.
\end{align*}
Recall that $\abs{\beta_j} \geq \{2r (\log p) /n\}^{1/2}$ for all $j \in \calS$. This is presumably the hardest setup as increasing signal strength can only decrease the error probability. Then $\norm{\beta}_2^2 \geq (2rs \log p)/n$. also note that $E^{ \mathrm{\scriptscriptstyle T} } P_{\calS} E\sim \chi^2_s$. Hence we have ,
\[
\frac{\norm{P_{\calS} Y }_2^2}{s \log p} \geq 2r \frac{\norm{\tilde{\varepsilon}}_2^2}{n} + 2 \left(\frac{2r}{s \log p}\right)^{1/2} \frac{\tilde{\varepsilon}^{ \mathrm{\scriptscriptstyle T} } E}{n^{1/2}} + \frac{E^{ \mathrm{\scriptscriptstyle T} } P_{\calS} E}{s \log p} \overset{\rm p}{\longrightarrow} 2r .
\]
Thus $\lim_{p \to \infty}\pr(\widehat{\calS}\neq \calS)\leq \pr \left ( 2r \leq \limsup_{p\to \infty}\max_{\D \neq \calS} \norm{P_{\D} Y}_2^2/(s \log p) \right) $. The limiting behaviour of the obtained maximal process turns out to be very challenging to analyze and hence we do not directly study this maximal process. Instead we focus on a related maximal process (will be defined shortly) derived from the earlier one and we use the results from \cite{fan2018discoveries} to study its asymptotic behaviour. Now let us denote the set $\calS \cap \D$ by $\I_0$ and $\D \setminus \calS$ by $\I_1$, i.e., $\D = \I_0 \cup \I_1$. Next define the class $\J_{\I_0} = \{ \I_1 \subseteq [p]: \I_1 \cap S = \emptyset, \abs{\I_1 \cup \I_0} = s \}$ for each $\I_0\subset S$. Note that $0\leq \abs{\I_0}\leq s-1$ from the construction (if $\abs{\I_0}=s$ then $\D= \calS$). The random variable of interest can be rewritten as follows:
\[
\max_{\D\neq \calS} \frac{\norm{P_{\D} Y}_2^2}{s \log p} = \max_{\I_0: \I_0 \subset \calS}\; \max_{\I_1: \I_1 \in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} Y}_2^2}{s \log p}.
\]
Using union bound we get,
\begin{equation}
\pr(\widehat{\calS}\neq \calS)\leq \sum_{\I_0\subset \calS} \pr \left(\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} Y}_2^2}{s \log p} > \frac{\norm{P_{\calS} Y}_2^2}{s \log p} \right).
\label{eq: 0-1_loss_decompositon_constant_sparsity}
\end{equation}
Now fix a subset $\I_0$ of the true support $\calS$. Similar to previous section define $\tilde{Y} = Y - X_{\I_0}\beta_{\I_0}$ and this independent of the features in $\I_0 \cup \I_1$. Also we have
\begin{align*}
&\norm{P_{\I_0 \cup \I_1} Y }_2^2 = \norm{P_{\I_0 \cup \I_1} \tilde{Y}}_2^2 + \norm{X_{\I_0} \beta_{\I_0}}_2^2 + 2 \beta_{\I_0}^{ \mathrm{\scriptscriptstyle T} } X_{\I_0}^{ \mathrm{\scriptscriptstyle T} } \tilde{Y},\\
&\hspace{.78cm}\norm{P_{\calS} Y }_2^2 = \norm{P_{\calS} \tilde{Y}}_2^2 + \norm{X_{\I_0} \beta_{\I_0}}_2^2 + 2 \beta_{\I_0}^{ \mathrm{\scriptscriptstyle T} } X_{\I_0}^{ \mathrm{\scriptscriptstyle T} } \tilde{Y}.
\end{align*}
Thus the summands in the right hand side of \eqref{eq: 0-1_loss_decompositon_constant_sparsity} can be written as the probability of the event $\{\max_{\I_1: \I_1\in \J_{\I_0}} \norm{P_{\I_0 \cup \I_1} g}_2^2/(s\log p) > \norm{P_{\calS} g}_2^2/(s \log p) \}$, where $g := (1 + \norm{\beta_{\calS \setminus\I_0}}_2^2)^{-1/2} \tilde{Y}$. Note that $g\sim {\sf N}_n(0,{\sf I}_n)$ and is independent of the features in $\mathcal{D}$.
Now fix a specific $\I_0$. In the analysis we encounter the maximal process
\[
\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0\cup \I_1} g}_2^2}{s\log p},
\]
Now consider the set of indices $F_{\I_0} = (\{ 1,\cdots, p\}\setminus \calS)\cup \I_0$. Hence it is easy to see that $\tilde{p}:=\abs{F_{\I_0}} = p -s + \abs{\I_0}$. Without loss of generality, let $F_{\I_0} = \{1, \ldots, \tilde{p}\}$. Also define $\tilde{s}:= s - \abs{\I_0}$. Let the set $\V_{\I_0} = \{\alpha \in {\mathbb{R}}^p: \norm{\alpha}_0= s, \norm{\alpha}_2=1, \I_0\subseteq \calS(\alpha), \alpha_{F_{\I_0}^c} = 0\}$. Here $\alpha_J$ denotes the sub-vector of $\alpha$ corresponding to the indices in $J \subseteq [p]$. Next we will focus on the random variable,
\begin{equation}
\widehat{L}_n:=\widehat{L}_n(\tilde{s},\tilde{p})= \sup_{\alpha \in \V_{\I_0}} \frac{1}{n^{1/2}} \sum_{i=1}^n \frac{\alpha^{ \mathrm{\scriptscriptstyle T} } (g_i x_i)}{(\alpha^{ \mathrm{\scriptscriptstyle T} } \widehat{\boldsymbol{\Sigma}}_n \alpha)^{1/2}},
\label{eq: hat_Ln}
\end{equation}
here $\widehat{\boldsymbol{\Sigma}}_n =\frac{1}{n} \sum_{i=1}^n x_i x_i^{ \mathrm{\scriptscriptstyle T} }$. Now recall that $\D = \I_0 \cup \I_1$ for all $\D\neq \calS$ with $\abs{\D} = s$. To see the connection, first note that the above optimization problem can be viewed as the following:
\begin{align*}
\widehat{L}_n &= \max_{\I_1\in \J_{\I_0}}\max_{\alpha \in \V_{\I_0 \cup \I_1}} \frac{\alpha_{\D}^{ \mathrm{\scriptscriptstyle T} } (\sum_{i=1}^n g_i x_{i,\D}/n^{1/2}) }{\{\alpha_{\D}^{ \mathrm{\scriptscriptstyle T} } (\widehat{\boldsymbol{\Sigma}}_{n,\D \D}) \alpha_{\D}\}^{1/2} }\\
&= \max_{\I_1\in \J_{\I_0}} \left\{(\sum_{i=1}^n g_i x_{i,\D}/n^{1/2})^{ \mathrm{\scriptscriptstyle T} } \widehat{\boldsymbol{\Sigma}}^{-1}_{n,\D \D}(\sum_{i=1}^n g_ix_{i,\D}/n^{1/2}) \right\}^{1/2}\\
&= \max_{\I_1\in \J_{\I_0}} \{g^{ \mathrm{\scriptscriptstyle T} } X_{\D} (X_{\D} ^{ \mathrm{\scriptscriptstyle T} } X_{\D})^{-1} X_{\D}^{ \mathrm{\scriptscriptstyle T} } g\}^{1/2}\\
& = \max_{\I_1\in \J_{\I_0}} \norm{P_{\I_0\cup \I_1
} g}_2.
\end{align*}
Thus it is essential to study the asymptotic property of $\widehat{L}_n$. Now we define the standardized version of $\widehat{L}_n$ as follows
\[
L_n :=L_n(\tilde{s},\tilde{p}) = \sup_{\alpha \in \V_{\I_0}} \frac{1}{n^{1/2}} \sum_{i=1}^n \alpha^{ \mathrm{\scriptscriptstyle T} } (g_i x_i).
\]
Let $\boldsymbol{Z}=(Z_1,\cdots, Z_{\tilde{p}})$ be $\tilde{p}-$variate Gaussian random variable with covariance matrix $I_{\tilde{p}\times \tilde{p}}$ and define the random variable $T^*:= T^*(\tilde{s},\tilde{p}) = \sup_{\alpha \in \V_{\I_0}} \alpha_{F_{\I_0}}^{ \mathrm{\scriptscriptstyle T} } \boldsymbol{Z}$.
\begin{lemma}
There exists universal constants $K_0,K_1$ such that for any $\delta_1\in(0, K_0 K_1]$,
\begin{equation}
\abs{L_n - T^*}\leq_c n^{-1} c_n^{1/2}(\tilde{s},\tilde{p}) + K_0 K_1 n^{-3/2} c_n^2(\tilde{s},\tilde{p}) +\delta_1
\label{eq: approximating-L_n&T*}
\end{equation}
holds with probability at least $1- C\Delta_n(s, \tilde{p};\delta_1)$ where $c_n(s,\tilde{p})= s \log(e \tilde{p}/s)\vee \log n$ and
\[
\Delta_n(\tilde{s},\tilde{p};\delta_1) = (K_0 K_1)^3 \frac{\{ \tilde{s} b_n(\tilde{s},\tilde{p})\}^2}{\delta_1^3 n^{1/2}} + (K_0 K_1)^4 \frac{\{ \tilde{s} b_n(\tilde{s},\tilde{p})\}^5}{\delta_1^4 n}
\]
with $b_n(\tilde{s}, \tilde{p}) = \log(\tilde{p}/\tilde{s})\vee \log n$.
\label{lemma: approximating-L_n&T*}
\end{lemma}
\begin{lemma}
Assume that the sample size satisfies $n\geq C_1(K_0\vee K_1)^4 c_n(\tilde{s},\tilde{p})$. then with probability at least $1- C_2 n^{-1/2} c_n^{1/2}(\tilde{s},\tilde{p})$,
\begin{equation}
\abs{\widehat{L}_n - L_n}\leq_c (K_0\vee K_1)^2 K_0 K_1 n^{-1/2} c_n(\tilde{s},\tilde{p}),
\label{eq: approximating-L_n&L_n_hat}
\end{equation}
where $c_n(\tilde{s},\tilde{p}) = \tilde{s}\log(e \tilde{p}/\tilde{s})\vee \log n$.
\label{lemma: approximating-L_n&L_n_hat}
\end{lemma}
Proof of the above two lemmas are omitted as it is in the same line of the proofs of \cite{fan2018discoveries}.
Now applying Lemma \ref{lemma: approximating-L_n&T*} and \ref{lemma: approximating-L_n&L_n_hat} with $$\delta_1 = \delta_1(s,\tilde{p}) = (K_0 K_1)^{3/4} \min[1, n^{-1/8}\{ \tilde{s} b_n(\tilde{s},\tilde{p})^{3/8}\}]$$ yields that with probability at least $1- C (K_0K_1)^{3/4} n^{-1/8} \{s b_n(s,\tilde{p})\}^{7/8}$,
\[
\abs{\widehat{L}_n - T^*}\leq_c (K_0K_1)^{3/4} n^{-1/8} \{\tilde{s} b_n(\tilde{s},\tilde{p})\}^{3/8}.
\]
Together with Lemma 2.3 from \cite{chernozhukov2014gaussian} we can conclude that
\begin{equation}
\sup_{t\in {\mathbb{R}}} \abs{\pr(\widehat{L}_n \leq t) - \pr(T^*\leq t)}\leq_c C (K_0K_1)^{3/4} n^{-1/8} \{\tilde{s} b_n(\tilde{s},\tilde{p})\}^{7/8}.
\label{eq: KS_distance-L_n&T*}
\end{equation}
Next by the definition of $T^*$ it follows that
$$T^{*2} = \max_{\I_1\in \J_{\I_0}} \norm{\boldsymbol{Z}_{\I_0 \cup \I_1}}_2 =\sum_{j \in \I_0} Z_j^2 + \max_{\I_1\in \J_{\I_0}} \sum_{k \in \I_1} Z_k^2.$$
Let $\boldsymbol{W}\sim {\sf N}_{(p-s)}(0, {\sf I}_{(p-s)})$ be a Gaussian vector independent of $\boldsymbol{Z}$. Thus it follows that
\[
T^{*2}\overset{d}{=} \sum_{j \in \I_0} Z_j^2 + \sum_{k= p-2s+\abs{\I_0}+1}^{p-s} W^2_{(k:p-s)}\leq \sum_{j \in \I_0} Z_j^2 + (s- \abs{\I_0}) W^2_{(p-s:p-s)}.
\]
From Equation \eqref{eq: KS_distance-L_n&T*} it also follows that
\begin{equation}
\sup_{t\geq 0} \abs{\pr(\widehat{L}^2_n \leq t) - \pr(T^{*2}\leq t)}\leq_c C (K_0K_1)^{3/4} n^{-1/8} \{\tilde{s} b_n(\tilde{s},\tilde{p})\}^{7/8}.
\label{eq: KS_distance- square_L_n&T*}
\end{equation}
Now from the assumption we have $r = 1+ 8 \delta_0$. Assume that $$ s\leq 0.5\min\{\delta_0, \frac{2\delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2}\} \log p.$$ Hence $\abs{\I_0}\leq s -1 \leq \delta_0 \log p$. Thus we have,
\begin{align*}
\scriptstyle
&\quad \pr(T^{*2} > 2(1+ 4\delta_0) (s-\abs{\I_0}) \log p) \\
&= \pr(\sum_{j \in \I_0} Z_j^2 + (s- \abs{\I_0})W^2_{(p-s:p-s)}> 2(1+ 4\delta_0) (s-\abs{\I_0}) \log p)\\
&\leq \pr \left(\sum_{j \in \I_0} Z_j^2 > 4\delta_0 (s-\abs{\I_0}) \log p \right) + \pr \left((s- \abs{\I_0}) W^2_{(p-s:p-s)} > 2(1+ 2\delta_0)(s-\abs{\I_0}) \log p \right)\\
& \overset{(a)}{\leq} \pr \left( \frac{\sum_{j \in \I_0} Z_j^2 - \abs{\I_0}}{\abs{\I_0}} > (4 \delta_0\log p - \abs{\I_0})/\abs{\I_0} \right) + \pr\left( W^2_{(p-s:p-s)} > 2(1+ 2\delta_0)\log p \right)\\
& \overset{(b)}{\leq} \pr \left( \frac{\sum_{j \in \I_0} Z_j^2 - \abs{\I_0}}{\abs{\I_0}} > (3 \delta_0\log p)/\abs{\I_0} \right) + \pr\left( W^2_{(p-s:p-s)} > 2(1+ 2\delta_0)\log p \right)\\
& \leq_c {\sf exp}(- 0.75\delta_0 \log p) + (p - s)\pr\left( W^2_1 > 2(1+ 2\delta_0)\log p \right)\\
& \leq_c p^{-0.75\delta_0} + C \frac{p^{-2\delta_0}}{\sqrt{\log p}}.
\end{align*}
Inequality $(a)$ uses $s - \abs{\I_0}\geq 1$ and inequality $(b)$ uses $\abs{\I_0}<s< \delta_0 \log p$.
Now define the event $ \calE_{\I_0} := \{ \norm{P_{\calS} {g}}_2^2/s\log p > 2(1 + 4\delta_0)R_{\I_0}\}$ where $R_{\I_0}:= (s- \abs{\I_0})/s$. Recall that
\begin{equation}
\begin{split}
\frac{\norm{P_{\calS} {g}}_2^2}{s\log p} &\geq \frac{\norm{ \frac{ \sum_{j \in \calS \setminus
\I_0} X_j \beta_j + P_{\calS} E }{(1 + \Vert\beta_{\calS \setminus \I_0}\Vert_2^2)^{1/2}} }_2^2}{s \log p}\\
& \geq \frac{ \left\{ \frac{ \norm{\sum_{j \in \calS \setminus
\I_0} X_j \beta_j }_2}{(1 + \Vert\beta_{\calS \setminus \I_0}\Vert_2^2)^{1/2}} -
\frac{\norm{ P_{\calS} E}_2}{(1 + \Vert\beta_{\calS \setminus \I_0}\Vert_2^2)^{1/2}} \right\}^2 }{s \log p} = (T_1^{1/2}- T_2^{1/2})^2.
\end{split}
\label{eq: E_I_0 lower bound}
\end{equation}
where $$ T_1:=\frac{ \norm{\sum_{j \in \calS \setminus
\I_0} X_j \beta_j }_2^2}{(1 + \norm{\beta_{\calS \setminus \I_0}}_2^2)s\log p}\geq \frac{2r R_{\I_0}}{1 + 2r (s- \abs{\I_0}) (\log p)/n} \frac{V_n}{n},$$ and $$V_n:= \frac{\Vert \sum_{j \in \calS \setminus
\I_0} X_j \beta_j \Vert_2^2} {\Vert \beta_{\calS \setminus \I_0}\Vert_2^2}$$ is an $\chi^2_n$ random variable. Also we have $$ T_2:=\frac{\norm{ P_{\calS} E}^2_2}{(1 + \norm{\beta_{\calS \setminus \I_0}}_2^2) s\log p}\leq V_s/s\log p$$ where $V_s:= \norm{P_{\calS} E}_2^2$ is an $\chi^2_s$ random variable independent of $X_S$. Next we state the following simple algebraic relationship:
\[
(1+ 6\delta_0)^{1/2} - \frac{\delta_0}{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}}\geq (1+ 4\delta_0)^{1/2}.
\]
In light of Equation \eqref{eq: E_I_0 lower bound} and using the above algebraic inequality we have the following:
\[
\calE_{\I_0}^c \subseteq \{ T_1 \leq 2(1 + 6\delta_0) R_{\I_0} \} \bigcup \{T_2 \geq 2 \delta_0^2 \{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^{-2} R_{\I_0}\}
\]
Next we have
\[
\pr(T_1 \leq 2(1+ 6\delta_0)R_{\I_0}) \leq \pr \left(\frac{V_n}{n} \leq \frac{1 + 6\delta_0}{1 + 8\delta_0}(1 + 2rs \log p/n)\right).
\]
Now choose large $n$ such that $(1 + 6\delta_0)(1 + 2rs \log p/n))< (1+ 7\delta_0)$. Then for large $n$ we have,
$$
\pr(T_1 \leq 2(1+ 6\delta_0)R_{\I_0})\leq \pr\left(\abs{V_n/n-1}\geq \frac{\delta_0}{1+ 8\delta_0} \right) \leq_c {\sf exp} \left(-C^* \frac{\delta_0}{1+ 8\delta_0} n \right),
$$
where $C^*$ is an universal constant.
Now we analyze the quantity $T_2$. We have
\begin{align*}
&\pr(T_2 \geq \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2} R_{\I_0})\\
& \leq \pr(V_s/s\geq \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2} R_{\I_0}\log p)\\
&\leq \pr(V_s\geq \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2} \log p)\\
&\leq \pr(\abs{V_s/s-1}\geq 0.5 \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2} (\log p)/s)\\
& \leq {\sf exp}(- C^{\prime} \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2} \log p)\\
& = p^{-C^{\prime} \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2}}.
\end{align*}
Ultimately it shows that $\pr(\calE_{\I_0}^c)\leq_c {\sf exp} \left(-C^* \frac{\delta_0}{1+ 8\delta_0} n \right)+ p^{-C^{\prime} \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2}}$.
Now we are ready to show that the error probability goes to $0$.
\begin{align*}
&\pr_\beta(\widehat{\calS}\neq \calS) \leq \sum_{\I_0\subset S} \pr_\beta \left(\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} Y}_2^2}{s \log p} > \frac{\norm{P_{\calS} Y}_2^2}{s \log p} \right)\\
&\leq \sum_{k =0}^{s-1}\sum_{\I_0: \abs{\I_0}=k} \pr_\beta \left(\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} Y}_2^2}{s \log p} > \frac{\norm{P_{\calS} Y}_2^2}{s \log p} \right)\\
&\leq
\sum_{k=0}^{s-1} \sum_{\I_0: \abs{\I_0}=k} \pr_\beta \left(\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} Y}_2^2}{s \log p} > \frac{\norm{P_{\calS} Y}_2^2}{s \log p} ,\calE_{\I_0}\right) + \pr(\calE_{\I_0}^c)\\
&
\leq
\sum_{k=0}^{s-1} \sum_{\I_0: \abs{\I_0}=k} \pr_\beta\left(\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} g}_2^2}{s \log p} > \frac{\norm{P_{\calS} g}_2^2}{s \log p} ,\calE_{\I_0}\right) + \pr(\calE_{\I_0}^c)\\
& \leq
\sum_{k=0}^{s-1} \sum_{\I_0: \abs{\I_0}=k} \pr_\beta\left(\max_{\I_1: \I_1\in \J_{\I_0}} \frac{\norm{P_{\I_0 \cup \I_1} g}_2^2}{s \log p} > 2(1 + 4\delta_0)R_{\I_0}\right) + \pr(\calE_{\I_0}^c)\\
& \overset{(a)}{\leq_c}
\sum_{k=0}^{s-1} \sum_{\I_0: \abs{\I_0}=k}\left[ \pr\left(T^{*2} > 2(1 + 4\delta_0)s R_{\I_0} \log p\right) + \pr(\calE_{\I_0}^c) + C (K_0K_1)^{3/4} n^{-1/8} \{s b_n(s,p)\}^{7/8}\right]\\
&
\leq_c \sum_{k=0}^{s-1} \sum_{\I_0: \abs{\I_0}=k} p^{-0.75\delta_0} + C \frac{p^{-2\delta_0}}{\sqrt{\log p}} + {\sf exp} \left(-C^* \frac{\delta_0}{1+ 8\delta_0} n \right)+ p^{-C^{\prime} \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2}}\\
& + n^{-1/8} \{s b_n(s,p)\}^{7/8} \\
& \leq_c \sum_{k=0}^{s-1} \binom{s}{k} \left[ C_{\delta_0} \frac{p^{-2\delta_0}}{\sqrt{\log p}} + {\sf exp} \left(-C^* \frac{\delta_0}{1+ 8\delta_0} n \right)+ p^{-C^{\prime} \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2}}
+ n^{-1/8} \{s b_n(s,p)\}^{7/8}\right]\\
& \leq_c 2^s \left[ C_{\delta_0} \frac{p^{-2\delta_0}}{\sqrt{\log p}} + {\sf exp} \left(-C^* \frac{\delta_0}{1+ 8\delta_0} n \right)+ p^{-C^{\prime} \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2}}+ n^{-1/8} \{s b_n(s,p)\}^{7/8}\right].
\end{align*}
Inequality $(a)$ uses $ \tilde{s} b_n(\tilde{s},\tilde{p})\leq s b_n(s,p)$ for large $p$. Thus if $$ s\leq_c \left( \frac{\delta_0}{1+ 8\delta_0} \wedge \frac{2 \delta_0^2}{\{(1+ 6\delta_0)^{1/2}+ (1+ 4\delta_0)^{1/2}\}^2} \wedge \frac{k}{16} \right)\log p $$, then error probability goes to 0 uniformly over $\beta \in {\mathcal{M}}_s^a$.
\bibliographystyle{apalike}
\section{Appendix}
\label{sec: appendix_01}
\subsection{Proof of Theorem \ref{thm: Information theoretic boundary}}
We first present a result form \cite{wang2010} which is gives us necessary condition for asymptotic exact recovery.
\begin{theorem}[\cite{wang2010}]
Consider the model \eqref{eq: base_model} with the design matrix $X\in {\mathbb{R}}^{n \times p}$ be drawn with i.i.d elements from any distribution with zero mean and unit variance. Let $a:= \min_{j \in \calS(\beta)}\abs{\beta_j}$, i.e., it denote the minimum signal strength of $\beta$. Define the function
\[
f_m(p,s, a) := \frac{\log \binom{p-s+ m}{m} -1}{\frac{1}{2} \log\left( 1 + m a^2 (1 - \frac{m}{p-s+m})\right)},\quad 1\leq m \leq s.\]
Then $n \geq \max\{ f_1(p,s,a), \ldots, f_s(p,s,a),s\}$ is necessary for asymptotic exact recovery.
\label{thm: wang information bound}
\end{theorem}
In the light of Theorem \ref{thm: wang information bound} the proof of Theorem \ref{thm: Information theoretic boundary} is follows immediately. To see this note that if $r<1$ then there exists $\alpha \in (0,1)$ such that $r= 1- \alpha$. Also recall that $a = \{2r (\log p)/n\}^{1/2}, s= O(\log p)$ and $n = \floor{p^k}$. Note that $f_1(p,s,a)/n \sim \frac{1}{1-\alpha}$. This shows that asymptotically the necessary condition in above theorem is violated and hence $r\geq 1$ is necessary.
\subsection{Proof of Theorem \ref{thm: IHT-scrrening exact recovery rate}}
\label{sec: Proof of main ETS}
We first introduce some standard assumptions analyzing ETS.
\begin{definition}[RSC property]
A differentiable function $F: {\mathbb{R}}^p \to {\mathbb{R}}$ is said to satisfy restricted strong convexity (RSC) at sparsity level $s= s_1+s_2$ with strong convexity constraint $\ell_s$ if the following holds for all $\theta_1,\theta_2$ s.t. $\norm{\theta_1}_0\leq s_1$ and $\norm{\theta_2}_0\leq s_2$:
\[
F(\theta_1) - F(\theta_2)\geq \innerprod{\theta_1- \theta_2, \nabla_\theta F(\theta_2)} + \frac{\ell_s}{2}\norm{\theta_1- \theta_2}_2^2.
\]
\label{def: RSC_property}
\end{definition}
\begin{definition}[RSS property]
A differentiable function $F: {\mathbb{R}}^p \to {\mathbb{R}}$ is said to satisfy restricted strong smoothness (RSS) at sparsity level $s= s_1+s_2$ with strong smoothness constraint $L_s$ if the following holds for all $\theta_1,\theta_2$ s.t. $\norm{\theta_1}_0\leq s_1$ and $\norm{\theta_2}_0\leq s_2$:
\[
F(\theta_1) - F(\theta_2)\leq \innerprod{\theta_1- \theta_2, \nabla_\theta F(\theta_2)} + \frac{L_s}{2}\norm{\theta_1- \theta_2}_2^2.
\]
\label{def: RSS_property}
\end{definition}
Now we quote an important theorem from \cite{Jain2014iterative} which quantifies the sub-optimality gap of Algorithm \ref{alg: two-stage IHT}.
\begin{theorem}[\cite{Jain2014iterative}]
Let $F$ has RSC and RSS parameters given by $\ell_{2 \hat{s} + s}(F)= \alpha$ and $L_{2 \hat{s} + \hat{\pi}}(F)= L$ respectively. Call Algorithm \ref{alg: two-stage IHT} with $\hat{s} \geq 32 L^2 \ell^{-2} s$ and $h = 2/(3L)$. Also let $\hat{\beta}= {\arg\min} _{\theta, \norm{\theta}_0\leq s} F(\theta)$. Then $t{\rm th}$ iterate of Algorithm \ref{alg: two-stage IHT} for $t = O(L \ell^{-1} \log (F(\beta^{(0)})/\epsilon))$ satisfies:
\[
F(\beta^{(t)}) - F(\hat{\beta}) \leq \epsilon.
\]
\label{thm: two_stage_IHT_sub_optimality}
\end{theorem}
In our setup, the observations $\{x_i\}_{i=1}^n$ are coming from i.i.d. mean zero isotropic Gaussian distribution. Thus, lemma 6 from \cite{agarwal2012fast} immediately tells that RSC and RSS at any sparsity level $m$ hold for $f_{n_1}(\cdot)$ with probability at least $1- {\sf exp}(- c_0 n_1)$ with $\ell_m = \frac{1}{2} - c_1 (m \log p)/n_1$ and $L_m = 2 + c_1 (m \log p)/n_1$, where $c_0,c_1$ are universal constants. Now set $m = 2 \hat{s} + s$ and recall that $n_1 \sim \gamma p^k$. If $n_1 > 4c_1 (2 \hat{s} + s) \log p$ then we have $\ell_{m} \geq 1/4$ and $L_{m}\leq 9/4$, which means that $L_m/(9\ell_m)\leq 1$. Thus to apply Theorem \ref{thm: two_stage_IHT_sub_optimality} it is enough to choose $\hat{s} = 2592 s$.
Also by the assumption on $n$ for large $p$ we have $n_1> 4c_1 (2 \hat{s}+ s) \log p$. Let $f_{n_1}(\theta): = n_1^{-1} \Vert Y^{(1)} -X^{(1)}\theta \Vert_2^2$ for $\theta \in {\mathbb{R}}^{p}$. Note that $f_{n_1}(0) = n_1^{-1}\Vert Y^{(1)}\Vert_2^2 \overset{\rm d}{=} (1 + \norm{\beta}_2^2) V_{n_1}/n_1$, where $V_{n_1}$ is chi-square random variable with $n_1$ degrees of freedom. Also by Bernstein's type inequality it follows that $\abs{(V_{n_1}/n_1) -1}\leq 1/2$ with probability at least $1 - {\sf exp}(-c_4 n_1)$, where $c_4$ is a universal positive constant. Thus if $t = O(L_m \ell_m^{-1} \log ((1 + \norm{\beta}_2^2)/\epsilon))$, then we have $f_{n_1}(\beta^{(t)}) - f_{n_1}(\hat{\beta})\leq \epsilon$. Thus by Theorem 3 of \cite{Jain2014iterative} it follows that with probability at lest $1 - {\sf exp}(-c_0 n_1) - {\sf exp}(- c_4 n_1) - c_2 p^{-c_3}$ ($c_2, c_3$ are universal constants) we have
\[
\Vert\beta^{(t)} - \beta\Vert_2 \leq C \left \{\left(\frac{s \log p}{n_1} \right)^{1/2} + \epsilon^{1/2} \right\},
\]
for some universal positive constant $C$.
In practice one can use the full data in both steps of ETS but for theoretical convenience we use data splitting in the algorithm. The key idea here is to use one subsample for obtaining $\hat{\beta}^{\rm iht}$ and the remaining subsample in the screening step. Due to this $\hat{\beta}^{\rm iht}$ becomes independent of the second subsample, thus making the analysis of ETS easy.
For notational brevity we write $\eta$ instead of $\eta_\beta$. Define the event $\his = \left\{ \Vert \hat{\beta}^{\rm iht} - \beta \Vert_2 \leq \xi_p \right\}$, where $\xi_p = C \left\{ (\frac{s \log p}{n_1})^{1/2} + \epsilon^{1/2} \right\}\leq (A \epsilon)^{1/2}$ (for some universal constant $A>0$) and $\hat{\beta}^{\rm iht}$ is based on the subsample $\D_1$. From Theorem 3 of \cite{Jain2014iterative}, it follows that $\pr(\his^c) \leq {\sf exp}(-c_0 n_1) + {\sf exp}(- c_4 n_1) + c_2 p^{-c_3}$.
Next, for algebraic convenience we again reparametrize $\delta$ as $8 \delta_0$ and set $\epsilon = 6 \delta_0/A, \varsigma = (1 + A\epsilon)^{1/2}$. Now note that for any $\beta \in {\mathcal{M}}_{a}^s$, we have
\begin{align*}
\pr_\beta (\hat{\eta} \neq \eta \vert \D_1) &\leq \sum_{j:\beta_j=0} \pr( \hat{\eta}_j =1, \his \vert \D_1) + \sum_{j: \beta_j \neq 0} \pr_\beta (\hat{\eta}_j \neq 1, \his\vert \D_1) + \pr(\his^c\vert \D_1)\\
& = \sum_{j: \beta_j =0 } \pr_\beta (\abs{\Delta_j} > \kappa_\varsigma(X_j^{(2)}), \his\vert \D_1) + \sum_{j: \beta_j \neq 0 } \pr_\beta (\abs{\Delta_j} \leq \kappa_\varsigma(X_j^{(2)}), \his\vert \D_1) + \pr(\his^c \vert \D_1),
\end{align*}
Using the fact that conditionally on $\hat{\beta}$ and $X_j$, the random variable $\Delta_j$ has the same distribution as \eqref{eq: alpha_distribution}, we conclude that for all $j \notin \calS(\beta)$ ,
\begin{align*}
\pr(\eta_j = 1, \his \vert \D_1)& \leq \pr \left( (1 + \xi_p^2)^{1/2}\abs{g_j}> \frac{a \norm{X_j^{(2)}}_2}{2} + \frac{(1+ A \epsilon) \log p}{a \norm{X_j^{(2)}}_2} , \his \big \vert \D_1 \right)\\
& \leq \pr \left( (1 + A \epsilon)^{1/2}\abs{g_j}> \frac{a \norm{X_j^{(2)}}_2}{2} + \frac{(1+ A \epsilon) \log p}{a \norm{X_j^{(2)}}_2} , \his \big \vert \D_1 \right)
\\
& = 2 \E \left \{ \overline{\Phi}\left ( \frac{a \norm{X_j^{(2)}}_2}{2 (1 + A\epsilon)^{1/2}} + \frac{(1+ A \epsilon)^{1/2} \log p}{a \norm{X_j^{(2)}}_2}\right) \right\}.
\end{align*}
Here $\overline{\Phi}(\cdot)$ denotes the survival function of standard Gaussian random variable. Now note that for each $j$ we have $\norm{X_j}_2^2 \overset{\rm d}{=} V_{n_2}$, where $V_{n_2}$ is a chi-squared random variable with $n_2$ degrees of freedom. Thus we have
$$\pr (\eta_j = 1, \his \vert \D_1) \leq 2 \E \left \{ \overline{\Phi}\left ( \frac{a V_{n_2}^{1/2}}{2 (1 + A\epsilon)^{1/2}} + \frac{(1+ A \epsilon)^{1/2} \log p}{a V_{n_2}^{1/2}}\right) \right\}.$$
Analogous argument and the fact that $\abs{\beta_j}\geq a$ for all $\beta_j \neq 0$, leads to the fact that for all $j\in \calS(\beta)$,
\[
\pr(\eta_j \neq 1, \his \vert \D_1)\leq 2 \mathbb{E} \left \{ \overline{\Phi}\left ( \max \left \{\frac{a V_{n_2}^{1/2}}{2 (1 + A\epsilon)^{1/2}} - \frac{(1+ A \epsilon)^{1/2} \log p}{a V_{n_2}^{1/2}} , 0 \right \}\right) \right\}.
\]
Now recall that $\epsilon = \frac{6\delta_0}{A}$ and $\gamma \in ( 0,\frac{\delta_0}{1 + 8\delta_0})$. With this choice of tuning parameters it is easy to see that $r (1-\gamma ) / (1 + A\epsilon) \geq \frac{1 + 7\delta_0}{ 1 + 6\delta_0}>1$ and hence as $p \to \infty$ we have
\begin{align*}
W_{n_2} & := \frac{1}{(\log p)^{1/2}}\left(\frac{a V_{n_2}^{1/2}}{2 (1 + A\epsilon)^{1/2}} - \frac{(1+ A \epsilon)^{1/2} \log p}{a V_{n_2}^{1/2}}\right)\\
&
\overset{\rm p}{\longrightarrow} \frac{1}{(2r)^{1/2}} \left\{ r \left(\frac{1 - \gamma}{ 1 + A\epsilon}\right) ^{1/2} - \left( \frac{1 + A \epsilon}{ 1 - \gamma}\right)^{1/2} \right\}> 0.
\end{align*}
The above display uses the fact that $n_2/n \to 1 - \gamma $ and $V_{n_2}/n_2 \overset{\rm p}{\to} 1$ as $p \to \infty$.
Next let is define the following quantity $q$: $$ q: = q(A, \epsilon, \delta_0, \gamma) = \frac{1}{\{2 (1+ 8 \delta_0)\}^{1/2}} \left\{ (1+ 8 \delta_0) \left(\frac{1 - \gamma}{ 1 + A\epsilon}\right)^{1/2} - \left(\frac{1 + A \epsilon}{ 1 - \gamma} \right)^{1/2} \right\}.$$
Due to choice of $\epsilon$ and $\gamma$ it is easy to show $q>0$.
Now define the event $G_{n_2}: = \{ W_{n_2} > q/2 \}$. Before we proceed it is useful to note the following:
\[
W_{n_2} = \frac{1}{(2 r)^{1/2}} \left( \frac{r \{V_{n_2}/(n(1-\gamma))\}^{1/2} (1-\gamma)^{1/2}}{(1 + A\epsilon)^{1/2}} - \frac{(1 + A\epsilon)^{1/2}}{\{V_{n_2}/(n(1-\gamma))\}^{1/2}(1-\gamma)^{1/2}}\right).
\]
Next define the function
\[
H (u) := \frac{1}{( 2+ 16 \delta_0)^{1/2}} \left\{ u (1+ 8 \delta_0) \left(\frac{1 - \gamma}{ 1 + A\epsilon}\right)^{1/2} - \frac{1}{u}\left( \frac{1 + A \epsilon}{ 1 - \gamma}\right)^{1/2} \right\}, \quad u >0.
\]
As $r = 1 + 8\delta_0$ we have $W_{n_2} = H( \{V_{n_2}/(n(1-\gamma))\}^{1/2} )$. It is also easy to see that $H(\cdot)$ is strictly increasing function on $(0,\infty)$ and $H(1) = q$. Hence $\lambda_{\delta_0}: = H^{-1}(q/2) \in (0,1)$. Now $G_{n_2}^c = \{W_{n_2} \leq q/2\}\subseteq \{ H(\{V_{n_2}/(n(1-\gamma))\}^{1/2}) \leq q/2 \}$. Thus a straight forward calculation shows that that,
\[
\pr (G_{n_2}^c ) \leq \pr \left( \frac{V_{n_2}}{n_2} \leq \frac{n (1-\gamma)}{n_2}\lambda_{\delta_0}^2
\right).
\]
Choose $p$ large enough such that $n_2/n > \lambda_{\delta_0} (1-\gamma)$ and hence we have,
\[
\pr (G_{n_2}^c) \leq \pr \left( \frac{V_{n_2}}{n_2} \leq\lambda_{\delta_0}\right) \leq_c {\sf exp}(- K_{\delta_0} n_2),
\]
where $K_{\delta_0}$ is constant depending on $\delta_0$. Note that $\overline{\Phi}(t) \leq e^{-t^2/2}$ for all $t>0$. Using this fact we have the following:
\[
\pr (\eta_j = 1, \his \vert \D_1)\leq \E \left[{\sf exp} \left\{- \left(1+ \frac{W_{n_2}^2}{2} \right) \log p \right\} {\sf ind}_{G_{n_2}}\right] + \pr(G_{n_2}^c) \leq_c p^{-(1+ q^2/8)} + {\sf exp}( - K_{\delta_0} n_2),
\]
for all $j \notin \calS(\beta)$.
Similarly,
\[
\pr (\eta_j \neq 1, \his \vert \D_1) \leq_c p^{- q^2/8} + {\sf exp}( - K_{\delta_0} n_2), \quad \forall j \in \calS(\beta).
\]
Thus marginalizing out $\D_1$ and summing over all $j$ we get,
\[
\sup_{\beta\in {\mathcal{M}}_{a}^s} \pr_\beta (\hat{\eta}\neq \eta) \leq_c p^{-q^2/8} \log p + p{\sf exp}(- K_{\delta_0} n_2) + {\sf exp}(-c_0 n_1) + {\sf exp}(-c_4 n_1) + c_2 p^{-c_3}.
\]
Now taking $p\to \infty$ we have the result.
\subsection{Proof of Corollary \ref{cor: model_consistecy_ETS_tilde}}
\label{sec: Proof of claim ets_tilde}
Similar to previous proofs, we reparametrize $\delta$ by $8 \delta_0$ and set $\epsilon = 6 \delta_0/A, \varsigma = (1 + A\epsilon)^{1/2}$. Now note that, it is enough to prove the following:
\[
\lim_{p \to \infty}\inf_{\beta \in {\mathcal{M}}_s^a}\pr_\beta \left( \max_{j \notin \calS(\beta)} \abs{\Delta_j} < \min_{j \in \calS(\beta)} \abs{\Delta_j} \right)\to 1
\]
as $p \to \infty$. To this end first define the following quantity:
\[
t_p := \frac{\left(\frac{2r n_2 \log p}{n} \right)^{1/2}}{2} + \frac{\varsigma^2\log p}{\left(\frac{2r n_2 \log p}{n} \right)^{1/2}}.
\]
We will show that $\lim_{p \to \infty} \inf_{\beta \in {\mathcal{M}}_s^a}\pr_\beta \left(\min_{j \in \calS(\beta) } \abs{\Delta_j} > t_p, \max_{j \notin \calS(\beta) } \abs{\Delta_j} \leq t_p\right) \to 1$ as $p \to \infty$. For convenience let us define the events $G_{\min} : = \{\min_{j \in \calS(\beta) } \abs{\Delta_j} > t_p \}$ and $G_{\max} : = \{ \max_{j \notin \calS(\beta) } \abs{\Delta_j} \leq t_p \}$. Let $\his$ be the event as defined in Section \ref{sec: Proof of main ETS}. First we will analyze $\pr_\beta (G_{\min}^c)$. Note that $\pr_\beta (G_{\min}^c ) \leq \pr_\beta (G_{\min}^c \cap \his) + \pr_\beta(\his^c)$. Now the second term goes to $0$ uniformly over $\beta \in {\mathcal{M}}_{s}^a$. Also using Equation \eqref{eq: alpha_distribution} under the event $\his$ we get
\begin{align*}
& \sup_{\beta \in {\mathcal{M}}_{s}^a}\pr_\beta (G_{\min}^c \cap \his)\\
&\leq \sup_{\beta \in {\mathcal{M}}_{s}^a} \pr_\beta \left(
\min_{j \in \calS(\beta)}\abs{\beta_j \Vert X_j^{(2)}\Vert_2 + (1 + A\epsilon)^{1/2} g_j} \leq t_p
\right)\\
& \leq \sup_{\beta \in {\mathcal{M}}_{s}^a}\pr_\beta \left(
\max_{j \in \calS(\beta) } \frac{\abs{g_j}}{(\log p)^{1/2}} \geq \frac{1}{(1 + A\epsilon)^{1/2} (\log p)^{1/2}} \left\{ a \min_{j \in \calS(\beta)} \Vert X_j^{(2)}\Vert_2 - t_p\right\}
\right)\\
& \leq \sup_{\beta \in {\mathcal{M}}_{s}^a} \pr_\beta \left(
\max_{j \in \calS(\beta) } \frac{\abs{g_j}}{(\log p)^{1/2}} \geq \frac{1}{(1 + A\epsilon)^{1/2} (\log p)^{1/2}} \left\{ a \min_{j \in [p]} \Vert X_j^{(2)}\Vert_2 - t_p\right\}
\right)
\end{align*}
where $\{g_j\}_{j \in \calS(\beta)}$ are non i.i.d. standard Gaussian. Note that $\abs{\calS(\beta)} = O(\log p)$. Hence $\max_{j \in \calS(\beta)} \abs{g_j} = O_{\pr}(\log \log p)$, which tells that $$\max_{j \in \calS(\beta) } \frac{\abs{g_j}}{(\log p)^{1/2}}\overset{\rm p}{\to} 0.$$
Also using lemma 3 from \cite{fletcher2009necessary} we have
\[
\frac{1}{(1 + A\epsilon)^{1/2} (\log p)^{1/2}}\left(a \min_{j \in [p]} \Vert X_j^{(2)}\Vert_2 - t_p \right) \overset{\rm p}{\to} \frac{1}{(2r)^{1/2}} \left\{ r \left(\frac{1 - \gamma}{ 1 + A\epsilon}\right) ^{1/2} - \left( \frac{1 + A \epsilon}{ 1 - \gamma}\right)^{1/2} \right\}.
\]
The right hand side of the above display is at least $q(A, \epsilon, \delta_0, \gamma)$ (defined in Section \ref{sec: Proof of main ETS}) which is strictly positive. Again for compactness we use $q$ instead of $q(A, \epsilon, \delta_0, \gamma)$. The above display motivates us to define the following event:
\[
\calE_p = \left\{ \frac{1}{(1 + A\epsilon)^{1/2} (\log p)^{1/2}} \left( a \min_{j \in [p]} \Vert X_j^{(2)}\Vert_2 - t_p\right) \geq q/2\right\},
\]
and it follows that $\pr(\calE_p^c)\to 0 $ as $p \to \infty$. This leads to the following inequality:
\begin{align*}
\sup_{\beta \in {\mathcal{M}}_{s}^a}\pr_\beta (G_{\min}^c \cap \his) \leq & \sup_{\beta \in {\mathcal{M}}_{s}^a} \pr_\beta \left(
\max_{j \in \calS(\beta) } \frac{\abs{g_j}}{(\log p)^{1/2}} \geq q/2\right) + \pr(\calE_p^c)\\
& \leq_c p^{-q^2/8} \log p + \pr(\calE_p^c) \to 0.
\end{align*}
Thus we have $\sup_{\beta \in {\mathcal{M}}_a^a}\pr_\beta (G_{\min}^c)\to 0$. Similarly it can be shown that $\sup_{\beta \in {\mathcal{M}}_s^a} \pr_\beta (G_{\max}^c)\to 0$ as $p \to \infty$. These two claims together completes the proof.
|
\section{Introduction}\label{sec:introduction
\IEEEPARstart{Y}{ukiko} is an 83-year-old Japanese woman living with her son Matsuo
in a traditional Japanese house.
Yukiko sleeps on a futon on the tatami floor, which, she says, is very good for her back. Recently, every time Yukiko has woken up at night to go to the bathroom, she has felt a little dizzy and confused, and sometimes she has had trouble finding the light switch.
Last week she fell in the dark, and nobody noticed she was lying on the floor until morning. Matsuo is worried about her safety
and decided to set up a Smart Environment composed of a small table-top robot assistant named Tetsuwan Atomu\footnote{The android boy popular in Japan, also known as ``Astro Boy".},
voice-activated lights, and cameras that can detect emergencies. Yukiko gladly agreed to have a camera installed in her room: ``I don’t want to be a burden for my son," she says. It’s Sunday evening: Tetsuwan greets her with a bow and then chats with her about cherry blossoms in Spring by showing pictures of beautiful trees in Kyoto on its
display. ``You know so many things," Yukiko says with a smile as she lies down on her futon on the tatami floor, preparing to sleep, ``Goodnight!" Tetsuwan receives an alert from one of the cameras that has recognized a person lying on the floor as an emergency, which ordinarily requires alerting Yukiko’s family. However, according to the robot's cultural database, sleeping on a tatami floor in a traditional Japanese house is common,
so this may not be an emergency. ``Switch off the light, please," says Yukiko, yawning and confirming Tetsuwan’s assessment. Tetsuwan relaxes: it would smile if it could.
The scenario above introduces a key concept:
how to provide personalized interaction by taking the cultural context into account. Indeed, the way Tetsuwan Atomu greets Yukiko, the topics it chooses for conversation, the pictures it shows on its screen, and how it interprets the situation when Yukiko lies on the floor show that the assistant is culturally competent.
The concept of ``culture" is complex \cite{Geertz73, Hofstede80, Leininger1988, Henare2007, Carrithers2010},
and there is no consensus among researchers in defining it.
A simple yet effective definition holds that culture is a shared representation of the world of a group of people.
Then, by ``culturally competent," we mean an intelligent system that can adapt its perceptions, plans, actions, and interaction style depending on the worldview of the person it is interacting with, including their beliefs, values, language, norms, and visibly expressed forms such as customs, art, clothing, food \cite{Papadopoulos2006, Bruno2017, Bruno2019}.
\begin{figure}[t]
\centering
\includegraphics[width=0.81\columnwidth]{Figure01}
\caption{Instructions for dipping the mop, wringing it, and mopping with different ``cultural brushstrokes."}
\label{fig:cleaning}
\end{figure}
Cultural factors in Robotics have been investigated in the last decade \cite{Bartneck2007, Rehm2013}. Most previous approaches focused on what can make the system more or less acceptable to people of different cultures, either concerning its appearance or its behavior \cite{Shibata2009, Rau2010, Lee2016}, including verbal and non-verbal interaction \cite{Andrist2015, Rau2009}
or social distance \cite{Eresha2013, Joosse2014}.
Recently, we defined\footnote{https://cordis.europa.eu/article/id/124441-the-worlds-first-culturally-sensitive-robots-for-elderly-care}
a conceptual framework to make Socially Assistive Robots (SAR) for elderly care culturally competent \cite{Bruno2017, Bruno2019} starting from research in Transcultural Nursing \cite{Leininger1988} and culturally competent Health Care \cite{Betancourt2003}.
However, cultural competence is crucial
in all domains that could benefit from Artificial Intelligence (AI),
including Education, Travel, and Business
\cite{Abrams2009, Buehler2009,
Johnson2006}.
In this general scenario, this article focuses on the visual representation of objects
in different cultures \cite{Knappett2005} and the impact on people. The anthropological study of material culture teaches us that everyday objects have different designs in diverse world areas, even when they share many similarities in terms of their functionalities. This is not a surprise: objects' design is related to the social tasks they intend to accomplish and the material enabling them to do it \cite{Robb2015}.
Consequently, within a given geographical area and culture, some visual features tend to coherently repeat with a higher frequency, even if the presence of aesthetic universal has been found \cite{Hekkert2008, Hardonk1999}.
With a metaphor, two objects may have the same functionalities or affordances \cite{Gibson1986}
and yet be given ``brushstrokes of cultural features" that make them uniquely recognizable.
Suppose now a robot interacting with a person, i.e., showing instructions on how to perform a given task such as shaving oneself or cleaning the floor: this capability may play a key role in SARs, e.g., when interacting with older people with mild forms of dementia.
In the case of floor cleaning, instructions are likely to go like this: (1) Take a bucket; (2) Take a mop; (3) Take a cleaning detergent; (4) Sweep or vacuum First; (5) Fill the bucket; (6) Dip the mop; (7) Wring the mop; (8) Begin mopping (steps 6, 7, and 8 in Figure \ref{fig:cleaning}).
How important is it that, during interaction with people,
robots and other intelligent systems show images with which the person is familiar? The advertising industry knows all too well that a product may need culturally appropriate advertisements to hit the market in diverse world areas.
An ad for a detergent in
Japan is unlikely to
show an Italian family in an Italian house unless a specific message in this sense is required, and the opposite is true in Italy. This concept is also the basis of cross-cultural design \cite{Chai2015, Hsu2011}, a process that requires designers to understand other cultures, select cultural elements, rethink and review them, and then integrate these into a new product to satisfy target users and incorporate them into product design.
Given these premises, the article's contribution is twofold.
\begin{itemize}
\item The article addresses the problem of ``culturizing" images and explores Generative Adversarial Networks (GAN) \cite{Goodfellow2014, Isola2017, Jun-Yan2017, Taeksoo2017, Mejjati2018, Sangwoo2018, Unit2017} for this purpose.
With the term ``culturization," we intuitively mean altering the ``brushstroke of cultural features" that make objects perceived as belonging to a given culture while preserving the perceived functionalities of objects. Specifically, we translate selected objects in the image from a source to a target cultural domain.
\item The article explores the research question: `to what extent do people like images that were ``culturized" to make them coherent with their culture?' To this end, we prepared an online questionnaire to be submitted to Italian participants. Then, we evaluated through statistical analysis how people perceive the cultural context of objects, their preferences for culturized and non-culturized objects, and how realistic they perceive environments patched with culturized objects.
\end{itemize}
Please notice that this work's novelty does not consist in proposing a new network for image-to-image translation but in introducing the culturization concept, choosing a state-of-the-art GAN and a pipeline for image culturization, and defining an experimental protocol for the subjective evaluation of culturized images with recruited participants.
We do not perform a comparison with previous methods because these concepts have not received much attention so far. To the best of our knowledge, the only work that shares similar ideas is \cite{Zhou2022}, which uses deep learning techniques for cross-cultural design. The tool proposed requires a designer to upload a design sketch and a set of cultural images used as a reference for the style; then, it generates a culture-specific image with the same content as the uploaded sketch and the cultural style of the selected style image. However, the focus is different from our work. First, the article explores the designers' perception of using the tool for cross-cultural design, but not the subjective perception of end-users that will ultimately use the designed object. Second, the style transfer module used in \cite{Zhou2022} can only alter the texture of images \cite{Gatys15}. In contrast, we search for solutions capable of changing all visual design elements, including color, material, pattern, and form.
Section \ref{sec:SOTA} describes the state-of-the-art related to image-to-image translation using GANs. Section \ref{Materials and Methods} presents the process to culturize objects and environments, and to evaluate peeople's preferences. Sections \ref{results} and \ref{Discussion} present and discuss results. Conclusions are in Section \ref{sec:conclusion}.
\section{State of the Art}
\label{sec:SOTA}
Image culturization can be described as an image-to-image translation problem.
Suppose having:
\begin{itemize}
\item a set of images related to each other by a shared characteristic that defines them as belonging to the same domain, referred to as the source domain $S$;
\item a set of images, different from the first one, which defines a target domain $T$.
\end{itemize}
The image-to-image translation problem is the problem of learning a mapping $G:S \to T$ such that the distribution of images from $G(S)$ is indistinguishable from the distribution $T$. Otherwise said, the objective is to modify images belonging to $S$ to make them ``similar" to those belonging to $T$.
Image-to-image translation \cite{Kaji2019} may benefit from GAN-based approaches.
GANs are not the only feasible approach to this problem \cite{pang2021imagetoimage}: however, we do not aim to find the most performing solution for image culturization. Instead, our objective is to find a feasible solution that meets all constraints,
to be later tested with human participants to confirm the relevance of this new concept. According to this rationale, and considering that GAN-based methods are the vast majority of solutions, we will limit our analysis to GANs.
\subsection{Generative Adversarial Networks}
\label{sec:GANstateoftheart}
GANs have been successful in several image-to-image translation domains, ranging from the creation of purely synthetic images (e.g., faces of people that do not exist in the real world \cite{Goodfellow2014, radford2015unsupervised}) to the modification of selected features in a pre-existing image (e.g., the age or facial attributes of a person \cite{Zhao2020, Huang2021, Tan2021}). Other GANs applications include adding semantic labels to photos
\cite{Cordts2016}, producing photorealistic images from sketched drawings \cite{Eitz2012, Yu2021},
translating photographs into paintings \cite{Jun-Yan2017}, or filling missing regions of images
\cite{Haofeng2019}.
GANs have two components \cite{Goodfellow2014}: a generator and a discriminator. Both the generator and the discriminator must be trained using a large set of data $x$ describing a probability distribution $p_{data}$: the generator is trained to generate samples with probability distribution $p_g = p_{data}$; the discriminator is trained to distinguish the real data from the generated ones. The idea is to create a competition between two networks.
In \cite{Goodfellow2014}, the discriminator and the generator are both multilayer perceptrons
with parameters $\theta_g$ (the weights of the generator) and $\theta_d$ (the weights of the discriminator).
The generator $x=G(z; \theta_g)$ takes as input a noise variable $z$ with probability distribution $p_z(z)$ and maps it to data space with a distribution $p_g(x )$.
The discriminator is instead a binary classifier $D(x; \theta_d)$ that takes as input a sample $x$ (that may either be from one of the samples of the data or the output of the generator) and outputs the probability that $x$ came from the data distribution $p_{data}(x)$ rather than from the generator. After training, both $G$ and $D$ will reach a point at which they cannot improve because $p_{data}(x)=p_{g}(x)$, and $D(x; \theta_d)=0.5$.
Optimal parameters $(\theta_g, \theta_d)$ are obtained by playing the following two-player minimax game with value function $V$, maximizing the cost function over $\theta_d$ and minimizing it over $\theta_g$
\begin{multline}
\min_{\theta_g} \max_{\theta_d} V(\theta_d, \theta_g) = \frac{1}{2}\mathbb{E}_{x\sim p_{data}(x)}[logD(x)]+\\
\frac{1}{2}\mathbb{E}_{z\sim p(z)}[1 - log(D(G(z)))]
\label{eq:ganlossoriginal}
\end{multline}
Starting from this idea, Deep Convolutional Generative Adversarial Networks (DCGANs) were developed to exploit the success of Convolutional Neural Networks (CNNs) \cite{radford2015unsupervised}: the generator still takes in input a latent noise variable $z$ but now generates an image through convolutional decoding operations.
Unfortunately, similar to the original GANs, DCGANs do not consent to specify additional constraints on the generated data samples: a relationship between the noise $z$ fed to the generator and the generated images exists (e.g., the network produces different faces) but we cannot control the outcome (e.g., we cannot choose to produce younger or older people's faces).
Conditional GANs (CGAN) \cite{mirza2014conditional} overcome this limitation by providing additional information to the generator and the discriminator during training to synthesize a fake sample with desired characteristics. The generator learns to produce realistic samples (e.g., faces) corresponding to a specific label (e.g., younger or older), whereas the discriminator learns to distinguish fake sample-label pairs from real sample-label pairs.
The value function $V$ is the same as Eq. \ref{eq:ganlossoriginal}, by substituting $G(z)$ with $G(z|y)$ and $D(x)$ with $D(x|y)$, where $y$ is the additional information provided to the network.
\subsection{Image-to-Image translation}
\label{sec:imagetoimagetranslation}
In the following, we will review a subset of GAN-based solutions for image-to-image translation \cite{Isola2017, Jun-Yan2017, Taeksoo2017, Mejjati2018, Sangwoo2018, Unit2017} that have interesting properties for image culturization. The analysis is limited to a small subset of the available solutions \cite{pang2021imagetoimage} since this article's main objective is to evaluate the impact of image culturization on people rather than comparing the performance of different GANs. However, given the general idea, we might consider other solutions in the future.
All approaches typically require two datasets, each dataset containing images with homogeneous content, style, and resolution. However, some methods have the additional requirements that the images need to be paired during training: for every image of the source domain $S$, there should be a corresponding image of the target domain $T$.
An example in this class is Pix2Pix \cite{Isola2017}. The model postulates a transformation that modifies the source image to make it ideally belong to the target domain while preserving some of its characteristics: the goal of Pix2Pix is to learn this transformation and perform it.
During training, Pix2Pix adopts a CGAN approach where the input image conditions the generator; the discriminator takes as input a generated or a real image belonging to the target domain and guesses whether the image is real or not.
The network is trained with the CGAN loss function by adding a weighted term that measures the difference between the generated and real images in the pair.
The main limitation is that Pix2Pix and similar approaches require large paired datasets to learn the mapping between $S$ and $T$: paired datasets are complex and expensive to build, especially if one aims to culturize objects into several different cultures. %
In order to overcome the constraint that datasets need to be paired during training, some approaches propose a new concept, referred to as \textit{consistency of the cycle}: the image $x_S$, after the first translation $G(x_S)$ from $S$ to $T$, is fed to a second generator that performs a backward translation $F(G(x_S))$ to remap the image to the starting domain $S$ and compares it with the original image.
This concept was first introduced by CycleGAN \cite{Jun-Yan2017},
trained by
adding a weighed, ``cycle-consistency" L1 loss measuring the difference between $x_S$ and $F(G(x_S))$.
CycleGAN has been used to re-create the style of impressionist painters starting from a photograph.
This concept fits well with our idea: in analogy with painters representing reality with their painting style, we want to capture how a human self-identifying with a given culture would represent a particular object through ``cultural brushstrokes''.
CycleGAN and similar approaches \cite{Zili2017, Taeksoo2017} allow for building datasets quickly by searching for images on the web (e.g., Greek and Japanese vases).
However, this solution also has negative sides. First, it cannot culturize an individual object in a complex scene and keep the background unaltered. Second, the image-to-image translation works better when changes occur in the texture (from a horse to a zebra, from an orange to an apple) than when they occur in shape: CycleGAN can alter the shape of an object, but this creates artifacts and less realistic images.
It would be helpful if the network could learn which parts of the scene to translate from $S$ to $T$ by isolating a region of interest in the image.
The solution proposed in attention-guided GAN \cite {Mejjati2018} achieves this:
based on the structure of CycleGAN and the concept of the consistency of the cycle,
two \textit{attention networks} are added, one for the generator $G: S \to T$ and the other for the generator $F: T \to S$, which learn to extract ``attention maps." Attention maps are trained in parallel with generators, and provide semantic information by segmenting images into regions that must or must not be translated from $S$ to $T$. This is achieved by assigning each pixel of the map a continuous value in the interval [0;1]:
CycleGAN is equivalent to an attention-guided GAN where the attention maps are equal to $1$ everywhere.
This approach and similar ones \cite{Chen2018, Tang2021}
allow modifying individual elements of a scene without further constraints on the datasets. However,
attention maps imply a higher computational load and training time. Additionally, attention-guided approaches have limitations when image-to-image translation involves changes in the objects' shape (typically required in image culturization, see the vases on the first row of Figure \ref{all}) since they aim to keep the image background unaltered.
Solutions exist to
to deal with datasets that exhibit more evident changes in the shape of objects or include multiple instances of objects. To this end, InstaGAN \cite{Sangwoo2018} requires information on the instances to be modified, obtained through segmentation masks a priori provided: $S$ and $T$ must contain both images and masks for all elements of interest.
A set of generators supplements the original generator:
each generator operates on a segmentation mask and produces a translated mask in the target domain.
The cross-entropy used in CGAN
is modified by adding weighted terms that include the cycle-consistency loss and a background-preserving loss.
InstaGAN produces high-quality images where only selected regions are modified.
However, it adds a new constraint: it requires a segmentation mask associated with each object. Once again,
this is problematic for image culturization based on custom-built datasets composed of images downloaded from the web.
Approaches exist that emphasize the dichotomy between high-level semantics (e.g., depending on the object's functionalities or affordances) and low-level features (e.g., ``brushstrokes of cultural features").
UNsupervised Image-to-image Translation (UNIT) \cite{Unit2017} is a representative of this class based on CoGAN \cite{Ming2016}. CoGAN consists of two GANs: each GAN takes a random vector as input without being conditioned to any input image. Generators are trained to produce pairs of images in the two domains $S$ and $T$; discriminators are trained to distinguish generated images from real ones. Very importantly, the weights of the first few layers of the generators and those of the last few layers of the discriminators (respondible of decoding and encoding high-level semantics) are shared. Thanks to this, CoGAN generates pairs of related images in $S$ and $T$ with the same high-level features but different low-level ones.
However, CoGAN is unsuitable for image-to-image translation because it is fed with random input $z$. UNIT makes the network conditioned by adding variational autoencoders \cite{VAeGAN}, trained to encode images from the two domains $S$ and $T$ into a shared latent code $z$, which is then fed to a CoGAN-like network to produce images in $S$ and $T$. This process
implies a cycle-consistency constraint comparable with CycleGAN's.
\subsection{Choosing a solution for GAN-based culturization}
\label{Solution}
GANs evaluation \cite{Proecons2018} is a challenging task. GANs are trained to reach an equilibrium situation where the generator can ``cheat" the discriminator. Still, the former, taken alone, is not associated with a cost function to be minimized. Under these conditions, it is hard to predict when the generator will produce samples that fit the target probability distribution at their best. Then, researchers have defined qualitative and quantitative tools to evaluate GANs.
Qualitative assessment of the generated images is often performed
using crowdsourcing platforms such as Amazon Mechanical Turk \cite{Isola2017, Jun-Yan2017, Zili2017, Mejjati2018}.
Quantitative evaluation can be performed using such metrics as the Inception score (IS) \cite{Tim2016} and the Fréchet Inception Distance (FID) \cite{FID2017} -- or other metrics \cite{pang2021imagetoimage}.
IS evaluates generated images based on how Inception v3, a widely-used image recognition model that attains greater than 78.1\% accuracy on the ImageNet \cite{Szegedy20162818}, classifies them.
IS also evaluates if there is sufficient diversity among the generated samples, addressing the so-called problem of the ``collapse of the model": the generator learns to generate a specific image of the target domain and keeps on generating the same image (even when varying the input) because the latter is very good in deceiving the discriminator.
FID also uses a pre-trained Inception v3 model to measure the quality of the generated images. Still, it does so differently, i.e., by evaluating the accuracy of the Inception v3 model in classifying GAN-generated images compared with real ones. Consequently, FID captures the similarity of generated images to real ones better than IS.
In principle, we might compare the different solutions for image culturization discussed in the previous section using FID as done in \cite{Zhao2020}, which shows how FID values can significantly vary depending on specific image-to-image translation task. However, we found it more appropriate to compare different GANs on a qualitative basis to assess the extent to which each solution can meet the requirements for image culturization (FID will play a role in the next section in tuning hyperparameters of the selected GAN solution).
Specifically, we observed that Pix2Pix and similar approaches require paired datasets, which is a limitation since image culturization should work on custom-built datasets, e.g., composed of images downloaded from the web. InstaGAN did not reveal a feasible solution for similar reasons since it requires a carefully designed dataset composed of images belonging to different cultural domains and the corresponding masks.
Following these initials considerations, we implemented CycleGAN, AttentionGAN, and UNIT that do not pose constraints on the dataset, and then we qualitatively evaluated the results. Tests confirmed that AttentionGAN has strong limitations when an image-to-image translation requires altering the image shapes, often needed in image culturization (see the vases in the first row of Figure \ref{all}).
Both CycleGAN and UNIT might be good choices: after qualitative evaluation, and since CycleGAN is the basis for many other approaches (including InstaGAN that might be reconsidered in the future), we finally selected it as the best candidate for image culturization. However, we do not see any obstacles in implementing the same procedure described in the next Sections with a different GAN.
\section{Materials and Methods}
\label{Materials and Methods}
This section describes the process adopted to culturize images using CycleGAN and the study design to explore our research question
with recruited participants.
\subsection{Culturization pipeline}
\label{sec:cult-pipeline}
In the following, we refer to $S$ as the set of images defining a start domain (e.g., Greek vases), $T$ as the set of images defining a target domain (e.g., Chinese vases), $S\to T$ as the images generated starting from $S$ (e.g., Greek vases changed to Chinese style), $T \to S$ as the images generated starting from $T$ (e.g., Chinese vases changed to Greek style).
The culturization process includes three phases, Figure \ref{fig:cultu}: the building of datasets (phase 1), training and validation (phase 2), and the production of culturized images for interaction with people (phase 3). We did not follow the usual training-validation-test pipeline since we are only interested in producing images of sufficient quality for evaluation with people, not in assessing how the CycleGAN model performs on a holdout test set. Additionally, following a preliminary investigation, we realized that the default hyperparameters in the CycleGAN implementation\footnote{Available at github.com/junyanz/pytorch-CycleGAN-and-pix2pix} are appropriate to produce images of sufficient quality (which is good because retraining the model takes tens of hours with our hardware/software configuration). Therefore, after collecting datasets describing a given object class (e.g., vases) in the source and target cultural domains $S$ and $T$ (e.g., Greek and Japanese), we split them into a training and validation set. Then, we performed training and validation a few times to tune the number of training epochs and the dimensions of the datasets based on a qualitative and quantitative evaluation of the culturized images.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Figure03b}
\caption{Culturization process.}
\label{fig:cultu}
\end{figure}
In phase 1, two fully automated steps are required by the system to build datasets, managed by a script in Python.
\begin{itemize}
\item \textit{Image collection}. The script requires to enter the name of an object class (e.g., vase, sofa, pillow, lamp, etc.), a cultural domain (e.g., Greek, Indian, Moroccan, Chinese, etc.),
and the number of images to download. The images are automatically downloaded from Google through the Selenium Web Driver API.
\item \textit{Preprocessing of images and object segmentation}. Downloaded images
are preprocessed to make them suitable for CycleGAN: first, images are resized to the required $256\times 256$ dimensions; then, objects are segmented using a network capable of identifying the desired objects within the image.
Object detection and segmentation rely on
the \textit{Faster-RCNN-inception-V2} pre-trained network. To this end, the candidate objects to extract need to be chosen,
and networks are trained by labeling around $3,000$ images via the labelImg tool.
If multiple instances of an object are present, they are all detected and segmented.
\end{itemize}
In phase 2, three steps are required to train and validate the model to map images from $S$ to $T$ and vice versa.
\begin{itemize}
\item \textit{Preparation of training and validation sets}. We consider two datasets of images $S$ and $T$
corresponding to the same object class in the two cultural domains
and split them into a training set (90\% of the dataset, built with an equal number of images from $S$ and $T$) and a validation set (10\% of the dataset).
\item \textit{Training}. The network is trained to generate images $S \to T$ starting from $S$ as well as images $T \to S$ starting from $T$.
We train the network with a GPU TESLA P100 using the GPU-accelerated NVIDIA CUDA Deep Neural Network library (about 30 hours for 200 epochs). \item \textit{Validation}. The validation set is used to generate images $S \to T$ starting from $S$ as well as images $T \to S$ starting from $T$. Only the trained generator is used, whereas the discriminator is no longer needed. The resulting images $S \to T$ and $T \to S$ undergo a qualitative analysis to visually judge their quality and the FID score is computed
to confirm that the results are acceptable.
If needed, the network is trained again by changing the number of training epochs and the dimensions of the dataset.
\end{itemize}
During training and validation,
we did not find a significant difference in the perceived quality and FID values when changing the number of training epochs from 150 to 400 (training time increasing from 20h to 50h and more). For example, when processing images of sofas, FID values ranged from 129.18 to 122.33 (from Indian to classic European) by increasing the number of epochs; when processing images of pillows, FID values ranged from 78.60 to 83.02 (from classic European to Indian).
It must also be remembered that the range of FID values depends on the application for which GANs are used: FID is typically used to compare different approaches rather than providing an absolute estimate of the quality of the process. The work in \cite{Zhao2020}, which focuses on human faces, reports FID values ranging from 23.72 to 48.71 for ``glass removal", from 16.63 to 36.17 for ``male to female", and from 93.58 to 102.92 for ``selfie to anime". Since CycleGAN has limitations in modifying the shape of objects, FID values tend to be quite high in our case: however, they do not change significantly with the number of training epochs. This result is also confirmed by the qualitative evaluation of culturized images, which motivated us to keep the default value of 200 training epochs in the chosen CycleGAN implementation. The same rationale motivated us to use training datasets of 1,000 images, which allows for a good compromise between FID values, perceived quality, and variability.
\begin{figure}
\centering
\includegraphics[width=0.45\columnwidth]{Figure04a}
\caption{Pepper explains how to clean the floor.}
\label{fig:pepper}
\end{figure}
In phase 3, to culturize individual objects, it is now sufficient to feed CycleGAN with an image belonging to $S$ or $T$.
\begin{itemize}
\item \textit{Choice of an object image}. The image of an object belonging to a cultural domain $S$ or $T$ is provided
\item \textit{Image culturization}. The image is translated from $S$ to $T$ or $T$ to $S$ using CycleGAN generator.
\end{itemize}
However, if the objects to be culturized are part of a complex environment (as in Figure \ref{fig:cleaning}), four steps are required.
\begin{itemize}
\item \textit{Choice of an environment image}. The image of an environment belonging to a domain $S$ or $T$ is provided
\item \textit{Object segmentation}. The image of the environment is preprocessed using \textit{Faster-RCNN-inception-V2} to extract all objects that are a candidate for culturization.
\item \textit{Object culturization}. CycleGAN generator is used to map each extracted object from the domain $S$ to $T$ or $T$ to $S$.
\item \textit{Environment patching}. The original image is patched with the culturized objects. Currently, this step is not automated but manually performed with a tool.
\end{itemize}
The images produced are finally used for a more rewarding human-robot interaction: see, for example, Pepper showing cleaning instructions in Figure \ref{fig:pepper}.
However, culturally competent real-time interaction was not the focus of the experiments performed in this work. Due to the novelty of the culturization concept, we decided explore our research question through an online questionnaire to measure the impact of image culturization on people, which allows for the involvement of a greater number of recruited participants.
\subsection{Study design}
\label{Experiments}
To start exploring our research question (`to what extent do people appreciate images that were ``culturized" to make them coherent with their culture?'), we decided to do experiments with Italian participants only to simplify recruitment. We will consider other nationalities in future works. The general idea was to show participants images of different objects and environments belonging to European and non-European cultural domains and ask them some questions.
To this end, we considered four object classes (vases, sofas, pillows, and lamps) and, for each class, four sets of images produced in different ways: one set was downloaded from the Internet as belonging to the European culture and tradition ($E$); one was downloaded from the Internet as belonging to a different culture and tradition (Chinese, Indian or Moroccan, $O$ for other); one was a non-European object ``culturized" using GANs ($O \to E$); one was a European object ``culturized" ($E \to O$) using GANs.
Specifically, we searched for evidence supporting the four hypotheses below through
data acquisition and statistical analysis.
\begin{itemize}
\item \textit{Hypothesis 1}: objects are correctly recognized by Italian participants as belonging to the $E$ culture. This hypothesis holds both when the object originally belonged to the $E$ culture and when it was culturized from $O$ to $E$.
\item \textit{Hypothesis 2}: European objects are generally preferred by Italian participants. This hypothesis holds both when the object originally belonged to the $E$ culture and when it was culturized from $O$ to $E$.
\item \textit{Hypothesis 3}:
Italian participants who prefer $E$ objects tend also to prefer objects culturized from $O$ to $E$.
\item \textit{Hypothesis 4}: Participants perceive environments as more realistic when objects in the image were culturized with GANs than when we patched them with unmodified objects downloaded from the Internet.
\end{itemize}
It is worth spending a word concerning Hypothesis 2. Each person has their preferences: hypothesizing that Italian people will prefer European objects incurs the risk of stereotyping and, generally speaking, is not true. Different persons may be more or less attracted by objects belonging to their own or other cultures, which also depends on individual objects. For example, the same person may adore colorful Moroccan lamps and, at the same time, be particularly attracted by the design of a lamp produced in Sweden or France. We are aware of this and do not intend to make stereotyped claims. The objective of Hypothesis 2, together with Hypothesis 3, is to explore if there are people with whom using images that match their cultural background can be a winning strategy.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{Figure05}
\caption{Pairwise comparison of objects. $E$: downloaded from the Internet as belonging to the European culture; $O$: downloaded from the Internet as belonging to a different culture; $O \to E$: non-European object culturized using GANs; $E \to O$: European object culturized using GANs. 1st row: $O$ vs. $O \to E$; $E$ vs. $E \to O$. 2nd row: $E$ vs. $E \to O$; $O$ vs. $O \to E$; 3rd row: $E \to O$ vs. $E$; $O \to E$ vs. $O$. 4th row: $O$ vs. $O \to E$; $E \to O$ vs. $E$.}
\label{all}
\end{figure}
After preparing images using the procedures in section \ref{sec:cult-pipeline}, we set up a Google Form questionnaire. The questionnaire was anonymous and included the four sections in Table \ref{Tableq}, showing pictures and asking related questions\footnote{Full questionnaire available at
{https://bit.ly/3zjVHWL}
}.
\begin{table}
\centering
\includegraphics[width=1\linewidth]{Figure06}
\caption{Questionnaire's structure: after personal questions, 16 objects are shown in the section \textit{Recognizing culture and expressing preferences}, questions Q1 and Q2; 16 pairwise comparisons of objects are shown in the section \textit{Comparing culturized and non-culturized objects}, questions Q1 and Q2; 8 pairwise comparisons of environments are shown in the section \textit{Comparing the realism of environments}, question Q1.}
\label{Tableq}
\end{table}
\begin{enumerate}
\item \textit{Personal questions}: the respondents had to declare their age (\textit{less than 20}, \textit{20--29}, \textit{30--39}, \textit{40--49}, \textit{50--59}, \textit{60--69}, \textit{70 and over}) and if how often travel out of Italy (\textit{never or very rarely}; \textit{yes, but mostly in Europe}; \textit{yes, both in Europe and outside Europe});
\item \textit{Recognizing culture and expressing preferences}: we showed respondents four vases, four sofas, four pillows, and four lamps. For each image, the respondents had to reply to the following two questions by assigning a score on a 5-points Likert scale. Q1: `How close is this object to European culture and tradition?' (from \textit{1:very far} to \textit{5:very close}); Q2: `Do you like this object?' (from \textit{1:not at all} to \textit{5: a lot}). For each object class (vases, sofas, pillows, and lamps), the four objects shown to the respondent belonged to different groups ($E$, $O \to E$, $E \to O$, and $O$).
\item \textit{Comparing culturized and non-culturized objects}: we showed respondents sixteen pairwise comparisons, four of which concerned vases, four sofas, four pillows, and four lamps. For each comparison, two objects were shown side by side: the original one (which can either be $E$ or $O$) and its corresponding GAN counterpart ($E \to O$ or $O \to E$, depending on the original one). The respondents had to reply to the following two questions. Q1: `Which object best represents European culture and tradition?' (\textit{the one on the left} or \textit{the one on the right}); Q2: `Which object do you like the most?' (\textit{the one on the left} or \textit{the one on the right}). We assigned a score of 1 to the winning object and 0 to the loser. The order in which original and modified objects appeared in the pair varied for each comparison. Figure \ref{all} shows some of the pairs presented to respondents.
\item \textit{Comparing the realism of environments}: we showed respondents eight pairwise comparisons. For each comparison, two environments were presented to the respondent: an environment that was modified by segmenting objects,
culturizing them with GANs, and then pasting them back onto the original image ($Gcult$);
an environment modified by substituting the original objects with objects of different cultures downloaded from the Internet, pasted onto the original image ($PCult$). The respondent had to reply to the following question. Q1: `Which image looks more realistic?' (\textit{the one on the top} or \textit{the one on the bottom}), Figure \ref{rooms}. We assigned a score of 1 to the winning environment and 0 to the loser
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Figure07b}
\caption{Top: $Pcult$ image patched with unmodified objects; Bottom: $Gcult$ image patched with GAN-modified objects.
}
\label{rooms}
\end{figure}
We use question Q1 in the questionnaire's sections 1 and 2 to collect data supporting Hypothesis 1, Q2 in sections 1 and 2 to collect data supporting Hypotheses 2 and 3, and Q1 in section 3 to collect data supporting Hypotheses 4.
Notice that the questionnaire showed respondents objects whose cultural belonging was emphasized. Greek vases were easily recognizable as belonging to the European culture; non-European vases were mostly Chinese and Japanese. Figure \ref{all} shows that, when culturizing vases through GANs, the original Chinese or Greek drawings are recognizable in their GAN-modified counterparts: a Chinese vase with greek warriors and a Greek vase with an oriental tavern can be spotted in the Figure.
Similarly, we chose European sofas designed in classic style. One could object that this kind of sofa is not common in ordinary European houses: however, we wanted to avoid international, modern-style sofas that may not be immediately recognizable as European. Pillows were probably less recognizable as
European / non-European
since colorful pillows are customary also in ordinary European houses. Finally, the cultural belonging of lamps is evident from shape and texture: see the moon-shaped Moroccan lamp versus its European counterpart, and the European ``night table" lamp rethought in a Moroccan style.
The Google Form was online from 18/09/2021 to 7/10/2021. We recruited participants through public social networks (Facebook and Twitter) and student networks of the University of Genova. The questionnaire included the following introduction in Italian: `The questionnaire will show you a sequence of images and a few questions. The images show different objects and environments associated with different cultures. We will ask you to rate how close the images are to your culture and express your tastes and preferences. Please look at the images and answer truthfully based on your feelings: some questions require you to respond with a score between 1 and 5, and other questions require you to choose between two images. The questionnaire should take between 10 and 20 minutes.'
\section{Results}
\label{results}
This section describes collected data and related analyses.
\subsection{Personal questions}
\label{Anagraphic data}
Overall N=392 participants filled the questionnaire, out of which: 7.9\% were less than 20 years old; 21.9\% were in the range 20-29; 13.8\% in the range 30-39; 20.4\% in the range 40-49; 20.9\% in the range 50-59; 11.5\% in the range 60-60; 3.6\% are 70 years or older. Concerning travels, 26\% of the participants declared they never travel out of Italy or very rarely; 48.5\% frequently travel out of Italy, but mostly in Europe; 25.5\% travel both in Europe and out of Europe.
\subsection{Recognizing culture and expressing preferences}
\label{Recognizing culture and expressing preferences}
Table \ref{Table1} shows the results of the second section of the questionnaire. For each question Q1 and Q2, the table reports the average score and its standard deviation obtained by objects (vase, sofa, pillow, or lamp, both individually taken and all together) belonging to different groups ($E$, $O \to E$, $E \to O$, $O$), computed over N respondents.
\begin{table}
\centering
\includegraphics[width=1\linewidth]{Figure08.eps}
\caption{Questionnaire's second section. For each question Q1 and Q2, the table reports the average score and its standard deviation obtained by objects (vase, sofa, pillow, or lamp, both individually taken and all together) belonging to different groups ($E$, $O \to E$, $E \to O$, $O$), N=391 respondents.}
\label{Table1}
\end{table}
\begin{table}
\centering
\includegraphics[width=1\linewidth]{Figure09.eps}
\caption{Questionnaire's second section, age$< 30$, N=117. Data are organized as in Table \ref{Table1}.
\label{Table2}
\end{table}
To search for evidence supporting Hypothesis 1, we considered Q1: `How close is this object to European culture and tradition?' To support Hypothesis 2, we considered Q2: `Do you like this object?' In both cases, we performed statistical tests to reject a set of null hypotheses in the form $H_0: \mu_1 = \mu_2$ in favor of the alternate hypothesis $H_1: \mu_1 > \mu_2$. Depending on each test, $\mu_1$ must be interpreted as the mean score $\mu_E$ or $\mu_{O \to E}$ assigned to $E$ or $O \to E$ objects, which we hypothesized to be higher than the mean score of other objects. To test
hypotheses, we initially performed an ANOVA test with $\alpha=0.05$ by analyzing all groups together ($E$, $O \to E$, $E \to O$, $O$) and a Tukey HSD test to check if pairwise groups are statistically different.
Next, we focused specifically on the two groups $E$ and $O \to E$, and performed a series of one-tail, two-sample Welch’s T-test with $\alpha=0.05$ to compare them with each other and other groups.
Due to the sample size N=392, we did not need to check distribution normality.
Table \ref{Table1} shows that, concerning Q1 (`How close is this object to European culture and tradition?'),
\begin{itemize}
\item for any object class, $E$ objects achieved the highest average score;
\item for any object class, $O \to E$ objects achieved the second-highest average score.
\end{itemize}
The ANOVA returned that we can reject the null hypothesis for all object classes. The Tukey HSD test returned that all groups' averages, when pairwise taken, are significantly different with the only exception of the couple ($E \to O$, $O$) in the vases, sofas, pillows classes.
To confirm this result, we then performed a one-tail, two-sample Welch’s T-test to compare $\mu_E$ (corresponding to the highest average score) with $\mu_{O \to E}$ (the second-highest), and then $\mu_{O \to E}$ with $\mu_O$ and $\mu_{E \to O}$. We can reject the null hypothesis in all tests in favor of the alternate hypothesis with $p<0.05$ for any object class.
In summarizing, with a focus on $\mu_{E}$ and $\mu_{O \to E}$ in Q1,
\begin{itemize}
\item $\mu_{E}>\mu_{O \to E} > \mu_{O}$ for vases, pillows, and lamps,
\item $\mu_{E}>\mu_{O \to E} > \mu_{E \to O}$ for sofas,
\item $\mu_{E}>\mu_{O \to E} > \mu_{O}$ when considering all classes together,
\end{itemize}
which supports Hypothesis 1.
Concerning Q2 (`Do you like this object?'),
\begin{itemize}
\item for any object class, $E$ objects achieved the highest average score;
\item for vases, the second-highest average score was achieved by $O \to E$ objects; for all other classes, the second-highest average score was achieved by $O$ objects.
\end{itemize}
The ANOVA returned that we can reject the null hypothesis for all object classes. The Tukey HSD test returned that all groups' averages, when pairwise taken, are significantly different with the only exception of the couple ($O \to E$, $E \to O$) in the sofa class, ($E \to O$, $O$) in the pillow class, ($E$, $O$) and ($O \to E$, $E \to O$) in the lamp class.
The one-tail, two-sample Welch's T-test returned that, for vases, the difference between $\mu_E$ (highest average score) and $\mu_{O \to E}$ (second-highest) is statistically significant with $p<0.05$, and the same holds for the difference between $\mu_{O \to E}$ and $\mu_O$ (third-highest). For sofas and pillows, the difference between $\mu_E$ (highest average score) and $\mu_O$ (second-highest) is statistically significant with $p<0.05$, and the same holds for the difference between $\mu_O$ and $\mu_{O \to E}$. For lamps, the difference between $\mu_E$ (highest average score) and $\mu_O$ (second-highest) is not statistically significant with $p<0.05$, but the difference between $\mu_O$ and $\mu_{E \to O}$ is.
In summarizing, with a focus on $\mu_{E}$ and $\mu_{O \to E}$ and Q2,
\begin{itemize}
\item $\mu_{E}>\mu_{O \to E} > \mu_{O}$ for vases,
\item $\mu_{E}>\mu_{O} > \mu_{O \to E}$ for sofas and pillows,
\item $\mu_{E}=\mu_{O} > \mu_{O \to E}$ for lamps,
\item $\mu_{E}>\mu_{O} = \mu_{O \to E}$ when considering all classes together,
\end{itemize}
which supports Hypothesis 2 only partially because, in some cases, $O$ objects are preferred to $O \to E$ ones.
We performed a stratified analysis controlling for age and travel habits. Despite minor differences, the results were similar. However, something interesting emerged when considering only respondents with age $< 30$ (N=117), Table \ref{Table2}.
Concerning Q1, results were mostly confirmed: statistical analyses (ANOVA, Tukey HSD, and Welch’s T-test, $p<0.05$) showed that, for any object individually taken and for all objects taken together, it always holds $\mu_{E}>\mu_{O \to E} > \mu_{O}$ with $p<0.05$.
However, concerning Q2, there are some differences compared to considering all age ranges: statistical analysis on vases returned, as previously, $\mu_{E}>\mu_{O \to E} > \mu_{O}$ with $p<0.05$; statistical analysis on sofas
returned, in this case, $\mu_{E}>\mu_{O \to E} > \mu_{O}$ with $p<0.05$; statistical analysis on all objects taken together returned $\mu_{E}>\mu_{O \to E} = \mu_{O}$ as in the previous case, but now the average score of $O \to E$ is slightly higher than
$O$, even if the difference is not statistically significant with $p<0.05$. Results support Hypothesis 2 to a higher degree, but still partially.
To search for evidence supporting Hypothesis 3, we considered Q2: `Do you like this object?' for the whole sample with N=391. Specifically, we considered the N participants and the four groups ($E$, $O \to E$, $E \to O$, and $O$). Next, for each participant and group, we computed the average score for all objects belonging to that group: this process yielded four N-sized vectors, one per group.
Next, we computed Pearson's $\rho$ to correlate the N-sized vector of $E$ scores with the vectors of $O \to E$, $E \to O$, and $O$ scores: please notice that $\rho$ ranges from $-1$ to $1$ and a higher correlation $\rho_{E, O \to E}$ (respectively, $\rho_{E, E \to O}$, $\rho_{E, O}$) means that respondents giving high scores to $E$ objects tend to give high scores to $O \to E$ (respectively, $E \to O$, and $O$) objects as well. The analysis returned $\rho_{E, O \to E}=0.48$, $\rho_{E, E \to O}=0.16$, $\rho_{E, O}=0.15$ supporting Hypothesis 3: respondents giving high scores to $E$ objects tend to privilege $O \to E$ objects.
\begin{table}
\centering
\includegraphics[width=1\linewidth]{Figure10.eps}
\caption{Questionnaire's third section. For each question Q1 and Q2, the table reports the average score and its standard deviation obtained by objects (vase, sofa, pillow, or lamp, both individually taken and all together) when performing pairwise comparisons $E$ vs. $E \to O$ and $O$ vs. $O \to E$, N=391 respondents. For each question, the sum of the $E$ and $E \to O$ columns (respectively, $O$ and $O \to E$ columns) is one.}
\label{Table 3}
\end{table}
\begin{table}
\centering
\includegraphics[width=1\linewidth]{Figure11.eps}
\caption{Questionnaire's third Section, age$<30$, N=117. Data are organized as in Table \ref{Table 3}.}
\label{Table 4}
\end{table}
\subsection{Comparing culturized and non-culturized objects}
\label{Comparing culturized and non-culturized objects}
Table \ref{Table 3} reports the scores of each object class (vase, sofa, pillow, and lamp) and group ($E$, $O \to E$, $E \to O$, $O$) in the third section of the questionnaire, averaged over N respondents.
To search for evidence supporting Hypothesis 1, we considered Q1: `Which object best represents European culture and tradition?' To support Hypothesis 2, we considered Q2: `Which object do you like the most?' In both cases, we showed each respondent two $E$ vs. $E \to O$ and two $O$ vs. $O \to E$ comparisons for each object class and assigned the score of $1$ to the winner of each comparison.
After averaging over two comparisons and N respondents per object class/group, the maximum score obtained in each cell of the table is 1 (which happens if a group, say $E$, is always selected by all respondents for that object), and the sum of two competing groups (say, $E$ and $E \to O$) is 1 as well.
Analyses were performed using a one-sample T-test: in the $E$ vs. $E \to O$ comparisons, we checked if the average number of times that $E$ was selected $\mu_{E} > 0.5$ with $\alpha=0.05$; in the $O$ vs. $O \to E$ comparison, we checked if $\mu_{O \to E} > 0.5$.
It can be observed that, concerning Q1 (`Which object best represents European culture and tradition?'),
\begin{itemize}
\item for any object class, $E$ objects were selected, on average, more than half of the time,
\item for any object class, $O \to E$ objects were selected, on average, more than half of the time.
\end{itemize}
The one-tail, one-sample T-test returned that
\begin{itemize}
\item $\mu_{E} > 0.5$ with $p<0.05$ for any objects individually taken and for all objects taken together,
\item $\mu_{O \to E} > 0.5$ with $p<0.05$ for any objects individually taken and for all objects taken together,
\end{itemize}
which supports Hypothesis 1.
Concerning Q2 (`Which object do you like the most?'),
\begin{itemize}
\item for vases, pillows, lamps, $E$ objects were selected, on average, more than half of the time,
\item the same happened for the $O \to E$ group, even if average scores are lower than the $E$ group,
\end{itemize}
The one-tail, one-sample T-test returned that
\begin{itemize}
\item $\mu_{E} > 0.5$ with $p<0.05$ for vases, pillows, lamps, as well as for all objects taken together,
\item $\mu_{O \to E} > 0.5$ with $p<0.05$ for vases and pillows, $\mu_{O \to E} = 0.5$ with $p<0.05$ for lamps, $\mu_{O \to E} > 0.5$ with $p=0.052$ for all objects together,
\end{itemize}
which partially supports Hypothesis 2.
We performed a stratified analysis controlling for age: results with age $< 30$ (N=117) are in Table \ref{Table 4}.
All scores totaled by the $E$ and $O \to E$ groups tend to increase. When analyzing Q2 for the $O \to E$ group, it holds $\mu_{O \to E} > 0.5$ with $p<0.05$ for vases, sofas, pillows, and for all objects taken together. However, it holds $\mu_{O \to E} = 0.5$ with $p<0.05$ for lamps. Results almost fully support Hypothesis 2.
To search for evidence supporting Hypothesis 3, we considered Q2: `Which object do you like the most?' for the whole sample with N=391.
Specifically, as we previously did in section \ref{Recognizing culture and expressing preferences}, we considered the N participants and the four groups ($E$, $O \to E$, $E \to O$, and $O$). For each participant, in this case, we computed only the average score for all objects belonging to the $E$ and $O \to E$ groups: this process yielded two N-sized vectors, one per group.
Next, it was sufficient to correlate the two N-sized vectors by computing Pearson's $\rho_{E, O \to E}$, since it holds $\rho_{E, E \to O}=-1$ and $\rho_{E, O}=-\rho_{E, O \to E}$ as a consequence of how scores are assigned to pairwise comparisons.
The analysis returned $\rho_{E, O \to E} = 0.66$ (and, consequently, $\rho_{E, O}=-0.66$) supporting Hypothesis 3: respondents giving high scores to $E$ objects tended to privilege $O \to E$ objects as well.
\begin{table}
\centering
\includegraphics[width=0.43\linewidth]{Figure12.eps}
\includegraphics[width=0.43\linewidth]{Figure13.eps}
\caption{Questionnaire's fourth Section. Left: all age ranges, N=391. Right: age $< 30$, N=117. The table reports the average score and its standard deviation relative to $Gcult$ and $Pcult$ environments, computed over N respondents. The sum of the $Gcult$ and $Pcult$ columns is one.}
\label{Table5}
\end{table}
\subsection{ Comparing the realism of environments }
\label{ Comparing realism of environments }
Table \ref{Table5} reports the scores of each environment ($Gcult$ and $Pcult$) in the fourth section of the questionnaire, averaged over N respondents. To search for evidence supporting Hypothesis 4, we asked Q1 (`Which image looks more realistic?'). Specifically, we showed each respondent eight $Gcult$ vs. $Pcult$ comparisons and assigned a score of $1$ to the winner of each comparison. After averaging the scores given by each respondent in the eight comparisons and then averaging over N respondents, the
sum of the $Gcult$ and $Pcult$ columns is 1.
Analyses were performed using a one-sample T-test: we checked if the average number of times that $Gcult$ environments were selected $\mu_{Gcult} > 0.5$ with $\alpha=0.05$.
It can be observed that:
\begin{itemize}
\item $Gcult$ environments were selected, on average, more than half of the time;
\item the average $Gcult$ score tends to increase when performing a stratified analysis for age; the maximum value corresponded to respondents whose age $< 30$.
\end{itemize}
The one-tail, one-sample T-test returned $\mu_{Gcult} > 0.5$ with $p<.001$ in all cases, fully supporting Hypothesis 4.
\section{Discussion}
\label{Discussion}
The data show that responses may vary significantly depending on the object class considered and different objects within the same class.
Also, in some cases, the original, non-European lamp is preferred to the GAN-modified version; in other cases, the opposite is true. This result is particularly evident when objects are not pairwise compared: people express their preferences, notwithstanding their cultural belonging, depending on their tastes or just because an object looks better than another, thanks to the photographer's skill. The pairwise comparisons in the second section somehow control for confounding variables by presenting two objects with very similar shapes, dimensions, perspective, lighting, and sometimes texture and altering only the cultural context. However, based on the results, we cannot exclude that choosing different objects for pairwise comparisons might produce different outcomes.
Given these considerations, we can draw some conclusions by limiting our claims to the specific set of pictures shown.
First, it is evident from responses to Q1 in the second and third questionnaire's sections that $E$ and $O \to E$ objects are rated as belonging to the European culture to a higher degree than their non-European counterpart, which is in line with Hypothesis 1. Interestingly, $O \to E$ objects achieve a lower average score than $E$ ones: this is likely because GANs preserve some elements of the source image in the translation. $O \to E$ objects are perceived by respondents as not perfectly matching their expectations for a European object.
Second, it is evident from the responses to Q2 in the second and third sections that $E$ objects are preferred, on average, to all other objects, which is in line with Hypothesis 2.
The only cases for which this result is not confirmed are the lamp class in the second section and the sofa class in the third section.
However, it is impossible to draw similar conclusions when focussing on $O \to E$ objects: $O \to E$ vases are always preferred to $O$ vases, but the results vary for other object classes.
The individual preferences of respondents may explain this result: due to immigration, it is very common in Italian shops and houses to see Moroccan, Indian, and Chinese pillows and lamps. People are accustomed to these objects and perceive them as familiar, and many respondents may have a favorable bias towards colorful objects coming from all around the world. A similar explanation may hold when considering the sofa class: the classic style sofas shown in the questionnaire are not very common in ordinary Italian houses where a more sober design usually prevails.
Also, it is interesting that the strongest bias towards $O \to E$ objects is observed in the vase class. Greek vases are correctly perceived as European artifacts, and, on average, respondents prefer them to Far-eastern vases. This result is in line with the explanation above. According to the authors' experience, decorated Far-eastern vases, as shown in Figure \ref{all}, are less frequently encountered in ordinary Italian houses than Moroccan and Indian pillows and lamps. Then, when comparing Greek and Far-eastern vases, the choice might be more determined by cultural belonging than individual preferences and familiarity.
Third, in the second and third questionnaire's section younger respondents $<30$ tend to prefer $E$ and $O \to E$ objects: this traditionalist attitude looks coherent with the lesser experience with other cultures and the limited interest younger people may have in furniture items. Many of them have never faced the problem of furnishing a house, possibly because they still live with their parents -- in Italy, the average age when people leave their parents' house is around 30.
Fourth, we found a positive correlation between respondents giving a high score to $E$ objects and $O \to E$ objects. Even if drawing strong conclusions about the preferences of Italians is not possible (and not desirable as it might lead to stereotypes), this result confirms the possibility to identify people that prefer culturized images that match their background during interactions with robots: a person that appreciates European objects, tend to favor both $E$ and $O \to E$ objects.
Fifth, it is evident from the responses to Q1 in the fourth section that environments modified by segmenting objects, modifying them with GANs, and re-inserting them into the original image are perceived as more realistic. Some respondents motivated their choice with a written comment to complain that objects in non-GAN images have a weird perspective and they do not merge well with the background. Not surprisingly, younger people $< 30$ appear more skilled in distinguishing between the GAN-modified and non-GAN environments, with a more marked preference for the former.
\section{Conclusion}
\label{sec:conclusion}
This article introduced the concept of image ``culturization," proposed a process for object culturization based on GANs, and posed the research question of whether people appreciate images that were culturized to make them coherent with their culture.
We preliminary explored this research question by
analyzing the preferences of Italian participants towards objects belonging to European and non-European cultures and the realism of culturized environments. Overall, experiments motivate our intention to proceed further along this path: even if, as expected, not all participants prefer objects belonging to their cultural background, those who prefer European over non-European objects also tend to have a positive attitude towards objects that we culturized to be perceived as European
Our work has three main limitations. First, from a technical standpoint, the culturization process is not entirely automated since culturized objects need to be manually reinserted into the environment. However, we do not expect this to impact the hypotheses tested with experiments. Second, we should take the experimental results with a grain of salt. When preparing the questionnaires, we selected a subset of objects and environments. However, we did not validate our questionnaire to capture a hypothetic construct ``positive attitude towards European object". Then, we cannot ensure that results may be generalized to any set of objects, even if the pairwise comparisons showing the original and culturized versions of the same object may somehow control confounding variables.
Third, we implicitly assumed that it is possible to culturize an environment by modifying the objects that it contains.
This conjecture needs to be revised in the future: the texture, shape, and colors of walls and other architectural features may also play a role. However, given current GAN technology, this is likely the best we can do.
Future work will include an evaluation with non-Italian participants and real-time human-robot interaction with culturized images, not yet tested
due to the Covid-19 pandemic.
\section*{Statements and Declarations}
\label{sec:statements}
The research involves the use of an anonymous online survey. The information obtained is recorded in such a manner that the identity of the human subjects cannot readily be ascertained, directly or through identifiers linked to the subjects.
The authors declare that they have no conflict of interest.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{sec:intro}
The framework of online learning or multi-armed bandit is developed to handle the situation where the agent has to interact repeatedly with an unknown environment.
The agent has to simultaneously learn the rewards of available decisions (or arms) and search for the optimal arm.
The framework has seen great success in many applications, including dynamic pricing, clinical trials, and inventory management \cite{bubeck2012regret,den2015dynamic,bouneffouf2019survey}.
The classic multi-armed bandit considers the stationary environment,
in which the reward distribution of the arms remains constant over time.
This is hardly true in many applications.
For example, in dynamic pricing, the market environment that determines the reward of each arm (charged price) evolves constantly.
It causes the reward of each arm to drift as well.
To address this issue, there are two major frameworks proposed in the literature that do not impose any additional structures on the nonstationarity.
\begin{itemize}
\item Adversarial bandit. The environment evolves in an arbitrary and adversarial way.
As a result, the agent has to prepare for the worst-case realization of the mean reward over the horizon.
To measure the performance of a policy, the \emph{regret} is calculated against a \emph{static oracle}, i.e., the best arm in hindsight over the whole period.
The well-known EXP3 algorithm proposed in \cite{auer1995gambling} achieves the near-optimal regret $\tilde O(\sqrt{T})$\footnote{We use $\tilde O(\cdot)$ to denote the asymptotic rate omitting the logarithmic factors.}, where $T$ is the length of the learning horizon.
\item Nonstationary bandit. The drift of the environment is controlled by a variational budget $V$.
Note that $V=0$ corresponds to the stationary setting and $V=O(T)$ corresponds to the adversarial setting.
The policy needs to adapt to $V$ and its regret is measured against the \emph{dynamic oracle}, i.e., the best arm in each period.
It has been shown in \cite{besbes2015non,besbes2019optimal} that the optimal regret is $\tilde O(V^{1/3}T^{2/3})$.
\end{itemize}
Comparing the two frameworks, there are two major differences.
First, in terms of the nonstationarity in the environment, the adversarial framework allows for more drastic changes.
In particular, the total variation over time can be linear in $T$.
On the other hand, the nonstationary bandit allows for a variational budget $V$ which must be sublinear in $T$; otherwise it leads to linear regret from the above discussion.
As a result, the adversarial bandit is harder to learn.
Second, despite the difficulty, one can achieve smaller regret ($T^{1/2}$ versus $V^{1/3}T^{2/3}$) in the adversarial bandit, because the oracle against which the regret is measured is weaker.
The static oracle is the best single arm over the horizon in hindsight, as opposed to the dynamic oracle, which pulls the best arm in each period.
Therefore, the learning in the adversarial bandit has a lower standard than that in the nonstationary bandit.
The two differences drive distinct algorithmic designs, analyses, and regret bounds in the two formulations.
\subsection{Our Results}
Motivated by the difference of the oracles in the two formulations,
we propose a windowed oracle that sits in between the static and dynamic oracles.
More precisely, we divide the horizon $T$ into non-intersecting windows $\ensuremath{\mathcal W}_1,\dots,\ensuremath{\mathcal W}_N$ and consider
the oracle in window $j$ to be $\argmax_{k} \sum_{t\in \ensuremath{\mathcal W}_j}\mu_{t,k}$, where $\mu_{t,k}$ is the mean reward of arm $k$ in period $t$.
Note that the oracle in different windows may choose different arms.
The regret is thus measured against the oracle
\begin{equation*}
\sum_{j=1}^N \max_k\{\sum_{t\in \mathcal{W}_j} \mu_{t,k}\}.
\end{equation*}
It is clear that if there is a single window, then this oracle reduces to the static oracle used in adversarial bandit;
if there are $T$ windows, then it reduces to the dynamic oracle used in nonstationary bandit.
\textbf{Motivation:} The window size has a natural interpretation in practice.
It essentially reflects the expectation the agent has for the algorithm.
For example, suppose the agent is a firm entering a new business and faces unknown market demand as well as seasonal
shifts in the market environment.
The arms are the prices the firm can charge and the reward is the profit.
If the firm is ambitious, then s/he may choose the window size to be a day and benchmark against the daily optimal prices.
Otherwise the firm may choose the window size to be a month and benchmark against the monthly optimal prices.
Therefore, the choice of the window sizes reflects the conservativeness of the agent.
The regret, as a result, shows how well the algorithm does measured by different degrees of conservativeness.
We design an algorithm for a given set of windows and a variational budget $V$.
The algorithm attains the following regret.
When all the windows are relatively small, i.e., $W_i V^{2/3}\leq T^{2/3}K^{1/3}$ ($W_i$ is the size of window $i$; $K$ is the number of arms), then the regret is bounded by
\begin{equation*}
\big ( 4 + 2\sqrt{e-1} (1 + \sqrt{2}) \big ) (K V \log K)^{1/3} T^{2/3}.
\end{equation*}
When all windows are relatively large, i.e., $W_i V^{2/3} > T^{2/3}K^{1/3}$,
then the regret is bounded by
\begin{equation*}
\sum_{j=1}^N 2\sqrt{e-1} \sqrt{W_j K \log K}.
\end{equation*}
The regret is shown to be near-optimal in the two regimes.
Our result implies that the regret has a phase transition.
When the windows are small, the regret doesn't depend on the window sizes and the regret is similar to that of nonstationary bandit;
when the windows are large, the regret doesn't depend on the variational budget and the regret is similar to that of adversarial bandit.
\subsection{Related Literature}
This paper is related to the nonstationary bandit literature.
It extends the classic stochastic multi-armed bandit problems by introducing time-varying reward distributions of the arms.
Earlier papers consider discrete changes or a finite set of change points, such as \cite{auer2002nonstochastic,Garivier2011}, and thus is sometimes referred to as switching bandit.
Recent literature of this stream includes \cite{auer2019adaptively,chen2019new}.
In \cite{besbes2015non,besbes2019optimal}, the changes of the mean rewards can be continuous and a variational budget is imposed.
This setting is further investigated in various extensions, including continuous arms \cite{besbes2015stopt}, unknown variational budget \cite{cheung2019learning}, MDPs \cite{cheung2019non} and contextual bandit \cite{luo2018efficient}.
The oracle we consider in this work encapsulates the nonstationary bandit formulation considered in \cite{besbes2015non} as a special case.
Our novelty is mainly in the formulation of the windowed oracle instead of the algorithmic design and analysis.
There are other formulations to incorporate the nonstationarity, such as \cite{wu2022performance,zhou2020regime}, which are not considered in this paper.
The framework of adversarial bandit is considered in \cite{auer1995gambling} and has been studied extensively in the last few decades.
\cite{bubeck2012regret} provides a review of the literature.
Our study is related to the literature on the design of algorithms that are agnostic to the faced bandit problem (adversarial or stochastic)
and can yet achieve optimal or near-optimal regret in both adversarial and stochastic bandit problems \cite{bubeck2012best,seldin2014one,auer2016algorithm}.
Our work differs from this literature in that the agent may face a continuous spectrum of formulations between the adversarial and stochastic (nonstationary) bandit problems.
The agent doesn't need to detect the specific formulation (associated with the sizes of the windows), which is given as prior information.
This paper is also related to the papers on stochastic bandit with adversarial corruption \cite{lykouris2018stochastic,gupta2019better,golrezaei2021learning}.
While this literature uses the dynamic oracle like the nonstationary bandit, we focus on the impact of the conservativeness of the oracle (window size) instead of adversarial changes in the distribution.
\section{Problem Formulation}\label{sec:formulation}
Let $[K] \triangleq \{1,2,\dots ,K\}$ be the set of arms and $T$ be the length of the horizon.
In each period $t\in [T]\triangleq \left\{1,\dots,T\right\}$, the agent chooses an action $A_t\in [K]$ and a random reward $Y_{t, A_t}$ is observed and collected.
Conditioned on $A_t=k$, the random reward $Y_{t,k}\in [0,1]$ in period $t$ is independent of everything else and has mean $\mu_{t,k}\in[0,1]$, without loss of generality.
The agent does not know $\{\mu_{t,k}\}_{t\in[T],k\in[K]}$ initially.
She implements a policy $\pi$ to guide the decision in each period.
That is, $A_t = \pi_t(A_1,Y_{1,A_1},A_2,Y_{2,A_2}\dots,A_{t-1},Y_{t-1,A_{t-1}})$ is a mapping from the history prior to period $t$ to an arm in period $t$.
The objective of the agent is to maximize the total reward over the horizon $\sum_{t=1}^T Y_{t,A_t}$.
In the classic multi-armed bandit (MAB) framework, the expected total reward is benchmarked against the ``expected reward of the best arm'' when $\{\mu_{t,k}\}_{t\in[T],k\in[K]}$ were known.
Since we investigate the nonstationary case in which $\mu_{t,k}$ may vary over time, there are typically two ways to define the reward of the best arm(s).
One is the dynamic oracle $\sum_{t=1}^T \max_{k\in [K]}\mu_{t,k}$.
For the dynamic oracle, it pulls the best arm in time $t$:
\begin{equation*}
a_{t}^*= \argmax_{k \in [K]} \mu_{t,k}.
\end{equation*}
The other is the best arm in hindsight or a static oracle $\max_{k\in [K]}\sum_{t=1}^T \mu_{t,k}$.
Clearly, the dynamic oracle is a higher standard than the static oracle, which is targeted at the best arm in each period: if $\mu_{t,k}$ varies arbitrarily, then the agent cannot earn nearly as much reward as the dynamic oracle.
In this study, we provide a framework to bridge the two types of oracles.
We introduce $N$ time windows that partition $[T]$.
In particular, consider $W\triangleq \left\{W_j\right\}_{j=1}^N$ be $N$ positive integers such that $\sum_{j=1}^NW_j=T$.
The $j$-th time window is defined as the following contiguous index set
$\mathcal{W}_j \triangleq \left\{\sum_{i=1}^{j-1}W_i+1,\sum_{i=1}^{j-1}W_i+2,\dots, \sum_{i=1}^{j}W_i\right\}$.
In other words, $\mathcal{W}_j$ consists of $W_j$ time periods.
We consider the best arm inside a time window as follows (when there are more than one optimal arm, we choose arbitrarily among them):
\begin{equation*}
a_{\mathcal{W}_j}^*= \argmax_{k \in [K]} \left\{ \sum_{t \in \mathcal{W}_j} \mu_{t,k} \right\},
\end{equation*}
Similarly, the optimal expected reward of a fixed arm in the same time window is defined as
\begin{equation*}
\mu_{\mathcal{W}_j}^{*} = \frac{1}{W_j}\max_{k \in [K]} \left\{ \sum_{t \in \mathcal{W}_j} \mu_{t,k} \right\}.
\end{equation*}
The ``windowed'' oracle we consider can be naturally defined as
\begin{equation}\label{eq:window-oracle}
\sum_{j=1}^N\sum_{t\in\mathcal{W}_j} \mu_{\mathcal{W}_j}^*=\sum_{j=1}^N W_j \mu_{\mathcal{W}_j}^*.
\end{equation}
To make the connection, notice that if $N=T$ and $W_{j}\equiv 1$, then it is equivalent to the dynamic oracle;
if $N=1$ and $W_1=T$, then it is equivalent to the static oracle.
Therefore, our setup has adversarial and nonstationary MAB as special cases.
The (pseudo) regret of a policy $\pi$ is defined on a given set of time windows $W$.
Consider
\begin{equation}\label{eq:regret}
\begin{split}
R^{\pi}(T; W,\left\{\mu_{t,k}\right\} ) = \EE\left[ \sum_{j=1}^N \sum_{t\in \mathcal{W}_j} (\mu_{\mathcal{W}_j}^*-\mu_{t,A_t})\right]
=\sum_{j=1}^N W_j\mu_{\mathcal{W}_j}^* - \EE \left[\sum_{j=1}^N \sum_{t\in \mathcal{W}_j} \mu_{t,A_t} \right].
\end{split}
\end{equation}
The expectation is taken with respect to the randomness of the history when implementing $\pi$ as well as the internal randomization of the policy itself.
As one may expect, the regret depends on how much and frequently the expected rewards $\mu_{t,k}$ change over time.
To control for that, we use a metric called the variational budget that has been introduced in the literature to track the cumulative variation of the mean rewards, denoted by $V$.
More precisely, consider the set of mean rewards
\begin{equation}
\begin{split}
\cL(V) = \Big \{ \{\mu_{t,k}\} \in [0,1]^{T \times K}: \sum_{t=1}^T\sup_{k\in [K]} |\mu_{t,k} - \mu_{t+1,k}|\leq V \Big \}
\end{split}
\end{equation}
whose variation over the horizon is bounded by $V$.
See, e.g., \cite{besbes2015non,besbes2019optimal} for similar setups.
In the rest of the paper, we investigate the regret
\begin{equation}
R^{\pi}(T; W,V) = \sup_{\{\mu_{t,k}\} \in \cL(V)} R^{\pi}\Big (T;W,\left\{\mu_{t,k}\right\} \Big )
\end{equation}
as a function of $T$, $W$ and $V$.
Before introducing the algorithm and the analysis, we explicitly state what the agent knows initially.
The agent is aware of $K$, $T$, $W$ and $V$, as well as the fact that $\mu_{t,k},Y_{t,k}\in[0,1]$ for all $t$ and $k$.
The knowledge of $K$ is commonly assumed in the literature.
The window size $W$ is usually an intentional choice of the agent instead of a parameter in the problem: it encodes how conservative the agent is and the balance she wants to strike between dynamic and static oracles.
Therefore, it is reasonable to assume its knowledge.
Without the knowledge of $T$, it may be hard to set the window sizes, and thus we also assume it.
(Note that the ``doubling trick'' \cite{besson2018doubling} that is commonly used to relax the requirement of $T$ may not work in this case because of the windows.)
We assume the knowledge of the variational budget in our model for simplicity.
It is possible to extend to unknown $V$ using the techniques developed in \cite{cheung2019learning,auer2019adaptively,cheung2019learning}.
We leave it for future research.
\section{The Windowed EXP3 (WE3) Algorithm}\label{sec:algorithm}
Next we present the WE3 Algorithm which applies EXP3 to a number of windows combined together based on their sizes.
The algorithm follows a similar design principle to \cite{besbes2019optimal}.
It is based on the EXP3 algorithm \cite{auer1995gambling}, which is commonly used to handle adversarial bandit problems.
For completeness, the detail of the EXP3 algorithm is presented in Algorithm~\ref{alg:exp3}.
\begin{algorithm}[t]\caption{Exploration-Exploitation with Exponential weights (EXP3) }\label{alg:exp3}
\begin{algorithmic}[1]
\REQUIRE{learning rate $\gamma$, arms $K$, horizon length $T$}
\STATE{Initilization: for any $k \in [K]$, set $w_{1,k} \gets 1$. }
\FOR{$t= 1,2,\cdots,T $}
\STATE{For each $k\in [K]$, set
$$
p_{t,k} = (1-\gamma) \frac{w_{t,k}}{\sum_{s=1}^K w_{t,s}} + \frac{\gamma}{K}.
$$}
\STATE Draw an arm $A_t$ from
the probability distribution $\{ p_{t,k} \}_{k=1}^K$, receive a reward $Y_{t,A_t}$.
\STATE Set $\hat X_{t,A_t} = \frac{Y_{t,A_t}}{p_{t,A_t}}$, and set $\hat X_{t,k} = 0$ for $k \neq A_t$. Update
$$
w_{t+1,k} = w_{t,k} \exp \Big ( \frac{ \gamma \hat X_{t,k}}{K} \Big ), \text{ for all }k \in [K].
$$
\ENDFOR
\end{algorithmic}
\end{algorithm}
Note that the regret of Algorithm~\ref{alg:exp3} against the \emph{static oracle} has been studied in the literature.
For example, the regret of Algorithm~\ref{alg:exp3} is bounded by
$2\sqrt{TK\log(K)}$ by choosing the learning rate $\gamma$ optimally.\footnote{We always use $\log$ for the natural logarithm in this paper.}
This serves as a building block for our analysis.
The main algorithm, Algorithm~\ref{alg:we3}, applies the EXP3 algorithm to a batch of time periods, and restart in each batch.
The batch sizes are defined based on the problem instance, including $T$, $K$ and the window sizes.
In particular, if all window sizes $W_i$ are small (Step~\ref{step:small-w}), then
we first group a number of windows together up to a batch of size $\Delta$, and run the EXP3 algorithm for the batch.
On the other hand, if all windows sizes are big (Step 9),
then the EXP3 algorithm is simply applied to each window individually, i.e., each window is a batch.
The intuition of the WE3 algorithm is as follows: for large windows (or equivalently, a small number of windows), the variation of the mean rewards from window to window is significant.
It is optimal to forget about the learned information in the last window and restart the EXP3 algorithm.
Note that it is never optimal to split a window into several batches and restart in each batch, regardless of the size.
This is because inside each window, the windowed oracle is the same as the static oracle, in which case the EXP3 algorithm for the whole window is optimal.
For small windows (or equivalently, a large number of windows), the variation budget implies that the mean rewards learned from the last window only have limited changes.
Therefore, by grouping a number of windows together and run the EXP3 algorithm, this information can be shared.
We remark that although Algorithm~\ref{alg:we3} is already shown to attain the optimal regret (Theorem~\ref{thm:ub} below), its design may be improved in practice using the ideas in the literature for nonstationary bandit, such as sliding-window UCB \cite{Garivier2011}.
Since our focus is the formulation and the optimal regret rate, we only study the basic version.
Moreover, Algorithm~\ref{alg:we3} doesn't work for windows with widely varying sizes, for example, when some windows are small and the others are big.
We discuss the challenge of this situation in Section~\ref{sec:future-direction}.
\begin{algorithm}[t]\caption{The WE3 Algorithm }\label{alg:we3}
\begin{algorithmic}[1]
\REQUIRE{arms $K$, horizon $T$, window size $\{ W_j\}_{j=1}^N$, variational budget $V$}
\IF{$W_i \leq (T/V)^{2/3}K^{1/3}$ for all $i \in [N]$}\label{step:small-w}
\STATE{Batch size $\Delta \gets \lceil K^{1/3} (T/V)^{2/3} \rceil$, batch index $j=1$}
\WHILE{$j \leq \lceil T/\Delta \rceil$}
\STATE{$\gamma_j \gets \min \Big \{ 1, \sqrt{\frac{K \log K}{(e-1)\min \{\Delta, T-(j-1)\Delta\}}} \Big \}$;}
\STATE{Apply Algorithm~\ref{alg:exp3} to periods $t=(j-1)\Delta +1, (j-1)\Delta +2,\dots, \min \{T, j \Delta\}$}; that is, EXP3($\gamma$, $K$, $\min\{\Delta, T-(j-1)\Delta\}$)
\STATE{ $j \gets j+1$}
\ENDWHILE
\ENDIF
\IF{\label{step:big-w}$W_i \geq (T/V)^{2/3}K^{1/3} $ for all $i \in [N]$}
\FOR{$j=1,2,\cdots,N$}
\STATE{$\gamma_j \gets \min \Big \{ 1, \sqrt{\frac{K \log K}{(e-1) W_j }} \Big \}$;}
\STATE{ Apply Algorithm~\ref{alg:exp3} to periods $t\in \mathcal{W}_j$}; that is, EXP3($\gamma$, $K$, $W_j$)
\ENDFOR
\ENDIF
\end{algorithmic}
\end{algorithm}
\subsection{Regret Analysis of Algorithm~\ref{alg:we3}}\label{sec:ub}
In this section, we provide an upper bound for the regret of Algorithm~\ref{alg:we3}.
\begin{theorem}\label{thm:ub}
Suppose $KV \leq T$ and $K \geq 2$.
The regret of Algorithm~\ref{alg:we3} is bounded by
\begin{equation*}
\begin{split}
R^{WE3}(T; W,V)
\leq \big ( 4 + 2\sqrt{e-1} (1 + \sqrt{2}) \big ) (K V \log K)^{1/3} T^{2/3}
\end{split}
\end{equation*}
if $W_i V^{2/3}\leq T^{2/3}K^{1/3}$ for all $ i \in [N]$, and the regret is bounded by
$$
R^{WE3}(T; W,V) \leq\sum_{j=1}^N 2\sqrt{e-1} \sqrt{W_j K \log K}
$$
if $W_i V^{2/3} > T^{2/3}K^{1/3}$ for all $ i \in [N]$.
\end{theorem}
When all the windows have a uniform size, Theorem~\ref{thm:ub} has a simple form.
\begin{corollary}
Suppose $W_i = W$ for all $i=1,2,\cdots,N$.
We have
\begin{itemize}
\item [(a)]
The regret of Algorithm~\ref{alg:we3} is bounded by
\begin{equation*}
\begin{split}
R^{WE3}(T; W,V) \leq \big ( 4 + 2\sqrt{e-1} (1 + \sqrt{2}) \big ) (K V \log K)^{1/3} T^{2/3}
\end{split}
\end{equation*}
if $W V^{2/3}\leq T^{2/3}K^{1/3}$, and the regret is bounded by
\begin{equation*}
R^{WE3}(T; W,V) \leq 2T\sqrt{e-1} \sqrt{ \frac{ K \log K} {W} }
\end{equation*}
if $W V^{2/3} > T^{2/3}K^{1/3}$.
\item[(b)] Suppose in addition that $W = \lceil K^{1/3} (T/V)^{2/3} \rceil $, then the regret of Algorithm \ref{alg:we3} is bounded by
\begin{equation*}
\begin{split}
R^{WE3}(T; W,V) \leq 2\sqrt{ (e-1) \log K} (VK)^{1/3} T^{2/3} .
\end{split}
\end{equation*}
\end{itemize}
\end{corollary}
There are a few implications of the results.
First, the two cases are separated by a threshold of $WV^{2/3}$.
When the window size or the variational budget is small, the regret is independent of $W$ and is of the same order as that in \cite{besbes2015non,besbes2019optimal}.
This corresponds to the regime of nonstationary MAB.
When the window size or the variational budget is large, the regret is independent of $V$
and decreasing in $W$.
When $W=T$, it is consistent with the regret in \cite{auer1995gambling},
corresponding to the regime of adversarial MAB.
The two regimes have distinct regret behavior and thus display a phase transition.
Second, the threshold of $W$ that separates the two cases, satisfies
\begin{equation*}
WV^{2/3} = T^{2/3}K^{1/3}.
\end{equation*}
We may investigate a few extreme cases: if $V=\Omega(T)$ or $W=\Omega(T^{2/3})$, then the problem always falls into the adversarial regime and the regret is ${\cal O}(T/\sqrt{W})$.
The variational budget doesn't affect the regret anymore and the regret is solely determined by the window size.
By part (b), it can be shown that at the boundary, the two regret expressions are of the same order except for a constant term and a term of order $\log K$.
\begin{proof}
To prove part (a), it suffices to derive the upper bound for the case that $W V^{2/3} > T^{2/3}K^{1/3}$ as the other case directly follows from Theorem~\ref{thm:ub}.
Note that in this case, since there are $N = \frac{T}{W}$ windows, we have
\begin{equation*}
\begin{split}
& R^{WE3}(T; W,V)
\leq \sum_{j=1}^N 2\sqrt{e-1} \sqrt{W K \log K}
\\
&\ = \frac{2T\sqrt{e-1}}{W} \sqrt{W K \log K} = 2T \sqrt{e-1} \sqrt{ \frac{K \log K}{W}}.
\end{split}
\end{equation*}
This completes the proof for part (a).
For part (b), when $W = \lceil K^{1/3} (T/V)^{2/3} \rceil $, the regret incurred in this case can be bounded by
\begin{equation*}
\begin{split}
R^{WE3}(T; W,V) & \leq 2T \sqrt{e-1} \sqrt{ \frac{K \log K}{W}} \\
& \leq 2T \sqrt{e-1} \sqrt{ \frac{K \log K}{ K^{1/3} (T/V)^{2/3} }}
\leq 2\sqrt{(e-1) \log K} (VK)^{1/3} T^{2/3} .
\end{split}
\end{equation*}
This completes the proof.
\end{proof}
\section{Lower Bound}\label{sec:lb}
In this section, we derive a lower bound for the regret of any policy for the problem.
We show that the regret derived in Theorem~\ref{thm:ub} cannot be further improved.
\begin{theorem}\label{thm:lower-bound}
For given $T$, $K$, $V$ and $K$ satisfying $T \geq K \geq 2$ and $V \leq T/K$, the regret for any policy $\pi$ is lower bounded by
\begin{equation*}
R^{\pi}(T; W,V) \geq
\frac{1}{3} \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right ) (KV)^{1/3} T^{2/3},
\end{equation*}
if $ W_i \leq K^{1/3} (T/V)^{2/3} $ for all $i \in [N]$, and the regret is lower bounded by
$$
R^{\pi}(T; W,V) \geq \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right ) \sum_{i=1}^N \sqrt{ KW_i},
$$
if
$W_i \geq K^{1/3}(T/V)^{2/3}$ for all $i \in [N]$.
\end{theorem}
We briefly discuss how we construct the problem instances under the two distinct scenarios for the lower bounds.
To do so, we group the time windows $\mathcal{W}_j$'s into $m$ batches $\cT_1,\cdots, \cT_m$, which will be specified later. We denote by $T_j$ the number of time periods within batch $\cT_j$.
We construct the following bandit instances:
for periods in batch $i$ ($t\in \cT_i$), we uniformly randomly pick an arm $k_i \in [K]$ and set $\mu_{t,k_i} = 1/2 + \epsilon_i$ and $\mu_{t,s} = 1/2$ for $s \neq k_i$ in the entire batch.
More precisely, we construct $K^m$ uniformly distributed bandit instances, denoted by $(k_1,\dots,k_m)\in [K]^m$.
For an instance $(k_1,\dots,k_m)$, the mean reward satisfies
\begin{equation}\label{eq:lb-mean}
\mu_{t,k} =\begin{cases}
1/2 + \epsilon_i & t\in \cT_i, k=k_i,\\
1/2 & \text{otherwise}.\\
\end{cases}
\end{equation}
Case one: $W_j \leq \Delta = K^{1/3} (T/V)^{2/3}$ for all $j$.
This case is similar to Theorem~1 of \cite{besbes2019optimal}.
We let each batch $\cT_j$ include a number of windows such that the total time periods are just more than $\Delta$.
More precisely, it can be done iteratively as follows:
given batches $\cT_1,\cdots, \cT_{j}$, let $t_0 = 0$ and $ \{ t_s \}_{s=1}^j$ be the indices of the last window within batches $\{\cT_s\}_{s=1}^j $, respectively.
To create batch $\cT_{j+1}$, we group the next $k$ windows such that
\begin{equation}\label{eq:lb_case1_construction}
\sum_{i=t_j+1}^{t_j+k} W_i \ge \Delta,\quad \sum_{i=t_j+1}^{t_j+k-1} W_i < \Delta
\end{equation}
and let $t_{j+1}=t_j+k$.
As a result, we have that the number of time periods in $\cT_{j+1}$ satisfies $\Delta\le T_j<2\Delta$, because $W_{t_j+k}\le \Delta$ in case one.
If the remaining windows have a total number of time periods less than $\Delta$, i.e., $\sum_{i=t_j+1}^{N} W_i < \Delta$,
then we combine the remaining windows with the last batch $\cT_j$ and terminate the operation.
After the batching operation, we have that for each batch $\cT_j$, the total number of time periods therein satisfies $\Delta\le T_j<2\Delta$,
except for the last one, which we have $\Delta\le T_m<3\Delta$.
For each batch $\cT_j$, we set $\epsilon_j = \min \{ \frac{1}{4} \sqrt{ \frac{K}{T_j}}, \frac{V T_j}{2T}\}$. It can be shown that $\epsilon_j \leq \frac{1}{4}$ so that the constructed instances satisfy $\mu_{t,k} \in [0,1]$. Moreover,
the variational budget is also satisfied.
Case two: $W_i > \Delta = K^{1/3} (T/V)^{2/3}$ for $i=1,2,\cdots, N$.
In this case, we split the windows into $N$ batches $\cT_1,\cdots, \cT_N$ with batch $\cT_i = \mathcal{W}_i$.
For batch $\cT_i$, we set
$\epsilon_i = \min \Big \{ \frac{1}{4} \sqrt{ K/W_i}, \frac{ V W_i}{2T} \Big \}$. Similar as in the previous case, it can also be shown that $\epsilon_j \leq \frac{1}{4}$, $\mu_{t,k} \in [0,1]$, and the variational budget is satisfied.
By combining Theorems \ref{thm:ub} and \ref{thm:lower-bound}, we observe that the upper and lower bounds match each other in both scenarios, respectively, if we ignore the $\log K$ factor and constant terms.
Setting a uniform window size, we obtain the following corollary.
\begin{corollary}
Suppose that $W_i = W$ for $i=1,2,\cdots, N$, then the followings hold.
\begin{itemize}
\item [(a)] The regret for any policy $\pi$ is lower bounded by
\begin{equation*}
R^{\pi}(T; W,V) \geq \frac{1}{3} \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right ) (KV)^{1/3} T^{2/3}
\end{equation*}
if $ W \geq K^{1/3}(T/V)^{2/3}$, and the regret is lower bounded by
\begin{equation*}
R^{\pi}(T; W,V) \geq
\left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right )
T \sqrt{ \frac{K}{W} },
\end{equation*}
if $W \geq K^{1/3}(T/V)^{2/3}$.
\item [(b)] Suppose in addition that $W = \lfloor K^{1/3}(T/V)^{2/3} \rfloor $, then the regret for any policy $\pi$ is bounded by
\begin{equation*}
R^{\pi}(T; W,V) \geq
\frac{1}{3} \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right ) T \sqrt{\frac{K}{W}}.
\end{equation*}
\end{itemize}
\end{corollary}
\begin{proof}
(a) To prove part (a), it suffices to derive the lower bound when $WV^{2/3} \geq K^{1/3}(T/V)^{2/3}$ because the other case directly follows from Theorem \ref{thm:lower-bound}. In this case, for any policy $\pi$, Theorem \ref{thm:lower-bound} implies that
\begin{equation*}
\begin{split}
& R^{\pi}(T; W,V) \geq \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right )
\sum_{i=1}^N \sqrt{ KW} \\
& \geq \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right )
\frac{T}{W} \sqrt{ KW}
\\
& \geq \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right )
T \sqrt{ \frac{K}{W}},
\end{split}
\end{equation*}
completing the proof.
\noindent (b) When $W = \lfloor K^{1/3}(T/V)^{2/3} \rfloor < K^{1/3}(T/V)^{2/3}$, we have
\begin{equation*}
(KV)^{1/3} T^{2/3} = T \sqrt{K} \big ( \frac{V}{T\sqrt{K}} \big )^{1/3} \geq T\sqrt{ \frac{K}{W}},
\end{equation*}
which implies that
\begin{equation*}
R^{\pi}(T; W,V) \geq
\frac{1}{3} \left (
\frac{1}{8} - \frac{ \sqrt{ \log(4/3)}}{8}
\right ) T \sqrt{\frac{K}{W}}.
\end{equation*}
This completes the proof.
\end{proof}
\section{Numerical Experiments}\label{sec:numerical}
In this section, we conduct numeric experiments to test the empirical performance of our WE3 algorithm.
For simplicity, we focus on the scenario where all windows have the same size $W$.
Guided by Theorems~\ref{thm:ub} and~\ref{thm:lower-bound}, we consider two regimes:
(i) $ W V^{2/3} \leq T^{2/3} K^{1/3}$, i.e., $W$ is relatively small; (ii) $ W V^{2/3} >T^{2/3} K^{1/3}$, i.e., $W$ is relatively large.
We report the experimental details in the rest of this section.
We first consider the regime $W V^{2/3} \leq T^{2/3} K^{1/3}$ and conduct three experiments to understand the performance of our WE3 algorithm against the variational budget $V$, the number of arms $K$, and the number of periods $T$, respectively.
In each experiment, given the problem parameters $K,V,W$, and $T$, we generate problem instances by using the construction of lower bounds for $W V^{2/3} \leq T^{2/3} K^{1/3}$ in following Theorem \ref{thm:lower-bound}.
Specifically, we calculate $\Delta = K^{1/3} (T/V)^{2/3}$ and group the time periods along horizon $T$ into different batches based on \eqref{eq:lb_case1_construction}, with $W_ i =W$. Within batch $\cT_k$, we uniformly pick one arm $i_k$ and set its expected reward to be $\mu_{t,i_k} = \frac{1}{2}+ \epsilon_k$ where $\epsilon_k = \min \{ \frac{1}{4} \sqrt{ \frac{K}{ T_k}}, \frac{V T_k}{2T}\}$, and set the expected rewards of the remaining arms to be $\mu_{t,s} = \frac{1}{2}$ for all $s \neq i_k$. Under such construction, the expected rewards of all arms remain unchanged within each epoch, and vary between batches.
For the regret versus $V$, we set $K = 10$, $V \in [50,400]$, $T \in \{ 50000,100000,200000,500000 \} $, and $W \in \{ 20, 40\}$, .
For each instance, we run 10 independent replications of our WE3 algorithm, and compute the empirical regret by taking their average. To understand the performance of our algorithm against the variation budget $V$, we plot $\log V$ against the log-regret for different choices of $T$. We provide one additional line of slope $1/3$ for benchmark comparison. We summarize the results in Figure \ref{fig:01}.
For the regret versus $K$, we set $V \in \{ 50,100\}$, $W= 10$, $K \in [20,100]$, and test the WE3 algorithm over $T \in \{ 5000,100000,200000,500000 \}$. We compute the empirical regrets by averaging over 10 independent replications, and plot $\log K$ against log-regret in Figure \ref{fig:02}. For benchmark comparison, we also provide one additional line of slope $1/3$.
For the regret versus $T$, we set $K=10$, $W \in \{ 20, 40\}$, $T \in [50000,500000]$, and $V \in \{ 50,100,150,200 \} $. Similar as in previous experiments, we plot $\log T$ against the log-regret, provide one additional line of slope $2/3$, and summarize the results in Figure \ref{fig:03}.
In Figures \ref{fig:01}, \ref{fig:02} and \ref{fig:03}, we observe that the slopes of $\log V$, $\log K$, and $\log T$ against the log-regret are close to $1/3$, $1/3$, and $2/3$, respectively, matching our theoretical rates in Theorem \ref{thm:ub} that the expected regret incurred by WE3 algorithm is bounded by $\tilde {\cal O}((KV)^{1/3}T^{2/3})$ when $W V^{2/3}\leq T^{2/3}K^{1/3}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.4\linewidth]{res3_W_20}
\includegraphics[width=0.4\linewidth]{res3_W_40}
\caption{ Empirical Log-Regret when $W V^{2/3}\leq T^{2/3}K^{1/3}$ for $W \in \{ 20,40\}$, $K = 10$, $V \in [50, 400]$ }
\label{fig:01}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.4\linewidth]{res7_V_50}
\includegraphics[width=0.4\linewidth]{res7_V_100}
\caption{ Empirical Log-Regret when $W V^{2/3}\leq T^{2/3}K^{1/3}$ for $V = 50$, $W=10$, $V \in \{ 50,100\}$, $K \in [20,100]$ }
\label{fig:02}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.4\linewidth]{res5_W_20}
\includegraphics[width=0.4\linewidth]{res5_W_40}
\caption{ Empirical Log-Regret when $W V^{2/3} \leq T^{2/3}K^{1/3}$ for $W \in \{ 20, 40\}$ , $K= 10$, $T \in [50000,500000]$}
\label{fig:03}
\end{center}
\vskip -0.2in
\end{figure}
We then consider the regime $W V^{2/3} > T^{2/3} K^{1/3}$. In this case, we conduct two sets of experiments to study the performance of WE3 algorithm against the window size $W$ and number of arms $K$, respectively. For regret versus $W$, we set $V \in \{ 50, 100\}$, $K = 10$, $W \in [400,1000]$, and generate problem instances by using the construction of lower bounds when $W V^{2/3} > T^{2/3} K^{1/3}$ in the proof of Theorem \ref{thm:lower-bound}, with $T \in \{ 20000,40000,60000,80000 \}$. For each instance, we run 10 independent replications of our WE3 algorithm, and compute the empirical regret by taking their average. We plot $\log V$ against the log-regret for different choices of $T$, and provide one additional one of slope $-1/2$ for benchmark comparison. We summarize the results in Figure \ref{fig:04}.
For regret versus $K$, we generated problem instances with $W =400$ , $K\in [20,100]$, $T \in [5000,40000]$, $V \in \{ 50,100 \}$. We plot $\log K$ against the log-regret averaged over 10 independent simulations, provide one additional line of slope $1/2$, and summarize the results in Figure \ref{fig:05}.
From Figures \ref{fig:04} and \ref{fig:05}, we observe that the slopes of $\log W$ and $\log K$ against log-regret are close to $-1/2$ and $1/2$, respectively, matching our theoretical claim in Theorem \ref{sec:lb} that the regret incurred by our WE3 algorithm is bounded by ${\cal O}(T \sqrt{K/W })$.
Furthermore, from Figures \ref{fig:02} and \ref{fig:05}, we observe that the number of arms $K$ exhibits distinct impacts on the regret in the two cases. In particular, the expected regret incurred by WE3 algorithm is upper bounded by ${\cal O}(K^{1/3})$ when $W$ is relatively small, and is bounded by ${\cal O}(K^{1/2})$ when $W$ is relatively large. This also matches the upper bounds provided by Theorem \ref{thm:ub} under the two cases.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.4\linewidth]{res4_V_50}
\includegraphics[width=0.4\linewidth]{res4_V_100}
\caption{ Empirical Log-Regret when $W V^{2/3} > T^{2/3}K^{1/3}$ for $V \in \{50$,100\}, $K = 10$, $W \in [400,1000]$ }
\label{fig:04}
\end{center}
\vskip -0.2in
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.4\linewidth]{res6_V_50}
\includegraphics[width=0.4\linewidth]{res6_V_100}
\caption{ Empirical Log-Regret when $W V^{2/3} > T^{2/3}K^{1/3}$ for $V \in \{ 50,100\}$, $W = 400$, $K \in [20,100]$ }
\label{fig:05}
\end{center}
\vskip -0.2in
\end{figure}
\section{Future Directions}\label{sec:future-direction}
This paper provides a general formulation that smoothly bridges the nonstationary and adversarial bandit problems studied in the literature.
The proposed windowed oracle allows the agent to fine tune the expectation and conservativeness of the algorithm.
We believe that it provides a flexible framework to handle nonstationary environment.
However, this study also points several important open directions.
\begin{itemize}
\item Does there exist a policy or an algorithm that can always achieve the optimal regret while being agnostic to the choice of window sizes?
Since the window sizes only affect the performance of the benchmark or oracle and do not affect the objective performance of the policy (expected total reward), one would expect that the policy may not need the input of the window sizes.
This is not the case of our WE3 algorithm and it remains an open direction.
\item Drastically non-uniform window sizes.
Our algorithm currently only allows for a limit degree of nonuniformity among the windows.
That is, although the window sizes can be different, all of them have to be either large or small relative to the threshold (Step~\ref{step:small-w} in Algorithm~\ref{alg:we3}).
To understand the challenge, consider the simplified case when the window sizes are monotone, e.g., $W_1\le W_2\le\dots\le W_N$.
One may naturally apply Step~\ref{step:small-w} for the first part of the horizon when the windows are small and Step~9 for the remaining.
However, in this case, because the regret of the second part doesn't depend on $V$ (see the second inequality of Theorem~\ref{thm:ub}), the worst-case regret would allocate all the variational budget to the first part.
As a result, the effective variational budget of the second part is zero and this would change the threshold of the window size ($WV^{2/3}=T^{2/3}K^{1/3}$).
One can see that the technical challenge comes from the fact that the threshold becomes endogenous as the variational budget is allocated to the two parts.
\end{itemize}
|
\section{Introduction}
One of the most common attributes among different organisms in nature is to dwell in groups or move in consensus and mimic the activities of their local neighbors, the reason of which can be traced back to the survival instinct of those organisms. The examples of which can be found in systems as small as bacterial aggregation \cite{levy2008stochastic, chavy2016local} to macro-organisms such as flock of birds, herd of sheep, and school of fish \cite{bialek2012statistical, garcimartin2015flow, barbaro2009discrete, couzin2007collective, sumpter2010collective, herbert2016understanding}. In all these systems, the individuals organize their positions in space to aggregate together or move in unison. This phenomenon, commonly known as swarming \cite{fetecau2011swarm, topaz2004swarming, mogilner1999non, reynolds1987flocks}, is widespread in coordinated movement of a group of animals. Swarming usually means self-organization of entities in space without considering the effect of the internal state. Another such collective behavior is synchronization \cite{winfree1967biological,uzuntarla2019synchronization, kuramoto1975international,chowdhury2019convergence,anwar2022stability, pikovsky2001universal,rakshit2021relay,uzuntarla2019firing, mirollo1990synchronization,chowdhury2019synchronization1}, which is more ubiquitous in nature and technology, where the units adjust their internal states to self-organize in time. Flashing of fireflies \cite{buck1988synchronous}, chorusing frogs \cite{aihara2008mathematical}, firing neurons \cite{rakshit2018synchronization, montbrio2015macroscopic}, phase-locking in Josephson junction \cite{wiesenfeld1996synchronization, vlasov2013synchronization} are some of the well-known instances where synchronization occur. Here, only the oscillator's internal phase dynamics receives the central focus without shedding much light on the spatial motion. Examples are also found in nature where oscillator's spatial and phase dynamics affect each other \cite{riedel2005self, yan2012linking, nguyen2014emergent}. Tree frogs, crickets, katydids synchronize their calling rhythms with nearby individuals, and their movements are believed to be influenced by relative phases of their calling \cite{walker1969acoustic, greenfield1994synchronous, greenfield1993katydid}. The study of ferromagnetic colloids, sperms, land-based robots, aerial drones, and other active entities involves both dynamics \cite{snezhko2011magnetic, yang2008cooperation, barcis2019robots, barcis2020sandsbots}.
\par Particles whose spatial and phase dynamics affect each other are commonly known as {\it swarmalators} \cite{o2017oscillators} in the essence that they swarm in space to self-organize their positions and simultaneously oscillate to adjust their internal state. Earlier, in the study of {\it mobile agents} or {\it moving oscillators}, the influence of agents' motion on their phase dynamics are considered, but their spatial dynamics are not affected by phase dynamics \cite{vicsek1995novel, stilwell2006sufficient, frasca2008synchronization, chowdhury2019synchronization, majhi2019emergence}. The first step towards the study of swarmalators was taken by Tanaka and others while describing the diverse phenomena of chemotactic oscillators \cite{tanaka2007general, iwasa2010hierarchical}. Recently, O'Keeffe et al. \cite{o2017oscillators} modeled the swarmalators by incorporating appropriate coupling functions where the influence of spatial and phase dynamics on each other was aptly illustrated. Swarmalator model for global interaction among the units is governed by the pair of equations \cite{o2018ring},
\begin{equation}
\label{eq.1}
\dot{\textbf{x}}_{i} = \textbf{v}_{i}+\frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N}\bigg[\text{I}_{\text{att}}(\textbf{x}_{ij}) \text{F}_{\text{att}}(\theta_{ij}) - \text{I}_{\text{rep}}(\textbf{x}_{ij}) \text{F}_{\text{rep}}(\theta_{ij})\bigg],
\end{equation}
\begin{equation}
\label{eq.2}
\dot{\theta}_{i} = \omega_{i}+\frac{\epsilon}{N-1}\sum_{\substack{j = 1\\j \neq i}}^{N} \text{H}(\theta_{ij})\text{G}(\textbf{x}_{ij}),
\end{equation}
for $i = 1,2,\ldots,N$, where $N$ is the total number of swarmalators, $\textbf{x}_i = (x_i, y_i) \in \mathbb{R}^2$ is the position vector of the $i$-th swarmalator, $\theta_i \in \mathbb{S}^1$ is its internal phase, while $\textbf{v}_i$ and $\omega_i$ denote its self-propulsion velocity and natural frequency, respectively with $\textbf{x}_{ij} \equiv \textbf{x}_j - \textbf{x}_i$ and $\theta_{ij} \equiv \theta_j - \theta_i$. The spatial attraction and repulsion are governed by the functions $\text{I}_\text{att}$ and $\text{I}_\text{rep}$. $\text{F}_{att}$ and $\text{F}_\text{rep}$ represent the influence of phase similarity on spatial attraction and repulsion, respectively. In Eq.\ \eqref{eq.2}, phase interaction between the swarmalators is controlled by the function $\text{H}$ and the influence of spatial proximity on phase dynamics is given by $\text{G}$. Swarmalators' phases are coupled with strength $\epsilon \in \mathbb{R}$. When $\epsilon$ is positive (attractive coupling), swarmalators try to minimize their phase differences while a negative value (repulsive coupling) of $\epsilon$ increases the incoherence of their phases. In Ref.~\cite{o2017oscillators}, three stationary (one each for attractive, repulsive, and absence of phase coupling) and two non-stationary states (both for repulsive phase coupling) of possible long-term aggregation are found. More new states are reported when the model of swarmalators is extended by addition of periodic forcing \cite{lizarraga2020synchronization}, noise \cite{hong2018active}, and finite cut-off interaction distance \cite{lee2021collective, jimenez2020oscillatory}. Finite size effects \cite{o2018ring} of the population of swarmalators and the well-posedness of solution in the mean-field limit \cite{ha2021mean, ha2019emergent} are also studied.
The collective behavior of the swarmalators positioned on the one-dimensional ring is studied analytically in Ref.\ \cite{o2021collective}. The sign of phase coupling determines the coherent or incoherent nature of the states.
\par Mixed influence of positive and negative couplings can be found in interactions in neuronal networked systems \cite{hopfield1982neural} and in the calling behavior of Japanese frogs \cite{aihara2008mathematical}. Coupling disorder with random interaction of swarmalators' phases is introduced in Ref.~\cite{hong2021coupling}, where chimera-like states are observed. The nature of coupling in most real world systems is often complex, which motivates us to study the behavior of swarmalators under the mixed coupling strategy. More importantly, earlier, most of the existing studies on coexisting attractive-repulsive interaction \cite{hong2011kuramoto,majhi2020perspective,hong2011conformists,chowdhury2021antiphase,yuan2018periodic,chowdhury2020effect} were performed on static network formalism. Such signed networks can display fascinating macroscopic dynamics \cite{hong2011kuramoto,iatsenko2014glassy,iatsenko2013stationary}, including the $\pi$ state, the traveling wave state, and the mixed state.
Recently, the impact of such competitive interactions through the concurrence of positive-negative coupling has been investigated on the time-evolving networks of mobile agents \cite{chowdhury2019synchronization,chowdhury2020distance}, leading to diverse peculiar dynamical states, including extreme events \cite{chowdhury2021extreme,chowdhury2021extrememap}. However, these studies consider only the unidirectional influence of spatial dynamics towards the oscillator's amplitude and phase dynamics. This present article focuses on the bidirectional interplay between swarmalators' phase values and spatial positions. To the best of our knowledge, in spite of the colossal importance of time-varying interaction \cite{holme2012temporal,nag2020cooperation,majhi2017amplitude,dixit2021dynamic,ghosh2022synchronized,dixit2021emergent,majhi2017synchronization} from diverse aspects, the study of swarmalators is less explored, particularly during the simultaneous presence of
attractive-repulsive temporal interaction in the phase dynamics
\par We are curious to investigate how competitive phase interaction induces long-term states of position and phase aggregation of the system. We design a particular coupling scheme to serve this purpose. A communication circle with a fixed and uniform radius is associated with each swarmalator so that the swarmalators within this vision range interact bidirectionally with positive coupling strength. Outside this attractive vision range, each swarmalator goes through repulsive phase coupling. Under this setup, suitable choices of parameters lead to various emergent states due to the interplay between attractive and repulsive phase couplings among the swarmalators with global spatial attraction and repulsion. We emphasize the role of attractive vision radius in realizing the collective dynamics and examine the possible routes for achieving the static sync state. The extreme limit of this attractive vision radius transforms the phase interaction into a global attractive or repulsive one and leads to different emergent patterns like static sync, static async, active phase wave, and splintered phase wave. We elaborately discuss the main features of these patterns depending on the phase-dependent spatial dynamics and position-dependent phase dynamics. The absence of repulsive coupling strength ensures the manifestation of static cluster synchronization. We are able to derive a sufficient condition for the transition from incoherence to static sync state. We also identify a novel state, viz.\ static inscribed cluster with local attractive coupling. When $F_{rep}$ depends on the phase dynamics explicitly, we trace out the static ring phase wave state under the influence of global repulsive coupling. We analytically derive the radius of this stationary state and validate the theoretical findings through numerical simulation. Carrying forward our interest in local phase interaction, we model swarmalators where the Stuart-Landau limit-cycle oscillators govern the phase dynamics.
\par The subsequent sections of this article are arranged as follows. In Sec.\ \ref{sec:level2}, we introduce a two-dimensional swarmalator model endowed with global co-existing spatial attraction-repulsion and competitive phase interaction. Section\ \ref{sec:level3} contains theoretical analysis where we ensure the non-existence of finite time collision and the existence of a minimal inter-particle distance between the swarmalators. We discuss in detail the main findings of our investigation in Sec.\ \ref{sec:level4}. The emergence of different collective behaviors and their characterization with order parameters are studied by considering several cases. Section\ \ref{sec:level5} includes the study of our model when the Stuart-Landau oscillator \cite{kuramoto2003chemical} is used for the phase dynamics with local attractive phase coupling. Finally, we summarize our work in Sec.\ \ref{sec:level6} and discuss the possible scope for future research. We include Appendix A \eqref{sec:level9} dealing with the swarmalators moving in three-dimensional space to inspect the impact of higher dimensional spatial dynamics on our proposed model.
\section{\label{sec:level2}Swarmalator model with competitive phase interaction}
\par Every swarmalator moves in space with an attractive vision radius $r$, and the closed circular region it covers being at the center, can be considered as its attractive range of interaction. In a sense, they can feel the presence of nearby swarmalators and couple their phases attractively with positive coupling strength. Outside the attractive vision range the positive connectivity gets lost, and they are coupled repulsively (see Fig.\ \ref{Fig.1}). The attractive-repulsive phase coupling and the all-to-all spatial attraction and repulsion make the swarmalator model more challenging as the competitive phase interaction increases the possibility of finding new states. Instead of choosing this circular vision shape of a uniform radius, one can select other polygons and heterogeneous communication radii. However, the results shown in this manuscript remain the same qualitatively for such choices. We propose the following model,
\begin{equation}
\label{eq.3}
\dot{\textbf{x}}_{i} = \textbf{v}_{i}+\frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N}\left[ \frac{\textbf{x}_{ij}}{|\textbf{x}_{ij}|^ \alpha} (1+J\cos(\theta_{ij})) - \frac{\textbf{x}_{ij}}{{|\textbf{x}_{ij}|}^{\beta}} \right],
\end{equation}
\begin{equation}
\label{eq.4}
\dot{\theta}_{i} = \omega_{i}+\frac{\epsilon_a}{N_{i}}\sum_{\substack{j = 1\\j \neq i}}^{N} \text{A}_{ij}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|^\gamma} + \frac{\epsilon_r}{N-1-N_{i}}\sum_{\substack{j = 1\\j \neq i}}^{N} \text{B}_{ij}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|^ \gamma}.
\end{equation}
Here, we consider the set $\Lambda_{i}(r) = \{j \in \{1,2,...,N\}\; \mbox{such that} \; |\textbf{x}_{ij}| \le r, j \ne i\}$ and $N_{i}$ is the number of elements in $\Lambda_{i}(r)$. $N_{i} \ne 0$ means there is at least one swarmalator inside the attractive vision range of the $i$-th swarmalator except itself and $N_{i} \ne N-1$ indicates the presence of at least one swarmalator outside its attractive vision range. $\text{A}=[\text{A}_{ij}]_{N \times N}$ and $\text{B}=[\text{B}_{ij}]_{N \times N}$ are the adjacency matrices for attractive and repulsive phase couplings, respectively, and are defined as,
\begin{equation*}
\text{A}_{ij} =
\begin{cases}
1 & \text{if $j \in \Lambda_{i}(r)$}\\
0 & \text{otherwise}
\end{cases}
;\hspace{5pt} \text{B}_{ij} =
\begin{cases}
1 & \text{if $j \not\in \Lambda_{i}(r) \cup \{i\}$}\\
0 & \text{otherwise}
\end{cases}
\end{equation*}
Equation \eqref{eq.4} remains well-defined, if $N_{i}$ $\neq$ $0$ and $(N-1)$. When $N_{i}$ attains these two values $0$ and $(N-1)$, then the matrices $\text{A}$ and $\text{B}$ become null matrices, respectively. We will study two extreme cases for $N_{i}$ $=$ $0$ and $N_{i} = (N-1)$ later in the subsection~\ref{subsec:level1} in terms of attractive vision radius $r$. We choose power law attraction and repulsion for $\text{I}_{\text{att}}$ and $\text{I}_{\text{rep}}$ with positive exponents $\alpha$ and $\beta$, respectively. Note that, Eq.\ \eqref{eq.3} can be written as
\begin{equation}
\dot{\textbf{x}}_{i} = \textbf{v}_{i}+\frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N} \text{F}(|\textbf{x}_{ji}|) \frac{\textbf{x}_{ji}}{|\textbf{x}_{ji}|},
\label{eq.5}
\end{equation}
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 0.70]{figure1.jpg}}
\caption{{\bf Schematic diagram of $N = 100$ swarmalators with circular attractive range of vision}: The swarmalators positioned randomly in the two-dimensional plane inside the bounded region $[-1,1] \times [-1,1]$ are represented by the small dots which are colored according to their phases, drawn from $[-\pi, \pi]$ uniformly at random. Two circular regions are the respective attractive vision ranges for two distinct particles, where the central swarmalators are attractively phase coupled with all other swarmalators inside and is repulsively phase coupled with all those who are outside the regions.}
\label{Fig.1}
\end{figure*}
where $\text{F}(|\textbf{x}_{ji}|) = \frac{1}{|\textbf{x}_{ji}|^{\beta-1}} - \frac{1+J\cos(\theta_{ji})}{|\textbf{x}_{ji}|^{\alpha-1}}$ is the net force exerted on the $i$-th swarmalator by the $j$-th one, along the direction $\textbf{x}_{ji}$. When this force is negative, swarmalators attract each other following Eq.\ \eqref{eq.5} and similarly positive force means repulsion between them. As a result, this force must be positive (or repulsive) for nearby swarmalators so that they do not collide, whereas it must be negative (or attractive) for swarmalators far away from each other so that they do not disperse indefinitely \cite{fetecau2011swarm}. To ensure this, we need to choose $\alpha$ and $\beta$ such that $1\le \alpha < \beta$ holds. In our work, without loss of any generality, we consider the influence of phase only on spatial attraction (i.e., spatial repulsion between swarmalators is independent of their phases) by our choice of functions, $\text{F}_{\text{att}}(\theta) = 1+J\cos(\theta)$ and $\text{F}_{\text{rep}} = 1$. However, we also study the behavior of swarmalators when phase similarity affects spatial repulsion for a particular case in Sec.\ \eqref{sec:level7}. The parameter $J$ measures how phase similarity influences spatial attraction. We choose $0 \leq J <1$, so that the attraction function is always positive. Positive value of $J$ indicates that swarmalators which are in nearby phases, attract themselves spatially and stay close to each other. If $J$ is negative, swarmalators are spatially attracted to all those in opposite phases. In Sec.\ \eqref{sec:level8}, we study our model by considering negative values of $J$ under the presence of local attractive coupling only. The phase interaction function $\text{H}$ is taken as the sine function inspired by the Kuramoto model \cite{kuramoto1975international}. To capture the spatial influence of the swarmalators on their phase dynamics, we choose $\text{G}(\textbf{x}) = \frac{1}{|\textbf{x}|^ \gamma} (\gamma > 0)$. Results in this paper are presented with a specific choice of parameters' values, viz.\ $\alpha = 1$, $\beta = 2$, and $\gamma = 1$, unless otherwise mentioned. These choices of $\alpha$, $\beta$, and $\gamma$ make sure that the solutions are bounded and prevent inter-particle collision.
For simplicity, we choose identical swarmalators so that $\textbf{v}_i = \textbf{v}$ and $\omega_{i} = \omega$, and by proper choice of reference frame we set both $|\textbf{v}|$ and $\omega$ to zero. In Fig.\ \ref{Fig.1}, a schematic diagram of the initial positions of the swarmalators in the two-dimensional plane is shown, where they are colored according to their initial phases. Initially, the swarmalators are positioned randomly inside the region $[-1,1] \times [-1,1]$ and their phases are selected randomly from $[-\pi,\pi]$.
\par It should be noted that the effect of spatial position on the internal phase has two folds. The swarmalator's position in space controls the phase coupling strategy as well as the effective interaction strength. The positions of the swarmalators in the two-dimensional plane determine whether their phases are coupled attractively or repulsively, depending on the attractive vision radius $r$. Also, the strength of phase interaction is multiplied by $\frac{1}{\textbf{x}_{ij}}$ which indicates that the swarmalators staying nearby in space are coupled with more strength than swarmalators who stay some distance apart. The parameters $\epsilon_a$ ($>0$) and $\epsilon_r$ ($<0$) control the strength of attractive and repulsive phase interactions, respectively. To avoid monotony, from hereon by vision radius and vision range, we will mean attractive vision radius and attractive vision range, respectively. For the sake of simplicity, we also include Table \ref{tab:table1} mentioning all the observed states and their corresponding abbreviations used in this article.
\begin{table}
\caption{\label{tab:table1} Various emergent collective states and their abbreviations:}
\centerline{
\begin{tabular}{ccccc}
Emergent collective states & Abbreviations\\ \hline
Static sync& SS \\
Static async& SA \\
Static phase wave& STPW \\
Splintered phase wave& SPPW \\
Active phase wave& APW \\
Static cluster sync& SCS \\
Static $\pi$ & SPI \\
Attractive mixed phase wave& AMPW \\
Repulsive mixed phase wave& RMPW \\
Static inscribed cluster & SIC\\
Static ring phase wave & SRPW
\end{tabular}}
\end{table}
\section{\label{sec:level3}Theoretical Analysis}
The proposed model defined by Eqs.\ \eqref{eq.3} and \eqref{eq.4} captures the spatial and phase dynamics of swarmalators where they are interconnected. This interplay between swarmalators' position and phase gives rise to different complex dynamical patterns which we will discuss in the following sections. Before moving into illustrating the asymptotic states of the system, first we discuss some basic properties of the proposed model. Let $\Gamma(\textbf{x}_{ij},\theta_{ij}) = \frac{\textbf{x}_{ij}}{|\textbf{x}_{ij}|^ \alpha} (1+J\cos(\theta_{ij})) - \frac{\textbf{x}_{ij}}{{|\textbf{x}_{ij}|}^{\beta}}$. So, Eq.\ (\ref{eq.3}) with $\textbf{v}_{i}=\bf{0}$ can be rewritten as,
\begin{equation}
\label{eq.6}
\dot{\textbf{x}}_{i} = \frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N} \Gamma(\textbf{x}_{ij},\theta_{ij}).
\end{equation}
It is easy to notice that, $\Gamma(\textbf{x}_{ij},\theta_{ij}) = - \Gamma(\textbf{x}_{ji},\theta_{ji})$, i.e., $\Gamma$ is skew-symmetric under the exchange of indices $i$ and $j$. If we take the total sum by considering all the particles, then it will be identically zero. As a result
\begin{equation}
\label{eq.7}
\sum_{\substack{i = 1}}^{N}\dot{\textbf{x}}_{i} = \frac{1}{N-1} \sum_{\substack{i = 1}}^{N}\sum_{\substack{j = 1\\j \neq i}}^{N} \Gamma(\textbf{x}_{ij},\theta_{ij}) = 0.
\end{equation}
So, the center of position of the system $\frac{1}{N}\sum_{\substack{i = 1}}^{N}{\textbf{x}}_{i}$ is always conserved. Hence, the arithmetic mean of the initial distribution of the spatial position of the system at $t=0$ from the box $[-1,1] \times [-1,1]$ remains invariant for future iterations. However, the conservation of mean phase $\frac{1}{N}\sum_{\substack{i = 1}}^{N}\theta_{i}$ can not be assured, as the phase interaction in Eq.\ (\ref{eq.4}) is split into two different components, viz.\ attractive and repulsive parts. These two separate portions do not always combine resulting into the vanishing of the total sum $\sum_{\substack{i = 1}}^{N}\dot{\theta}_{i}$.
\par One of the main difficulties in studying the system theoretically is the presence of singular terms like $|\textbf{x}_{ij}|^ \eta$ in the denominators of the coupling functions. The question that arises is whether the system is well-defined for the case of a collision at any finite time between two or more swarmalators or not.
To encounter this, we chose coupling functions so that the finite time collision avoidance between the swarmalators can be assured. To avoid notation complexity, we set $\mathcal{N} = \{1,2,...,N\}$.
\begin{theorem}\label{Theo:1}
Suppose $1 \le \alpha < \beta$ and the initial data is chosen for which there is no collision among the swarmalators, i.e., for all $i,j \in \mathcal{N}$ and $i \ne j$,
\[ \min_{1\le i,j \le N} |\mathbf{x}_{i}(0)-\mathbf{x}_{j}(0)| > 0. \]
Then, in finite time, the swarmalators will never collide, i.e., there exists a global solution of Eqs.\ \eqref{eq.3} and \eqref{eq.4} with
\[\mathbf{x}_{i}(t) \ne \mathbf{x}_{j}(t),\]
for all $i \ne j$, $t \in (0, \infty)$. Moreover, there exists a positive lower bound for the minimal inter-particle distance, $\delta$ such that,
\[\inf_{0\le t< \infty} \min_{i,j} |\textbf{x}_{i}(t)-\textbf{x}_{j}(t)| \ge \delta.\]
\end{theorem}
\paragraph*{ Sketch of the proof:}
This theorem mainly focuses on how the swarmalators move in space without colliding with each other. To prove these results, we have adopted the method of Ref.~\cite{ha2019emergent} where a general swarmalator model with global space and phase coupling was considered. Collision avoidance and minimal inter-particle distance are spatial properties of the swarmalators, irrespective of their phase. Eq.\ (\ref{eq.3}) of our model corresponds to the spatial dynamics of (1.1) in Ref.~\cite{ha2019emergent} with the choices $\omega_i = 0, \forall i \in \mathcal{N}$, $\Gamma_a(\theta) = 1+J\cos(\theta)$, and $\Gamma_r(\theta) = 1$. The main difference of our model given by Eqs.\ \eqref{eq.3}-\eqref{eq.4} and the one in Ref.~\cite{ha2019emergent} is in the phase dynamics of the swarmalators. Since the phase equation does not play any role in proving these results, we can proceed in the same way to prove this theorem.
The last theorem eliminates the possibility of collision among particles in finite time. So, the solution of this system is always well-posed. Now, we can proceed to study the collective states of their position aggregation and phase synchronization. To investigate the phase synchrony of the swarmalators, we use the complex order parameter of Kuramoto model defined by
\begin{equation}
R e^{l \bar{\Theta}} = \frac{1}{N} \sum_{\substack{j = 1}}^{N} e^{l \theta_j}, (l = \sqrt{-1})
\label{eq.8}
\end{equation}
where $0 \le R \le 1$ measures the coherence of swarmalators' phases and $\bar{\Theta}$ is their mean phase. Here $R \ll 1$ represents there is no convergence into a unique phase among the swarmalators. On the other hand, when swarmalators' phases are fully synchronized, then we have complete phase coherence with $R = 1$.
\par Swarmalators organize themselves in space and adjust their phases with others. The correlation between swarmalator's phase $\theta$ and spatial angle $\phi = \tan ^{-1}(y/x)$ is measured by another order parameter defined by
\begin{equation}
S_{\pm} e^{l \Psi_{\pm}} = \frac{1}{N} \sum_{\substack{j = 1}}^{N} e^{l(\phi_j \pm \theta_j)}.
\label{eq.9}
\end{equation}
When the phase and spatial angle of each swarmalator is perfectly correlated, we have $\phi_i = \pm \theta_i + C$ (for some constant $C$ which depends on the initial conditions). In this case, we have $S_\pm = 1$, and this value of $S_\pm$ decreases when the correlation is reduced. The maximum of $S_\pm$ is defined by $S = \max(S_+,S_-)$ and can be chosen to measure this correlation.
\par For numerical simulation in our study, we use FORTRAN 90 compiler and the integration is done using Runge Kutta method of order 4 with step size 0.01. Swarmalators are initially distributed inside the bounded region $[-1,1] \times [-1,1]$ and their phases are drawn from $[-\pi, \pi]$ randomly.
\begin{figure*}[hpt]
\includegraphics[scale = 0.36]{figure2.pdf}
\caption{{\bf Scatter plots of swarmalators in the $(x,y)$ plane for different static and active states}: Here, the term `active' reflects movement in space and continuous change in the phases of the swarmalators in the asymptotic states. In contrast, the term `static' demonstrates stationarity in spatial as well as in the phase dynamics. The numerical simulations are done with $N = 500$ swarmalators over 30000 iterations with integration step size 0.01. The snapshots shown here are taken at the last $(300)$ time unit. (a) Static sync (SS) for $(J, r, \epsilon_a, \epsilon_r) = (0.1,5.0,0.5,-0.5)$, (b) static async (SA) for $(J, r, \epsilon_a, \epsilon_r) = (0.1, 10^{-5}, 0.5, -0.5)$, (c) static phase wave (STPW) for $(J, r, \epsilon_a, \epsilon_r) = (0.9, 10^{-5}, 0.5, 0.0)$, (d) splintered phase wave (SPPW) for $(J, r, \epsilon_a, \epsilon_r) = (0.9, 10^{-5}, 0.5, -0.1)$, and (e) active phase wave (APW) for $(J, r, \epsilon_a, \epsilon_r) = (0.75, 10^{-5}, 0.5, -0.5)$. We classify all these states depending on their respective phase values as well as spatial positions. Please see the main text for detailed descriptions of these states. }
\label{Fig.2}
\end{figure*}
\section{\label{sec:level4}Results}
In this section, we consider different coupling topologies. The behaviors of the swarmalators when vision radius is very large and very small are studied in Sec.~\ref{subsec:level1}. Local attractive coupling is considered in Sec.~\ref{subsec:level2} for phase interaction among swarmalators. Finally, in Sec.~\ref{subsec:level3}, asymptotic states of the swarmalators under attractive-repulsive phase coupling are discussed.
\subsection{\label{subsec:level1}Extreme limits of vision radius ($r$)}
First, we want to study the two cases when the vision range of the swarmalators is either very large or very small. When the swarmalators move in space with an infinite vision radius, they will sense the presence of every other swarmalators within its vision range. In that case, $\Lambda_i (r) = \{1,\ldots,i-1,i+1,\ldots,N\}$ and as a result the repulsive matrix $\text{B}$ becomes null. This means swarmalators' phases are attractively coupled only with the coupling strength $\epsilon_a >0$. So, the phase dynamics given by Eq.\ \eqref{eq.4} with $\omega_i = 0$ effectively becomes,
\begin{equation}
\dot{\theta}_{i} =\frac{\epsilon_a}{N-1}\sum_{\substack{j = 1\\j \neq i}}^{N}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|^\gamma} \hspace{0.2cm}(\epsilon_a>0).
\label{eq.10}
\end{equation}
The positive phase coupling strength along with the global interaction (note that, the attractive coupling matrix $\text{A}$ becomes $\text A_{ij} = 1$ ($i \ne j$), $\text A_{ii} = 0$ for $i,j \in \mathcal{N}$) makes the swarmalators minimize their phase difference. Phase coherence among the swarmalators is found, where they lie
inside a circular disc in two-dimensional plane. The formation of this disc structure is influenced by the force function $\text{F}$ (cf.\ Eq.\ \ref{eq.5}), which leads to uniform density of particles inside a disc in the absence of $J$. Due to the complete coherence in swarmalators' phases, the phase influence on spatial position $1+J\cos(\theta_{ij})$ becomes $1+J$, a constant, and the disc-like structure is sustained with the radius being scaled by $\frac{1}{1+J}$. The order
parameter R gives the value 1 here, justifying the swarmalators' totally phase
synchronized state. In sufficiently long time, swarmalators stop moving in space, and remain spatially static. Besides, their phases become static as well. This state is called static synchrony (SS) (See Fig.\ \ref{Fig.2}(a)).
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 0.85]{figure3.jpg}}
\caption{{\bf $\epsilon_r$-$J$ parameter space when $r<\delta$}: We are able to map different emergent collective states in the two-dimensional $\epsilon_r$-$J$ parameter space for a small attractive vision radius $r=10^{-5} < \delta$. We perform the numerical simulations with $N=100$ swarmalators for $500$ time units with step size $0.01$ and take last half of the data to calculate $T$ and last $10$$\%$ data points for $S$. For each point, 10 realizations are taken. In subfigure (a), the red dots are points where $S$ bifurcates from zero to non-zero value and blue dots are the points, where $T$ bifurcates from zero when $S$ is non-zero. The points have been joined by dashed line to separate the region of occurrence of SA ($S = T = 0$), APW ($S \ne 0$, $T \ne 0$), and SPPW ($S \ne 0$, $T = 0$) states. To demonstrate these bifurcations more accurately, we plot two subfigures (b) and (c). The subfigure (b) is classified based on the order parameter $S$, where the red region signifies non-zero $S$ values indicating the emergence of either SPPW or APW. The blue region contemplates the SA state with $S=0$. To distinguish between SPPW and APW states, we utilize the order parameter $T$ in subfigure (c), where $T=0$ (blue region) highlights the SA state or SPPW state. The non-zero values of $T$ (red region) reveal the manifestation of the APW state. For further information, please see the main text.}
\label{Fig.3}
\end{figure*}
\par On the extreme opposite to the previous scenario, when the vision radius ($r$) is very small (less than the lower bound of minimal inter-particle distance, $\delta$), every swarmalator lies outside the vision range of every other swarmalators. They lose their connection with others which is required for attractive phase coupling. The presence of only repulsive phase interaction is validated by the fact that $\text{A}$ is the zero matrix now. The phase dynamics in this case is governed by the equation
\begin{equation}
\dot{\theta}_{i} =\frac{\epsilon_r}{N-1}\sum_{\substack{j = 1\\j \neq i}}^{N}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|^\gamma} \hspace{0.2cm}(\epsilon_r<0).
\label{eq.11}
\end{equation}
Here, we observe three different asymptotic behaviors of swarmalators depending on the parameter values of $J$ and $\epsilon_r$ of Eqs.\ \eqref{eq.3} and \eqref{eq.11}.
One of them is a static state where the positions and phases of swarmalators become static and their phases are distributed over $[-\pi, \pi)$, called static asynchrony (SA) (in Fig.\ \ref{Fig.2}(b)). In the other two states, called splintered phase wave (SPPW) and active phase wave (APW), swarmalators move. For a small value of $|\epsilon_r|$, swarmalators break into clusters of distinct phases and inside each cluster, they execute small oscillations about their mean values of both position and phase. But they do not move from one cluster to another once this SPPW state is achieved (Fig.\ \ref{Fig.2}(d)). The order parameter $S$ gives nonzero value here indicating some correlation between their phase and spatial angle. If the value of $\epsilon_r$ is decreased, the SPPW pattern vanishes and swarmalators start to rotate in space and their phases execute full cycle from $-\pi$ to $\pi$ (Fig.\ \ref{Fig.2}(e)). In this APW state, $S$ also gives nonzero value which makes it impossible to distinguish these two states from each other only using $S$. Following the behavior of swarmalators in the APW state another order parameter $T$ is defined, which is the fraction of swarmalators that execute at least one full cycle in space and phase after discarding the transient. In Fig. \ref{Fig.3}, we investigate the $\epsilon_r$-$J$ parameter space to identify the regions of existence of SA, SPPW, and APW states, respectively. In the SA state, both the order parameters $S$ and $T$ give zero values. $S$ is non-zero and $T$ is zero in SPPW, whereas both non-zero in the APW state. We summarize all this information in Table \ref{tab:table2}. The stationary nature of the SA state trivially makes $T = 0$ as swarmalators are static in space, and the asynchronous behavior of their phases gives $S = 0$ which can be seen in Fig.\ \ref{Fig.3}. There is some correlation between swarmalators' spatial positions and internal phases both in the SPPW and APW states and as a result, $S$ is non-zero in both these states (see Figs.\ \ref{Fig.3}(b) and \ref{Fig.3}(c)). The inability of $S$ to distinguish these two states calls for the order parameter $T$ which is non-zero only in the APW state (see Fig.\ \ref{Fig.3}(c)).
\begin{table}
\caption{\label{tab:table2} Different values of $S$ and $T$ are used to separate the $\epsilon_r$-$J$ space in Fig.\ \ref{Fig.3}.}
\centerline{
\begin{tabular}{ccccc}
$S$&$T$&Emerging state\\ \hline
$\approx 0$&$\approx 0$& Static async (SA) \\
$\ne 0$&$\approx 0$& Splintered phase wave (SPPW) \\
$\ne 0$&$\ne 0$& Active phase wave (APW) \\
\end{tabular}}
\end{table}
\subsection{\label{subsec:level2}Local attractive coupling}
In the absence of negative phase coupling (i.e., $\epsilon_r = 0$), the swarmalators are only allowed to couple their phases with positive strength ($\epsilon_a$) among nearby neighbors, i.e., with other swarmalators which lie inside their range of vision. This is the case where swarmalators staying far apart from each other in space lose the connection between them to have an influence on each other's phase. For a suitable low value of vision radius $r$, the attractive adjacency matrix $\text{A}$ becomes null and both $\epsilon_a = \epsilon_r = 0$ effectively. The swarmalators are locked to their initial phases and they rearrange themselves in space following their spatial dynamics. For $J = 0$, the influence of phase on position is absent and as a result swarmalators arrange themselves inside a disc following only swarming dynamics while their phases are uniformly distributed in the range $[-\pi, \pi)$. They become static in space and phase settling into the SA state. For a non-zero value of $J$, the swarmalators in similar phase attract each other and form an annulus like structure, whose inner and outer radii depend on the values of $J$, and eventually become static in space and phase. In this state, known as static phase wave (STPW) (see Fig.\ \ref{Fig.2}(c)), spatial angle $\phi$ and phase $\theta$ of every swarmalator is perfectly correlated, which gives $S = 1$. This state is a static state and $T$ gives zero value.
\begin{figure}[hpt]
\centerline{
\includegraphics[scale = 0.55]{figure4.pdf}}
\caption{{\bf Emergence of static cluster sync state for $r = 0.3$, $J = 0.8$, and $\epsilon_a = 0.5$ in the absence of repulsive coupling}: (a) Snapshot of the swarmalators at $t = 50$ time units where the swarmalators are colored according to their phase. (b) Phase time series of the swarmalators over $t = 50$ time units. Swarmalators converge into five different stationary phases. The phase difference of $2\pi$ between two clusters represents the identical phase. Thus, the system settles down here into three distinct clusters. Evolution of a static cluster synchrony state from initial positions is shown by a video (Figure4a.gif) in the GitHub repository \cite{web_5}.}
\label{Fig.4}
\end{figure}
When the radius of vision is increased beyond a critical value $r \approx 0.2$, swarmalators start to feel the presence of others inside their interaction range and couple their phases attractively to minimize their respective phase differences. This enhances phase similarity between nearby swarmalators. For large $r$, all their phases get fully synchronized and the emergence of static synchrony is observed. However, for a large value of $J$, the spatial attraction strength between swarmalators, given by $1 + J \cos(\theta_j - \theta_{i})$, is much bigger for the ones who are in similar phases than for those who share a relatively higher phase difference. This forces them to form groups among themselves where in each group, swarmalators' phases get totally synchronized. Between two distinct groups there is always a difference of phases and in space they are at least $r$ distance away from one another. We name this new state as {\it static cluster synchrony} (SCS), as eventually swarmalators cease to move in space and their phases become static. Note that, this formation of clusters in space happens because of the fact that swarmalators in similar phases attract themselves in space with more strength than others, even when there is global spatial attraction and repulsion among them. We validate the phase dependence of position of swarmalators in Fig.\ \ref{Fig.4}(a) by snapshot of this state where swarmalators are grouped into three clusters and they are colored according to their phases. Figure \ref{Fig.4}(b) shows their phase time series which confirms the fact that their phases become static and are divided into three groups. The phases of swarmalators remain bounded between $-2\pi$ to $2\pi$. However this range varies depending on the choice of initial conditions. It is worth mentioning that the number of such clusters and number of swarmalators inside each cluster depend on the initial spatial position and phase distributions. The emergence of these states is independent of the values of $\epsilon_a$ when it is varied within $0$ and $1$. Generally, a natural tendency among the coupled swarmalators is to reduce their respective phase differences beyond a critical value of attractive coupling strength $\epsilon_a > 0$. However, once the clusters are formed, they remain beyond their attractive vision radius for the observed static cluster sync state. As a result, they maintain their phase differences and are never able to merge into a single cluster, irrespective of $\epsilon_a \in (0,1]$. We include a video in the GitHub repository \cite{web_5} showing the formation of static cluster sync state from initial configurations as mentioned in Fig.\ \ref{Fig.4}.
\begin{figure}[hpt]
\centerline{
\includegraphics[scale = 0.5]{figure5.png}}
\caption{{\bf Emergent collective behavior based on the phase synchronization error in the $J$-$r$ parameter plane}: Simulation is done with $N = 100$ swarmalators for $t = 500$ time units with step size 0.01 and $\theta_{err}$ is calculated over last 10 time units. Each point is achieved using 20 realizations. We plot the analytically calculated critical curve $r = \frac{1}{1-J}$ (red) beyond which one can anticipate the occurrence of SS state. Note that this theoretically calculated curve provides a sufficient condition for the manifestation of the SS state. The black dashed lines are used to distinguish three observed fascinating states. Here $\epsilon_a = 0.5$.
\label{Fig.5}
\end{figure}
It is evident that whether swarmalators will end up inside a single disc (SS) or break into two or more clusters (SCS) is determined by the interplay between vision radius ($r$) and the parameter $J$. We try to approach this numerically by introducing the phase synchronization error defined by
\begin{equation}
\theta_{err} = \Bigg \langle \frac{\sum_{i =2}^{N} |\theta_i - \theta_1|}{(N-1)} \Bigg \rangle_t,
\label{eq.12}
\end{equation}
where $\langle \cdots \rangle_t$ stands for time average. For calculating $\theta_{err}$, we bring the phase values of swarmalators between $0$ and $2\pi$ so that swarmalators whose phase difference is an integer multiple of $2\pi$ represent the same phase. Under this scenario, our numerical simulations suggest the phase synchronization error $\theta_{err}$ takes minimum value $0$ and maximum value $\pi$. In Fig.\ \ref{Fig.5}, we plot the $J$-$r$ parameter space with the color bar indicating the phase synchronization error ($\theta_{err}$). When $r < 0.2$ (in the SA or STPW states), the value of $\theta_{err}$ is very high since the phases vary uniformly between $-\pi$ to $\pi$. For $r > 0.2$, phase synchrony starts to occur either among all units or in groups. In SS state, the synchronization error gives zero value while intermediate non-zero values of $\theta_{err}$ are observed in SCS state as synchrony is present only among the elements of each cluster but not globally.
\par Now we investigate the sufficient condition for global synchronization of the swarmalators. It is to be noted that global synchronization occurs when two or more clusters overcome their phase differences and coincide in a unique phase. For a fixed $J $ if we gradually increase $r$, cluster synchrony states are found after the initial asynchronous SA or STPW states. Two or more clusters establish connection for phase attraction among them and merge into a single cluster by minimizing their phase differences when $r$ is further increased. The static sync state is achieved when two separate clusters merge to form a single one where their phases are fully coherent.
Let $\mathcal{C}_i$ be the set of indices of the swarmalators belonging to the $i$-th cluster for $i \in \{1,2\}$. We define
\begin{equation}
\textbf{x}_{c_i} = \frac{1}{|\mathcal{C}_i|} \sum_{j \in \mathcal{C}_i} \textbf{x}_{j},
\label{eq.13}
\end{equation}
as the center of position of $i$-th cluster. When the static two-cluster state is formed, we can consider the swarmalators as a two-particle system where these two particles are positioned at the center of position of their respective clusters. Let their positions be denoted by $\textbf{x}_{c_1}$ and $\textbf{x}_{c_2}$ where $\textbf{x}_{c_1} \ne \textbf{x}_{c_2}$ and their phases be $\theta_{c_1}$ and $\theta_{c_2}$ ($\theta_{c_1} \ne \theta_{c_2}$), respectively. Since the system becomes stationary, following Eq.\ \eqref{eq.3}, we can write
\begin{equation}
{\bf 0} = \left[ \frac{\textbf{x}_{c_2} - \textbf{x}_{c_1}}{|\textbf{x}_{c_2} - \textbf{x}_{c_1}|} (1+J\cos(\theta_{c_2} - \theta_{c_1})) - \frac{\textbf{x}_{c_2} - \textbf{x}_{c_1}}{{|\textbf{x}_{c_2} - \textbf{x}_{c_1}|^2}} \right],
\label{eq.14}
\end{equation}
which gives
\begin{equation}
|\textbf{x}_{c_2} - \textbf{x}_{c_1}| = \frac{1}{1+J\cos(\theta_{c_2} - \theta_{c_1})} \le \frac{1}{1-J}.
\label{eq.15}
\end{equation}
So, the maximum spatial distance between the centers of the clusters is less than or equal to $\frac{1}{1-J}$. If $r$ is chosen to be greater than this value then the clusters synchronize their phases and as a result will merge to a single cluster to give the static synchrony state, i.e., $r>\frac{1}{1-J}$ is the sufficient condition for static sync state to take place which has been validated in Fig.\ \ref{Fig.5} by the red curve.
\subsection{\label{subsec:level3}Attractive-repulsive phase coupling}
We have modeled the swarmalators (Eqs.\ \eqref{eq.3} and \eqref{eq.4}) such that at short distance ($|\textbf{x}_{j}-\textbf{x}_{i}| \le r$) their phases are coupled positively and at long distance ($|\textbf{x}_{j}-\textbf{x}_{i}| > r$), they interact with negative phase coupling strength. It is easy to see, how swarmalators couple their phases depends strictly on the choice of vision radius $r$. We have already discussed the two cases (Sec.~\ref{subsec:level1}) for $r$, when it is very small and infinitely large. For small $r$, the repulsive coupling prevails over attractive coupling. When vision radius is increased, number of swarmalators inside the interaction range of each increases and attractive coupling between them takes place. For a large value of $r$, this attractive coupling completely dominates the repulsive coupling in the sense that more number of swarmalators are coupled attractively than repulsively. So, increment in the value of $r$ ensures the transition from repulsive dominance to attractive dominance in the phase coupling.
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 0.65]{figure6.jpg}}
\caption{{\bf Static $\pi$ state}: (a) Time series of $N = 100$ swarmalators are shown over $t =300$ time units after transient time for $r = 0.8$ and $J = 0.5$. (b) We display the corresponding snapshot of node index ($i$) vs phase ($\theta_i$) at $t = 300$ time unit. (c) Distance between the centers of positions of the clusters, $d_\pi$ as a function of $J$ is plotted here for $r = 0.75$. There exists a critical value of $J$ depending other system parameters beyond which one can expect such SPI state. Red dots represent simulated data, while the black curve shows theoretical result given by Eq.\ \eqref{eq.16}. Evolution of the static $\pi$ state for $r = 0.8, J = 0.5, \epsilon_a = 0.5, \epsilon_r = -0.5$ is shown by a video (Figure6.gif) in the GitHub repository \cite{web_5}.}
\label{Fig.6}
\end{figure*}
\paragraph*{Static $\pi$ state:} The co-existence of attractive and repulsive couplings in the system makes phase interaction between swarmalators complex and gives rise to richer collective behaviors. Like in the case of only attractive coupling, here also swarmalators group themselves into clusters depending on the values of $r$ and $J$. But the presence of repulsive coupling between the units of clusters forces them to maximize their phase difference and eventually the phase difference becomes $\pi$ for a stable solution. For static state, our numerical simulations assure the phase difference between the clusters is either 0 or $\pi$ (modulo $2\pi$). If clusters share a phase difference $0$ (modulo $2\pi$) then the units inside those clusters attract each other in space (note that, the spatial attraction strength is $1+J$ if phase difference $\theta_{ij} = 0$) and they form a single cluster. As a result, number of clusters is exactly two and the phase difference between them is $\pi$. We designate this new state as {\it static $\pi$ state} (SPI) since swarmalators become stationary in phase and space eventually. As the swarmalators inside each cluster are fully synchronized but not with the ones inside the other cluster, $R$ gives a non-zero value and is less than $1$. The order parameter $S$ also gives an intermediate non-zero value between $0$ and $1$ in this state, demonstrating the presence of correlation between swarmalators' spatial angles and phases. In Fig.\ \ref{Fig.6}(a), the temporal behavior of 100 swarmalators is shown, where the phases become static and are clustered maintaining exactly a phase-difference $\pi$. At a particular time instant after the transient time, their phases are plotted versus their respective node indices in Fig.\ \ref{Fig.6}(b), which validates the clustering of swarmalators in two groups. In Fig.\ \ref{Fig.6}(a)-(b), the phases of swarmalators are plotted after taking modulo $2\pi$ so that swarmalators at a phase difference of integer multiple of $2\pi$ represent the same phase. A video showing the origination of static $\pi$ state is included for visual understating as highlighted in Fig.\ \ref{Fig.6}.
In the SPI state, we can consider the swarmalators as a two-particle system where they are positioned at the center of position of their respective cluster having a phase difference $\pm \pi$. Let their positions be denoted by $\textbf{x}_{c_1}$ and $\textbf{x}_{c_2}$, where $\textbf{x}_{c_1} \ne \textbf{x}_{c_2}$. Then from Eq.\ \eqref{eq.14}, we can write
\begin{equation}
|\textbf{x}_{c_2} - \textbf{x}_{c_1}| = \frac{1}{(1-J)} = d_\pi,
\label{eq.16}
\end{equation}
since $\theta_{c_2} - \theta_{c_1} = \pm \pi$ and $\textbf{x}_{c_2} - \textbf{x}_{c_1} \ne {\bf 0}$.
Equation\ \eqref{eq.16} gives the distance between the centers of position of the clusters ($d_\pi$) in SPI state which increases when we increase the value of $J$. Figure \ref{Fig.6}(c) shows the change of $d_\pi$ with increment of $J$ in the interval where SPI state is achieved.
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 0.55]{figure7.pdf}}
\caption{{\bf Snapshots of mixed phase wave states}: Simulations are performed for $N = 100$ swarmalators over 500 time units with step size 0.01 and $J = 0.1$, and phases are changing over $0$ to $2\pi$. Attractive mixed phase wave is observed for (a) $\epsilon_a = 1.0$, $\epsilon_r = -0.1$, $r = 0.21$, (b) $\epsilon_a = 0.5$, $\epsilon_r = -0.5$, $r = 1.38$, and (c) $\epsilon_a = 0.1$, $\epsilon_r = -1.0$, $r = 1.83$. Swarmalators show the tendency of forming clusters with the ones in nearby phases, but can not achieve separate cluster formation for small values of $J$. Repulsive mixed phase wave emerges for (d) $\epsilon_a = 0.5$, $\epsilon_r = -0.5$, $r = 0.21$, and (e) $\epsilon_a = 0.1$, $\epsilon_r = -1.0$, $r = 0.36$. Repulsive coupling is dominant on attractive coupling for sufficiently small $r$ and when $\epsilon_a \ll |\epsilon_r|$. This state is like merging of separate clusters in splintered phase wave state. Evolution of these states corresponding to (a)-(e) are shown by five videos (Figure7a.gif-Figure7e.gif) in the GitHub repository \cite{web_5}.}
\label{Fig.7}
\end{figure*}
\paragraph*{Mixed phase wave state:} Significantly, for a small value of $J$, swarmalators accumulate themselves in space with the ones in nearby phase, but they do not form distinct clusters. The positive phase coupling induces minimization of phase difference between spatially nearby swarmalators, and at the same time, negative coupling ensures that there is some phase difference between swarmalators who lie outside the vision range. They form a deformed disc like structure, where swarmalators move inside it maintaining phase similarity with nearby units. This active state is different from the two previously mentioned active states (SPPW and APW states), as swarmalators neither break into disjoint clusters nor execute full cycle in space and phase. In this state, the swarmalators are neither completely phase synchronized nor fully desynchronized, rather they show an intermediate behavior. We name this state as {\it mixed phase wave} (MPW) state. In Fig.\ \ref{Fig.7}, snapshots of this state are shown for different values of parameters. Like in the previous cases, again we plot the swarmalators after bringing their phases in the interval $[0, 2\pi]$. Corresponding to Fig.\ \ref{Fig.7}, five videos showing the evolution of mixed phase wave states are incorporated in the GitHub repository \cite{web_5}. This kind of behavior can be seen when swarmalators are on the verge of forming splintered phase wave or static $\pi$ but they can not attract other swarmalators in nearby phases with enough strength to make disjoint clusters. We classify this mixed phase wave state into two categories. One is repulsive mixed phase wave (RMPW) state (where repulsive coupling dominates and qualitatively similar to SPPW) and the other is attractive mixed phase wave (AMPW) state (which is dominated by attractive coupling and qualitatively similar to SPI). In both the AMPW and RMPW states, $S$ is non-zero indicating the presence of some correlation among swarmalators' phases and spatial angles. In the attractive coupling influenced AMPW state, the order parameter $R$ gives a non-zero value showing some coherence among swarmalators' phases. The value of $R$ is zero in the repulsive coupling dominated RMPW state. So by using the value of $R$, we separate these two states from each other.
\begin{figure*}[hpt]
\centerline{
\includegraphics[width = 16cm,height = 12cm]{figure8.pdf}}
\caption{{\bf Transition of states with varying $r$ for different choices of $J$, $\epsilon_a$, and $\epsilon_r$}: Behavior of order parameters $R$ (red), $S$ (black), $T$ (blue), and $V$ (yellow) with varying $r$ are shown. $\epsilon_a = 1.0$ and $\epsilon_r = -0.1$ in (a)-(c), $\epsilon_a = 0.5$, and $\epsilon_r = -0.5$ in (d)-(f), and $\epsilon_a = 0.1$ and $\epsilon_r = -1.0$ in (g)-(i). $J = 0.1$ in (a), (d), (g), $J = 0.5$ in (b), (e), (h), and $J = 0.9$ in (c), (f), (i). Simulations are done for $N = 100$ swarmalators over 500 time units with integration step size 0.01. Last 10$\%$ of the data are chosen for calculating $R$ and $S$, whereas last 50$\%$ data are used to calculate $T$. $V$ is calculated by taking time average over last 50 time units. For each point, a mean of 20 realizations is taken. $R$, $S$, and $T$ are plotted along left y axis and $V$ is plotted along the right y axis.}
\label{Fig.8}
\end{figure*}
\par To separate the stationary states from the non-stationary ones, we measure the mean velocity $V$ defined as,
\begin{equation}
V = \Bigg \langle\frac{1}{N} \sum_{i=1}^{N} \sqrt{\dot{x}_i^2 + \dot{y}_i^2 + \dot{\theta}_i^2} \Bigg \rangle_t,
\label{eq.17}
\end{equation}
where the time average is taken after discarding the transients. A finite non-zero value of the mean velocity $V$ suggests that swarmalators move in space, and their phases evolve within the interval $[0,2\pi)$, while in stationary states $V$ is zero. Depending on the values of $r$, $J$, $\epsilon_a$, and $\epsilon_r$, the swarmalators exhibit different long-term states. In Fig.\ \ref{Fig.8}, we vary $r$ over the range $(0,3]$ keeping the other three parameters fixed and examine the transition of states. {We consider three different scenarios for three different values of $J = 0.1, 0.5$, and $0.9$. Depending on the values of $\epsilon_a$ and $\epsilon_r$, three different cases can be implemented:}
\paragraph*{Case-1:} $\epsilon_a > |\epsilon_r|$. This condition implies that the attractive coupling strength between the swarmalators within their vision range is larger than the repulsive coupling strength with outer swarmalators.
\paragraph*{Case-2:} $\epsilon_a = |\epsilon_r|$. The swarmalators interact with same strength with others, both inside and outside their vision range but with opposite sign.
\paragraph*{Case-3:} $\epsilon_a < |\epsilon_r|$. Here, the attractive coupling strength between swarmalators within their vision range is smaller than the repulsive coupling strength for interaction outside it.
\begin{table*}[ht]
\caption{\label{tab:table3}This table shows how the emerging states of swarmalator system are identified with the order parameters $R$, $S$, $T$, and $V$. }
\centerline{
\begin{tabular}{ccccc}
$R$&$S$&$T$&$V$&Emerging state\\ \hline
$\approx$ 1&$0 < S < R$&$\approx 0$ &$\approx 0$& Static sync (SS) \\
$\approx 0$&$\approx 0$&$\approx 0$&$\approx 0$&Static async (SA)\\
$\approx 0$&$\approx 1$&$\approx 0$&$\approx 0$&Static phase wave (STPW)$^{a}$
\\
$R <1 (\ne 0)$&$S \ne 0$&$\approx 0$&$\approx 0$&Static $\pi$ (SPI)\\
$\approx 0$&$\ne 0$ &$\approx 0$&$ \ne 0$ &Splintered phase wave (SPPW)\\
& & & &or Repulsive mixed phase wave (RMPW)$^b$
\\
$\approx 0$&$\ne 0$ &$\ne 0$&$ \ne 0$ &Active phase wave (APW)\\
$\ne 0$&$\ne 0$ &$\approx 0$&$ \ne 0$ &Attractive mixed phase wave (AMPW)\\
\end{tabular}}
$^{a}${In Fig.\ \ref{Fig.8}(d) for STPW state $R$ and $S$ are slightly deviated form $0$ and $1$, respectively due to the deformed structure.
$^{b}${RMPW state is qualitatively similar to the SPPW state when they are studied with these four order parameters. RMPW state takes place when $J$ is not large enough for formation of clusters. If we increase the value of $J$ keeping the other parameters fixed, evolution of SPPW state is seen from the RMPW state}.}
\end{table*}
\par We use different order parameters to classify various collective patterns and briefly encapsulate all possible emergent states in Table \ref{tab:table3}. For the top row of Fig.\ \ref{Fig.8}, attractive phase coupling strength $\epsilon_a$ is chosen larger in modulus than the repulsive coupling strength $\epsilon_r$. In the middle row, these two are same in modulus, while $\epsilon_a < |\epsilon_r|$ for the bottom row. The value of $J$ is same in every column and $J = 0.1$, $0.5$, and $0.9$ for column 1, column 2, and column 3, respectively. In Fig.\ \ref{Fig.8}(a), $\epsilon_a = 1.0$ is bigger than the modulus of $\epsilon_r = -0.1$. Very small $r$ ensures phase coupling is repulsive and as a result SA is seen. Since $\epsilon_a$ is very larger than $|\epsilon_r|$, synchrony is achieved quiet fast while varying $r$. Before achieving complete synchrony in SS state, over a small range of values of $r$, clustering tendency is seen, but for small value of $J = 0.1$, the swarmalators are unable to break into distinct clusters and as a result AMPW state is found (see Fig.\ \ref{Fig.7}(a)). The possible route of SS state by changing the vision radius $r$ is
\begin{equation}
\text{SA} \rightarrow \text{AMPW} \rightarrow \text{SS}.
\end{equation}
When $J$ is increased to 0.5 in Fig.\ \ref{Fig.8}(b), emergence of SPI state is seen due to the tendency of swarmalators to group themselves. With increasing $r$, the transition of states can be seen as
\begin{equation}
\text{SPPW} \rightarrow \text{SPI} \rightarrow \text{SS}.
\end{equation}
Qualitative same behaviors of the order parameters are seen in Fig.\ \ref{Fig.8}(c) where $J$ is further increased to 0.9.
\par The absolute values of $\epsilon_a$ and $\epsilon_r$ are same in Fig.\ \ref{Fig.8}(d)-\ref{Fig.8}(f). When $J = 0.1$, the swarmalators stay nearby with swarmalators in similar phases but fail to make disjoint clusters. In Fig.\ \ref{Fig.8}(d), when $r$ is very small, $\epsilon_r$ dominates the phase coupling and swarmalators are found to settle into SA state. When $r$ is increased, effect of $\epsilon_a$ starts to take place but still repulsive coupling continue to dominate. Emergence of RMPW (see Fig.\ \ref{Fig.7}(d)) is seen for a small range of $r$. This state is analogous to the SPPW state, the only difference, $J$ is not large enough for the swarmalators to splinter into disjoint clusters. Further, increasing the value of $r$, a situation is achieved, where the effect of $\epsilon_a$ and $\epsilon_r$ neutralizes each other, and the swarmalators settle in the STPW state. In some cases, this STPW state is formed in a slightly deformed annular structure and, as a result, the correlation between phases and spatial angles of the swarmalators deteriorates little. This is the reason why $S$ is not exactly 1 here, otherwise $S=1$ in the STPW state. Moving to the right with increasing $r$, phase attraction starts to dominate the phase repulsion. AMPW state is noticed which is close to the SPI state (found for bigger $J$ values). Here, $J = 0.1$ being small, the swarmalators can not break into separate clusters, but they coexist (see Fig.\ \ref{Fig.7}(b)). Finally, for reasonable larger $r$ values where the phase attraction dominates, SS state emerges. So, the transition which is seen in Fig.\ \ref{Fig.8}(d) is
\begin{equation}
\text{SA} \rightarrow \text{RMPW} \rightarrow \text{STPW} \rightarrow \text{AMPW} \rightarrow \text{SS}.
\end{equation}
When $J$ is increased to 0.5, formation of clusters happens which was not seen in the case of $J = 0.1$. For very small $r$, the swarmalators are effectively phase coupled only with $\epsilon_r$ and as a result, APW is found which is shown in Fig.\ \ref{Fig.8}(e). Here, due to the ability to break into clusters, SPPW state is seen in place of RMPW in Fig.\ \ref{Fig.8}(d). Increment in the value of $r$ increases phase attraction among the swarmalators and SPI state is found to occur before the ultimate SS state. The route to achieve SS state in Fig.\ \ref{Fig.8}(e) is
\begin{equation}
\text{APW} \rightarrow \text{SPPW} \rightarrow \text{SPI} \rightarrow \text{SS}.
\end{equation}
For $J = 0.9$ in Fig.\ \ref{Fig.8}(f), the transitions are same as Fig.\ \ref{Fig.8}(e).
\par In the bottom row of Fig.\ \ref{Fig.8} $\epsilon_a = 0.1$ is relatively small than $|\epsilon_r| = 1.0$. For $J = 0.1$ in Fig.\ \ref{Fig.8}(g), the evolution of states takes place in the same manner as in Fig.\ \ref{Fig.8}(d). In Fig.\ \ref{Fig.8}(h), the route for SS state is
\begin{equation}
\text{SA} \rightarrow \text{APW} \rightarrow \text{SPPW} \rightarrow \text{SPI} \rightarrow SS,
\end{equation}
where $J = 0.5$. With $J = 0.9$ in Fig.\ \ref{Fig.8}(i), we can see that the SPI state exists over a long range of $r$ and SS is yet not achieved.
\subsection{\label{sec:level8} Effect of negative $J$}
The parameter $J$ which stands for the effect of phase similarity on spatial attraction, plays a significant role in determining the asymptotic state of the swarmalator system defined by Eqs.\ \eqref{eq.3} and \eqref{eq.4}. So far the results in our work are accomplished by considering $J$ to be strictly positive. A positive value of $J$ makes sure that the spatial attraction strength between nearby phased swarmalators is increased and as a result they stay nearby to each other. A negative value of the parameter $J$ completely alters the scenario. Now swarmalators which are in opposite phases attract each other spatially with more strength than those which are in similar phases. This reverse phenomena changes the complexion of the swarmalator system altogether. At the same time it increases the complexity of our model. To capture the effect when a negative $J$ is considered, we study the case of local attractive coupling with $J<0$. Also to show the generic nature of our model for the exponents $\alpha$, $\beta$, and $\gamma$ (as long as $\alpha < \beta$ holds), here we consider $\alpha = 0$ (note that, all the studies so far has been done with $\alpha=1$). Now, with these assumptions we write down the governing equations as
\begin{equation}
\label{eq.39}
\dot{\textbf{x}}_{i} = \frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N}\left[ \textbf{x}_{ij} (1+J\cos(\theta_{ij})) - \frac{\textbf{x}_{ij}}{{|\textbf{x}_{ij}|}^2} \right],
\end{equation}
\begin{equation}
\label{eq.40}
\dot{\theta}_{i} = \frac{\epsilon_a}{N_{i}}\sum_{\substack{j = 1\\j \neq i}}^{N} \text{A}_{ij}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|},
\end{equation}
where $\text{A}$ is the adjacency matrix for local attractive phase coupling described in Sec.\ \ref{sec:level2}. This model inherits contrasting spatial and phase interaction. The local attractive phase coupling among the swarmalators minimizes the phase difference among spatially nearby ones. On the other hand, swarmalators which are in nearby phases reduce the spatial attraction among them (since $J$ is negative).
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 0.48]{figure11.png}}
\caption{{\bf Emerging asymptotic states with local attractive phase coupling when $J<0$}: The emergence of static async (SA), static inscribed cluster (SIC), and static sync (SS) is seen by considering negative $J$ under local attractive phase coupling. (a) The $J$-$r$ parameter plane is plotted based on the values of phase synchronization error $\theta_{err}$. It is maximum in the SA state and minimum in the SS state, whereas it gives intermediate values in the SIC state. (b) Snapshot of the SIC state at $t=500$ time unit for $J=-0.8$ and $r=0.4$. Swarmalators are colored according to their phases and it is seen that phases are divided into two clusters. Swarmalators whose phases are synchronized, all lie either within an annulus or a disc where the disc is inscribed within the annulus. (c) The evolution of order parameter $R$ is shown varying the value of $r$ from $0$ to $1$ for a fixed $J$ ($=-0.8$). The transition from SA state to SS state via the SIC state is seen with increasing $r$. For $R \approx 0$, the SA state is perceived. While for $R \approx 1$, the system settles down to the SS state. For any intermediate non-zero values of $R$ and $\theta_{err}$, we confirm the emergence of the SIC state by plotting the snapshot at a specific time after the initial transient.}
\label{Fig.11}
\end{figure*}
When the vision radius $r$ is small ($r<0.2$), swarmalators can not feel the presence of one another inside their vision range and their phases remain decoupled. But negative value of $J$ induces disorder in the system where swarmalators' phases are totally desynchronized and they become stationary both in phase and space. In this static async (SA) state, the value of order parameter $R$ is close to zero. The phase synchronization error $\theta_{err}$ defined by eq.\ \eqref{eq.12} is very large (see Fig.\ \ref{Fig.11}(a)). Keeping the value of $J$ fixed when we increase the value of $r$, swarmalators start to couple their phases with spatially nearby ones and minimize their phase difference. We observe the emergence of a new pattern where swarmalators' phases are settle in two different clusters. Contrary to the static cluster state found with a positive value of $J$ for local attractive coupling in subsection~\ref{subsec:level2}, here swarmalators do not form separate clusters in the two-dimensional plane. Rather they settle themselves inside an annulus and a disc where the disc remains inscribed within the annulus. Inside both of these annulus and disc swarmalators are totally phase synchronized but they maintain a phase difference between them. Swarmalators become static after transient period both in space and phase. Following the nature of this state, we name it as the {\it static inscribed cluster} (SIC) state. In Fig.\ \ref{Fig.11}(b) we have shown the snapshot of this state in the two-dimensional plane taking $J=-0.8$ and $r = 0.4$. The order parameter $R$ gives intermediate values between $0$ and $1$ due to the presence of phase synchrony among the units of each cluster. The value of $\theta_{err}$ also decreases for the same reason compared to the SA state (see Fig.\ \ref{Fig.11}(a)). This SIC state loses its stability when we further increase the value of $r$. A large value of $r$ ensures the dominance of attractive phase coupling among all the swarmalators. As a result globally phase synchronized static sync (SS) state emerges. $R$ gives the maximum value whereas $\theta_{err}$ acquires the minimum value in this stationary state.
In Fig.\ \ref{Fig.11}(a) we have plotted the $J$-$r$ parameter space based on the values of phase synchronization error $\theta_{err}$. It is seen that $\theta_{err}$ is very large in the SA state and is almost zero in the SS state. In the SIC state it gives intermediate values between $0$ and $\pi$. To observe the transition of these states, we have plotted the order parameter $R$ against increasing $r$ in Fig.\ \ref{Fig.11}(c) where a particular value of $J$ ($=-0.8$) has been chosen. It is seen that for small $r$ ($<0.2$) in the SA state the value of $R$ is near $0$. When $r$ is increased beyond $0.2$, phase coherence among units of each clusters starts to take place in the SIC state and as a result the value of $R$ increases. Another critical value of $r$ ($\approx 0.58$) is found where this SIC state loses its stability and emergence of SS state is seen beyond that.
\subsection{\label{sec:level7}Effect of phase dynamics on spatial repulsion}
\par Till now, we have considered $\text{F}_{\text{rep}} = 1$, i.e., when there is no effect of phase similarity on spatial repulsion. Here, we investigate the scenario when swarmalators' phases affect the spatial repulsion among them. We specifically choose $\text{F}_{\text{rep}}(\theta) = 1 - K\cos(\theta)$ so that the spatial repulsion among nearby phase swarmalators is reduced when $K>0$. We consider $0<K<1$ which keeps the function $\text{F}_{\text{rep}}$ strictly positive (If $K>1$, then depending on the value of $\theta_j - \theta_i$ the function $\text{F}_{\text{rep}}$ can take negative value, which makes $\text{I}_\text{rep}$ attractive in nature). It is worth mentioning here that Theorem \eqref{Theo:1} holds for any choices of the function $\text{F}_{\text{att}}$ and $\text{F}_{\text{rep}}$ as long as they are even and bounded. In this case, $\text{F}_{\text{att}}(\theta) = 1 + J\cos(\theta)$ and $\text{F}_{\text{rep}}(\theta) = 1 - K\cos(\theta)$ are both even functions of their arguments and are also bounded. So we can safely say that Theorem \eqref{Theo:1} holds with these choices of functions which eliminates the inter-particle collision among the swarmalators. This extra interaction term in the swarmalator model can establish novel states for position aggregation and phase synchronization especially in the presence of competitive phase interaction, which requires a sincere inspection of the model. For our purpose we only consider such effect of phase similarity on spatial repulsion when the vision radius is very small. In that case our model with the choice of $\text{F}_{\text{rep}}$ can be written as
\begin{equation}
\label{eq.26}
\begin{split}
\dot{\textbf{x}}_{i} = \frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N} \bigg[ \frac{\textbf{x}_{j}-\textbf{x}_{i}}{|\textbf{x}_{j}-\textbf{x}_{i}|} (1+J\cos(\theta_{j}-\theta_{i})) \\ -\frac{\textbf{x}_{j}-\textbf{x}_{i}}{{|\textbf{x}_{j}-\textbf{x}_{i}|}^2}(1-K\cos(\theta_{j}-\theta_{i})) \bigg],
\end{split}
\end{equation}
\begin{equation}
\label{eq.27}
\dot{\theta}_{i} = \frac{\epsilon_r}{N-1}\sum_{\substack{j = 1\\j \neq i}}^{N} \frac{\sin(\theta_{j}-\theta_{i})}{|\textbf{x}_{j}-\textbf{x}_{i}|}.
\end{equation}
Note that, here we have expressed $\textbf{x}_{ij}$ and $\theta_{ij}$ as $\textbf{x}_{j}-\textbf{x}_{i}$ and $\theta_{j}-\theta_{i}$, respectively, which is simpler to deal with while carrying out mathematical calculation. It is found that, for certain values of $J$, $K$, and $\epsilon_r$ the swarmalators organize themselves on a ring where their phases ($\theta_i$) are perfectly correlated with their spatial angles ($\phi_i$). This state is a stationary state, where after the transient period, swarmalators become static in phase as well as in the spatial positions, and we call this state as the static ring phase wave state. This static state was reported in \cite{o2018ring} for $\alpha=0$ and $\gamma=2$. In Fig.\ \ref{Fig.10}(a) swarmalators position on a ring centered around the origin is shown where they are colored according to their phases. In the static ring phase wave state, the spatial angle and phase of the swarmalators are perfectly correlated. Figure \ref{Fig.10}(b) highlights the correlation between swarmalators' spatial angles and phases. The position and phase of the $k$-th swarmalator in this state can be written as
\begin{equation}
\textbf{x}_k = R\cos(2\pi k / N) \hat{i} + R\sin(2\pi k/N) \hat{j},
\label{eq.28}
\end{equation}
\begin{equation}
\theta_k = 2\pi k/N + C,
\label{eq.29}
\end{equation}
where $R$ is the radius of the ring state, $\hat{i}$ and $\hat{j}$ are unit vectors in $x$ and $y$ directions, respectively, and the constant $C$ depends on the initial conditions. The radius $R$ can be calculated analytically. The structure of static ring phase wave state leads us to use convenient complex notation where the two-dimensional vector $\textbf{x}_{k} = (x_k,y_k)$ is identified as a point in the complex plane $z_k = (x_k,y_k)$. To calculate the radius $R$, we consider a general model of swarmalators \cite{o2018ring}
\begin{equation}
\begin{split}
\dot{z_i} = \frac{1}{N-1}\sum_{j=1}^{N} \big[ f(|z_i-z_j|^2)(z_i-z_j) \\+h(|z_i-z_j|^2)(z_i-z_j)\cos(\theta_i-\theta_j)\big],
\label{eq.30}
\end{split}
\end{equation}
\begin{equation}
\dot{\theta_i} = \frac{1}{N-1}\sum_{j=1}^{N} \sin(\theta_i-\theta_j) g(|z_i-z_j|^2).
\label{eq.31}
\end{equation}
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 0.45]{figure10.pdf}}
\caption{{\bf Static ring phase wave state}: We do simulations with $N=100$ swarmalators for $J=0.5$, $K=1.0$, and $\epsilon_r=-0.1$. (a) Position of swarmalators on a ring centered around the origin at $t=150$ time unit. (b) We plot the phases ($\theta$) of the swarmalators against their spatial angles ($\phi$) where perfect correlation among them is found. (c) The radius ($R$) of the ring phase wave state is calculated numerically for several values $J$ while keeping $K$ fixed at $1.0$. Red dots are the numerically calculated values of $r$. The black curve represents analytical expression given by Eq.\ \eqref{eq.38} for $K=1.0$. It is seen that, our numerical and analytical results match satisfactorily well.}
\label{Fig.10}
\end{figure*}
The swarmalator model defined by Eqs.\ \eqref{eq.26} and \eqref{eq.27} corresponds to Eqs.\ \eqref{eq.30} and \eqref{eq.31} when we specifically choose
\begin{equation}
f(r) = \frac{1}{r} - \frac{1}{\sqrt{r}}; \hspace{10pt} h(r) = -\frac{K}{r} - \frac{J}{\sqrt{r}}; \hspace{10pt} g(r) = -\frac{\epsilon_r}{\sqrt{r}}.
\label{eq.32}
\end{equation}
In complex plane the ring phase wave state is given by
\begin{equation}
\begin{split}
z_k = R u^k, \hspace{10pt} \text{where} \hspace{10pt} u= exp(2\pi l / N), \hspace{10pt} \\ \theta_k = 2 \pi k /N + C \hspace{10pt} (l = \sqrt{-1}).
\end{split}
\end{equation}
Although Eq.\ \eqref{eq.31} is satisfied by any values of $R$, Eq.\ \eqref{eq.30} is satisfied only if
\begin{equation}
\begin{split}
\sum_{i \ne 0} \big[ f(R^2|1-u^i|^2)(1-u^i) \\ + h(R^2|1-u^i|^2)(1-u^i) \cos(2\pi i/N) \big]= 0 .
\label{eq.33}
\end{split}
\end{equation}
For the choices of the functions $f$, $h$, and $r$ given by Eq.\ \eqref{eq.32}, we get an expression for $R$ from Eq.\ \eqref{eq.33} using the identities
\begin{align}
\sum_{i \ne 0} \frac{1}{1-u^i} &= \frac{N-1}{2},
\end{align}
\begin{equation}
\sum_{i \ne 0} \frac{u^i+u^{-i}}{1-u^{-i}} = -1,
\end{equation}
\begin{equation}
\sum_{i \ne 0} \frac{1-u^i}{|1-u^i|} = \frac{\sin(\pi/N)}{1-\cos(\pi/N)} = a,
\end{equation}
\begin{equation}
\sum_{i \ne 0} \frac{(u^i+u^{-i})(1-u^i)}{|1-u^i|} = \frac{\sin(3\pi/N)}{1-\cos(3\pi/N)} - \frac{\sin(\pi/N)}{1-\cos(\pi/N)} = b.
\end{equation}
This gives the expression for $R$ in the form
\begin{equation}
R = \frac{N-1+K}{2a + Jb}.
\label{eq.38}
\end{equation}
In Fig.\ \ref{Fig.10}(c) we plot the values of $R$ varying $J$ with $K=1.0$ (black curve). We also numerically calculate the values of $R$ for some values of $J$ (red dotted points) to show that numerical values replicate the analytical ones.
\section{\label{sec:level5}Swarmalators with the phase dynamics of the Stuart-Landau oscillator}
Having studied various emerging states of the swarmalators, we now want to find out what happens when a different kind of phase dynamics is considered in the swarmalator model in place of the vastly used Kuramoto-like dynamics. For each swarmalator moving in the two-dimensional plane, we associate a Stuart-Landau (SL) oscillator with it. The SL oscillator is an amplitude oscillator with two state variables and the dynamics is represented by
\begin{equation}
\dot{\textbf{u}}_i = \begin{pmatrix} [1-(u_{i}^{2}+v_{i}^{2})]u_{i} - \omega_{i}v_{i}\\
[1-(u_{i}^{2}+v_{i}^{2})]v_{i} + \omega_{i}u_{i} \end{pmatrix}\\ = F(\textbf{u}_i),
\label{eq.23}
\end{equation}
where $\textbf{u}_i = (u_i,v_i)$ is the two-dimensional state variable and $\omega_i$ is the intrinsic frequency of the $i$-th swarmalator for $i = 1, 2, \ldots,N$. We consider symmetry preserving diffusive type coupling $K(\textbf{u}_i,\textbf{u}_j) = (u_j-u_i,v_j-v_i)^T$ for the interaction among the SL oscillators. As before, the spatial coupling is taken to be all-to-all, whereas we study only the local attractive coupling for the phase interaction between the swarmalators. Here we adopt the same coupling scheme as described in subsection~\ref{subsec:level2}. The swarmalator model with Stuart-Landau oscillator is governed by the pair of equations
\begin{equation}
\label{eq.24}
\dot{\textbf{x}}_{i} = \frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N}\left[ \frac{\textbf{x}_{j}-\textbf{x}_{i}}{|\textbf{x}_{j}-\textbf{x}_{i}|} (1+J\cos(\theta_{j}-\theta_i)) - \frac{\textbf{x}_{j}-\textbf{x}_{i}}{{|\textbf{x}_{j}-\textbf{x}_{i}|}^2} \right],
\end{equation}
\begin{equation}
\label{eq.25}
\dot{\textbf{u}}_i = \begin{cases}
F(\textbf{u}_i) + \dfrac{\epsilon_a}{N_i} \sum_{\substack{j = 1\\j \neq i}}^{N} \text{A}_{ij} \frac{K(\textbf{u}_i,\textbf{u}_j)}{|\textbf{x}_{j}-\textbf{x}_{i}|} &\mbox{if } N_i \ne 0 \\ F(\textbf{u}_i) &\mbox{if } N_i = 0 \end{cases}
\end{equation}
where $J, \epsilon_a, N_i,$ and $\text{A}_{ij}$ are same as described in Section~\ref{sec:level2}. In Eq.\ \eqref{eq.24}, by the phase $\theta_i$ of the $i$-th SL oscillator, we mean the principal value of argument of the complex number $Z_i = u_i + l v_i$ ($l = \sqrt{-1}$), where $u_i,v_i$ are the state variables of the corresponding oscillator. Evidently, $\theta_i \in [-\pi, \pi)$, which fulfills the requirement of the phase variable of the swarmalators. We choose identical oscillators with $\omega_i = \omega = 3.0$. The initial conditions for the state variables are chosen randomly from $[-1,1]$.
\begin{figure*}[hpt]
\centerline{
\includegraphics[scale = 1.1]{figure9.pdf}}
\caption{{\bf Swarmalators with the phase dynamics of SL oscillator}: Position in the $(u,v)$ plane of $N=100$ swarmalators at $t=300$ time unit in (a) static phase wave state for $r = 10^{-5}$, (b) static cluster state for $r = 0.3$, and (c) static sync state for $r = 1.0$, where $J = 0.8$. Color bar indicates the phase of the SL oscillators. The $\phi$-$\theta$ plots for the same values of $r$ and $J$ at the same time instant are shown in (d), (e), and (f), respectively. $\epsilon_a = 0.5$.}
\label{Fig.9}
\end{figure*}
\par It is seen that all the states (namely, STPW, SCS, and SS), which have been found with Kuramoto oscillator with local attractive phase coupling, take place for SL oscillator as well. Attractive phase coupling invokes coherence among swarmalators' phases, and swarmalators in nearby phases attract themselves spatially with more strength, caused by the influence of phase similarity on spatial attraction. Due to the choice of non-zero intrinsic frequency, swarmalators' phases continue to evolve even when synchrony is achieved. Their positions in the $(x,y)$ plane become static as before since the difference of phase between any two swarmalators remains constant over time after transient period, and as a result, change of phase does not affect the spatial position. The Stuart-Landau oscillators execute limit cycle oscillation on a unit circle in the $(u,v)$ plane. In Figs.\ \ref{Fig.9}(a)-(c), positions of the oscillators on the limit cycle are shown for different emergent states, where they are colored according to their phase $\theta_i$. The relations between spatial angle $\phi$ and phase $\theta$ are shown in panel \ref{Fig.9}(d)-(f). For a very small choice of attractive vision radius $r$, phase coupling is effectively absent in the system and static phase wave (STPW) takes place. In the STPW state, the swarmalators' phases are distributed over $[-\pi,\pi)$, where their spatial angle $\phi$ and phase $\theta$ are perfectly correlated, which can be seen from the linear relation between $\theta$ and $\phi$ in Fig.\ \ref{Fig.9}(d). In the static cluster synchrony (SCS) state, the swarmalators form clusters both in $(x,y)$ (for position) and $(u,v)$ (for phase) plane. We show the formation of one such SCS state where the number of clusters is two (in Figs.\ \ref{Fig.9}(b) and \ref{Fig.9}(e)). The two horizontal bands in Fig.\ \ref{Fig.9}(e) clearly indicate that for the two clusters there are two different phases, whereas, their spatial angles are distributed over $[-\pi,\pi)$. The number of such clusters in the SCS state depends on the initial choices of spatial position and phase distribution of the swarmalators. The two clusters merge and form a single group when $r$ is increased and static sync (SS) state is achieved. It can be seen in Figs.\ \ref{Fig.9}(c) and \ref{Fig.9}(f) that all the oscillator' phases are same and as a result, the phase influenced spatial attraction among them increases to form a single cluster in space. Here, by using the SL oscillator for the phase dynamics of swarmalators, we validate the results in local attractive phase coupling scenario which we have found for swarmalators with the Kuramoto phase dynamics.
\section{Discussion and Conclusion} \label{sec:level6}
\par Swarmalators have been studied from the perspective of multi-agent systems for their position aggregation and phase dynamics. The bidirectional influence of agents' spatial position and internal phase on each other gives rise to fascinating collective patterns. Some of these stationary and non-stationary patterns are significantly found while studying the behaviors of real-world multi-agent systems. It is worth mentioning that in Ref.~\cite{lee2021collective} the collective behaviors of swarmalators have been studied with a finite-cutoff interaction distance, which is similar to the vision radius in our model. In this finite-cutoff interaction, the distance controlled spatial coupling among the swarmalators while keeping the phase coupling globally attractive or repulsive. Local spatial coupling among the swarmalators resulted in the formation of groups with spatially nearby ones. They found the emergence of multiple static sync discs (for globally attractive phase coupling) or multiple static async discs (for globally repulsive phase coupling).
\par In contrast, we have taken our spatial coupling to be all-to-all and phase coupling to be competitive and dependent on the vision radius. We have encountered cluster states in our model even when the spatial coupling is global. Swarmalators with same phase attract themselves and stay nearby in space. As a result, the cluster formation is influenced by the effect of phase similarity on spatial attraction rather than the direct local spatial coupling in Ref.~\cite{lee2021collective}. In another work \cite{hong2021coupling}, the coupling strength between the swarmalators was chosen randomly from a two-peaks distribution where the peaks correspond to positive and negative coupling strengths. The phase transition was studied while varying the probability of attractive phase coupling. The randomness in the phase coupling strength allowed the swarmalators to get repulsive or attractive coupling, irrespective of their spatial distance. But, in our model, the coupling strategy is deterministic in the sense that swarmalators' phases are coupled attractively if they lie inside a specific vision range, and otherwise, the coupling is repulsive. While both these works represent competitiveness in the phase interaction, the underlying strategies and their emergent states are different.
\par In this article, we have modeled swarmalators in such a way that their phases are attractively coupled only when they lie inside each other's attractive vision range, otherwise their phases are coupled repulsively. It has been assured that the swarmalators avoid collisions in finite time and they maintain a minimal distance among them in space which is uniform in time. Our model comprises four parameters $r, J, \epsilon_a$, and $\epsilon_r$, which control the long-term behavior of the swarmalator system. Considering attractive vision radius $r$ to be very large and very small, phase coupling has been seen to become globally attractive and globally repulsive, respectively. A new static state (static cluster synchrony) where swarmalators form disjoint clusters is found, when $\epsilon_r$ is taken to be zero. We have analytically found a sufficient condition, which, when satisfied, leads to the static sync state. In the presence of attractive-repulsive phase coupling in the system, competitive phase interaction gives rise to several asymptotic states. Varying $r$, we have studied the transition of these states for different cases of $J, \epsilon_a$, and $\epsilon_r$ in Fig.\ \ref{Fig.8}. The two-clustered static $\pi$ state is a novel state for swarmalators in the two dimensional plane. This state is different from the static $\pi$ state found in Ref.~\cite{o2021collective} where swarmalators were positioned on a 1D ring. In the mixed phase wave states, the effect of competitive phase interaction is prominent where attractive phase coupling minimizes the phase difference between swarmalators at short distances and repulsive phase coupling increases the phase difference between the swarmalators who are outside each other's vision range. Four order parameters $R, S, T$, and $V$ unfold the qualitative behavior of the emergent dynamical states. To study swarmalators with a different oscillator, we have used the phase dynamics of the Stuart-Landau oscillator in our system and studied their behavior in the presence of local attractive phase coupling.
\par
Our model has introduced a particular coupling strategy to incorporate competitive phase interaction. It remains to be seen what happens when this competitive interaction takes place following a different rule. In fact, a different set of values for the plausible choices of $\alpha$, $\beta$, and $\gamma$ make a significant variation in the observed dynamics of swarmalators, leading an important direction for future exploration. A step to study swarmalators with the Stuart-Landau oscillator in place of the conventional Kuramoto phase dynamics is considered in our work. Another scope of the study is to consider some other phase dynamics while modeling the swarmalator systems, resulting in different long-term behaviors. A possible step towards this direction is considering chaotic and hyperchaotic intrinsic dynamics \cite{sayeed2020behavioral,rossler1976equation,nag2020hidden,lorenz1963deterministic} with the positive Lyapunov exponent(s). This interesting inclusion may reveal further systematic insights; however, drawing a conclusion with chaotic dynamics requires more detailed rigorous analysis. The repulsive coupling may lead to an unbounded solution blowing out the dynamics away from the invariant manifold. It would be fascinating to investigate this future generalization expecting a broad spectrum of various novel dynamical states.
\par Moreover, we have considered the effect of phase similarity only on spatial attraction while studying the cumulative impact of attractive-repulsive interaction. It would be interesting if the effect of phase similarity on spatial repulsion is also considered under such competitive interactions. Although we scrutinize the whole investigation with only swarmalators moving in two dimensions, the study on $3$D provides the same equations in (\ref{eq.3}-\ref{eq.4}), except the exponent of the repulsion will now be $\beta=3$ for a physically meaningful and analytically tractable model. We expect comparable results just like in $2$D. The static sync and async states will become spheres. We present a brief section devoted to understanding the influence of three-dimensional spatial movement in Appendix A \eqref{sec:level9}. A detailed analysis with three-dimensional spatial dynamics will be a theoretical avenue to explore. We believe that the results reported in this work with time-varying phase interaction will enrich the exiting study of swarmalators and emphasize competitive interaction in the swarmalator systems.
\color{black}
\section*{Acknowledgements}
\par We thank and gratefully acknowledge anonymous reviewers for their helpful comments and insightful suggestions that helped in considerably improving the manuscript. S.N.C acknowledges the financial support by the Council of Scientific \& Industrial Research (CSIR) under Project No.\ 09/093(0194)/2020-EMR-I. M.P. acknowledges funding from the Slovenian
Research Agency (Grant nos P1-0403, J1-2457, and J1-9112).
\section*{Data availability statement}
The data that support the findings of this study are openly available in the GitHub repository \cite{web_5}.
\section*{\label{sec:level9}Appendix A: $3$D model of swarmalators}
\begin{figure}[hpt]
\centerline{
\includegraphics[scale = 0.35]{figure12.pdf}}
\caption{{\bf Swarmalators in $3$D}: Snapshots of the swarmalators in three-dimensional space at $t = 200$ time unit. (a) Static sync state for $\epsilon_a = 0.5$, $\epsilon_r = -0.5$, $r = 10.0$, and $J = 0.8$. The static sync state takes spherical structure in $3$D where all the swarmalators lie inside this sphere and they are completely phase synchronized. (b) Static async state for $\epsilon_a = 0.5$, $\epsilon_r = -1.0$, $r = 10^{-5}$, and $J = 0.8$. In this state also the swarmalators form a spherical structure analogous to their $2$D disc shape where phases remain desynchronized. Here, $N=500$.}
\label{Fig.12}
\end{figure}
One of the motivations of studying the dynamics of swarmalators is to mimic the behavior of real-world entities where external and internal dynamics are influenced by each other. Although the real-world systems are generally three-dimensional, for simplicity, we have introduced our model with two-dimensional spatial dynamics. After investigating the long-range behavior of the swarmalators under different interaction functions and coupling schemes, it is now suitable to find out if the results with $2$D spatial dynamics are generalized when the spatial dynamics are extended to $3$D. Now the spatial position of the $i$-th swarmalator is denoted by the vector $\textbf{x}_{i} = (x_i,y_i,z_i) \in \mathbb{R}^3$. In $3$D the swarmalator model is as follows
\begin{equation}
\label{eq.41}
\dot{\textbf{x}}_{i} = \frac{1}{N-1} \sum_{\substack{j = 1\\j \neq i}}^{N}\left[ \frac{\textbf{x}_{ij}}{|\textbf{x}_{ij}|} (1+J\cos(\theta_{ij})) - \frac{\textbf{x}_{ij}}{{|\textbf{x}_{ij}|}^{3}} \right],
\end{equation}
\begin{equation}
\label{eq.42}
\dot{\theta}_{i} = \frac{\epsilon_a}{N_{i}}\sum_{\substack{j = 1\\j \neq i}}^{N} \text{A}_{ij}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|^2} + \frac{\epsilon_r}{N-1-N_{i}}\sum_{\substack{j = 1\\j \neq i}}^{N} \text{B}_{ij}\frac{\sin(\theta_{ij})}{|\textbf{x}_{ij}|^ 2},
\end{equation}
where $|\textbf{x}_{ij}|$ denotes the Euclidean distance between the $i$-th and the $j$-th swarmalators in three-dimensional space. These equations are similar to the governing equations of $2$D swarmalator model defined by Eqs.\ \ref{eq.3} and \ref{eq.4}. Although, here we have chosen $\alpha = 1$, $\beta = 3$, and $\gamma=2$ which differ significantly from our previous choice of $\alpha = 1$, $\beta = 2$, and $\gamma=1$ for the earlier results of our work.
The states in the earlier simulations with the two-dimensional spatial dynamics now form their three-dimensional structures. In Fig.\ \ref{Fig.12}, we plot the snapshot of the swarmalators in the static sync and static async state. We find complete phase synchrony among swarmalators' phases in the static sync state where the two-dimensional disc structure is now converted into a three-dimensional sphere (see Fig.\ \ref{Fig.12}(a)). The same phenomenon takes in the static async state in Fig.\ \ref{Fig.12}(b) where swarmalators' phases remain desynchronized. The radii of these spheres depend on the choices of $\alpha$, $\beta$ and $\gamma$. All the other states found in the $2$D model can be found in the $3$D model too with appropriate choices of parameter values. Moreover, a different set of values for ($\alpha$, $\beta$, $\gamma$) does not change the qualitative behavior of the emerging asymptotic states significantly, as observed here.
\bibliographystyle{apsrev4-1}
\providecommand{\noopsort}[1]{}\providecommand{\singleletter}[1]{#1
|
\section{Introduction} \label{sec: introduction}
Tidal disruption events (TDEs) occur when a star passes within the tidal radius $R_{t} = R_{*} (\text{M}_{\text{BH}}/\text{M}_{*})^{1/3}$ of a supermassive blackhole (SMBH), where $R_{*}$ and $\text{M}_{*}$ are the radius and mass of the star, respectively, and $\text{M}_{\text{BH}}$ is the mass of the SMBH. Within this radius, the star's self gravity is overwhelmed by tidal forces, causing the star to be torn apart \citep{Hills1975}. Roughly half of the stellar debris is accreted onto the SMBH and powers a luminous flare, whilst the rest of the material is ejected from the system \citep{Rees1988, Phinney1989}.
The physical processes involved in generating the luminous flare are still uncertain. In the simplest, "canonical", scenario, the flare is a consequence of bound stellar material forming a quasi-circular accretion disc through which material is transported at typically super-Eddington rates \citep[see, e.g.][]{Shiokawa_2015, Hayasaki2016}. The light observed is then thermal emission from this hot accretion disc. The emission from TDEs is expected to be dominated by soft X-ray emission, with a weak optical and ultraviolet (UV) component \citep{Rees1988, Strubbe2009}.
\begin{figure*}
\centering
\includegraphics[scale=0.58]{img/figure1_observations.pdf}
\caption{Rest frame optical and UV spectra for a sample of TDEs. The spectra have been coloured depending on their \citet{VanVelzen2020} spectral class. Blue corresponds to TDE-Bowen, orange to TDE-H and green to TDE-He. \textit{Top}: normalized UV and optical spectra of ASASSN-14li \citep{Holoien2016}, AT2018zr \citep{Hung2019} and iPTF15af \citep{Blagorodnova_2019}. The optical spectra have been re-scaled for the blue edge to roughly match the red edge of the UV spectra. The re-scaling amounts, as well as observation phase of the spectra are labelled in the figure. Specifically visible is how some TDEs display BELs (e.g. ASASSN-14li), whilst others exhibit BALs (e.g. iPTF15af). \textit{Bottom}: optical spectra of AT2018iih \citep{VanVelzen2020}, AT2018zr \citep{Hung2019}, ASASSN-14ae \citep{Holoien2014}, ASASSN-14li \citep{Holoien2016}, AT2019dsg \citep{VanVelzen2020}, iPTF15af \citep{Blagorodnova_2019} and ASASSN-14li \citep{Holoien2014}. Important transitions have been labelled at the top of the figure. None of the spectra have been host galaxy subtracted and have been smoothed using a boxcar filter.}
\label{fig: dual_p_tde_uv_opt_obs}
\end{figure*}
Contrary to this expectation, an increasing number of UV and optically bright TDEs with a weaker than expected soft X-ray component have been discovered over the past decade. The exact source of the optical emission is still uncertain, but it is clearly inconsistent with the spectral energy distribution (SED) of a hot accretion disc. Two mechanisms have been proposed to explain this ``optical excess''. In the first, the optical emission is powered by shocks associated with collisions between infalling stellar debris streams \citep[e.g.][]{Dai_2015, Piran2015, Shiokawa_2015}. {In the second, the optical excess is due to the reprocessing -- and consequent softening -- of the disc's X-ray radiation in a surrounding optically thick medium, such as an outflow \citep[e.g.][]{Strubbe2009, Metzger2016a, Roth2016, Dai_2008, roth_what_2018, Lu_2020, Piro_2020, bonnerot_2020}.}
Observationally, the optical spectra of TDEs are characterised by broad (FWHM $\sim 10^{4} \text{ km s}^{-1}$) hydrogen and/or helium lines (see bottom panel of Figure \ref{fig: dual_p_tde_uv_opt_obs}). In some TDEs, a complex series of emission lines due to Bowen fluorescence are also present. In the taxonomy proposed by \citet {VanVelzen2020}, TDEs can be broadly split into three distinct sub classes: i) TDE-H: broad H$\alpha$~ and H$\beta$~ emission lines; ii) TDE-Bowen: broad H$\alpha$~ and H$\beta$~ emission lines with a complex of emission lines around \atomictransition{He}{ii} $\lambda 4686$ with most due to Bowen fluorescence \citep[see][]{Bowen1934, Bowen1935}, and; iii) TDE-He: no broad Balmer features but a broad emission feature near \atomictransition{He}{ii} $\lambda 4686$. The most common class of optical TDE (to date) is TDE-H, followed by TDE-Bowen and TDE-He.
The discovery of blue-shifted broad absorption lines (BALs) in the ultraviolet spectra of some TDEs \citep[e.g.][]{Cenko_2016, Blagorodnova_2019, Hung2019, hung_discovery_2020} provides unambiguous observational evidence for the presence of fast and powerful outflows in these systems (top panel of Fig. \ref{fig: dual_p_tde_uv_opt_obs}). Additional evidence for powerful sub-relativistic outflows in TDEs also comes from blueshifted broad \textit{emission} lines (BELs) \citep{arcavi14, roth_what_2018, Hung2019, Nicholl2020} and from radio observations \citep{vanvelzen16, alexander16, alexander17, anderson20}. Not all TDEs exhibit BALs at the same stage in their outburst evolution, however. For example, ASASSN-14li instead exhibits BELs (cf. the top panel of Fig. \ref{fig: dual_p_tde_uv_opt_obs}). Acknowledged by, e.g., \citet{Blagorodnova_2019, hung_discovery_2020}, the BEL vs. BAL dichotomy of TDEs is reminiscent of Type I quasars (QSOs). Most QSOs exhibit BELs in their UV spectra, but $\simeq 20$ per cent of the population instead display prominent BALs \citep[the so-named Broad Asborption Line QSOs][]{Weymann1991, Dai_2008, Knigge2008, Allen2010}. In QSOs, line formation within outflows has, in the past, been invoked to explain the BEL vs. BAL dichotomy, following as a consequence of the orientation of an observer \citep[e.g.][]{Murray1995, Elvis2000, Higginbottom2013}.
In \citet[][]{Parkinson2020}, hereafter referred to as \citetalias{Parkinson2020}, we showed that line formation in an accretion disc wind could explain the existence of both BELs and BALs in the UV spectra of TDEs. Specifically, we proposed that the BEL vs. BAL dichotomy arises naturally as a consequence of the orientation of an observer. We found that BALs are preferentially seen for sight lines looking into the outflow, whereas BELs are more likely to be observed for viewing angles which are above or below the wind cone. The work presented here extends our modelling to longer wavelengths. We find that accretion disc emission reprocessed by an accretion disc wind is a promising mechanism for the formation of the distinctive \textit{optical} emission line spectra seen in TDEs. Our goal, then, is to test if the outflow reprocessing model, for the optical excess in TDEs, can naturally produce the optical (and ultraviolet) emission line spectra and broadband SED seen in TDEs. The specific type of outflow we consider is a simple biconical accretion disc wind. This is quite a natural scenario, since strong, radiation-driven mass loss is almost inevitable, given the extreme luminosities of TDEs.
The paper is structured as follows. In Section~\ref{sec: method}, we describe our wind models and radiative transfer calculations. We then present the results of our modelling in Section~\ref{sec: results} and discuss their implications in Section~\ref{sec: discussion}. Finally, we summarize our findings in Section~\ref{sec: conclusion}.
\section{Radiative transfer and model setup} \label{sec: method}
\subsection{Radiative transfer and ionization}
Numerical simulations were conducted using {\sc python}\footnote{{\sc python~} is a collaborative open-source project available at \hyperlink{https://github.com/agnwinds/python}{github.com/agnwinds/python}.}, a state-of-the-art Monte Carlo radiative transfer and ionization code for moving media using the Sobolev approximation \citep[e.g.][]{Sobolev1957, Rybicki1978}. {\sc python~} was originally described by \citet{Long2002} and subsequent improvements to the code have been described multiple times in the literature \citep{Higginbottom2013, Higginbottom2014, Matthews2015, Matthews2016a}. Here, we provide only a brief description.
\subsubsection{Basic operation}
{\sc python~} consists of two separate calculation stages. The first stage concerns itself with calculating the ionization state, level populations and temperature structure of an outflow on a spatially discretised grid. This is done iteratively by tracking a population of Monte Carlo energy quanta (``photon packets'') and simulating their random walk through the grid. Photon packets are randomly generated over a wide frequency range, sampled from the spectral energy distribution (SED) of the radiation sources included in the simulation. As photon packets travel through the grid, they interact with the plasma and update Monte Carlo estimators which are used to describe the radiation field in each grid cell. The heating effect of photon packets is recorded and used to iterate the temperature towards thermal equilibrium, where the amount of heating and cooling in each grid cell is eventually balanced.
Once the photon packets have been transported through the grid, updated temperature and radiation field estimators are used to recalculate level populations and an updated ionization state of the outflow. This process is repeated until the simulation has converged. We consider a grid cell to be converged when i) the electron and radiation temperature have stopped changing between iterations to within 5 per cent, and, ii) when the heating and cooling rates are balanced to within 5 per cent. It is usually not necessary, or expected, for all grid cells to converge. Typically, cells with poor photon statistics or noisy Monte Carlo estimators tend to not converge. These cells are usually located near the outer edge of the computational domain and are relatively unimportant; they typically contribute little to the final result.
The second calculation stage produces synthetic spectra for a converged simulation. Additional populations of photon packets are generated, typically over a narrow frequency range to give high signal-to-noise, and are flown through the \textit{converged} grid to generate spectra for a selection of sight lines.
\subsubsection{Atomic data}
By default {\sc python~} assumes solar abundances in the modelled outflow, following \cite{1994A&AS..108..287V}. We use the same atomic data as outlined by \citet{Long2002}, with subsequent improvements as described by \citet{Higginbottom2013} and \citet{Matthews2015}. Hydrogen and helium are described with a multi-level model atom, treated using the ``macro-atom'' formalism of \citet{Lucy2002, Lucy2003}. Metals, however, are treated with the two-level atom formalism described by \citet{Long2002}. The resulting hybrid approach is explained by \citet{sim_two-dimensional_2005} and \citet{Matthews2015}.
\subsubsection{Radiation sources} \label{sec: radiation_sources}
Multiple radiation sources can be included in a {\sc python~} simulation. In the models presented here, we include two radiation sources: an accretion disc and the outflow itself. However, the accretion disc is the only \textit{net} source of photons, as all emission produced in the outflow is reprocessed disc radiation.
Modulo adiabatic cooling, the outflow is assumed to be in radiative equilibrium. Any energy absorbed by it is reprocessed and either re-radiated or converted to $PdV$ work. {Radiative reprocessing in the outflow takes place via radiative recombination, free-free and line emission and Compton scattering (which is also included in the heating/cooling balance of the plasma).} As these processes depend on the temperature and ionization state, the number of photon packets generated by the outflow is updated as these quantities change. The accretion disc is assumed to be geometrically thin and optically thick. It it treated as an ensemble of annuli, each radiating as a black body with a standard $\alpha$-disc temperature profile \citep{Shakura1973}. The disc SED is therefore specified entirely by the mass accretion rate, $\dot{\text{M}}_{\text{disc}}$, and the mass of the black hole, $\text{M}_{\text{BH}}$. In principle, both the inner and outer radii of the disc are free parameters, but the former is typically set to the innermost stable circular orbit for a Schwarzschild black hole. Both fore-shortening and limb-darkening are included, resulting in a highly anisotropic radiation field.
In reality, the inner disc region in TDEs are likely radiation dominated and vertically extended. However, the structure, evolution and stability of such discs is an area of intense research and still highly uncertain \citep[e.g.][]{Hirose2009, Jiang_2013, blaes14, Shen_2014}. In the near-Eddington critical accretion regime, where $\dot{\text{M}}/\dot{\text{M}}_{\text{Edd}} \approx 1$, so-called ``slim'' disc models are often used to describe the disc structure \citep[see, e.g.][]{Abramowicz1988, Strubbe2009}. In these slim disc models, for large Eddington fractions the temperature for the inner disc region is lowered due to decreasing radiative efficiency and consequently increased radiation trapping. In the accretion regime of interest, i.e. $\dot{\text{M}}/\dot{\text{M}}_{\text{Edd}} \leq 0.5$, we find that both a slim disc and a standard $\alpha$-disc temperature profile result in very similar disc SEDs. Our reliance on a simple $\alpha$-disc treatment, then, is one borne out of practicality in the absence of a more physically realistic description.
\subsubsection{Clumping}
Whilst {\sc python~} was originally developed to model smooth outflows, in reality various instabilities within the flow are likely to break the smooth flow up into clumps. Multiple clumping mechanisms have already been identified in various astrophysical settings \citep[e.g.][]{Owocki1983, McCourt2018, dannen_clumpy_2020}. From a computational standpoint, addressing this is difficult. Not only does clumping introduce additional parameters into already complex models, but, ideally, the spatial resolution of the computational grid would be high enough to resolve each individual clump. This is not yet feasible, and so {\sc python~} implements a necessarily simple approximation known as \textit{micro-clumping}, which is commonly used in stellar-wind modelling \citep[e.g.][]{Hamann1998, Hamann2013}. Micro-clumping in {\sc python~} has been discussed in previous work by \citet{Matthews2016a, Matthews2020} and, in the context of the problem at hand by \citetalias{Parkinson2020}. Within the micro-clumping framework, clumps are assumed to be optically thin, smaller than all relevant length-scales and embedded in a vacuum. In this limit, clumping can be parameterised by a single parameter $f_{v}$, the volume filling factor. The density of a clump is enhanced by a factor $D = 1 / f_{v}$ relative to the density of the equivalent smooth flow. Consequently, at a fixed temperature, processes which scale linearly with density (e.g. electron scattering) remain unchanged, whereas processes which scale with the square of the density (e.g. recombination) are enhanced.
One of the long-standing challenges for wind models in QSOs is overionization. This can occur when an outflow is exposed to an (E)UV or X-ray radiation field near the central engine \citep{Proga2002, Higginbottom2013}. In line-driven winds, if the wind is too highly ionized it can become impossible to sustain the wind \citep[e.g.][]{Proga2000, Proga2002}. Irrespective of the driving mechanism of the wind, overionization can also prevent the formation of both absorption and emission features. Clumping is one natural way to overcome the so-called ``overionization problem'' and can moderate the ionization state of the outflow (e.g. \citealt{Hamann2013, Matthews2016a}; another solution is self-shielding, e.g. \citealt{Murray1995,Proga2000, Proga2004}). Clumping is therefore also an important effect to consider in TDEs. {In previous QSO models, \citet{Matthews2016a} and \citet{Matthews2020} assumed clumping factors of both $f_{v} = 0.1$ and $f_{v} = 0.01$.} Throughout this work we have adopted a volume filling factor $f_{v} = 0.1$, corresponding to a density enhancement by a factor of ten.
\subsubsection{Special relativity}
Prior to this work, {\sc python~} made the assumption that the outflow velocities were small compared to the speed of light. Doppler shifts were therefore treated "classically" (proportional to linear order in $v/c$), and no other corrections associated with relativistic effects were considered. For this work, however, we have updated {\sc python~} to fully incorporate special relativistic effects. {\sc python~} now explicitly distinguishes and correctly transforms between the co-moving (local/fluid) and observer frames. We ensure that energy is conserved in the co-moving frame and that the correct special relativistic transforms are applied when converting between frames \citep[see, e.g.,][]{Castor1972, Mihalas_Mihalas1984}.
The synthetic spectra are generated in the observer frame, the same frame in which photons are transported. Photon interactions, however, are handled in the co-moving frame. This means photons are transformed into the local fluid frame when they interact with the outflow, e.g. during resonance scattering. All atomic processes, as well as heating and cooling, are treated in the co-moving frame.
\subsection{A biconical wind model}
\begin{figure}
\centering
\includegraphics[scale=0.35]{img/figure2_sv93_wind.png}
\caption{A cartoon showing the basic geometry of the \citet{Shlosman1993} biconical disc wind model, adapted from \citet{Long2002}.}
\label{fig: sv93_cartoon}
\end{figure}
In this work, we model a disc wind outflow using a kinematic biconical wind prescription presented originally by \citet{Shlosman1993}. This model is illustrated in Figure \ref{fig: sv93_cartoon}. In this prescription, streamlines emerge from disc radii between $r_{\text{min}}$ and $r_{\text{max}}$ at angles relative to the disc normal given by
\begin{equation}
\theta(r_{0}) = \theta_{\text{min}} + (\theta_{\text{max}} - \theta_{\text{min}}) x^{\gamma},
\label{eq: sv93_theta}
\end{equation}
where
\begin{equation}
x = \frac{r_{0} - r_{\text{min}}}{r_{\text{max}} - r_{\text{min}}}.
\label{eq: sv93_x}
\end{equation}
Here, $r_{0}$ is the launch radius of a streamline, $\theta_{\text{min}}$ and $\theta_{\text{max}}$ are the minimum and maximum opening angles of the wind, and $\gamma$ controls the concentration of streamlines towards either of the two boundaries $r_{\text{min}}$ and $r_{\text{max}}$.
The launch velocity, $v_{0}$, and terminal velocity, $v_{\infty}$, of a streamline are set to the local sound
speed $c_{s}$ and a multiple of the escape velocity, $v_{\rm esc}$, at the streamline footpoint, respectively. The poloidal velocity, $v_{l}$, at a distance $l$ along a streamline, is given by the following power law,
\begin{equation}
v_{l}(r_{0}) = v_{0} + (v_{\infty} - v_{0})\left[ \frac{(l / R_{v})^{\alpha}}{(l / R_{v})^{\alpha} + 1} \right],
\label{eq: sv93_velocity}
\end{equation}
where $R_{v}$ is the acceleration length scale, and $\alpha$ controls the acceleration along the streamline. The rotational velocity, $v_{\phi}$, is Keplerian at the footpoint of a streamline and is assumed to conserve specific angular momentum as it rises above the disc. At a cylindrical radius $r$, $v_{\phi}$ is given by
\begin{equation}
v_{\phi}(r) = v_{k} r_{0}/r,
\label{eq: sv93_rotational_vel}
\end{equation}
where $v_{k}$ is the Keplerian velocity at $r_{0}$. The density at any point in the wind is obtained by enforcing mass continuity, giving
\begin{equation}
\rho(r, z) = \frac{r_{0}}{r} \frac{dr_{0}}{dr} \frac{\dot{m}^{\prime}}{v_{z}(r, z)},
\end{equation}
where $\dot{m}^{\prime}$ is the mass-loss rate per unit surface area of the accretion disc,
\begin{equation}
\dot{m}^{\prime} \propto \dot{\text{M}}_{\text{wind}} r_{0}^{\lambda} \cos[\theta(r_{0})].
\label{eq: sv93_mass_loss_disc_surf}
\end{equation}
The parameter $\lambda$ controls where mass is lost from the disc, and $\dot{\text{M}}_{\text{wind}}$ is the total mass loss rate of the wind. A value of $\lambda = -2$ results in roughly uniform wind densities at the base of the outflow, across the disc surface.
\subsection{Simulation grid setup} \label{sec: model_grid_setup}
\begin{table}
\centering
\resizebox{\columnwidth}{!}{ \begin{tabular}{ccccc}
\hline
\multicolumn{5}{l}{Grid parameters} \\
\hline
Parameters & \multicolumn{3}{c}{Values} & Units \\
\hline
$\dot{\text{M}}_{\text{disc}}$ & 0.05 & 0.15 & 0.50 & $\dot{\text{M}}_{\text{Edd}}$ \\
$\dot{\text{M}}_{\text{wind}}$ & 0.1 & 0.3 & 1.0 & $\dot{\text{M}}_{\text{disc}}$ \\
$v_{\infty}$ & 0.1 & 0.3 & 1.0 & $v_{\text{esc}}$ \\
\hline
\multicolumn{5}{l}{Geometry parameters} \\
\hline
Parameters & \multicolumn{3}{c}{Values} & Units \\
\hline
$\text{M}_{\text{BH}}$ & $10^{6}$ & $3 \times 10^{6}$ & $10^{7}$ & $\text{M}_{\odot}$ \\
\hline
$\text{R}_{\text{co}}, r_{\text{min}}$ & $8.85 \times 10^{11}$ & $2.66 \times 10^{12}$ & $8.85 \times 10^{12}$ & cm \\
& 6 & 6 & 6 &
$\text{R}_{\text{g}}$ \\
$\text{R}_{\text{disc}}, r_{\text{max}}$ & $7 \times 10^{13}$ & $10^{14}$ & $1.5 \times 10^{15}$ & cm \\
& 473 & 228 & 102 &
$\text{R}_{\text{g}}$ \\
$\alpha$ & 4 & 4 & 4 & - \\
$R_{v}$ & $3\times10^{14}$ & $8.9 \times 10^{14}$ & $3\times10^{15}$ & cm \\
& $2000$ & $2000$ & $2000$ &
$\text{R}_{\text{g}}$ \\
$\theta_{\text{min}}$ & 20 & 20 & 20 & $^{\circ}$ \\
$\theta_{\text{max}}$ & 65 & 65 & 65 & $^{\circ}$ \\
$\text{R}_{\text{max}}$ & $5 \times 10^{17}$ & $5 \times 10^{17}$ & $5 \times 10^{17}$ & cm \\
& $3.39 \times 10^{6}$ & $1.13 \times 10^{6}$ & $3.39 \times 10^{5}$ &
$\text{R}_{\text{g}}$ \\
$\gamma$ & 1 & 1 & 1 & - \\
$\lambda$ & 2 & 2 & 2 & - \\
$f_{v}$ & 0.1 & 0.1 & 0.1 & - \\
\hline
\multicolumn{5}{l}{Eddington ratio conversions} \\
\hline
Ratio & \multicolumn{3}{c}{Values} & Units \\
\hline
$\text{M}_{\text{BH}}$ & $10^{6}$ & $3 \times 10^{6}$ & $10^{7}$ & $\text{M}_{\odot}$
\\
\hline
0.05 $\dot{\text{M}}_{\text{Edd}}$ & $1.11 \times 10^{-3}$ & $3.33 \times 10^{-3}$ & $1.11 \times 10^{-2}$ &
$\text{M}_{\odot}~\text{yr}^{-1}$ \\
0.15 $\dot{\text{M}}_{\text{Edd}}$ & $3.33 \times 10^{-3}$ & $9.99 \times 10^{-3}$ & $3.33 \times 10^{-2}$ &
$\text{M}_{\odot}~\text{yr}^{-1}$ \\
0.50 $\dot{\text{M}}_{\text{Edd}}$ & $1.11 \times 10^{-2}$ & $3.33 \times 10^{-2}$ & $1.11 \times 10^{-1}$ &
$\text{M}_{\odot}~\text{yr}^{-1}$ \\
\hline
\end{tabular}}
\caption{{Key parameters for the models presented. The top sub-table shows the values for the parameter grid as described in Section \ref{sec: model_grid_setup}, whilst the middle sub-table defines the parameters which control the geometry for each black hole mass. The bottom table shows conversion values of the Eddington accretion rate into solar masses per year for each black hole mass.}}
\label{table: key_parameters}
\end{table}
In our previous paper we explored the formation of UV lines in accretion disc winds associated with TDEs \citepalias{Parkinson2020}. In this paper, we now present both optical and UV spectra from a broader parameter search.
The masses of the black holes in TDEs are expected to fall within a broad, but well defined mass range. If the black hole is too large, i.e. $\gtrsim 10^{8}~\text{M}_{\odot}$, then the disruption radius will be within the event horizon of the black hole. In this case the star will be completely swallowed before it is disrupted \citep{MacLeod_2012}, resulting in no observable TDE. At the opposite end of the mass scale, the lowest inferred black hole masses are $\simeq 5 \times 10^{5}~\text{M}_{\odot}$ \citep{Wevers_2017, Mockler_2019, ryu2020, Zhou_2021}. It is worth noting that several different methods are being used to estimate black hole masses in TDEs. For example, if the velocity dispersion can be measured from the stellar kinematics of the host galaxy, the $\text{M}_{\text{BH}}-\sigma$ relation can be used \citep[e.g.][]{Wevers_2017}. Alternatively, a multi-parameter model can be fit to the observed light curve to extract the physical properties of the TDE \citep{Guillochon_2018, Mockler_2019}.\footnote{The widely used {\sc mosfit} light curve modelling package is available at \hyperlink{https://github.com/guillochon/MOSFiT}{https://github.com/guillochon/MOSFiT}}
In order to settle on a reasonable range of black hole masses and Eddington ratios for our simulation grid, we used the largest compilation of these parameters by \citep{ryu2020}. Their estimates are based on {\sc tdemass}\footnote{\hyperlink{https://github.com/taehoryu/TDEmass}{https://github.com/taehoryu/TDEmass}} \citep{ryu2020}, which assumes a \textit{slow circularization} scenario for the in-falling stellar debris. In this model, the optical emission is generated from stream-stream collisions; both black hole and stellar mass can then be estimated from the optical luminosity and colour temperature at peak flare. Even though their underlying physical picture is quite different from that envisaged in our simulations -- where the reprocessed accretion disc radiation dominates the optical light -- the black hole masses inferred via this method are in agreement with those obtained by \textsc{mosfit}, in which the luminosity is assumed to be due to a disc \citep{Mockler_2019}.
Fig. \ref{fig: Mbh_parameter_space} shows the black hole masses and Eddington fractions for a sample of 20 TDEs, as provided by \citet{ryu2020}. We also include the TDE spectra shown in Figure \ref{fig: tde_model_comparison}, at the times the spectra were obtained. Based on Fig. \ref{fig: Mbh_parameter_space}, we consider three black hole masses in our simulations: $\text{M}_{\text{BH}} = 10^{6},~3\times10^{6},\text{ and } 10^{7}~\text{M}_{\odot}$. These cover the vast majority of the observationally inferred black hole parameter space.
{
In terms of accretion rates, we only consider the sub-Eddington regime in the present work: the highest accretion we consider is $\dot{\text{M}}_{\text{acc}} = 0.5 ~ \dot{\text{M}}_{\text{Edd}}$. Early on in their evolution, some TDEs with $\text{M}_{\text{BH}} \lesssim 10^{7}~\text{M}_{\odot}$ are expected to undergo a super-Eddington accretion phase, transitioning eventually to sub-critical accretion rates as the fallback rate of stellar debris diminishes {\citep{Strubbe2009, Wu2018, roth_radiative_2020}}. However, most TDEs in the \citet{ryu2020} sample appear to be sub-critical even at peak. Moreover, the TDE spectra to which we compare our models have typically been obtained $\simeq 2$ months post-peak (cf Fig.~\ref{fig: dual_p_tde_uv_opt_obs}). Many super-Eddington TDEs are expected to have transitioned to the sub-critical regime by this point \citep{Strubbe2009}. Last, but not least, the structure and radiative properties of super-Eddington accretion discs are still poorly understood. By restricting ourselves to the sub-Eddington regime we are able to avoid some of the resulting uncertainties.} We have made the assumption that all the luminosity generated in the TDE is (initially) emitted by the accretion disc. For the model with the largest accretion rate (with $\text{M}_{\text{BH}} = 10^{7}~\text{M}_{\odot}$), this places the maximum disc luminosity of the grid at $L_{\text{disc}} \approx 5 \times 10^{44}~\text{ergs s}^{-1}$.
\begin{figure}
\centering
\includegraphics[scale=0.52]{img/figure3_Mbh_parameters.pdf}
\caption{The Eddington luminosity fraction against inferred black hole masses for a sample of 20 TDEs taken from \citet{ryu2020} and the spectra included in Fig. \ref{fig: tde_model_comparison}, labelled in the legend. Each point from the \citet{ryu2020} sample has been coloured depending on the mass return time $p$ since disruption, measured in days. The Eddington luminosity is estimated using the inferred black hole masses. The bounded region represents the black hole mass parameter space covered by the grid in this work.}
\label{fig: Mbh_parameter_space}
\end{figure}
We not only explore the black hole mass dependence on the optical spectra with our simulation grid, but also the effects of three key parameters: (1) the disc accretion rate, (2) the wind mass-loss rate, and, (3) the terminal velocity of wind streamlines. For each parameter, we explore three values for three black hole masses. The values of these parameters are contained within the upper section of Table \ref{table: key_parameters}.
The inner radius of the accretion disc in each model is set to the location of the innermost stable circular orbit (ISCO). {The outer disc radius is set to $R_{\text{disc}} = 10 R_{t}$, inspired by hydrodynamic simulations that follow the disruption of a solar-like star \citep{bonnerot_2020}.} The outflow is wide-angled and characterised by opening angles $\theta_{\text{min}} = 20^{\circ}$ and $\theta_{\text{max}} = 65^{\circ}$, emanating from the entire accretion disc surface. {The mass-loss rate per unit area from the disc is slightly enhanced at larger disc radii ($\lambda = 2$), meaning the density at the base of the wind increases with radius.}
{The acceleration length scale for each wind model is set to $2000~R_{g}$, a value motivated by the radiation-hydrodynamic simulations presented by \citet{Dai2018}. In their model, the outflow is still accelerating at distances of $\sim 1500-3000~R_{g}$. The launch velocities of the streamlines are set to the the local sound speed at the footpoint of the streamline and are uniformly spaced at the base of the outflow ($\gamma = 1$).}
{The outflow is logarithmically discretised onto a 2D cylindrical rotationally symmetric grid. Photons are transport in 3D, to account for effects due to the rotational velocity field and structure.} The grid has a resolution of $100 \times 100$ grid cells with a greater concentration of cells at smaller radii, where line formation typically occurs. The total volume covered by the grid is a sphere of radius $5 \times 10^{17}$ cm. Previous tests from \citetalias{Parkinson2020} indicate that the grid resolution and spherical extent of the wind are sufficient to resolve the line forming region. A complete summary of the grid and parameter values is shown in Table \ref{table: key_parameters}.
\section{Results} \label{sec: results}
In the following we present the results of our simulations. We first describe the physical characteristics and synthetic spectra for a fiducial model that lies at the center of the parameter space covered by our grid. We then explore how the synthetic spectra depend on the four key parameters covered by our grid (accretion rate, mass-loss rate, terminal velocity and black hole mass).\footnote{The synthetic spectra and {\sc python~} parameter files for the entire grid are available online at \hyperlink{https://github.com/saultyevil/tde_optical_reprocessing}{github.com/saultyevil/tde\_optical\_reprocessing} or upon request.}
\subsection{A fiducial model} \label{section: fiducial_model}
We begin by considering a fiducial "benchmark" model, in which all parameters are set to their central values on the grid: $\text{M}_{\text{BH}} = 3 \times 10^{6}~\text{M}_{\odot}$; $\dot{\text{M}}_{\text{disc}} = 0.15~\dot{\text{M}}_{\text{Edd}}$; $\dot{\text{M}}_{\text{wind}} = 0.3~\dot{\text{M}}_{\text{disc}}$ and $v_{\infty} = 0.3~v_{\text{esc}}$. We now present and describe both the physical properties of the outflow and its optical spectrum.
\subsubsection{Physical properties}
\begin{figure*}
\centering
\includegraphics[scale=0.54]{img/figure4_wind_properties.pdf}
\caption{Colour plots showing a selection of physical properties for the fiducial model on logarithmic axes. The $r-z$ plane, where $r$ is the cylindrical radius, is shown and the wind is rotationally symmetric about the $z$-axis. The lines drawn over the wind show sight lines for an observer for inclination angles indicated in the legend. The spatial scales and colour maps are both logarithmic. \textit{Top left:} the electron temperature. {\textit{Top right:} the Hydrogen number density.} \textit{Middle left:} the ionization parameter. {\textit{Middle right:} H \textsc{i} ion fraction.} The contour lines show the origin of H$\alpha$~ photons contributing to the spectra. From darkest to lightest colour, the contours are regions where the photon count is 0 $N_{\text{tot}}$, 0.25 $N_{\text{tot}}$, 0.5 $N_{\text{tot}}$, 0.75 $N_{\text{tot}}$ and 1.0 $N_{\text{tot}}$ and where $N_{\text{tot}}$ is the total number of H$\alpha$~ photons. \textit{Bottom left:} the polodial velocity. \textit{Bottom right:} the rotational velocity.}
\label{fig: fiducial_wind_plot}
\end{figure*}
In Fig. \ref{fig: fiducial_wind_plot}, we show colour plots for a selection of physical parameters in the fiducial model and five lines showing how and where different sight lines intersect the wind, as well as additional contours showing the relative number density of H$\alpha$~ photon emission.
At the base of the wind, the velocity field is rotation dominated and effectively Keplerian, with $v_{\phi} \sim 10^{5}$ \kms~near the inner wind edge and $v_{\phi} \sim 10^{4}$ \kms~at the outer edge. {The launch velocities of streamlines are set as the local sound speed, therefore at the inner and outer edges of the wind the launch velocity is $v_{0} \sim 40~\text{km s}^{-1}$ and $v_{0} \sim 10~\text{km s}^{-1}$, respectively. And by $\sim 10^{15}$ cm, the outflow has accelerated to $10^{4}$ km s$^{-1}$. The velocity at the base of the wind is far lower than that in the \citet{Dai2018} radiation-hydrodynamic model, from which the acceleration length scale is inspired. In this model, the velocity of the outflow is of order $\sim 10^{4}$ km s$^{-1}$ at its base. However, at large radii the velocities are comparable between the two models.}
{Along the disc plane, the matter density increases with radius (due to the choice of $\lambda = 2$) and is $n_{\rm H} \sim 10^{9}$ \pcmc\ at the inner edge which increases to $n_{\rm H} \sim 10^{12}$ \pcmc\ at the outer edge. At increasing cylindrical radii, both the hydrogen density and rotational velocity decrease. As the electron density is also decreasing, line formation processes which scale with the electron density, such as collisional excitation, also decrease in strength.} The rotational velocity decreases linearly with radius, since material in the outflow is assumed to conserve specific angular momentum. However, the poloidal velocity of the wind increases with radius. The result of this is that Doppler broadening of lines is dominated by rotation close to the disc surface, but by the poloidal velocity field at larger radii.
At the base of the wind, where it is exposed directly to the radiation generated by the accretion disc, the electron temperature is greatest. The temperature at the very inner edge is in excess of $T_{e} \sim 3 \times 10^{5}$ K, but is closer to $T_{e} \sim 3 \times 10^{4}$ K at the outer edge. At large radii, near the far outer edges of the wind, adiabatic cooling dominates, and the temperature is much cooler ($T_{e} \sim 10^{3}$ K). Since {\sc python~} does not implement any dust or molecular physics, the treatment of this region of the wind is highly approximate. However, the majority of the line formation we are interested in does not occur in this region. Thus our neglect of this physics should not affect the emergent spectrum to a significant degree.
To measure the ionization state of the wind, we define the ionization parameter
\begin{equation}
U_{\text{H}} = \frac{4\pi}{n_{\text{H}}c} \int_{13.6 \frac{\text{eV}}{h}}^{\infty} \frac{J_{\nu}}{h\nu}~d\nu,
\label{eq: ion_param}
\end{equation}
where $\nu$ denotes frequency, $n_{\text{H}}$ is the number density of hydrogen, $h$ is Planck's constant and $J_{\nu}$ is the monochromatic mean intensity. The ionization parameter measures the ratio of the number density of hydrogen ionizing photons to the number density of hydrogen, making $U_{\text{H}}$ a useful predictor of the global ionization state. However, $U_{\text{H}}$ has no knowledge of the SED shape, meaning it is a poor indicator for the ionization state of ionic species such as C \textsc{iv}.
Fig. \ref{fig: fiducial_wind_plot} shows that the ionization parameter is fairly constant throughout the wind, with {$U_{\text{H}} \sim 10^{5}$ meaning hydrogen is ionized almost everywhere.} At the very top of the wind, the wind is very highly ionized, with {$U_{\text{H}} \sim 10^{7}$}. There is a portion of the wind where {$U_{\text{H}} \lesssim 10$. This part of the wind is located close the the dense base of the wind, at the outer edge. In this region, the wind has been somewhat shielded by the large column of wind material closer in, meaning it is exposed to a softer, reprocessed SED. This results in this part of the wind being cooler. The reduced ionizing flux also means that the neutral hydrogen population is enhanced in this region. This is where the majority of H$\alpha$~ photons are produced, as shown by the contour lines on the middle right panel of Fig. \ref{fig: fiducial_wind_plot}.}
\subsubsection{Optical spectra}
\begin{figure*}
\centering
\includegraphics[scale=0.58]{img/figure5_fiducial_optical_spectrum.pdf}
\caption{Synthetic optical spectra of the fiducial model. The spectra for five inclination angles are shown, each coloured as labelled by the legend. For each inclination angle, the model produces broad Balmer and helium recombination features. At low inclinations, the emission lines are single peaked but transition into double peaked lines as the inclination angle increases. Notably, the intermediate and high inclination spectra are almost identical as to the base of the wind, which these sight lines cut through, dominates spectrum generation. The spectra are plotted on a log-linear axes and have not been smoothed. Labelled at the top of each panel are important line transitions.}
\label{fig: fiducial_optical_spectra}
\end{figure*}
The synthetic optical spectra produced by the fiducial model for five sight lines are shown in Fig. \ref{fig: fiducial_optical_spectra}. {For each inclination angle, the model produces broad, double peaked, recombination emission in H$\alpha$~ and H$\beta$. The spectra also include other \atomictransition{He}{ii} recombination lines commonly seen in the spectra of TDEs. The \atomictransition{He}{ii} emission line is clearest at low inclination; it becomes very broad and blended with H$\beta$~ for intermediate and edge-on inclinations, as these probe higher projected velocities.}
{The line widths and peak-to-peak separations of the double-peaked emission lines increase with inclination. This happens because rotation is the dominant line broadening mechanism, and edge-on sightlines see higher projected rotational velocities. This also explains the double-peaked nature of the emission lines. Rotational kinematics dominate in the line-forming regions. The emissivity of recombination lines scales with density squared, so these features are formed preferentially in the dense base of the wind where the poloidal velocities are still relatively low. In fact, the lines are mostly formed near the outer edge of the wind base, where the outflow is both dense and relatively more neutral (since it is exposed to the redder, reprocessed SED, instead of the ionizing disc SED).}
The emission lines are asymmetric, displaying an extended red wing. TDEs, such as ASASSN-14ae \citep{Holoien2014}, AT2019qiz \citep{hung_discovery_2020} or AT2018zr \citep{holoien_ps18kh_2019}, often exhibit the same asymmetry, deviating from a Gaussian line profile. \citet{roth_what_2018} proposes that the origin of this asymmetry is from line formation in an optically thick outflow. Photons which scatter in an optically thick expanding outflow lose energy through adiabatic work on the outflow, redshifting their frequency resulting in an extended red wing in emission lines \citep[e.g.][]{laurent_effects_2007, Strubbe2009, Metzger2016a, roth_what_2018}, in a process \citet{roth_what_2018} coins as \textit{adiabatic reprocessing}. Other than through energy losses due to adiabatic work, extended red wings could instead form due to (general) relativistic effects \citep[e.g.][]{ERACLEOUS2009133}.
\begin{table}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{cccccc}
\hline
Inclination & Column density & \multicolumn{4}{c}{Optical depths} \\
($^{\circ})$ & (cm$^{-2}$) & & & & \\
\hline
& Hydrogen & Lyman & \atomictransition{He}{i} & \atomictransition{He}{ii} & Electron\\
\hline
35 & $1.02 \times 10^{25}$ & $1.13 \times 10^{-3}$ & $2.4 \times 10^{-4}$ & $4.00 \times 10^{-4}$ & 0.82\\
60 & $2.46 \times 10^{26}$ & 2.75 & 0.58 & $4.01 \times 10^{3}$ & 19.8\\
75 & $8.19 \times 10^{26}$ & 190 & 5.5 & $3.41 \times 10^{5}$ & 56.2\\
85 & $8.42 \times 10^{26}$ & 5.49 & 1.17 & $1.48 \times 10^{4}$ & 54.2\\
\hline
\end{tabular}
}
\caption{{The hydrogen column density and the integrated optical depth for three frequencies corresponding to three photo-ionization edges and electron scattering for the fiducial model in Section \ref{section: fiducial_model}. In cases where the optical depth are constant across multiple frequencies, the electron scattering opacity is dominating the opacity from photo-absorption.}}
\label{table: fiducial_opacity_parameters}
\end{table}
The spectra produced by our fiducial model {at high inclinations are very similar}, with similar line strengths and widths. These sight lines intersect the dense base of the wind, and are probing parts of the wind which have very similar conditions. In Table \ref{table: fiducial_opacity_parameters}, we tabulate the hydrogen column density and {the integrated optical depths for three photo-ionization edges} along four of the sight lines shown in the top left panel of Fig. \ref{fig: fiducial_wind_plot}. Along these sight lines, the integrated optical depths and column densities are very high ($\text{N}_{\text{H}} \sim 10^{26}$ cm$^{-2}$). They are so high, in fact, that this region of the wind actually dominates the formation of the optical spectrum.
As an example, for inclinations close to the disc plane, {the optical depth of the \atomictransition{He}{ii} edge is around $\tau \sim 10^{4} - 10^{5}$}, suggesting that photo-ionization of helium ions in the base of the wind is contributing significantly to reprocessing the disc emission. {In fact, this is the main reprocessing mechanism in our outflows}. Electron scattering optical depths are also fairly large for these inclinations, with {$\tau \simeq 55$}. By contrast, at low inclinations, the electron scattering optical depth is only {$\tau \lesssim 1$}, and the optical depths associated with the photo-ionization edges are broadly comparable. Even so, these sight lines can still intersect a small fraction of the dense wind base and thus see a substantial column density (e.g. $\text{N}_{\text{H}} \sim 10^{25}$ cm$^{-2}$ for $i=35^{\circ}$). Electron scattering is the dominant global source of opacity along low inclination sight lines, since the reduction in photo-ionization opacity means electron scattering reprocesses across a wider frequency range.
\subsection{Synthetic spectra across the simulation grid} \label{section: spectra_results}
\begin{figure*}
\centering q
\includegraphics[scale=0.525]{img/figure6_optical_spectra.pdf}
\caption{Synthetic optical spectra for the four parameter grids on log-linear axes. The first three panels show spectra for the mass accretion rate ($\dot{\text{M}}_{\text{disc}}$), the wind mass-loss rate ($\dot{\text{M}}_{\text{wind}}$) and the terminal velocity of wind streamlines ($v_{\infty}$) grids, using a black hole mass of $\text{M}_{\text{BH}} = 3 \times 10^{6}~\text{M}_{\odot}$. The bottom right panel shows spectra of the fiducial model for three black hole masses ($\text{M}_{\text{BH}}$). In each panel, the orange spectrum corresponds to the fiducial model in Section \ref{section: fiducial_model}. All of the spectra exhibit broad Balmer and helium recombination features, strongly reminiscent of TDE-Bowen objects. The emission lines are double peaked and feature an extended red wing. Each spectrum is for an intermediate, in-wind, inclination angle of $60^{\circ}$. The spectra have not been smoothed. Labelled at the top of each panel are important line transitions.}
\label{fig: optical_line_spectra}
\end{figure*}
In Fig. \ref{fig: optical_line_spectra}, we show an intermediate-inclination spectrum for each model in our simulation grid (as labelled in the legend). Note that the viewing angle adopted for these plots, $i = 60^{\circ}$, corresponds to a sight line which looks into the wind cone. Additionally, in Fig. \ref{fig: fiducial_mass_reprocessing}, we show angle-integrated disc and emergent reprocessed SEDs for the same models, along with the continuum optical depths they present along several sight lines.
\begin{figure*}
\centering
\includegraphics[scale=0.53]{img/figure7_model_reprocessing.pdf}
\caption{The continuum optical depth as a function of frequency for various sight lines (right axis) and both the emergent and disc SEDs for the same models as in Fig. \ref{fig: optical_line_spectra}. The middle column corresponds to the fiducial model from Section \ref{section: fiducial_model}. The spectra have been plotted on logarithmic axes and important photo-ionization edges have been labelled at the top. Each model is producing a redder (reprocessed) SED relative to the input \textit{disc} SED. In general, denser winds are more efficiently reprocessing the disc SED and result in optically bright spectra. The main reprocessing mechanism is photo-ionization of \atomictransition{He}{i} and \atomictransition{He}{ii}, but in some winds H \textsc{i} also contributes. The electron scattering optical depth (corresponding to the ``flat'' region of the optical depth spectra) is roughly the same between models.}
\label{fig: fiducial_mass_reprocessing}
\end{figure*}
For all of the models, the optical spectra feature broad Balmer and helium recombination lines that are strongly reminiscent of the TDE-Bowen class of spectra.\footnote{Note that {\sc python~} does not currently include the Bowen fluorescence process, so the models cannot produce the Bowen features themselves.} {However, the \atomictransition{He}{ii} and H$\beta$~ emission lines can often become blended with one another, making them difficult to distinguish from each other.}
All the synthetic spectra shown in Fig. \ref{fig: optical_line_spectra} feature double-peaked lines and asymmetric red-skewed line wings, for the same reasons as the fiducial model discussed earlier in Section \ref{section: fiducial_model}. No model on our simulation grid creates the pure Balmer recombination spectra (without helium lines) that are characteristic of the TDE-H class. We suspect this is a signature of the radiation field in the optical line-forming regions. More specifically, the absence of helium recombination features in TDE-H systems suggests a lack of photons with energies sufficient to ionize helium (24.6 eV). By contrast, all of our models contain regions whose ionization state and emission measure are both high enough to create observable \atomictransition{He}{i} and \atomictransition{He}{ii} features.
In any case, Fig. \ref{fig: fiducial_mass_reprocessing} shows that each of our wind models, regardless of its detailed parameters, strongly reprocesses the accretion disc continuum (labelled as "disc"), producing a much redder observed SED (labelled as "emergent"). Thus reprocessing of the disc continuum in an outflow can, rather naturally, result in an optically bright emergent SED.
Qualitatively, the line spectra and SEDs do not change drastically with different parameter values or with inclination. More specifically, all the synthetic line spectra produce broadly the same set of transitions, accompanied by a continuum with roughly similar colour temperatures. The ionization state of the wind, in particular that of hydrogen and helium, is insensitive to moderate changes in the input SED, the mass-loading of the wind or its kinematics. {The dense base of the wind, which is the main line-forming region, does not change all that much between models.} This suggests that reprocessing, and the production of Balmer and helium recombination features, can happen over a fairly broad range of conditions in TDEs. Quantitatively, the spectra do, of course, depend on the model parameters. In the following sections, we outline the main effect of the different parameters we have explored on the resulting line spectra and emergent SEDs.
\subsubsection{Disc mass-accretion rate}
The top left panel of Fig. \ref{fig: optical_line_spectra} shows three models with different disc accretion rates. In all of these models, the wind mass-loss rate is 30\% of the accretion rate. The main difference between the spectra is the continuum normalization and the strength of the H$\alpha$~ line. The highest accretion rate model produces the most luminous spectrum. {The increase in the continuum level is mostly due to a more luminous disc, but also because of greater reprocessing associated with the higher density of the wind. This is clear by looking at the top row of Fig. \ref{fig: fiducial_mass_reprocessing}, which shows the reprocessed SED relative to the input SED for the three accretion rates. In the highest accretion rate model, the high frequency region of the disc continuum is slightly less attenuated, but there is significant optical depth from lower frequency edges contributing to reprocessing the SED. So while the higher continuum luminosity can mostly be attributed with the higher accretion rate, it does appear this wind reprocesses more radiation and re-emits it at longer wavelengths.}
{
The spectrum produced with the lowest accretion rate in Fig. \ref{fig: fiducial_optical_spectra} has the strongest emission lines. In general, the strength of the emission lines decreases with accretion rate, even though the higher accretion rate models are reprocessing (and re-emitting via recombination) more emission and that the emissivity of these lines scales with the density squared. In this situation, the lines appear weaker because of the stronger optical continuum due to stronger reprocessing decreasing the line-to-continuum contrast.
}
\subsubsection{Wind mass-loss rate}
The top right panel in Fig. \ref{fig: optical_line_spectra} shows how changing the mass-loss rate of the wind affects the line spectrum, and the second row of Fig. \ref{fig: fiducial_mass_reprocessing} shows how the reprocessed SED changes. The model featuring the highest mass-loss rate (and thus highest densities) has the strongest recombination features, since, stating this once again, the emissivity of recombination scales with the density squared. This model also produces the brightest optical spectrum, since it reprocesses the largest fraction of the disc emission, although the middle mass-loss model is not far behind.
Quite generally, denser winds tend to be more efficient at reprocessing disc emission, generating optically brighter spectra. This derives from the dominant reprocessing mechanisms being photo-ionization and free-free absorption, and since the optical depth of edges, specifically \atomictransition{He}{ii}, and the electron/ion density, which free-free opacity scales with, are enhanced in a denser wind, the reprocessing becomes stronger. In the lowest mass-loss rate model, the input SED is reprocessed less (refer to Fig. \ref{fig: fiducial_mass_reprocessing}) {and the emergent SED is the least red. The opacity of the Lyman and \atomictransition{He}{i} photo-ionization edges are considerably diminished and comparable to the scattering optical depth, resulting in a reduction in the overall amount of reprocessing via photo-ionization. However, as the \atomictransition{He}{ii} edge is still comparably strong, there is still enough reprocessing via photo-absorption to create an optically enhanced spectrum. It appears that the \atomictransition{He}{ii} edge is very important in reprocessing high frequency disc emission.}
\subsubsection{Terminal velocity}
The terminal velocity is one of the parameters which controls the wind kinematics. Larger terminal velocities imply higher velocities \textit{along} streamlines, as well as lower densities (via the continuity equation). Thus slower outflows are denser and produce redder reprocessed SEDs (third row down of Fig. \ref{fig: fiducial_mass_reprocessing}). However, the effect of the terminal velocity on the optical continuum luminosity and colour temperature is relatively modest.
The bottom left panel of Fig. \ref{fig: optical_line_spectra} shows the synthetic spectra for the three different values of $v_{\infty}$ on our grid. Since the density is modified by the velocity, {the two models with the lowest terminal velocities, i.e. which have largest densities at the base of the wind, produce the strongest recombination lines and the brightest optical continuum. Much like changing the mass-loss rate of the wind, varying $v_{\infty}$ from $0.1-1 ~ v_{\rm esc}$ does not significantly affect the overall optical spectrum, producing only modest changes in the normalisation. However, as noted previously, changing the acceleration length scale has a more significant impact.}
\subsubsection{Black hole mass}
The bottom right panel of Fig. \ref{fig: optical_line_spectra} shows the optical spectra produced by the fiducial wind model for three different black hole masses. Changing the black hole mass indirectly affects other parameters of the models as well, such as the three other parameters comprising our grid (since these parameters all scale with black hole mass).
{
The most significant differences between the spectra, associated with different black hole masses, is their continuum luminosities and their line strengths. The model with the most massive black hole produces the strongest lines, as expected since recombination scales with the density squared. The same model produces the brightest optical spectrum, in this case because the accretion disc is the brightest in the grid, rather than because of any significantly greater level of reprocessing relative to the rest of the grid. In the case of the highest black hole mass model, there are simple ``more'' photons to reprocess. Fig. \ref{fig: optical_line_spectra} also shows that the same Balmer and helium recombination features are produced for all three black hole masses. Otherwise, the spectra remain relatively unchanged in terms of distinct spectral features.
}
\section{Discussion} \label{sec: discussion}
\subsection{The origin of optical emission}
In early TDE models, the stellar debris is assumed to fall back onto the black hole, typically, at a super-Eddington rates. The luminosity of these events are then generated via thermal blackbody radiation from a hot, qausi-circular accretion disc \citep[see, e.g.,][]{Rees1988}. The prevalence of optically bright TDEs with a weak soft X-ray component is, then, somewhat inconsistent with these early models. Although the early models of TDEs assumed that the radiation observed would be that observed directly from the hot accretion disc, most modern interpretations of these events involve a significant amount of radiative reprocessing to explain the optical brightness of some TDEs.
{Reprocessing is a core component in a number of TDE models \citep{Strubbe2009, Metzger2016a, Roth2016, Dai_2008, roth_what_2018, Lu_2020, Piro_2020, bonnerot_2020}.} Notably, \citet{Roth2016} showed that reprocessing in an optically thick outflow plays a key role in the formation of the optical spectrum. They identified two distinct reprocessing regimes, which can explain why some TDEs can still retain their soft X-ray component. Reprocessing also plays a key role in the unification scenario proposed by \citet{Dai2018}. In their model, a given TDE could be observed as either optically or X-ray bright, depending on the inclination of the observer. Intermediate and high-inclination observers see an optically bright SED, since the disc emission is reprocessed via adiabatic cooling and photo-ionization along these sight lines. By contrast, observers viewing the TDE face-on see an X-ray bright TDE, since they are looking directly at the exposed inner accretion disc.
In contrast to the dense, spherically symmetric model of \citet{Roth2016}, our biconical disc wind models typically do not not completely absorb the soft X-ray flux. This difference is likely a geometric effect, as photons are able to escape along paths with lower optical depths in our outflows. Thus, similar to \citet{Dai2018}, we find that the ratio of optical to X-ray fluxes increases with inclination angle. At polar angles, the SED is more akin to a underlying disc spectrum, since the optical depth is lower -- and reprocessing less efficient -- along these sight lines.
In Section \ref{sec: results}, we show how reprocessing in an optically thick accretion disc wind, with plausible parameters, results in an optically bright spectrum, significantly redder than the original accretion disc SED. The main reprocessing mechanism in our winds is photo-ionization, which is also the case for the \citet{Dai2018} unification model. {Even so, free-free absorption is \textit{still} important with an optical depth of $\tau_{\rm ff} \sim 10^{2}$ for intermediate and high inclinations in the fiducial model.} The majority of the photo-ionization reprocessing is associated with the ionization of \atomictransition{He}{ii}; in some models, \atomictransition{He}{i} and H \textsc{i} are also important (Fig. \ref{fig: fiducial_mass_reprocessing}). When the opacity in both helium edges is significant, high frequency radiation is efficiently absorbed and reprocessed, softening the SED and reducing the EUV/X-ray brightness. The absorbed luminosity is re-emitted at longer wavelengths via recombination, free-free and line emission.
{
The photo-ionization of \atomictransition{He}{ii} seems to be particularly important for the reprocessing efficiency. For instance, the optical brightness in the wind mass-loss rate sub-grid (second row down in Fig. \ref{fig: fiducial_mass_reprocessing}) increases as the optical depth (and hence absorption) of this edge increases, as is clear by comparing the $0.1~\dot{\rm M}_{\rm disc}$ and $0.3~\dot{\rm M}_{\rm disc}$ models. This is, in-fact, a common trend across all of the models in Fig. \ref{fig: fiducial_mass_reprocessing}. But additionally, models with high optical depths for lower frequency edges (e.g. \atomictransition{He}{i} and Lyman) also have optically brighter spectra (see the top right panel of the first and second row in Fig. \ref{fig: fiducial_mass_reprocessing}). It seems that reprocessing of not just the soft X-ray photons, but also of the EUV photons is important for the production of optically bright spectra.
}
{
As mentioned previously, photons undergoing repeated scattering events in a divergent flow lose energy and are redshifted as they do work on the outflow \citep[see, e.g.,][]{laurent_effects_2007, roth_what_2018}. This effect, sometimes referred to as ``adiabatic reprocessing'', can be important \citep[see Fig. 14 in][]{roth_what_2018}. The integrated electron scattering optical depth in our winds are significant ($\tau \gtrsim 50$ for intermediate and high inclinations in the fiducial model), and photons are almost certain to undergo multiple scatters before reaching an observer. However, compared to outflow models in which adiabatic reprocessing is known to dominate, the scattering optical depths in our winds are fairly modest. For example, \citet{roth_what_2018} report a scattering optical depth of $\tau \simeq 130$ for their spherical outflow, in which repeated scatterings in the outflow play a major role in redshifting the SED. Additionally, as the models here are 2D, photons are able to scatter around highly optically thick regions and do not become trapped, as they would in a spherical model such as by \citet{roth_what_2018}.
}
{
In our fiducial model, the role played by adiabatic reprocessing is minor. The scattering optical depths are similar between the models in Fig. \ref{fig: fiducial_mass_reprocessing}, but the amount of reprocessing is not. It is clear, then, that whilst adiabatic reprocessing can modify the SED, in our models it is not an important process for the creation optically bright spectra. To see this, consider the $\dot{\text{M}}_{\text{wind}} = 0.1~\dot{\text{M}}_{\text{disc}}$ panel in Fig. \ref{fig: fiducial_mass_reprocessing}. Here, the scattering optical depth is similar to other models, but the optical depth of the \atomictransition{He}{i} edge is greatly reduced and comparable to the scattering optical depth. The optical luminosity in this model is low, showing that it tracks the photo-ionization optical depth, rather than the scattering one. Therefore, adiabatic reprocessing is not a dominant effect in our outflow models.
}
\begin{figure*}
\centering
\includegraphics[scale=0.55]{img/figure8_fiducial_reprocessing_comparison.pdf}
\caption{The continuum optical depth as a function of frequency for various sight lines (right axis) and both the emergent and disc SEDs (left axis) for the fiducial model in \citetalias{Parkinson2020} and and in this work (Section \ref{section: fiducial_model}). Both axes are logarithmically spaced and important photo-ionization edges have been labelled at the top. Relative to the \citetalias{Parkinson2020} fiducial model, reprocessing is more efficient in the fiducial model presented in this work. Crucially whilst the both models have similar optical depth for the same photo-ionization edges, the disc continuum peaks at a higher frequency in the ``new'' fiducial model. In the \citetalias{Parkinson2020} model, the disc SED does not include enough \textit{ionizing} photons resulting in reprocessing via photo-ionization becoming inefficient.}
\label{fig: old_new_reprocessing}
\end{figure*}
The emergent optical luminosity is also affected by the SED of the accretion disc. If the SED does not have enough high frequency flux, then there is nothing to reprocess to create a red optical spectrum. This effect is illustrated in Fig. \ref{fig: old_new_reprocessing}, with the fiducial model we used in \citetalias{Parkinson2020} and the fiducial model used in the present work. Clearly, the new fiducial model reprocesses a great deal more of the underlying disc continuum. This is despite the fact that the photo-ionization optical depth for \atomictransition{He}{ii} is not too dissimilar between the models. {But there is a disparity between the optical depths for the \atomictransition{He}{i} and Lyman edges in the old fiducial model, where they are larger.}
So what is responsible for the higher reprocessing efficiency of our new fiducial model, compared to our \citetalias{Parkinson2020} model? Crucially, the disc continuum in this model peaks at a higher frequency, which means a higher fraction of its luminosity is susceptible to reprocessing by the \atomictransition{He}{ii} edge. The reprocessing efficiency is not just related to the physical properties of the wind, notably its (column) density, but it also depends on the underlying SED. If this SED does not include enough \textit{ionizing} photons, reprocessing via photo-ionization becomes inefficient (as does free-free reprocessing). Of course, adiabatic scattering can still reprocess the continuum if the outflow is optically thick, even if this is a relatively small effect in our models.
\subsection{Formation of optical lines} \label{sec: formation_of_optical_lines}
{
When the density of the wind is low, the Balmer and \atomictransition{He}{ii} optical features become weak. If the density is too low to shield or soften the disc emission via reprocessing, the recombination features are greatly diminished as the wind becomes overionized, preventing the formation of absorption and emission lines \citep[e.g.][]{Proga2000, Proga2002, Higginbottom2013}. More generally, however, the strength of recombination features scales with emission measure, $\text{EM} = \int n_{e}^{2}~dV$. This depends not only the local density, but also on the emitting volume, so where in the outflow these lines form is also important. For example, emission lines will preferentially form near the base of the wind where the density is largest. But, if H$\alpha$~ photons were being emitted over an extended region in the sparse, outer portion of the outflow, this emission could still be significant, since the volume of this region is so large.}
{
One way to modify the density of the outflow in our kinematic approach (other than by modifying the mass loss rate) is by changing the acceleration length scale or the acceleration exponent. Together, these variables control how quickly the outflow accelerates. By increasing the acceleration length and increasing the exponent, the outflow density increases and the ionization state is lowered. Both effects result in stronger recombination features, so the appearance of the optical spectra is sensitive to these kinematic model parameters. However, some lines, such as H$\alpha$, are insensitive to changes in the velocity structure. This suggests these line photons are produced in the very base of the wind, whose kinematic properties remain fairly constant with the change in parameters.}
\subsection{The effect of electron scattering on the line profiles}
\begin{figure*}
\centering
\includegraphics[scale=0.55]{img/figure9_spectrum_breakdown.pdf}
\caption{Synthetic spectra for two inclinations (labelled in the top right of each panel) of photons contributing toward the emergent spectrum (labelled as full as in the figure) of the H$\alpha$~ line profile for the fiducial model. Included are ``pure'' continuum photons (free-free, bound-free and electron scattered), as well as H$\alpha$~ photons and H$\alpha$~ photons which have electron scattered at least once between emission and escape (H$\alpha$~ w/ ES). The spectra have been smoothed by a six pixel boxcar filter. Importantly, electron scattering is responsible for the broad wings of the emission lines. It also smooths out the double peaked line profile at low inclinations. \textit{Left:} a low inclination spectrum (10$^{\circ}$). \textit{Right:} an intermediate inclination spectrum (60$^{\circ}$).}
\label{fig: spec_breakdown}
\end{figure*}
In Fig. \ref{fig: spec_breakdown}, we break down the contribution of different physical processes on the spectrum produced by our fiducial model in the vicinity of H$\alpha$, isolating the contribution of ``continuum'' photons which have never undergone any bound-bound interaction in the outflow. This category includes photons emitted directly by the disc and photons produced via bound-free or free-free processes. We also isolate ``line'' photons which have undergone one or more electron scatters since their bound-bound interaction (H$\alpha$~ w/ES). The final contribution is "pure" H$\alpha$~ photons, i.e. photons whose last interaction in the wind was in the bound-bound H$\alpha$~ transition.
Fig. \ref{fig: spec_breakdown} shows very clearly that electron scattering is responsible for the broad wings of the emission line, in line with the results of \citet{roth_what_2018}. In their simulations, higher electron scattering optical depths resulted in broader emission lines. In our models, electron scattering can also smooth out the double peaked line profile, at least at low inclinations (left panel).
\subsection{Comparison to observations}
In Fig. \ref{fig: tde_model_comparison}, we show comparisons of two of our models to the optical spectra of ASASSN14-li and AT2019qiz. Other than the choice of inclination angles and re-scaling the spectra to the relevant distance, the model parameters used to generate the synthetic spectra have not been tailored and/or fine-tuned to fit either object. The observed spectra were taken at roughly the same time post peak, at $ \Delta t \approx 50$ days, and have been corrected for foreground extinction assuming a \citet{Cardelli1989} extinction curve with $R_{v} = 3.1$ with $E(B-V) = 0.0225$ and $E(B-V) = 0.0939$ for ASASSN14-li and AT2019qiz, respectively. {For ASASSN-14li, we have also included photometry, provided by \citet{Holoien2016}, from both the \textit{Swift} UltraViolet and Optical Telescope \citep{roming_swift_2005} and from the 2-m Liverpool telescope \citep{steele2004proc}. For the comparison with ASASSN-14li, we use a black hole with mass $\text{M}_{\rm BH} = 3 \times 10^{6}~\text{M}_{\odot}$ and a mass loss rate of $\dot{\rm M}_{\rm wind} = \dot{\rm M}_{\rm disc}$. For AT2019qiz we use a model with $\dot{\rm M}_{\rm disc} = 0.5~\dot{\rm M}_{\rm Edd}$ and a black hole mass $\text{M}_{\rm BH} = 10^{6}~\text{M}_{\odot}$. The rest of the model parameters are set to their ``fiducial'' values.}
Both TDEs belong to the TDE-Bowen spectroscopic class in the \citet{VanVelzen2020} taxonomy, as they display broad Balmer and helium emission, along with complex of lines due to Bowen fluorescence. TDE-Bowen objects present the strongest case for reprocessing, as Bowen fluorescence requires both large densities and a high flux of EUV photons. As noted earlier, none of our models produce pure Balmer emission spectra, so we have not made any comparisons to TDE-H objects, such as AT2018zr. We also do not make any comparisons to TDE-He objects. These transients are probably associated with the disruption of helium-rich stars, so their pure-He spectra are likely to be an abundance effect. {Note, however, that the disruption of a regular main sequence can in principle also result in hydrogen free spectra as a result of ionization effects \citep{Guillochon_2014,Roth2016}.}
\begin{figure*}
\centering
\includegraphics[scale=0.55]{img/figure10_tde_14-li_2019qiz_models.pdf}
\caption{Rest frame synthetic optical spectra for two models (labelled in the legend) plotted on linear axes and against the optical spectra of ASASSN14-li bat $\Delta t = 54$ days \citep{Holoien2016} and AT2019qiz at $\Delta t = 49$ days \citep{Nicholl2020}. The spectra for the objects have been corrected for foreground extinction assuming a \citet{Cardelli1989} extinction curve with $R_{v} = 3.1$ for both TDEs and $E(B-V) = 0.0225$ and $E(B-V) = 0.0939$ for ASASSN14-li and AT2019qiz respectively. {Photometry for ASASSN-14li, taken from \citet{Holoien2016}, has also been included with the filters labelled in the legend.} The distance of the object and inclination of the model are labelled in the top right of each panel. The spectra have not been smoothed. Important line transitions are marked at the top of each panel. The synthetic spectra are in rather excellent agreement with the observations, especially considering neither model has been tailored to fit the specific observation, producing the same recombination features with similar continua.}
\label{fig: tde_model_comparison}
\end{figure*}
Qualitatively, the models do a good job for matching both of the observed spectra in Fig. \ref{fig: tde_model_comparison}. They exhibit the same recombination features, and some with similar line strengths and line profiles, as well as the optical continua which agree quite well. From a more quantitative perspective, there is good agreement between the fiducial model and ASASSN-14li's continuum and emission line spectrum. The only feature(s) missing entirely from the model are Bowen fluorescence feature, which (as noted previously) is produced by a process that is not yet included in {\sc python}. {The agreement is, of course, far from perfect: for example, the Balmer and the \atomictransition{He}{ii} emission lines are too weak, and the continuum is slightly too blue at UV wavelengths.}
{There is, however, potential for improving the fit of this model by fine-tweaking the model parameters. For example, parameters which modify the density structure of the wind will change the strength of the Balmer lines, since denser winds tend to produce stronger emission as well as brighter optical continua (see \S \ref{sec: formation_of_optical_lines} and Fig. \ref{fig: optical_line_spectra}). Additionally, emission lines can also change by increasing/decreasing the volume of the line emitting region. This could be done by, for example, increasing the covering factor of the wind, i.e. the solid angle subtended by the wind as seen by the central engine, or increasing the outflow outer radius $R_{\rm max}$. Finally, the continuum SED could also be changed to something more realistic, which better approximates the underlying emission in ASASSN14-li. This would likely cause changes to both the line profiles, as well as the overall shape of the continuum.}
{The match between model and observation is slightly less good for AT2019qiz. Here, the synthetic spectrum is slightly redder than the data. Increasing the mass accretion rate of the disc (and scaling the mass loss of the wind appropriately) would probably fix this and make the spectrum slightly bluer. But the model produces the similar emission features, although the Balmer lines are significantly weaker than the data, with the higher order Balmer features missing entirely.}
Overall, given that we have not fine-tuned our models to precisely match the data, we consider the qualitative agreement with observations extremely promising. More specifically, it confirms that reprocessed disc emission can naturally produce both the continuum and emission lines that are commonly seen in the optical spectra of TDE-Bowen objects.
\subsection{UV spectral lines as geometry and orientation indicators} \label{section: pcygni_discussion}
\begin{figure*}
\centering
\includegraphics[scale=0.55]{img/figure11_Pcyg_model.pdf}
\caption{Synthetic spectra of the fiducial model (blue) and a modified fiducial model with a truncated inner disc radius (orange, $R_{\rm co} = 3.8~R_{\rm ISCO}$). The fiducial model in this work (blue) is too highly ionized to produce the UV BAL features reported by \citetalias{Parkinson2020}. By truncating the inner disc radius of the fiducial model, the disc SED is softened and the ionization state of the wind is lowered. The result is that the BEL vs. BAL behaviour reappears for the truncated disc model (orange). Absorption features, such as in C \textsc{iv}, form for sight lines which look into the wind cone (top panel) and emission features form for sight lines with do not look into the wind cone (bottom panel). \textit{Top panel:} an intermediate sight line looking into the wind cone ($60^{\circ}$). \textit{Bottom panel:} a sight line looking below the wind cone ($75^{\circ}$). }
\label{fig: pcygni_grid}
\end{figure*}
In \citetalias{Parkinson2020}, we showed how the BEL vs. BAL dichotomy found in the UV spectra of TDEs \citep[see the top panel in Fig. \ref{fig: dual_p_tde_uv_opt_obs}, and, e.g.,][]{Cenko_2016, Blagorodnova_2019, Hung2019, hung_discovery_2020} could be explained via line formation in an accretion disc wind. In this picture, BALs are preferentially produced for sight lines which look "into" the wind cone. For other orientations, BELs are observed. The fraction of BAL vs. BEL TDEs can then provide an estimate of the covering fraction of outflows in TDEs. Just as importantly, UV lines could be used to infer our orientation with respect to the TDE disc plane.
The models presented in this work all produce the UV transitions commonly see in the spectra of TDEs. However, they do not have the same inclination dependence as the synthetic UV spectra in \citetalias{Parkinson2020}. Specifically, none of the in-wind sight lines give rise to BALs in our models. Instead, BELs are produced for all inclinations.
The absence of BALs in the UV spectra of our models is due to their higher ionization state (relative to the models explored by \citetalias{Parkinson2020}). This is the result of the lower black hole masses we consider, which give rise to harder SEDs. For example, in the fiducial model of \citetalias{Parkinson2020}, the adopted black hole mass is $\text{M}_{\rm BH} = 3 \times 10^{7}~\text{M}_{\odot}$. By contrast, our fiducial model here assumes a black hole mass of $\text{M}_{\rm BH} = 3 \times 10^{6}~\text{M}_{\odot}$. As discussed in Section \ref{sec: model_grid_setup}, this choice, and the range we explore, is based on the data in Fig. \ref{fig: Mbh_parameter_space}.
Given that BALs are observed in the UV spectra of at least some TDEs, it is worth asking whether disc wind models could still produce these features, even at the lower black hole masses we have considered here. Fundamentally, the answer is clearly yes. For example, the ionization state of a wind model can be lowered by stronger clumping (i.e. a lower volume filling factor and denser clumps), by self-shielding \citep[see, e.g.][]{Murray1995, Proga2000, Proga2004}, or by a softer underlying SED.
How significant a change is required in order for our fiducial model to produce BALs? In order to address this question, we have carried out tests with both lower filling factors and modified SEDs. These revealed that a reduction in the filling factor by a factor of 10, to $f_V = 0.01$, is not always sufficient enough to produce BALs. By contrast, only a modest change to the disc SED is required before absorption reliably reappears in the UV spectrum.
{The modified SED test is illustrated in Fig. \ref{fig: pcygni_grid}. For this test, the inner disc radius was simply truncated to $3.8~R_{\rm ISCO}$, which decreases the maximum temperature and softens the SED. This models illustrates that these models can recover the BEL vs. BAL behaviour, even if the absorption is significantly narrower here than found by \citetalias{Parkinson2020}. More specifically, absorption is seen for inclinations looking into the wind cone ($60^{\circ}$), whereas only BELs are observed for sight lines which do not look into the wind cone ($75^{\circ}$ in Fig. \ref{fig: pcygni_grid}).}
Our blackbody $\alpha$-disc SED is clearly a highly idealized description of the spectrum produced by the central engine (see also Section \ref{sec: limitations} below). Since BAL formation in our models depends sensitively on the shape of this spectrum, the presence or absence of these features cannot be used as a "pure" orientation indicator. However, the fundamental point made by \citetalias{Parkinson2020} remains: BALs can only be observed for specific sight lines, typically those looking into the wind cone.
\subsection{Limitations} \label{sec: limitations}
As is the case for any numerical simulation, our modelling involves a number of limiting assumptions and simplifications. First, whilst {\sc python~} now takes into account special relativistic effects, we still neglect general relativistic effects, i.e. whilst {\sc python~} simulates the full random walk of photons, they still travel along straight lines between interaction points. We also do not include any concept of an event horizon within the computational domain. In \citetalias{Parkinson2020}, we partially tested the importance of general relativistic effects by imposing an absorbing spherical boundary at $R_{\text{ISCO}}$. The ionization state and the emergent spectrum were changed only marginally in this test. We have repeated this test for our fiducial model here, with the same results.
Second, all {\sc python~} simulations are time-independent. We have made the implicit assumption that the outflow is in a steady state, which, especially at early times in the evolution of TDEs, may not be true. More specifically, the steady state assumption is valid so long as photon travel times through the flow are short relative to the time-scale over which the luminosity, the SED or the outflow itself can evolve significantly.
In some TDEs, i.e. ASASSN-14ae, AT2018zr and AT2019qiz, the optical spectra have been observed to evolve significantly over time-scales as short as 10-20 days at early times. In our modelled outflows, photons typically take up to 1-3 days to escape from the simulation grid. Although, some photons may stay within the wind cone for almost 231 days. However, since we are not modelling the initial super-Eddington phase, our steady state assumption is likely to be quite reasonable. In line with this, simulations have shown that TDEs are able to reach a quasi-steady state after their initial super-Eddington phase \citep[e.g.][]{Dai2018}.
Third, reiterating another limitation mentioned previously, we have assumed that the accretion disc is in a steady state, and is geometrically thin and optically thick. Specifically, we model the accretion disc SED by assuming a steady state $\alpha$-disc, with constant $\dot{\rm M}_{\rm acc}$, that locally emits as a blackbody \citep{Shakura1973}. Strictly speaking, our use of the standard $\alpha$-disc temperature profile -- which assumes $\dot{\rm M}_{\rm acc}$ as a function of radius -- is not self-consistent, since our discs suffer significant mass loss from the surface. In principle, this can be modelled by adjusting the effective temperature distribution \citep[see, e.g.][]{knigge_effective_1999}. Such models have, in fact, already been considered for TDEs. Specifically, \citet{miller_flows_2015} outlined such an idea, constructed from work by \citet{laor_line-driven_2014} for AGN. In this scheme, \citet{miller_flows_2015} show that mass losses from the disc, via a disc wind, reduces the accretion rate at the inner edge of the disc, decreasing the temperature and regulating the thermal emission. In practice, the net effect is likely to be similar to the inner disc truncation we explored in Section \ref{section: pcygni_discussion}.
Another limitation of our disc model is that the inner regions of TDE discs are dominated by radiation pressure and probably vertically extended. The most important effect of the thin-disc approximation in this context is on the angular distribution of the radiation field. Due to foreshortening and limb darkening, geometrically thin and optically thick discs generate highly anisotropic radiation fields. This degree of anisotropy is likely an overestimate for the vertically extended inner discs in TDEs. However, the structure, evolution and even the stability of such radiation dominated discs is the subject of on-going intense research \citep[e.g.][]{Hirose2009, Jiang_2013, blaes14}. Our choice to use a simple disc model is, then, one borne out of practically. Given that the emission from the disc is reprocessed by the outflow, which isotropizes the radiation field, we do not expect our qualitative results to change significantly.
Fourth and finally, we only model hydrogen and helium self-consistently with full, multi-level model atoms. By contrast, bound-bound transitions in metals are treated using a 2-level atom approximation. Whilst this is perfectly reasonable for resonance lines, it is not reasonable for transitions involving excited and/or meta-stable states. With the exception of C \textsc{iii}] $\lambda$1909 (which we treat in an approximate manner via a modified 2-level atom approximation), our simulations do not include realistic treatments for semi-forbidden transitions such as N \textsc{iii}] $\lambda$1750. Our simplified treatment of metals also currently prevents us from modelling the fluorescence process that gives rise to the Bowen blend between 4630~\AA~and 4660~\AA, commonly seen in observations.
\section{Conclusions} \label{sec: conclusion}
We have shown that the reprocessing of disc radiation by an accretion disc wind can naturally produce the optical line and continuum spectra seen in TDEs. In order to achieve this, we have run a grid of Monte Carlo radiative transfer and ionization simulations to produce synthetic UV and optical spectra of wind- and disc-hosting TDEs.
The disc winds we model are rotating, biconical and clumpy, and they are illuminated by a geometrically thin and optically thick accretion disc. Our model grid covers a realistic range of wind kinematics, black hole masses and accretion states, inspired by observations. Using this grid, we explore how the mass of the black hole, the accretion rate and the wind properties affect the broadband SED and optical line spectra. Our main results are as follows:
\begin{enumerate}
\item The hydrogen and helium emission lines commonly seen in the optical line spectra of TDEs can be produced by an optically thick accretion disc wind that reprocesses the SED of the underlying disc. Reprocessing in such an outflow can also naturally produce the UV line spectra seen in TDEs.
\item Our models produce excellent matches to the optical spectra of TDEs of the TDE-Bowen class, reproducing the correct continuum shapes and emission line properties. However, our models do not produce the pure Balmer spectra seen in the TDE-H spectroscopic class.
\item All of our outflow models are optically thick. There is significant opacity in all of the modelled outflows at the \atomictransition{He}{ii} ($\approx 54$ eV) photo-ionization edge, with typical integrated optical depths of $\tau \sim 10^{7}$ at high and intermediate inclinations. At low inclinations, the optical depth of this edge is still large, but can be of similar magnitude to the electron scattering optical depth. Our models also present significant opacity at the \atomictransition{He}{i} (24.6 eV) and, in some models, the hydrogen Lyman edges. The electron scattering optical depth is similar in all of our models, {with $\tau \simeq 1~-~55$ depending on inclination.}
\item The optical colour temperature of the reprocessed SED is much redder than that of the input accretion disc SED. Changes to the black hole mass, accretion rate and the kinematics of the wind affect the degree of reprocessing and the properties of the line spectra. In general, slower and denser winds result in a greater amount of reprocessing, as well as stronger Balmer and helium recombination features.
\item The main reprocessing mechanism is photo-ionization of \atomictransition{He}{i} and \atomictransition{He}{ii} near the base of the wind. The absorbed luminosity is then re-radiated at longer wavelengths via recombination, or as free-free or line emission. {Based on our simulations, we expect \atomictransition{He}{ii} bound-free interactions to be a critical source of opacity for reprocessing the continuum emission in TDEs. Free-free absorption can also play an important role, but affects the spectra to a lesser degree.}
\item The optical emission lines are often double peaked, since the kinematics of the line-forming regions are dominated by rotation. Some lines exhibit asymmetric red wings due to adiabatic reprocessing. Electron scattering also dramatically affects the line profiles, both by increasing the width of the lines and, for low inclinations, by smoothing out the double peaked shape.
\item Our models produce synthetic UV spectra with the same set of atomic transitions commonly seen in TDEs. However, we find that the winds are often too highly ionized to produce UV absorption features. Truncating the inner radius of the accretion disc at $\simeq 4~R_{\text{ISCO}}$ softens the SED, lowers the ionization state of the wind and recovers UV broad absorption lines for sight lines that look into the wind cone.
\end{enumerate}
\section*{Data availability}
The data underlying this article will be shared on reasonable request to the corresponding author, or can be accessed online at \hyperlink{https://github.com/saultyevil/tde_optical_reprocessing}{github.com/saultyevil/tde\_optical\_reprocessing}.
\section*{Acknowledgements}
{We thank the anonymous reviewer for their careful reading of the manuscript, and their helpful and insightful feedback which significantly improved the quality and usefulness of this work.} Figures were prepared using \texttt{matplotlib} \citep{Hunter2007}. The authors acknowledge the use of the IRIDIS High Performance Computing Facility, and associated support services at the University of Southampton. The authors acknowledge the use of the \texttt{GNU Science Library} \citep{gsl}. EJP would like to acknowledge financial support from the EPSRC Centre for Doctoral Training in Next Generation Computational Modelling grant EP/L015382/1. JHM acknowledges a Herchel Smith Research Fellowship at Cambridge. NSH and CK acknowledge support from the Science and Technology Facilities Council grant ST/V001000/1. KSL acknowledges partial support for this project by NASA through grant number HST-GO-15984, 16058 and 16066 from the Space Telescope Science Institute, which is operated by AURA, Inc., under NASA contract NAS 5-26555.
\bibliographystyle{mnras}
|
\section{Introduction and Main Results}
We set $ \Delta = \partial_{11} + \partial_{22} $ on the annulus $ \Omega = C(1,1/2,0) $ of $ {\mathbb R}^2 $ of radii $ 1 $ and $ 1/2 $ centered at the origin.
\smallskip
We consider the following system:
$$ (P) \left \{ \begin {split}
-\Delta u & = (1+|x-x_0|^{2\beta}) V e^{v} \,\, &\text{in} \,\, & \Omega \subset {\mathbb R}^2, \\
- \Delta v & = W e^{u} \,\, &\text{in} \,\, & \Omega \subset {\mathbb R}^2, \\
u & = 0 \,\, & \text{on} \,\, &\partial \Omega,\\
v & = 0 \,\, & \text{on} \,\, &\partial \Omega.
\end {split}\right.
$$
Here: $ C(1) $ the unit circle and $ C(1/2) $ the circle of radius $ 1/2 $ centered at the origin.
$$ \beta \in (0, 1/2), \,\, x_0 \in C(1/2), $$
and,
$$ u \in W_0^{1,1}(\Omega), \,\, e^u \in L^1({\Omega}) \,\, {\rm and} \,\, 0 < a \leq V \leq b, $$
and,
$$ v \in W_0^{1,1}(\Omega), \,\, e^v \in L^1({\Omega}) \,\, {\rm and} \,\, 0 < c \leq W \leq d. $$
This is a system with regular H\"olderian weight not Lipschitz in $ x_0 $ but have a weak derivative.
This problem $ (P) $ is defined in the sense of the distributions, see [10]. The system was studied by many authors, see [14, 16, 27], also for Riemannian surfaces, see [1-27], where one can find some existence and compactness results. In [9] we have an interior estimate for elliptic equations with exponential nonlinearity.
In this paper we try to prove that we have on all $ \Omega $ the boundedness of the volume of the solutions of $ (P) $ if we add the assumption that $ V $ and $ W $ are uniformly Lipschitz with particular Lispchitz numbers.
Here we have:
\begin{theorem} Assume that $ u $ is a solution of $ (P) $ relative to $ V $ and $ W $ with the following conditions:
$$ x_0 \in C(1/2) \subset \partial \Omega, \,\, \beta \in (0,1/2), $$
and,
$$ 0 < a \leq V \leq b,\,\, ||\nabla V||_{L^{\infty}} \leq A=\dfrac{a}{2(1+2^{2\beta})}, $$
and,
$$ 0 < c \leq W \leq d,\,\, ||\nabla W||_{L^{\infty}} \leq B=\dfrac{c}{2}, $$
we have,
$$ \int_{\Omega} e^u \leq c(a, b, c, d, \beta, x_0, \Omega), \,\, {\rm and}, \,\, \int_{\Omega} e^v \leq c'(a, b, c, d, \beta, x_0, \Omega) $$
\end{theorem}
We have the same result if we consider a system with boundary singularity. On the annulus $ \Omega = C(1,1/2,0) $ of $ {\mathbb R}^2 $ of radii $ 1 $ and $ 1/2 $ centered at the origin.
We consider the following system:
$$ (P_{\beta}) \left \{ \begin {split}
-\Delta u & = |x-x_0|^{2\beta} V e^{v} \,\, &\text{in} \,\, & \Omega \subset {\mathbb R}^2, \\
- \Delta v & = |x-x_0|^{2\beta} W e^{u} \,\, &\text{in} \,\, & \Omega \subset {\mathbb R}^2, \\
u & = 0 \,\, & \text{on} \,\, &\partial \Omega,\\
v & = 0 \,\, & \text{on} \,\, &\partial \Omega.
\end {split}\right.
$$
Here: $ C(1) $ the unit circle and $ C(1/2) $ the circle of radius $ 1/2 $ centered at the origin.
$$ \beta \in (-1/2, +\infty), \,\, x_0 \in C(1/2), $$
and,
$$ u \in W_0^{1,1}(\Omega), \,\, |x-x_0|^{2\beta} e^u \in L^1({\Omega}) \,\, {\rm and} \,\, 0 < a \leq V \leq b, $$
and,
$$ v \in W_0^{1,1}(\Omega), \,\, |x-x_0|^{2\beta} e^v \in L^1({\Omega}) \,\, {\rm and} \,\, 0 < c \leq W \leq d. $$
Here we have:
\begin{theorem} Assume that $ u $ is a solution of $ (P_{\beta}) $ relative to $ V $ and $ W $ with the following conditions:
$$ x_0 \in C(1/2) \subset \partial \Omega, \,\, \beta \in (-1/2, +\infty), $$
and,
$$ 0 < a \leq V \leq b,\,\, ||\nabla V||_{L^{\infty}} \leq A=\dfrac{(\beta+1)a}{2}, $$
and,
$$ 0 < c \leq W \leq d,\,\, ||\nabla W||_{L^{\infty}} \leq B=\dfrac{(\beta+1)c}{2}, $$
we have,
$$ \int_{\Omega} |x-x_0|^{2\beta} e^u \leq c(a, b, c, d, \beta, x_0, \Omega), \,\, {\rm and}, \,\, \int_{\Omega} |x-x_0|^{2\beta} e^v \leq c'(a, b, c, d, \beta, x_0, \Omega) $$
\end{theorem}
We have the same result if we consider a system with boundary singularity. On the annulus $ \Omega = C(1,1/2,0) $ of $ {\mathbb R}^2 $ of radii $ 1 $ and $ 1/2 $ centered at the origin.
We consider the following system:
$$ (P_{\beta}) \left \{ \begin {split}
-\Delta u & = |x-x_0|^{2\beta} V e^{v} \,\, &\text{in} \,\, & \Omega \subset {\mathbb R}^2, \\
- \Delta v & = W e^{u} \,\, &\text{in} \,\, & \Omega \subset {\mathbb R}^2, \\
u & = 0 \,\, & \text{on} \,\, &\partial \Omega,\\
v & = 0 \,\, & \text{on} \,\, &\partial \Omega.
\end {split}\right.
$$
Here: $ C(1) $ the unit circle and $ C(1/2) $ the circle of radius $ 1/2 $ centered at the origin.
$$ \beta \in (-1/2, +\infty), \,\, x_0 \in C(1/2), $$
and,
$$ u \in W_0^{1,1}(\Omega), \,\, e^u \in L^1({\Omega}) \,\, {\rm and} \,\, 0 < a \leq V \leq b, $$
and,
$$ v \in W_0^{1,1}(\Omega), \,\, |x-x_0|^{2\beta} e^v \in L^1({\Omega}) \,\, {\rm and} \,\, 0 < c \leq W \leq d. $$
Here we have:
\begin{theorem} Assume that $ u $ is a solution of $ (P_{\beta}) $ relative to $ V $ and $ W $ with the following conditions:
$$ x_0 \in C(1/2) \subset \partial \Omega, \,\, \beta \in (-1/2, +\infty), $$
and,
$$ 0 < a \leq V \leq b,\,\, ||\nabla V||_{L^{\infty}} \leq A=\dfrac{(\beta+1)a}{2}, $$
and,
$$ 0 < c \leq W \leq d,\,\, ||\nabla W||_{L^{\infty}} \leq B=\dfrac{c}{2}, $$
we have,
$$ \int_{\Omega} e^u \leq c(a, b, c, d, \beta, x_0, \Omega), \,\, {\rm and}, \,\, \int_{\Omega} |x-x_0|^{2\beta} e^v \leq c'(a, b, c, d, \beta, x_0, \Omega) $$
\end{theorem}
\section{Proof of the Theorems:}
\smallskip
{\it Proof of the theorem 1.1:}
\smallskip
By corollary 1 of the paper of Brezis-Merle, we have: $ e^{ku}, e^{kv} \in L^1(\Omega) $ for all $ k >2 $ and the elliptic estimates and the Sobolev embedding imply that: $ u,v \in W^{2,k}(\Omega) \cap C^{1,\epsilon}(\bar \Omega), \epsilon >0 $. By the maximum principle $ u,v \geq 0 $.
\smallskip
Step 1: We use the first eigenvalue and the first eigenfunction with Dirichlet boundary condition to bound the volumes locally uniformly. Thus the solutions are locally uniformly bounded by Brezis-Merle result. The solutions $ u,v >0 $ are locally uniformly bounded in $ C^{1,\epsilon}(\Omega) $ for $ \epsilon $ small.
\smallskip
By Cauchy-Schwarz inequality, applied to $ u \sqrt {\phi_1} $ and $ {\sqrt \phi_1} $ for the following equality:
$$ \int_{\Omega} (1+|x-x_0|^{2\beta}) Ve^v \phi_1dx =\lambda_1 \int_{\Omega} u\phi_1 dx \leq c_1 {(\int_{\Omega} u^2 \phi_1)}^{1/2} \leq c_2 {(\int_{\Omega} W e^u \phi_1)}^{1/2}, $$
and for $ v $,
$$ \int_{\Omega} We^u \phi_1dx =\lambda_1 \int_{\Omega} v\phi_1 dx \leq c_3 {(\int_{\Omega} v^2 \phi_1)}^{1/2} \leq c_4 {(\int_{\Omega} (1+|x-x_0|^{2\beta })V e^v \phi_1)}^{1/2} , $$
Thus,
$$ (\int_{\Omega} (1+|x-x_0|^{2\beta})Ve^v \phi_1dx)^{3/4} \leq c_5, $$
and,
$$ (\int_{\Omega} We^u \phi_1dx)^{3/4} \leq c_6, $$
We can use Brezis-Merle arguments to prove that for all subdomain $ K $ of $ \Omega $: the two integrals, $ \int_{K} (1+|x-x_0|^{2\beta})Ve^v dx, $ and, $ \int_{K} We^u dx $ converge to nonegative measues $ \mu_1, \mu_2 $ without nonregular points.
\smallskip
By contradiction, suppose that $ \max_K u_i \to +\infty $ and $ \max_K v_i \to +\infty $.
\smallskip
Since $ (1+|x-x_0|^{2\beta})V_ie^{v_i} $ and $ W_ie^{u_i} $ are bounded in $ L^1(K) $, we can extract from those two sequences two subsequences which converge to two nonegative measures $ \mu_1 $ and $ \mu_2 $. (This procedure is similar to the procedure of Brezis-Merle, we apply corollary 4 of Brezis-Merle paper, see [9]).
If $ \mu_1(y_0) < 4 \pi $, by a Brezis-Merle estimate for the first equation, we have $ e^{u_i} \in L^{1+\epsilon} $ uniformly around $ y_0 $, by the elliptic estimates, for the second equation, we have $ v_i \in W^{2, 1+\epsilon} \subset L^{\infty} $ uniformly around $ y_0 $, and , returning to the first equation, we have $ u_i \in L^{\infty} $ uniformly around $ y_0 $.
If $ \mu_2(y_0) < 4 \pi $, then $ u_i $ and $ v_i $ are also locally uniformly bounded around $ y_0 $.
Thus, we take a look to the case when, $ \mu_1(y_0) \geq 4 \pi $ and $ \mu_2(y_0) \geq 4 \pi $. By our hypothesis, those points $ y_0 $ are finite.
We will see that inside $ K $ no such points exist. By contradiction, assume that, we have $ \mu_1(y_0) \geq 4 \pi $. Let us consider a ball $ B_R(y_0) $ which contain only $ y_0 $ as nonregular point. Thus, on $ \partial B_R(y_0) $, the two sequence $ u_i $ and $ v_i $ are uniformly bounded. Let us consider:
$$ \left \{ \begin {split}
-\Delta z_i & = (1+|x-x_0|^{2\beta})V_i e^{v_i} \,\, &\text{in} \,\, & B_R(y_0) \subset {\mathbb R}^2, \\
z_i & = 0 \,\, & \text{in} \,\, &\partial B_R(y_0).
\end {split}\right.
$$
By the maximum principle:
$$ z_i \leq u_i, $$
and $ z_i \to z $ almost everywhere on this ball, and thus,
$$ \int e^{z_i} \leq \int e^{u_i} \leq C, $$
and,
$$ \int e^z \leq C.$$
but, $ z $ is a solution in $ W_0^{1,q}(B_R(y_0)) $, $ 1\leq q <2 $, of the following equation:
$$ \left \{ \begin {split}
-\Delta z & = \mu_1\,\, &\text{in} \,\, & B_R(y_0) \subset {\mathbb R}^2, \\
z & = 0 \,\, & \text{in} \,\, &\partial B_R(y_0).
\end {split}\right.
$$
with, $ \mu_1 \geq 4 \pi $ and thus, $ \mu_1 \geq 4\pi \delta_{y_0} $ and then, by the maximum principle in $ W_0^{1,1}(B_R(y_0)) $:
$$ z \geq -2 \log |x-y_0|+ C $$
thus,
$$ \int e^z = + \infty, $$
which is a contradiction. Thus, there is no nonregular points inside $ K $. Thus $(u_i)$ and $(v_i)$ are uniformly bounded in $ K $ and also in $ C^{1,\epsilon}(K) $ by the elliptic estimates, for all $ K \subset \subset \Omega $.
\smallskip
Step 2: Let's consider $ C_1=C(1,3/4,0) $ and $ C_2=C(3/4,1/2,0) $ the two annulus wich are the neighborhood of the two components of the boundary.
We multiply the equation by $ (x-x_0)\cdot \nabla u $ on $ C_1 $ and $ C_2 $ and use the Pohozaev-Rellich identity and Stokes theorem, see [26]. We use the fact that $ u $ and $ v $ are uniformly bounded around the circle $ C(3/4) $. We obtain:
1) We have on $ C_1 $:
$$ \int_{C_1} (\Delta u)[(x-x_0)\cdot \nabla v] dx=\int_{C_1} -[(1+|x-x_0|^{2\beta}) V (x-x_0)\cdot \nabla (e^v)]dx, $$
and,
$$ \int_{C_1} (\Delta v)[(x-x_0)\cdot \nabla u] dx=\int_{C_1} -[W (x-x_0)\cdot \nabla (e^u)]dx, $$
Thus, by integration by parts,
$$ \int_{C_1} (\Delta u)[(x-x_0)\cdot \nabla v]+ (\Delta v)[(x-x_0)\cdot \nabla u] dx = $$
$$
\int_{\partial C_1} [(x-x_0)\cdot \nabla u ] (\nabla v \cdot \nu)+[(x-x_0)\cdot \nabla v ] (\nabla u \cdot \nu) -[(x-x_0)\cdot \nu] (\nabla u \cdot \nabla v)= $$
$$ = \int_{C_1} (2+2(\beta + 1)|x-x_0|^{2\beta}) Ve^v dx + \int_{C_1} (1+|x-x_0|^{2\beta}) (x-x_0)\cdot \nabla V e^v dx+ $$
$$ + \int_{C_1} 2We^u dx + \int_{C_1} (x-x_0)\cdot \nabla W e^u dx+ $$
$$ - \int_{\partial C_1} (1+|x-x_0|^{2\beta}) [(x-x_0)\cdot \nu ] V e^v d\sigma +$$
$$ - \int_{\partial C_1} [(x-x_0)\cdot \nu ] W e^u d\sigma $$
We can write, ($ u= 0 $ on $ C(1) $):
$$ \int_{C(1)} [(x-x_0)\cdot \nu ] (\partial_{\nu} u)(\partial_{\nu} v) d\sigma + O(1)= $$
$$ \leq k_1 \int_{C_1} (1+|x-x_0|^{2\beta}) Ve^v dx +k_2 \int_{C_1} We^u dx+ O(1) = $$
$$ = k_1\int_{C(1)} \partial_{\nu} u d\sigma + k_2\int_{C(1)} \partial_{\nu} v d\sigma +O(1), $$
with $ k_1,k_2>0 $ not depends on $ u $.
Be cause $ \nu = x, ||x||=1, ||x_0||=1/2 $ and then by Cauchy-Schwarz, $ (x-x_0)\cdot x=||x||^2-x_0 \cdot x \geq 1/2 $, we obtain:
\be 0 < \int_{C(1)} (\partial_{\nu} u ) (\partial_{\nu} v) d\sigma \leq k_1\int_{C(1)} \partial_{\nu} u d\sigma + k_2\int_{C(1)} \partial_{\nu} v d\sigma +O(1), \ee
Let $ \epsilon =\inf (1, \frac{a}{d}) $, then: $ -\Delta (\epsilon v)=\epsilon W e^u \leq a e^u \leq (1+|x-x_0|^{2\beta})V e^u $, and $ \epsilon \leq 1, v\geq 0 \Rightarrow (1+|x-x_0|^{2\beta} ) V e^v \geq (1+|x-x_0|^{2\beta})V e^{\epsilon v} $ and then: $ -\Delta u \geq (1+|x-x_0|^{2\beta})V e^{\epsilon v} $. (We can remove $ (1+|x-x_0|^{2\beta})$ ) We obtain:
$$ -\Delta (u-\epsilon v)= -\Delta u+ \Delta (\epsilon v) \geq (1+|x-x_0|^{2\beta})V(e^{\epsilon v}-e^u), $$
Thus,
$$ -\Delta (u-\epsilon v) + (1+|x-x_0|^{2\beta})V(e^u-e^{\epsilon v}) \geq 0, $$
(For the theorem 1.2, we have the weight $ |x-x_0|^{2\beta} $, in the two equations, we can compare $ \epsilon v $ and $ u $).
We return to the proof of theorem 1.1. Let's consider the fonction:
$$ c(x)=(1+|x-x_0|^{2\beta})V\dfrac{(e^u-e^{\epsilon v})}{u-\epsilon v}, \, {\rm if} \, u\not = \epsilon v, \, and, \, c(x)=(1+|x-x_0|^{2\beta})Ve^u,\, {\rm if} \, u=\epsilon v. $$
The function $ c \geq 0 $ is $ C^{\beta}(\bar \Omega) $, and $ -c \leq 0 $.(For the theorem 1.2, if $ -1/2< \beta <0$, $ c $ is $ C^{-\beta}(\Omega)\cap L^2(\Omega) $, but this is sufficient to apply the weak maximum principle, see the book of Gilbarg-Trudinger).
We can write:
$$ \Delta (\epsilon v-u)- c(x) (\epsilon v-u) \geq 0, \, {\rm in } \, \Omega, \, and, \, u-\epsilon v=0 \, {\rm on} \,\, \partial \Omega, $$
The operator $ L= \Delta + (-c) =\Delta + \tilde c $ satisfies the maximum principle because $ \tilde c =-c \leq 0 $, we obtain:
by the weak maximum principle for $ \epsilon v-u \in C^2(\Omega) \cap C^1(\bar \Omega) $, see the book of Gilbarg-Trudinger, we obtain (and for the outer normal):
$$ \epsilon v - u \leq 0, \, {\rm and \, then }, \, \partial_{\nu} (\epsilon v-u) \geq 0, $$
Thus, for the inner normal, we have:
\be \partial_{\nu} u \geq \epsilon \cdot \partial_{\nu} v >0 \,\, {\rm on } \,\, \partial \Omega. \ee
By the same argument, if we set $ \bar \epsilon= \inf (1, \frac{c}{(1+2^{2\beta})b}) $, we obtain, for the inner normal:
\be \partial_{\nu} v \geq \bar \epsilon \cdot \partial_{\nu} u >0 \,\, {\rm on } \,\, \partial \Omega. \ee
{\bf Remark:} For theorems 1.2 and 1.3: we can remove the fact that we have the weight $ |x-x_0|^{2\beta} $ in the two equations of the system, we can assume that we have one equation with weight and the other equation without weight for example. Indeed, remark that $ x_0 \in C(1/2) $ and here we consider $ C(1) $, we can apply the weak maximum principle on $ C_1 $ after choosing $ 1> \epsilon >0 $ and $ 1 > \bar \epsilon >0 $ independent of $ u $ and $ v $, such that: $ \epsilon v - u \leq 0 $ on $ C(3/4) $ and $ \bar \epsilon u-v \leq 0 $ on $ C(3/4) $, because we have the uniform interior estimate around $ C(3/4) $ and the maximum principle (by contradiction if we assume there are $ (t_i) \in C(3/4), t_i \to t_0 $ and $ u_i, v_i $ such that $ u_i(t_i) \to 0 $, $ (u_i), (v_i) $ converge to $ u_0,v_0 $ on $ B(t_0, r_0), r_0 >0 $ with $ u_0(t_0)=0 $, but $ -\Delta u_0 >0 $ and the maximum principle applied to $ -u_0 $ imply that $ u_0 >0 $ around $ t_0 $. This fact imply that $ u, v $ have positive lower bounds on $ C(3/4)$ and we can choose $ \epsilon, \bar \epsilon $ independent of $ u, v $ on $C(3/4)$. Finally we can apply the weak maximum principle on $ C_1 $).
Thus, we use $ (1), (2), (3) $, we obtain the same result as for one equation;
$$ \int_{C(1)} (\partial_{\nu} v)^2 d\sigma \leq C_3 \int_{C(1)} \partial_{\nu} v d\sigma + C_4, $$
with, $ C_3, C_4 >0 $ and not dpend on $ u $ and $ v $, and,
$$ \int_{C(1)} (\partial_{\nu} u)^2 d\sigma \leq C_5 \int_{C(1)} \partial_{\nu} u d\sigma + C_6, $$
with, $ C_5, C_6 >0 $ and not dpend on $ u $ and $ v $, and,
Applying the Cauchy-Schwarz inequality, we have:
$$ \int_{C(1)} (\partial_{\nu} u)^2 d\sigma =O(1), \,\, \int_{C(1)} (\partial_{\nu} v)^2 d\sigma = O(1) $$
and thus,
$$ \int_{C(1)} \partial_{\nu} u d\sigma =O(1), \,\, \int_{C(1)} \partial_{\nu} v d\sigma = O(1) $$
2) We have on $ C_2 $, we use again, the uniform boundedness of $ u $ and $ v $ in $ C^1 $ norm around $ C(3/4) $:
$$ \int_{C_2} (\Delta u)[(x-x_0)\cdot \nabla v] dx=\int_{C_2} -[(1+|x-x_0|^{2\beta}) V (x-x_0)\cdot \nabla (e^v)]dx, $$
and,
$$ \int_{C_2} (\Delta v)[(x-x_0)\cdot \nabla u] dx=\int_{C_2} -[W (x-x_0)\cdot \nabla (e^u)]dx, $$
Thus, by integration by parts,
$$ \int_{C_2} (\Delta u)[(x-x_0)\cdot \nabla v]+ (\Delta v)[(x-x_0)\cdot \nabla u] dx = $$
$$
\int_{\partial C_2} [(x-x_0)\cdot \nabla u ] (\nabla v \cdot \nu)+[(x-x_0)\cdot \nabla v ] (\nabla u \cdot \nu) -[(x-x_0)\cdot \nu] (\nabla u \cdot \nabla v)= $$
$$ = \int_{C_2} (2+2(\beta + 1)|x-x_0|^{2\beta}) Ve^v dx + \int_{C_2} (1+|x-x_0|^{2\beta}) (x-x_0)\cdot \nabla V e^v dx+ $$
$$ + \int_{C_2} 2We^u dx + \int_{C_2} (x-x_0)\cdot \nabla W e^u dx+ $$
$$ - \int_{\partial C_2} (1+|x-x_0|^{2\beta}) [(x-x_0)\cdot \nu ] V e^v d\sigma +$$
$$ - \int_{\partial C_2} [(x-x_0)\cdot \nu ] W e^u d\sigma $$
But here, $ \nu=-2x$ and $ (x-x_0)\cdot \nu= -2(x-x_0)\cdot x \leq 0 $ and thus:
$$ \int_{C(1/2)} (x-x_0)\cdot \nu (\partial_{\nu} u ) (\partial_{\nu} v) d\sigma + O(1) = $$
$$ = \int_{C_2} (2+2(\beta + 1)|x-x_0|^{2\beta}) Ve^v dx + \int_{C_2} (1+|x-x_0|^{2\beta}) (x-x_0)\cdot \nabla V e^v dx+ $$
$$ + \int_{C_2} 2We^u dx + \int_{C_2} (x-x_0)\cdot \nabla W e^u dx+ $$
$$ - \int_{\partial C_2} (1+|x-x_0|^{2\beta}) [(x-x_0)\cdot \nu ] V e^v d\sigma +$$
$$ - \int_{\partial C_2} [(x-x_0)\cdot \nu ] W e^u d\sigma $$
thus:
$$ \int_{C_2} (2+2(\beta +1)|x-x_0|^{2\beta}) Ve^v dx + \int_{C_2} (1+|x-x_0|^{2\beta}) (x-x_0)\cdot \nabla V e^v dx + $$
$$ + \int_{C_2} 2We^u dx + \int_{C_1} (x-x_0)\cdot \nabla W e^u dx+ $$
$$ + \int_{C(1/2)} \frac{1}{2}[-(x-x_0)\cdot \nu ] (\partial_{\nu} u)(\partial_{\nu} v) d\sigma = O(1) $$
If we choose:
\be \dfrac{|(x-x_0)\cdot \nabla V|}{V} \leq \dfrac{1}{2} \inf_{x \in \bar \Omega } \dfrac{(2+2(\beta+1)|x-x_0|^{2\beta})}{1+|x-x_0|^{2\beta}}, \ee
and,
\be \dfrac{|(x-x_0)\cdot \nabla W|}{W} \leq \dfrac{1}{2} \cdot 2, \ee
(For the theorem 1.2, we choose: $ \dfrac{|(x-x_0)\cdot \nabla V|}{V} \leq \dfrac{1}{2} \inf_{x \in \Omega } \dfrac{(2(\beta+1)|x-x_0|^{2\beta})}{|x-x_0|^{2\beta}}=\beta+1,$ and also for $ W $, we choose $ \dfrac{|(x-x_0)\cdot \nabla W|}{W} \leq \dfrac{1}{2} \inf_{x \in \Omega } \dfrac{(2(\beta+1)|x-x_0|^{2\beta})}{|x-x_0|^{2\beta}} =(\beta+1) $).
We obtain:
$$ \int_{C_2} (2+2(\beta +1)|x-x_0|^{2\beta}) Ve^v dx + \int_{C_2} (1+|x-x_0|^{2\beta}) (x-x_0)\cdot \nabla V e^v dx \geq $$
$$ \geq \frac{1}{2} \int_{C_2} (2+2(\beta +1)|x-x_0|^{2\beta}) Ve^v dx \geq 0 $$
and,
$$ \int_{C_2} 2 W e^u dx + \int_{C_2} (x-x_0)\cdot \nabla W e^u dx \geq $$
$$ \geq \frac{1}{2} \int_{C_2} 2 W e^u dx \geq 0 $$
thus,
$$ \int_{C_2} [(2+2(\beta +1)|x-x_0|^{2\beta}) Ve^v+ 2W e^u] dx =O(1), $$
we obtain:
$$ \int_{C_2} (1+|x-x_0|^{2\beta}) Ve^v dx =O(1), $$
and,
$$ \int_{C_2} W e^u dx =O(1), $$
and thus, if $ A\leq \dfrac{a}{2(1+2^{2\beta})} $ and $ B \leq \dfrac{c}{2} $, we obtain :
$$ \int_{C(1/2)} \partial_{\nu} u d\sigma = O(1), \,\, \int_{C(1/2)} \partial_{\nu} v d\sigma = O(1), $$
Thus, if we use 1) and 2), we obtain: if $ A\leq \dfrac{a}{2(1+2^{2\beta})} $ and $ B \leq \dfrac{c}{2} $:
$$\int_{C(1)} (\partial_{\nu} u) d\sigma =O(1),\,\, {\rm and} \,\, \int_{C(1/2)} \partial_{\nu} u d\sigma = O(1), $$
and,
$$\int_{C(1)} (\partial_{\nu} v) d\sigma =O(1),\,\, {\rm and} \,\, \int_{C(1/2)} \partial_{\nu} v d\sigma = O(1), $$
and, thus:
$$ \int_{\Omega } [(1+|x-x_0|^{2\beta}) Ve^v] dx=\int_{\partial \Omega} (\partial_{\nu} u) d\sigma = O(1). $$
and,
$$ \int_{\Omega } We^u dx=\int_{\partial \Omega} (\partial_{\nu} v) d\sigma =O(1). $$
For the theorem 1.2. we obtain the same result if $ A\leq \dfrac{(\beta+1) a}{2} $ and $ B \leq \dfrac{(\beta+1)c}{2} $.
We have the same result for theorem 1.3.
|
\section{Motivation\label{sec:motivations}}
The Ly-$\alpha$\ and Ly-$\beta$\ lines of neutral hydrogen respectively at 121.6~nm and 102.5~nm and the Ly-$\alpha$\ line of singly ionized helium at 30.4~nm are among the brightest lines of the ultraviolet (UV) spectrum of the Sun, and their study is of importance in many areas of solar physics. These lines are observed by several instruments on board the Solar Orbiter mission \citep{Muller2020}. In this paper, we derive empirical relationships between these lines with, as described below, three main applications in mind: (i) the modeling of the resonance scattering emission of these lines in the corona, taking the nonuniformity of the chromospheric source into account, (ii) the observational constraint of nonlocal thermodynamic equilibrium (NLTE) models of chromospheric structures, and (iii) the modeling of the UV and extreme UV (EUV) irradiances.
The bright coronal Ly-$\alpha$\ line is mainly formed by the resonant scattering of the underlying chromospheric radiation by residual neutral hydrogen \citep{Gabriel1971}. The efficiency of the resonant scattering process depends on how much the chromospheric source profile and coronal scattering profile overlap. In a static atmosphere, the central wavelength of these two profiles lines up perfectly, whereas in a region with solar wind flow the scattering profile is Doppler-shifted relative to the disk profile, which results in a less efficient scattering and, therefore, in a reduced (dimmed) line intensity. Doppler dimming observations at Ly-$\alpha$\ and in other lines that have a significant resonantly scattered component, such as the \ion{O}{vi} doublet at 103.2 and 103.7 nm., have been extensively used to estimate the outflow velocities of the coronal plasma \citep[e.g.,][]{Kohl1997, Antonucci2000}. The interpretation of the Ly-$\alpha$\ resonance scattering measurements relies on iterative forward modeling and thus requires independent knowledge of the chromospheric intensity, which is often taken from in-ecliptic irradiance measurements. However, \citet{Auchere2005b} has shown that using disk-integrated values and thus not taking into account the anisotropy of the illumination of the corona resulting from the distribution of bright features (e.g., active regions) and dark features (e.g., coronal holes) in the chromosphere can lead to significant systematic over- or underestimations of the coronal intensity, especially in polar regions. Recently, \cite{Dolei2018, Doleil2019} estimated that this translates to a $50\ \mathrm{km.s}^{-1}$ error on the outflow velocities derived from Doppler dimming methods. The same effect applies to all resonantly scattered coronal lines (i.e., the Ly-$\alpha$\ line of \ion{He}{II} at 30.378~nm). Calibrated Ly-$\alpha$\ disk images are thus in principle necessary for interpreting the coronal observations\footnote{This is also true of the modeling of the Ly-$\alpha$\ radiation back-scattered by interplanetary neutral hydrogen \citep{Cook1981}.}. While 30.4~nm disk images are routinely available from narrowband telescopes such as the Extreme-ultraviolet Imaging Telescope \citep[EIT;][]{Delaboudiniere1995}, the Extreme Ultra Violet Imager \citep[EUVI;][]{Wuelser2004}, or the Atmospheric Imaging Assembly \citep[AIA;][]{Boerner2012, Lemen2012}, full-disk Ly-$\alpha$\ observations are rare \citep[e.g.,][]{Bonnet1980}. \cite{Auchere2005b} thus introduced the possibility of using scaled 30.4~nm images as proxies for Ly-$\alpha$\ data, which requires the derivation of an empirical relationship between the intensities of the two lines.
The chromospheric Ly-$\alpha$\ line of hydrogen was first observed by \cite{Purcell1960} in a 1959 rocket flight and since then has been the subject of many more observations (see below), along with theoretical and modeling works. However, the profile of the line (very often self-reversed) has led to questions concerning both its characterization and interpretation. The line is optically thick and formed in non-local thermodynamic equilibrium \citep[NLTE;][]{Jefferies1961, Morton1961}, a situation that complicates the spectroscopic diagnostic. Nonetheless, the comparison of model outputs and observed line profiles of Ly-$\alpha$ , Ly-$\beta$\ \citep[which is also self-reversed;][]{Tousey1963}, and other strong UV lines was at the root of one-dimensional (1D) models such as the Vernazza, Avrett, Loeser \citep[VAL;][]{Vernazza1981} and later the Fontenla, Avrett, Loeser~\citep[FAL;][]{Fontenla1990} models. Observations from the Orbiting Solar Observatory 8 (OSO 8) spacecraft and from sounding rockets \citep{Bonnet1978, Gouttebroze1978, Bonnet1981, Bonnet1982} of the Ly-$\alpha$\ and Ly-$\beta$\ lines have led to the modeling of the chromosphere, active regions \citep[e.g.,][]{Lemaire1981}, sunspots \citep{Kneer1981, Lites1982}, and prominences \citep{Vial1982}. In our present theoretical effort, the Ly-$\beta$\ line was also used because it was simultaneously observed with the Ly-$\alpha$\ line and provided a complementary diagnostic.
$\mathrm{H}^0$ and $\mathrm{He}^+$ have similar atomic structures, and NLTE modeling has shown that their Ly-$\alpha$\ photons come from close-by but different layers of the solar atmosphere, although the details of the processes themselves may differ from one line to the other. Ly-$\alpha$\ and Ly-$\beta$\ lines of \ion{H}{I} are thermally produced (electronic collisions) in the chromosphere up to 30~000~K, and the Ly-$\alpha$ \ of \ion{He}{II} is formed above 30~000~K. On the contrary, in prominences, the Ly-$\alpha$\ photons are mainly radiatively produced by resonance scattering. Radiation also plays an important role in the formation of the line of \ion{He}{II} since the existence of the $\mathrm{He}^+$ ion is determined by the incident EUV coronal radiation \citep[e.g.,][]{Andretta2003}. Moreover, the three lines play a very important role in the radiative losses of the various solar features \citep[see the VAL and FAL models for the
chromosphere or the case of radiative equilibrium for prominences in ][]{Heinzel2012}. This means that simultaneous observations of the three abovementioned lines are (or would be) very useful for modeling the various solar structures. Despite the difficulties in recording the Ly-$\alpha$\ line, nearly simultaneous observations of the Ly-$\alpha$\ and Ly-$\beta$\ lines of \ion{H}{I} were possible with the Solar Ultraviolet Measurement of Emitted Radiation \citep[SUMER;][]{Wilhelm1995} spectrograph, and~\cite{Lemaire2012} were able to constrain the ratio between the intensities of the two lines in various solar structures. A correlation between the intensities of the Ly-$\alpha$\ lines of \ion{H}{I} and \ion{He}{II} has been obtained by \citet{Auchere2005b}, but, to the best of our knowledge, no relation has yet been established between the Ly-$\beta$\ line of \ion{H}{I} and the Ly-$\alpha$\ line of \ion{He}{II}.
Irradiance modeling is closely related to the two previous objectives. Semiempirical models of the solar spectral radiance adjust the variation in temperature with height in the solar atmosphere to obtain optimum agreement between calculated and observed continuum intensities, line intensities, and line profiles \citep{Avrett2008}. Major constraints to these models come from the observed intensities and profiles of the Ly-$\alpha$\ and Ly-$\beta$\ lines (and other strong UV lines) and in fact led to the development of the seminal 1D VAL and FAL models. For instance, the observations of the self-reversal of the Ly-$\beta$\ line \citep{Tousey1963} constrained the models with the existence of a temperature plateau around 20 000~K.
Using this approach, the reconstruction of the spectral irradiance is obtained by a linear combination of a family of models in which the physical parameters are adjusted to match observations for a number of solar structures: coronal holes, quiet Sun, plage, sunspot umbra, penumbra, etc. It is thus necessary to use not only irradiance values, but also resolved measurements of the lines intensities.
In Sect.~\ref{sec:lya_hi_vs_heii} of the paper, we derive a new relationship between the intensities of the Ly-$\alpha$\ lines of \ion{H}{I} and \ion{He}{II} using two different sets of spatially resolved data. In Sect.~\ref{sec:lyb_hi_vs_heii} we derive a relationship between the intensities of the Ly-$\beta$\ line of \ion{H}{I} and the Ly-$\alpha$\ line of \ion{He}{II}. In Sect.~\ref{sec:lyb_hi_vs_lya_hi} we derive a relationship between the intensities of the Ly-$\beta$\ and Ly-$\alpha$\ lines of \ion{H}{I}. We summarize our findings in Sect.~\ref{sec:conclusions}.
\section{H\textsc{I} 121.6~nm versus He\textsc{II} 30.4~nm\label{sec:lya_hi_vs_heii}}
\cite{Auchere2005b} derived an empirical relation between the \ion{H}{I} 121.6~nm and \ion{He}{II} 30.4~nm irradiances using data from the EIT telescope, the Solar Stellar Irradiance Comparison Experiment \citep[SOLSTICE;][]{Rottman1993}, and the Solar EUV Experiment \citep[SEE;][]{Woods2000}. It matched spatially resolved but episodic observations from Skylab \citep{Vernazza1978}. In the present study, we revisit this relationship using two new sets of spatially resolved observations that cover all types of solar structures and different activity levels. The first set is full-disk observations made on November 2, 1998, by EIT and the Multiple XUV Imager \citep[MXUVI;][]{Auchere1999}. The second set corresponds to disk-center and limb observations made on September 3, 2015, by the AIA and the Chromospheric Lyman-Alpha Spectro Polarimeter \citep[CLASP;][]{Kano2012, Kobayashi2012, Kano2017}.
\subsection{MXUVI and EIT data}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{eit304-mxuvilya-images_2.eps}
\caption{Quasi-simultaneous images of the Sun taken on November 2, 1998, in \ion{H}{I} 121.6~nm (left) and \ion{He}{II} 30.4~nm (center, right). The Lyman-$\alpha$ image was obtained by the rocket-borne MXUVI instrument at 18:20 UT. The 30.4~nm image was recorded by EIT at 18:26 UT. The EIT image was degraded to match the resolution of the MXUVI. Solar north is up. The intensity scales are logarithmic in both cases, and the contrast is the same. There is a remarkable morphological similarity for all types of structures, but, as is also visible in Fig.~\ref{fig:eit-mxuvi-correl}, they are less contrasted at Lyman~$\alpha$. In the right image, we overlaid different contours corresponding to different solar regions: active regions (red), filaments (green), the coronal hole (yellow), and prominences (purple).}
\label{eit-mxuvi}
\end{figure*}
The second flight of the MXUVI sounding-rocket-borne telescope occurred on November 2, 1998, at 18:20 UT. The Ly-$\alpha$\ channel of the MXUVI uses a 256 $\times$ 256 detector with a 10 arcsec/pixel sampling. Isolation of the Ly-$\alpha$\ line is obtained using Al/MgF2 filters and selective mirror coating for a resulting passband of 10~nm. All the data acquired during the flight were co-registered and added into a single image. The flat field and dark level were corrected according to \citet{Auchere1999}. Radiometric calibration was obtained by scaling the resulting image to the simultaneous composite Ly-$\alpha$\ irradiance, which for this date includes data from SOLSTICE \citep{Rottman1993, Woods1993}. The composite passband has a width of 1.0~nm.
The EIT telescope has been providing Sun full-disk observations since 1996. We used the November 2, 1998, image taken at 18:26 UT, which is the 30.4~nm image closest in time to the MXUVI second flight and corresponds to a specific EIT campaign for this flight. The 30.4~nm channel uses a 1024 $\times$ 1024 detector with a 2.627 arcsec/pixel sampling \citep{Auchere2000a, Auchere2004}. The flat field, dark noise, and degradation were corrected with the \texttt{eit\_prep} procedure. Radiometric calibration was obtained by using the \texttt{eit\_parms} function. The instrument passband has a width of 7.3~nm.
\subsection{CLASP and AIA data}
The first flight of the CLASP sounding-rocket-borne spectro-polarimeter occurred on September 3, 2015, between 17:03 UT and 17:08 UT. The Ly-$\alpha$\ images used are provided by the CLASP Slit-Jaw camera (CLASP-SJ). The CLASP-SJ uses a 528 $\times$ 536 pixel sensor with a 1.03 arcsec/pixel sampling. The instrument passband has a width of 3.5~nm. The flat field and dark level were corrected. The radiometric calibration was obtained by scaling the resulting image to the simultaneous composite Ly-$\alpha$\ irradiance, which for this date includes data from SOLSTICE\citep{McClintock2005a, McClintock2005b, Snow2005}. The composite passband has a width of 1.0~nm. During this flight, images from the Sun center and limb were taken.
The AIA telescope has been providing full-disk observations at high resolution since 2010. The 30.4~nm channel of AIA uses a 4096 $\times$ 4096 detector with a 0.63 arcsec/pixel sampling. The 12-second cadence permitted 25 AIA images to be obtained during the CLASP flight. The radiometric calibration was obtained by using the \texttt{aia\_get\_response} function. The instrument passband has a width of 4.5~nm.
\subsection{Data co-registration\label{sec:co-registration}}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{CLASP_AIA_cen_rot.eps}
\caption{Sun center \ion{H}{I} 121.6~nm observations from September 3, 2015, from the CLASP-SJ (left) and simultaneous \ion{He}{II} 30.4~nm observations from AIA degraded to match the CLASP-SJ resolution (right). The black band in the center of the CLASP-SJ image is the entrance slit of the spectrometer. Solar north is up.}
\label{aia-clasp-obs-cen}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{CLASP_AIA_limb_rot.eps}
\caption{Same as Fig.~\ref{aia-clasp-obs-cen}, but for the CLASP-SJ limb pointing.}
\label{aia-clasp-obs-limb}
\end{figure*}
We co-aligned the \ion{He}{II} 30.4~nm (EIT or AIA) and \ion{H}{I} 121.6~nm (MXUVI or CLASP-SJ) images to compare and find a relation between the intensities. We used a reduced $\chi^{2}$ cross-correlation method in translation and rotation. The spatial scaling was assumed to be known from the independently derived plate scales and from the distances between the instruments and the Sun. When scanning the parameter space, the highest resolution images (EIT and AIA) were resampled using bilinear interpolation in the reference frame of the lowest resolution images (MXUVI and CLASP), the latter being left untouched. For CLASP-SJ, each image was co-registered with the AIA image that was closest in time. We validated our procedure by co-aligning an EIT 30.4~nm image with an AIA 30.4~nm image taken on August 1, 2011, at 01:07 UT. Since the pointing information of the two instruments is known independently from the headers of the images, this test shows that the co-alignment error is about 0.15 pixels rms. This translates to 0.16 arcsec for AIA-CLASP pairs and 1.6 arcsec for EIT-MXUVI pairs. These sub-pixel co-alignment errors contribute negligible dispersion to the correlations described in Sect.~\ref{sec:heii_lya_results}. Figure~\ref{eit-mxuvi} shows the \ion{H}{I} 121.6~nm MXUVI and the \ion{He}{II} 30.4~nm EIT images after co-registration. The first two images were used to find the relationship between the intensities of the two lines and correspond to MXUVI and EIT images with a 10 arcsec pixel resolution. In the third image, which shows the original full resolution EIT data, colored contours correspond to different solar regions. Figures~\ref{aia-clasp-obs-cen} and \ref{aia-clasp-obs-limb} show the \ion{H}{I} 121.6~nm CLASP-SJ and the \ion{He}{II} 30.4~nm AIA images after co-registration for, respectively, Sun center and limb observations.
\subsection{Isolation of the He\textsc{II} 30.4~nm line\label{sec:heii_isolation}}
In addition to the main line of \ion{He}{II} at 30.378~nm, the 30.4~nm passbands of EIT and AIA include several coronal and transition region lines. In particular, the \ion{Si}{XI} line at 30.332~nm cannot be suppressed by the multilayer coating technology of these instruments. In order to estimate the contribution of all the contaminating spectral lines from the EIT and AIA, we constructed a differential emission measure (DEM) curve for each pixel of the images. Then, knowing the spectral response of the instruments, we computed the intensity of all the spectral lines included in the passbands except for that of \ion{He}{II} at 30.378~nm. The sum of their contributions was then removed from the original 30.4~nm image to obtain the intensity of the \ion{He}{II} 30.378~nm line alone. For EIT, the DEM curve was constructed using the code developed specifically for EIT by \citet{Cook1999} and used in \citet{Auchere2005} to compute the \ion{He}{II} irradiance. For AIA, we used the Gaussian DEM inversion from \citet{Guennou2012a, Guennou2012b}. As a validation, we verified that the computed intensity of the \ion{Si}{XI} line represents 5 to 20\% of that of the \ion{He}{II} line, in agreement with the spectroscopically derived values of \citet{Thompson2000}.
\subsection{Results\label{sec:heii_lya_results}}
In Fig.~\ref{fig:eit-mxuvi-correl} we have represented the intensity of the \ion{H}{I} 121.6~nm line measured by MXUVI as a function of that of the \ion{He}{II} line at 30.4~nm measured by EIT. We fitted the data points with the following function:
\begin{equation}
I_{121.6} = C_{1}(1-C_{2}e^{-C_{3}I_{30.4}}) \ [\mathrm{W}.\mathrm{m}^{-2}.\mathrm{sr}^{-1}]
\label{eq1}
.\end{equation}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{correl2.eps}
\caption{Correlation between \ion{He}{II} 30.4~nm and \ion{H}{I} 121.6~nm for the November 2, 1998, data set (MXUVI and EIT). The cross on the right indicates the error bars for each data point. The different colored circles on the plot correspond to the intensities of the colored contours in the EIT image (Fig. \ref{eit-mxuvi}). The dark squares are the Vernazza \& Reeves (1978) data from Skylab, shown with their error bars. The solid line fits the MXUVI and EIT data. The dashed line corresponds to the fit from \cite{Auchere2005b}.}
\label{fig:eit-mxuvi-correl}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth ]{clasp_correl_bis.eps}
\caption{Correlation between \ion{He}{II} 30.4~nm and \ion{H}{I} 121.6~nm for the September 3, 2015, data set (CLASP-SJ and AIA). The plot in the left panel corresponds to the 1.03 arcsec/pixel CLASP-SJ resolution, whereas the plot in the right panel has the 10 arcsec/pixel MXUVI resolution. The cross on the right of each plot indicates the error bars for each data point. The solid lines correspond to the fit shown in Fig.~\ref{fig:eit-mxuvi-correl}. The dot-dashed lines correspond to the fit of the CLASP (at MXUVI resolution) and AIA data. The dashed lines show the fit for both the MXUVI-EIT and CLASP-AIA data. The colored contours outline the intensity distributions for the different solar regions from Fig.~\ref{fig:eit-mxuvi-correl}.}
\label{aia-clasp-correl}
\end{figure*}
This relation was introduced by \cite{Vourlidas2001}, and it was used by \cite{Auchere2005b} to fit the correlation between the \ion{H}{I} 121.6~nm and \ion{He}{II} 30.4~nm irradiances. The best-fit parameters are given in Table~\ref{Coefs} along with those obtained by \cite{Auchere2005b}. The solid and dashed curves in Fig.~\ref{fig:eit-mxuvi-correl} correspond respectively to the present fit and to that of \cite{Auchere2005b}. The colored circles correspond to the contribution of the different solar regions encircled in the third image of Fig.~\ref{eit-mxuvi}. As observed by \cite{Auchere2005b}, we notice in active regions (red circles) a saturation effect at high intensities. The prominence data points have a distinct distribution because of the predominantly radiative formation process of that line compared to other types of structures. It should be noted that the data points corresponding to prominences were not taken into account for the fit. Indeed, for the primary application of coronal modeling, prominences represent a negligible contribution. The uncertainty for EIT intensities is about 25\%, and the uncertainty for the Ly-$\alpha$\ irradiance from SOLSTICE is about 5\% for November 2, 1998 \citep{Snow2016}. The offset between the two fits, at most 30\%, can thus be explained by calibration uncertainties. We also plot the data from Skylab \citep{Vernazza1978} (squares), which match with the two relations found. This shows that the correlation between \ion{He}{II} 30.4~nm intensity and \ion{H}{I} 121.6~nm intensity is stable with time.\\
\begin{table}
\centering
\begin{tabular}{l|l|l|l}
& $C_{1}$ & $C_{2}$ & $C_{3}$\\
\hline MXUVI - EIT & 613 & 0.968 & 0.00991\\
\hline CLASP - AIA & 1461 & 1.002 & 0.00422\\
\hline Both data sets & 820 & 0.975 & 0.00681\\
\hline \cite{Auchere2005b} & 436 & 0.955 & 0.0203\\
\end{tabular}
\caption{Comparison between the fitted coefficients of Eq.~\ref{eq1} for the data sets considered in the present work and those in \cite{Auchere2005b}.}
\label{Coefs}
\end{table}
Figure~\ref{aia-clasp-correl} shows the intensity of CLASP-SJ \ion{H}{I} 121.6~nm as a function of AIA \ion{He}{II} 30.4~nm intensity. The uncertainties for AIA 30.4~nm are about 25\% \citep{Boerner2012}. Regarding the Ly-$\alpha$\ irradiance from SOLSTICE, the uncertainties are about 8\% for September 3, 2015 \citep{Snow2016}. The two panels correspond to two spatial resolutions: the CLASP-SJ resolution (1.03~arcsec, left) and binned to the MXUVI resolution (10~arcsec, right). The solid lines in the two plots correspond to the relation derived in the previous section from MXUVI and EIT data. The dotted-dashed line corresponds to the relation derived from CLASP and AIA images, and the dashed line corresponds to the relation derived from both data sets. These fits have been calculated with images at the MXUVI resolution. The colored contours correspond to the clusters of colored circles in Fig.~\ref{fig:eit-mxuvi-correl}. They show that the relations we obtained from CLASP-SJ and AIA images can also describe the correlation between the \ion{H}{I} 121.6~nm and \ion{He}{II} 30.4~nm lines from MXUVI-EIT images. This confirms, as mentioned above, that the relation between \ion{He}{II} 30.4~nm and \ion{H}{I} 121.6~nm intensities is stable in time.
The best-fit relationships of Figs.~\ref{fig:eit-mxuvi-correl} and \ref{aia-clasp-correl} are consistent given the uncertainties of each data set. While being thus equivalent, the corresponding uncertainties must be propagated when being used for the possible applications listed in Sect.~\ref{sec:motivations}. We also notice that there is more dispersion in the left plot of Fig.~\ref{aia-clasp-correl}, which corresponds to the higher resolution images. It shows that the correlation is stronger when one observes at lower spatial resolution. This is consistent with the extreme case of the disk-integrated data used by \cite{Auchere2005b}, for which the dispersion was even smaller. Consequently, the empirical relationships given here are particularly suited for applications that do not require high spatial resolution, such as the coronal and irradiance modeling described in Sect.~\ref{sec:motivations}. The increase in dispersion with increased resolution is possibly a signature of the differences between the formation processes of the two lines. We also notice that the dispersion is larger in \ion{He}{II} than in \ion{H}{I}. This could be explained by the sensitivity of the Planckian (collisional) contribution to the source function, since its sensitivity to temperature is proportional to the frequency. There is also a clearly different behavior between the quiet Sun and prominences for a given \ion{He}{II} intensity (see Fig.~\ref{fig:eit-mxuvi-correl}). The study of these differences is potentially important for chromospheric modeling but is beyond the scope of this paper.
\section{H\textsc{I} 102.5~nm versus He\textsc{II} 30.4~nm\label{sec:lyb_hi_vs_heii}}
\subsection{SUMER and EIT data}
In order to derive this relationship, we used data from the SUMER instrument. We used a Ly-$\beta$\ line raster measured on April 8, 1996, between 06:09 UT and 06:36 UT, close to the Sun center. The spectra we used were recorded by the SUMER/A detector, which has 1024 (spectral) $\times$ 360 (spatial) pixels$^{2}$. The slit used for the raster was the 1 $\times$ 120 arcsec$^{2}$ one. The raster contains 313 spectra at the Ly-$\beta$\ line with a step of 0.38~arcsec between each spectrum. Each spectrum covers a 0.7~nm width. We used the standard preparation routine, \texttt{sum\_read\_corr\_fits}, which includes the radiometric calibration. For the 30.4~nm data, we used the April 8, 1996, EIT image taken at 06:21 UT, which is the image closest in time to the SUMER raster.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{sum_fig.eps}
\caption{Observations from SUMER at \ion{the H}{I} Lyman-$\beta$ line on April 8, 1996, at 06:09:27 UT (left) and EIT at \ion{the He}{II} 30.4~nm line on April 8, 1996, at 06:21:51 UT, corresponding to the SUMER field of view (middle) and the full-disk observation (right). The white square on the full-disk EIT image shows the position of the SUMER field of view.}
\label{eit-lb obs}
\end{figure*}
\subsection{Data processing}
In order to build the SUMER raster image, we integrated the intensity along the wavelength axis to obtain the total intensity at each point of the field of view. We then degraded the resulting SUMER image by smoothing it down to the EIT 2.627~arcsec resolution. We used the same co-alignment procedure used for the study of the correlation between \ion{He}{II} 30.4~nm and \ion{H}{I} 121.6~nm intensities (Sect.~\ref{sec:co-registration}). Figure \ref{eit-lb obs} shows the resulting \ion{H}{I} 102.5~nm SUMER image (left) along with the co-registered EIT image (middle). These two images were used to find the relationship between the two lines. The third image (right) shows the original EIT data, the white square indicating the region covered by the SUMER raster.
\subsection{Results}
Figure \ref{eit-lb correl} shows the SUMER \ion{H}{I} 102.5~nm intensity as a function of the EIT \ion{He}{II} 30.4~nm intensity. The solid curve corresponds to a linear fit to the data with the following relation:
\begin{equation}
I_{102.5} = 0.132 I_{30.4} + 0.0233 \ [\mathrm{W}.\mathrm{m}^{-2}.\mathrm{sr}^{-1}]
\label{eq2}
.\end{equation}
The uncertainties for SUMER intensities are about 20\%. Equation~\ref{eq2} was obtained using only quiet-Sun data, covering a limited range of intensities, which justifies the linear fit. We note that the ratio ${I_{30.4}}/{I_{102.5}}$ is about 8 in the quiet Sun, to be compared to the irradiance ratio \citep[in the range 5-10, e.g., in June 2010;][]{ssisee}, the \cite{Vernazza1978} ratio of 11, or the X-flare ratio \citep[about 5;][]{Milligan2012}.
\section{H\textsc{I} 102.5~nm versus H\textsc{I} 121.6~nm\label{sec:lyb_hi_vs_lya_hi}}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{lbcorrel.eps}
\caption{Correlation between \ion{He}{II} 30.4~nm from EIT and \ion{H}{I} 102.5~nm from SUMER. The cross at the bottom right indicates the error bars for each data point. The solid line is an estimation of the correlation with a linear fit.}
\label{eit-lb correl}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{lyalybratio.eps}
\caption{Distribution of the \ion{H}{I} Ly-$\alpha$/Ly-$\beta$\ intensity ratio from the MXUVI image and the EIT image converted to Ly-$\beta$\ using Eq. ~\ref{eq2}. The dots and colored circles represent the measurements, which were performed at a period of minimum activity. The colored circles correspond to the different solar regions defined in Fig.~\ref{fig:eit-mxuvi-correl}. In order to maintain SI units, Ly-$\alpha$\ intensities are expressed in $\mathrm{mW}.\mathrm{m}^{-2}.\mathrm{sr}^{-1}$, providing numerical values equivalent to the centimeter-gram-second units of $\mathrm{erg}.\mathrm{s}^{-1}.\mathrm{cm}^{-2}.\mathrm{sr}^{-1}$ used by \cite{Lemaire2012}. The three-dotted-dashed line represents the \ion{H}{I} Ly-$\alpha$ /Ly-$\beta$\ intensity ratio deduced from Eqs.~\ref{eq1} and \ref{eq2}. The stars with error bars give an estimate of the accuracy of the ratio. The three other lines are the fits from Fig. 2 of \cite{Lemaire2012}. The solid and dashed lines correspond to two quiet-Sun observations, whereas the dotted-dashed line corresponds to a coronal hole. The squares with error bars on these lines give an estimation of the error for different ratio values on the three fits. The large black circles correspond to the ratio deduced from \cite{Patsourakos1998}.}
\label{la-lb ratio}
\end{figure*}
\cite{Lemaire2012} measured the Ly-$\alpha$ /Ly-$\beta$\ ratio from SUMER data, which were obtained in coronal holes, the quiet Sun, and active regions. As we have two relationships, between \ion{He}{II} and \ion{H}{I} Ly-$\alpha$\, and between \ion{He}{II} and \ion{H}{I} Ly-$\beta$, we can indirectly obtain a relationship between \ion{H}{I} Ly-$\alpha$\ and \ion{H}{I} Ly-$\beta$.
In order to do this, we used the November 2, 1998, EIT 30.4~nm image converted to Ly-$\beta$\ using Eq.~\ref{eq2}. This requires the assumption that Eq.~\ref{eq2}, which was derived for April 8, 1996, still holds for November 2, 1998. We also suppose that Eq.~\ref{eq2}, obtained for the quiet Sun, can be extrapolated to regions of higher intensities.
In order to compare our results to the ones of \cite{Lemaire2012}, we plot in Fig.~\ref{la-lb ratio} the Ly-$\alpha$/Ly-$\beta$\ intensity ratio as a function of the Ly-$\alpha$\ intensity. We chose a range that matches the range of validity of Eq. \ref{eq2}. Using Eqs. \ref{eq1} and \ref{eq2}, we can derive a relation for the Ly-$\alpha$/Ly-$\beta$\ intensity ratio. This relation is represented in Fig. \ref{la-lb ratio} by the three-dotted-dashed line. As in Sect.~\ref{sec:lya_hi_vs_heii}, prominences are not taken into consideration to derive Eq.~\ref{eq1}. We trace the three fits from \cite{Lemaire2012}, which correspond to one coronal hole and two quiet-Sun observations. We notice higher ratios for coronal holes and lower ratios for active regions than for the quiet Sun. We also plot values that can be derived from the intensities obtained by \citet{Patsourakos1998} from OSO 8 spectra. These values are represented by the large circles and correspond respectively to, from left to right, cell, network, bright point, and plage. This shows that our Ly-$\alpha$/Ly-$\beta$\ ratio is in the range of values measured by \cite{Lemaire2012} and deduced from \cite{Patsourakos1998}. However, they do not agree with the higher values derived by \cite{Tian2009b} and \cite{Tian2009a} for the quiet Sun (about 200) and a coronal hole (130). An explanation of these discrepancies is given in \cite{Lemaire2012}. Given the uncertainties of about 50\% on the Ly-$\alpha$/Ly-$\beta$\ ratio, represented by the stars with error bars, our values for coronal holes are compatible with those of \cite{Lemaire2012}. However, as pointed out by \cite{Bocchialini1994, Bocchialini1996} and \cite{Patsourakos1997}, the Ly-$\alpha$\ and Ly-$\beta$\ lines appear to have a singular behavior in equatorial coronal holes. For active regions, the ratio is lower. This can be explained if the linear relationship between \ion{H}{I} 102.5~nm and \ion{He}{II} 30.4~nm is not valid for active regions. To be fully compatible with \cite{Lemaire2012}, including for active regions, we can predict that the relationship between \ion{He}{II} 30.4~nm and \ion{H}{I} 102.5~nm in fact saturates at high intensities, such as the one found between \ion{He}{II} 30.4~nm and \ion{H}{I} 121.6~nm (Sect.~\ref{sec:lya_hi_vs_heii}). We note that the Ly-$\alpha$/Ly-$\beta$\ ratio is much higher (210) in a streamer at about 1~R$_{\odot}$ above the surface \citep{Giordano2013}. Such a high value could be explained by the importance of resonance scattering in both lines. We finally mention that in the case of a coronal mass ejection (CME) found a ratio of about 4, while \cite{Ciaravella2003} report a ratio of 450 in the pre-CME and CME phases.
\section{Conclusions\label{sec:conclusions}}
Using full-disk images, we derived an empirical relationship between \ion{H}{I} Ly-$\alpha$\ and \ion{He}{II} 30.4~nm intensities. Considering the uncertainties and the scatter on the intensities, our results are compatible with those obtained by \cite{Auchere2005b} from irradiance measurements, as well as with the spatially resolved results of \cite{Vernazza1978}. As the observations span four decades and various levels of activity, we conclude that this relationship is stable in time. We also observed that the dispersion around the average relationship increases with increasing resolution, which could have implications for NLTE modeling. We derived a new relation between the \ion{H}{I} Ly-$\beta$\ and \ion{He}{II} 30.4~nm intensities for two quiet-Sun regions. Combined with the \ion{H}{I} Ly-$\alpha$\ and \ion{He}{II} 30.4~nm relationship, we also obtained a Ly-$\alpha$/Ly-$\beta$\ intensity ratio compatible with previous results in the quiet Sun, which can be used to constrain chromospheric emission models. Our results are not only important for the modeling of various solar features, but also for the reconstruction of the spectral irradiance in \ion{H}{I} and \ion{He}{II} lines, which are important radiative input components in the Earth's ionosphere and thermosphere.
The three lines studied in this paper can be observed simultaneously by the Extreme Ultraviolet Imager~\citep[EUI;][]{Halain2015, Rochus2020} telescope, the Spectral Imaging of the Coronal Environment~\citep[SPICE;][]{Spice2020} spectrometer, and the Metis \citep{Antonucci2012, Antonucci2020} coronagraph on Solar Orbiter~\citep{Muller2013a, Muller2020}. The derived relationships will help in the joint analysis of the observations of these three instruments~\citep{Auchere2020}. In particular, the relationships derived in this paper are needed to analyze the Metis observations of the resonantly scattered emission of \ion{H}{I} in the corona. Since Metis does not measure the Ly-$\alpha$\ disk intensity, it is possible to instead use 30.4~nm data from the full-Sun channel of the EUI telescope \citep{Auchere2005c} to compute the illumination of the corona at 121.6~nm.
\bibliographystyle{aa}
|
\section{Introduction} \label{int01}
\subsection{Homogeneous approximation and singular vectors} \label{s1.1}
Throughout, let $m$ be a positive integer, mostly $m\geq 2$,
and let $\Vert.\Vert$ be any norm on $\Rm$.
By Dirichlet's Theorem, for any $\ux\in\Rm$ and any $Q\geq 1$ the homogeneous
system of inequalities
\begin{equation} \label{eq:1}
1\leq q\leq Q, \qquad \Vert q\ux-\underline{p}\Vert \leq cQ^{-1/m}
\end{equation}
has a solution in an integer vector $(q,\underline{p})=(q,p_1,\ldots,p_m)$
for $c=1$.
As customary we call a vector $\ux\in\Rm$ singular
if \eqref{eq:1} has a solution for
arbitrarily small $c>0$ and all large $Q\geq Q_0(c)$.
We will frequently restrict to totally irrational vectors,
i.e. to the set $\Rr\subseteq \Rm$ of $\ux=(\xi_1,\ldots,\xi_m)$ for which $\{1,\xi_1,\ldots, \xi_m\}$ is $\mathbb{Q}$-linearly independent,
or equivalently that do not lie in a rational affine hyperplane.
Denote by $\mathcal{S}_m\subseteq \Rm$ resp. $\mathcal{S}_m^{\ast}=\mathcal{S}_m\cap \Rr$
the set of singular resp.
totally irrational singular vectors in $\Rm$.
Clearly $\mathcal{S}_m^{\ast}\subseteq \mathcal{S}_m$, and for $m=1$ we have $\mathcal{S}_1=\mathbb{Q}$
and $\mathcal{S}_1^{\ast}=\emptyset$ due to Khintchine~\cite{khintchine}.
An important problem in Diophantine approximation that has received
much attention
in recent years is to determine the Hausdorff and packing dimensions
(for a definition see see~\cite{falconer})
of the set of singular vectors (matrices).
Building up on a landmark paper by
Cheung~\cite{che} for $m=2$, Cheung and Chevallier~\cite{cheche}
determined the Hausdorff dimension of $\mathcal{S}_m$ for $m\geq 2$.
An upper bound in a more general matrix setting
was later obtained by
Kadyrov, Kleinbock, Lindenstrauss and Margulis~\cite{kklm}.
Using their deep variational principle,
Das, Fischman, Simmons, Urba\'nski~\cite{dfsu1, dfsu2} showed that this upper bound is sharp
and thereby extended~\cite{cheche}, with a new proof.
In~\cite{dfsu1, dfsu2} it is
also shown that the packing dimensions
attain the same value (in the general matrix setting). Denoting by $\dim_H(A)$ resp. $\dim_P(A)$ the Hausdorff and packing dimension
of $A\subseteq \Rm$, the claims read as follows.
\begin{theorem}[Cheung, Chevallier; Das, Fishman, Simmons, Urba\'nski] \label{DFSU}
For any $m\geq 2$, we have
\[
\dim_{H}(\mathcal{S}_m)=\dim_{H}(\mathcal{S}_m^{\ast})=\dim_{P}(\mathcal{S}_m)=\dim_{P}(\mathcal{S}_m^{\ast})=m-1+\frac{1}{m+1}.
\]
\end{theorem}
For $w\in[0,1)$, define $\mathcal{S}_m(w)$ resp. $\mathcal{S}_m^{\ast}(w)$
as the set of $\ux$ for which
\begin{equation} \label{eq:c=1}
1\leq q\leq Q, \qquad \Vert q\ux-\underline{p}\Vert \leq cQ^{-w}
\end{equation}
is soluble in integers for any $c>0$ and $Q\geq Q_0(c)$,
for $\ux\in \Rm$ resp. $\ux\in\Rr$.
Then
\begin{equation} \label{eq:GONG}
\mathcal{S}_m=\mathcal{S}_m(1/m), \qquad \mathcal{S}_m^{\ast}=\mathcal{S}_m^{\ast}(1/m),
\end{equation}
and we need not consider $w\geq 1$ since then $\mathcal{S}_m(w)=\mathbb{Q}^m$, $\mathcal{S}_m^{\ast}(w)=\emptyset$
by~\cite{khintchine} mentioned above.
Despite Theorem~\ref{DFSU}, the following problem
discussed in a slightly altered form (and in the general matrix setting) in~\cite{dfsu1},
remains open for $m>2$.
\begin{problem} \label{p1}
For $m\geq 2$ and $w\in[1/m,1)$, determine the Hausdorff and packing dimensions of the sets
$\mathcal{S}_{m}(w)$ and $\mathcal{S}_m^{\ast}(w)$.
\end{problem}
In view of \eqref{eq:GONG}, for $w=1/m$ Problem~\ref{p1}
is solved by Theorem~\ref{DFSU}.
For $m=2$, after several preceeding papers,
it is completely solved as well
in view of Theorem~1.9 and the subsequent Remark in~\cite{dfsu1}, however the explicit formulas are cumbersome.
Lower bounds for general $m$ and the packing dimensions
given in~\cite[Theorem~3.8]{dfsu2} can be phrased
\begin{equation} \label{eq:dfsuoben}
\dim_P(\mathcal{S}_m(w))\geq \max\left\{ m-1+\frac{1}{m+1}- \frac{(2m+1)(mw-1)}{(m+1)(w+1)}, m-\frac{m(m-1)w}{m-w} \right\}.
\end{equation}
For $m=2$ and regarding Hausdorff dimension, see also the independent
work by Bugeaud, Cheung and Chevallier~\cite{bcc} who showed that
\begin{equation} \label{eq:bchch}
\dim_H(\mathcal{S}_2(w))\geq 2-2w, \qquad w\in(1/2,1),
\end{equation}
and also settled equality when $w>1/\sqrt{2}$.
For $m>2$, the precise formulas for both Hausdorff and packing
dimension of $\mathcal{S}_m(w)$ remain unknown, however
several other estimates complementing the stated claims
are given in~\cite{dfsu1, dfsu2}.
Next we introduce
the classical homogeneous exponents of ordinary approximation $\om_m(\ux)$ respectively of uniform approximation
$\wo_m(\ux)$ as the supremum of $w$ such that
the system \eqref{eq:c=1} for $c=1$
admits a solution in integers for arbitrarily large resp. all large $Q$.
For any $\ux\in\Rm$, by \eqref{eq:1} we obviously have
\begin{equation} \label{eq:kontrast}
\om_m(\ux) \geq \wo_m(\ux)\geq \frac{1}{m}.
\end{equation}
For any $w$ the sets $\mathcal{S}_m(w)$
have a simple relation to superlevelsets of
$\wo_m$ given by
\begin{equation} \label{eq:accext}
\{ \ux\in\Rm: \wo_m(\ux)>w \}\subseteq
\mathcal{S}_m(w)\subseteq \{ \ux\in\Rm: \wo_m(\ux)\geq w \},
\end{equation}
and similarly for $\mathcal{S}_m^{\ast}(w)$.
In~\cite{dfsu1, dfsu2} the notation $Sing(m,1)$ means
the singular vectors in $\Rm$, which is our $\mathcal{S}_m$.
Thus our sets $\mathcal{S}_m^{\ast}$ coincide with $Sing(m,1)\cap \Rr$.
The sets $\widetilde{Sing}_{m,1}^{\ast}(w):=\{ \ux\in\Rr: \wo_m(\ux)\geq w \}$ studied in~\cite{dfsu1} in view of \eqref{eq:accext}
just subtly differ
from our sets $\mathcal{S}_m^{\ast}(w)$.
It will be convenient for us to only impose lower bounds for the
order of uniform approximation. However,
certain results below can be extended with some effort
to sets $Sing_{m,1}^{\ast}(w):=\{ \ux\in\Rr: \wo_m(\ux)=w \}$
or $Sing_{m,1}(w):=\{ \ux\in\Rm: \wo_m(\ux)=w \}$ from~\cite{dfsu1} as well, see Remark~\ref{hirsch} below. Presumably, the metrical formulas
for levelsets and superlevelsets coindice, see~\cite[Theorem~4.9]{dfsu2}
and subsequent comments.
The natural restriction to $\Rr$, inducing the exponents
with the asterisk ''$\ast$'', is crucial at some places.
This applies in particular to our Theorem~\ref{cantor} on (homogeneously) singular vectors in certain fractal sets.
Indeed, for $1\leq k\leq m$ integers, $\underline{\zeta}=(\zeta_1,\ldots,\zeta_k)\in\mathbb{R}^k$
and $\ux=(\zeta_1,\ldots,\zeta_k,\zeta_k,\ldots,\zeta_k)\in\Rm$
have the same irrationality exponents, i.e. $\om_k(\underline{\zeta})=\om_m(\ux)$ and $\wo_k(\underline{\zeta})=\wo_m(\ux)$, and similarly for
the $b$-ary exponents. From Dirichlet's Theorem
and since the Hausdorff dimension of a set never exceeds its packing dimension~\cite{falconer} we get
\begin{equation} \label{eq:tr}
\dim_P(\mathcal{S}_m(w)\setminus \Rr) \geq \dim_H(\mathcal{S}_m(w)\setminus \Rr) \geq k, \qquad w<\frac{1}{k}.
\end{equation}
Similar results can also be derived from transference arguments,
considering a linear form in $m$ variables.
At some places \eqref{eq:tr}
exceeds our lower bounds, thereby directly implying some of
our (homogeneous, non-$b$-ary) results without restriction to $\Rr$.
\subsection{Inhomogeneous and $b$-ary approximation} \label{inhomo}
Consider now the classical inhomogeneous approximation problem,
where for given $\ux, \ut\in \Rm$,
we study small values of $\Vert q\ux-\underline{p}-\ut\Vert$
in terms of $Q$, where $1\leq q\leq Q$.
The theory again splits in ordinary and uniform approximation.
Define the ordinary exponent $\om_{m,\ut}(\ux)$
resp. uniform exponent $\wo_{m,\ut}(\ux)$ of inhomogeneous approximation
as the supremum
of $w$ such that
\begin{equation} \label{eq:F}
1\leq q\leq Q, \qquad \Vert q\ux-\underline{p}-\ut\Vert \leq
c Q^{-w}
\end{equation}
has a solution in a non-zero integer vector $(q,\underline{p})=(q,p_1,\ldots,p_m)$
for
$c=1$ and
certain arbitrarily large $Q$ resp. all large $Q$.
The homogeneous exponents from \S~\ref{s1.1}
represent the special case
\begin{equation} \label{eq:konserve}
\om_{m}(\ux)= \om_{m,\underline{0}}(\ux), \qquad \wo_{m}(\ux)= \wo_{m,\underline{0}}(\ux).
\end{equation}
In the inhomogeneous case, hugely contrasting \eqref{eq:kontrast},
even the ordinary exponents may
vanish. Yet,
from a metrical point of view,
for ordinary approximation
we still have a very satisfactory understanding
by the
inhomogeneous Khintchine-Groshev Theorem, analogous
to the homogeneous theory.
See~\cite{allen, hss} for recent
subtle improvements and further references. On the other hand,
little is known when extending Problem~\ref{p1} and related metrical questions on uniform approximation to the
inhomogeneous setting. These problems are
explictly discussed in~\cite[\S~5.8]{dfsu2}.
However, we should mention
a special case of a metrical result by T. Kim and W. Kim~\cite[Corollary~1.5]{kimkim}.
\begin{theorem}[Kim, Kim] \label{blau}
For $m\geq 2$ and every $\ut\notin \mathbb{Q}^m$ and
every $w>0$, we have
\[
\dim_H( \{ \ux\in\Rm: \wo_{m,\ut}(\ux) \leq w \} ) = \min\{ m-\frac{1-mw}{1+w} , m\}.
\]
\end{theorem}
The same is true when taking equality in the left hand side set.
See also Bakhtawar
and Simmons~\cite{baksim} for refinements, and for
further contributions to Problem~\ref{p1} Bugeaud and Laurent~\cite{buglau}
and Kim and Liao~\cite{kimliao}, the last treating
$m=1$ and fixing $\ux$ and letting $\ut$ vary. Note that
we are primarily interested in sets as in Problem~\ref{p1}
with the reverse inequality and $w>1/m$, where Theorem~\ref{blau}
gives little information.
In this paper we develop some metrical theory
for inhomogeneous, uniform approximation in arbitrary dimension $m$.
In contrast to the homogeneous case in \S~\ref{s1.1},
for general $\ut$, we are unable to provide an affine hyperplane that
consists of vectors with good (even ordinary)
approximation of order greater than $1/m$, so it seems to us that
no inhomogeneous analogue of \eqref{eq:tr} exists.
Nevertheless, to be on the safe side, we will often restrict to
$\Rd$ which we introduce
as the complement of any given countable union of affine hyperplanes
in $\Rm$.
Besides the inhomogeneous setting, in this paper we also consider vectors with good $b$-ary approximation, i.e.
where $q=b^N, N\in\mathbb{N}$, for some $b\geq 2$. Then the base $b$ expansion of all coordinates of $b^N\ux-\underline{p}-\ut\in\Rm$
start with long blocks of consecutive $0$ or $b-1$ digits. Define the inhomogeneous,
$b$-ary exponents $\om_{m,\ut}^{(b)}(\ux)$
and $\wo_{m,\ut}^{(b)}(\ux)$ as above but where we take
$q=b^N$ integer
powers of $b$ in \eqref{eq:F}. The homogeneous, ordinary $b$-ary exponent $\om_1^{(b)}(\xi)=\om_{1,0}^{(b)}(\xi)$
equals the exponent $v_b(\xi)$
defined and studied by Amou and Bugeaud~\cite{ab10}. It was also implicitly studied by Levesley, Salp and Velani~\cite{lsv} for $b=3$ and numbers in Cantor's middle third set. In contrast to \eqref{eq:kontrast}, the $b$-ary exponent $\om_m^{(b)}$
generically (for Lebesgue almost all
$\ux\in\Rm$)
vanishes.
In the
inhomogeneous setting,
the inequality $\wo_{1,\theta}(\xi)>1$ occurs
for certain $\xi,\theta\in\mathbb{R}$ not inducing the trivial
case given by $q\xi-p-\theta=0$ for some
$(q,p)\in\mathbb{Z}^{2}$, see Kim and Liao~\cite{kimliao} for metrical results. It is however unclear to us whether
$\wo_{1,\theta}^{(b)}(\xi)>1$ may occur unless in the trivial situation.
We extend some more notions from \S~\ref{s1.1}
to our inhomogeneous, $b$-ary setting.
\begin{definition} \label{d}
For $w\in[0,\infty)$, define the nested sets
\[
\mathcal{S}_{m,\ut}(w)= \{ \ux\in\Rm:\; \eqref{eq:F} \; \text{has a solution for every} \;\; c>0 \;\; \text{and all large Q} \},
\]
and let $\mathcal{S}_{m,\ut}^{\ast}(w)=\mathcal{S}_{m,\ut}(w)\cap \Rd $ with $\Rd$ defined above. Further let
\[
\mathcal{S}_{m,\ut}= \mathcal{S}_{m,\ut}(1/m),\qquad
\mathcal{S}_{m,\ut}^{\ast}= \mathcal{S}_{m,\ut}^{\ast}(1/m).
\]
Derive accordingly
\[
\mathcal{S}_{m,\ut}^{(b)}(w)\subseteq \mathcal{S}_{m,\ut}(w),\quad
\mathcal{S}_{m,\ut}^{(b)}\subseteq \mathcal{S}_{m,\ut}, \quad
\mathcal{S}_{m,\ut}^{(b)\ast}(w)\subseteq \mathcal{S}_{m,\ut}^{\ast}(w),
\quad \mathcal{S}_{m,\ut}^{(b)\ast}\subseteq \mathcal{S}_{m,\ut}^{\ast}
\]
with respect to $b$-ary approximation, i.e. where we restrict
to $q=b^N$.
\end{definition}
As a consequence of \eqref{eq:konserve}, the according extension of \eqref{eq:accext} holds, and
we find the classical singular vectors as a special
case of our definition via
\[
\mathcal{S}_m(w)= \mathcal{S}_{m,\underline{0}}(w), \qquad \mathcal{S}_m= \mathcal{S}_{m,\underline{0}}, \qquad \mathcal{S}_m^{\ast}(w)= \mathcal{S}_{m,\underline{0}}^{\ast}(w), \qquad \mathcal{S}_m^{\ast}= \mathcal{S}_{m,\underline{0}}^{\ast},
\]
and similarly for the $b$-ary sets.
Results from~\cite{buglau}
indicate that $w=1/m$ is a metrical threshold value as
in the homgeneous case, so considering $\mathcal{S}_{m,\ut}$ is plausible.
\subsection{Organisation of the paper} The remainder of the paper is
organized as follows. In \S~\ref{se2}, we establish lower bounds
for the packing dimension of the set of inhomogeneously ($b$-ary) singular vectors,
and also address the case of multiple
inhomogenities $\ut$ simultaneously.
In \S~\ref{se3} we provide independent
results on sumsets of singular vectors, as well as topological
results. In particular we show that $\mathcal{S}_m$ is not a comeagre set.
In \S~\ref{se4} we deal with Cartesian products of missing digit sets (the Cantor middle third set is a special case), classical fractals. In the homogeneous case $\ut=\underline{0}$, we show that totally irrational vectors with
arbitrary order of singularity $w<1$ exist in these fractals
and again provide lower bounds for their packing and Hausdorff dimension.
\section{Metrical results for inhomogeneous, $b$-ary singular vectors in $\Rm$} \label{se2}
\subsection{Metrical claims in $\Rm$} \label{s2.1}
Our first new result gives lower bounds for the packing dimension
of the set of inhomogeneously $b$-ary singular vectors in $\Rd$.
\begin{theorem} \label{H}
Let $m\geq 1$, $b\geq 2$ integers and $\ut\in\Rm$. Then for
any $w\in[0,1)$ we have
\begin{equation} \label{eq:unfrei}
\dim_{P}(\mathcal{S}_{m,\ut}^{\ast}(w)) \geq \dim_{P}(\mathcal{S}_{m,\ut}^{(b)\ast}(w)) \geq (1-w)m>0.
\end{equation}
In particular
\begin{equation} \label{eq:frei}
\dim_{P}(\mathcal{S}_{m,\ut}^{\ast}) \geq \dim_{P}(\mathcal{S}_{m,\ut}^{(b)\ast}) \geq m-1.
\end{equation}
\end{theorem}
\begin{remark} \label{ureh}
If $\ut=\underline{0}$,
then \eqref{eq:unfrei}, \eqref{eq:frei} for
$\mathcal{S}_{m}^{\ast}(w)$ and
$\mathcal{S}_{m}^{\ast}$ instead
are superseeded by \eqref{eq:dfsuoben}, and regarding the sets $\mathcal{S}_{m}(w)$ and $\mathcal{S}_{m}$ also the bound in
\eqref{eq:tr} is larger. Moreover, the Hausdorff (thus packing) dimension of the non-$b$-ary sets $\mathcal{S}_{m,\ut}^{\ast}(w)$
is in fact full when $w<1/m$ by Theorem~\ref{blau}. So the non-$b$-ary claims are of interest for $\ut\neq \underline{0}$ and $w\geq 1/m$.
\end{remark}
\begin{remark}
When combining our method with the one of~\cite[Theorem~2.3]{arxiv3}
based on~\cite[Example~4.6]{falconer},
we also obtain a positive lower bound for the Hausdorff dimension, which is however considerably weaker. A calculation yields $\dim_H(\mathcal{S}_{m,\ut}^{(b)\ast}(w))\geq \left(\frac{1-w}{1+w}\right)^2$.
\end{remark}
Note that when $\ut=\underline{0}$, we cannot have anything larger than $m-1+\frac{1}{m+1}$
in \eqref{eq:frei} by Theorem~\ref{DFSU}, so up to $O(m^{-1})$
it is optimal.
We stress that the setup in Theorem~\ref{H} is more general
than Theorem~\ref{DFSU} in the sense
that it holds in inhomogeneous and $b$-ary setting. In particular, we
should expect strict inequality $\dim_{P}(\mathcal{S}_{m}^{(b)\ast})<\dim_{P}(\mathcal{S}_{m}^{\ast})$, as is also
suggested by \eqref{eq:jarnik} below. It is thus
natural that our new estimate \eqref{eq:unfrei}
is weaker than \eqref{eq:dfsuoben}.
We further notice that for $m=2$ the bound is identical
with the upper bound for the Hausdorff dimension in \eqref{eq:bchch}.
Our proof of Theorem~\ref{H} and in fact all claims
below are surprisingly easy and completely
unrelated to any previous approach, as for example in~\cite{dfsu2}.
We remark that our method a fortiori
readily extends to general inhomogeneous systems of $m$ linear
forms in $n$ variables,
however the results turn out most interesting
in our setting $n=1$. Moreover, for $n>1$ the exact
value of the packing dimensions of levelsets in the homogeneous
problem are known~\cite[Theorem~3.8]{dfsu1}.
We also obtain a slightly weaker lower bound for the Hausdorff dimension.
\begin{theorem} \label{ax3}
With the assumption and notation of Theorem~\ref{H}, we have
\[
\dim_H(\mathcal{S}_{m,\ut}^{\ast}(w))\geq \dim_H(\mathcal{S}_{m,\ut}^{(b)\ast}(w))\geq m\left(\frac{1-w}{1+w}\right)^2.
\]
In particular
\[
\dim_{H}(\mathcal{S}_{m,\ut}^{\ast}) \geq \dim_{H}(\mathcal{S}_{m,\ut}^{(b)\ast}) \geq m\left(1-\frac{2}{m+1}\right)^2.
\]
\end{theorem}
Note that the latter bound is of order $m-4+O(m^{-1})$. See
also~\cite[Theorem~2.3]{arxiv3}.
\subsection{Singularity with respect to several $\ut$ simultaneously} \label{se2.2}
We now investigate sets in $\Rm$ that belong to $\mathcal{S}_{m,\ut}$
simultaneously for several $\ut\in\Rm$. To our knowledge, this topic
has not been studied before, see however the related remarks
on~\cite{bglasgow, ts, kimliao} in the last paragraph
of this section. Denote by
$e=2.71\ldots$ Euler's number.
\begin{theorem} \label{simu}
Let $m\geq 1$, $k\geq 1, b\geq 2$ be integers
and $\Theta=\{ \ut_1,\ldots,\ut_k \}$
with $\ut_s\in\Rm, 1\leq s\leq k$ arbitrary. For any $w\in[0,1)$
we have
\begin{equation} \label{eq:bound}
\dim_P(\bigcap_{\ut\in \Theta} \mathcal{S}_{m,\ut}^{\ast}(w))\geq
\dim_P(\bigcap_{\ut\in \Theta} \mathcal{S}_{m,\ut}^{(b)\ast}(w))\geq (1-kew)m.
\end{equation}
In particular, if $k<m/e$,
then for any $\Theta=\{ \ut_1,\ldots,\ut_k \}$ we have
\[
\dim_P(\bigcap_{\ut\in \Theta} \mathcal{S}_{m,\ut}^{\ast})\geq \dim_P(\bigcap_{\ut\in \Theta} \mathcal{S}_{m,\ut}^{(b)\ast})>0.
\]
\end{theorem}
Again for $w<1/m$ the claim for the non-$b$-ary sets is implied by Theorem~\ref{blau} since
full measure is preserved under countable intersections.
When $k=1$, compared to Theorem~\ref{H} our bound \eqref{eq:bound} is poorer due to the appearance of the factor $e$. However,
as the proof will show, this factor can be improved for
given $k$, and rather significantly when $k$
is small.
We further remark that for any given element $\Theta$
of the power set of $\Rm$, we have $\cap_{\ut\in\Theta} \mathcal{S}_{m,\ut}(w)=\cap_{\ut\in\scp{\Theta}}\mathcal{S}_{m,\ut}(w)$
where $\scp{\Theta}$ denotes
the $\mathbb{Z}$-module in $\Rm$ spanned by $\Theta$ and the canonical
base vectors $\underline{e}_i=(0,\ldots,0,1,0\ldots,0)$, $1\leq i\leq m$.
Regarding $b$-ary approximation, we can present a reverse result.
We restrict to $m=1$ and $k=2$ for simplicity, it can however be generalized.
\begin{theorem} \label{reverse}
Let $\theta_1= \sum_{N\geq 1} 3^{-N!}, \theta_2=2\theta_1$.
Then for any $\xi\in\mathbb{R}$ we have
\[
\wo_{1,\theta_1}^{(3)}(\xi)+ \wo_{1,\theta_2}^{(3)}(\xi) \leq 1.
\]
In particular, for any $w>1/2$, we have
\[
\mathcal{S}_{1,\theta_1}^{(3)}(w)\cap \mathcal{S}_{1,\theta_2}^{(3)}(w)= \emptyset.
\]
\end{theorem}
The proof admits plenty of freedom in the choice of $\theta_i$ in the theorem,
and there is no significance of the base $3$ either.
We next want to establish below a new result on ordinary,
inhomogeneous approximation.
Complementing the sets $\mathcal{S}_m(w)$ and $\mathcal{S}_m^{(b)}(w)$, regarding ordinary approximation we define similarly the nested sets
\[
\mathcal{W}_{m,\ut}(w)=\{ \ux\in\Rm: \om_{m,\ut}(\ux)\geq w \},\qquad
\mathcal{W}_{m,\ut}^{(b)}(w)=\{ \ux\in\Rm: \om_{m,\ut}^{(b)}(\ux)\geq w \},
\]
for $w\in[0,\infty]$. We may introduce variants
where we restrict to $\Rr$ or $\Rd$ as we did for singular vectors, however the distinction is of no relevance here. For any parameter $w$, we have
the obvious inclusions $\mathcal{W}_{m,\ut}^{(b)}(w)\subseteq \mathcal{W}_{m,\ut}(w)$ and
\[
\mathcal{S}_{m,\ut}(w)\subseteq \mathcal{W}_{m,\ut}(w), \qquad
\mathcal{S}_{m,\ut}^{(b)}(w)\subseteq \mathcal{W}_{m,\ut}^{(b)}(w).
\]
See Marnat and Moshchevitin~\cite{mamo} for refinements of the left inclusion when $\ut=\underline{0}$, in which case we get the classical sets $\mathcal{W}_m(w)=\mathcal{W}_{m,\underline{0} }(w)$ and $\mathcal{W}_m^{(b)}(w)=\mathcal{W}_{m,\underline{0} }^{(b)}(w)$ given as
\[
\mathcal{W}_m(w)= \{ \ux\in\Rm: \omega_m(\ux)\geq w \}, \qquad \mathcal{W}_m^{(b)}(w)= \{ \ux\in\Rm: \omega_m^{(b)}(\ux)\geq w \}.
\]
Dirichlet's Theorem implies
$\mathcal{W}_m(1/m)=\Rm$. Moreover
\begin{equation} \label{eq:jarnik}
\dim_H(\mathcal{W}_m^{(b)}(w))=\frac{m}{w+1},\; w\in[0,\infty], \qquad \dim_H(\mathcal{W}_m(w))= \frac{m+1}{w+1}, \; w\in[1/m,\infty],
\end{equation}
where the left identity follows from Borosh and Fraenkel~\cite{bf72},
and the right identity is a famous result due to Jarn\'ik~\cite{jarnik}. The packing dimension of both sets is full,
a stronger result can be found in~\cite[Corollary~4]{arxiv}.
For the larger
sets $\mathcal{W}_m(w)$, alternatively this is consequence of independent
results by Marnat~\cite{marnat} deduced
from the variational principle established in~\cite{dfsu1, dfsu2}.
Theorem~\ref{heep} below extends these claims and shows that
for ordinary approximation,
we get considerably stronger metrical results than in Theorem~\ref{simu}.
\begin{theorem} \label{heep}
Let $m\geq 1, b\geq 2$ integers and $\Theta=\{ \ut_1,\ldots \}$ be countable with $\ut_i\in\Rm$. Then
\[
\dim_P(\bigcap_{\ut\in\Theta} \mathcal{W}_{m,\ut}(\infty))= \dim_P(\bigcap_{\ut\in\Theta} \mathcal{W}_{m,\ut}^{(b)}(\infty))=m.
\]
\end{theorem}
Combining our proof below with the argument in Remark~\ref{hirsch}
below, it follows that any countable intersection over sets
$\widetilde{W}_{m,\ut_i}^{(b)}(w_i):=
\{ \ux\in\Rm: \om_{m,\ut_i}^{(b)}(\ux)=w_i \}$
of exact $b$-ary order of ordinary, inhomogeneous
approximation $w_i\in[0,\infty]$, has full
packing dimension. However, this claim for the
accordingly defined non-$b$-ary sets $\widetilde{W}_{m,\ut_i}(w_i)$ is not obvious.
Notice that the sets $\Theta$ in Theorems~\ref{simu},~\ref{heep}
are countable.
Regarding uncountable intersections, when $m=1$, some
information can be obtained from
the papers by Bugeaud~\cite{bglasgow} or Troubetzkoy and Schmeling~\cite{ts} for ordinary approximation,
and Kim and Liao~\cite{kimliao}
for uniform approximation. A particular consequence of~\cite{kimliao} is that for any given $w<\infty$, there is certain $\Theta\subseteq \mathbb{R}$ of positive Hausdorff
dimension, so that the intersection over $\theta\in\Theta$
of the sets $\mathcal{S}_{1,\theta}(w)$
is non-empty.
\section{Topological properties of singular vectors} \label{se3}
\subsection{Sumsets of singular vectors}
Our method for the proofs of claims in \S~\ref{se2}
is based on investigation of sumsets.
We state some related result directly addressing sumsets,
obtained from a similar method. Denote $A+B=\{a+b: a\in A, b\in B\}$ for $A,B\subseteq \Rm$.
\begin{theorem} \label{t3}
Let $m\geq 1, b\geq 2$ be integers and $\ut_0,\ut_1\in\Rm$.
If $w\in(0,1)$ and
\begin{equation} \label{eq:idne}
w^{\prime}= (1-\sqrt{w})^2= w+1-2\sqrt{w},
\end{equation}
then for any pair of real numbers $(w_0, w_1)$ satisfying
$0\leq w_0<w$ and $0\leq w_1<w^{\prime}$, the identity of sets
\begin{equation} \label{eq:G0}
\mathcal{S}_{m,\ut_0}^{(b)}(w_0) + \mathcal{S}_{m,\ut_1}^{(b)}(w_1)=
\mathcal{S}_{m,\ut_0}(w_0) + \mathcal{S}_{m,\ut_1}(w_1)= \Rm
\end{equation}
holds. Moreover the sets
\begin{equation} \label{eq:fullm}
\mathcal{S}_{m,\ut_0}^{(b)\ast}(w_0) + \mathcal{S}_{m,\ut_1}^{(b)\ast}(w_1), \qquad
\mathcal{S}_{m,\ut_0}^{\ast}(w_0) + \mathcal{S}_{m,\ut_1}^{\ast}(w_1)
\end{equation}
have full $m$-dimensional Lebesgue measure.
\end{theorem}
\begin{remark} \label{rmark}
We believe that the set identities in fact hold for the sets with $\ast$ as well, but technical obstructions on small sets occur in our proof.
\end{remark}
We want to highlight the special, symmetric case $w=1/4$.
\begin{corollary} \label{mko}
With notation and assumptions of Theorem~\ref{t3}, for any $w<1/4$ we have
\begin{equation} \label{eq:B}
\mathcal{S}_{m,\ut_0}^{(b)}(w) + \mathcal{S}_{m,\ut_1}^{(b)}(w)=
\mathcal{S}_{m,\ut_0}(w) + \mathcal{S}_{m,\ut_1}(w)=\Rm,
\end{equation}
and the according sumsets with $\ast$ throughout have full $m$-dimensional Lebesgue measure.
\end{corollary}
The claim is most
interesting for $m\geq 5$, otherwise
the sets $\mathcal{S}_{m,\ut_i}(w)$ with $w<1/4$ have full $m$-dimensional Lebesgue measure by Theorem~\ref{blau}, consequently the
implication for the (larger) set
$\mathcal{S}_{m,\ut_0}(w) + \mathcal{S}_{m,\ut_1}(w)$ follows from
an easy argument.
From \eqref{eq:B} with $\ut_0=\ut_1=\ut$ and \eqref{eq:AB} below
we could deduce the lower bound $m/2$ for
the packing dimension of $\mathcal{S}_{m,\ut}^{(b)}(w)$ for
any $\ut\in\Rm$ and $w<\frac{1}{4}$, however this is superseeded by Theorem~\ref{H}.
From the full measure claim \eqref{eq:fullm}, we infer some information
on the Hausdorff dimension of the homogeneous sets
$\mathcal{S}_m^{\ast}(w)$ for certain $w$ close to $1$. Corollary~\ref{coro}
below may be regarded
complementary to~\cite[Theorem~1.7]{dfsu1}, in fact it shows that
the interval for the validity of its left estimate
has length $\ll 1/\sqrt{m}$.
\begin{corollary} \label{coro}
For $m\geq 2$ and any $w<(\sqrt{m}-1)^2/m$, we have
\[
\dim_H(\mathcal{S}_{m}^{\ast}(w)) \geq\dim_H(\mathcal{S}_{m}^{(b)\ast}(w)) \geq 1-\frac{1}{m+1}.
\]
\end{corollary}
We prove the corollary in \S~\ref{pko}.
The analogous claim for $\mathcal{S}_m(w)$ instead would be
trivially implied by \eqref{eq:tr}.
Similar to \eqref{eq:B},
Corollary~\ref{coro} is most interesting for $m\geq 5$, otherwise the
parameter bound does not exceed the critical value $1/m$.
We next notice that $1/4$ in Corollary~\ref{mko}
cannot be replaced by $1$, at least in the
homogeneous case $\ut_0=\ut_1=\underline{0}$. This is an easy
corollary of results in~\cite{dfsu1}.
\begin{theorem} \label{t4}
For $m\geq 2$ and $\epsilon>0$,
for $w=w(m,\epsilon)<1$ close enough to $1$ we have
\[
\dim_H(\mathcal{S}_m(w) + \mathcal{S}_m(w))= 2, \qquad
\dim_H(\mathcal{S}_m^{\ast}(w) + \mathcal{S}_m^{\ast}(w))\leq 1+\epsilon.
\]
\end{theorem}
The reverse claims in Corollary~\ref{mko} and Theorem~\ref{t4}
motivate to determine the threshold exponents where the
topological behavior of the sumsets change.
\begin{problem}
Determine $\mathcal{T}_m$ resp. $\mathcal{T}_{m}^{\ast}$ defined as
the supremum of $w$ so that the sumset
$\mathcal{S}_m(w) + \mathcal{S}_m(w)$ resp. $\mathcal{S}_{m}^{\ast}(w) + \mathcal{S}_{m}^{\ast}(w)$ equals $\Rm$.
\end{problem}
It is clear that $\mathcal{T}_1=\mathcal{T}_{1}^{\ast}=1$.
By considering two rational lines in $\mathbb{R}^2$ as in the proof
of Theorem~\ref{t4} below,
we still easily get $\mathcal{T}_2=1$.
However, we can deduce from the results above Dirichlet's Theorem
that
$\mathcal{T}_m\in [ \max\{ \frac{1}{4}, \frac{1}{m}\},1)$ when $m\geq 3$,
and moreover $\mathcal{T}_m^{\ast}\in [\frac{1}{m},1)$ when $m\geq 2$ (conjecturally $1/4$ is also a lower bound). We expect
analogous results
for accordingly defined inhomogeneous suprema $\mathcal{T}_{m,\ut}, \mathcal{T}_{m,\ut}^{\ast}$.
\subsection{On the Baire category of singular vectors} \label{top}
We next turn towards topological claims on the set
of singular vectors. Recall a set is called comeagre (or residual)
if it is the intersection
of countably many open dense sets, or equivalently its complement
is of first Baire category. The set of Liouville numbers $\mathcal{W}_1^{\ast}(\infty)=W_1(\infty)\cap \mathcal{R}_1=W_1(\infty)\setminus \mathbb{Q}$ is well-known to be comeagre~\cite{oxtoby}. By a very similar
argument this extends to Liouville vectors $\mathcal{W}_m^{\ast}(\infty)$, and further
clearly to $\mathcal{W}_m(\infty)$ as the difference set
is contained in a countable union of affine hyperplanes.
We show that this is not true for $\mathcal{S}_m$. This relies
on our next very general observation
linking toplogy and measure theory.
\begin{theorem} \label{T2}
If $A\subseteq \Rm$ is comeagre, then $\dim_P(A)=m$.
\end{theorem}
To our knowledge this has not been noticed before.
The claim is far from being true for the Hausdorff dimension, as the set $\mathcal{W}_m(\infty)\subseteq \Rm$ (and $\mathcal{W}_m^{(b)}(\infty)$ as well) provides a counterexample of Hausdorff dimension $0$ in view
of \eqref{eq:jarnik}.
As
the inhomogeneous sets $\mathcal{W}_{m,\ut}(\infty)$ and $\mathcal{W}_{m,\ut}^{(b)}(\infty)$
are comeagre as well by the same argument as for $\ut=\underline{0}$,
Theorem~\ref{T2} further leads to an alternative, unconstructive proof of Theorem~\ref{heep}. Moreover, from Theorem~\ref{DFSU} we immediately infer topological information on the sets $\mathcal{S}_m$.
\begin{corollary} \label{T1}
For any $m\geq 1$, the set $\mathcal{S}_m$ is not comeagre in $\Rm$.
\end{corollary}
It can be shown that $\mathcal{S}_m$ has non-empty intersection with the comeagre set
of Liouville vectors $\mathcal{W}_m^{\ast}(\infty)$ (follows
for example from~\cite[Theorem 2.5]{ichostrava}),
however Corollary~\ref{T1} shows that the intersection is small
in a topological sense.
We strongly believe that the claim extends to the inhomogeneous sets $\mathcal{S}_{m,\ut}$
as well, however from our results above we cannot exclude
that it has full packing dimension for certain $\ut$.
We wonder if the Corollary~\ref{T1} remains true for the larger set
$Di_m\subseteq \Rm$ of Dirichlet improvable vectors,
defined as the set of $\ux$ for which \eqref{eq:1} is soluble
for some $c<1$ and all large $Q$ (we let $\Vert.\Vert$ be
the maximum norm here). We state
an open problem.
\begin{problem}
Let $m\geq 2$. Decide whether the set $\mathcal{S}_m$ is meagre or not. What can we say topologically about $Di_m$?
\end{problem}
For $m=1$ we have $\mathcal{S}_1=\mathbb{Q}$ and $Di_1=\mathcal{S}_1\cup Bad_1$, where $Bad_m$ denotes the set of badly approximable vectors,
consisting of $\ux\in\Rm$ for which \eqref{eq:1} admits no
integer solution for some $c>0$ and any $Q>1$.
Since $Bad_m\subseteq W_m(\infty)^c$ are meagre, indeed so are
$Di_1$ and $\mathcal{S}_1$.
Generally $Di_m\supseteq \mathcal{S}_m\cup Bad_m$ holds~\cite[Theorem~2]{ds},
however for $m\geq 2$, the set $Di_m\setminus (Bad_m\cup \mathcal{S}_m)$
is non-empty~\cite{beretc},
and conjecturally has full Hausdorff dimension.
Hence no conclusion from Theorem~\ref{T2} can be drawn.
\subsection{Sets with remarkable sum and product sets}
We want to add another
result that is partly motivated
by the proof of Theorem~\ref{T2}, however some claims
heavily rely on
recent results from~\cite{art}. For $A,B\subseteq \Rm$
and $\circ\in\{ +,-,\cdot,/\}$ we denote by
\begin{align*}
A\circ B=\{ (a_1\circ b_1,\ldots,a_m\circ b_m):\;\; (a_1,\ldots,a_m)\in A,\; (b_1,\ldots,b_m)\in B\},
\end{align*}
their coordinatewise sum, difference, product and quotient (where we
restrict to $b_i\neq 0$ in the last case).
\begin{theorem} \label{C3}
Let $m\geq 1$. There exist sets $A,B\subseteq \Rm$ that
for any operation $\circ\in\{ +,-,\cdot,/\}$ satisfy
\begin{equation} \label{eq:sumse}
A\circ A=B\circ B=\Rm, \qquad \dim_H(A\circ B)<m.
\end{equation}
For $m\geq 5$, claim \eqref{eq:sumse} for $\circ\in\{ +,-\}$
holds in particular for $A=\mathcal{W}_m(\infty)$
and $B=\mathcal{S}_m$.
\end{theorem}
While it is not hard to construct $A,B$ with $A\pm A=B\pm B=\Rm$ and
$A\pm B\neq \Rm$, it suffices to take $A=\mathcal{W}_m^{\ast}(\infty)$
the set of Liouville
numbers and $B=A^c$ its complement,
we are unaware of any previous notice of
the stronger claim \eqref{eq:sumse} for $\circ=+$ in Corollary~\ref{C3}.
We want to formulate a problem regarding refinements of Theorem~\ref{C3}.
\begin{problem}
Do there exist sets $A,B\subseteq \Rm$ for which
\begin{equation} \label{eq:Stern}
A+A=B+B=\Rm, \qquad \dim_H(A+B)=0
\end{equation}
or
\begin{equation} \label{eq:Katz}
A+A=B+B=\Rm, \qquad \dim_P(A+B)<n
\end{equation}
holds?
\end{problem}
We cannot take packing dimension in \eqref{eq:Stern}.
If $A+A=\Rm$
then $\dim_P(A)\geq m/2$ in view of
$\dim_P(A+C)\leq \dim_P(A\times C)\leq \dim_P(A)+\dim_P(C)$
for all $A,C\subseteq\Rm$ (see~\cite{falconer}),
applied to $C=A$. Thus since clearly $B\neq \emptyset$
we also have $\dim_P(A+B)\geq \dim_P(A)\geq m/2$.
We cannot have both claims \eqref{eq:Stern}, \eqref{eq:Katz} simultaneously either, since by \eqref{eq:tricot}
\begin{align*}
m&=\dim_H(\Rm)=\dim_H(A+A)=\dim_H((A+A)+B)=\dim_H(A+(A+B))\\ &\leq \dim_H(A+B)+\dim_P(A)\leq \dim_H(A+B)+\dim_P(A+B).
\end{align*}
\section{Singular vectors in fractal sets in the homogeneous setting} \label{se4}
Singular vectors on fractals have recently received much attention.
We only mention a few important results.
Kleinbock and Weiss showed in~\cite{KW}
that for a wide class of fractals $K\subseteq \Rm$, the
set $\mathcal{S}_m\cap K$ is a nullset with respect to the natural
measure on $K$. On the other hand,
Kleinblock, Moshchevitin and Weiss~\cite{kmw} established that for a large class of fractals $K\subseteq \Rm$,
the set of singular vectors in $K$ is non-empty,
more precisely $\mathcal{S}_m^{\ast}(w)\cap K$
is uncountable for any $w<\frac{1}{m-1}$.
For $K$ the attractor of an iterated function system (upon some
regularity conditions),
upper bounds for the Hausdorff dimension of $\mathcal{S}_m\cap K$ were
obtained by Khalil~\cite{khalil}, contributing to a problem
posed by Bugeaud, Cheung and Chevallier~\cite[Problem~6]{bcc} who were
interested in $K\subseteq \mathbb{R}^2$ the Cartesian product of two copies of the Cantor
middle third set. We complement these results by lower bounds for the packing dimension of $\mathcal{S}_m^{\ast}\cap K$, for $K\subseteq \Rm$ certain
Cartesian products of classical missing digit Cantor sets. This generalizes the setting raised in~\cite{bcc} and is
also explicitly addressed in~\cite{khalil}. We must
restrict ourselves to the homogeneous case $\ut=\underline{0}$ here.
For $b\geq 2$ an integer and $W\subseteq \{0,1,\ldots,b-1\}$ with $|W|\geq 2$,
let $C_{b,W}\subseteq \mathbb{R}$ be the
missing digit Cantor set of real numbers that admit a
base $b$ representation
\[
\xi= \sum_{i\geq 1} a_ib^{-i}, \qquad a_i\in W.
\]
For the sets $C_{b,W}$ the Hausdorff and packing dimensions coincide
and take the value
\begin{equation} \label{eq:haupack}
\dim_H(C_{b,W})=\dim_P(C_{b,W})=\frac{\log |W|}{\log b},
\end{equation}
see Falconer~\cite{falconer}.
Below we just write $\dim(A)$ without index
for any set $A$ with this property, meaning either dimension.
We consider Cartesian products
\begin{equation} \label{eq:wiro}
K=C_{b,W_1}\times \cdots\times C_{b,W_m}
\end{equation}
with arbitrary $W_i\subseteq \{0,1,\ldots,b-1\}$, $|W_i|\geq 2$, and uniform $b$.
A well-known consequence of \eqref{eq:haupack} reads
\begin{equation} \label{eq:kpro}
\dim(K)= \sum_{i=1}^{m} \dim(C_{b,W})=\frac{\sum_{i=1}^{m} \log |W_i|}{\log b}.
\end{equation}
Indeed, this can be deduced
from
\begin{equation} \label{eq:kproS}
\dim_H(A) + \dim_H(B)\leq \dim_H(A\times B)\leq \dim_P(A\times B)\leq \dim_P(A) + \dim_P(B)
\end{equation}
valid for any Euclidean sets $A,B$
(see Falconer~\cite{falconer}). For brevity let us
write
\[
\mathcal{S}_K^{\ast}(w)= \mathcal{S}_m^{\ast}(w)\cap K,\quad \mathcal{S}_K^{(b)\ast}(w)= \mathcal{S}_m^{(b)\ast}(w)\cap K, \qquad\qquad w\in[0,1),
\]
and
\[
\mathcal{S}_K(w)= \mathcal{S}_m(w)\cap K,\quad \mathcal{S}_K^{(b)}(w)= \mathcal{S}_m^{(b)}(w)\cap K, \qquad\qquad w\in[0,1).
\]
The set $C_{b,W}$ supports a natural probability measure $\mu_{b,W}$,
essentially the restriction of the $(\log |W|/\log b)$-dimensional
Hausdorff measure to $C_{b,W}$,
and consequently the product measure $\mu=\mu_{b,W_1}\times \cdots\times \mu_{b,W_m}$
is the natural measure on $K$. Variants of Theorem~\ref{H} and Theorem~\ref{t3} for homogeneous approximation in $K$ read as follows.
\begin{theorem} \label{cantor}
Let $m\geq 1$, $b\geq 2$ be integers and $W_i$, $K$ and $\mathcal{S}_K(w), \mathcal{S}_K^{(b)}(w), \mathcal{S}_K^{\ast}(w)$ and $\mathcal{S}_K^{{(b)}\ast}(w)$ as above. Then the following two claims hold:
\begin{itemize}
\item[(i)] Assume $0\in W_i$ for $1\leq i\leq m$. Then for any $w\in[0,1)$ we have
\begin{align} \label{eq:csr}
\dim_P(\mathcal{S}_K^{\ast}(w))\geq \dim_P(\mathcal{S}_K^{(b)\ast}(w))
\geq (1-w)\dim(K)=
(1-w)\frac{\sum_{i=1}^{m} \log |W_i|}{\log b}>0.
\end{align}
In particular
\begin{equation} \label{eq:habicht}
\dim_P(\mathcal{S}_K^{\ast}(w))\geq \dim_P(\mathcal{S}_K^{(b)\ast}(w))\geq
\left(1-\frac{1}{m}\right)\dim(K)=\left(1-\frac{1}{m}\right)\frac{\sum_{i=1}^{m} \log |W_i|}{\log b}.
\end{equation}
For arbitrary $W_i$, for the non-$b$-ary sets still we have
the same estimate
\begin{equation} \label{eq:nonedda}
\dim_P(\mathcal{S}_K^{\ast}(w))\geq (1-w)\dim(K)=
(1-w)\frac{\sum_{i=1}^{m} \log |W_i|}{\log b}>0.
\end{equation}
\item[(ii)] Assume $0\in W_i$ for $1\leq i\leq m$. Let $w\in(0,1)$ and $w^{\prime}=(1-\sqrt{w})^2$
as in \eqref{eq:idne}. Then for any $w_0<w$ and $w_1<w^{\prime}$ the sets
\[
(\mathcal{S}_K^{\ast}(w_0) + \mathcal{S}_K^{\ast}(w_1))\cap K, \qquad
(\mathcal{S}_K^{(b)\ast}(w_0) + \mathcal{S}_K^{(b)\ast}(w_1))\cap K
\]
have full $\mu$-measure and we have the inclusion
\[
\mathcal{S}_K(w_0) + \mathcal{S}_K(w_1)\supseteq
\mathcal{S}_K^{(b)}(w_0) + \mathcal{S}_K^{(b)}(w_1)\supseteq K.
\]
\end{itemize}
\end{theorem}
\begin{remark}
Similar to Remark~\ref{rmark},
we believe that the stronger inclusion property in (ii)
holds for the sets restricted to $\Rr$ (denoted with asterisks $\ast$) as well. Similar to Remark~\ref{ureh}, without
restriction to $\Rr$ in (i), the non-$b$-ary claims are implied by the according estimate \eqref{eq:tr} up to a factor $\dim(K)$ on the right hand side, obtained by the same argument.
\end{remark}
\begin{remark}
The claim in particular implies the existence of
vectors in $K\cap \Rr$ singular to any order $w<1$, for $K$
as in \eqref{eq:wiro}.
In fact, we can also obtain $\wo_m(\ux)=1$ by a minor twist of the proof, however then we cannot guarantee positive packing dimension. This topic is discussed in~\cite[\S~1.6]{kmw}, see the first paragraph of this section.
We point out that a construction
of $\ux\in K\cap \Rr$ with any prescribed exponent $\wo_m(\ux)=\wo_m^{(b)}(\ux)\in[1/m,1]$, is implicitly contained
in~\cite[Theorem~2.5]{ichostrava}, upon
a suitable choice of involved $q_{i,j}$ and $\eta_i$ (see also~\cite[Corollary~2.11]{ichostrava} when $\wo_m(\ux)\in(1/2,1]$). However, no metrical claims are given in~\cite{ichostrava}.
The problem
if $\mathcal{S}_m^{\ast}(w)\cap K$
is non-empty for $w\geq \frac{1}{m-1}$ for $K$ more generally a product of perfect sets in which $\mathbb{Q}$ is dense treated in~\cite{kmw} remains open as well.
\end{remark}
The claims simplify to the homogeneous case of Theorem~\ref{H} and
Theorem~\ref{t3} if $W=\{0,1,\ldots,b-1\}$.
We enclose the special case of the product of two copies
of the Cantor middle
third set, as proposed in~\cite[Problem~6]{bcc}, and further
treat $\mathbb{R}\times C_{3,\{0,2\}}$.
\begin{corollary} \label{uh}
Let $K=C_{3,\{0,2\}}\times C_{3,\{0,2\}}$. Then
\[
\dim_P(\mathcal{S}_K^{\ast})\geq \dim_P(\mathcal{S}_K^{(b)\ast})\geq \frac{\dim(K)}{2}= \frac{\log 2 }{\log 3}.
\]
For $\tilde{K}= \mathbb{R}\times C_{3,\{0,2\}}$ we have
\[
\dim_P(\mathcal{S}_{\tilde{K}}^{\ast})\geq \dim_P(\mathcal{S}_{\tilde{K}}^{(b)\ast})\geq \frac{\dim(\tilde{K})}{2}=\frac{\log 2}{2\log 3}+ \frac{1}{2}.
\]
\end{corollary}
We also obtain a generalization of
Theorem~\ref{ax3} on Hausdorff dimension of singular vectors in the fractals, however the result is not as clean.
\begin{theorem} \label{khr}
Let $K=\prod C_{b,W_i}$ as above and assume $0\in W_i$ for all $1\leq i\leq m$. Write $d_i= \dim(C_{b,W_i})=\log |W_i|/\log b$ for $1\leq i\leq m$. Then $\dim_H(\mathcal{S}_K^{(b)\ast})$ is bounded from below by
the maximum of the function
\[
t\longmapsto \sum_{i=1}^{m} d_i \cdot
\frac{ -wt^2 + (w+1)t -1 }{ (w+1)(1-d_i) t^2 + ((w+2)d_i -1) t - d_i }
\]
over $t>1$.
If $K=C_{3,\{0,2\}}\times C_{3,\{0,2\}}$, then
\[
\dim_H(\mathcal{S}_K^{\ast})\geq \dim_H(\mathcal{S}_K^{(3)\ast}) > 0.1255.
\]
If we drop the assumption $0\in W_i$, the estimates remain
true for the non-$b$-ary sets $\mathcal{S}_K^{\ast}$.
\end{theorem}
If $d_1=d_2=\cdots=d_m=1$ it can be shown that the claim simplifies to Theorem~\ref{ax3}.
Theorem~\ref{khr} complements an estimate by Khalil~\cite{khalil} for the same set $K$ given by
\begin{equation} \label{eq:kalil}
\dim_H(\mathcal{S}_K) \leq
\frac{2}{3}\dim(K)= \frac{4\log 2}{3\log 3}.
\end{equation}
Some variant of Corollary~\ref{coro} can also be inferred
from Theorem~\ref{cantor} for general $K$ as in \eqref{eq:wiro},
however its statement turns out to be considerably weaker
(in many cases trivial) than the
expected analogous claim
\begin{equation} \label{eq:FR}
\dim_H(\mathcal{S}_K^{\ast}(w))\geq \dim_H(\mathcal{S}_K^{(b)\ast}(w))\geq (1-\frac{1}{m+1})\dim(K), \qquad w<(\sqrt{m}-1)^2/m,
\end{equation}
since we lack an
analogue of Theorem~\ref{DFSU} for Cantor sets.
Lower bounds for $\dim_P(\mathcal{S}_K^{(b)\ast}(w))$
could be obtained when combining Tricot's estimate
\eqref{eq:tricot} below, claim (ii) in Theorem~\ref{cantor} and~\cite[Theorem~A]{khalil},
upon evaluating the expressions $\alpha_{\ell}(\mu)$
defined in~\cite{khalil},
where $\mu$ is the natural measure on $K$.
This computation is only done explicitly for $m=2$ and $K=C_{3,\{0,2\}}\times C_{3,\{0,2\}}$ in~\cite[Corollary~1.4]{khalil}, from which
\eqref{eq:kalil} was deduced.
However, only for $m\geq 5$ the induced bounds are not implied
by Theorem~\ref{blau}, so \eqref{eq:kalil} is not helpful in this matter. We further mention
that a general inhomogeneous, $b$-ary version of Theorems~\ref{cantor},~\ref{khr} cannot hold as soon as $W\subsetneq \{0,1,\ldots,b-1\}$, as then
$\{ b^{N}\ux-\underline{p}-\ut: \ux\in K, \underline{p}\in\mathbb{Z}^m, N\geq 1\}$
avoids a neighborhood of $\underline{0}$
for any $\ut\subseteq [0,1)^m\setminus K$.
It remains however unclear to us for the sets $\mathcal{S}_{m,\ut}^{\ast}(w)\cap K$ where
we do not restrict to $b$-ary setting.
\section{Proofs}
\subsection{Preparation for metrical results} \label{sec2.1}
Our proofs below are based on a result by Tricot~\cite{tricot} on Cartesian products.
\begin{theorem}[Tricot]
Any measurable sets $A\subseteq \mathbb{R}^{d_1},B\subseteq \R^{d_2}$ satisfy
\begin{equation} \label{eq:tricot}
\dim_H(A\times B) \leq \dim_P(A) + \dim_H(B).
\end{equation}
\end{theorem}
See also Bishop and Peres~\cite{bishop}.
It is well-known that Lipschitz map $\Psi: \mathbb{R}^{d_1}\to \mathbb{R}^{d_2}$ satisfies
$\dim_H(\Psi(A))\leq \dim_H(A)$ for any $A\subseteq \mathbb{R}^{d_1}$,
see~\cite{falconer}. Applied to $d_1=2m, d_2=m$ and
the sum map $\Psi(A\times B)=A+B$ for $A,B\subseteq \Rm$ we see
$\dim_H(A+B) \leq \dim_H(A\times B)$.
Combining with \eqref{eq:tricot} we obtain
\begin{equation} \label{eq:thor}
\dim_P(A) \geq \dim_H(A+B) - \dim_H(B)
\end{equation}
for any measurable $A,B\subseteq \Rm$. Moreover it is well-known (see again Falconer~\cite{falconer}) that
any measurable $A\subseteq \Rm$ satisfies
$\dim_H(A)\leq \dim_{P}(A)$.
Specializing to $A=B\subseteq \Rm$ and combining
with above estimates, we see that
any $A\subseteq \Rm$ satisfies
\begin{equation} \label{eq:AB}
\dim_P(A) \geq \frac{\dim_P(A) + \dim_H(A)}{2}\geq
\frac{\dim_H(A\times A)}{2}\geq \frac{\dim_H(A+A)}{2}.
\end{equation}
Due to our restriction to $\Rd$ in several claims, we
require the following lemma whose statement appears rather
technical but which admits a straightforward proof.
We formulate it in a general
Cantor set setting that allows us to apply it
to Theorem~\ref{cantor} as well as any any other instance.
Recall the notation $\mu=\prod \mu_{b,W_i}$
for the natural measure on $K=\prod C_{b,W_i}$.
\begin{lemma} \label{daslemma}
Let $\mathcal{A}\subseteq \mathbb{N}$ be infinite and denote its complement by
$\mathcal{B}=\mathbb{N}\setminus \mathcal{A}$. Let
$b\geq 2$ an integer and $W_i\subseteq \{0,1,\ldots,b-1\}$, $|W_i|\geq 2$ for $1\leq i\leq m$,
and $K$ as in \eqref{eq:wiro}. Given fixed sequences $(c_{i,n})_{n\geq 1}\in W_i^{\mathbb{N}}$ for $1\leq i\leq m$
and any given partial maps
\begin{align*}
&\chi_i: \mathcal{B}\to W_i, \\
&\chi_i(n)= c_{i,n}
\end{align*}
writing $\ux=(\xi_1,\ldots,\xi_m)\in K$ consider the diagonal map
\begin{align*}
&\varphi: \prod_{i=1}^{m} (C_{b,W_i}\setminus \mathbb{Q}) \to K, \\
&\varphi(\ux)=(\varphi_1(\xi_1),\ldots,\varphi_m(\xi_m)),
\end{align*}
with coordinate functions
\[
\varphi_i(\xi_i)= (0.e_{i,1}e_{i,2}\ldots )_b=\sum_{n\geq 1} \frac{e_{i,n}}{b^n}, \qquad 1\leq i\leq m,
\]
where if $\xi_i=(0.d_{i,1}d_{i,2}\ldots)_b$, $d_{i,n}\in W_i$, then
\[
e_{i,n}= \begin{cases}
d_{i,n}, \qquad n\in \mathcal{A},\\ c_{i,n}, \qquad n\in\mathcal{B}.
\end{cases}
\]
Then, if $\mathcal{T}$ is any countable union of affine hyperplanes in $\Rm$, then for almost
all $\ux\in K$ with respect to $\mu$ we have
$\varphi(\ux)\notin \mathcal{T}$.
\end{lemma}
The map $\varphi$ changes the $b$-ary digits of $\xi_i\in C_{b,W_i}$
to prescribed digits on some set $\mathcal{B}$ and preserves the digits
on the infinite complement set $\mathcal{A}$, in each coordinate.
The special case $W_i=\{0,1,\ldots,b-1\}$ for $1\leq i\leq m$
gives general $\ux\in K=\Rm$.
The restriciton to irrational $\xi_i$ is necessary only
if $\{0,b-1\}\subseteq W_i$, but if true then certain rational
numbers have two different base $b$
expansions, and then indeed our map would not be well-defined.
However, from a metrical point of view, neglecting countably many
affine hyperplanes does not make a difference for our claim.
\begin{proof}[Proof of Lemma~\ref{daslemma}]
We proceed by induction on $m$. Let $m=1$. Then $\mathcal{T}$
consists of countably many singleton points.
Consider the set $\varphi^{-1}(\mathcal{T})$ defined
on $\mathcal{T}\cap \varphi(K)$,
which consists of numbers in $K$ with the same digital representation
of some element in $\mathcal{T}$ within intervals $\mathcal{A}$
and arbitrary digits in $W=W_1$ at the remaining places in $\mathcal{B}$.
Since $\mathcal{A}$ is infinite and choosing a fixed base
$b$ digit within $W$ at some given position shrinks the $\mu_1=\mu_{b,W_1}$ measure of a set by a factor $|W|^{-1}<1$, the preimage of any singleton has $\mu_1$ measure $0$.
Thus, as $\mathcal{T}$ is only countable, the set $\varphi^{-1}(\mathcal{T})$
has $\mu_1$ measure $0$. Hence
$K\setminus \varphi^{-1}(\mathcal{T})$ has full measure, in other
words for some full measure set $F\subseteq K$ we have $\varphi(F)\subseteq \mathcal{T}^c$.
The case $m=1$ follows.
We explain the induction step from $m-1$ to $m$. Recall that $\mu=\prod \mu_i$
where $\mu_i:=\mu_{b,W_i}$ is the natural Cantor measure on $C_{b,W_i}\subseteq \mathbb{R}$.
Since $\mathcal{T}$ consists of only countably many affine
hyperplanes, there
are in particular only countably many of them of the form
$H_{\xi_1}: x_1=\xi_1$, i.e. consisting of vectors $(\xi_1,x_2,\ldots,x_m)$
with prescribed first coordinate $x_1$. Since affine hyperplanes
have $\mu$ measure $0$ and sigma-additivity of measures,
we can omit this countable union.
Now for any other $\xi_1$ with $H_{\xi_1}$ not among
the affine hyperplanes defining $\mathcal{T}$, the affine hyperplane
$H_{\xi_1}$ intersects any affine hyperplane in $\mathcal{T}$ in
a lower dimensional affine space.
We apply the induction hypothesis to any such $H_{\xi_1}$
to see that
$\mu_{(m-1)}(\{ x\in H_{\xi_1}: \varphi(x)\in \mathcal{T}\})=0$, where
$\mu_{(m-1)}=\mu_2\times \cdots\times \mu_m$.
Writing $\mu=\mu_1\times \mu_{(m-1)}$ and integrating
the constant $0$ function over
the first variable $\xi_1$ with respect to $\mu_1$,
we see that the for a full $\mu$ measure set $F$ we have $\varphi(F)\notin \mathcal{T}$, equivalent to the claim of the theorem.
\end{proof}
The proofs of our main results combine above observations,
in particular \eqref{eq:thor}, with
ideas on sumsets similar as in~\cite{arxiv}. Before
we provide a short proof of Theorem~\ref{t4}.
\subsection{Proof of Theorem~\ref{t4}}
By combination of~\cite[Theorem~1.7, Theorem~2.9 and Remark~1.8]{dfsu1}, for $m\geq 2$ and $w<1$ sufficiently close to $1$,
we have
\[
\dim_H(\mathcal{S}_m^{\ast}(w))\leq \epsilon, \qquad \dim_P(\mathcal{S}_m^{\ast}(w))=1.
\]
Hence, by \eqref{eq:thor} we have
\begin{align*}
\dim_H(\mathcal{S}_m^{\ast}(w)+\mathcal{S}_m^{\ast}(w))
\leq
\dim_P(\mathcal{S}_m^{\ast}(w))+\dim_H(\mathcal{S}_m^{\ast}(w))\leq 1+\epsilon,
\end{align*}
the right claim. For the left identity, the lower bound follows from considering the sum
of two rational non-parallel lines in $\Rm$
given by equations $x_j=r_jx_1$ resp.
$x_j=s_jx_1$ with $r_j, s_j\in\mathbb{Q}$, $2\leq j\leq m$,
which becomes a two-dimensional subspace of $\Rm$ consisting of vectors
with $\wo_m(\ux)\in \{ 1,\infty\}$ depending on
whether $\ux\in\mathbb{Q}^m$ or not. For
the upper bound, consider
for any rational affine subspace $\mathscr{V}\subseteq \Rm$
the projection of $\mathcal{S}_m(w)\cap \mathscr{V}$ to a space spanned
by coordinate axes that induce a maximum $\mathbb{Q}$-linearly
independent set (i.e. all other coordinates of $\ux\in \mathscr{V}$
are $\mathbb{Q}$-linear combinations, inducing a rational subspace).
The image consists of vectors in $\mathcal{S}_k^{\ast}(w)$ for some $1\leq k<m$
where the first claim applies. Clearly the projection is bi-Lipschitz for any fixed rational subspace and keeps thus the dimensions invariant.
By sigma-additivity of measures,
this
exhausts the Hausdorff/packing dimension of $\mathcal{S}_m(w)\cap \mathscr{V}$.
This argument shows that
\[
\dim_H(\mathcal{S}_m(w))\leq\dim_P(\mathcal{S}_m(w))\leq \max_{1\leq k\leq m} \dim_P(\mathcal{S}_k^{\ast}(w)).
\]
For $k=1$ this is clearly bounded from above by $1$,
and by our observations above for $2\leq k\leq m$ as well when $w$ is close enough to $1$.
Similarly as above via \eqref{eq:thor} we conclude
\[
\dim_H(\mathcal{S}_m(w)+\mathcal{S}_m(w))
\leq
\dim_P(\mathcal{S}_m(w))+\dim_H(\mathcal{S}_m(w))\leq 2.
\]
\subsection{Proof of Theorem~\ref{t3}}
The claims follow easily from the next lemma.
\begin{lemma} \label{le}
Let $\nu_0>1, \nu_1>1$ and $\Lambda=\nu_0\nu_1$.
If $w_0, w_1$ satisfy $w_i<y_i:=(\nu_i-1)/\Lambda$
for $i=0,1$, then
we have the identity of sets
\begin{equation} \label{eq:G}
\mathcal{S}_{m,\ut_0}^{(b)}(w_0) + \mathcal{S}_{m,\ut_1}^{(b)}(w_1) =
\mathcal{S}_{m,\ut_0}(w_0) + \mathcal{S}_{m,\ut_1}(w_1)=\Rm.
\end{equation}
Morever, the sets
\begin{equation} \label{eq:G5}
\mathcal{S}_{m,\ut_0}^{(b)\ast}(w_0) + \mathcal{S}_{m,\ut_1}^{(b)\ast}(w_1), \qquad
\mathcal{S}_{m,\ut_0}^{\ast}(w_0) + \mathcal{S}_{m,\ut_1}^{\ast}(w_1)
\end{equation}
still have full $m$-dimensional Lebesgue measure.
\end{lemma}
Assume the lemma is true. Then putting
\[
\nu_0=\frac{1+\sqrt{w}}{1-w}>1, \qquad \nu_1=\frac{1}{\sqrt{w}}>1,
\]
a calculation verifies $y_0=w$, $y_1=w^{\prime}$ and thus the sets
in \eqref{eq:G0} equal $\Rm$, and those in \eqref{eq:fullm} still
have full measure.
To finish the proof of the theorem,
we verify Lemma~\ref{le}.
Let $\ux=(\ux_1,\ldots,\ux_m)\in\Rm$ arbitrary.
For $i=0,1$,
we have to find $\underline{x}_i\in \mathcal{S}_{m,\ut_i}^{(b)}(w_i)$
for every $w_i<y_i$,
that sum up to $\ux$, with $y_i=(\nu_i-1)/\Lambda$
where $\Lambda=\nu_0\nu_1$. The full measure claim will
be treated later.
We first assume the homogeneous case
$\ut_1=\ut_1=\underline{0}$ and later explain
how to pass to general $\ut_i\in\Rm$.
Define a sequence of intervals $I_j=[g_j,h_j]\cap \mathbb{Z}$, $j\geq 0$, that partitions the positive integers, inductively as follows: Take $M>0$ a large integer. Then set $g_0=1, h_0=M, g_1=M+1$ and for $j\geq 1$ let
\[
h_j= \lfloor \nu_i g_j\rfloor, \qquad g_{j+1}=h_j+1,
\]
where $i\in\{0,1\}$ is so that $j\equiv i\bmod 2$. Notice $h_j/g_j=\nu_i+o(1)$.
Let $b\geq 2$ be any integer.
Now consider the vector $\underline{x}_0=(x_{0,1},\ldots,x_{0,m})\in\Rm$
derived from $\ux$ as follows. For each $x_{0,\ell}$ the base $b$ representation
has digit has $0$ in intervals $I_j$ for even $j$ and the same digits
as $\xi_{\ell}$ for odd $j$, for $1\leq \ell\leq m$. Conversely, let $\underline{x}_1=(x_{1,1},\ldots,x_{1,m})$ where each $x_{1,\ell}$
has base $b$ digit has $0$ in intervals $I_j$ for odd $j$ and the same digits
as $\xi_{\ell}$ for even $j$. Obviously $\underline{x}_0+\underline{x}_1=\ux$.
We must show that $\underline{x}_i\in \mathcal{S}_{m}^{(b)}(w_i)$ for every $w_i<y_i$, to finish the proof of \eqref{eq:G0}
for the homogeneous case.
Let $Q=b^R$ be any large parameter. For $i=0,1$,
let $j=j(Q,i)$ be the largest index
with $j\not\equiv i\bmod 2$ so that $R\geq h_{j}$.
Thus $b^{h_j}\leq Q\ll b^{h_{j+2}}\ll b^{\Lambda h_j}$.
By construction, the first non-zero base $b$ digit
of $b^{h_j}x_{i,\ell}$ after the comma is not before position $h_{j+1}-h_{j}$,
in each coordinate $1\leq \ell\leq m$.
Hence, if we define for $i=0,1$
the integer vectors $\underline{p}_j=\lfloor b^{h_j}\underline{x}_i\rfloor$,
where we mean that the floor function is applied in each coordinate, then
\[
\Vert b^{h_j}\underline{x}_i - \underline{p}_j\Vert \ll b^{-(h_{j+1}-h_j) } \ll
b^{-(\nu_i-1) h_j }\ll Q^{-\frac{\nu_i-1}{ \Lambda }}, \qquad i=0,1.
\]
Since $Q$ was arbitrary indeed $\wo_m^{(b)}(\underline{x}_i)\geq (\nu_i-1)/\Lambda=y_i$, hence $\underline{x}_i\in \mathcal{S}_m^{(b)}(w_i)$ for the induced
$\underline{x}_i$ and any $w_i<y_i$, as required.
The full measure claim \eqref{eq:fullm}
of the theorem follows from Lemma~\ref{daslemma} applied
to $\mathcal{A}$ the union over intervals $I_j$
with $j\not\equiv i\bmod 2$ and the complement $\mathcal{B}$ the union
over those $I_j$ with $j\equiv i\bmod 2$, and $K=\Rm$.
Indeed, it directly implies that Lebesgue almost all $\ux\in\Rm$
in fact induce $\underline{x}_i\in \mathcal{S}_{m,\ut_i}^{(b)\ast}(w_i)$,
$i=0,1$.
Now we explain the modification for general $\ut_i=(\theta_{i,1},\ldots,\theta_{i,m})$, $i=0,1$. Instead of
letting the base $b$ digits of the $x_{i,\ell}$, $1\leq \ell\leq m$, be $0$ in the respective intervals for $j\equiv i\bmod 2$,
in these intervals we follow the base $b$ representation of $\theta_{i,\ell}$ from its start.
We give more details. Without loss of generality assume
$\ut_i\in[0,1)^m, \ux_i\in[0,1)^m$ and let
\[
\theta_{i,\ell}= (0.a_{i,1}^{\ell}a_{i,2}^{\ell}\ldots)_b=
\frac{a_{i,1}^{\ell}}{b}+ \frac{a_{i,2}^{\ell}}{b^2} + \cdots , \qquad\qquad i\in\{ 0,1\},\;\; 1\leq \ell\leq m,
\]
and
\[
\xi_{i,\ell}= (0.d_{i,1}^{\ell}d_{i,2}^{\ell}\ldots)_b= \frac{d_{i,1}^{\ell}}{b}+ \frac{d_{i,2}^{\ell}}{b^2} + \cdots , \qquad\qquad i\in\{ 0,1\},\;\; 1\leq \ell\leq m,
\]
be the base $b$ representation of $\theta_{i,\ell}$ and
$\xi_{i,\ell}$, respectively.
For $i=0,1$ we define $\underline{\alpha}_i=(\alpha_{i,1},\ldots,\alpha_{i,m})$
with coordinates $\alpha_{i,\ell}\in\mathbb{R}$ for $1\leq \ell\leq m$
defined as follows.
For every $j\equiv i\bmod 2$,
in $I_j=\{ g_j, g_j +1 , \ldots, h_j\}$ as above, we put
the base $b$ digit of $\alpha_{i,\ell}$ at any position $g_j+u-1\in I_j$
equal to the $u$-th base $b$
digit $a_{i,u}^{\ell}$ of $\theta_{i,\ell}$.
Define further for $i=0,1$ and $1\leq \ell\leq m$ the rational numbers
\[
r_{j}^{i,\ell}= \frac{ d_{i,g_j}^{\ell}b^{h_{j-1}} + \cdots+ d_{i,h_j}^{\ell}b^{h_j} }{b^{h_j}}- \frac{ a_{i,1}^{\ell}b^{h_{j-1}} + \cdots+ a_{i,|I_j|}^{\ell}b^{h_j} }{b^{h_j}}\in\mathbb{Q},\quad\; j\geq 1.
\]
Note that $r_{j}^{i,\ell}$ has common denominator $b^{h_j}$.
To the numbers $\alpha_{i,\ell}$ we add the sum of
$r_j^{1-i,\ell}$ over each $j\equiv i\bmod 2$ to derive $x_{i,\ell}$, i.e.
\[
x_{i,\ell}= \alpha_{i,\ell}+ \sum_{j\geq 1,\\ j\equiv i\bmod 2} r_{j}^{1-i,\ell}=\alpha_{i,\ell}+ \sum_{j\geq 1,\\ j\not\equiv i\bmod 2} r_{j}^{i,\ell}, \qquad \quad i=0,1, \; 1\leq \ell\leq m.
\]
If we write $\underline{x}_i=(x_{i,1},\ldots,x_{i,m})$, then by construction $\underline{x}_0+\underline{x}_1=\ux$.
Moreover, defining $Q$ and the induced $j$ as above, similarly as above
the base $b$ digits of $b^{h_j}\underline{x}_i-\underline{p}_{i,j}$
for $\underline{p}_{i,j}=\lfloor b^{h_j}\underline{x}_i-\ut_i\rfloor$
in any coordinate vanish up to position $h_{j+1}-h_j$.
Thus we have
\[
\Vert b^{h_j}\underline{x}_i - \underline{p}_{i,j}- \ut_i\Vert \ll b^{-(h_{j+1}-h_j) } \ll
b^{-(\nu_i-1) h_j }\ll Q^{-\frac{\nu_i-1}{ \Lambda } },
\]
for $i\in\{0,1\}$, so again $\wo_{m,\ut_i}^{(b)}(\underline{x}_i)\geq (\nu_i-1)/\Lambda=y_i$ and thus, as above $\underline{x}_i\in \mathcal{S}_{m,\ut_i}^{(b)}(w_i)$ for any $w_i<y_i$.
For the full measure measure claim we again apply Lemma~\ref{daslemma}.
Lemma~\ref{le} and thus the theorem are proved. We mention
that the special case \eqref{eq:B} of Corollary~\ref{mko}
corresponds to $\nu_0=\nu_1=2$.
\subsection{Proof of Corollary~\ref{coro}} \label{pko}
Let $w=1/m+\varepsilon_0$ and
$w^{\prime}=w+1-2\sqrt{w}=(\sqrt{m}-1)^2/m-\varepsilon_1$ for small $\varepsilon_0>0, \varepsilon_1>0$,
so that identity \eqref{eq:idne}
in Theorem~\ref{t3} holds.
Now the extension~\cite[Theorem~1.5]{dfsu1} of
Theorem~\ref{DFSU} implies that $w\to \dim_P(\mathcal{S}_m^{\ast}(w))$ is right-continuous
at $w=1/m$. Hence, since the $\varepsilon_i$ can be chosen
arbitrarily small, from the sets in \eqref{eq:fullm} having full measure and a rearrangement of \eqref{eq:thor} we get
\begin{align*}
\dim_H(\mathcal{S}_{m}^{(b)\ast}(w_1)) &\geq m - \dim_P(\mathcal{S}_m^{(b)\ast})\\ &\geq
m - \dim_P(\mathcal{S}_m)=
m - (m-1+\frac{1}{m+1})= 1-\frac{1}{m+1},
\end{align*}
for any $w_1<w^{\prime}$, the claim of the lemma.
\subsection{Proof of Theorem~\ref{H}}
We modify the construction of the proof of Theorem~\ref{t3}.
Let $w^{\prime}>0$ and $\nu_0>1, \nu_1>1$ be real numbers to be
chosen later related by the identity
\begin{equation} \label{eq:satisfy}
w^{\prime}=\frac{\nu_1-1}{\nu_0\nu_1}.
\end{equation}
Take any $w<w^{\prime}$. We show that the set
\begin{equation} \label{eq:RITZ}
\mathcal{S}_{m,\ut}^{(b)\ast}(w) + \mathcal{W}_m^{(b)}(\nu_0-1)
\end{equation}
has full $m$-dimensional Lebesgue measure (in fact removing the $\ast$
the sumset equals $\Rm$).
Provided this is true, from \eqref{eq:thor} and formula \eqref{eq:jarnik} we get
\begin{align*}
\dim_{P}( \mathcal{S}_{m,\ut}^{(b)\ast}(w) )
\geq \dim_H(\mathcal{S}_{m,\ut}^{(b)\ast}(w) + \mathcal{W}_m^{(b)}(\nu_0-1)) - \dim_H(\mathcal{W}_m^{(b)}(\nu_0-1)) \geq m- \frac{m}{\nu_0}.
\end{align*}
If we choose $\nu_1$ large enough then $\nu_0=(\nu_1-1)/(w^{\prime}\nu_1)$ will be arbitrarily close
to $w^{\prime -1}$, and since $w$ can be taken arbitrarily
close to $w^{\prime}$, the lower bound $m-mw$ follows for $w>0$.
Finally, if $w=0$, it follows from an obvious inclusion argument.
To show \eqref{eq:RITZ},
we again first restrict to the homogeneous case $\ut=\underline{0}$,
and describe the generalization later.
Take an arbitrary
vector $\ux=(\xi_1,\ldots,\xi_m)\in \mathbb{R}^{m}$.
Consider the intervals $I_j$ partitioning the positive integers
constructed as follows:
Set $g_0=1, h_0=M, g_1=M+1$ and for $j\geq 1$ let
\[
h_j= \lfloor \nu_i g_j\rfloor, \qquad g_{j+1}=h_j+1,
\]
where $i\in\{0,1\}$ is taken so that $j\equiv i\bmod 2$, i.e.
the same parity as $j$. Notice $h_j/g_j=\nu_i+o(1)$.
Now define $\underline{x}_0=(x_{0,1},\ldots,x_{0,m})$ as follows:
For $1\leq \ell\leq m$, let the digits of $x_{0,\ell}$ be $0$
in intervals $I_j$ with $j\equiv 0\bmod 2$, and equal to those of $\xi_{\ell}$
if $j\equiv 1 \bmod 2$.
Conversely, let the coordinates $x_{1,\ell}$ of $\underline{x}_1=(x_{1,1},\ldots,x_{1,m})$
have digit $0$ in intervals $I_j$ when $j\equiv 1\bmod 2$, and equal to those of $\xi_{\ell}$
if $j\equiv 0 \bmod 2$, for $1\leq \ell\leq m$.
Then clearly $\ux=\underline{x}_0+\underline{x}_1$.
To finish the proof of the homogeneous case, we need to show that
\begin{equation} \label{eq:OP}
\wo_m^{(b)}(\underline{x}_1) \geq w^{\prime}, \qquad \omega_m^{(b)}(\underline{x}_0)\geq \nu_0-1,
\end{equation}
as by definition $\underline{x}_0\in \mathcal{W}_m^{(b)}(\nu_0-1)$ and
$\underline{x}_1\in \mathcal{S}_m^{(b)}(w)$ for any $w<w^{\prime}$.
By Lemma~\ref{daslemma} we still have $\underline{x}_1\in \mathcal{S}_m^{(b)\ast}(w)$ for any $w<w^{\prime}$ induced by almost all $\ux\in\Rm$ we started with.
We proceed similar to the proof of Theorem~\ref{t3}. Let $Q=b^{R}$ be large and $j$ be the largest
even integer such that $b^{h_j}\leq Q$. Then $Q\ll b^{h_{j+2}}\ll b^{\Lambda h_j}$, where $\Lambda=\nu_0\nu_1$.
It is again not hard to see that again not before position $h_{j+1}-h_j$
non-zero digits of $b^{h_j}\underline{x}_1$ after the comma occur, hence writing $\underline{p}_{1,j}= \lfloor b^{h_j}\underline{x}_1\rfloor$, we infer
\[
\Vert b^{h_j}\underline{x}_1 - \underline{p}_{1,j}\Vert \ll b^{-(h_{j+1}-h_j) } \ll
b^{-(\nu_1-1) h_j }\ll Q^{-\frac{\nu_1-1}{ \Lambda }}=Q^{-w^{\prime}}.
\]
This shows the left inequality in \eqref{eq:OP}. Similarly for odd $j$ with $\underline{p}_{0,j}= \lfloor b^{h_j}\underline{x}_0\rfloor$ we have
\[
\Vert b^{h_j}\underline{x}_0 - \underline{p}_{0,j}\Vert \ll b^{-(h_{j+1}-h_j) } \ll
b^{-(\nu_0-1) h_j }= (b^{h_j})^{-(\nu_0-1)}.
\]
This shows the right inequality in \eqref{eq:OP}.
Finally, if $\ut$ does not vanish, in the intervals $I_j$
with odd $j$, we follow with $\underline{x}_1$
the initial $b$-ary expansion of $\ut$ instead of letting
the digits be $0$,
and then alter $\underline{x}_0$ accordingly in the same intervals
to preserve $\underline{x}_0+\underline{x}_1=\ux$, very similarly
to the proof of Theorem~\ref{t3}. We conclude likewise that
$\underline{x}_0\in \mathcal{W}_m^{(b)}(\nu_0-1)$ and almost all $\ux\in\Rm$ induce $\underline{x}_1\in \mathcal{S}_{m,\ut}^{(b)\ast}(w)$,
thus the set in \eqref{eq:RITZ} has full measure in the general case.
\begin{remark} \label{hirsch}
Define $b$-ary sets
$\widetilde{\mathcal{S}}_{m,\ut}^{(b)\ast}(w)\subseteq \mathcal{S}_{m,\ut}^{(b)\ast}(w)$
of exact singular order $w$ by
\[
\widetilde{\mathcal{S}}_{m,\ut}^{(b)\ast}(w)= \mathcal{S}_{m,\ut}^{(b)\ast}(w) \setminus \bigcup_{t>w} \mathcal{S}_{m,\ut}^{(b)\ast}(t)= \{ \ux\in \mathcal{S}_{m,\ut}^{(b)\ast}(w): \wo_{m,\ut}^{(b)\ast}(\ux)=w \}.
\]
It is not hard to show that the sumset
$\widetilde{\mathcal{S}}_{m,\ut}^{(b)\ast}(w) + \mathcal{W}_m^{(b)}(\nu_0-1)$ with $\nu_0$
as in the proof above still contains a full measure
subset of the set
$\widetilde{\mathcal{W}}_m^{(b)}(0)=\{ \ux\in\Rm: \om_m^{(b)}(\ux)=0 \}$ (and dropping the ''$\ast$'' the entire set, proceed
similar to the proof of \cite[Theorem~3.1]{arxiv}).
The latter set has full $m$-dimensional
Lebesgue measure as follows from more general results of Fraenkel~\cite{franky} (see
also \eqref{eq:jarnik} obtained in~\cite{bf72}). Thus the claim of Theorem~\ref{H} for the smaller sets $\widetilde{\mathcal{S}}_{m,\ut}^{(b)\ast}(w)$
follows from the same line of arguments in the proof above.
However, it seems
considerably more difficult to get analogous results for the accordingly altered non-$b$-ary sets $\widetilde{\mathcal{S}}_{m,\ut}^{\ast}(w)$. Similarly,
other claims may be extended to exact order of approximation.
\end{remark}
\subsection{Proof of Theorem~\ref{simu}}
Let $\ux\in\Rm$ and $\Theta=\{ \ut_1,\ldots,\ut_k \}$ with $\ut_s\in\Rm$
for $1\leq s\leq k$
arbitrary and $\nu_0>1, \nu_1>1$ real numbers.
For $i=0,1$ and $j\geq 1$
define again iteratively a partition of the positive integers
into intervals $I_j=[g_j,h_j]\cap \mathbb{Z}$ with $h_j= \lfloor\nu_i g_j\rfloor$
and $g_{j+1}=h_j+1$,
where $i=1$ if $j\not\equiv 0\bmod k$ and $i=0$
if $j\equiv 0\bmod (k+1)$.
Then assume in $I_j$ the real vector
$\underline{x}_1$ follows the base $b$ representation
of $\ut_s$ where $j\equiv s\bmod k$ with $s\in\{ 1,2,\ldots,k\}$
for $j\not\equiv 0\bmod (k+1)$,
and the digits are those of $\ux$ in intervals
with $j\equiv 0\bmod (k+1)$.
Define $\underline{x}_0=\ux-\underline{x}_1$ and note
that it has $0$ base $b$ digits in $I_j$ for $j\equiv 0\bmod (k+1)$. Then for any $1\leq s\leq k$ and any large $Q_s$,
let $j=j(s,Q_s)$ be the largest index that satisfies
$j+1\equiv s\bmod (k+1)$
and $b^{h_{j}}\leq Q$. Then with $\Lambda=\nu_0\nu_1^k$
and $\underline{p}_{1,j}=\lfloor b^{h_j}\underline{x}_1\rfloor$,
we have $b^{h_j}\leq Q<b^{h_{j+k+1}}\ll b^{\Lambda h_j }$
and we check similar to the proof above that
\[
\Vert b^{h_j}\underline{x}_1-\underline{p}_{1,j}-\ut_s\Vert \ll
b^{-(h_{j+1}-h_j)} \ll b^{-(\nu_1-1)h_j}
\ll Q^{-\frac{\nu_1-1}{\Lambda}}.
\]
Moreover, with
$\underline{p}_{0,j}=\lfloor b^{h_j}\underline{x}_0\rfloor$,
for any $j\equiv -1\bmod (k+1)$ we have
\[
\Vert b^{h_j}\underline{x}_0-\underline{p}_{0,j}\Vert \ll
b^{-(h_{j+1}-h_j)} \ll (b^{h_j})^{-(\nu_0-1)}.
\]
By construction
$\wo_m^{(b)}(\underline{x}_1,\ut_s)\geq (\nu_1-1)/\Lambda= w^{\prime}$
for any $s\in\{ 1,2,\ldots,k\}$. Thus
we have $\underline{x}_1\in \cap_{\ut\in\Theta} \mathcal{S}_{m,\ut}^{(b)}(w)$ for any $w<w^{\prime}$, and by
Lemma~\ref{daslemma} still $\underline{x}_1\in \cap_{\ut\in\Theta} \mathcal{S}_{m,\ut}^{(b)\ast}(w)$ for almost all $\ux$ we started with.
Finally choosing $\nu_1=\frac{k}{k-1}$ and if we identify
\[
w^{\prime}=\frac{\nu_1-1}{\Lambda}
\]
we also have
\[
\om_m^{(b)}(\underline{x}_0)\geq \nu_0-1=\frac{\nu_1-1}{w^{\prime}\nu_1^k}-1=\frac{(k-1)^{k-1}}{ w^{\prime}k^k}-1>\frac{1}{kew^{\prime}}-1,
\]
where we used the inequality
\[
(1-\frac{1}{k})^{k-1} > e^{-1}, \qquad k\geq 2 \quad\; \Longleftrightarrow \quad\;
(1+\frac{1}{k-1})^{k-1}<e, \qquad k\geq 2,
\]
a well-known fact. Hence if we write
\[
A=\mathcal{W}_m^{(b)}((kew^{\prime})^{-1}-1),\qquad B=\bigcap_{\ut\in \Theta} \mathcal{S}_{m,\ut}^{(b)\ast}(w),
\]
then again $A+B$ has full $m$-dimensional Lebesgue measure for any $w<w^{\prime}$, and
by \eqref{eq:thor} and formula \eqref{eq:jarnik} we infer
\begin{align*}
m=\dim_H(A+B)\leq \dim_H(A)+\dim_P(B)\leq mkew^{\prime}+ \dim_P(B).
\end{align*}
By a continuity argument again we may write $w$ instead of $w^{\prime}$, and subtracting $mkew$ from both sides the proof is finished.
\subsection{Proof of Theorem~\ref{heep} }
Let $h_j/g_j\to\infty$ and $g_{j+1}=h_j+1$ and define again the sequence
of neighboring intervals $I_j=[g_j,h_j]\cap \mathbb{Z}$
that partitions $\mathbb{N}$. Further take any map
$\varphi: \mathbb{N}\to \mathbb{N}$ such that each $s\in\mathbb{N}$
has infinitely many preimages both among
even and odd $j$, for example by $\{ \varphi(1), \varphi(2),\ldots \}=\{1,1,2,1,2,3,1,2,3,4,1,\ldots\}$. Then, for any $s\geq 1$,
assume $\underline{x}_0$ follows the base $b$
representation of $\ut_s$ in an interval $I_j$ whenever $\varphi(j)=s$
and $j$ is even. Similarly, let $\underline{x}_1$ follow
when $j$ is odd and $\varphi(j)=s$.
For given $\ux\in\Rm$, define the digits in the remaining intervals
so that $\underline{x}_0+\underline{x}_1=\ux$, very similar
to the proof of Theorem~\ref{t3}. Then
for arbitrarily large $N$, any $s\geq 1$ and $j\in \{ \varphi^{-1}(s)-1\}$ with $j$ large enough, for $p_{i,j}=\lfloor b^{h_j}\underline{x}_i \rfloor$
we have
\[
\Vert b^{h_j}\underline{x}_i-\underline{p}_{i,j}-\ut_s\Vert
\ll b^{-(h_{j+1}-h_j) }\ll (b^{h_j})^{-N}, \qquad i=0,1.
\]
Hence $\om_{m,\ut_s}^{(b)}(\underline{x}_i)=\infty$ for $i=0,1$ and all
$s\geq 1$. Thus
with $A=\cap_{\ut\in\Theta} \mathcal{W}_{m,\ut}^{(b)}(\infty)$ we have $A+A=\Rm$. We may without loss of generality assume $\underline{0}\in\Theta$
so that $\dim_H(A)=0$ by \eqref{eq:jarnik}.
Consequently from \eqref{eq:thor} we obtain
\[
m=\dim_H(\Rm)=\dim_H(A+A)\leq \dim_H(A)+ \dim_P(A)= \dim_P(A),
\]
hence $\dim_P(A)=m$.
\subsection{Proof of Theorem~\ref{reverse}}
The claim follows from combination of two results below.
\begin{proposition} \label{prop}
For any $\xi$ and $\theta_i$, $i=1,2$, as in the theorem we have
\[
\wo_{1,\theta_i}^{(3)}(\xi)\leq 1, \qquad \om_{1,\theta_i}^{(3)}(\xi) \geq \frac{\wo_{1,\theta_i}^{(3)}(\xi)}{1-\wo_{1,\theta_i}^{(3)}(\xi)},
\]
where we interpret $1/0=\infty$.
\end{proposition}
\begin{proof}
We only show the claim for $i=1$, the other case is analogous.
We show the equivalent claim
\[
\wo_{1,\theta_1}^{(3)}(\xi) \leq \frac{\om_{1,\theta_1}^{(3)}(\xi)}{\om_{1,\theta_1}^{(3)}(\xi)+1}\leq 1.
\]
We may assume $\om_{1,\theta_1}^{(3)}(\xi)>0$, otherwise
the claim is clear by the trivial estimate $\wo_{1,\theta_1}^{(3)}(\xi)\leq \om_{1,\theta_1}^{(3)}(\xi)$.
By definition of $\om_{1,\theta_1}^{(3)}(\xi)$, for any $\tau\in (0,\om_{1,\theta_1}^{(3)}(\xi))$,
arbitrarily large integers $t$ and $p_1=p_1(t)$ we have
\begin{equation} \label{eq:tauin}
|3^t\xi-p_1-\theta_1| < (3^t)^{-\tau}.
\end{equation}
By an easy digital argument,
using that only at the beginning of the expansion of $\theta_1$ we have two consecutive base $3$ digits $1$,
for $u\in (t,\lfloor t (\tau+1)\rfloor]$ we
have $|3^{u}\xi-p_5-\theta_1|\gg 1$.
For $u\leq t$, arbitrarily small $\varepsilon>0$ and any integer $p_2$ we estimate
\[
|3^{u}\xi-p_2-\theta_1|\gg (3^{u})^{-(\om_{1,\theta_1}^{(3)}(\xi) +\varepsilon) }\geq
(3^{t})^{-(\om_{1,\theta_1}^{(3)}(\xi) +\varepsilon) }
\]
from the definition of $\om_{1,\theta_1}^{(3)}(\xi)$.
Hence, the last estimate holds for any $u\leq \lfloor t (\tau+1)\rfloor$. Hence
with $Q=3^{\lfloor t(\tau+1)\rfloor}$ we get
\[
\wo_{1,\theta_1}^{(3)}(\xi)\leq \frac{t(\om_{1,\theta_1}(\xi)+\varepsilon)}{t(\tau+1)}=
\frac{\om_{1,\theta_1}(\xi)+\varepsilon}{\tau+1}.
\]
Now $\tau$ can be chosen arbitrarily
close to $\om_{1,\theta_1}^{(3)}(\xi)$ and $\varepsilon$ arbitrarily
small and the claim follows.
\end{proof}
\begin{proposition} \label{pp1}
For $\theta_1, \theta_2$ as in the theorem and any $\xi\in\mathbb{R}$ we have
\[
\wo_{1,\theta_2}^{(3)}(\xi)\leq \frac{1}{\om_{1,\theta_1}^{(3)}(\xi)+1}\leq 1.
\]
\end{proposition}
\begin{proof}
We may assume $\om_{1,\theta_1}(\xi)>0$ as otherwise the claim
follows from Proposition~\ref{prop}.
By definition, for any $\tau\in (0,\om_{1,\theta_1}^{(3)}(\xi))$
and arbitrarily large integers $t, p_1=p_1(t)$ we have
\eqref{eq:tauin}.
Then the base $3$ digits of $\xi$ at
places $t+1,t+2,\ldots, \lfloor t(\tau+1)\rfloor$ are the
same as the initial digit sequence of $\theta_1$, which
has digit $1$ at places of the form $N!$ and $0$ otherwise.
Thus, since $\theta_2$ starts with very different
digits $(0.220)_{3}$ in base $3$,
it is easy to see that for any integer $u\in [t,\lfloor t (\tau+1)\rfloor]$ and any integer $p_2$ we have
\[
|3^u\xi-p_2-\theta_2|\gg 1.
\]
For similar reasons, if $u<t$ then for any integer $p_3$ we have
\[
|3^u\xi-p_3-\theta_2|\gg 3^{u-t}.
\]
Hence for $u\leq \lfloor t(\tau+1)\rfloor$ we have $\vert 3^u\xi-p_4-\theta_2|\gg 3^{u-t}\gg 3^{-t}$,
thus considering again $Q=3^{\lfloor t(\tau+1)\rfloor}$ we get $\wo_{1,\theta_2}^{(3)}(\xi)\leq \frac{t}{t(\tau+1)}=\frac{1}{\tau+1}$.
The claim follows as $\tau$ can be chosen arbitrarily
close to $\om_{1,\theta_1}^{(3)}(\xi)$.
\end{proof}
We deduce the first claim of the theorem by combining the propositions via
\[
\wo_{1,\theta_2}^{(3)}(\xi)\leq \frac{1}{\om_{1,\theta_1}^{(3)}(\xi)+1}\leq \frac{1}{\frac{\wo_{1,\theta_1}^{(3)}(\xi)}{1-\wo_{1,\theta_1}^{(3)}(\xi)}+1}= 1-\wo_{1,\theta_1}^{(3)}(\xi).
\]
The latter claim of the theorem
follows since by the first not both $\wo_{1,\theta_i}^{(3)}(\xi)$
can exceed $1/2$.
\subsection{Proof of Theorem~\ref{cantor}}
Additionally to Lemma~\ref{daslemma},
we further require upper bounds on the Hausdorff dimension
of the set of $b$-ary, ordinary $\tau$-approximable vectors in $K$.
A supposedly sharp bound reads as follows.
\begin{lemma} \label{lemus}
Let $m\geq 1$ be an integer and $K$ as in \eqref{eq:wiro}. We have
\[
\dim_H(\mathcal{W}_m^{(b)}(\tau)\cap K) \leq \frac{\dim(K)}{\tau+1}= \frac{\sum_{i=1}^{m} \log |W_i|}{(\tau+1)\log b}, \qquad \tau\in[0,\infty].
\]
\end{lemma}
The proof of Lemma~\ref{lemus} follows from a very standard Borel-Cantelli
covering approach, mimicking the short proof of the
special case $m=1$, $K=C_{3,\{0,2\}}$
in~\cite[\S~5]{lsv}. We leave the details to the reader.
\begin{remark}
Lemma~\ref{lemus} can alternatively proved with Proposition~\ref{falke} below
from~\cite{falconer}.
Showing equality in Lemma~\ref{lemus} is a considerably harder task. It was proved rigorously for $m=1$
and $K=C_{3,\{0,2\}}$ the Cantor middle third set by Levesley, Salp, Velani~\cite{lsv}. The right identity in \eqref{eq:jarnik} is
a special case
when we let $W=\{ 0,1,\ldots,b-1\}$.
\end{remark}
Now we prove Theorem~\ref{cantor}.
Assume first $0\in W_h$ for all $1\leq h\leq m$.
Starting with $\ux\in K$, then for $\ut=\underline{0}$
the contructions in Theorem~\ref{H} and Theorem~\ref{t3} obviously
lead to $\underline{x}_i\in K$, $i=0,1$, as well. Thus, following the proof of Theorem~\ref{H} and letting
\[
A(w)= \mathcal{S}_{m}^{(b)}(w) \cap K, \qquad B= \mathcal{W}_m^{(b)}(\nu_0-1) \cap K,
\]
with the same assumptions and notation,
we get the inclusion
\begin{equation} \label{eq:EE}
A(w)+B\supseteq K.
\end{equation}
We next show that for
$A^{\ast}(w)=A(w)\cap \Rr$ the restriction of $A(w)$ to $\Rr$, the set
$(A^{\ast}(w)+B)\cap K$
still has full Cantor measure, in particular
\begin{equation} \label{eq:FF}
\dim_H((A^{\ast}(w)+B)\cap K)= \dim(K).
\end{equation}
This is immediate from \eqref{eq:EE}
and Lemma~\ref{daslemma}
with $\mathcal{A}$ the union of the $I_j$ over $j\not\equiv i\bmod 2$,
as it precisely claims that for almost all $\ux\in K$ we started with
the constructed $\underline{x}_i$ do not lie in a rational
affine hyperplane.
Combining \eqref{eq:FF} with \eqref{eq:thor} and \eqref{eq:kpro}
we get
\begin{align*}
\dim_P(\mathcal{S}_m^{(b)\ast}(w)) &\geq \dim_H( A^{\ast}(w)+ B)- \dim_H(B)\\
&= \dim(K)-\dim_H(B)= \frac{\sum_{h=1}^{m} \log |W_h|}{\log b}-\dim_H(B).
\end{align*}
An upper bound for $\dim_H(B)$ comes from Lemma~\ref{lemus}.
Our claim \eqref{eq:csr} and thus also \eqref{eq:habicht}
in (i) follow directly with
identification $w^{-1}=\nu_0-1$ and letting $\tau=\nu_0-1$.
As similar argument based on
the proof of Theorem~\ref{t3} again yields
the inclusion in (ii), and with Lemma~\ref{daslemma} we get the
metrical claim in (ii).
Finally, if $0\notin W_h$ for some $h$, we reduce \eqref{eq:nonedda}
to the case $0\in W_h$ in \eqref{eq:csr}
via applying the rational shift $t\to t- \min W_h/(b-1)$ in coordinate $h$. It maps general $C_{b,W}$ into some
$C_{b,W^{\prime}}$ with $0\in W^{\prime}$ and $|W^{\prime}|=|W|$.
Moreover, it both keeps the
property of belonging to some set $\mathcal{S}_{m}^{\ast}(w)$
invariant and, as a bi-Lipschitz map, preserves the Hausdorff
and packing dimensions. The proof is finished.
\subsection{Proof of Theorem~\ref{ax3} }
We use~\cite[Example~4.6]{falconer}.
\begin{proposition}[Falconer] \label{falke}
Let $[0,1] = E_0 \supseteq E_1 \supseteq E_2 \supseteq \cdots$
be a decreasing sequence of sets, with each $E_k$ a union of a finite number of
disjoint closed intervals (called $k$-th level basic intervals), with each interval of
$E_{k-1}$ containing $m_{k}\geq 2$ intervals of $E_{k}$,
which are separated by gaps of length at least $\epsilon_k$,
with $0<\epsilon_{k+1}<\epsilon_{k}$ for each $k$,
which tend to $0$ as $k\to\infty$. Then the set
\[
F= \bigcap_{i\geq 1} E_i
\]
satisfies
\[
\dim_H(F)\geq \liminf_{k\to\infty} \frac{ \log(m_1m_2 \ldots m_{k-1})}{-\log m_k \epsilon_k}.
\]
\end{proposition}
Let $A\asymp B$ denote $A\ll B\ll A$.
Let $H_n=b^{h_{2n}}$ with $h_j$ as in the notation of the proof
of Theorem~\ref{H}. As in the proof of the homogeneous case
of Theorem~\ref{H},
we consider the set $\mathscr{Y}$, corresponding to coordinates of $\underline{x}_1$,
of real numbers where we can choose the base $b$ digits freely
in intervals $I_{2n}=[h_{2n}, \nu_0 h_{2n}]$, and take digit $0$ in the remaining
intervals of the form $I_{2n+1}=[\nu_0 h_{2n}, h_{2n+2}]$, where
$h_{2n+1}\asymp \nu_0 h_{2n}$ and $h_{2n+2}\asymp \nu_0\nu_1 h_{2n}$.
It fits the Cantor type interval construction of
Proposition~\ref{falke} with parameters
\[
m_k\asymp H_k^{\nu_0-1}, \qquad \epsilon_k\asymp H_k^{-\nu_0}.
\]
It yields after a short calculation via the geometric sum formula that
\[
\dim_H(\mathscr{Y})\geq \liminf_{k\to\infty}
\frac{ (\nu_0-1)\sum_{j=1}^{k-1} \frac{1}{(\nu_0 \nu_1)^j} h_k }{ h_k } =
\frac{\nu_0-1}{\nu_0\nu_1-1}.
\]
One may compare the method with the proof of~\cite[Theorem~2.3]{arxiv3} for more details.
On the other hand,
if $\nu_0, \nu_1$ are related by identity \eqref{eq:satisfy},
then the proof of Theorem~\ref{H} shows
$\mathcal{S}_m^{(b)}(w)\supseteq \mathscr{Y}^m=\mathscr{Y}\times \cdots\times \mathscr{Y}$, and with repeated application of \eqref{eq:kproS} and the aid of Lemma~\ref{daslemma}
still $\dim_H(\mathcal{S}_m^{(b)\ast}(w))\geq m\dim_H(\mathscr{Y})$.
Hence $\dim_H(\mathcal{S}_m^{(b)\ast}(w))\geq m\dim_H(\mathscr{Y})\geq m(\nu_0-1)/(\nu_0\nu_1-1)$, where we used that $w^{\prime}$ is arbitrarily close to $w$.
Upon \eqref{eq:satisfy}, which is basically $\nu_1= (1-\nu_0w)^{-1}$, we maximize this with choice of parameters
\[
\nu_0= \frac{2}{w+1},\qquad \nu_1=\frac{1+w}{1-w},
\]
inserting we derive the claimed bound. The extension to the
general inhomogeneous case
is obtained by very similar ideas as in the proof of Theorem~\ref{H} again.
\subsection{Proof of Theorem~\ref{khr}}
With notation from the proof of Theorem~\ref{ax3}
and Theorem~\ref{khr}, the according sets $\mathscr{Y}_i\subseteq \mathbb{R}$, $1\leq i\leq m$,
corresponding to the coordinates for $\underline{x}_1$ constructed in the proof of Theorem~\ref{cantor} are again as in~Proposition~\ref{falke} with parameters
\[
m_k\asymp H_k^{(\nu_0-1)d_i}, \qquad \epsilon_k\asymp H_k^{-\nu_0}.
\]
Proceeding as in the proof of Theorem~\ref{ax3}
leads to a lower bound of the form
\[
\dim_H(\mathcal{S}_K^{(b)\ast}(w))\geq \sum_{i=1}^{m} \dim_H(\mathscr{Y}_i)\geq \sum_{i=1}^{m} d_i \cdot
\frac{ -w\nu_0^2 + (w+1)\nu_0 -1 }{ (w+1)(1-d_i) \nu_0^2 + ((w+2)d_i -1) \nu_0 - d_i }.
\]
The first claim follows when identifying $\nu_0$ with $t$.
If $K=C_{3,\{0,2\}}\times C_{3,\{0,2\}}$, inserting $m=2, d_1=d_2=\log 2/\log 3, w=1/2$ and choosing the optimal parameter $\nu_0=1.2994\ldots$,
we calculate a bound larger than $0.1255$. Finally
as in the proof of Theorem~\ref{cantor} we can always reduce the problem to the case $0\in W_i$ by a rational transformation to infer
the same estimates for the non-$b$-ary sets $\mathcal{S}_K^{\ast}(w)$.
\subsection{Preparation for topological results}
The proof of Theorem~\ref{T2} relies on
an observation that is essentially due to Erd\H{o}s~\cite{erdos}.
\begin{lemma}[Erd\H{o}s] \label{erd}
If $A,B\subseteq \Rm$ are comeagre, then $A+B=A\cdot B=\Rm$.
\end{lemma}
A short proof for $A=B$ and $m=1$ based on Baire's category Theorem
is described in~\cite{erdos}. We present the analogous proof for
the sum in the general case.
For $\underline{t}\in\Rm$, we may write the translate
$\underline{t}-A= \cap_{j\geq 1} O_{1,j}$ of $-A$
as countable intersection of open dense sets $O_{1,j}=\underline{t}-O_{1,j}^{\prime}$.
Similarly $B=\cap_{j\geq 1} O_{2,j}$ for open dense sets $O_{2,j}$. Then
the intersection $(\underline{t}-A)\cap B=\cap_{j\geq 1} O_{1,j} \cap_{j\geq 1} O_{2,j}$ is still a countable
intersection of open dense sets, thus by Baire's theorem
dense, in particular non-empty. This means $\underline{t}\in A+B$.
Since $\underline{t}\in\Rm$ was arbitrary, the claim follows.
The proof for the product works very similarly, considering
the componentwise quotient
$\underline{t}/A=\{ \underline{t}\}/A$ in place of $\underline{t}-A$.
\subsection{Proof of Theorem~\ref{T2}}
Since the set $\mathcal{W}_m(\infty)$ is comeagre
as recalled in \S~\ref{top},
by Lemma~\ref{erd}, if $A\subseteq \Rm$ is a comeagre set as well, then we the sumset
$\mathcal{W}_m(\infty)+ A$ must be the entire space $\Rm$. On the other hand, from \eqref{eq:thor} and \eqref{eq:jarnik}
we see that
\begin{align*}
\dim_H(\mathcal{W}_m(\infty)+ A)\leq
\dim_H( \mathcal{W}_m(\infty) ) + \dim_P(A)= \dim_P(A),
\end{align*}
hence indeed we conclude $\dim_P(A)=m$.
\subsection{Proof of Theorem~\ref{C3} } \label{s5.12}
Let $m\geq 1$. For
$A,B\subseteq \Rm$, write $A\pm B$ for either $A+B$ or
$A-B$ and $A^2=\{ (a_1^2, \ldots,a_m^2): (a_1, \ldots,a_m)\in A\}$.
Now let $A=\mathcal{W}_m(\infty)$ throughout.
We have $A\circ A=\Rm$ for any $\circ\in\{ +,-,\cdot,/ \}$
by Lemma~\ref{erd} and since $A=-A$ and $A=A^{-1}$.
The latter identity needs some explanation. Assume $\ux=(\xi_1,\ldots,\xi_m)\in A$.
Then
\[
\Vert q\ux-\underline{p}\Vert \leq q^{-N}
\]
has a solution for arbitrarily large $N$. Since $|p_i|\asymp q$, $1\leq i\leq m$,
are all of the same magnitude, we check that if we let
$\underline{p}=(p_1,\ldots,p_m)$ and $P=p_1p_2\cdots p_m$ then
\[
\Vert P\xi_i^{-1}-\frac{qP}{p_i}\Vert \ll q^{-N+m-1}\ll
|P|^{-\frac{N-m+1}{m} }, \qquad 1\leq i\leq m.
\]
Since the term $(N-m+1)/m$ tends to infinity with $N$,
we infer $\ux^{-1}=(\xi_1^{-1}, \ldots,\xi_m^{-1})\in A$ as well, and the claim follows.
Moreover, from Theorem~\ref{DFSU}, the argument in the proof of Theorem~\ref{T2}
and since $\mathcal{S}_m=-\mathcal{S}_m$, we see
\[
\dim_H(A\pm \mathcal{S}_m)\leq \dim_P(\mathcal{S}_m)<m, \qquad m\geq 1.
\]
From \eqref{eq:B},
for $m\geq 5$ and $\epsilon\in(0,1/20)$ we infer $\mathcal{S}_m\pm \mathcal{S}_m\supseteq \mathcal{S}_m(1/4-\epsilon)\pm \mathcal{S}_m(1/4-\epsilon)=\Rm$.
Now let again $m\geq 1$ be arbitrary.
Write $C=C_{3,\{0,2\}}$ for Cantor's middle
third set and let
$B=((C\cup C^2)+\mathbb{Z})^m$,
the $m$-fold Cartesian product of the union of all
shifts by integers of Cantor's middle third set and its square.
It is well-known that $C+C=[0,2]$ and $C-C=[-1,1]$,
see~\cite[\S1]{art} for a short proof, and hence $B+B=B-B=\Rm$.
In~\cite[Theorem~1]{art} it is shown that $C^2\cdot C=\{ x^2y: x,y\in C\}=[0,1]$, thus $B\cdot B=\Rm$. Moreover, again as a consequence
of~\cite[Theorem~1]{art}, the set $C/C$
contains arbitrarily large intervals (in fact its precise structure
as a countable union of closed intervals is determined).
Hence we have $B/B=\Rm$.
Since $C^2$ is a locally Lipschitz image of $C$ we
have $\dim_H(C^2)=\dim_{H}(C)$ as well as $\dim_P(C^2)=\dim_P(C)$,
hence by \eqref{eq:haupack} these values all coincide,
and since $B$ is only a countable union,
from \eqref{eq:kproS} we finally get
\begin{equation} \label{eq:pp}
\dim_P(B)=\dim_H(B)=\dim(C^m)=\frac{\log 2}{\log 3}m<m.
\end{equation}
Thus, on the other hand, since our componentwise operations $\circ\in\{+,-,\cdot,/\}$ are
locally Lipschitz maps, by \eqref{eq:tricot}, \eqref{eq:jarnik} and \eqref{eq:pp}
we have
\[
\dim_H(A\circ B)\leq \dim_H(A\times B) \leq \dim_H(A) + \dim_P(B) \leq 0+\frac{\log 2}{\log 3}m<m.
\]
For $\circ=+$ see alternatively
\eqref{eq:thor}. All claims are proved.
|
\section{Introduction\label{sec:intro}}
The remote sensing of magnetic fields and other physical quantities of the outer solar atmosphere is a notoriously difficult problem. The new generation of solar telescopes such as DKIST \cite[currently in the commissioning phase;][]{2020SoPh..295..172R} and EST \citep[currently in the preparatory phase;][]{2019AdSpR..63.1389J} will provide spectropolarimetric observations of unprecedented quality. In addition, there is an urgent need for\ plasma diagnostic tools to aid in their correct interpretation. However, our modeling techniques lag behind the capabilities of such new observational facilities.
While the equations governing the intensity and polarization of spectral lines in solar prominences, filaments, and the chromosphere are mostly well known, the process of forward modeling remains difficult mainly for numerical reasons: accounting for all the relevant processes requires the solution of the problem of the generation and transfer of spectral line polarization in three-dimensional (3D) geometry. Since many spectral lines of interest are formed out of local thermodynamic equilibrium (NLTE), a complicated non-linear and non-local problem needs to be solved. This is the case for a number of spectral lines of high diagnostic potential whose optical thickness can approach or exceed unity, such as H$\alpha$ at 6563\,\AA\ or the \ion{He}{i} triplet at 10830\,\AA. Under NLTE conditions, we can expect a significant impact on the part of the radiative transfer within the medium and, as we show below, the full 3D\,NLTE radiative transfer already becomes inevitable for small optical thicknesses on the order of one.
The ultimate goal of solar spectropolarimetry is to reliably infer the plasma properties from the observed data. This so-called inverse problem is even more difficult than that of forward modeling. In principle, it is necessary to explore the space of all possible model parameters and to eliminate the models that do not agree with the observations. This is not possible in practice because such parameter space is too large and a single 3D\,NLTE model evaluation already takes up a substantial amount of computing (CPU) time. We therefore need to approach the 3D\,NLTE inverse problem (3DNIP) in a different way.
A common approach to solve the inverse problem in the photosphere, chromosphere, prominences, and filaments is to use the so-called pixel-by-pixel approach in which every column of matter behind an observed pixel is treated as independent from any other. In the case of the photosphere and chromosphere, a number of techniques have been developed over the years \citep[e.g.,][]{1992ApJ...398..375R,2000ApJ...530..977S,2015A&A...577A...7S,2019A&A...626A.102A,2019A&A...623A..74D}. For recent reviews on inversion methods, see \citet{2016LRSP...13....4D} and \citet{2017SSRv..210..109D}. A number of inversion tools have also been developed for the case of prominences and filaments \citep[e.g.,][]{2003ApJ...598L..67C,2008ApJ...683..542A,2009ASPC..415..327L}. Some algorithms go beyond the pure pixel-by-pixel approach by taking into account the spatial correlation of the data \citep[e.g.,][]{2012A&A...548A...5V,2015A&A...577A.140A,2019A&A...626A.102A} but these solutions do not take into account NLTE radiative coupling among different regions of the plasma.
In the pixel-by-pixel approach to prominences, it is challenging to go beyond models in which the physical quantities are constant along the line of sight (LOS), namely, the so-called constant-property slab approximation because there is simply no justification for considering and constructing more refined models. This approximation is indeed very rough: if we assume the prominence properties to be constant along the LOS, we would expect the properties to be also constant along a perpendicular direction, namely, in the plane of the sky -- however, the spectra are indeed changing in the plane-of-sky and in the pixel-by-pixel inversions we are interpreting due to the changing physical quantities. This approach is clearly inconsistent. An additional problem of the oversimplified models is the existence of a number of ambiguous solutions. Spectral lines that are typically optically thin, such as \ion{He}{i}\,D$_3$ at 5877\,\AA, do not suffer so much from the effects of NLTE radiative transfer \citep{2015ASSL..415..179L}, even though as subordinate lines, they can be affected by such effects due to their coupling with optically thick transitions. However, due to their negligible optical thickness, they do not provide sufficient information on the variation of the plasma parameters along the LOS. The pixel-by-pixel approach can be successfully used but only under certain physical conditions. The observed plasma structure needs to be optically thin in all directions and it must be sufficiently homogeneous along the LOS.
To the best of our knowledge, the 3DNIP has not yet been seriously studied and it has not been even clarified if it is practically possible. The existing inversion methods solve the problem by neglecting the effects of NLTE radiative transfer between different parts of the medium in the direction perpendicular to the LOS, but this leads to the neglecting of a crucial ingredient of spectral line formation. In this paper, we introduce a new framework for solving the 3DNIP problem that takes into account the 3D NLTE radiative transfer effects. We reformulate the problem in a meshfree manner and we consider it to be an unconstrained global minimization problem. The method can provide both exact and approximate results depending on the available CPU time. Apart from the fact that it can be relatively easily implemented, we show that it can lead to orders-of-magnitude acceleration with respect to grid-based methods. In addition, it is less prone to end up within the local minima.
In contrast to other methods, we do not consider the requirement of NLTE coupling to be an obstacle but rather to be a very strong natural regularization. The fact that all parts of the domain are coupled by a nontrivial set of NLTE equations leads to a much more robust method than if the individual LOS were considered separately. While our method is very efficient, our focus here is more on the physical consistency rather than on computational speed.
This paper is the first in a series. It mainly focuses on the inversion problem of solar prominence and filament data, but a similar approach can be applied, after some modifications, to the case of the chromosphere. In the papers to follow, we plan to address a number of details regarding the method and its practical applications.
The structure of the current paper is as follows. In Sects.~\ref{sec:formul} and~\ref{sec:method0}, we briefly review the key ingredients of the numerical NLTE spectral synthesis and inverse problems, respectively, and we discuss the specifics of the 3D solution. We also recall the concept of sparse approximations of physical quantities and we discuss the benefits of such representations. In Sect.~\ref{sec:method1}, we recall the standard minimization procedure of the inversion algorithms and we develop a new meshfree approach to the 3D inverse problem. In order to make this possible, we reformulated the inversion problem as a global optimization in which deviation from the NLTE consistency is used as a natural regularization together with other penalizations due to different physical inconsistencies. In the resulting unconstrained minimization algorithm, the atomic state is treated as independent of the magneto-hydrodynamical (MHD) state of the atmosphere. We complete the algorithm in Sect.~\ref{sec:stoch} by using a stochastic approach to the loss function minimization in order to reduce the risk of convergence to a local minimum and to make the solution more accurate and less demanding in terms of computer memory. We demonstrate the usability of the algorithm in Sect.~\ref{sec:example}, where we apply the method to invert the thermal and magnetic structure of an academic 3D\,NLTE problem and showing that the total inversion time is by two orders of magnitude smaller than it would be when using the grid-based techniques. We summarize our results in Sect.~\ref{sec:concl}, where we also provide some future prospects. Two appendices provide technical details on the implementation of the method and on the numerical example discussed in Sect.~\ref{sec:example}.
\section{Brief overview of the forward NLTE modeling\label{sec:formul}}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig-1d3d}
\end{center}
\caption{Spatial coupling between lines of sight in a 3D (left) and in a 1.5D (right) model atmosphere. In 3D, the radiation field at a given point results from radiation transfer from different regions of the domain, both in the vertical and in the horizontal directions. In 1.5D, the horizontal variation is neglected. The 1.5D model can be sufficiently accurate in some cases but, in general, it often leads to serious errors in calculation of the line scattering polarization.}
\label{fig:1d3d}
\end{figure}
In this work, we develop a framework for the numerical inference of the 3D distribution of physical quantities in the solar atmosphere from spectropolarimetric data. We assume that our spectropolarimetric observation consists of a data cube with ${N_{\rm pix}}=N^2$ spatial dimensions and ${N_{\lambda}}$ wavelengths for each of the four Stokes parameters $(I,Q,U,V)=(I_0,I_1,I_2,I_3)$, namely, the specific intensity ($I$), the two components of the linear polarization ($Q$ and $U$), and the circular polarization component ($V$). We also assume that the data are affected by Gaussian noise, but we do not consider any other instrumental degradation in this paper.
Although the main objective of this paper is to develop and describe a meshfree inference method, it is useful to compare its properties with a Cartesian grid-based method. To our knowledge, no such grid-based method has been developed before, apart from the 2.5D attempt by \v{S}t\v{e}p\'an et al. (2019; talk at the Solar Polarization Workshop~9\footnote{\url{https://www.mps.mpg.de/spw9}}). However, the numerical properties of such a method can be easily estimated. In this section, we consider (without any loss of generality) that the dimensions of the grid in the grid-based examples is $N^3$, that is, the mesh has the same spatial resolution as the data. However, take into account that in practice, it may be necessary to use a finer mesh depending on the particular spectral lines and the spatial extension of their formation region \citep[e.g.,][]{2003ASPC..288..405A}.
Given the spatial distribution of physical quantities defining the thermal and magnetic properties of the medium (such as temperature, density, magnetic field, etc.), the forward NLTE solution gives the atomic-level populations and quantum coherence that are consistent with the radiation field at every point, $\vec r$, of the medium. Hereafter, we use the term MHD-like quantities to describe this set of quantities, $\{\psi(\vec r;k)\}_{k=1}^{N_\psi}$, where ${N_\psi}$ is the number of such physical quantities. For example, if we consider the kinetic temperature of the plasma to be our first ($k=1$) MHD-like quantity, we have $\psi(\vec r;1)=T(\vec r)$.
In the standard unpolarized NLTE forward problem, the specific intensity of the radiation at any given point, propagation direction, and wavelength can be formally expressed as\footnote{For clarity, we omit the explicit mentioning of the term on the right hand side of the equation standing for the external illumination of the domain, as its inclusion does not contribute to the explanation.}
\begin{equation}
I(\vec r,\vec\Omega,\lambda)=\Lambda(\vec r,\vec\Omega,\lambda)[n(\vec{r}')]\,,
\label{eq:fslam}
\end{equation}
where $\Lambda$ is an operator which depends on the MHD-like quantities and $n(\vec {r}')$ stands for the atomic level populations within the medium. The simplicity of Eq.~(\ref{eq:fslam}) is only formal, as it represents a non-local coupling of the radiation field and the state of the matter. This equation is often referred to as the formal solution of the radiative transfer equation \citep[RTE,][]{2014HubenyMihalas}, or simply as the formal solution. It is worth noticing that, although this form of the forward problem in terms of the $\Lambda$ operator is useful in theoretical derivations, the operator is usually not explicitly constructed; it represents the process of solving the RTE.
The mean radiation field in the co-moving reference frame at every point in the medium, that is, the specific intensity integrated over the unit sphere and over the line absorption profile of spectral lines of interest is
\begin{equation}
\bar J(\vec r)=\oint \frac{\mathrm{d}\vec\Omega}{4\pi} \int \mathrm{d}\lambda\; I(\vec r,\vec\Omega,\lambda) \varphi(\vec r,\lambda)\,,
\label{eq:jbardef}
\end{equation}
where $\varphi(\vec r,\lambda)$ is the line's absorption profile. Likewise, we can define an averaged $\bar\Lambda$ operator that takes the above integration into account:
\begin{equation}
\bar J(\vec r)=\bar\Lambda(\vec r)[n(\vec {r}')]\,.
\label{eq:jbar}
\end{equation}
The atomic populations, the local mean radiation field, and the local MHD-like quantities are related via the equations of statistical equilibrium \citep[ESE,][]{2014HubenyMihalas}:
\begin{equation}
n(\vec r\,)={\mathcal{E}}[\bar J(\vec r\,)]\,.
\label{eq:ese}
\end{equation}
Together with Eq.~(\ref{eq:jbar}), these equations make a system of coupled NLTE equations for which we must find the self-consistent distribution of atomic populations.
The general NLTE problem can only be solved numerically with iterative methods. The simplest of such methods is the so called $\Lambda$-iteration, which can be summarized in a compact form, namely:
\begin{align}
\bar J_{k+1}(\vec r) &= \bar\Lambda(\vec r) [n_k(\vec r')]\;, \label{eq:lamit1} \\
n_{k+1}(\vec r) &= {\mathcal{E}}[\bar J_{k+1}(\vec r)]\;, \label{eq:lamit2}
\end{align}
where the radiation field in iteration $k+1$ is calculated, at every spatial point, using the atomic populations from the previous iteration $k$ (Eq.~\ref{eq:lamit1}), and the atomic populations are, in turn, updated using this new radiation field (Eq.~\ref{eq:lamit2}). Solving Eq.~\eqref{eq:lamit1} involves the solution of the RTE at all the domain points, propagation directions, and wavelengths, and it is usually the most computationally expensive part of the NLTE problem solution (see the next section for a discussion of the numerical aspects of this step). The ESE operator ${\mathcal{E}}$ in Eq.~(\ref{eq:lamit2}), which depends on the MHD-like quantities, $\{\psi(\vec r;k)\}^{N_\psi}_{k=1}$, acts locally, and the computational cost of its application is rather negligible in comparison with the cost of Eq.~\eqref{eq:lamit1} for all the relevant solar physics applications.
There are numerical techniques to accelerate the solution of the NLTE problem \citep[Jacobi, Gauss-Seidel/SOR, multigrid; see, e.g.,][and references therein]{1991A&A...245..171R,1991A&A...242..290S, 1995ApJ...455..646T, 1997A&A...324..161F} that can significantly reduce the number of necessary iterations. Recently, a very promising approach to the NLTE problem acceleration has been proposed by \citet{2021arXiv211011861J} and \citet{2021arXiv211011873B}. However, the evaluation of the $\Lambda$ operator is always a key ingredient of the forward problem solution. For brevity, we use the term $\Lambda$ iteration to describe an iteration of any such numerical technique involving the evaluation of the $\Lambda$ operator, regardless of a particular acceleration scheme.
In this work, we solve the more general problem of the generation and transfer of polarized radiation. We consider the stationary NLTE problem in complete frequency redistribution (CRD) discussed in the monograph of \citet{LL04} (the more general partial frequency distribution (PRD) problem is beyond the scope of this paper). The specific intensity is then replaced by the vector of Stokes parameters, the mean radiation field by the radiation field tensors, $\bar{J}^K_Q$, and the populations of the atomic levels by the atomic density matrix, $\rho^K_Q$. For the sake of simplicity, we only consider bound-bound transitions, but it is straightforward to generalize the method to include the bound-free and free-free transitions as well. We take into account scattering polarization, the Hanle and Zeeman effects, the macroscopic velocity gradients, and the full 3D radiative transfer, that is, all the relevant physical ingredients leading to the breaking of the symmetry of the scattering processes that critically affect the emergent polarization \citep[see][]{2015lya,2015IAUS..305..360S,2018ApJ...863..164D,2021ApJ...909..183J}.
Analogously to the MHD-like variables, we define the set of ${N_\xi}$ atomic-like quantities $\{ \xi(\vec r;k) \}_{k=1}^{N_\xi}$. The variables in this set are essentially the solution of the the NLTE forward problem. There is no unique way of defining these quantities, but there are two especially useful representations. In the first, it is necessary to specify the components of the atomic density matrix, in which case the $\{\xi(\vec r;k)\}_{k=1}^{N_\xi}$ quantities correspond to the real and imaginary parts of the density matrix elements $\rho^K_Q(\vec r; \alpha J)$ in a multi-level atom \citep[see Sect.~7.2 of][for details]{LL04}. In the second representation, we instead specify the components of the averaged radiation field tensors, $\bar J^K_Q(\vec r; \ell),$ for each of the spectral lines, $\ell$. Because the only condition that the atomic-like quantities must fulfill is to fully describe the atomic state, the two given representations are equivalent, since the radiation field tensors, $\bar J^K_Q(\vec r; \ell)$, fully determine the atomic density matrix via the ESE. In Sect.~\ref{sec:method1}, we demonstrate why it is useful to introduce the $\{ \xi(\vec r;k) \}_{k=1}^{N_\xi}$ variables and why some representations can be more advantageous than others, depending on the specific inverse problem.
\section{Grid-based approach to the NLTE inversion}\label{sec:method0}
\subsection{Scaling of grid-based numerical forward solutions}\label{ssec:mesh}
The multi-dimensional NLTE problem of the generation and transfer of polarized radiation is typically solved on a discrete grid using the short characteristics method for the RTE \citep{1988JQSRT..39...67K}. For the sake of simplicity, we only consider Cartesian meshes in this work. As mentioned in Sect.~\ref{sec:formul}, in solar physics problems, the evaluation of the $\Lambda$ operator is typically the most computationally intensive part of the NLTE forward solution, while the evaluation time of the ESE is negligible in comparison.
In 1D problems, the computing time per $\Lambda$ iteration is proportional to the number of grid points, $N$, rays in the angular quadrature, ${N_{\vec\Omega}}$, and wavelengths, ${N_{\lambda}}$. The number of accelerated $\Lambda$ iterations of the Jacobi method (one of the most commonly used techniques) needed for convergence down to the fixed error tolerance is proportional to $N$ \citep{hackbusch85}. Consequently, the total computing time for the NLTE forward solution is on the order of $O(N^2 {N_{\vec\Omega}} {N_{\lambda}})$.
In 3D geometry, the computing time per $\Lambda$ iteration is also proportional to the number of grid points, $N^3$. However, regarding the number of iterations, the convergence rate is determined by the distance in the units of grid steps between the most distant points. This distance is on the order of $\sqrt[3]{N^3}=N$, hence the number of accelerated $\Lambda$ iterations in 3D is usually similar to that of the 1D case. Therefore, the solution time of the 3D\,NLTE problem is typically on the order of $O(N^4{N_{\vec\Omega}}{N_{\lambda}})$.
In order to provide a computing time example, we note that in the 3D\,NLTE code PORTA \citep{2013PORTA} the CPU time needed for evaluating the $\Lambda$ operator per grid point, propagation direction, wavelength, and Stokes parameter can be, typically, on the order of $10^{-6}$\,seconds in a common CPU. Considering a small 3D model with $N=100$ (total $100^3$ grid points), angular quadrature with 100 propagation directions, and the four Stokes parameters of a single spectral line sampled in 50 wavelengths, a single $\Lambda$ iteration would take about $5.5$~hours. With the typically expected 100 Jacobi iterations, the total computation time of the NLTE forward solution is about 550~hours. The computing time in more realistic problems of the solar chromosphere, with typical meshes of about $500^3$ grid points, becomes correspondingly larger (by two orders of magnitude) and the application of high-performance computing (HPC) becomes a requirement \citep[see][]{2015lya,2016caii}.
In order to bypass these high requirements in computational resources, it is a common practice to split the 3D problem of the $N^3$ grid into $N^2$ independent 1D problems with $N$ grid points (the so-called 1.5D approximation). When scattering polarization, the Hanle effect, and the macroscopic velocity fields do not play a significant role in the polarization of the emergent spectral line radiation, such a 1.5D solution is faster because the ensuing axial symmetry of each 1D model of every pixel or column allows us to neglect the azimuthal dependence of the radiation field. Such approach would still be approximate, as the radiative interaction between such models is neglected, but it can still provide sufficiently accurate approximations in some applications. However, when the cylindrical symmetry is broken, both 1.5D and 3D solutions have similar computing demands and, moreover, the 1.5D approach can lead to significant errors, especially when accounting for scattering polarization (see Fig.~\ref{fig:1d3d}).
\subsection{Parameterization and discrete representation of variables}\label{sssec:discrete}
In the inverse problem, we rarely try to directly find the $\{\psi(\vec r;k)\}_{k=1}^{N_\psi}$ values in all the grid points because, even in 1D, the number of unknowns would be too large and, moreover, such an approach leads to underdetermined problems. Instead, we can parameterize the spatial distribution of variables using a smaller number, ${M_\psi}$, of parameters $\{\psi_i(k)\}_{i=1}^{M_\psi}$ from which the value of each individual variable can be reconstructed at any point within the medium:
\begin{equation}
\psi(\vec r;k) = f(\vec r; \psi_1(k), \cdots, \psi_{M_\psi}(k)) \,.
\label{eq:psi-f}
\end{equation}
The form of the function $f(\cdot)$ depends on the particular chosen representation of the variables. Examples of this approach are the node-based interpolation in 1D inversions \citep{1992ApJ...398..375R} and the wavelet expansion in 2D geometry proposed by \citet{2015A&A...577A.140A}.
Henceforth, we assume that each of the MHD-like quantities is represented by a set of ${M_\psi}$ parameters. For the simplicity of notation, here, we chose to parameterize every MHD-like quantity using the same number ${M_\psi}$ of coefficients. Therefore, the whole model, $\psi$, can be defined as an ${N_\psi} {M_\psi}$-elements vector of these sets of parameters,
\begin{align}
\psi &= (\psi_1, \cdots, \psi_{{N_\psi} {M_\psi}})\nonumber\\
& = ( \psi_1(1), \cdots, \psi_{M_\psi}(1), \psi_1(2), \cdots, \psi_{M_\psi}({N_\psi}) )\,.
\label{eq:psidef}
\end{align}
However, in practice, it can be advantageous to use different numbers of parameters for different quantities. We postpone the discussion of the particular parameterization of Eq.~\eqref{eq:psi-f} to Sect.~\ref{ssec:expans}. For convenience, we assume that the elements of $\psi$ are dimensionless parameters.
It is important to emphasize that under certain conditions, we can aim to fulfill the condition ${M_\psi}\ll N^3$, which leads to a significant reduction of the number of unknowns and makes this underdetermined problem a more well-defined one. Moreover, it can lead to a very significant reduction of the total CPU time needed for the inversion. We discuss and apply this condition in the following sections.
\subsection{Inverse problem definition and solution}\label{ssec:invstd}
\begin{algorithm}
\caption{
Inversion as the minimization of $\chi^2$ via successive solutions of self-consistent NLTE forward problems. Here, we show the gradient descent method with step $h$.
$\psi(i)$ stands for the $i$-th iteration of the $\psi$ vector in Eq.~\eqref{eq:psidef} and $\psi_j$ stands for its $j$-th element.
The symbol $\bar J$ is a condensed notation for the averaged radiation field tensors calculated in the grid points and $\bar\Lambda_\psi$ corresponds to the $\bar\Lambda$ operator, constructed using the $\psi$ vector.
\label{alg:inv1}}
\begin{algorithmic}[1]
\STATE \textbf{Initialization}: Randomly initialize the model's MHD-like quantities $\psi(0)$.
\STATE Given $\psi(0)$, solve the NLTE forward problem for $\bar J(0)$ via $\Lambda$ iteration.
\STATE $i\leftarrow 0$
\REPEAT [\textbf{Descent along the negative $\chi^2$ gradient.}]
\STATE $i\leftarrow i+1$
\FOR[\textbf{Loop over the model MHD-like variables.}]{$j=1$ to ${N_\psi} {M_\psi}$}
\STATE Calculate j-th element gradient
$\nabla_j \chi^2=\partial \chi^2(D;\psi(i-1))/\partial\psi_j$.
\{{\bf $\Lambda$ iteration till NLTE consistency.}\}
\ENDFOR
\STATE $\psi(i)\leftarrow \psi(i-1)-h \nabla\chi^2$. \{{\bf New estimate of the model parameters.}\}
\STATE Given $\psi(i)$, solve the NLTE forward problem for $\bar J(i)$ via $\Lambda$ iteration.
\UNTIL{$\chi^2(D;\psi(i))\approx 1$.}
\end{algorithmic}
\end{algorithm}
Assuming that our data, $D$, consist of a square matrix of ${N_{\rm pix}}=N^2$ pixels and ${N_{\lambda}}$ wavelengths for each of the four Stokes parameters and that they are contaminated with Gaussian noise. The inverted 3D model is parameterized by the $\psi$ vector defined in Eq.~\eqref{eq:psidef}. The measure of the goodness of the fit of the model to the data is the familiar $\chi^2$ function:
\begin{equation}
\chi^2(D;\psi) = \frac{1}{{N_{\rm pix}}{N_{\lambda}} } \sum_{i=1}^{N_{\rm pix}} \sum_{j=1}^{N_{\lambda}}\sum_{k=0}^3 w_k \frac{(I_{ijk}^{\rm M}(\psi)-I_{ijk}^{\rm O})^2}{\sigma_{ijk}^2}\,,
\label{eq:chi2}
\end{equation}
where $I_{ijk}^{\rm M}(\psi)$ is the $k$-th Stokes parameter at pixel, $i,$ and wavelength index, $j,$ calculated from the model's vector, $\psi$. Similarly, $I_{ijk}^{\rm O}$ is the corresponding observed signal, for the same $k$-th Stokes parameter at the same $i$-th pixel and $j$-th wavelength, with the noise variance $\sigma_{ijk}^2$. The $w_k$ quantities are weights for the different Stokes parameters which must fulfill the normalization condition $\sum_{k=0}^3 w_k=1$ (see Appendix \ref{app:normal}). The goal of the inversion process is to find an estimate $\hat\psi$ of the model parameters leading to the best fit to the observed data,
\begin{equation}
\hat\psi = \underset{\psi}{\argmin}\; \chi^2(D;\psi)\,.
\label{eq:psimin}
\end{equation}
Given the normalization of the signals to their noise variance, it follows that the optimal fit fulfills $\chi^2=1$. The restriction of the number of model parameters to ${M_\psi}$ per quantity is an example of regularization of the solution, namely, additional constraints that cannot be solely derived from the observed data. Examples of such a regularization in the general inversion theory include the requirement on the number of non-zero parameters (i.e., sparsity regularization) or on the smoothness of the solution. Instead of Eq.~\eqref{eq:psimin}, we then solve a problem of the following form:
\begin{equation}
\hat\psi = \underset{\psi}{\argmin}\; \left[ \chi^2(D;\psi) + g(\psi) \right]\,,
\label{eq:regular}
\end{equation}
with $g(\psi)$ being the regularization condition. Regularization can help make an ill-defined problem to become well defined and it can be understood as an Occam's razor condition.
It is important to emphasize that in contrast to pixel-by-pixel inversion, in the minimization of the $\chi^2$ in Eq.~\eqref{eq:chi2}, all the pixels are inverted together because, in general, every component of the $\psi$ vector affects the model parameters in the whole computational domain (depending on the particular form of the right-hand side of Eq.~\ref{eq:psi-f}). This leads to a more robust family of methods where correlations among the MHD-like quantities in different spatial points can be naturally taken into account, such as in \citet{2015A&A...577A.140A}. Moreover, it also allows us to take full 3D radiative transfer into account, as well as other physical constraints, such as the condition of zero divergence of the magnetic field, which are practically impossible to implement in the pixel-by-pixel approaches (see below).
Algorithm~\ref{alg:inv1} shows a pseudocode for the traditional inversion process based on the minimization of the $\chi^2$ function via the gradient descent method. Even though more efficient algorithms than that of the gradient descent can be used to reduce the number of required iterations, the general structure remains the same, involving the calculation of the $\chi^2$ function gradient with respect to each model parameter (loop in lines 6--8). To calculate each gradient component $\nabla\chi^2_j$ (line 7) using a rather simple first-order method, we need to modify the $\psi_j$ variable by a small amount $\delta$, resulting in the model vector $\psi'=(\psi_1, \psi_2, \cdots, \psi_{j-1}, \psi_j+\delta, \psi_{j+1},\cdots,\psi_{{N_\psi} {M_\psi}})$, then calculate a new self-consistent solution for $\psi'$, and calculate the new $\chi^2$ value. Thus, we obtain:
\begin{equation}
\frac{\partial \chi^2(\psi)}{\partial \psi_j}\approx\frac { \chi^2(\psi') - \chi^2(\psi)} {\delta}\,.
\label{eq:grad}
\end{equation}
Evaluating Eq.~\eqref{eq:grad} requires a NLTE forward solution, $O({N_{\vec\Omega}}{N_{\lambda}} N^3)$, and evaluating $\chi^2(\psi')$, $O({N_{\lambda}} N^3)$. Therefore, for ${N_\psi}{M_\psi}$ components the calculation of the gradient (loop 6--8) scales as $O({N_\psi}{M_\psi}({N_{\vec\Omega}}{N_{\lambda}} N^3+{N_{\lambda}} N^3))=O({N_\psi}{M_\psi}{N_{\vec\Omega}}{N_{\lambda}} N^3)$, which is completely dominated by the NLTE forward solution. Assuming that the number of inversion iterations is ${N_{\rm it}}$ (loop 4--11) and assuming that evaluating Eq.~(\ref{eq:grad}) requires a single $\Lambda$ iteration, the total number of $\Lambda$ iterations required for the inverse solution according to Algorithm~\ref{alg:inv1} is equal to ${N_{\rm it}} {N_\psi} {M_\psi}$, implying a total scaling of $O({N_{\rm it}}{N_\psi}{M_\psi}{N_{\vec\Omega}}{N_{\lambda}} N^3)$.
We can now easily estimate the wall-clock time of a 3D inversion. Assuming, conservatively, that the number of inversion iterations is ${N_{\rm it}}=100$, considering that the number of MHD-like quantities would typically be on the order of ${N_\psi}=10$ and that the number of coefficients per quantity, ${M_\psi}$, is on the order of $10^2$ (for very simple models) to $10^4$ (more refined models; see below for further details), as well as that the CPU time for each $\Lambda$ iteration is about $5.5$~hours for a model with $N^3=100^3$ (Sect.~\ref{ssec:mesh}), the total CPU time is ${N_{\rm it}} {N_\psi} {M_\psi} \cdot 5.5$\,hours; namely, between 0.5 and 55 millions of CPU hours. Although enormous, such computing resources are available in today's HPC facilities. However, our experience shows that the approach sketched in Algorithm~\ref{alg:inv1} often tends to end up in a local minimum of the $\chi^2$ function and the calculation needs to be restarted. This already happens in the 2.5D problems we tested and it can be expected that in full 3D geometry, this problem is only going to get worse. In addition, due to the possible non-uniqueness of the solution, it is nevertheless worthwhile to repeat the calculations in order to discover alternative compatible solutions. Taking into account the fact that spectropolarimetric observations with $100^2$ pixels are very coarse by today standards, the CPU time necessary for grid-based 3D inversions quickly becomes unacceptable in practice.
\section{Meshfree approach to the 3D NLTE inversion}\label{sec:method1}
In the multi-dimensional NLTE forward problem in solar-physics, the input model atmosphere is often the result of an MHD simulation, and the large-scale simulations of the outer solar atmosphere that are used in the NLTE synthesis are often computed using rectilinear Cartesian grids \citep[e.g.,][]{2011A&A...531A.154G}. Such regular discretization is advantageous in the NLTE forward problem, but the existence of a grid is a complication in the sense that the radiation transfer needs to be performed in the particular topological order of the grid nodes and, consequently, the parallelization of the NLTE solution becomes a non-trivial task.
In the inverse problem, we are not constrained by any a priori model to work with and, therefore, we are not obliged to adopt any particular spatial discretization. In this section, we introduce the basic ingredients and algorithms for a meshfree method for solving the 3DNIP. As shown below in this section, abandoning the space discretization entails a number of advantages, even though it is necessary to reformulate the standard inversion Algorithm~\ref{alg:inv1}.
One of the main properties of the new method is that it allows us to avoid
the costly and repetitive application of the $\bar\Lambda$ operator (Sect.~\ref{sec:formul}), at
the expense of not automatically guaranteeing the full NLTE consistency
of the problem at every step of the inversion.
\subsection{Expansion of variables into a basis}\label{ssec:expans}
In Sect.~\ref{sssec:discrete}, we show how we parameterized the spatial variation of each of the ${N_\psi}$ MHD-like quantities, $\{\psi(\vec r;k)\}_{k=1}^{N_\psi}$, with ${M_\psi}$ coefficients, $\{\psi_i(k)\}_{i=1}^{M_\psi}$, which, together with the function $f(\cdot)$ in Eq.~\eqref{eq:psi-f}, allow us to compute the physical quantities at any point, $\vec r$, within the medium. Henceforth, we particularize the general function $f(\cdot)$ to the linear function
\begin{equation}
\psi(\vec r;k)=\sum_{i=1}^{M_\psi} \psi_i(k) \phi_i(\vec r)\,,
\label{eq:psiexp}
\end{equation}
where $\{\phi_i(\vec r)\}_{i=1}^{M_\psi}$ is a certain basis of orthonormal functions in the 3D computational domain, namely,
\begin{equation}
(\phi_i,\phi_j) = \int \mathrm{d}^3 r \, \phi_i(\vec r)\phi_j(\vec r)=\delta_{ij} \,,
\end{equation}
where $(\cdot,\cdot)$ stands for the inner product of the functions and the integration is done over the computational domain.
The choice of the basis functions is generally arbitrary, but different bases can have different degrees of suitability for approximating the distribution of quantities in different models and coordinate systems. A good approximation should fulfill ${M_\psi}\ll N^3$, so that the number of free parameters per quantity is much smaller than the number of mesh points in an equivalent 3D grid, and should allow the approximation of the spatial variation of the quantities on both large and small scales. Such an approximation is often possible because the physical quantities are at least piece-wise continuous. However, due to the nature of the NLTE inverse problem, it is not possible to know the most suitable basis a priori and, at the same time, it is usually not possible either to find an optimal sparse subset of the basis functions as in \citet{2015A&A...577A.140A} due to the high CPU demands of such a process. Nevertheless, it is possible to improve the choice of the basis by considering heuristic methods and by using the fact that different quantities are often spatially (anti)correlated.
In this paper, we restrict ourselves to a fixed basis consisting of a set of typically smooth functions up to a certain order (such as in Appendix~\ref{ssec:bases}, where we provide an example in the 3D Cartesian coordinates). In this sense, it is important to understand that the term ``smooth model'' that we occasionally use is often adequate but generally overly restrictive; our constraint is actually on the number of basis functions, ${M_\psi}$, rather than on their smoothness in any particular mathematical sense.
\subsection{Sampling the radiation field in pilot points}\label{ssec:sampling}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig-lc}
\end{center}
\caption{Example of three pilot points and their associated long characteristics in a computational domain.}
\label{fig:lc}
\end{figure}
What makes the 3DNIP problem more difficult than other inverse problems is that, even if we find the $\psi$ correctly reproducing the observed data $D$, we still need to guarantee that the radiation and the atomic state are consistent at every point within the medium (see Sect.~\ref{sec:formul}). Similarly to the expansion of the MHD-like variables in Sect.~\ref{ssec:expans}, we can do the same with the atomic-like quantities:
\begin{equation}
\xi(\vec r;k)=\sum_{i=1}^{M_\xi} \xi_i(k) \phi_i(\vec r)\,,
\label{eq:xiexp}
\end{equation}
where ${M_\xi}$ is the number of basis functions. Analogously to Eq.~\eqref{eq:psidef}, we can represent the information on these variables with a ${N_\xi}{M_\xi}$ elements vector,
\begin{align}
\xi &= (\xi_1, \cdots, \xi_{{N_\xi} {M_\xi}}) \nonumber\\
&=( \xi_1(1), \cdots, \xi_{M_\xi}(1), \xi_1(2), \cdots, \xi_{M_\xi}({N_\xi}) ) \,.
\label{eq:xidef}
\end{align}
It is important to emphasize that since we assume that in the family of models of interest, it is also ${N_\xi}\ll N^3$, we can fully determine the $\xi$ vector from a relatively small number of samples in the 3D domain. To show this, let $y(\vec r)$ be an arbitrary real function that can be expanded in the $\{\phi_i(\vec r)\}_{i=1}^M$ basis,
\begin{equation}
y(\vec r) = \sum_{i=1}^M q_i \phi_i(\vec r),
\label{eq:ydef}
\end{equation}
where the $q_i = (y,\phi_i)$ are the expansion coefficients, and let $\{\vec {r}_j\}_{j=1}^M$ be a set of points randomly distributed in the model domain. At each of these points Eq.~\eqref{eq:ydef} becomes $y(\vec{r}_j) = \sum_{i=1}^M q_i\phi_i(\vec{r}_j)$ or, equivalently, $\vec{y} = \vec\Phi\vec{q}$, where $\vec{y}$ and $\vec{q}$ are column vectors and $\vec{\Phi}$ is a $M$ by $M$ matrix with elements $\Phi_{ji}=\phi_i(\vec r_j)$. In our approach, we restrict ourselves to the sets of random points for which the matrix $\vec\Phi$ can be inverted which are, from a practical point of view, the vast majority of cases. Due to isomorphism of $\vec q$ and $\vec y$, any random sample $\vec y$ uniquely determines the vector $\vec{q} = \vec{\Phi}^{-1}\vec{y}$.
Consequently, if we apply the same reasoning to the atomic-like variable $\xi(\vec r;k)$ sampled in ${N_{\Lambda}}={M_\xi}$ random points\footnote{We introduce here the new variable ${N_{\Lambda}}$ to represent the number of sampling points and we make it equal to the number of basis functions, although we go on to show in Sect.~\ref{sec:stoch} that it can be advantageous to give different values to these variables.}, these samples are sufficient to determine the $\xi(\vec r;k)$ vector in the whole 3D domain. Because ${N_{\Lambda}}\ll N^3$, such a condensed description could lead to a significant optimization of the numerical solution in comparison to the full $\Lambda$ iteration. From now on, we refer to the random points used for sampling the atomic-like quantities as pilot points.\footnote{We note that the same term is used in the geophysical inverse problem research in a slightly different context \citep{doherty10}.}
Regardless of the particular representation of the atomic-like quantities, their sampling in the pilot points needs to be done via the calculation of the mean radiation field tensors $\bar J^K_Q(\vec r_j;\ell),$ as described in Sect.~\ref{sec:formul}. From these, the density matrix elements are easily obtained by solving the ESE.
Given an estimate of the $\psi$ and $\xi$ vectors, we can get the mean radiation field at the $\{\vec r_j\}_{j=1}^{N_{\Lambda}}$ points by solving the RTE using the long characteristics (LC) method \citep{2014HubenyMihalas,2021ApJ...912...63D} with a suitable angular quadrature with ${N_{\vec\Omega}}$ propagation directions (see Fig.~\ref{fig:lc}). Conservatively assuming, for simplicity, that the number of discrete steps along the LC is $N$, then the time required for the calculation of the mean radiation field tensors is $O({N_{\Lambda}}{N_{\vec\Omega}}{N_{\lambda}} N)$, while the full $\Lambda$ iteration is of $O({N_{\vec\Omega}}{N_{\lambda}} N^3)$. To calculate the $\chi^2$ function, we need to solve the RTE in $N^2$ pixels in order to get the emerging Stokes parameters in the whole field of view. This can also be done with LC parallel to the LOS, once per pixel, and this formal solution is then on the order of $O(N^2{N_{\lambda}})$.
The first step towards the meshfree method discussed in this subsection is still conceptually similar to the traditional approach based on the $\Lambda$-iteration approaches: instead of determining the radiation field and atomic state (represented formally by $\vec y$ in Eq.~\ref{eq:ydef}) in all the points of a fine-spaced rectilinear grid, we may consider it solely in a smaller set of pilot points. Then the inverse of the $\vec\Phi$ matrix provides us with the model parameters $\vec q$. These can be used to reconstruct the atomic state at all points in the domain. This information can subsequently be used in the radiative transfer calculations along the long characteristics going through the whole domain. In the following section, we show that this naive generalization does not bring many benefits over the traditional $\Lambda$-iteration approach and a more radical approach is needed.
\subsection{3DNIP as an unconstrained global minimization problem}\label{ssec:unconstr}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig-cons}
\end{center}
\caption{Schematic visualization of the inversion process in the space spanned by the MHD-like and atomic-like variables. The background color indicates deviation from the NLTE consistency, $\mathcal{L}_\Lambda$, for every combination of the MHD- and atomic-like variables, with the brightest color representing the minimum value. The red line shows the common approach to the NLTE inversion as a sequence of physically consistent NLTE models of decreasing $\chi^2$ value (constrained minimization). The blue line shows an unconstrained minimization with allowed inconsistent solutions during the inversion process.}
\label{fig:unconstr}
\end{figure}
The solution of the 3DNIP in the meshfree representation of quantities could be done with minor modifications to Algorithm~\ref{alg:inv1}. The difference would be in the application of the $\bar\Lambda$ operator at lines 2, 7, and 10 of the algorithm. This operator would be evaluated just in ${N_{\Lambda}} = {M_\xi}$ pilot points instead of in the whole 3D grid. However, a closer inspection of this approach reveals that it is seriously deficient.
At line 7 in Algorithm~\ref{alg:inv1}, we modify $\psi_j$ by a small amount $\delta$, we solve the ESE in the pilot points, calculate the new radiative transfer coefficients along the LC, and then we solve the RTE in order to get the new values of the $\bar J^K_Q(\vec r_j)$ tensors. If $\delta$ is small, only a few such iterations are necessary. Finding $\bar J^K_Q(\vec r_j)$ in the pilot points is on the order of $O({N_{\Lambda}} {N_{\vec\Omega}} {N_{\lambda}} N)=O({M_\xi} {N_{\vec\Omega}} {N_{\lambda}} N)$. The problem is that once the $\bar J^K_Q(\vec r_j)$ in the pilot points have been found (or, equivalently, the new values of $\xi(\vec r_j,k)$, as in Sect.~\ref{ssec:sampling}), it is necessary to find the expansion coefficients $\xi_i(k)$ of Eq.~(\ref{eq:xiexp}) that are necessary for evaluating the modified $\chi^2$ function. This operation involves the $\vec\Phi^{-1}$ matrix multiplication (see the previous section). With ${N_\xi}$ atomic-like quantities, this operation is on the order of $O({N_\xi}{M_\xi}^2)$. Moreover, evaluating the radiative transfer coefficients for the $\chi^2$ calculation is on the order of $O({N_\xi} {M_\xi} N^3)$ because the RTE coefficients are linear combinations of ${M_\xi}$ basis functions of ${N_\xi}$ quantities evaluated in $N^3$ points. Finally, the calculation of the $\chi^2$ is then on the order of $O(N^3{N_{\lambda}})$.
Consequently, the loop 6--8 would be $O( {N_\psi}{M_\psi}({M_\xi} {N_{\vec\Omega}} {N_{\lambda}} N + {N_\xi}{M_\xi}^2 + {N_\xi}{M_\xi} N^3 + N^3{N_{\lambda}}))=O( {N_\psi}{M_\psi}({M_\xi} {N_{\vec\Omega}} {N_{\lambda}} N + {N_\xi}{M_\xi} N^3 + N^3{N_{\lambda}}))$, taking into account that ${M_\xi}<N^3$. In any problem of reasonable complexity, the dominant term is ${N_\xi}{M_\xi} N^3$, hence the computing demands of the method can be written as $O( {N_\psi}{M_\psi} {N_\xi}{M_\xi} N^3 )$.
In Sect.~\ref{ssec:invstd} we estimated the CPU time demands per iteration of the grid-based method to be $O({N_\psi}{M_\psi}{N_{\vec\Omega}}{N_{\lambda}} N^3)$. Since we can expect ${N_\xi}{M_\xi}>{N_{\vec\Omega}}{N_{\lambda}}$
in realistic models, the meshfree approach seems to bring no advantage with respect to the grid-based methods. In order to make the mesh-free method really efficient, we need to avoid the approach based on the inversion of the $\vec\Phi$ matrix and to develop a different technique.
Algorithm~\ref{alg:inv1} is an example of a constrained minimization of $\chi^2(D;\psi,\xi)$ in which we impose on $\xi$ the following constraint: the problem is always NLTE consistent. At every step of the solution, the $\xi$ variables are fully determined by $\psi$: $\xi=\xi(\psi)$, hence, we can write $\chi^2(D;\psi,\xi)=\chi^2(D;\psi)$ with the implicit condition on the NLTE consistency. The inversion process can be understood as a sequence of solutions in the space spanned by $\psi$ and $\xi$ such that, at every step, the model is NLTE consistent. In Fig.~\ref{fig:unconstr}, we represent such a process with the red curve, whose thickness is proportional to the value of $\chi^2$. In the context of Algorithm~\ref{alg:inv1}, the horizontal axis stands for the independent variables while the vertical axis stands for the dependent ones. Algorithm~\ref{alg:inv1} follows the consistency curve until a model with $\chi^2\approx 1$ is found. In order to benefit from the meshfree formulation, we need a more significant change in the inversion algorithm.
\subsubsection{NLTE consistency regularization}
In order take advantage of the meshfree formulation of the inverse problem, we need to substantially revise the traditional inversion algorithm. Let us define a vector of all model variables, $\theta$, as a union of the $\psi$ and $\xi$ vectors,
\begin{equation}
\theta=
\{\theta_1, \cdots, \theta_{{N_{\rm V}}} \}
=\{\psi_1, \cdots, \psi_{{N_\psi}{M_\psi}}, \xi_1, \cdots, \xi_{{N_\xi}{M_\xi}} \}\,,
\label{eq:thetadef}
\end{equation}
where
\begin{equation}
{N_{\rm V}}={N_\psi} {M_\psi} + {N_\xi}{M_\xi}
\label{eq:nvar}
\end{equation}
is the total number of variables in the model, and let us formulate the inverse problem as an unconstrained minimization in which we let all the $\theta$ variables acquire any value independently of each other. Instead of the inverse problem of Eq.~\eqref{eq:psimin}, which is based on the minimization of the $\chi^2$ function, we can base the inverse problem on the minimization of a more general loss function,
\begin{equation}
\mathcal{L}(D;\theta)=\chi^2(D;\theta) + \lambda \mathcal{L}_\Lambda(\theta)\,,
\label{eq:ll1}
\end{equation}
where $\chi^2(D;\theta)$ is the $\chi^2$ function of Eq~(\ref{eq:chi2}) but now calculated using the whole set of generally physically inconsistent parameters $\theta$ and then $\lambda$ is some positive constant, while $\mathcal{L}_\Lambda(\theta)$ is a non-negative differentiable function that stands for a penalty for deviation of $\theta$ from the NLTE consistency in the pilot points. In this formulation, we allow the model parameters to take unphysical values during the inversion, but deviations from self-consistency are penalized. The goal of the minimization is to find the lowest value of $\mathcal{L}$, which is equivalent to finding the minimum value with $\chi^2=1$ and $\mathcal{L}_\Lambda=0$, corresponding to a NLTE-consistent solution that simultaneously fits the observed data. In Fig.~\ref{fig:unconstr}, such an inversion process is schematically shown with the blue curve. Starting from a physically inconsistent model that does not fit the data, we aim to find the solution that is simultaneously consistent and fits the observations. In comparing Eqs.~\eqref{eq:ll1} and \eqref{eq:regular}, we can interpret the $\lambda\mathcal{L}_\Lambda$ term as a regularization of the $\chi^2$ minimization problem that is due to the requirement of NLTE consistency.\footnote{This analogy needs to be taken with care because regularizations are usually conditions imposed regardless of the physical model, but we can understand this particular regularization in the context of models neglecting the NLTE transfer effects --- in this regard, the $\lambda\mathcal{L}_\Lambda$ term can be seen as a natural physical regularization imposing NLTE consistency.}
As mentioned above, the loss function $\mathcal{L}_\Lambda(\theta)$ should be differentiable and its minimum value should be equal to zero for solutions that are NLTE-consistent at every pilot point, $\vec r_j$, namely:\
\begin{equation}
\xi(\vec r_j;k)=\tilde\xi(\vec r_j,\theta; k)\,,
\end{equation}
where $\xi(\vec r_j,k)$ is the atomic-like $k$-th variable calculated at the pilot point, $\vec r_j$, from the current guess of the $\xi$ vector and $\tilde\xi(\vec r_j,\theta; k)$ is the same atomic-like variable, at the same pilot point, resulting from the radiative transfer via LC and for the given model parameterization, $\theta$. A suitable choice for $\mathcal{L}_\Lambda$ is the norm
\begin{equation}
\mathcal{L}_\Lambda(\theta) = \frac{1}{{N_{\Lambda}}{N_\xi}} \sum_{j=1}^{N_{\Lambda}} \sum_{k=1}^{{N_\xi}}
\left| \xi(\vec r_j;k) - \tilde\xi(\vec r_j,\theta;k) \right|^2\,.
\label{eq:llamdef}
\end{equation}
The $\lambda$ parameter in Eq.~\eqref{eq:ll1} needs to be estimated based on the particular problem and the required accuracy. If we define an acceptable error, $\Delta\xi^2$, $\lambda$ should fulfill $\lambda\Delta\xi^2 > 1$ in order to have the second term in Eq.~\eqref{eq:ll1} of at most the same order of magnitude as $\chi^2$ in the final solution. Therefore, a rough estimate of $\lambda$ can be made so that $\lambda > 1/\Delta\xi^2$. In practice, this choice needs to be made with care and with respect to each particular model and its parameterization. Moreover, the $\lambda$ parameter does not have to be constant: the decrease of its value as the solution approaches convergence can reduce the oscillatory behavior of the loss function and lead to better convergence, although we do not further discuss this aspect in this work.
In addition to the NLTE consistency, the form of the loss function in Eq.~\eqref{eq:ll1} allows us to include additional penalties, such as deviations from consistency with local physical-laws (in the sense that they are independently fulfilled at every point, in contrast with the radiatively coupled NLTE consistency). We include these penalties in an additional term $\gamma\mathcal{L}_{\rm L}$ on the right-hand side of Eq.~\eqref{eq:ll1}, namely:
\begin{equation}
\mathcal{L}(D;\theta)=\chi^2(D;\theta) + \lambda \mathcal{L}_\Lambda(\theta) + \gamma \mathcal{L}_{\rm L}(\theta)\,,
\label{eq:ll2}
\end{equation}
where $\gamma$ is another positive parameter of the problem (and similar considerations to those about the $\lambda$ parameter above can be made about its value). Examples of what we mean by deviation from local physical consistency are the existence of negative atomic populations, the non-zero divergence of the magnetic field vector $\nabla\cdot\vec B$, or the deviation from magnetohydrostatic equilibrium of the plasma. In the stationary models that we consider in this paper, it is also straightforward to include a penalty for deviations from the equation of continuity $\nabla\cdot(\rho\vec v)=0$ or other MHD equations.
The term $\mathcal{L}_{\rm L}$ in the loss function can thus be understood as an additional natural regularization due to physical consistency. It should be evaluated in a random set of ${N_{\rm L}}$ points that can typically be set as ${N_{\rm L}}\gg{N_{\Lambda}}$, while still requiring a negligible amount of CPU time in comparison to $\mathcal{L}_\Lambda$ because there is no need to perform radiative transfer calculations in order to evaluate it.
Likewise, $\mathcal{L}_{\rm L}\to 0$ as we are reaching a physically consistent solution. In Appendix~\ref{sec-app:lloc} we briefly discus the construction of the $\mathcal{L}_{\rm L}$ penalty. In general, we want this function to be a convex differentiable function of $\theta$ with the minimum $\mathcal{L}_{\rm L}=0$ for the locally consistent solutions.
In summary, with the loss function given by Eq.~\eqref{eq:ll2}, the 3DNIP problem can be formulated as
\begin{equation}
\hat \theta =\underset{\theta}{\rm arg\,min} \;\mathcal{L}(D;\theta)\,,
\end{equation}
that is, to find the estimate $\hat\theta$ for the given data $D$ such that the $\mathcal{L}(D;\theta)$ loss function has its minimum value equal to $1$. As we will see below, the condition $\mathcal{L}_\Lambda=\mathcal{L}_{\rm L}=0$ will be rarely satisfied exactly in practice, hence, assuming a suitable choice of $\lambda$ and $\gamma$ a typical solution fulfills $\chi^2\approx 1$, $\mathcal{L}_\Lambda\ll 1$, and $\mathcal{L}_{\rm L}\ll 1$.
\subsubsection{New inversion algorithm\label{sssec:newalg}}
\begin{algorithm}
\caption{3DNIP solution as an unconstrained $\mathcal{L}$ minimization.
\label{alg:inv2}}
\begin{algorithmic}[1]
\STATE \textbf{Initialization}: Randomly initialize the model parameters $\theta(0)$.
\STATE $i\leftarrow 0$
\REPEAT [\textbf{Descent along the negative $\mathcal{L}$ gradient.}]
\STATE $i\leftarrow i+1$
\FOR[\textbf{Loop over the model parameters.}]{$j=1$ to ${N_{\rm V}}$}
\STATE Calculate the gradient element
$\nabla_j \mathcal{L}=\partial \mathcal{L}(D;\theta(i-1))/\partial\theta_j$.
\ENDFOR
\STATE $\theta(i)\leftarrow \theta(i-1)-h \nabla\mathcal{L} $. \{{\bf New estimate of the model parameters.}\}
\UNTIL{$\mathcal{L}\approx 1$ (or, explicitly, $\chi^2(D;\theta(i))\approx 1$ and $\mathcal{L}_\Lambda\ll 1$ and $\mathcal{L}_{\rm L}\ll 1$).}
\end{algorithmic}
\end{algorithm}
We are now ready to devise a new inversion algorithm for the unconstrained NLTE inversion, namely, Algorithm~\ref{alg:inv2}. Even though it seems similar in structure to Algorithm~\ref{alg:inv1}, there are some very important differences. First, we randomly initialize all the variables $\theta$, including the atomic-like variables $\xi$. Secondly, the main loop (lines 3--8) runs until a good fit to the data is found that is simultaneously physically consistent. Finally, the gradient of the loss function in the inner loop (lines 5--7) is calculated with respect to all $\theta$ variables, that is, to both MHD-like, $\psi$, and atomic-like, $\xi$, variables.
In contrast to the step in line 7 in Algorithm~\ref{alg:inv1}, there is no need to evaluate the exact $\xi$ variables from the radiation field tensors in the pilot points because the $\xi$ variables are explicitly known at every step. A small perturbation of any of the $\psi_j$ or $\xi_j$ coefficients by $\delta$ leads to a minor modification of the RTE coefficients along the LC of the pilot points, whose calculation scales as $O({N_{\Lambda}} {N_{\vec\Omega}}{N_{\lambda}} N)=O({M_\xi} {N_{\vec\Omega}}{N_{\lambda}} N)$ (we recall here that we imposed ${N_{\Lambda}}={M_\xi}$ in Sect.~\ref{ssec:sampling}). We emphasize that in this new approach the change of $\psi$ variables is not followed by the calculation of the corresponding changes of $\xi$ variables because all the variables are now independent.
The calculation of the loss function gradient consists of three terms:
\begin{equation}
\frac{\partial\mathcal{L}}{\partial\theta_j}=\frac{\partial\chi^2}{\partial\theta_j}+\lambda\frac{\partial\mathcal{L}_\Lambda}{\partial\theta_j}+\gamma\frac{\partial\mathcal{L}_{\rm L}}{\partial\theta_j}\,.
\label{eq:lgrad}
\end{equation}
Regarding the scaling, the first term on the right-hand side is on the order of $O({N_{\lambda}} N^3),$ as described in Sect.~\ref{ssec:invstd}, the second one is on the order of $O({N_{\Lambda}}{N_{\vec\Omega}}{N_{\lambda}} N)=O({M_\xi}{N_{\vec\Omega}}{N_{\lambda}} N)$, and the last term is on the order of $O({N_{\rm L}})$. Even though the inner loop in Algorithm~\ref{alg:inv2} is over ${N_{\rm V}}$ variables (which is larger than ${N_\psi}{M_\psi}$ in Algorithm~\ref{alg:inv1} because we need to explicitly consider the atomic-like variables), the order of magnitude for ${N_{\rm V}}$ and ${N_\psi}{M_\psi}$ is expected to be the same in practical applications if the number of spectral lines in the problem is relatively small. The comparison with the grid-based scaling of $O({N_\psi}{M_\psi}{N_{\vec\Omega}}{N_{\lambda}} N^3)$ shows that if the number of iterations of both inversion methods is similar, the meshfree algorithm can be faster because all three derivatives on the right-hand side of Eq.~\eqref{eq:lgrad} are significantly faster than the NLTE-consistent derivative $\partial\chi^2(D;\psi)/\partial\psi_j$ of Algorithm~\ref{alg:inv1}.
In this work, we have assumed that the set of basis functions (and, in particular, its dimension) of the MHD-like and atomic-like quantities (Eqs.~\eqref{eq:psiexp} and \eqref{eq:xiexp}, respectively) are known. In some particular cases they can be estimated empirically but, in general, neither ${M_\psi}$ nor ${M_\xi}$ are known a priori and, thus, they need to be determined during the inversion process. Therefore, ${M_\psi}$ must be such that the model is capable to fit the observed data and ${N_{\rm L}}$ must be such that the computational domain is sufficiently finely sampled so that we can guarantee consistency of the physical quantities. This work is meant to be a discussion of the general framework of the inversion method rather than an exhaustive guideline for practical inversions and, consequently, we do not discuss the values of ${M_\psi}$ and ${N_{\rm L}}$ in detail here. Instead, we simply assume that optimal values of ${M_\psi}$, ${N_{\Lambda}}={M_\xi}$, and ${N_{\rm L}}$ are known and fixed.
However, what we do want to stress here is one important aspect related to the NLTE-consistency criteria. In the grid-based methods, it is possible to reach NLTE consistency with a grid of any coarseness. Indeed, this solution is only approximate due to the finite discretization of the medium: it does provide the solution in the grid points, but the values of the atomic-like variables are only approximate and dependent on the discretization \citep{1994A&A...292..599A}. Similarly, if we approximate the spatial distribution of the atomic-like variables by an expansion in ${M_\xi}$ basis functions, we can find a consistent solution in ${N_{\Lambda}}={M_\xi}$ pilot points but due to the non-linearity of the NLTE problem, the consistency is not fully guaranteed in other points unless ${M_\xi}$ is so high (potentially infinite) that the solution is practically exact. In both grid-based and meshfree methods, we need to find the right compromise between accuracy and computation time.
In grid-based methods, we can estimate the error in the self-consistent solution by considering grids with different spatial resolutions \citep[see][]{1995ApJ...455..646T}. Similarly, in the meshfree method, we can increase the number of basis functions and of pilot points. When a sufficiently large value of ${N_{\Lambda}}={M_\xi}$ is reached so that any increase leads to a negligible change of $\mathcal{L}_\Lambda$, it is an indication that a sufficient NLTE consistency has been achieved. We come back to this problem in Sect.~\ref{sec:stoch}.
\subsection{Parallelization and memory requirements}\label{ssec:parmem1}
The parallelization of multi-dimensional grid-based NLTE codes is difficult because the RTE needs to be solved following a particular order of grid points that depends on the radiation propagation direction. Different techniques based on domain decomposition and parallelization in the propagation directions and wavelengths have been developed in the past, such as the PORTA code \citep{2013PORTA}, which implements domain decomposition only in the vertical direction, resulting in good scaling with the number of CPUs but relatively high memory constraints, while the MULTI3D code \citep{2009ASPC..415...87L} implements a 3D domain decomposition, allowing for a larger distribution of sub-domains among CPUs but resulting in relatively worse scaling due to the need of iterating the boundary conditions of the domains.
One of the advantages of our meshfree method is that parallelization of the inversion is straightforward: each of the ${N_{\rm V}}$ variables in loop 5--7 in Algorithm~\ref{alg:inv2} can be treated independently. Moreover, the RTE can also be solved independently for every LC associated with the pilot points or field of view pixels. This allows us to fully parallelize the solution with up to ${N_{\rm V}} {N_{\Lambda}} {N_{\vec\Omega}}$ (or ${N_{\rm V}} N^2$ in the case of the $\chi^2$ evaluation) processes and the scaling with the number of processes will be practically linear. Given that these numbers exceed the tens of thousands in any practical application, the parallelization can be massive.
It follows, based on its parameterization, that the amount of memory needed to store the model is proportional to ${N_{\rm V}}$. For models that are computable in a reasonable time with current supercomputers, ${N_{\rm V}}$ would probably not exceed the order of a million significantly, which implies tens of megabytes needed to store the model. Consequently, the whole model parameterization can be stored in memory at every parallel process and no domain decomposition is ever needed.
However, from the computational point of view, it is advantageous to keep in the computer memory the LC of the pilot points and of the output pixels during the $\nabla\mathcal{L}$ calculations, as well as some information on the RTE coefficients. This guarantees that fast corrections of the RTE coefficients can be calculated in the inner loop of Algorithm~\ref{alg:inv2} and the scaling discussed in Sect.~\ref{ssec:unconstr} is guaranteed. The amount of these data can become too large with an increasing number of pilot points, ${N_{\Lambda}}$, or the improved resolution of the observation. We provide a solution for this drawback of the meshfree method in Sect.~\ref{sec:stoch}.
\subsection{Suitability of the meshfree method}\label{ssec:suitability}
\begin{figure}
\begin{center}
\includegraphics[width=0.65\columnwidth]{fig-suit}
\end{center}
\caption{Comparison between the efficiency of the meshfree and grid-based methods. The horizontal axis shows the grid resolution, while the vertical axis shows the number of coefficients per model parameter. The gray shaded area below the red curve $M=N^2$ shows the models more efficiently inverted using the meshfree method. The black curves connect models with identical required CPU time per iteration of a grid-based model. See the main text for further details.}
\label{fig:suit}
\end{figure}
Let us study the suitability of the meshfree method by comparing the required computation time of both meshfree and grid-based methods for a broad set of models. For this analysis, we assume that the number of iterations ${N_{\rm it}}$ needed by both methods is similar. It is difficult to analytically show that this is the general case, but our numerical experiments with 2.5D grid-based and 3D meshfree models indicate that this is likely the case in many situations. Under this assumption, the comparison of the solution with both methods is equivalent to compare just one iteration.
One straightforward condition that the meshfree method must fulfill to be more efficient is that the first and second terms on the right-hand side of Eq.~\eqref{eq:nvar} must be on a comparable order, that is, the number of $\theta$ variables should not be much larger than the number of $\psi$ variables. Given that the number of spectral lines in the model is not too great, this is usually expected to be the case.
By comparing the scaling of the calculation of each gradient component between the meshfree ($O({N_{\lambda}} N^3 + {M_\xi}{N_{\vec\Omega}}{N_{\lambda}} N + {N_{\rm L}})$, described in Sect.~\ref{sssec:newalg}, and the grid-based ($O({N_{\vec\Omega}}{N_{\lambda}} N^3)$, described in Sect.~\ref{ssec:invstd}, methods, we can roughly estimate that the former will be faster if
\begin{equation}
\frac{1}{{N_{\vec\Omega}}} + \frac{{M_\xi}}{N^2} + \frac{{N_{\rm L}}}{{N_{\vec\Omega}}{N_{\lambda}} N^3} \sim \frac{{M_\xi}}{N^2} < 1
\label{eq:cond}
\end{equation}
is satisfied, where we have taken into account that the first and third terms on the left-hand side of the inequality are clearly negligible.\footnote{In the 3D radiative transfer we always have ${N_{\vec\Omega}}\sim 10^2$.} In Fig.~\ref{fig:suit}, we show this resulting condition with the shaded region below the red curve in a diagram of number of basis functions, $M,$ per quantity versus the grid resolution. Here, we are assuming ${M_\xi}\approx{M_\psi}$, $M$ represents any of the MHD- or atomic-like quantities. The gray area where the condition in Eq.~\eqref{eq:cond} holds shows the models that are more efficiently calculated using the meshfree method.
We can thus conclude that the meshfree method is superior for models and observations with high resolution and somewhat limited spatial variability of the parameters (smaller $M$). Likewise, the grid-based methods can be more efficient in the case of low-resolution models and observations with abrupt spatial variability of the parameters.
The black curves in Fig.~\ref{fig:suit} correspond to curves with constant $MN^3$, proportional to the CPU time per iteration of the grid-based method. The area below each of them corresponds to the models that can be inverted within a given maximum available CPU time. One important observation is that if a limited CPU time is given (i.e., a particular black curve in the plot) and a particular model or observation resolution, $N$, then it is always possible to find a meshfree solution for certain values of $M$ below the red curve. Given the parabolic shape of the red curve, it follows that grid-based methods are increasingly less efficient compared to the meshfree method as the model or observation resolution increases.
\section{Stochastic inversion}\label{sec:stoch}
In the previous section, we introduce several concepts leading to the formulation of a meshfree method to solve the 3DNIP. Even though the method appears promising, there are some problems that need to be solved to achieve a truly efficient algorithm, such as the fact that with ${N_{\Lambda}}={M_\xi}$ fixed pilot points, we cannot fully guarantee the NLTE consistency in the whole computational domain or the memory allocation of a large number of LC. Additionally, in this section, we discuss the problem of local minima for $\mathcal{L}$ and the artifacts introduced by the use of fixed angular quadratures in 3D.
\subsection{Reshuffling of the pilot points and pixels: The stochastic algorithm}
One of the issues of the meshfree method, as introduced in Sect.~\ref{sec:method1} and mentioned in Sect.~\ref{sssec:newalg}, is the impossibility of ensuring the NLTE consistency in the whole domain by testing this consistency in just ${M_\xi}\ll N^3$ random pilot points. One possible solution to this problem could be (after the solution has converged) to change the set of pilot points and to check whether the solution is still NLTE consistent. While this approach would result in a slower method (we would be increasing the number of required iterations without changing the cost per iteration), it directs us toward a different approach: we can generate a new set of pilot points after every $n<{N_{\rm it}}$ iterations, even before the convergence is reached for a given set of pilot points. In particular, we can use a small number of pilot points, ${N_{\Lambda}}\ll{M_\xi}$, which are reshuffled after every $n$ iterations (with a lower limit of $n=1$). This allows us to eventually sample the whole 3D domain. Of course, changing the set of pilot points before reaching convergence leads to a much greater final number of iterations, but these are also much faster than in the standard algorithm because of the condition ${N_{\Lambda}}\ll{M_\xi}$.
With this approach, estimating the $\mathcal{L}_\Lambda$ gradient goes to the order of $O({N_{\Lambda}}{N_{\vec\Omega}}{N_{\lambda}} N)$, which much smaller than $O({M_\xi}{N_{\vec\Omega}}{N_{\lambda}} N)$. The most time-consuming part of the $\mathcal{L}$ gradient estimation would now be the $\chi^2$ function gradient, which is still on the order of $O({N_{\lambda}} N^3)$. We would have a very fast although inaccurate estimation of the $\mathcal{L}_\Lambda$ gradient, but an accurate and very slow estimation of the $\chi^2$ gradient. We can thus balance the computing time of these quantities by applying a similar strategy to the calculation of the $\chi^2$ gradient, namely, instead of evaluating $\chi^2$ in $N^2$ pixels, we randomly generate ${N_{\chi}}<N^2$ pixels, to be reshuffled after every iteration, in which we calculate an approximation of this quantity. The balancing of the scaling of the $\mathcal{L}_\Lambda$ and $\chi^2$ gradients can be done by taking ${N_{\Lambda}}{N_{\vec\Omega}}\approx{N_{\chi}}$.
Naturally, it is also possible to decrease ${N_{\rm L}}$ and to apply the same reshuffling strategy to the estimation of the $\mathcal{L}_{\rm L}$ gradient.
This procedure can only provide an approximation to the gradient $\nabla\mathcal{L}$, but every iteration becomes significantly faster than before. This approach is used in the so-called stochastic gradient descent (SGD) class of methods and it provides several benefits over the traditional gradient descent (see below). The approximate loss function now takes the form:
\begin{eqnarray}
\mathcal{L}(D;\theta;{N_{\chi}},{N_{\Lambda}},{N_{\rm L}})=\nonumber\\
\chi^2(D;\theta;{N_{\chi}})+\lambda \mathcal{L}_\Lambda(\theta;{N_{\Lambda}}) + \gamma \mathcal{L}_{\rm L}(\theta;{N_{\rm L}})\,.
\label{eq:llapprox}
\end{eqnarray}
The expression for $\chi^2$ is identical to that in Eq.~\eqref{eq:chi2}, except for ${N_{\rm pix}}$ being replaced by ${N_{\chi}}$. The general structure of Algorithm~\ref{alg:inv2} remains valid, but the loss function at line 6 is replaced by that of Eq.~\eqref{eq:llapprox}. Moreover, the convergence criteria at line 9 must be modified in such a way that the convergence is guaranteed not only for a single sample of ${N_{\Lambda}}$ and ${N_{\chi}}$ points, but for the whole domain. To this end, the stopping criteria should not only involve the current value of the loss function components and of their gradients, but also their values over time. We leave the discussion of this technical issue for later works.
One last change is necessary at line 8 of Algorithm~\ref{alg:inv2}, where new values of the model parameters are calculated using the estimation of the gradient. A number of SGD algorithms have been developed in the recent years and led to much better convergence rates than the naive approach outlined in Algorithm~\ref{alg:inv2}. In our calculations, we have used the ADAM algorithm \citep{2014arXiv1412.6980K}, which uses running averages of the gradient and of its second moment in order to estimate the new values of the problem variables. We have found that it provides good convergence results for the test presented in this work (see Sect.~\ref{sec:example} for details).
In the stochastic approach outlined in this section, we replaced the small number of computationally intensive iterations from the method described in Sect.~\ref{sec:method1} by a large number of very fast iterations. With a sufficient number of iterations, we can guarantee that the solution is consistent in the whole domain and not just in ${M_\xi}$ fixed points. As we show in the example below (Sect.~\ref{sec:example}), the number of stochastic iterations exceeds the thousands or tens of thousands, hence, the computational domain can be effectively sampled with a small number of ${N_{\Lambda}}$, ${N_{\chi}}$, and ${N_{\rm L}}$ points.
We note that even though we are testing the NLTE consistency in just ${N_{\Lambda}}$ pilot points during each iteration, we are also effectively probing the model along the LC themselves (cf. Fig.~\ref{fig:lc}). Due to the non-local character of the radiative transfer problem, an error accumulated along the LC will produce an error at the pilot points, in contrast to the loss function of local quantities, $\mathcal{L}_{\rm L}$.
Finally, the convergence analysis in the SGD is more difficult and, thus, we leave its discussion to future publications. In this work, we restrict ourselves to an empirical convergence test in Sect.~\ref{sec:example} which shows that the method can converge very quickly. The numbers ${N_{\Lambda}}$, ${N_{\chi}}$, and ${N_{\rm L}}$ determine how noisy the gradient and the convergence are, with higher values leading to a better convergence rate with a less noisy gradient at the cost of larger CPU time and memory requirements per iteration.
\subsection{Convexity, local minima, and an analogy with deep learning}
The NLTE problem is strongly non-linear. We are not aware of any existing rigorous mathematical analysis of the equations for multilevel systems, but it is very likely that the $\chi^2$ and $\mathcal{L}$ functions are also non-convex; hence, there are a number of local minima in which the inversion algorithm can end up. In fact, our numerical experiments with deterministic inversion algorithms show that it is not a rare occurrence.
The use of a SGD method helps to partially solve this problem because unlike deterministic methods, the stochastic crawling through the parameter space in the SGD method is not slated to remain in a local minimum. This is simply due to the fact that the exact shape of the loss function landscape (see also the background color in Fig.~\ref{fig:unconstr}) is unknown and its estimation changes between iterations; hence, the local minima may be passed through, in contrast to what happens in the standard gradient descent method with a more accurate and unchanging estimation of the loss function.
Recently, SGD methods have become important in the context of machine learning techniques. While our method is not based on these methods, an analogy can be made with the training of deep neural networks: as in our method, deep learning can be understood as a global optimization process with a substantial number of parameters. The network training proceeds by feeding a large example data set and evaluating the gradient of the loss function. In practice, it is unfeasible to use the whole database of examples in every iteration of the training. Instead, the network is fed with a relatively small number of randomly chosen examples and the loss function gradient is only approximately calculated (so-called mini-batch training). The use of a small number of random pilot points per iteration in our inversion method can be seen as analogous to the mini-batch training, while the calculation of the loss function gradient in all domain points would correspond to using the whole database of examples in every iteration of the training.
\subsection{Angular quadratures artifacts}\label{ssec:quad}
\begin{figure}
\begin{center}
\includegraphics[width=0.65\columnwidth]{fig-quad}
\end{center}
\caption{Geometry of the illumination of the computational domain by the underlying solar surface using an angular quadrature. If the angular quadrature is identical at every point (see the inclined lines), artificial sharp jumps in the illumination at the boundary, which do not actually exist, will appear in the model. Such artifacts could make a full 3D inversion nearly impossible. The solution to this problem is to use a randomized quadrature at every pilot point.}
\label{fig:quad}
\end{figure}
The accurate calculation of the $\bar J^K_Q$ tensors in the pilot points requires us to solve the RTE along the LC in the directions of a sufficiently accurate angular quadrature. For a discussion of optimal quadratures for the transfer of polarized radiation, see \cite{2020quad1} and \cite{2021quad2}.
Under the physical conditions in the solar atmosphere, the number of propagation directions of a good quadrature is typically on the order of ${N_{\vec\Omega}}\sim 10^2$. As with the truncation error in the grid-based methods, the angular discretization necessarily leads to some degree of numerical error, but if an appropriate quadrature is chosen, this error can be negligible. However, a problem arises with the interaction between the quadrature and the boundary conditions of a 3D domain in the case of plasma structures embedded in the solar corona, that is, in structures such as prominences and filaments. Given a fixed quadrature and the illumination from the underlying solar surface, there will be a sharp and artificial jump in the boundary conditions, as demonstrated in Fig.~\ref{fig:quad}.
An easy solution to this problem is possible within the framework of the stochastic approach: in the same way that the pilot points are randomized in every iteration, the orientation of the quadrature rays can also be randomized in each of these points. As a result, there are no preferable directions in the radiation transfer and artifacts similar to that in Fig.~\ref{fig:quad} cannot appear. For obvious reasons, this randomization of the angular quadrature orientation cannot be implemented in the grid-based methods relying on the short-characteristics method.
In our 2.5D grid-based experiment we have found that these artifacts can actually appear in practice and they significantly complicate the solution. As a byproduct of the stochastic approach, we have the means to solve another critical problem that would actually make the full 3D inversion practically impossible.
\subsection{Parallelization and memory requirements}\label{ssec:parmem2}
The parallelization of the stochastic method is as straightforward as in the deterministic method described in Sect.~\ref{sec:method1}. Given that in every iteration we need to solve the radiative transfer along ${N_{\rm V}}({N_{\vec\Omega}}{N_{\Lambda}}+{N_{\chi}})$ independent long characteristics. Since, in practice, we always have ${N_{\rm V}},\,{N_{\rm pix}}>10^2$ and ${N_{\vec\Omega}}\approx 10^2$, and ${N_{\Lambda}}$ is such that, typically, ${N_{\vec\Omega}}{N_{\Lambda}}\approx{N_{\chi}}\gtrsim 10^2$, it is clear that the method can be massively parallelized and that the scaling with the number of CPUs should be practically linear.
One of the most significant benefits of the stochastic method over the deterministic one is its random access memory requirement. As discussed in Sect.~\ref{ssec:parmem1}, it is necessary to store the data of the LC during the $\nabla\mathcal{L}$ estimation, and increasing ${M_\xi}$ and $N$ could dramatically increase the memory requirements to the point of this becoming a limiting factor in the method. However, in the stochastic method ${N_{\Lambda}}\ll{M_\xi}$ and ${N_{\chi}}\ll N^2$ and thus the issue is non-existent. Moreover, we can always choose ${N_{\Lambda}}$ and ${N_{\chi}}$ such that the method fits any memory constraint.
\section{Example application}\label{sec:example}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig-example}
\end{center}
\caption{Geometry of the model and the observation. The computational domain (black cube) is illuminated by the underlying Sun (gray surface). The limb darkening of the incident continuum radiation depends on the $\theta$ heliocentric angle (see the text for details). The LOS, which is in the direction of the $z$ axis, and the local solar vertical direction ($y$ axis) form a 90$^\circ$ angle. In this academic example, the solar surface is approximated by an infinite plane.}
\label{fig:geom}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=0.7\columnwidth]{fig-mg}
\includegraphics[width=0.7\columnwidth]{fig-mg-inv}
\end{center}
\caption{Magnetic field vector in the synthetic (left) and inverted (right) models.}
\label{fig:mg}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{fig-observation} \\
\includegraphics[width=\textwidth]{fig-inversion}
\end{center}
\caption{
Synthetic observation (top row) and emergent radiation from the inverted model (bottom row). From left to right: $I$, $Q$, $U$ in the line center, and $V$ at around the wavelength $\lambda\approx 1.4$. The Stokes parameters are in the disk-center intensity units, $I(\mu=1)$. The positive $Q$ direction is parallel to the solar limb (parallel to the $x$ axis, cf. Fig.~\ref{fig:geom}). The synthetic signal is contaminated with Gaussian noise with $\sigma=4\cdot 10^{-4}$, in the disk-center intensity units.
}
\label{fig:obs}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{fig-observation-B0}
\end{center}
\caption{Emergent radiation from the same model as in the top row in Fig.~\ref{fig:obs}, but after switching off the magnetic field. This figure demonstrates that even though the maximum optical thickness of the model is only around $\tau\approx 1$, the symmetry breaking effects due to 3D radiative transfer play a significant role and the Stokes~$U$ signal is far from zero. Neglecting 3D radiative transfer could lead to serious errors in the interpretation of the observations.}
\label{fig:obsb0}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig-conv}
\end{center}
\caption{Convergence of the stochastic meshfree method. The individual curves correspond to the respective terms in Eq.~(\ref{eq:llapprox}) shown in the legend. The bottom horizontal axis shows the number of stochastic iterations, while the top horizontal axis shows the CPU time in units of full $\Lambda$-iterations in a $N^3=64^3$ grid-based model.}
\label{fig:conv}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{fig-errors}
\end{center}
\caption{Scatter plots of the true model (horizontal axes) and inverted (vertical axes) model quantities in 5\,000 randomly distributed points within the domain. From left to right: Line opacity, line Doppler width, and $x$-, $y$-, and $z$-components of the magnetic field vector.}
\label{fig:err}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{fig-profs}
\end{center}
\caption{Example of the inversion fit (orange curve) to the observed (blue dots) Stokes profiles at the spatial point $(x,y)=(0.75,-0.25)$}
\label{fig:profs}
\end{figure*}
In this section, we apply the previously developed stochastic meshfree method to invert the physical properties of a solar prominence-like 3D structure. For the sake of simplicity in the demonstration of the method, and without loss of generality, we use a dimensionless academic model. The chosen model is intentionally very simple and its purpose is to test the inversion algorithm.
\subsection{Model definition and synthetic observations}
The computational domain is a 3D cube with dimensions $[-1,1]^3$ suspended above the solar surface and observed along an LOS perpendicular to one of its faces (see Fig.~\ref{fig:geom}). In this example, we assume that the scattering geometry is fully known a priori during the inversion process and that we can expect the observed plasma to be completely confined within the cubic box. Needless to say, none of these assumptions would be generally satisfied in an actual observation. The boundary conditions for the illumination are chosen such that they resemble the irradiation from the underlying solar photosphere. This unpolarized spectrally flat radiation is limb-darkened according to the rule:
\begin{equation}
I(\mu)=
\begin{dcases*}
1-\tfrac 12 (1-\mu^2) & for $\mu>0$ \\
0 & for $\mu\le 0$
\end{dcases*}\;,
\label{eq:clv}
\end{equation}
where $\mu=\cos\theta$, with $\theta$ the heliocentric angle (i.e., the angle between the propagation direction and the normal to the solar surface).
At this point, we are considering an academic problem of a normal Zeeman triplet susceptible to the Hanle and Zeeman effects. Further details on the spectral line are given in Appendix \ref{ssec:app-atom}. For the sake of simplicity, we consider a spherically symmetric plasma distribution with the line opacity decreasing and the line Doppler width increasing with the distance from the domain center. This model qualitatively represents cold prominence plasma embedded in a hot surrounding corona. Details of the particular parameterization can be found in Appendix \ref{sec:app-plasma}, together with the simple, albeit non-trivial, configuration of the magnetic field vector shown in the left panel of Fig.~\ref{fig:mg}. For additional details on the model, see Appendix~\ref{sec:app-model}.
Assuming that our observation consists of the four Stokes profiles, sampled in ${N_{\lambda}}=47$ wavelengths, for a ${N_{\rm pix}}=64\times 64$ field of view, we synthesized the observation with the PORTA code in an atmospheric model with a spatial grid of $N^3=64^3$ points. In order to mimic more realistic observations, we added random Gaussian noise with a standard deviation of $\sigma=4\cdot 10^{-4}$ in units of the disk center intensity. Given the maximum observed intensity, the noise-to-signal ratio is always larger than $10^{-3}$ in the observations. The synthetic observations are shown in the top panel of Fig.~\ref{fig:obs}.
In Fig.~\ref{fig:obsb0}, we show the same synthetic data, calculated in the same model atmosphere but without any magnetic field. While the circular polarization (Stokes~$V$) is obviously zero and the intensity is practically unaffected, the linear polarization components (Stokes $Q$ and $U$) are significantly different from the magnetized case due to the missing Hanle effect. Even though the optical thickness of the plasma structure is on the order of 1 (see Eq.~\ref{eq:tau}), the symmetry breaking due to the 3D radiative transfer within the medium is sufficiently strong to produce non-zero $U$ signals of the same order as the Stokes $Q$ ones. The inversion of this data using the pixel-by-pixel approach would lead to erroneous conclusions about the presence of magnetic fields.
\subsection{Basis functions and the inversion setup}\label{ssec:basinv}
The MHD-like $\psi$ quantities to invert are: $\chi_{\rm L},\,\Delta\lambda_{\rm D},\,\Gamma_x,\,\Gamma_y,$ and $\Gamma_z$, namely: the line opacity, the line Doppler width, and the Cartesian magnetic field components, defined in Appendix~\ref{ssec:app-atom}. While in this particular model, all the quantities are dimensionless by definition, in realistic applications, a suitable normalization should be applied. The model atmosphere is static, hence we are not considering any macroscopic velocity field. Both elastic and inelastic collisions can also be (and are) neglected. In this simple model, the line opacity is equivalent to the density, while the Doppler width is equivalent to the kinetic temperature.
Because we adopted a two-level atom with unpolarized lower state, for this example, the upper-level density matrix, $\rho^K_Q(u),$ and the mean radiation field tensors, $\bar J^K_Q$, are completely equivalent in describing the atomic state. In this example, we chose the $\bar J^K_Q$ components as our atomic-like $\xi$ variables: $\bar J^0_0,\,\bar J^2_0,\,\Re \bar J^2_1,\,\Im \bar J^2_1,\,\Re \bar J^2_2,$ and $\Im \bar J^2_2$, where $\Re$ and $\Im$ stand for the real and imaginary parts, respectively. The components with rank $K=1$ remain identically zero because of the adopted model. We normalized the atomic like variables with the disk-center intensity and, in addition, we scaled the components with $K=2$ with a factor $20$ because we can expect the anisotropy of the radiation field to be on the order of few percent of the mean intensity.
For the function basis to approximate the spatial distribution of the problem quantities, we chose the Chebyshev polynomials of the first kind, using the tetrahedron subset (see the central panel of Fig.~\ref{fig:baf}). The maximum orders of the basis functions for the MHD-like quantities are $p(\chi_{\rm L}) = p(\Delta\lambda_{\rm D}) = 2$ and $p(\Gamma_x) = p(\Gamma_y) = p(\Gamma_z) = 1$ -- which, according to Eq.~\eqref{eq:mp}, entails ${M_\psi}(\chi_{\rm L})={M_\psi}(\Delta\lambda_{\rm D})=10$ and ${M_\psi}(\Gamma_x) = {M_\psi}(\Gamma_y) = {M_\psi}(\Gamma_z)=4,$ with the total number of MHD-like variables equal to 32. In this example, we can afford to determine the basis from the a priori knowledge of the parameterization of the model (cf. Eqs.~\ref{eq:chir} to \ref{eq:bz}) but, in general, we would have to use adequate techniques to determine them (see Sect.~\ref{sssec:newalg}).
We empirically found via numerical experimentation that the order of the basis of the atomic-like variables sufficient to reach NLTE consistency in our example is $p(\bar J^K_Q)=3$ -- which, using Eq.~\eqref{eq:mp}, gives us ${M_\xi}=20$ and the total number of atomic-like variables is ${N_\xi}{M_\xi}=5\cdot 20=120$.
Given the relative amplitudes of the Stokes signals in the observations (see top panels in Fig.~\ref{fig:obs}), we set the weights to $w_I=1$, $w_Q=w_U=20$, and $w_V=200$ in the $\chi^2$ function in Eq.~(\ref{eq:chi2}). We note that these are the values of $w_k$ before their normalization to 1. The weight of the NLTE regularization factor is set to $\lambda=10^4$ with the aim to reduce the residual NLTE error below $10^{-4}$. We set $\gamma=1$ for the $\mathcal{L}_{\rm L}$ term, a value sufficient to end up with a locally consistent solution in this example, as we show below.
The model was initialized with random values of the variables and we used ${N_{\Lambda}}=3$ pilot points, ${N_{\chi}}=10$ probing pixels, and ${N_{\rm L}}=10$. The angular quadrature in the pilot points is the 88-point $L=11$ quadrature of \citet{2020quad1}. Alternatively, it might be preferable to use even more accurate quadratures by \citet{2021quad2}. In this example, we do not consider the randomization of the quadrature orientation described in Sect.~\ref{ssec:quad} because the boundary conditions of Eq.~(\ref{eq:clv}) are such that the problem does not suffer from disk-edge artifacts.
We have implemented the meshfree stochastic method in a parallel C code and solved the inversion problem using the ADAM algorithm with parameters $\alpha=0.001$, $\beta_1=0.9$, $\beta_2=0.999$, and $\epsilon=10^{-8}$ \citep[see][for details]{2014arXiv1412.6980K}.
\subsection{Results}
In Fig.~\ref{fig:conv}, we show the convergence of the total loss function, $\mathcal{L}$, as well as of the penalty functions, $\mathcal{L}_\Lambda$ and $\mathcal{L}_{\rm L}$, and of the $\chi^2$ function. The $\chi^2$ eventually converges almost to 1 and $\mathcal{L}_\Lambda$ and $\mathcal{L}_{\rm L}$ decrease to the values corresponding to small deviations from the full consistency. It is worth to say that we have not used any quantitative stopping criteria nor did we extensively experimented with the hyperparameters affecting the convergence. The results in Fig.~\ref{fig:conv} are given only to demonstrate the general behavior of the convergence in the numerical experiments we performed and we note that the problem requires a more thorough analysis.
In terms of CPU time, every iteration shown in Fig.~\ref{fig:conv} requires about 5 seconds using a common contemporary CPU. The total computing time of this inversion with 16\,000 iterations is of about 22 hours. In other words, this is equivalent to about 20 seconds per pixel. It is interesting to compare this time with an estimation of the inversion time of the very same model using the standard $\chi^2$ minimization of a grid-based method. In the PORTA code, that could be used as a $\Lambda$-iteration engine, the CPU time per mesh point, per wavelength, and per quadrature propagation direction is of about $t=10^{-6}$ seconds using the same CPU. One $\Lambda$ iteration with four Stokes parameters therefore requires about $4 N^3 N_\lambda N_\Omega t \approx 4\,300$\,seconds. Assuming, being very optimistic, that we would only need one $\Lambda$ iteration to calculate a derivative of $\chi^2$ with respect to each of the $\psi$ variables and if we neglect the computing time required to solve the NLTE problem at the end of every iteration and the time needed to synthesize the emergent radiation in every step of the $\chi^2$ gradient calculation, we would need about $32\cdot 4\,300\approx 38$\,hours per iteration because the number of the MHD-like variables is 32. Since a realistic number of such iterations is at least ${N_{\rm it}}=100$, the inversion time using such method would require about 3\,800 hours, that is, about a factor 170 more than our solution.\footnote{However, in practice, the number of $\Lambda$ operator evaluations would be larger than one per gradient component in the grid-based methods, hence the speedup of the solution in the meshfree method would be $\gtrsim 340$.} We note that this significant difference is not just due to the fact that our model is quite smooth because in both approaches the CPU time scales with the level of smoothness quantified by ${M_\psi}$.
The scatter plots in Fig.~\ref{fig:err} show the correlation of the synthetic and the inverted model quantities at the same spatial points. The diagonal red lines in every panel show the span of the correct values and, in the case of a perfect inversion, all the black points should be on these diognals. Even though the match is not perfect due to the limited accuracy of the basis, the presence of noise, and the slightly premature stopping of the iterative process, the agreement seems to be quite satisfying for all five quantities in the whole 3D domain. In the right panel of Fig.~\ref{fig:mg} we visualize the inferred magnetic field vector, which is indistinguishable at the same plot level as the one in the original model (left panel of the same figure).
The bottom row in Fig.~\ref{fig:obs} shows the emergent radiation corresponding to the inferred model, and compared with the top row in the same figure, we can see that the agreement is remarkable. An example of the good quality of the fit to the observed Stokes profiles at a particular pixel in the field of view is shown in Fig.~\ref{fig:profs}.
\section{Discussion and conclusions\label{sec:concl}}
The unsolved problem of 3DNIP is generally considered to be one of the greatest challenges to face in the theory of radiation transfer. In this paper, we present a first attempt to solve it. Our approach is not to generalize in a brute-force manner the standard 1D methods into the 3D geometry because such an approach is doomed to failure for a number of reasons.
In inversion methods based on the pixel-by-pixel approach, the NLTE radiative coupling of different regions (in the direction perpendicular to the LOS) of the plasma is usually considered to be a complication. We have developed a meshfree approach that takes this coupling as an advantageous natural regularization which leads to more robust, more physically correct, and computationally faster solutions.
The new method employs the idea that 3DNIP can be solved as an unconstrained minimization problem in which unphysical solutions are allowed during the iterative process. We show that the method has the potential to be much faster than methods based on 3D grids with the recurring evaluations of the $\Lambda$ operator. This approach promises to provide fast solutions, especially in case of relatively smooth models, but it can also provide at least approximate solutions in complex models that would be completely unsolvable using grid-based techniques.
We can summarize the main advantages of the proposed framework as follows:
\begin{enumerate}
\item Consistent 3D\,NLTE solution with scattering polarization, Hanle, and Zeeman effects fully taken into account.
\item Additional conditions for physical consistency, such as $\nabla\cdot\vec B=0$ or the equation of continuity, can be naturally incorporated as penalty terms to the loss function.
\item Since all the pixels are inverted together, the solution is more robust and less sensitive to noise in the data than the pixel-by-pixel methods.
\item Stochastic angular quadratures make it possible to avoid unphysical discontinuities in the boundary illumination in the case of the prominence or filament geometry.
\item Long characteristics lead to more accurate calculation of the radiation field in the pilot points than short characteristics in the grid-based models.
\item There is no numerical error due to the finite spacing of the 3D mesh.
\item The method based on the modern algorithms of stochastic gradient descent is less prone to end up in a local minimum of the loss function.
\item The method can be trivially parallelized for massive HPC facilities. On the other hand, even a single desktop computer can be used to infer at least some information about the global structuring of the plasma properties --- something that would not be possible with grid-based methods.
\item Much smaller memory requirements than in the grid-based methods. There is no need for domain decomposition in order to reduce the memory demands.
\item Given the lack of a grid, the implementation of the code is easier than in grid-based methods.
\end{enumerate}
We go on to list the following disadvantages of the method:
\begin{enumerate}
\item The method is less suitable for problems with a large number of spectral lines or atomic levels because the number of variables to be inferred could grow and slow down the calculation (see Eq.~\ref{eq:nvar}).
\item In the case of abrupt changes of physical parameters, the parameterized model can have difficulties to accurately describe such discontinuities. This problem is inherent not only to the method presented here, but to every method using an expansion of the physical parameters into a predefined basis.
\item Given the representation of the atomic-like quantities in terms of an expansion in a finite set of basis functions, there may be certain residual NLTE inconsistency in the solution. However, this inconsistency can be reduced to any desired level of precision.
\end{enumerate}
We intentionally do not discuss a number of important issues related to the 3D inversion problem. Among these issues, we have the problems of possible ambiguities, as well as of the local minima, location of the observed structures along the LOS, the convergence criteria, various possibilities for adaptivity of the inversion algorithm, and many others. Last but not least, applications with realistic spectral lines need to be studied.
\begin{acknowledgements}
J.\v{S}. acknowledges the financial support of the grants \mbox{19-20632S} and \mbox{19-16890S} of the Czech Grant Foundation (GA\v{C}R) and the support from project \mbox{RVO:67985815} of the Astronomical Institute of the Czech Academy of Sciences. We acknowledge the funding received from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (ERC Advanced Grant agreement No~742265). We are grateful to Roberto Casini and Luca Belluzzi for their insightful comments which helped us to improve the content of the paper.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Social bots are social media accounts controlled in part by software that can post content and interact with other accounts programmatically and possibly automatically~\cite{ferrara2016rise}.
While many social bots are benign, malicious bots can deceptively impersonate humans to manipulate and pollute the information ecosystem.
Such malicious bots are involved with all types of online discussions, especially controversial ones.
Studies have identified interference of social bots in U.S. elections~\cite{shao_spread_2018,gorodnichenko_social_2021,bessi_social_2016,ferrara_characterizing_2020}, French elections~\cite{ferrara_disinformation_2017}, the Brexit referendum~\cite{bastos_public_2018,bastos_brexit_2019,gorodnichenko_social_2021,duh_collective_2018}, German elections~\cite{keller_social_2019}, and the 2017 Catalan referendum~\cite{stella_bots_2018}.
Bots also actively participate in public health debates~\cite{jamison_malicious_2019} including those about vaccines~\cite{broniatowski_weaponized_2018,yuan_examining_2019}, the COVID-19 pandemic~\cite{ferrara_what_2020,shi_social_2020,uyheng_bots_2020,yang2020prevalence}, and cannabis~\cite{allem_cannabis_2020}.
Research has also reported on the presence of social bots in discussions about climate change~\cite{marlow_twitter_2020,marlow_bots_2021,chen_social_2021}, cryptocurrency~\cite{nizzoli_charting_2020}, and the stock market~\cite{cresci_cashtag_2019,fan_social_2020}.
Malicious social bots demonstrate various behavioral patterns in their actions.
They may simply generate a large volume of posts to amplify certain narratives~\cite{marlow_bots_2021,keller2020political} or to manipulate the price of stocks~\cite{cresci_cashtag_2019,fan_social_2020} and cryptocurrencies~\cite{nizzoli_charting_2020}.
They can also disseminate low-credibility information strategically by getting involved in the early stage of the spreading process and targeting popular users through mentions and replies~\cite{shao_spread_2018}.
Some bots act as fake followers to inflate the popularity of other accounts~\cite{bilton_social_2014,confessore_follower_2018,varol2020journalists}.
In terms of content, malicious bots are found to engage other accounts with negative and inflammatory language~\cite{stella_bots_2018} or hate speech~\cite{albadi_hateful_2019,uyheng_bots_2020}.
In some cases, bots form dense social networks to boost engagement and popularity metrics and to amplify each other's messages~\cite{caldarelli_role_2020,torres2020trains,chen_neutral_2021}.
Most existing reports and studies on social bots focus on Twitter, largely because its data can be easily accessed.
Although Twitter strengthened their efforts to contain malicious actors in recent years,\footnote{\url{blog.twitter.com/common-thread/en/topics/stories/2021/the-secret-world-of-good-bots}} deceptive bots remain prevalent and display evolving tactics to evade detection~\cite{yang2019arming}.
This has two implications for researchers.
First, characterizing the behavior of and assessing the impact of social bots remains an interesting research topic~\cite{rahwan_machine_2019}.
Second, researchers need to properly handle bots in their data since their presence may distort analyses~\cite{jamison_malicious_2019,ledford_social_2020}.
It is therefore crucial for researchers to have access to a reliable tool for detecting social bots.
This practicum aims to provide a tutorial for Botometer, a machine learning tool for bot detection on Twitter.
Although other bot detection tools such as tweetbotornot\footnote{An R package for classifying Twitter accounts as bot or not available at \url{github.com/mkearney/Tweetbotornot}} and Bot Sentinel\footnote{A platform that classifies and tracks inauthentic accounts and toxic trolls available at \url{botsentinel.com}} exist, we focus on Botometer for several reasons.
First, it is well maintained and has been serving the community for the past seven years without major outages.
It has also been routinely upgraded to stay accurate and relevant.
Second, Botometer is easily accessible through both a web interface and an application programming interface (API).
Anyone with a Twitter account can use the web version for free; researchers with Twitter developer accounts can use the API endpoints to analyze large-scale datasets. The API has a nominal fee for heavy use, which discourages abuse and partially offsets infrastructure and maintenance costs.
Third, Botometer is quite popular. It handles around a quarter million daily queries---over half a billion in total since its inception.
Finally, Botometer has been extensively validated in the field.
Many researchers have applied Botometer in their studies to directly investigate social bots and their impact~\cite{broniatowski_weaponized_2018,keller_social_2019,allem_cannabis_2020,fan_social_2020}, or to distinguish human accounts and bot-like accounts in order to better address their questions of interest~\cite{vosoughi_spread_2018,grinberg_fake_2019,bovet_influence_2019}.
This tutorial is designed for data scientists and computational social scientists who might not be familiar with Botometer, the machine learning methods behind it, its programmatic interface, or how to interpret its results.
We start with an introduction to how Botometer works and how users can access it.
We then present a case study to demonstrate Botometer usage.
The source code for this case study is shared through a public repository for readers to replicate this analysis and use it as a template for their own research.
We finally discuss recommended practice.
\section{How Botometer works}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{botometer_history.pdf}
\caption{
The timeline of Botometer versions.
}
\label{fig:botometer_history}
\end{figure}
Figure~\ref{fig:botometer_history} presents the timeline and key characteristics of successive Botometer versions over the years.
Since the behaviors of bot and human accounts evolve over time, version upgrades are necessary for Botometer to stay accurate and relevant.
Upgrades typically included adding new training data and updating model features. The most recent version also involved major architectural changes.
Users of Botometer should be aware that results from different versions are usually not comparable and the format of input and output might change as well.
For details of early versions such as V2~\cite{varol2017online} and V3~\cite{yang2019arming}, readers can refer to the corresponding papers.
This tutorial focuses on V4~\cite{sayyadiharikandeh2020detection}.
In addition to new training data and new features, this version introduced a new architecture.
We will also briefly cover a recently added model for fast bot detection~\cite{yang_scalable_2020}.
\subsection{Supervised machine learning for bot detection}
Under the hood, Botometer is a supervised machine learning classifier that distinguishes bot-like and human-like accounts based on their features (i.e., characteristics). Unsupervised methods have also been proposed in the literature~\cite{chavoshi2016debot,echeverria2017discovery1}, but they only allow for the detection of specific, predefined behaviors. Therefore they are not suitable to build a general detection tool.
Botometer considers over 1,000 features that can be categorized into six classes: user profile, friends, network, temporal, content and language, and sentiment~\cite{varol2017online}.
For example, the user profile category includes features such as the length of the screen name, whether the account uses the default profile picture and background, the age of the account, etc.
The content and language category consists of features such as the number of verbs, nouns, and adjectives in the tweets.
For a given account, these features are extracted and encoded as numbers.
This way the account can be represented by a vector of feature numbers, enabling machine learning classifiers to process the information.
\begin{table}
\caption{
Annotated datasets of human and bot accounts used to train Botometer.
}
\centering
\begin{tabular}{lrrll}
\hline
Dataset & Bots & Humans & Annotation method & Ref. \\
\hline
\texttt{varol-icwsm} & 733 & 1,495 & Human annotation & \cite{varol2017online} \\
\texttt{cresci-17} & 7,049 & 2,764 & Various methods & \cite{cresci2017paradigm} \\
\texttt{pronbots} & 17,882 & 0 & Spam bots & \cite{yang2019arming} \\
\texttt{celebrity} & 0 & 5,918 & Celebrity accounts & \cite{yang2019arming} \\
\texttt{vendor-purchased} & 1,087 & 0 & Fake followers &\cite{yang2019arming} \\
\texttt{botometer-feedback} & 139 & 380 & Human annotation & \cite{yang2019arming} \\
\texttt{political-bots} & 62 & 0 & Human annotation & \cite{yang2019arming} \\
\texttt{gilani-17} & 1,090 & 1,413 & Human annotation & \cite{gilani2017bots} \\
\texttt{cresci-rtbust} & 353 & 340 & Human annotation &\cite{mazza_rtbust_2019} \\
\texttt{cresci-stock} & 7,102 & 6,174 & Signs of coordination & \cite{cresci_fake_2018} \\
\texttt{botwiki} & 698 & 0 & Self-declared & \cite{yang_scalable_2020} \\
\texttt{midterm-2018} & 0 & 7,459 & Human annotation & \cite{yang_scalable_2020} \\
\texttt{astroturf} & 505 & 0 & Human annotation & \cite{sayyadiharikandeh2020detection} \\
\texttt{kaiser} & 875 & 499 & Politicians + bots & \cite{harvardDataset} \\
\hline
\end{tabular}
\label{table:dataset}
\end{table}
Supervised machine learning algorithms such as Botometer depend on the availability of training data---accounts labeled as either human or bot. These labels usually come from human annotation \cite{varol2017online}, automated methods (e.g., honey pots \cite{lee2011seven}), or botnets that display suspicious behaviors \cite{echeverria2017discovery2,echeverria2017discovery1}. A critical issue with existing datasets is the lack of ground truth. There is no objective, agreed-upon, operational definition of social bot. A further complicating factor is the prevalence of accounts that lie in the gray area between human and bot behavior, where even experienced researchers cannot easily discriminate. Nevertheless, datasets do include many typical bots; using the training labels as proxies for ground truth makes it possible to build practically viable tools.
Botometer-V4 is trained on a variety of datasets shown in Table~\ref{table:dataset}, which are publicly available in a Bot Repository.\footnote{\url{botometer.osome.iu.edu/bot-repository}}
With all training accounts being represented as feature vectors, a classifier can learn the characteristics of bot and human accounts.
Botometer uses a classification model called Random Forest, which consists of many rules learned from the training data.
To evaluate a Twitter account, Botometer first fetches its 200 most recent tweets and tweets mentioning it from Twitter, extracts its features from the collected data, and represents this information as a feature vector.
Each model rule uses some of the features and provides a vote on whether an account is more similar to bot or human accounts in the training data. Based on how many rules vote for the bot or human class, the model provides a ``bot score'' between zero and one:
a score close to one means the account is highly automated, while a score near zero means a human is likely handling the account.
Some accounts may demonstrate the characteristics of both humans and bots.
For instance, a bot creator might generate content like a regular user but uses a script to control many accounts.
These cases can be confusing for the classifier, which would then produce scores around 0.5.
While human accounts tend to behave similarly, different types of bots usually have unique behavioral patterns. Based on this observation, Botometer-V4 uses several specialized Random Forest classifiers: one for each type of bots in the training data and one for humans.
The results of this Ensemble of Specialized Classifiers (ESC) are aggregated to produce a final result.
More details about the ESC architecture can be found in the original paper~\cite{sayyadiharikandeh2020detection}.
At the end of the day, the ESC architecture is still a machine learning classifier, which yields scores between 0 and 1.
Different from a single Random Forest, the scores generated by ESC tend to have a bimodal distribution.
It is worth mentioning that the content and language features and sentiment features are based on English.
When a non-English account is passed to Botometer, these features become meaningless and might affect the classification.
As a workaround, Botometer also returns a language-independent score, which is generated without any language-related features.
Users need to be aware of the account language and choose the most appropriate Botometer score.
\subsection{Model accuracy}
The accuracy of the model is evaluated through 5-fold cross-validation on the annotated datasets shown in Table~\ref{table:dataset}.
Simply speaking, the classifier is trained on part of the annotated datasets and tested on the rest to provide a sense of its accuracy.
In the experimental environment, Botometer works really well. V4 has an AUC (area under the receiver operating characteristic curve) of 0.99, suggesting that the model can distinguish bot and human accounts in Table~\ref{table:dataset}---as well as accounts in the wild that resemble those in the training datasets---with very high accuracy.
However, Botometer is not perfect and may misclassify accounts due to several factors.
For example, the training datasets might have conflicts because they were created by different people with different standards.
In some cases Botometer fails to capture the features that can help distinguishing different accounts.
Botometer sometimes struggles with inactive accounts since not enough data is available for evaluation.
The accuracy of the model may further decay when dealing with new accounts different from those in the training datasets.
These accounts might come from a different context, use different languages other than English~\cite{rauchfleisch_false_2020,martini_bot_2021}, or show novel behavioral patterns~\cite{cresci2017paradigm,yang2019arming,Dimitriadis_multiclass_2021}.
These limitations are inevitable for all supervised machine learning algorithms, and are the reasons why Botometer has to be upgraded routinely.
Some critics exploit these limitations to undermine the entire
field of study devoted to social bots.
For example, one might select small sets of accounts with large false-positive error rates to argue that no bot detection tool is valid or that social bots do not exist at all. These arguments use fallacies such as cherry-picking and strawman in disingenuous ways. Validation through manual annotations is extremely valuable, especially when highlighting cases where existing machine learning models perform poorly, but should be used in constructive ways. New manually-annotated datasets should be made available, ideally via the public Bot Repository, to support the development of improved models.
\subsection{Results interpretation}
Early versions of Botometer returned to users raw scores in the unit interval, produced by the Random Forest classifiers.
Although users often treated them as probabilities, such interpretation is inaccurate.
Consider Twitter accounts $a$ and $b$ and their respective scores 0.7 and 0.3 produced by a Random Forest classifier.
We can say that $a$ is more bot-like than $b$, but it is inaccurate to say that there is a 70\% chance that $a$ is a bot or that $a$ is 70\% bot.
Since Botometer-V3, the scores displayed in the web interface are rescaled to the range 0--5 to discourage inaccurate probabilistic interpretations.
For users who need a probabilistic interpretation of a bot score, the Complete Automation Probability (CAP) represents the probability that an account with a given score or greater is automated. CAP scores have also been available since Botometer-V3.
The CAP scores are Bayesian posteriors that reflect both the results from the classifier and prior knowledge of the prevalence of bots on Twitter, so as to balance false positives with false negatives.
For example, suppose an account has a raw bot score of 0.96/1 (equivalent to 4.8/5 display score on the website) and a CAP score of 90\%.
This means that 90\% of accounts with a raw bot score above 0.96 are labeled as bots, or, as indicated on the website, 10\% of accounts with a bot score above 4.8/5 are labeled as humans.
In other words, if you use a threshold of 0.96 on the raw bot score (or 4.8 on the display score) to classify accounts as human/bot, you would wrongly classify 10\% of accounts as bots---a false positive rate of 10\%. This helps researchers determine an appropriate threshold based on acceptable false positive and false negative error rates for a given analysis.
\subsection{Fast bot classification}
When Botometer-V4 was released, a new model called BotometerLite was added to the Botometer family~\cite{yang_scalable_2020}.
BotometerLite was created to enable fast bot detection for large scale datasets.
The speed of bot detection methods is bounded by the platform's rate limits.
For example, the Twitter API endpoint used by Botometer-V4 to fetch an account's most recent 200 tweets and recent mentions from other users has a limit of 43,200 accounts per app key, per day.
Many studies using Twitter data have millions of accounts to analyze; with Botometer-V4, this may take weeks or even months.
To achieve scalability, BotometerLite relies only on features extracted from user metadata, contained in the so-called user object from the Twitter API.
The rate limit for fetching user objects is over 200 times the rate limit that bounds Botometer-V4.
Moreover, each tweet collected from Twitter has an embedded user object.
This brings two extra advantages.
First, once tweets are collected, no extra queries to Twitter are needed for bot detection.
Second, the user object embedded in each tweet reflects the user profile at the moment when the tweet is collected.
This makes bot detection on archived historical data possible.
In addition to the improved scalability, BotometerLite employs a novel data selection mechanism to ensure its accuracy and generalizability.
Instead of throwing all training data into the classifier, a subset is selected by optimizing three evaluation metrics: cross-validation accuracy on the training data, generalization to holdout datasets, and consistency with Botometer.
This mechanism was inspired by the observation that some datasets are contradictory to each other.
After evaluating the classifiers trained on all possible combinations of candidate training sets, the winning classifier only uses five out of eight datasets but performs well in terms of all evaluation metrics.
BotometerLite allows researchers to analyze large-volume streams of accounts in real time, while the limited training data may involve a compromise in accuracy on certain bot classes compared to Botometer-V4.
In terms of how to choose between the two endpoints, we still recommend using Botometer-V4 when feasible since it analyzes more data and produces more detailed results.
\section{Botometer interface}
Although the machine learning model might seem complicated, the interface of Botometer is designed to be easy to use.
Botometer has a website and API endpoints with similar functionality.
The website\footnote{\url{botometer.org}} is handy for users who need to quickly check several accounts.
With a Twitter account, users can access the Botometer website from any web browsers, even on their mobile devices.
The website is straightforward to use: after authorizing Botometer to fetch Twitter data, users just need to type a Twitter handle of interest and click the ``Check user'' button.
The Botometer Pro API\footnote{\url{rapidapi.com/OSoMe/api/botometer-pro}} can be more useful for research since it allows to programmatically check accounts in bulk.
The API is hosted by RapidAPI, a platform that helps developers manage API rate limits and user subscriptions.
Using the Botometer API requires keys associated with a Twitter app, which can be obtained through Twitter's developer portal.\footnote{\url{developer.twitter.com}} One also needs a RapidAPI account and a subscription to one of the API usage plans.
When querying the API, users are responsible to send the required data (i.e., 200 most recent tweets by the account being checked and tweets mentioning this account) in a specified format through HTTPS requests.
The Botometer API will process the data and return the results.
While queries can be sent through any programming language, we recommend using Python and the official \code{botometer-python} package that we maintain.\footnote{\url{github.com/IUNetSci/botometer-python}}
The package can fetch data from Twitter, format the data, and query the API on behalf of the user with a few lines of code:
\lstset{
showspaces=false,
basicstyle=\ttfamily\footnotesize
}
\begin{lstlisting}
import botometer
bom = botometer.Botometer(
rapidapi_key="XYZ",
consumer_key="XYZ",
consumer_secret="XYZ",
access_token="XYZ",
access_token_secret="XYZ"
)
result = bom.check_account("@yang3kc")
print(f"Bot score={result['display_scores']['english']['overall']}/5")
print(f"CAP score={result['cap']['english']:.2f}")
\end{lstlisting}
BotometerLite is also available as an endpoint through the Botometer Pro APIs.
We list the the input, output, and limitations of the API endpoints for Botometer-V4 and BotometerLite side by side in Table~\ref{table:api_specs}.
We also summarize the common resources for using Botometer in Table~\ref{table:links} to help the readers navigate these resources.
\begin{table}
\centering
\caption{
Comparison of Botometer-V4 and BotometerLite APIs.
}
\begin{threeparttable}
\centering
\begin{tabular}{p{3cm}||p{3.0cm}|p{4.5cm}}
\hline
Model & Botometer-V4 & BotometerLite \\
\hline
Endpoint & Check account & Check account in bulk \\
\hline
Query payload & User object, 200 most recent tweets, mentions & List of user objects and timestamps \\
\hline
Response & Raw bot scores, sub-scores, CAP scores, basic account information, etc. & BotometerLite scores \\
\hline
Daily number of accounts allowed\tnote{*} & 43,200 & $\sim 8.6$ million \\
\hline
Corresponding \code{botometer-python} method(s) & \code{check\_account} & \code{check\_accounts\_from\_tweets}, \code{check\_accounts\_from\_user\_ids}, \code{check\_accounts\_from\_screen\_names} \\
\hline
\end{tabular}
\begin{tablenotes}
\item[*] The values represent the upper bounds based on Twitter's API rate limit when using a single app key. The actually numbers depend on other factors such as internet speed as well.
\end{tablenotes}
\end{threeparttable}
\label{table:api_specs}
\end{table}
\begin{table}
\caption{
Common resources for using Botometer.
}
\centering
\begin{tabular}{p{3cm}|p{3.5cm}|p{4cm}}
\hline
Resource name & Resource & Note \\
\hline
Botometer website & \url{botometer.org} & Web interface of Botometer: useful for checking a small amount of accounts \\
\hline
Botometer Pro API & \url{rapidapi.com/OSoMe/api/botometer-pro} & API of Botometer: useful for checking accounts in bulk programmatically \\
\hline
\code{Botometer-python} package & \url{github.com/IUNetSci/botometer-python} & Python package to access Botometer Pro API \\
\hline
Botometer case study & \url{github.com/osome-iu/Botometer101} & Case study using Botometer with source code\\
\hline
Bot repository & \url{botometer.osome.iu.edu/bot-repository} & Annotated training datasets for Botometer \\
\hline
\end{tabular}
\label{table:links}
\end{table}
Note that both Botometer and Twitter APIs have rate limits, meaning that users can only make a certain number of queries in a given time period.
Please check the respective websites for detailed documentation.
Getting familiar with the rate limits can help researchers better estimate the time needed for their analysis.
\section{Case study}
Since some readers may not be familiar with programming, querying the API could be challenging.
Moreover, analyzing the results returned by Botometer API is not trivial.
In this section, we provide a simple case study as a demonstration.
Different ways of analyzing the data are shown with recommended practice.
We share the code for this case study in a public repository\footnote{\url{github.com/osome-iu/Botometer101}} so that readers can use it as a template for their own research.
Next we outline the data collection and analysis steps implemented in this software repository.
\subsection{Data collection}
Let us consider two cryptocurrency cashtags, \cashtag{FLOKI} and \cashtag{SHIB}, and the cashtag of Apple Inc., \cashtag{AAPL}, and attempt to quantify which is more amplified by bot-like accounts.
A cashtag works like a hashtag but consists of a dollar sign ``\$'' and a stock or cryptocurrency symbol to help users track related discussions.
We use \code{Tweepy},\footnote{\url{tweepy.org}} a Python package that helps access the Twitter API, to search tweets containing these cashtags.
For each cashtag, we only collect 2,000 tweets, which are sufficient for the demonstration.
\begin{table}
\caption{
Numbers of tweets and unique accounts mentioning different cashtags in raw data and analytical sample.
}
\centering
\begin{tabular}{l|r|r|r|r}
\hline
& \multicolumn{2}{c|}{Raw data} & \multicolumn{2}{c}{Analytical sample} \\
\hline
Cashtag & Tweets & Unique accounts & Tweets & Unique accounts \\
\hline
\cashtag{SHIB} & 2,000 & 1,241 & 1,819 & 1,111 \\
\cashtag{FLOKI} & 2,000 & 937 & 1,893 & 860 \\
\cashtag{AAPL} & 2,000 & 1,107 & 1,864 & 1,006 \\
\hline
\end{tabular}
\label{table:number_of_tweets_accounts}
\end{table}
First, let us count the number of unique accounts in each dataset, as shown in Table~\ref{table:number_of_tweets_accounts}.
The number of unique accounts is much smaller than the number of tweets in all three datasets, suggesting that some accounts tweeted the same cashtag multiple times.
The next step is to query the Botometer API for bot analysis.
Instead of going through each tweet and check every user encountered, researchers can keep a record of accounts already queried to avoid repetition and increase efficiency.
The Botometer API returns rich information about each account.
We recommend storing the full results from Botometer for flexibility.
\begin{figure}
\centering
\includegraphics[width=0.5\linewidth]{language_freqency.pdf}
\caption{
Percentage of accounts using each language in the three datasets combined.
}
\label{fig:lang_frequency}
\end{figure}
As mentioned above, Botometer generates an overall score and a language-independent score.
Since the two scores come from different classifiers, they are not comparable and should not be mixed together. To decide which one to use, let us calculate the proportion of accounts using each language. We can see in Figure~\ref{fig:lang_frequency} that the majority of accounts in our raw data tweet in English.
Therefore we only include English-speaking accounts and their tweets in our analytical sample (see Table~\ref{table:number_of_tweets_accounts} for summary statistics) and use the overall bot score.
\subsection{Analysis}
We plot the bot score distribution for tweets mentioning each cashtag in Figure~\ref{fig:bot_score_distribution}(a).
Here we base our analysis on the raw scores in the unit interval.
Since we are interested in the bot activity level of each cashtag, we use tweets (as opposed to accounts) as the units of analysis.
This means that accounts tweeting the same cashtag multiple times have a larger contribution.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{bot_score_dist.pdf}
\caption{
(a) Bot score distributions for tweets mentioning different cashtags.
(b) Percentage of tweets posted by likely bots using 0.5 as a threshold.
(c) Box plots of the bot scores for tweets mentioning different cashtags.
The white lines indicate the median values; the white dots indicate the mean values.
(d) Similar to (b) but using a bot score threshold of 0.7.
Statistical tests are performed for pairs of results in (b--d).
Significance level is represented by the stars: ***$p \le 0.001$, **$p \le 0.01$, *$p \le 0.05$, NS$=p>0.05$.
}
\label{fig:bot_score_distribution}
\end{figure}
In all three cases, the distribution has a bimodal pattern, a result of the ESC architecture of Botometer-V4.
We can observe some spikes in all cases, which are caused by accounts tweeting the same cashtag repeatedly.
For example, the spike near 0.89 for \cashtag{SHIB} and \cashtag{FLOKI} comes from a bot-like account that replied the same message promoting cryptocurrency tokens to a large number of tweets containing the keyword ``NFT''; see the screenshot of the message in Figure~\ref{fig:bot_screen_shot}.
\begin{figure}
\centering
\includegraphics[width=0.45\linewidth]{bot_screen_shot.png}
\caption{
Screenshot of a bot-like account replying to a tweet containing the keyword ``NFT'' with a message promoting cryptocurrencies.
The same message was replied by this account to a large number of tweets.
}
\label{fig:bot_screen_shot}
\end{figure}
To address our research question, we need to quantify the bot activity level for each cashtag and compare them.
The first approach is to compare their bot score distributions with two-sided Mann–Whitney U tests (see results in Figure~\ref{fig:bot_score_distribution}(c)).
The bot score distributions of \cashtag{SHIB} and \cashtag{FLOKI} are not significantly different from each other ($p=0.56$), but both of them have a higher bot activity level than \cashtag{AAPL} (\cashtag{SHIB} vs. \cashtag{AAPL}: $p<0.001$; \cashtag{FLOKI} vs. \cashtag{AAPL}: $p<0.001$).
The second approach dichotomizes the bot scores and considers the accounts with scores higher than a threshold as likely bots.
Then the proportion of tweets from likely bots can be calculated and compared.
In this approach, a threshold has to be chosen.
In the literature, 0.5 is the most common choice~\cite{shao_spread_2018,vosoughi_spread_2018,bessi_social_2016}; higher values, such as 0.7~\cite{grinberg_fake_2019} and 0.8~\cite{broniatowski_weaponized_2018}, are also used.
One may also consider running the same analysis with different threshold values to test the robustness of the findings~\cite{shao_spread_2018}.
Here we use both 0.5 and 0.7 as thresholds and show the results in Figure~\ref{fig:bot_score_distribution}(b) and (d), respectively.
We apply two-proportions $z$-tests to estimate the significance level of the differences.
When using 0.5 as the threshold, the percentage of tweets from likely bots that mentioned \cashtag{SHIB} is significantly higher than those in the \cashtag{FLOKI} ($p=0.009$) and \cashtag{AAPL} datasets ($p<0.001$).
The percentage of tweets from likely bots that mentioned \cashtag{FLOKI} is also significantly higher than that in the \cashtag{AAPL} dataset ($p<0.001$).
However, when using 0.7 as the threshold, the results change: percentages of tweets from likely bots in \cashtag{SHIB} and \cashtag{FLOKI} datasets are no longer significantly different from each other ($p=0.38$); both of them are lower than that in the \cashtag{AAPL} dataset (\cashtag{SHIB} vs. \cashtag{AAPL}: $p<0.001$; \cashtag{FLOKI} vs. \cashtag{AAPL}: $p<0.001$).
In other studies, different approaches or threshold choices may yield consistent results.
However, they lead to seemingly different conclusions in this case.
This is because different measures represent different properties of the bot score distribution.
If we revisit Figure~\ref{fig:bot_score_distribution}(a), we can see that although the distributions of \cashtag{SHIB} and \cashtag{FLOKI} scores have more mass in the $(0.5, 1]$ region than that of \cashtag{AAPL} scores, the mass tends to concentrate around 0.6, while the distribution of \cashtag{AAPL} scores has more mass near 1.
This nuanced difference causes the contradictory results when using different threshold values.
By reconciling the results from different approaches, we can answer our research question now.
It appears that discussions about the cryptocurrencies \cashtag{SHIB} and \cashtag{FLOKI} show more automated activities than that about \cashtag{AAPL}, but among the accounts tweeting \cashtag{AAPL}, we find more highly automated bot-like accounts.
Note that the analysis here is mainly for demonstrating the use of Botometer; the samples of tweets analyzed are small and not representative of the entire discussion, so the conclusions only reflect the status of the collected data and should not be generalized.
\section{Recommended practice}
The sections above cover some recommended practice such as being careful when interpreting raw bot scores, being mindful about user language, and being aware of different versions of Botometer.
Here we make a few more recommendations to help avoid common pitfalls.
\subsection{Transient nature of Botometer scores}
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{bot_score_change.pdf}
\caption{
Time series of bot scores of an account from September 2020 to November 2021.
The queries were not made regularly, so the time intervals between consecutive data points vary.
}
\label{fig:bot_score_change}
\end{figure}
Recall that Botometer uses the 200 most recent tweets by an account and other tweets mentioning the account for analysis.
This means that the results of Botometer change over time, especially for very active accounts.
To demonstrate this, we plot the time series of the overall bot score of an account in Figure~\ref{fig:bot_score_change}.
This account posts roughly 16 tweets each week and gets mentioned by others frequently.
We can see that the bot score fluctuates over time.
In some other cases, an account might be suspended or removed after a while, making it impossible to analyze.
Due to the transient nature of Botometer scores, a single bot score only reflects the status of the account at the moment when it is evaluated.
Users should be careful when drawing conclusions based on the bot scores of individual accounts.
For researchers, a common practice is to collect tweets first, then perform bot detection later.
To reduce the effect of unavailable accounts and to keep the bot scores relevant, bot analysis should be conducted right after data collection.
\subsection{Evaluating bot score distributions}
Whenever possible, we recommend collecting large datasets and use statistical analyses to evaluate bot activity based on comparisons of score distributions across different groups of accounts.
As demonstrated in the case study, bot score distributions can reveal rich information about the data.
Using distributions for analysis also reduces the uncertainty level of Botometer due to its imperfection and transient nature.
Most importantly, comparing distributions of scores---e.g., for accounts tweeting about a given topic versus a suitable baseline---allows for statistical tests that are impossible at the level of individual accounts.
\subsection{Validating thresholds}
In some analyses, dichotomizing the bot scores based on a threshold is necessary.
In these cases, we recommend validating the choice of threshold.
For researchers with the ability and resources, the ideal approach is to manually annotate a batch of bot and human accounts in their datasets.
Such a preliminary analysis could be used, first, to determine whether Botometer is a helpful tool to evaluate a given scenario. Assuming it is, one can then vary the threshold and select the value that optimizes some appropriate metric on the annotated accounts.
Depending on the desire to maximize accuracy, minimize false positive errors, minimize false negative errors, or some combination, one can use metrics such as accuracy, precision, recall, or F1.
When annotating additional accounts is not feasible, we suggest running multiple analyses using different threshold choices to confirm the robustness of the findings.
\subsection{Using Botometer in a civil way}
We have noticed that Botometer has been used to attack others.
For example, some users may call others with whom they disagree ``bots'' and use the results of Botometer as justification.
This is a misuse of Botometer. Users should keep in mind that any classifier such as Botometer can mislabel individual accounts. Furthermore, even if an account is automated, it does not mean it is deceptive or malicious.
Most importantly, such name calling is not helpful for creating healthy and informative conversations.
\bibliographystyle{unsrt}
|
\section{Introduction}
A commutative Hopf algebroid is somehow the dual of a groupoid, in the spirit of Hopf algebras versus groups.
One is extending the scalar, similarly to the passage from Hilbert space to Hilbert module:
the ground field $k$ gets replaced by an algebra $B$ which could be noncommutative.
The result is a bi-algebra over a noncommutative base algebra. In fact, in general not all structures survive:
there is a notion of coproduct and counit but in general there is no antipode.
The notions of source and target maps are still present.
An important groupoid used in gauge theory,
is the gauge groupoid associated with a principal bundle \cite{Mac05}.
In \cite{HL21}, as a preliminary step to study the gauge group of a noncommutative principal bundle, we considered the Ehresmann--Schauenburg bialgebroid of the noncommutative bundle which, in a sense, is the quantization of the classical gauge groupoid. For a monopole bundle over a quantum Podle\'s sphere
and a not faithfully flat Hopf--Galois extension of commutative algebras we gave a suitable invertible antipode
so that the corresponding bialgebroids got upgraded to Hopf algebroids.
In the present paper we study two classes of examples of Hopf algebroids associated with noncommutative principal bundles.
The first comes from deforming the principal bundle while leaving unchanged the structure Hopf algebra.
The prototype for this is the bundle over the noncommutative four-sphere $S_\theta^{4}$ with classical $\SU(2)$
as structure group.
The second class is associated to deformations of quantum homogeneous spaces.
It is known that one needs a careful deformation of the multiplication in a Hopf algebra in order to preserve the compatibilities between the Hopf algebra structures. And this attention is needed also for deforming homogeneous spaces. Examples of the second class are the principal bundles over the noncommutative spheres $S_\theta^{2n}$ with noncommutative orthogonal group $\SO_\theta(2n, \mathbb{R})$ as structure group.
This paper is organised as follows.
In \S\ref{se;pr} we give a recap of algebraic preliminaries and notation, and of the relevant concepts for noncommutative principal bundles (Hopf--Galois extensions), bialgebroids and Hopf algebroids.
We devote \S\ref{se:hge} to two well know examples of Hopf--Galois extensions
for which in \S\S ~\ref{sec:ev-q-shp} and \ref{se:alg-su2sym} we construct the corresponding Hopf algebroids; these are a $\SU(2)$-bundle over the sphere $S^4_\theta$ and $\SO_\theta(2n)$ bundles over even spheres $S^{2n}_\theta$.
In \S\ref{alg-def} we review the general scheme of deforming by the action of tori.
This is done via $\mathbb{Z}^n$-graded spaces and deforming relevant structures by means of a bi-character.
The discussion is developed along two scenarios to cover the constructions of both
\S\ref{subsec:ScI}, where the structure Hopf algebra is not changed, and \S\ref{subsec:ScII}
where attention is payed to a suitable deformation of the multiplication that is compatible with all Hopf algebra
operations, in order to get new Hopf algebras with related comodule
algebras. The latter framework accommodates deformed homogeneous spaces. The noncommutative principal bundles that result from both schemes of deformation have natural Ehresmann--Schauenburg bialgebroids. In the context of the present paper the flip map will preserve the bialgebrois and will satisfy all properties for an invertible algebrois antipode. All of these last parts and the examples are worked out in \S\ref{se:had}.
\bigskip
\noindent
\subsubsection*{Acknowledgements}
We are grateful to Chiara Pagani for useful discussions.
XH was partially supported by Marie Curie Fellowship HADG - 101027463 agreed between QMUL and the European Commission, and by an Assistant Professor fellowship at IMPAN.
GL acknowledges partial support from INFN, Iniziativa Specifica GAST,
from INdAM-GNSAGA and from the INDAM-CNRS IRL-LYSM.
YL would like to thank SISSA for the postdoctoral fellowship and excellent working environment.
\section{Preliminary results} \label{se;pr}
To be definite we work over the field $\mathbb{C}$ of complex numbers. Algebras (coalgebras) are assumed to be unital and associative (counital and coassociative) with morphisms of algebras taken to be unital (of coalgebras taken counital).
Tensor product over $\mathbb{C}$ is denoted ${\otimes}$ while the symbol $\stackrel{.}{\otimes}$ implies also a matrix sum: for matrices $M=(m_{jk})$
and $N=(n_{kl})$ the product $M \stackrel{.}{\otimes} N$ have components $M \stackrel{.}{\otimes} N = (\sum_k m_{jk} {\otimes} n_{kl})$.
\subsection{Rings and corings over an algebra}
For an algebra $B$ a {\em $B$-ring} is a triple $(A,\mu,\eta)$. Here $A$ is a $B$-bimodule with $B$-bimodule maps
$\mu:A{\otimes}_ {B} A \to A$ and $\eta:B\to A$, satisfying associativity and unit conditions:
\begin{equation
\mu\circ(\mu{\otimes} _{B} \id_A)=\mu\circ (\id_A {\otimes} _{B} \mu),
\quad
\mu\circ(\eta {\otimes} _{B} \id_A)=\id_A=\mu\circ (\id_A{\otimes} _{B} \eta).
\end{equation}
A morphism of $B$-rings $f:(A,\mu,\eta)\to (A',\mu',\eta')$ is an
$B$-bimodule map $f:A \to A'$, such that
$f\circ \mu=\mu'\circ(f{\otimes}_{B} f)$ and $f\circ \eta=\eta'$.
From \cite[Lemma 2.2]{Boehm} there is a bijective correspondence between $B$-rings $(A,\mu,\eta)$ and algebra automorphisms
$\eta : B \to A$. Starting with a $B$-ring $(A,\mu,\eta)$, one obtains a multiplication map $A {\otimes} A \to A$ by composing the canonical surjection $A {\otimes} A \to A{\otimes}_B A$ with the map $\mu$. Conversely, starting with an algebra map $\eta : B \to A$, a $B$-bilinear associative multiplication $\mu:A{\otimes}_ {B} A \to A$ is obtained from the universality of the coequaliser $A {\otimes} A \to A{\otimes}_B A$ which identifies an element $ a r {\otimes} a'$ with $ a {\otimes} r a'$.
Dually, for an algebra $B$ a {\em $B$-coring} is a
triple $(C,\Delta,\varepsilon)$. Here $C$ is a $B$-bimodule with $B$-bimodule maps
$\Delta:C\to C{\otimes}_{B} C$ and $\varepsilon: C\to B$ that satisfy coassociativity and counit conditions,
\begin{align
(\Delta{\otimes} _{B} \id_C)\circ \Delta = (\id_C {\otimes} _{B} \Delta)\circ \Delta, \quad
(\varepsilon {\otimes} _{B} \id_C)\circ \Delta = \id_C =(\id_C {\otimes} _{B} \varepsilon)\circ \Delta.
\end{align}
A morphism of $B$-corings $f:(C,\Delta,\varepsilon)\to
(C',\Delta',\varepsilon')$ is a $B$-bimodule map $f:C \to C'$, such that
$\Delta'\circ f=(f{\otimes}_{B} f)\circ \Delta$ and
$\varepsilon' \circ f =\varepsilon$.
Let $B$ be an algebra.
A {\em left $B$-bialgebroid} $\mathcal{L}$ consists of a $(B{\otimes} B^{op})$-ring
together with a $B$-coring structures on the same vector space $\mathcal{L}$, with mutual compatibility conditions
\cite{Take77}.
From what said above, a $(B{\otimes} B^{op})$-ring $\mathcal{L}$ is the same as an algebra map $\eta : B {\otimes} B^{op} \to \mathcal{L}$.
Equivalently, one may consider the restrictions
$$
s := \eta ( \, \cdot \, {\otimes}_B 1_B ) : B \to \mathcal{L} \quad \mbox{and} \quad t := \eta ( 1_B {\otimes}_B \, \cdot \, ) : B^{op} \to \mathcal{L}
$$
which are algebra maps with commuting ranges in $\mathcal{L}$, called the \emph{source} and the \emph{target} map of the
$(B{\otimes} B^{op})$-ring $\mathcal{L}$. Thus a $(B{\otimes} B^{op})$-ring is the same as a triple $(\mathcal{L},s,t)$ with $\mathcal{L}$ an algebra and $s: B \to \mathcal{L}$
and $t: B^{op} \to \mathcal{L}$ both algebra maps with commuting range.
For a left $B$-bialgebroid $\mathcal{L}$ the compatibility conditions are required to be the following.
\begin{itemize}
\item[i)] The bimodule structures in the $B$-coring $(\mathcal{L},\Delta,\varepsilon)$ are
related to those of the $B{\otimes} B^{op}$-ring $(\mathcal{L},s,t)$ via
\begin{equation}\label{eq:rbgd.bimod}
b\triangleright a \triangleleft \tilde{b}:= s(b) t(\tilde{b})a , \quad \textrm{for} \,\, b, \tilde{b}\in B, \, a\in \mathcal{L}.
\end{equation}
\item[ii)] Considering $\mathcal{L}$ as a $B$-bimodule as in \eqref{eq:rbgd.bimod},
the coproduct $\Delta$ corestricts to an algebra map from $\mathcal{L}$ to
\begin{equation}\label{eq:Tak.prod}
\mathcal{L} \times_{B} \mathcal{L} := \left\{\ \sum\nolimits_j a_j{\otimes}_{B} \tilde{a}_j\ |\ \sum\nolimits_j a_jt(b) {\otimes}_{B} \tilde{a}_j =
\sum\nolimits_j a_j {\otimes}_{B} \tilde{a}_j s(b), \,\,\, \forall \, b \in B\ \right\},
\end{equation}
where $\mathcal{L} \times_{B} \mathcal{L}$ is an algebra via component-wise multiplication.
\\
\item[iii)]
The counit $\varepsilon : \mathcal{L} \to B$
satisfies the properties,
\begin{itemize}
\item[1)] $\varepsilon(1_{\mathcal{L}})=1_{B}$,
\item[2)] $\varepsilon(s(b)a)=b\varepsilon(a) $,
\item[3)] $\varepsilon(as(\varepsilon(\tilde{a})))=\varepsilon(a\tilde{a})=\varepsilon(at (\varepsilon(\tilde{a})))$, \qquad
for all $b\in B$ and $a,\tilde{a} \in \mathcal{L}$.
\end{itemize}
\end{itemize}
An automorphism of the left bialgebroid $(\mathcal{L}, \Delta, \varepsilon,s,t)$ over the algebra $B$ is a pair $(\Phi, \varphi)$
of algebra automorphisms, $\Phi: \mathcal{L} \to \mathcal{L}$, $\varphi : B \to B$ such that:
\begin{align}
\Phi\circ s & = s\circ \varphi , \qquad \Phi\circ t = t\circ \varphi , \label{amoeba(i)} \\
(\Phi{\otimes}_{B} \Phi)\circ \Delta & = \Delta \circ \Phi , \qquad
\varepsilon\circ \Phi= \varphi\circ \varepsilon \label{amoeba(ii)} .
\end{align}
In fact, the map $\varphi$ is uniquely determined by $\Phi$ via $\varphi = \varepsilon \circ \Phi \circ s$
and one can just say that $\Phi$ is a bialgebroid automorphism. Automorphisms of a bialgebroid $\mathcal{L}$
form a group $\Aut(\mathcal{L})$ by map composition. A \textit{vertical} automorphism is one of the type
$(\Phi, \varphi=\id_{B})$.
From the conditions \eqref{amoeba(i)}, $\Phi$ is a $B$-bimodule map:
$\Phi(b \triangleright c\triangleleft \tilde{b})=b \triangleright_{\varphi}\Phi(c)\triangleleft_\varphi \tilde{b}$.
The first condition \eqref{amoeba(ii)} is well defined once the conditions \eqref{amoeba(i)} are satisfied (the balanced tensor product is induced by $s':=s\circ \varphi $ and $t':=t\circ \varphi $). Conditions \eqref{amoeba(i)} imply $\Phi$ is a coring map, therefore $(\Phi, \varphi)$ is an isomorphism between the starting and the new bialgebroid.
Finally, we recall from \cite[Def. 4.1]{BS04} the conditions for a Hopf algebroid with invertible antipode.
Given a left bialgebroid $(\mathcal{L}, \Delta, \varepsilon,s,t)$ over the algebra $B$, an invertible antipode
$S : \mathcal{L} \to \mathcal{L} $ in an algebra anti-homomorphism with inverse $S^{-1} : \mathcal{L} \to \mathcal{L} $ such that
\begin{equation}\label{hopbroid1}
S \circ t = s
\end{equation}
and satisfying compatibility conditions with the coproduct:
\begin{align}\label{hopbroid2}
\onet {(S^{-1}\two{h})} {\otimes}_B \twot {(S^{-1}\two{h})}\one{h} & = S^{-1} h {\otimes}_B 1_\mathcal{L} \nonumber \\
\onet {(S\one{h})} \two{h} {\otimes}_B \twot {S(\one{h})} & = 1_\mathcal{L} {\otimes}_B S h ,
\end{align}
for any $h\in \mathcal{L} $. These then imply
$
S(\one{h}) \, \two{h} = t \circ \varepsilon \circ S h.
$
\subsection{Hopf--Galois extensions}\label{sec:hge}
We give a brief recall of Hopf--Galois extensions
as noncommutative principal bundles. These extensions are $H$-comodule algebras
$A$ with a canonically defined map $\chi: A{\otimes}_B A\to A{\otimes} H$
which is required to be invertible \cite{HJSc90}.
\begin{defi} \label{def:hg}
Let $H$ be a Hopf algebra and let $A$ be a $H$-comodule algebra with coaction $\delta^A$.
Consider the subalgebra $B:= A^{coH}=\big\{b\in A ~|~ \delta^A (b) = b {\otimes} 1_H \big\} \subseteq
A$ of coinvariant elements
with balanced tensor product $A {\otimes}_B A$.
The extension $B\subseteq A$ is called a $H$-\textup{Hopf--Galois extension} if the
\textit{canonical Galois map}
\begin{align*}
\chi := (m {\otimes} \id) \circ (\id {\otimes} _B \delta^A ) :
A {\otimes} _B A \longrightarrow A {\otimes} H ,
\quad \tilde{a} {\otimes}_B a &\mapsto \tilde{a} \zero{a} {\otimes} \one{a}
\end{align*}
is an isomorphism.
\end{defi}
\begin{rem}\label{fun-rem}
For a Hopf--Galois extension $B\subseteq A$, we take the algebra $A$ to be \emph{faithfully flat} as a right $B$-module.
One possible way to state this property is that for any left $B$-module map $F : M \to N$,
the map $F$ is injective if and only if the map $\id_A {\otimes}_B F : A {\otimes}_B M \to A {\otimes}_B N$
is injective; injectivity of $F$ implying the injectivity of $\id_A {\otimes}_B F$ would state that $A$ is flat as a right $B$-module (see \cite[Chap.~13]{Wa79}.
\qed\end{rem}
Since the canonical Galois map $\chi$ is left $A$-linear, its inverse is
determined by the restriction $\tau:=\chi^{-1}_{|_{1_A {\otimes} H}}$, named \textit{translation map},
\begin{align}
\label{eq:tau-defn}
\tau =\chi^{-1}_{|_{1_A {\otimes} H}} : H\to A{\otimes} _B A ~ ,
\quad h \mapsto \tau(h) = \tuno{h} {\otimes}_B \tdue{h} \, .
\end{align}
Thus by definition:
\begin{equation}
\label{p7}
\tuno{h}\zero{\tdue{h}}{\otimes} \one{\tdue{h}} = 1_{A} {\otimes} h \, .
\end{equation}
The translation map enjoys a number of properties \cite[3.4]{HJSc90b}
that we listed here for later use.
For any $h, k \in H$ and $a\in A$, $b\in B$:
\begin{align}
\tuno{h} {\otimes}_B \zero{\tdue{h}} {\otimes} \one{\tdue{h}} &= \tuno{\one{h}} {\otimes}_B \tdue{\one{h}} {\otimes} \two{h} \label{p4} \, , \\
\zero{\tuno{h}} {\otimes}_B {\tdue{h}} {\otimes} \one{\tuno{h}} &= \tuno{\two{h}} {\otimes}_B \tdue{\two{h}} {\otimes} S(\one{h}) \, , \label{p1}
\end{align}
\begin{align}
\label{p5}
\tuno{h}\tdue{h} &= \varepsilon(h)1_{A} \, , \\
\label{p3}
\zero{a}\tuno{\one{a}}{\otimes}_{B}\tdue{\one{a}} &= 1_{A} {\otimes}_{B}a \, , \\
\label{p8}
b\, \tuno{h} {\otimes}_B \tdue{h} &= \tuno{h} {\otimes}_B \tdue{h} \, b \, ,
\end{align}
\begin{align}
\label{p2}
\tuno{(h k)}{\otimes}_{B}\tdue{(h k)} &= \tuno{k}\tuno{h}{\otimes}_{B}\tdue{h}\tdue{k} \, , \\
\label{p6}
\tuno{\one{h}}{\otimes}_{B}\tdue{\one{h}}\tuno{\two{h}}{\otimes}_{B}\tdue{\two{h}} &
=\tuno{h}{\otimes}_{B}1_{A}{\otimes}_{B}\tdue{h} \, .
\end{align}
\subsection{Ehresmann--Schauenburg bialgebroids}
\label{sec:ES-bialerd}
To any Hopf--Galois extension $B=A^{co \, H}\subseteq A$ one associates a $B$-coring
and a bialgebroid \cite{schau2} (see \cite[\S 34.13 and \S 34.14]{BW}).
These can be viewed as a quantization of the gauge groupoid that is associated to a (classical) principal fibre bundle (see \cite{Mac05}).
The coring can be given in a few equivalent ways.
Let $B=A^{co \, H}\subseteq A$ be a
Hopf--Galois extension with right coaction $\delta^{A} : A \to A {\otimes} H$. This extends to
a diagonal coaction,
\begin{equation}\label{AAcoact}
\delta^{A{\otimes} A}: A{\otimes} A\to A{\otimes} A{\otimes} H, \quad a{\otimes} \tilde{a} \mapsto
\zero{a}{\otimes} \zero{\tilde{a}} {\otimes} \one{a}\one{\tilde{a}} , \quad \textup{for} \quad a, \tilde{a} \in A .
\end{equation}
Let $\tau$ be the translation map of the Hopf--Galois extension. We have the following:
\begin{lem}\label{lem-2vers}
The $B$-bimodule of coinvariant elements for the diagonal coaction,
\begin{equation} \label{ec2}
(A{\otimes} A)^{coH} = \{a{\otimes} \tilde{a}\in A{\otimes} A \, ; \,\, \zero{a}{\otimes} \zero{\tilde{a}}{\otimes} \one{a}\one{\tilde{a}}=a{\otimes} \tilde{a}{\otimes} 1_H \}
\end{equation}
is the same as the $B$-bimodule
\begin{equation}\label{ec1}
\mathcal{C}(A, H) :=\{a{\otimes} \tilde{a}\in A{\otimes} A : \,\, \zero{a}{\otimes} \tau(\one{a})\tilde{a}=a{\otimes} \tilde{a}{\otimes} _B 1_A\}.
\end{equation}
\end{lem}
\begin{proof}
This is a direct check: using properties of the canonical map $\chi$ and of the translation map $\tau$, one shows the two inclusions.
\end{proof}
We have then the following definition \cite{schau2} (see \cite[\S 34.13]{BW}).
\begin{defi}\label{def:ec}
Let $B=A^{co \, H}\subseteq A$ be a
faithfully flat
Hopf--Galois extension with translation map $\tau$.
Then the $B$-bimodule $\mathcal{C}(A, H)$ in \eqref{ec1} is a $B$-coring with %
coproduct,
\begin{equation}\label{copro}
\Delta(a{\otimes} \tilde{a}) = \zero{a}{\otimes} \tau(\one{a}){\otimes} \tilde{a}
= \zero{a} {\otimes} \tuno{\one{a}} {\otimes}_B \tdue{\one{a}} {\otimes} \tilde{a} ,
\end{equation}
and counit,
\begin{equation}
\varepsilon(a{\otimes} \tilde{a}) = a\tilde{a} . \label{counit}
\end{equation}
\end{defi}
\noindent
Applying the map $m_A{\otimes} \id_H$ to elements of \eqref{ec2} one gets $a\tilde{a}\in B$.
The above $B$-coring is called the \textit{Ehresmann} or \textit{gauge coring}; we denote it $\mathcal{C}(A, H)$.
Also, using the well know relation between the coinvariants of a tensor product of comodules and their
cotensor product \cite[Lemma 3.1]{HJSc90b}, the coring $\mathcal{C}(A, H)$ can be given as a
cotensor product $A\, \square \, {}^H\!\!A$.
The Ehresmann coring of a Hopf--Galois extension is
in fact a bialgebroid \cite{schau2}, called the \textit{Ehresmann--Schauenburg bialgebroid} (see \cite[34.14]{BW}).
One see that $\mathcal{C}(A, H) = (A{\otimes} A)^{coH}$ is a subalgebra of $A {\otimes} A^{op}$; indeed, given
$x {\otimes} \tilde{x}, \, y {\otimes} \tilde{y} \in (A{\otimes} A)^{coH}$, one computes
\begin{align*}
\delta^{A{\otimes} A}(x y {\otimes} \tilde{y} \tilde{x})
&=
\zero{x} \zero{y} {\otimes} \zero{\tilde{y}} \zero{\tilde{x}}
{\otimes} \one{x}\one{y} \one{\tilde{y}} \one{\tilde{x}} \\
&=
\zero{x} y {\otimes} \tilde{y} \zero{\tilde{x}} {\otimes} \one{x} \one{\tilde{x}} \\
&=
x y {\otimes} \tilde{y} \tilde{x} {\otimes} 1_H.
\end{align*}
\begin{defi}\label{def:reb}
Let $\mathcal{C}(A, H)$ be the coring associated with a faithfully flat Hopf--Galois extension
$B=A^{co \, H}\subseteq A$. Then $\mathcal{C}(A, H)$
is a (left) $B$-bialgebroid with product
$$
(x {\otimes} \tilde{x}) \bullet_{\mathcal{C}(A, H)} ({y}{\otimes} \tilde{y}) = x y {\otimes} \tilde{y} \tilde{x} ,
$$
for all $x {\otimes} \tilde{x}, \, y {\otimes} \tilde{y} \in \mathcal{C}(A, H)$ (and unit $1_A{\otimes} 1_A$). The target and the source maps are
$$
t(b)=1_A {\otimes} b \quad \textup{and} \quad s(b)=b{\otimes} 1_A.
$$
\end{defi}
\noindent
We refer to \cite[34.14]{BW} for the checking that all defining properties are satisfied. When there is no risk of confusion we drop the decoration $\bullet_{\mathcal{C}(A, H)}$ in the product.
\section{Two examples of Hopf--Galois extension}\label{se:hge}
We review two well know examples of Hopf--Galois extensions for which in
\S\S ~\ref{sec:ev-q-shp} and \ref{se:alg-su2sym}
we shall explicitly construct the corresponding algebroids.
\subsection{The $\SU(2)$ principal fibration} \label{sec:HG-SU2}
Consider the sphere $S_\theta^4$ constructed in \cite{CL01}.
With $\theta$ a real parameter, the algebra $A(S_\theta^4)$ of polynomial functions on the sphere $S_\theta^4$ is
generated by elements $\zeta_0=\zeta_0^*$ and $\zeta_j, \zeta_j^*$, $j=1,2$, subject to relations
\begin{equation}\label{s4t}
\zeta_\mu \zeta_\nu = \lambda_{\mu\nu} \zeta_\nu \zeta_\mu, \quad \zeta_\mu \zeta_\nu^* = \lambda_{\nu\mu} \zeta_\nu \zeta_\mu^*,
\quad \zeta_\mu^* \zeta_\nu^* = \lambda_{\mu\nu} \zeta_\nu^* \zeta_\mu^*, \quad \mu,\nu = 0,1,2 ,
\end{equation}
with deformation parameters given by
\begin{equation}
\lambda_{1 2} = \bar{\lambda}_{2 1} =: \lambda=e^{2\pi {\,{\rm i}\,} \theta},
\quad \lambda_{j 0} = \lambda_{0 j } = 1, \quad j=1,2 ,
\end{equation}
and together with the spherical relation $\sum_\mu \zeta_\mu^* \zeta_\mu=1$. For $\theta=0$ one recovers
the
$*$-algebra of complex polynomial functions on the usual sphere $S^4$.
On the sphere $S_\theta^4$ there is an $\SU(2)$ noncommutative principal fibration
$S_{\theta'}^7 \to S_\theta^4$ given in \cite{LS04}. Firstly, with $\lambda'_{a b} = e^{2 \pi {\,{\rm i}\,} \theta'_{ab}}$ and $(\theta'_{ab})$ a real antisymmetric matrix,
the algebra $A(S_{\theta'}^7)$ of polynomial functions on the sphere $S_{\theta'}^7$ is generated by elements
$\psi_a, \psi_a^*$, $a=1,\dots,4$, subject to relations
\begin{equation}\label{s7t}
\psi_a \psi_b = \lambda'_{a b} \, \psi_b \psi_a, \quad \psi_a \psi_b^* = \lambda'_{b a} \, \psi_b^* \psi_a,
\quad \psi_a^*\psi_b^* = \lambda'_{a b} \, \psi_b^* \psi_a^* ,
\end{equation}
and with the spherical relation $\sum_a \psi_a^* \psi_a=1$.
At $\theta=0$, it is the $*$-algebra of
complex polynomial functions on the sphere $S^7$. For the noncommutative Hopf bundle over the given 4-sphere $S_\theta^4$, we need to select
a particular noncommutative 7 dimensional sphere $S_{\theta'}^7$. We take the one corresponding to the
following deformation parameters
\begin{equation}\label{lambda7}
\lambda'_{ab}=
\begin{pmatrix} 1 & 1 & \bar{\mu} & \mu \\
1 & 1 & \mu & \bar{\mu} \\
\mu & \bar{\mu} &1 & 1\\
\bar{\mu} & \mu &1 & 1
\end{pmatrix}, \quad \mu = \sqrt{\lambda} \qquad \mathrm{or} \qquad
\theta'_{ab}=\frac{\theta}{2}\begin{pmatrix} 0 & 0 & -1 & 1 \\
0 & 0 & 1 & -1 \\
1 & -1 & 0 & 0 \\
-1 & 1 & 0 & 0 \end{pmatrix}.
\end{equation}
The previous choice is essentially the only one that allows the algebra $A(S_{\theta'}^7)$ to carry an action
of the group $\SU(2)$ by automorphisms and such that the invariant subalgebra coincides with
$A(S_\theta^4)$. The best way to see this is by means of the matrix-valued function on $A(S_{\theta'}^7)$ (we are
changing notations with respect to \cite{LS04})
\begin{equation}\label{Psi}
\Psi =
\begin{pmatrix}
\psi_1 & - \psi^*_2 \\
\psi_2 & \psi^*_1 \\
\psi_3 & -\psi^*_4 \\
\psi_4& \psi^*_3
\end{pmatrix}.
\end{equation}
Then, the commutation relations of the algebra $A(S_{\theta'}^7)$, with deformation parameter in \eqref{lambda7}, gives that $\Psi^\dagger \Psi = \mathbb{I}_2 $. As a consequence, the matrix-valued function $p = \Psi \Psi^\dagger$ is a projection, $p^2=p=p^\dagger$, and its entries rather that functions in $A(S_{\theta'}^7)$ are (the generating) elements of $A(S_\theta^4)$. Indeed, the right coaction of $A(\SU(2))$ on $A(S_{\theta'}^7)$ is simply given by
\begin{equation} \label{actionSU2}
\delta (\Psi) = \Psi \stackrel{.}{\otimes} w , \qquad w = \begin{pmatrix} w_1 & - w^*_2 \\ w_2 & w^*_1
\end{pmatrix} \in A(\SU(2)), \quad w w^\dagger = 1 = w^\dagger w .
\end{equation}
If $\sigma(a {\otimes} b) = b {\otimes} a$ is the flip, this gives
$$ \delta (\Psi^\dagger) = \sigma(w^\dagger \stackrel{.}{\otimes} \Psi^\dagger) ,
$$
and the invariance of the entries of $p$ follows at once:
\begin{equation}\label{once}
p \mapsto \delta (\Psi) \, \delta (\Psi^\dagger) = p \stackrel{.}{\otimes} w w^\dagger = p \stackrel{.}{\otimes} 1.
\end{equation}
The generators of $A(S^4)$, the independent entries of $p$, are identified as bilinears expressions in the $\psi,\psi^*$'s. Explicitly,
\begin{equation}\label{proj}
p = \Psi \cdot \Psi^\dagger =
\begin{pmatrix}
\zeta_0 & 0 & \zeta_1 & - \bar{\mu} \zeta_2^* \\
0 & \zeta_0 & \zeta_2 & \mu \zeta_1^* \\
\zeta_1^*& \zeta_2^* & 1-\zeta_0 & 0\\
-\mu \zeta_2 & \bar{\mu} \zeta_1 & 0 & 1-\zeta_0
\end{pmatrix} ,
\end{equation}
with
\begin{align} \label{sub}
\zeta_1 &= \psi_1 \psi^*_3 + \psi^*_2 \psi_4 , \qquad
\zeta_2 = \psi_2 \psi^*_3 - \psi^*_1 \psi_4 , \nonumber \\
\zeta_0 &= \psi_1 \psi^*_1 + \psi^*_2 \psi_2 = 1 - \psi_3 \psi^*_3 - \psi^*_4 \psi_4 .
\end{align}
By using the commutation relations of the $\psi$'s, one computes the commutation rules
$\zeta_1 \zeta_2 = \lambda \zeta_2 \zeta_1$, $\zeta_1 \zeta_2^* = \bar{\lambda} \zeta_2^* \zeta_1$,
and that $\zeta_0$ is central and hermitian and $\zeta_1$, $\zeta_2$ are normal. The spherical relation for
$S_{\theta'}^7$ gives an analogous one,
$\zeta_1^* \zeta_1 + \zeta_2^* \zeta_2 = \zeta_0 (1-\zeta_0)$, for $S_\theta^4$.
There are compatible toric actions on $S_\theta^4$ and $S_{\theta'}^7$ (see e.g. \cite[\S 2.3]{Brain:2013wp}.)
With a slight change of notation, the torus $\IT^2$ acts on $A(S_\theta^4)$ as
\begin{equation}\label{eq:act-S4}
\sigma_s(\zeta_0, \zeta_1, \zeta_2) = (\zeta_0, e^{2\pi i s_1} \zeta_1, e^{2\pi i s_2} \zeta_2), \quad s\in \IT^2 .
\end{equation}
This action is lifted to a double cover action on $A(S_{\theta'}^7)$. The double cover map $p: \widetilde \IT^2 \to \IT^2$ is
given explicitly by $p:(s_1,s_2) \mapsto (s_1+s_2,-s_1+s_2)$.
Then $\widetilde \IT^2$ acts on the $\psi_a$'s as:
\begin{equation} \label{eq:lift-S7}
\widetilde \sigma: \left( \psi_1, \psi_2, \psi_3, \psi_4 \right)
\mapsto \left(e^{2\pi i s_1}~\psi_1, ~e^{-2\pi i s_1}~\psi_2, ~e^{-2\pi i s_2}~\psi_3, ~e^{2\pi i s_2}~\psi_4 \right)
\end{equation}
The sense in which the algebra inclusion $A(S_\theta^4) \subset A(S_{\theta'}^7)$ is a nontrivial (faithfully flat)
noncommutative $\SU(2)$ principal bundle is explained in \cite{LS04}. Here we mention
that there is a canonical Galois maps $\chi : A(S_{\theta'}^7) {\otimes}_{A(S_\theta^4)} A(S_{\theta'}^7) \to A(S^{7}_{\theta})\otimes A(SU(2))$ which is invertible.
The corresponding translation map $\tau : A(\SU(2)) \to A(S_{\theta'}^7) {\otimes}_{A(S_\theta^4)} A(S_{\theta'}^7)$ on generators is
\begin{equation}
\tau(w) = \Psi^\dagger \stackrel{.}{\otimes}_{A(S_\theta^4)} \Psi
\end{equation}
Indeed, $\chi \circ \tau(w) = \chi(\Psi^\dagger \stackrel{.}{\otimes}_{A(S_\theta^4)} \Psi) = \Psi^\dagger \delta (\Psi) = \Psi^\dagger \Psi \stackrel{.}{\otimes}
w = 1 {\otimes} 1_2 w = 1 {\otimes} w$.
There is also a copy of the projection $p$ in the opposite algebra:
\begin{equation}
q = \Psi \cdot_{op} \Psi^\dagger =
\begin{pmatrix}
\zeta_0 & 0 & \bar{\mu} \zeta_1 & - \zeta_2^* \\
0 & \zeta_0 & \mu \zeta_2 & \zeta_1^* \\
\mu \zeta_1^*& \bar{\mu} \zeta_2^* & 1- \zeta_0 & 0\\
- \zeta_2 & \zeta_1 & 0 & 1- \zeta_0
\end{pmatrix}
\end{equation}
The difference between $p$ and ${q}$ is due to the multiplication in $A(S_{\theta'}^7)$ versus the one in $A(S_{\theta'}^7)^{op}$. Indeed:
\begin{equation}\label{rojopprojp}
p_{mn} = \sum_r \Psi_{mr} \Psi^\dagger{}_{rn} , \qquad {q}_{mn} = \sum_r\Psi_{mr} \cdot_{op} \Psi^\dagger{}_{rn} = \sum_r \Psi^\dagger{}_{rn} \Psi_{mr} .
\end{equation}
With the commutation relations \eqref{s7t}, the condition $\Psi^\dagger \cdot \Psi = \mathbb{I}_2 $ leads also to $\Psi^\dagger \cdot_{op} \Psi = \mathbb{I}_2$.
\subsection{Principal bundles over even quantum spheres} \label{sec:-q-shp}
Even noncommutative spheres $S_\theta^{2n}$, introduced in \cite{CL01}, were shown in \cite{Var01} to be homogeneous spaces of quantum groups $\SO_\theta(2n+1, \mathbb{R})$.
The algebra of coordinate functions of the latter $A=\O(\SO_\theta(2n+1, \mathbb{R}))$ is the total space algebra of a principal bundle over the algebra $B=\O(S_\theta^{2n})$ for the Hopf (structure) algebra
$H=\O(\SO_\theta(2n, \mathbb{R}))$.
These bundles were worked out in details in \cite[\S 4.1.1]{ABPS17} that we follows with changes.
Start with the commutative torus $\IT^n$ with generators $t_j, t_j^*$ and relations $t_j t_j^*=t_j^* t_j=1$.
Consider the bi-character $\gamma : \IT^n \times \IT^n \to \U(1)$ defined on generators by
$$
\gamma(t_j, t_k) = e^{{\,{\rm i}\,} \pi \theta_{jk}} , \qquad \theta_{j k} = - \theta_{k j} .
$$
We shall denote $\lambda_{jk} = \gamma(t_j, t_k)^2 =e^{2i\pi \theta_{jk}}$.
In order for the deformed algebra to still be a Hopf algebra one needs
a left and a right action of $\IT^n = \diag(t_1, \dots, t_n, t_1^*, \dots, t_n^*)$ (or of
$\IT^n \times \IT^n$).
This action then allows one to deform the algebra $\O(\SO(2n))$ into
an algebra $\O(\SOt(2n))$ described as follows. It has generators $\b{a} = (a_{jk})$, $\b{b} = (b_{jk})$,
$\b{a}^* = (a_{jk}^*)$, $\b{b}^* = (b_{jk}^*)$ with commutation relations computed to be
\begin{align}\label{thetaCR}
a_{ij} a_{kl} & = \lambda_{ik}\lambda_{lj} ~a_{kl} a_{ij} , \qquad
a_{ij} b^*_{kl} = \lambda_{ki}\lambda_{lj} ~b^*_{kl} a_{ij} \nonumber
\\
a_{ij} b_{kl} & = \lambda_{ik}\lambda_{jl} ~b_{kl} a_{ij} , \qquad
a_{ij} a^*_{kl} = \lambda_{ki}\lambda_{jl} ~a^*_{kl} a_{ij} \nonumber
\\
b_{ij} b_{kl} & = \lambda_{ik}\lambda_{lj} ~b_{kl} b_{ij} , \qquad
b_{ij} b^*_{kl} = \lambda_{ki}\lambda_{jl} ~b^*_{kl} b_{ij}
\end{align}
together with their $*$-conjugated. In fact, the Hopf algebra structure of $\O(\SO(2n))$ survives the quantization.
In matrix notation the deformed $\O(\SOt(2n))$ has coproduct and counit given by
\begin{equation}\label{eq:mtxM}
M = (M_{J K}) =
\begin{pmatrix}
\b{a} & \b{b} \\
\b{b}^* & \b{a}^*
\end{pmatrix} , \qquad \Delta(M) = M \stackrel{.}{\otimes} M , \qquad \varepsilon(M) = \mathbb{I} .
\end{equation}
To define an antipode there is a suitable determinant ${\det}_{\theta}(M)$
and one can pass to the quotient by the $*$-bialgebra ideal given by
\begin{equation}\label{idealM}
I_Q = <M^t Q M - Q , \, M Q M^t - Q , \, {\det}_{\theta}(M) - 1 > , \qquad
Q = \begin{pmatrix}
0 & \mathbb{I}_n \\
\mathbb{I}_n & 0
\end{pmatrix} = Q^{-1}.
\end{equation}
The $*$-structure is then $*M=Q M Q$ while the antipode is $S(M)= Q M^t Q = M^\dagger $.
The previous conditions reads then $M^\dagger M = M M^\dagger = \mathbb{I}_{2n}$.
The odd case of $\O(\SOt(2n+1))$ is defined in a similar fashion by deforming the
left and right actions of the torus
$\IT^n= \diag(t_1, \dots, t_n, t_1^*, \dots, t_n^*, 1)$ on $\O(\SO(2n+1))$. In matrix notation
$$
N = (N_{J K}) =
\begin{pmatrix}
\b{a} & \b{b} & \b{u} \\
\b{b}^* & \b{a}^* & \b{u}^* \\
\b{v} & \b{v}^* & x
\end{pmatrix} ,
$$
with $n$-component column vectors $\b{u} = (u_j), \b{u}^* = (u_j^*)$ and row vectors $\b{v} = (v_j), \b{v}^* = (v_j^*)$ and a hermitian scalar $x$. The commutation relations are found to be given by
\begin{equation}\label{crN}
N_{I J} \, N_{K L} = \lambda_{IK} \lambda_{L J} N_{K L} \, N_{I J} .
\end{equation}
Now the coproduct and antipode are as before by $\Delta(N) = N \stackrel{.}{\otimes} N$ and $\varepsilon(N) = \mathbb{I}$ and one verifies ideal conditions analogue to the ones in \eqref{idealM}:
\begin{equation}\label{ocN}
N^t Q N = Q , \quad N Q N^t = Q , \quad {\det}_{\theta}(N) = 1, \qquad
Q = \begin{pmatrix}
0 & \mathbb{I}_n & 0 \\
\mathbb{I}_n & 0 & 0 \\
0 & 0 & 1
\end{pmatrix} = Q^{-1}.
\end{equation}
The $*$-structure is $*N=Q N Q$ while the antipode is $S(N)= Q N^t Q = N^\dagger$. Then the previous condution read $N^\dagger N = N N^\dagger = \mathbb{I}_{2n+1}$.
The Hopf algebra $\O(\SOt(2n))$ is a quantum subgroup of $\O(\SOt(2n+1))$
with surjective Hopf algebra morphism
\begin{align}\label{subha}
\pi : \O(\SOt(2n+1)) & \to \O(\SOt(2n)), \nonumber \\
\begin{pmatrix}
\b{a} & \b{b} & \b{u} \\
\b{b}^* & \b{a}^* & \b{u}^* \\
\b{v} & \b{v}^* & x
\end{pmatrix} & \mapsto
\begin{pmatrix}
\b{a} & \b{b} & 0 \\
\b{b}^* & \b{a}^* & 0 \\
0 & 0 & 1
\end{pmatrix} =:
\begin{pmatrix}
{\bf h} & {\bf k} & 0 \\
{\bf k}^* & {\bf h}^* & 0 \\
0 & 0 & 1
\end{pmatrix} = \b{w} .
\end{align}
\noindent
This results into a right coaction of $\O(\SOt(2n))$ on $\O(\SOt(2n+1))$:
\begin{align}
\delta^A : \O(\SOt(2n+1)) & \to \O(\SOt(2n+1)) {\otimes} \O(\SOt(2n)) , \nonumber \\
\delta^A(N) & = N \stackrel{.}{\otimes} \pi(N).
\end{align}
The subalgebra $B$ of coinvariant elements, generated by the last column of the matrix $N$: $(u_j, u_j^*, x)$,
is the algebra $\O(S_\theta^{2n})$ of coordinate functions on a quantum $2n$-sphere $S_\theta^{2n}$.
The commutation relations of the generators follows from \eqref{crN}:
\begin{equation}
u_i u_j= \lambda_{ij} \, u_j u_i \, , \qquad
u_i^* u^*_j= \lambda_{ij} \, u^*_j u^*_i \, ,
\qquad u_i u_j^*= \lambda_{ji} \, u_j^* u_i \, ,
\end{equation}
and $x$ central. The orthogonality conditions \eqref{ocN} imply the sphere
relation
$$
\sum_{j=1}^{n} 2 u_j^* u_j + x^2 = 1 ,
$$
(each generator is normal $u_j^* u_j = u_j u_j^*$).
The algebra extension $\O(S_\theta^{2n}) \subset \O(\SOt(2n+1))$ is a Hopf Galois extension
for the Hopf algebra $H=\O(\SOt(2n))$ (cf. \cite[\S 4.1.1]{ABPS17}).
In particular we record the form of the translation map to be used later on.
In components
\begin{align}
\label{eq:tranmap-h}
\tau(\b{h}) &= \b{a}^\dagger \stackrel{.}{\otimes}_B \b{a} + (\b{b}^*)^\dagger \stackrel{.}{\otimes}_B \b{b}^* + \b{v}^\dagger \stackrel{.}{\otimes}_B \b{v} , \nonumber \\
\tau(\b{k}) &= \b{a}^\dagger \stackrel{.}{\otimes}_B \b{b} + (\b{b}^*)^\dagger \stackrel{.}{\otimes}_B \b{a}^* + \b{v}^\dagger \stackrel{.}{\otimes}_B \b{v}^* .
\end{align}
\section{Algebraic $\theta$-deformations} \label{alg-def}
In this section we review the general scheme of deforming by the action of tori.
This will be done in the crudest way via $\mathbb{Z}^n$-graded spaces and deforming relevant structures
by means of a bi-character.
The role of $\mathbb{Z}^n$ comes from it being the Pontryagin dual of the torus $\IT^n$ and one is effectively deforming
objects with a torus action. More details are e.g. in \cite{Brain:2012} and \cite{Brain:2013wp}. In particular we shall deform principal bundles and associated Hopf algebroids.
A general scheme of deformations of noncommutative principal bundles via convolution invertible 2-cocycles
$\gamma : H {\otimes} H \to \mathbb{C}$ on a Hopf algebra $H$ is in \cite{ABPS17}.
Let $\mathcal T_n$ be the category of $\mathbb{Z}^n$-graded complex vector spaces whose objects are written as (finite) sums of the kind
\begin{align*}
V = \bigoplus_{ r \in \mathbb{Z}^n } V_r, \qquad
p_r : V \to V_r .
\end{align*}
Here $p_r$ is the projection onto the $r$-th component, and most of the time
we simply use a subscript to indicate the projection
$ v_r = p_r (v)$ for $v \in V$.
Morphisms $\psi \in \Hom(V , W)$ are linear maps that preserve homogeneity,
but not necessarily the degree.
More precisely, there always exists a group homomorphism
$\rho_\psi : \mathbb{Z}^n \to \mathbb{Z}^n$ such that
\begin{align}
\label{eq:mor-N}
\psi (V_r) \subset V_{ \rho_\psi (r) }.
\end{align}
In the $\theta$-deformation literature, one starts with a smooth
action of a $n$-torus on a Fr\'echet space $V$, $t \in \IT^n \mapsto \alpha_t \in \mathrm{Aut} (V)$.
The induced $\mathbb{Z}^n$-grading, based as mentioned on Pontryagin duality, is given by
projections $p_r : V \to V_r$, $r \in \mathbb{Z}^n$, taking the $r$-th
Fourier coefficients of the vector-valued function $ t \to \alpha_t (v)$,
\begin{align*}
p_r (v) = \int_{ \IT^n } e^{- 2 \pi i r \cdot t} \alpha_t (v)\, \mathrm{d} t , \, \, \, \, v \in V,
\end{align*}
where $\mathrm{d} t$ is the normalized Lebesgue measure on $\IT^n$.
Morphisms as in \eqref{eq:mor-N} corresponds to linear maps
$ \widetilde \psi : V \to W$ which are $\IT^n$-equivariant up-to a
group homomorphism $ \widetilde \rho_{ \widetilde \psi} : \IT^n \to \mathbb T^n$
so that the diagram commute:
\begin{equation}
\begin{tikzcd}
\mathbb T^n \times V
\arrow[d, " \rho_{ \widetilde \psi} \times \widetilde \psi "] \arrow[r]
& V \arrow[d, " \widetilde \psi"] \\
\mathbb T^n \times W \arrow[r] & W
\end{tikzcd} .
\end{equation}
The parameter $\theta$ in a $\theta$-deformation is a
$n \times n$ skew-symmetric matrix and what is actually needed for the
deformation is the induced bi-character on $\mathbb{Z}^n$, that is a map
\begin{align*}
\lambda_\theta : \mathbb{Z}^n \times \mathbb{Z}^n \to \IT , \quad
(r, l) \mapsto \lambda_\theta (r , l)
:= e^{ \pi i \abrac{ \theta r , l} } ,
\end{align*}
which is a $2$-cocycle in the sense of
\begin{align}
\label{eq:2-coc-e-theta}
\lambda_\theta ( r , l)
\lambda_\theta ( r + l , s) =
\lambda_\theta ( l , s)
\lambda_\theta ( r , s + l ) , \quad r, s, l \in \mathbb{Z}^n .
\end{align}
The tensor products functor
$ \otimes : \mathcal T_n \times \mathcal T_n \to \mathcal T_n$
makes $ \mathcal T_n$ into a monoidal category, in which
the $\mathbb{Z}^n$-grading is assigned in the usual way,
\begin{equation}\label{tot-gra}
( V \otimes W)_s = \bigoplus_{ s = r + l } V_r \otimes V_l, \qquad
s , r, l \in \mathbb{Z}^n ,
\end{equation}
by taking the total degree of the natural bi-grading.
One can deform the tensor functor via the following natural transformation
$c^\theta$, for any $V , W \in \mathcal T_n$,
\begin{align}
c^\theta_{ V , W} : V \otimes W \to V \otimes_\theta W , \quad
v_r \otimes w_l \to v_r \otimes_\theta w_l :=
\lambda_\theta (r, l) v_r \otimes w_l
\label{eq:cVW-dfn}
\end{align}
which is defined firstly on homogeneous elements and then extended by linearity.
It is not difficult to see that it has an inverse given by
$( c^\theta_{ V , W})^{-1} = c^{- \theta}_{ V , W}$.
Given an algebra $(A , m)$ in $ \mathcal T_n$, with multiplication
$m: A \otimes A \to A$ preserving the grading,
\begin{align}
\label{eq:A-grd}
m \brac{ A_r \otimes A_l } \subset A_{ r + l} , \, \, \, \,
r, l \in \mathbb{Z}^n,
\end{align}
its deformation $A_\theta = ( A , m_\theta )$ maintains the underlying
(graded) vector space unchanged, but endowed with a new multiplication:
\begin{align}
\label{eq:m_theta-defn}
m_\theta = m \circ c^\theta_{ A , A} :
A \otimes A \xrightarrow{ c^\theta_{ A , A}} A \otimes A
\xrightarrow{m} A .
\end{align}
As shown in \eqref{eq:cVW-dfn},
$m$ is twisted by a phase factor on homogeneous elements:
\begin{align}
\label{eq:m_theta-r-l}
m_\theta ( a_r , \tilde a_l ) = \lambda_\theta (r , l)
m ( a_r , \tilde a_l) ,
\end{align}
which, provided that $m$ is commutative, leads to the commutation relations:
\begin{align}
\label{eq:com-rltn-A}
m_\theta ( a_r , \tilde a_l ) =
\lambda_\theta (r , l)^2
m_\theta ( a_l , \tilde a_r ) .
\end{align}
The required associativity for $m_\theta$ follows directly
from the $2$-cocycle condition in \eqref{eq:2-coc-e-theta}.
For easy of notation in the following we shall denote $m_\theta ( a_l, \tilde a_r )= a_l \cdot_\theta \tilde a_r$.
Clearly, $\lambda_\theta (r , \pm r) = 1$ since $\theta$ is skew-symmetric.
We record this simple observation as a lemma which will be used often later on.
\begin{lem}
\label{lem:mtheta-m}
For homogeneous element $a , \tilde a \in A$ of the same degree
or of the opposite degree, that is $\deg a \pm \deg \tilde a =0$,
the deformed multiplication agrees with the original one:
\begin{align}
\label{eq:mtheta-m}
m_\theta (a , \tilde a) = m( a , \tilde a) .
\end{align}
In particular, \eqref{eq:mtheta-m} holds whenever the product
$m(a , \tilde a) \in A_{ 0 } $ belongs to the degree zero component, in this case,
$a , \tilde a$ are not required to be homogeneous.
\end{lem}
In a similar manner, for an $A$-module $V$ in $\mathcal T_n$ such that
the action $ \rhd : A \otimes V \to V$ preserves the grading as in
\eqref{eq:A-grd}, the deformation
$ \rhd_\theta := \rhd \circ c^\theta_{ A , V} $ makes $V_\theta$ into an
$A_\theta$-module.
The `associativity' (the action properties) for $\rhd_\theta$ again follows directly
from the $2$-cocycle condition in \eqref{eq:2-coc-e-theta}. There is clearly a right-module version of this.
And finally, if $(C,\Delta)$ is a coalgebra in $\mathcal T_n$ with $\Delta: C \to C{\otimes} \mathbb{C}$ that preserves the degree in the sense of \eqref{tot-gra}: $\Delta(c_s) = \sum_{r+l=s} \one{c_r} {\otimes} \two{c_l}$ the deformation
$\Delta_\theta := c^{-\theta}_{C, C} \circ \Delta$ makes $C_\theta$ into a coalgebra with co-associativity again
following from the $2$-cocycle condition.
The next step in deforming a bialgebra (or even a Hopf algebra) structures needs some extra care.
Also, for deforming a Hopf--Galois extension with structure Hopf algebra $H$, and aiming at including
both examples in \S\S ~\ref{sec:HG-SU2} and \ref{sec:-q-shp},
it turns out that the construction of gradings on the algebra involved and the related assumptions are
quite different depending on whether the Hopf algebra is deformed or not.
We will break the discussion into two scenarios to cover the constructions
of both \S\S ~\ref{subsec:ScI} and \ref{subsec:ScII}
in which our aim is to get a (possible new) structure Hopf algebra with related comodule
algebras out of the $\theta$-deformation scheme.
After that, the deformation of the Ehresmann--Schauenburg bialgebroids can be handled
in a uniform way, and will be carried out in \S\S ~\ref{subsec:ES-alg} and \ref{subsec:flip-atpd}.
\subsection{Scenario I: No Hopf algebra is deformed}
\label{subsec:ScI}
We start with a setting in which the Hopf algebra $H$ is not touched.
Thus, we assume that $H$ has trivial $\mathbb{Z}^n$-grading and a $H$-comodule algebra
$A$ is $\mathbb{Z}^n$-graded so that the multiplication preserves the grading as in
\eqref{eq:A-grd},
and the coaction $\delta^A : A \to A \otimes H$ also behaves the same way:
\begin{equation}
\label{eq:coprd-grd}
\delta^A ( A_r ) \subset A_r \otimes H .
\end{equation}
Thus when writing $\delta^A(a) = \zero{a} {\otimes} \one{a}$ one has $\deg a = \deg \zero{a} $.
The following is an almost free version of \cite[Cor. 3.16]{ABPS17}.
\begin{prop} \label{prop:comodalg-theta}
Let $A_\theta = ( A , m_\theta)$ be the deformation of $A$ as in
\eqref{eq:m_theta-defn}. It is still a $H$-comodule algebra with
the same coaction treated as
$\delta^A: A_\theta \to A_\theta \otimes H$.
Then, the coinvariant subspace $B = A^{\mathrm{co} H}$ remains the
same and the $\theta$-multiplication can be restricted onto $B$ to form
$B_\theta = (B, m_\theta) $.
Moreover, if the starting pair $(A , H)$ is a Hopf--Galois extension with algebra of coinvariant elements $B$,
such is its deformation $( H , A_\theta)$, with algebra of coinvariant elements $B_\theta$.
\end{prop}
\begin{proof}
The first part is evident. As for the final (almost evident) statement, consider the starting canonical map
$\chi : A {\otimes} _B A \longrightarrow A {\otimes} H , \quad \chi (\tilde{a} {\otimes}_B a) = \tilde{a} \zero{a} {\otimes} \one{a}$ and define
$$
\chi_\theta : A_\theta {\otimes} _{B_\theta} A_\theta \longrightarrow A_\theta {\otimes} H , \quad \chi_\theta (\tilde{a} {\otimes}_B a) = \tilde{a} \cdot_\theta \zero{a} {\otimes} \one{a}.
$$
Then, for $h \in H$ consider the starting canonical map $\tau(h) = \tuno{h} {\otimes}_B \tdue{h}$
with components (sum of terms) of opposite degree $\deg \tuno{h} = - \deg \tdue{h}$ since
$H$ has zero degree which is preserved by $\tau$.
Then, from Lemma \ref{lem:mtheta-m}, \begin{align*}
\chi_\theta (\tuno{h} {\otimes}_B \tdue{h}) & = \tuno{h} \cdot_\theta \zero{\tdue{h}} {\otimes} \one{\tdue{h}} =
\tuno{h} \zero{\tdue{h}} {\otimes}_B \one{\tdue{h}} \\
& = \chi(\tuno{h}{\otimes}_B \tdue{h})
\end{align*}
(the latter being just $1 {\otimes} h$ from \eqref{p7}) and $\chi_\theta$ is invertible if and only if $\chi$ is.
Thus the translation map of $\chi_\theta$ is the same as the starting
undeformed one that can be considered as a map
$\tau : H \to A_\theta {\otimes}_{B_\theta} A_\theta$.
\end{proof}
\begin{rem}[On the degree of the translation map]
The fact that, in writing for the translation map $\tau(h) = \tuno{h} {\otimes}_B \tdue{h}$, one can take
$\deg \tuno{h} = - \deg \tdue{h}$ does not depend on the representatives: suppose
$\tuno{h} {\otimes}_B \tdue{h} = \tuno{\tilde h} b {\otimes}_B \tdue{h} =
\tuno{\tilde h} {\otimes}_B b \tdue{h} = \tuno{\tilde h} {\otimes}_B \tdue{\tilde h}$. Then it follows that
$\deg \tuno{h} = - \deg \tdue{h}$ if and only if $\deg \tuno{\tilde h} = - \deg \tdue{\tilde h}$.
\end{rem}
\begin{exa}[Noncommutative Hopf-fibration] \label{rem-dtm}
As mentioned, the $\mathbb{Z}^n$-grading we consider is derived from a torus action.
To construct the $\SU(2)$-fibration $S^7_\theta \to S^4_\theta$ in \S\ref{sec:HG-SU2},
one begins with a two torus action defined in \eqref{eq:act-S4} and \eqref{eq:lift-S7},
in which all generators in \ref{s4t} and \ref{s7t} are $\IT^2$-eigenfunctions.
Then,
\begin{align*}
& \deg \zeta_0 = \deg \zeta_0^* = 0
\\
& \deg\zeta_1 = (1, 0), \, \, \, \, \deg \zeta_2 = (0 , 1)
\\
& \deg \psi_1 = - \deg \psi_2
= (1,0) , \, \, \, \,
\deg \psi_4 = -\deg \psi_3
= (0,1) .
\end{align*}
The deformation matrix just reads
$\begin{bmatrix} 0& - \theta \\ \theta & 0
\end{bmatrix} $, with $\theta \in \mathbb R$.
We have, according to \eqref{eq:m_theta-r-l},
\begin{align*}
\zeta_\mu \cdot_\theta \zeta_\nu = \sqrt{ \lambda_{ \mu \nu}}
\zeta_\mu \zeta_\nu, \, \, \, \,
\psi_a \cdot_\theta \psi_b = \sqrt{ \lambda'_{a b}}
\psi_a \psi_b ,
\end{align*}
so that the commutation relations \ref{s4t} and \ref{s7t}
follow immediately from \eqref{eq:com-rltn-A}.
It is also worth noting that the double covering between
the two actions \eqref{eq:act-S4} and \eqref{eq:lift-S7}
is exactly dual to the following map $\mathbb{Z}^2 \to \mathbb{Z}^2$:
\begin{align*}
(1,0) &= \deg \zeta_1 \mapsto (1, 1) = \deg \psi_1 \psi^*_3
= \deg \psi_2^* \psi_4, \\
(0,1) &= \deg \zeta_2 \mapsto (-1,1) = \deg \psi_2 \psi^*_3
= \deg \psi_1^* \psi_4.
\end{align*}
revealed in the embedding $A(S^4_\theta) \to A(S^7_\theta)$ given by
\ref{sub}.
\qed\end{exa}
The details of the construction of the Ehresmann--Schauenburg bialgebroid related to the Hopf--Galois extension
$(H, A_\theta)$, are postponed to \S\ref{subsec:ES-alg}.
\subsection{Scenario II: Deforming Hopf algebras and homogeneous spaces}
\label{subsec:ScII}
Unlike the previous section, in order to deform a Hopf algebra $H$
(or in a more accurate context, to only deform the algebra structure of $H$),
in a way that the all compatibilities axioms for Hopf algebras remains,
one needs a more delicate setup for the grading and the $\theta$-matrix.
To motivate the long list of requisites below,
the reader is referred to App. \ref{sec:Qgrp}, where we recall the
original formulation in terms of torus actions due to Rieffel
\cite{Rieffel:1993tw}.
Let $H = \bigoplus_{ r, s \in \mathbb{Z}^n} H_{ (r,s)} $
be a Hopf algebra with a bi-grading of $\mathbb{Z}^n$ (in particular, a grading of $\mathbb{Z}^{2n}$),
such that the group homomorphism $\rho_{ \psi}$ on
gradings in \eqref{eq:mor-N} induced via the structure maps of $H$ are given as follows:
\begin{itemize}
\item[i)] the multiplication preserves the grading as in \eqref{eq:A-grd}
\begin{align}
\label{eq:prd-bigrd}
m ( H_{ (r ,s)} \otimes H_{ (p , q)} )
\subset H_{ (r + p , s + q)} \, ,
\end{align}
\item[ii)] the coproduct $\Delta : H \to H \otimes H$, is required to be such that
\begin{align}
\label{eq:coprd-bigrd}
\Delta \brac{ H_{ (r, l )}} \subset
\bigoplus_{ s \in \mathbb{Z}^n} H_{ (r , s)} \otimes H_{ (s, l)} \, ,
\end{align}
\item[iii)]
the counit factors through the projection:
\begin{align}
\label{eq:counit-bigrd}
\varepsilon: H \to \bigoplus_{ s \in \mathbb{Z}^n} H_{ (s , s)} \to \mathbb{C},
\end{align}
that is $ \varepsilon ( h_{ (r , l)} ) = 0$ for all homogeneous elements
$h_{ (r , l)} $ with $r \neq l$ ,
\item[iv)] for the antipode and the $*$-operator (if $H$ has one), one assumes
\begin{align}
\label{eq:antp-star-gd}
S ( H_{ (r , l )} ) \subset H_{ (-l , - r) } ,
\qquad
* ( H_{ (r , l)} ) \subset H_{ (-r , -l) } .
\end{align}
\end{itemize}
\begin{rem}[On condition \eqref{eq:coprd-bigrd}]
From the general assigning of the total degree in \eqref{tot-gra}, on the right hand side of \eqref{eq:prd-bigrd} one would have
$\bigoplus_{ a+c = r, b+d = l } H_{ (a , b)} \otimes H_{ (c, d)}$.
The subspaces
$\bigoplus_{ s \in \mathbb{Z}^n } H_{ (r , s)} \otimes H_{ (s, l)} $ of
$H \otimes H$ when summed on the indices $r,l$ corresponds to
the subspace $ \mathcal D$ in \eqref{eq:subspace-C} in which the coproduct lands.
\qed\end{rem}
Next,
let $\theta$ be a $n \times n$ skew-symmetric matrix and put
\begin{align}
\label{eq:Theta-matrix}
\Theta =
\begin{bmatrix} \theta & 0 \\ 0 & - \theta \end{bmatrix} ,
\end{align}
so that their $2$-cocycles are related as follows:
for $r = (r_1 , r_2 )$ and $l = (l_1 , l_2)$,
\begin{align*}
\lambda_{\Theta} ( r , l )
=
\lambda_\theta ( r_1 , l_1 )
\lambda_{ -\theta } ( r_2 , l_2).
\end{align*}
Denote by $H_{\Theta} = (H , \cdot_{\Theta} )$ the deformed algebra, with the new
multiplication given, on homogeneous elements $h_r , g_l $ of degree
$r , l \in \mathbb{Z}^n$ respectively, by
\begin{align}
\label{eq:Theta-hr-gl}
h_r \cdot_{\Theta} g_l =
\lambda_\Theta (r ,l)
h_r g_l =
\lambda_\theta ( r_1 , l_1 )
\lambda_{ - \theta } ( r_2 , l_2)
h_r g_l .
\end{align}
\begin{lem}
\label{lem:coprd-prd}
With the condition in \eqref{eq:coprd-bigrd},
the (undeformed) coproduct $\Delta$ is still an algebra homomorphism for the product $ \cdot_{\Theta}$:
\begin{align*}
\Delta ( h \cdot_{\Theta} g) =
\one{h} \cdot_\Theta \one{g} \otimes \two{h} \cdot_\Theta \two{g}.
\end{align*}
\end{lem}
\begin{proof}
It suffices to work with homogeneous elements.
Take $h , g \in H$, with $\deg h = (r , l)$ and $\deg g = (p , q)$ with
their components in the coproduct, $\Delta x = \one{x} {\otimes} \two{x}$ in Sweedler
notation,
that can be assumed to be homogeneous as well:
\begin{align*}
\deg \one{h} = (r , s) , \quad
\deg \two{h} = (s , l) , \quad
\deg \one{g} = (p , k) , \quad
\deg \two{g} = (k , q) .
\end{align*}
where only $s, k$ vary within to the components.
Then,
\begin{align*}
\brac{ \one{h} \otimes \two{h} } \cdot_\Theta
\brac{ \one{g} \otimes \two{g} }
& =
\one{h} \cdot_\Theta \one{g} \otimes \two{h} \cdot_\Theta \two{g}
\\
& =
\lambda_\theta ( r, p )
\lambda_{-\theta} (s , k)
\lambda_{\theta} (s , k)
\lambda_{ - \theta} ( l, q)
\one{h} \one{g} \two{h} \two{g}
\\
& =
\lambda_{ - \theta} ( l, q)
\lambda_\theta ( r, p )
\one{h} \one{g} \two{h} \two{g} \\
& =
\lambda_{ - \theta} ( l, q)
\lambda_\theta ( r, p )
\Delta ( h g) \\
& =
\Delta ( h \cdot_\Theta g) ,
\end{align*}
as stated. \end{proof}
\begin{prop}
\label{prop:H-Theta}
By $\theta$-deforming the multiplication of $H$ as in \eqref{eq:Theta-hr-gl},
we obtain a new Hopf algebra
$H_\Theta = (H , \cdot_\Theta , \Delta , \varepsilon , S)$
with the same coproduct, counit and antipode.
\end{prop}
\begin{proof}
The compatibility between the algebra $ \cdot_\Theta$ and the coalgebra
$\Delta$ structures has been dealt with in Lemma \ref{lem:coprd-prd}.
The coproduct $\Delta$ and counit $\varepsilon$ are not deformed at all,
thus property
$( \varepsilon \otimes 1) \Delta = 1 = ( 1 \otimes \varepsilon ) \Delta$
remains. We are left to verify
\begin{align}
\label{eq:S-ve-2chk}
S ( \one{h}) \cdot_\Theta \two{h} = \varepsilon ( h ) =
\one{h} \cdot_\Theta S ( \two{h}) , \quad \forall h \in H.
\end{align}
Suppose $h$, $\one{h}$ and $\two{h}$ are homogeneous of degree $(r , l)$,
$(r , s)$ and $(s , l)$ respectively. By the assumptions in
\eqref{eq:antp-star-gd}, $S ( \one{h})$ is of degree $(-s , -r)$, thus
\begin{align*}
S ( \one{h}) \cdot_\Theta \two{h}
& =
\lambda_\theta (-s ,s) \lambda_{ - \theta } ( r , l)
S ( \one{h}) \cdot \two{h} =
\lambda_{ - \theta } ( r , l)
S ( \one{h}) \two{h} \\
& =
\lambda_{ - \theta } ( r , l)
\varepsilon ( h) = \varepsilon ( h ) .
\end{align*}
For the last step, we need to invoke \eqref{eq:counit-bigrd}, so that
$ \varepsilon (h) = 0$ whenever $l \neq r$, while for $r = l$, we have
$ \lambda_{ - \theta } ( r , l) = 1$.
\end{proof}
Next,
let $ \mathcal M^H$ be the category of $H$-comodule with a bi-grading of $\mathbb{Z}^n$
and such that the coaction $\delta^V : V \to V \otimes H$ with $V \in \mathcal M^H$,
behaves in a similar way to the coproduct in
\eqref{eq:coprd-bigrd} as regarding the grading:
\begin{align}
\label{eq:coact-bigrd}
\delta^V \brac{ V_{ (r , l)}} \subset \bigoplus_{ s \in \mathbb{Z}^n}
V_{ (r ,s) } \otimes H_{ ( s , l)} .
\end{align}
The co-representations $ \mathcal M^{H_\Theta}$ of $H_\Theta$,
keep the same objects and morphisms as $ \mathcal M^H$.
Modification only occurs on the coaction on the monoidal
structure. Namely, in the coaction on $V \otimes W$, where $V , W$ are in $ \mathcal M^H$, we must
use of the multiplication of $H_\Theta$:
\begin{align}
\label{eq:coact-HTheta}
\delta^{V \otimes_\Theta W}: V \otimes W \to V \otimes W \otimes H_\Theta,
\quad v \otimes w \mapsto
\zero{v} \otimes \zero{w} \otimes
\one{v} \cdot_\Theta \one{w} .
\end{align}
When deforming a comodule algebra $A$ in $ \mathcal M^H$, which play
the role of function algebra on the noncommutative principal bundle,
we have to impose similar conditions. That is, we have that
$A$ also admits a bi-grading of $\mathbb{Z}^n$ such that
\begin{enumerate}
\item the product of $A$ preserves the bi-grading as in \eqref{eq:prd-bigrd};
\item[]
\item the coaction $\delta^A : A \to A \otimes H$ satisfies
\eqref{eq:coact-bigrd} on the bi-grading.
\end{enumerate}
The first condition allows one to form the deformed algebra $A_\Theta$ and the
second one makes sure we still have a comodule algebra after deformation.
\begin{prop}
\label{prop:comodalg-Theta}
Consider a Hopf--Galois extension $(A , H)$ with both $H$ and $A$ endowed with a bi-grading of $\mathbb{Z}^n$, and algebra of coinvariants $B= A^{co H}$ (with a heredity bi-grading from $A$). Then,
their bi-grading leads to the deformed algebras
$H_\Theta = ( H , \cdot_\Theta )$ and $A_\Theta = ( A , \cdot_\Theta)$
according to \eqref{eq:Theta-hr-gl}.
Moreover, $A_\Theta$ is a $H_\Theta$-comodule algebra with the same coaction
viewed as a map $\delta^A_\Theta : A_\Theta \to A_\Theta \otimes H_\Theta$.
Also, the coinvariant subspace
$B_\Theta = A_\Theta^{\mathrm{co} H_\Theta } = (A^{\mathrm{co}H} , \cdot_\Theta) = (B, \cdot_\Theta)$
maintains its starting vector space sitting inside $A_\Theta$ as a subalgebra.
\end{prop}
\begin{proof}
Observe that, from \eqref{eq:coprd-bigrd} and \eqref{eq:coact-bigrd}, the coaction
$\delta^A$ and coproduct $\Delta$ of $H$ change the bi-grading in a similar manner,
hence compatibility between the coaction and multiplication of $A$ can be
proved along the lines of Lemma \ref{lem:coprd-prd}.
Since the coaction is taken directly from $(A , H)$, the coinvariant
subspace remains the same as a vectors space. Moreover,
the $\theta$-multiplication differs from the original one by a phase factor
on homogeneous elements, thus it maps $B \otimes B$ into $B$.
In other words, $ \cdot_\Theta$ can be restricted onto the coinvariant
subspace $B$ to form $B_\Theta$.
\end{proof}
\begin{exa}
\label{eg:Q-SO}
Let us specialize the discussion in Appendix \ref{sec:Qgrp} to the case
$G = \SO(2n)$ and $ \widetilde G = \SO ( 2n +1 ) $ and
discuss the bi-grading behind the quantum spheres in \eqref{sec:ev-q-shp}
in great detail.
The torus action $\alpha$ of $\IT^n \times \IT^n $ is now given by matrix multiplications
from two sides so that all the generators in \eqref{thetaCR} and \eqref{crN}
are eigenfunctions:
\begin{align}
\label{eq:eigfuns-eigvals}
\begin{split}
\alpha_{t, \tilde t} ( a_{ j k } ) = t_j \tilde t_k a_{ jk} ,
\, \, \, \,
\alpha_{t, \tilde t} ( b_{ j k } ) = t_j \tilde t_k^* b_{ jk} ,
\, \, \, \,
\alpha_{t, \tilde t} ( u_{ j } ) = t_j u_{ j} ,
\, \, \, \,
\alpha_{t, \tilde t} ( v_{ k } ) = \tilde t_k v_{ k} ,
\\
\alpha_{t, \tilde t} ( a^*_{ j k } ) = t_j^* \tilde t^*_k a_{ jk} ,
\, \, \, \,
\alpha_{t, \tilde t} ( b^*_{ j k } ) = t_j^* \tilde t_k^* b_{ jk} ,
\, \, \, \,
\alpha_{t, \tilde t} ( u^*_{ j } ) = t_j^* u_{ j} ,
\, \, \, \,
\alpha_{t, \tilde t} ( v^*_{ k } ) = \tilde t_k^* v_{ k} .
\end{split}
\end{align}
Therefore, we can reconstruct the algebras
$ H_\Theta = \mathcal O (\SO_\theta (2n))$ and
$ \widetilde H_\Theta = \mathcal O (\SO_\theta (2n + 1))$ by
assigning the following degrees to generators:
\begin{align}
\begin{split}
& \deg a_{ i j } = (e_i , e_j) = - \deg a^*_{ i j} ,
\, \, \, \,
\deg b_{ i j } = (e_i , - e_j) = - \deg b^*_{ i j},
\\
& \deg u_i = - \deg u_i^* = ( e_i, 0),
\, \, \, \,
\deg v_i = - \deg v_i^* = (0 , e_i) ,
\end{split}
\label{eq:gnt-degs}
\end{align}
where $\set{e_j, \, j =1, \cdots, n}$ is the standard basis of $\mathbb{Z}^n$,
and extents to the whole algebra according to \eqref{eq:prd-bigrd}.
For homogeneous elements, the new multiplication differs from the commutative one by
the phase factors as in \eqref{eq:Theta-hr-gl} instance,
\begin{align*}
a_{ i j } \cdot_\theta a_{ k l } =
\lambda_{\theta} ( e_i , e_k )
\lambda_{ - \theta } ( e_j , e_l )
a_{ i j } a_{ k l } =
\sqrt{ \lambda_{ i k} \lambda_{l j} }
a_{ i j } a_{ k l } ,
\end{align*}
and similarly,
$ a_{ i j } \cdot_\theta b_{ k l } =
\sqrt{ \lambda_{ i k} \lambda_{ j l} } b_{ i j } b_{ k l }$, while
for generators $\mathbf u$ and $ \mathbf v$ in \eqref{crN},
\begin{align*}
u_i \cdot_\theta u_ j =
\lambda_\theta (e_i , e_j) u_i u_j =
\sqrt{ \lambda_{ i j} } u_i u_j ,
\\
v_i \cdot_\theta v_ j =
\lambda_{-\theta} (e_i , e_j)v_i v_j =
\sqrt{ \lambda_{ j i} } v_i v_j .
\end{align*}
One recovers the commutation relations in \eqref{thetaCR} and \eqref{crN}
by taking \eqref{eq:com-rltn-A} into account.
Let us sample the assumptions
\eqref{eq:coprd-bigrd} - \ref{eq:antp-star-gd} on
some of generators. For the coproduct:
\begin{align*}
\Delta ( a_{ j l} ) =
\sum_{ s } a_{ j s} \otimes a_{ s l} + b_{ j s} \otimes b^*_{ s l} ,
\end{align*}
the right hand side indeed fulfils
$ a_{ j s} \otimes a_{ s l} \in H_{ (j ,s)} \otimes H_{ ( s , l)}$
and
$ b_{ j s} \otimes b^*_{ s l} \in H_{ ( j , -s)} \otimes H_{ (-s , l)}$.
For the counit $\varepsilon$ defined by $ \varepsilon (N) = \mathbb I$,
only the diagonal entries of $N$ will survive after applying $\varepsilon$
and they indeed belong to
$\bigoplus_{ s \in \mathbb{Z}^n } H_{ (s,s)}$ as required in \eqref{eq:counit-bigrd}.
For the $*$-operator $*N = Q N Q$, we see, for instance,
that $ \deg ( b_{ i j} )^* = \deg b^*_{ i j} = - \deg b_{ i j} $.
For the antipode $ S (N) = N^{\dagger}$, we would like to check on, say $u_j$:
$\deg u_j = ( e_j , 0)$ compared with
$ \deg S ( u_j ) = \deg v^*_j = (0, - e_j)$.
Lastly, the analysis of \eqref{eq:coact-bigrd} for the coaction $\delta^{ \widetilde H}$ is just the same as that for the coproduct $\Delta$.
\qed\end{exa}
\begin{exa}
The matrix representation of $\SO(2n)$ in \eqref{eq:mtxM} and \eqref{idealM}
require an extra structure on
$\mathbb{R}^{2n}$, that is a choice of polarization. Concretely, one identifies
$\mathbb{R}^{2n} \cong \mathbb{C}^n$ and choose a basis formed by complex
coordinates $\set{z_j, \bar z_j, \, j = 1 \cdots n}$, with respect to which
the coefficient matrix of the Euclidean inner product is of the form $Q$ in
\eqref{idealM}. We recall a remark made in \cite[\S 8]{Connes:2002wh}
which further motivates the bigrading setting of our scenario II in connection with
the grading in scenario I.
At the level of the endomorphisms $M(2n,\mathbb{R})$,
the identification is achieved by realising
$ M(2n, \mathbb{R}) \subset \mathrm{End} ( \mathbb{C}^n)
\cong (\mathbb{C}^n)^* \otimes \mathbb{C}^n$.
One first applies the $\theta$-deformation to
$(\mathbb{C}^n)^* \otimes \mathbb{C}^n$ following the setting in scenario I,
which gives rise to two deformed algebra
$A( \mathbb{R}^{2n}_{ \theta}) = A(\mathbb{C}^n_{ \theta}) $
and $A(\mathbb{R}^{2n}_{ -\theta}) = A(\mathbb{C}^n_{ -\theta})$, with
generators
$\set{z^j , \bar z^j := (z^j)^* , \, j=1, \cdots, n}$ for $ A(\mathbb{R}^{2n}_{ \theta})$ and
$\set{z_j , \bar z_j := (z_j)^*, \, j=1, \cdots, n}$ for $ A(\mathbb{R}^{2n}_{ -\theta})$.
The $n$-torus action $\alpha$ is the standard one:
\begin{align}
\label{eq:z_j-actn}
\alpha_t ( z_j ) = t_j z_j , \, \,
\alpha_t ( \bar z_j ) = \bar t_j \bar z_j , \qquad
\alpha_t ( z^j ) = t_j z^j , \, \,
\alpha_t ( \bar z^j ) = \bar t_j \bar z^j , \, \,
\end{align}
which leads to the $\mathbb{Z}^n$-grading:
\begin{align}
\label{eq:z_j-deg}
e_j = \deg z^j = \deg z_j = - \deg \bar z^j = - \deg \bar z_j,
\end{align}
where $\set{e_j, \, j=1, \cdots, n}$ is the standard basis of $\mathbb{Z}^n$.
The algebra structure of $A(\mathbb{R}^{2n}_{ \theta })$ is determined by
the commutation relations:
\begin{align*}
z_j z_k = \lambda_{ j k} z_k z_j, \, \quad
\bar z_j z_k = \lambda_{ k j} z_k \bar z_j .
\end{align*}
Those for $A(\mathbb{R}^{2n}_{ - \theta })$ are obtained by replacing
$\lambda_{ j k}$ with $ \bar \lambda_{ j k}$.
Now, the deformed $*$-algebra
$\mathcal O(M_\theta (2n , \mathbb{R}))$
(One can forget the coalgebra structure for the time being.)
has already been defined in \eqref{eq:mtxM} and \eqref{thetaCR}
in terms of generators and relations.
A key point is that there is
a $*$-algebra homomorphism $\varphi: \mathcal O(M_\theta (2n , \mathbb{R})) \to
A(\mathbb{R}^{2n}_{ \theta}) \otimes A(\mathbb{R}^{2n}_{ - \theta})$
induced by
\begin{align*}
\varphi (a_{ i j}) = z^i \otimes z_j , \, \quad
\varphi (b_{ i j}) = z^i \otimes \bar z_j .
\end{align*}
Furthermore, the map $\varphi$ is injective and transfers the torus action
$\alpha \otimes \alpha $ (cf. \eqref{eq:z_j-actn}),
or equivalently, the bi-grading of
$ A(\mathbb{R}^{2n}_{ \theta}) \otimes A(\mathbb{R}^{2n}_{ -\theta})$
(cf. \eqref{eq:z_j-deg}), to those described in the Example \ref{eg:Q-SO}: see
\eqref{eq:eigfuns-eigvals} and \eqref{eq:gnt-degs}.
\qed
\end{exa}
Let us now take a closer look at the algebra of coinvariants and at the balanced product.
\begin{lem} \label{lem:balancedts-deg}
With the assumptions on $H$ and $A$ as before,
the coinvariant subalgebra $B = A^{ \mathrm{co} H}$
is contained in
\begin{align}
\label{eq:B-deg}
B \subset \bigoplus_{ r \in \mathbb{Z}^n } A_{ (r , 0) } .
\end{align}
\end{lem}
\begin{proof}
The (algebra of functions on) the torus $\IT^n$ acting on the right is contained in $H$ and gets washed away when passing to the coinvariant elements for the coaction of $H$.
Explicitly, consider a homogeneous element $b \in B$ with $\deg b = (r, l)$.
From \eqref{eq:coact-bigrd} we have
$\deg b_{ (0)} = (r , s) $ and $\deg b_{ (1)} = (s , l)$ where $s \in \mathbb{Z}^n$
depends on the components.
The condition of being coinvariant
$ b_{ (0)} \otimes b_{ (1)} = b \otimes 1 $ forces that
$(r, s ) = ( r , l)$ and $ (s, l) = (0 , 0)$, hence $\deg b$ is always
of the form $(r , 0)$ for some $r \in \mathbb{Z}^n$.
\end{proof}
The $2n$-sphere $B = \mathcal O (S^{2n})$ and its deformation
in \eqref{eg:Q-SO} indeed satisfy \eqref{eq:B-deg}:
the generators of $B$ (or $B_\Theta$) are
$\set{u_j , u^*_j , \, j = 1, \cdots, n}$ which are of degree $(\pm e_j ,0)$.
When forming the balanced tensor product
$a \otimes_B \tilde a$, where $a , \tilde a \in A$,
the degrees of $a $ and $ \tilde a$ (assumed to be homogeneous)
depend on the choice of the representative. However, from the previous lemma,
the action of $B$ only varies the left degree.
As we shall see in next lemma,
by slightly abusing the notation, for the translation map we can write
\begin{align}
\label{eq:tau-bigrd-Theta}
\tau \brac{ H_{ (r , l)} } \subset \bigoplus_{ p \in \mathbb{Z}^n}
A_{ ( - p , - r)} {\otimes}_B A_{ (p , l)} .
\end{align}
Let us check this on $\tau (\mathbf h)$ and $\tau (\mathbf k)$ in \eqref{eq:tranmap-h}.
For the $(r,l)$-entry of $\mathbf h$, we have
\begin{align}\label{tauhexp}
\tau ( h_{ r l} )
& =
\sum_{ s } (\mathbf a)_{ r s}^{\dagger} {\otimes}_B \mathbf a_{ s l} +
(\mathbf b^*)_{ r s}^{\dagger} {\otimes}_B (\mathbf b^*)_{ s l} +
(\mathbf v)^{\dagger}_r {\otimes}_B \mathbf v_l
\nonumber \\
& =
\sum_{ s }
a^*_{ s r} {\otimes}_B a_{ s l } +
b_{ s r} {\otimes}_B b^{*}_{ s l} +
v^*_{ r } {\otimes}_B v_l .
\end{align}
We see that
$ a^*_{ s r} \otimes a_{ s l }
\in \widetilde H_{ (-e_s ,-e_r)} \otimes \widetilde H_{ ( e_s ,e_l )} $,
$ b_{ s r} \otimes b^{*}_{ s l}
\in \widetilde H_{ (e_s ,-e_r)} \otimes \widetilde H_{ ( -e_s ,e_l )} $
as well as
$ v^*_{ r } \otimes v_l
\in \widetilde H_{ (0 ,-e_r)} \otimes \widetilde H_{ ( 0 ,e_l )} $
all satisfy \eqref{eq:tau-bigrd}.
Similarly, for the $(r,l)$-entry of $\mathbf k$, we have
\begin{align}\label{taukexp}
\tau ( k_{ r l} )
& =
\sum_{ s } (\mathbf a)_{ r s}^{\dagger} {\otimes}_B \mathbf b_{ s l} +
(\mathbf b^*)_{ r s}^{\dagger} {\otimes}_B (\mathbf a^*)_{ s l} +
(\mathbf v)^{\dagger}_r {\otimes}_B \mathbf v^*_l
\nonumber \\
& =
\sum_{ s }
a^*_{ s r} {\otimes}_B b_{ s l } +
b_{ s r} {\otimes}_B a^{*}_{ s l} +
v^*_{ r } {\otimes}_B v^*_l .
\end{align}
We see that
$ a^*_{ s r} \otimes b_{ s l }
\in \widetilde H_{ (-e_s ,-e_r)} \otimes \widetilde H_{ ( e_s , -e_l )} $,
$ b_{ s r} \otimes a^{*}_{ s l}
\in \widetilde H_{ (e_s ,-e_r)} \otimes \widetilde H_{ ( -e_s , -e_l )} $
as well as
$ v^*_{ r } \otimes v^*_l
\in \widetilde H_{ (0 ,-e_r)} \otimes \widetilde H_{ ( 0 , -e_l )} $
and again they all satisfy \eqref{eq:tau-bigrd}.
We also have
\begin{align*}
m_\Theta ( \tau (\mathbf h )) & =
\mathbf a^{\dagger} \cdot_\Theta \mathbf a +
\mathbf b^{t} \cdot_\Theta \mathbf b^* +
\mathbf v^{\dagger} \cdot_\Theta \mathbf v = \varepsilon( \mathbf h ) \mathbb{I} = \mathbb{I}, \\
m_\Theta ( \tau (\mathbf k )) & =
\mathbf a^{\dagger} \cdot_\Theta \mathbf b +
\mathbf b^{t} \cdot_\Theta \mathbf a^* +
\mathbf v^{\dagger} \cdot_\Theta \mathbf v^* = \varepsilon( \mathbf k ) \mathbb{I} = 0
\end{align*}
and both agree with \eqref{p5} and Corollary \ref{cor:tau-bigrd} below.
\begin{lem} \label{lem:balancedts-deg-2}
Let $(A , H)$ be a Hopf--Galois extension fulfilling
all assumptions of earlier.
For any homogeneous elements $h \in H_{ (r , l)} $, there are suitable
representatives for the translation map $\tau ( h ) = h^{\abrac{ 1 }} {\otimes}_B h^{\abrac{ 2 }}$
such that
\begin{equation}
\label{eq:tau-bigrd}
\deg h^{\abrac{ 1 }} = ( -p , - r ) , \, \, \, \,
\deg h^{\abrac{ 2 }} = ( p , l) ,
\end{equation}
where, by taking \eqref{eq:B-deg} into account, the left degree $p$ depends on
the components $ h^{\abrac{ 1 }}, h^{\abrac{ 2 }}$ and the choice
of the representatives (cf. also Remark \ref{rem-dtm}).
\end{lem}
\begin{proof}
The constraint on degrees in \eqref{eq:tau-bigrd} follows from \eqref{p7}:
$ h^{\abrac{ 1 }} (h^{\abrac{ 2 }})_{ (0)}
\otimes (h^{\abrac{ 2 }})_{ (1)} = 1 \otimes h $.
Suppose $\deg h = (r , l)$, $ \deg h^{\abrac{ 2 }} = (p ,q)$
and $\deg h^{\abrac{ 1 }} = (p' ,q')$, so that
$ \deg (h^{\abrac{ 2 }})_{ (0)} = ( p ,s)$ and
$\deg (h^{\abrac{ 2 }})_{ (1)} = (s , q)$ for some $s \in \mathbb{Z}^n$.
By comparing the two sides of \eqref{p7}, we have
$ ( s , q) = (r , l)$ and $ (p' , q') = - ( p , s)$. Thus
\eqref{eq:tau-bigrd} follows:
$q = l$, $p' = - p$ and $q' = -s = -r$.
\end{proof}
\begin{cor}
\label{cor:tau-bigrd}
Let $ h \in H_{ (r , l)} $ be a homogeneous element with $\tau ( h ) = h^{\abrac{ 1 }} {\otimes}_B h^{\abrac{ 2 }}$. Then,
\begin{itemize}
\item[i)] for $r = l $ one has $\deg h^{\abrac{1 }} + \deg h^{ \abrac{ 2 }} = 0$;
\item[]
\item[ii)] for $ r \neq l$, one has $\tau ( h) = 0$.
\end{itemize}
\end{cor}
\begin{proof}
With $r = l$, the first statement follows from \eqref{eq:tau-bigrd}.
The latter also says that $ h^{\abrac{1 }} h^{\abrac{ 2 }} \in A_{ (0 , l - r) }$,
which is non-zero unless $ h^{\abrac{1 }} \otimes h^{\abrac{ 2 }} = 0$.
\end{proof}
This result is in accordance with \eqref{tauhexp} and \eqref{taukexp} by recalling that $
h_{jl}$ has bi-degree $((e_j, 0), (e_l, 0))$ while $k_{jl}$ has bi-degree $((e_j, 0), (-e_l, 0))$.
It allows one to repeat the second part of Proposition \ref{prop:comodalg-theta}
and deform the starting Hopf--Galois extension into a new one.
\begin{prop}\label{prop:H-G-Theta}
Consider the deformed pair $(H_\Theta , A_\Theta)$ obtained in {\rm Proposition
\ref{prop:comodalg-Theta}} and define a deformed canonical Galois map $\chi_\Theta$ by
\begin{align}
\label{eq:chi-Theta}
\chi_\Theta: A_\Theta \otimes_{ B_\Theta } A_\Theta
\to A_\Theta \otimes H_\Theta , \quad
a' \otimes_{ B_\Theta} a \mapsto a' \cdot_\Theta a_{ (0)} \otimes a_{ (1)} .
\end{align}
This is invertible if and only if the starting canonical Galois map is with
the same translation map, but viewed as a map $\tau : H_\Theta \to A_\Theta \otimes_{ B_\Theta } A_\Theta$.
\end{prop}
\begin{proof}
Let $ h \in H_{ (r , l)} $ be a homogeneous element with $\tau ( h ) = h^{\abrac{ 1 }} {\otimes}_B h^{\abrac{ 2 }}$,
the starting translation map.
From Corollary \ref{cor:tau-bigrd}, $\deg h^{\abrac{1 }} = - \deg h^{ \abrac{ 2 }}$. Then, with a slight abuse of notation
($B = B_\theta$ as a vector space), from Lemma \ref{lem:mtheta-m},
\begin{align*}
\chi_\Theta (\tuno{h} {\otimes}_B \tdue{h}) & = \tuno{h} \cdot_\Theta \zero{\tdue{h}} {\otimes} \one{\tdue{h}} =
\tuno{h} \zero{\tdue{h}} {\otimes}_B \one{\tdue{h}} \\
& = \chi(\tuno{h}{\otimes}_B \tdue{h})
\end{align*}
and $\chi_\Theta$ is invertible if and only if $\chi$ is. Or, the pair $(H_\Theta , A_\Theta)$
is a Hopf--Galois extension if and only if the
pair $(H, A)$ is such.
\end{proof}
\section{Hopf algebroids} \label{se:had}
We are ready for the Hopf algebroid structure.
We start with a bialgebroid $ \mathcal{C} ( H_\Theta , A_\Theta)$ associated to the
Hopf--Galois extension $(H_\Theta, A_\Theta)$ of the previous section.
We next show that the flip can serve as an antipode.
In \S\S ~\ref{sec:ev-q-shp} and \ref{se:alg-su2sym}, we present two examples.
Firstly an algebroid for the principal $\SU(2)$-principal bundle
over the four-sphere $S_\theta^4$ decribed in \S\ref{sec:-q-shp} followed
by the one for the bundles over the even spheres of \S\ref{sec:HG-SU2}.
\subsection{The bialgebroid $ \mathcal{C} ( H_\Theta , A_\Theta)$} \label{subsec:ES-alg}
We know from Lemma \ref{lem:balancedts-deg} that the coinvariant elements for the action of
$H_\Theta$ have trivial right grading. This will clearly be the case also for the coinvariant elements
for the diagonal action that is needed for the Ehresmann--Schauenburg bialgebroid
Thus the construction of the bialgebroid will be the same for the Hopf--Galois extension $(H_\Theta, A_\Theta)$
in Proposition \ref{prop:H-G-Theta} of our scenario II and for the pair $( H , A_\theta)$
discussed for Scenario I in Proposition \ref{prop:comodalg-theta}. We describe the former here.
Consider then the Hopf--Galois extension $(H_\Theta, A_\Theta)$. The diagonal coaction
is in \eqref{eq:coact-HTheta}:
\begin{equation}\label{eq:coact-Alg}
\delta^{ A \otimes_\Theta A}:
A_\Theta \otimes A_\Theta \to A_\Theta \otimes A_\Theta \otimes H_\Theta ,
\quad
\delta^{ A \otimes_\Theta A} (a{\otimes} \tilde a) =
\zero{a} \otimes \zero{\tilde a} \otimes
\one{a} \cdot_\Theta \one{\tilde a}.
\end{equation}
with $V = W = A$.
From the analysis before, and in particular from the fact that the canonical map and translation
maps are the same as maps between vector spaces, the conclusion is that all the structure
equations listed in
\S\ref{sec:ES-bialerd} hold true after deformation (which means
replacing every occurrence of multiplication by the deformed one).
\begin{lem}
\label{lem:ES-ald}
Let $(H, A)$ be a Hopf--Galois extension that fulfil the assumptions
on the bigradings of earlier and let $(H_\Theta , A_\Theta)$ be
the deformed $(H_\Theta , A_\Theta)$ Hopf--Galois extension obtained in
Proposition \ref{eq:chi-Theta}.
Then the deformed coaction $\delta^{ A \otimes_\Theta A}$ in \eqref{eq:coact-Alg},
gives rise to the same coinvariant subspace
as that of $\delta^{A \otimes A}$:
\begin{align*}
\brac{ A_\Theta \otimes A_\Theta }^{\mathrm{co} H_\Theta }
= \brac{ A \otimes A }^{\mathrm{co} H} .
\end{align*}
Also, the deformed Ehresmann--Schauenburg bialgebroid
\begin{align*}
\mathcal{C} ( H_\Theta , A_\Theta)
= \brac{ \brac{ A_\Theta \otimes A_\Theta }^{\mathrm{co} H_\Theta } ,
\bullet_\Theta }
\end{align*}
with respect to $\mathcal{C} (A , H)$ in Def. \ref{def:reb}), has only the algebra structure changed, given by:
\begin{align}
\label{eq:ES-m-Theta}
(x \otimes y) \bullet_\Theta ( \tilde x \otimes \tilde y) :=
x \cdot_\theta \tilde x \otimes \tilde y \cdot_\theta y \, .
\end{align}
\end{lem}
\begin{proof}
The results follows from the identification in Lemma \ref{lem-2vers}, which uses only the translation map that is unchanged (as a map between vector spaces) when deforming.
\end{proof}
\subsection{The flip map as the antipode} \label{subsec:flip-atpd}
The bialgebroids of the previous section gets in fact a structure of Hopf algebroid with a suitable antipode.
Now, when the structure Hopf algebra $H$ is commutative, the flip map
preserves the coinvariant elements of the diagonal coaction. Indeed, given
\begin{align}
\label{eq:flip-S}
S: A \otimes A \to A \otimes A , \quad
a \otimes \tilde a \mapsto
\tilde a \otimes a ,
\end{align}
for any coinvariant $a \otimes \tilde a \in A \otimes A$,
by swapping $a$ and $ \tilde a$ in
$ a \otimes \tilde a \otimes 1 =
a_{ (0)} \otimes \tilde a_{ (0)} \otimes a_{ (1)} \tilde a_{ (1)} $,
we see that $ \tilde a \otimes a$ is coinvariant as well:
$$
\tilde a \otimes a \otimes 1 =
\tilde a_{ (0)} \otimes a_{ (0)} \otimes a_{ (1)} \tilde a_{ (1)}
=
\tilde a_{ (0)} \otimes a_{ (0)} \otimes \tilde a_{ (1)} a_{ (1)} ,
$$
where the last equal sign invokes the commutativity of $H$. Therefore,
when restricted to the coinvariant subspaces the flip is a candidate for the antipode of
$\mathcal{C}( H, A)$ and $\mathcal{C} ( H, A_\theta)$.
In the more general situation, despite $H_\Theta$ needs no longer stay commutative after the $\theta$-deformation,
the flip $S$ still maps $\mathcal{C} ( H_\Theta , A_\Theta)$ into itself since we have shown in Lemma \ref{lem:ES-ald}
that $\mathcal{C} ( H_\Theta , A_\Theta)$ and $ \mathcal{C} (H, A)$ are identical as vector spaces
(This fact will be explicitly seen for the example in \S\ref{sec:ev-q-shp} below.)
The main result of this section is that the flip $S$ makes
$\mathcal{C} ( H_\Theta , A_\Theta)$ into a Hopf algebroid.
\begin{thm}
By only deforming multiplication related structures of the Hopf
algebroid $\mathcal{C}( H, A)$ over $B$, the resulting $\mathcal{C} ( H_\Theta , A_\Theta)$
is a Hopf algebroid, but with base algebra $B_\Theta$.
\end{thm}
\begin{proof}
One needs to verify the compatibility conditions in
\eqref{hopbroid1} and \eqref{hopbroid2}. The latter one is the
less nontrivial one and is handled in Lemma \ref{lem:anti-comp-I} below.
We point out that the computations below work for
both $\mathcal{C}( H, A)$ and $\mathcal{C} ( H_\Theta , A_\Theta)$ since they do not rely
on the commutativity of the underlying algebra structures in the Hopf--Galois
extension.
\end{proof}
\begin{lem}
\label{lem:anti-comp-I}
The flip $ S : \mathcal{C} ( H_\Theta , A_\Theta) \to \mathcal{C} ( H_\Theta , A_\Theta)$
with $S^{-1} = S $ fulfils
the compatibility conditions in \eqref{hopbroid2},
that is, for all $ h \in \mathcal{C} ( H_\Theta , A_\Theta)$:
\begin{align*}
(S^{-1} h_{ (2)} )_{ (1)} \otimes_{B_\Theta}
(S^{-1} h_{ (2)} )_{ (2)} \bullet_\Theta h_{ (1)}
&=
S^{-1} h \otimes_{B_\Theta} 1 ,
\\
(S h_{ (1)} )_{ (1)} \bullet_\Theta h_{ (2)} \otimes_{B_\Theta}
(S h_{ (1)} )_{ (2)}
&=
1 \otimes_{B_\Theta} S(h) .
\end{align*}
\end{lem}
\begin{proof}
We shall prove the first one as an example and leave the second one to avid readers.
Write $h = a \otimes \tilde a \in \mathcal{C} ( H_\Theta , A_\Theta)$,
where $a, \tilde a \in A_\Theta$, then
the coproduct in \eqref{copro} reads
\begin{align*}
\Delta (h) = h_{ (1)} \otimes_{B_\Theta} h_{ (2)} =
\brac{ a_{ (0)} \otimes (a_{ (1)})^{\abrac{ 1 } } }
\otimes_{B_\Theta} \brac{ (a_{ (1)})^{\abrac{ 2 } } \otimes \tilde a } .
\end{align*}
We compute:
\begin{align*}
(S^{-1} h_{ (2)} )_{ (1)}\otimes_{B_\Theta} &
(S^{-1} h_{ (2)} )_{ (2)} \bullet_\Theta h_{ (1)}
\\ &=
\brac{ \tilde a \otimes (a_{ (1)})^{\abrac{ 2 } } }_{ (1)}
\otimes_{B_\Theta}
\brac{ \tilde a \otimes (a_{ (1)})^{\abrac{ 2 } } }_{ (2)}
\bullet_\Theta
\brac{ a_{ (0)} \otimes (a_{ (1)})^{\abrac{ 1 } } }
\\ &=
\tilde a_{ (0)} \otimes (\tilde a_{ (1)})^{\abrac{ 1 }}
\otimes_{B_\Theta}
\brac{ ( \tilde a_{ (1)} )^{\abrac{ 2 }} \otimes (a_{ (1)})^{\abrac{ 2 } } }
\bullet_\Theta
\brac{ a_{ (0)} \otimes (a_{ (1)})^{\abrac{ 1 } } }
\\ &=
\tilde a_{ (0)} \otimes (\tilde a_{ (1)})^{\abrac{ 1 }}
\otimes_{B_\Theta}
( \tilde a_{ (1)} )^{\abrac{ 2 }} \cdot_\Theta a_{ (0)} \otimes
(a_{ (1)})^{\abrac{ 1 } } \cdot_\Theta (a_{ (1)})^{\abrac{ 2 } }
\\ &=
\tilde a_{ (0)} \otimes (\tilde a_{ (1)})^{\abrac{ 1 }}
\otimes_{B_\Theta}
( \tilde a_{ (1)} )^{\abrac{ 2 }} \cdot_\Theta a_{ (0)} \otimes
\varepsilon ( a_{ (1)} ) 1_{ A_\Theta}
\\ &=
\tilde a_{ (0)} \otimes (\tilde a_{ (1)})^{\abrac{ 1 }}
\otimes_{B_\Theta}
( \tilde a_{ (1)} )^{\abrac{ 2 }} \cdot_\Theta a \otimes 1 ,
\end{align*}
where, in the last two steps,
we have used \eqref{p5} and the compatibility between the counit
$ \varepsilon: H_\Theta \to \mathbb{C}$ and the coaction $\delta^A$.
To continue:
\begin{align*}
(S^{-1} h_{ (2)} )_{ (1)} \otimes_{B_\Theta}
(S^{-1} h_{ (2)} )_{ (2)} \bullet_\Theta h_{ (1)}
&=
\brac{
\tilde a_{ (0)} \otimes (\tilde a_{ (1)})^{\abrac{ 1 }}
} \otimes_{B_\Theta}
\brac{
( \tilde a_{ (1)} )^{\abrac{ 2 }} \bullet_\Theta a \otimes 1
}
\\ & =
\tilde a_{ (0)} \otimes \tau \brac{ \tilde a_{ (1)} } \cdot_\Theta a
\\ & =
\tilde a \otimes a \otimes_{ B_\Theta } 1
= S^{-1} h \otimes_{ B_\Theta } 1 ,
\end{align*}
where we need \eqref{ec1}, which is an equivalent description for
$a \otimes \tilde a \in \mathcal{C} ( H_\Theta , A_\Theta)$,
to complete the second line.
\end{proof}
\subsection{The algebroid with $\SU(2)$-symmetry}\label{se:alg-su2sym}
With respect to the example in \S\ref{sec:HG-SU2},
denote $A=A(S_{\theta'}^7)$, $H=A(\SU(2))$ and $B=A(S_\theta^4) = A^{co H}$ the subalgebra of invariants and,
as usual $\delta^A(a) = \zero{a} \stackrel{.}{\otimes} \one{a}$ and $\tau(h) = \tuno{h} {\otimes}_B \tdue{h}$.
Consider then the diagonal coaction of $H$ on the tensor product algebra $A{\otimes} A$:
$$
\delta^{A{\otimes} A}: A{\otimes} A \to A{\otimes} A{\otimes} H, \quad a{\otimes} \tilde{a} \mapsto
\zero{a}{\otimes} \zero{\tilde{a}} {\otimes} \one{a}\one{\tilde{a}} \, .
$$
\begin{lem}
The $B$-bimodule $\mathcal{C}(A,H)$ of coinvariant elements for the diagonal coaction is generated
by elements of the tensor products $p {\otimes} 1$ and $1{\otimes} {q}$ together with
$$
V = \Psi \stackrel{.}{\otimes} \Psi^\dagger.
$$
\begin{proof}
It is clear that elements of $p {\otimes} 1$ and $1{\otimes} {q}$ are coinvariants. For $V = \Psi \stackrel{.}{\otimes} \Psi^\dagger$:
$$
\delta^{A {\otimes} A}(V) = \zero{\Psi} \stackrel{.}{\otimes} \zero{\Psi^\dagger} \stackrel{.}{\otimes} \one{\Psi} \one{\Psi^\dagger}
= \Psi {\otimes} \Psi^\dagger \stackrel{.}{\otimes} (w w^\dagger)= V \stackrel{.}{\otimes} (w w^\dagger) = V \stackrel{.}{\otimes} \mathbb{I}_2 ,
$$
in parallel with the coinvariance \eqref{once}.
\end{proof}
\end{lem}
With the flip $\sigma(a {\otimes} b) = b {\otimes} a$ we define
\begin{equation}\label{antipode1case}
S_\mathcal{C}(V) := \sigma(\Psi \stackrel{.}{\otimes} \Psi^\dagger) = V^\dagger, \quad \textup{or} \quad S_\mathcal{C}(V_{mn})
= V^\dagger_{mn} = \sum_r \Psi^\dagger_{rn} \stackrel{.}{\otimes} \Psi_{mr}.
\end{equation}
Then, a direct computation shows that
\begin{align} \label{alter}
S_\mathcal{C}(V) V &= V^\dagger V = 1 {\otimes} \, \Psi \cdot_{op} \Psi^\dagger = 1 {\otimes} \, {q} \nonumber \\
V S_\mathcal{C}(V) &= V V^\dagger = \Psi \cdot \Psi^\dagger \, {\otimes} 1 = p \, {\otimes} 1 .
\end{align}
These then are relations among the elements of $V$ as generators of the
$B$-bimodule $\mathcal{C}(A,H)$. The latter has the structure of a Hopf algebroid.
Firstly,
the projections $V V^\dagger = p {\otimes} 1$ and $V^\dagger V= 1{\otimes} {q}$ are the two embedded copies of the
4-sphere $A(S_\theta^4)$ in $\mathcal{C}(A,H)$: $A(S_\theta^4) {\otimes} 1$ and $1 {\otimes} A(S_\theta^4)$, via source and target map respectively,
as explicitly described in Lemma \ref{source-target} below.
Next,
according to the definition \eqref{copro}
a coproduct $\Delta : \mathcal{C}(A,H) \to \mathcal{C}(A,H) {\otimes}_B \mathcal{C}(A,H)$, is given on the matrix $V$ of generators by
\begin{align} \label{ccop}
\Delta (V) = \zero{\Psi} \stackrel{.}{\otimes} \tuno{\one{\Psi}} \stackrel{.}{\otimes} _B \tdue{\one{\Psi}} \stackrel{.}{\otimes} \Psi^\dagger
= \Psi {\otimes} \Psi^\dagger \stackrel{.}{\otimes}_B \Psi {\otimes} \Psi^\dagger = V \stackrel{.}{\otimes} _B V.
\end{align}
In components this reads:
\begin{equation}\label{copcom}
\Delta(V_{m n}) = \sum_r \, V_{m r} {\otimes}_B V_{r n}.
\end{equation}
From \eqref{ccop} one also gets
\begin{equation}\label{ccops}
\Delta(S_\mathcal{C}(V)) = \Delta(V^\dagger) = \sigma(V^\dagger \stackrel{.}{\otimes} _B V^\dagger)
\end{equation}
or, in components:
\begin{equation}\label{copcomop}
\Delta(V^\dagger_{mr}) = \sum_r \, V^\dagger_{r n} {\otimes}_B V^\dagger_{m r} .
\end{equation}
Finally, the map $S_\mathcal{C}$ in \eqref{antipode1case} is indeed an antipode for $\mathcal{C}(A,H)$.
Since $S_\mathcal{C}$ is the flip, condition \eqref{hopbroid1} is obvious.
We are left to show condition \eqref{hopbroid2}. For this, take $h=V$.
With expressions \eqref{ccop} and \eqref{ccops} for the coproducts, and using \eqref{alter}:
\begin{align} \label{sverif}
\onet {(S_\mathcal{C}\one{h})} \two{h} {\otimes}_B \twot {S_\mathcal{C}(\one{h})}
& = \onet {(S(\one{V}))} \, \two{V} \stackrel{.}{\otimes}_B \twot {(S(\one{V}))} \nonumber \\
& = \onet {(S_\mathcal{C}(V))} \, V \stackrel{.}{\otimes}_B \twot {(S_\mathcal{C}(V))} \nonumber \\
& = \onet{(V^\dagger)} \, V \stackrel{.}{\otimes}_B \twot{(V^\dagger)}
= V^\dagger \, V \stackrel{.}{\otimes}_B V^\dagger \nonumber \\
& = 1 {\otimes} {q} \stackrel{.}{\otimes}_B \, V^\dagger = 1 {\otimes} 1 {\otimes}_B \, {q} \, V^\dagger \nonumber \\
& = 1 {\otimes} 1 {\otimes}_B V^\dagger = 1 {\otimes} 1 {\otimes}_B S_\mathcal{C}(V).
\end{align}
Since elements of ${q}$ are in $B$ they can be crossed over $B$-tensor products, and we used the relation
$q \,V^\dagger = V^\dagger$.
The other condition in \eqref{hopbroid2} is similar since $S_\mathcal{C}^{-1} = S_\mathcal{C}$.
\noindent
In components of $V$ this works as follows.
Take $h=V_{mn}$, with $S_\mathcal{C}(h) = V^\dagger_{mn}$. Using the expressions for the coproduct, we compute:
\begin{align}\label{sverif-bis}
\onet {(S_\mathcal{C}\one{h})} \two{h} {\otimes}_B \twot {S_\mathcal{C}(\one{h})}
& = \sum_{r} \onet {(S_\mathcal{C}(V)_{m r})} V_{rn} {\otimes}_B \twot {(S_\mathcal{C}(V)_{mr})} \nonumber \\
& = \sum_{r} \onet {(V^\dagger_{m r})} V_{rn} {\otimes}_B \twot {(V^\dagger_{m r})} \nonumber \\
& = \sum_{rs} V^\dagger_{s r} V_{rn} {\otimes}_B V^\dagger_{ms} = \sum_{s} (V^\dagger V)_{s n} {\otimes}_B V^\dagger_{ms} \nonumber \\
& = \sum_{s} 1 {\otimes} (\Psi \cdot_{op} \Psi^\dagger)_{s n}{\otimes}_B V^\dagger_{ms} = \sum_{s} 1 {\otimes} 1 {\otimes}_B (\Psi \cdot_{op} \Psi^\dagger)_{s n} V^\dagger_{ms}
\nonumber \\
& = \sum_{s, j, k} 1 {\otimes} 1 {\otimes}_B \, \Psi^\dagger_{j n} \Psi_{s j} \Psi^\dagger_{ks} {\otimes} \Psi_{mk} \nonumber \\
& = \sum_{s, j,k} 1 {\otimes} 1 {\otimes}_B \, \Psi^\dagger_{j n} (\Psi^\dagger_{ks} \cdot_{op} \Psi_{s j} ) {\otimes} \Psi_{mk} \nonumber \\
& = \sum_{j,k} 1 {\otimes} 1 {\otimes}_B \, \Psi^\dagger_{j n} ( \delta_{k j} 1 ) {\otimes} \Psi_{mk}
= \sum_{k} 1 {\otimes} 1 {\otimes}_B \, \Psi^\dagger_{k n} {\otimes} \Psi_{mk} \nonumber \\
& = 1 {\otimes} 1 {\otimes}_B V^\dagger_{mn}
= 1 {\otimes} 1 {\otimes}_B S_\mathcal{C}(V)_{mn} = 1 {\otimes} 1 {\otimes}_B S_\mathcal{C}(h) .
\end{align}
Here we crossed elements of $(\Psi \cdot_{op} \Psi^\dagger)$
over the $B$-tensor product, since they are in $B$, and the relation
$ \sum_{s} \Psi^\dagger_{ks} \cdot_{op} \Psi_{s j} = \delta_{k j} 1$.
\subsubsection{Generators and relations}
In term of generators and relations let us write
\begin{align}\label{4matrices}
V = \begin{pmatrix}
P_1 & Q_2 \\
Q_1 & P_2
\end{pmatrix} \qquad \textup{with} \qquad
P_1 &= \begin{pmatrix}
Z_0 & - \widetilde{X}_0 \\
X_0 & \widetilde{Z}_0
\end{pmatrix}, \quad
Q_2 = \begin{pmatrix}
Z_2 & - \widetilde{W}_2 \\
W_2 & \widetilde{Z}_2
\end{pmatrix}, \nonumber \\
Q_1 & = \begin{pmatrix}
Z_1 & - \widetilde{W}_1 \\
W_1 & \widetilde{Z}_1
\end{pmatrix}, \quad
P_2 = \begin{pmatrix}
W_0 & - \widetilde{Y}_0 \\
Y_0 & \widetilde{W}_0
\end{pmatrix} .
\end{align}
An explicit computation leads to
\begin{align} \label{subtens}
Z_0 & = \psi_1 {\otimes} \psi^*_1 + \psi^*_2 {\otimes} \psi_2 , \qquad \widetilde{Z}_0 = \psi^*_1 {\otimes} \psi_1 + \psi_2 {\otimes} \psi^*_2
= Z^*_0 , \nonumber \\
X_0 , & = \psi_2 {\otimes} \psi^*_1 - \psi^*_1 {\otimes} \psi_2 , \qquad \widetilde{X}_0 = \psi^*_2 {\otimes} \psi_1 - \psi_1 {\otimes} \psi^*_2
= X^*_0 \nonumber \\
W_0 & = \psi_3 {\otimes} \psi^*_3 + \psi^*_4 {\otimes} \psi_4 , \qquad \widetilde{W}_0 = \psi^*_3 {\otimes} \psi_3 + \psi_4 {\otimes} \psi^*_4
= W^*_0 , \nonumber \\
Y_0 & = \psi_4 {\otimes} \psi^*_3 - \psi^*_3 {\otimes} \psi_4 \qquad \widetilde{Y}_0 = \psi^*_4 {\otimes} \psi_3 - \psi_3 {\otimes} \psi^*_4
= Y^*_0, \nonumber \\
Z_1 & = \psi_3 {\otimes} \psi^*_1 + \psi_4^* {\otimes} \psi_2 , \qquad \widetilde{Z}_1 = \psi^*_3 {\otimes} \psi_1 + \psi_4 {\otimes} \psi^*_2
= Z^*_1, \nonumber \\
W_1 & = \psi_4 {\otimes} \psi^*_1 - \psi^*_3 {\otimes} \psi_2 , \qquad \widetilde{W}_1 = \psi^*_4 {\otimes} \psi_1 - \psi_3 {\otimes} \psi^*_2
= W^*_1, \nonumber \\
Z_2 &= \psi_1 {\otimes} \psi^*_3 + \psi^*_2 {\otimes} \psi_4 , \qquad \widetilde{Z}_2 = \psi_2 {\otimes} \psi^*_4 + \psi^*_1 {\otimes} \psi_3
= Z^*_2, \nonumber \\
W_2 & = \psi_2 {\otimes} \psi^*_3 - \psi^*_1 {\otimes} \psi_4 , \qquad \widetilde{W}_2 = \psi^*_2 {\otimes} \psi_3 - \psi_1 {\otimes} \psi^*_4
= W^*_2.
\end{align}
It is then immediate to check that
$$
S_\mathcal{C}(V) = V^\dagger .
$$
We know that the generators are not independent. Indeed:
\begin{lem} There are four sphere relations:
\begin{align}
\widetilde{Z}_0 Z_0 + \widetilde{X}_0 X_0 &= Z_0 \widetilde{Z}_0 + X_0 \widetilde{X}_0 = \zeta_0 {\otimes} \zeta_0 , \nonumber \\
\widetilde{W}_0 W_0 + \widetilde{Y}_0 Y_0 &= W_0 \widetilde{W}_0 + Y_0 \widetilde{Y}_0 = (1-\zeta_0) {\otimes} (1-\zeta_0) , \nonumber \\
\widetilde{Z}_1 Z_1 + \widetilde{W}_1 W_1 &= Z_1 \widetilde{Z}_1 + W_1 \widetilde{W}_1 = (1-\zeta_0) {\otimes} \zeta_0 , \nonumber \\
\widetilde{Z}_2 Z_2 + \widetilde{W}_2 W_2 &= Z_1 \widetilde{Z}_1 + W_1 \widetilde{W}_1 = \zeta_0 {\otimes} (1-\zeta_0) .
\end{align}
\begin{proof}
One computes these from the relations \eqref{alter}.
\end{proof}
\end{lem}
In a sense this says that the four matrices in \eqref{4matrices} are all equivalent and for the generators of the $B$-bimodule
$\mathcal{C}(A,H)$ of coinvariant elements one can take any one of those together with $A(S_\theta^4) {\otimes} 1$ and $1 {\otimes} A(S_\theta^4)$.
Alternatively, one could express the generators of the latter spheres in terms of the generators in \eqref{subtens}.
\begin{lem}\label{source-target}
The source map:
\begin{align}
& \widetilde{Z}_0 Z_0 + \widetilde{X}_0 X_0 + \widetilde{Z}_2 Z_2 + \widetilde{W}_2 W_2
=\zeta_0 {\otimes} 1 , \nonumber \\
& Z_0 \widetilde{Z}_1 + \widetilde{X}_0 W_2 + Z_2 \widetilde{W}_0 + \widetilde{W}_2 Y_0
= \zeta_1 {\otimes} 1 , \nonumber \\
& X_0 \widetilde{Z}_1 + W_2 \widetilde{W}_0 - \widetilde{Z}_0 W_1 - \widetilde{Z}_2 Y_0
= \zeta_2 {\otimes} 1 .
\end{align}
and the target map:
\begin{align}
& \widetilde{Z}_0 Z_0 + \widetilde{X}_0 X_0 + \widetilde{Z}_1 Z_1 + \widetilde{W}_1 W_1 = 1 {\otimes} \zeta_0, \nonumber \\
& W_2 \widetilde{X}_0 + Z_2 \widetilde{Z}_0 + Y_0 \widetilde{W}_1 + W_0 \widetilde{Z}_1 = 1{\otimes} \zeta_1, \nonumber \\
& W_2 Z_0 - Z_2 X_0 + Y_0 Z_1 - W_0 W_1 = 1 {\otimes} \zeta_2 .
\end{align}
\begin{proof}
The direct way for these is just to use again the relations in \eqref{alter}.
\end{proof}
\end{lem}
\subsection{A Hopf algebroid with quantum orthogonal symmetry} \label{sec:ev-q-shp}
Let us denote $A=\O(\SOt(2n+1))$, $H=\O(\SOt(2n))$ and $B=\O(S_\theta^{2n})$.
With the notations of \S\ref{sec:-q-shp} consider the matrix valued function
\begin{equation}
\Phi = (\Phi_{J K}) =
\begin{pmatrix}
\b{a} & \b{b} \\
\b{b}^* & \b{a}^* \\
\b{v} & \b{v}^*
\end{pmatrix} .
\end{equation}
Then, the orthogonality conditions $N^\dagger N = \mathbb{I}$ gives that $\Phi^\dagger \cdot \Phi = \mathbb{I}_{2n}$.
Moreover, the entries of the matrix $\Phi \cdot \Phi^\dagger$
(a projection from the condition $\Phi^\dagger \cdot \Phi = \mathbb{I}_{2n}$) are coinvariants
for the coaction \eqref{subha}.
In fact, one computes explicitly that
\begin{equation}\Phi \cdot \Phi^\dagger =
\begin{pmatrix}
1 - \b{u} \b{u}^\dagger & - \b{u} \b{u}^t & - \b{u} x \\
- \b{u}^* \b{u}^\dagger & 1 - \b{u}^* \b{u}^t & - \b{u}^* x \\
- x \b{u}^\dagger & - x \b{u}^t & 1 - x^2
\end{pmatrix}
\end{equation}
Let us denote
\begin{equation}
\b{w}=\begin{pmatrix}
{\bf h} & {\bf k} \\
{\bf k}^* & {\bf h}^*
\end{pmatrix},
\end{equation}
the defining matrix of $\O(\SOt(2n))$, with $\b{w}^\dagger \b{w} = \b{w} \b{w}^\dagger = \mathbb{I}_2$.
Then the coaction \eqref{subha} reduces to a coaction
\begin{equation}\label{cored}
\delta^A (\Phi) = \Phi \stackrel{.}{\otimes} \b{w}
\end{equation}
or
\begin{equation}\label{cored-bis}
\begin{matrix}
\delta^A (\b{a}) = \b{a} \stackrel{.}{\otimes} \b{h} + \b{b} \stackrel{.}{\otimes} \b{k}^* \\
\delta^A (\b{b}) = \b{a} \stackrel{.}{\otimes} \b{k} + \b{b} \stackrel{.}{\otimes} \b{h}^* \\
\delta^A (\b{v}) = \b{v} \stackrel{.}{\otimes} \b{h} + \b{v}^* \stackrel{.}{\otimes} \b{k}^*
\end{matrix} \qquad \mbox{and} \qquad
\begin{matrix}
\delta^A (\b{a}^*) = \b{a}^* \stackrel{.}{\otimes} \b{h}^* + \b{b}^* \stackrel{.}{\otimes} \b{k} \\
\delta^A (\b{b}^*) = \b{a}^* \stackrel{.}{\otimes} \b{k}^* + \b{b}^* \stackrel{.}{\otimes} \b{h} \\
\delta^A (\b{v}^*) = \b{v}^* \stackrel{.}{\otimes} \b{h}^* + \b{v} \stackrel{.}{\otimes} \b{k} \, .
\end{matrix}
\end{equation}
In turns this gives
\begin{equation}
\delta^A (\Phi^\dagger) = \sigma(\b{w}^\dagger \stackrel{.}{\otimes} \Phi^\dagger).
\end{equation}
The translation map is easily seen to be given by
\begin{equation} \label{tramap2}
\tau(\b{w}) = \Phi^\dagger \stackrel{.}{\otimes}_B \Phi .
\end{equation}
One could show the coinvariance of the entries of $\Phi \cdot \Phi^\dagger$ by using the explicit form \eqref{cored-bis} of the coaction.
Then in exactly the same way, one shows the following:
\begin{lem}
The $B$-bimodule $\mathcal{C}(A,H)$ of coinvariant elements for the diagonal coaction of $H$ on $A {\otimes} A$ is generated
by elements $1{\otimes} \b{u}, 1{\otimes} \b{u}^*, 1 {\otimes} x$, and $\b{u} {\otimes} 1, \b{u}^* {\otimes} 1, x {\otimes} 1$,
together with the entries of the matrix
$$
\b{V} = \Phi \stackrel{.}{\otimes} \Phi^\dagger =
\begin{pmatrix}
\b{a} & \b{b} \\
\b{b}^* & \b{a}^* \\
\b{v} & \b{v}^*
\end{pmatrix} \stackrel{.}{\otimes}
\begin{pmatrix}
\b{a} & \b{b} \\
\b{b}^* & \b{a}^* \\
\b{v} & \b{v}^*
\end{pmatrix}^\dagger.
$$
Moreover, the flip $S_\mathcal{C}(x {\otimes} y) = y {\otimes} x$ leaves unchanged the space of coinvariants, and in particular
$$
S_\mathcal{C} (\b{V}) = \sigma(\Phi \stackrel{.}{\otimes} \Phi^\dagger) = \b{V}^\dagger .
$$
\end{lem}
\begin{proof}
For $\b{V} = \Phi \stackrel{.}{\otimes} \Phi^\dagger$:
$$
\delta^{A {\otimes} A}(\b{V}) = \zero{\Phi} \stackrel{.}{\otimes} \zero{\Phi^\dagger} \stackrel{.}{\otimes} \one{\Phi} \one{\Phi^\dagger}
= \Phi \stackrel{.}{\otimes} \Phi^\dagger \stackrel{.}{\otimes} (\b{w} \b{w}^\dagger)= \b{V} \stackrel{.}{\otimes} (\b{w} \b{w}^\dagger) = \b{V} \stackrel{.}{\otimes} \mathbb{I}_{2n}
$$
and directly:
$\delta^{A {\otimes} A}(\b{V}^\dagger) = \sigma(\zero{\Phi} \stackrel{.}{\otimes} \zero{\Phi^\dagger}) \stackrel{.}{\otimes} \one{\Phi^\dagger} \one{\Phi}
= \sigma(\Phi \stackrel{.}{\otimes} \Phi^\dagger) \stackrel{.}{\otimes} (\b{w}^\dagger \b{w}) = \b{V^\dagger} \stackrel{.}{\otimes} \mathbb{I}_{2n}
$.
\end{proof}
\begin{lem} Using the conditions $N^\dagger N = \mathbb{I} = N N^\dagger$ one finds the following relations
\begin{align}
\Phi^\dagger \cdot_{op} \Phi =
\begin{pmatrix}
(\Phi^\dagger \cdot \Phi )_{22} & (\Phi^\dagger \cdot \Phi )_{12} \\
(\Phi^\dagger \cdot \Phi )_{21} & (\Phi^\dagger \cdot \Phi )_{11}
\end{pmatrix} =
Q (\Phi^\dagger \cdot \Phi)^t Q =
\begin{pmatrix}
\mathbb{I}_n & 0 \\
0 & \mathbb{I}_n
\end{pmatrix} = \mathbb{I}_{2n}
\end{align}
\begin{align}
\Phi \cdot_{op} \Phi^\dagger =
\begin{pmatrix}
1 - \b{u}^* \b{u}^t & - \b{u} \b{u}^t & - \b{u}^* x\\
- \b{u}^* \b{u}^\dagger & 1 - \b{u} \b{u}^\dagger & - \b{u} x \\
- x \b{u}^t & - x \b{u}^\dagger & 1 - x^2
\end{pmatrix} =
Q (\Phi \cdot \Phi^\dagger)^t Q
\end{align}
It is evident that $(\Phi \cdot \Phi^\dagger)_{J K} \in B$; as well as $(\Phi \cdot_{op} \Phi^\dagger)_{J K} \in B$.
Also,
\begin{equation}
\b{V} \b{V}^\dagger = (\Phi \cdot \Phi^\dagger) {\otimes} 1 , \qquad \b{V}^\dagger \b{V} = 1 {\otimes} (\Phi \cdot_{op} \Phi^\dagger),
\end{equation}
which express relations among the generators of $\mathcal{C}(A,H)$.
\end{lem}
\noindent
In parallel with the projection $\Phi \cdot \Phi^\dagger$, the matrix $\Phi \cdot_{op} \Phi^\dagger$ is a projection due to $\Phi^\dagger \cdot_{op} \Phi = \mathbb{I}_{2n}$.
\medskip
We are ready for the Hopf algebroid structure.
\begin{prop}
On $\mathcal{C}(A,H)$, the coproduct $\Delta : \mathcal{C}(A,H) \to \mathcal{C}(A,H) {\otimes}_B \mathcal{C}(A,H)$, according to the definition \eqref{copro} and using the translation map \eqref{tramap2},
is given by
\begin{align} \label{ccopor}
\Delta (\b{V}) = \zero{\Phi} \stackrel{.}{\otimes} \tuno{\one{\Phi}} \stackrel{.}{\otimes} _B \tdue{\one{\Phi}} \stackrel{.}{\otimes} \Phi^\dagger
= \Phi \stackrel{.}{\otimes} \Phi^\dagger \stackrel{.}{\otimes}_B \Phi \stackrel{.}{\otimes} \Phi^\dagger = \b{V} \stackrel{.}{\otimes} _B \b{V}.
\end{align}
In components this reads:
\begin{equation}\label{copcom-1}
\Delta(V_{J K}) = \sum_L V_{J L} {\otimes}_B V_{L K}.
\end{equation}
Also,
\begin{equation}
\Delta (S_\mathcal{C}(\b{V})) = \sigma(S_\mathcal{C}(\b{V}) \stackrel{.}{\otimes} _B S_\mathcal{C}(\b{V})) , \quad
\end{equation}
or
\begin{equation}
\Delta(S_\mathcal{C}(V)_{J K}) = \sum_L S_\mathcal{C}(V)_{L K} {\otimes}_B S_\mathcal{C}(V)_{J L} , \qquad \Delta(V^\dagger_{J K}) = \sum_L V^\dagger_{L K} {\otimes}_B V^\dagger_{J L}.
\end{equation}
\end{prop}
Very much in the lines of the proof \eqref{sverif-bis}, we have the following:
\begin{prop}
The flip $S_\mathcal{C}$ is the antipode of $\mathcal{C}(A,H)$.
\end{prop}
\begin{proof}
Since $S_\mathcal{C}$ is just the flip, condition \eqref{hopbroid1} is obvious. For conditions \eqref{hopbroid2} take $h=V_{J K}$, with $S_\mathcal{C}(h) = V^\dagger_{J K}$ and use the explicit form of the coproduct \eqref{copcom-1}. Then the proof proceeds verbatim as in the proof of \eqref{sverif-bis}.
\end{proof}
|
\section{Linear Advection and Advection Diffusion\label{sec:linearadvdiff}}
In order to perform a series of numerical experiments, two test cases have been chosen --- the linear advection of a sine wave, on a domain $\Omega = [-\pi,\pi]$, and the linear advection-diffusion of a Gaussian bump, on a domain $\Omega = [-10,10]$, according to the equation:
\begin{align}\label{eq:linearAdvection}
\px{u}{t} + a \px{u}{x} = D \pxi{2}{u}{x}
\end{align}
For both test cases, the advection speed, $a$, is equal to unity, and for the advection-diffusion problem, the diffusion coefficient, $D$, is set to $0.1$. This coefficient is set to zero for the pure advection of the sine wave. Periodic boundary conditions were used in both cases.
To perform these analyses, an initial condition was specified for both cases:
\begin{subequations}\label{eq:LAD_Initial}
\begin{align}
\boldsymbol{u}_0 & = \boldsymbol{u}(t=0) = \sin{\boldsymbol{x}} \qquad \text{Sine Wave Advection} \\
\boldsymbol{u}_0 & = \boldsymbol{u}(t=0) = \exp{\left(-\boldsymbol{x}^2\right)} \qquad \text{Gaussian Bump Advection Diffusion}
\end{align}
\end{subequations}
From these initial conditions, the calculations were advanced in time, and the resulting solutions from the calculation at time $t$, $\boldsymbol{u}^\delta(t)$, are compared with the exact analytical solution, $\boldsymbol{u}(t)$, across the entire domain. The difference between the approximate and analytical solutions is reported as an error vector. A variety of ways of measuring the magnitude of this error exist. Here, the error is reported as an $\mathcal{L}^2$ norm of this vector, \emph{i.e.}:
\begin{equation}
\|\mathcal{E}\|_2 = \sqrt{\sum^N_{j=1}\sum^{n_s}_{i=1}\left(\mathbf{u}^\delta_j(x_i) - u(x_{j,i})\right)^2}
\end{equation}
for a case with $N$ elements and with each element composed of $n_s$ solution points. In addition to the $\mathcal{L}^2$ norm, the $\mathcal{L}^1$ and $\mathcal{L}^\infty$ vector norms were also calculated. These were both found to behave similarly to the $\mathcal{L}^2$ norm across all cases, and so they are further reported here.
By making use of these two test cases, the various factors which might influence the performance of radial basis flux reconstruction are explored in this section.
\subsection{Layout of the Solution Points\label{ssec:solnpointlocs}}
The layout of the solution points within the reference cell is known to have an influence on the performance of flux reconstruction [REF]. It is theorised that this effect is likely to be more significant with radial basis interpolation, because the bases themselves are directly dependent on the point location in a way that the polynomials are not.
\begin{figure}[htbp]
\centering
\adjustbox{width=0.6\linewidth, valign=b}{\input{Figs/tikz/InterpolationPoints}}
\caption{\label{fig:pointLayouts}Distributions of solution point layout within the reference element.}
\end{figure}
\cref{fig:pointLayouts} shows the solution point locations that have been tested here. The Legendre, Lobatto, and Chebyshev points are familiar from polynomial quadrature theory as the locations of the roots of the relevant functions. They have different advantages and disadvantages in terms of function approximation, and the Lobatto points in particular have some advantages in terms of flux reconstruction. As can be seen from the figure, the Lobatto points include the end points. This means that the flux points are collocated with the terminal solution points, avoiding the need for projection to the flux points, and thereby offering a potential saving in compute time. In addition to these three familiar point layouts, two others have been tested --- one, the `internal' uniform points, which are uniformly distributed over the inside of the cell, with half a spacing between the terminal solution points and the flux points, and two, the `full' uniform points, which are uniformly distributed over the entire cell, from flux point to flux point, and sharing the potential computational benefit of collocated flux and solution points with the Lobatto points.
\begin{figure}[tbhp]
\centering
\subfloat[$n=3$.]{\label{fig:n3_gad_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_pointLocations}}}
~
\subfloat[$n=4$.]{\label{fig:n4_gad_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_pointLocations}}}
\newline
\subfloat[$n=5$.]{\label{fig:n5_gad_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_pointLocations}}}
\caption{\label{fig:gad_locations}The effect of point layout on $L_2$ error for linear advection-diffusion of a Gaussian bump.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n=3$.]{\label{fig:n3_sa_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_pointLocations}}}
~
\subfloat[$n=4$.]{\label{fig:n4_sa_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_pointLocations}}}
\newline
\subfloat[$n=5$.]{\label{fig:n5_sa_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_pointLocations}}}
\caption{\label{fig:sa_locations}The effect of point layout on $L_2$ error for linear advection of a sine wave.}
\end{figure}
\cref{fig:gad_locations,fig:sa_locations} show the performance of the RBF-FR method using different point locations for a range of mesh spacings. \cref{fig:gad_locations} shows behaviour of the advection diffusion of the Gaussian bump, and~\cref{fig:sa_locations} shows performance on the pure advection of the sine wave, both at a modest shape function value of $\varepsilon=0.5$, and across three solution point numbers, of $n_s = 3$, $4$, and $5$. For comparison, a set of solutions computed with the polynomial bases using the Legendre is supplied. It is immediately clear that, as expected, there is a strong dependence of accuracy on the point layouts, with close to an order of magnitude of error separating the best performing points from the worst performing points. Typically, the full uniform points and the Lobatto points perform least well, and, in general, the Legendre points perform best. For the advection diffusion equation, there is some evidence in \cref{fig:gad_locations} that the better performing solution point layouts are capable of outperforming the polynomial bases on the coarsest meshes, but that this does not hold true as the meshes are refined. The RBF-FR also seems to perform better, relative to the polynomial baseline, with $3$ and $5$ solution points, performing noticeably worse for the $n_s = 4$ case, for both the advection diffusion and pure advection test cases.
\cref{fig:gad_locations_01,fig:sa_locations_01} show the same runs as \cref{fig:gad_locations,fig:sa_locations}, but with the shape factor of the radial bases reduced from $0.5$ to $0.1$. As would be expected, as the shape factor is reduced towards zero, the performance of the fit follows the polynomial results more closely --- in the flat limit of $\varepsilon = 0$, the Gaussian RBF fit can be shown to be equivalent to a polynomial fit [REF]. In terms of the point locations, similar trends are observed with the wider Gaussian bases --- there is still close to an order of magnitude of difference between the best and worst performing. Again, the functions with collocated solution and flux points (the full uniform and Lobatto points) are seen to perform least effectively, and the Legendre points are again found to be the most effective of those tested. With the advection diffusion of the Gaussian bump, \cref{fig:gad_locations_01}, there is evidence of the best performing point locations outperforming the polynomial fit across the entire range tested. With the pure advection of the sine wave, \cref{fig:sa_locations_01}, the behaviour of RBF-FR is seen to be more complex. There are pronounced peaks and troughs in the error curves, which typically moves the RBF-FR accuracy either side of the polynomial accuracy. Again, the general trend is that RBF-FR is able to outperform the polynomial method on the coarser meshes, but is overtaken as the meshes are refined.
\begin{figure}[tbhp]
\centering
\subfloat[$n=3$.]{\label{fig:n3_gad_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_pointLocations_01}}}
~
\subfloat[$n=4$.]{\label{fig:n4_gad_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_pointLocations_01}}}
\newline
\subfloat[$n=5$.]{\label{fig:n5_gad_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_pointLocations_01}}}
\caption{\label{fig:gad_locations_01}The effect of point layout on $L_2$ error for linear advection-diffusion of a Gaussian bump.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n=3$.]{\label{fig:n3_sa_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_pointLocations_01}}}
~
\subfloat[$n=4$.]{\label{fig:n4_sa_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_pointLocations_01}}}
\newline
\subfloat[$n=5$.]{\label{fig:n5_sa_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_pointLocations_01}}}
\caption{\label{fig:sa_locations_01}The effect of point layout on $L_2$ error for linear advection of a sine wave.}
\end{figure}
\subsection{Effect of the Shape Function\label{ssec:shapefunction}}
The scale of the shape function, $\varepsilon$, or equivalently, the relative width of the basis, is known to have a significant effect on the performance of RBF interpolation [REF]. In this section, the effect of reducing this scaling parameter is explored on the two test cases. In all of these cases, the best performing point locations from the previous section were used --- the Legendre points.
\cref{fig:gad_varEpsilon,fig:sa_varEpsilon} show the effect of reducing this parameter on the performance of RBF-FR, again with a polynomial FR curve for comparison, for $n_s = 3$, $4$, and $5$. Generally, the accuracy of the RBF-FR solution improves as the shape parameter is reduced. \cref{fig:gad_varEpsilon}, which shows the error incurred in the evolution of the Gaussian bump under both advection and diffusion, suggests that each of the shape factors are capable of outperforming the polynomial fit at some grid density. Larger values of $\varepsilon$ tend to outperform the polynomial on extremely coarse grids, but as the grid is refined, the polynomial approach overtakes them. As the value of the shape parameter is reduced, the range of mesh densities at which RBF-FR outperforms polynomial FR moves towards finer meshes. By the time a value of $\varepsilon = 0.1$ is reached, the RBF-FR outperforms the polynomial across the entire range of mesh densities tested, with close to a $50\%$ reduction in the error on the finest mesh tested for $n_s = 5$. The other observable trend in general is that as the shape parameter is reduced, the performance becomes increasingly like that of the polynomial FR.
\cref{fig:sa_varEpsilon} shows the same tests, but for the pure advection of a sine wave. The RBF-FR performs much less competitively here, but similar trends can be observed --- higher values of shape parameter, so narrower radial bases, tend to outperform the polynomial on more coarse meshes, and wider bases tend to outperform the polynomial FR on finer meshes. However, in all of these cases the polynomial FR outperforms RBF-FR for the finest meshes. Despite this, there are large sections of mesh densities in which the RBF-FR is highly competitive with the polynomial FR in terms of $L_2$ error norm. As seen with the advection diffusion of the Gaussian bump, as the shape parameter is reduced, the curves gradually trend towards the baseline polynomial $L_2$ error curve. However, with the baseline Gaussian Radial Basis function, it was not possible to use shape parameters significantly smaller than this, as the condition number of system of equations requiring to be solved to find $\psi(r)$ from $\phi(r)$ in \cref{eq:RBF_AlternantMatrix} grows extremely rapidly, and numerical errors swamp the solution. A method to ameliorate this is covered in detail in SECTION.
\begin{figure}[tbhp]
\centering
\subfloat[$n=3$.]{\label{fig:n3_gad_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_varEpsilon}}}
~
\subfloat[$n=4$.]{\label{fig:n4_gad_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_varEpsilon}}}
\newline
\subfloat[$n=5$.]{\label{fig:n5_gad_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_varEpsilon}}}
\caption{\label{fig:gad_varEpsilon}The effect of shape parameter, $\varepsilon$, on $L_2$ error for linear advection-diffusion of a Gaussian bump.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[Three point $L_2$ error, $n=3$, Legendre points.]{\label{fig:n3_sa_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_varEpsilon}}}
~
\subfloat[Four point $L_2$ error, $n=4$, Legendre points.]{\label{fig:n4_sa_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_varEpsilon}}}
\newline
\subfloat[Five point $L_2$ error, $n=5$, Legendre points.]{\label{fig:n5_sa_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_varEpsilon}}}
\caption{\label{fig:sa_varEpsilon}The effect of shape parameter, $\varepsilon$, on $L_2$ error for linear advection of a sine wave.}
\end{figure}
\subsection{Effect of the RBF Centre Locations}
As well as their shape, radial basis functions are also defined by the locations from which the Euclidean distance which gives $r$ is measured --- their centres. In the previous sections, the centres of the bases have been placed at the same location as the solution points. Here, we investigate the effect of separating them.
\cref{fig:gad_cenLocations} shows this, at two different values of shape parameter. The solution points are kept at the Legendre nodes, and the RBF centres are changed from the Legendre nodes to the full uniform and interal uniform distributions. From the plot, it appears that any influence made by moving the RBF centres is marginal, and, as might be expected, this influence reduces even further as the shape parameter decreases and the bases increasingly accurately approximate $\phi(r)=1$. In general, the results shown here for the advection diffusion of the Gaussian bump suggest that there may be a small benefit in using the full uniform point set as the centres, but this effect is again marginal. It is possible that this small benefit is related to the effect of the functional centre locations on the condition number of the alternant matrix, and thus the numerical accuracy of the fit. This is outlined in more detail in \cref{subsec:smallShapeParameters}. The result, however, was not repeated for the pure advection of the sine wave, and due to the marginal nature of the effect seen, further results are omitted here for brevity.
\begin{figure}[tbhp]
\centering
\subfloat[Three point $L_2$ error, $n=3$, Legendre points.]{\label{fig:n3_gad_cenLocations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_cenLocations}}}
~
\subfloat[Four point $L_2$ error, $n=4$, Legendre points.]{\label{fig:n4_gad_cenLocations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_cenLocations}}}
\newline
\subfloat[Five point $L_2$ error, $n=5$, Legendre points.]{\label{fig:n5_gad_cenLocations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_cenLocations}}}
\caption{\label{fig:gad_cenLocations}The effect of varying the RBF centre locations on $L_2$ error for linear advection-diffusion of a Gaussian bump.}
\end{figure}
\subsection{Small Values of Shape Parameter \label{subsec:smallShapeParameters}}
So far, it has been shown that the reduction of $L_2$ error with increased mesh density has been shown to trend towards the polynomial case as the shape parameter is reduced towards zero. However, the RBF alternant matrix shown in \cref{eq:RBF_AlternantMatrix} becomes increasingly ill-conditioned as the the shape parameter decreases, becoming equal to the matrix of ones in the limit $\varepsilon \to 0$. \cref{fig:condNumber_RBFDirect} shows this variation in condition number with shape parameter, which is also affected by the location of the centres, related to the mean distance between the functional centres. A general rule-of-thumb is that a for every order of magnitude the condition number increases by, a digit of precision is lost in the solution. This means that the system of equations cannot be reliably solved for for smaller values of shape parameter. However, various sources have pointed out [REF, REF, REF] that it is not the basis itself that is ill-conditioned --- it is merely an intermediate step in calculating it which is to blame.
\begin{figure}[tbhp]
\centering
\adjustbox{width=0.6\linewidth, valign=b}{\input{Figs/tikz/conditionNumber_01}}
\caption{\label{fig:condNumber_RBFDirect}The effect of reducing $\varepsilon$ on alternant matrix condition number when using direct Gaussian RBFs.}
\end{figure}
To this end, various methods have been developed which circumvent this step to allow an effective basis to be produced when using very flat radial bases. Examples of ways that this has been achieved are the Contour-Padé method [REF] and the RBF-QR method [REF]. Here, since the investigation is restricted to Gaussian radial bases, an approach known as the RBF-GA method is used to build the basis functions without the need to directly solve the system of equations arising from the alternant matrix. The details of the approach are omitted for brevity, but it involves expanding the original Gaussian functions into a Taylor series, and then selecting linear combinations which cause most of these terms to cancel out. The result is a new basis which, critically, spans exactly the same space as the original, but which is well conditioned even at very low values of $\varepsilon$.
\cref{fig:condNumber_RBFGA} shows the effect on the condition number of the solution point alternant matrix of calculating the basis via the RBF-GA method, rather than the directly. The same dependence on the mean spacing between the functional centres is observed, but in general the condition number of the alternant is below $10^3$, making it apparent that by side-stepping the use of the alternant matrix, it becomes numerically feasible to use far smaller values of the shape parameter.
\begin{figure}[tbhp]
\centering
\adjustbox{width=0.6\linewidth, valign=b}{\input{Figs/tikz/conditionNumber_02}}
\caption{\label{fig:condNumber_RBFGA}The effect of reducing $\varepsilon$ on alternant matrix condition number when using RBF-GA.}
\end{figure}
\cref{fig:gad_smallEpsilon,fig:sa_smallEpsilon} show the effect that making use of the RBF-GA approach to compute bases has on the mesh density sweep for the two test cases. It is possible to use lower values of $\varepsilon$, but $\varepsilon=0.1$ is repeated with the RBF-GA approach, and is found to give the same results as the direct RBF method for the test cases --- as expected, given that both span the same functional basis space.
\begin{figure}[tbhp]
\centering
\subfloat[Three point $L_2$ error, $n=3$, Legendre points.]{\label{fig:n3_gad_smallEpsilonc}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_smallEpsilon}}}
~
\subfloat[Four point $L_2$ error, $n=4$, Legendre points.]{\label{fig:n4_gad_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_smallEpsilon}}}
\newline
\subfloat[Five point $L_2$ error, $n=5$, Legendre points.]{\label{fig:n5_gad_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_smallEpsilon}}}
\caption{\label{fig:gad_smallEpsilon}The effect of very small shape parameters on $L_2$ error for linear advection-diffusion of a Gaussian bump.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[Three point $L_2$ error, $n=3$, Legendre points.]{\label{fig:n3_sa_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_smallEpsilon}}}
~
\subfloat[Four point $L_2$ error, $n=4$, Legendre points.]{\label{fig:n4_sa_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_smallEpsilon}}}
\newline
\subfloat[Five point $L_2$ error, $n=5$, Legendre points.]{\label{fig:n5_sa_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_smallEpsilon}}}
\caption{\label{fig:sa_smallEpsilon}The effect of very small shape parameters on $L_2$ error for linear advection of a sine wave.}
\end{figure}
\section{Conclusions}\label{sec:conclusions}
In this paper, the Gaussian RBF has been used to replace polynomials as the method of approximating functions within elements as part of the broader FR framework for solving PDEs The effect that this has on the performance of the simulations is investigated both analytically and numerically.
The shape parameter, $\varepsilon$, is found to be of vital importance. Lower values of shape parameter trend towards the polynomial performance, but there tends to be a range of mesh densities at which the RBF-FR approach is capable of outperforming the polynomial, for linear problems. This range of mesh densities becomes finer as the shape parameter is reduced.
Although directly fitting RBFs to scattered points becomes untenable for very low values of $\varepsilon$, a method was discussed to avoid this. Although this makes the operators more expensive to compute, the nature of flux reconstruction means that these operator matrices are calculated \emph{a priori}, and do not need to be recomputed for each run, meaning there is no increase in simulation run time. Very low values of shape parameter are proven to make RBF-FR effective for both linear and nonlinear problems.
The position of the solution points is also demonstrated to have a considerable impact on the performance of the simulation, with the Legendre points being found to be the most effective for linear problems. This difference was less pronounced for the nonlinear numerical experiments. Although this paper is restricted to studying the effects in one dimension, the increased geometrical flexibility of RBFs make them potentially attractive in higher dimensions. Future work will explore the potential of this geometric flexibility in improving the performance of FR.
The choice of functional centres for the RBFs was found to have a very small influence on the performance --- particularly as the shape parameter was reduced and the RBFs became flat.
\section{\label{sec:fourier}Fourier Analysis}
Initially, the properties of RBF-FR schemes are examined analytically. Subsequent sections will illustrate their behaviour through numerical experiments. Fourier analysis provides a powerful tool for examining spatial operators through their frequency responses. Previously in the FR literature, Fourier analysis has been used to understand the dispersion and dissipation of schemes in one and two dimensions for advection equations~\citep{Huynh2007,Trojak2020}. Further works have explored schemes' CFL limits, diffusion effects, and rates of convergence~\citep{Vincent2011,Watkins2016,Asthana2017}.
As discussed, the FR method generates a solution through the superposition of several modes. When viewed in discrete Fourier space, these have often been divided into primary and secondary modes or into physical and spurious modes. The primary mode describes the most energised, whereas the physical mode is that which most accurately represents the projection from the approximation space into Fourier space. This approach has provided considerable insight into the behaviour of spatial methods. However, an alternative `holistic' approach, in which the combined effect of all the modes is considered, has recently been presented \citep{Alhawwary2020,Asada2020,Fernandez2019}.
We begin by introducing the discrete eigenmode approach. Take the linear advection equation:
\begin{equation}\label{eq:lin_adv}
\px{u}{t} + \px{u}{x} = 0, \quad \forall u(x,t): \Omega\times\mathbb{R}_+ \mapsto \mathbb{C}
\end{equation}
Here, we allow the solution to be complex. If this is taken as a Cauchy problem --- by setting the domain $\Omega$ to be either toroidal or infinite --- then \cref{eq:lin_adv} can support Bloch wave solutions:
\begin{subequations}
\label{eq:bloch}
\begin{align}
u &= \exp{\left(ik(x - ct)\right)} \\
\boldsymbol{u}_j &= \exp{\left(ik\left(x_j + \frac{h}{2}(\pmb{\xi}+1) + ct\right) \right)} \quad \forall \boldsymbol{u}_j : \boldsymbol{x}_j \times \mathbb{R}_+ \mapsto \mathbb{C}^n
\end{align}
\end{subequations}
where $\boldsymbol{x}_j \in \Omega^n_j$ for $\Omega_j = [x_j,x_{j+1}]$, where the mesh is taken to be uniform with spacing $h=x_{j+1}-x_j$. The latter equation here represents the discrete solution vector within the $j^\mathrm{th}$ element.
Constructing the flux reconstruction operator, \cref{eq:lin_adv} is first written as:
\begin{equation}\label{eq:linadv_approx}
\px{\boldsymbol{u}_j}{t} = -\px{\boldsymbol{u}_j}{x} \approx \mathbf{Q}\boldsymbol{u}_j
\end{equation}
then, following the method of \citet{Trojak2020}, we have:
\begin{equation}
\mathbf{Q} = -\frac{2}{h}\left(\mathbf{C}_-\exp{(-ikh)} + \mathbf{C}_0 + \exp{(ikh)}\mathbf{C}_+\right)
\end{equation}
where:
\begin{subequations}
\begin{align}
\mathbf{C}_{-} = \mathbf{C}\mathbf{K}_{-}\mathbf{P}^T, \quad &\mathbf{K}_{-} = \begin{bmatrix}0 & 0 \\ 1-\alpha & 0 \end{bmatrix} \\
\mathbf{C}_{0} = \mathbf{D} + \mathbf{C}\mathbf{K}_{0}\mathbf{P}^T, \quad &\mathbf{K}_{0} = -\text{diag}{(\alpha,1-\alpha)} \\
\mathbf{C}_{+} = \mathbf{C}\mathbf{K}_{+}\mathbf{P}^T, \quad &\mathbf{K}_{+} = \begin{bmatrix}0 & \alpha \\ 0 & 0 \end{bmatrix}
\end{align}
\end{subequations}
Here, $\alpha$ is the interface upwinding parameter, with $\alpha=1$ providing a fully upwinded interface and $\alpha=0.5$ giving centrally differenced interfaces.
By substituting \cref{eq:bloch} into \cref{eq:linadv_approx} and rearranging, the modified angular frequency is obtained as:
\begin{equation}
c^\prime k\mathbf{w} = \omega^\prime\mathbf{w} = i\mathbf{Q}\mathbf{w}
\end{equation}
where $\mathbf{w}$ is an eigenvector, and $c^\prime$ and $\omega^\prime$ are the modified wavespeed and modified angular frequency, respectively. The implication of this is that for a given wavenumber, $k$, the solution to the Bloch wave problem actually given by an FR scheme is equivalent to:
\begin{equation}
u_\text{mod}=\exp{(ikx-i\omega^\prime t)}
\end{equation}
\begin{figure}[tbhp]
\centering
\subfloat[Dispersion relation, $n_s=3$.]{\label{fig:n3_disp}\adjustbox{width=0.49\linewidth,valign=b}{\input{Figs/tikz/n3_gauss_dispersion}}}
~
\subfloat[Dissipation relation, $n_s=3$.]{\label{fig:n3_diss}\adjustbox{width=0.49\linewidth,valign=b}{\input{Figs/tikz/n3_gauss_dissipation}}}
\newline
\subfloat[Dispersion relation, $n_s=4$.]{\label{fig:n4_disp}\adjustbox{width=0.49\linewidth,valign=b}{\input{Figs/tikz/n4_gauss_dispersion}}}
~
\subfloat[Dissipation relation, $n_s=4$.]{\label{fig:n4_diss}\adjustbox{width=0.49\linewidth,valign=b}{\input{Figs/tikz/n4_gauss_dissipation}}}
\caption{\label{fig:disp_diss}Upwind RBF-FR dispersion and dissipation plots.}
\end{figure}
The results presented in \cref{fig:disp_diss} show the real and imaginary components of this modified angular frequency, which correspond to the dispersion and dissipation, respectively, for an RBF-FR scheme. For these calculations, the solution points have been placed at the Gauss--Legendre points, as have the centres of the radial bases. Only the physical mode is shown. These results show that the change of basis has a significant impact on the dispersion and dissipation characteristics, which trend towards the behaviour of the polynomial basis scheme as the shape parameter, $\varepsilon$, is reduced and the RBFs flattened. Furthermore, as $\varepsilon$ is increased the characteristic dispersion over shoot and high dissipation at high wavenumbers is reduced. Which is most likely linked to a flattening of the correction function as $\varepsilon$ increases. Ultimately, continued increases to the shape parameter lead the imaginary part to become positive for some wavenumbers, suggesting a degree of instability.
\begin{figure}[tbhp]
\centering
\subfloat[Gaussian RBF-FR max dissipation.]{\label{fig:diss_max}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/dmax_plot}}}
~
\subfloat[SBP test of Gaussian RBF-FR. Solid lines test conservation (\ref{eq: SBP_conservation}), and dashed lines test linear stability (\ref{eq: SBP_stability}).]{\label{fig:sbp}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/sbp_eps}}}
\caption{Testing the stability of Gaussian RBF-FR schemes.}
\end{figure}
Since changing the RBF width can lead to significant changes in the numerical properties of the method, a sweep was taken over a range of $\varepsilon$ values, and the maximum value of the dissipation recorded. The results for Legendre functional centres and Legendre solution points with a Gaussian RBF are presented in \cref{fig:diss_max}. This shows that the change in stability with $\varepsilon$ is dependent on degree of the basis, with higher degree bases permitting higher $\varepsilon$ values.
Pursuing this further, the SBP framework introduced in \cref{ssec:sbp} gives a means to test for conservation and linear stability in a manner which is independent of the initial condition used. Using this, the norm of the error has been calculated and plotted in \cref{fig:sbp} for Legendre solution points and centres with Gaussian RBFs of varying width. The two norms investigated are:
\begin{subequations}
\begin{align}
\|\mathbf{1}^T\mathbf{MC}-\mathbf{1}^T\mathbf{P}^T\mathbf{B}\|_2 &= \text{Conservation Error} \label{eq: SBP_conservation}\\
\|\mathbf{M}^{-1}\mathbf{P}^T\mathbf{B}-\mathbf{C}\|_2 &= \text{Linear Stability Error} \label{eq: SBP_stability}
\end{align}
\end{subequations}
These show that for small values of $\varepsilon$ the linear stability errors scale with $\varepsilon^2$ and conservation errors scale with $\varepsilon^6$. The loss of linear stability this shows is consistent with the rising value of maximum $\Im{(\hat{c}^\prime(\hat{k}))}$ shown in \cref{fig:diss_max}
To gain further insight, we pursue the combined analysis of \citet{Alhawwary2018}. Here we will use the diagonalisation of $\mathbf{Q}$:
\begin{equation}
\mathbf{Q} = \mathbf{W\Lambda W}^{-1},
\end{equation}
which is valid in 1D provided the RBF centres and solution points are unique.
If we take the initial condition as being:
\begin{equation}
\boldsymbol{u}_j(t=0) =\boldsymbol{u}_{j,0} = \exp{\left(ik\left(x_j + \frac{h}{2}(\pmb{\xi}+1)\right) \right)} = \pmb{\beta}\mathbf{W}
\end{equation}
then exponential integration can be used to find the approximate solution at time $t$ as:
\begin{equation}
\tb{u}_j(t) = \mathbf{W}\exp{(-it\mathbf{\Lambda})}\mathbf{W}^{-1}\boldsymbol{u}_{j,0}
\end{equation}
Following \citet{Alhawwary2018}, we may then define the combined amplification factor as:
\begin{equation}
G(k) = \frac{\|\tb{u}_{j}(t)\|_{L^2(\hat{\Omega})}}{\|\mathbf{u}_{j}(t)\|_{L^2(\hat{\Omega})}}
\end{equation}
and the phase difference as:
\begin{equation}
\Delta\phi(k) = t - \angle\left(\int_{\hat{\Omega}} \tilde{u}_j(t)\overline{u}_j(t) \mathrm{d}\xi\right)
\end{equation}
where $\overline{z}$ and $\angle(z)$ are the complex conjugate and argument, respectively, of the complex number $z$.
\begin{figure}[tbhp]
\centering
\subfloat[Combined Amplitude, $n_s=3$.]{\label{fig:cd_amp3}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_combined_fourier_amp}}}
~
\subfloat[Combined Phase Lag, $n_s=3$.]{\label{fig:cd_phase3}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_combined_fourier_phase}}}
\newline
\subfloat[Combined Amplitude, $n_s=4$.]{\label{fig:cd_amp4}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_combined_fourier_amp}}}
~
\subfloat[Combined Phase Lag, $n_s=4$.]{\label{fig:cd_phase4}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_combined_fourier_phase}}}
\caption{\label{fig:cd}Combined semi-discrete Fourier analysis (solid $t=1$, dashed $t=2$).}
\end{figure}
\cref{fig:cd} presents the results of the amplification factor and phase lag at two instants in time, $t=1$ and $t=2$. These amplification factors show that the RBF bases produce lower dissipation, and that this difference becomes larger for longer integration times. This `holistic' approach confirms what was seen in the Fourier analysis results presented earlier. The holistic phase lag results present a more complex picture. Generally, the RBF basis is shown to have lower phase lag than the polynomial basis; however, at low wavenumbers for $n_s=4$, \cref{fig:cd_phase4}, a small constant phase error is observed, which could be due to the increasing numerical sensitivity of the RBFs at higher degrees. Numerical investigations will be required to determine what impact, if any, this has on the ultimate behaviour of the scheme.
\section{Introduction}\label{sec:intro}
The rapid and accurate numerical solution of partial differential equations (PDEs) is critically important across all disciplines of engineering and the sciences. Key applications include, amongst many others: aerospace engineering~\citep{Temam2001}; astrophyics~\citep{Harris1957}; and geophysical engineering~\citep{Morency2008}. The numerical solution of partial differential equations involves converting the continuous equations into a discrete difference form, which can then be solved approximately. The details of this discrete form --- its derivation --- are what distinguishes between different approaches to solving PDEs.
In recent years, the attention of the numerical modelling community has been turning towards high order methods, which are thought to offer the potential for computing highly accurate solutions at much reduced computational cost~\citep{Hesthaven2011}. One particularly promising area appears to be the family of Discontinuous Galerkin (DG) or Flux Reconstruction (FR) schemes \citep{Reed1973, Huynh2007}. These schemes have become popular because of their efficiency of implementation on modern computational hardware~\citep{Vermeire2015}, and the elegance of their mathematical representation.
Modern high performance computing facilities have continued to increase their theoretical peak count of floating point operations per second (FLOPS), broadly in line with Moore's Law~\citep{Schaller1997}. However, over the past decade or two these performance gains have largely come not from the increase in individual processor clock speeds, but from the inclusion of increased numbers of computational cores per processor~\citep{top500, Watson2016}, a trajectory which has been taken to extreme levels in clusters based on graphical processing units~\citep{Nvidia2020}. This has caused a new challenge: the memory bandwidth has not kept pace with FLOPS across many architectures, and consequently methods such as FR, which have a high ratio of local computation to remove communication, are well suited to this new computational paradigm~\citep{Iyer2021}.
Alongside the development of these methods, there has been an increased interest in the use of Radial Basis Functions (RBFs) to interpolate functions. Indeed, a number of approaches have used RBFs to solve complex partial differential equations across various disciplines of engineering~\citep{Flyer2007, Larsson2003, Hon1998}. These applications of RBFs generally fall into the category of so-called meshless methods, with two seminal works on the topic introducing the RBF finite difference (RBF-FD) method~\citep{Tolstykh2003} and the meshless Petrov--Galerkin approach~\citep{Atluri1998}. More recently, in the unpublished work of \citet{Glaubitz2021}, a global RBF method that draws upon the family of FR methods has been constructed and proven to be linearly stable. In this paper, we investigate the potential use of RBFs as an underlying method of functional approximation within the broader framework of solving PDEs via Flux Reconstruction.
This investigation is structured as follows. In \cref{sec:rbf}, we introduce RBFs and the techniques used to form both the basis and the subsequent interpolations. \cref{sec:fr} introduces the FR spatial discretisation and detail how RBFs can be applied to FR. In \cref{sec:fourier}, Fourier and combined Fourier analysis are used to investigate the linear stability and effect of RBFs on the numerical properties of FR. In \cref{sec:numerical}, a range of numerical experiments are performed to assess the effects on error and order, which culminate in an investigation of Burgers' turbulence. Finally, the conclusions of this work are drawn in \cref{sec:conclusions}, where insights into the future applications of this methodology are outlined.
\section*{Acknowledgements}\label{sec:ack}
\section*{Data Availability}
The data that informed the findings of this study are available from the corresponding author upon reasonable request.
\bibliographystyle{plainnat}
\section{Numerical Experiments}\label{sec:numerical}
As well as the Fourier-inspired analysis of the equations, the effect of the RBF functional approximation was also investigated using numerical experiments of two kinds --- first, linear problems, typified by the linear advection--diffusion equation, and secondly nonlinear problems, using Burgers' equation.
\subsection{Linear Advection-Diffusion}\label{sec:linearadvdiff}
The basic one dimensional linear advection--diffusion equation is given by:
\begin{equation}\label{eq:linearAdvection}
\px{u}{t} + a\px{u}{x} = \mu\pxi{2}{u}{x}, \quad \forall\; u(x,t):\Omega\times\mathbb{R}_+\mapsto\mathbb{R}
\end{equation}
Two sets of equation constants and initial conditions are used:
\begin{subequations}
\begin{align}
u(x,t=0) = \sin{x}&, \quad x\in\Omega=[-\pi,\pi), \quad \mathrm{with} \quad a=1, \mu=0\\
u(x,t=0) = \exp{\left(-x^2\right)}&, \quad x\in\Omega=[-10,10), \quad \mathrm{with} \quad a=1, \mu=0.1
\end{align}
\end{subequations}
which represent pure advection of a sine wave and advection--diffusion of a Gaussian bump. In both cases, the domain, $\Omega$, is taken as periodic.
To assess the resulting solutions, the point-mean $L^2$-norm of the error is used, with the approximate solution, $u^\delta(x,t)$, compared to the exact solution, $u(x,t)$. The following definition of the point-mean $L^2$ error norm is utilised:
\begin{equation}
\|\mathcal{E}\|_2 = \sqrt{\frac{1}{N \times n_s} \sum^N_{j=1}\sum^{n_s}_{i=1}\left(\boldsymbol{u}^\delta_j(x_i) - u(x_{j,i})\right)^2}
\end{equation}
for a case with $N$ elements and with each element composed of $n_s$ solution points.
Along with the the $L^2$ norm, the $L^1$ and $L^\infty$ norms were also calculated. These were both found to behave similarly to the $L^2$ norm across all the cases tested, and so they are not further reported here.
\subsubsection{Layout of the Solution Points\label{ssec:solnpointlocs}}
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$.]{\label{fig:n3_gad_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_pointLocations}}}
~
\subfloat[$n_s=4$.]{\label{fig:n4_gad_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_pointLocations}}}
\newline
\subfloat[$n_s=5$.]{\label{fig:n5_gad_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_pointLocations}}}
\caption{\label{fig:gad_locations}Effect of point layout on $L^2$ error for linear advection--diffusion, $\varepsilon = 0.5$.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$.]{\label{fig:n3_sa_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_pointLocations}}}
~
\subfloat[$n_s=4$.]{\label{fig:n4_sa_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_pointLocations}}}
\newline
\subfloat[$n_s=5$.]{\label{fig:n5_sa_locations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_pointLocations}}}
\caption{\label{fig:sa_locations}Effect of point layout on $L^2$ error for linear advection, $\varepsilon = 0.5$.}
\end{figure}
\cref{fig:gad_locations,fig:sa_locations} show the performance of the RBF-FR method using different point locations for a range of mesh spacings. \cref{fig:gad_locations} shows behaviour of the advection--diffusion of the Gaussian bump, and~\cref{fig:sa_locations} shows the performance of pure advection of the sine wave, both at a modest shape parameter value of $\varepsilon=0.5$, and across three solution point numbers, of $n_s = 3$, $4$, and $5$. The functional centres were located at the solution points for all the cases. For comparison, a set of solutions computed with the polynomial bases using the Legendre points is also supplied. As expected, there is a strong dependency of the accuracy on the point layouts, with close to an order of magnitude of error separating the best performing points from the worst performing sets. Typically, the full uniform points and the Lobatto points perform least well, and, in general, the Legendre points perform best. For the advection--diffusion equation, there is some suggestion in \cref{fig:gad_locations} that the better performing solution point layouts are capable of outperforming the polynomial bases on the very coarsest meshes, but this does not hold true as the meshes are refined. The RBF-FR also seems to perform better, relative to the polynomial baseline, with $3$ and $5$ solution points, performing noticeably more poorly on finer meshes for the $n_s = 4$ case, for both the advection--diffusion and pure advection test cases.
\cref{fig:gad_locations_01,fig:sa_locations_01} show the same set of runs as \cref{fig:gad_locations,fig:sa_locations}, but with the shape parameter of the radial bases reduced from $0.5$ to $0.1$. As would be expected, as the parameter is reduced towards zero, the performance of the fit follows the polynomial results more closely --- in the flat limit of $\varepsilon = 0$, the Gaussian RBF fit can be shown to be equivalent to a polynomial fit [REF]. In terms of the point locations, similar trends are observed with the wider Gaussian bases --- there is still close to an order of magnitude of difference between the best and worst performers. Again, the functions with colocated solution and flux points (the full uniform and Lobatto points) are seen to perform least effectively, and the Legendre points are again found to be the most effective of those tested. With the advection diffusion of the Gaussian bump, \cref{fig:gad_locations_01}, there is evidence that the best performing point locations outperform the polynomial FR scheme across the entire range tested. With the pure advection of the sine wave, \cref{fig:sa_locations_01}, the behaviour of RBF-FR is seen to be more complex. There are pronounced peaks and troughs in the error curves, which typically move the RBF-FR accuracy either side of the polynomial accuracy. Again, the general trend is that RBF-FR is able to outperform the polynomial method on the coarser meshes, but is overtaken as the meshes are refined.
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$.]{\label{fig:n3_gad_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_pointLocations_01}}}
~
\subfloat[$n_s=4$.]{\label{fig:n4_gad_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_pointLocations_01}}}
\newline
\subfloat[$n_s=5$.]{\label{fig:n5_gad_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_pointLocations_01}}}
\caption{\label{fig:gad_locations_01}Effect of point layout on $L^2$ error for linear advection--diffusion, $\varepsilon = 0.1$.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$.]{\label{fig:n3_sa_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_pointLocations_01}}}
~
\subfloat[$n_s=4$.]{\label{fig:n4_sa_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_pointLocations_01}}}
\newline
\subfloat[$n_s=5$.]{\label{fig:n5_sa_locations_01}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_pointLocations_01}}}
\caption{\label{fig:sa_locations_01}Effect of point layout on $L^2$ error for linear advection, $\varepsilon = 0.1$.}
\end{figure}
\subsubsection{Effect of the RBF Centre Locations}
In the previous sections, the functional centres of the RBFs have been placed at the same location as the solution points. Here, the effect of separating them is investigated.
\cref{fig:gad_cenLocations} shows these results, at two different values of shape parameter. The solution points are kept at the Legendre nodes, and the RBF centres are changed from the Legendre nodes to the full uniform and internal uniform distributions. From the plot, it appears that any influence made by moving the RBF centres is marginal, and, as might be expected, this influence reduces even further as the shape parameter decreases and the bases increasingly accurately approximate $\phi(r)=1$. In general, the results shown here for the advection--diffusion of the Gaussian bump suggest that there may be a small benefit in using the full uniform point set as the centres, but this effect is again marginal. It is possible that this small benefit is related to the effect of the functional centre locations on the condition number of the alternant matrix, and thus the numerical accuracy of calculating the operator matrices. This result, however, was not observed to be repeated for the pure advection of the sine wave, and, due to the marginal nature of any effect seen, further results are omitted here for brevity.
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$, Legendre points.]{\label{fig:n3_gad_cenLocations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_cenLocations}}}
~
\subfloat[$n_s=4$, Legendre points.]{\label{fig:n4_gad_cenLocations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_cenLocations}}}
\newline
\subfloat[$n_s=5$, Legendre points.]{\label{fig:n5_gad_cenLocations}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_cenLocations}}}
\caption{\label{fig:gad_cenLocations}Effect of varying the RBF centre locations on $L^2$ error for linear advection--diffusion.}
\end{figure}
\subsubsection{Effect of the Shape Parameter\label{ssec:shapefunction}}
Given its importance to RBF interpolation, the effect of reducing this scaling parameter is explored through the two test cases. In all cases, the best performing point locations from the previous section were used --- the Legendre points --- and the functional centres were located at the solution points.
\cref{fig:gad_varEpsilon,fig:sa_varEpsilon} show the effect of reducing this parameter on the performance of RBF-FR, again with a polynomial FR curve for comparison, for $n_s = 3$, $4$, and $5$. Generally, the accuracy of the RBF-FR solution was found to improve as the shape parameter is reduced. \cref{fig:gad_varEpsilon}, for the bump advection--diffusion case, suggests that each of the shape parameter values are capable of outperforming the polynomial fit across some range of grid densities. Larger values of $\varepsilon$ tend to outperform the polynomial on extremely coarse grids, but as the grid is refined, the polynomial approach overtakes them. As the value of the shape parameter is reduced, the range of mesh densities at which RBF-FR outperforms polynomial FR moves towards finer meshes. By the time a value of $\varepsilon = 0.1$ is reached, the RBF-FR outperforms the polynomial across the entire range of mesh densities tested, with close to a $50\%$ reduction in the error on the finest mesh tested for $n_s = 5$. The other observable trend is that as the shape parameter is reduced, the performance becomes increasingly like that of the polynomial FR.
\cref{fig:sa_varEpsilon} shows the results of the same range of tests, but for the pure advection of a sine wave. The RBF-FR performs much less competitively here, but similar trends can be observed --- higher values of shape parameter, so narrower bases, tend to outperform the polynomial scheme on more coarse meshes, and wider bases tend to outperform the polynomial FR on finer meshes. However, in all of these cases the polynomial FR outperforms RBF-FR for the finest meshes. Despite this, there are still large ranges of mesh densities in which the RBF-FR is highly competitive with the polynomial FR in terms of $L^2$ error norm, outperforming it by an order of magnitude in some cases. As seen with the advection diffusion of the Gaussian bump, as the shape parameter is reduced, the curves gradually trend towards the baseline polynomial error curve. With the baseline Gaussian RBF, it was not possible to use shape parameters significantly smaller than $\varepsilon = 0.1$, as the condition number of the system of equations requiring solution became too large, and numerical errors swamped the solution.
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$, Legendre points.]{\label{fig:n3_gad_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_varEpsilon}}}
~
\subfloat[$n_s=4$, Legendre points.]{\label{fig:n4_gad_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_varEpsilon}}}
\newline
\subfloat[$n_s=5$, Legendre points.]{\label{fig:n5_gad_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_varEpsilon}}}
\caption{\label{fig:gad_varEpsilon}Effect of shape parameter, $\varepsilon$, on $L^2$ error for linear advection--diffusion.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$, Legendre points.]{\label{fig:n3_sa_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_varEpsilon}}}
~
\subfloat[$n_s=4$, Legendre points.]{\label{fig:n4_sa_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_varEpsilon}}}
\newline
\subfloat[$n_s=5$, Legendre points.]{\label{fig:n5_sa_varEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_varEpsilon}}}
\caption{\label{fig:sa_varEpsilon}Effect of shape parameter, $\varepsilon$, on $L^2$ error for linear advection.}
\end{figure}
\subsubsection{Small Values of Shape Parameter\label{subsec:smallShapeParameters}}
Using RBF-Direct to compute the necessary operators limits the minimum possible value of $\varepsilon$. This issue can be sidestepped by making use of the RBF-GA method to compute the operators.
\cref{fig:gad_smallEpsilon,fig:sa_smallEpsilon} show the effect of using smaller values of $\varepsilon$ has on the performance of the scheme, made possible by RBF-GA. To test for consistency between the two methods, a value of $\varepsilon=0.1$ is repeated from the RBF-Direct approach, and is found to give the same results as the RBF-GA method for all cases --- this is as expected, given that both have been deliberately constructed to span the same functional approximation basis space.
At lower values of shape parameter, the sweep continues to trend towards the polynomial result, and pushes the range of mesh densities at which RBF-FR outperforms the polynomial scheme further to the left. For the pure advection of the sine wave, where RBF-FR has previously performed poorly relative to the polynomial scheme, very low values of shape parameter tend to produce more competitive results, with values of $\varepsilon \leq 0.025$ outperforming polynomial approximation across the entire range tested. For $n_s=5$, all values of $\varepsilon \leq 0.05$ outperform or match the polynomial scheme.
The use of RBF-GA to calculate the bases gives RBF-FR the ability to outperform polynomial FR on both test cases at no additional runtime cost, while retaining the geometric flexibility of both RBFs and FR in higher dimensions.
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$, Legendre points.]{\label{fig:n3_gad_smallEpsilonc}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_gad_smallEpsilon}}}
~
\subfloat[$n_s=4$, Legendre points.]{\label{fig:n4_gad_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_gad_smallEpsilon}}}
\newline
\subfloat[$n_s=5$, Legendre points.]{\label{fig:n5_gad_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_gad_smallEpsilon}}}
\caption{\label{fig:gad_smallEpsilon}Effect of very small shape parameters on $L^2$ error for linear advection--diffusion.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$, Legendre points.]{\label{fig:n3_sa_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n3_sa_smallEpsilon}}}
~
\subfloat[$n_s=4$, Legendre points.]{\label{fig:n4_sa_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n4_sa_smallEpsilon}}}
\newline
\subfloat[$n_s=5$, Legendre points.]{\label{fig:n5_sa_smallEpsilon}\adjustbox{width=0.49\linewidth, valign=b}{\input{Figs/tikz/n5_sa_smallEpsilon}}}
\caption{\label{fig:sa_smallEpsilon}Effect of very small shape parameters on $L^2$ error for linear advection.}
\end{figure}
\subsection{\label{ssec:bt}Viscous Burgers' Equation}
The viscous Burgers' equation represents a meaningful step towards the full Navier--Stokes equations. Burgers' equation is nonlinear, and is able to display turbulence-like phenomena, making it an insightful numerical test case as it exhibits integral, inertial, and dissipation ranges. Several other works have studied the phenomenon of Burgers' turbulence, notably those of \citet{San2016,Alhawwary2018,Frisch2013}. The primary relevant difference compared to the Navier--Stokes equations is that the energy cascade of Burgers' turbulence follows a $k^{-2}$ profile, compared to the famous $k^{-5/3}$ relation seen in Navier--Stokes.
For the viscous Burgers' equation:
\begin{equation}\label{eq:burgers}
\px{u}{t} + \frac{1}{2}\px{u^2}{x} = \mu\pxi{2}{u}{x} \quad \text{for} \quad u(x,t) : \Omega\times\mathbb{R}_+ \mapsto \mathbb{R}
\end{equation}
the initial condition proposed by \citet{San2016} is used, where energy is distributed among modes according to the spectrum given by:
\begin{equation}
E_0(k) = \frac{2}{3\sqrt{\pi}\rho}(k\rho)^4\exp{\left(-(k\rho)^2\right)}, \quad \mathrm{where} \quad \rho = \frac{1}{10}
\end{equation}
Use of these values gives the most energised mode as corresponding to $k=13$. For a periodic domain of $\Omega=[0,2\pi)$, the solution field may then be reconstructed from the spectrum as:
\begin{equation}
u_0(x) = \sum^{k_\mathrm{max}}_{k=0}\sqrt{2E_0(k)}\cos{(kx + 2\pi\Psi(k))}
\end{equation}
where $k_\mathrm{max} = 2048$ is found to be sufficient for double-precision floating-point accuracy. The term $\Psi(k)\in[0,1)$ is used to produce a random phase angle for each wavenumber, and for reproducibility the Mersenne twister algorithm~\citep{Matsumoto1998} is used.
The quantity of interest is the frequency response of the system, and to understand this the resulting energy spectrum at $t=0.1$ is studied. Time integration is performed using an RK4 scheme and, to enable discrete Fourier transforms, data is interpolated via the underlying basis to a set of equispaced points. For a uniform one-dimensional mesh with spacing $h$, uniform reference points are located at $\hat{x}\in\{-1+\hat{h}, -1+3\hat{h},\dots, -1+(2n-1)\hat{h}\}$, for $\hat{h}=2/(2n)$. The discrete Fourier transform results were averaged over 1000 runs, with the same 1000 random seeds used for the different initial states tested. Using the same 1000 seeds is important for objective comparison, as without this the behaviour at high frequencies can appear to be quite different between tests.
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$]{\label{fig:bt_pnt_03}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_points_n03}}}
~
\subfloat[$n_s=4$]{\label{fig:bt_pnt_04}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_points_n04}}}
~
\subfloat[$n_s=5$]{\label{fig:bt_pnt_05}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_points_n05}}}
\caption{\label{fig:bt_pnt}Effect of changing solution and collocated centre locations for Gaussian RBF with $\varepsilon=0.01$.}
\end{figure}
Investigating first the effect of varying the solution points (with colocated RBF centres), it can be seen from \cref{fig:bt_pnt} that there seems to be no significant impact on the energy spectra of the Burgers' turbulence for the point layouts of \cref{fig:pointLayouts}. In particular, it is known that the turbulent bottleneck~\citep{Fernandez2019} is caused by higher order numerical diffusion terms resulting from the discretisation. Therefore, if the scale of these terms was meaningfully changed by using either an RBF approximation space or by the point locations, it would show as a change in the bottleneck effect. The bottleneck is clearly visible in the approximate range $100<f<250$, and is unchanged here.
These tests were repeated, but fixing the solution points at the Legendre points and varying the locations of the RBF centres. The results of these tests are presented in \cref{fig:bt_ctr}. Once again, no significant variation in the Burgers' turbulence spectra is seen.
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$]{\label{fig:bt_ctr_03}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_centre_n03}}}
~
\subfloat[$n_s=4$]{\label{fig:bt_ctr_04}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_centre_n04}}}
~
\subfloat[$n_s=5$]{\label{fig:bt_ctr_05}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_centre_n05}}}
\caption{\label{fig:bt_ctr}Effect of changing solution and collocated centre locations for Gaussian RBF with $\varepsilon=0.01$.}
\end{figure}
\begin{figure}[tbhp]
\centering
\subfloat[$n_s=3$]{\label{fig:bt_eps_03}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_eps_n03}}}
~
\subfloat[$n_s=4$]{\label{fig:bt_eps_04}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_eps_n04}}}
~
\subfloat[$n_s=5$]{\label{fig:bt_eps_05}\adjustbox{width=0.32\linewidth, valign=b}{\input{Figs/tikz/bt_eps_n05}}}
\caption{\label{fig:bt_eps}Effect of varying $\varepsilon$ for Legendre solution points and internal uniform centres.}
\end{figure}
The final test undertaken with Burgers' turbulence was to again vary the RBF shape parameter, $\varepsilon$. The results shown in \cref{fig:bt_eps} were produced with the Legendre solution points and the internal uniform functional centres. This test clearly shows that $\varepsilon$ can have a profound impact on the behaviour of the method. As $\varepsilon$ is increased, these results appear to show that the method becomes unstable at high frequencies, with the width of the peak caused by this instability spreading. This is in agreement with the dissipation results of the Fourier analysis. Furthermore, as $\varepsilon$ is increased the location of the bottleneck is shifted towards higher frequencies. This is consistent with an overall reduction in the dissipation, and is again in agreement with the results of the Fourier analysis presented earlier.
\section{Radial Basis Functions}\label{sec:rbf}
In finite element methods it is common to form an approximation space in some subset of polynomial space. This has several advantages. Polynomials have been widely studied and analysed for centuries, and their strengths and weaknesses are fairly well understood (\textit{e.g.}~\citep{Brutman1978, Gasca2000, Boor1992}). Polynomials are generally quite flexible, and many useful finite element spaces can be formed with them --- such as Lagrange elements, Raviart--Thomas elements, and N\'ed\'elec elements. For most of these methods, it is advantageous to make use of a basis formed from an orthogonal set of polynomials, such as the Jacobi polynomials, $P_n^{\alpha, \beta} \left ( \boldsymbol{x} \right )$.
One possible alternative to the use of spaces built on polynomials is to make use of RBFs as the basis. RBFs are a class of functions whose value at depends solely on a distance metric --- typically the Euclidean norm --- of the evaluation point, $\boldsymbol{x}$, from the functional centre, $\boldsymbol{r_c}$. This distance metric is usually denoted as a radius, $r = \left \| \boldsymbol{x} - \boldsymbol{r_c} \right \|$. RBFs come in several flavours, and some of the most popular are shown in Table~\ref{tab:rbftypes} below.
\begin{table}[hbt!]
\centering
\caption{\label{tab:rbftypes}Common Radial Basis Functions.}
\begin{tabular}{llc}
\toprule
RBF & & $\phi(r)$ \\ \midrule \\[-10pt]
Gaussian & (GA) & $\exp{\left(- \varepsilon^2 r^2\right)}$ \\[5pt]
Multiquadratic & (MQ) & $\left(1 + \varepsilon^2 r^2\right)^{0.5}$ \\[5pt]
Inverse Quadratic & (IQ) & $\left(1 + \varepsilon^2 r^2\right)^{-1}$ \\[5pt]
Inverse Multiquadratic & (IMQ) & $\left(1 + \varepsilon^2 r^2\right)^{-0.5}$ \\[5pt]
Wendland$_{1,3}$ & (W13) & $\left ( 1 - r \right ) ^4_+ (4r + 1)$ \\[5pt]
\bottomrule
\end{tabular}
\end{table}
Radial Basis Functions have become popular for use in approximating functions both because they are conceptually simple and because they have been shown to be extremely effective at interpolating accurately from randomly scattered points~\citep{Wu1993}. Indeed, there are few theoretical limits to their ability to interpolate over arbitrary sets of points~\citep{Buhmann2003}. Owing to their flexibility, several schemes --- such as a mesh-free Petrov--Galerkin method which made use of the MQ-RBF as a basis~\citep{Atluri1998,Duan2008,Li2019} --- have been developed to exploit their properties. Other works have studied the properties of RBFs in the approximation of functions in various spaces and with reference to Discontinuous Galerkin, such as \citet{Wendland1997} and \citet{Wendland1999}, and these have shown that RBFs can be used to form an effective approximation space under certain conditions.
\cref{fig:Types_of_RBF} shows the approximate shape of various commonly seen members of the family of RBFs. All of the functions shown here, with the exception of the Wendland functions~\citep{Wendland1995}, are parameterised by $\varepsilon$, a constant known as the \emph{shape parameter}. Typically, as $\varepsilon$ is reduced, the radial bases become increasingly wide, and tend towards an approximation of $\phi(r) = 1$. The effect of decreasing the shape parameter on the shape of the different radial basis functions is shown for the Gaussian in \cref{fig:ShapeFunction_EffectOnRBF}.
\begin{figure}[htb]
\centering
\subfloat[Types of continuous RBF.]{\label{fig:Types_of_RBF}\adjustbox{width=0.49\textwidth,valign=b}{\input{./Figs/tikz/Types_of_RBF_Shape.tex}}}
~
\subfloat[Effect of the shape parameter, $\varepsilon$, on RBFs.]{\label{fig:ShapeFunction_EffectOnRBF}\adjustbox{width=0.49\textwidth,valign=b}{\input{./Figs/tikz/RBF_ShapeParameter_Shape.tex}}}
\caption{The behaviour of Radial Basis Functions.}
\end{figure}
To use radial bases to generate an analytic approximation to a function, $u(\boldsymbol{x})$, which has been evaluated at $N$ points, an $N \times N$ system of equations must be solved:
\begin{align}
\label{eq:RBF_AlternantMatrix}
\begin{pmatrix} \phi \left ( \left \| \boldsymbol{x}_1 - \boldsymbol{r_c}_1 \right \| \right) & \phi \left ( \left \| \boldsymbol{x}_1 - \boldsymbol{r_c}_2 \right \| \right) & \dots & \phi \left ( \left \| \boldsymbol{x}_1 - \boldsymbol{r_c}_N \right \| \right) \\ \phi \left ( \left \| \boldsymbol{x}_2 - \boldsymbol{r_c}_1 \right \| \right) & \phi \left ( \left \| \boldsymbol{r_c}_2 - \boldsymbol{x}_2 \right \| \right) & \dots & \phi \left ( \left \| \boldsymbol{x}_2 - \boldsymbol{r_c}_N \right \| \right) \\ \vdots & \vdots & \ddots & \vdots \\ \phi \left ( \left \| \boldsymbol{x}_N - \boldsymbol{r_c}_1 \right \| \right) & \phi \left ( \left \| \boldsymbol{x}_N - \boldsymbol{r_c}_2 \right \| \right) & \dots & \phi \left ( \left \| \boldsymbol{x}_N - \boldsymbol{r_c}_N \right \| \right) \end{pmatrix} \begin{pmatrix} \boldsymbol{\psi}_1 \\ \boldsymbol{\psi}_2 \\ \vdots \\ \boldsymbol{\psi}_N \end{pmatrix} = \begin{pmatrix} u(\boldsymbol{x}_1) \\ u(\boldsymbol{x}_2) \\ \vdots \\ u(\boldsymbol{x}_N) \end{pmatrix}
\end{align}
which is analogous to using a Vandermonde matrix to compute polynomial functional approximations, and which results in the functional approximation:
\begin{align}
\label{eq:RBF_Decomposition}
u(\boldsymbol{x}) \approx \sum^{N}_{i=1} \psi_i \phi \left ( \left \| \boldsymbol{x} - \boldsymbol{r_c}_i \right \| \right)
\end{align}
As the shape parameter is decreased towards zero, the alternant matrix governing the system of equations in \ref{eq:RBF_AlternantMatrix} approaches the matrix of ones, and so becomes increasingly ill-conditioned. However, there are several ways to avoid this ill-conditioning, which will be discussed more extensively later in the paper.
\section{Flux Reconstruction}\label{sec:fr}
To provide some necessary background, the basic procedure for Flux Reconstruction as applied to the linear advection--diffusion equation, as originally developed by \citet{Huynh2007}, is outlined in this section. The approach taken here largely follows that outlined by \citet{Witherden2014}. To begin, the linear advection--diffusion equation is written in strong conservation form:
\begin{align}
\label{eq:advectionDiffusionEquation}
\px{\boldsymbol{u}}{t} + \px{\boldsymbol{f}_i}{x_i} = 0
\end{align}
where:
\begin{align}
\boldsymbol{u} = \begin{pmatrix} u \end{pmatrix} \quad , \quad \boldsymbol{f} = \begin{pmatrix} a_i u - \mu_i \frac{\partial u}{\partial x_i} \end{pmatrix}
\end{align}
The computational domain, $\Omega$, is partitioned into $N$ compatible non-overlapping elements, $\Omega_n$, such that:
\begin{align}
\Omega = \bigcup^N_{n=1} \Omega_n \qquad \text{and} \qquad \bigcap^N_{n=1} \Omega_n = \emptyset
\end{align}
Each of these subdomains is, for computational efficiency, mapped onto a standard reference domain, $\Omega_n \to \Omega_s$, in which $\Omega_s \in \mathbb{R}^d$, where $d$ denotes the dimensionality of the problem. For example, the typical computational reference domain used for quadrilateral elements is $\Omega_s=[-1,1]^2$. Solution points are placed at specified locations within the domain, and flux points around its edges, as shown in \cref{fig:SolutionAndFluxLayout} for a one dimensional element. To formalise the transform between the \emph{real} and \emph{transformed} domains, $\Omega_n \to \Omega_s$, a Jacobian matrix, $\boldsymbol{J}_n$, is defined:
\begin{subequations}
\begin{align}
\boldsymbol{u} = \boldsymbol{u}(\hat{\boldsymbol{x}}, t) = \left | \boldsymbol{J}_n \right | u^\delta (\boldsymbol{x}, t) \\
\hat{\boldsymbol{f}}^\delta = \hat{\boldsymbol{f}}^\delta (\hat{\boldsymbol{x}}, t) = \left | \boldsymbol{J}_n \right | \boldsymbol{J}_n^{-1} \boldsymbol{f}^\delta (\boldsymbol{x}, t)
\end{align}
\end{subequations}
where the hat indicates transformed variables.
\begin{figure}[htbp]
\centering
\adjustbox{width=0.6\linewidth, valign=b}{\input{Figs/tikz/SolutionFluxPoints}}
\caption{\label{fig:SolutionAndFluxLayout}A one dimensional example of solution point and flux point layouts, $n_s=5$}
\end{figure}
The problem is assumed to be formulated as an initial value problem, so that the initial solution at $t=0$ is known at all the solution points. The conserved variables are approximated within each subdomain as a weighted sum of basis functions:
\begin{align}
\label{eq:SumofBasisFunctions}
\hat{u}^\delta (\hat{\boldsymbol{x}}) = \sum^{n_s}_{i=1} \hat{u}_i^\delta \theta_i(\hat{\boldsymbol{x}})
\end{align}
In one spatial dimension a Lagrange polynomial, for example, can be used as the basis function:
\begin{align}
\label{eq:LagrangeInterpolation}
\theta_k (\hat{x}) = \prod^{n_s}_{i=1,i \ne k} \frac{\hat{x} - \hat{x}_i}{\hat{x}_k - \hat{x}_i}
\end{align}
When working in higher dimensions, it is straightforward to extend this approach to tensor-product elements with maximal order bases, but other bases are better suited for simplices, such as the Dubiner basis for triangles~\citep{Dubiner1991}.
Once the conserved variables are functionally approximated within the subdomain based on their solution point values, these bases are projected to the flux points. This allows the values of the conserved variables at the flux points to be determined:
\begin{align}
\label{eq:SolutionToFluxConserved}
\hat{u}^\delta_F = \hat{u}^\delta (\hat{\boldsymbol{x}}_F)
\end{align}
The flux points of the compatible subdomains are colocated with those of their neighbours. As the projection from the solution points has been performed separately for each subdomain, the two sets of values of the conserved variables at each pair of colocated flux points are likely to be different.
The gradient of the conserved variables at each of the solution points is calculated in the transformed domain by simply differentiating the bases:
\begin{align}
\px{\hat{u}^\delta}{\hat{x}_j} = \sum^{n_s}_{i=1} \hat{u}_i^\delta \px{\theta_i(\hat{x})}{\hat{x}_j}
\end{align}
This gradient is formed from a continuous solution within each element, however, in order for the resulting scheme to be conservative between elements, it is necessary to correct the gradient using a common interface solution, $\hat{u}^c$, and then to propagate this into both cells by using a correction function, $g$. In one dimension, incorporating the corrections from the colocated flux points at both the left and right sides of the cell, the process is given by:
\begin{align}
\px{\hat{u}}{\hat{x}} = \sum^{n_s}_{i=1} \hat{u}_i^\delta \px{\theta_i(\hat{x})}{\hat{x}} + \left ( \hat{u}^c_L - \hat{u}^\delta_{F,L}\right ) \dx{g_L(\hat{x})}{\hat{x}} + \left ( \hat{u}^c_R - \hat{u}^\delta_{F,R}\right ) \dx{g_R(\hat{x})}{\hat{x}}
\end{align}
Applying this correction ensures that gradient approximation is in the point-wise $H^0_\mathrm{div}$ space, which is a regularity constraint that can be derived from the weak form of the equations.
A number of approaches for calculating the common interface values exist~\citep{Huynh2009}, but here a simple arithmetic mean of the two discontinuous flux point values at the collocated interface is suggested, so that, in the $i^{th}$ cell:
\begin{align}
\hat{u}^c_{L,i} = \frac{1}{2} \left ( \hat{u}^\delta_{F,L,i} + \hat{u}^\delta_{F,R,i-1} \right ) \qquad \text{and} \qquad \hat{u}^c_{R,i} = \frac{1}{2} \left ( \hat{u}^\delta_{F,R,i} + \hat{u}^\delta_{F,L,i+1} \right )
\end{align}
Now, the flux at each of the solution points can be calculated in the computational domain from the conserved variables, $\hat{\boldsymbol{f}}^\delta_s = \hat{\boldsymbol{f}}^\delta \left ( \hat{u}_s^\delta, \frac{\partial \hat{u}}{\partial \hat{x}_j} \big\vert_s \right )$, with the advective-diffusive fluxes being dependent on both the conserved variables themselves and their corrected gradients.
Once there is a functional approximation of the fluxes within the subdomain, these can be evaluated at the flux points, and simultaneously projected in the direction of the computational normal at the corresponding flux point to get the discontinuous normal computational fluxes, $\left ( \hat{\boldsymbol{f}} \cdot \hat{\boldsymbol{n}} \right )^\delta_F = \hat{F}^\delta_n $.
To ensure conservation is enforced between cells, a common interface normal flux is calculated at the collocated flux points, $\left ( \hat{\boldsymbol{f}} \cdot \hat{\boldsymbol{n}}\right )^c_F = \hat{F}^c_n$. As with the common interface value, there is a choice in how these fluxes are computed. Typically, the fluxes are divided into advective and diffusive components, such that $\left ( \hat{\boldsymbol{f}} \cdot \hat{\boldsymbol{n}}\right )^c_F = \hat{F}^c_n = \hat{F}^c_{n,\text{Adv}} + \hat{F}^c_{n,\text{Dif}}$. The contribution of the advective fluxes can then be estimated with an approximate Riemann solver, such as a Roe flux~\citep{Roe1981}, while a typical approach for the diffusive is taking the arithmetic mean of those calculated on the two discontinuous sides. When this approach is coupled to the use of the arithmetic mean for the common interface value, it is generally referred to as BR2~\citep{Bassi2000}.
Equipped with the common interface normal fluxes, the divergence of the fluxes can estimated within each cell in a manner which enforces conservation between cells. This is given in one dimension by:
\begin{align}
\label{eq:FR_advectionDiffusionEquation}
\frac{\partial \hat{\boldsymbol{f}}}{\partial \hat{x}} = \sum^{n_s}_{i=1} \hat{\boldsymbol{f}}_i^\delta \frac{\partial \theta_i(\hat{x})}{\partial \hat{x}} + \left ( \hat{F}^c_{n,L} - \hat{F}^\delta_{n,L}\right ) \frac{\partial g_L(\hat{x})}{\partial \hat{x}} + \left ( \hat{F}^c_{n,R} - \hat{F}^\delta_{n,R}\right ) \frac{\partial g_R(\hat{x})}{\partial \hat{x}}
\end{align}
With the approximation of the divergence of the fluxes from~\cref{eq:advectionDiffusionEquation} given by~\cref{eq:FR_advectionDiffusionEquation}, it is possible to advance the conserved variables in time through the use of a classical ODE integrator, such as a Runge--Kutta scheme.
\subsection{Correction Functions for FR}\label{ssec:correctionfunctions}
One of the defining features of flux reconstruction is its use of a correction function, $g$, to tie the individual cells together, thereby enforcing intercell continuity. The shape of this correction function has been shown to have a dramatic influence on the behaviour of the scheme~\citep{Vincent2011}.
There have been several investigations into novel families of correction functions, usually based on bounding the growth of some kind of energy norm~\citep{Castonguay2013,Vincent2010,Trojak2019}. The resulting schemes have proven to be highly effective in improving accuracy and maximising stability, but can be tricky to extend to non-tensor product shapes, because of their complex structures in higher dimensions. For example, the VCJH scheme~\citep{Vincent2011} on simplices in two dimensions:
\begin{equation}
\label{eq:VCJHcorrection}
c \sum^{n_s}_{k=1} \sigma_{f,k} \sum^{p+1}_{m=1} \begin{pmatrix} p \\ m - 1 \end{pmatrix} \left ( D^{(m,p)} \xi_i \right ) \left ( D^{(m,p)} \xi_k \right ) = -\sigma_{f,i} + \int_{\boldsymbol{\Gamma}_S} \left ( \boldsymbol{g}_f \cdot \bar{\hat{\boldsymbol{n}}} \right ) \xi_i d\Gamma
\end{equation}
where the operator $D^{(m,p)}$ is defined as:
\begin{equation*}
D^{(m,p)} = \frac{\partial^p}{\partial r^{p-m+1}\partial s^{m-1}}
\end{equation*}
and where $r$ and $s$ are some orthogonal coordinates. The coefficients of the divergence of the correction function corresponding to the $f^\mathrm{th}$ flux point are given by \cref{eq:VCJHcorrection}, such that:
\begin{equation}
\label{eq:correctionFunctionSum}
\nabla \cdot \boldsymbol{g}_f = \sum^{n_s}_{j=1} \sigma_{f,j} \xi_j \left ( \hat{\boldsymbol{x}} \right )
\end{equation}
where $\xi_j$ forms an orthonormal basis with support on the reference domain.
When the parametrisation constant, $c$, is set to zero, this function recovers a colocated nodal Discontinuous Galerkin scheme, but non-zero values can give enhanced performance. It has also been shown that modifying the correction function can be thought of as applying a filter to the collocated nodal Discontinuous Galerkin scheme~\citep{Zwanenburg2016}. However, when setting $c$ to zero in \cref{eq:VCJHcorrection} for computing the VCJH correction function, it collapses to the far simpler:
\begin{equation}
\label{eq:VCJHcorrection_DG}
\sigma_{f,i} = \int_{\boldsymbol{\Gamma}_S} \left ( \boldsymbol{g}_f \cdot \bar{\hat{\boldsymbol{n}}} \right ) \xi_i d\Gamma
\end{equation}
which gives the required coefficients for \cref{eq:correctionFunctionSum} directly. A further step eliminates the dependency of the right hand side on the correction function by noting that, to ensure continuity, $\left ( \boldsymbol{g}_f \cdot \bar{\hat{\boldsymbol{n}}} \right )$ must equal unity at flux point $f$ and zero at the others. This allows the term to be replaced by a flux point identity basis, which takes a value of one at its own flux point, and zero at every other:
\begin{equation}
\zeta^F_k \left ( \hat{\boldsymbol{x}}_f \right ) = \delta_{f,k}
\end{equation}
This finally results in an easily solvable set of coefficients for the correction functions:
\begin{equation}
\label{eq:VCJHcorrection_DG_simp}
\sigma_{f,i} = \int_{\boldsymbol{\Gamma}_S} \zeta^F_f \xi_i d\Gamma
\end{equation}
which simply require knowledge of an orthonormal basis over the transformed domain, and an identity basis over the flux points.
No requirement is placed on the functional form of the bases, $\xi_j$, used to compute the correction function field --- except for the requirement for orthonormality. However, the Gram--Schmidt process provides a tool to generate an orthonormal set of bases from any spanning set.
\subsection{Summation-by-parts}\label{ssec:sbp}
Before introducing the application of RBFs to FR, it is worth first taking a brief diversion into some approaches for testing the stability of the numerical method. From the work of \citet{Ranocha2016} and others, the FR method can be cast into the summation-by-parts (SBP) framework~\citep{Strand1994,DelReyFernndez2014}. Here, a one dimensional definition of the SBP framework is used, although higher dimensionality versions are also available~\citep{DelReyFernndez2014,Hicken2016}.
\begin{definition}[Summation-by-parts]
For a set of operators, $\mathbf{M}, \mathbf{D}, \mathbf{P}$, and $\mathbf{B}$, defined by:
\begin{subequations}\label{eq:sbp}
\begin{align}
\int_{\hat{\Omega}} \hat{v}\hat{w}\mathrm{d}x &\approx \langle\hb{v},\hb{w}\rangle_M \equiv \hb{v}^T\mathbf{M}\hb{w} \\ \px{\hat{w}}{x}\bigg|_{x=\hb{x}} &\approx \mathbf{D}\hb{w} \\
\mathbf{B} &= \mathrm{diag}{(-1,1)} \\
\mathbf{P}\hb{w} &\approx [\hb{w}_L, \hb{w}_R]^T
\end{align}
\end{subequations}
then these operators are said to be summation-by-parts operators if:
\begin{subequations}
\begin{align}
\int_{\hat{\Omega}}w\px{v}{x}\mathrm{d}x + \int_{\hat{\Omega}} v\px{w}{x}\mathrm{d}x = vw\big|_{\partial\hat{\Omega}} &\approx \hb{w}^T\mathbf{MD}\hb{v} + \hb{w}^T\mathbf{D}^T\mathbf{M}\hb{v}\\
\mathbf{MD} + \mathbf{D}^T\mathbf{M} &= \mathbf{P}^T\mathbf{BP}
\end{align}
\end{subequations}
\end{definition}
From \citet{Ranocha2018}, we state two lemmas on the conservation and stability of the FR method for linear advection equations:
\begin{lemma}[Conservation]\label{thm:sbp_cons}
Given a set of SBP operators, $\{\mathbf{M}, \mathbf{D}, \mathbf{P}, \mathbf{B}\}$ and the FR lifting operator, $\mathbf{C}$, then for the one dimensional linear advection equation this set is conservative if:
\begin{equation}
\mathbf{1}^T\mathbf{MC} = \mathbf{1}^T\mathbf{P}^T\mathbf{B}
\end{equation}
\end{lemma}
\begin{lemma}[Linear Stability]\label{thm:sbp_stab}
Given a set of SBP operators, $\{\mathbf{M}, \mathbf{D}, \mathbf{P}, \mathbf{B}\}$, and the FR lifting operator, $\mathbf{C}$, then for the one dimensional linear advection equation this set is linearly stable if:
\begin{equation}
\mathbf{C} = \mathbf{M}^{-1}\mathbf{P}^T\mathbf{B}
\end{equation}
\end{lemma}
These lemmas prove to be useful in testing the linear stability and conservation of schemes by casting the operators in the form of \cref{eq:sbp}, and will be used later.
\subsection{RBFs for Flux Reconstruction}\label{ssec:method}
RBFs can be used to form the basis of the approximation within the transformed cell in the same way as polynomials. For computational efficiency, most of the functional approximation and evaluation operations in FR are done on the transformed elements, $\boldsymbol{\Omega}_s$, where they can be expressed as matrix operations of the form:
\begin{align}
\hat{u}^\delta_F = \mathbf{M}_1 \hat{u}^\delta_S
\end{align}
which applies the $\mathbf{M}_1$ operator to compute the flux point conserved variables from the solution point conserved variables --- steps~\cref{eq:SumofBasisFunctions,eq:LagrangeInterpolation,eq:SolutionToFluxConserved} above. Critically, the operators themselves do not change during the calculation. To find $\mathbf{M}_1$, systems of equations are solved \emph{a priori} on the transformed domain, $\Omega_s$, and the required matrices are generated, which can then be applied at each time step.
Making the change from FR based on polynomial approximation to FR based on radial basis approximation essentially involves swapping out the polynomial operators and replacing them with operators derived using radial bases. The RBF-derived operators themselves are computed by manipulating the alternant matrix of~\cref{eq:RBF_AlternantMatrix}.
\begin{figure}[htb]
\centering
\adjustbox{width=0.8\linewidth, valign=b}{\input{Figs/tikz/AllRBFs_GaussianBump.tex}}
\caption{\label{fig:AllRBF_AdvGaussian}Advection of a Gaussian with RBF functional approximation.}
\end{figure}
\cref{fig:AllRBF_AdvGaussian} shows the use of various different radial bases, along with the baseline polynomial case, to advect and diffuse a Gaussian bump on a one dimensional domain. It can be seen in the figure that all of the different bases used are able to capture the basic behaviour of the wave as it travels across the domain. This suggests that RBFs are in principle able to serve as the underlying functional approximation method for flux reconstruction. Over the following sections, the behaviour and performance of radial basis function flux reconstruction (RBF-FR) will be explored in more detail.
To keep the scope manageable, the rest of this paper will focus solely on the use of Gaussian RBFs, where the interpolating function is given by $\phi(r) = \exp{(-\varepsilon^2 r^2)}$. The Gaussian is chosen because of its familiarity, its ease of manipulation, and because it has been shown to have some particularly intriguing properties in the flat limit~\citep{Fornberg2004a}, as the shape parameter $\varepsilon \to 0$.
\subsection{Investigating RBF-FR}
By changing the method of approximation within the transformed domains from polynomials to RBFs, various modifications to the simulation are possible. Here, we detail the modifications investigated in this work.
\begin{description}
\item[Solution Point Locations.]
The layout of the solution points within the reference cell is known to have an influence on the performance of flux reconstruction~\citep{Witherden2014a, Witherden2021}.
\begin{figure}[htbp]
\centering
\adjustbox{width=0.6\linewidth, valign=b}{\input{Figs/tikz/InterpolationPoints}}
\caption{\label{fig:pointLayouts}Canonical point layouts within the reference element.}
\end{figure}
\cref{fig:pointLayouts} shows the one dimensional solution point locations that have been tested here. The Legendre, Lobatto, and Chebyshev points are familiar from polynomial quadrature theory as the roots of the relevant functions. They have different advantages and disadvantages in terms of function approximation, and the Lobatto points in particular have some computational advantages in terms of FR. As seen in the figure, the Lobatto points include the end points. This means that the flux points are colocated with the terminal solution points, avoiding the need for projection to the flux points, and thereby offering a saving in compute time. In addition to these three familiar point layouts, two others have been included --- one, the `internal' uniform points, which are uniformly distributed over the inside of the cell, with half a spacing between the terminal solution points and the flux points, and two, the `full' uniform points, which are uniformly distributed over the entire cell, from flux point to flux point, and which also share the potential computational benefit of colocated flux and solution points with the Lobatto points.
Although distinct points are not able to form non-unisolvent sets in one dimension, this is something which may occur in higher dimensions. RBFs are more geometrically flexible than polynomials, so in two and higher dimensions, they are able to interpolate a wider number of scattered points. Two and higher dimensional cases are outside the scope of the present paper, but form a strong motivation for the development of RBF-FR as outlined here.
\item[RBF Functional Centre Locations.]
Along with their shape, radial basis functions are defined by the locations from which the distance $r$ is measured --- the functional centres of the bases. Frequently in the use of RBFs, the functional centres are chosen to be colocated with the measuring points, but this is not a requirement. To investigate the influence of different sets of functional centres, the same set of points shown in \cref{fig:pointLayouts} are used.
\item[Shape Parameter.]
The shape parameter, $\varepsilon$, or equivalently, the width of the basis, is known to have a significant effect on the performance of RBF interpolation. Indeed, a number of investigations have been undertaken to tune the value of the shape parameter to give the best interpolating performance~\citep{Fasshauer2007,Bayona2011,Davydov2011}, and so the influence of this parameter is explored here. It is also possible to have different bases take different shape parameters within the same functional interpolation. This kind of local variation in $\varepsilon$ is not considered further within this paper, but it remains something which may be of interest for future research.
\item[Small Values of Shape Parameter.]
Although the value of the shape parameter is critical to the performance of an RBF interpolation, the RBF alternant matrix shown in \cref{eq:RBF_AlternantMatrix} becomes increasingly ill-conditioned as the the shape parameter decreases, becoming equal to the matrix of ones in the limit $\varepsilon \to 0$. \cref{fig:condNumber_RBFDirect} shows this variation in condition number with shape parameter. The condition number is also more weakly affected by the location of the functional centres, an effect which is related to the mean distance between them. A general rule-of-thumb is that a for every order of magnitude the condition number increases by, a digit of precision is lost in the solution. This means that the system of equations cannot reliably be solved for smaller values of shape parameter. However, it is not the basis \emph{itself} that is ill-conditioned --- rather, it is an intermediate step in calculating it which is to blame.
\begin{figure}[tbhp]
\centering
\adjustbox{width=0.6\linewidth, valign=b}{\input{Figs/tikz/conditionNumber_01}}
\caption{\label{fig:condNumber_RBFDirect} Effect of reducing $\varepsilon$ on alternant matrix condition number when using direct Gaussian RBFs.}
\end{figure}
To this end, various methods have been developed which circumvent this step to allow an effective basis to be produced when using very flat radial bases. Examples of ways in which this has been achieved are the Contour-Padé method~\citep{Fornberg2004} and the RBF-QR method~\citep{Fornberg2011}. Here, since the investigation is restricted to Gaussian radial bases, an approach known as the RBF-GA method~\citep{Fornberg2013} is used to build the basis functions without the need to directly solve the system of equations. The details of the approach are omitted here for brevity, but it involves expanding the original Gaussian functions into a Taylor series, and then selecting linear combinations which cause most of these terms to cancel out. The result is a new basis which, critically, spans exactly the same space as the original, but which is well conditioned even at very low values of $\varepsilon$.
\cref{fig:condNumber_RBFGA} shows the effect on the condition number of the solution point alternant matrix of calculating the basis via the RBF-GA method, rather than the RBF-Direct approach. The same dependence on the mean spacing between the functional centres is observed, but in general the condition number of the alternant is below $10^3$, meaning that by side-stepping the direct solution, it becomes numerically feasible to use far smaller values of the shape parameter.
\begin{figure}[tbhp]
\centering
\adjustbox{width=0.80\linewidth, valign=b}{\input{Figs/tikz/conditionNumber_02}}
\caption{\label{fig:condNumber_RBFGA} Effect of reducing $\varepsilon$ on alternant matrix condition number when using RBF-GA.}
\end{figure}
\end{description}
|
\section{Introduction} \label{sec1}
Work in nonequilibrium systems is a fundamental research topic and has stimulated many interesting research directions in statistical physics \cite{Nonequilibrium1,Nonequilibrium2,Nonequilibrium3}. Quantum work distribution in a thermally isolated system can be effectively determined by the beginning-time and end-time energy measurements. Previously, there are many explorations on quantum work distribution in the nonequilibrium quantum system, both theoretically \cite{QHO,QP,work-distribution,work-distribution1,work-distribution4} and experimentally \cite{work-distribution2,work-distribution3,work-distribution5}. However, these results mainly focused on the single-particle systems.
In recent years, multiparticle work distribution in nonequilibrium processes has received more and more attentions \cite{many-particle,many-particle0,many-particle1,goold2018role}.
The calculation of work distribution for an identical many-particle system involves the transition probability between multiparticle eigenstates, which could be formidable difficulty due to the interference influence of these particles \cite{Interference0,Interference1,Interference2}. The transition probability between eigenstates of multiple identical bosons (fermions) associates with a permanent (determinant) of the corresponding transition amplitude matrix. The determinant can be efficiently computed, while evaluating the permanent is a so-called `\#P-complete' hard \cite{P-complete,P-complete1}. It has long been believed to be classically intractable even using the state-of-the-art simulation algorithms \cite{algorithm1,algorithm2}.
Quantum boson sampling, a remarkably quantum computational supremacy candidate \cite{quantum-supremacy}, was proposed by Aaronson and Arkhipov in 2011 \cite{Aaronson}. Boson sampling emerges as a powerful paradigm to efficiently estimate the permanent of a matrix, and provides several potential practical applications in graph theory \cite{graph1,graph2}, decision and function problems \cite{Decision,Decision1}, and quantum chemistry \cite{chemistry1,chemistry2}. An ideal boson sampling machine can be created by injecting indistinguishable bosons into a linear optical network, then measuring and recording the photon count in each output mode by the photon-number-resolving detectors. The output probability distribution to be sampled is proportional to the permanent of a related sub-matrix given by the linear optical interferometer. The early proof-of-principle demonstration of boson sampling has been realized experimentally \cite{QBS1,QBS2,QBS3,QBS4,QBS5}. The calculated
matrix permanent is consistent with the classical simulation. However, calculating the permanent of a unitary matrix on a classical computer takes the time scaling exponentially of the size of the matrix \cite{Aaronson}. It has been theoretically estimated that computing a 20$\times$20 (30$\times$30) matrix permanent requires billions (trillions) of operations \cite{Aaronson}, which goes beyond the capability of today's most powerful classical computer.
But for boson sampling experiments, it is believed to be possible in a near future.
Photons are considered as an outstanding candidate to realize boson sampling because of the long coherence times, single-particle level flexibility, and robustness against the decoherence. So far, the boson sampling machines with high-performance single-photon sources \cite{scource1,scource2,scource3} and ultra-low-loss photonic circuits \cite{Reck,Clements,2021potential} have been well-developed and they improved the multi-photon sampling rates \cite{QBS5,QBS6,QBS7}. In 2018, Wang \emph{et al.} proposed boson sampling with photon loss \cite{QBS6}. Recently, boson sampling involved 20-photon and 60-mode in a higher-dimensional Hilbert space was experimentally realized \cite{QBS7}. The scaling of boson sampling later reached 76 detected photons and 100 modes that demonstrated a quantum computational advantage over the classical computers \cite{supremacy}. Nowadays, the largest Gaussian boson sampling experiment with up to 113 photon detection events and 144 modes has been experimentally reported \cite{Interofeter}. Scalable implementations of boson sampling utilizing phononic modes of trapped ions have also be proposed \cite{Ions1,Ions2}.
In this paper, we investigate how to calculate the work distribution of multiple non-interaction identical bosons in a one-dimensional quantum piston by a boson sampling machine. We first present a general theory for the work distribution of multiple bosons and establish a connection between the work distribution and the boson sampling. The transition amplitude matrix of the multi-boson eigenstates can be obtained by the numerical calculation, and we extend it into a near-unitary matrix according to its normalization property. The matrix then is programmed into an optical network to estimate the transition probability between the eigenstates through the measurement of output photons. As an example, we design a specific scheme to calculate the work distribution of three identical bosons. Our scheme can be easily extended to calculate the work distribution of tens of bosons.
As the number of bosons scales up, the polynomial resource costs are required in terms of the number of optical elements. Scaling up to larger number of bosons would provide a strong evidence for quantum-enhanced computation.
\section{Results} \label{Sec2}
\subsection{Work distribution for multiple identical bosons} \label{Sec2.1}
\begin{figure
\begin{center}
\includegraphics[width=8.1 cm,angle=0]{quantum-piston.pdf}
\caption{Schematic diagram of performing work on the multiple identical bosons in a one-dimensional quantum piston. \textbf{a} Multiple identical bosons are initially prepared in a thermal equilibrium state, and trapped in a keep-temperature quantum piston with an initial length $\lambda_0$. The piston performs work on the bosons by pulling towards the right at a constant speed $v$. \textbf{b} The length of piston becomes $\lambda_\tau$ at final time $t=\tau$. \textbf{c} Initial bosons distribution over energy levels at time $t=0$. \textbf{d} Final bosons distribution over energy eigenstates at $t=\tau$ after the work being performed.} \label{quantum-piston}
\end{center}
\end{figure}
We first review some preliminaries about quantum work distribution of many identical particles \cite{many-particle,many-particle0,many-particle1}. As shown in Fig. \ref{quantum-piston}, we consider that $N$ identical bosons in a stretchable one-dimensional quantum piston, whose Hamiltonian $\hat{H}^\lambda$ has an energy spectrum
\begin{align} \label{eq1}
\hat{H}^\lambda|i^\lambda\rangle=E_i^\lambda|i^\lambda\rangle.
\end{align}
First, suppose that the system is initially prepared in an equilibrium thermal state with a fixed piston length $\lambda(0)=\lambda_0$ at inverse temperature $\beta=1/(k_BT)$, see Fig. \ref{quantum-piston}a. Here $k_B$ is the Boltzmann constant and $T$ is the temperature of the system. The initial bosons distribution over energy levels is described in Fig. \ref{quantum-piston}c. After performing a projective measurement on energy eigenbasis \cite{two-time}, the system collapses into a multi-boson eigenstate $|i_k^{\lambda_0}: n_{i_k}\rangle$. The probability of the eigenstate is
\begin{align} \label{eq2}
P(|i_k^{\lambda_0}: n_{i_k}\rangle)=\frac{1}{Z^{\lambda_0}}\text{exp}\big[{-\beta\big(\sum_{k} n_{i_k}E_{i_k}^{\lambda_0}\big)}\big],
\end{align}
where $Z^{\lambda_0}$ is a partition function
\begin{align} \label{eq3}
Z^{\lambda_0}=\sum_{i_k}\text{exp}\big[{-\beta\big(\sum_{k} n_{i_k}E_{i_k}^{\lambda_0}\big)}\big].
\end{align}
Here, $i_k^{\lambda_0}$ denotes the $k$th boson in the $i$th eigenstate at initial time, and $n_{i_k}$ represents the occupation number of bosons in the $i_k$th eigenstate. $E_{i_k}^{\lambda_0}$ is the $i_k$th eigenenergy.
\begin{figure*
\begin{center}
\includegraphics[width=17 cm,angle=0]{sampling.pdf}
\caption{Schematic diagram for calculating the work distribution of multiple identical bosons in a boson sampling machine. \textbf{a} The transition amplitude matrix $\Lambda$ is encoded into a boson sampling device by tuning the interferometer parameters. The transition probability between the multi-boson eigenstates is simulated by reading out and counting the expected photon distribution probability. The work distribution can be calculated according to Eq. (\ref{eq7}). \textbf{b} Variable beam splitter, a key building block for establishing the programmable linear optical interferometer. Each line is an optical mode and the crossing line between two modes denotes a variable beam splitter. The variable beam splitter acting on modes $a$ and $a+1$ is represented by a matrix $T_{a,a+1}(\theta,\varphi)$, which can be realized by two 50:50 beam splitters and two phase shifters with rotated angles $\varphi$ and $\theta$.} \label{sampling}
\end{center}
\end{figure*}
Then, as shown in Fig. \ref{quantum-piston}b, the piston is pulled to $\lambda(\tau)=\lambda_\tau$ at a constant speed $v$ under an external drive. At time $t=\tau$, the energy of the system is measured again, and the bosons distribution over energy eigenstates is presented in Fig. \ref{quantum-piston}d. We assume that the system collapses into the final multi-boson eigenstate $|f_l^{\lambda_\tau}: n_{f_l}\rangle$ with a probability of $P\big(|i_k^{\lambda_0}: n_{i_k}\rangle\rightarrow|f_l^{\lambda_\tau}: n_{f_l}\rangle\big) $, which has a complex mathematical structure
for the classical computer. The work performed during this process can be written as
\begin{align}
\begin{split} \label{eq6}
&W\big(|i_k^{\lambda_0}: n_{i_k}\rangle\rightarrow|f_l^{\lambda_\tau}: n_{f_l}\rangle\big)\\&=\sum_{l} n_{f_l}E_{f_l}^{\lambda_\tau}-\sum_{k} n_{i_k}E_{i_k}^{\lambda_0}.
\end{split}
\end{align}
In this way, we can define the the work distribution as \cite{Tasaki,Talkner}
\begin{align}
\begin{split} \label{eq7}
\rho(W)=&\sum_{i_k}\sum_{f_l}P\big(|i_k^{\lambda_0}: n_{i_k}\rangle\big) \\&
\times P\big(|i_k^{\lambda_0}: n_{i_k}\rangle\rightarrow|f_l^{\lambda_\tau}: n_{f_l}\rangle\big) \\&
\times \delta\big(W-\sum_{l} n_{f_l}E_{f_l}^{\lambda_\tau}+\sum_{k} n_{i_k}E_{i_k}^{\lambda_0}\big).
\end{split}
\end{align}
Next, we show how to compute numerically the transition probability $P\big(|i_k^{\lambda_0}: n_{i_k}\rangle\rightarrow|f_l^{\lambda_\tau}: n_{f_l}\rangle\big)$. As described above, the piston system evolving from time $t=0$ to $t=\tau$ follows the time-dependent Schr\"{o}dinger equation $i\hbar\partial_t\hat{U}=H(t)\hat{U}$. $\hat{U}$ denotes an evolutionary unitary operator. A complete orthogonal solution set of this Schr\"{o}dinger equation can be expressed as \cite{Schrodinger}
\begin{align}
\begin{split} \label{eq8}
\Phi_j(x,t)=&\text{exp}\big[\frac{i}{\hbar\lambda(t)}\big(\frac{1}{2}Mvx^{2}-E_j^{\lambda_0}\lambda_0t\big)\big]\\&\times\phi_j(x,\lambda(t)).
\end{split}
\end{align}
The time-dependent Schr\"{o}dinger equation has the general solution of the following form
\begin{align}
\begin{split} \label{eq9}
\Psi(x,t)=\sum_j^\infty c_j\Phi_j(x,t).
\end{split}
\end{align}
Here, $j=1, 2, \cdots$, and $x$ is the length of the piston with the time change, $0\leq x\leq \lambda(t)$. The $j$th eigenenergy $E_{j}^{\lambda_0}$ is given by $E_{j}^{\lambda_0}=\frac{(j\pi\hbar)^2}{2M{\lambda_0}^2}$, and $M$ is the mass of the boson.
The $j$th eigenstate of a boson in the piston system is
\begin{align} \label{eq10}
\phi_j(x,\lambda)=\sqrt{\frac{2}{\lambda}}\text{sin}\big(\frac{j\pi x}{\lambda}\big).
\end{align}
The coefficients $c_j$ of the solution of Schr\"{o}dinger equation in Eq. (\ref{eq9}) can be determined by the initial condition. That is, under the initial condition $\Psi(x,0)=\phi_{i_k}(x,\lambda_0)=\langle x|i_k^{\lambda_0} \rangle$ (taking $M=\hbar=1$), the coefficients become
\begin{align} \label{eq11}
c_j(i_k)=\frac{2}{\lambda_0}\int_0^{\lambda_0}e^{-i\frac{vx^2}{2\lambda_0}}\text{sin}\big(\frac{j\pi x}{\lambda_0}\big)\text{sin}\big(\frac{i_k\pi x}{\lambda_0}\big)dx.
\end{align}
The transition amplitude between eigenstates of single-particle from time $t=0$ to $t=\tau$ can be expressed as
\begin{align}
\begin{split} \label{eq12}
\langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle=\sum_{j=1}^\infty c_j(i_k)\int_0^{\lambda_\tau}\Phi_j(x,\tau)\phi_{f_l}^{*}(x,\lambda_\tau)dx.
\end{split}
\end{align}
We therefore can obtain the transition probability between the multi-boson eigenstates by the single-particle one \cite{Interference0,Interference1}
\begin{align}
\begin{split} \label{eq13}
&P\big(|i_k^{\lambda_0}: n_{i_k}\rangle\rightarrow|f_l^{\lambda_\tau}: n_{f_l}\rangle\big)\\&=\prod_{k=1}^K\frac{1}{n_{i_k}!}\prod_{l=1}^L\frac{1}{n_{f_l}!}\bigg|\text{Per}\big(\Lambda^\text{(I,F)})\bigg|^2.
\end{split}
\end{align}
Here, the function Per(M) represents the permanent of a matrix $M$. $\Lambda$=$\big(\langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle\big)$ is an ${L}\times {K}$ matrix. $I$=$(0,\cdots,0,n_{i_k},0,\cdots,0)$ is an initial arrangement of the boson number in eigenstate ${i_k}$ and $F=(0,\cdots,0,n_{f_l},0,\cdots,0)$ is the final arrangement of the boson number in eigenstate ${f_l}$ in the system. $\Lambda^\text{(I,F)}$ denotes a sub-matrix of $\Lambda$ by taking $n_{f_l}$ copies of the $f_l$th row and $n_{i_k}$ copies of the $i_k$th column of $\Lambda$, which occupies a dimension of $n_{f_l}\times n_{i_k}$. Since the total number of bosons is conserved, we have the identity relation $\sum_{k=1}^{K}n_{i_k}=\sum_{l=1}^{L}n_{f_l}=N$.
\subsection{Calculating the transition amplitude matrix with a boson sampling}\label{Sec2.2}
As we discussed above, it is necessary for the calculation of work distribution to estimate the transition probability between the multi-boson eigenstates, but the transition probability relates to the permanent of a matrix. Calculating the matrix permanent with classical computer requires exponential scale in time with respect to the size of the matrix. The matrix permanent cannot be efficiently simulated by any classical computers when the matrix dimension exceeds 20-30 \cite{Aaronson}. Fortunately, the intractable problem could be handled by boson sampling \cite{QBS1,QBS2,QBS3,QBS4,QBS5}.
As shown in Fig. \ref{sampling}, we design a schematic setup for calculating the work distribution of multiple identical bosons by boson sampling. We first construct the transition amplitude matrix $\Lambda$ from the transitions of single-boson eigenstates in the quantum piston system. The matrix elements $\langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle$ described in Eq. (\ref{eq12}) can be computed numerically, which depend on the parameters $\lambda_0, \lambda_\tau$, and $v$. The transition amplitude matrix $\Lambda$ has the normalization property, because of
\begin{align}
\begin{split} \label{eq14}
&\sum_{i_k}\big|\langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle\big|^2\\&=\sum_{i_k} \langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle\langle i_k^{\lambda_0}|\hat{U}^{\dag}|f_l^{\lambda_\tau}\rangle=1,
\end{split}
\end{align}
\begin{align}
\begin{split} \label{eq15}
&\sum_{f_l}\big|\langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle\big|^2\\&=\sum_{f_l} \langle f_l^{\lambda_\tau}|\hat{U}|i_k^{\lambda_0}\rangle\langle i_k^{\lambda_0}|\hat{U}^{\dag}|f_l^{\lambda_\tau}\rangle=1.
\end{split}
\end{align}
To encode the matrix $\Lambda$ into a linear optical interferometer composed of beam splitters and phase shifters, the matrix $\Lambda$ should be further restricted to a unitary matrix (the product of $\Lambda$ and its Hermitian conjugate is an identity matrix). Based on the normalized property of the matrix $\Lambda$, the dimension of matrix can be determined in principle. We evaluate the unitary property of the matrix $\Lambda$ according to the fidelity, defined as \cite{Distance-measure}
\begin{align}
\begin{split} \label{eq16}
\mathcal{F}=\frac{1}{d}\bigg|tr\sqrt{I_d^{1/2}\sigma I_d^{1/2}}\bigg|.
\end{split}
\end{align}
Here, $d$ is the dimension of $\Lambda$ and $I_d$ is a $d$-order identity matrix. $\sigma=\Lambda\Lambda'$ and $\Lambda'$ is the Hermitian conjugate of $\Lambda$. We numerically simulate the matrix elements and truncate the size of matrix $\Lambda$ when the fidelity reaches 0.995 with fixed parameters $\lambda_0=1$, $\lambda_\tau=2$ and a change speed $v$. The relationship between the matrix dimension and expansion speed $v$ is plotted in Fig. \ref{dim-v}. We also present the relationship between the matrix dimension and the final length of piston in Fig. \ref{dim-length}. One can see that the dimension of the matrix $\Lambda$ increases almost linearly with the acceleration of the piston speed $v$ or the final stretched length of piston $\lambda_\tau$.
\begin{figure
\begin{center}
\includegraphics[width=8 cm,angle=0]{dim-v.pdf}
\caption{The relation between the dimension of the matrix $\Lambda$ and the piston expansion speed $v$ when the unitary fidelity reaches $\mathcal{F}$= 0.995. The parameters $\lambda_0=1$ and $\lambda_\tau=2$ are taken.} \label{dim-v}
\end{center}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[width=8.2 cm,angle=0]{dim-final-length.pdf}
\caption{The relation between the dimension of the matrix $\Lambda$ and the final length $\lambda_\tau$ of piston when the unitary fidelity reaches $\mathcal{F}$= 0.995. The parameters $\lambda_0=1$ and $v=1$ are taken. } \label{dim-length}
\end{center}
\end{figure}
The next step is to program the transition amplitude matrix $\Lambda$ into a square-shaped photonic interferometer, see Fig. \ref{sampling}a. The symmetry design of the interferometer is more robust against photon loss tolerant, and has high stability and minimal optical depth \cite{Clements}. As shown in Fig. \ref{sampling}b, the crossing between two optical modes $a$ and $a+1$ in the interferometer consists of two 50:50 beam splitters and two phase shifters, and it can be expressed mathematically by a matrix $T_{a,a+1}(\theta,\varphi)$ \cite{Clements}. $T_{a,a+1}(\theta,\varphi)$ is an elimination matrix, which can be obtained by replacing the entries of an identity matrix with the same size as $\Lambda$ at the $a$th and $(a+1)$th rows and the $a$th and $(a+1)$th columns with
\begin{align} \label{eq17}
\left(\begin{array}{cc}
e^{i\varphi}\cos\theta & -\sin\theta\\
e^{i\varphi}\sin\theta & \cos\theta\\
\end{array}\right),
\end{align}
and the rest of the other entries remain unchanged. Based on Gaussian elimination method, the matrix $\Lambda$ can be diagonalized into a diagonal matrix $D$ by multiplying a series of $T_{a,a+1}$ and its inverse matrixs $T_{a,a+1}^{-1}$. The matrix $\Lambda$ is realized physically in an optical network by choosing suitable values of parameters $\theta$ and $\varphi$ of $T_{a,a+1}$ and the phase values of diagonal matrix $D$ at each output port. The resource overhead for programming a $d\times d$ unitary matrix into the interferometer requires $d(d-1)$ 50:50 beam splitters and $d^2$ phase shifters. From Fig. \ref{dim-v} and Fig. \ref{dim-length}, one can see that the total resource costs present at a polynomial hierarchy with the stretching speed $v$ or the final length of piston $\lambda_\tau$ in terms of the number of required optical elements.
\subsection{Correspondence and an example}\label{Sec4}
\begin{table*} [htb]
\centering \caption{The analogies between transition probability between multiple bosons eigenstates and the boson sampling.}
\begin{tabular}{lp{7.7cm}p{7.7cm}
\hline
\hline
&Transition probability & Boson sampling \\
\cline{1-2}
\hlin
&Transition amplitude matrix $\Lambda$ & Optical interferometer \\
&Dimension of the $\Lambda$ & Modes of interferometer \\
&Photon distribution $I$ at $t=0$ & Output target state $|S\rangle$ \\
&Photon distribution $F$ at $t=\tau$ & Input photon state $|T\rangle$ \\
&Per\big($\Lambda^\text{(I,F)}$\big) &Per\big($\Lambda^\text{(S,T)}$\big) \\
&Transition probability between multi-boson eigenstates &Frequency of target $|S\rangle$ in $n_{f_l}$-fold coincidences \\
\hline\hlin
\end{tabular}\label{table1}
\end{table*}
We are allowed to get a sub-matrix $\Lambda^\text{(I,F)}$ of the transition amplitude matrix $\Lambda$ to calculate the work distribution of multiple bosons in the piston. To determine the sub-matrix $\Lambda^\text{(I,F)}$, the projective energy measurement on photons and the standard photon-distribution measurement are performed to find the energy eigenstates and the distribution of the photon count in the realistic experiment before and after the system evolves \cite{work-distribution3,two-time}. We suppose that there are $N$ bosons are initially prepared in the piston system, with $n_{i_k}$ bosons in the $i_k$th eigenstate. The occupation distribution of particles in the each eigenstate is $X=(n_{i_1}, n_{i_2}, \cdots, n_{i_k}, \cdots)$. After the work is done, the boson-distribution measurement is performed again, and the final occupation distribution in each eigenstate is denoted by $Y=(n_{f_1}, n_{f_2}, \cdots, n_{f_l}, \cdots)$, where $\sum_{k}n_{i_k}=\sum_{l}n_{f_l}=N$. During this process, the transition probability from the $i_k$th eigenstate to the $f_l$th eigenstate can be obtained by computing the permanent of a sub-matrix $\Lambda^\text{(I,F)}$. Here, $I=(0,\cdots,0,n_{i_k},0,\cdots,0)$ and $F=(0,\cdots,0,n_{f_l},0,\cdots,0)$.
To calculate the permanent of the $\Lambda^\text{(I,F)}$ with boson sampling, we first prepare the input photons in the Fock state $|\text{T}\rangle=|t_1, t_2, \cdots, t_m\rangle$. Here, $m$ is the input bosonic modes which is equal to the dimension of matrix $\Lambda$. Each $t_i$ denotes boson occupation-number in the $i$th optical mode and $|\text{T}\rangle$ describes $n_{f_l}=\sum_{i=1}^{m}t_i$ bosons distribution. The input boson number in each mode should be consistent with the number of distributed bosons $F$ at final time. After $n_{f_l}$ identical bosons are scattered and propagate through the linear optical interferometer that is programmed as matrix $\Lambda$, the $n_{f_l}$-fold coincident events are collected and recorded by the single-photon counting detectors connected to the output modes. The measured relative frequency of the desired event $|\text{S}\rangle=|s_1, s_2, \cdots, s_m\rangle$ for every allowed outcome is just the probability \cite{QBS1,QBS2,QBS3,QBS4}
\begin{align}
\begin{split} \label{eq18}
P\big(S\,|\,T\big)=\prod_{j=1}^m\frac{1}{s_{j}!}\prod_{i=1}^m\frac{1}{t_{i}!}\bigg|\text{Per}\big(\Lambda^\text{(S,T)})\bigg|^2.
\end{split}
\end{align}
Each $s_j$ is the measured photon number in the $j$th output mode, and the desired event $S$ corresponds to $I$. The results of Eqs. (\ref{eq13}) and (\ref{eq18}) show a highly good agreement between the transition probability and the experimental boson sampling. Therefore, the complex calculation for the transition probability between the multi-boson eigenstates can be solved well by a boson sampling system. A clear relationship between the transition probability of multiple bosons eigenstates and the boson sampling concepts is listed in Tab. \ref{table1}.
As an example, we calculate the work distribution of three bosons in details. We consider three bosons are trapped in a piston with the initial length $\lambda_0=1$, constant stretching speed $v=0.4$, and final length $\lambda_\tau=2$.
We calculate the matrix elements of $\Lambda$ based on Eq. (\ref{eq12}) and obtain a $5\times5$ dimensional near-unitary matrix $\Lambda_{5}$ (the fidelity of $\Lambda_{5}$ to a unitary matrix is $\mathcal{F}=0.9992$). The calculated result of matrix $\Lambda_{5}$ is given by
\begin{widetext}
\setlength{\arraycolsep}{2.7pt}
\begin{eqnarray} \label{eq19}
\small{
\Lambda_{5}=\left(\begin{array}{ccccc}
0.9843 + 0.1712i &0.0300 - 0.0273i & -0.0120 - 0.0017i & 0.0041 + 0.0051i & 0.0003 - 0.0039i\\
-0.0047 - 0.0401i & 0.8639 + 0.4990i & 0.0504 - 0.0012i &-0.0108 - 0.0147i & -0.0017 + 0.0096i\\
0.0030 + 0.0119i & 0.0054 - 0.0494i & 0.4535 + 0.8874i & 0.0394 + 0.0452i & 0.0070 - 0.0216i\\
-0.0011 - 0.0069i &-0.0021 + 0.0186i & 0.0338 - 0.0475i &-0.3230 + 0.9414i & -0.0232 + 0.0659i\\
-0.0006 + 0.0044i & 0.0039 - 0.0101i & -0.0166 + 0.0171i & 0.0662 - 0.0114i & -0.9723 + 0.2054i\\
\end{array}\right).}
\end{eqnarray}
\end{widetext}
We decompose the matrix $\Lambda_{5}$ into the product of a diagonal matrix $D$ and a series of elimination matrices $T_{a,a+1}$ based on Gaussian elimination method \cite{Clements}. The result of decomposition result is written as
\begin{align}
\begin{split} \label{eq20}
\Lambda_{5}=DT_{3,4}^{(5)}T_{4,5}^{(4)}T_{1,2}^{(5)}T_{2,3}^{(4)}T_{3,4}^{(3)}T_{4,5}^{(2)}T_{1,2}^{(3)}T_{2,3}^{(2)}T_{3,4}^{(1)}T_{1,2}^{(1)}.
\end{split}
\end{align}
As shown in Fig. \ref{Lambda_5}, the matrix $\Lambda_{5}$ is programmed into an optical network in a boson sampler. We calculate the values of phase shifter angles $\theta$ and $\varphi$, and the calculated results are presented in Tab. \ref{table2}. The diagonal matrix $D$ can be realized physically by setting some phase shifters on all individual channel outputs of the interferometer. The angles of these phase shifters to realize the diagonal matrix $D$ from top to bottom in Fig. \ref {Lambda_5} are 4.7590, 4.8586, 2.4069, 5.7315, 2.9009.
\begin{figure} [!h]
\begin{center}
\includegraphics[width=7.8 cm,angle=0]{Lambda_5.pdf}
\caption{Schematic setup for implementing a matrix $\Lambda_{5}$ in an optical network. The numbers (left) are the spatial modes of the optical interferometer. The number ($n$) (top) denotes the $n$th time. That is, $T_{a,a+1}^{(n)}$ represents a variable beam splitter acting on modes $a$ and $a+1$ at time $n$. } \label{Lambda_5}
\end{center}
\end{figure}
\begin{table}[htb]
\centering \caption{The results for calculating the parameters $\theta$ and $\varphi$ of the elimination matrix $T_{a,a+1}$. }
\begin{tabular}{cccc}
\hline
\hline
& Matrix &\qquad\qquad $\theta$ & \qquad\qquad $\varphi$ \\
\cline{1-4}
\hlin
& $T_{1,2}^{(1)}$ & \qquad\qquad 0.3892 &\qquad\qquad 2.9086 \\
& $T_{3,4}^{(1)}$ & \qquad\qquad 0.0852 &\qquad\qquad 5.4230 \\
& $T_{2,3}^{(2)}$ & \qquad\qquad 0.3795 & \qquad\qquad 1.9238 \\
& $T_{1,2}^{(3)}$ &\qquad\qquad 0.0226 &\qquad\qquad 3.1496 \\
& $T_{4,5}^{(2)}$ &\qquad\qquad 0.5999 &\qquad\qquad 5.5247 \\
& $T_{3,4}^{(3)}$ &\qquad\qquad 0.0524 &\qquad\qquad 3.3019 \\
& $T_{2,3}^{(4)}$ & \qquad\qquad 0.3493 &\qquad\qquad 0.0315 \\
& $T_{1,2}^{(5)}$ &\qquad\qquad 0.3798 &\qquad\qquad 1.8981 \\
& $T_{4,5}^{(4)}$ &\qquad\qquad 0.6472 &\qquad\qquad 3.2375 \\
& $T_{3,4}^{(5)}$ &\qquad\qquad 0.0889 &\qquad\qquad 2.5562 \\
\hline\hlin
\end{tabular}\label{table2}
\end{table}
Finally, we calculate the cumulative work distribution by the definition as follows
\begin{align}
\begin{split} \label{eq21}
\chi(W)=\int^{W}\rho(W')dW'.
\end{split}
\end{align}
The numerical result of cumulative work distribution for three identical bosons based on Eq. \eqref{eq20} is presented in Fig. \ref{Work}, which is consistent with the solution of Eq. \eqref{eq13}.
In real experiments, the noise and error are inevitable. Therefore, we evaluate the effect of the noise on the cumulative work distribution by adding a random noise $\mathcal{N}\in(-0.01,0.01)$ to the angles of the beam splitters and phase shifters. As shown in Fig. \ref{Work}, we randomly choose $100$ groups of noise terms, and calculate the cumulative work distribution under the noise effects. We find that
the ratio between the error bar and the cumulative work distribution curve is $1$\% to $2$\%.
\begin{figure
\begin{center}
\includegraphics[width=8.1 cm,angle=0]{Work.pdf}
\caption{The cumulative work distribution for three bosons in an expanding quantum piston system. The parameters $\lambda_0=1$, $\lambda_\tau=2$, $v=0.4$, and $\beta=0.1$ are taken. Error bars represent
the standard deviations in the $100$ numerical simulations with random noise terms $\mathcal{N}$ on the beam splitters and phase shifters. The evaluated work points are (-48, -42, -36, -27, -16, -8, 4) and the corresponding standard deviations are (6.6$\times10^{-5}$, 1.5$\times10^{-4}$, 4.0$\times10^{-4}$, 8.6$\times10^{-4}$, 3.6$\times10^{-3}$, 2.1$\times10^{-2}$, 2.1$\times10^{-2}$). } \label{Work}
\end{center}
\end{figure}
\section{Discussion and conclusion}\label{Sec3}
In conclusion, we have presented an efficient method for the calculation of the work distribution of multiple identical bosons where the permanent of the transition amplitude matrix is simulated by boson sampling.
As we know, calculating the permanent of a matrix has been quantified as a \#P-hard problem \cite{P-complete,P-complete1}. The well-known algorithm to compute the permanent of any a $d\times d$ complex matrix is Ryser's $\mathcal{O}(d^22^d)$ running time algorithm \cite{algorithm1}. Ryser's algorithm has shown an upper bound for computing $d=50$ matrix permanent on a classical supercomputer \cite{Tianhe}. Later, the scaling was slightly upgraded to the currently best $\mathcal{O}(d2^d)$ result \cite{Nijenhuis,algorithm2,Clifford}, but it is still an exponential time scale.
Boson sampling has been proved that can efficiently estimate the permanent. Recently, photonic boson sampling has demonstrated the computational advantages of the hard-to-simulate quantum behaviour over their classical counterparts \cite{QBS7,supremacy}, even in presence of noise \cite{Error}. Boson sampling is considered to be the first step for universal photonic quantum computing. In order to realize boson sampling in photonic experiments, we only need passive devices, such as high purity and indistinguishable single-photon source based on quantum dot \cite{scource1,scource2,scource3}, multi-mode optical network with robustness against photon losses \cite{Clements,Interofeter}, and high-efficient photon-number-resolving detectors \cite{detector1,detector2,detector3}. To calculate the work distribution of multiple identical bosons with boson sampling, the optical devices need to be programmable, which have been realized in the silicon photonics \cite{qiang2018large}.
As a feasible example, we have designed a compact optical architecture to calculate the work distribution of multiple bosons in detail. We estimate the resources need to calculate the work distribution of multiple identical bosonic system, and find that
the size of the optical network in the boson sampling increases approximately linearly with the acceleration of the expansion speed or the final stretched length of piston. The total resource cost (the number of optical elements) of our scheme is reserved at the polynomial hierarchy.
The approach we developed here is also suitable to the calculation of work distribution of multiple identical bosons in a contraction piston or a harmonic oscillator system.
We anticipate that the boson sampling could have potential applications in calculating chemical structure and social networks with quantum advantage.
\section*{ACKNOWLEDGMENTS}
We thank valuable discussions with Haitao Quan, Xianmin Jin, and Yuanhao Wang.
This work was supported by National Natural Science Foundation of China under Grant No. 61771278 and Beijing Institute of Technology Research Fund Program for Young Scholars.
\bibliographystyle{unsrtnat}
|
\section{INTRODUCTION}
The Accelerator Test Facility (ATF)~\cite{Aryshev:2742899} is a \SI{1.3}{GeV} electron beamline complex located at the High Energy Research Organisation (KEK) in Tsukuba, Japan. The facility is a test-bed for technologies required for a future linear electron-positron collider. The ATF comprises a linear accelerator, damping ring (DR) and final focus system. Ultra-low emittance beams can be produced with the \SI{138.6}{\m} circumference DR via the process of radiation damping. After extraction from the DR, the beam passes through an extraction line and a final focus system (ATF2~\cite{walker2005atf2, atf22006atf2}) (Fig.~\ref{fig:ATFDiagram2}) which is a scaled prototype of the final focus system of the International Linear Collider (ILC)~\cite{TDR_Vol1} or the Compact Linear Collider (CLIC)~\cite{CLICCDR}. The vertical focal point of the ATF2 beamline is designated the `interaction point' (IP). The two main goals~\cite{bambade2010present, GlenATF} of the ATF2 Collaboration are to produce nano-beams with an IP beam size of 37~nm and position stabilization at the nanometer level.
The ATF is typically operated with an extracted beam pulse repetition rate of 3.12 Hz, a beam charge in the range $0.1-1.0\times10^{10}~e$, and with one bunch per pulse.
Multi-bunch trains can also be produced by accumulating two or three bunches in the DR and extracting them as a single train with one pulse of the DR extraction kicker.
The Feedback On Nanosecond Timescales (FONT) group~\cite{FONT} has developed several generations of prototype bunch-by-bunch beam-stabilization feedback systems which have been tested at the ATF. A feedback system was deployed in the upstream section of the ATF2 extraction line, using high-resolution bunch-position measurements from stripline beam-position monitors (BPMs)~\cite{PhysRevSTAB.18.032803}, to demonstrate~\cite{PhysRevAccelBeams.21.122802} the resolution, correction-range and latency requirements for the ILC IP beam collision feedback system~\cite{burrows2016fast}. An extended feedback system based on this hardware was recently used to stabilize the beam trajectory before its entrance to the final-focus region, and yielded a significant reduction in the impact of `wakefields' on the beam-size growth~\cite{Bett_2021}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\linewidth]{ATFSchematic.png}
\caption[Schematic of the ATF2 layout]{Schematic of the ATF layout, with the ATF2 focal point indicated as the IP~\cite{Bett_2021}.}
\label{fig:ATFDiagram2}
\end{figure}
Here we report the design and performance of a high-resolution, high-precision, low-latency, beam-position feedback system located around the ATF2 IP (Fig.~\ref{fig:ATFDiagram}), which is aimed at stabilizing directly the IP vertical beam position to the nanometer level. This system incorporates five cavity BPMs similar to those reported in~\cite{Inoue}, but with a much lower `quality factor'. The down-mixed BPM signals are digitized using a custom FPGA-based feedback controller, the `FONT5A' board~\cite{PhysRevAccelBeams.21.122802, Christian:2312282}, and the feedback calculation is performed on an FPGA mounted on the board. An analogue correction signal is output from the board, amplified using a custom power amplifier with a fast rise-time (35~ns)~\cite{TMD}, and used to drive a stripline kicker, IPK.
The cavity-BPM system is described in Section II, and its resolution performance is presented in Section III. The bunch-by-bunch feedback system is described in Section IV, and its beam stabilization performance is reported in Section V. A summary of results, and conclusions, is given in Section VI.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99\linewidth]{ATF_IPSchematic.png}
\caption{Schematic of the ATF2 IP region, showing the final-focus magnets and the elements of the FONT IP feedback system including dipole cavity BPMs IPA, IPB and IPC, reference cavity BPMs Ref $x$ and Ref $y$, and the stripline kicker IPK.}
\label{fig:ATFDiagram}
\end{figure}
\section{IP CAVITY BPM SYSTEM}
The IP BPM system incorporates five C-band cavity BPMs~\cite{jang2013development,kim2012cavity}, IPA, IPB, IPC, Ref $x$ and Ref $y$ (Fig.~\ref{fig:ATFDiagram}). Throughout this paper $x$ and $y$ refer respectively to the horizontal and vertical beam position coordinates in the plane transverse to the beam propagation direction. For beam-size measurements using the IP Beam Size Monitor~\cite{Suehara:2010zz} the IP is placed longitudinally between IPB and IPC. However, for the nano-beam stabilization studies reported here, the IP can instead be placed at any one of IPA, IPB or IPC; this is discussed in Section IV. The cavity BPM design and operation is described below.
\subsection{Cavity BPM design and operation}
As a bunch of charged particles passes through a cavity BPM, its electromagnetic eigenmodes are excited~\cite{Boogert}. The transverse magnetic (TM) modes can be used to determine both the bunch charge and the bunch offset w.r.t.\ the cavity's electrical axis. Separate cavities were designed for sensitivity to the monopole and dipole TM modes, referred to as `reference' and `dipole' cavities respectively; Ref $x$ and Ref $y$ are reference cavities and IPA, IPB and IPC are dipole cavities (Fig.~\ref{fig:ATFDiagram}).
In the circular cylindrical $x$ and $y$ reference cavities the dominant excited mode is the monopole mode, illustrated in Fig.~\ref{fig:Extraction}(a), which is sensitive to the bunch charge and, for small offsets, insensitive to its position offset w.r.t.\ the cavity electrical center. A schematic of the reference cavity is shown in Fig.~\ref{fig:CavityDiagram}(a), with the coupling slot and antenna indicated~\cite{NevenThesis}. The $x$ and $y$ reference cavities have diameters of 42.95 and 38.65~mm respectively, designed to yield monopole-mode frequencies equal to the respective dipole-mode frequencies of the dipole cavities (see below). Using dedicated tuning pins, the monopole-mode frequency of each reference cavity was fine-tuned to match the respective dipole-mode frequency of the dipole cavities (see Table~\ref{CavityFrequencies}).
\begin{figure}[htbp]
\centering
\subfigure[]{\includegraphics[width=0.73\linewidth]{MonopoleMode.png}}
\hfill
\subfigure[]{\includegraphics[width=0.73\linewidth]{Eigenmodes.png}}
\caption{Schematic of the electric and magnetic field lines of (a) the TM010 mode for
a circular cylindrical cavity BPM and (b) the TM210 mode (or $x$-dipole mode) for a rectangular cylindrical cavity BPM. The
waveguides which couple to the TM210 mode are shown. There are also corresponding waveguides which couple to the $y$-dipole TM120 mode.}
\label{fig:Extraction}
\end{figure}
The dipole cavity principle is illustrated in Fig.~\ref{fig:Extraction}(b). The cavity design is of rectangular cylinder form and uses spatial filtering to suppress the dominant monopole mode so that the higher-frequency dipole mode can be extracted~\cite{SelectiveCoupling}; this mode is sensitive to the bunch position offset as well as its charge. A schematic of the cavities used is shown in Fig.~\ref{fig:CavityDiagram}(b), with the coupling slots and waveguides indicated. The cavities were designed with different vertical and horizontal dimensions so as to decouple the horizontal and vertical dipole modes; the positioning of the coupling slots and waveguides allows these modes to be extracted separately from the same BPM. There are pairs of output $x$-ports and $y$-ports in each cavity; the respective output signals are combined (see Fig.~\ref{fig:4ProcessingElectronics}) so as to double the signal from the anti-symmetric dipole mode and cancel the unwanted symmetric monopole mode. A \SI{700}{MHz} bandwidth band-pass filter (BPF) removes the residual monopole signal.
The cavities are fabricated from aluminium and were designed~\cite{jang2013development} to have ultra-low quality-factor values so as to be suitable for resolving in time individual particle bunches in trains with bunch separations of order 100~ns. The design and measured values of the cavity resonant frequencies are given in Table~\ref{CavityFrequencies}. Also given are the measured signal exponential decay times, which are around 25~ns for the dipole cavities and 14~ns for the reference cavities.
\begin{figure}[htbp]
\centering
\subfigure[]{\includegraphics[width=0.4\textwidth]{MonopoleExtraction.png}{}}\hfill
\subfigure[]{\includegraphics[width=0.4\textwidth]{CavityExtraction.png} }\hfill
\caption{Schematic of the (a) cylindrical reference cavity BPMs and (b) rectangular dipole
cavity BPMs~\cite{Inoue}.}
\label{fig:CavityDiagram}
\end{figure}
\begin{table}[htbp]
\begin{center}
\centering
\newcolumntype{Y}{>{\centering\arraybackslash\hsize=0.3\hsize}X}
\newcolumntype{Z}{>{\centering\arraybackslash\hsize=0.8\hsize}X}
\begin{tabularx}{\linewidth}{Z Y Y Y}\toprule
BPM cavity & Design& Measured & Decay\\
& frequency & frequency & time \\
& (GHz) & (GHz) & (ns)\\\hline\\[-0.9em]
Dipole IPA ($x$-port) & 5.712 & 5.705 & 25\\
Dipole IPB ($x$-port) & 5.712 & 5.706 & 25\\
Dipole IPC ($x$-port) & 5.712 & 5.704 & 23\\\hline\\[-0.9em]
Dipole IPA ($y$-port) & 6.426 & 6.428 & 26\\
Dipole IPB ($y$-port) & 6.426 & 6.427 & 22\\
Dipole IPC ($y$-port) & 6.426 & 6.428 & 21\\\hline\\[-0.9em]
Reference ($x$-cavity) & 5.711 & 5.705 & 14\\
Reference ($y$-cavity) & 6.415 & 6.428 & 14\\
\toprule
\end{tabularx}
\end{center}
\caption[Resonant frequencies of the dipole and reference cavity BPMs]{Design and measured values of the resonant frequencies of the dipole~\cite{jang2013development} and reference~\cite{TalithaThesis} cavity BPMs, and measured signal decay times.}
\label{CavityFrequencies}
\end{table}
Since the ATF2 is designed to focus the beam to c. 37~nm in the vertical ($y$) plane, and our aim is nano-beam stabilization in this plane, for the remainder of this paper we consider only vertical beam position measurements and hence discuss only those signals from the reference $y$-cavity BPM (Ref $y$) and the dipole BPM (IPA, IPB, IPC) $y$-ports.
The position resolution of a dipole cavity BPM is primarily limited by the signal-to-noise ratio, where the signal level is determined both by how much energy is transferred from the beam to the dipole modes and also how well this mode is coupled out of the BPM through the waveguides. Sources of noise in the system include thermal and electronic noise, as well as signal contamination from the monopole mode~\cite{Boogert}. The resolution performance is discussed in Section~\ref{sec:resolution}.
A variable attenuator on the combined output of each dipole BPM (Fig.~\ref{fig:4ProcessingElectronics}) can be used to increase the dynamic range of the position measurement but at the expense of the resolution. For typical operating bunch charges of 1~nC, 10~dB attenuation was added to the dipole signal, yielding a dynamic range for vertical position measurements of~\SI{\pm 3}{\um}.
The dipole cavity BPMs are mounted on two piezo-mover systems (Fig.~\ref{fig:Submovers}) within the vacuum chamber which allow horizontal, vertical and angular BPM alignments w.r.t.\ the beam trajectory~\cite{Blanco}. IPA and IPB are mounted on a single `IPAB' mover block and, therefore, cannot be moved independently. The IPAB movers were manufactured by Cedrat Technologies and have a working range of \SI{248}{\um}, while the IPC mover was manufactured by PI and has a working range of \SI{300}{\um}. The movers incorporate feedback systems designed to ensure a position stability of better than 2~nm.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.98\linewidth]{IPBPMMovers}
\caption{Schematic of the IP BPM configuration, showing the `IPAB' mover block, with submovers $\mathrm{m_1}$, $\mathrm{m_2}$ and $\mathrm{m_3}$, on which IPA and IPB are mounted, and the IPC mover block with submovers $\mathrm{m_C}$, $\mathrm{m_D}$ and $\mathrm{m_E}$. The nominal IP location, as used for beam-size measurements, is indicated.}
\label{fig:Submovers}
\end{figure}
\subsection{BPM analogue signal processing}
The cavity BPM signals undergo two stages of frequency down-mixing~\cite{Inoue, NevenThesis} (see Fig.~\ref{fig:4ProcessingElectronics}) so as to produce baseband signals that can be digitized with the FONT5A board. In the first stage, both the reference and dipole cavity signals are down-mixed to an intermediate frequency (IF) centered at 714~MHz using a common Local Oscillator (LO) signal so as to retain the phase relation between the signals. The 5.712~GHz LO signal is generated using frequency multiplication of the Master Oscillator signal~\cite{TimingSystem} and hence is phase-locked to the beam. The LO signal can be written
\begin{equation}\label{VLOInput}
V_{LO} \sim L\sin (2\pi f_{\textup{LO}}t+\Delta\phi_{\textup{LO}}),
\end{equation}
where $f_{\textup{LO}}$ = 5.712~GHz, $\Delta\phi_{\textup{LO}}$ is the phase difference between the LO signal and the dipole signal, and $L$ is a constant.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{TwoStageMixing.png}
\caption[Simplified block diagram of two-stage processing electronics]{Simplified block diagram of the two-stage down-mixing process of the dipole and reference cavity signals from GHz-level to baseband. Diagram adapted from~\cite{kim2012cavity}.}
\label{fig:4ProcessingElectronics}
\end{figure*}
The output signals from the $y$-port of a dipole cavity ($V_{\textup{dip}}$) and from the $y$ reference cavity ($V_{\textup{ref}}$) can be represented by:
\begin{subequations}
\begin{equation} \label{VdipInput}
V_{\textup{dip}} \sim q(D_{y}y\sin (2\pi f_{\textup{dip}}t) +(D_{y'}y'-D_{\alpha}\alpha)\cos(2\pi f_{\textup{dip}}t))
\end{equation}
and
\begin{equation}\label{VrefInput}
V_{\textup{ref}} \sim q(\sin (2\pi f_{\textup{ref}}t+\Delta\phi)),
\end{equation}
\end{subequations}
where $q$ is the bunch charge, $y$ is the vertical beam position offset w.r.t.\ the cavity electrical axis, $y'$ and $\alpha$ are the bunch pitch-angle and angle-of-attack, respectively, $f_{\textup{dip}}$ or $f_{\textup{ref}}$ is the respective signal frequency, and $\Delta\phi$ is the difference in phase between the monopole and dipole signals; $D_y$, $D_{y'}$ and $D_{\alpha}$ are constants. It can be seen that the signals excited by a $y'$ or $\alpha$ offset are \ang{90} out of phase with those excited by a $y$ offset. For small bunch offsets the reference-cavity signal is independent of the beam position.
After the first stage of down-mixing, signals $V_{\textup{dip}}\otimes V_{\textup{LO}}$ and $V_{\textup{ref}}\otimes V_{\textup{LO}}$ are produced (Fig.~\ref{fig:4ProcessingElectronics}) at both the IF (714~MHz) and
the higher frequencies $f_{\textup{dip}}+ f_{\textup{LO}}$ or $f_{\textup{ref}}+ f_{\textup{LO}}$, respectively.
In the second-stage processing, the latter are removed with a \SI{150}{MHz} bandwidth band-pass filter centerd at c. 700 MHz. The $V_{\textup{ref}}\otimes V_{\textup{LO}}$ IF signal is then split, with one of the outputs passing through a diode detector to produce a pulse whose magnitude is proportional to the bunch charge. This signal, subsequently denoted $q$, is used for bunch-charge normalization to obtain the bunch position (see below). The other $V_{\textup{ref}}\otimes V_{\textup{LO}}$ output passes through a limiting amplifier to remove its charge dependence. This signal is then used as the LO signal for the second stage of down-mixing of the dipole signals, from the IF to baseband.
In the second stage, the reference and dipole signals are mixed in-phase and in-quadrature to produce $I$ and $Q$ signals, respectively. These signals are orthogonal components that together include the full amplitude and phase information of the BPM waveform~\cite{IPACCavityNeven}. If the BPM is well-aligned in $y'$ and $\alpha$, the contributions to $V_{\textup{dip}}$ from these terms are much smaller than those from $y$, such that
\begin{equation}\label{VI}
\begin{split}
I&=(V_{\textup{dip}}\otimes V_{\textup{LO}})\otimes (V_{\textup{ref}}\otimes V_{\textup{LO}})\\
&\propto qy\cos(\theta_{IQ}),
\end{split}
\end{equation}
and
\begin{equation}\label{VQReduced}
Q \propto qy\sin(\theta_{IQ}),
\end{equation}
where the phase angle, $\theta_{IQ}$, corresponds to
\begin{equation}\label{DefineTheta1}
\theta_{IQ}=2\pi (f_{\textup{dip}}-f_{\textup{ref}})t-\Delta\phi.
\end{equation}
Since the reference cavity is tuned such that $f_{\textup{ref}}\simeq f_{\textup{dip}}$ (see Table~\ref{CavityFrequencies}), the $I$ and $Q$ signals are at baseband. Before digitization these signals are amplified so as to reduce the effect of quantization noise.
\subsection{Signal digitization and digital processing}
The signal digitization is performed on a FONT5A board~\cite{PhysRevAccelBeams.21.122802}, a custom feedback controller with a Xilinx Virtex-5 XC5VLX50T FPGA at its core~\cite{Virtex5}. The primary inputs and outputs of this board are shown in Fig.~\ref{fig:4FONTBoard}. The FPGA firmware is written in the Verilog hardware description language, and the configuration bitstream is stored on a non-volatile Xilinx XCF32P Programmable Read-Only Memory (PROM), from where it is loaded on power-up or system reset. The inputs and outputs of the PCB are via Micro-Coaxial connectors (MCX) which patch to BNC connectors on the case which houses the board~\cite{ConstanceThesis}. The FONT5A board is shown in Fig.~\ref{fig:FONTBoardPhoto} with the case removed.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.48\textwidth]{FONTboard}
\caption{Block diagram of the FONT5A digital board showing the primary input and output signals used.}
\label{fig:4FONTBoard}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth,trim=1.0cm 0cm 1.0cm 0.8cm,clip]{FONTboardPhoto}
\caption[Photograph of the FONT5A digital board]{Photograph of the FONT5A digital board with the case removed~\cite{ConstanceThesis,christian2011latest}.}
\label{fig:FONTBoardPhoto}
\end{figure}
The board contains nine Texas Instruments 14-bit ADS5474~\cite{TexasInstruments} analogue-to-digital converters (ADCs) grouped into separately-clocked banks of three. Seven ADCs are used to digitize the $I$ and $Q$ waveforms from IPA, IPB and IPC, and the $q$ waveform from the Ref~$y$ cavity; the least significant bit is removed as it corresponds to the noise level of the signals~\cite{ConstanceThesis}. The ADC channels contain an inherent offset on their baseline signal which can be zeroed by coupling each with the output of a 16-bit DAC, referred to as a trim DAC~\cite{BettThesis}. The values used for the trim DAC can be set using the associated FONT LabVIEW DAQ which is used to transmit values to the board through an RS-232 Universal Asynchronous Receiver/Transmitter (UART) via an Ethernet serial device server.
A clock at \SI{357}{\MHz} is used for the time-critical FPGA logic. It is derived from the LO, meaning it is phase-locked to the beam, and used to clock the ADCs, so that the $I$, $Q$ and $q$ signals are digitized at 357~MHz. The start of the sampling window is set with respect to the trigger, which is internally delayed on the board and can be adjusted. The sampling window can be varied within the firmware but for one- or two-bunch operation typically consists of 164 samples each separated by 2.8~ns, meaning that a complete DR beam-circulation period (462~ns) can be digitized within a single window. Representative digitized waveforms for 2-bunch-train operation (Section~\ref{sec:accel}) are shown in Fig.~\ref{fig:IQWaveforms}. The difference in the $I$ and $Q$ signals between the two bunches derives from the transverse position offset between them.
The firmware includes the functionality to provide a constant offset to the $I$, $Q$ and $q$ signals before they are used to calculate the bunch position. This is used to remove the position-independent baseline signals that are generated on each $I$ and $Q$ waveform at the second stage of the signal processing (Fig.~\ref{fig:IQWaveforms}).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.48\textwidth]{Waveform3.png}
\caption{Representative digitized $I$, $Q$ and $q$ waveforms from IPC, for two-bunch-train operation with a bunch spacing of 280~ns. The waveforms were sampled at intervals of 2.8~ns.}
\label{fig:IQWaveforms}
\end{figure}
\subsection{Position measurement}
\label{sec:position}
A linear combination of $I$ and $Q$ can be chosen to produce a signal, $I'$, with an amplitude proportional to the bunch position $y$~\cite{Inoue}:
\begin{equation}\label{DefineVIPrime}
I'=I\cos(\theta_{IQ})+Q\sin(\theta_{IQ}).
\end{equation}
By substituting Eqs.~\ref{VI} and~\ref{VQReduced} into Eq.~\ref{DefineVIPrime},
\begin{equation}
\label{AfterExpPrimed2}
y=\frac{1}{k}\frac{I'}{q},
\end{equation}
where $k$ [$\SI{}{\um}^{-1}$] is a constant, found by calibration of the BPM.
A signal orthogonal to $I'$ can also be generated, $Q'$, that is proportional to the beam pitch:
\begin{equation}\label{DefineVQPrime}
Q'=-I\sin(\theta_{IQ})+Q\cos(\theta_{IQ}).
\end{equation}
Each dipole BPM is calibrated w.r.t.\ position by vertically scanning the beam across a known range by changing the position of quadrupole QD0FF (Fig.~\ref{fig:ATFDiagram}) and measuring the corresponding BPM response. Calibrations w.r.t.\ the beam angle $y'$ are performed by tilting the BPMs through a known range using the submovers shown in Fig.~\ref{fig:Submovers}.
For each measured bunch in the beam, the calibration calculation can be performed using either single or multiple samples of the $I$ and $Q$ waveforms. For convenience a single sample of the $q$ signal from Ref $y$ is used for charge normalization of the $I'$ and $Q'$ signals from all three dipole BPMs. The requirements for low-latency feedback preclude the direct implementation of division for the charge normalization within the firmware and, instead, a method of lookup tables (LUTs) is employed using block RAM resources in the FPGA. The charge, $q$, is used as an address to the LUTs, for which the elements are preloaded with $\frac{1}{q}$ scaled by the appropriate feedback coefficient $C_i$,
\begin{equation}\label{LUTCalc}
q\xrightarrow{\text{$LUT_i$}} \frac{C_i}{q},
\end{equation}
where $C_i$ incorporates the terms involving $\theta_{IQ}$, $k$ and the feedback gain $G$ (see Section~\ref{feedback_calculation}); there are four instances of the LUT logic ($1\leq i\leq 4$), each loaded with the respective value of $G$, allowing for up to any two of the BPMs to be used as input to the feedback system~\cite{RamjiawanThesis}.
The position resolution can be significantly improved (see Section~\ref{sec:resolution}) by integrating over multiple samples of the $I$ and $Q$ signals as this both increases the signal level and averages over thermal and electronic noise. The integration range is chosen around the peak of the $I$ and $Q$ signals, as samples significantly in advance of the peak may contain transient effects from unwanted modes and samples late in the waveform have a poorer signal-to-noise ratio. This integration is performed in real time on the FONT5A board: on every rising fast-clock edge within the selected integration window, the most recent $I$ and $Q$ value is summed with the previous respective sum. As an example, an IPA position calibration using 11-sample integration is shown in Fig.~\ref{fig:f4IPA1CalQPrime}.
\begin{figure*}[htbp]
\centering
\subfigure[]{\includegraphics[width=0.30\textwidth]{IPACal11.png}\label{fig:f1IPACalAQD}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{IPACal22.png}{ }\label{fig:f2IPACalIQAQD}}
\subfigure[]{\includegraphics[width=0.30\textwidth]{IPACal33.png}\label{fig:f3IPACalkAQD}}
\caption[IPA position calibration]{Example vertical position calibration of IPA, using an 11-sample integration range: (a) $\frac{I}{q}$ and $\frac{Q}{q}$ versus trigger number; (b) ${\frac{Q}{q}~\mathrm{versus}~\frac{I}{q}}$ (points); the line shows a least-squares fit to determine ${\theta_{IQ} = -1.093\pm0.006~\mathrm{radians}}$; (c) the data points show $\frac{I'}{q}$ versus QD0FF mover position, the red error bars show the standard error on the mean values at each QD0FF setting and the red line shows a least-squares fit, which yields ${k = 0.184 \pm 0.002~\SI{}{\um^{-1}}}$.}
\label{fig:f4IPA1CalQPrime}
\end{figure*}
Representative position and angle calibration constants for the three dipole BPMs, calculated using 11-sample integration, are presented in Tab.~\ref{t:AngFit}. It can be seen that IPA and IPB have similar sensitivities, whereas IPC has a lower sensitivity; this is due to a minor fabrication difference.
\setlength{\tabcolsep}{12pt}
\begin{table*}[htbp]
\begin{center}
\begin{tabular}{c c c }\toprule \\[-0.8em]
BPM & Position calibration const. & Angle calibration const.
\\\\[-0.9em]\hline \\[-0.8em]
IPA & $0.184\pm0.002$~${\upmu\mathrm{m}}^{-1}$ & $0.277\pm0.003$~m${\mathrm{rad}}^{-1}$ \\
IPB & $0.168\pm0.002$~${\upmu\mathrm{m}}^{-1}$ & $0.253\pm0.003$~m${\mathrm{rad}}^{-1}$ \\
IPC & $-0.110\pm0.001$~${\upmu\mathrm{m}}^{-1}$ & $-0.157\pm0.002$~m${\mathrm{rad}}^{-1}$\\ \toprule
\end{tabular}
\end{center}
\caption[Position and angular calibration constants]{BPM position and angle calibration constants.}
\label{t:AngFit}
\end{table*}
\section{Cavity BPM system position resolution}
\label{sec:resolution}
The resolution of the BPM system was evaluated using measurements of the bunch trajectory at all three dipole BPMs. Since the bunch follows a straight-line trajectory which can be characterized with measurements from only two BPMs, measurements from the third BPM can be used to estimate the resolution of the system.
The beam position at BPM $i$, $y_i$, can be represented as a linear combination of the positions of the beam at the other two BPMs, $y_j$ and $y_k$:
\begin{equation}\label{ResPredLinearRep2}
y_{i} = A_{ij} y_{j} + A_{ik} y_{k},
\end{equation}
where $A_{ij}$ and $A_{ik}$ are `geometric' coefficients defined by the relative separations of the three BPMs (Fig.~\ref{fig:Submovers}). The predicted beam position at BPM $i$, $y_{i}^{\mathrm{pred}}$, can therefore be written in terms of the measured positions at BPMs $j$ and $k$:
\begin{equation}\label{ResPredLinearRep}
y_{i}^{\mathrm{pred}} = A_{ij} y_{j}^{\mathrm{meas}} + A_{ik} y_{k}^{\mathrm{meas}}.
\end{equation}
The difference between this and the measured position, $y_{i}^{\mathrm{meas}}$, yields a residual. Under the assumption that all three BPMs have the same resolution, $\sigma_\mathrm{res.}$, the resolution is derived from the standard deviation of the distribution of residuals measured over a batch of sequential beam pulses:
\begin{equation}\label{ResAssumeAllSame}
\sigma_\mathrm{res.} = \textup{std} \bigg\{\frac{(y_i^\mathrm{meas}-y_i^\mathrm{pred})}{\sqrt{1+A_{ij}^2 + A_{ik}^2}}\bigg\}_{ijk}.
\end{equation}
Detailed studies of the experimental setup to optimize the resolution, including the BPM alignment procedure, are given in~\cite{RamjiawanThesis}. For a data set with bunch charge $0.5\times10^{10}~e$, Fig.~\ref{fig:3ResolutionAsFnSampleWindow} shows the resolution as a function of the number of $I$ and $Q$ samples integrated in real time for the position calculation. It can be seen that the resolution improves from 41~nm (single sample) to an optimal value of 19~nm with 11 samples. No improvement is seen by integrating additional later samples as the BPM waveforms have decayed and the signal levels are low.
As a cross-check, an alternative, `fitting', method was employed. Here the coefficients $A_{ij}$ and $A_{ik}$ (Eq.~\ref{ResPredLinearRep2}) are fitted to the measured position data set so as to minimise empirically the resolution (Eq.~\ref{ResAssumeAllSame}).
The fitting method may be applied separately to each of the three BPMs, giving three correlated estimates of the resolution. Were the resolution effectively degraded via the influence of uncontrolled correlated parameters, the empirical fit could yield an improvement over the geometric method~\cite{RamjiawanThesis}. The fitted resolution results for the same data set are also given in Fig.~\ref{fig:3ResolutionAsFnSampleWindow}; the results are in good agreement with the geometric method and confirm that the real-time resolution of 19~nm is the best that could be obtained for these BPMs with the given beam conditions. The resolution results are summarised in Table~\ref{t:IntegratedResolutionImprovement}.
\begin{figure}[htbp]
\centering
\includegraphics[width=1\linewidth]{ResolutionvsSampleWindow_All.png}
\caption{Resolution vs. number of samples integrated; the location of each integration window was chosen so as to optimize the geometric resolution. Results for the `geometric' (black) and `fitting' method for each BPM (green, gold, blue) are shown. The error bars represent the statistical uncertainty on the resolution. }
\label{fig:3ResolutionAsFnSampleWindow}
\end{figure}
\begin{table}[htbp]
\newcolumntype{Y}{>{\centering\arraybackslash}X}
\begin{tabularx}{\linewidth}{X Y Y}\toprule
Resolution & Single-sample (nm) & 11-sample (nm)\\\hline
Geometric & $40.6\pm1.0$ & $19.0\pm0.4$\\
IPA fitting & $40.6\pm1.0$ & $19.2\pm0.6$ \\
IPB fitting & $40.8\pm1.0$ & $19.4\pm0.6$\\
IPC fitting & $62.8\pm1.3$ & $17.6\pm0.4$ \\\toprule
\end{tabularx}
\caption[Single-sample and integrated-sample resolution]{The best single-sample and integrated-sample resolution measurements for the geometric and fitting methods.}
\label{t:IntegratedResolutionImprovement}
\end{table}
\section{ATF2 IP Bunch-by-bunch Feedback System}
\subsection{System design}
The high-resolution real-time vertical beam position information from the cavity BPM system was used as input to a closed-loop feedback. For two-bunch trains (see Section~\ref{sec:accel}) the position of the first bunch was measured and used to correct the position of the second bunch. Two feedback operating modes were used, represented functionally in Fig.~\ref{fig:FeedbackLoops}. In single-BPM mode (Fig.~\ref{fig:FeedbackLoops}(a)) the position signal from one BPM was used to derive the correction signal supplied to the kicker IPK, such that the vertical beam position was stabilised at the chosen BPM. For this mode the IP was moved longitudinally from the nominal IP to the center of the chosen BPM so as to directly stabilise the vertical position there. In two-BPM mode (Fig.~\ref{fig:FeedbackLoops}(b)) the IP was placed longitudinally at one BPM; the position signals from the other two BPMs were used to derive a correction signal such that the nano-beam was stabilised vertically at the chosen BPM, which hence served as an independent witness of both the corrected and uncorrected beam positions. Four multiplexers within the firmware allow selection among the three BPMs for their input either individually (mode (a)) or as a pair (mode (b)).
For both modes the correction signal to the kicker is output from the FONT5A board (Fig.~\ref{fig:4FONTBoard}) via a Linear Technology 14-bit LTC2624~\cite{LinearTechnology} digital-to-analogue converter (DAC).
\begin{figure}[htbp]
\centering
\subfigure[]{\includegraphics[width=0.48\linewidth]{IPFB_1BPM.png}}\hfill
\subfigure[]{\includegraphics[width=0.48\linewidth]{IPFB_2BPM.png}}\hfill
\caption{Diagrams of feedback loops showing dipole cavity BPMs (IPA, IPB and IPC) and stripline kicker (IPK). (a) Single-BPM feedback with beam measurement and stabilization illustrated at IPC (red). (b) Two-BPM feedback, illustrated for position measurements at IPA and IPC (purple) with beam stabilization at IPB (red).}
\label{fig:FeedbackLoops}
\end{figure}
\subsection{Kicker and kicker amplifier}
\label{StriplineKickers}
The correction signal from the FONT5A board requires amplifying before it can be used to drive the kicker (see Fig.~\ref{fig:FeedbackLoops}). The stripline kicker (Fig.~\ref{fig:stripline}) is a modified stripline BPM~\cite{PhysRevSTAB.18.032803} and consists of two conducting strips, $\sim$\SI{12.5}{\cm} in length and separated by \SI{24}{\mm}, at the top and bottom of the inside of the beam-pipe. The custom-made kicker amplifier (see eg.~\cite{ConstanceThesis}) was manufactured by TMD Technologies Ltd~\cite{TMD}. In order to meet the low-latency requirements the amplifier was designed with a fast rise time of 35~ns to reach 90\% of the peak output. The amplifier is capable of providing a drive current of $\pm$30~A.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\linewidth]{StriplinePhoto.png}
\caption{Photograph of the IP stripline kicker.}
\label{fig:stripline}
\end{figure}
\subsection{Feedback calculation}
\label{feedback_calculation}
For either feedback mode, the signal sent to the kicker, $V$ (DAC counts), is derived from the measured position offset at the chosen BPM(s) as
\begin{equation}\label{FBAlgorithm}
V=-G\frac{y}{M}+c,
\end{equation}
where $G$ is the feedback gain, $M$ (\SI{}{\um}/DAC counts) is the kicker response calibration constant, and $c$ is an arbitrary offset. If the beam is being stabilized at a location in between BPMs, $y$ refers to the interpolated position. For the case in which the vertical positions of bunches 1 and 2 are 100\% correlated, optimal stabilization of bunch-2 is obtained with $G$ set to unity. For the case of uncorrelated position components $G$ can be adjusted empirically so as to achieve optimal stabilization of bunch-2. The value of $c$ can be controlled via the firmware settings so as to place the stabilised bunch-2 at any desired vertical position within the feedback dynamic range.
The firmware is designed so that the kicker drive signal is output at the same time relative to the beam arrival regardless of the number of samples integrated in the digital signal processing, up to a maximum of 15 samples. The firmware is also set up to allow a selectable constant kicker drive signal from the DAC with the same timing structure as for a real feedback pulse; this feature is used to evaluate $M$ directly by measuring the response of the beam as a function of the kicker drive signal.
\subsection{Latency Measurement}
The closed-loop feedback latency is defined as the time interval between bunch-1 passing through the longitudinal center of IPK and the derived kicker correction pulse (for bunch-2) reaching 90\% of its final output value. The latency was measured directly with the beam by adding a controlled delay to a constant kicker drive signal (of 2000 DAC counts) and measuring the resulting position deflection of the second bunch. The principle is illustrated in Fig.~\ref{fig:4LatencyDesc}. For large added delay (small $\Delta t$) the kick arrives too late and bunch-2 is undeflected. For small added delay (large $\Delta t$) the kick arrives in time to fully deflect bunch-2. When the kick arrives in time to kick the bunch by 90\% of the maximum value, then $\Delta t$ is equal to the latency. Sequential triggers were toggled between feedback `off' and `on' to allow running baseline subtraction.
Fig.~\ref{fig:4Latency} shows the beam deflection as a function of $\Delta t$ from which the latency is measured to be 83 samples, i.e. 232~ns.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.8\linewidth]{Latency.png}
\caption{Schematic illustrating the principle of the direct latency measurement by adding a controlled delay to the kicker drive output signal.}
\label{fig:4LatencyDesc}
\end{figure}
\begin{figure}[hbtp]
\centering
\includegraphics[width=1\linewidth]{LatencyScanApril2}
\caption[Bunch deflection versus feedback signal delay]{The deflection of bunch-2 as a function of $\Delta t$ (\# 2.8 ns samples) defined in Fig.~\ref{fig:4LatencyDesc}. The red line shows a sigmoid fit of the form
$f(\Delta t)=p_1+\frac{p_2-p_1}{1+10^{p_3-p_4 \Delta t}}$,
where $p_1$, $p_2$, $p_3$ and $p_4$ are fit parameters. The dashed lines show the latency definition at 90\% of maximum deflection.}
\label{fig:4Latency}
\end{figure}
\section{IP Feedback System Performance}
\subsection{Accelerator and feedback setup}
\label{sec:accel}
For the operation of the IP bunch-by-bunch feedback system, the ATF DR was configured to deliver two-bunch trains to ATF2 with a bunch separation of \SI{280}{ns}. The train repetition rate was 1.56~Hz. This setup provides a high degree of correlation between the vertical positions of the two bunches in each train~\cite{NevenThesis}, which yields the conditions for optimal feedback performance in stabilizing the second bunch.
The limited dynamic range of the dipole BPMs for optimal resolution necessitates both their good transverse centering w.r.t.\ the beam trajectory and, ideally, small beam jitter at each BPM. Each final-focus-system quadrupole is mounted on transverse movers, which allows for adjustments to both the incoming beam position and angle. In particular, moving QD0FF (see Figure~\ref{fig:ATFDiagram}) vertically adjusts the vertical IP position, while moving QF1FF (or the upstream QF7FF) adjusts the vertical beam incoming angle~\cite{BeamWaist2, BeamWaist1}. The beam trajectory is first globally aligned with the electrical centers of the IP BPMs, and fine adjustments are then made to center each BPM w.r.t.\ the beam by using the BPM movers (Fig.~\ref{fig:Submovers}).
For single-BPM feedback (Fig.~\ref{fig:FeedbackLoops}(a)), small beam jitter is achieved by setting the IP at the longitudinal center of the feedback BPM~\cite{Transients}. For two-BPM feedback (Fig.~\ref{fig:FeedbackLoops}(b)) the situation is more difficult as the extreme IP angular divergence produces increasingly large beam jitter as longitudinal distance from the IP increases.
With the nominal optics configuration the jitter at the feedback BPMs can exceed the dynamic range for best resolution. Therefore, for two-BPM feedback operation an optics configuration with a reduced angular divergence at the IP was used. This yields a reduced beam jitter at the feedback BPMs, although at the expense of increasing the IP beam jitter. These optics are designed such that the ATF2 beamline has the same magnet strength as for the nominal optics except within the matching section. With these optics, the vertical $\beta$-function at the IP is \SI{12}{\cm}.
The BPMs were set up for optimal performance, and calibrated, as described in Section~\ref{sec:position}. In order to make a direct comparison between the data with feedback `on' and `off' within a given dataset, the feedback was toggled between on and off on alternate bunch trains.
\subsection{Single-BPM IP Feedback Results}
Single-BPM feedback was operated with a bunch charge of $0.8\times10^{10}e^-$, with the IP set at IPC.
The feedback gain was set to 0.8 to account for the imperfect bunch-to-bunch position correlation, as determined from correlation measurements taken at the start of the shift (Table~\ref{tab:50nmHist}). Further analysis has suggested, however, that the correlation decreased during the shift. A 10-sample integration window was found empirically to optimize the resolution.
The feedback performance is illustrated in Fig.~\ref{fig:50nmHist} and summarised in Table~\ref{tab:50nmHist}, where we compare feedback-on and feedback-off results. Since bunch-1 provides the input to the feedback its position is unaffected by the correction. By contrast, the bunch-2 mean position is zeroed by the feedback and its jitter is substantially reduced, from \SI{119}{\nm} to \SI{50}{\nm}.
The same dataset is used in Fig.~\ref{fig:51Correl}, which shows the effect of the feedback on the bunch-to-bunch correlation as well as on the time-sequence of the bunch-2 position.
\begin{figure}[!bhtp]
\centering
\includegraphics[width=1\linewidth]{IPC51nmHist}
\caption{Distributions of bunch positions measured at IPC, for bunch-1 (left) and bunch-2 (right) with feedback off (blue) and feedback on (red).}
\label{fig:50nmHist}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure[]{\includegraphics[width=0.4\textwidth]{IPC51nmCorrel.png}}
\hfill
\subfigure[]{\includegraphics[width=0.4\textwidth]{IPC51nmJit.png}}
\caption{(a) IPC bunch-2 position versus bunch-1 position and (b) bunch-2 position versus trigger number; for feedback off (blue) and feedback on (red).}
\label{fig:51Correl}
\end{figure}
\begin{table}[htbp]
\caption{Position jitters and bunch-to-bunch position correlation with feedback off and on, for single-BPM feedback.}
\newcolumntype{Y}{>{\centering\arraybackslash}X}
\newcolumntype{M}{>{\centering\arraybackslash}m{0.8cm}}
\begin{tabularx}{1\linewidth}{M Y Y Y}
\toprule
\textbf{FB} & \multicolumn{2}{c}{\textbf{Position jitter}} & \textbf{Correlation}\\
& \multicolumn{2}{c}{{(nm)}} & {(\%)}
\\ \hline
& \textbf{Bunch-1} & \textbf{Bunch-2} &
\\\hline
Off & 109 $\pm$ 11 & 119 $\pm$ 12 & $85.1^{+2.5}_{-3.5}$\\
On & 118 $\pm$ 12 & 50 $\pm$ 5 & $-26.0^{+9.8}_{-8.8}$\\
\toprule
\end{tabularx}
\label{tab:50nmHist}
\end{table}
The expected level of beam stabilization can be computed from the bunch jitter and the incoming bunch-to-bunch correlation. The corrected bunch-2 position, $Y_2$, in terms of the uncorrected bunch-1 and bunch-2 positions, $y_1$ and $y_2$, respectively, is
\begin{equation}
Y_2=y_2- G y_1+c.
\label{4CorrectedPos}
\end{equation}
Taking the variance of Eq.~\ref{4CorrectedPos} gives
\begin{equation}
\label{4FBPred}
\sigma_{Y_{\mathrm{2}}}^2 = \sigma_{y_{\mathrm{1}}}^2 + G\sigma_{y_{\mathrm{2}}}^2 - 2G\sigma_{y_{\mathrm{1}}}\sigma_{y_{\mathrm{2}}}\rho_{12},
\end{equation}
where $\rho_{12}$ is the bunch-to-bunch correlation and $\sigma_{Y_\mathrm{2}}$, $\sigma_{y_\mathrm{1}}$ and $\sigma_{y_\mathrm{2}}$ represent the jitters on positions $Y_\mathrm{2}$, $y_\mathrm{1}$ and $y_\mathrm{2}$, respectively.
The measured incoming position correlation between bunches 1 and 2 (feedback off) is about 85\% (Table~\ref{tab:50nmHist}); hence, from Eq.~\ref{4FBPred}, the expected feedback-corrected jitter for bunch-2 is $65\pm11$ nm, which is in reasonable agreement with the measured performance. With feedback on, the measured correlation between bunches 1 and 2 is $-26.0^{+9.8}_{-8.8}$ (Table~\ref{tab:50nmHist}), which implies a slight over-correction. This naively suggests that an improved feedback-corrected jitter would have been possible with a slightly lower gain. Limited beam operation availability at the facility did not allow this to be verified at the time, but it could be investigated in future beam studies.
The theoretically optimum performance is obtained for 100\% correlation between bunches 1 and 2, i.e. $\rho_{12}=1$, comparable bunch jitters, $\sigma_{y_{\mathrm{1}}}=\sigma_{y_{\mathrm{2}}}$, and feedback gain $G=1$. With these conditions fulfilled, the ultimate limit to stabilization is determined by the BPM resolution, $\sigma_{\mathrm{res.}}$:
\begin{equation}
\label{eq:resolution}
\sigma_{Y_{\mathrm{2}}}=\sqrt{2}\sigma_{\mathrm{res.}}.
\end{equation}
For a real-time BPM resolution of c. \SI{19}{\nm} (Section~\ref{sec:resolution}) the ultimate feedback performance in single-BPM mode (Eq.~\ref{eq:resolution}) would hence be stabilization of bunch-2 to c. 27~nm, so there is in principle still a margin for improvement of the feedback performance reported here, subject to improved beam conditions.
\subsection{Two-BPM IP Feedback Results}
Two-BPM feedback was operated with a bunch charge of $0.5\times10^{10}e^-$, with the IP set at IPB and with IPA and IPC used as inputs to the feedback; hence IPB was used as an independent witness of the feedback performance. The longitudinal separations of IPA and IPC from the beam waist yield much larger position signal levels and higher signal-to-noise ratios. The sample window was chosen empirically to optimize the resolution, here with a measured resolution of $\sim$\SI{31.2}{\nm}, for a five-sample window. Feedback was operated with a gain of 0.8 to account for the differences in the position jitter between the two bunches and the imperfect bunch-to-bunch position correlation (Table~\ref{2BPMFBPerformance}).
The feedback performance is illustrated in Fig.~\ref{fig:40nmHist2} and summarised in Table~\ref{2BPMFBPerformance}, where we compare feedback-on and feedback-off results. Since bunch-1 provides the input to the feedback its position is unaffected by the correction. By contrast, the bunch-2 jitter is substantially reduced, from 96~nm to 41~nm. The measured incoming position correlation between bunches 1 and 2 (feedback off) is about 92\% (Table~\ref{2BPMFBPerformance}); hence, from Eq.~\ref{4FBPred}, the expected feedback-corrected jitter for bunch-2 is $40\pm11$~nm, which is in excellent agreement with the measured value.
\begin{figure}[htbp]
\centering
\includegraphics[width=1\linewidth]{IPB40FB}
\caption{Distributions of bunch-1 (left) and bunch-2 (right) positions measured at IPB, with feedback off (blue) and feedback on (red). Feedback was performed in 2-BPM mode, stabilizing at IPB using beam position measurements from IPA and IPC.}
\label{fig:40nmHist2}
\end{figure}
\begin{table}[htbp]
\newcolumntype{Y}{>{\centering\arraybackslash}X}
\newcolumntype{M}{>{\centering\arraybackslash}m{0.8cm}}
\centering
\caption{Position jitters and bunch-to-bunch position correlation with feedback (FB) off and on, for 2-BPM feedback.}
\begin{tabularx}{1\linewidth}{M Y Y Y}
\toprule\\[-0.9em]
\textbf{FB} & \multicolumn{2}{c}{\textbf{Position jitter}} & \textbf{Correlation}\\
& \multicolumn{2}{c}{(nm)} & (\%)
\\\hline \\[-0.9em]
& \textbf{Bunch-1} & \textbf{Bunch-2} & \\\\[-0.9em]
\hline \\[-0.9em]
Off & 106 $\pm$ 11 & 96 $\pm$ 10 & $91.6^{+1.8}_{-3.1}$ \\
On & 100 $\pm$ 10 & 41 $\pm$ 4 & $41.3^{+9.1}_{-12.3}$\\\\[-0.9em]
\toprule
\end{tabularx}
\label{2BPMFBPerformance}
\end{table}
In Fig.~~\ref{fig:40nmHist2} it can be seen that the mean corrected bunch-2 position was at $\sim$\SI{0.5}{\um}, which simply arises from the residual relative transverse offsets between IPB and IPA/IPC; if desired this offset can trivially be removed with a compensating constant offset term, shown as $c$ in the feedback algorithm (Eq.~\ref{FBAlgorithm}).
With feedback on the measured correlation between bunches 1 and 2 is about 41\% (Table~\ref{2BPMFBPerformance}), which implies an under-correction. This suggests that an improved feedback performance would have been possible with a higher gain. As previously noted, at the time, beam operation availability was limited and this could not be verified but it could be confirmed with further beam studies.
In this feedback mode, with stabilization at IPB, the feedback BPMs, IPA and IPC, contribute position information in the ratio 32:68, determined by their relative distances from IPB (Fig.~\ref{fig:Submovers}). Hence, the theoretically best-possible resolution on the corrected beam position at IPB is given by:
\begin{equation}
\sigma_{\mathrm{IPB}} = \sqrt{0.32^2\sigma_{\mathrm{res.}}^2+0.68^2\sigma_{\mathrm{res.}}^2} = 0.75\sigma_{\mathrm{res.}};
\end{equation}
i.e. beam stabilization at IPB as low as c. 23~nm would have been achievable in principle given the measured resolution of 31~nm. Correspondingly, with the best achieved resolution of 19~nm, and a perfect feedback correction, stabilization to 15~nm would be theoretically possible. Hence there is still a margin for improvement of the feedback performance reported here, subject to improved beam conditions.
\section{SUMMARY AND CONCLUSIONS}
We have reported the design, operation and performance of a high-resolution, low-latency, bunch-by-bunch feedback system for beam stabilization. The system includes high-resolution cavity BPMs, two stages of analogue signal down-mixing system, and a digital board incorporating an FPGA. The FPGA firmware allows for the real-time integration of up to fifteen samples of the BPM waveforms so that feedback can be performed within a latency of \SI{232}{\ns}. We have shown that this real-time sample integration improves the beam position resolution, with measured resolutions as good as \SI{19}{\nm}, which consequently improves the feedback performance.
In~\cite{Inoue} results were reported using similar cavity BPMs, but with a higher design quality factor: data were recorded and the resolution was determined in a subsequent offline analysis using a function that included 10 free parameters to account for uncontrolled effects; a resolution of $\sim$9~nm was thereby obtained. In addition, the position-calibration constant was not measured directly at the most sensitive resolution setting, but was interpolated from measurements made with added signal attenuation, at lower position sensitivity. Furthermore, no attention was paid to signal processing latency as the BPMs were not used for bunch-by-bunch feedback.
We have made several significant advances since this earlier study: 1) the BPMs were calibrated w.r.t.\ position directly at the most sensitive resolution setting and the respective calibration factors were applied in the subsequent BPM operations; 2) the signal processing was done in real-time and with low latency, so as to permit the BPMs to be used for bunch-by-bunch feedback; 3) the resolution was measured directly, in real-time, without fitting any extra parameters. The high BPM resolution was hence utilised directly for stabilization of the beam, and is not merely an impressive offline performance figure of merit.
The feedback was operated in two complementary modes to stabilise the vertical position of the ultra-small beam produced at the focal point of the ATF2 beamline at KEK. In single-BPM feedback mode, beam stabilization to $50\pm5$~nm was demonstrated. In two-BPM feedback mode, beam stabilization to $41\pm4$~nm was achieved, in good agreement with the predicted value, given the incoming beam conditions, of \SI{40}{\nm}.
Some margin remains to improve the feedback performance by increasing the degree of bunch-to-bunch position correlation in the incoming beam, and suitably optimising the gain. For the best achieved position resolution to date, and for 100\% bunch-to-bunch correlation, an ultimate beam stabilization to about 15~nm is in principle achievable with the current hardware. Should ATF/ATF2 beam operations resume, this will be the subject of future feedback studies.
\section{ACKNOWLEDGEMENTS}
We thank the KEK ATF staff for their outstanding logistical support and for providing the beam time and the necessary stable operating conditions for this research. In addition, we thank our colleagues from the ATF2 collaboration for their help and support. In particular, we thank the KNU group for fabricating the low-quality-factor BPMs, the LAL group from the Paris-Saclay University for providing the BPM mover system, and the KEK group for making available the analogue signal-processing down-mixing electronics.
We acknowledge financial support for this research from the United Kingdom Science and Technology Facilities Council via the John Adams Institute, University of Oxford, and CERN, CLIC-UK Collaboration, Contract No. KE1869/DG/CLIC. The research leading to these results has received funding from the European Commission under the Horizon 2020/Marie Sklodowska-Curie Research and Innovation Staff Exchange (RISE) project E-JADE, Grant Agreement No. 645479.
\providecommand {\noopsort}[1]{}\providecommand {\singleletter}[1]{#1}%
|
\section{Introduction}
The detections of gravitational waves from black hole (BH) binary mergers by the Ligo and Virgo Scientific Collaborations~\cite{ligo1_2016,ligo2_2016}, together with the first image of a BH shadow~\cite{EHT_sombra}, mark the beginning of a new era for gravitational physics. In particular, the physics around BHs has become even more important, since these phenomena may now be better constrained by the experimental data available.
BH spacetimes possess a nontrivial causal structure and their strong gravitational field implies that they are important pieces not only in General Relativity (GR), but also in alternative theories of gravity~\cite{wald_1984,berti_2015}.
On one hand, the classical aspects of BH interactions with matter have importance in the astrophysical context, since these compact objects are believed to exist in the center of almost all galaxies, being one of the main sources of power in them~\cite{Hoyle1963,Salpeter1964,Lynden-Bell1969,Kormendy2013}.
On the other hand, BHs causal structure provides an interesting setting to study the quantum aspects of matter, particularly fields, since we can use the quantum field theory (QFT) in curved spaces framework. The theoretical implications of this analysis for the quantum fields are interesting in their own right, but they might also provide insights to the quantum nature of gravity.
Ever since its formulation more than a hundred years ago, GR has passed several experimental tests (see Ref.~\cite{Turyshev2008} for a review) and is regarded as the established theory to describe the physics of spacetime. However, it is a classical theory plagued with singularities~\cite{hawking_1996}, which breaks down in the Planck scale, such as in the very beginning of our universe or inside a BH. A satisfactory quantum theory of gravity is still missing, being one outstanding issue in fundamental physics~\cite{kiefer_2005}. In an effort to understand possible quantum aspects of gravity, we may use an effective approach, in which quantum fields are coupled to the gravitational field represented by a fixed classical background spacetime~\cite{birrel_1982,parker_2009}. Even if QFT in curved spacetimes is not a fundamental theory, important results have been obtained using this framework, namely the particle creation in expanding universes~\cite{parker_1969} and BH spacetimes (Hawking radiation)~\cite{hawking_1975}, as well as the Unruh effect, which establishes the observer-dependent nature of the particle content in field theory~\cite{unruh_1976}. Hence, although being an effective theory, QFT in curved spacetimes sheds light on important aspects of the interface between gravity and quantum physics.
Accretion disks are expected to surround most BHs found in the Universe. As matter spirals around these compact objects, it emits radiation through several channels. As previously mentioned, BHs and their surrounding matter are generally a main source of power in galaxies and the associated radiation processes play an important role in high-energy astrophysics.
QFT may be used to compute radiation emission by accelerated sources in flat~\cite{alves_2004,alves_2010,castineiras_2011} and curved spacetimes~\cite{castineiras_2002}.
In the context of radiation processes near BHs, QFT can be used to analyze the interaction of sources with Hawking radiation~\cite{crispino_1998} and the radiation emission by sources in geodesic motion.
We may analyze these phenomena by considering a source moving along circular geodesics. For unstable orbits, there is the possibility of radiation emission of the synchrotron type, hence this emission is usually called the geodesic synchrotron radiation. This was initially analyzed for a scalar field in the Schwarzschild background using the Green function framework~\cite{misner_1972,misner_et_al_1972}. We can treat this phenomenon with a semiclassical approach using the QFT in curved spacetimes framework. The scalar radiation was investigated in asymptotically flat spacetimes, e. g., in Refs.~\cite{crispino_2000,castineiras_2007,crispino_2008,crispino_2009,macedo_2012,bernar_2019}, the electromagnetic radiation was studied in Ref.~\cite{castineiras_2005} and the gravitational radiation was analyzed in Refs.~\cite{bernar_2017,bernar_2018}. In the context of spacetimes with nonvanishing cosmological constant, the scalar radiation was studied in the Schwarzschild--de Sitter (SdS) spacetime in Ref.~\cite{brito_2020} and in the Schwarzschild--anti--de Sitter (SAdS) spacetime, using the Green function method, in Ref.~\cite{cardoso_2002}. In the SAdS spacetime an investigation of the geodesic scalar radiation using QFT in curved spacetimes has not been previously carried out.
In particular, one motivation to study dynamical processes in asymptotically anti--de--Sitter spacetimes is concerned with the conjectured anti--de--Sitter/conformal field theory (AdS/CFT) correspondence, in which physics in the bulk of AdS is related to the physics in the conformal field theory boundary (see, e.g., Ref.~\cite{aharony_2000} and the references therein).
Solutions of the GR field equations with a nonvanishing cosmological constant are not asymptotically flat. In particular, for the SAdS solution ($\Lambda < 0$), in contrast to an asymptotically flat solution, the spacetime boundary at spatial infinity is a timelike surface.
The SAdS solution describes a nonglobally hyperbolic spacetime and, therefore, does not possess a Cauchy surface on which one can give initial data. In order to have a well defined QFT, we must impose suitable boundary conditions~\cite{isham_1978,morley_2020,ishibashi_2004,wald_1980}. In this paper, we use QFT in curved spacetimes at tree level to analyze the scalar radiation emitted by a source in geodesic circular motion around a SAdS BH, considering reflective boundary conditions at infinity.
The remainder of this paper is organized as follows. In Sec.~\ref{sec_SdS_black_hole}, we review some general features of the SAdS spacetime, analyzing the geodesic structure and its associated conserved quantities. In Sec.~\ref{sec_scalar_radiation}, we study the scalar field dynamics in the curved background and we canonically quantize the field, choosing reflective boundary conditions at $r \to \infty,$ where the field effective potential diverges. In Sec.~\ref{Sec_radiation}, we compute the power emitted by the rotating source considering first order in perturbation theory and using numerical methods to solve the differential wave equation. In Sec.~\ref{sec_results}, we exhibit some selected results and we present our final remarks in Sec.~\ref{Sec_remarks}. In this paper, we adopt natural units such that $c=G=\hbar=1$ and the metric signature ($-,+,+,+$).
\section{Schwarzschild--anti--de Sitter black holes}
\label{sec_SdS_black_hole}
In this section, we present some important properties of the SAdS spacetime, which is a spherically symmetric vacuum solution of Einstein field equations with negative cosmological constant, characterized by a central black hole with geometric mass $M.$ The SAdS line element is given by~\cite{stuchlik_1999}
\begin{equation}
\label{SAdS_line_element}
ds^2 = -f_{\Lambda}(r)dt^2 + \frac{dr^2}{f_{\Lambda}(r)} + r^2(d\theta^2 + \sin^2 \theta d\phi^2),
\end{equation}
with
\begin{equation}
\label{f}
f_{\Lambda}(r) \equiv 1 - \frac{2 M}{r} - \frac{\Lambda}{3}r^2.
\end{equation}
The line element, given by Eq.~\eqref{SAdS_line_element}, is static for $r>r_e,$ where $r_e$ is the radial location of the black hole event horizon ($H_e$). The radial position $r_e$ is obtained by solving
\begin{equation}
\label{f_zero}
f_{\Lambda}(r_{e}) = 0,
\end{equation}
from which the real solution is given by
\begin{equation}
\label{event_horizon}
r_e = -\frac{\Lambda + \zeta^{2/3}}{\Lambda \zeta^{1/3}},
\end{equation}
where
\begin{equation}
\label{defined_function}
\zeta \equiv \sqrt{\Lambda^3 \left(9 M^2 \Lambda -1 \right)} + 3 M \Lambda^2.
\end{equation}
Another useful quantity is the so-called anti-de Sitter radius, which is the scale associated with the cosmological constant and given by
\begin{equation}
R_{AdS} = \sqrt{\frac{3}{\abs{\Lambda}}}.
\end{equation}
The Schwarzschild solution is obtained in the limit $\Lambda \rightarrow 0,$ in Eq.~\eqref{f}, for which $r_e \rightarrow 2M.$ The anti--de Sitter (AdS) solution~\cite{hawking_1973} is obtained in the limit $M \rightarrow 0,$ for which $r_e \rightarrow 0.$
For a fixed nonvanishing $M$, in the limit $\Lambda \rightarrow - \infty$ we have that $r_e \rightarrow 0$.
The behavior of $f_{\Lambda}(r)$ is illustrated in Fig.~\ref{fig_radial_function}.
\begin{figure}
\includegraphics[scale=0.42]{radial_function.pdf}
\caption{The function $f_{\Lambda}(r),$ given by Eq. (\ref{f}). The gray region covers all possible values of $\Lambda$ in the interval $-\infty < \Lambda < 0.$}
\label{fig_radial_function}
\end{figure}
We shall consider a source that orbits the black hole along circular geodesics in the equatorial plane.
The choice of motion in the equatorial plane can be done, without loss of generality, due to the spherical symmetry of the SAdS spacetime.
In the next subsection, we review some features of the SAdS circular geodesics.
\subsection{Geodesics in SAdS spacetime}
The dynamics of test-particles in a curved background can be derived from the Lagrangian
\begin{equation}
\label{lagrangian}
\mathcal{L}_P = \frac{1}{2}g_{\mu \nu} \dot{x}^{\mu} \dot{x}^{\nu},
\end{equation}
where the overdot denotes differentiation with respect to an affine parameter (in the case of timelike paths, we identify the affine parameter as the particle's proper time). The components of the metric $g_{\mu \nu}$ are obtained from Eq.~(\ref{SAdS_line_element}).
There are two integrals of motion along the geodesics in the equatorial plane ($\theta = \pi/2$ and $\dot{\theta} = 0$) due to the time and axial Killing vectors, $\partial_t$ and $\partial_{\phi},$ respectively.
The corresponding conserved quantities can be written as
\begin{eqnarray}
\label{E_circular}
P_t &=& -\frac{\partial \mathcal{L}_P}{\partial \dot{t}} = f_{\Lambda}(r) \dot{t} \equiv E, \label{integral_motion_E}\\
\label{L_circular}
P_{\phi} &=& \frac{\partial \mathcal{L}_P}{\partial \dot{\phi}} = r^2 \dot{\phi} \equiv L. \label{integral_motion_L}
\end{eqnarray}
The quantities $E$ and $L$ are identified as the specific energy and angular momentum of the test-particle, as seen by a static observer at the radial position $r_{obs} \equiv (6M/\abs{\Lambda})^{1/3}.$ Note that $r_{obs} \rightarrow \infty$ as $\abs{\Lambda} \rightarrow 0.$
Given that $2 \mathcal{L}_P \equiv \epsilon = -1$ ($0$) for timelike (null) geodesics, using Eqs.~\eqref{integral_motion_E} and \eqref{integral_motion_L} one finds that the test-particle motion is determined by the first order non-linear differential equation:
\begin{equation}
\label{energy}
\dot{r}^2 = E^2 - 2V_P(r),
\end{equation}
where we have defined the central potential as
\begin{equation}
\label{central_potenctial}
V_P(r) \equiv \frac{1}{2}f_{\Lambda}(r) \left(-\epsilon + \frac{L^2}{r^2} \right).
\end{equation}
For timelike geodesics, the potential $V_P(r)$ presents points of maximum and minimum. It vanishes at the black hole horizon $r_e$ and diverges at infinity ($r=\infty$). For null geodesics, the potential $V_P(r)$ has a maximum at the radial position $r_0 \equiv 3M$ (unstable photon sphere). It vanishes at the event horizon and tends to $-L^{2}\Lambda/6$ [$= V_P(2M)$] at infinity. Accordingly, massless particles can escape to infinity, in contrast to massive particles, which are confined due to the diverging potential barrier as $r \to \infty.$
Let us analyze circular geodesics that are constrained by the conditions $\dot{r} = 0$ and $\ddot{r} =0.$ From Eqs.~\eqref{energy} and \eqref{central_potenctial}, we obtain the conserved quantities along the timelike circular geodesics:
\begin{equation}
\label{circular_E_L}
E^2 = r\frac{f_{\Lambda}(r)^2}{r-3M}, \hspace{1 cm} L^2 = r^2 \frac{M - r^3 \Lambda /3}{r - 3M}.
\end{equation}
Noting that $E$ and $L$ must be real quantities, it follows that circular geodesics must exist in the region
\begin{equation}
\label{circular_range}
3M < r < +\infty.
\end{equation}
The lower limit of the interval given by Eq.~\eqref{circular_range} is the radial position of the circular null geodesic, $r_0$, which is independent of $\Lambda.$
By analyzing the central potential given by Eq.~\eqref{central_potenctial}, it is straightforward to obtain the condition that stable circular orbit parameters in the SAdS spacetime must obey, namely
\begin{equation}
\label{stable_condition}
S(\Lambda,r) \equiv -4 \Lambda r^4 + 15 \Lambda M r^3 + 3Mr - 18 M^2 \geq 0.
\end{equation}
In Fig.~\ref{fig_F_stability}, we show the surface function $S(\Lambda, r)$ defined in Eq.~\eqref{stable_condition}. The stable orbits are associated with the parameters $\Lambda$ and $r$ of the surface on and above the hatched plane [$S(\Lambda,r) \geq 0$].
We see that, as the value of $\Lambda$ decreases, the radial position of the \textit{innermost stable circular orbit}, $r_{ISCO}$, also decreases, since it is associated to the locus $S(\Lambda,r) = 0$ (dashed curve).
As $\Lambda \rightarrow -\infty,$ we have $r_{ISCO} \rightarrow r^{min}_{ISCO} \equiv 15 M/4$ (solid line). For $\Lambda = 0$, we have $S(\Lambda,r) \geq 0$ for $r \geq 6M$ and $r_{ISCO}=6M$.
\begin{figure}
\includegraphics[scale=0.4]{F_stability.pdf}
\caption{The function $S(\Lambda,r),$ defined in Eq.~(\ref{stable_condition}). The dashed curve is the intersection of the surface $S$ with the gray plane ($S=0$). The solid line is the asymptote of the dashed curve along $\Lambda.$}
\label{fig_F_stability}
\end{figure}
The angular velocity of a massive particle moving along circular geodesics, as seen by the static observer at $r_{obs},$ is given by
\begin{equation}
\label{angular_velocity}
\Omega \equiv \frac{d \phi}{dt} = \frac{\dot{\phi}}{\dot{t}} = \sqrt{\frac{M}{r^3} - \frac{\Lambda}{3}},
\end{equation}
which, for $\Lambda<0,$ tends to a minimum positive constant value , $\Omega_{min} = \sqrt{-\Lambda/3}$, as the orbit radial position $r \rightarrow \infty$.
The angular velocity $\Omega$ tends to a maximum value, $\Omega_{max}$,
as $r \rightarrow 3M$.
We obtain the power of scalar radiation emitted by the source in terms of observable quantities such as $\Omega$, $M$ and $\Lambda$. One can invert Eq.~\eqref{angular_velocity} to write the radial position $r$ of the circular orbit as a function of $\Omega$ and $\Lambda.$
In asymptotically AdS spacetimes, we must specify, beyond the initial configuration, suitable boundary conditions at infinity~\cite{hawking_1983} in order to have a well defined field theory. In the next section, we analyze the massless scalar field in the SAdS spacetime. To canonically quantize the field, we follow the prescription similar to that of Schwarzschild~\cite{birrel_1982,parker_2009,boulware_1975,
ashtekar_1975} and SdS spacetimes~\cite{higuchi_1987,brito_2020}.
\section{Scalar field in Schwarzschild--anti--de Sitter}
\label{sec_scalar_radiation}
We consider a minimally coupled massless real scalar field $\Phi(x).$ Its dynamics in a curved background can be derived from the following action:
\begin{equation}
\label{action}
S =-\frac{1}{2}\int \nabla_{\mu} \Phi(x) \nabla^{\mu} \Phi(x) \sqrt{-g} \, d^4x.
\end{equation}
The equation of motion can be written as
\begin{equation}
\label{KG}
\nabla_{\mu}\nabla^{\mu}\Phi(x) \equiv \frac{1}{\sqrt{-g}}\partial_{\mu} \left( \sqrt{-g} g^{\mu \nu} \partial_{\nu} \Phi \right) = 0,
\end{equation}
where $g = - r^4 \sin^2 \theta$ is the determinant of the SAdS metric.
We may define the positive-frequency modes, solutions of Eq.~\eqref{KG}, associated to the timelike Killing vector field $\partial_t,$ in the form given by
\begin{equation}
\label{Mode_positive}
u_{\omega l m}(x) = \sqrt{\frac{\omega}{\pi}} \frac{\Psi_{\omega l}(r)}{r} Y_{l m}(\theta, \phi) e^{- i \omega t} \hspace{0.4 cm} (\omega > 0),
\end{equation}
where the quantities $Y_{l m}(\theta, \phi)$ are the scalar spherical harmonics~\cite{NIST_handbook}. Using Eq.~\eqref{Mode_positive} and the eigenvalue equation of $Y_{l m}(\theta, \phi),$ it is possible to simplify Eq.~\eqref{KG} to the ordinary differential equation:
\begin{equation}
\label{dif_radial}
\left[ - f_{\Lambda}\frac{d}{dr} \left( f_{\Lambda}\frac{d}{dr}\right) + V_{eff}\right]\Psi_{\omega l}(r) = \omega^2 \Psi_{\omega l}(r),
\end{equation}
with the effective potential $V_{eff}$ defined by
\begin{equation}
\label{effective_potential}
V_{eff}(r) \equiv f_{\Lambda}(r)\left( \frac{l(l+1)}{r^2} + \frac{2M}{r^3} - \frac{2 \Lambda}{3} \right).
\end{equation}
The potential \eqref{effective_potential} is illustrated in Fig.~\ref{fig_effective_potential}. We see that $V_{eff}(r)$ vanishes at the horizon $r_{e}$ and diverges at infinity. The potential also increases as the parameter $\Lambda$ decreases from $0$. For sufficiently small values of $\abs{\Lambda}$ (or sufficiently large values of $l$) the potential presents points of local maximum and minimum.
In contrast, the potential in the Schwarzschild spacetime presents only a point of maximum near the radial position of the circular null geodesic.
The local minimum of the effective potential in the SAdS spacetime gives rise to quasibound states, which are long-lived quasinormal modes~\cite{grain_2006,festuccia_2009,daghigh_2009,berti_2009}.
Expanding the potential $V_{eff}(r),$ defined in Eq.~\eqref{effective_potential}, near infinity, its leading order asymptotic approximation can be given by
\begin{equation}
\label{potential_at_infinity}
V_{eff}(r \rightarrow \infty) \sim \frac{2 \Lambda ^2 r^2}{9}-\frac{1}{3} \Lambda (l (l+1)+2).
\end{equation}
\begin{figure}
\includegraphics[scale=0.42]{fig_effective_potential.pdf}
\caption{The effective potential $M^2V_{eff},$ defined in Eq.~(\ref{effective_potential}), with $l=5$ and five different values of the cosmological constant $M^2\Lambda.$ The vertical gray line denotes the radial position of the photon sphere.}
\label{fig_effective_potential}
\end{figure}
In principle, we can derive asymptotic solutions of Eq.~\eqref{dif_radial} using Eq.~\eqref{potential_at_infinity}. However, in order to analyze the behavior of the field at infinity, we perform a change of variable in the radial coordinate valid outside the horizon [$r \in (r_e,+\infty) \rightarrow \rho \in (- \infty, \infty$)], given by~\cite{winstaley_2001}
\begin{equation}
\label{R_coordenate}
\rho \equiv \ln(r-r_e),
\end{equation}
together with
\begin{equation}
\label{fomega}
\Psi_{\omega l}(r) = \sqrt{\frac{r-r_e}{f_{\Lambda}(r)}} \psi_{\omega l}(\rho).
\end{equation}
From Eq.~\eqref{R_coordenate}, we see that $\rho \rightarrow -\infty$ as $r \rightarrow r_e$ and $\rho \rightarrow +\infty$ as $r \rightarrow +\infty$. Substituting Eqs.~\eqref{R_coordenate} and \eqref{fomega} into Eq.~\eqref{dif_radial}, the Klein--Gordon equation becomes
\begin{equation}
\label{dif_with_rho}
\frac{d^2}{d \rho^2}\psi_{\omega l}(\rho) + V(r)\psi_{\omega l}(\rho) = 0,
\end{equation}
with the potential function $V(r)$ defined by
\begin{equation}
\label{potential_for_dif_with_rho}
V(r) \equiv - \frac{1}{4} + \frac{\left(r-r_e\right){}^2}{f_{\Lambda }(r){}^2} \left(\omega^2 - V_{\text{eff}}(r) - \frac{\chi}{4} \right),
\end{equation}
where
\begin{equation}
\label{function_defined_2}
\chi \equiv \left(2 f_{\Lambda }(r) f_{\Lambda }''(r)-f_{\Lambda}'(r){}^2 \right)
\end{equation}
and $r$ should be seen as a function of $\rho$, defined by the inversion of Eq.~(\ref{R_coordenate}).
At infinity ($r \rightarrow + \infty$), the function $V(r)$ tends to a constant value:
\begin{equation}
\label{potential_rho_inf}
V(r \rightarrow\infty) = - 9/4.
\end{equation}
At the event horizon $r_e$ ($r \rightarrow r_{e}$), the potential is approximated to
\begin{equation}
\label{potential_rho_r_e}
V(r \rightarrow r_e) \sim \frac{\omega^{2}}{f'_{\Lambda}(r_{e})^{2}} \equiv \bar{\omega}^2.
\end{equation}
With Eqs.~\eqref{potential_rho_inf} and \eqref{potential_rho_r_e}, we obtain the asymptotic solutions to Eq.~\eqref{dif_with_rho}. Near the event horizon, the solution is expressed by
\begin{equation}
\label{asymp_rho_re}
\psi_{\omega l} (\rho) \sim A_{\omega l} e^{i \bar{\omega} \rho} + B_{\omega l} e^{- i \bar{\omega} \rho},
\end{equation}
where the modes with coefficients $A_{\omega l}$ correspond to flux incoming from the past event horizon $H_e^-$ and those with coefficients $B_{\omega l}$ correspond to the flux reflected back down to the future event horizon $H_e^+.$
At infinity, we can write the solution as
\begin{equation}
\label{asymp_rho_inf}
\psi_{\omega l} (\rho) =\frac{C_{\omega l}}{\sqrt{2}} \left( e^{\frac{3}{2} \rho} - i e^{-\frac{3}{2} \rho} \right) + \frac{D_{\omega l}}{\sqrt{2}} \left( e^{\frac{3}{2} \rho} + i e^{-\frac{3}{2} \rho} \right),
\end{equation}
where the modes with coefficients $C_{\omega l}$ are associated to outgoing flux at infinity and those with coefficients $D_{\omega l}$ are associated to incoming flux (see Ref.~\cite{winstaley_2001}). We impose reflective boundary conditions at infinity, which means that the radial flux must be zero. This condition is satisfied if $|C_{\omega l}|=|D_{\omega l}|,$ i.e. if $C_{\omega l}=D_{\omega l}$ or $C_{\omega l}=-D_{\omega l}.$
Hence, there are two linearly independent solutions to Eq.~\eqref{dif_with_rho}, with asymptotic forms given by
\begin{equation}
\label{asymptotic-1}
\psi^1_{\omega l} \sim \mathcal{A}^1_{\omega l} \begin{cases}
e^{i \bar{\omega} \rho} + \mathcal{B}^1_{\omega l}e^{-i\bar{\omega} \rho} , & \rho \rightarrow -\infty, \\
\mathcal{N}^1_{\omega l} e^{\frac{3}{2} \rho}, & \rho \rightarrow +\infty,
\end{cases}
\end{equation}
and
\begin{equation}
\label{asymptotic-2}
\psi^2_{\omega l} \sim \mathcal{A}^2_{\omega l} \begin{cases}
e^{i \bar{\omega} \rho} + \mathcal{B}^2_{\omega l}e^{-i\bar{\omega} \rho} , & \rho \rightarrow -\infty, \\
\mathcal{N}^2_{\omega l} e^{-\frac{3}{2} \rho}, & \rho \rightarrow +\infty,
\end{cases}
\end{equation}
where $\mathcal{A}^k_{\omega l} $ (with $k=1,2$) are overall normalization constants.
The constants $\mathcal{N}^k_{\omega l}$ and $\mathcal{B}^k_{\omega l}$ are complex quantities. Note that we have defined the basis modes such that they are associated with a unitary flux coming up out of $H_e^-$.
By considering the Wronskian of the solutions, given by Eqs.~\eqref{asymptotic-1} and \eqref{asymptotic-2}, we can show that $\abs{\mathcal{B}^{k}_{\omega l}} = 1,$ due to conservation of the flux.
The function $\psi^{1}_{\omega l}$ diverges as $\rho \to \infty$ or, equivalently, $r \to \infty$. However, the actual scalar mode $\Psi^{1}_{\omega l m}/r$ is proportional to $R_{{AdS}}$ as $r \to \infty$.
To normalize the modes $u_{\omega l m},$ we use the inner product defined by
\begin{equation}
\left(\Psi_1, \Psi_2 \right) \equiv i \int_{\Sigma} d\Sigma^{\mu}\left( \Psi_1^* \left(\nabla_{\mu} \Psi_2 \right) - \left( \nabla_{\mu} \Psi_1^* \right)\Psi_2 \right),
\label{inner_product}
\end{equation}
where $d\Sigma^{\mu} = d\Sigma n^{\mu},$ with $n^{\mu}$ being a future directed unit vector orthogonal to a hypersurface $\Sigma$, defined by $t=\text{constant}$~\cite{hawking_1973, parker_2009}. It can be shown that this inner product does not depend on the choice of hypersurface if $\Psi_{1}$ and $\Psi_{2}$ satisfy the field equations~\cite{parker_2009}.
In globally hyperbolic spacetimes, such as the Kerr--Newman spacetime family, the surface $\Sigma$ is a Cauchy surface.
Imposing the orthogonality conditions
\begin{equation}
\label{u_orthogonality}
\left(u_{\omega l m},u_{\omega' l' m'} \right) = \delta_{l l'} \delta_{m m'} \delta(\omega - \omega')
\end{equation}
and
\begin{equation}
\label{u_ortogonality_null}
\left( u_{\omega l m},u^{*}_{\omega' l' m'} \right) = \left( u^{*}_{\omega l m},u_{\omega' l' m'} \right) = 0,
\end{equation}
we can obtain the absolute value of the overall normalization constants $\mathcal{A}^k_{\omega l},$ namely
\begin{equation}
\label{normali_A_1}
\abs{\mathcal{A}^k_{\omega l}} = \frac{1}{2 \sqrt{\omega \bar{\omega}}}.
\end{equation}
We canonically quantize the scalar field following the prescription for a static spacetime~\cite{birrel_1982,parker_2009,higuchi_1987,
ashtekar_1975}, in which one expands the quantum field operator $\hat{\Phi}(x)$ in terms of negative and positive frequencies associated to creation ($\hat{a}^{\dagger}_{\omega l m}$) and annihilation ($\hat{a}_{\omega l m}$) operators. The quantum field is thus given by
\begin{equation}
\label{field_expansion}
\hat{\Phi}(x) = \sum_{l,m, k} \int_0^{\infty} \left[u^k_{\omega l m}(x)\hat{a}^k_{\omega l m} + u^{k *}_{\omega l m}(x) \hat{a}^{k \dagger}_{\omega l m} \right] d\omega,
\end{equation}
where the index $k=1,2$ stands for modes associated to the two solutions given by Eqs.~\eqref{asymptotic-1} and \eqref{asymptotic-2}. The creation and annihilation operators satisfy the usual nonvanishing commutation relations, namely
\begin{equation}
\label{comutation_relation}
\left[\hat{a}^k_{\omega l m},\hat{a}^{k \dagger}_{\omega' l' m'} \right] = \delta_{l l'} \delta_{m m'} \delta(\omega - \omega').
\end{equation}
The vacuum state $|0\rangle$ can be defined as the quantum state that vanishes when acted upon by all annihilation operators~\cite{fulling_1973}. Thus,
\begin{equation}
\label{vacuum}
\hat{a}^k_{\omega l m} \ket{0} \equiv 0, \hspace{1 cm} \forall \hspace{0.3 cm} (\omega, l, m, k),
\end{equation}
and the one-particle-states are obtained by applying a creation operator on the vacuum state, that is
\begin{equation}
\label{one_parti_state}
\ket{k;\omega l m} = \hat{a}^{k \dagger}_{\omega l m} \ket{0}.
\end{equation}
In the next section, we compute the one-particle-emission amplitude by considering the massless scalar field coupled to a classical matter source in a timelike geodesic circular orbit on the equatorial plane of a SAdS black hole.
\section{Scalar radiation and emitted power}
\label{Sec_radiation}
The orbiting scalar source at $r=R,$ with angular velocity $\Omega$, is described by the normalized current
\begin{equation}
\label{current}
j(x) = \frac{\varrho}{\sqrt{-g} u^0} \delta(r-R) \delta(\theta - \pi/2)\delta(\phi - \Omega t),
\end{equation}
such that $\int d\beta^{(3)}j(x) = \varrho,$ where $\beta^{(3)}$ is a hypersurface orthogonal to the source's $4-$velocity $u^{\mu},$ and $\varrho$ is the magnitude of the scalar source.
The associated $4-$velocity is given by
\begin{equation}
\label{four_velocity}
u^{\mu}(R) = u^0 \left(1,0,0,\Omega \right),
\end{equation}
with $u^0$ fixed by the condition $u^{\mu}u_{\mu} = -1.$ Therefore,
\begin{equation}
\label{gamma_factor}
u^0 = \left( f_{\Lambda}(R) - R^2 \Omega^2 \right)^{-1/2} = \left( 1 - \frac{r_0}{R} \right)^{-1/2}.
\end{equation}
The interaction between the field and the source is carried out by adding to the action \eqref{action} the following term:
\begin{equation}
\label{interaction_action}
\hat{S}_{I} = \int d^4x \sqrt{-g} j(x) \hat{\Phi}(x),
\end{equation}
so that the scalar source $\varrho$ can be regarded as a coupling constant that determines the magnitude of the interaction.
There is a nonvanishing probability for the source to radiate scalar quanta. To first order in perturbation theory, the transition amplitude from the vacuum-state, defined in Eq.~(\ref{vacuum}), to the one-particle-state, with energy $\omega$ and quantum numbers $l,$ $m,$ expressed in Eq.~\eqref{one_parti_state}, is written as~\cite{itzykson_1980}
\begin{equation}
\label{probability_amplitude}
A^{k; \omega l m}_{em} = \bra{k; \omega l m} i \hat{S}_I \ket{0} = i \int d^4x \sqrt{-g} j(x)u^{k *}_{\omega l m}(x).
\end{equation}
As can be seen by evaluating the integral, the probability amplitude, given by Eq.~(\ref{probability_amplitude}), is proportional to $\delta(\omega - m\Omega),$ i.e. the emitted scalar particles have frequency proportional to the angular velocity of the source ($\omega_m \equiv m\Omega$).\footnote{Although we are considering only lowest order in perturbation theory, non-perturbative calculations in Minkowski spacetime for the scalar radiation emitted by a uniformly accelerated source show that only zero-energy Rindler particles compose the final state~\cite{Landulfo2019,Fulling2020}.
The present case shares most of its technical details with the uniformly accelerated source case presented in~\cite{Landulfo2019}. Thus, we might expect that only particles obeying the condition $\omega = m \Omega$ will be emitted.}
Furthermore, there is no emission of observable scalar particles with $m=0.$ We note that $\omega$ and $\Omega$ are positive quantities, hence we must have $m \geq 1.$
The emitted power, with fixed $l$ and $m$ for each mode $k$, is given by
\begin{equation}
\label{partial_power_implicit}
W^{k;l m}_{em} = \int_{0}^{\infty} d\omega \ \omega \frac{\abs{A^{k; \omega l m}_{em}}^2}{T},
\end{equation}
where $T=\int dt = 2 \pi \delta(0)$ is the total time during which the source radiates, assumed to be the whole range of coordinate time $t,$ with $-\infty < t < \infty$~\cite{breuer_1975,crispino_1998}.
With Eqs.~\eqref{probability_amplitude} and \eqref{partial_power_implicit}, we readily obtain the partial emitted power, namely
\begin{equation}
\label{partial_power}
W^{l m}_{em} = 2 \varrho^2 \omega_m^2 \gamma \left( \abs{\psi^1_{\omega_m l}}^2 + \abs{\psi^2_{\omega_m l}}^2\right) \abs{Y_{l m}\left(\frac{\pi}{2}, 0 \right)}^2,
\end{equation}
with
\begin{equation}
\label{GR_factor}
\gamma \equiv \left( 1 - \frac{r_0}{R} \right)\frac{R-r_e}{R^2 f_{\Lambda}(R)}.
\end{equation}
From Eq.~\eqref{GR_factor}, we see that $W^{l m}_{em} \rightarrow 0$ as $R \rightarrow r_0$ and $R \rightarrow \infty.$
Furthermore, there is no emission for odd values of $l+m,$ because the time independent quantity $\abs{Y_{l m}(\pi/2, 0)}$ vanishes in this case. For even values of $l+m,$ we can write~\cite{gradshteyn_1980}
\begin{equation}
\label{harmonic_even_values}
\abs{Y_{l m}\left(\frac{\pi}{2}, 0 \right)}^2 = \frac{2l+1}{4\pi}\frac{(l+ m - 1)!!(l - m - 1)!!}{(l+m)!!(l-m)!!}.
\end{equation}
The total emitted power $W_{em}$ is obtained by summing the contributions of all multipoles, namely
\begin{equation}
\label{total_power}
W_{em} = \sum_{l=1}^{\infty} \sum_{m=1}^{l} W^{l m}_{em}.
\end{equation}
In order to obtain the emitted power, we solve numerically the differential equation for $\psi^k_{\omega l},$ given by ~\eqref{dif_with_rho}. We follow a procedure similar to that stated in Schwarzschild and SdS spacetimes. In the next section, we present some representative results.
\begin{figure*}
\centering
\includegraphics[scale=0.45]{Pot_Parc_SAdS_Lamb1_2.pdf} \hspace{0.3 cm}
\includegraphics[scale=0.45]{Pot_Parc_SAdS_Lamb1.pdf}
\includegraphics[scale=0.45]{Pot_Parc_SAdS_Lamb3.pdf}
\hspace{0.3 cm}
\includegraphics[scale=0.45]{Pot_Parc_SAdS_Lamb5.pdf}
\caption{The emitted power as a function of $M\Omega,$ given by Eq.~\eqref{partial_power}, for $M^2 \Lambda = -2^{-1}$ (top-left), $M^2 \Lambda = -1$ (top-right), $M^2 \Lambda = -3$ (bottom-left), $M^2 \Lambda = -5$ (bottom-right) and different choices of $l=m,$ as indicated. The curves are plotted from $\Omega_{min},$ corresponding to the radial position $R \rightarrow + \infty,$ up to the value $\Omega_{max},$ corresponding to the radial position $R=3M$.}
\label{Pot_Parc_SAdS_Lamb1}
\end{figure*}
\section{Results for the emitted power}
\label{sec_results}
Using methods outlined in Refs.~\cite{bernar_2017,brito_2020}, we compute numerically the scalar power emitted by the source for different choices of the parameters. We start with the integration of Eq.~\eqref{dif_with_rho}, using boundary conditions given by Eqs.~\eqref{asymptotic-1} and \eqref{asymptotic-2}. The integration is performed in the region $r_e + \delta^{-1}r_e \leq r \leq \delta r_e,$ where $\delta \equiv 10^{5}.$ We compute the coefficients $\abs{\mathcal{N}^{k}_{\omega l}}$ by matching the numerically obtained solutions with the asymptotic forms, given by Eqs.~\eqref{asymptotic-1} and \eqref{asymptotic-2}, near the event horizon.
Note that we are free to choose either $M$ or $R_{AdS}$ equal to the unity. This choice corresponds to rescaling the radial coordinate as $r \rightarrow r/M$ or $r\rightarrow r/R_{AdS},$ respectively. In this section, we display our results considering $M$ as the fundamental scale, making use of the $R_{AdS}$ scale when convenient.
We shall analyze the radiation scenario in which $M^2\Lambda \leq -1/2,$ such that $r_e > 0.6 R_{AdS},$ which corresponds to a wide variety of black hole sizes, when compared to the anti--de Sitter radius.
We note that, for the values of $\Lambda$ and $l$ considered in this section, more than $98\%$ of the emitted power comes from the $l=m$ mode. This percentage decreases with decreasing $\Lambda.$
In Fig.~\ref{Pot_Parc_SAdS_Lamb1}, we show the partial emitted power for four choices of the cosmological constant $\Lambda$ and different choices of the multipole number $l=m.$ We see, as a characteristic behavior, that the power for each value of $l=m$ increases from zero at $\Omega_{min}$ ($R \rightarrow \infty$), reaches a maximum and decreases to zero as $\Omega \rightarrow \Omega_{max}$ ($R \rightarrow r_0$).
For all values of $\Lambda$ considered, we have found that the peak of the partial emitted power increases with $l=m$.
Additionally, the value of $\Omega$ where this peak is located gets closer to $\Omega_{max}$ as we increase $l=m$.
This shows the dominance of higher multipole modes in the power for orbits closer to the photon sphere.
We also see that emission of higher multipole modes is enhanced with decreasing $M^2\Lambda$. For spacetimes with $|\Lambda| \geq 3/M^{2}$, even orbits far from the black hole, i.e. stable orbits, present this behavior of increasing partial emitted power with $l=m$. We note that $|\Lambda| = 3/M^{2}$ is the point in the parameter space where $r_{e}=R_{AdS}$. Thus, the region $|\Lambda| \geq 3/M^{2}$ in the parameter space denotes the large black hole regime.
This is in agreement with the classical result reported in Ref.~\cite{cardoso_2002}, where the Green's function approach was used.
In contrast, the analogous case in SdS spacetime ($\Lambda > 0$) presents an enhancement of lower multipole modes as $M^2\Lambda$ increases~\cite{brito_2020}.
For $\Lambda = -1/(2M^{2}),$ the last stable circular orbit occurs at $r \approx 3.8 M$ (or $M \Omega \approx 0.43$).
We see, on Fig.~\ref{fig:partial_power_stable_orbits_Lamb_1_2}, that at this position the power is dominated by the contribution from modes with $l \geq 2$.
Furthermore, this does not occur in the Schwarzschild (cf. Fig.~2 of Ref.~\cite{crispino_2000}) or SdS (cf. Fig.~7 of Ref.~\cite{brito_2020}) spacetime scenarios.
Numerically we find that, for a circular orbit in $r_{ISCO}$, we have $W_{em}^{2 2} \geq W_{em}^{1 1}$ when $M^2 \Lambda \lesssim - 0.0791$.
\begin{figure}[h!]
\includegraphics[scale=0.45]{Pot_Parc_SAdS_Lamb1_2_Estaveis.pdf}
\caption{The partial emitted power as a function of $M\Omega,$ for $M^2\Lambda = -2^{-1}$ and different choices of $l=m$. The angular velocity range is from $\Omega_{min}$ up to the maximum allowed value for stable orbits.}
\label{fig:partial_power_stable_orbits_Lamb_1_2}
\end{figure}
In Fig.~\ref{Fig_Total_Power}, we show the total emitted power, where the characteristic synchrotronic behavior can be readily observed for unstable orbits ($M \Omega > 0.43$).
The summation in $l$ was truncated at a given $l=l_{max}$.
We can see that, as we include the contribution from higher multipoles to the emitted power (by changing $l_{max}$), the total emitted power increases. For unstable orbits, a substantial portion of the multipoles higher than the fundamental dipole mode have non negligible contributions to the emitted power. We note, however, that for any finite value $R < 3M$ of the source's radial position, the multipole contributions to the total emitted power are negligible past a certain $R$-dependent value of $l$. This also happens in asymptotically flat spacetimes.
In fact, our numerical results suggest that sufficiently high multipole modes do not contribute significantly to the total emitted power, as can be seen in Figs.~\ref{Pot_Parc_SAdS_Lamb1} and~\ref{fig:partial_power_stable_orbits_Lamb_1_2}, namely: There is an initial enhancement in the contribution of increasing multipole numbers $l,$ but there is a maximum $l$ beyond which the contribution to the emitted power starts to decrease. The peak in the total emitted power, plotted in Fig.~\ref{Fig_Total_Power}, approaches $R = 3M$ as we sum higher multipole contributions, analogously to what happens in asymptotically flat spacetimes. This occurs because closer to the black hole photon sphere, the scalar source emits radiation of the synchrotron type, i.e, the contribution of higher multipole modes becomes dominant as $R \rightarrow 3M$ ($\Omega \rightarrow \Omega_{max}$).
In Fig.~\ref{Fig_Total_Power}, the peaks in the total power occur at $10^2 M \Omega \approx \{43.99,44.27,44.44,44.56\}$ for $l_{max} = \{5, 10, 15, 20 \},$ respectively.
\begin{figure}[h!]
\includegraphics[scale=0.45]{Pot_Tot_SAdS_Lamb1_2.pdf}
\caption{The total emitted power as a function of $M\Omega,$ given by Eq.~\eqref{total_power}, for $M^2\Lambda = -2^{-1}$ and different choices of $l_{max},$ as indicated.}
\label{Fig_Total_Power}
\end{figure}
\section{Final Remarks}
\label{Sec_remarks}
We have investigated the synchrotron scalar radiation in the Schwarzschild--Anti--de Sitter (SAdS) spacetime using the semiclassical approach of quantum field theory in curved spacetimes at tree level.
We have presented results for the power of scalar radiation emitted by the source as a function of its angular velocity $\Omega$ for some values of the cosmological constant $\Lambda$.
In the region of the parameter space analyzed in the present investigation, we have obtained good agreement between our semiclassical results and the classical ones previously reported in the literature~\cite{cardoso_2002}.
Different values of $\abs{\Lambda}$ can be associated to different black hole sizes, when compared to the anti-de Sitter radius. Given that the SAdS spacetime is not globally hyperbolic, we must impose suitable boundary conditions at spatial infinity. We considered reflective boundary conditions such that there is no net radial flux at the AdS boundary. This gives rise to two linearly independent sets of modes, which are normalizable and can be used to construct a quantum theory for the massless scalar field. The emitted power can be obtained by computing the one-particle emission amplitude when the quantum scalar field is excited by the source in a geodesic circular orbit.
Our results show that for $M^2 \Lambda \leq -1/2,$ there is an enhancement in the contribution of the high multipole modes to the emitted power. The radiation exhibits a synchrotron-like behavior for orbits much farther from the black hole, when compared to the corresponding case in Schwarzschild spacetime. For sufficiently large $\abs{\Lambda}$, this behavior happens even for stable circular geodesics.
Results in the Schwarzschild--de Sitter (SdS) spacetime case show the opposite behavior. There is an enhancement in the contribution of the lower multipole modes as $M^2 \Lambda$ increases~\cite{brito_2020}.
\begin{acknowledgments}
The authors thank Funda\c{c}\~ao Amaz\^onia de Amparo a Estudos e Pesquisas (FAPESPA), Conselho Nacional de Desenvolvimento Cient\'ifico e Tecnol\'ogico (CNPq) and Coordena\c{c}\~ao de Aperfei\c{c}oamento de Pessoal de N\'{\i}vel Superior (Capes) - Finance Code 001, in Brazil, for partial financial support.
This work has been further supported by the European Union's Horizon 2020 research and innovation (RISE) programme H2020-MSCA-RISE-2017 Grant No.~FunFiCO-777740.
\end{acknowledgments}
|
\section{Introduction}
It has been stated that computation by means of a dynamical system requires that the system have memory and be able to transmit signals \cite{langton1990}. Memory is necessary for computation, but the amount of memory depends on the particular computation. One way to arrange a dynamical system to do computation is as a reservoir computer, also known as an echo state machine or a liquid state machine \cite{jaeger2001, natschlaeger2002}. In a reservoir computer, a set of nonlinear nodes is connected into a network. The nodes are driven by a common input signal, and the time series responses of all of the nodes are recorded. In some ways the reservoir computer resembles a recurrent neural network, but unlike a neural network, the connections between nodes in a reservoir computer are not changed. Rather, training proceeds by fitting the node outputs to a training signal. The fit coefficients are the output of the training procedure. The fact that the connections between nodes are never changed means that reservoir computers may be constructed from analog hardware where altering the connections between nodes may be difficult. Keeping a fixed network also removes constraints on the node activation function imposed by training requirements, so many different nonlinear devices may be used as reservoir computer nodes.
The memory in a reservoir computer must be fading memory. Boyd and Chua \cite{boyd1985} define fading memory in a way that resembles the definition of the largest Lyapunov exponent for a dynamical system.
Examples of reservoir computers so far include photonic systems \cite{appeltant2011,larger2012, van_der_sande2017}, analog circuits \cite{schurmann2004}, mechanical systems \cite{dion2018} and field programmable gate arrays \cite{canaday2018}. Many other examples are included in the review paper \cite{tanaka2019}, which describes hardware implementations of reservoir computers that are very fast, and yet consume little power, while being small and light. Reservoir computers have been shown to be useful for solving a number of problems, including reconstruction and prediction of chaotic attractors \cite{lu2018,zimmerman2018,antonik2018,lu2017,jaeger2004}, recognizing speech, handwriting or other images \cite{jalavand2018} or controlling robotic systems \cite{lukosevicius2012} . Reservoir computers have also been used to better understand the function of neurons in the brain \cite{stoop2013}. Several groups have been using theory to better understand reservoir computers; in \cite{hart2020}, the authors show that there is a positive probability that a reservoir computer can be an embedding of the driving system, and therefore can predict the future of the driving system within an arbitrary tolerance, while Grigoryeva et al. \cite{grigoryeva2021} show conditions under which a reservoir computer can be in strong generalized synchronization with the driving system. Lymburn et al. \cite{lymburn2019} study the relation between generalized synchronization and reconstruction accuracy, while Herteux and R{\"a}th examine how the symmetry of the activation function affects reservoir computer performance \cite{herteux}.
Because the network in a reservoir computer is not trained, there are a number of parameters that must be chosen ahead of time to optimize the performance of the reservoir computer. Motivated by work such as that of Langton \cite{langton1990}, some researchers have studied ways to adjust the memory of a reservoir computer to the problem being solved. It has been shown that in some situations a reservoir computer can have too much memory \cite{carroll2020b}. In a different example that demonstrated that only a short memory may be necessary, Gauthier et al. \cite{gauthier2021} were able to predict the signals from a Lorenz chaotic system with a series approximation that had a memory of only one time step. Luko{\v s}evi{\v c}ius \cite{lukosevicius2012a} points out that adjusting the memory capacity of the reservoir computer for the particular problem is necessary. In \cite{verstraeten2010} the authors show how bias, input scaling and spectral radius affect the memory capacity of a reservoir computer using tanh nodes.
{\subsection{Is Memory Capacity a Useful Statistic?}
In addressing memory, it is necessary to define what is meant by "memory". The Boyd and Chua work \cite{boyd1985} gives a definition, but this definition is not the one most commonly used in discussing memory capacity. The most widely used definition, which measures the ability to predict previous values of a white noise signal, was introduced in \cite{jaeger2002}. This definition does not depend on the actual signal driving the reservoir computer, but in a nonlinear system the dynamics depends both on the system and the driving signal, so two other memory measures are introduced here; one which like the Jaeger memory capacity depends on the change in correlation over time within a reservoir computer but using the actual driving signal, and another definition which resembles the Boyd and Chua definition.}
{While the real measure of performance in a reservoir computer is how accurately it can compute or classify a desired signal, some papers on reservoir computing only examine memory capacity without also measuring testing error. Still, in some cases memory by itself can be a useful statistic, so concentrating on optimizing memory can be helpful in improving the performance of a reservoir computer. The standard model of memory in computation \cite{langton1990} states that longer memory leads to higher computational capacity, and in many cases this is true. On the other hand, \cite{carroll2020b} showed that a reservoir computer can have too much memory. These dynamical effects can be measured by comparing the Lyapunov exponent spectrum of the reservoir computer to the Lyapunov spectrum of the driving system, but in an analog reservoir computer the equations describing the reservoir may not be known. In addition, there is a range of Lyapunov exponents for the reservoir computer, and they may all be important. Memory capacity can be a way to describe the reservoir computer dynamics using only one number, and two of the three memory statistics described in this work do not require knowledge of the dynamical system.}
\section{Reservoir Computer}
\label{rescomp}
The reservoir computer I use in this work is based on a tanh function:
\begin{equation}
\label{tanhnode}
{\bf{R}}\left( {n + 1} \right) = g\tanh \left( {{\bf{AR}} + \varepsilon s\left( n \right)} \right)
\end{equation}
where ${\bf R}$ is the vector of reservoir variables, ${\bf A}$ is the adjacency matrix and $s(n)$ is the input signal. All the entries in the adjacency matrix were set to values drawn from a Gaussian random distribution with mean 0 and standard deviation 1, and then ${\bf A}$ was renormalized to have a spectral radius of 1. For the work in this paper, ${\bf A}$ had $M=100$ nodes. Some papers using a tanh node add a bias within the tanh function, but in order to reduce the number of parameters, I set the bias to zero. {For small signals the tanh function is approximately linear, so a bias will move small signals into the nonlinear region of the tanh. The effects of the nonlinearity may also be increased by increasing the input constant $\varepsilon$. }
In the training stage, the reservoir computer is driven with the input signal $s(n)$ to produce the reservoir computer output signals $r_i(n)$. The first 1000 points from the $r_i(n)$ time series are discarded and the next 10,000 points are used to fit the training signal $f(n)$. The tanh nonlinearity is asymmetric about zero, so in order to add terms symmetric about zero, a matrix is constructed from the reservoir signals as
\begin{equation}
\label{fitmat}
\Omega = \left[ {\begin{array}{*{20}{c}}
{{r_1}\left( 1 \right)}& \cdots &{{r_M}\left( 1 \right)}&{r_1^2\left( 1 \right)}& \cdots &{r_M^2\left( 1 \right)}\\
{{r_1}\left( 2 \right)}&{}&{{r_M}\left( 2 \right)}&{r_1^2\left( 2 \right)}&{}&{r_M^2\left( 2 \right)}\\
\vdots &{}& \vdots &{}&{}& \vdots \\
{{r_1}\left( N \right)}& \cdots &{{r_M}\left( N \right)}&{r_1^2N}& \cdots &{r_M^2\left( N \right)}
\end{array}} \right]
\end{equation}
where the reservoir has $M$ nodes and the time series have $N$ points. The fit to the training signal is
\begin{equation}
\label{train_fit}
{h(t)} ={\Omega } {{\bf C}}
\end{equation}
where ${\bf C}$ is a vector of training coefficients. The fit is usually done by a ridge regression to avoid overfitting
The training error is
\begin{equation}
\label{train_err}
{\Delta _{RC}} = {{\left\langle {f\left( n \right) - h\left( n \right)} \right\rangle } \mathord{\left/
{\vphantom {{\left\langle {f\left( n \right) - h\left( n \right)} \right\rangle } {\left\langle {f\left( n \right)} \right\rangle }}} \right.
\kern-\nulldelimiterspace} {\left\langle {f\left( n \right)} \right\rangle }}
\end{equation}
where $\left\langle \; \right\rangle $ indicates a standard deviation.
In the testing stage, a new input signal $\tilde s\left( n \right)$ is generated from the same dynamical system that generated $s(n)$, but with different initial conditions. The corresponding test signal is $\tilde f\left( n \right)$. The input signal $\tilde s\left( n \right)$ drives the same reservoir to produce the output signals ${\tilde r_i}\left( n \right)$, which are arranged in a matrix $\tilde \Omega$. The testing error is
\begin{equation}
\label{test_err}
{\Delta _{tx}} = {{\left\langle {\tilde f\left( n \right) - \tilde \Omega {\bf{C}}} \right\rangle } \mathord{\left/
{\vphantom {{\left\langle {\tilde f\left( n \right) - \tilde \Omega {\bf{C}}} \right\rangle } {\left\langle {\tilde f\left( n \right)} \right\rangle }}} \right.
\kern-\nulldelimiterspace} {\left\langle {\tilde f\left( n \right)} \right\rangle }}
\end{equation}
where the coefficient vector ${\bf C}$ was found in the training stage.
\section{Lyapunov Exponents}
The Lyapunov exponents were estimated by the Gram-Schmidt method, as described in \cite{Parker:1989}. To estimate the largest $n_{\lambda}$ Lyapunov exponents for the reservoir computer, an $M \times M$ variational matrix $\Theta$ was created. The matrix $\Theta$ was initially set to the identity. The variational matrix was propagated in time as
\begin{equation}
\label{vareq}
{ \Theta \left(k+1 \right) = {D_r}f\left( {{\bf{r}}\left( k \right)} \right)\Theta \left( k \right)}
\end{equation}
where ${D_r}f\left( {{\bf{r}}\left( k \right)} \right)$ is the Jacobian of eq. (\ref{tanhnode}).
A perturbation $\delta{\bf r}(0)$ of the initial condition ${\bf r}(0)$ evolves as $\delta {\bf{r}}\left( 0 \right) = \Theta \left( k \right)\delta {\bf{r}}\left( 0 \right)$. To find the largest $n_{\lambda}$ Lyapunov exponents for the reservoir computer, $\delta{\bf r}(0)$ is initially set to a random $M \times n_{\lambda}$ matrix and the columns are orthonormalized. To find the Lyapunov exponents for the different directions in the reservoir computer phase space, after each time step $k$ the columns of $\delta{\bf r}(k)$ are orthogonalized using the Gram-Schmidt procedure. The Lyapunov exponents at this time step are obtained from the logarithms of the norms of the orthogonal vectors. The variational equation is then propagated to the next time step, $k+1$. If the norm of the variational matrix becomes too large or too small, so that numerical accuracy is affected, the variational matrix is reset to a random orthonormal matrix.
The Gram-Schmidt method is not as accurate as the QR decomposition method \cite{eckmann1985}, but the full Lyapunov exponent spectrum of the reservoir computer was not required. In this case, only the largest Lyapunov exponent was used, so the Gram-Schmidt method could be used to avoid decomposing very large matrices.
\section{Measures of Memory}
\label{memstat}
\subsection{Memory Capacity}
The standard calculation of memory in reservoir computers is the memory capacity introduced in \cite{jaeger2002}. {This memory capacity is a measure of how well the reservoir computer can predict previous values of the input signal, with the quality of fit being measured by the cross correlation between the input signal $s(n-\tau)$ and the reservoir computer fit to this signal. To avoid confusing correlations induced by the reservoir computer with correlations in the input signal}, the reservoir input signal $s(n)$ is a random noise signal- in this work, the noise is Gaussian with a standard deviation of 1. The memory capacity as a function of delay is calculated as
\begin{equation}
\label{memdel}
{\rm{M}}{{\rm{C}}_\tau } = \frac{{\sum\limits_{n = 1}^N {{{\left( {\left[ {s\left( {n - \tau } \right) - \bar s} \right]\left[ {{h_\tau }\left( n \right) - \overline {{h_\tau }} } \right]} \right)}^2}} }}{{\sum\limits_{n = 1}^N {{{\left[ {s\left( {n - \tau } \right) - \bar s} \right]}^2}\sum\limits_{n = 1}^N {{{\left[ {{h_\tau }\left( n \right) - \overline {{h_\tau }} } \right]}^2}} } }}
\end{equation}
with $N=10 000$ and the overbar indicator indicates the mean. The signal $h_{\tau}(n)$ is the fit of the reservoir signals $r_i(n)$ to the delayed input signal $s(n-\tau)$. The extra squared terms used in the fitting matrix $\Omega$ in eq. (\ref{fitmat}) do not add to the memory capacity, so they are not included in the fit for the purpose of calculating memory capacity. The memory capacity as a function of delay, $MC_{\tau}$, is plotted in figure \ref{memtau}.
\begin{figure}
\centering
\includegraphics[scale=0.8]{memtau.pdf}
\caption{ \label{memtau} Memory capacity $MC_{\tau}$ as defined in eq. (\ref{memdel}) for the tanh reservoir as a function of the delay $\tau$. For this plot, $\varepsilon=1$ and $g=1$.}
\end{figure}
The total memory capacity is
\begin{equation}
\label{memcap}
{\rm{MC}} = \sum\limits_{\tau = 1}^{{\tau _{\max }}} {{\rm{M}}{{\rm{C}}_\tau }}
\end{equation}
where $\tau_{max}$ was set to 100 because at that value ${\rm MC}_{\tau}$ was small.
There are some drawbacks to this definition of memory. Changing the input signal for a nonlinear system will change its properties- even changing the amplitude of the random input signal will change the effect of the nonlinearities in the reservoir computer, so the memory capacity as defined in eq. (\ref{memcap}) may not be a true reflection of the memory of the reservoir computer. Also, the memory capacity calculation requires that one fit a signal, but memory capacity is often used as an independent metric to define how well the reservoir computer will fit signals. Characteristics of the reservoir computer that are independent of the memory but lead it to be better or worse at reproducing signals can affect the memory capacity estimate. Because of these drawbacks, two alternate estimates of memory are used in this paper.
\subsection{Norm of the Variation}
Inubushi and Yoshimura \cite{inubushi2017} use the variational equation for the reservoir computer to prove that the memory for a nonlinear system is less than the memory for a linear system. This statistic also is similar to the Boyd and Chua definition of memory \cite{boyd1985}. Following their example, the variational equation may be used to measure the amount of memory in a reservoir computer. If the initial perturbation is $\delta_0$, the perturbation at $n$ steps later for a nonlinear system is,
\begin{equation}
\label{nljac}
\delta_n = \left[ {\prod\limits_{j = 0}^n {{D_r}f\left( {{\bf{r}}\left( j \right)} \right)} } \right]{\delta _0}
\end{equation}
where $f ({\bf R})$ is the node nonlinearity and $D_ff({\bf R})$ is the derivative of $f$ with respect to the node variables.
To create a statistic based on the variational equation, for a system with $M$ nodes I created a random $M \times M$ matrix for $\delta_0$ and then made the rows orthonormal. I propagated this matrix with the nonlinear variational equation. {I propagate the variation along a known trajectory of the reservoir computer and label the initial variation with $j$ to indicate the particular initial condition. I then calculated the norm of $\delta_n(j)$ after $n$ steps. The nonlinear variational equation depends on the reservoir variables, so I randomly picked 100 different starting points on the reservoir attractor ($j = 1, \ldots, 100$) and averaged the resulting norm. }In order to make the statistic similar to the memory capacity statistic, I chose a maximum delay $\tau_{max}$, which I set equal to 100 to match the delay for the memory capacity calculation. A typical perturbation as a function of number of time steps $n$ for the tanh reservoir computer is plotted in figure \ref{perturb}.
\begin{figure}
\centering
\includegraphics[scale=0.8]{perturb.pdf}
\caption{ \label{perturb} Norm of the perturbation to the tanh reservoir computer as a function of number of time steps $n$. For this plot, $\varepsilon=1$ and $g=1$.}
\end{figure}
I then take the sums
\begin{equation}
\label{varsum}
\begin{array}{l}
{\delta_{{\mathop{\rm sum}} } = \sum\limits^{{\tau _{\max }}}_{n=1} {\left\| {\delta _n(j)} \right\|} }
\end{array}
\end{equation}
where the $\left\| \; \right\|$ operator returns the 2-norm of a matrix. The norm of the variation, $D_{var}$, is the mean of the individual sums
\begin{equation}
\label{varnorm}
{D_{var}=\frac{1}{100} \sum\limits_{j=1}^{100 }\delta_{{\mathop{\rm sum}}} (j)}
\end{equation}
While this statistic, which I call the norm of the variation, is similar to the largest Lyapunov exponent, it does not follow exactly the same pattern. At short delays, the decay of the variation will be dominated by the most negative Lyapunov exponents, while the less negative exponents will govern the decay at later times.
While the norm of the variation does not depend on signal fitting and uses the actual signal driving the reservoir, it can only be found if one has the equations defining the reservoir. The delay capacity method described in the next section can be used when only experimental data is available.
\subsection{Delay Capacity}
The delay capacity statistic is adapted from the consistency capacity developed in \cite{jungling2021}. The consistency capacity method used the auxiliary system approach to detecting generalized synchronization \cite{Abarbanel:1996} to estimate the different computational capacities of a reservoir computer due to different input signals. J{\"u}ngling et al. \cite{jungling2021} apply a whitening transformation to the matrix ${\bf R}$ of signals from the reservoir computer. Typically they create two copies of the reservoir and use two different input signals, $s_1(n)=s(n)+\eta$ or $s_2(t)=s(n)+\eta '$, where $\eta$ and $\eta'$ are different noise signals. They find the covariance matrices for the two reservoirs, add a small regularization constant and then use the covariance matrices to create whitening transforms for the two reservoirs. After whitening, they find the cross covariance between the whitened reservoirs and use the trace of the cross covariance as a measure of capacity.
To estimate the delay capacity, rather than create two copies of a reservoir computer, I compared the reservoir computer signals at two different times. I whiten both of these sets of signals, calculate the cross covariance and find the trace. Again, to be consistent with the definition of memory capacity I sum the capacities for all delays and divide by the number of delays.
The signals from a reservoir with $M$ nodes and $N$ time series points may be arranged in an $M \times N$ matrix ${\bf R}_0$
\begin{equation}
\label{omega}
{{\bf R} _0} = \left[ {\begin{array}{*{20}{c}}
{{r_1}\left( 1 \right) - {{\bar r}_1}}& \cdots &{{r_1}\left( N \right) - {{\bar r}_1}}\\
{{r_2}\left( 1 \right) - {{\bar r}_2}}&{}&{{r_2}\left( N \right) - {{\bar r}_2}}\\
\vdots &{}& \vdots \\
{{r_M}\left( 1 \right) - {{\bar r}_M}}& \cdots &{{r_M}\left( N \right) - {{\bar r}_M}}
\end{array}} \right]
\end{equation}
where $\bar r_i$ is the mean of $r_i$. This matrix is similar to the matrix $\Omega$ in eq. (\ref{fitmat}) except that it does not contain any terms in $r_i^2$ because these do not contribute to the memory.
The covariance matrix is then formed as:
\begin{equation}
\label{cmat}
{\bf C} = \frac{{{{\bf R}_0(t) }{\bf R}_0^T(t) }}{N}+L_{reg}{\bf I} .
\end{equation}
The regularization factor $L_{reg}=10^{-10}$ is added because the covariance matrix can be near singular.
The covariance matrix is then decomposed by a singular value decomposition, ${\bf C} = {\bf{US}}{{\bf{V}}^{\bf{T}}}$ and the reservoir signals are normalized as
\begin{equation}
\label{resnorm}
{\widetilde {\bf{R}}_0} ={\left( {{{\bf{V}}^T}{{\bf{R}}_0}} \right)}\left({\sqrt {\bf{S}} }\right)^{-1}
\end{equation}
As a measure of memory, a second matrix is created, ${{\bf R} _\tau }\left( t \right) = {{\bf R} _0}\left( {t - \tau } \right)$ and normalized in the same manner to produce ${\tilde {\bf R} _\tau }\left( t \right)$. The cross covariance between the normalized versions of the regular and delayed matrices is then found
\begin{equation}
\label{crosscorrmat}
{\bf{C}}\left( \tau \right) = \frac{{{{\tilde {\bf R} }_0}\tilde {\bf R} _\tau ^T}}{N} .
\end{equation}
The delay capacity statistic is calculated as
\begin{equation}
\label{delcap}
{\Theta _d} = \frac{{\sum\limits_{\tau = 0}^{{\tau _{\max }}} {{\rm{Trace}}\left| {{\bf{C}}\left( \tau \right)} \right|} }}{{{\tau _{\max }}}}
\end{equation}
where the $\left| \; \right|$ operator indicates an absolute value. The trace of the cross covariance matrix ${\bf C}(\tau)$ is shown in figure \ref{cortrace}.
\begin{figure}
\centering
\includegraphics[scale=0.8]{cortrace.pdf}
\caption{ \label{cortrace} Trace of the cross covariance matrix ${\bf C}(\tau)$ as a function of the delay $\tau$, for the tanh reservoir computer. In this plot, $\varepsilon=1$ and $g=1$.}
\end{figure}
{
\subsection{Nonlinear Index}
The nonlinearity in a reservoir computer is important in reproducing signals from nonlinear systems, so as an additional statistic I used the nonlinear index described in \cite{dambre2012}. The reservoir was driven with sine waves with 100 different periods ranging from $T_j=10 , j=1$ to $T_j=50, j=100$. The individual node signals were Fourier transformed into ${\mathcal F}_i(f)$. For a fundamental frequency $f_j=2\pi T_j$, the nonlinear index is
\begin{equation}
\Gamma \left( {{f_j}} \right) = \frac{1}{M}\sum\limits_{i = 1}^M {\left( {\frac{{\sum\limits_{f > {f_j}} {\left| {{\mathcal F_i}(f)} \right|} }}{{\left| {{\mathcal F_i}({f_j})} \right|}}} \right)}
\end{equation}
where the $\left| \; \right|$ operator returns the absolute magnitude. The mean nonlinear index is
\begin{equation}
\label{nlindex}
\Gamma = \frac{1}{{100}}\sum\limits_{j = 1}^{100} {\Gamma \left( {{f_j}} \right)} .
\end{equation}
}
\section{Simulations}
\subsection{Fitting the Lorenz System}
In order to see how the reservoir parameters affect the actual performance, the Lorenz chaotic system was used to drive the reservoir.
The Lorenz system is described by \cite{lorenz1963}
\begin{equation}
\label{loreq}
\begin{array}{l}
\frac{{dx}}{{dt}} = {p_1}\left( {y - x} \right)\\
\frac{{dy}}{{dt}} = x\left( {{p_2} - z} \right) - y\\
\frac{{dy}}{{dt}} = xy - {p_3}z
\end{array}
\end{equation}
with $p_1=10$, $p_2=28$ and $p_3=8/3$. The equations were numerically integrated with a time step of 0.02. The input to the reservoir computer was $x$, and the reservoir computer was trained on $z$. The testing error for fitting the Lorenz $z$ signal is plotted in figure \ref{lormemerr}.
{
The three different memory statistics or the Lorenz system are also shown in figure \ref{lormemerr}. }The delay capacity $\Theta_d$ in figure \ref{lormemerr} roughly echos the memory capacity MC , but because the delay capacity depends on the actual input signal, it is not the same as the memory capacity. The delay capacity is a measure of how long slowly the autocorrelation of signals in a reservoir computer drops off with time, while memory capacity is measured by how well the reservoir computer fits a delayed noise signal. The delay capacity is affected by the autocorrelation of the input signal.
The variation of the norm in figure \ref{lormemerr} does more closely resemble the plot of memory capacity. The variation of the norm measures how quickly a perturbation to the reservoir trajectory decays, but unlike the memory capacity, it depends on the signal that drives the reservoir. The decay of a perturbation would seem to be an excellent way to quantify fading memory, but calculating the norm of the variation requires that the reservoir computer equations are known, which may not be true in an experiment.
\begin{figure*}
\centering
\includegraphics[scale=0.8]{lormemerr.jpg}
\caption{ \label{lormemerr} {Testing error and memory statistics for the tanh reservoir driven by the Lorenz $x$ signal and trained on the Lorenz $z$ signal. (A) is the memory capacity MC, (B) is the delay capacity $\Theta_d$ as defined in eq. (\ref{delcap}) , (C) is the variation of the norm $D_{var}$ (eq (\ref{varnorm})) while (D) is the testing error $\Delta_{tx}$.}}
\end{figure*}
The testing error for the Lorenz system is not minimized at the largest values of the memory capacity, shown in figure \ref{lormemerr}. The different memory statistics in figure \ref{lormemerr} are all maximized for large values of the feedback constant $g$ but small values of the input multiplier $\varepsilon$.
The minimum training error comes at larger values of $\varepsilon$ than the memory capacity. A reason for this may be seen in figure \ref{lornonlinly}, which shows the nonlinearity index $\Gamma$ and the largest Lyapunov exponent for the reservoir computer driven by the Lorenz $x$ signal.
\begin{figure}
\centering
\includegraphics[scale=0.8]{lornonlinly.jpg}
\caption{ \label{lornonlinly} {The top plot is the nonlinear index $\Gamma$ for the reservoir computer driven by the Lorenz $x$ signal, while the bottom plot is the largest Lyapunov exponent. } }
\end{figure}
{Figure \ref{lornonlinly} shows that the nonlinear index $\Gamma$ increases as the input multiplier $\varepsilon$ increases. For very small signals the tanh nonlinearity is approximately linear, while fitting the Lorenz $z$ signal requires some nonlinearity, so increasing the nonlinearity in the reservoir computer by increasing the input multiplier decreases the testing error. This is one reason why the smallest testing error in figure \ref{lormemerr} does not come for the largest memory capacity. Some work such as \cite{inubushi2017,verstraeten2010} has suggested that increasing nonlinearity can decrease memory capacity. }
The largest Lyapunov exponent for the Lorenz driven tanh reservoir computer, also in figure \ref{lornonlinly}, is calculated from the variational equation as is the norm of the variation, but the two statistics follow different patterns. For a reservoir computer with $M$ nodes, there are $M$ Lyapunov exponents. The initial decay of a perturbation may be governed by the most negative of these exponents; only after the degrees of freedom governed by the most negative exponents have decayed away does the largest exponent dominate. The norm of the variation drops off faster than what would be expected from the largest Lyapunov exponent because of the more negative exponents, but as expected, the largest value of all the memory statistics comes when the largest reservoir Lyapunov exponent is the least negative.
\section{Increasing Memory}
\label{moremem}
One way to increase the memory capacity of a reservoir computer is to operate near the edge of stability, where the largest Lyapunov exponent for the reservoir computer is just below 0. Working too close to the edge of stability can create problems, however. In \cite{carroll2020b}, it was demonstrated that as the reservoir computer Lyapunov exponents increased, they overlapped with the Lyapunov exponent spectrum of the training system, increasing the fractal dimension of the reservoir signals and causing a larger training error. In a different approach, in \cite{farkas2016} the memory capacity was increased by making the columns in the adjacency matrix orthogonal. In \cite{verstraeten2007} the memory capacity of different types of nodes were compared as the reservoir computer parameters were varied.
\subsection{Sparsity}
Increasing the sparsity of the adjacency matrix; that is, decreasing the number of connections between nodes- could possibly increase the memory of the reservoir computer. Conventional rules for designing a reservoir computer often call for using a sparse adjacency matrix, although no justification is usually given for this rule. Using an adjacency matrix with fewer connections means that the feedback path between nodes will be longer, which may increase memory.
Farka{\v s} et al. \cite{farkas2016} examined adjacency matrices with different amounts of sparsity, but found that changing the sparsity did not change the maximum memory capacity. While figure \ref{memcapsparse} below agrees with this conclusion, I show that changing the sparsity of the adjacency matrix also alters the strength of the interaction between nodes. If the strength of this interaction is compensated for, then the effect of sparsity on the memory capacity can be seen in figure \ref{lorfixlen}.
A series of adjacency matrices were created for the tanh reservoir of eq. (\ref{tanhnode}) by randomly choosing a fraction $\eta_f$ of the connections between nodes and setting them to a number drawn from a Gaussian random distribution with mean 0 and standard deviation 1. For the smallest values of $\eta_f$ it was possible for some nodes to have no connections to any other nodes, so the adjacency matrix was required to have at least one entry in each row and at least one entry in each column.
The path length between nodes should depend on $\eta_f$. The length of the path between two nodes in a network is a well known statistic \cite{albert2002}, and may be found by a breadth-first search of the adjacency matrix \cite{Skiena2008}. The unweighted path length between any nodes i and j is designated as $L_U(i,j)$. For node $i_0$, the following algorithm puts the unweighted path length to all the other nodes into the vector distanceList:
\begin{algorithmic}
\STATE distanceList($1 \ldots M$) $\gets \infty$
\STATE distanceList$(i_0)\gets 0$
\STATE queue $\gets i_0$
\WHILE{queue is not empty}
\STATE queue2 is empty
\FOR{$k=1$ to length of queue}
\STATE $i_k \gets$ queue($k$)
\STATE inList $\gets A_{i_k,1 \ldots M}$
\STATE outList $\gets A_{1 \ldots M, i_k}$
\STATE nodeList $\gets$ inList $\cup$ outList
\FOR {$j=1$ to length of nodeList}
\STATE $i_M \gets$ nodeList($j$)
\IF{distanceList($i_M$) $\ne \infty$}
\STATE add $i_M$ to queue2
\STATE distanceList($i_M$) $\gets$ distanceList($i_k$) $+ 1$
\ENDIF
\ENDFOR
\ENDFOR
\STATE queue $\gets$ queue2
\ENDWHILE
\end{algorithmic}
The mean unweighted path length is
\begin{equation}
\label{unwpath}
\left\langle {{L_U}} \right\rangle = \frac{1}{M}\sum\limits_{i = 1}^M {{\rm{distanceList}}\left( i \right)}
\end{equation}
If there is no path between two nodes, the path length is $\infty $ and it is not used in calculating the mean path length. Distances with values of 0 are also not included in the mean unweighted path length. Figure \ref{unwpathfig} shows the mean unweighted path length $<L_U>$ as a function of $\eta_f$.
\begin{figure}
\centering
\includegraphics[scale=0.8]{unwpathfig.pdf}
\caption{ \label{unwpathfig} Mean unweighted path length $<L_U>$ for adjacency matrices versus the fraction $\eta_f$ of the entries that are nonzero. }
\end{figure}
Figure \ref{memcapsparse} shows the memory capacity for the tanh reservoir as the fraction $\eta_f$ varies. The amplitude parameter $g$ was set to 1.0.
\begin{figure}
\centering
\includegraphics[scale=0.8]{memcapsparse.jpg}
\caption{ \label{memcapsparse} The memory capacity for the tanh reservoir of eq. (\ref{tanhnode}) as the fraction of nonzero entries in the adjacency matrix. $\eta_f$, varies. }
\end{figure}
The memory capacity in figure \ref{memcapsparse} is not correlated with the unweighted path length in figure \ref{unwpathfig}. As the number of connections between nodes changes, the strength of the interaction between nodes also changes, so just changing the sparsity of the adjacency matrix does not change the memory capacity.
In \cite{carroll2020a} a weighted path length statistic was derived that was a measure of the strength of the interaction between nodes. The weighted path length statistic $L_W(i,j)$ was calculated in a similar fashion to the unweighted path length statistic, except that the distance between two adjacent nodes is not 1 but
\begin{equation}
\label{dist2}
{\delta _{ij}} = \ln \left[ {\frac{{1}}{{\left| {{A_{ij}}} \right| + \left| {{A_{ji}}} \right|}}} \right].
\end{equation}
The distance was defined in this way so that the weighted distance between nodes that were more strongly coupled was smaller. Because of the natural log in eq. (\ref{dist2}) the weighted path length can have negative values, so it is not a true distance. The log factor was included because the weighted path length can vary over several orders of magnitude.
The interaction between nodes may be held constant by varying the spectral radius so that the mean weighted path length $<L_W>$ is constant. The resulting spectral radius $\rho$ is shown in figure \ref{spectrad}. Figure \ref{lorfixlen} shows the memory capacity when the spectral radius for the adjacency matrix was varied to fix the mean weighted path length $<L_W>$ at 2.0. This value was chosen because it was the mean weighted path length when $\eta_f=0.5$.
\begin{figure}
\centering
\includegraphics[scale=0.8]{spectrad.pdf}
\caption{ \label{spectrad} The spectral radius for the tanh reservoir of eq. (\ref{tanhnode}) as the fraction of nonzero entries in the adjacency matrix. $\eta_f$, varies. The spectral radius was varied to keep the mean weighted path length $<L_W>$ fixed at 2.0. }
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=0.8]{lorfixlen.jpg}
\caption{ \label{lorfixlen} {(A) The memory capacity for the tanh reservoir computer of eq. (\ref{tanhnode}), driven by the Lorenz $x$ signal and trained on the Lorenz $z$ signal, as the fraction of nonzero entries in the adjacency matrix. $\eta_f$, varies. (B) is the delay capacity, (C) is the norm of the variation, while (D) is the testing error. The mean weighted path length $<L_W>$ for these plots was fixed at 2.0. } }
\end{figure*}
When the mean weighted path length $<L_W>$ was fixed, the memory capacity for the tanh reservoir showed a variation with the node fraction $\eta_f$, but the relationship was not monotonic. Figure \ref{lorfixlen} shows that for small values of the input multiplier $\varepsilon$ the memory capacity reaches a maximum for $\eta_f$ near 0.3. The maximum is less prominent as $\varepsilon$ increases. The other measures of memory plotted in figure \ref{lorfixlen} also had maxima near $\eta_f=0.3$. In \cite{farkas2016} a similar pattern was seen; the memory capacity went through a maximum as the spectral radius was varied. The testing error, by contrast, was a maximum at the largest values of the input constant $\varepsilon$. It was seen in figure \ref{lornonlinly} that larger values of nonlinearity led to lower testing error for the Lorenz system.
\subsection{R{\" o}ssler System}
The effect of the reservoir computer memory depends on the problem being solved; as a demonstration, the same reservoir computer was driven with the $x$ signal from the Rossler chaotic system and trained on the $z$ signal. The R{\"o}ssler system is described by \cite{rossler1976}
\begin{equation}
\label{rosseq}
\begin{array}{*{20}{l}}
{\frac{{dx}}{{dt}} = - y - {p_1}z}\\
{\frac{{dy}}{{dt}} = x + {p_2}y}\\
{\frac{{dz}}{{dt}} = {p_3} + z\left( {x - {p_4}} \right)}
\end{array}
\end{equation}
These equations were numerically integrated with a time step $t_s$=0.3, and parameters $p_1=1$, $p_2=0.2$, $p_3=0.2$, $p_4=5.7$.
\begin{figure*}
\centering
\includegraphics[scale=0.8]{rossfixlen.jpg}
\caption{ \label{rossfixlen} {(A) The memory capacity for the tanh reservoir computer of eq. (\ref{tanhnode}), driven by the R{\"o}ssler $x$ signal and trained on the R{\"o}ssler $z$ signal, as the fraction of nonzero entries in the adjacency matrix. $\eta_f$, varies. (B) is the delay capacity, (C) is the norm of the variation, while (D) is the testing error. The mean weighted path length $<L_W>$ for these plots was fixed at 2.0. } }
\end{figure*}
The pattern of memory as a function of $\varepsilon$ and $\eta_f$ when the reservoir computer is driven with a R{\"o}ssler $x$ signal is similar to when it was driven with a Lorenz $x$ signal, but the pattern of testing error is very different. For the R{\"o}ssler system the minimum training error does come when the memory is maximized. A possible reason the patterns of testing error in the Lorenz and R{\" o}ssler systems are so different is because of their respective autocorrelation functions. The top plot in figure \ref{xcorr} shows the autocorrelation for the Lorenz $x$ signal, while the bottom plot shows the autocorrelation for the R{\"o}ssler $x$ signal.
\begin{figure}
\centering
\includegraphics[scale=0.8]{xcorr.pdf}
\caption{ \label{xcorr} Top plot is the autocorrelation of the Lorenz $x$ signal, $C_{xx}$, while the bottom plot is the autocorrelation of the R{\"o}ssler $x$ signal. The horizontal axis is labeled in the number of time series points $n$. For the top plot, each time step was 0.02, while in the bottom plot is was 0.3. }
\end{figure}
Figure \ref{xcorr} shows that the Lorenz autocorrelation drops quickly to near 0, while for the R{\"o}ssler the autocorrelation is large for large times. The R{\"o}ssler signal is almost periodic, so the testing error will be small when the reservoir signals maintain a larger correlation over time. For the Lorenz signal, with its autocorrelation that quickly decreases, a reservoir computer with a long memory will mix together parts of the signal that are uncorrelated in time, resulting in larger errors. In \cite{carroll2020b} a similar effect was attributed to an overlap in the Lorenz exponent spectrum; these are two different ways to explain the same thing. For the R{\"o}ssler system, whose autocorrelation persists over long times, more memory leads to lower error.
The different measures of memory in figures \ref{lorfixlen} and \ref{rossfixlen} show that the memory capacity goes through a maximum when the fraction of occupied nodes is about 0.3. The next section provides a possible cause of this maximum.
\subsection{Approximate Delay Coefficients}
\label{approxcoeff}
The reason for this maximum in the memory capacity may be shown with a linear model of a reservoir computer. Consider a simple linear reservoir computer:
\begin{equation}
\label{linres}
{\bf{R}}\left( {n + 1} \right) = \rho {\bf{AR}}\left( n \right) + {\bf{W}}s\left( n \right)
\end{equation}
where ${\bf R}$ is the vector of reservoir variables, ${\bf A}$ is the adjacency matrix, $\rho$ is used to rescale the spectral radius of ${\bf A}$, ${\bf W}$ is the input vector and the input signal is $s(n)$. The input vector ${\bf W}$ was set to all ones. The adjacency matrix in eq. (\ref{linres}) was rescaled so that the absolute magnitude of its largest eigenvalue was 1.0, so the value of $\rho$ will set the desired spectral radius.
If the reservoir computer output is truncated after five iterations with ${\bf R}(0)=0$, it looks like
\begin{equation}
\label{liniter}
\begin{array}{*{20}{l}}
{{\bf{R}}\left( {n + 1} \right) = {\bf{W}}s\left( n \right) + {{\bf{b}}_1}s(n - 1) + {{\bf{b}}_2}s\left( {n - 2} \right) + {{\bf{b}}_3}s\left( {n - 3} \right) + {{\bf{b}}_4}s\left( {n - 4} \right)}\\
{{{\bf{b}}_1} = {\rho ^{}}{\bf{AW}}}\\
{{{\bf{b}}_j} = {\rho ^{j - 1}}{{\bf{A}}^{j - 1}}{{\bf{b}}_1}\quad j > 1}
\end{array}
\end{equation}
The delay coefficients from eq. (\ref{liniter}) were calculated by varying the fraction of occupied nodes $\eta_f$ and substituting the spectral radius from figure \ref{spectrad} for $\rho$. The coefficients for the delayed input signals $s(n-1)$ to $s(n-4)$, averaged over all nodes, are shown in figure \ref{delcoeff}.
\begin{figure}
\centering
\includegraphics[scale=0.8]{delcoeff.pdf}
\caption{ \label{delcoeff} Values of the delay coefficients ${\bf b}_j$ from eq. (\ref{liniter}), averaged over all nodes. }
\end{figure}
The delay coefficients all peak for $\eta_f$ near 0.3. The coefficient on the undelayed signal $s(n)$ is 1.0, so the memory capacity is partly determined by the ratio of the delayed signal to the undelayed signal. The approximation of eqs. (\ref{linres}-\ref{liniter}) does not exactly match the behavior of the tanh reservoir because the nonlinearity will also affect the memory capacity, but the linear approximation can help explain why the memory capacity goes through a maximum.
\section{Multidimensional Nodes}
\label{mdtanh}
Changing the sparsity of the adjacency matrix can change the memory capacity of a reservoir computer, but only to a limited extent. In a nonlinear system, changing the sparsity causes changes in other quantities, such as the mean weighted path length between nodes. A different way to alter the memory capacity is to add extra dynamical dimensions to the nodes.
Multidimensional nodes may be created by adding delayed coordinates to the tanh nodes from eq. (\ref{tanhnode}):
\begin{equation}
\label{ndtanh}
\begin{array}{l}
{r_{i,1}}\left( {n + 1} \right) = g\tanh \left( {\sum\limits_{j = 1}^M {{A_{i,j}}{r_{i,1}}\left( n \right)} + \varepsilon s\left( n \right) + 0.5{r_{i,{d_e}}}} \right)\\
{r_{i,j}}(n + 1) = {r_{i,j - 1}}(n)\quad j = 2 \ldots {d_e}
\end{array}
\end{equation}
where $d_e$ is the dimension of the node activation function and $r_{i,j}(n)$ is the $j$th component of the $i$'th node. One could create similar nodes for an ordinary differential equation system by using a delayed signal.
Figure \ref{tanhvardmem} shows the three different measures of memory used in this work for the multidimensional tanh nodes of eq. (\ref{ndtanh}) as the node dimension $d_e$ was scanned and the parameter $g$ was fixed at 0.35, $\varepsilon=0.5$ and the spectral radius $\rho=1$. {These parameter values were set by scanning through a number of parameter values and node dimensions and choosing the values of $g$ and $\varepsilon$ that gave a minimum testing error for a range of $d_e$ between 5 and 10. }The input signal for all three statistics was a Lorenz $x$ signal. To make it possible to plot all three statistics on one axis, all statistics were normalized by their maximum values. The memory variations were similar when the input signal was the R{\" o}ssler $x$ signal.
\begin{figure}
\centering
\includegraphics[scale=0.8]{tanhvardmem.pdf}
\caption{ \label{tanhvardmem} {The memory capacity MC, the delay capacity $\theta_d$ and the norm of the variation $D_{var}$ for the reservoir computer with multidimensional nodes. All statistics were normalized by their maximum values so the could be plotted on one axis. The input signal was the Lorenz $x$ signal.}}
\end{figure}
All three memory statistics in figure \ref{tanhvardmem} are consistent, and all three show that increasing the node dimension $d_e$ increased the memory for the reservoir computer.
\subsection{Lorenz and R{\"o}ssler systems}
\label{ndnode}
Figure \ref{lorerrvard} shows the testing error for the Lorenz system as the dimension $d_e$ of the reservoir computer was increased. There are two curves in figure \ref{lorerrvard}: the full reservoir computer produced $d_e \times M$ signals, so one curve shows the testing error found by using the full set of signals. Using more signals in testing the reservoir computer can by itself decrease the error, so to remove this effect, for the other curve only the signals corresponding to the first component from each node, $r_{i,1}(n)$, were used in the fit.
\begin{figure}
\centering
\includegraphics[scale=0.8]{lorerrvard.pdf}
\caption{ \label{lorerrvard} Testing error $\Delta_{tx}$ when the multidimensional reservoir computer is driven by the Lorenz $x$ signal and trained on the Lorenz $z$ signal. The dimension of the nodes in the reservoir computer was $d_e$. For the curve labeled $d_e$ signals, only the first component from each node, or $r_{i,1}(n)$, was used in the fit, while for the curve labeled $d_e \times M$ signals, all $d_e$ components from each node were used.}
\end{figure}
When only the first component from each node is used, the minimum testing error for the Lorenz system came for $d_e=6$, while when all components were used minimum was at $d_e=7$. The vertical axis on figure \ref{tanhvardmem} confirms that the different measures of memory all increase as the reservoir computer dimension $d_e$ increases, so there is an optimal amount of memory for fitting the Lorenz $x$ signal- more or less memory produces a larger training error.
Figure \ref{rosserrvard} shows the testing error for the multidimensional reservoir computer when driven with the R{\"o}ssler system, as a function of the reservoir computer dimension, for both types of fits: when all $d_e \times M$ signals are used and when only the $M$ signals from $r_{i,1}$ are used.
\begin{figure}
\centering
\includegraphics[scale=0.8]{rosserrvard.pdf}
\caption{ \label{rosserrvard} Testing error $\Delta_{tx}$ when the multidimensional reservoir computer is driven by the R{\"o}ssler $x$ signal and trained on the R{\"o}ssler $z$ signal. The dimension of the nodes in the reservoir computer was $d_e$. For the curve labeled $d_e$ signals, only the first component from each node, or $r_{i,1}(n)$, was used in the fit, while for the curve labeled $d_e \times M$ signals, all $d_e$ components from each node were used.}
\end{figure}
The testing error for the R{\"o}ssler system does not go through a minimum, as did the testing error for the Lorenz system. The testing error for the R{\"o}ssler system when all $d_e \times M$ signals are used barely decreases as the reservoir computer dimension increases, while when only the first component for each node is used, there is scatter within the data but no decreasing trend. Increasing memory is not harmful in the R{\"o}ssler system, but there appears to be no significant improvement in the testing error from increasing memory.
The autocorrelations for the Lorenz and R{\"o}ssler systems were shown in figure \ref{xcorr}. The autocorrelation for the Lorenz system drops off quickly to near zero. If the reservoir computer memory is too long when attempting to fit the Lorenz signal, the reservoir computer will mix together parts of the Lorenz signal that are uncorrelated in time, degrading the testing error; as a result, there is an optimum memory capacity for fitting the Lorenz system. In contrast, the R{\"o}ssler system is nearly periodic, so its autocorrelation drops to zero only over a very long time.
\subsection{NARMA System}
\label{narmadat}
The nonlinear autoregressive moving average (NARMA) system was developed as a way to test the ability of neuromorphic systems to reproduce a signal with memory \cite{atiya2000}. As used here the NARMA system is described by
\begin{equation}
\label{narmavar}
y\left( {n + 1} \right) = 0.3y\left( n \right) + 0.05y\left( n \right)\sum\limits_{j = 1}^{{N_N}} {y\left( {n - j} \right)} + 1.5u\left( {n - {N_N} + 1} \right)u\left( n \right) + 0.1
\end{equation}
where the order of the model is $N_N$. The input signal $u(n)$ is drawn from a uniform random distribution between 0 and 0.5.
The dependance of the training error on both memory capacity and the memory required to reproduce a particular signal may be investigated by using NARMA systems of varying orders. Equation \ref{narmavar} described a NARMA system of order $N_N$; in this section, $N_N$ varies from 1 to 10.
Figure \ref{vartanhvarnarma} shows the testing error when both the node dimension $d_e$ and the NARMA order $N_N$ are varied. For this plot, $g=0.35$ and $\varepsilon=0.35$ were chosen by varying both of these parameters for a NARMA system with $N_N=10$ and choosing the values that gave the lowest training error.
\begin{figure}
\centering
\includegraphics[scale=0.8]{vartanhvarnarma.jpg}
\caption{ \label{vartanhvarnarma} Testing error for the multidimensional tanh nodes of eq. (\ref{ndtanh}) for the NARMA system of variable order $N_N$ as the node dimension $d_e$ varied.}
\end{figure}
For NARMA systems with $N_N$ from 1 to 4, the lowest training error occurred for $d_e=1$. For higher order NARMA systems, the lowest training error was seen when $N_d=d_e-1$, demonstrating that the lowest training error came when the memory capacity for the reservoir matched the memory required by the problem being solved.
\section{Summary}
Memory is necessary for a reservoir computer, but there is an optimal amount of memory in a reservoir computer; too much or too little memory can lead to increased training errors. I have used Jaeger's standard definition of memory and I have introduced two alternate methods to measure memory; one that was suggested by \cite{inubushi2017} was based on using the variational equation for the reservoir computer to determine how the size of a perturbation changed with time, while the other, which was based on the ideas of capacity described in \cite{lymburn2020, jungling2021}, measured how the reservoir computers lost correlation over time.
It is necessary to tune the memory of a reservoir computer to get optimal results. This paper showed two methods; changing the sparsity of the adjacency matrix while maintaining the interaction strength between nodes, or constructing a multidimensional reservoir computer specifically to have more or less memory. Different measures of memory may be more or less useful for optimizing the reservoir computer for different problems. When building reservoir computers from analog hardware this sort of flexibility in choosing parameters may not always be available, so more work on designing reservoir computers to have a specified amount of memory is necessary. It would be useful in optimizing the reservoir computer to know beforehand how much memory was optimal for the particular problem. It is probable that systems whose autocorrelation drops quickly in time require less memory capacity than systems whose autocorrelation stays large for long times, but currently I know of no good way to quantify the optimal amount of memory short of simulating the full reservoir computer.
\section{Data Availability}
All necessary data is included in this paper.
\section{Acknowledgements}
I would like to thank Thomas Lymburn and Thomas J{\" u}ngling for useful discussions and feedback.
This work was supported by the Naval Research Laboratory's Basic Research Program.
|
\section{Introduction}
Let $X_{1:n} \leq X_{2:n} \leq \dots \leq X_{n:n}$ denote the order statistics arising from random variables $X_1, X_2,\dots, X_n$.
The kth order statistic $X_{k:n}$ is well known for being the lifetime of a $(n-k+1)$-out-of-$n$ system, a prominent structure of redundancy in fault-tolerant systems in reliability theory that has been extensively explored in the literature; see, for example, \cite{2006Some} and \cite{2013Comparisons}.
A parallel system's lifetime corresponds to the largest order statistics $X_{n:n}$, while a series system's lifetime corresponds to the smallest order statistic $X_{1:n}$. Order statistics are thus equivalent to the study of $k$-out-of-$n$ systems, which is a suitable research area in reliability.
Many fields, including reliability theory, goodness-of-fit tests, auction theory, insurance, actuarial science, operations research, and life testing, use order statistics and statistics closely linked to order statistics.
One may refer to \cite{2004Order} , and \cite{199810,1998Order} for comprehensive discussions.
The method of stochastic ordering is useful for measuring the magnitude and variability of random variables. It has been frequently utilized in reliability theory to compare the lifetimes of different systems, in economics to conduct capital allocation and choose the optimal portfolios, and in actuarial science to optimize risk measures or expected utility functions.
We refer \cite{2010Some}'s excellent monograph for more details on stochastic ordering. \cite{1971Comparisons} were among the first to examine stochastic order statistics comparisons deriving from heterogeneous exponential variables. \cite{1976Stochastic}, \cite{1996Some},\cite{1997Stochastic}, \cite{Khaledi2000SOME}, \cite{2006COMPARISON}, \cite{2007Stochastic,2011On}, \cite{2013ORDERING}, \cite{2014On}, and \cite{2016Comparisons} are just a few of the many researchers who have worked on this topic since then.
The second-order statistic, which characterizes the lifetime of the $(n-1)$-out-of-$n$ system
(referred to as the fail-safe system in reliability theory, see Barlow and Proschan, [18]) and yields the
winner’s price for the bid in the second-price reverse auction (see, for example, \cite{2004Mean,2005A}),
has attracted the attention of many researchers for studying its ordering properties.
A fail-safe system is a significant $k$-out-of-$n$ system, this fault tolerance system is widely applied to the various critical safety system.
In the literature, there has been an abundant study on stochastic properties of fail-safe systems. For instance,
\cite{Paltanea2008} established the comparison of fail-safe systems with respect to hazard rate order.
\cite{Zhao2009} obtained the comparison results for two fail-safe systems consisting of heterogeneous and homogeneous components in terms of the likelihood ratio order.
\cite{Zhao2010} provided the dispersive order of fail-safe systems with heterogeneous exponential components.
\cite{Zhao2011} given the right spread order of the fail-safe systems from heterogeneous exponential components.
\cite{Balakrishnan2015} improved the ordering results for fail-safe systems with exponential components.
\cite{cai2017hazard} compared the hazard rate functions of two fail-safe systems consisting of independent and multiple-outlier proportional hazard rate components. For more the investigations of order statistics, one can may refer to \cite{torrado2015magnitude,torrado2015stochastic,barmalzan2016likelihood,barmalzan2019ordering,das2021some,das2021orderings}.
In reliability theory, to model the lifetime data with different hazard shapes, it is desirable to introduce flexible families of distributions, and to this end, there are two methods have been commonly
used to characterize lifetime distribution with considerable flexibility. One method is to adopt the well-known families of distributions, for example, Gamma, Weibull, and Log-normal, which have been
studied quite extensively in the literature, for more discussions on this topic, we refer readers to \cite{1994Continuous,1995Continuous,1993Exponentiated}.
\cite{MARSHALL1997A} developed a new method to introduce one parameter to a base distribution results in a new family of distribution with more flexibility. For example, for a baseline distribution function $F$ with support $\mathbb{R}^+$ and corresponding survival function $\overline{F}$, the new distribution functions can be defined as
\begin{equation}
G(x;\alpha)=\frac{F(x)}{1-\overline{\alpha}\overline{F}(x)},\quad x,\alpha\in \mathbb{R}^+,\quad \overline{\alpha}=1-\alpha,
\end{equation}
\begin{equation}
H(x;\alpha)=\frac{\alpha F(x)}{1-\overline{\alpha}F(x)},\quad x,\alpha\in \mathbb{R}^+,\quad \overline{\alpha}=1-\alpha.
\end{equation}
\cite{MARSHALL1997A} originally proposed the family of distributions in (1) and studied it for the case when $F$ is a Weibull distribution. When $F$ has probability density and hazard rate functions as $f$
and $r_F$, respectively, then the hazard rate function of $G$ is given by
\begin{equation*}
r_F(x;\alpha)=\frac{1}{1-\overline{\alpha}\overline{F}(x)}r_F(x), \quad x,\alpha\in \mathbb{R}^+,\quad \overline{\alpha}=1-\alpha.
\end{equation*}
Therefore, one can observe that if $r_F(x)$ is decreasing (increasing) in x, then for $0<\alpha\leq1(\alpha\geq1)$,
$r_F(x;\alpha)$ is also decreasing in x. Moreover, one can observe that $r_F(x)\leq r_F(x;\alpha)$ for $0<\alpha\leq1$, and $r_F(x;\alpha)\leq r_F(x)$ for $\alpha\geq1$. For this reason, the parameter $\alpha$ in (1) is referred to as a tilt parameter (see \cite{2007Life}). Note that (1) is equivalent to (2) if $\alpha$ in (1) is changed to $\frac{1}{\alpha}$. The proportional hazard rates (PHR) have important applications in reliability and survival analysis. The random variables $X_1,\cdots, X_n$ are said to follow: PHR model if $X_i$ has the survival function $\overline{F}_{X_i}(x)=\overline{F}^{\lambda_i}(x),i=1,\dots,n$ where $\overline{F}$ is the baseline survival function and $(\lambda_1,\dots, \lambda_n)$ is the frailty vector. It is well-known that the Exponential, Weibull, Lomax and Pareto distributions are special cases of the PHR model. Balakrishnan et al. \cite{2017Modified} introduced two new statistical models by adding a parameter to PHR models, which are regarded as the baseline distributions in $G(x;\alpha)$. The new model is referred to as the modified proportional hazard rates (MPHR). It is given by
\begin{equation}
G(x;\alpha,\lambda)=\frac{1-\overline{F}^\lambda(x)}{1-\overline{\alpha}\overline{F^\lambda}(x)},\quad x,\alpha\in \mathbb{R}^+,\quad \overline{\alpha}=1-\alpha,
\end{equation}
where $\lambda$ is the proportional hazard rate parameters. We denote $X\sim MPHR(\alpha,\lambda;\overline{F})$ if $X$ has the distribution
functions $G(x;\alpha,\lambda)$. For the case $\lambda=1$, (3) simply reduce
to (1). For the case $\alpha=1$, (3) simply reduce to the PHR model. According to the Theorem 2.1 of Navarro et al. [14], (3) can
be rewritten the distorted distribution of $h$, where
\begin{equation}
h(u;\alpha,\lambda)=\frac{1-(u)^\lambda}{1-\overline{\alpha}(u)^\lambda},\quad u=\overline{F}(x),\quad x,\alpha\in\mathbb{R}^+,\quad \overline{\alpha}=1-\alpha,
\end{equation}
if $\lambda=1$, (4) just as the distorted distributions of (1). For some special
models, please refer to multiple-outlier models (\cite{2016Comparisons,2017Comparisons}), extended exponential and extend Weibull
distribution (\cite{MARSHALL1997A}, \cite{0Stochastic}), extended Pareto distribution ( \cite{2006Marshall}) and extended Lomax distribution (\cite{2007Marshall}).
However, to the best of our knowledge, there is no study on the ordering properties of the second-order statistics from the heterogeneous and dependent MPHR model. Our purpose in this paper is to investigate the existence of stochastic orderings between the second-order statistics. The rest of the paper is laid out as follows.
The remaining part of the paper is organized as follows: Section 2 recalls some basic concepts
and notations that will be used in the sequel. Section 3 obtains the usual stochastic order between the second-order statistics under the assumption that the observations are dependent on Archimedean copula. Section 4
deals with the hazard rate order under the assumption that the observations are independent. Section 5 concludes this paper and gives some remarks.
\section{Preliminaries}\label{pre}
In this section, we recall some pertinent definitions and lemmas in the sequel. Throughout, the term ``increasing'' and ``decreasing'' are used in a non-strict sense. $ h^{-1} (u)$ denotes the inverse function of $h(u)$, and $h^{'}(u)$ denotes the derivative function of $h(u)$ when they appear. Let $\mathbb R=(-\infty, +\infty)$, $\mathbb R_+=[0, +\infty)$, $\mathbb R_{++}=(0, +\infty)$, $\mathbb N=\{1,2, \dots, n\}$, $\mathcal D =\{\bm x: x_1\ge x_2 \ge \dots\ge x_n\}$ and $\mathcal I =\{\bm x: x_1\leq x_2 \leq \dots\leq x_n\}$.
Stochastic order is a very useful tool to compare random variables arising from reliability theory, operations research, actuarial science, economics, finance, and so on. Let $X$ and $Y$ be two random variables with distribution functions $F(t)$ and $G(t)$,
survival functions $\overline F(t) =1- F(t)$ and $\overline G(t) =1- G(t)$, probability density functions $f(t)$ and $g(t)$, hazard rate functions $h_X(t) = f(t) /\overline F(t) $ and $h_Y(t) = g(t)/\overline G(t)$, and reversed hazard rate functions $\tilde r_X(t) = f(t) / F(t) $ and $\tilde r_Y(t) = g(t)/ G(t)$, respectively.
\begin{definition}
A random variable $X$ is said to be smaller than $Y$ in the
\begin{enumerate} [\rm (i)]
\item usual stochastic order (denoted by $X \le_{\rm st} Y$) if $\overline F(t) \le \overline G(t)$, for $t \in \mathbb R;$
\item { hazard rate order} (denoted by $X\leq_{\rm hr}Y$) if
$h_{X}(t)\ge h_{Y}(t)$ for all $t\in\mathbb{R}$, or equivalently, if $\overline {G}(t)/\overline {F}(t)$ is increasing in $t\in\mathbb{R};$
\item { reversed hazard rate order} (denoted by $X\leq_{\rm rh}Y$) if
$\tilde r_{X}(t)\le \tilde r_{Y}(t)$ for all $t\in\mathbb{R}$, or equivalently, if $ {G}(t)/ {F}(t)$ is increasing in $t\in\mathbb{R}$.
\end{enumerate}
\end{definition}
It is well known that the (reversed) hazard rate order implies the usual stochastic order, but the reversed statement is not true in general. For more comprehensive discussions on various stochastic orders and their applications, one may refer to the classical monograph \cite{Shaked2007} and \cite{belzunce2015introduction}.
Majorization order is an useful tool in establishing various inequalities arising from many research areas. For any two real-valued vectors $\bm{x}=(x_{1}, x_{2},\ldots, x_{n})$ and $\bm{y}=(y_{1}, y_{2},\ldots, y_{n})$, let $x_{1:n}\leq x_{2:n}\leq \dots\leq x_{n:n}$ and $y_{1:n}\leq y_{2:n}\leq \dots\leq y_{n:n}$ be their increasing arrangements, respectively.
\begin{definition}
A vector $\bm{x} $ is said to
\begin{enumerate}[{\rm (i)}]
\item majorized $\bm{y} $ {\rm (denoted by $\bm{x}\overset{\rm{m}}{\succeq}\bm{y}$)},
if
$\sum_{i=1}^{j}x_{i:n}\leq\sum_{i=1}^{j}y_{i:n}$ for every $j=1,2,\ldots,n-1$, and $\sum_{i=1}^{n}x_{i:n}=\sum_{i=1}^{n}y_{i:n}${\rm;}
\item weakly supermajorized $\bm{y} $ {\rm (denoted by $\bm{x}\stackrel{{\rm w}}{\succeq}\bm{y}$)}, if $\sum^{j}_{i=1}x_{i:n}\leq \sum^{j}_{i=1}y_{i:n}$ for every $j=1,\ldots, n${\rm;}
\item weakly submajorized $\bm{y} $ {\rm (denoted by $\bm{x}\underset{{\rm w}}{\succeq}\bm{y}$)}, if $\sum^{n}_{i=j}x_{i:n}\geq \sum^{n}_{i=j}y_{i:n}$ for every $j=1,\ldots, n${\rm;}
\end{enumerate}
\end{definition}
It is known that
\begin{equation*}
\bm{x}\underset{{\rm w}}{\succeq}\bm{y}\Longleftarrow\bm{x}\overset{\rm{m}}{\succeq}\bm{y}\Longrightarrow\bm{x}\overset{\rm{w}}{\succeq}\bm{y}\Longrightarrow\bm{x}\stackrel{{\rm rm}}{\succeq}\bm{y}
\end{equation*}
for any two non-negative vectors $\bm{x} $ and $ \bm{y}$, while the reverse is not true in general. The notion
of majorization is quite useful in establishing various inequalities. For more detailed discussion on the theory of majorization and its applications, one may refer to \cite{Marshall2011} and \cite{Balakrishnan2013}.
The following lemma is very helpful to establish
inequalities for asymmetrical multivariate functions in terms of weak majorization order.
\begin{lemma}{\rm (\cite{Marshall2011})}\label{wconvex}
Let $\phi$ be a real valued function defined on $\mathcal D_n$ and continuously differential on the interior of $\mathcal D_n$. Denote the partial derivative of $\phi$ with respect to its $k$-th argument by $\phi_{(k)}(\bm z)=\partial \phi(\bm z) /\partial z_k$, $k=1,\dots, n$. Then, $\phi (\bm x) \le \phi (\bm y)$ whenever
$ \bm{x}\overset{{\rm w}}{\preceq}\bm{y} \text{ on } \mathcal D_n$ if and only if $ 0 \ge \phi_{(1)}(\bm z) \ge \phi_{(2)}(\bm z) \ge \cdots \ge \phi_{(n)}(\bm z).$ Similarly, $\phi (\bm x) \le \phi (\bm y)$ whenever
$ \bm{x}{\preceq}_{{\rm w}}\bm{y} \text{ on } \mathcal D_n$ if and only if $ \phi_{(1)}(\bm z) \ge \phi_{(2)}(\bm z) \ge \cdots \ge \phi_{(n)}(\bm z) \ge 0.$
\end{lemma}
\begin{lemma}\label{kundu}
\cite{2016Some}Let $h:\mathcal{I_+}\rightarrow\mathbb{R}$ be a function, continuously differentiable on the interior of $\mathcal{I_+}.$ Then, for ${\bm x}$,${\bm y} \in \mathcal{I_+}$, ${\bm x}\overset{\rm m}{\succeq}{\bm y}$ implies $h({\bm x})\geq(respectively, \leq) h({\bm y})$ if and only if, $h_{(k)}({\bm z})$ is increasing (respectively, decreasing) in $k=1,\dots,n,$ where $h_{(k)}({\bm z})=\partial h_({\bm z})/\partial z_k$ denotes the partial derivative of $h$ with respect to its $k$th argumemt.
\end{lemma}
\begin{lemma}\label{proof}
Let the function $T(x)=\frac{x^2}{(1-p+px)^2}$, where $p\in(0,1]$. Then $T(x)$ is increasing in $x\in [0,1]$.
\end{lemma}
\begin{proof}
Taking the derivative of $T(x)$ with respect to $X\in (0,1]$, we immediately get
\begin{eqnarray*}
T^\prime(x)&\overset{\rm sgn}{=}&2x(1-p+px)[1-p+px-px]\\
&=&2x(1-p+px)[1-p]\geq0.
\end{eqnarray*}
Then,we get the required result.
\end{proof}
Next, let us review the concept of copula. For a random vector $\bm{X}=(X_1,X_2,\ldots,X_n)$ with the joint distribution function $H$ and respective marginal distribution functions $F_1,F_2, \ldots,F_n$, the \emph{copula} of $X_1,X_2,\ldots,X_n$ is a distribution function $C:[0,1]^n\mapsto [0,1]$, satisfying $$H(\bm{x})=\mathbb{P}(X_1\le x_1,X_2\le x_2,\ldots,X_n\le x_n)=C(F_1(x_1),F_2(x_2),\ldots,F_{n}(x_n)).$$ Similarly, a \emph{survival copula} of $X_1,X_2,\ldots,X_n$ is a survival function $\hat{C} :[0,1]^n\mapsto [0,1]$, satisfying
\begin{equation*}
{\overline H}(\bm{x})=\mathbb{P}(X_1>x_1,X_2>x_2,\ldots,X_n>x_n)=\hat{C}( {\overline F}_1(x_1), {\overline F}_2(x_2),\ldots, {\overline F}_n(x_n)),
\end{equation*}
where $ {\overline H}(\bm{x})$ is the joint survival function.
Notably, since Archimedean copula has well the mathematical properties,
which is applied to the various areas. In this paper, we will sometimes employ Archimedean copulas to model the dependence structure among components lifetimes.
\begin{definition}{\rm\citep[Expression 4.6.1 on Page 151 of][]{Nelsen2006}}
For a decreasing and continuous function $\phi:[0,1]\mapsto [0,+\infty]$ such that $\phi(0)=+\infty $ and $\phi(1)=0$, and let $\psi=\phi^{-1}$ be the pseudo-inverse of $\phi$. Then
$$C_\phi (u_1, u_2,...,u_n)=\psi \Big(\sum_{i=1}^n\phi (u_i)\Big), \text{ for all } u_i\in [0,1], \quad i=1,2,\dots,n $$
is said to be an Archimedean copula with generator $\phi$ if $(-1)^k\phi^{(k)}(x)\ge 0$ for $k=0,1, \dots,n-2$ and $(-1)^{n-2}\phi^{(n-2)}(x)$ is decreasing and convex.
\end{definition}
\section{Usual stochastic order of second-order statistics from dependent and heterogeneous observations}\label{st}
In this section, we carry out the usual stochastic order of second order statistics from dependent and heterogeneous MPHR samples. Let $\bm{X}=(X_1,X_2,\dots,X_n)$ and $\bm{Y}=(Y_1,Y_2,\dots,Y_n)$ be two n-dimensional vectors of heterogeneous dependent random observations, with $\bm{X}\thicksim MPHR(\bm{\alpha},\bm{\lambda};\overline{F},\psi)$ and $\bm{Y}\thicksim MPHR(\bm{\beta},\bm{\mu};\overline{F},\psi)$, where $\bm{\alpha}=(\alpha_1,\alpha_2,\dots,\alpha_n),\ \bm{\lambda}=(\lambda_1,\lambda_2,\dots,\lambda_n),\ \bm{\beta}=(\beta_1,\beta_2,\dots,\beta_n),$ and $\bm{\mu}=(\mu_1,\mu_2,\dots,\mu_n)$. Here, $X_i\thicksim MPHR(\alpha_i,\lambda_i;\overline{F},\psi)$ and $Y_i\thicksim MPHR(\beta_i,\mu_i;\overline{F},\psi)$, for $i=1,2,\dots,n$. We recall that $\overline{F}$ is the baseline survival function, $\psi$ is generator of the associated Archimedean survival copula.
Next, we establish sufficient conditions for the usual stochastic order, whenever the modified proportional hazard rate parameters may be different with the tilt parameters being equal.
\begin{theorem}\label{tp}
Let $X_1,X_2\dots,X_n$ be n statistically dependent heterogeneous random variables with $X_i\thicksim MPHR(\alpha,\lambda_i;\overline{F};\psi)$, where $0<\alpha\leq1,$ and let $N_1$ be a non-negative integer-valued random variable independent of $ X_{i}'$s. Let $Y_1,Y_2\dots,Y_n$ be another n statistically dependent heterogeneous random variables with $Y_i\thicksim MPHR(\alpha,\mu_i;\overline{F};\psi)$, and let $N_2$ be another non-negative integer-valued random variable independent of $ Y_{i}'$s. Suppose $\lambda,\mu\in\mathcal D_+$ $( or
\ \mathcal{I_+})$, if $N_1\geq_{\rm st}N_2,$ and $\psi$ is log-concave. Then we have
\begin{equation*}
\lambda\overset{\rm w}{\succeq}\mu\Rightarrow X_{2:N_1}\geq_{\rm st}Y_{2:N_2}
\end{equation*}
\end{theorem}
\begin{proof}
We present the proof when $\lambda,\mu\in\mathcal D_+$. The proof is similar when these vectors belong to $\mathcal I_+$. The survival function of $X_{2:n}$ can be express as
\begin{eqnarray*}
\overline{F}_{X_{2:n}}(x)=\sum\limits_{i=1}^{n}\psi\left( \sum\limits_{j\neq i}^{n}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(n-1)\psi\left(\sum\limits_{i=1}^{n}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)
\end{eqnarray*}
If $N_1\geq_{\rm st}N_2$, then
\begin{eqnarray*}
\overline{F}_{X_{2:N_1}}(x)&=&\sum\limits_{m=1}^{n}P(X_{2:N_1}> x|N_1=m)\mathbb{P}(N_1=m)\\
&=&\sum\limits_{m=1}^{n}P(X_{2:m}> x)\mathbb{P}(N_1=m)\\
&\geq&\sum\limits_{m=1}^{n}P(X_{2:m}> x)\mathbb{P}(N_2=m).
\end{eqnarray*}
To obtain the desired result, it is then sufficient to show that $\overline{F}_{X_{2:m}}(x)\geq\overline{F}_{Y_{2:m}}(x),\ m=1,\dots,n. $ For any $k= {1,2,\dots,m}$, taking the partial derivative of $\overline{F}_{X_{2:m}}(x)$ with respect to $\lambda_k,$ we have
\begin{eqnarray*}
\frac{\partial\overline{F}_{X_{2:m}}(x)}{\partial \lambda_k}
&=& \phi^\prime\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\frac{\alpha {\rm log}{\overline{F}(x)}\overline{F}^{\lambda_k}(x)}{[1-\overline{\alpha}\overline{F}^{\lambda_k}(x)]^2}\\
&&\times\left[ \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)\right] \\
&\leq&0
\end{eqnarray*}
where the inequality is due to the fact that the decreasing and convex properties of $\phi$.
Hence, we have
\begin{eqnarray*}
&&\frac{\partial\overline{F}_{X_{2:m}}(x)}{\partial \lambda_k}-\frac{\partial\overline{F}_{X_{2:m}}(x)}{\partial \lambda_l}\\
&=&\phi^\prime\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\frac{\alpha {\rm log}{\overline{F}(x)}\overline{F}^{\lambda_k}(x)}{[1-\overline{\alpha}\overline{F}^{\lambda_k}(x)]^2}\\
&&\times\left[ \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)\right] \\
&&-\phi^\prime\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\frac{\alpha {\rm log}{\overline{F}(x)}\overline{F}^{\lambda_l}(x)}{[1-\overline{\alpha}\overline{F}^{\lambda_l}(x)]^2}\\
&&\times\left[ \sum\limits_{i\neq l}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)\right] \\
&\overset{\rm sgn}{=}&\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right)}\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\\
&&\times\left[ \sum\limits_{i\neq l}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)\right]\\
&&-\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right)}\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\\
&&\times\left[ \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)\right]\\
&=:&P_1Q_1-U_1V_1=P_1(Q_1-V_1)+(P_1-U_1)V_1.
\end{eqnarray*}
For any $1\leq k<l\leq m$, $\lambda_k\geq \lambda_l$, by the log-concave and decreasing properties of $\psi$, we have
\begin{eqnarray*}
\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right)}\geq\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right)}.
\end{eqnarray*}
It holds that
\begin{eqnarray*}
(P_1-U_1)V_1&=&\left[ \frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right)}\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}-\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right)}\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right] \\
&&\times\left[ \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_i}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_i}(x)}\right) \right)\right]\\
&\overset{\rm sgn}{=}&\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right)}\left(\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}-\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)} \right)\\
&&+\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)} \left(\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_k}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_k}(x)}\right)\right)}-\frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right)} \right)\\
&\geq&0
\end{eqnarray*}
For any $\lambda_k\geq\lambda_l$, owing to $\phi$ is decreasing and convex,we have
\begin{eqnarray*}
\psi^\prime\left( \sum\limits_{j\neq k}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)\leq\psi^\prime\left( \sum\limits_{j\neq l}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)
\end{eqnarray*}
As a result, we have
\begin{eqnarray*}
P_1(Q_1-V_1)
&=& \frac{\psi\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\alpha\overline{F}^{\lambda_l}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\right)\right)}\frac{1}{1-\overline{\alpha}\overline{F}^{\lambda_l}(x)}\\
&&\times\left[ \sum\limits_{i\neq l}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)- \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right) \right]\\
&\overset{\rm sgn}{=}&\sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)-\sum\limits_{i\neq l}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)\\
&=&\psi^\prime\left( \sum\limits_{j\neq l}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)-\psi^\prime\left( \sum\limits_{j\neq k}^{m}\phi\left( \frac{\alpha\overline{F}^{\lambda_j}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_j}(x)}\right) \right)\\
&\geq&0
\end{eqnarray*}
Therefore, $P_1(Q_1-V_1)+(P_1-U_1)V_1\geq0$, and
the desired result obtains by lemma 1.
\end{proof}
\begin{remark}
The result of Theorem \ref{tp} improves Theorem 1 of \cite{zhang2021ordering} from the extreme order statistics to the case of the second-order statistics.
\end{remark}
Next, we give a numerical example to illustrate the result of Theorem \ref{tp}.
\begin{example}
Consider the case of $n=4$. Let $\overline{F}(x)=e^{-(ax)^b}, a>0, b>0$, and generator $\psi(x)=e^{\frac{1-e^x}{\theta}}, 0<\theta\leq1.$ Set $\alpha=0.8, a=1.2, b=0.5, \theta=0.1,\lambda=(0.2,0.4,0.8,1.3)\overset{\rm w}{\succeq}(0.3,0.3,1.5,1.6)=\mu.$ Let $N_1$ be an integer-valued random variable having the probability distribution $\mathbb{P}(N_1=1)=0.05, \mathbb{P}(N_1=2)=0.2, \mathbb{P}(N_1=3)=0.3, \mathbb{P}(N_1=4)=0.45$, and $N_2$ be another integer-valued random variable with the probability distribution $\mathbb{P}(N_2=1)=0.05, \mathbb{P}(N_2=2)=0.2, \mathbb{P}(N_2=3)=0.35, \mathbb{P}(N_2=4)=0.4$. It is easy to check that the conditions of Theorem \ref{tp} are all statisfied. The plots of survival functions of $X_{2:4}$ and $Y_{2:4}$, denoted as $\overline{F}_{X_{2:4}}$ and $\overline{F}_{Y_{2:4}}$, are plotted in figure 1, all $x=-{\rm ln}(u)$ and $u\in (0,1]$. Thus, the effectiveness of Theorem \ref{tp} are vaidated.
\begin{figure}[H]
\centering
{\includegraphics[width=7cm]{mphr_st1_-eps-converted-to.pdf}\label{fig2a}}
\caption{Plots of survival function of $X_{2:4}[Y_{2:4}]$, for all $x=-\ln u$ and $u\in(0,1]$.
}
\end{figure}
\end{example}
The following theorem presents the usual stochastic order between the second-order statistics arising from two sets of random variables with Archimedean copula, here, we assume that the two samples have common modified proportional hazard rates parameters.
\begin{theorem}\label{tla}
Let $X_1,X_2\dots,X_n$ be n statistically dependent heterogeneous random variables with $X_i\thicksim MPHR(\alpha_i,\lambda;\overline{F};\psi)$, where $0<\alpha_i\leq1,$ for all $i=1,2,\dots,n,$ and let $N_1$ be a non-negative integer-valued random variable independent of $ X_{i}'$s. Let $Y_1,Y_2\dots,Y_n$ be another n statistically dependent heterogeneous random variables with $Y_i\thicksim MPHR(\beta_i,\lambda;\overline{F};\psi)$, where $0<\beta_i\leq1,$ for all $i=1,2,\dots,n,$ and let $N_2$ be another non-negative integer-valued random variable independent of $ Y_{i}'$s. Suppose $\alpha,\beta\in\mathcal D_+$ $( or
\ \mathcal{I_+})$, if $N_1\geq_{\rm st}N_2,$ and $\psi$ is log-concave. Then we have
\begin{equation*}
\frac{1}{\alpha}\overset{\rm w}{\succeq}\frac{1}{\beta}\Rightarrow X_{2:N_1}\geq_{\rm st}Y_{2:N_2}
\end{equation*}
\end{theorem}
\begin{proof}
We present the proof for $\alpha,\beta\in \mathcal{I}_+$. The proof for the other case is similar, and hence not presented here. The survival function of $X_{2:n}$ can be expressed as
\begin{equation*}
\overline{F}_{X_{2:n}}(x)=\sum\limits_{i=1}^{n}\psi\left( \sum\limits_{j\neq i}^{n}\phi\left( \frac{\alpha_j\overline{F}^{\lambda}(x)}{1-\overline{\alpha_j}\overline{F}^{\lambda}(x)}\right) \right) -(n-1)\psi\left(\sum\limits_{i=1}^{n}\phi\left( \frac{\alpha_i\overline{F}^{\lambda}(x)}{1-\overline{\alpha_i}\overline{F}^{\lambda}(x)}\right) \right).
\end{equation*}
Similarly, to obtain the desired result, it is then sufficient to show that $\overline{F}_{X_{2:m}}(x)\geq\overline{F}_{Y_{2:m}}(x),\ m=1,\dots,n.$
Let $a_k=\frac{1}{\alpha_k}$, for $k=1,\dots,m$. On differentiating this partially with respect to $a_k,$ we get
\begin{eqnarray*}
&&\frac{\partial\overline{F}_{X_{2:m}}(x)}{\partial a_k}\\
&=&\frac{\psi\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right)}\frac{-\frac{1}{a_k}[1-\overline{F}^{\lambda}(x)]}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\\
&&\times\left[ \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\frac{1}{a_i}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_i})\overline{F}^{\lambda}(x)}\right)\right)\right]\\
&\leq&0.
\end{eqnarray*}
Notice that
\begin{eqnarray*}
&&\frac{\partial\overline{F}_{X_{2:m}}(x)}{\partial a_k}-\frac{\partial\overline{F}_{X_{2:m}}(x)}{\partial a_l}\\
&=&\frac{\psi\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right)}\frac{-\frac{1}{a_k}[1-\overline{F}^{\lambda}(x)]}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\\
&&\times\left[ \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\frac{1}{a_i}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_i})\overline{F}^{\lambda}(x)}\right)\right)\right]\\
&&-\frac{\psi\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right)}\frac{-\frac{1}{a_l}[1-\overline{F}^{\lambda}(x)]}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\\
&&\times\left[ \sum\limits_{i\neq l}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right) -(m-1)\psi^\prime\left(\sum\limits_{i=1}^{m}\phi\left( \frac{\frac{1}{a_i}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_i})\overline{F}^{\lambda}(x)}\right)\right)\right]\\
&=:&P_2Q_2-U_2V_2=P_2(Q_2-V_2)+(P_2-U_2)V_2.
\end{eqnarray*}
For any $1\leq k<l\leq m$, $a_k\geq a_l$, by the log-concave and decreasing properties of $\psi$, we have
\begin{eqnarray*}
\frac{\psi\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right)}\geq \frac{\psi\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right)}.
\end{eqnarray*}
It holds that
\begin{eqnarray*}
&&(P_2-U_2)V_2\\
&\overset{\rm sgn}{=}&\frac{\psi\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right)}\frac{\frac{1}{a_k}}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}-\frac{\psi\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right)}\frac{\frac{1}{a_l}}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\\
&=&\frac{\psi\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right)}\left[ \frac{\frac{1}{a_k}}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}-\frac{\frac{1}{a_l}}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right]\\
&&+\frac{\frac{1}{a_l}}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\left[ \frac{\psi\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_k}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x)}\right)\right)}-\frac{\psi\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right) }{\psi^\prime\left( \phi\left( \frac{\frac{1}{a_l}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x)}\right)\right)}\right]\\
&\geq&0.
\end{eqnarray*}
For any $a_k\geq a_l$, owing to $\phi$ is convex and decreasing, we have
\begin{eqnarray*}
\psi^\prime\left( \sum\limits_{j\neq k}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right)\leq\psi^\prime\left( \sum\limits_{j\neq l}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right).
\end{eqnarray*}
As a result, we have
\begin{eqnarray*}
P_2(Q_2-V_2)
&\overset{\rm sgn}{=}& \sum\limits_{i\neq k}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right) - \sum\limits_{i\neq l}^{m}\psi^\prime\left( \sum\limits_{j\neq i}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right)\\
&=&\psi^\prime\left( \sum\limits_{j\neq l}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right)-\psi^\prime\left( \sum\limits_{j\neq k}^{m}\phi\left( \frac{\frac{1}{a_j}\overline{F}^{\lambda}(x)}{1-(1-\frac{1}{a_j})\overline{F}^{\lambda}(x)}\right) \right)\\
&\geq&0.
\end{eqnarray*}
Therefore, the desires result completed by lemma 1.
\end{proof}
\begin{remark}
The result of Theorem \ref{tla} improves Theorem 2 of \cite{zhang2021ordering} from the extreme order statistics to the case of the second-order statistics.
\end{remark}
The next example illustrates result of Theorem \ref{tla}.
\begin{example}
Consider the case of $n=4$. Let $\overline{F}(x)=e^{-(ax)^b}, a>0, b>0$, and generator $\psi(x)=e^{1-(1+x)^\theta}, \theta>0.$ Taking $\lambda=0.4, a=0,5, b=0.8, \theta=7,\alpha=(\frac{1}{3},\frac{1}{3},\frac{1}{5},\frac{1}{8})\ and\ \beta=(\frac{1}{5},\frac{1}{6},\frac{1}{7},\frac{1}{9}).$ Evidently, $(\frac{1}{\alpha_1},\frac{1}{\alpha_2},\frac{1}{\alpha_3},\frac{1}{\alpha_4})\overset{\rm w}{\succeq}(\frac{1}{\beta_1},\frac{1}{\beta_2},\frac{1}{\beta_3},\frac{1}{\beta_4}).$ Let $N_1$ be an integer-valued random variable having the probability distribution $\mathbb{P}(N_1=1)=0.05, \mathbb{P}(N_1=2)=0.2, \mathbb{P}(N_1=3)=0.3, \mathbb{P}(N_1=4)=0.45$, and $N_2$ be another integer-valued random variable with the probability distribution $\mathbb{P}(N_2=1)=0.05, \mathbb{P}(N_2=2)=0.2, \mathbb{P}(N_2=3)=0.35, \mathbb{P}(N_2=4)=0.4$, it is apparently that $N_1\geq_{\rm st}N_2$. The plots of survival functions of $X_{2:4}$ and $Y_{2:4}$, denoted as $\overline{F}_{X_{2:4}}$ and $\overline{F}_{Y_{2:4}}$, are plotted in figure 2, all $x=-{\rm ln}(u)$ and $u\in (0,1]$. Thus, the effectiveness of Theorem \ref{tp} are vaidated.
\begin{figure}[H]
\centering
{\includegraphics[width=7cm]{mphr_st2_-eps-converted-to.pdf}\label{fig2a}}
\caption{Plots of survival function of $X_{2:4}[Y_{2:4}]$, for all $x=-\ln u$ and $u\in(0,1]$.
}
\end{figure}
\end{example}
\section{Hazard rate order of second-order statistics from independent and heterogeneous observations}\label{hr}
The next theorem gives sufficient conditions guaranteeing the hazard rate order between two
modified proportional hazard rates models with the same modified proportional hazard rates parameters
and the heterogeneous tilt parameters.
\begin{theorem}\label{tpla}
Let $X_1,X_2\dots,X_n$ be n independent heterogeneous random variables with $X_i\thicksim MPHR(\alpha_i,\lambda;\overline{F})$, where $0<\alpha_i\leq1,$ for all $i=1,2,\dots,n.$ Let $Y_1,Y_2\dots,Y_n$ be another n independent heterogeneous random variables with $Y_i\thicksim MPHR(\beta_i,\lambda;\overline{F})$, where $0<\beta_i\leq1,$ for all $i=1,2,\dots,n$. Suppose $\alpha,\beta\in\mathcal D_+$ $( or
\ \mathcal{I_+})$. Then we have
\begin{equation*}
\frac{1}{\alpha}\overset{\rm m}{\succeq}\frac{1}{\beta}\Rightarrow X_{2:n}\geq_{\rm hr}Y_{2:n}.
\end{equation*}
\end{theorem}
\begin{proof}
The idea of the proof is borrowed from Theorem 3.7 of \cite{das2021comparison}.
The survival function of $X_{2:n}$ can be rewritten
\begin{eqnarray*}
\overline{F}_{X_{2:n}}(x)&=&\sum\limits^n_{i=1}\prod\limits^n_{j\neq i}\frac{\alpha_j\overline{F}^{\lambda}(x)}{1-\overline{\alpha_j}\overline{F}^{\lambda}(x)}-(n-1)\prod\limits^n_{i=1}\frac{\alpha_i\overline{F}^{\lambda}(x)}{1-\overline{\alpha_i}\overline{F}^{\lambda}(x)}\\
&=&\prod\limits^n_{i=1}\frac{\alpha_i\overline{F}^{\lambda}(x)}{1-\overline{\alpha_i}\overline{F}^{\lambda}(x)}\left[ \sum\limits^n_{i=1}\frac{1-\overline{\alpha_i}\overline{F}^{\lambda}(x)}{\alpha_i\overline{F}^{\lambda}(x)}-(n-1)\right]\\
&=&\prod\limits^n_{i=1}\frac{\alpha_i\overline{F}^{\lambda}(x)}{1-\overline{\alpha_i}\overline{F}^{\lambda}(x)}\left[ \sum\limits^n_{i=1}\frac{1-\overline{F}^{\lambda}(x)}{\alpha_i\overline{F}^{\lambda}(x)}+1\right].
\end{eqnarray*}
Thus, the hazard rate function of the second-order statistic $X_{2:n}$ can be obtained as
\begin{eqnarray*}
r_{X_{2:n}}(x)=\sum\limits^n_{i=1}\frac{\lambda r(x)}{1-\overline{\alpha}_i\overline{F}^{\lambda}(x)}-\frac{\sum\limits^n_{i=1}\frac{\lambda r(x)}{\alpha_i\overline{F}^{\lambda}(x)}}{\sum\limits^n_{i=1}\frac{1-\overline{F}^\lambda(x)}{\alpha_i\overline{F}^{\lambda}(x)}+1},
\end{eqnarray*}
where $r(x)$ is the hazard rate function of the baseline distribution. Define
\begin{eqnarray}\label{one}
\Phi(a)=r_{X_{2:n}}(x)=\sum\limits^n_{i=1}\frac{\lambda r(x)}{1-(1-\frac{1}{a_i})\overline{F}^{\lambda}(x)}-\frac{\sum\limits^n_{i=1}\frac{a_i\lambda r(x)}{\overline{F}^{\lambda}(x)}}{\sum\limits^n_{i=1}\frac{a_i(1-\overline{F}^\lambda(x))}{\overline{F}^{\lambda}(x)}+1},
\end{eqnarray}
where $a_i=\frac{1}{\alpha_i}$, $i=1,\dots,n.$ After differentiating equation (\ref{one}) partially with respect to $a_k,$ we obtain
\begin{eqnarray}\label{two}
\frac{\partial \Phi(a) }{\partial a_k}=\frac{\frac{1}{a^2_k}\overline{F}^{\lambda}(x)\lambda r(x)}{\left[1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x) \right]^2}-\frac{\frac{\lambda r(x)}{\overline{F}^{\lambda}(x)}}{\sum\limits^n_{i=1}\frac{a_i(1-\overline{F}^\lambda(x))}{\overline{F}^{\lambda}(x)}+1}+\frac{\sum\limits^n_{i=1}\frac{a_i\lambda r(x)}{\overline{F}^{\lambda}(x)}\frac{1-\overline{F}^{\lambda}(x)}{\overline{F}^{\lambda}(x)}}{\left[ \sum\limits^n_{i=1}\frac{a_i(1-\overline{F}^\lambda(x))}{\overline{F}^{\lambda}(x)}+1\right]^2 }.
\end{eqnarray}
To prove the result, it is sufficient to show that $\Phi(a)$ is schur-concave with respect to $a\in\mathcal{I_+}$ $or\mathcal{D_+}$. Without loss of generality, let $a\in\mathcal{I_+}$, for any $1\leq k<l\leq n,$ $a_k\leq a_l$. Using (\ref{two}), we have
\begin{eqnarray*}
\frac{\partial \Phi(a) }{\partial a_k}-\frac{\partial \Phi(a) }{\partial a_l}&\overset{\rm sgn}{=}&\frac{\frac{1}{a^2_k}}{\left[1-(1-\frac{1}{a_k})\overline{F}^{\lambda}(x) \right]^2}-\frac{\frac{1}{a^2_l}}{\left[1-(1-\frac{1}{a_l})\overline{F}^{\lambda}(x) \right]^2}\geq0,
\end{eqnarray*}
where the last inequality is due to lemma \ref{proof}. Thus, the result follows from Lemma \ref{kundu}.
\end{proof}
The following example illustrates the result of Theorem \ref{tpla}.
\begin{remark}
The result of Theorem \ref{tpla} completes Theorem 3 of \cite{zhang2021ordering} from the extreme order statistics to the case of the second-order statistics.
\end{remark}
\begin{example}
Let $\overline{F}(x)=e^{-(ax)^b}, a>0, b>0$, Set $n=4, \lambda=0.5, a=0.15, b=1.2, \alpha=(\frac{1}{4},\frac{1}{3},\frac{1}{2},1)\ and\ \beta=(\frac{1}{3},\frac{1}{3},\frac{1}{2},\frac{1}{2}).$ Evidently, $(\frac{1}{\alpha_1},\frac{1}{\alpha_2},\frac{1}{\alpha_3},\frac{1}{\alpha_4})\overset{\rm m}{\succeq}(\frac{1}{\beta_1},\frac{1}{\beta_2},\frac{1}{\beta_3},\frac{1}{\beta_4}).$ One can verify that the conditions of Theorem \ref{tpla} are all statisfied. Figure 3 plots the hazard rate functions of $X_{2:4}$ and $Y_{2:4}$, denoted as $r_{X_{2:4}}(x)$ and $r_{Y_{2:4}}(x)$, all $x=-{\rm ln}(u)$ and $u\in (0,1]$. Thus, we can get $X_{2:4}\geq_{\rm hr}Y_{2:4}$.
\begin{figure}[H]
\centering
{\includegraphics[width=7cm]{mphr_hr3_-eps-converted-to.pdf}\label{fig2a}}
\caption{Plots of hazard rate functions of $X_{2:4}[Y_{2:4}]$, for all $x=-\ln u$ and $u\in(0,1]$.
}
\end{figure}
\end{example}
In the next, we discuss stochastic comparisons on the second-order statistics arising from multiple-outlier MPHR samples in the sense of the hazard rate order. We first present the following result.
\begin{theorem}\label{thrp}
Let $X_{1}, X_{2}, \dots,X_{n}$ be independent random variables following the multiple-outlier MPHR moder with survival functions$(\frac{\alpha\overline{F}^{\lambda_1}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_1}(x)}{\rm I}_p,\frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}{\rm I}_q),$ where $p+q=n,p,q\geq1,$ $\overline{F}(x)$ is the baseline survival function, and ${\rm I}_p$ stand for a p-dimensional vector with all its components equal to 1.Let $Y_{1}, Y_{2}, \dots,Y_{n}$ be another set of independent random variables following the multiple-outlier MPHR moder with survival functions$(\frac{\alpha\overline{F}^{\lambda_2}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_2}(x)}{\rm I}_p,\frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}{\rm I}_q).$ If $\lambda\geq\lambda_2\geq\lambda_1,$ then $X_{2:n}\geq_{\rm hr}Y_{2:n}.$
\end{theorem}
\begin{proof}
The idea of the proof is borrowed from Theorem 3.1 of \cite{cai2017hazard}.
The survival function of $X_{2:n}$ and $Y_{2:n}$ are given by
\begin{eqnarray*}
\overline{F}_{X_{2:n}}(x)&=&p\left[ \frac{\alpha\overline{F}^{\lambda_1}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_1}(x)}\right]^{p-1}\left[ \frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}\right]^q+q\left[ \frac{\alpha\overline{F}^{\lambda_1}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_1}(x)}\right]^p\left[ \frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}\right]^{q-1}\\
&&-(n-1)\left[ \frac{\alpha\overline{F}^{\lambda_1}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_1}(x)}\right]^p\left[ \frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}\right]^q,
\end{eqnarray*}
and
\begin{eqnarray*}
\overline{F}_{X^*_{2:n}}(x)&=&p\left[ \frac{\alpha\overline{F}^{\lambda_2}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_2}(x)}\right]^{p-1}\left[ \frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}\right]^q+q\left[ \frac{\alpha\overline{F}^{\lambda_2}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_2}(x)}\right]^p\left[ \frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}\right]^{q-1}\\
&&-(n-1)\left[ \frac{\alpha\overline{F}^{\lambda_2}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_2}(x)}\right]^p\left[ \frac{\alpha\overline{F}^{\lambda}(x)}{1-\overline{\alpha}\overline{F}^{\lambda}(x)}\right]^q,
\end{eqnarray*}
where $\overline{\alpha}=1-\alpha$ and $p+q=n.$ Let us set $\overline{F}^{\lambda}(x)=e^{-\lambda(-{\rm ln}\overline{F}(x))}$ and $t=-{\rm ln}\overline{F}(x)$. We can then rewritten the survival function $X_{2:n}$ as:
\begin{eqnarray*}
\overline{F}_{X_{2:n}}(t)&=&p\left[ \frac{\alpha}{e^{\lambda_1t}-\overline{\alpha}}\right]^{p-1}\left[ \frac{\alpha}{e^{\lambda t}-\overline{\alpha}}\right]^q+q\left[ \frac{\alpha}{e^{\lambda_1t}-\overline{\alpha}}\right]^p\left[ \frac{\alpha}{e^{\lambda t}-\overline{\alpha}}\right]^{q-1}\\
&&-(n-1)\left[ \frac{\alpha}{e^{\lambda_1t}-\overline{\alpha}}\right]^p\left[ \frac{\alpha}{e^{\lambda t}-\overline{\alpha}}\right]^q, t\geq0.
\end{eqnarray*}
Thus, its hazard rate function is given by
\begin{eqnarray*}
r_{X_{2:n}}(t)&=&\frac{{\rm d}[-{\rm ln}\overline{F}_{X_{2:n}}(t)]}{{\rm d}t} =\frac{pA_1\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +qA_2\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)A}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)},
\end{eqnarray*}
where $A_1=(p-1)\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$, $A_2=p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$ and $A=p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$.
It suffices to show $r_{X_{2:n}}(t)\leq r_{X^*_{2:n}}(t)$ for $\lambda\geq \lambda_2\geq \lambda_1>0,$ i.e.,
\begin{eqnarray*}
\frac{pA_1\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +qA_2\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)A}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}\leq\frac{pB_1\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right) +qB_2\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)B}{p\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}.
\end{eqnarray*}
where $B_1=(p-1)\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$, $B_2=p\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$ and $B=p\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$, or equivalently
\begin{eqnarray*}
&&\Bigg \{ p\left[ (p-1)\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)
+q\left[ p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right)\\
&&
- (n-1)\left[ p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \Bigg \}\times \frac{1}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}\\
&\leq& \Bigg \{ p\left[ (p-1)\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)
+q\left[ p\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) \\
&&- (n-1)\left[ p\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \Bigg \}\times \frac{1}{p\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}.
\end{eqnarray*}
Denote by
\begin{eqnarray*}
M_1&=&\frac{1}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}\times \Bigg \{ p\left[ (p-1)\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)\\
&&+q\left[ p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right)- (n-1)\left[ p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right]\Bigg \},
\end{eqnarray*}
\begin{eqnarray*}
M_2&=&\frac{1}{p\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}\times \Bigg \{ p\left[ (p-1)\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)\\
&&+q\left[ p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right)- (n-1)\left[ p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right]\Bigg \},
\end{eqnarray*}
\begin{eqnarray*}
M_3&=&\frac{1}{p\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right) -(n-1)}\times \Bigg \{p\left[ (p-1)\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)\\
&&+q\left[ p\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right] \left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right)
-(n-1)\left[ p\frac{\lambda_2e^{\lambda_2t}}{e^{\lambda_2t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right]\Bigg \}.
\end{eqnarray*}
For $\lambda\geq\lambda_2\geq\lambda_1>0$, by the increasing properties of $\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}$ with respect to $\lambda$, we have
\begin{eqnarray*}
\frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}\geq\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}.
\end{eqnarray*}
It holds that
\begin{eqnarray*}
M_3-M_2&\overset{\rm sgn}{=}&p(p-1)\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)\left[ \frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}-\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right]+pq\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right)\times \left[ \frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}-\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right]\\
&&-p(n-1)\left[ \frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}-\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right]\\
&=&p\left[ \frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}-\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right]\left[ (p-1)\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)+q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha}\right)-(n-1)\right]\\
&\geq&p\left[ \frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}-\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right]\left[p+q-1-(n-1) \right]\\
&=&0.
\end{eqnarray*}
Which implies that $M_2\leq M_3$. Let
\begin{eqnarray*}
&&P=q\left[p \frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}+(q-1)\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}} \right]\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha} \right)-(n-1)\left[ p\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}\right]\\
&&Q= p\left[ (p-1)\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}+q\frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}} \right]\\
&&U=q\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha} \right)-(n-1) \quad and \quad V=p.
\end{eqnarray*}
So, we have
\begin{eqnarray*}
QU-PV&=&pq\left( \frac{e^{\lambda t}-\overline{\alpha}}{\alpha} \right)\left[ \frac{\lambda e^{\lambda t}}{e^{\lambda t}-\overline{\alpha}}-\frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right]+p(n-1)\left( \frac{\lambda_1 e^{\lambda_1 t}}{e^{\lambda_1 t}-\overline{\alpha}}\right)\geq0.
\end{eqnarray*}
So,we have
\begin{eqnarray*}
M_2-M_1&=&\frac{Q\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)+P}{V\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)+U}-\frac{Q\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)+P}{V\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)+U}\\
&=&\frac{\left[Q\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)+P \right]\left[V\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)+U \right]-\left[Q\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)+P \right]\left[ V\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)+U\right] }{\left[ V\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}\right)+U\right] \left[ V\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)+U\right] } \\
&\overset{\rm sgn}{=}&(QU-PV)\left( \frac{e^{\lambda_2t}-\overline{\alpha}}{\alpha}- \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right)\\
&\geq&0,
\end{eqnarray*}
which means that $M_1\leq M_2$. Thus, $M_1\leq M_3$, and the theorem follows.
\end{proof}
In the next, we turn to discuss the effect generated by the discrepancy among sample sizes on the hazard rate function of the second-order statistic arising from multiple-outlier MPHR samples.
\begin{theorem}\label{thrf}
Let $X_{1}, X_{2}, \dots,X_{p},X_{p+1},\dots,X_{n}$ be independentrandom variables following the multiple-outlier MPHR moder with survival functions$(\frac{\alpha\overline{F}^{\lambda_1}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_1}(x)}{\rm I}_p,\frac{\alpha\overline{F}^{\lambda_2}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_2}(x)}{\rm I}_q),$ where $p+q=n,p,q\geq1$ and $\overline{F}(x)$ is the baseline survival function.Let $Y_{1}, Y_{2}, \dots,Y_{p^*},Y_{p^*+1},\dots,Y_{n^*}$ be another set of independent random variables following the multiple-outlier MPHR moder with survival functions$(\frac{\alpha\overline{F}^{\lambda_1}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_1}(x)}{\rm I}_p^*,\frac{\alpha\overline{F}^{\lambda_2}(x)}{1-\overline{\alpha}\overline{F}^{\lambda_2}(x)}{\rm I}_q^*),$ where $p^*+q^*=n^*,p^*,q^*\geq1$. Denote by $X_{2:n}$ and $Y_{2:n^*}$ the second-order statistic arising from these two sets of multiple-outlier MPHR models, respectively. Suppose that $p^*\leq p\leq q\leq q^*$ and $\lambda_1\leq\lambda_2$.Then, we have
\begin{equation*}
(p,q)\preceq_{\rm w}(p^*,q^*)\Rightarrow X_{2:n}\geq_{\rm hr} Y_{2:n^*}
\end{equation*}
\end{theorem}
\begin{proof}
The idea of the proof is borrowed from Theorem 3.4 of \cite{cai2017hazard}.
Following the notation in the proof of Theorem 4, the hazard rate function of $X_{2:n}$ can be written as follows:
\begin{eqnarray*}
r_{X_{2:n}}(t)=\frac{pT_1\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +qT_2\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n-1)T}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n-1)},
\end{eqnarray*}
where $T_1=(p-1)\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda_{2} e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}$, $T_2=p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+(q-1)\frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}$ and $T=p\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q\frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}$. Denote by $r_{Y_{2:n^*}}(t)$ the hazard rate function of $Y_{2:n^*}$. We need to show that $r_{X_{2:n}}(t)\leq r_{Y_{2:n^*}}(t)$, i.e.,
\begin{eqnarray*}
\frac{pT_1\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +qT_2\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n-1)T}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n-1)}
\leq\frac{p^*T^*_1\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q^*T^*_2\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n^*-1)T^*}{p^*\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q^*\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n^*-1)}.
\end{eqnarray*}
where $T^*_1=(p^*-1)\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q^*\frac{\lambda_{2} e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}$, $T^*_2=p^*\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+(q^*-1)\frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}$ and $T^*=p^*\frac{\lambda_1e^{\lambda_1t}}{e^{\lambda_1t}-\overline{\alpha}}+q^*\frac{\lambda_2 e^{\lambda_2 t}}{e^{\lambda_2 t}-\overline{\alpha}}$. Let $a_i=\frac{\lambda_ie^{\lambda_it}}{e^{\lambda_it}-\overline{\alpha}}$, $b_i=\left( \frac{e^{\lambda_it}-\overline{\alpha}}{\alpha}\right)$ and $c_{ij}=a_ib_j$ for $i,j=1,2.$ Denote
\begin{eqnarray*}
\phi(p,q)&=&\frac{pT_1\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +qT_2\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n-1)T}{p\left( \frac{e^{\lambda_1t}-\overline{\alpha}}{\alpha}\right) +q\left( \frac{e^{\lambda_2 t}-\overline{\alpha}}{\alpha}\right) -(n-1)}\\
&=&\frac{p(p-1)c_{11}+pqc_{21}+pqc_{12}+q(q-1)c_{22}-(n-1)(pa_1+qa_2)}{pb_1+qb_2-(n-1)}.
\end{eqnarray*}
It suffices to prove $\phi(p,q)\leq\phi(p^*,q^*)$ under the condition $p^*\leq p\leq q\leq q^*$ and $(p,q)\preceq_{\rm w}(p^*,q^*)$. Taking the derivative of $\phi(p,q)$ with respect to $p$, we have
\begin{eqnarray*}
\frac{\partial\phi(p,q)}{\partial p}&\overset{\rm sgn}{=}&\left[(2p-1)c_{11}+q(c_{21}+c_{12})-(n-1)a_1-pa_1-qa_2 \right]\times\left[ pb_1+qb_2-(n-1)\right]\\
&&-\left[ p(p-1)c_{11}+pqc_{21}+pqc_{12}+q(q-1)c_{22}-(n-1)(pa_1+qa_2)\right]\times(b_1-1)\\
&=&a_1\left[pb_1+qb_2-(n-1)\right]\times\left[ (p-1)b_1+qb_2-(n-1)\right]+(pc_{11}+qc_{22})(b_1-1) \\
&\geq&0.
\end{eqnarray*}
Similarly,
\begin{eqnarray*}
\frac{\partial\phi(p,q)}{\partial q}&\overset{\rm sgn}{=}&a_2\left[pb_1+qb_2-(n-1)\right]\times\left[ pb_1+(q-1)b_2-(n-1)\right]+(pc_{11}+qc_{22})(b_2-1) \\
&\geq&0.
\end{eqnarray*}
From the proof of Theorem 4, we get $a_2\geq a_1\geq0$ and $b_2\geq b_1\geq1$.Thus, we have
\begin{eqnarray*}
\frac{\partial\phi(p,q)}{\partial q}-\frac{\partial\phi(p,q)}{\partial p}&\overset{\rm sgn}{=}&\left[ pc_{21}+(q-1)c_{22}-(p-1)c_{11}-qc_{12}-(n-1)(a_2-a_1)\right]\\
&&\times \left[pb_1+qb_2-(n-1) \right]+(pc_{11}+qc_{22})(b_2-b_1)\\
&\geq&\left[ pc_{21}+(q-1)c_{22}-(p-1)c_{11}-qc_{12}-(n-1)(a_2-a_1)\right]\\
&&\times \left[pb_1+qb_2-(n-1) \right]+(pb_1+qb_2)(a_1b_2-a_1b_1)\\
&=&\left[ pc_{21}+(q-1)c_{22}-pc_{11}-(q-1)c_{12}-(n-1)(a_2-a_1)\right]\\
&&\times \left[pb_1+qb_2-(n-1) \right]+(n-1)(c_{12}-c_{11})\\
&\geq&(a_2-a_1)\left[ pb_1+(q-1)b_2-(n-1)\right]+(n-1)(c_{12}-c_{11})\\
&\geq&0.
\end{eqnarray*}
Now, the desired result follows from lemma \ref{wconvex}.
\end{proof}
\begin{remark}
The results of theorem \ref{thrp} and theorem \ref{thrf} improve Theorem 3.1 and theorem 3.4 of \cite{cai2017hazard} from the PHR model to the case of the MPHR model.
\end{remark}
The following numerical example is provided as an illustration of Theorem \ref{thrf}.
\begin{example}
Suppose $\overline{F}(x)=e^{-(ax)^b}, a>0, b>0$, Set $n=7, n^*=9, a=1.5, b=0.2, \alpha=0.05, \lambda_1=0.1, \lambda_2=0.3, p=3, q=4, p^*=1, q^*=8.$ It is easy to show that conditions $\lambda_1\leq\lambda_2, p^*\leq p\leq q\leq q^*$ and $(p,q)\preceq_{\rm w}(p^*,q^*)$ are all statisfied in theorem \ref{thrf}. Figure 4 shows the hazard rate functions of $X_{2:7}$ and $X_{2:9}$, from which it can be obversed that $r_{X_{2:7}}(x)$ is less than $r_{Y_{2:9}}(x)$, for $x>0$, thus validating the result in theorem \ref{thrf}.
\begin{figure}[H]
\centering
{\includegraphics[width=7cm]{mphr_hr4_-eps-converted-to.pdf}\label{fig2a}}
\caption{Plots of hazard rate functions of $X_{2:7}[Y_{2:9}]$, for all $x=-\ln u$ and $u\in(0,1]$.
}
\end{figure}
\end{example}
\section{Concluding remarks}\label{con}
In this paper, we study stochastic comparisons on the second-order statistics from heterogeneous dependent or independent MPHR samples. Some ordering results are established for the usual stochastic, hazard rate orderings on the second-order statistics. These new
results established here provide theoretical guidance both for the winner’s prize for bid in the second-price reverse auction in auction theory and fail-safe system design in reliability theory.
\section*{Funding}
This research is supported by the National Natural Science Foundation of China.
%
%
%
%
%
%
%
%
|
\section{Introduction}
Federated learning (FL) is a machine learning paradigm that performs decentralized training of models on mobile devices (e.g., smartphones) with locally stored data~\cite{mcmahan01}. FL trains on a large corpus of private user data without collecting them, with only the model weight updates communicated externally from the user’s device~\cite{bonawitz01}. With FL, researchers have proposed to improve AI in diverse domains: human mobility prediction~\cite{feng01}, RF localization~\cite{ciftler01}, traffic sign classification~\cite{albaseer01}, tumour detection~\cite{li02, jimenez01}, and Clinical Decision Support (CDS) model for COVID-19~\cite{dayan01}. FL also has product deployments as large companies such as Google or Taobao deploy language processing and item recommendation tasks across millions of real-world devices~\cite{niu01, yang02}.
One of the key objectives in FL is to optimize \textit{time-to-accuracy} performance, which is a wall clock time for a model to achieve the target accuracy~\cite{lai01}. Achieving high time-to-accuracy performance is important as FL consumes significant computation and network resources on edge-user devices~\cite{dinh01}. For model developers who prototype a mobile AI with FL without a proxy dataset, achieving faster convergence on thousands to millions of devices is desired to efficiently test multiple model architectures and hyperparameters~\cite{kairouz01}. Service providers who frequently update a model with continual learning with FL require to minimize the user overhead with better time-to-accuracy performance~\cite{le01}.
In realistic FL scenarios, however, heterogeneous data being distributed over isolated users becomes the main challenge in achieving high time-to-accuracy performance~\cite{zhao01, li03}. While data engineering techniques such as
\textit{importance sampling}~\cite{katharopoulos01, alain01, loshchilov01, schaul01} are widely adopted in centralized learning to optimize the training process, applying them in FL is infeasible as it requires private user data sharing. For this reason, previous FL algorithms~\cite{mcmahan01, li01, lai01} mostly treat every client data equally, which could result in a waste of computational resources and slow convergence. We conducted a preliminary study to examine this phenomenon and found that the ratio of informative samples is reduced from 93.2\% to 20\% as the FL progresses.
This could further exacerbate with heterogeneous hardware of FL clients, as low-end devices might fail to send model updates while training a large portion of unimportant samples.
To address this problem, we propose \textit{FedBalancer}{}, a systematic FL framework with sample selection. The sample selection of \textit{FedBalancer}{} prioritizes more ``informative'' samples of clients to efficiently utilize their computational effort. This allows low-end devices to contribute to the global training within the round deadline by focusing on smaller but more important training samples. To achieve high time-to-accuracy performance, the sample selection is designed to operate without additional forward or backward pass for sample utility measurement at FL rounds. Lastly, \textit{FedBalancer}{} can coexist and collaborate with orthogonal FL approaches to further improve performance.
To realize \textit{FedBalancer}{}, we addressed the following challenges: (1) Simply reducing the training data of a client with random sampling could lead to lower model accuracy as the statistical utility of the training data would decrease. As such, \textit{FedBalancer}{} selects samples based on their statistical utility measurement. (2) Collecting sample-level statistical utility for sample selection might break the privacy guarantee of FL. To address this problem, we propose client-server coordination to maintain \textit{loss threshold}, which allows clients to effectively select important samples while only exposing differentially-private statistics of their data. (3)
The sample selection itself might not directly lead to time-to-accuracy performance improvement when the FL round deadline remains fixed.
To formulate the benefit of selecting different deadlines, we propose a metric \textit{deadline efficiency} (DDL-E)
that calculates the number of round-completed clients per time. This allows \textit{FedBalancer}{} to predict the optimal deadline with varying client training data.
We implemented \textit{FedBalancer}{}
and conducted experiments on five datasets from three domains that contain real-world user data. Compared with existing FL aggregation algorithms with deadline configuration methods, \textit{FedBalancer}{} improves the time-to-accuracy performance by 1.20$\sim$4.48$\times$. \textit{FedBalancer}{} achieves this improvement without sacrificing model accuracy; in fact, it improves the accuracy by 1.1\%$\sim$5.0\%. In addition, we implement \textit{FedBalancer}{} on top of three orthogonal FL algorithms to demonstrate that \textit{FedBalancer}{} is easily applicable to other FL approaches and improves their time-to-accuracy performance.
We summarize our contributions as follows:
\begin{itemize}[topsep=0pt, align=left, labelwidth=10pt, leftmargin=15pt]
\item We propose a systematic framework for FL with sample selection, which actively selects high utility samples at each round without collecting privacy-invasive sample-level information from clients.
\item We propose a deadline control strategy for each round of FL based on the newly defined metric \textit{deadline efficiency}, which optimizes the time-to-accuracy performance along with our sample selection.
\item We implement \textit{FedBalancer}{} jointly with existing FL algorithms, showing improvement in both time-to-accuracy and model accuracy.
\end{itemize}
\section{Background and Motivation}
\subsection{Federated Learning}
\label{sec:fedlearning}
Federated Learning (FL) operates across multiple mobile devices to globally train a model from the distributed client data. FedAvg~\cite{mcmahan01}, the most commonly used FL approach~\cite{jiang01}, operates as follows: (i) Suppose there are $N$ clients in an FL system. For each round of FL, the server randomly selects $K$ clients ($K<<N$) who participate in model training. (ii) At the \textit{R}-th round, the server transmits the current model weights $\displaystyle {w}_R$ to the selected clients. (iii) Each client then performs model training for $E$ epochs with their local data and generates $\displaystyle {w}^k_{R+1}$, where $k$ denotes the client index. (iv) Clients upload the updated model parameters to the server, and (v) the server aggregates different clients' updates and generate the updated model as ${w}_{R+1} \leftarrow \sum^K_{k=1} {\frac{n_k}{n}}{{w}^k_{R+1}}$,
where $n_k$ indicates the number of data points of client $k$ and $n$ the number of all data points.
A key objective in FL is to optimize time-to-accuracy performance. FL tasks typically require hundreds to thousands of rounds to converge~\cite{caldas01, lai01}, and clients participating at a round undergo substantial computational and network overhead~\cite{dinh01}. Deploying FL across thousands to millions of devices should be done efficiently, quickly reaching the model convergence while not sacrificing the model accuracy. This becomes more important when FL has to be done multiple times, as often the case when model developers prototype a new model with FL without a proxy dataset or periodically update a deployed model to new domain via continual learning or online learning with FL.
However, the heterogeneities of the real-world mobile clients make it challenging to achieve good time-to-accuracy performance. \newline \noindent \textbf{Data heterogeneity}: The client-generated training data are imbalanced and not independent and identically distributed (non-IID) due to each user's different mobile device usage or physical and mental characteristics~\cite{mcmahan01, wu01}. While the training data in centralized learning could be filtered and organized with data engineering techniques~\cite{katharopoulos01, alain01, loshchilov01, schaul01} to address data heterogeneity, applying the same solution is infeasible in FL as it is dealing with distributed
private data on clients. Such data heterogeneity of FL clients results in slow convergence and suboptimal performance~\cite{li01, zhao01, li03}. \newline \noindent \textbf{Hardware heterogeneity}: The client devices
have distinct computational capabilities and network connectivity, resulting in up to 12$\times$ more round completion time between different clients~\cite{li04}.
Thus, waiting for every client to complete its task at a round might significantly delay the training process. A common practice for such an issue is to set a \textit{deadline} for a round duration and drop clients that fail to send the model updates before the deadline~\cite{nishio01, abdelmoniem01}. However, this results in less contribution from clients with low-end devices, which could result in delayed convergence or biased model training~\cite{yang01, li01}.
\subsection{Motivational Study}
\label{sec:motivation}
To address the heterogeneity problems, various approaches have been proposed. Researchers proposed FL algorithms that allow clients to train different numbers of local epochs~\cite{li01} or different subsets of model weights~\cite{diao01, horvath01} based on clients' hardware capabilities. Model personalization has also been proposed to tackle data heterogeneity~\cite{pillutla01, ouyang01, fallah01, tu01, li06, li07, liu2020pmc, liu2021distfl}. In addition, client selection strategies for FL training have been proposed to optimize the convergence speed on heterogeneous clients~\cite{lai01, cho01, cho02}. Although these techniques have improved the convergence speed or accuracy, they treat all data of clients equally, which could lead to a waste of computational resources for training non-important samples and result in suboptimal time-to-accuracy performance.
We investigate such limitation of previous FL approaches and discuss how \textit{FedBalancer}{} should be designed based on our experiments that simulate FL on heterogeneous clients. We used widely used datasets to benchmark FL methods: FEMNIST~\cite{cohen01} and Shakespeare~\cite{shakespeare01}.
We simulated heterogeneous training latency and network connectivity on the real-world clients as described in Section~\ref{sec:experimentalsetup}.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/02Background/motivation-1-v2-square.pdf}
\caption{FEMNIST dataset.}
\label{fig:motivation-1}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/02Background/motivation-4-v2-square.pdf}
\caption{Shakespeare dataset.}
\label{fig:motivation-2}
\end{subfigure}
\caption{Ordered gradient norm (GN) of samples from FL training rounds on two different datasets
}
\vspace{-0.2cm}
\label{fig:motivation-gn}
\vspace{-0.2cm}
\end{figure}
\textbf{Inefficiency of Full Data Training.} As in FedAvg~\cite{mcmahan01}, most FL approaches let clients to fully train their entire data at a training round. Other approaches that samples a subset of client training data for each round use equal weights on all data~\cite{lai01, fallah01, jiang02}. However, previous studies in centralized machine learning~\cite{katharopoulos01, alain01, loshchilov01} found that the importance of all samples are not equal; A large portion of samples are learned quickly after few training rounds and could be ignored afterwards. Thus, we conducted an experiment to verify if the same phenomenon also applies in FL and how significant it is. We suspect that the inefficiency problem would be more serious in FL as limited computational resources of mobile clients could be wasted on non-important samples.
In this experiment, we measured the Gradient Norm (GN) of each sample on federated clients to evaluate sample-level contribution on a model update. For each training round, we collected and sorted the GN of each sample from the selected clients. We removed top 5\% of samples to avoid evaluating noisy samples and applied min-max scaling to [0.0, 1.0] interval. The experiment ran until the model converged. The results from each dataset are shown in Figure~\ref{fig:motivation-1} and~\ref{fig:motivation-2}. Each column of graphs indicates the sorted GN of samples from a training round, where the largest value is located at the bottom. The x-shaped points illustrate where the gradient norm with scaled value of 0.2 exists at each training round.
From both datasets, we observe that the GN of samples are mostly high at the early stage of FL, but only small portion of samples show high GN afterwards, meaning that the number of samples that contribute knowledge to the model are reducing as the training progresses. In FEMNIST dataset, only 6.8\% of the samples from early training rounds (round index 0$\sim$150) have less scaled GN value than 0.2, but 80.0\% of the samples are less than 0.2 for the last 150 training rounds. Similarly in the Shakespeare dataset, 6.7\% of the samples from early training rounds (round index 0$\sim$25) have less scaled GN value than 0.2, but 54.8\% of the samples are less than 0.2 for the last 25
training rounds. This result also implies that the samples are not equally important during these FL tasks and current FL approaches could spend large portion of training time for samples that have negligible contribution to the model update.
This experiment motivates \textit{FedBalancer}{} to start training with all the samples at the beginning, and then gradually remove samples that the model has already learned. In Section~\ref{sec:cds}, we further illustrate how \textit{FedBalancer}{} selects a subset of samples of each client at training rounds based on our findings in this experiment.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/02Background/motivation-3-v2-square.pdf}
\caption{FEMNIST dataset.}
\label{fig:motivation-3}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/02Background/motivation-5-v2-square.pdf}
\caption{Shakespeare dataset.}
\label{fig:motivation-4}
\end{subfigure}
\caption{FL on two datasets with different deadline configuration methods
}
\vspace{-0.5cm}
\vspace{-0.5cm}
\label{fig:motivation-ddl}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth]{figures/03FedBalancer/overview13.pdf}
\vspace{-0.1cm}
\caption{Overview of \textit{FedBalancer}{} architecture and its operation in an FL round in seven steps.
}
\label{fig:overview}
\vspace{-0.1cm}
\end{figure*}
\textbf{Importance of Deadline Selection.} While the existence of \textit{optimal deadline} for achieving shortest training time in FL has been studied~\cite{yang01}, controlling the deadline for high time-to-accuracy performance has been largely overlooked.
To understand the performance of existing deadline configuration methods, we conducted an experiment on the two datasets
with SmartPC~\cite{li04} and four different fixed deadlines --- 0.5$T$, 1.0$T$, 1.5$T$, and 2.0$T$ --- where $T$ indicates the mean of round completion time on every participating client. For SmartPC, we implemented a training round to last until 80\% of the clients complete their task, where 80\% is suggested by Li et al.~\cite{li04}.
Figure~\ref{fig:motivation-3} and ~\ref{fig:motivation-4} illustrate the results on two datasets. Our takeaway from this experiment is twofold: (1) The deadline is a significant factor in achieving fast convergence speed and high accuracy, and (2) no single method achieved the best performance for both FL tasks.
In FEMNIST dataset, deadline 1.0$T$ achieved the highest final accuracy (.815), while being 43.6\%, 47.0\% and 77.9\% faster than SmartPC, deadline 1.5$T$, and deadline 2.0$T$, respectively, in achieving the test accuracy of .750. On the other hand, in Shakespeare dataset, SmartPC achieved the highest final accuracy (.488) while being 61.7\%, 42.8\%, and 45.4\% faster than deadline 1.0$T$, 1.5$T$, and 2.0$T$, respectively, in achieving the test accuracy of .450. Deadline 0.5$T$ could not achieve high accuracy in either task, as most clients failed to upload their model update within the deadline.
In a training round of FL, the computation time has been shown to be the bottleneck~\cite{wang02, pilla01, yang01}. As the
amount of client computation changes with the sample selection of \textit{FedBalancer}{}, finding an optimal deadline could be an important problem in achieving high time-to-accuracy performance. To this end, in Section~\ref{sec:adc}, we propose how \textit{FedBalancer}{} finds the optimal deadline for each training round for efficient FL on heterogeneous clients.
\section{\textit{FedBalancer}{}}
\subsection{Overview}
\label{sec:overview}
For each round of FL, \textit{FedBalancer}{} adaptively selects the client training data and controls the deadline to achieve high time-to-accuracy performance. We first provide an overview of how \textit{FedBalancer}{} operates in an FL round and then describe how each component of \textit{FedBalancer}{} is designed.
Figure~\ref{fig:overview} depicts the \textit{FedBalancer}{} architecture during an FL round. The main functionality of \textit{FedBalancer}{} is to actively control two variables for each round: \textit{loss threshold} ($lt$) and \textit{deadline} ($ddl$). The loss threshold works as a parameter that determines each client's training data (Section~\ref{sec:cds}) and the \textit{deadline} determines the round termination time. The numbers inside a circle show the seven steps of an FL round with \textit{FedBalancer}{}.
The server first transmits the current model weights $\displaystyle {w}_R$, the loss threshold $lt_R$, and the deadline $ddl_R$ (\textit{R} indicates the \textit{R}-th round) to the selected clients of the round (Step 1). The \textit{sample selection module} at each client selects the partial training data with the received loss threshold (Step 2) and trains the received model (Step 3). The client transmits the model update and the \textit{metadata} collected from the sample selection and model training (Step 4), and the server aggregates these responses from all clients (Step 5). Based on the metadata from clients, the \textit{loss threshold selection module} and the \textit{deadline selection module} each selects the loss threshold $lt_{R+1}$ and the deadline $ddl_{R+1}$ for the next round (Steps 6 and 7).
\textbf{Challenges: } We aim to address the following challenges to realize \textit{FedBalancer}{}:
\begin{enumerate}[topsep=3pt, align=left, labelwidth=10pt, leftmargin=15pt]
\item \textit{Sample selection without accuracy drop:}
Simply reducing the training data with random sampling could result in degradation of model accuracy due to the decreased statistical utility. \textit{FedBalancer}{} should thus prioritize samples based on their statistical utility.
\item \textit{Privacy-preserving sample selection in FL}: While we aim to select an optimal set of client training data for each round, requiring up-to-date sample-level information from clients could harm the privacy guarantee of FL. \textit{FedBalancer}{} should select client training samples without collecting privacy-invasive information from clients.
\item \textit{Predicting optimal deadline with varying data}: As computation could be the bottleneck of an FL round~\cite{wang01, pilla01, yang01}, applying sample selection strategy of \textit{FedBalancer}{} might greatly change the round completion time of clients. \textit{FedBalancer}{} should adaptively predict the optimal deadline based on the sample selection status of heterogeneous clients for each round of FL.
\end{enumerate}
\subsection{Client Sample Selection
\label{sec:cds}
In Section~\ref{sec:motivation}, we observed that existing FL methods consume large portion of time to train samples that contribute only small gradient to the model. As these samples are quickly learned after few rounds, we design \textit{FedBalancer}{} to start training with all samples and gradually remove \textit{already-learned} samples. This enables \textit{FedBalancer}{} to efficiently focus on more \textit{important} samples at each round while optimizing the training process of FL. However, implementing such design in FL is non-trivial as the following question needs to be addressed: \textit{How should \textit{FedBalancer}{} distinguish between \textit{important} and \textit{non-important} samples at each stage of FL?}
A straightforward solution is to collect every sample-level importance information from all clients to a server at each round, and derive a criteria that determines more important samples to a current model. However, such approach is hardly applicable in FL as sharing information of every sample could break the privacy guarantee of FL and reveal the clients' data. This approach also incurs significant network overhead in communicating all sample's information at each round. An alternative approach is to have clients classify more important samples within their local data without exposing any information. However, as client data distributions are heterogeneous in FL~\cite{mcmahan01, wu01, li01, zhao01, li03}, clients could struggle to determine important samples without knowing the global data distribution.
To address this issue, we propose a client-server coordination to maintain a \textit{loss threshold} variable, which enables clients to effectively select important samples without exposing private sample-level information.
\textit{FedBalancer}{} actively controls the loss threshold based on the collected metadata from clients with \textit{loss threshold selection module}, where the metadata consists of differentially-private statistics of sample-level information.
Note that \textit{FedBalancer}{} uses the \textit{loss} of a sample to measure the statistical utility (and thus the importance) of a sample to the current model,
similar to \textit{Importance Sampling}~\cite{loshchilov01, schaul01}. While other studies have also leveraged \textit{gradient norm} or \textit{gradient norm upper bound}~\cite{katharopoulos01, alain01, li08} to achieve the same goal, we use loss as it is more widely applicable to FL tasks with non gradient-based optimizations~\cite{rios01}.
\subsubsection{Sample selection module}
\begin{algorithm}[t]
\caption{Client $i$ at $R$-th round: Sample selection
}\label{alg:ssm}
\begin{algorithmic}[1]
\Procedure{SelectSample}{$D^i$, $B^i$, $loss^i$, $lt_{R}$, $ddl_{R}$, $\displaystyle {w}_{R}$, $p$, $E$}
\State $S \gets $\hspace{0.1cm}\texttt{maxTrainableSize}$(mean(B^i), ddl_{R}, E)$ \label{algline:ssm-line2}
\If{$S \geq len(D^i)$} \label{algline:ssm-line3}
\State $D^i_{R} \gets D^i$ \label{algline:ssm-line4}
\Else
\State $D^i_{R}, OT^i, UT^i \gets \emptyset$
\For{$x \gets 1$ to $len(D^i)$} \label{algline:ssm-line8}
\State \textbf{if} $loss^i[x] \geq \displaystyle lt_{R}$ \textbf{then} $OT^i.insert(D^i[x])$ \label{algline:ssm-line9}
\State \textbf{else} $UT^i.insert(D^i[x])$ \label{algline:ssm-line10}
\EndFor \label{algline:ssm-line11}
\State $L \gets max(S, len(OT^i))$ \label{algline:ssm-linenew}
\State $D' \gets $\hspace{0.1cm}\texttt{randSample}$(OT^i, L \cdot p))$ \label{algline:ssm-line16}
\State $D'' \gets $\hspace{0.1cm}\texttt{randSample}$(UT^i, L \cdot (1 - p))$\label{algline:ssm-line17}
\EndIf \label{algline:ssm-line18}
\State $D^i_{R} \gets $\texttt{concatenate}$(D', D'')$
\State \textbf{return} $D^i_{R}$
\EndProcedure
\end{algorithmic}
\end{algorithm}
Algorithm~\ref{alg:ssm} describes how the \textit{sample selection module} of a client selects the samples at each round after it receives the loss threshold from the server. Let us assume that client $i$'s sample selection module is working at round $R$, with a given loss threshold $lt_R$.
First, the module measures if a sample selection on a client is required for this round --- i.e., it measures if a client is fast enough to train its full dataset within the given deadline $ddl_R$. While \textit{FedBalancer}{} makes clients focus on more important samples for efficient training, it allows clients to fully utilize its statistical utility if feasible. To this end, we calculate the maximum number of samples $S$ which client $i$ can train for $E$ epochs before the deadline, and verify if it is larger than the size of the client dataset $D^i$ (Line~\ref{algline:ssm-line2} -~\ref{algline:ssm-line4}). As the computational ability of a client can change according to the device runtime conditions~\cite{yang03}, \textit{FedBalancer}{} collects the batch training latency of a client as $B^i$ during FL and uses the mean latency to estimate the max samples it can process. To calculate the mean latency from the first round, \textit{FedBalancer}{} asks clients to sample the latency for $k$ times before FL begins. We used 10 for $k$ in our evaluation.
If the client $i$ is incapable of training its full dataset, the \textit{sample selection module} determines which samples to train at the FL round by using the \textit{list of sample loss} ($loss^i$). The loss list shows the statistical utility of all samples on the current model. While such loss list could be obtained by inferring all samples on up-to-date model at each round, it requires additional forward pass latency that could degrade the time-to-accuracy performance. Therefore, clients of \textit{FedBalancer}{} perform whole-dataset forward pass only once when they are first selected at a round to generate a loss list. Then, whenever they train the subset of data, they update the loss values of selected samples that are obtained from the training procedure. We discuss the trade-off between latency reduction and obtaining up-to-date information in Section~\ref{sec:latencytradeoff}.
\textit{FedBalancer}{} selects client samples based on the list of sample loss as follows: First, it divides client i's samples into two groups: Under-Threshold ($UT^i$) and Over-Threshold ($OT^i$). Samples that have smaller loss than the loss threshold $lt_R$ are put in $UT^i$ and otherwise in $OT^i$ (Line~\ref{algline:ssm-line8} -~\ref{algline:ssm-line11}). We regard samples in $OT^i$ to be more \textit{important} samples in training the current model and prioritize them in sample selection. We sample $L\cdot p$ samples from $OT^i$ and $L\cdot(1 - p)$ samples from $UT^i$ where $L$ indicates the number of selected samples and $p$ is a parameter in an interval of [0.5, 1.0] (Line~\ref{algline:ssm-linenew} -~\ref{algline:ssm-line18}). The intuition of sampling a portion of data from $UT^i$ is to avoid \textit{catastrophic forgetting}~\cite{kirkpatrick01, yoon01} of the model on \textit{already-learned} samples.
The number of selected samples, $L$, is determined as the number of samples in $OT^i$ ($len(OT^i)$). The loss threshold gradually increases (explained in Section~\ref{sec:ltsm}), which allows clients to efficiently focus on samples with high statistical utility. However, if $S$ is larger than $len(OT^i)$, a client instead uses $S$
to maximize the statistical utility within the deadline (Line~\ref{algline:ssm-linenew}). As \textit{FedBalancer}{} is built on top of Prox~\cite{li01} that allows clients to train less number of epochs, clients with $S$ less than $len(OT^i)$ could still contribute to the model update.
\subsubsection{Loss threshold selection module}
\label{sec:ltsm}
\begin{algorithm}[t]
\caption{$lt$ selection for next ($R+1$)-th round}\label{alg:ps}
\begin{algorithmic}[1]
\Procedure{SelectLossThreshold}{$LLow_{R}$, $LHigh_{R}$, $ltr$} \label{algline:ps-line1}
\State $ll \gets min(LLow_{R})$ \label{algline:ps-line2}
\State $lh \gets mean(LHigh_{R})$ \label{algline:ps-line3}
\State $lt_{R+1} \gets ll + (lh - ll) \cdot ltr$ \label{algline:ps-line4}
\State \textbf{return} $lt_{R+1}$ \label{algline:ps-line5}
\EndProcedure\label{algline:ps-line6}
\end{algorithmic}
\end{algorithm}
The \textit{loss threshold selection module} determines a loss threshold that effectively distinguishes the \textit{important} and \textit{non-important} samples. As the loss distribution of samples changes as FL proceeds, it is essential for the module to be knowledgeable of the current distribution. To respect privacy, the server collects few statistical information from the loss list of clients as a metadata at the end of each round.
Specifically, client $i$ at the $R$-th round provides $LLow^i_R$ and $LHigh^i_R$ values to the server, which indicate the low and high loss value of its current samples, respectively. We use the min loss value as $LLow^i_R$ and use 80\% percentile loss
from the list as $LHigh^i_R$ instead of the max value, as noisy samples could have abnormally high loss~\cite{song01} and make \textit{FedBalancer}{} to misjudge the loss distribution. As such values directly indicate a loss value of a specific sample, we apply Gaussian noise to the values to protect user privacy, as in differential privacy~\cite{lai01, mcmahan02, wei01}. These values from clients get further aggregated on the server into a list as $LLow_R$ and $LHigh_R$. We report the performance of \textit{FedBalancer}{} when different levels of noise is applied in Section~\ref{sec:parametersensitivityanalysis}.
With these metadata, the loss threshold selection module selects a loss threshold as described in Lines~\ref{algline:ps-line1} -~\ref{algline:ps-line6} in Algorithm~\ref{alg:ps}. The module measures the \textit{loss low} value ($ll$) and \textit{loss high} value ($lh$) to estimate the current range of sample loss values of the clients (Lines~\ref{algline:ps-line2} and ~\ref{algline:ps-line3}). The module then outputs the loss threshold of the next round ($R+1$) with variable \textit{loss threshold ratio} ($ltr$), which calculates the linear interpolation between $ll$ and $lh$ (Line~\ref{algline:ps-line4}) as $lt_{R+1} \leftarrow ll + (lh - ll) \cdot ltr$.
The loss threshold ratio $(ltr)$ enables \textit{FedBalancer}{} to start training with all samples and gradually remove \textit{already-learned} samples. \textit{FedBalancer}{} initialize $ltr$ as 0.0 and gradually increases the value by \textit{loss threshold step size} $(lss)$ as shown in Algorithm~\ref{alg:rc}. Note that the \textit{deadline ratio} $(ddlr)$, which controls the deadline of each round (described in Section~\ref{sec:adc}), is also controlled with $ltr$.
To control $ltr$ and $ddlr$, \textit{FedBalancer}{} evaluates the benefit of the current configuration (loss threshold and deadline) at each round based on the statistical utility. For the $R$-th round, it is defined as $U_R \leftarrow \frac{LSum_R}{L_R \cdot ddl_R}$
where $LSum_R$ is the loss sum of the selected samples, $L_R$ is the sum of the number of selected samples, and $ddl_R$ is the chosen deadline for the round. Note that $LSum_R$ and $L_R$ are calculated only from the clients who completed their task and succeeded in sending the model updates. The calculated $U_R$ value is added to the list $U$. \textit{FedBalancer}{} compares the $U_R$ values in the past rounds to the recent rounds (Line~\ref{algline:rc-line5}). If the past rounds have higher value than the recent rounds, \textit{FedBalancer}{} considers the model training to be stable and increases the $ltr$ value by $lss$ to further optimize the training process (Line~\ref{algline:rc-line6}). Otherwise, \textit{FedBalancer}{} decreases the $ltr$ value by $lss$ (Line~\ref{algline:rc-line9}). Note that $ddlr$, which is initialized as 1.0, is controlled in opposite direction with $dss$ (Lines~\ref{algline:rc-line7} and ~\ref{algline:rc-line10}). Such control of $ltr$ and $ddlr$ happens every $w$ round (Line~\ref{algline:rc-line4}).
\begin{algorithm}[t]
\caption{$ltr$ and $ddlr$ control at $R$-th round}\label{alg:rc}
\begin{algorithmic}[1]
\Procedure{Ctrl}{$U$, $LSum_R$, $L_R$, $ddl_R$, $ltr$, $ddlr$, $lss$, $dss$, $w$} \label{algline:rc-line1}
\State $U_R \gets \frac{LSum_R}{L_R \cdot ddl_R}$\label{algline:rc-line2
\State $U.insert(U_R)$ \label{algline:rc-line3}
\If{$R \bmod w \equiv 0$}\label{algline:rc-line4}
\If{$\sum U(R-2w:R-w) > \sum U(R-w:R)$}\label{algline:rc-line5}
\State $ltr \gets $\texttt{min}$(ltr + lss, 1.0)$\label{algline:rc-line6}
\State $ddlr \gets $\texttt{max}$(ddlr - dss, 0.0)$\label{algline:rc-line7}
\Else\label{algline:rc-line8}
\State $ltr \gets $\texttt{max}$(ltr - lss, 0.0)$\label{algline:rc-line9}
\State $ddlr \gets $\texttt{min}$(ddlr + dss, 1.0)$\label{algline:rc-line10}
\EndIf\label{algline:rc-line11}
\EndIf\label{algline:rc-line12}
\State \textbf{return} $ltr$, $ddlr$\label{algline:rc-line13}
\EndProcedure\label{algline:rc-line14}
\end{algorithmic}
\end{algorithm}
\subsubsection{Client selection with sample selection}
\label{sec:clientselection}
Researchers have studied on \textit{how to select a group of clients for a training round} to optimize convergence speed and model performance in heterogeneous FL~\cite{lai01, cho01, cho02}. While these approaches prioritize clients with higher statistical utility from the data, applying them along with \textit{FedBalancer}{} is non-trivial as the samples are dynamically selected with the loss threshold. To address this issue, we propose a new formulation to calculate the statistical utility of a client $i$ along with the sample selection strategy of \textit{FedBalancer}{} as follows:
\[ StatUtil(i) \leftarrow \lvert len(OT^i) \rvert \sqrt{\frac{1}{\lvert len(OT^i) \rvert}\sum_{s \in OT^i}Loss(s)^2} .\]
This is based on the formulation of statistical utility of state-of-the-art client selection method~\cite{lai01}, which we only calculate the statistical utility from the $OT^i$ group. Thus, the sum of loss squares in $OT^i$, $\sum_{s \in OT^i}Loss(s)^2$ and the number of samples in $OT^i$, $len(OT^i)$ are also collected from clients as a differentially-private metadata.
\subsection{Adaptive Deadline Control}
\label{sec:adc}
We explain how \textit{FedBalancer}{} finds an optimal deadline for each round when the clients' training time changes with the sample selection.
\subsubsection{Efficiency of a deadline}
\label{sec:ddle}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/03FedBalancer/adc-1.pdf}
\caption{FEMNIST dataset.}
\label{fig:adc-1}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/03FedBalancer/adc-2.pdf}
\caption{Shakespeare dataset.}
\label{fig:adc-2}
\end{subfigure}
\vspace{-0.6cm}
\caption{Deadline efficiency (DDL-E) evaluation on different deadlines on two FL tasks.}
\vspace{-0.5cm}
\label{fig:adc}
\end{figure}
In order to find the best deadline for each round, we define a metric named \textit{deadline efficiency (DDL-E)} for deadline $t$ as follows:
\[ DDL\text{-}E(t) \leftarrow \frac{\text{\# of completed clients before } t}{t}.\]
Our definition of DDL-E formulates the benefit of using deadline $t$ by measuring the amount of completed clients per time. Finding a deadline with high DDL-E value allows the system to avoid choosing too long or too short deadlines. Setting a long deadline with a large $t$ value would have more completed clients but have low efficiency. On the other hand, configuring an extremely short deadline with a small $t$ would result in almost no completed clients and low efficiency.
To understand how DDL-E is distributed at different deadlines at FL task, we profiled the DDL-E based on a large-scale smartphone datase
~\cite{yang01}. It contains the downlink and uplink network connectivity data and the model training latency data from real-world clients using heterogeneous hardware. We measured the DDL-E for deadlines in range of $[1sec, 1200sec]$ on two FL tasks: FEMNIST and Shakespeare.
Figure~\ref{fig:adc} presents the DDL-E measurements on two FL tasks. From both tasks, we observe that there exists a specific deadline that shows the peak DDL-E value. From the FEMNIST dataset, $t = 172$ seconds shows the max efficiency with DDL-E value of $2.57$, while in the Shakespeare dataset, $t = 157$ seconds shows the max efficiency with DDL-E value of $0.13$. The distribution shape of DDL-E values showing sharp peak around the max value implies that finding an optimal deadline for FL with \textit{FedBalancer}{} could enable more completed clients and improved convergence speed.
\begin{algorithm}[t]
\caption{$ddl$ selection for next ($R+1$)-th round}\label{alg:ps2}
\begin{algorithmic}[1]
\Procedure{FindPeakDDL-E}{$\{DL^i, UL^i, B^i\}_{i=1}^{N}, numEpoch$}\label{algline:ps-line8}
\State $completeTime \gets \emptyset$\label{algline:ps-line9}
\State $DDL\text{-}E \gets \emptyset$ \Comment{initialize a list of DDL-E values}\label{algline:ps-line10}
\State $c \gets 0$\label{algline:ps-line11}
\State $t \gets 1$\label{algline:ps-line12}
\For{$i \gets 1$ to $N$}\label{algline:ps-line13}
\State $T_{network} \gets mean(DL^i) + mean(UL^i)$\label{algline:ps-line14}
\State $T_{train} \gets $\hspace{0.1cm}\texttt{getTrainTime}$(mean(B^i), numEpoch)$\label{algline:ps-line16}
\State $completeTime.insert(T_{network} + T_{train})$\label{algline:ps-line17}
\EndFor\label{algline:ps-line18}
\While{$c \neq N$}\label{algline:ps-line19}
\State $c \gets 0$\label{algline:ps-line20}
\For{$i \gets 1$ to $N$}\label{algline:ps-line21}
\If{$t \geq completeTime(i)$}\label{algline:ps-line22}
\State $c \gets c + 1$\label{algline:ps-line23}
\EndIf\label{algline:ps-line24}
\EndFor\label{algline:ps-line25}
\State $DDL\text{-}E.insert(\frac{c}{t})$\label{algline:ps-line26}
\State $t \gets t + 1$\label{algline:ps-line27}
\EndWhile\label{algline:ps-line28}
\State \textbf{return} \texttt{maxIndex}$(DDL\text{-}E)$\label{algline:ps-line29}
\EndProcedure\label{algline:ps-line30}
\Procedure{SelectDeadline}{$\{DL^i, UL^i, B^i\}_{i=0}^{N}, E, ddlr$}\label{algline:ps-line32}
\State $dl \gets $\textsc{FindPeakDDL-E}$(\{DL^i, UL^i, B^i\}_{i=1}^{N}, 1)$\label{algline:ps-line33}
\State $dh \gets $\textsc{FindPeakDDL-E}$(\{DL^i, UL^i, B^i\}_{i=1}^{N}, E)$\label{algline:ps-line34}
\State $ddl_{R+1} \gets dl + (dh - dl) \cdot ddlr$\label{algline:ps-line35}
\State \textbf{return} $ddl_{R+1}$\label{algline:ps-line36}
\EndProcedure\label{algline:ps-line37}
\end{algorithmic}
\end{algorithm}
We designed \textit{FedBalancer}{} to select a deadline based on finding the best DDL-E values. Lines~\ref{algline:ps-line8}- ~\ref{algline:ps-line30} of Algorithm~\ref{alg:ps2} shows how \textit{FedBalancer}{} finds the deadline with max DDL-E value. This is based on the clients' hardware capability information: for client $i$, they are DownLink speed ($DL^i$), UpLink speed ($UL^i$), and Batch training latency ($B^i$) (Line ~\ref{algline:ps-line8}). When there are a total of $N$ clients involved, we assume clients have already profiled $\{DL^i, UL^i\}_{i=1}^{N}$ before FL and \textit{FedBalancer}{} collects $\{B^i, I^i\}_{i=1}^{N}$ during FL. \textit{FedBalancer}{} first iterates over N clients to measure their completion time with the mean value of $\{DL^i, UL^i, B^i\}_{i=1}^{N}$ (Lines~\ref{algline:ps-line13}- ~\ref{algline:ps-line18}). \textit{FedBalancer}{} then measures the DDL-E from the smallest deadline $t$ (we use $t=1sec$)
and increments $t$ until all clients complete before $t$ (Lines~\ref{algline:ps-line19}-~\ref{algline:ps-line28}) and outputs the deadline with the max DDL-E value (Line~\ref{algline:ps-line29}).
As different subset of clients are selected for each round, \textit{FedBalancer}{} finds the max DDL-E value among the selected clients of each round. Moreover, as clients use different size of training data with sample selection, we estimate the training time of client $i$ (\texttt{getTrainTime} from Line~\ref{algline:ps-line16}) as follows:
\[ \frac{len(OT^i) - 1 }{batchSize} \cdot mean(B^i) \cdot numEpoch. \]
We use the length of $OT^i$ in measuring the training time of a client to reflect the number of samples being selected. This allows \textit{FedBalancer}{} to reliably find the deadline with the best DDL-E along with the sample selection strategy.
\subsubsection{Deadline selection module}
The \textit{deadline selection module} of \textit{FedBalancer}{} determines the deadline that optimizes the training process and convergence speed. The module selects the deadline as shown in Lines~\ref{algline:ps-line32} -~\ref{algline:ps-line37} of Algorithm~\ref{alg:ps}. The module measures \textit{deadline low} value $dl$ and \textit{deadline high} value $dh$, which are the deadlines with the max DDL-E value when clients are running 1 epoch and $E$ epochs, respectively (Lines~\ref{algline:ps-line33} and ~\ref{algline:ps-line34}).
The module then outputs the deadline of the next round (R+1) with parameter \textit{deadline ratio} ($ddlr$) that calculates the linear interpolation between $dl$ and $dh$ (Line~\ref{algline:ps-line35}) as $ddl_{R+1} \leftarrow dl + (dh - dl) \cdot ddlr$.
The reason of selecting a value between $dl$ and $dh$ is because \textit{FedBalancer}{} is built on top of Prox~\cite{li01} that allows clients to train various number of epochs within the deadline. With an aim to optimize the training efficiency, \textit{FedBalancer}{} initially configures $ddlr$ as 1.0 and gradually decreases the value by the parameter $dss$, as explained in Section~\ref{sec:ltsm}.
\subsection{Collaboration with FL Methods}
One advantage of \textit{FedBalancer}{} is its applicability to orthogonal FL approaches that do not perform sample selection on clients or control the deadline. Applying \textit{FedBalancer}{} could be achieved by simply adding the sample selection and deadline control strategies on top of other methods. We demonstrate the collaboration capability of \textit{FedBalancer}{} with its implementation on top of three existing FL methods and improved performances in Section~\ref{sec:evaluationcollaboration}.
Some recent FL approaches, such as Oort~\cite{lai01}, use one batch instead of the full dataset for training in each local epoch, making them non-trivial to be directly integrated with \textit{FedBalancer}{}. To address this issue,
we propose \textit{OortBalancer}, which
is built on top of \textit{Oort}, where the sample selection strategy of \textit{FedBalancer}{} is adopted with one adjustment: $L$ from Algorithm~\ref{alg:ssm} is fixed to the batch size. Intuitively, while \textit{Oort} trains one randomly selected batch for each local epoch, \textit{OortBalancer} selects samples for one batch that focuses on more important samples and thereby optimizes the training process. We demonstrate the performance of \textit{OortBalancer} in Section~\ref{sec:evaluationcollaboration} as one of the three examples of \textit{FedBalancer}{} collaboration.
\section{Evaluation}
\label{sec:evaluation}
We evaluate \textit{FedBalancer}{} to answer the following key questions: 1) How much performance improvement (in terms of time-to-accuracy and model accuracy) does \textit{FedBalancer}{} achieves over existing FL methods? 2) How sensitive is \textit{FedBalancer}{} with different choice of parameters?
3) How much performance improvement does each component of \textit{FedBalancer}{} achieves? 4) How does \textit{FedBalancer}{} perform when it jointly operates with orthogonal FL approaches?
\subsection{Experimental Setup}
\label{sec:experimentalsetup}
\begin{table*}[t]
\caption{\rev{Speedup and accuracy on five datasets with the real-world user data.}}
\label{table:speedup-and-accuracy}
\begin{center}
\scalebox{0.82}{
\begin{tabular}{ccccccccccc}
\toprule
\multicolumn{1}{c}{Task}&\multicolumn{4}{c}{CV}&\multicolumn{4}{c}{NLP}&\multicolumn{2}{c}{HAR} \\
\cmidrule(lr){2-5}
\cmidrule(lr){6-9}
\cmidrule(lr){10-11}
\multicolumn{1}{c}{Dataset} &\multicolumn{2}{c}{FEMNIST}&\multicolumn{2}{c}{Celeba}&\multicolumn{2}{c}{Reddit}&\multicolumn{2}{c}{Shakespeare}&\multicolumn{2}{c}{UCI-HAR} \\ \midrule
Methods&Speedup&Acc.&Speedup&Acc.&Speedup&Acc.&Speedup&Acc.&Speedup&Acc. \\
\midrule
FedAvg+1$T$&1.00$\pm$0.00&.796$\pm$.007&0.97$\pm$0.05&.851$\pm$.005&0.08$\pm$0.12&.090$\pm$.001&0.24$\pm$0.22&.399$\pm$.020&0.49$\pm$0.36&.814$\pm$.029 \\
FedAvg+2$T$&0.59$\pm$0.01&.763$\pm$.009&0.59$\pm$0.04&.824$\pm$.010&0.58$\pm$0.07&.104$\pm$.000&0.39$\pm$0.10&.373$\pm$.046&0.68$\pm$0.07&.819$\pm$.014 \\
FedAvg+SPC&0.71$\pm$0.02&.777$\pm$.004&0.80$\pm$0.14&.829$\pm$.013&0.87$\pm$0.08&.112$\pm$.001&0.56$\pm$0.10&.416$\pm$.017&0.91$\pm$0.13&.840$\pm$.016 \\
FedAvg+WFA&0.33$\pm$0.20&.594$\pm$.205&0.54$\pm$0.04&.813$\pm$.020&1.00$\pm$0.00&.113$\pm$.001&1.00$\pm$0.00&.439$\pm$.017&0.68$\pm$0.07&.819$\pm$.014 \\
Prox+1$T$&0.99$\pm$0.02&.795$\pm$.008&1.05$\pm$0.03&.855$\pm$.002&2.87$\pm$0.43&.121$\pm$.005&1.14$\pm$0.16&.476$\pm$.003&0.96$\pm$0.09&.849$\pm$.008 \\
Prox+2$T$&0.65$\pm$0.02&.767$\pm$.006&0.75$\pm$0.04&.833$\pm$.010&3.63$\pm$0.43&.127$\pm$.005&1.00$\pm$0.14&.457$\pm$.003&0.68$\pm$0.07&.819$\pm$.014 \\
SampleSelection&1.01$\pm$0.02&.799$\pm$.006&1.03$\pm$0.06&.852$\pm$.002&1.52$\pm$0.38&.118$\pm$.001&0.85$\pm$0.24&.439$\pm$.021&0.90$\pm$0.05&.845$\pm$.008 \\
\midrule
\textit{FedBalancer}{}&1.57$\pm$0.03&.815$\pm$.006&1.43$\pm$0.07&.862$\pm$.006&4.48$\pm$0.23&.146$\pm$.001&1.20$\pm$0.10&.489$\pm$.004&1.56$\pm$0.28&.855$\pm$.034\\
\textit{FedBalancer}{}-A&1.60$\pm$0.06&\textbf{.820$\pm$.003}&1.52$\pm$0.07&\textbf{.873$\pm$.004}&4.08$\pm$0.76&\textbf{.154$\pm$.000}&1.31$\pm$0.28&\textbf{.505$\pm$.004}&1.39$\pm$0.43&\textbf{.893$\pm$.008}\\
\textit{FedBalancer}{}-S&\textbf{1.71$\pm$0.01}&.819$\pm$.002&\textbf{1.67$\pm$0.04}&.859$\pm$.006&\textbf{4.99$\pm$0.42}&.148$\pm$.003&\textbf{1.83$\pm$0.14}&.488$\pm$.001&\textbf{1.98$\pm$0.48}&.863$\pm$.010\\
\bottomrule
\end{tabular}
}
\end{center}
\end{table*}
\textbf{Implementation.} We developed \textit{FedBalancer}{} on FLASH~\citep{yang01}, a heterogeneity-aware benchmarking framework for FL based on LEAF~\citep{caldas01}. FLASH provides a simulation of heterogeneous computational capabilities and network connectivity from a large-scale real-world trace dataset collected over 136k smartphones that span one thousand types of devices. We implement \textit{FedBalancer}{} with the state-of-the-art FL aggregation method Prox~\cite{li01}. Our implementation is based on Python 3.6 and TensorFlow 1.14 with 2,062 lines of code on top of FLASH.
\rev{The source code of our \textit{FedBalancer}{} implementation are available at \url{https://github.com/jaemin-shin/FedBalancer}.}
\textbf{Datasets.} To simulate FL tasks in our evaluation, we use five datasets that contain data generated by real-world users, which are categorized in three different domains as follows:
\begin{itemize}[topsep=3pt, align=left, labelwidth=10pt, leftmargin=15pt]
\item {\textit{Computer Vision (CV)}}: For CV, we evaluated \textit{FedBalancer}{} on two image recognition datasets: FEMNIST~\cite{cohen01} and Celeba~\cite{liu01}. FEMNIST dataset contains images of handwritten digits and characters from 712 users with total 157,132 samples. Celeba dataset contains face attributes of 915 users with 19,923 samples. We use CNN models for both datasets as in previous work~\cite{yang01}.
\item {\textit{Natural Language Processing (NLP)}}: We evaluate \textit{FedBalancer}{} on two NLP tasks each on different dataset: next-word prediction on Reddit~\cite{caldas01} dataset and next-character prediction on Shakespeare~\cite{shakespeare01} dataset. The Reddit dataset contains reddit posts from 813 users with 32,680 samples, and the Shakespeare dataset contains 845,231 samples separated into 171 users. We use LSTM models for both datasets as in previous work~\cite{yang01}.
\item {\textit{Human Activity Recognition (HAR)}}: We use UCI-HAR dataset~\cite{anguita01} that contains six types of activity data on accelerometer and gyroscope from 30 users with 10,299 samples. As in previous work, we use \rev{CNN models~\cite{ek01}}.
\end{itemize}
\textbf{Metrics.} As in the previous work~\citep{lai01} that evaluated on heterogeneous FL clients, we mainly evaluate \textit{time-to-accuracy} performance and \textit{final model accuracy} on the experiments. Here, the \textit{time-to-accuracy} performance indicates the wall clock time that is required for a model training task to reach an accuracy target. We repeat each experiment for three times with different random seeds and report the average and standard deviation of these evaluation metrics.
\textbf{Baselines.} We use the following list of approaches as a baseline to compare with \textit{FedBalancer}{} in our evaluation:
\begin{itemize}[topsep=3pt, align=left, labelwidth=10pt, leftmargin=15pt]
\item {\textit{Aggregation algorithms}}: We use FedAvg~\cite{mcmahan01} and Prox~\cite{li01}, the most widely used aggregation algorithms for FL. Prox offers an optimizer with convergence guarantee on heterogeneous clients, while allowing clients to train various number of local epochs within the deadline. For the $\mu$ parameter of Prox, we tested $\mu$ in $\{0.0, 0.001, 0.01,$ $0.1, 1.0\}$ as suggested by the paper and pick one with the best final accuracy. All the datasets showed the best accuracy with $0.0$ except Celeba with $0.1$.
\item {\textit{Deadline configuration methods}}: We use four different deadline configuration methods in the evaluation. We configure two different fixed deadlines as a baseline, which are $1T$ and $2T$. Before the training begins, we sample the round completion time of all participating clients and calculate the mean value as $T$. Thus, $2T$ uses double of that mean value as a fixed deadline. We also adopt SmartPC (SPC)~\cite{li04}, and implemented it to involve the certain portion $U_{required}$ of users to complete at a training round. As suggested in the paper, we use 80\% for $U_{required}$. Lastly, we use a method that waits every client to finish a round, which we named as WaitForAll (WFA).
\item \rev{{\textit{Sample selection method}}: We implement the baseline sample selection method that is a combination of the following: (1) We determined \textit{how many} samples to select based on FedSS~\cite{cai01}, which controls the training dataset size on clients with larger datasets for each training round. (2) There were several approaches that propose \textit{which} samples to select based on \textit{loss}~\cite{loshchilov01, schaul01}, \textit{gradient}~\cite{alain01}, or \textit{gradient norm upper bound}~\cite{katharopoulos01, li08} of samples. As in \textit{FedBalancer}{}, we use \textit{loss} to select samples for the baseline experiment.}
\end{itemize}
For baseline experiments, we use the combination of the aggregation algorithms and the deadline configuration methods of above.
We do not test SPC with Prox as it is nontrivial to accept stragglers' model update with less number of epochs when $U_{required}$ users complete a training round. Moreover, we do not test WFA with Prox as it is identical with FedAvg when $\mu = 0$, which is used by most datasets. \rev{We test the sample selection method with the best performing deadline configuration methods with Prox, which is Prox+$2T$ for Reddit dataset and Prof+$1T$ otherwise.}
\begin{table*}[t]
\caption{\rev{$\{w, lss, dss, p\}$ parameters from the \textit{FedBalancer}{} experiments in Table~\ref{table:speedup-and-accuracy}. Parameters with forward slashes indicate the different parameters from the multiple runs of experiments with three different random seeds.}}
\vspace{-0.2cm}
\label{table:parameters}
\begin{center}
\scalebox{0.815}{
\begin{tabular}{c|cccc|cccc|cccc}
\toprule
\multicolumn{1}{c}{}&\multicolumn{4}{c}{\textit{FedBalancer}{}}&\multicolumn{4}{c}{\textit{FedBalancer}{}-A}&\multicolumn{4}{c}{\textit{FedBalancer}{}-S}\\
\cmidrule(lr){2-5}
\cmidrule(lr){6-9}
\cmidrule(lr){10-13}
\multicolumn{1}{c}{}&\multicolumn{1}{c}{$w$}&\multicolumn{1}{c}{$lss$}&\multicolumn{1}{c}{$dss$}&\multicolumn{1}{c}{$p$}&\multicolumn{1}{c}{$w$}&\multicolumn{1}{c}{$lss$}&\multicolumn{1}{c}{$dss$}&\multicolumn{1}{c}{$p$}&\multicolumn{1}{c}{$w$}&\multicolumn{1}{c}{$lss$}&\multicolumn{1}{c}{$dss$}&\multicolumn{1}{c}{$p$}\\
\midrule
FEMNIST&&&&&20&0.01/0.1/0.1&0.10&1.00&20&0.05/0.05/0.01&0.25/0.1/0.05& 1.00\\
Celeba&&&&&20/5/5&0.05/0.1/0.01&0.1/0.1/0.05&1.00/0.75/1.00&20/20/5&0.01/0.01/0.1&0.1/0.1/0.25&1.00/1.00/0.75\\
Reddit&20&0.05&0.05&1.00&5/20/20&0.01&0.1/0.25/0.1&1.00&5&0.1& 0.25/0.1/0.05&0.75\\
Shakespeare&&&&&5&0.01/0.05/0.01&0.1/0.25/0.25&1/0.75/1&5/20/5&0.05/0.01/0.1&0.1/0.05/0.1&0.75\\
UCI-HAR&&&&&5/5/20&0.1/0.01/0.05&0.1/0.1/0.25&1/1/0.75&5&0.1/0.1/0.05&0.25&0.75/0.75/1\\
\bottomrule
\end{tabular}
}
\end{center}
\vspace{-0.3cm}
\end{table*}
\begin{table*}[t]
\vspace{0.2cm}
\caption{\rev{Speedup and accuracy on different choice of \textit{FedBalancer}{} parameters.}}
\vspace{-0.2cm}
\label{table:paramter-analysis}
\begin{center}
\scalebox{0.83}{
\begin{tabular}{cccc|ccc|ccc|cc}
\toprule
\multicolumn{2}{c}{\multirow{2}{*}{Parameter}}&\multicolumn{2}{c}{$w$}&\multicolumn{3}{c}{$lss$}&\multicolumn{3}{c}{$dss$}&\multicolumn{2}{c}{$p$} \\
\cmidrule(lr){3-4}
\cmidrule(lr){5-7}
\cmidrule(lr){8-10}
\cmidrule(lr){11-12}
&&$5$&$20$&$0.01$&$0.05$&$0.10$&$0.05$&$0.10$&$0.25$&$0.75$&$1.00$ \\ \midrule
\multirow{2}{*}{FEMNIST}&Speedup&1.39$\pm$0.16&\textbf{1.52$\pm$0.15}&1.37$\pm$0.15&1.49$\pm$0.18&\textbf{1.50$\pm$0.14}&\textbf{1.54$\pm$0.07}&1.47$\pm$0.16&1.35$\pm$0.18&1.41$\pm$0.15&\textbf{1.50$\pm$0.17} \\
&Accuracy&.813$\pm$.005&\textbf{.816$\pm$.004}&.813$\pm$.004&.815$\pm$.006&\textbf{.816$\pm$.006}&\textbf{.817$\pm$.003}&.814$\pm$.004&.812$\pm$.005&.814$\pm$.004&\textbf{.815$\pm$.005} \\
\midrule
\multirow{2}{*}{Reddit}&Speedup&3.40$\pm$0.53&\textbf{3.70$\pm$0.23}&3.60$\pm$0.29&\textbf{3.61$\pm$0.31}&3.44$\pm$0.61&\textbf{3.66$\pm$0.44}&3.60$\pm$0.33&3.40$\pm$0.48&\textbf{3.71$\pm$0.30}&3.39$\pm$0.49 \\
&Accuracy&.146$\pm$.007&\textbf{.147$\pm$.003}&\textbf{.150$\pm$.002}&.147$\pm$.002&.143$\pm$.008&\textbf{.147$\pm$.004}&.146$\pm$.007&.147$\pm$.005&\textbf{.149$\pm$.002}&.144$\pm$.007 \\
\bottomrule
\end{tabular}
}
\end{center}
\end{table*}
\textbf{Method.} We first ran FedAvg+$1T$ on five datasets until convergence, with the number of rounds that are suggested by previous works~\cite{caldas01, yang01, ek01, sozinov01}: 1000, 100, 600, 40, and 50 rounds for FEMNIST, Celeba, Reddit, Shakespeare, and UCI-HAR. Based on the user trace data of FLASH, we measured the wall clock time which FedAvg+$1T$ ran for each dataset, and ran experiments with other baselines and \textit{FedBalancer}{} until the same wall clock time. Among the four FedAvg baselines, we pick the one with best accuracy for each dataset and configure it as the target accuracy of that task. Then, we measured the speedup of other methods in achieving the target accuracy and their final model accuracy achieved within the same wall clock time.
We tested the following combination of parameters for \textit{FedBalancer}{}: $w$ in $\{5,20\}$, $lss$ in $\{0.01, 0.05, 0.10\}$, $dss$ in $\{$0.05, 0.10, 0.25$\}$, and $p$ in $\{0.75, 1.00\}$. Among the results, we report the performance of \textit{FedBalancer}{} with only one set of parameter: $\{w, lss, dss, p\} = \{20, 0.05, 0.05, 1.00\}$. This is a set of parameters that we recommend FL developers to try with their task when they are not knowledgeable of which parameter performs the best. We used this parameter for all the experiments in Section~\ref{sec:evaluation}. In Section~\ref{sec:speedupandaccuracy}, we also report the parameter set with the best final accuracy for each dataset
as \textit{FedBalancer}{}-A \rev{and the best speedup as \textit{FedBalancer}{}\textit{-S}
to demonstrate the maximum performance \textit{FedBalancer}{} could achieve.}
\textbf{Other configurations.} As in previous study~\cite{yang01}, we use batch size of 100 for Shakespeare and 10 for rest of the datasets. We select 100 clients at each round for datasets with more than 500 users, and otherwise select 10 users for Shakespeare and 5 users for UCI-HAR. We configured the clients to train five local epochs per round. We use learning rate of 0.001 for FEMNIST and Celeba, 2 for Reddit, 0.8 for Shakespeare, and 0.005 for UCI-HAR.
\subsection{Speedup and Accuracy on Five FL Tasks}
\label{sec:speedupandaccuracy}
Table~\ref{table:speedup-and-accuracy} shows the performance of \textit{FedBalancer}{} on five datasets compared with the baseline methods, and Table~\ref{table:parameters} shows the parameters used for \textit{FedBalancer}{}, \textit{FedBalancer}{}-A, and \rev{\textit{FedBalancer}{}-S} for each dataset. We observed that \textit{FedBalancer}{} shows improved time-to-accuracy performance over the baselines on every dataset: \textit{FedBalancer}{} reaches the target accuracy \rev{1.43$\sim$1.57$\times$} faster than the FedAvg-based baselines on CV datasets, \rev{1.36$\sim$1.58$\times$} faster than the Prox-based baselines\rev{, and 1.39$\sim$1.55$\times$ faster than the \textit{SampleSelection} baseline}. \textit{FedBalancer}{} achieves \rev{1.20$\sim$4.48$\times$} speedup and \rev{1.05$\sim$1.23$\times$} speedup compared with the FedAvg and Prox-based methods respectively\rev{, while achieving 1.41$\sim$2.95$\times$ speedup over the \textit{SampleSelection} baseline} on NLP datasets. \textit{FedBalancer}{} achieves speedup of \rev{1.56$\times$, 1.63$\times$, and 1.73$\times$} compared with FedAvg, Prox, and \textit{SampleSelection} baselines on a HAR dataset, respectively.
We noticed that \textit{FedBalancer}{} consistently shows high time-to-accuracy performance on all datasets, while the performance of baselines was inconsistent across datasets. For example, among FedAvg-based baselines, FedAvg+$1T$ shows the best time-to-accuracy performance on CV tasks but shows extremely low performance on NLP tasks. In contrast, FedAvg+WFA shows the best time-to-accuracy performance on NLP tasks among FedAvg-based baselines but shows low performance on CV tasks. In UCI-HAR, FedAvg+$1T$ and FedAvg+SPC resulted in the best performance at different experiments with different random seeds. Prox+$1T$ shows the best performance among baselines on Celeba, Shakespeare, and UCI-HAR, but shows worse performance than \rev{\textit{SampleSelection}} and Prox+$2T$ on FEMNIST and Reddit respectively.
We observed that \textit{FedBalancer}{} achieves this improvement of time-to-accuracy performance without sacrificing model accuracy; in terms of the final model accuracy, \textit{FedBalancer}{} showed improvement over the baselines on all datasets. Compared to the FedAvg-based methods, \textit{FedBalancer}{} achieved \rev{1.1$\sim$5.0\%} accuracy improvement on different datasets. \textit{FedBalancer}{} achieved \rev{0.6$\sim$3.2\%} and \rev{1.0$\sim$5.0\%} accuracy improvement over Prox-based methods and the \textit{Sampleselection} baseline.
\textit{FedBalancer}{}-A, which marks the best accuracy of \textit{FedBalancer}{}, also shows time-to-accuracy performance improvement over baselines at all datasets --- showing further improvement in speedup on FEMNIST, Celeba, and Shakespeare. \rev{On the other hand, \textit{FedBalancer}{}-S, which reports the best time-to-accuracy performance of \textit{FedBalancer}{}, shows accuracy improvement over baselines at all datasets, with further improvement in accuracy on FEMNIST, Reddit, and UCI-HAR.} This suggests that the performance of \textit{FedBalancer}{} could be further improved with a carefully selected parameter for an FL task, while the fixed parameter set we recommend still shows improved performance.
\subsection{Parameter Sensitivity Analysis}
\label{sec:parametersensitivityanalysis}
Table~\ref{table:paramter-analysis} shows the time-to-accuracy performance and final accuracy of \textit{FedBalancer}{} on different choice of parameters $\{w, lss, dss, p\}$. For each type of parameter, we fixed it to a certain value and averaged the performance from experiments with different combination of other parameters. We used speedup compared to the best FedAvg-based baseline to measure the time-to-accuracy performance. We chose FEMNIST and Reddit to explore the effect of different parameters at different domains of FL tasks (CV and NLP).
For $w$, \textit{FedBalancer}{} shows similar final accuracy (\rev{81.3\% and 81.6\%}) performance on both of the candidate values on FEMNIST, but reports better time-to-accuracy performance (\rev{1.52$\times$ over 1.39$\times$}) with $w=20$. On Reddit, $w=20$ showed better speedup (\rev{3.71$\times$ over 3.42$\times$}) and accuracy (\rev{14.8\% over 13.3\%}). In terms of $lss$, \textit{FedBalancer}{} achieves faster training and higher accuracy when $lss\geq0.05$ on FEMNIST, but achieves better performance when $lss\leq0.05$ on Reddit. With $dss$, \textit{FedBalancer}{} performs the better when $dss\leq0.1$ on both datasets, where both showed the best time-to-accuracy and accuracy with smaller $dss=0.05$. \rev{Lastly, \textit{FedBalancer}{} performs better with $p=1.00$ on FEMNIST but reports better performance with $p=0.75$ on Reddit. For the different trend of performance with $lss$ and $p$ parameters on two datasets, we suspect that Reddit requires more rounds with data sampled from full dataset in the early stage of training to achieve better performance, and big $lss$ and $p=1.00$ performs worse as it might quickly remove the low-loss samples from training. On the other hand, FEMNIST training performs better with high-loss samples from the early stage of training.} This suggests that the best set of parameters could be selected based on how the training at an FL task proceeds.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/undernoise-femnist.pdf}
\caption{On Different Level of Differential Privacy.}
\label{fig:differentialprivacy}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/pipelining-femnist.pdf}
\caption{With and Without the Forward Pass at Each Round.}
\label{fig:forwardpass}
\end{subfigure}
\caption{Performance evaluation of \textit{FedBalancer}{} with different options on FEMNIST.}
\vspace{-0.2cm}
\label{fig:discussion}
\vspace{-0.2cm}
\end{figure}
Other than the algorithm parameter of \textit{FedBalancer}{}, we study the effect of different level of noise on differential privacy that we applied to mask the metadata shared by the clients. Our implementation of differential privacy is based on the previous work~\cite{lai01}, adding the noise drawn from Gaussian distribution on the metadata, with the mean as zero and the standard deviation as Noise Factor (NF). Figure~\ref{fig:differentialprivacy} shows the effect of different NFs on the performance of \textit{FedBalancer}{}. We observe that the performance of \textit{FedBalancer}{} degrades as the NF increases, as NF 0.0 achieves 81.9\% accuracy but NF 0.5 and 5.0 each achieves 81.4\% and 81.1\% accuracy. However, NF 0.5 and 5.0 still achieves better time-to-accuracy performance and accuacy compared to Prox, while NF of 5.0 is considered to be very large noise~\cite{abadi01}. This result implies that \textit{FedBalancer}{} could achieve performance improvement over the baselines while applying the differential privacy.
\subsection{Effect of \textit{FedBalancer}{} Components}
\label{sec:ablationstudy}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/ablation-femnist.pdf}
\caption{FEMNIST dataset.}
\label{fig:ablation-femnist}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/ablation-reddit.pdf}
\caption{Reddit dataset.}
\label{fig:ablation-reddit}
\end{subfigure}
\caption{Performance breakdown of \textit{FedBalancer}{} into Sample Selection (SS) and Deadline Control (DC).}
\vspace{-0.2cm}
\label{fig:ablation}
\end{figure}
We conducted an experiment to understand the performance brought by each component of \textit{FedBalancer}{}: Sample Selection (SS) and Deadline Control (DC). As \textit{FedBalancer}{} is built on top of Prox~\cite{li01}, we add the components one by one to observe how the performance changes when each component is introduced.
Figure~\ref{fig:ablation} reports the result of the experiment on FEMNIST and Reddit dataset. On FEMNIST dataset, we observe that the performance drops when SS is introduced, but gets further improved when DC is added. On Reddit dataset, however, the accuracy escalates as each component is added. The reason of performance degradation on FEMNIST with SS is due to the reduced statistical utility trained at each round with the same deadline. In contrast on Reddit dataset, the performance improved with SS as it allowed more clients to successfully send their model updates within the deadline with selected client data. Moreover, we suspect that SS has brought more performance improvement on Reddit than FEMNIST by effectively selecting more important samples and the clients that have such data. The result with DC on both dataset shows its effectiveness with SS in time-to-accuracy performance improvement.
\subsection{Collaboration with FL Algorithms}
\label{sec:evaluationcollaboration}
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/collaboration-oort.pdf}
\caption{Oort~\cite{lai01}.}
\label{fig:collaboration-oort}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/collaboration-qffl.pdf}
\caption{q-FFL~\cite{li05}.}
\label{fig:collaboration-qffl}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/collaboration-su-aftermobisys.pdf}
\caption{Structured Updates~\cite{konevcny01}.}
\label{fig:collaboration-su}
\end{subfigure}
\vspace{-0.2cm}
\caption{Collaboration of \textit{FedBalancer}{} with three FL algorithms on FEMNIST dataset.}
\vspace{-0.2cm}
\label{fig:collaboration}
\end{figure*}
To demonstrate the applicability of \textit{FedBalancer}{} on orthogonal FL algorithms, we implement \textit{FedBalancer}{} on top of three widely used FL approaches from different categories: \textit{Oort}~\cite{lai01} as a client selection algorithm, \textit{q-FFL}~\cite{li05} as an aggregation algorithm, and \textit{Structured Updates}~\cite{konevcny01} as a gradient compression algorithm.
Figure~\ref{fig:collaboration} reports the experiment result, which we observe improvement in time-to-accuracy performance and model accuracy from all three cases. Collaboration with \textit{FedBalancer}{} achieved 1.84$\times$, 1.19$\times$, and 1.31$\times$ speedup, while achieving 2.6\%, 2.5\%, 1.7\% accuracy improvement on three algorithms respectively. These results suggest that \textit{FedBalancer}{} could be implemented on top of various advanced FL algorithms to achieve further performance improvement.
\begin{table}[t]
\caption{\rev{Android devices for the testbed experiments.}}
\vspace{-0.2cm}
\label{table:android}
\begin{center}
\scalebox{0.83}{
\begin{tabular}{ccccc}
\toprule
Builder&Year&Device&Processor&Quantity\\
\midrule
\multirow{5}{*}{Google}&2016&Pixel&Snapdragon 821&2\\
&2017&Pixel 2&Snapdragon 835&1\\
&2017&Pixel 2 XL&Snapdragon 835&2\\
&2018&Pixel 3&Snapdragon 845&1\\
&2020&Pixel 5&Snapdragon 765G&1\\
\midrule
\multirow{3}{*}{Samsung}&2016&Galaxy S7&Exynos 8890&1\\
&2017&Galaxy J7&Exynos 7870&1\\
&2019&Galaxy Fold&Snapdragon 855&1\\
\midrule
\multirow{2}{*}{Huawei}&2015&Nexus 6P&Snapdragon 810&3\\
&2018&P20 Lite&Kirin 659&1\\
\midrule
\multirow{1}{*}{Motorola}&2014&Nexus 6&Snapdragon 805&2\\
\midrule
\multirow{1}{*}{LG}&2015&Nexus 5X&Snapdragon 808&4\\
\midrule
\multirow{1}{*}{Essential}&2017&Essential Phone&Snapdragon 835&1\\
\bottomrule
\end{tabular}
}
\end{center}
\vspace{-0.3cm}
\end{table}
\subsection{\rev{Testbed Experiments with Android Clients}}
\label{sec:realevaluation}
\begin{table}[t]
\vspace{0.2cm}
\caption{\rev{Results from the testbed experiments.}}
\vspace{-0.2cm}
\label{table:testbedresults}
\begin{center}
\scalebox{0.85}{
\begin{tabular}{ccc}
\toprule
Method&Speedup&Accuracy\\
\midrule
FedAvg+1T&0.99$\pm$0.03&.852$\pm$.020\\
FedAvg+2T&0.75$\pm$0.30&.800$\pm$.034\\
FedAvg+SPC&0.61$\pm$0.35&.849$\pm$.013\\
FedAvg+WFA&0.92$\pm$0.07&.846$\pm$.007\\
FedProx+1T&1.03$\pm$0.23&.860$\pm$.007\\
FedProx+2T&0.90$\pm$0.20&.860$\pm$.013\\
SampleSelection&0.99$\pm$0.13&.846$\pm$.014\\
\midrule
\textit{FedBalancer}{}&\textbf{1.33$\pm$0.08}&\textbf{.885$\pm$.017}\\
\bottomrule
\end{tabular}
}
\end{center}
\vspace{-0.3cm}
\end{table}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/pixel5-latency.pdf}
\caption{\rev{Google Pixel 5.}}
\label{fig:pixel5}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.235\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/05Evaluation/huawei-latency.pdf}
\caption{\rev{Huawei P20 Lite.}}
\label{fig:p20lite}
\end{subfigure}
\vspace{-0.4cm}
\caption{\rev{On-device training latency (train) and communication latency (network) on different devices.}}
\vspace{-0.2cm}
\label{fig:devicelatency}
\end{figure}
\rev{We further conducted experiments on Android clients to understand the effectiveness of \textit{FedBalancer}{} on real hardware devices. We implemented the server using Flower~\cite{beutel01}, which is an open-source FL framework that communicates with clients via gRPC (Google Remote Procedure Call) and Protocol Buffers~\cite{protocolbuffers}. We used Ubuntu 18.04 server with Intel Xeon Gold 6254 Processor @ 3.10GHz and 512GB RAM. On-device training on Android devices were implemented with the model personalization feature of Tensorflow Lite~\cite{tflite}. We used UCI-HAR dataset and used the same experimental configuration as illustrated in Section~\ref{sec:experimentalsetup}.}
\rev{For 21 client devices in UCI-HAR experiments, we used 13 different Android models to simulate the hardware heterogeneity in the real-world as illustrated in Table~\ref{table:android}. We placed the devices in an office room of a laboratory building, where the devices were connected to a campus Wi-Fi. To configure fixed deadlines for the deadline configuration methods $1T$ and $2T$, we sampled the training round completion time of each device 10 times before the experiments.}
\rev{Table~\ref{table:testbedresults} shows the performance of \textit{FedBalancer}{} compared with the baseline methods in our testbed experiments. \textit{FedBalancer}{} shows higher time-to-accuracy performance and final model accuracy over the baselines. \textit{FedBalancer}{} showed 1.34$\times$ speedup and 3.3\% accuracy improvement over the FedAvg-based baselines. \textit{FedBalancer}{} also achieved the target accuracy 1.29$\times$ and 1.34$\times$ faster than Prox and \textit{SampleSelection} baselines, with 2.5\% and 3.9\% accuracy improvements.}
\rev{One of the unique challenges we discovered in the testbed experiments was dynamically changing round completion times of the client devices. To further understand the cause of such a phenomenon, we separately sampled the on-device training latency and the communication latency on the client devices at a training round, as shown in Figure~\ref{fig:devicelatency}. Compared with the on-device training latency that remained constant at different rounds, the communication latency showed high variability at each round, showing up to 3.91$\times$ increase over the mean communication latency. Our testbed experiment had higher variability than our previous simulated experiments, as the mean CV (Coefficient of Variance) of communication latency was 1.5$\times$ larger (0.59 over 0.40).}
\rev{Such variable communication latency affected each FL method differently. For baselines that use fixed deadlines (e.g., 1$T$), clients often failed to send their model updates due to long communication latency. For SmartPC (SPC) and Wait-for-All (WFA) baselines, the server had to wait for a prolonged duration when the clients were experiencing poor network connectivity. As the network conditions were not identical at different experiments,
some baselines (FedAvg+2$T$, FedAvg+SPC, Prox+1$T$, and Prox+2$T$) yielded performance with huge variance ($\geq$ 0.20 in speedup). While client failures also negatively affected \textit{FedBalancer}{}, it achieved superior time-to-accuracy and accuracy performance over the baselines due to its adaptive deadline configuration with DDL-E measurement on the sampled clients at a training round.}
\rev{We suspect the variability of the communication latency would be higher in real deployments as users with mobility and unstable network conditions would be involved. Moreover, the impact of variable communication latency would be more significant when we train larger models with >100M parameters (e.g., BERT~\cite{devlin01}) in FL. As \textit{FedBalancer}{} is not designed to actively respond to the network connectivity changes of clients in real-time, we expect \textit{FedBalancer}{} could be improved further if the client-side network condition analysis system is integrated to accurately predict the round completion time of selected clients at each round. We leave this as future work.}
\section{Discussion}
\textbf{Local Epoch Training Policies.}
While \textit{FedBalancer}{} was originally designed for FL based on FedAvg~\cite{mcmahan01} that performs full data training per local epoch of client, recent studies such as Oort~\cite{lai01} propose to perform single batch training per local epoch. There are pros and cons in both approaches; single batch training offers more frequent global model update with shorter training time on clients, but this could lead to excessive communication overhead when training large models. Full data training exploits full statistical utility of client data per round, but offers less frequent global model update with longer round. While it is the model developer's role to determine the option, \textit{FedBalancer}{} is readily applicable and improves time-to-accuracy performance on both as we demonstrated in the evaluation.
\textbf{Effect of Forward Pass at a Client.}
\label{sec:latencytradeoff}
When \textit{FedBalancer}{} selects a clients' samples, it uses a list of sample loss that is maintained on a client from the beginning of FL, instead of performing a forward pass on the client data at each round. A sample loss at the maintained list is only updated whenever the sample is selected and trained by a client, which may result in containing outdated information. We conducted an experiment to understand the effect of forward pass on performance of \textit{FedBalancer}{}, which is shown in Figure~\ref{fig:forwardpass}. On FEMNIST dataset, integrating the forward pass with \textit{FedBalancer}{} resulted in slower training, which supports our design without the forward pass. This is because that the outdated loss values are generally larger than the newly-updated ones, which encourages \textit{FedBalancer}{} to select more diverse samples while prioritizing informative samples.
\textbf{Robustness of Sample Selection.}
One of the possible limitation of \textit{FedBalancer}{} is that it might perform worse on FL tasks with noisy data, as noisy samples are highly likely to be selected by the sample selection module that prioritizes high loss. As we observed the performance improvement with \textit{FedBalancer}{} on five real-world user datasets which may already have certain noise level, we expect \textit{FedBalancer}{} would be helpful on most FL tasks. To improve further, we could systematically involve robust training approaches at centralized learning~\cite{shen01, roh01, song02} to actively deal with noisy data. This is part of our future work.
\rev{\textbf{Potential Bias of Sample Selection.}
While FedBalancer prioritizes more ``informative'' samples for training, it might integrate more samples from certain classes or sensitive groups than others, potentially leading to performance degradation on less sampled entities. To address this issue, we could integrate our sample selection strategy with other sample selection or reweighting approaches~\cite{roh01, yan01} that are designed to achieve unbiased model training. We leave this as future work.}
\section{Related Work}
\label{sec:relatedwork}
We survey closely related work with \textit{FedBalancer}{} other than the FL approaches on heterogeneous clients which we discussed earlier in Section~\ref{sec:motivation}.
\textbf{Sample Selection in Machine Learning.} There are several sample selection approaches in the field of machine learning research that could be arranged in threefold: (1) \textit{Curriculum Learning (CL)}~\cite{wu01, bengio01, graves01, hacohen01, huang01}:
CL is a sample ordering technique which trains a network with easier samples in early training stage and gradually increase the difficulty to improve convergence speed and model generalization.
However, applying it in FL is challenging as CL require a \textit{reference model} to determine the difficulty of samples, which hardly exists in FL scenarios. (2) \textit{Active Learning (AL)}~\cite{settles01, balcan01, gal01, kirsch01}: AL is a sample selection technique on unlabeled data, which interactively queries the user to label new data points that is likely to be more informative to the given task. Applying AL in FL could be non-trivial, as the training data is isolated and the labels are known but not shared externally from the clients.
(3) \textit{Importance Sampling}~\cite{katharopoulos01, alain01, loshchilov01, schaul01}: Being motivated by the fact that the importance of each training samples is different, researchers have proposed importance sampling techniques to accelerate the model training.
While their idea could be brought to FL to prioritize samples during training, determining \textit{how many} and \textit{which} samples to use for each training round and \textit{when} to calculate the sample importance is yet unknown.
\textbf{Sample Selection in FL.} Tuor et al.~\cite{tuor01} proposed a scheme that selects relevant clients' data to the given FL task, but it only selects the dataset before the FL starts. Moreover, it requires an example dataset, which is hardly applicable at FL scenarios where the client data distributions are usually unknown.
\rev{Li et al.~\cite{li08}
proposes how we can prioritize client training samples with higher importance in FL using \textit{gradient norm upper bound}~\cite{katharopoulos01}. However, their approach does not provide \textit{how many} samples should be selected per each round. While other methods such as FedSS~\cite{cai01} determines the amount of client training samples during FL, it does not specify which samples to select, simply adopting random sampling of the data. Moreover, combining FedSS with Li et al. is nontrivial, as FedSS assumes random sampling of the data. Unlike previous approaches, \textit{FedBalancer}{} is the first systematic framework that actively determines (1) \textit{how many} and (2) \textit{which} samples to select during FL to improve time-to-accuracy performance. We believe such design of \textit{FedBalancer}{} enables the use of client sample selection to improve time-to-accuracy performance.}
\textbf{Deadline Control in FL.}
Determining an optimal deadline has been largely overlooked by previous approaches; only
SmartPC determines a deadline to enable a specified proportion of the devices to complete a training round. \rev{It assumes that every client uses the same set of data for each round of FL.
\textit{FedBalancer}{} on the other hand utilizes a new deadline control strategy for FL that enables high convergence speed where client training samples dynamically change during FL due to sample selection.}
\section{Conclusion}
We presented \textit{FedBalancer}{}, a systematic FL framework with sample selection for optimized training process. \textit{FedBalancer}{} actively selects the samples with high statistical utility through client-server coordination at each FL round without exposing private information of users. To further accelerate FL with our sample selection, we design adaptive deadline control strategy for \textit{FedBalancer}{} to predict the optimal deadline for each round with client sample selection. Our evaluation of on five real-world datasets from three different domains reveal that \textit{FedBalancer}{} achieves \rev{1.20$\sim$4.48$\times$} speedup over existing FL algorithms with different deadline configuration methods, while improving the model accuracy by \rev{1.1$\sim$5.0\%}. Our design of \textit{FedBalancer}{} is easily applicable on top of orthogonal FL methods, that we demonstrate the joint implementation of \textit{FedBalancer}{} with three existing FL algorithms and report the improved time-to-accuracy performance and model accuracy.
|
\section{Introduction and notations}
\subsection{Introduction}
In parallel with many studies of classical stochastic processes, oscillating random walks, which was introduced systematically by \textsc{Kemperman}\cite{Kemperman}, have been found to be good models with several applications, see \cite{KeilsonServi} for instance. This paper deals with the homogeneous Markov chain $\mathcal{X}^{(\alpha)}=(X_{n}^{(\alpha)})_{n\geq 0}$ indexed by a parameter $\alpha \in [0,1]$ such that $X_0^{(\alpha)}=x_0$ with some fixed $x_0 \in \mathbb Z$ and for $n\geq 1$,
\begin{align} \label{def_alpha}
X_{n+1}^{(\alpha)}:=X_n^{(\alpha)}+ \left(\xi_{n+1}\mathbb 1_{\{X_n^{(\alpha)} \leq -1\}} + \eta_{n+1}\mathbb 1_{\{X_n^{(\alpha)}=0\}} + \xi_{n+1}'\mathbb 1_{\{X_n^{(\alpha)} \geq 1\}} \right),
\end{align}
where
\begin{itemize}
\item the $\xi_n, n\geq 1$, have common distribution $\mu$,
\item the $\xi'_n, n\geq 1$, have common distribution $\mu'$,
\item the $\eta_n, n\geq 1$, have common distribution given linearly by
$$\mathbb P[\eta_n=y]:= \alpha \mu(y) +(1-\alpha) \mu'(y) \text{ for any } y \in \mathbb Z,$$
\item $(\xi_n, \xi'_n, \eta_n)_{n \geq 1}$ is a sequence of independent and identically distributed (abbreviate i.i.d.) random variables.
\end{itemize}
When we want to emphasize the dependence of $\mathcal{X}^{(\alpha)}$ on the distributions $\mu$ and $\mu'$, the process $(X_n^{(\alpha)})_{n\geq 0}$ is denoted by $\mathcal{X}^{(\alpha)}(\mu,\mu')$ instead $\mathcal{X}^{(\alpha)}$.
It is merely to say that the excursion will be directed by $\mu$ (resp. $\mu'$) as long as the process stays on the negative (resp. positive) side and therefore, this present model is often called the \textit{oscillating random walk} with respect to (w.r.t.) zero level. The choice of zero is arbitrary and can be replaced by any fixed level. In case of $\alpha \in\{0;1\}$, we use the terminology \say{\textit{crossing}} to mean the point $0$ belongs to only one of the two half lines and it belongs to both if $0<\alpha<1$. We are mainly interested in recurrence of this process on its essential (i.e. maximal irreducible)
classes.
The case $\mu=\mu'$ is well-treated by $\textsc{Mijatovi\'c}$ and \textsc{Vysotsky}\cite{MijaVysot}, except providing detailed illustrations of the trajectory. The highlight result is an invariant measure (and a probability) constructed in a probabilistic manner and under the additional assumption of (topological) recurrence of the chain, it is up to a multiplicative constant finite invariant measure. For the sake of completeness, we will study all irreducible classes of $\mathcal{X}^{(\alpha)}$ from the simple case when $\xi_n\geq 0$ and $\xi_n'\leq 0$ to the general one. Appropriately refining the formula in \cite{MijaVysot}, we obtain the exact invariant measure and then apply the idea of \textsc{Knight}\cite{Knight78} to get its discrete version, see Section $2$.
Section $3$ is devoted to such an important sub-process of $\mathcal{X}^{(0)}$ evolving within a definite state space, whose elements are recorded at corresponding successive crossing times. A particular interest will be put on the structure of (essential) irreducible classes, especially on $\mathcal{N}$, the set of isolated states which are impossible to reach from any opposite states in a single step. Theorem \ref{condition} stipulates some mild conditions for $\mathcal{N}$ to be empty and also yields an expression for it. In analogy to the approach for reflected random walks, we finally compute the invariant measure for the sub-process based on some arguments developed in the previous work of \textsc{Peign\'e} and \textsc{Woess}\cite{MarcWoess}.
The recurrence of the general oscillating random walk $\mathcal{X}^{(\alpha)}$ is dealt with in Section $4$. Some powerful tools such as the \textsc{Kemperman}'s criterion \cite{Kemperman} (a divergent series represented in term of renewal functions) or the integral criterion of \textsc{Rogozin} and \textsc{Foss} \cite{RogozinFoss} (a transformation established with the help of Wiener-Hopf factorization) are mentioned for reference. We also furnish a new approach coming from the fact that the recurrence of the crossing sub-process implies to the recurrence of the full process. To do this, we first show that the process of crossing is (positive) recurrent if the tail distribution condition
$$\displaystyle \sum_{n = 0}^{+\infty} \mathbb P(\xi_1 >n)\,\mathbb P(\xi_1'<-n) < +\infty$$
holds when $\xi_n \geq 0, \xi_n'\leq 0$. Moreover, it can be attained under the hypothesis $\mathbb{E}[\xi_1^p]<+\infty$ and $\mathbb{E}[(-\xi_1')^q] <+\infty$, where $p, q \in ]0,1[$ satisfying $p+q=1$. When jumps are generalized on $\mathbb Z$, there may have different possibilities, for instance, both $\xi_n$ and $\xi_n'$ are either drifted (positive and negative, respectively) or centered as well as mixed and by Theorem \ref{extend_condition}, we will address suitable conditions to each corresponding situation.
\subsection{Notations}
Throughout this paper, we fix some frequently used notations
\begin{itemize}
\item $S_{\mu}\, (\text{resp. } S_\mu'):$ the support of $\mu \, (\text{resp. } \mu')$.
\item $D\, (\text{resp. } D'): \text{ the maximum of }\mu \, (\text{resp. the minimum of } \mu')$.\\
We adhere to the convention that $D=+\infty $ (resp. $D'=-\infty$) when $S_\mu$ (resp. $S_{\mu'}$) is unbounded from above (resp. from below).
\item $d \,(\text{resp. }d'):$ the greatest common divisor of $S_\mu \, (\text{resp. } S_\mu')$.
\item $\mathbb Z^{+}/\mathbb Z^{-}:$ the set of positive/negative integers $(\text{and }\mathbb Z_0^+/\mathbb Z_0^- \text{ if } 0 \text{ is included})$.
\item $r_x$: the remainder of $x$ in the Euclidean division by $\delta$ (i.e. $0\leq r_x<\delta$).
\end{itemize}
Let us end this paragraph dedicated to notations by reminding that for any fixed $0 \leq \alpha \leq 1$, the chain \eqref{def_alpha} is denoted by
$ \mathcal{X}^{(\alpha)} (\mu, \mu')$ (and simply $\mathcal{X}^{(\alpha)} $ when there is no ambiguity on the choices of $\mu$ and $\mu'$).
\section{Irreducible classes and invariant measure of $ \mathcal{X}^{(0)} (\mu, \mu')$}
It is easy to check that if $\mu=\mu'$ (in this case $\mathcal{X}^{(0)} (\mu, \mu')$ becomes an ordinary random walk on $\mathbb Z$ with the unique jump measure $\mu$) then $d=d'$ and the irreducible classes of $\mathcal{X}^{(0)} (\mu, \mu')$ are the sets $r+d\mathbb Z$ with $0\leq r <d$. In this section, we describe the essential classes of $\mathcal{X}^{(0)} (\mu, \mu')$ when $\mu\neq \mu'$.
\subsection{ The chain $ \mathcal{X}^{(0)}(\mu, \mu')$ when $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$}
For any $x \in \mathbb Z$, let ${\mathcal I}(x)$ be the irreducible class of $x$. It holds $ {\mathcal I}(0)\subset \{D', \ldots, D-1\}$. Furthermore, for any starting point $x$, after finitely many steps, the chain $\mathcal X ^{(0)} $ stays for ever a.s. in the subset $\{D', \ldots, D-1\}$.
\begin{theo}\label{irreducibleclass} { We suppose that $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$.}
$\bullet$ Assume first $D$ and $D'$ are finite. If $d \wedge d'=\delta$, then
i) there exist $\delta$ irreducible essential classes
\[
\{D', \ldots, D-1\}\cap (r+\delta \mathbb Z) \quad with \quad 0\leq r <\delta;
\]
(in particular the irreducible class of $0$ equals $\mathcal I(0)= \{D', \ldots, D-1\}\cap \delta \mathbb Z$);
ii)
if $x \geq D$ or $x<D'$ then $x$ is transient and, after finitely many steps, reaches $\mathbb P$-a.s. the essential class $ \{D', \ldots, D-1\}\cap (r_x+\delta \mathbb Z)$.
$\bullet$ If $D'=-\infty$ and $D$ is finite then
i) there exist $\delta$ irreducible essential classes
\[
]-\infty, D-1]\cap (r+\delta \mathbb Z) \quad with \quad 0\leq r <\delta;
\]
ii)
the $x \geq D$ are all transient and, after finitely many steps, reaches $\mathbb P$-a.s. the essential class $ ]-\infty, \ldots, D-1]\cap (r_x+\delta \mathbb Z)$.
(similar dual statement follows when $D'$ is finite and $D=+\infty$).
$\bullet$ If $D=+\infty$ and $D'=-\infty$ then
there exist $\delta$ irreducible essential classes, which are all essential:
\[
r+\delta \mathbb Z \quad with \quad 0\leq r <\delta.
\]
\end{theo}
\begin{proof}
The proof of \textsc{Kemperman} based on the theory of semi-groups of $\mathbb Z$ entirely solved for the chain $ \mathcal{X}^{(\alpha)}, 0\leq \alpha \leq 1$ providing that $D=-D'=+\infty$ (see Remark \ref{ess_general}).
Back to the current model, we will prove by induction, but let us first fix some notations.
Let $T$ be a finite subset of $ S_\mu\cup S_{\mu'}$ s.t. $ T\cap S_\mu \neq \emptyset $ and $T\cap S_{\mu'}\neq \emptyset$; without loss of generality, we assume $0\notin T$.
For any $x \in \mathbb Z$, we denote by $\mathcal O_{T}(x)$ the ``{\it orbit of $x$ under $T$}'', that is the set of sequences ${\bf x}= (x_i)_{i \geq 0}$ defined by induction as follows: $x_0=x$ and, for any $i\geq 1$,
- if $x_i\leq -1$ then $x_{i+1}= x_i+s $ for some $s \in T \cap S_\mu$;
- if $x_i\geq 0$ then $x_{i+1}= x_i+s'$ for some $s' \in T\cap S_{\mu'}.$
Notice that all the $x_i$ but finitely many do belong to $\{ \min(S_{\mu'}), \ldots, \max (S_\mu)-1\}$.
For any $x, y \in \mathbb Z$, we write $x\ {\stackrel{T}{\rightarrow}}\ y$ if there exists ${\bf x} \in \mathcal O_T(x)$ s.t. $x_0=x$ and $x_n= y$ for some $n\geq 0$. When there exists no such sequence $\bf x$, we write $x\ {\stackrel{T}{\not\rightarrow}}\ y$.
The notation $x\ {\stackrel{T}{\leftrightarrow}}\ y$ means $x\ {\stackrel{T}{\rightarrow}}\ y$ and $y\ {\stackrel{T}{\rightarrow}}\ x$.
The relation $ {\stackrel{T}{\leftrightarrow}} $ is an equivalence relation on $\mathbb Z$ whose classes are called {\it $T$-irreducible classes. }
The $T$-irreducible class of $x$ is denoted $\mathcal I_{T}(x)$.
The relation $ {\stackrel{T}{\rightarrow}} $ induces a partial order relation on $\mathcal I_T$, denoted again $ {\stackrel{T}{\rightarrow}}$. The maximal irreducible classes for this relation are called $T${\it -essential}; a non essential irreducible class is said $T$-transient.
We now describe ${\mathcal I}_T$, by induction on the cardinality of $T$.
\noindent \underline{{\bf Step 1-} {\it Case when $T= \{s, s'\}$ with $s \in \mathbb Z^{+}$ and $s'\in \mathbb Z^{-}$. }}
$\bullet$ {\it We assume first $s\wedge s'=1, x=0$ and prove that $\{s', \ldots, s-1\}$ is the unique $T$-essential class. Furthermore, $\mathcal I_T(x)$ equals $\{x\}$ and is $T$-transient when $x\geq s$ or $x<s'$.}
Indeed, for any ${\bf \omega} = ( \omega_i)_{i \geq 0} $ in $\mathcal O_T(0)$,
the $\omega_i$ all belong to $\mathbb N s+\mathbb N s'$ and to $\{s', \ldots, s-1\}$. Hence, there exist $j>i\geq 1$ such that $\omega_j=\omega_i$. Since $\omega_{i}\neq \omega_{i+1}, \omega_{j-1}\neq \omega_j$ and $\omega_i\ {\stackrel{T}{\rightarrow}} \ \omega_j=\omega_i$, there exists $k, \ell \geq 1$ such that $ks+\ell s' = 0$. The condition $s\wedge s'=1$ yields $k=0$ mod$(s')$ and $\ell = 0$ mod$(s$), hence $k\geq |s'|$ and $\ell \geq s$. Consequently, the sub-orbit $\{\omega_i, \omega_{i+1}, \ldots, \omega_{j-1}\}$ contains at least $s+|s'|$ elements; since it is included in $\{s', \ldots, s-1\}$, it holds in fact $\{\omega_i, \omega_{i+1}, \ldots, \omega_{j-1}\}=\{s', \ldots, s-1\}.$ This proves that $x\ {\stackrel{T}{\rightarrow}} \ y$ for any $x, y \in \{s', \ldots, s-1\}$, hence $\{s', \ldots, s-1\} \subset \mathcal I_T(0)$. Eventually, $\{s', \ldots, s-1\}= \mathcal I_T(0)$ since the elements of all the orbits of $0$ remain in $\{s', \ldots, s-1\}$. This also implies that $\{s', \ldots, s-1\}$ is $T$-essential.
As a consequence of this argument, the orbit $\mathcal O_T(0)$ contains a unique sequence $\omega$, which is is periodic with period $ \omega_0, \omega_1, \ldots, \omega_{s +|s'|-1}$ where $\omega_i \in \{s', \ldots, s-1\}$ and $\omega_i\neq \omega_j$ for any $0\leq i<j< s +|s'|.$ We write for short $ \omega = \overline {\omega_0, \ldots, \omega_{s +|s'|-1}}$ and emphasize that $\{\omega_0, \ldots, \omega_{s +|s'|-1}\}=\{s', \ldots, s-1\}$.
Now, if $x\geq s$ or $x<s'$ and ${\bf x} = (x_i)_{i \geq 0} \in \mathcal O_T(x)$, then $x_i \neq x$ for any $i\geq 1$; in other words, $\mathcal I_T(x) = \{x\}$ and $x$ is $T$-transient.
$\bullet$ {\it Assume now $s\wedge s'= \delta \geq 2$. Then, the $T$-essential classes are $\{s', \ldots, s-1\}\cap (x+\delta \mathbb Z)$. Furthermore, if $x\geq s$ or $x<s'$, then $\mathcal I(x)$ equals $\{x\}$ and is $T$-transient.}
In this case, the set $\mathcal O_T(0)$ still contains a unique sequence
$\omega= \overline {\omega_0, \ldots, \omega_{k-1}}$ with $k= {s+ |s'| \over \delta} $ and $\{\omega_0, \ldots, \omega_{ k-1}\}=\{s', \ldots, s-1\}\cap \delta \mathbb Z$. As a direct consequence, for any $x \in \mathbb Z$, the set $\mathcal O_T(x)$ is included in $x+\delta \mathbb Z$ and also contains a unique sequence ${\bf x}$, which is ultimately periodic with period $r_x+\omega_0, \ldots, r_x+\omega_{k-1}$. The description of the $T$-irreducible classes follows immediately.
\vspace{3mm}
\noindent \underline{{\bf Step 2-}\it Case when $T\subset \mathbb Z$ is finite and satisfies $T\cap \mathbb Z^{-} \neq \emptyset$ and $T\cap \mathbb Z^{+} \neq \emptyset$}.
The proof is made by induction, from $T$ to ${\bf T}:= T\cup \{t'\}$ with $t'\in \mathbb Z^{-}$ ; the case when ${\bf T}:= T\cup \{t\}$ with ${t}\in \mathbb Z^{+}$ is studied in the same way. By Step 1, the property is true for $T=\{s, s'\}, s>0$ and $s'<0$.
\noindent \underline{Hypothesis of induction:} {\it
Let $T$ be a set of non-zero integers s.t. $ T\cap \mathbb Z^{-} \neq \emptyset$ and $ T\cap \mathbb Z^{+} \neq \emptyset$. We set $\delta_T:=\text{\rm gcd}(T)$ and denote $s$ (resp. $s'$) the largest (resp. smallest) element of $T$. We assume that
- the $T$-essential classes are $\{s', \ldots, s-1\}\cap (r+\delta_T \mathbb Z)$ with $0\leq r<\delta_{ T}$;
- when $x\geq s$ or $x<s'$, then $\mathcal I_T(x)=\{x\}$ and $x$ is $T$-transient; furthermore, for any ${\bf x} = (x_i)_{i \geq 0} \in \mathcal O_T(x)$, all $x_i$ but finitely many belong to the $T$-essential class $\{s', \ldots, s-1\}\cap (x+\delta_T \mathbb Z)$.
\noindent \underline{Conclusion:} For $t' \in \mathbb Z^{-}$ then the same property holds for ${\bf T}= T\cup\{t'\}$. In other words, setting $\delta_{\bf T}:= \text{\rm gcd}({\bf T})$ and $m_{\bf T}:= \min(s', t')$,
- the ${\bf T}$-essential classes are $\{m_{\bf T}, \ldots, s-1\}\cap (r +\delta_{\bf T} \mathbb Z)$ with $0\leq r <\delta_{\bf T}$;
- if $x \geq s$ or $x<m_{\bf T}$, then $\mathcal I_{\bf T}(x)= \{x\}$ and $x$ is ${\bf T}$-transient; furthermore, for any ${\bf x} = (x_i)_{i \geq 0} \in \mathcal O_{\bf T}(x)$, all $x_i$ but finitely many belong to the $\bf T$-essential class $\{m_{\bf T}, \ldots, s-1\}\cap (x+\delta_{\bf T} \mathbb Z)$.
}
The same argument as in Step 1 works to deduce the case $\delta_{\bf T}\geq 2$ from the case $\delta_{\bf T}=1$; thus, we only consider the case $\delta_{\bf T}=1$.
Notice that $\{m_{\textbf{T}},\dotsi,s-1\}$ is absorbing; in other words, for any $x \in \mathbb{Z}$ and any $\bf{x}=(x_i)_{i \geq 0} \in \mathcal{O}_{\bf{T}}(x)$, all $x_i$ but finitely many belong to this set. In particular, $x$ is transient when $x\geq s$ or $x<m_{\bf T}$. It thus remains to check that $\{m_{\bf T},\dotsi,s-1\}$ is irreducible.
Let us first prove that
\begin{equation}\label{zlibekjf}
\{s', \ldots, s-1\} \ {\stackrel{\bf T}{\rightarrow}}\
\{s', \ldots, s-1\}
\end{equation}
For any $x \in \{s', \ldots, s-1\} $, we choose $\ell_x\geq 0$ s.t. $x+t'+\ell_x s\in\{0, \ldots, s-1\}$ and notice that $x\ {\stackrel{\bf T}{\rightarrow}}\ x+t'+\ell_x s$. Now, for any ${\bf y} = (y_i)_{i\geq 0} \in \mathcal O_T(x+t'+\ell_x s)$, all $y_i$ but finitely many belong to the $T$-essential class $\mathcal I_T(x+t')=\{s, \ldots, s'-1\}\cap (x+t'+\delta_T \mathbb Z)$; thus, $\{x\}\ {\stackrel{\bf T}{\rightarrow}}\ \{s', \dots, s-1\} \cap (x+t'+\delta_T \mathbb Z).$ Reiterating the argument, we get, for $k\geq 1$ \begin{equation}\label{libve}
\{x\}\ {\stackrel{\bf T}{\rightarrow}}\ \{s', \dots, s-1\} \cap (x+kt'+\delta_T \mathbb Z).
\end{equation}
Since $\text{\rm gcd}(t', \delta_T)=\delta_{\bf T}=1$, the class of $t'\ \text{\rm mod}(\delta_T)$ generates $\mathbb Z/\delta_T\mathbb Z$, so
$$
\bigcup_{0\leq k<\delta_T}\{s', \ldots, s-1\}\cap (x+kt'+\delta_T \mathbb Z)=\{s', \ldots, s-1\}.
$$
This yields immediately $\{x\} \ {\stackrel{\bf T}{\rightarrow}}\ \{s', \ldots, s-1\}$: indeed, for any $y \in \{s', \ldots, s-1\}$ there exists $k_y\geq 1$ s.t. $y \in \{s', \ldots, s-1\}\cap (x+k_y t'+\delta_T \mathbb Z)$ and (\ref{libve}) readily implies $ x \ {\stackrel{\bf T}{\rightarrow}} \ y$. This holds for any $x \in \{s', \ldots, s-1\}$ and proves (\ref{zlibekjf}).
This implies $\{m_{\bf T}, \ldots, s-1\}\ {\stackrel{\bf T}{\rightarrow}} \{ m_{\bf T}, \ldots, s-1\}$ when $s' < t'$. It remains to consider the case when $t'<s'$; we fix $x \in \{t', \ldots, s'-1\}$.
The same argument as above proves that $ \{x\} \ {\stackrel{\bf T}{\rightarrow}} \ \{s', \ldots, s-1\}$.
Conversely, we decompose $x$ as $x=t'+k_x s+r_x$ with $k_x \geq 0$ and $0\leq r_x<s$. For any $y \in \{s', \ldots, s-1\}$, property (\ref{zlibekjf}) yields $y\ {\stackrel{\bf T}{\rightarrow}} \ r_x$ since $0\leq r_x <s$; now, immediately, $r_x
\ {\stackrel{\bf T}{\rightarrow}} \ t'+k_x s+r_x=x$ so that $y\ {\stackrel{\bf T}{\rightarrow}} \ x$ as expected.
\vspace{3mm}
\noindent \underline{{\bf Step 3 -}{\it Proof of Theorem \ref{irreducibleclass} } }
When $D$ and $D'$ are finite, we set $T=S_\mu\cup S_{\mu'}$ and apply Step 2.
If $D'=-\infty$ and $D$ is finite, we set $T=T_{s'}= S_{\mu}\cup (S_{\mu'}\cap \{s', \ldots, -1\})$ with $s' \leq -1$, apply Step 2 then let $s'\to -\infty$. The two other cases are treated in the same way.
\end{proof}
It is worth remarking that one may extend the above to adapt for $\mathcal{X}^{(\alpha)}$, that is to say, the chain (starting at any initial point) will be absorbed after finitely many steps by the essential class
\begin{align*}
\left\{
\begin{array}{lll}
\{D',\ldots, D-1\}& {\rm if} & \alpha=0,\\
\{D',\ldots, D\}&{\rm if} & 0<\alpha<1,\\
\{D'+1, \ldots, D\}& {\rm if} & \alpha=1.
\end{array}
\right.
\end{align*}
under the assumption of bounded jumps. The remaining statements of the theorem are proved in an analogous way.
\subsection{ The chain $ \mathcal{X}^{(0)} $ in the general case}
We start by considering the irreducible classes of $ \mathcal X^{(0)} $ on the additional assumptions that $S_\mu \cap \mathbb Z^- \neq \emptyset$ and $S_{\mu'} \cap \mathbb Z^+ \neq \emptyset$. Intuitively, the oscillating random walk $\mathcal{X}^{(0)}$ starting from $0$ can visit arbitrarily large integers and so, it is quite natural to think that $ {\mathcal I}(0)$ contains the whole line $\mathbb Z$ in this case, under the hypothesis $d\wedge d'=1$. In fact, it is false and depends deeply on the structure of $S_\mu$ and $S_\mu'$. The following theorem, which is based on the ideas developed in Step 2 above, clarifies this point.
\begin{theo}\label{generalcase}
We write $S_{\mu}^{+}$ (resp. $S_{\mu'}^{+}$) and $S_{\mu}^{-}$ (resp. $S_{\mu'}^{-}$) as the positive and negative components of $S_\mu$ (resp. $S_\mu'$), respectively. Suppose that these subsets are all non-empty. If $d \wedge d'=\delta$ and $d \neq d'$ then
$\bullet$ \underline{Case when $D <d'$}\\
i) if $x \in \{D,\ldots,d'-1\}+d'\mathbb Z^{+}_0$ then $x$ is transient and its irreducible class is $$\mathcal{I}(x)=(x+d' \mathbb Z) \cap [D,+\infty[;$$
ii) otherwise, $x$ is essential and its essential class is given by
$$\mathcal{I}(x)=(r_x+\delta \mathbb{Z}) \setminus (\{D,\ldots,d'-1\}+d' \mathbb Z^{+}_0).$$
$\bullet$ \underline{Case when $D'>-d$}
i) if $x \in \{-d,\ldots,D'-1\}+d\mathbb Z^{-}_0$ then $x$ is transient and its irreducible class is $$\mathcal{I}(x)=(x+d\mathbb{Z}) \,\cap \,]-\infty,D'-1];$$
ii) otherwise, $x$ is essential and its essential class is given by
$$\mathcal{I}(x)=(r_x+\delta \mathbb{Z}) \setminus (\{-d,\ldots,D'-1\}+d\mathbb Z^{-}_0).$$
$\bullet$ \underline{Case when $D \geq d'$ and $D' \leq -d$}
There is/are $\delta$ irreducible class(es), which is/are all essential:
$$r+\delta\mathbb Z \text{ with } 0\leq r <\delta.$$
\end{theo}
\begin{proof}
Before delving into details, we shall pay more attention to the fact that these two conditions cannot be attained at the same time due to $-D<d'<D'<d$ and we thus arrive at a contradiction. In particular, if $D<d'$ then $-D' \geq d$ and vice versa.
\newline
$\bullet$ \underline{Case when $D'<d$} \\
i) The converse is easily done since $x\ {\stackrel{S_\mu \cup S_{\mu'}}{\longleftrightarrow}}\ x+kd'$ for any $k \in \mathbb{Z}$ satisfying $x+kd' \geq D$. Indeed, the assumption of $S_{\mu'}$ leads to the semi-group generated by $S_{\mu'}$, says $T_{\mu'}$, is equal to $d' \mathbb{Z}$. One can write $kd'=\displaystyle \sum s'_{i}$ as the finite sum of elements in $S_{\mu'}$. Selecting first the positive elements (if any) and then the negative ones, it immediately follows that $(x+d'\mathbb{Z}) \cap [D,+\infty[\, \subset C(x)$.
Now, we will show by contraposition that $C(x) \subset (x+d'\mathbb{Z}) \cap [D,+\infty[$. Suppose $z \in ]-\infty,D-1] \cap C(x)$.
Let $\tau$ be the last time entering to the set $\{0,\ldots,D -1\}$ of $\mathcal{X}^{(0)}$ before visiting $x$ for the first time. Since $z\ {\stackrel{S_\mu \cup S_{\mu'}}{\longrightarrow}}\ x$, we have $\mathbb{P}_{z}[\tau < +\infty]=1$ and then put $X_\tau^{(0)}=y \in \{0,\ldots,D-1\}$. Observe that $x-y \in \{1,\ldots,d'-1\} +d'\mathbb Z^{+}_0$, i.e. $x-y \notin d'\mathbb{N}$ and thus, $z\ {\stackrel{S_\mu \cup S_{\mu'}}{\not\longrightarrow}}\ x$ (contradiction). When $z \geq D$, the crossing process starting at $z$ is directed only by $S_{\mu'}$ and therefore, $z \in x+ d' \mathbb Z$. \\
ii) A reasoning similar to the above yields that
\begin{align*}
\ y\ {\stackrel{S_{\mu} \cup S_{\mu'}}{\longleftrightarrow}} \ y+kd' \ \text{ \rm if } y\in \{0,\ldots,D-1\}+d' \mathbb Z^{+}_0,
\end{align*}
where $k \in \mathbb Z \text{ s.t. } y+kd' \geq 0$. \\
Moreover, by Theorem \ref{irreducibleclass}
\begin{align}\label{abc}
\{D',\ldots,D-1\}\cap (r_x+\delta \mathbb{Z}) \ {\stackrel{S_{\mu}^{+} \cup S_{\mu'}^{-}}{\longrightarrow}} \{D',\ldots,D-1\}\cap (r_x+\delta \mathbb{Z}).
\end{align}
Now, we fixed $t \in ]-\infty,D'[ \,\cap \,(r_x+\delta \mathbb Z)$, then there is some $z_t \in \{D',\dotsi,-1\} \cap (r_x+\delta \mathbb Z)$ s.t. $z_t \equiv t $(mod $d$) due to $-D' \geq d$. Combining \eqref{abc} with the fact that $T_\mu=d \mathbb Z$, we get
\begin{align*}
\{t\} \ {\stackrel{S_{\mu} \cup S_{\mu'}}{\longleftrightarrow}} \{D',\dotsi,D-1\}\cap (r_x+\delta \mathbb{Z}).
\end{align*}
This property holds for every choice of $t$, so
\begin{align*}
\{ ]-\infty,D'[ \,\cap \,(r_x+\delta \mathbb Z)\} \ {\stackrel{S_{\mu} \cup S_{\mu'}}{\longleftrightarrow}} \{D',\dotsi,D-1\}\cap (r_x+\delta \mathbb{Z})
\end{align*}
which achieves the proof.\\
$\bullet$ We finish by mentioning that, suitably modified, the above argument applies
to the other cases.
\end{proof}
\begin{remark}{\label{ess_general} Suppose that $S_\mu$ is unbounded from above $(D=+\infty)$ and $S_{\mu'}$ is unbounded from below $(D'=-\infty)$. In such situation, all states connect and it is in fact possible that the process $\mathcal{X}^{(0)}$ with a positive probability will never reach a given neighborhood of $0$ due to infinitely many extremely large jumps across $0$. Consequently, there exists (with a positive probability) an orbit between every two points, which has no intermediary state belonging to a given finite set $F$, even if $0 \notin F$ (see \cite{Kemperman}). }
\end{remark}
\subsection{ On the invariant measure for $ \mathcal{X}^{(0)}$ when $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$}
The concept of invariant measure plays a crucial role in the study in the long-time behaviour and asymptotic properties of a Markov chain. To adapt for the current situation, a proper adjustment to the invariant measure in \cite{MijaVysot} is indispensable and what we found is the following
\begin{lem} { Assume $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$}. The measure $\lambda$ on $\mathbb{R}$ given by
$$\lambda(dx) = \left(\mathbb{1}_{]-\infty,0[}(x)\, \mathbb{P}[\xi_{1}' < x] + \mathbb{1}_{[0,+\infty[}(x)\, \mathbb{P}[\xi_{1} > x] \right)dx,$$
where $dx$ is Lebesgue measure, is invariant for $\mathcal{X}^{(0)}$.
\end{lem}
\begin{proof}
For any $a \geq 0$, it follows without difficulty that
\begin{align*}
\lambda P ]a, +\infty[ &= \displaystyle \lim_{N \to +\infty} \displaystyle \int_{-N}^{0} \mathbb{P}[\xi'_{1} <x] \, \mathbb{P}[x+\xi_{1} >a] \, dx + \displaystyle \lim_{N \to +\infty} \displaystyle \int_{0}^{N} \mathbb{P}[\xi_1>x] \, \mathbb{P}[ x+\xi'_{1} > a] \, dx \\
&= \displaystyle \lim_{N \to +\infty} \displaystyle \int_{-N}^{0} \mathbb{P}[\xi'_{1} <x] \, \mathbb{P}[\xi_{1} >a-x] \, dx + \displaystyle \lim_{N \to +\infty} \displaystyle \int_{0}^{N} \mathbb{P}[\xi_1>x] \, \mathbb{P}[\xi'_{1} > a-x] \, dx \\
&= \displaystyle \lim_{N \to +\infty} \displaystyle \int_{a}^{N+a} \mathbb{P}[\xi'_{1} <a-y] \, \mathbb{P}[\xi_{1} >y] \, dy + \displaystyle \lim_{N \to +\infty} \displaystyle \int_{a}^{N} \mathbb{P}[\xi_1>y] \, \mathbb{P}[\xi'_{1} > a-y] \, dy \\
&= \displaystyle \lim_{N \to +\infty}\displaystyle \int_{a}^{N} \mathbb{P}[ \xi_{1} > y] \, dy+ \displaystyle \lim_{N \to +\infty}\displaystyle \int_{N}^{N+a}\mathbb{P}[\xi'_{1} <a-y] \, \mathbb{P}[\xi_{1} >y] \, dy\\
&=\lambda ]a, +\infty[
\end{align*}
since $\displaystyle \int_{N}^{N+a}\mathbb{P}[\xi'_{1} <a-y] \, \mathbb{P}[\xi_{1} >y] \, dy \leq \displaystyle \int_{0}^{a} \mathbb{P}[\xi_1 > z+N]\,dz \rightarrow 0$ as $N \rightarrow +\infty$.\\
The same computation yields $\lambda P ]-\infty, -a[ = \lambda ]-\infty, -a[ $ and thus, the proof is complete.
\end{proof}
As a direct consequence, we obtain the following result.
\begin{coro} \label{positive-recurrent}
Assume $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$. Then $\mathcal{X}^{(0)}$ is positive recurrent on each essential class iff both measures $\mu$ and $\mu'$ have finite first moment.
\end{coro}
\begin{lem} Assume $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$. The discrete measure $\nu$ on $\mathbb{Z}$ given by
\begin{align}\label{discrete}
\nu(m):= \left\{
\begin{array}{lll}
\mu']-\infty,m] &{\rm if} &m \leq -1\\
\mu [m+1, +\infty[ &{\rm if} &m\geq 0.
\end{array}
\right.
\end{align}
is invariant for the homogeneous random walk $\mathcal X^{(0)}$. Moreover, for arbitrary $x_0 \in \mathbb{Z}$, it induces a corresponding invariant measure $\nu_{x_0}$ on $\mathcal{I}(r_{x_0})$ by its restriction on this essential class (in other words, if $A \subset \mathcal{I}(r_{x_0})$ then $\nu_{x_0}(A)= \displaystyle \sum_{x \in A}\nu(x)$).
\end{lem}
\begin{proof}
At first, we briefly outline the idea of the result. Let $X$ be a measurable space that is compatible with the $\text{Borel } \sigma-$algebra $\mathcal{B}$. Suppose we have successfully founded an invariant measure $\lambda$ on $(X,\mathcal{B})$ with the corresponding transition operator $P$. Let $\sim$ be an equivalence relation on $X$ and denote by $\overset{\sim}{X}:=X_{/_{\sim}}$ the quotient space of $X$ (whose equivalence classes belong to $\mathcal{B}$) under this relation. We assume $\overset{\sim}{X}$ is countable and holds for any elements $C_i, C_j \in \overset{\sim}{X}$,
\begin{align}\label{quotientproperty}
P(x,C_j)=P(y,C_j) \text{ \rm for any }x,y \in C_i. \tag{*}
\end{align}
Then the kernel $P$ induces a Markov transition $\overset{\sim}{P}$ on $\overset{\sim}{X}$ s.t. $\overset{\sim}{P}(C_i,C_j):=P(x,C_j)$ with $x \in C_i$; furthermore, the measure $\overset{\sim}{\lambda}$ on $\overset{\sim}{X}$ defined by $\overset{\sim}{\lambda}(C_i):=\lambda{(C_i)}$ is $\overset{\sim}{P}$- invariant. Indeed, for any $C' \in \overset{\sim}{X}$
\begin{align*}
\overset{\sim}{\lambda}\overset{\sim}{P}(C')&= \displaystyle \sum_{C}\overset{\sim}{\lambda}(C)\overset{\sim}{P}(C,C')\\
&=\displaystyle \sum_{C} \displaystyle \int_{C}P(x,C')\lambda(dx)\\
&=\displaystyle \int_{X}P(x,C')\lambda(dx)\\
&=\overset{\sim}{\lambda}(C').
\end{align*}
Let us now explain how to apply this general principle to get the exact formula of the invariant measure for the oscillating random walk. Consider the following equivalence relation
$$x\sim y \Longleftrightarrow \exists n \in \mathbb{Z} \text{ s.t. }x,y \in [n,n+1[,$$
which apparently satisfies the condition \eqref{quotientproperty}. Taking $X=\mathbb R$ and $C_j:=[j,j+1[$ for any $j\in \mathbb Z$, one admits $\nu(m):=\lambda(C_m)$ (compare \eqref{discrete}) as the discrete invariant measure of $\mathcal{X}^{(0)}$.
\end{proof}
\begin{remark} When the state $0$ is supposed to be merged to the negative side $(\alpha=1)$, we replace $C_j$ by $C_j^*=]j-1,j]$ and the resulting invariant measure $\nu^*$ is given by
\[
\nu^*(m):= \left\{
\begin{array}{lll}
\mu']-\infty,m-1] &{\rm if} &m \leq 0\\
\mu [m, +\infty[ &{\rm if} &m\geq 1.
\end{array}
\right.
\]
\end{remark}
\section{ On the crossing sub-process of $ \mathcal{X}^{(0)}$ when $S_\mu\subset \mathbb Z^+$ and $S_{\mu'}\subset \mathbb Z^-$}
In this section, we would like to study the recurrence and the invariant measure of the embedding process of $\mathcal{X}^{(0)}$, which contains only states at the crossing times. For convenience, let us define the random variables $(S_n)_{n \geq 0}$ and $(S'_n)_{n \geq 0}$, the simple random walks associated with laws $\mu$ and $\mu'$ respectively by $S_0=S_0'=0$ and for $n\geq 1$,
$$S_{n} = \xi_{1} + \xi_{2} + \dots + \xi_{n},$$
and
$$S'_{n} = \xi'_{1} + \xi'_{2} + \dots + \xi'_{n}.$$
Denote by $\mu^{*n}$ the $n$-fold convolution of $\mu$ with itself (also the distribution of $S_n$) and $\mathcal{U} = \displaystyle \sum_{n \geq 0}\mu^{*n}$ its potential kernel; similarly for $\mu'^{*n}$ and $\mathcal{U'}$.
Now we consider the sequence of crossing times ${\bf C}=(C_k)_{k \geq 0}$ at which the process changes its sign whenever crossing $0$. Assume that $C_0=0$ and we designate $C_k$ as the time of $k^{th}$- crossing given by
\begin{align}\label{crossing_time}
C_{k+1} := \inf \{n > C_k: X_{C_k}^{(0)} + (\xi_{C_k + 1} + \xi_{C_k + 2} + \dotsi +\xi_{n}) \geq 0 \text{ \rm if } X_{C_k}^{(0)} \leq -1 \\
\text{or } X_{C_k}^{(0)} + (\xi'_{C_k + 1} + \xi'_{C_k + 2} + \dotsi + \xi'_{n}) < 0 \text{ \rm if } X_{C_k}^{(0)} \geq 0 \} \nonumber
\end{align}
This forms a sequence of stopping times with respect to the filtration $\mathbb{F}:=(F_n)_{n \geq 0}$ where $F_n :=\sigma \left(\xi_{k}, \xi'_{k}\mid k \leq n \right)$. By the law of large number, one gets $S_{n} \to +\infty$ and $S'_{n} \to -\infty$ $\mathbb{P}$-almost surely and thus, $\mathbb{P}_{x}[C_k <+\infty]=1$ for all $x \in \mathbb{Z}$.
\begin{lem} { Assume $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$.} The sub-process $(X_{C_k}^{(0)})_{k \geq 0}$ is a (time-homogeneous) Markov chain on $\mathbb{Z}$ with its transition kernel determined by
\begin{align}\label{matrixC}
C(x,y)=\left\{
\begin{array}{lll}
\displaystyle \sum_{t = 0}^{-x - 1} \mu(y - x -t)\,\mathcal{U}(t) &\text{\rm if } x<0 \text{ \rm and } y\geq 0,\\
\displaystyle \sum_{t= -x}^{0} \mu'(y - x - t) \, \mathcal{U'}(t) &\text{\rm if } x\geq 0 \text{ \rm and } y<0,\\
0 &{\rm otherwise}.
\end{array}
\right.
\end{align}
\end{lem}
The process $\mathcal X^{(0)}_{\bf C}:= (X_{C_k}^{(0)})_{k \geq 0}$ is called the {\bf crossing sub-process} of $\mathcal X^{(0)}$.
\begin{proof}
The Markov property is obvious.\\
If $x < 0$ and $y \geq 0$ (similar to $x\geq 0,\, y<0$) then we have
\begin{align*}
C(x, y) &= \mathbb{P} [X_{C_1} = y \mid X_{0} = x] \\
&= \displaystyle \sum_{n=1}^{+\infty} \underbrace{\mathbb{P}[ x + S_{n-1} \leq -1,\, x + S_{n} = y]}_{\text{since }(S_n)_{n\geq 1} \text{ \rm is increasing}}\\
&= \displaystyle \sum_{n=1}^{+\infty} \displaystyle \sum_{t = 0}^{-x-1} \mathbb{P}[S_{n-1} = t] \, \mathbb{P}[\xi_{n} = y - x - t]\\
&= \displaystyle \sum_{t =0}^{-x-1} \mathbb{P}[\xi_{1} = y - x - t]\, \displaystyle \sum_{n=1}^{+\infty} \mathbb{P}[S_{n-1} = t]\\
&= \displaystyle \sum_{t= 0}^{-x-1} \mu(y - x - t) \, \mathcal{U}(t).
\end{align*}
\end{proof}
\subsection{Irreducible classes of $\mathcal X_{\bf C}^{(0)}$}
In case of reflected random walk, it is well-known in \cite{MarcWoess} that the full process and its process of reflections possess the common essential classes.
Since the reflected random walk is regarded as the anti-symmetric case of our general model in which we identify the points themselves and their mirror images relative to $0$, it comes naturally a question whether this phenomenon possibly occurs. There is no solid information to give an exact answer other than the intuitive relationship $\mathcal{I}_{\bf C}(x) \subset \mathcal{I}(x)$, where $\mathcal{I}_{\bf C}(x)$ represents the irreducible class of $x$ with respect to the crossing sub-process $\mathcal X^{(0)}_{\bf C}$ starting at any given $x\in \mathbb Z$. Thus it is reasonable to attempt, using the below construction, to gain an understanding of the structure of $\mathcal{I}_{\bf C}$.
\begin{lem} Assume $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$, and, for any fixed $0 \leq r <\delta$, let us decompose $\mathcal{I}(r)$ into $\mathcal{I}^{+}(r)\cup \mathcal{I}^{-}(r)$ where $\mathcal{I}^{+}(r):= \mathcal{I}(r) \cap \mathbb Z_0^+$ and $\mathcal{I}^{-}(r):= \mathcal{I}(r) \cap \mathbb Z^-$. Set
\begin{align}\label{plusorbit}
\mathcal{I}_{\bf C}^+(r):= \{y \in \mathcal{I}^+(r): (y-S_{\mu})\cap \mathcal{I}^-(r) \neq \emptyset\},
\end{align}
and
\begin{align} \label{negativeorbit}
\mathcal{I}_{\bf C} ^-(r):= \{y \in \mathcal{I}^-(r): (y-S_{\mu'})\cap \mathcal{I}^+(r) \neq \emptyset\}.
\end{align}
Then $\mathcal{I}_{\bf C}(r)=\mathcal{I}_{\bf C}^+(r) \cup \mathcal{I}_{\bf C}^-(r)$ is an essential class of the crossing sub-process $\mathcal X_{\bf C}^{(0)}$. Furthermore, all the $X_{C_k}^{(0)}$ but finitely many belong to $\mathcal{I}_{\bf C}(r_{x_0})$ $\mathbb P$-a.s for any initial point $x_0\in \mathbb Z$.
\end{lem}
\begin{proof}
For any $x,y \in \mathcal{I}_{\bf C}(r)$, we write $x \leadsto y$ to indicate that the crossing process $\mathcal{X}^{(0)}$ starting at $x$, reaches $y$ (with a positive probability) at certain crossing time $C_k$. Equivalently, there is such $z \in \mathcal{I}(r)$ and $n \in \mathbb{N}$ that $x \overset{n}{\rightarrow} z \overset{1}{\rightarrow} y$ \footnote{The notation means that $p^{(n)}(x,z)>0$ and $p(z,y)>0$.}, where $y$ and $z$ have the opposite signs.
The attractive property is immediate from the definition, so it remains to check that $\mathcal{I}_{\bf C}(r)$ is an irreducible class for $\mathcal{X}_{\bf C}^{(0)}$, i.e. $x \leadsto y$ for any given $x,y \in \mathcal{I}_{\bf C}(r)$. Without loss of generality, we suppose $x,y \in \mathcal{I}_{\bf C}^+(r)$. There exists some $s \in S_{\mu}$ and $n\geq 0$ s.t. $y-s \in \mathcal{I}^{-}(r)$ and $p^{(n)}(x,y-s)>0$. Then in a single step from $y-s$, the crossing process $\mathcal{X}^{(0)}$ reaches $y$ with the probability $\mu(s)>0$ at a crossing time as desired.
\end{proof}
\begin{theo} \label{condition} Let $(a_i)_{i\geq 1}$ and $(b_j)_{j\geq 1}$ be strictly increasing sequences of positive integers. Set $S_\mu=(a_i)_{i\geq 1}$ and $S_{\mu'}=(-b_j)_{j\geq 1}$. For any $0\leq r <\delta$, the structure of $\mathcal{I}_{\bf C}(r)$ and its complement will be completely revealed in view of the following conditions:
\begin{enumerate}[(i).]
\item $D=+\infty \Longrightarrow \mathcal{I}_{\bf C}^-(r)=\mathcal{I}^-(r)$.
\item $D'=-\infty \Longrightarrow \mathcal{I}_{\bf C}^+(r)=\mathcal{I}^+(r)$.
\item If $-D'<+\infty$ and $D=+\infty$ then
$$\mathcal{I}_{\bf C}^+(r)=\mathcal{I}^+(r) \Longleftrightarrow \sup_{k \geq 1} \{a_k-a_{k-1}\} \leq -D' \text{ with } a_0=0.$$
\item If $D'=-\infty$ and $D<+\infty$ then
$$\mathcal{I}_{\bf C}^-(r)=\mathcal{I}^-(r) \Longleftrightarrow \sup_{\ell \geq 1}\{b_{\ell}-b_{\ell-1}\} \leq D \text{ with } b_0=0.$$
\item If $D<+\infty$ and $-D'<+\infty$ then
$$\mathcal{I}_{\bf C}^+(r)=\mathcal{I}^+(r) \Longleftrightarrow \underset{1\leq k \leq m}{\max}\{a_{k}-a_{k-1}\} \leq -D', $$
and
$$\mathcal{I}_{\bf C}^-(r)=\mathcal{I}^-(r) \Longleftrightarrow \underset{1\leq \ell \leq n}{\max}\{b_{\ell}-b_{\ell-1}\} \leq D, $$
where $a_0=b_0=0; D=a_m$ and $D'=-b_n$ for some $m,n\geq 1$.
\item Set $I^+:=\{k\geq 1 \mid a_k-a_{k-1} > -D'\}$ and $I^-:=\{\ell \geq 1 \mid b_{\ell}-b_{\ell-1}>D\}$ in case (v) is violated. For every choice of $k \in I^+$ and $\ell \in I^-$, we define
$$\mathcal{N}_k^+(r):=\left\{a_{k-1}+r+\delta s: 0\leq s\leq \dfrac{a_{k}-a_{k-1}+D'}{\delta}-1\right\},$$
and
$$\mathcal{N}_\ell^-(r):=\left\{-b_{\ell-1}+r+\delta s: \dfrac{b_{\ell-1}-b_\ell+D}{\delta}\leq s \leq -1\right\}.$$
Then $\mathcal{N}(r):= \underset{(k, \ell) \in I^+ \times I^-}{\bigcup} \mathcal{N}_k^+(r) \cup \mathcal{N}_\ell^-(r)$ is the set of \text{non-crossing} points.
\end{enumerate}
\end{theo}
\begin{proof} (i)-(ii). By definition.\\
(iii)-(iv). Note that \eqref{plusorbit} and \eqref{negativeorbit} can be rewritten as
\[
\mathcal{I}_{\bf C}^+(r)= \mathcal{I}^+(r) \cap \bigcup_{k\geq 1} A_k \text{ \rm with } A_k:=\{a_k+D'+r,\ldots,a_k+r-\delta\}
\]
and
\[
\mathcal{I}_{\bf C}^-(r)= \mathcal{I}^-(r) \cap \bigcup_{\ell\geq 1}B_{\ell} \text{ \rm with } B_{\ell}:=\{-b_{\ell}+r,\ldots,-b_{\ell}+ D+r-\delta\}.
\]
To cover $\mathcal{I}^+(r)$ by countably many same length sub-intervals, it requires $r \in A_1$ and further, no point in the form $r+\delta \mathbb Z$ stays inside the gap between $A_k$ and $A_{k+1}$ since $\{a_k+D'+r\}_{k\ge 1}$ is a strictly increasing sequence. More precisely,
\begin{align*}
\mathcal{I}_{\bf C}^+(r)=\mathcal{I}^+(r) &\Longleftrightarrow \left\{
\begin{array}{lll}
a_1+D'+r \leq r \\
a_{k+1}+D'+r\leq a_k+r, \forall k\geq 1 \\
\end{array}
\right.
&\Longleftrightarrow \left\{
\begin{array}{lll}
a_1\leq -D' \\
a_{k+1}-a_k\leq -D', \forall k\geq 1 \\
\end{array}
\right.
\end{align*}
Identically, we also infer
\begin{align*}
\mathcal{I}_{\bf C}^-(r)=\mathcal{I}^-(r) &\Longleftrightarrow \left\{
\begin{array}{lll}
-b_1+D+r-\delta \geq r-\delta \\
-b_\ell+r \leq -b_{\ell+1}+D+r, \forall \ell \geq 1 \\
\end{array}
\right.
&\Longleftrightarrow \left\{
\begin{array}{lll}
b_1\leq D \\
b_{\ell+1}-b_\ell \leq D, \forall \ell\geq 1. \\
\end{array}
\right.
\end{align*}
(v). This is a direct consequence of (iii) and (iv).\\
(vi). A straightforward argument yields
$$(\mathcal{N}_k^+(r)-S_\mu) \cap \mathcal{I}^-(r)= (\mathcal{N}_\ell^-(r)-S_{\mu'}) \cap \mathcal{I}^+(r)=\emptyset$$
for every pair $(k,\ell) \in I^+\times I^-$ and it is enough to end the proof.
\end{proof}
\begin{exam} We simply deal with the case when $D, -D'<+\infty$ by, for instance, taking $S_\mu:=\{2, 4, 10\}$ and $S_\mu':=\{-4, -1\}$. Obviously, $D=10, D'=-4, d=2, d'=1$ and the chain $\mathcal{X}^{(0)}$ has the unique essential class $\mathcal{I}(0):=\{-4,\ldots, 9\}$. An easy verification on (v) would lead to $\mathcal{I}_{\bf C}^-(0)=\mathcal{I}^-(0)$ while $\mathcal{I}_{\bf C}^+(0) \varsubsetneq \mathcal{I}^+(0)$ and its complement $\mathcal{N}_3^{+}=\{4, 5\}$ according to (vi). Hence, equality is achieved only by replacing $S_\mu$ by $\{2, 6, 10\}$.
\end{exam}
\begin{remark} We do emphasize that it fails to let $\mathcal{I}_{\bf C}^+(r) \varsubsetneq \mathcal{I}^+(r)$ and $\mathcal{I}_{\bf C}^-(r) \varsubsetneq \mathcal{I}^-(r)$ simultaneously occur due to (v). In other words, $\mathcal{I}_{\bf C}(r)$ always has at least one side which coincides with $\mathcal{I}(r)$. The case when $S_{\mu} \subset \mathbb Z$ and $S_{\mu'} \subset \mathbb Z$ is much complicated since the behaviour of the chain now is significantly affected by many factors. However, in connection with Theorem \ref{generalcase} and the above theorem, one may derive some properties of $\mathcal{I}_{\bf C}$, for instance, $\mathcal{I}_{\bf C}(r)=\mathcal{I}(r)$ if and only if $D=-D'=+\infty$.
\end{remark}
\subsection{Invariant measure for $\mathcal X_{\bf C}^{(0)}$}
Use of the explicit formula \eqref{discrete} of $\nu$ enables us to derive the invariant measure of the sub-process $\mathcal X_{\bf C}^{(0)}$. In particular
\begin{theo}
Assume $S_\mu \subset \mathbb Z^+$ and $S_{\mu'} \subset \mathbb Z^-$. Let $\rho$ be the measure on $\mathbb Z$ defined by
\begin{align}\label{formula_rho}
\rho(n):=\left\{
\begin{array}{lll}
\displaystyle \sum_{k=1}^{+\infty} \mu(k)\,\mu'[n-k+1,n] &{\rm if} &n \leq -1, \\
\displaystyle \sum_{k=1}^{+\infty} \mu'(-k)\,\mu[n+1,n+k] &{\rm if} &n \geq 0.
\end{array}
\right.
\end{align}
Then, for any $x_0\in \mathbb Z$, the restriction $\rho_{x_0}$ of $\rho$ to $\mathcal{I}_{\bf C}(r_{x_0})$
is an invariant measure for $\mathcal X_{\bf C}^{(0)}$.
\end{theo}
\begin{proof}
Consider the signed measures $\mathcal{A}$ and $\mathcal{A}'$ defined by $\mathcal{A}(m) = \delta_{0}(m) - \mu(m)$ if $m \geq 0$ and $\mathcal{A}'(m) = \delta_{0}(m) - \mu'(m)$ if $m \leq 0$. It is easily seen that
\begin{align} \label{convolution}
\mathcal{A}* \mathcal{U} = \mathcal{U}*\mathcal{A} = \delta_{0} \text{ and } \mathcal{A}'*\mathcal{U}' = \mathcal{U}'* \mathcal{A}' = \delta_{0}.
\end{align}
In addition, we also have
\begin{align} \label{rho-nu}
\rho(n) := \left\{
\begin{array}{lll}
\displaystyle \sum_{k = 0}^{+\infty}\mathcal{A}(k)\,\nu(n-k) &{\rm if} &n \leq -1,\\
\displaystyle \sum_{k = 0}^{+\infty}\mathcal{A}'(-k)\,\nu(n+k) &{\rm if} &n \geq 0.
\end{array}
\right.
\end{align}
Indeed, if $n \geq 0$ then we get
\begin{align*}
\displaystyle \sum_{k = 0}^{+\infty}\mathcal{A}'(-k)\,\nu(n+k) &= \displaystyle \sum_{k=0}^{+\infty} \left(\delta_{0}(-k) - \mu'(-k)\right)\, \nu(n + k)\\
&= (1 - \mu'(0))\,\nu(n) - \displaystyle \sum_{k=1}^{+\infty} \mu'(-k)\nu(n + k)\\
&= \displaystyle \sum_{k=1}^{+\infty} \mu'(-k)\,\left(\mu[n+1, +\infty[ - \mu[n+k+1, +\infty[ \right)\\
&=\displaystyle \sum_{k=1}^{+\infty} \mu'(-k)\,\mu[n+1,n+k].
\end{align*}
Conversely, $\nu$ can be represented in terms of $\rho$ and $\mathcal{U}'$ directly from \eqref{convolution} and \eqref{rho-nu}, that is, for $n \geq 0$
\begin{align*}
\displaystyle \sum_{k = 0}^{+\infty}\mathcal{U}'(-k)\,\rho(n + k) &= \displaystyle \sum_{k = 0}^{+\infty}\mathcal{U}'(-k)\,\displaystyle \sum_{\ell = 0}^{+\infty}\mathcal{A}'(-\ell)\,\nu(n + k + \ell) \\
&=\displaystyle \sum_{k=0}^{+\infty}\mathcal{U}'(-k)\,\displaystyle \sum_{s=k}^{+\infty}\mathcal{A}'(k-s)\nu(n+s)\\
&= \displaystyle \sum_{s = 0}^{+\infty} \nu(n+s)\, \underbrace{\displaystyle \sum_{k=0}^{s}\mathcal{U}'(-k)\,\mathcal{A}'(-s+k)}_{\delta_0(-s)} \\
&= \nu(n)
\end{align*}
and the same property holds for $n \leq -1$. Briefly, one may write
\begin{align} \label{nu-rho}
\nu(n) := \mathbb{1}_{]-\infty, 0[}(n)\left(\displaystyle \sum_{k = 0}^{+\infty}\mathcal{U}(k)\,\rho(n - k)\right)+\mathbb{1}_{[0, +\infty[}(n)\left(\displaystyle \sum_{k = 0}^{+\infty}\mathcal{U}'(-k)\,\rho(n + k)\right).
\end{align}
We claim that
\begin{align} \label{mainformula}
\nu_{x_0}(n) := \mathbb{E}_{\rho_{x_{0}}}\left(\displaystyle \sum_{j = 0}^{C_1 - 1} \mathbb{1}_{n}(X_j^{(0)}) \right),\,\,\,\,\,\,\,\,\, \text{if } n \in \mathcal{I}(r_{x_0})
\end{align}
where $\mathbb{E}_{\rho_{x_0}}(.) = \displaystyle \sum_{w \in \mathcal{I}_{\bf C}(r_0)} \rho_{x_0}(w)\,\mathbb{E}_{w}(.)$ indicates the expectation governed by $\rho_{x_0}$.\\
If $n \in \mathcal{I}^-(r_{x_0})$ then the crossing process can reach $n$ before the first crossing time if and only if $X_0^{(0)}=\omega \in \mathcal{I}_{\bf C}^-(r_{x_0})$ and $\omega \leq n$. In other words, there is $k \in \mathbb{Z}_0^+$ and $i \geq 1$ s.t. $S_i=k$ and $\omega=n-k$. Since $\displaystyle \sum_{j = 0}^{C_1 - 1} \mathbb{1}_{n}(X_j^{(0)}) = \displaystyle \sum_{j=0}^{+\infty} \mathbb{1}_{n}(X_j^{(0)})\,\mathbb{1}_{\{C_{1}>j\}}$,
\begin{align*}
\mathbb{E}_{\rho_{x_0}} \left(\displaystyle \sum_{j = 0}^{C_1 - 1} \mathbb{1}_{n}(X_j^{(0)}) \right) &= \displaystyle \sum_{k=0}^{+\infty} \rho_{x_0}(n - k)\, \mathbb{E}_{n - k} \left(\displaystyle \sum_{j=0}^{+\infty} \mathbb{1}_{n}(X_j^{(0)})\,\mathbb{1}_{\{C_{1}>j\}}\right) \\
&= \displaystyle \sum_{k=0}^{+\infty}\rho_{x_0}(n-k) \displaystyle \sum_{j=0}^{+\infty} \mathbb{P}_{n-k}[X_j^{(0)}=n, C_1 > j]\\
&=\displaystyle \sum_{k=0}^{+\infty}\rho_{x_0}(n-k) \displaystyle \sum_{j=0}^{+\infty} \mathbb{P}[S_j=k]\\
&= \displaystyle \sum_{k=0}^{+\infty}\rho_{x_0}(n - k)\, \mathcal{U}(k) \\
&= \nu_{x_0}(n).
\end{align*}
Hence, \eqref{mainformula} is true for every $n \in \mathcal{I}(r_{x_0})$ and yields that
$$\displaystyle \sum_{m \in \mathcal{I}(r_{x_0})} \nu_{x_0}(m)\,p(m, n) = \mathbb{E}_{\rho_{x_0}}\left(\displaystyle \sum_{j = 1}^{C_1} \mathbb{1}_{n}(X_j^{(0)}) \right).$$
Since $\nu_{x_0}$ is invariant on $\mathcal{I}(r_{x_0})$, i.e. $\nu_{x_0} = \nu_{x_0} \, P$, we again apply \eqref{mainformula} and simplify as
$$\mathbb{E}_{\rho_{x_0}} \left(\mathbb{1}_{n}(X_0^{(0)})\right) = \mathbb{E}_{\rho_{x_0}}\left(\mathbb{1}_{n}(X_{C_1}^{(0)})\right).$$
The left hand side is $\rho_{x_0}(n)$ and the right hand side is the sum $\displaystyle \sum_{m \in \mathcal{I}_{\bf C}(r_{x_0})} \rho_{x_0}(m)\, C(m, n)$, which prove that $\rho_{x_0}$ is an invariant measure for $ \mathcal X_{\bf C}^{(0)}$ on $\mathcal{I}_{\bf C}(r_{x_0})$.
\end{proof}
\section{Criteria for the recurrence of $ \mathcal{X}^{(\alpha)}$, $ 0\leq \alpha \leq 1$}
{\bf From now on, we assume $d=d'=1$} and keep in mind that the recurrence of $\mathcal X^{(\alpha)} $ always means the recurrence of the state $0$ since {\bf $\mathcal X^{(\alpha)} $ has an unique irreducible class in this case}.
\subsection{Classical approach}
In this subsection, we consider the first passage of $(S_n)_{n\geq 0}$ to the subset $]0, +\infty[$ and of $(S_n')_{n\geq 0}$ to the subset $]-\infty, 0[$, namely
\begin{align*}
\ell_+:=\inf\{k>0: S_k>0\} \text{ and } \ell_-':=\inf\{k>0: S_k'<0\}
\end{align*}
(with the convention $\inf \emptyset=+\infty$). The random variables $\ell^+$ and $\ell_-'$ are stopping times with respect to the canonical filtration $(\sigma(\xi_k, \xi_k', 1\leq k \leq n))_{n\geq 1}$ In the sequel, we
only consider cases when these random variables are $P$-as finite i.e. equivalently when
$\mathbb{P}[\limsup S_n=+\infty]= \mathbb{P}[\liminf S_n'=-\infty] =1$; hence, the random variables $S_{\ell_+}$ and $S_{\ell_-'}'$
are well defined in these cases and \textbf{we denote by $ \mu_+$ and $ \mu'_-$ their respective distributions
Let us define also, for $h\geq 1$, the renewal functions associated with the ladder heights $S_{\ell^+}$ and $S_{\ell^-}$, respectively by
\begin{align*}
& C(h):= \displaystyle \sum_{n=1}^{+\infty} \mathbb P[S_n=h, \min_{1\leq i\leq n}S_i >0],\\
& C'(-h):= \displaystyle \sum_{n=1}^{+\infty} \mathbb P[S_n'=-h, \min_{1\leq j\leq n}S_j' >0].
\end{align*}
We now get a glimpse of the following well-known criterion of recurrence of $ \mathcal{X}^{(\alpha)}$
\begin{theo} \text{(Kemperman)} The general oscillating random walk $\mathcal{X}^{(\alpha)}$ is recurrent if and only if
\begin{align}\label{Kemperman_cond}
\displaystyle \sum_{h=1}^{+\infty} C(h)\,C'(-h)=+\infty.
\end{align}
\end{theo}
However, it is quite theoretical and difficult to check in several cases. Next, we will take into consideration an equivalent condition to \eqref{Kemperman_cond}, which has the additional advantage of being easily computable.
\begin{theo} (\text{Rogozin-Foss}) If for some $\epsilon>0$
\begin{align}
\displaystyle \int_{-\epsilon}^{\epsilon} \dfrac{1}{\vert 1-\mathbb E[e^{itS_{\ell_+}}]\vert \, \vert 1-\mathbb E[e^{itS_{\ell_-'}'}]\vert }\, dt <+\infty,
\end{align}
then the original crossing process is transient.\\
If, in addition, $\text{Re }((1-\mathbb E[e^{itS_{\ell_+}}])(1-\mathbb E[e^{itS_{\ell_-'}'}])) \geq 0$ for $\vert t \vert <\epsilon$ for some $\epsilon>0$ and the below condition holds
\begin{align}
\displaystyle \int_{-\epsilon}^{\epsilon} \text{Re }\left(\dfrac{1}{(1-\mathbb E[e^{itS_{\ell_+}}]) \, (1-\mathbb E[e^{itS_{\ell_-'}'}])} \right) \,dt =+\infty,
\end{align}
then $\mathcal{X}^{(\alpha)}$ is recurrent.
\end{theo}
We also refer to the recent paper \cite{Bremont}, Proposition $4.4$ for a luminous proof.
In the simple case when $\mu=\mu$, in other words when $\mathcal{X}^{(\alpha)}$ is an homogeneous classical
random walk on $\mathbb Z$, these conditions turn into the above conditions turn into
\begin{align}
\displaystyle \int_{\epsilon}^{\epsilon} \vert 1-\hat{\mu}(t) \vert^{-1} dt < +\infty,
\end{align}
and
\begin{align}\label{Kesten}
\displaystyle \int_{\epsilon}^{\epsilon} Re\left(\dfrac{1}{1-\hat{\mu}(t)}\right) dt = +\infty,
\end{align}
In fact, \eqref{Kesten} is a necessary and sufficient
condition for $(S_n)_{n\geq 0}$ be recurrent; see \textsc{Kesten}, \textsc{Spitzer} \cite{KestenSpitzer}, \cite{Revuz} and \cite{Bremont} for proofs and comments. In \cite{Bremont}, Proposition $2.2$, the reader will find an explicit and simple relation between the integral of the function $\text{Re}\left(\dfrac{1}{1-\hat{\mu}(t)}\right)$ and the Green function of the random walk $(S_n)_n$ which enlightens the above statement.
In the next subsection we develop another approach to identify quite general conditions which ensure that $\mathcal X^{(\alpha)}$ is recurrent. We first consider the case when $S_\mu\subset \mathbb Z^+$ and $S_{\mu'}\subset \mathbb Z^-$ then the general case, replacing the couple $(\mu, \mu')$ by $( \mu_+, \mu'_-)$.
\subsection{ Tail condition criterion for the recurrence of $ \mathcal{X}^{(0)}$ when $S_\mu\subset \mathbb Z^+$ and $S_{\mu'}\subset \mathbb Z^-$}
An easy observation gives that the crossing sub-process $ \mathcal X_{\bf C}^{(0)}$ is positive recurrent on $\mathcal{I}_{\bf C}(0)$ $(\text{equivalently, }\rho(\mathcal{I}_{\bf C}(0))<+\infty)$ when $D$ and $D'$ are both finite. Thus, it is reasonable to study the recurrence of $ \mathcal X_{\bf C}^{(0)}$ in the non-trivial cases.
\begin{prop} \label{tail} Assume that $S_\mu\subset \mathbb Z^+$ and $S_{\mu'}\subset \mathbb Z^-$. Then the total mass of $\rho$ on $\mathcal{I}_{\bf C}(0)$ is finite if and only if
\begin{align}\label{totalmass}
\displaystyle \sum_{n = 0}^{+\infty} H(n)\,H'(-n) < +\infty,
\end{align}
where $H(n) = \mu]n,+\infty[$ and $H'(-n) = \mu']-\infty, -n[$ respectively stands for the tail distributions of $\mu$ and $\mu'$.
In this case, the Markov chain $\mathcal X_{\bf C}^{(0)}$ is positive recurrent and $\mathcal X^{(0)}(\mu, \mu')$ is recurrent on its essential class.
\end{prop}
\begin{proof}
We compute $\rho(\mathbb Z^+)$ by substituting the formula \eqref{formula_rho}
\begin{align*}
\displaystyle \sum_{n=0}^{+\infty}\rho(n) &= \displaystyle \sum_{n=0}^{+\infty}\displaystyle \sum_{k=1}^{+\infty}\mu'(-k)\,\mu[n+1,n+k]\\
&=\displaystyle \sum_{k=1}^{+\infty}\mu'(-k)\displaystyle \sum_{n=0}^{+\infty}[H(n)-H(n+k)] \\
&= \displaystyle \sum_{k=1}^{+\infty}\mu'(-k) \left[\displaystyle \sum_{n=0}^{k-1}H(n)- \displaystyle \lim_{N \to +\infty}\left(H(N+1)+ \ldots +H(N+k)\right)\right] \\
&= \displaystyle \sum_{n=0}^{+\infty}H(n) \displaystyle \sum_{k=n+1}^{+\infty}\mu'(-k)\\
&= \displaystyle \sum_{n= 0}^{+\infty}H(n)\,H'(-n).
\end{align*}
One also obtains $\displaystyle \sum_{n=-\infty}^{-1}\rho(n)=\displaystyle \sum_{n=0}^{+\infty}H(n)\,H'(-n)$ which immediately implies \eqref{totalmass}.
\end{proof}
Apparently, \eqref{totalmass} holds providing that the first moment of either $\xi_n$ or $-\xi_n'$ is finite. This assumption can be sharpen by constraining finite H{\"o}lder moments as below
\begin{coro} \label{Holder} Assume that $S_\mu\subset \mathbb Z^+, S_{\mu'}\subset \mathbb Z^-$ and $\mathbb{E}[\xi_1^{p}], \mathbb{E}[(-\xi_1')^q] <+\infty$ with $p, q \in ]0,1[$ satisfying $p+ q=1$. Then \eqref{totalmass} holds and the Markov chain $\mathcal X^{(0)}$ is recurrent on its unique essential class (and positive recurrent when $\mathbb{E}[\xi_1], \mathbb{E}[-\xi_1'] < +\infty$ by Corollary \ref{positive-recurrent}).
\end{coro}
\begin{proof}
The formula $\mathbb{E}[X^k]:=k \displaystyle \int_{0}^{+\infty}t^{k-1} \mathbb{P}[X\geq t]\,dt$ yields
\begin{align}\label{moment}
k\displaystyle \sum_{n=0}^{+\infty}(n+1)^{k-1}\mathbb{P}[X\geq n+1] \leq \mathbb{E}[X^k]
\end{align}
so that, by the \textit{Markov's inequality} for $H^{p}(n)$ and $H'^{q}(n)$, we get
\begin{align*}
\displaystyle \sum_{n=0}^{+\infty} H(n)\, H'(-n)&= \displaystyle \sum_{n=0}^{+\infty}H^{q}(n)H'^{p}(-n)\left[H^{p}(n)H'^{q}(-n)\right] \\
&\leq \mathbb{E}\left[\xi_1^p\right]^p\,\mathbb{E} [(-{\xi_1'})^q]^q \displaystyle \sum_{n=0}^{+\infty} \left[(n+1)^{-q^2}H^{q}(n) \, (n+1)^{-p^2}H'^{p}(-n) \right].
\end{align*}
The product inside the bracket can be transformed into sum by using the \textit{Young's inequality} and then together with \eqref{moment}, it yields
\begin{align*}
\displaystyle \sum_{n=0}^{+\infty} H(n)\, H'(-n) &\leq \mathbb{E}\left[\xi_1^p\right]^p\,\mathbb{E} [(-\xi_1')^q]^q \left(q \displaystyle \sum_{n=0}^{+\infty} (n+1)^{-q}H(n)+p \displaystyle \sum_{n=0}^{+\infty} (n+1)^{-p}H'(-n) \right)\\
&\leq \mathbb{E}\left[\xi_1^p\right]^p\,\mathbb{E} [(-\xi_1')^q]^q \left(\dfrac{q}{p} \mathbb{E}\left[\xi_1^p\right]+\dfrac{p}{q} \mathbb{E} [(-\xi_1')^q] \right)<+\infty.
\end{align*}
\end{proof}
\begin{remark}
The condition $\mathbb{E}[\xi_1^{p}], \mathbb{E}[(-\xi_1')^q] <+\infty$ with $p+ q=1$ is a sufficient condition for the recurrence of $\mathcal{X}^{(0)}$. Notice that it is not far to be sharp. We refer to Proposition $5.12$ in \cite{MarcWoess} for an example in the case of the reflected random walk on $\mathbb N$, which corresponds to the antisymmetric case, i.e. $S=-S'$ (with $p=q= 1/2$ there). The reader can find other examples in \cite{RogozinFoss} Theorem $2$ in the case when $S$ and $S'$ are stable random walks on $\mathbb Z$ but $S \neq -S'$.
\end{remark}
\subsection{Recurrence of $ \mathcal{X}^{0)}$ in the general case}
To treat this model, let us first introduce the basic decomposition of $\xi_n$ and $\xi_n'$, namely
$$\xi_n=\xi_n^+-\xi_n^- \quad \text{and } \quad \xi_n'=\xi_n'^+-\xi_n'^-,$$
where $\xi_n^{\pm}=\max\{\pm \xi_n,0\}$ and $\xi_n'^{\pm}=\max\{\pm \xi_n',0\}$.
Consider the following assumptions\\
$\bf{(H)} \,\, \left(\mathbb E[\xi_1^-]<\mathbb E[\xi_1^+] \leq +\infty\right)$ or $\left(\mathbb E[\xi_1^-]=\mathbb E[\xi_1^+]<+\infty\right);$\\
$\bf{(H')} \,\, \left(\mathbb E[(\xi_1')^+]<\mathbb E[(\xi_1')^-] \leq +\infty\right)$ or $\left(\mathbb E[(\xi_1')^-]=\mathbb E[(\xi_1')^+]<+\infty\right).$
\begin{lem} \label{abcd} If $\bf H\, (\text{resp. } \bf{H'})$ is satisfied then $\limsup\limits S_n=+\infty\, (\text{resp. } \liminf\limits S_n'=-\infty)$ almost surely.
\end{lem}
Hence, when both $\bf H$ and $\bf{H'}$ hold, the random variables $\ell_+$ and $\ell_-'$ are $\mathbb P$-a.s. finite; more generally, there are infinitely many $\mathbb P$-a.s. finite crossing times in this case. Let us introduce the ladder times $(t_k)_{k\geq 0}$ defined recursively by: $t_0=0$ and, for $k\geq 1$,
\begin{align}
t_{k+1}:=\left\{
\begin{array}{lll}
\inf\{n>t_k \mid \xi_{t_k+1}+\ldots+\xi_n>0\}& {\rm if} & X_{t_k}^{(0)}\leq -1, \\ \\
\inf\{n>t_k \mid \xi_{t_k+1}'+\ldots+\xi_n'<0\}&{\rm if} & X_{t_k}^{(0)} \geq 0.
\end{array}
\right.
\end{align}
Notice that, in the first line, the random variable $t_{k+1}$ is an ascending ladder epoch of $S$, while, in the second line, it is a descending ladder epoch of $S'$. These random times are $\mathbb P$-a.s. finite and the increments $(t_{k+1}-t_k)_{k\geq 0}$ form a sequence of independent random variables with laws
\begin{align*}
\mathcal{L}(t_{k+1}-t_k \mid X_{t_k}^{(0)}<0) =\mathcal{L}(t_1 \mid X_0^{(0)}=x) \quad \text{and} \quad
\mathcal{L}(t_{k+1}-t_k \mid X_{t_k}^{(0)}\geq 0) =\mathcal{L}(t_1 \mid X_0^{(0)}=y)
\end{align*}
for any $x < 0 \leq y$. Similarly,
$$ \mathcal{L}(S_{t_{k+1}}-S_{t_k} \mid X_{t_k}^{(0)}<0) =\mathcal{L}(S_{t_1} \mid X_0^{(0)}=x)= \mu_+, $$
and
$$ \mathcal{L}(S_{t_{k+1}}'-S_{t_k}' \mid X_{t_k}^{(0)}\geq 0) =\mathcal{L}(S_{t_1}' \mid X_0^{(0)}=y)= \mu'_-.
$$
It is perhaps worth remarking that, by setting $Y_k:=S_{t_k}-S_{t_{k-1}}$ when $X_{t_k}^{(0)} < 0$ and $Y_k':=S_{t_k}'-S_{t_{k-1}}'$ when $X_{t_k}^{(0)} \geq 0$, the sub-process $(X_{t_k}^{(0)})_{k\geq 0}$ turns out to be an crossing process associated with the distributions $ \mu_+$ and $ \mu'_-$ of $Y_k$ and $Y_k'$ respectively. In other words, the process $(X_{t_k}^{(0)})_{k\geq 0}$ has the same distribution as $\mathcal{X}^{(0)}( \mu_+, \mu'_-)$.
\begin{lem}\label{lem} Assume that both $\bf H$ and $\bf{H'}$ hold. Then, the oscillating process $ \mathcal X^{(0)}(\mu, \mu')$ is recurrent if and only if the oscillating process $\mathcal{X}^{(0)}( \mu_+, \mu'_-)$ is recurrent.
\end{lem}
A proof of this statement for the process $\mathcal{X}^{(1)}$ appears in the recent paper \cite{Bremont} of \textsc{J. Bremont}, lemma $4.2 \, (ii)$. For the sake of completeness, we detail the argument below,
introducing the first return time at $0$ of $\mathcal{X}^{(\alpha)}, 0 \leq \alpha \leq 1$, which will be useful latter on.
\begin{proof}
For any $0\leq \alpha \leq 1$, let $\uptau^{(\alpha)}$ be the first return time at $0$ of $\mathcal{X}^{(\alpha)}$ given by
$$\uptau^{(\alpha)}:=\inf\{n\geq 1: X_n^{(\alpha)}=0\}.$$
In the present proof, we only consider the case when $\alpha = 0$.
Starting at $0$, we know that $\uptau^{(0)}<+\infty$ almost surely and since $X_{\uptau^{(0)}}^{(0)}=0$, there are only two possibilities: if $X_{\uptau^{(0)}-1}^{(0)} \geq 1$ then $\uptau^{(0)}$ must be a ladder time while the case $X_{\uptau^{(0)}-1}^{(0)} \leq -1$ will imply the existence of some $k\geq 1$ s.t. $\uptau^{(0)}=C_k$, the $k^{th}$-crossing time of $\mathcal{X}^{(0)}$ and of course that $\uptau^{(0)}$ is also a ladder time. This means $(X_{t_k}^{(0)})_{k\geq 0}$ is recurrent ans so does $\mathcal{X}^{(0)}( \mu_+, \mu'_-)$. The converse is obvious.
\end{proof}
It is easily seen that $\mathcal{X}^{(0)}$ and $(X_{t_k}^{(0)})_{k\geq 0}$ admit a common crossing sub-process since there is at most a crossing moment between two consecutive ladder times $t_k$ and $t_{k+1}$ happening when $X_{t_k}^{(0)} <0$ and $X_{t_{k+1}}^{(0)}\geq 0$ or vice versa. Therefore, we can take advantage of Corollary \ref{Holder} (applied to the process $\mathcal{X}^{(0)}( \mu_+, \mu'_-)$) to deduce the recurrence of $(X_{t_k}^{(0)})_{k\geq 0}$ and finally that of $\mathcal{X}^{(0)}$ by Lemma \ref{lem}.
\begin{theo} \label{extend_condition} Let $p, q\in ]0,1[$ s.t. $p+q=1$. Then each of the following is sufficient for the oscillating process $\mathcal{X}^{(0)}(\mu, \mu')$ to be recurrent on its essential class
\begin{enumerate}[(a)]
\item $\left(\mathbb E[\xi_1^-]<\mathbb E[\xi_1^+],\, \mathbb E[(\xi_1^+)^p]<+\infty \right)$ and $\left(\mathbb E[(\xi_1')^+]<\mathbb E[(\xi_1')^-],\, \mathbb E[(\xi_1'^{-})^q]<+\infty \right)$;
\item $\left(\mathbb E[\xi_1^-]=\mathbb E[\xi_1^+],\, \mathbb E[(\xi_1^+)^{1+p}]<+\infty \right)$ and $\left(\mathbb E[(\xi_1')^+]=\mathbb E[(\xi_1')^-],\, \mathbb E[(\xi_1'^{-})^{1+q}]<+\infty \right)$;
\item $\left(\mathbb E[\xi_1^-]<\mathbb E[\xi_1^+],\, \mathbb E[(\xi_1^+)^p]<+\infty \right)$ and $\left(\mathbb E[(\xi_1')^+]=\mathbb E[(\xi_1')^-],\, \mathbb E[(\xi_1'^{-})^{1+q}]<+\infty \right)$.\\
The similar condition holds when swapping the roles of $\xi_1$ and $\xi_1'$.
\end{enumerate}
\end{theo}
\begin{proof}
As mentioned above, it remains to check that $\mathbb E[(Y_n)^p]<+\infty$ and $\mathbb E[(-Y_n')^q]<+\infty$. The set of conditions $(a)$ means that the chain moves with positive drift on the left and negative drift on the right while $(b)$ represents the center case which was already done by \textsc{Chow} and \textsc{Lai} (see \cite{ChowLai}). The others are partly mixed from both of $(a)$ and $(b)$, so we will leave the proof only for the first case.
Notice that $0<p<1$ and by the Wald's identity, we obtain
\begin{align*}
\mathbb E[(\xi_1+ \ldots+\xi_{t_1})^p \mid X_0^{(0)}=x<0] &\leq \mathbb E[({\xi_1^+}+ \ldots+{\xi_{t_1}^+})^p\mid X_0^{(0)}=x<0]\\
&\leq \mathbb E[(\xi_1^+)^p+\ldots +(\xi_{t_1}^+)^p\mid X_0^{(0)}=x<0]\\
&= \mathbb E[(\xi_1^+)^p] \, \mathbb E[t_1\mid X_0^{(0)}=x<0]<+\infty
\end{align*}
due to $\mathbb E[t_1 \mid X_0^{(0)}=x<0]<+\infty$. Indeed, if $\mathbb E[\xi_1^+]<+\infty$ then $\mathbb E[\xi_1]>0$ and $\mathbb E[\vert \xi_1\vert]<+\infty$ and the \textsc{Feller'}s result tells us that $\mathbb E[t_1\mid X_0^{(0)}=x <0]<+\infty$ (see \cite{Feller}). On the other hands, if $\mathbb E[\xi_1^+]=+\infty$ then there is $L>0$ s.t. $\xi_n^{(L)}=\min\{\xi_n, L\}$ (which has finite first moment) satisfies $\mathbb E[\xi_n^{(L)}]=\mathbb E[\xi_1^{(L)}]>0$. The first ascending ladder time $t_1^{(L)}$ associated with $S_n^{(M)}= \xi_1^{(L)}+\ldots + \xi_n^{(L)}$ has finite expectation by what we just said. Therefore, $t_1$ is integrable since $t_1 \leq t_1^{(L)}$. The argument showing $\mathbb E[(-\xi_1'-\ldots -\xi_{t_1}')^q \mid X_0^{(0)}=x\geq 0] <+\infty$ goes exactly the same line.
\end{proof}
\begin{remark}
As a direct consequence of the above proof and Corollary \ref{positive-recurrent}, we deduce that when $\mathbb E[\xi_1^-]<\mathbb E[\xi_1^+]<+\infty$ and $\mathbb E[(\xi_1')^+]<\mathbb E[(\xi_1')^-]<+\infty$ then the process $\mathcal{X}^{(0)}$ is positive recurrent. Indeed, in this case, $\mathbb{E}[\tau^{(0)}] <+\infty$ and the result follows by a classical theorem of induced processes.
\end{remark}
\subsection{Recurrence of $ \mathcal{X}^{(\alpha)}$ with $0\leq \alpha\leq 1$}
We end this section by proving our main result
\begin{coro} \label{recurrence_alpha} If at least one of the assumptions of Theorem \ref{extend_condition} is satisfied, then the general oscillating process $\mathcal X^{(\alpha)}(\mu, \mu')$ is recurrent.
\end{coro}
\begin{proof}
It is clear that $\mathcal{X}^{(0)}$ and $\mathcal{X}^{(1)}$ (suitably modified) are recurrent. Now, we have
\begin{align*}
\mathbb{P}_0[\uptau^{(\alpha)}=n] &=\mathbb{P}_0 [X_1^{(\alpha)}\neq 0, X_2^{(\alpha)} \neq 0,\ldots,X_{n-1}^{(\alpha)} \neq 0, X_n^{(\alpha)} =0 ]\\
&= \displaystyle \sum_{k=1}^{+\infty} \mathbb{P}_0[X_1^{(\alpha)}=k]\,\mathbb{P}_k [X_1^{(\alpha)}\neq 0, X_2^{(\alpha)} \neq 0,\ldots,X_{n-1}^{(\alpha)} \neq 0, X_{n-1}^{(\alpha)} = 0 ] \\
&+ \displaystyle \sum_{k=-\infty}^{-1} \mathbb{P}_0[X_1^{(\alpha)}=k]\,\mathbb{P}_k [X_1^{(\alpha)}\neq 0, X_2^{(\alpha)} \neq 0,\ldots,X_{n-1}^{(\alpha)} \neq 0, X_{n-1}^{(\alpha)} = 0]\\
&= \alpha\, \underbrace{\displaystyle \sum_{k\neq 0} \mu(k)\,\mathbb{P}_k[\uptau^{(1)}=n-1]}_{\mathbb{P}_0[\uptau^{(1)}=n]} +(1-\alpha)\,\underbrace{\displaystyle \sum_{k\neq 0} \mu'(k)\,\mathbb{P}_k[\uptau^{(0)}=n-1]}_{\mathbb{P}_0[\uptau^{(0)}=n]}.
\end{align*}
Summing over $n\geq 1$ in both sides, it readily implies $\mathbb{P}_0[\uptau^{(\alpha)}<+\infty]=1$ as expected.
\end{proof}
|
\section{Introduction}
In this paper, we investigate the graphs in which all balls are convex and the groups acting on them geometrically. The (local and global) convexity of balls is one of the fundamental features of
geodesic metric spaces, which are (locally or globally) nonpositively curved \cite{BrHa}. The graphs with convex
balls have been introduced and characterized in
\cite{FaJa,SoCh} as graphs without embedded isometric cycles of lengths different from 3 and 5 and in which all neighbors of a
vertex $u$ on shortest $(u,v)$-paths form a clique. One of their important subclass is the class of bridged
graphs: these are the graphs without embedded isometric cycles of length greater than 3 and they are exactly the graphs in which
the balls around convex sets are convex \cite{FaJa,SoCh}.
CAT(0) (alias nonpositively curved geodesic metric spaces), introduced by Gromov in his seminal paper \cite{Gr},
and groups acting on them are fundamental objects of study in metric geometry and geometric group theory.
Graphs with strong metric properties often arise as 1-skeletons of CAT(0) cell complexes. Gromov \cite{Gr} gave a nice
combinatorial characterization of CAT(0) cube complexes as simply connected cube complexes in which the links of vertices
are simplicial flag complexes. Based on this
result, \cite{Ch_CAT, Roller1998} established a bijection between the 1-skeletons of CAT(0) cube
complexes and the median graphs, well-known in metric graph theory \cite{BaCh_survey}. A similar characterization of CAT(0) simplicial
complexes with regular Euclidean simplices as cells seems impossible. Nevertheless, Chepoi \cite{Ch_CAT} characterized the simplicial complexes having bridged graphs as 1-skeletons as the simply connected simplicial complexes
in which the links of vertices are flag complexes without embedded 4- and 5-cycles. Januszkiewicz and Swiatkowski \cite{JS}, Haglund \cite{Haglund}, and Wise \cite{Wise}
rediscovered this class of simplicial complexes, called them {\it systolic complexes}, and used them in the context of geometric group theory. The convexity of balls around convex sets, which characterizes the 1-skeletons of
systolic complexes, is a fundamental geometric property of CAT(0) spaces.
Systolic complexes turned out to be good combinatorial analogs of CAT(0) metric spaces; cf. \cite{Elsner2009-flats,Elsner2009-isometries,Haglund,JS,JanuszkiewiczSwiatkowski2007,Pr3,Wise}. One of the main results
of \cite{JS} is that systolic groups (i.e., groups acting geometrically on systolic complexes) are biautomatic. For other results about systolic groups, see the papers \cite{HuaOsa1,Osajda2007,OsaPrzy2009,OsajdaPrytula,Pr2}. From the results of \cite{Ch_CAT,JS} it follows that systolic complexes are contractible. Metrically systolic complexes, which are metric analogs of systolic complexes, have been introduced and investigated in \cite{HuOs_ms}. Bridged graphs have also been further investigated in several graph-theoretical papers; cf.
\cite{AnFa, BaCh95, Ch_bridged,Po_bridged1,Po_bridged2} and the survey \cite{BaCh_survey}.
Weakly bridged/systolic graphs and complexes were introduced by Osajda \cite{Osajda} and have been thoroughly investigated in \cite{ChepoiOsajda}. The initial motivation of \cite{Osajda} was to exibit a class of simplicial complexes with some kind of
simplicial nonpositive curvature that will include the systolic complexes and some other classes of complexes appearing in geometric group theory. Examples
of weakly systolic groups (i.e., groups acting geometrically on weakly systolic complexes) which are not systolic were presented in \cite{Osajda}. The results of \cite{ChepoiOsajda}
show that weakly systolic complexes behave much like systolic complexes: their 1-skeletons are dismantlable and thus contractible, they satisfy the fixed simplex property, and can be characterized in the local-to-global way. It was shown in \cite{ChepoiOsajda} that the 1-skeletons of weakly bridged graphs are exactly the weakly modular graphs with convex balls. Weakly modular graphs are defined by two metric conditions, the triangle and quadrangle conditions \cite{BaCh95,Ch_metric} and contain as subclasses most of the classes of graphs investigated in metric graph theory \cite{BaCh_survey}. For example, bridged graphs are exactly the weakly modular graphs without induced $C_4$ and $C_5$ \cite{Ch_metric}. For a general theory of weakly modular graphs and their cell complexes, see the recent paper \cite{CCHO}.
Except the papers \cite{FaJa} and \cite{SoCh}, there are no other papers investigating graphs with convex balls in full generality (notice however that convexity of balls
occurs in the characterization of 1-hyperbolic graphs \cite{BaCh_hyp}). We believe that this is due to the technical difficulties arising from the presence in such graphs of pentagons that are not paved by triangles (such pentagons do not exist in bridged and weakly bridged graphs). For example, all $C_4$-free graphs of diameter 2 are graphs with convex balls. Hoffman and Singleton \cite{HoSi} proved that regular $C_3$-free such graphs exist only for degrees $2,3,7$, and possibly for degree $57$. The graph of degree 3 is the \emph{Petersen graph} and the graph of degree 7 is nowadays called the \emph{Hoffman-Singleton graph} and has 50 vertices and 175 edges. Notice also that any $C_4$-free graph may occur as an induced subgraph of a graph with convex balls (and diameter 2). This shows that at small scale the graphs with convex balls can be quite arbitrary.
In this paper, we present a systematic structural study of graphs with convex balls (which we abbreviately call \emph{CB-graphs}) and of groups acting on them geometrically
(which we call \emph{CB-groups}). Roughly speaking, we show that at large scale the CB-graphs behave like
weakly modular and bridged graphs. Similarly to weakly modular graphs, we characterize the CB-graphs via two metric conditions, the \emph{Triangle-Pentagon} and
the \emph{Interval Neighborhood} conditions $\mathrm{TPC}$ and $\mathrm{INC}$. In analogy with weakly modular graphs, which were characterized as graphs with strongly equilateral metric triangles \cite{Ch_metric}, we show
that in CB-graphs all metric triangles are either strongly equilateral or pentagons. We use this result to extend the approach of \cite{BaCh95} for weakly modular graphs
and prove a local-to-global Helly-type theorem for convex sets in CB-graphs. Namely, we show that the Helly number of such a graph $G$ equals the size of a largest Helly independent set of diameter 1 or 2 of $G$.
These metric properties and characterizations of CB-graphs are used to provide geometric and
local-to-global (topological) properties and characterizations. Namely, we characterize CB-graphs $G$ as graphs whose triangle-pentagonal complexes
$X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ are simply connected and balls of radius at most $3$ are convex. Consequently, we obtain the following discrete Cartan-Hadamard result:
if small balls (i.e. balls of radius 2 and 3) of a graph $G$ are convex, then the 1-skeleton of the universal cover of $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is a
CB-graph. Similarly to bridged and weakly bridged graphs, we prove
a dismantlability result for CB-graphs $G$: we show that their squares $G^2$ are dismantlable. This implies that the Rips complexes of CB-graphs
are contractible. Finally, we adapt and extend the approach of \cite{JS} for systolic groups and of \cite{ChChGeHiOs} for Helly groups, to show that the CB-groups
are biautomatic. Such a result was not yet known for weakly systolic groups. Summarizing, here is the list of
the main results of the paper and the sections where they are proved:
\begin{itemize}
\item $G$ is a CB-graph iff $G$ satisfies the conditions $\mathrm{TPC}$ and $\mathrm{INC}$ (Section \ref{sec: INCTPC}).
\item $G$ is a CB-graph iff its triangle-pentagon complex $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is simply connected and balls of radius at most $3$ in $G$ are convex (Section~\ref{sec: loc2glob}).
\item if $G$ is a CB-graph, then $G^2$ is dismantlable. Consequently, all Rips complexes $X_k(G)$, $k\ge 2$ are contractible and any automorphism of $G$ stabilizes a convex set of diameter 2 (Section~\ref{sec: dism}).
\item CB-groups are biautomatic (Section~\ref{sec: biauto}).
\item if $G$ is a CB-graph, then all metric triangles are strongly equilateral or pentagons (Section~\ref{sec: metriangle}).
\item if $G$ is a CB-graph and $h(G)$ is its Helly number, then $h(G)=h_2(G)$, where $h_2(G)$ is the size of a largest $h$-independent set of diameter $\le 2$ (Section~\ref{sec: helly}).
\end{itemize}
Additionally, we show that CB-graphs satisfy the falsification by fellow traveler property (FFTP). Namely, we show that for any
non-geodesic $(u,v)$-path $\gamma$ of a CB-graph $G$ there exists a shorter $(u,v)$-path $\eta$, such that $\gamma$ and $\eta$ 2-fellow travel. FFTP was initially introduced for Cayley graphs of
groups by Neumann and Shapiro \cite{NeSh}. In \cite{NeSh} and \cite{Elder02} it was shown that the groups satisfying FFTP have many strong properties, in particular, they are almost convex in
the sense of Cannon \cite{Cannon}. While convexity of balls obviously implies almost convexity, we do not know if any CB-group admits a set of generators with respect to which the Cayley graph
of the group is almost convex or satisfies FFTP. Finally note that CB-graphs and CB-groups are subclasses of shortcut graphs and shortcut groups, introduced and investigated by
Hoda \cite{Ho_shortcut}. Shortcut graphs are graphs for which there is an upper bound on the lengths of isometrically embedded cycles. Shortcut groups are groups acting geometrically
on shortcut graphs.
\section{Preliminaries}
\subsection{Graphs}
Recall that a graph $G=(V,E)$ consists of a set of vertices $V$ and a set of edges $E \subseteq \binom{V}{2}$. In this article, all graphs are simple, undirected and connected, but are not necessarily finite or locally-finite.
Nevertheless, in all results we will consider only \emph{graphs in which all cliques are finite}.
In some results, we will additionally assume that the graphs have uniformly bounded degrees. We write $u \sim v$ if the vertices $u$ and $v$ are adjacent in $G$ and $u\nsim v$ if $u$ and $v$ are not adjacent.
For every $n \geq 0$, $C_n$ denotes the cycle on $n$ vertices. We call an induced $C_3$ a \emph{triangle}, an induced $C_4$ a \emph{square} and an induced $C_5$ a \emph{pentagon}.
For a set $A\subseteq V$, we denote by $G[A]$ the subgraph of $G$ induced by $A$. Given a graph $H$, the graph $G$ is called \emph{$H$-free} if $G$ does not
contain an induced subgraph isomorphic to $H$.
We suppose that $G$ is endowed with the standard \emph{graph-distance} $d=d_G$ so that for every $u, v \in V$, $d(u,v)$ is the length of a shortest path between $u$ and $v$ in $G$.
The \emph{ distance} from a vertex $u$ to a set $C\subseteq V$ is $d(u,C):=\min \{ d(u,x): u\in C\}$.
For any two vertices $u,v\in V$, the \emph{interval} between $u$ and $v$ is defined by $I(u,v):=\{x \in V: d(u,v) = d(u,x) + d(x,v)\}$.
For every integer $r\geq 0$ and every $u\in V$, $B_r(u) := \{v \in V: d(u, v) \leq r\}$ denotes the \emph{ball } of radius $r$ centered at $u$, and $S_r(u) := \{v \in V: d(u, v) = r\}$ denotes
the \emph{sphere} of radius $r$ centered at $u$. We say that a subset $C \subseteq V$ is at \emph{uniform distance} from a vertex $u\in V$ if for every $x \in C$, we have $d(u,x) = d(u,C)$. More generally, two sets of vertices $C,C'\subseteq V$ are at uniform distance one from the other if for every $(x,y), (x',y')\in C\times C'$ we have $d(x,y)=d(x',y')$.
The \emph{ball of radius $r$ around a set $C$} is the set $B_r(C)=\{ v\in V: d(v,C)\le r\}$. The \emph{diameter} of a set $C\subset V$ is $\mathrm{diam}(C)=\sup\{ d(u,v): u,v\in C\}$. Recall also that the {\it $p$-power} of a graph $G$ is the graph $G^p$ having the same set of vertices as $G$ and two vertices $u,v$ are adjacent in $G^p$ if and only if $d_G(u,v) \le p$. Three vertices $u,v,w$ of a graph $G$ form a \emph{metric triangle} $uvw$ if $I(u,v)\cap I(u,w)=\sg{u},$ $I(u,v)\cap I(v,w)=\sg{v}$, and $I(u,w)\cap I(v,w)=\sg{w}$. A metric triangle $uvw$ is \emph{equilateral} when $d(u,v)=d(v,w)=d(w,u)$.
Let $H$ be a subgraph of $G$. We say that $H$ is an \emph{isometric subgraph} of $G$ if $d_H(u,v)=d_G(u,v)$ for any two vertices $u,v$ of $H$, i.e., if any two vertices $u, v$ of $H$ can be connected by a shortest path of $G$ totally included in $H$. A cycle $C$ is an \emph{isometric cycle} if $C$ is an isometric subgraph of $G$. A subset of vertices $C \subseteq V$ (or the subgraph $G(C)$ induced by $C$) is \emph{convex} in $G$ if for every two vertices $u, v \in C$, we have $I(u,v) \subseteq C$.
Since the intersection of convex sets is convex, for any set of vertices $S$ of $G$, there exists the smallest convex set $\mathrm{conv}(S)$ containing $S$, called the \emph{convex hull} of $S$. A weaker condition is $k$-convexity: for an integer $k\geq 2$, we say that $C \subseteq V$ is \emph{$k$-convex} in $G$ if for every two vertices $u, v \in C$ such that $d(u,v) \leq k$, we have $I(u, v) \subseteq C$. We will call a graph $G$ a \emph{graph with convex balls} (or a \emph{convex balls graph}, abbreviately, a \emph{CB-graph}) if all balls $B_r(v)$ of $G$ are convex for every $v\in V$ and $r\geq 1$. Similarly, we will call graphs whose all balls are $k$-convex for some fixed $k\geq0$ \emph{graphs with $k$-convex balls}. A graph is called \emph{bridged}
if all balls $B_r(C)$ around convex sets $C$ are convex. Finally, we will say that the convexity of a graph $G$ \emph{preserves the diameters of sets}
if for any set $S\subset V$, $\mathrm{diam}(\mathrm{conv}(S))=\mathrm{diam}(S)$. Clearly, bridged graphs have convex balls. It is also easy to show that if $G$ is a graph with convex balls, then the convexity of $G$ preserves the diameters of sets \cite{SoCh}.
\subsection{Cell complexes and group actions}
All complexes considered in this paper are CW complexes. Following
\cite[Chapter 0]{Hat}, we call them \emph{cell complexes} or just
\emph{complexes}. If all cells are simplices and the nonempty
intersections of two cells is their common face, then $X$ is called a
{\em simplicial complex}. For a cell complex $X$, by $X^{(k)}$ we
denote its \emph{$k$--skeleton}. All cell complexes considered in
this paper will have graphs (that is, one-dimensional simplicial
complexes) as their $1$--skeleta. Therefore, we use the notation
$G(X):=X^{(1)}$. The \emph{star} of a vertex $v$ ($0$--dimensional
cell) in a complex $X$, denoted $\mathrm{St}(v,X)$, is the set of all
cells containing $v$. Note that the star of a vertex is not
necessarily a cell complex. The \emph{closed star}
$\mathrm{cSt}(v,X)$ of $v$ in $X$ is the smallest subcomplex of
$X$ containing $\mathrm{St}(v,X)$. It consists of all cells $\sigma'$
such that there exists $\sigma \in \mathrm{St}(v,X)$ with
$\sigma' \subseteq \sigma$.
An {\em abstract simplicial complex} $X$ on a set $V$
is a set of nonempty subsets of $V$ such that each member of $X$, called a {\em simplex},
is a finite set, and any nonempty subset of a simplex is also a simplex.
A simplicial complex $X$ naturally gives rise
to an abstract simplicial complex $X'$ on the set of vertices of $X$
by: $U \in X'$ if and only if there is a simplex in $X$ having $U$ as its vertices.
Let $X$ be an abstract simplicial complex with ground set $V$. The \emph{barycentric subdivision} $\beta(X)$ of $X$
is the abstract simplicial complex on ground set $2^V \backslash \sg{\emptyset}$, whose vertices are the simplices
of $X$, and whose simplices are chains $\sigma_1 \subseteq \ldots \subseteq \sigma_k$ of simplices of $X$.
The {\it clique complex} of a graph $G$ is the abstract simplicial complex $X(G)$ having the cliques (i.e.,
complete subgraphs) of $G$ as simplices. A simplicial complex $X$ is a {\it flag complex} if $X$ is the clique
complex of its $1$--skeleton. The {\it $p$-Rips complex} $X_k(G)$ of a graph $G$ has the subsets $\sigma$
of vertices of $G$ such that $d_G(u,v)\le p$ for all $u,v\in \sigma$, as simplices.
Obviously, the $p$-Rips complex of $G$ is exactly the clique complex $X(G^p)$ of $G^p$. We also
define the \emph{triangle-pentagon complex} $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$
of a graph $G$ as a two-dimensional cell complex with $1$--skeleton $G$, and such that the two-cells
are (solid) triangles and pentagons whose boundaries are identified by isomorphisms
with (graph) triangles and pentagons in $G$.
As morphisms between cell complexes we consider all \emph{cellular maps}, i.e.,
maps sending (linearly) cells to cells. An \emph{isomorphism} is a bijective
cellular map being a linear isomorphism (isometry) on each cell. A
\emph{covering (map)} of a cell complex $X$ is a cellular surjection $p\colon
\widetilde{X} \to X$ such that $p|_{\mbox{St}(\ensuremath{\widetilde{v}},\widetilde{X})}\colon \mbox{St}(\ensuremath{\widetilde{v}},\widetilde{X})\to \mbox{St}(p(\ensuremath{\widetilde{v}}),X)$ is
an isomorphism for every vertex $\ensuremath{\widetilde{v}}$ in $\ensuremath{\widetilde{X}}$; compare \cite{Hat}*{Section 1.3}.
The space $\widetilde{X}$ is then called a \emph{covering space}.
A \emph{universal cover} of $X$ is a simply connected covering space $\widetilde{X}$. It
is unique up to isomorphism. In particular, if $X$ is simply connected, then
its universal cover is $X$ itself. Note that $X$ is connected iff $G(X)=X^{(1)}$ is connected, and $X$ is
simply connected (i.e., every continuous map $S^1\to X$ is null-homotopic) iff
$X^{(2)}$ is so.
A group $\Gamma$ \emph{acts by automorphisms} on a cell complex $X$ if there is a
homomorphism $\Gamma\to \mbox{Aut}(X)$ called an \emph{action of $\Gamma$}. The action is
\emph{geometric} (or \emph{$\Gamma$ acts geometrically}) if it is proper (i.e., cells
stabilizers are finite) and cocompact (i.e., the quotient $X/\Gamma$ is compact). In the current paper we usually consider geometric
actions on graphs, viewed as one-dimensional complexes. Namely, we say that a group $\Gamma$ acts on a graph $G$ when it acts on $G$ if we consider it as a $1$-dimensional simplicial complex. Observe that if $\Gamma$ acts on $G$, then it induces a group action of $\Gamma$ on its clique complex $X(G)$, and thus we also get an induced action of $\Gamma$ on $\beta(X(G))$. Note that this also induces an action of $\Gamma$ on $G^p$ for any $p \geq 1$, as well as an action of $\Gamma$ on the triangle-pentagon $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$. We call a group $\Gamma$ a \emph{$CB$-group} if $\Gamma$ acts geometrically on a CB-graph.
\subsection{The characterization of graphs with convex balls from \cite{SoCh} and \cite{FaJa}}
We recall the known structaral characterization of CB-graphs:
\begin{theorem}{\cite{SoCh,FaJa}}
\label{thm: well-bridged}
For a graph $G=(V,E)$ the following conditions are equivalent:
\begin{itemize}
\item[(i)] $G$ has convex balls;
\item[(ii)] $G$ has $3$-convex balls;
\item[(iii)] any isometric cycle of $G$ has length 3 or 5 and for any two vertices $u,v$ of $G$ the neighbors of $u$ in $I(u,v)$ form a clique. \end{itemize}
\end{theorem}
Although CB-graphs have no isometric
cycles of length other than $3$ and $5$, the converse is not always true. For example, the graph obtained by gluing a triangle and a pentagon
along an edge has some balls which are not convex. Hence, the condition (iii) of Theorem \ref{thm: well-bridged} cannot be relaxed.
\section{A metric characterization}
\label{sec: INCTPC}
In this section, we provide a new metric characterization of graphs with convex balls. this characterization will be very useful in subsequent sections. We also characterize graphs with 2-convex balls.
\subsection{Triangle-Pentagon and Interval Neighborhood Conditions}
Consider the following metric conditions on a graph $G$:
\begin{itemize}
\item \emph{Triangle Condition} ($\mathrm{TC}$): for any three vertices $v,x,y$ such that $d(v,x)=d(v,y)=k$ and $x\sim y$, there exists a vertex $z\in B_{k-1}(v)$ such that $xzy$ is a triangle of $G$;
\item\emph{Quadrangle Condition} ($\mathrm{QC}$): for any four vertices $v,x,y,u$ such that $d(v,x)=d(v,y)=k=d(v,u)-1$ and $u\sim x,y$, $x\nsim y$, there exists
a vertex $z\in B_{k-1}(v)$ such that $xzyu$ is a square of $G$;
\item \emph{Pentagon Condition} ($\mathrm{PC}$)=($\mathrm{PC}^0$): for any three vertices $v,x,y$ such that $d(v,x)=d(v,y)=k$ and $x\sim y$, there exist vertices $z,w,w'$ such that $z\in B_{k-2}(v)$ and $xwzw'y$ is a pentagon of $G$.
\end{itemize}
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\begin{scope}[xshift = 3cm, yshift= 6cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1)[draw=white, fill=white] {$\mathrm{TC}(v, xy)$};
\node (v) at (0,0) [label=below: $v$] {};
\node (x) at (-1,4) [label=above: $x$] {};
\node (y) at (1,4) [label=above: $y$] {};
\node (z) at (0,3) [red, label=right: $z$] {};
\draw[very thick] (x)--(y);
\draw[very thick, red] (x)-- (z)-- (y);
\draw[very thick, red, dashed] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, dashed] (v) -- (x) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (y) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k$};
\end{scope}
\begin{scope}[xshift = 7cm, yshift= 6cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (v) at (0,0) [label=below: $v$] {};
\node (u) at (0,4) [label=above: $u$] {};
\node (w) at (-1,3) [label=left: $x$] {};
\node (w') at (1,3) [label=right: $y$] {};
\node (z) at (0,2) [red, label=right: $z$] {};
\node (lab) at (0, -1)[draw=white, fill=white, opacity =0, text opacity=1] {$\mathrm{QC}(u,v)$};
\draw[very thick] (u)--(w);
\draw[very thick] (u)--(w');
\draw[very thick, red] (w') -- (z) -- (w);
\draw[very thick, dashed, red] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k$};
\end{scope}
\begin{scope}
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (v) at (0,0) [label=below: $v$] {};
\node (u) at (0,4) [label=above: $u$] {};
\node (w) at (-1,3) [label=left: $w$] {};
\node (w') at (1,3) [label=right: $w'$] {};
\node (lab) at (0, -1)[draw=white, fill=white, opacity =0, text opacity=1] {$\mathrm{INC}^0(u,v)$};
\draw[very thick] (u)--(w);
\draw[very thick] (u)--(w');
\draw[very thick, red] (w)-- (w');
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill = white, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (w') node [draw=white, right, midway, fill=white, opacity =0, text opacity=1]{$k$};
\end{scope}
\begin{scope}[xshift=5cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (v) at (0,0) [label=below: $v$] {};
\node (u) at (0,4) [label=above: $u$] {};
\node (w) at (-1,3) [label=left: $w$] {};
\node (w') at (1,3) [label=right: $w'$] {};
\node (z) at (0,2) [red, label=right: $z$] {};
\node (lab) at (0, -1)[draw=white, fill=white, opacity =0, text opacity=1] {$\mathrm{INC}(u,v)$};
\draw[very thick] (u)--(w);
\draw[very thick] (u)--(w');
\draw[very thick, red] (w)-- (w') -- (z) -- (w);
\draw[very thick, dashed, red] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k$};
\end{scope}
\begin{scope}[xshift=10cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (v) at (0,0) [label=below: $v$] {};
\node (u) at (0,4) [label=above: $u$] {};
\node (w) at (-1,3) [] {};
\node (w') at (1,3) [] {};
\node (z) at (0,2) [red, label=right: $z$] {};
\node (w'') at (0.33,3) [] {};
\node (w''') at (-0.33,3) [] {};
\node (lab) at (0, -1)[draw=white, fill=white, opacity =0, text opacity=1] {$\mathrm{INC}^+(u,v)$};
\draw[very thick] (u)--(w);
\draw[very thick] (u)--(w');
\draw[very thick] (u)--(w'');
\draw[very thick] (u)--(w''');
\draw[very thick, red] (z)--(w'');
\draw[very thick, red] (z)--(w''');
\draw[very thick, red] (w) -- (w''') -- (w'') -- (w') -- (z) -- (w);
\draw[very thick, red] (w) to[bend right] (w'');
\draw[very thick, red] (w''') to[bend right] (w');
\draw[very thick, red] (w) to[bend right] (w');
\draw[very thick, dashed] (v) -- (w'') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (w''') node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k$};
\draw[very thick, dashed, red] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\end{scope}
\end{tikzpicture}
\caption{Illustration of the Triangle and Quadrangle conditions and the variants of the Interval Neighborhood Condition. The red vertices and edges are implied by these conditions.} \label{fig: TCQC}
\end{figure}
We define the \emph{Interval Neighborhood Condition} $\mathrm{INC}^0$ and its stronger versions:
\begin{itemize}
\item $\left(\mathrm{INC}^0\right)$: for any two distinct vertices $u,v\in V$, the neighbors of $u$ in $I(u,v)$ form a clique.
\item $\left(\mathrm{INC}\right)$: $G$ satisfies $\mathrm{INC}^0$ and for any
two distinct vertices $u,v$ and any $w,w'\in S_1(u) \cap I(u,v)$ there exists a vertex $z\in I(w,v)\cap I(w',v)$ such that $z\sim w,w'$.
\item $\left(\mathrm{INC}^+\right)$: $G$ satisfies $\mathrm{INC}^0$ and for any
two vertices $u,v$ with $d(u,v)=k\ge 2$ there exists a vertex $z\in I(u,v)$ at distance $k-2$ from $v$ and adjacent to all vertices of
the clique $S_1(u) \cap I(u,v)$.
\end{itemize}
Note that $\mathrm{INC}^0$ occurs in the characterization (iii) of CB-graphs in Theorem \ref{thm: well-bridged} and that $\mathrm{INC}^0$ is equivalent to $2$-convexity of balls of $G$.
Note also that $u$ and $v$ do not play the same role in the above definitions. To make clear how we use these conditions, we will write that $G$ satisfies $\mathrm{INC}^0(u,v)$ (respectively $\mathrm{INC}(u,v)$ or $\mathrm{INC}^+(u,v)$)
when the condition $\mathrm{INC}^0$ (respectively $\mathrm{INC}$ or $\mathrm{INC}^+$) is satisfied for the pair of vertices $u$ and $v$. We will use the notation $\mathrm{INC}^0(v)$ (respectively $\mathrm{INC}(v)$ or $\mathrm{INC}^+(v)$) when $\mathrm{INC}(u,v)$ (respectively $\mathrm{INC}(u,v)$ or $\mathrm{INC}^+(u,v)$) is satisfied for every $u\in V\backslash \sg{v}$. Finally for every $k\geq 1$, we will use the notation $\mathrm{INC}_{\leq k}$ for the ``local version of $\mathrm{INC}$'', whenever the condition $\mathrm{INC}(u,v)$ holds for every pair of vertices $u,v$ such that $d(u,v)\leq k$.
We will also consider the following stronger versions of $\mathrm{PC}^0$:
\begin{itemize}
\item $\left(\mathrm{PC}^1\right)$: for any three vertices $v,x,y$ such that $d(v,x)=d(v,y)=k\geq 2$ and $x\sim y$ and for any neighbor $w\in I(x,v)$ of $x$ there exists a neighbor $w'\in I(y,v)$ of $y$ and $z \in B_{k-2}(v)$ such that $xwzw'y$ is a pentagon of $G$.
\item $\left(\mathrm{PC}^2\right)$: for any $v,x,y$ such that $d(v,x)=d(v,y)=k\geq 2$ and $x\sim y$, we have:
$B_2(x)\cap B_{k-2}(v)= B_2(y)\cap B_{k-2}(v).$
\item $\left(\mathrm{PC}^+\right)$: for any three vertices $v,x,y$ such that $d(v,x)=d(v,y)=k\geq 2$ and $x\sim y$, there exist a neighbor $w'\in I(y,v)$ of $y$ and $z \in B_{k-2}(v)$ such that for all neighbors $w\in I(x,v)$ of $x$, $xwzw'y$ is a pentagon of $G$.
\end{itemize}
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\begin{scope}[]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1)[draw=white, fill=white] {$\mathrm{PC}^0(v, xy)$};
\node (v) at (0,0) [label=below: $v$] {};
\node (x) at (-1,4) [label=above: $x$] {};
\node (y) at (1,4) [label=above: $y$] {};
\node (w) at (-1,3) [red, label=left: $w$] {};
\node (w') at (1,3) [red, label=right: $w'$] {};
\node (z) at (0,2) [red, label=right: $z$] {};
\draw[very thick] (x)--(y);
\draw[very thick, red] (x)-- (w) -- (z)-- (w') -- (y);
\draw[very thick, red, dashed] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-2$};
\draw[very thick, red, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, red, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\end{scope}
\begin{scope}[xshift=4cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1)[draw=white, fill=white] {$\mathrm{PC}^1(v, xy)$};
\node (v) at (0,0) [label=below: $v$] {};
\node (x) at (-1,4) [label=above: $x$] {};
\node (y) at (1,4) [label=above: $y$] {};
\node (w) at (-1,3) [label=left: $w$] {};
\node (w') at (1,3) [red, label=right: $w'$] {};
\node (z) at (0,2) [red, label=right: $z$] {};
\draw[very thick] (w) -- (x)--(y);
\draw[very thick, red] (w) -- (z)-- (w') -- (y);
\draw[very thick, red, dashed] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-2$};
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, red, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\end{scope}
\begin{scope}[xshift=8cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1)[draw=white, fill=white] {$\mathrm{PC}^2(v, xy)$};
\node (v) at (0,0) [label=below: $v$] {};
\node (x) at (-1,4) [label=above: $x$] {};
\node (y) at (1,4) [label=above: $y$] {};
\node (w) at (-1,3) [label=left: $w$] {};
\node (w') at (1,3) [red, label=right: $w'$] {};
\node (z) at (0,2) [label=right: $z$] {};
\draw[very thick] (z) -- (w) -- (x)-- (y);
\draw[very thick, red] (z)-- (w') -- (y);
\draw[very thick, red, dashed] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-2$};
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, red, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\end{scope}
\begin{scope}[xshift=12cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1)[draw=white, fill=white] {$\mathrm{PC}^+(v, xy)$};
\node (v) at (0,0) [label=below: $v$] {};
\node (x) at (-1,4) [label=above: $x$] {};
\node (y) at (1,4) [label=above: $y$] {};
\node (w) at (-1,3) [] {};
\node (w'') at (-1.5,3) [] {};
\node (w''') at (-0.5,3) [] {};
\node (w') at (1,3) [red, label=right: $w'$] {};
\node (z) at (0,2) [red, label=right: $z$] {};
\draw[very thick] (z) -- (w) -- (x)-- (y);
\draw[very thick] (z) -- (w'') -- (x)-- (w''') -- (z);
\draw[very thick, red] (z)-- (w') -- (y);
\draw[very thick, red, dashed] (v) -- (z) node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-2$};
\draw[very thick, dashed] (v) -- (w) node [draw=white, left, midway, fill, opacity =0, text opacity=1]{$k-1$};
\draw[very thick, dashed] (v) -- (w'') node [draw=white, left, midway, fill, opacity =0, text opacity=1]{};
\draw[very thick, dashed] (v) -- (w''') node [draw=white, left, midway, fill, opacity =0, text opacity=1]{};
\draw[very thick, red, dashed] (v) -- (w') node [draw=white, right, midway, fill, opacity =0, text opacity=1]{$k-1$};
\end{scope}
\end{tikzpicture}
\caption{Illustration of the four variants of the Pentagon Condition. The red vertices and edges are implied by these conditions.} \label{fig: TPC}
\end{figure}
Recall that a graph $G$ is called \emph{weakly modular} \cite{BaCh95, Ch_metric} if it satisfies $\mathrm{TC}$ and $\mathrm{QC}$. For any $i\in \sg{0,1,2,+}$, it will be useful to write $\mathrm{PC}^i(v,xy)$ if the condition $\mathrm{PC}^i$ is satisfied for the vertex $v$ and edge $xy$.
We say that a graph $G$ satisfies the \emph{Triangle-Pentagon Condition} $\mathrm{TPC}^i$ if for any three vertices $v,x,y$ such that $d(v,x)=d(v,y)=k$ and $x\sim y$ either $\mathrm{TC}(v,xy)$ or $\mathrm{PC}^i(v,xy)$ holds. Thus we will write $\mathrm{TPC}^i(v,xy)$ when either $\mathrm{TC}(v,xy)$ or $\mathrm{PC}^i(v,xy)$ holds. We will use the notation $\mathrm{TPC}^i(v)$ (respectively $\mathrm{TC}(v)$, $\mathrm{PC}^i(v)$) whenever $\mathrm{TPC}^i(v,xy)$ (respectively $\mathrm{TC}(v,xy)$, $\mathrm{PC}^i(v,xy)$) holds for $v$ and any edge $xy$ at uniform distance from $v$.
Finally for every $k\geq 1$, we will use $\mathrm{TPC}^i_{\leq k}$ to denote the ``local version of $\mathrm{TPC}^i$, i.e. whenever $\mathrm{TPC}^i(v,xy)$ holds for every $v,x,y$ such that the edge $xy$ is at uniform distance at most $k$ from $v$.
\subsection{Graphs with $2$-convex balls}
We start with a simple characterization of graphs with $2$-convex balls.
\begin{theorem}\label{thm: INC}
For a graph $G=(V,E)$, the following conditions are equivalent:
\begin{itemize}
\item[(i)] $G$ has $2$-convex balls;
\item[(ii)] $G$ satisfies $\mathrm{INC}$;
\item[(iii)] $G$ satisfies $\mathrm{INC}^+$.
\end{itemize}
\end{theorem}
\begin{proof}
By definition, a graph satisfying $\mathrm{INC}$ has $2$-convex
balls, thus (ii)$\Rightarrow$(i). Suppose now that $G$ has $2$-convex balls and we will show that
$G$ satisfies $\mathrm{INC}$. Pick any vertices
$u,v$ at distance $k+1\ge 2$ and two neighbors $w,w'$ of $u$ in the
interval $I(u,v)$. We proceed by induction on
$k=d(v,u)-1\geq1$. The $2$-convexity of $B_k(v)$ implies
that $w\sim w'$. It remains to show that $w$ and $w'$ have a common
neighbor $z$ at distance $k-1$ from $v$. If $k=1$, we can set
$z:=v$. If $k\geq 2$, we may assume thanks to the induction
hypothesis that $I(v,w)\cap I(v,w')=\sg{v}$. Let
$a\in I(v,w), b\in I(v,w')$ such that $a,b\sim v$. The $2$-convexity
of the ball $B_k(u)$ implies that $a\sim b$. Let $c$ be a neighbor
of $w'$ in $I(b,w')$. Then $d(a,w)=k-1$ and $d(a,c)\leq k-1$. Since
$I(v,w)\cap I(v,w')=\sg{v}$, we also have $d(a,w')=k$. By
$2$-convexity of the ball $B_{k-1}(a)$, we conclude that $c\sim w$,
which contradicts the assumption that $I(v,w)\cap
I(v,w')=\sg{v}$. This establishes that (i)$\Rightarrow$(ii).
Clearly, any graph satisfying $\mathrm{INC}^+$ satisfies $\mathrm{INC}$, thus (iii)$\Rightarrow$(ii).
Therefore, it suffices to show that (ii)$\Rightarrow$(iii).
Let $G$ be a graph without infinite cliques that satisfies $\mathrm{INC}$. To show that $G$ satisfies $\mathrm{INC}^+$, we use a
maximality argument: let $u,v\in V$ be two distinct vertices at distance $k+1:=d(u,v)\geq 2$ and assume by way of contradiction that there is no
vertex at distance $k-1= d(v,u)-2$ from $v$ which is adjacent to all
vertices of the clique $C:=S_1(u) \cap I(u,v)$.
Choose $z\in I(v,u)$ at distance $k-1$ from $v$ which is adjacent to a
maximum number of vertices of $C$. Denote by $C'$ the set of
neighbors of $z$ in $C$. By $\mathrm{INC}(u,v)$, $C'$ contains at least
two vertices, unless $|C|\leq 1$, in which case we are trivially done. On the other hand, there is a vertex
$x \in C\setminus C'$. Pick any $y \in C'$. By $\mathrm{INC}(u,v)$,
there exists a vertex $z'\sim x,y$ at distance $k-1$ from
$v$. Applying $\mathrm{INC}^0(y,v)$, we conclude that $z\sim z'$.
We assert that $z'$ is adjacent to all vertices of $C'$. Indeed, if
$y'\in C'$, then $zy'xz'$ is a $4$-cycle, which cannot be
induced. Since $z\nsim x$, necessarily $z'\sim y'$. Consequently,
$z'$ is adjacent to $x$ and all vertices of $C'$, contrary to the
maximality choice of $z$. Thus we proved that $G$ satisfies $\mathrm{INC}^+$, establishing
(ii)$\Rightarrow$(iii). \end{proof}
\begin{remark} A graph with 2-convex balls has no isometric cycles of even length $2k$: if $G$ contains such a cycle $C$, then for any vertex $v$ of $C$ the
ball $B_{k-1}(v)$ is not 2-convex. Indeed, if $u$ is the vertex of $C$ antipodal to $v$ and $x,y$ are its neighbors in $C$, then $x,y\in B_{k-1}(v)$ and $u\notin B_{k-1}(v)$.
The converse however is not true: let $G$ be obtained by gluing two 5-cycles along an edge $e$. Then the unique isometric cycles of $G$ are the two 5-cycles.
Also $G$ has diameter 4 and the unique diametral pair is the pair $u,v$, where $u$ and $v$ are the vertices of the two 5-cycles opposite to $e$. Then the ball $B_3(v)$ is not 2-convex because
it contains the neighbors of $u$ but not the vertex $u$.
\end{remark}
\subsection{Graphs with convex balls}
Now, we show that in full analogy with weakly modular graphs, the conditions $\mathrm{INC}$ and $\mathrm{TPC}^0$ as well as
their stronger versions characterize the CB-graphs.
\begin{theorem}
\label{thm: TPC}
For a graph $G$, the following conditions are equivalent:
\begin{itemize}
\item[(i)] $G$ has convex balls;
\item[(ii)] $G$ satisfies $\mathrm{INC}$ and $\mathrm{TPC}^0$;
\item[(iii)] $G$ satisfies $\mathrm{INC}$ and $\mathrm{TPC}^1$;
\item[(iv)] $G$ satisfies $\mathrm{INC}$ and $\mathrm{TPC}^2$.
\item[(v)] $G$ satisfies $\mathrm{INC}^+$ and $\mathrm{TPC}^+$.
\end{itemize}
\end{theorem}
\begin{proof}
We first prove the implication $(i) \Rightarrow (iii)$. Let $G$ be a CB-graph. By Theorem \ref{thm: INC}, $G$ satisfies $\mathrm{INC}$. We show that it also satisfies $\mathrm{TPC}^1$: let $v \in V$ be a vertex of $G$ and $x, y \in S_k(v)$ such that $x \sim y$. We may assume that $k\geq 2$. Suppose that $\mathrm{TC}(v,xy)$ does not hold, i.e., there is no $z \in B_{k-1}(v)$ such that $z \sim x, y$. We assert that in this case $\mathrm{PC}^1(v,xy)$ holds. Let $w$ be any neighbor of $x$ in $I(x,v)$. If $k=2$, then we can set $z:=v$ and take as $w'$ any common neighbor of $y$ and $v$.
Let now $k \geq 3$ and proceed by induction on $k$. Let $a$ be a neighbor of $v$ in $I(w,v)$. If $a\in I(v,y)$, then $d(a,x)=d(a,y)=k-1$ and we can apply the induction hypothesis to the vertex $a$ and the edge $xy$. Therefore, $d(a,y)=k$. Let $w'$ be a neighbor of $y$ in $I(y,v)$. Since $\mathrm{TC}(v,xy)$ does not hold, $w\ne w'$. Since $w,w'\in B_{k-1}(v)$ and $x,y\notin B_{k-1}(v)$, by convexity of $B_{k-1}(v)$, $(w,x,y,w')$ cannot be a shortest path. Thus $d(w,w')\le 2$. If $w \sim w'$, then we obtain a 4-cycle $xww'y$, which cannot be induced. Hence, either $w \sim y$ or $w'\sim y$, contrary to the assumption that $\mathrm{TC}(v,xy)$ does not hold. Therefore, $d(w,w')=2$. Now, let $c$ be a neighbor of $a$ in $I(a,w)$ (if $k=3$, then $c=w$). Then $d(w',c)\le d(w',w)+d(w,c)=2+k-3=k-1$. Since $d(v,w')=k-1$ and $a\in I(c,v)$, the convexity of the ball $B_{k-1}(w')$ implies that $d(w',a)\leq k-1$. Since $w'\sim y$ and $w'\nsim x$, we have $y\in I(w',x)$.
Since $d(a,y)=k$ and $x,w'\in B_{k-1}(a)$ we obtain a contradiction with the convexity of the ball $B_{k-1}(a)$. This establishes $\mathrm{TPC}^1$.
The implication $(iii)\Rightarrow (ii)$ is trivial. Now we prove the
implication $(ii) \Rightarrow (i)$. Let $G$ be a graph satisfying
$\mathrm{INC}$ and $\mathrm{TPC}^0$. By Theorem \ref{thm: well-bridged} $(iii)$, it
suffices to show that $G$ has no isometric cycles of length different
from $3$ and $5$ (as $\mathrm{INC}^0$ follows immediately from $\mathrm{INC}$). Let
$C$ be an isometric cycle of $G$ of length $n$. By $\mathrm{INC}^0$, $n$ must be
odd, say $n=2k+1\geq 7$ and let
$C=(v_0,v_1,\ldots,v_k,v_{k+1},\ldots,v_{2k})$. Pick $v=v_0$ of $C$
and let $x=v_k, y=v_{k+1}$, i.e., $xy$ is the (unique) edge of $C$
opposite to $v$. Then $d(v,x)=d(v,y)=k$. Let $w=v_{k-1}$ and
$w'=v_{k+2}$ be the second neighbors of $x$ and $y$ (respectively) in
$C$. Since $C$ is an isometric cycle, $d(v,w)=d(v,w')=k-1$ and
$d(w,w')=3$. By $\mathrm{TPC}^0(v,xy)$, there
exists a vertex $z\in I(x,v)\cap I(y,v)$ which has distance 1 or 2 to
both vertices $x$ and $y$. If $z\sim x,y$, then $w,z\in I(x,v)$ and
$w',z\in I(y,v)$, thus by $\mathrm{INC}^0$, either $z$ is adjacent to both
$w,w'$ or $z$ coincides with one of the vertices $w,w'$ and is
adjacent to the second one. In both cases, we conclude that
$d(w,w')\le 2$, which is impossible. Thus $d(z,x)=d(z,y)=2$ and
$d(z,v)=k-2$ and we can suppose that there is no vertex in
$I(x,v)\cap I(y,v)$ adjacent to $x$ and $y$. .
First suppose that $z$ is adjacent to one of the vertices $w,w'$, say $z\sim w$. Since $d(w,w')=3$, $z\nsim w'$. Let $t'$ be a common neighbor of $z$ and $y$.
Let $p=v_{k-2}$. Since $z,p$ belong to $I(w,v)$ and are both adjacent to $w$, by $\mathrm{INC}^0$, the vertices $z$ and $p$ are adjacent or coincide. If $z=p$,
then $d(p,y)=2$, which is impossible because $y=v_{k+1}$ and $p=v_{k-2}$ and whence $d(p,y)=3$. Hence $z\sim p$. Since $x,t'\in I(y,p)$ and are both adjacent to $y$,
by $\mathrm{INC}^0$, $x\sim t'$. Since $d(t',v)=k-1$, $t'$ belongs to $I(x,v)\cap I(y,v)$ and is adjacent to $x$ and $y$, contrary to our assumption.
Thus, we can suppose that any vertex of $I(x,v)\cap I(y,v)$ at distance 2 from $x$ and $y$ is not adjacent to one of the vertices $w$ and $w'$. Let $t\ne w$
be a common neighbor of $x$ and $z$ and $t'\ne w'$ be a common neighbor of $y$ and $z$. By our assumption, $t\nsim y$ and $t'\nsim x$. Since $w,t\in I(x,v)$ and $x\sim w,t$, by $\mathrm{INC}$, $w\sim t$ and there exists a common neighbor $q$ of $w$ and $t$ at distance $k-2$ from $v$. Since $q,z \in I(t,v)$, by $\mathrm{INC}^0$, $q \sim z$. Then $d(q,x)=2$ because $w\sim x,q$ and $d(q,t')=2$ because $z\sim q,t'$. Therefore, if $d(q,y)=3$, then $x,t'\in I(y,q)$ and $x,t'\sim y$, thus by $\mathrm{INC}^0$ the vertices $x$ and $t'$ must be adjacent.
Since we proved that $x\nsim t'$, we conclude that $d(y,q)=2$. Since $d(q,v)=k-2$, $q$ is a vertex of $I(x,v)\cap I(y,v)$ at distance 2 from $x$ and $y$ and adjacent to $w$, contrary to our assumption
that such vertex does not exists. This concludes the proof of the implication $(ii) \Rightarrow (i)$ and proves the equivalence $(i)\Leftrightarrow (ii)\Leftrightarrow (iii)$.
The implication $(iv)\Rightarrow (iii)$ is immediate. We now prove $(iii)\Rightarrow (iv)$. We only have to show that if $v,x,y \in V$ are such that $d(v,x)=d(v,y)=k\geq 2$, $x\sim y$ and $\mathrm{TC}(v,xy)$ does not hold, then $\mathrm{PC}^2(v,xy)$ holds. We let $z\in B_2(x)\cap B_{k-2}(v)$ and $w\sim x,w$. We will show that $z \in B_2(y)\cap B_{k-2}(v)$. As we assumed $(iii)$, $\mathrm{PC}^1(v,xy)$ holds, so there exists $z', w'\in V$ such that $xwz'w'y$ form a pentagon and $z' \in B_{k-2}(v)$. If $z=z'$, then we are immediately done. Otherwise by $\mathrm{INC}^0(w,v)$, $z'\sim z$. In particular we have $x,w'\in B_2(z)$. As $xwzw'y$ is a pentagon, $x\nsim w'$. Thus by $\mathrm{INC}^0(z)$ we must have $d(z,y)\leq 2$, which is the desired result. We proved that $B_2(x)\cap B_{k-2}(v)\subseteq B_2(y)\cap B_{k-2}(v)$, and by symmetry we are done.
Since the implication $(v)\Rightarrow (iii)$ is trivial, it remains to
show that $(iv) \Rightarrow (v)$. By Theorem~\ref{thm: INC}, $G$
satisfies $\mathrm{INC}^+$. Let $v \in V$ be a vertex of $G$ and
$x, y \in S_k(v)$ such that $x \sim y$. Suppose that $\mathrm{TC}(v,xy)$ does
not hold. By $\mathrm{INC}^+$, there exists a vertex $z$ at distance $k-2$
from $v$ adjacent to all vertices in $S_1(x)\cap I(x,v)$. By $(iv)$,
$d(z,y) = 2$ and thus there exists $w' \sim y,z$ such that $xwzw'y$ is
a pentagon for every $w \in S_1(x)\cap I(x,v)$.
This establishes $\mathrm{TPC}^+$ and concludes the
proof of the theorem.
\end{proof}
\begin{remark}
Observe that to prove the implication $(ii)\Rightarrow (i)$, we only needed to use that $G$ has $3$-convex balls. Thus we also get an alternative proof of the equivalence between items $(i)$ and $(ii)$ of Theorem \ref{thm: well-bridged}.
\end{remark}
\begin{remark}\label{rem: PC+_tight}
The condition $\mathrm{TPC}^1$ is tight, in the sense that we cannot require $z \in B_{k-2}(v)$ to be adjacent to arbitrary $w\in I(x,v), w\sim x$ and $w'\in I(y,v), w'\sim y$. The example drawn in Figure \ref{fig: ctreex} illustrates this. To see that the described graph has convex balls, note that the only pairs of vertices at distance $>2$ are $(v,x)$ and $(v,y)$. From this remark one can easily deduce that balls of radius at least $2$ are convex. On the other hand, balls of radius $1$ are also convex because $G$ does not contain induced $C_4$.
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\tikzstyle{every node}=[draw,circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (v) at (-1,0) [label=left: $x$] {};
\node (w) at (1,0) [label=right: $y$] {};
\node (11) at (-2,1) [label=left: $w$] {};
\node (12) at (-1,1) {};
\node (13) at (0,1) {};
\node (14) at (1,1) {};
\node (15) at (2,1) [label=right: $w'$] {};
\node (21) at (-1,2) {};
\node (22) at (1,2) {};
\node (u) at (0,3) [label=left: $v$] {};
\draw[thick] (v) -- (11) -- (12) -- (13) -- (11) -- (21) -- (22) -- (12) -- (21) -- (13) -- (14) -- (21) -- (u) -- (22) -- (14) -- (w) -- (15) -- (22) -- (13);
\draw[thick] (w) -- (v) -- (12);
\draw[thick] (14) -- (15);
\draw[thick] (11) to[bend right] (13);
\draw[thick] (13) to[bend right] (15);
\end{tikzpicture}
\caption{A graph with convex balls and $k:=d(v,x)=d(v,y)=3$. $\mathrm{TC}(v,xy)$ does not apply and for vertices $w\in I(x,v), w\sim x$ and $w'\in I(y,v), w'\sim y$ there is no $z \in B_{k-2}(v)$ adjacent to $w$ and $w'$.} \label{fig: ctreex}
\end{figure}
\end{remark}
\subsection{Pairs of intersecting pentagons}
As observed earlier, two pentagons glued along one common edge do not define a CB-graph. In particular, such pairs of pentagons cannot occur as isometric subgraphs
of CB-graphs. In this subsection, we metrically describe the subgraphs of CB-graphs induced by pairs of pentagons intersecting in at least two vertices. If $\pi=abcde$ is a pentagon of a graph $G$, then we call a vertex $x$ a \emph{universal vertex of $\pi$} if $x$ is adjacent to all the vertices of $\pi$.
We denote by $\mathrm{PT}$ the graph obtained by gluing a pentagon and a triangle along a common edge. Similarly we denote by $\mathrm{PP}_1$ the graph obtained by gluing two pentagons together along a common edge and by $\mathrm{PP}_2$ the graph obtained by gluing two pentagons together along a common path of length 2.
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\begin{scope}[xshift = 0cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1.5)[draw=white, fill=white] {$\mathrm{PT}$};
\node (1) at (0:1) {};
\node (2) at (60:1) {};
\node (3) at (120:1) {};
\node (4) at (180:1) {};
\node (5) at (240:1) {};
\node (6) at (300:1) {};
\draw[very thick] (1)--(2)--(3)--(4) --(5) --(6)--(1);
\draw[very thick] (3)--(5);
\end{scope}
\begin{scope}[xshift = 6cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (-0.5, -1.5)[draw=white, fill=white] {$\mathrm{PP}_1$};
\node (a) at (-1.75,1) {};
\node (b) at (-2.5,0) {};
\node (c) at (-1.75,-1) {};
\node (d) at (-0.5,-0.6) {};
\node (e) at (-0.5,0.6) {};
\node (f) at (0.75,1) {};
\node (g) at (1.5,0) {};
\node (h) at (0.75,-1) {};
\draw[very thick] (a)--(b)--(c)--(d)--(e)--(a);
\draw[very thick] (e)--(f)--(g)--(h)--(d);
\end{scope}
\begin{scope}[xshift = 11cm]
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (lab) at (0, -1.5)[draw=white, fill=white] {$\mathrm{PP}_2$};
\node (1) at (30:1) {};
\node (2) at (90:1) {};
\node (3) at (150:1) {};
\node (4) at (210:1) {};
\node (5) at (270:1) {};
\node (6) at (330:1) {};
\node (7) at (0:0) {};
\draw[very thick] (1)--(2)--(3)--(4) --(5) --(6)--(1);
\draw[very thick] (2)--(7)--(5);
\end{scope}
\end{tikzpicture}
\caption{The graphs $\mathrm{PT}$, $\mathrm{PP}_1$ and $\mathrm{PP}_2$.} \label{fig: PTPP}
\end{figure}
\begin{lemma}
\label{lem: deuxpent} Let $G$ be a graph in which $\mathrm{TPC}^0_{\leq3}$, $\mathrm{INC}_{\leq 3}$ hold and such that every induced subgraph isomorphic to $\mathrm{PP}_1$ has diameter at most $3$.
Let $\pi_1$ and $\pi_2$ be two pentagons of $G$ intersecting in at least
two vertices. Then either $\mathrm{diam}(\pi_1\cup \pi_2)=2$ or one of the pentagons $\pi_1$ or $\pi_2$ contains a universal vertex. \end{lemma}
\begin{proof}
Observe that if we look back at the proof of Theorem \ref{thm: TPC}, we can assume that the variants of $\mathrm{TPC}^0$ also apply when the distances involved are at most $3$. We also can use the $2$-convexity of balls of radius at most $2$.
If $\pi_1$ and $\pi_2$ intersect in two non-adjacent vertices, then using the $2$-convexity of balls of radius 2, one can conclude that $d(x,y)\le 2$ for any vertex $x\in \pi_1\setminus \pi_2$ and $y\in \pi_2\setminus \pi_1$.
Now, suppose that the pentagons $\pi_1$ and $\pi_2$ intersect in exactly two adjacent vertices, say $\pi_1=abcde$ and $\pi_2=defgh$.
First observe that $\pi_1\cup \pi_2$ is induced and isomorphic to $\mathrm{PP}_1$, hence it must have diameter at most $3$. If $\mathrm{diam}(\pi_1\cup \pi_2)=2$, then we are done. Thus, suppose that $\mathrm{diam}(\pi_1\cup \pi_2)=3$. If $d(b,g)=2$, by convexity of $B_2(b)$ and $B_2(g)$ we will conclude that $d(b,f)=d(b,h)=d(g,a)=d(g,c)=2$. Then, by $2$-convexity of $B_2(a)$ and $B_2(c)$, we conclude that $d(c,f) = d(a,h) = 2$, whence $\pi_1\cup \pi_2$ has diameter 2. Thus, $d(b,g)=3$. The $2$-convexity of $B_2(b)$ implies that $d(b,f)=3$ or $d(b,h)=3$. Analogously, the $2$-convexity of $B_2(g)$ implies that
$d(a,g)=3$ or $d(c,g)=3$.
First suppose that $d(b,f)=d(b,h)=3$. This implies that $d(g,a)=d(g,c)=3$. Indeed, suppose by way if contradiction that $d(g,c)=2$ and $d(g,a)=3$. Then, $e,g\in B_2(c)$ and $f\in I(e,g)$. By $2$-convexity of $B_2(c)$ we conclude that $d(c,f)\le 2$. But this is impossible because $a,c\in B_2(f)$ and $b\in I(a,c)\setminus B_2(f)$, contrary to the $2$-convexity of $B_2(f)$. Consequently, $d(g,a)=d(g,c)=3$. We apply $\mathrm{TPC}^1(b,fg)$. If $\mathrm{TC}$ applies, then there exists $z\sim f,g$ such that $d(b,z)=2$. By $\mathrm{INC}^0(b)$ this means that $z \sim e$. Hence we get $d(z,d)\leq 2$ and $d(b,z)=2$, so by $2$-convexity of $B_2(z)$ we must have $d(c,z)\leq 2$. Thus by $\mathrm{INC}^0(c)$ we get $z\sim h$. We get the last adjacency $z\sim d$ because the $4$-cycle $zedh$ cannot be induced, showing that $z$ is universal for pentagon $\pi_2=defgh$. If $\mathrm{PC}^1$ applies for neighbor $e$ of $f$, then there exists $z \sim b, e$ with $d(z,g)=2$. Then the $4$-cycle $abze$ cannot be induced and we get $z\sim a$. As $d(z,g)=2$ and $d(z,d)\leq 2$, $2$-convexity of $B_2(z)$ implies that $d(z,h)=2$. Thus by $\mathrm{INC}^0(h)$ we must have $z \sim c$. As the $4$-cycle $zbcd$ cannot be induced we get $z\sim d$, so $z$ is universal for pentagon $\pi_1=abcde$.
Therefore, we can suppose without loss of generality that $d(b,f)=3$ and $d(b,h)=2$. By analogy, we also conclude that one of the distances $d(g,a),d(g,c)$ is 3 and another one is 2. First, let $d(g,a)=3$ and $d(g,c)=2$ and apply $\mathrm{TPC}^1(b,fg)$. If $\mathrm{TC}$ applies, then there exists $z\sim f,g$ such that $d(b,z)=2$. By $\mathrm{INC}^0(b)$ this means that $z \sim e$ and $z\sim h$. As the $4$-cycle $ezhd$ cannot be induced, we conclude that $z\sim h$, showing that $z$ is universal for $\pi_2$. If $\mathrm{PC}^1$ applies with respect to the neighbor $e$ of $f$, then there exists $z \sim b, e$ with $d(z,g)=2$. Again, when considering the $4$-cycle $aezb$ we get $z\sim a$. Then $\mathrm{INC}^0(g)$ implies that $z \sim c$. Since the $4$-cycle $zcde$ cannot be induced, so $z\sim d$, whence $z$ is universal for $\pi_1$.
By previous cases and symmetric arguments, it remains to deal with the case where $d(b,f)=d(b,g)=d(c,g)=3$ and $d(b,h)=d(a,g)=2$.
First note that $d(c,f)=3$, otherwise $d(c,f)=2$ and the $2$-convexity of $B_2(c)$ would imply that $d(c,g)=2$, contradicting our hypothesis.
Hence we may apply $\mathrm{TPC}^+(b,fg)$ and $\mathrm{TPC}^+(c,fg)$. Assume first that $\mathrm{TC}(b,fg)$ applies, i.e. that there exists some $z\sim f,g$ with $d(b,z)=2$. Then because $d(b,h)=d(b,e)=2$, $\mathrm{INC}^0(b)$ gives us $z\sim h$ and $z\sim e$, and the last adjacency $z\sim d$ immediately follows when considering the $4$-cycle $ezhd$, hence $z$ is a universal vertex for $defgh$. Observe that we can make a similar reasonning when $\mathrm{TC}(c,fg)$ applies. By symmetry we are also done when $\mathrm{TC}(g,bc)$ or $\mathrm{TC}(f,bc)$ apply. Hence we may assume now that $\mathrm{PC}^+$ and $\mathrm{PC}^1$ respectively hold when we apply $\mathrm{TPC}^+(b,fg), \mathrm{TPC}^+(c,fg),\mathrm{TPC}^1(g,bc)$ and $\mathrm{TPC}^1(h,bc)$.
The application of $\mathrm{PC}^+(b,fg)$ gives us a vertex $s\sim b$ such that $d(s,f)= d(s,g)=2$ and $s$ is adjacent to every vertex of the clique $B_1(g)\cap B_2(b)$. As $d(a,g)=2$, $\mathrm{INC}(g)$ gives us $s\sim a$. Applying $\mathrm{PC}^+(c,fg)$, we similarly find a vertex $t\sim c$ such that $d(t,f)=d(t,g)$ and $t$ is adjacent to every vertex of $B_1(f)\cap B_2(c)$. Now the applications of $\mathrm{PC}^1(g,bc)$ and $\mathrm{PC}^1(f,bc)$ to $b \sim c$ with respect to $t$ and $s$ give us vertices $u\sim g,t$ and $v\sim f,s$ such that $d(u,b)=d(u,c)=d(v,b)=d(v,c)=2$. Thanks to one of the previous cases (when $\mathrm{TC}$ applies somewhere), we may assume that $s\neq t$ and $u\neq v$. By definition of $s$, we must have $s\sim u$. By definition of $t$ we get $t\sim v$.
Hence $svtu$ form a $4$-cycle, which implies that $s\sim t$ or $u\sim v$. Observe that these two adjacencies create the $4$-cycles $bstc$ or $uvfg$, which ultimately imply that $s\sim c$ or $t\sim b$ or $v\sim g$ or $u\sim f$. Any of these four adjacencies correspond to one of the applications of $\mathrm{TC}$ that we already covered, concluding the proof; see Figure \ref{fig: lastcase} for an illustration of the last case.
\end{proof}
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\tikzstyle{every node}=[draw,circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (a) at (-2,2) [label=above: $a$] {};
\node (b) at (-3,1) [label=left: $b$] {};
\node (c) at (-2,0) [label=below: $c$] {};
\node (d) at (-0.5,0.25) [label=below: $d$] {};
\node (e) at (-0.5,1.75) [label=above: $e$] {};
\node (f) at (1,2) [label=above: $f$] {};
\node (g) at (2,1) [label=right: $g$] {};
\node (h) at (1,0) [label=below: $h$] {};
\node (s) at (-1.5,1.25) [label=above: $s$] {};
\node (t) at (-1.5,0.75) [label=below: $t$] {};
\node (u) at (0.5,0.75) [label=below: $u$] {};
\node (v) at (0.5,1.25) [label=above: $v$] {};
\draw (e) -- (a) -- (b) -- (c) -- (d) -- (e) -- (f) -- (g) -- (h) -- (d);
\draw (b) -- (s) -- (v) -- (t) -- (c);
\draw (f) -- (v);
\draw (s) -- (u) -- (g);
\draw (t) -- (u);
\draw (s) -- (t)[dashed];
\draw (u) -- (v)[dashed];
\draw (b) -- (t)[dashed];
\draw (c) -- (s)[dashed];
\draw (v) -- (g)[dashed];
\draw (u) -- (f)[dashed];
\end{tikzpicture}
\caption{The case $d(b,f)=d(g,c)=3$ and $d(b,h)=d(g,a)=2$ of the proof of Lemma \ref{lem: deuxpent}.
The dashed edges occur at the end of the proof (not necessarily all of them).} \label{fig: lastcase}
\end{figure}
\begin{example} \label{CB-pentagons-not-in-W5}
Lemma \ref{lem: deuxpent} is tight in the following sense: there exist CB-graphs with two pentagons glued together along one common edge having diameter $3$ and
such that only one of the two pentagons has a universal vertex; see Figure \ref{fig: diameter3notwm} for an illustration of such a graph. The pentagon $deabc$
is not included in any 5-wheel.
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.75]
\tikzstyle{every node}=[draw,circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (a) at (0,0) [label=below: $a$] {};
\node (b) at (3,0) [label=below: $b$] {};
\node (c) at (3,2) [label=right: $c$] {};
\node (d) at (3,4) [label=right: $d$] {};
\node (e) at (0,2) [label=left: $e$] {};
\node (f) at (0,4) [label=left: $f$] {};
\node (g) at (0,6) [label=above: $g$] {};
\node (h) at (3,6) [label=above: $h$] {};
\node (s) at (1.5,4) {};
\node (t) at (1.5,2) {};
\draw[very thick, magenta] (e) -- (a) -- (b) -- (c) -- (d) -- (e) -- (f) -- (g) -- (h) -- (d);
\draw[thick] (g)--(s)--(f);
\draw[thick] (e)--(s)--(d);
\draw[thick] (h)--(s)--(t);
\draw[thick] (s)--(c)--(t)--(f);
\draw[thick] (t)--(b);
\end{tikzpicture}
\caption{A CB-graph of diameter $3$ consisting of two pentagons (in magenta) and of two additional vertices. Only the upper pentagon has a universal vertex.} \label{fig: diameter3notwm}
\end{figure}
\end{example}
\begin{example}\label{CB-pentagons-infinite}
The first graph from Figure \ref{fig: infinitegraph} shows that a property similar to Lemma \ref{lem: deuxpent} does not hold if we glue two pentagons to two opposite edges of a $C_4$ with a diagonal: it has diameter $4$, it has convex balls, and none of its pentagons admit a universal vertex. This graph is also interesting because we can extend this construction to find an infinite $2$-connected CB-graph with an infinite number of pentagons and such that no pentagon has a universal vertex. Clearly, all such CB-graphs are not weakly systolic. The second graph from Figure \ref{fig: infinitegraph} shows how to get such a CB-graph. It it obtained by merging two copies of the first graph by identifying the vertices of one common pentagon from each copy. One can extend this construction by gluing an infinite number of pentagons, in a path-like way, where each connection between three successive pentagons is described by the second graph of Figure \ref{fig: infinitegraph}.
\end{example}
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\tikzstyle{every node}=[draw,circle,fill=black,minimum size=4pt,
inner sep=0pt]
\begin{scope}[yshift=0cm]
\node (a) at (-2,2) {};
\node (b) at (-3,1) {};
\node (c) at (-2,0) {};
\node (d1) at (-0.5,0.25) {};
\node (d2) at (0.5,0.25) {};
\node (e1) at (-0.5,1.75) {};
\node (e2) at (0.5,1.75) {};
\node (f) at (2,2) {};
\node (g) at (3,1) {};
\node (h) at (2,0) {};
\node (s) at (-1.5,1) {};
\node (v) at (1.5,1) {};
\draw (e1) -- (a) -- (b) -- (c) -- (d1) -- (e1) -- (e2) -- (d2) --(d1);
\draw (e1) -- (d2) -- (h) -- (g) -- (f) -- (e2);
\draw (b) -- (s) -- (a) -- (e1) -- (s) -- (d2) -- (e1) -- (v) -- (d2) -- (h) -- (v) -- (g);
\end{scope}
\begin{scope}[xshift=-1.5cm, yshift=-5cm]
\node[draw,circle,fill=green,minimum size=4pt, inner sep=0pt] (a) at (-2,2) {};
\node[draw,circle,fill=green,minimum size=4pt, inner sep=0pt] (b) at (-3,1) {};
\node[draw,circle,fill=green,minimum size=4pt, inner sep=0pt] (c) at (-2,0) {};
\node[draw,circle,fill=green,minimum size=4pt, inner sep=0pt] (d1) at (-0.5,0.25) {};
\node[draw,circle,fill=magenta,minimum size=4pt, inner sep=0pt] (d2) at (0.5,0.25) {};
\node[draw,circle,fill=green,minimum size=4pt, inner sep=0pt] (e1) at (-0.5,1.75) {};
\node[draw,circle,fill=magenta,minimum size=4pt, inner sep=0pt] (e2) at (0.5,1.75) {};
\node[draw,circle,fill=magenta,minimum size=4pt, inner sep=0pt] (f1) at (2,2) {};
\node[draw,circle,fill=blue,minimum size=4pt, inner sep=0pt] (f2) at (2.5,2.5) {};
\node[draw,circle,fill=magenta,minimum size=4pt, inner sep=0pt] (g1) at (3,1) {};
\node[draw,circle,fill=blue,minimum size=4pt, inner sep=0pt] (g2) at (3.5,1.5) {};
\node[draw,circle,fill=blue,minimum size=4pt, inner sep=0pt] (i) at (5,2.25) {};
\node[draw,circle,fill=blue,minimum size=4pt, inner sep=0pt] (j) at (4.75,3.75) {};
\node[draw,circle,fill=blue,minimum size=4pt, inner sep=0pt] (k) at (3.25,3.75) {};
\node[draw,circle,fill=magenta,minimum size=4pt, inner sep=0pt] (h) at (2,0) {};
\node[draw,circle,fill=green,minimum size=4pt, inner sep=0pt] (s) at (-1.5,1) {};
\node[draw,circle,fill=magenta,minimum size=4pt, inner sep=0pt] (v) at (1.5,1) {};
\node[draw,circle,fill=blue,minimum size=4pt, inner sep=0pt] (w) at (3.85,2.75) {};
\draw (e1) -- (a) -- (b) -- (c) -- (d1) -- (e1) -- (e2) -- (d2) --(d1);
\draw (e1) -- (d2) -- (h) -- (g1) -- (f1) -- (e2);
\draw (b) -- (s) -- (a) -- (e1) -- (s) -- (d2) -- (e1) -- (v) -- (d2) -- (h) -- (v) -- (g1);
\draw (f1)--(f2)--(k)--(j)--(i)--(g2)--(g1)--(f2)--(g2);
\draw (g1)--(w)--(f2)--(k)--(w)--(j);
\draw (v) -- (f2);
\end{scope}
\end{tikzpicture}
\caption{The above graph shows an elementary gluing of two pentagons with the help of one layer of triangles. The beneath graph shows how to merge two copies of the first one by identifying $6$ common vertices, that are represented in magenta in the figure.}\label{fig: infinitegraph}
\end{figure}
\subsection{Triangle-free CB-graphs}
As mentioned in the introduction, every graph of diameter $2$ and girth $5$ is a CB-graph. In particular, the Petersen graph and the Hoffman-Singleton graph are CB-graphs. It is also easy to construct CB-graphs of arbitrary diameter and of girth $5$. Namely, if $G_1$ and $G_2$ are two CB-graphs, then the \emph{wedge graph} $G_1\bigvee G_2$
obtained by identifying a single vertex of $G_1$ with a single vertex of $G_2$ also has convex balls. However, the graphs constructed in this way are not $2$-connected. The next result shows that this is the unique way
of producing CB-graphs of girth $5$ and arbitrary diameter.
\begin{proposition}
\label{thm: Moore}
The finite $2$-connected triangle-free CB-graphs are exactly the Moore graphs of diameter $2$. Consequently, every block in a finite triangle-free CB-graph is either a pentagon or the Petersen graph
or the Hoffman-Singleton graph or (if it exists) a Moore graph of diameter $2$ and degree $57$.
\end{proposition}
\begin{proof} To prove Theorem \ref{thm: Moore} we will use the following result of Stemple \cite{Stemple} about geodetic graphs of diameter $2$ (recall that a graph $G$ is \emph{geodetic} if there exists a unique shortest path
between any pair of vertices of $G$). We use its formulation from \cite{Blokhuis1988}.
\begin{theorem}[\cite{Stemple}]
\label{thm: Stemple}
Let $G$ be a diameter $2$ geodetic graph. Then $G$ is of one of the following three types:
\begin{itemize}
\item[$(i)$] $G$ contains a vertex adjacent to all other vertices.
\item[$(ii)$] $G$ is regular.
\item[$(iii)$] exactly two different degrees occur in $G$, say $a<b$, and if $B$ denotes the set of vertices of degree $b$, then it contains a clique of size $b-a+2>2$.
\end{itemize}
\end{theorem}
It is easy to observe that if $G$ satisfies item $(i)$, then it is either a star, in which case it is not $2$-connected, or it has a triangle. Moreover if item $(iii)$ is satisfied, then $G$ clearly has a triangle.
Hence to prove Theorem \ref{thm: Moore}, it remains to show the following claim:
\begin{claim}\label{diameter2}
Any finite 2-connected triangle-free CB-graph $G$ has diameter at most $2$.
\end{claim}
\begin{proof} Suppose by way of contradiction that $\mathrm{diam}(G)\ge 3$ and let $v_0$ be a vertex of eccentricity at least 3, i.e., a vertex for which there exists a vertex $u$ with $d(v_0,u)\ge 3$.
Consider a BFS-ordering of the vertices of $G$ starting from $v_0$ (see Section \ref{sec: dism} for definitions). For a vertex $x$, we denote by $f(x)$ its parent in this BFS-order.
We call an edge $uv$ of $G$ \emph{horizontal} if $d(v_0,u)=d(v_0,v)$ and \emph{vertical} if $d(v_0,u)\ne d(v_0,v)$. We call a path $P$ of $G$ \emph{horizontal} (respectively, \emph{vertical})
if all its edges are horizontal (respectively, vertical).
Since $G$ is a triangle-free CB-graph, $G$ satisfies the following properties:
\begin{itemize}
\item[(1)] If $d(v_0,v)=i$ for some $v\in V$, then $v$ has a unique neighbor $f(v)$ at distance $i-1$ to $v_0$.
\item[(2)] If $\pi_1$ and $\pi_2$ are two pentagons of $G$ sharing an edge, then $\mathrm{diam}(\pi_1\cup \pi_2)=2$.
\item[(3)] If $Q=(x_1,x_2,\ldots,x_k)$ is a horizontal path, then $f^2(x_1)=f^2(x_2)=\ldots=f^2(x_k)$.
\item[(4)] If $P=(v_0=u_1,u_2,\ldots,u_k=u)$ is a shortest $(v_0,u)$-path, then all edges of $P$ are vertical and no consecutive vertices $u_i,u_{i+1}$ of $P$ are incident to horizontal edges of $G$.
\end{itemize}
Indeed, (1) is an immediate consequence of the geodecity of $G$, (2) follows from Lemma \ref{lem: deuxpent}, and (3) is obtained by applying ($\mathrm{PC}$) to the edges of the path $P$ and by property (1). The first part of (4) is trivial. If $u_iv_i$ and $u_{i+1}v_{i+1}$ are horizontal edges of $G$, then applying ($\mathrm{PC}$) to these edges, we conclude that the vertices $u_{i-2},u_{i-1},u_i,u_{i+1}$ are included in two pentagons sharing the edge $u_{i-1}u_i$. By (2) we deduce that $d(u_{i-2},u_{i+1})=2$, a contradiction.
Let $u$ be a furthest from $v_0$ vertex of $G$ and let $P=(v_0=u_0,u_1,\ldots,u_k=u)$ be the unique shortest $(v_0,u)$-path of $G$. Since $v_0$ has eccentricity $\ge 3$, $k\ge 3$. By (4), all edges of $P$ are vertical. Since $G$ is 2-connected,
$u$ is adjacent to a vertex $v\ne u_{k-1}$. From the maximality choice of $u$, necessarily $uv$ is a horizontal edge. By (3), we conclude that $f^2(v)=u_{k-2}$ and there exists a pentagon $uvv'u_{k-1}u_{k-1}$. Since $G$ is 2-connected, the vertices $u_{k-1}$ and $u_{k-3}$ are connected in $G$ by a path $Q=(u_{k-1}=x_1,x_2,\ldots,x_{m-1},x_m=u_{k-3})$ not passing via $u_{k-2}$. Since $u_kv$ is a horizontal edge, by (4) $u_{k-1}x_2$ must be vertical. Since $x_2\ne u_{k-2}$, $d(v_0,x_2)=d(v_0,u)$. From the choice of $u$, $d(v_0,x_3)\le d(v_0,u)$. Since $x_3\ne u_{k-1}$, we conclude that $x_2x_3$ is a horizontal edge. Let $Q'=(x_2,x_3,\ldots,x_i)$ be a maximal by inclusion subpath of $Q$ in which all edges are horizontal. By (3), for each vertex $x_i$ of $Q'$, we have $f^2(x_i) = f^2(x_2) = u_{k-2}$ and by ($\mathrm{PC}$) each edge $x_{j-1}x_j$ of $Q'$ is included in a pentagon $x_{j-1}y_{j-1}u_{k-2}y_jx_j$ (where $y_2=u_{k-1}$). Consider the next edge $x_ix_{i+1}$ of $Q$. From the maximality choice of $u$ and $Q'$, we conclude that $x_ix_{i+1}$ is vertical with $d(v_0,x_{i+1})<d(v_0,x_i)$. Hence $x_{i+1}=y_i$ and $u_{k-2}$ is the parent of $x_{i+1}$. Since $(v_0=u_1,u_2,\ldots,u_{k-2},y_i,x_i)$ is a shortest $(v_0,x_i)$-path and $x_i$ is incident to the horizontal edge $x_ix_{i-1}$ and $x_{i+2}\ne u_{k-2}$, by (4) we conclude that $x_{i+1}x_{i+2}$ is a vertical edge with $d(v_0,x_{i+2})=d(v_0,x_i)$. Again, from the maximality choice of $u$ we deduce that the edge $x_{i+2}x_{i+3}$ is horizontal. By considering
the maximal subpath $Q''=(x_{i+2},x_{i+3},\ldots,x_j)$ of $Q$ consisting of horizontal edges, by (3) we conclude that all vertices of $Q''$ have $u_{k-2}$ as their grand parent and that $x_{j+1}$ has $u_{k-2}$ as its parent. Continuing in this way, we obtain that all vertices of the path $Q$ have $u_{k-2}$ either as their grand parent or as their parent. But this is impossible since $Q$ is a $(u_{k-1},u_{k-3})$-path and $u_{k-3}$ is the parent of $u_{k-2}$. This contradiction establishes that the
eccentricity of any vertex of $G$ is 2 and thus that $\mathrm{diam}(G)=2$.
\end{proof}
This concludes the proof of Proposition \ref{thm: Moore}.
\end{proof}
\section{A local-to-global characterization}
\label{sec: loc2glob}
The goal of this section is to prove the following topological characterization of CB-graphs:
\begin{theorem}
\label{thm: triangle-pentagon}
For a graph $G$ the following conditions are equivalent:
\begin{itemize}
\item[(i)] $G$ is a graph with convex balls;
\item[(ii)] the triangle-pentagon complex $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is simply connected, $\mathrm{TPC}^0_{\leq 3}$ and $\mathrm{INC}_{\leq 3}$ hold and every induced subgraph isomorphic to $\mathrm{PP}_1$ has diameter at most $3$.
\item[(iii)] the triangle-pentagon complex $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is simply connected and every ball of radius at most $3$ is convex.
\end{itemize}
Furthermore, if $G$ is a graph in which all balls of radius at most $3$ are convex, then the 1-skeleton $\widetilde{G}$ of the universal cover of
$X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is a graph with convex balls.
\end{theorem}
The implication (i)$\Rightarrow$(iii) is proved in the next subsection. The implication (iii)$\Rightarrow$(ii) follows from the first part of the proof of Theorem \ref{thm: TPC}. The proof of the implication (ii)$\Rightarrow$(i) is much harder and is based on the
proof of the second assertion of the theorem. We adapt the proof of similar local-to-global results from the papers \cite{Bresaretal2013,ChalopiChepoiOsajda2015,CCHO,Osajda}.
The main difference with those proofs is the technical difficulty in dealing with pentagons in the inductive construction of the universal
cover of $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$. The proof is given in the subsequent four subsections.
\subsection{Simple connectivity of the triangle-pentagon complex}The following lemma establishes the implication (i)$\Rightarrow$(iii) of Theorem \ref{thm: triangle-pentagon}:
\begin{lemma}\label{lem: triangle-pentagon-simply-connected}
Let $G$ be a graph such that $\mathrm{INC}(v_0)$ and $\mathrm{TPC}(v_0)$ hold for some vertex $v_0$. Then $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is simply connected.
\end{lemma}
\begin{proof}
We show that any circuit $C=(u_1, \ldots, u_k, u_1)$ of $G$ is homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ to the constant circuit $(v_0)$. We assume without loss of generality that $u_1$ maximizes the distance $d(v_0,u_i)$ among all $u_i\in C$. In what follows, the operations over the indices of a cycle of length $k$ will be done modulo $k$. Let $\delta(C):=\sum_{i=1}^k 5^{d(v_0,u_i)+d(v_0, u_{i+1})} \geq k$. Then $\delta(C)=k$ if and only if
$C$ is the constant circuit $(v_0)$. If $\delta(C)>k$, we assert that $C$ is homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ to a circuit $C'$ with $\delta(C')<\delta(C)$.
First we note that we may assume that for every $i$, we have $u_i\ne u_{i+1}$ and $u_{i-1}\ne u_{i+1}$, otherwise we are trivially done. In particular, $k\geq 3$. Let $r:=d(v_0, u_1)$. First suppose that both $u_{k-1}$ and $u_2$ are at distance $r-1$ from $v_0$. By $\mathrm{INC}(v_0)$, we get $u_{k-1}\sim u_2$, so $C$ is homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ to the circuit
$C'=(u_2, u_3, \ldots, u_k, u_2)$. Moreover, $\delta(C')=\delta(C)-2 \cdot5^{2r-1} + 5^{2(r-1)} <\delta(C),$ and we are done.
Otherwise, from the maximality choice of $u_1$, we can suppose without loss of generality that $d(v_0, u_2)=r$. Then $\mathrm{TPC}(v_0, u_1u_2)$ holds.
If $\mathrm{TC}(v_0, u_1u_2)$ applies, then there exists a vertex $z$ such that $z\sim u_1, u_2$ and $d(v_0, z)=r-1$. Then $C$ is homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}$ to the circuit $C':=(u_1, z, u_2, u_3, \ldots, u_k, u_1)$ and we have $\delta(C')=\delta(C) - 5^{2r} + 2\cdot 5^{2r-1} < \delta(C).$
Otherwise, $\mathrm{PC}(v_0, u_1u_2)$ applies and there exist $w, w', z$ such that $d(v_0, w)=d(v_0, w')= r-1$, $d(v_0, w)=r-2$ and $u_1wzw'u_2$ is a pentagon of $G$. Then $C$ is homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ to the circuit $C':=(u_1, w, z, w', u_2, u_3, \ldots, u_k, u_1)$ and we have
$\delta(C')= \delta(C) - 5^{2r} + 2\cdot5^{2r-1} + 2 \cdot 5^{2r-3} \leq \delta(C) - 5^{2r} + 4 \cdot 5^{2r-1} < \delta(C).$
In both cases we have the inequality $\delta(C')<\delta(C)$, so we are done.
\end{proof}
\subsection{The inductive construction of the universal cover} The proof of the implication (ii)$\Rightarrow$(i) of Theorem \ref{thm: triangle-pentagon} is technically involved and first we outline it. Notice, that this implication follows from the second assertion of the theorem. To establish it, let $G$ be a graph in which all balls of radius at most 3 are convex and set $X:= X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$. We construct the universal cover $\ensuremath{\widetilde{X}}$ of $X$ inductively, as the union $\bigcup_{i\geq 0}\ensuremath{\widetilde{B}}_i$, where each $\ensuremath{\widetilde{B}}_i$ is a ball of radius $i$ centered at an arbitrary (but fixed) basepoint $\ensuremath{\widetilde{v}}_0$. We define the graph $\ensuremath{\widetilde{G}}_i$ having $\ensuremath{\widetilde{B}}_i$ as the vertex set and we let $\ensuremath{\widetilde{X}}_i:= X_{\scaleobj{0.8}{\triangle}, \pentagon}(\ensuremath{\widetilde{G}}_i)$ be the triangle-pentagon complex $\ensuremath{\widetilde{G}}_i$. Our construction will ensure that the properties $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ and $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0)$ are satisfied for every $i$, i.e. when the distance involved are at most $i$. We will also construct inductively in a coherent way the maps $f_i : \ensuremath{\widetilde{X}}_i\to X$ such that $f := \bigcup_{i\geq 0}f_i : \ensuremath{\widetilde{X}} \to X$ is well defined and is a covering map. As the universal cover of $X$ is unique up to (cell-preserving) isomorphisms, this implies that the 1-skeleton $\ensuremath{\widetilde{G}}$ of $\ensuremath{\widetilde{X}}$ satisfies $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ and $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0)$ for any choice of the basepoint $\ensuremath{\widetilde{v}}_0$. By Theorem \ref{thm: TPC}, $\ensuremath{\widetilde{G}}$ is a graph with convex balls, and by Lemma \ref{lem: triangle-pentagon-simply-connected} the complex $\widetilde X = X_{\scaleobj{0.8}{\triangle}, \pentagon}(\ensuremath{\widetilde{G}})$ is simply connected.
If the input complex $X$ is simply connected, then we must have $X=\ensuremath{\widetilde{X}}$ and thus $G=\ensuremath{\widetilde{G}}$, showing that $G$ is a CB-graph.
For any vertex $\ensuremath{\widetilde{v}}$ of $\ensuremath{\widetilde{G}}_i$, we denote by $v$ the image $f_i(\ensuremath{\widetilde{v}})$ of $\ensuremath{\widetilde{v}}$ in $G$.
Let $v_0$ be an arbitrary but fixed vertex of $G$ and set $\ensuremath{\widetilde{B}}_0:= \sg{\ensuremath{\widetilde{v}}_0}$ and define $f_0(\ensuremath{\widetilde{v}}_0):= v_0$.
We let $\ensuremath{\widetilde{S}}_{i+1}:= \ensuremath{\widetilde{B}}_{i+1}\backslash \ensuremath{\widetilde{B}}_i$ for every $i \geq 0$.
We prove the following properties by induction:
\begin{itemize}
\item[($P_i$)] $B_j(\ensuremath{\widetilde{v}}_0)= \ensuremath{\widetilde{B}}_j$ for every $0\leq j \leq i$.
\item[($Q_i$)] $\ensuremath{\widetilde{G}}_i$ satisfies $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ and $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0)$.
\item[($R_i$)] for any $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_{i-1}$, $f_i$ defines an isomorphism between the subgraphs $\ensuremath{\widetilde{G}}_i[B_1(\ensuremath{\widetilde{u}})]$ and $G[B_1(f_i(\ensuremath{\widetilde{u}})]$.\item[($S_i$)] for any five vertices $\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{x}}, \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{y}}, \ensuremath{\widetilde{v}} \in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{u}} \ensuremath{\widetilde{x}} \ensuremath{\widetilde{w}} \ensuremath{\widetilde{y}} \ensuremath{\widetilde{v}}$ is a path in $\ensuremath{\widetilde{G}}_i$, if $uxwyv$ is a 5-cycle in $G$, then $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{v}}$ in $\ensuremath{\widetilde{G}}_i$.
\item[($T_i$)] for any $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{S}}_i$, $f_i$ defines an isomorphism between the subgraphs $\ensuremath{\widetilde{G}}_i\left[B_1(\ensuremath{\widetilde{u}}) \right]$ and $G\left[f_i(B_1(\ensuremath{\widetilde{u}})) \right]$.
\end{itemize}
Assume that everything was constructed and proved for every $j=0, \ldots, i$.
We define the set of all the vertices ``which can be seen from $\ensuremath{\widetilde{B}}_i$ but which were not reached yet'' by setting
$$Z :=\sg{(\ensuremath{\widetilde{w}}, z): \ensuremath{\widetilde{w}} \in \ensuremath{\widetilde{S}}_i \text{ and } z\in B_1(w)\backslash f_i(B_1(\ensuremath{\widetilde{w}}, \ensuremath{\widetilde{G}}_i))}.$$
On $Z$ we define the binary relation $\equiv $ by setting $(\ensuremath{\widetilde{w}}, z)\equiv (\ensuremath{\widetilde{w}}', z')$ if and only if $z= z'$ and either
$\ensuremath{\widetilde{w}}=\ensuremath{\widetilde{w}}'$ or $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{w}}'$. We will later show that $\equiv$ is an equivalence relation.
First we establish two useful lemmas.
\begin{lemma}\label{lem-voisinZ}
For any $(\ensuremath{\widetilde{w}},z) \in Z$ and any $\ensuremath{\widetilde{x}} \in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{w}}$ and $f_i(\ensuremath{\widetilde{x}}) = x \sim z$, $\ensuremath{\widetilde{x}} \in \ensuremath{\widetilde{S}}_i$ and $(\ensuremath{\widetilde{x}},z) \in Z$.
\end{lemma}
\begin{proof}
If $\ensuremath{\widetilde{x}} \in \ensuremath{\widetilde{B}}_{i-1}$, then by ($R_i$) , there exists $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{x}}$
such that $f_i(\ensuremath{\widetilde{z}}) = z$. Similarly, by the definition of $Z$, if
$\ensuremath{\widetilde{x}} \in \ensuremath{\widetilde{B}}_i$ and $(\ensuremath{\widetilde{x}},z) \notin Z$, then there exists
$\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{x}}$ such that $f_i(\ensuremath{\widetilde{z}}) = z$. In both cases, by ($R_i$) or
($T_i$) , we have $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{z}}$ and thus $(\ensuremath{\widetilde{w}},z) \notin Z$, a
contradiction.
\end{proof}
The following lemma imply that $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ inductively holds:
\begin{lemma}
\label{lem: INCcongr}
For any $(\ensuremath{\widetilde{w}}, z), (\ensuremath{\widetilde{w}}', z') \in Z$ such that $z=z'$ and $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{w}}'$, there exists a vertex $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}'$. Consequently, $(\ensuremath{\widetilde{w}}, z)\equiv (\ensuremath{\widetilde{w}}', z')$ if and only if $z= z'$ and $\ensuremath{\widetilde{w}} = \ensuremath{\widetilde{w}}'$ or $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{w}}'$ and there exists some $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}'$.
\end{lemma}
\begin{proof}
Assume that there is no such vertex $\ensuremath{\widetilde{u}}$. Then by ($Q_i$) we can apply $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0, \ensuremath{\widetilde{w}}\tw')$. Since the triangle condition does not applies, there exist vertices $\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}' \in \ensuremath{\widetilde{B}}_{i-1}, \ensuremath{\widetilde{y}} \in \ensuremath{\widetilde{B}}_{i-2}$ such that $\ensuremath{\widetilde{w}} \ensuremath{\widetilde{u}} \ensuremath{\widetilde{y}} \ensuremath{\widetilde{u}}' \ensuremath{\widetilde{w}}'$ is a pentagon in $\ensuremath{\widetilde{G}}_i$. By ($R_i$) and ($T_i$) this means that $wuyu'w'$ is a pentagon in $G$. As $z\sim w, w'$, this implies that $z\notin \sg{w, w', u, u', y}$. By convexity of $B_2(z)$ in $G$, we conclude that $1\le d(z,y)\leq 2$.
By Lemma~\ref{lem-voisinZ}, $z \nsim u, u'$.
Thus $d(z,y)=2$ and there exists $t \sim z, y$ different from $w,w',u,u',y$. By ($R_i$) there exists $\ensuremath{\widetilde{t}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that $f_i(\ensuremath{\widetilde{t}})=t$ and $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{y}}$. By ($R_i$) again there exists $\ensuremath{\widetilde{z}} \in \ensuremath{\widetilde{B}}_i$ such that $f_i(\ensuremath{\widetilde{z}})=z$ and $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{t}}$. We assert that $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{w}}$. Indeed, either $t \sim u$ and by convexity of $B_1(t)$ in $G$ this is equivalent to $t\sim w$ and we can apply ($R_i$) , otherwise, $t\nsim u, w$ and $ztyuw$ is a pentagon of $G$, so we can apply ($S_i$) . In both cases, $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{w}}$ contradicts the fact that $(\ensuremath{\widetilde{w}}, z)\in Z$ and we are done.
\end{proof}
\begin{lemma}
\label{lem: equivcongr}
The relation $\equiv$ is an equivalence relation on $Z$.
\end{lemma}
\begin{proof}
Reflexivity and symmetry of $\equiv$ immediately follow from its
definition. Hence, we only have to show transitivity. Let
$(\ensuremath{\widetilde{w}}, z), (\ensuremath{\widetilde{w}}', z'), (\ensuremath{\widetilde{w}}'', z'') \in Z$ such that
$(\ensuremath{\widetilde{w}}, z)\equiv (\ensuremath{\widetilde{w}}', z')$ and $(\ensuremath{\widetilde{w}}', z') \equiv (\ensuremath{\widetilde{w}}'', z'')$.
Then $z=z'=z''$. We assume that $\ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}', \ensuremath{\widetilde{w}}''$ are pairwise
distinct, otherwise we are done. Hence $\ensuremath{\widetilde{w}}' \sim \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}''$. We
also assume that $\ensuremath{\widetilde{w}} \nsim \ensuremath{\widetilde{w}}''$, otherwise we are done. By ($T_i$)
applied to $\ensuremath{\widetilde{w}}'$, this implies that $w \nsim w''$. By Lemma
\ref{lem: INCcongr} there exist $\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}' \in \ensuremath{\widetilde{B}}_{i-1}$ such that
$\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}'$ and $\ensuremath{\widetilde{u}}'\sim \ensuremath{\widetilde{w}}', \ensuremath{\widetilde{w}}''$. If $\ensuremath{\widetilde{u}}=\ensuremath{\widetilde{u}}'$, then
by convexity of $B_1(u)$ in $G$ we have $u \sim z$, which is
impossible by Lemma~\ref{lem-voisinZ}. Hence $\ensuremath{\widetilde{u}}\neq \ensuremath{\widetilde{u}}'$ and
$z \nsim u,u'$. By ($Q_i$) we can use $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ to get
$\ensuremath{\widetilde{u}}\sim\ensuremath{\widetilde{u}}'$ and a vertex $\ensuremath{\widetilde{s}} \in \ensuremath{\widetilde{B}}_{i-2}$ such that
$\ensuremath{\widetilde{s}} \sim \ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}'$. From ($R_i$) and the definition of $Z$ we conclude
that the vertices $z,w,w',w'',u,u'$ are pairwise distinct. Moreover,
$s\notin \sg{u, u', w, w', w'', z}$ by ($R_i$) , so by convexity of
$B_2(s)$ in $G$, we get $d(s,z)\leq 2$. If $d(s,z)=1$, then by
convexity of $B_1(s)$ and because $s\nsim w$, we have $u \sim z$,
which gives a contradiction. Hence $d(s,z)=2$ and let $t\sim
z,s$. Then $t\neq u$ and by ($R_i$) there exists $\ensuremath{\widetilde{t}} \in \ensuremath{\widetilde{B}}_{i-1}$
and $\ensuremath{\widetilde{z}} \in \ensuremath{\widetilde{B}}_i$ such that $f_i(\ensuremath{\widetilde{z}})=z$ and such that
$\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{z}}, \ensuremath{\widetilde{s}}$. Depending whether $t \sim u, w$ or not, by ($R_i$)
and ($S_i$) we get $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{w}}$, which contradicts that
$(\ensuremath{\widetilde{w}}, z)\in Z$.
\end{proof}
Let $\ensuremath{\widetilde{S}}_{i+1}$ denote the set of equivalence classes of $\equiv$, i.e., $\ensuremath{\widetilde{S}}_{i+1}=Z/_{\equiv}$. For a pair $(\ensuremath{\widetilde{w}},z)\in Z$,
we denote by $[\ensuremath{\widetilde{w}},z]$ the equivalence class of $\equiv$ containing $(\ensuremath{\widetilde{w}},z).$
We let $\ensuremath{\widetilde{B}}_{i+1}:= \ensuremath{\widetilde{B}}_i \sqcup \ensuremath{\widetilde{S}}_{i+1}$ and define the edges of $\ensuremath{\widetilde{G}}_{i+1}$ in the following way: (a) the adjacencies between the
vertices of $\ensuremath{\widetilde{B}}_i$ do not change, (b) we let $\ensuremath{\widetilde{w}} \sim \cro{\ensuremath{\widetilde{w}}, z}$, and (c) if $\ensuremath{\widetilde{a}} = \cro{\ensuremath{\widetilde{w}}, z}, \ensuremath{\widetilde{b}} = \cro{\ensuremath{\widetilde{w}}', z'}\in \ensuremath{\widetilde{S}}_{i+1}$,
then we let $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{b}}$ in $\ensuremath{\widetilde{G}}_{i+1}$ if $z\sim z'$ in $G$ and one of the following two conditions is satisfied:
\begin{itemize}
\item[(1)] there exists $\ensuremath{\widetilde{w}}\in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{a}} = \cro{\ensuremath{\widetilde{w}}, z}$ and $\ensuremath{\widetilde{b}} = \cro{\ensuremath{\widetilde{w}}, z'}$,
\item[(2)] there exist $\ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}' \in \ensuremath{\widetilde{B}}_i$ and $\ensuremath{\widetilde{y}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that $\ensuremath{\widetilde{a}} = \cro{\ensuremath{\widetilde{w}}, z}$, $\ensuremath{\widetilde{b}} = \cro{\ensuremath{\widetilde{w}}', z'}$ and $\ensuremath{\widetilde{y}} \sim \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}'$.
\end{itemize}
From ($R_i$) and the fact that balls of $G$ of radius $1$ are convex, we can equivalently require in (2) that $awyw'b$ is a pentagon of $G$, otherwise (1) holds.
Finally, we define $f_{i+1}$ on $\ensuremath{\widetilde{B}}_{i+1}$ by setting $f_{i+1}(\ensuremath{\widetilde{u}}):=f_{i}(\ensuremath{\widetilde{u}})$ when $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_i$ and $f_{i+1}(\cro{\ensuremath{\widetilde{w}}, z}):=z$ for any $\cro{\ensuremath{\widetilde{w}}, z} \in \ensuremath{\widetilde{S}}_{i+1}$.
We now prove that if an edge $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$ between vertices of $\ensuremath{\widetilde{S}}_{i+1}$
is defined only by condition (2) above, then $\mathrm{PC}^2(\ensuremath{\widetilde{v}}_0,\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}})$ holds.
\begin{lemma}\label{lem-ind-PCdist2}
Consider an edge $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$ of $\ensuremath{\widetilde{G}}_{i+1}$ with
$\ensuremath{\widetilde{a}},\ensuremath{\widetilde{b}} \in \ensuremath{\widetilde{S}}_{i+1}$. If there is no $\ensuremath{\widetilde{w}} \in \ensuremath{\widetilde{B}}_i$ such that
$\ensuremath{\widetilde{a}} = [\ensuremath{\widetilde{w}},a]$ and $\ensuremath{\widetilde{b}} = [\ensuremath{\widetilde{w}},b]$, then for any
$\ensuremath{\widetilde{t}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that $d(\ensuremath{\widetilde{a}},\ensuremath{\widetilde{t}}) = 2$, we have
$d(\ensuremath{\widetilde{b}},\ensuremath{\widetilde{t}}) = 2$.
\end{lemma}
\begin{proof}
By the definition of the edges of $\ensuremath{\widetilde{G}}_{i+1}$ and since $\ensuremath{\widetilde{a}},\ensuremath{\widetilde{b}}$
have no common neighbor in $\ensuremath{\widetilde{B}}_i$, there exist $\ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}' \in \ensuremath{\widetilde{B}}_i$
and $\ensuremath{\widetilde{y}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that $\ensuremath{\widetilde{a}} = \cro{\ensuremath{\widetilde{w}}, a}$,
$\ensuremath{\widetilde{b}} = \cro{\ensuremath{\widetilde{w}}', b}$ and $\ensuremath{\widetilde{y}} \sim \ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}'$. We distinguish two
cases, depending whether $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{w}}$ or not.
\medskip\noindent
{\bf Case 1.} $\ensuremath{\widetilde{t}}$ is a neighbor of $\ensuremath{\widetilde{w}}$.
\begin{proof} If $\ensuremath{\widetilde{t}}$ is a neighbor of $\ensuremath{\widetilde{w}}'$, we are done. Thus, we can assume
that $\ensuremath{\widetilde{t}} \nsim \ensuremath{\widetilde{w}}'$ and $\ensuremath{\widetilde{t}} \neq \ensuremath{\widetilde{y}}$. By $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$, we
have $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{y}}$ and there exists $\ensuremath{\widetilde{s}} \in \ensuremath{\widetilde{B}}_{i-2}$ such that
$\ensuremath{\widetilde{s}} \sim \ensuremath{\widetilde{t}},\ensuremath{\widetilde{y}}$. In $G$, $abw'yw$ is a pentagon and since
$t \sim w,y$, we have $t \notin \sg{a,b,w',y,w}$. By ($R_i$) , we have
$t \nsim w'$. By Lemma~\ref{lem-voisinZ}, we have $t \nsim a$. If
$t \sim b$, then $wabt$ is an induced 4-cycle in $G$, which is
impossible. Since $B_2(t)$ is convex, we have $d(t,b) = 2$. Let
$x \sim b,t$ and note that $x \notin \sg{a,b,w',y,w,t}$. By ($R_i$) ,
there exists $\ensuremath{\widetilde{x}} \in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{t}}$ and
$f_i(\ensuremath{\widetilde{x}}) = x$. We claim that $\ensuremath{\widetilde{x}} \in \ensuremath{\widetilde{S}}_i$. Indeed, if
$\ensuremath{\widetilde{x}} \notin \ensuremath{\widetilde{S}}_i$, there exists $\ensuremath{\widetilde{b}}' \in \ensuremath{\widetilde{B}}_i$ such that
$f_i(\ensuremath{\widetilde{b}}') = b$ and $\ensuremath{\widetilde{b}}' \sim \ensuremath{\widetilde{x}}$. By ($S_i$) applied to the vertices
$\ensuremath{\widetilde{b}}',\ensuremath{\widetilde{x}},\ensuremath{\widetilde{t}},\ensuremath{\widetilde{y}},\ensuremath{\widetilde{w}}'$, we have $\ensuremath{\widetilde{b}}'\sim \ensuremath{\widetilde{w}}'$, contradicting the
fact that $(\ensuremath{\widetilde{w}}',b) \in Z$. Since $G$ does not contain induced
4-cycles, $x \sim w'$ if and only if $x \sim y$.
If $x \sim w',y$, by ($R_i$) , $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{w}}'$ and by
Lemma~\ref{lem-voisinZ}, we have $(\ensuremath{\widetilde{x}},b) \in Z$. By the definition of $\equiv$, we have
$\ensuremath{\widetilde{b}} = [\ensuremath{\widetilde{w}}',b] = [\ensuremath{\widetilde{x}},b]$. Consequently, in this case, we have
$d(\ensuremath{\widetilde{b}},\ensuremath{\widetilde{t}}) = 2$. Let now $x \nsim w',y$. Let $s = f_i(\ensuremath{\widetilde{s}})$ and observe that
$s \notin \sg{a,b,x}$ since $s \sim y$. Moreover, by ($R_i$) , we have
$s \notin \sg{t,w,w',y}$ and $s \nsim w,x,w'$. Consequently,
$d(s,x) = d(s,w') = 2$ and by the convexity of $B_2(s)$, we get
$d(s,b) \leq 2$. If $s \sim b$, then $syw'b$ is a forbidden
induced $4$-cycle, thus $d(s,b) = 2$. Let
$r \sim b,s$. By ($R_i$) , there exists $\ensuremath{\widetilde{r}} \in \ensuremath{\widetilde{B}}_{i-1}$ and
$\ensuremath{\widetilde{b}}' \in \ensuremath{\widetilde{B}}_i$ such that $f_i(\ensuremath{\widetilde{r}}) = r$, $f_i(\ensuremath{\widetilde{b}}') = b$, and
$\ensuremath{\widetilde{r}} \sim \ensuremath{\widetilde{s}},\ensuremath{\widetilde{b}}'$. By ($S_i$) applied to the vertices
$\ensuremath{\widetilde{b}}',\ensuremath{\widetilde{r}},\ensuremath{\widetilde{s}},\ensuremath{\widetilde{y}},\ensuremath{\widetilde{w}}'$, we have $\ensuremath{\widetilde{b}}'\sim \ensuremath{\widetilde{w}}'$, contradicting the
fact that $(\ensuremath{\widetilde{w}}',b) \in Z$.
\end{proof}
\medskip\noindent
{\bf Case 2.} $\ensuremath{\widetilde{t}}$ is not a neighbor of $\ensuremath{\widetilde{w}}$.
\begin{proof} Since $d(\ensuremath{\widetilde{t}},\ensuremath{\widetilde{a}}) = 2$, there exists
$\ensuremath{\widetilde{s}} \sim \ensuremath{\widetilde{a}},\ensuremath{\widetilde{t}}$. Since $\ensuremath{\widetilde{t}} \in \ensuremath{\widetilde{B}}_{i-1}$ and
$\ensuremath{\widetilde{a}} \in \ensuremath{\widetilde{S}}_{i+1}$, we have $\ensuremath{\widetilde{s}} \in \ensuremath{\widetilde{S}}_i$ and
$[\ensuremath{\widetilde{s}},a] = \ensuremath{\widetilde{a}} = [\ensuremath{\widetilde{w}},a]$. By the definition of $\equiv$, this
implies that $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{s}}$ and by Lemma~\ref{lem: INCcongr},
there exists $\ensuremath{\widetilde{r}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that $\ensuremath{\widetilde{r}} \sim \ensuremath{\widetilde{w}},\ensuremath{\widetilde{s}}$. By
applying Case~1 to $\ensuremath{\widetilde{r}}$, we have that $d(\ensuremath{\widetilde{r}},\ensuremath{\widetilde{b}}) = 2$ and thus,
there exists $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{r}},\ensuremath{\widetilde{b}}$. Consequently, we can replace
$\ensuremath{\widetilde{y}}$, $\ensuremath{\widetilde{w}}$ and $\ensuremath{\widetilde{w}}'$ by $\ensuremath{\widetilde{r}}$, $\ensuremath{\widetilde{s}}$ and $\ensuremath{\widetilde{x}}$ and we can apply
Case~1 to $\ensuremath{\widetilde{t}}$.
\end{proof}
Cases 1 and 2 establish the lemma.
\end{proof}
\subsection{Properties ($P_{i+1}$), ($Q_{i+1}$), ($R_{i+1}$), and ($T_{i+1}$)}
We start with the proof of ($P_{i+1}$) and ($Q_{i+1}$):
\begin{lemma}
\label{lem: Qi}
$\ensuremath{\widetilde{G}}_{i+1}$ satisfies ($P_{i+1}$).
\end{lemma}
\begin{proof}
We show that for every $0\le j\le i+1$ we have $\ensuremath{\widetilde{B}}_j= B_j(\ensuremath{\widetilde{v}}_0)$. This is true when $j \leq i$ by induction hypothesis.
By construction, every vertex in $\ensuremath{\widetilde{S}}_{i+1}$ is adjacent to at least one vertex of $\ensuremath{\widetilde{B}}_i$ and is only adjacent to vertices of $\ensuremath{\widetilde{S}}_i$, hence it is
at distance $i+1$ from $\ensuremath{\widetilde{v}}_0$ and we are done.
\end{proof}
\begin{lemma}
$\ensuremath{\widetilde{G}}_{i+1}$ satisfies ($Q_{i+1}$).
\end{lemma}
\begin{proof}
$\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ is obtained by induction hypothesis and by Lemma \ref{lem: INCcongr}.
Now we prove that $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0)$ holds. Let $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{w}}' \in \ensuremath{\widetilde{B}}_{i+1}$, both at distance $k$ from $\ensuremath{\widetilde{v}}_0$. If $k\leq i$, then we conclude by induction hypothesis.
Otherwise, note that the adjacency $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{w}}'$ either comes from (1) or (2), which correspond to the triangle or the pentagon condition, respectively.
\end{proof}
The following two lemmas are trivial.
\begin{lemma}
\label{lem: l313}
If $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{b}}$ in $\ensuremath{\widetilde{G}}_{i+1}$, then $a\sim b$ in $G$.
\end{lemma}
\begin{lemma}
\label{lem: l314}
If $a \sim b$ in $G$ and $\ensuremath{\widetilde{a}} \in \ensuremath{\widetilde{B}}_i$, then there exists $\ensuremath{\widetilde{b}} \in \ensuremath{\widetilde{B}}_{i+1}$ such that $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{b}}$ and $f_{i+1}(\ensuremath{\widetilde{b}})=b$.
\end{lemma}
Now we show that $f_{i+1}$ is ``locally injective''.
\begin{lemma}
\label{lem: injloc}
Let $\ensuremath{\widetilde{a}},\ensuremath{\widetilde{b}},\ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_{i+1}$, such that $\ensuremath{\widetilde{b}} \neq \ensuremath{\widetilde{c}}$ and $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}}$, then $b\neq c$.
\end{lemma}
\begin{proof}
First, if $\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$, then $b \sim c$ by Lemma \ref{lem: l313} and we are done.
Hence, let $\ensuremath{\widetilde{b}} \nsim \ensuremath{\widetilde{c}}$. Moreover, if $\ensuremath{\widetilde{a}}, \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_i$, then we are done by ($R_i$) . From the definition of
$\ensuremath{\widetilde{S}}_{i+1}$ we are also done when $\ensuremath{\widetilde{a}} \in \ensuremath{\widetilde{B}}_i$ and $\ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{S}}_{i+1}$. Suppose by way of contradiction that $b=c$.
If $\ensuremath{\widetilde{a}}, \ensuremath{\widetilde{b}} \in \ensuremath{\widetilde{B}}_i$ and $\ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{S}}_{i+1}$, then we
have $\ensuremath{\widetilde{c}} = \cro{\ensuremath{\widetilde{a}}, b}$, and we get a contraction with the fact that $(\ensuremath{\widetilde{a}}, b)\in Z$. Hence, further we
may assume that $\ensuremath{\widetilde{a}} \in \ensuremath{\widetilde{S}}_{i+1}$. If $\ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_i$, then by definition of $\ensuremath{\widetilde{S}}_{i+1}$ we have $[\ensuremath{\widetilde{b}},a] = \ensuremath{\widetilde{a}} = [\ensuremath{\widetilde{c}},a]$, and by definition of $\equiv$, we have $\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$, a contradiction. Now, let $\ensuremath{\widetilde{b}} \in \ensuremath{\widetilde{S}}_{i+1}$ and $\ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_i$.
Then $\ensuremath{\widetilde{a}} = \cro{\ensuremath{\widetilde{c}}, a}$.
If condition (1) applies to the edge $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$, then there exists $\ensuremath{\widetilde{w}} \in \ensuremath{\widetilde{B}}_i$
such that $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{a}}$. Necessarily, $\ensuremath{\widetilde{w}} \neq \ensuremath{\widetilde{c}}$ because
$\ensuremath{\widetilde{b}} \nsim \ensuremath{\widetilde{c}}$. By $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$, we get $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{c}}$.
Replacing $\ensuremath{\widetilde{a}}$ by $\ensuremath{\widetilde{w}}$, we are in the previous case where $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}}$,
which was shown to be impossible. Therefore assume that there is no such $\ensuremath{\widetilde{w}}\in \ensuremath{\widetilde{B}}_i$ and thus only condition (2) applies. Consider a neighbor $\ensuremath{\widetilde{w}} \in \ensuremath{\widetilde{B}}_{i-1}$ of $\ensuremath{\widetilde{c}}$ and observe that by Lemma~\ref{lem-ind-PCdist2}, there exists $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{b}},\ensuremath{\widetilde{w}}$. By ($R_i$) , we get that $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{c}}$, and replacing $\ensuremath{\widetilde{a}}$ by $\ensuremath{\widetilde{u}}$
we are again in a previous case where $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}}$.
Finally, let $\ensuremath{\widetilde{a}}, \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{S}}_{i+1}$. First assume that
condition (1) applies to both $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$. Then there exist
$\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}'\in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{a}}, \ensuremath{\widetilde{b}}$ and
$\ensuremath{\widetilde{u}}' \sim \ensuremath{\widetilde{a}}, \ensuremath{\widetilde{c}}$. If $\ensuremath{\widetilde{u}} = \ensuremath{\widetilde{u}}'$, replacing $\ensuremath{\widetilde{a}}$ by $\ensuremath{\widetilde{u}}$, we
are in a previous case where $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{b}},\ensuremath{\widetilde{c}}$. So, assume that
$\ensuremath{\widetilde{u}} \neq \ensuremath{\widetilde{u}}'$. This implies that $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{u}}'$ since
$\ensuremath{\widetilde{a}} = [\ensuremath{\widetilde{u}},a] = [\ensuremath{\widetilde{u}}',a]$. By definition of $\equiv$, we thus have
$\ensuremath{\widetilde{b}} = [\ensuremath{\widetilde{u}}, b] = [\ensuremath{\widetilde{u}}', b] = \ensuremath{\widetilde{c}}$, a contradiction.
Assume now that condition (1) applies to $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$, while $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$ is
defined only by condition (2). Then, there exists $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_i$ such
that $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{a}},\ensuremath{\widetilde{b}}$. Pick any neighbor $\ensuremath{\widetilde{y}}$ of $\ensuremath{\widetilde{u}}$ in
$\ensuremath{\widetilde{B}}_{i-1}$. By Lemma~\ref{lem-ind-PCdist2}, there exists
$\ensuremath{\widetilde{u}}' \in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{u}}'\sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{c}}$. Since condition (1)
does not apply to the edge $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$, $acu'yu$ should be a pentagon of
$G$, contradicting the fact that $c=b \sim u$.
Consequently, we may assume that both $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$ are
defined only by condition (2). Pick any vertex $\ensuremath{\widetilde{y}} \in \ensuremath{\widetilde{B}}_{i-1}$ such
that $d(\ensuremath{\widetilde{y}},\ensuremath{\widetilde{a}}) = 2$. By Lemma~\ref{lem-ind-PCdist2}, there exists
$\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{w}}' \sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{c}}$. Consequently, $bwyw'$ is a
4-cycle of $G$ that cannot be induced. By Lemma~\ref{lem-voisinZ},
$b \nsim y$ and thus $w\sim w'$. By ($R_i$) , we have $\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{w}}'$ and
consequently, $\ensuremath{\widetilde{b}} = [\ensuremath{\widetilde{w}},b] = [\ensuremath{\widetilde{w}}',b] = [\ensuremath{\widetilde{w}}',c] = \ensuremath{\widetilde{c}}$, a contradiction.
\end{proof}
Now we show that $f_{i+1}$ is ``locally an isomorphism''.
\begin{lemma}
\label{lem: isoloc}
Let $\ensuremath{\widetilde{a}}, \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_{i+1}$ be such that $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}}$. Then $\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$ if and only if $b \sim c$.
\end{lemma}
\begin{proof}
If $\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$, then $b\sim c$ by Lemma \ref{lem:
l313}. Conversely, suppose that $b\sim c$ in $G$. Suppose first
that $\ensuremath{\widetilde{a}} \in \ensuremath{\widetilde{B}}_i$. If $\ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_i$, then we get the
result by ($R_i$) or by ($T_i$) . If $\ensuremath{\widetilde{b}} \in \ensuremath{\widetilde{B}}_i$ and $\ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{S}}_{i+1}$,
then by Lemma~\ref{lem-voisinZ}, $(\ensuremath{\widetilde{b}},c) \in Z$ and
$\ensuremath{\widetilde{b}} \sim [\ensuremath{\widetilde{b}},c] = [\ensuremath{\widetilde{a}},c] = \ensuremath{\widetilde{c}}$. If $\ensuremath{\widetilde{b}},\ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{S}}_{i+1}$,
$\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$ by the definition of $E(\ensuremath{\widetilde{G}}_{i+1})$.
Suppose now that $\ensuremath{\widetilde{a}} \in \ensuremath{\widetilde{S}}_{i+1}$. If $\ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_i$, then
the result follows from $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$. Assume that
$\ensuremath{\widetilde{b}} \in \ensuremath{\widetilde{S}}_{i+1}$ and $\ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{B}}_i$. If $(\ensuremath{\widetilde{c}},b) \in Z$, then
$\ensuremath{\widetilde{a}} = [\ensuremath{\widetilde{c}},a] \sim [\ensuremath{\widetilde{c}},b]$ and, by Lemma~\ref{lem: injloc}, we
have $[\ensuremath{\widetilde{c}},b] = \ensuremath{\widetilde{b}}$, implying that $\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$. If
$(\ensuremath{\widetilde{c}},b) \notin Z$, then there exists $\ensuremath{\widetilde{b}}' \in \ensuremath{\widetilde{B}}_i$ such that
$\ensuremath{\widetilde{b}}' \sim \ensuremath{\widetilde{c}}$ and $f_i(\ensuremath{\widetilde{b}}') = b$. By Lemma~\ref{lem-voisinZ},
$\ensuremath{\widetilde{b}}' \in \ensuremath{\widetilde{S}}_i$, $(\ensuremath{\widetilde{b}}',a) \in Z$, and $[\ensuremath{\widetilde{b}}',a] = [\ensuremath{\widetilde{c}},a] =
\ensuremath{\widetilde{a}}$. Consequently, $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{b}}'$, contradicting
Lemma~\ref{lem: injloc}.
Now, let $\ensuremath{\widetilde{a}}, \ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}} \in \ensuremath{\widetilde{S}}_{i+1}$. Observe that if $\ensuremath{\widetilde{b}}, \ensuremath{\widetilde{c}}$
have a common neighbor $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_i$, we are done by the previous
cases. Assume that such a common neighbor does not exist.
First suppose that both $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$ are defined by
condition (1). Thus there exists $\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}' \in \ensuremath{\widetilde{B}}_i$ such that
$\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{a}},\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{u}}' \sim \ensuremath{\widetilde{a}},\ensuremath{\widetilde{c}}$. Then,
$\ensuremath{\widetilde{a}} = [\ensuremath{\widetilde{u}},a] = [\ensuremath{\widetilde{u}}',a]$ and $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{u}}'$. Since
$\ensuremath{\widetilde{u}} \nsim \ensuremath{\widetilde{c}}$ and $\ensuremath{\widetilde{u}}' \nsim \ensuremath{\widetilde{b}}$, by the previous cases, we have
$u \nsim c$ and $u'\nsim b$, and thus $bcu'u$ is an induced 4-cycle
of $G$, a contradiction.
Assume now that condition (1) applies to $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$, while $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$ is
defined only by condition (2). Then, there exists $\ensuremath{\widetilde{u}} \in \ensuremath{\widetilde{B}}_i$ such
that $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{a}},\ensuremath{\widetilde{b}}$. Pick any neighbor $\ensuremath{\widetilde{y}}$ of $\ensuremath{\widetilde{u}}$ in
$\ensuremath{\widetilde{B}}_{i-1}$. By Lemma~\ref{lem-ind-PCdist2}, there exists
$\ensuremath{\widetilde{u}}' \in \ensuremath{\widetilde{B}}_i$ such that $\ensuremath{\widetilde{u}}'\sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{c}}$. But then,
$\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$ by condition (2).
Finally, assume that both $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{c}}$ are defined only by
condition (2). Pick any vertex $\ensuremath{\widetilde{y}} \in \ensuremath{\widetilde{B}}_{i-1}$ such that
$d(\ensuremath{\widetilde{y}},\ensuremath{\widetilde{a}}) = 2$. By Lemma~\ref{lem-ind-PCdist2}, there exists
$\ensuremath{\widetilde{w}} \sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{b}}$ and $\ensuremath{\widetilde{w}}' \sim \ensuremath{\widetilde{y}},\ensuremath{\widetilde{c}}$. Again, $\ensuremath{\widetilde{b}} \sim \ensuremath{\widetilde{c}}$ by
condition (2).
\end{proof}
Now we have everything at hand to prove ($R_{i+1}$) and ($T_{i+1}$) .
\begin{lemma}
\label{lem: RipTip}
The properties ($R_{i+1}$) and ($T_{i+1}$) hold.
\end{lemma}
\begin{proof}
Let $\ensuremath{\widetilde{w}}\in \ensuremath{\widetilde{B}}_{i+1}$. By Lemmas~\ref{lem: injloc} and~\ref{lem:
isoloc}, $f_{i+1}$ induces an isomorphism between $G[B_1(w)]$ and
$G[f_{i+1}(B_1(w))]$ and thus ($T_{i+1}$) holds. Moreover, by
Lemma~\ref{lem: l314}, for any $\ensuremath{\widetilde{w}} \in B_i$, $f_{i+1}$ induces a
surjection between $B_1(w)$ and $(B_1(f_{i+1}w))$ and consequently,
an isomorphism between $G[B_1(w)]$ and $G[f_{i+1}(B_1(w))]$. Thus
($R_{i+1}$) holds.
\end{proof}
From properties ($R_{i+1}$) and ($T_{i+1}$) , we immediately obtain the following lemma:
\begin{lemma}
\label{lem: morphism}
For any triangle $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{v}}\ensuremath{\widetilde{w}}$ of $\ensuremath{\widetilde{G}}_{i+1}$, its image $uvw$ under $f_{i+1}$ is a triangle of $G$.
Similarly, the image of every pentagon $\ensuremath{\widetilde{u}} \ensuremath{\widetilde{x}} \ensuremath{\widetilde{w}} \ensuremath{\widetilde{y}} \ensuremath{\widetilde{v}}$ of $\ensuremath{\widetilde{G}}_{i+1}$ is a pentagon of $G$.
\end{lemma}
\begin{lemma}\label{lemma: virtC4} Let $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}\ensuremath{\widetilde{c}}\ensuremath{\widetilde{d}}$ be a path of $\ensuremath{\widetilde{G}}_{i+1}$. Then the following holds:
\begin{itemize}
\item if $a=d$ in $G$, then $\ensuremath{\widetilde{a}} = \ensuremath{\widetilde{d}}$;
\item if $a\sim d$ in $G$, then $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{d}}$.
\end{itemize}
\end{lemma}
\begin{proof}
If $a=d$, then by applying ($R_{i+1}$) or ($T_{i+1}$) to $\ensuremath{\widetilde{b}}$, we conclude that $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{c}}$. By applying ($R_{i+1}$) or ($T_{i+1}$) to $\ensuremath{\widetilde{c}}$, we must have $\ensuremath{\widetilde{a}}=\ensuremath{\widetilde{d}}$.
If $a\sim d$, then by ($R_{i+1}$) or ($T_{i+1}$) , $abcd$ is a $4$-cycle of $G$. Since this cycle cannot be induced, we can assume without loss of generality that $a\sim c$.
By applying ($R_{i+1}$) or ($T_{i+1}$) to $\ensuremath{\widetilde{b}}$, we get $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{c}}$. Then applying ($R_{i+1}$) or ($T_{i+1}$) to $\ensuremath{\widetilde{c}}$, we get $\ensuremath{\widetilde{a}} \sim \ensuremath{\widetilde{d}}$.
\end{proof}
\begin{comment}
\begin{proof}
By convexity of $B_3(b)$, we get that $d(b,g)\leq 3$.
If $d(b,g)=2$, then $h$ belongs to a shortest path between $g$ and $d$, hence by convexity of $B_2(b)$ we have $d(b,h)\leq 2$ and we are in the first situation.
Now assume that $d(b,g)=d(b,h)=3$ and consider new vertices $p,q$ with $p \sim b,q$ and $q\sim p, g$. By $\mathrm{TPC}(b)$ applied on the edge $g\sim h$, there are two cases:
either the triangle condition applies and there exists $l_1,l_2$ with $l_1\sim b, l_2$ and $l_2\sim g,h$. Because we assumed that $d(b,h)>
2$, convexity of the balls $B_2(b)$ and $B_2(h)$ repectively imply that $l_2\sim h$ and $l_1\sim c$, hence we are in the second situation.
Otherwise the triangle condition does not apply, but the pentagon condition does so there exists $l_1,l_2,l_3$ such that $l_1\sim b, l_2, l_3$ and $l_2\sim h$ and $l_3\sim g$. Because the triangle condition does not apply, we also get $l_2\nsim g$ and $l_3\nsim h$. Eventually the same argument that in the previous case allows us to get $l_1 \sim c$ and $l_2 \sim d$. Eventually because $l_3\nsim h$, we also get $d(c,g)\leq 2$ by convexity of $B_2(c)$. Hence we are in the third situation.
\end{proof}
\end{comment}
\subsection{Property ($S_{i+1}$)}
In this subsection, we establish the property ($S_{i+1}$), whose proof is the most involved. We call a path $\ensuremath{\widetilde{\pi}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ of $\ensuremath{\widetilde{G}}_{i+1}$ a \emph{virtual 5-cycle} if it image $\pi=uxwyv$ in $G$ is a 5-cycle.
\begin{proposition} \label{Si+1} The property ($S_{i+1}$) holds, i.e., if $\ensuremath{\widetilde{\pi}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ of $\ensuremath{\widetilde{G}}_{i+1}$ is a virtual 5-cycle, then $\ensuremath{\widetilde{u}}\sim\ensuremath{\widetilde{v}}$.
\end{proposition}
\begin{proof} We say that a vertex $\ensuremath{\widetilde{a}}$ of $\ensuremath{\widetilde{B}}_{i+1}$ has \emph{height} $j$ (notation $h(\ensuremath{\widetilde{a}})=j$) if $\ensuremath{\widetilde{a}}\in \ensuremath{\widetilde{S}}_j$. We call an edge $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$
\emph{horizontal of height $j$} if $h(\ensuremath{\widetilde{a}})=h(\ensuremath{\widetilde{b}})=j$ and \emph{vertical} if $h(\ensuremath{\widetilde{a}})\ne h(\ensuremath{\widetilde{b}})$.
From the definition of edges of the graph $\ensuremath{\widetilde{G}}_{i+1}$, $|h(\ensuremath{\widetilde{a}})-h(\ensuremath{\widetilde{b}})|\le 1$ for any edge $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}$.
The \emph{height} $h(\ensuremath{\widetilde{\pi}})$ of a virtual 5-cycle $\ensuremath{\widetilde{\pi}}$ is the sum of heights of its vertices. We proceed by induction
on the height $h(\ensuremath{\widetilde{\pi}})$ of $\ensuremath{\widetilde{\pi}}$. If all vertices of $\ensuremath{\widetilde{\pi}}$ belong to $\ensuremath{\widetilde{B}}_i$, then applying ($S_i$) we get $\ensuremath{\widetilde{u}}\sim\ensuremath{\widetilde{v}}$.
Thus, we will assume that $\ensuremath{\widetilde{\pi}}$
contains a vertex of $\ensuremath{\widetilde{S}}_{i+1}$.
\medskip\noindent
{\bf Case 1.} The cycle $\pi=uxwyv$ has a chord.
\begin{proof} First, let $u \sim w$ or $v\sim w$, say the first. Then applying Lemma \ref{lemma: virtC4} to the path $\ensuremath{\widetilde{u}} \ensuremath{\widetilde{w}} \ensuremath{\widetilde{y}} \ensuremath{\widetilde{v}}$ we get $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{v}}$. Now, let
$w\nsim u,v$. Since $G$ does not contain induced $4$-cycles, we conclude that $y\sim x$. By ($R_{i+1}$) or ($T_{i+1}$) , we get $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{y}}$. Applying Lemma \ref{lemma: virtC4}
to the path $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$, we get $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{v}}$.
\end{proof}
In the remaining part of the proof we assume that $\pi=uxwyv$ is a pentagon of $G$. This implies that $\ensuremath{\widetilde{\pi}}$ does not contain \emph{peaks}, i.e., vertices whose heights are larger than the heights of
their neighbors. Indeed, if say $h(\ensuremath{\widetilde{x}})<h(\ensuremath{\widetilde{w}})>h(\ensuremath{\widetilde{y}})$, then by $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ we have $\ensuremath{\widetilde{x}}\sim \ensuremath{\widetilde{y}}$ and thus $x\sim y$, a contradiction.
\medskip\noindent
{\bf Case 2.} $|h(\ensuremath{\widetilde{u}})-h(\ensuremath{\widetilde{v}})|\ge 2$.
\begin{proof}
Let $h(\ensuremath{\widetilde{u}})\ge h(\ensuremath{\widetilde{v}})+2$. By ($R_i$) there exists $\ensuremath{\widetilde{u}}'\sim \ensuremath{\widetilde{v}}$. Since $h(\ensuremath{\widetilde{u}}')\le h(\ensuremath{\widetilde{v}})+1<h(\ensuremath{\widetilde{u}})$, the height of the virtual 5-cycle $\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}\ensuremath{\widetilde{u}}'$ is less than $h(\ensuremath{\widetilde{\pi}})$. By induction hypothesis, $\ensuremath{\widetilde{x}}\sim \ensuremath{\widetilde{u}}'$, contradicting Lemma \ref{lem: injloc}.
\end{proof}
Therefore, further we can suppose that $|h(\ensuremath{\widetilde{u}})-h(\ensuremath{\widetilde{v}})|\le 1$.
\medskip\noindent
{\bf Case 3.} All edges of $\ensuremath{\widetilde{\pi}}$ are vertical.
\begin{proof}
Since $\ensuremath{\widetilde{\pi}}$ does not contain peaks and horizontal edges and $|h(\ensuremath{\widetilde{u}})-h(\ensuremath{\widetilde{v}})|\le 1$, necessarily $h(\ensuremath{\widetilde{u}})>h(\ensuremath{\widetilde{x}})>h(\ensuremath{\widetilde{w}})<h(\ensuremath{\widetilde{y}})<h(\ensuremath{\widetilde{v}})$.
Applying rule (2) of the definition of edges of $\ensuremath{\widetilde{G}}_{i+1}$, we get $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{v}}$.
\end{proof}
Thus, we can assume that at least one edge of $\ensuremath{\widetilde{\pi}}$ is horizontal. We pick as $\ensuremath{\widetilde{e}}$ the second
or the third edge of $\ensuremath{\widetilde{\pi}}$ if such an edge is horizontal and the first or the fourth edge if this edge is horizontal and $\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$
and $\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}$ are vertical. Then $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0)$ applies to $\ensuremath{\widetilde{e}}$. Denote by $j$ the height of the edge $\ensuremath{\widetilde{e}}$. Observe that if $\ensuremath{\widetilde{e}} \notin \{\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}},\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\}$, then $\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}, \ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}$ are vertical and since $\ensuremath{\widetilde{\pi}}$ does not contain peaks and $|h(\ensuremath{\widetilde{u}})-h(\ensuremath{\widetilde{v}})|\leq 1$, we necessarily have $h(\ensuremath{\widetilde{w}}) = j-1$.
\medskip\noindent
{\bf Case 4.} $\mathrm{TC}$ applies to the edge $\ensuremath{\widetilde{e}}$.
\begin{proof} First, let $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$. By $\mathrm{TC}$ there exists
$\ensuremath{\widetilde{z}} \in \ensuremath{\widetilde{B}}_{j-1}$ such that $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{x}}, \ensuremath{\widetilde{w}}$. Since $\ensuremath{\widetilde{\pi}}$ is a pentagon, $\ensuremath{\widetilde{z}}$ does not belong to $\ensuremath{\widetilde{\pi}}$.
If $z$ is adjacent to $y$ or to $u$, by applying ($T_{i+1}$) to $\ensuremath{\widetilde{w}}$
or to $\ensuremath{\widetilde{x}}$ we conclude that $\ensuremath{\widetilde{z}}\sim \ensuremath{\widetilde{y}}$ or $\ensuremath{\widetilde{z}}\sim
\ensuremath{\widetilde{u}}$. Consequently, we obtain a virtual 5-cycle $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ or
$\ensuremath{\widetilde{u}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ with height $h(\ensuremath{\widetilde{\pi}})-1$, and by induction hypothesis
we conclude that $\ensuremath{\widetilde{u}}\sim\ensuremath{\widetilde{v}}$. Suppose now that $z \nsim u,y$. Observe
that if $h(\ensuremath{\widetilde{u}}) = j-1$ (respectively, $h(\ensuremath{\widetilde{y}}) = j-1$), then by $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ we have $\ensuremath{\widetilde{z}}\sim \ensuremath{\widetilde{u}}$ (respectively, $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{y}}$)
and by ($R_{i+1}$) we have $z \sim u$ (respectively, $z \sim y$). Consequently, we have $h(\ensuremath{\widetilde{u}}) \geq j$ and $h(\ensuremath{\widetilde{y}}) \geq j$.
Consider the ball $B_2(z)$ of $G$. Since $u,y\in B_2(z)$, $u\nsim y$,
and $v\sim u,y,$ by convexity of $B_2(z)$, we have $d(z,v)\le 2$.
Since $z\nsim u,y$ and since $G$ does not contain squares, we
deduce that $d(z,v)=2$. Let $t$ be a common neighbor of $z$ and $v$.
By ($R_i$) , there exists a corresponding $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{z}}$ in
$\ensuremath{\widetilde{B}}_{j}$. Then $\pi'=\ensuremath{\widetilde{t}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ is a virtual 5-cycle. Since
$h(\ensuremath{\widetilde{u}})\ge h(\ensuremath{\widetilde{x}}) = j$, $h(\ensuremath{\widetilde{t}})\le j$, and $h(\ensuremath{\widetilde{z}})=j-1$, we have
$h(\ensuremath{\widetilde{\pi}}')\le h(\ensuremath{\widetilde{\pi}})-1$. By induction hypothesis, $\ensuremath{\widetilde{t}}\sim
\ensuremath{\widetilde{v}}$. Consequently, $\pi''=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{t}}\ensuremath{\widetilde{v}}$ is a virtual 5-cycle and since
$h(\ensuremath{\widetilde{y}})\ge h(\ensuremath{\widetilde{w}}) = j$, $h(\ensuremath{\widetilde{t}})\le j$, and $h(\ensuremath{\widetilde{z}})=j-1$, we have
$h(\ensuremath{\widetilde{\pi}}'')\le h(\ensuremath{\widetilde{\pi}})-1$. By induction hypothesis, $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{v}}$.
Now, let $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$.
In this case, $h(\ensuremath{\widetilde{u}}) = h(\ensuremath{\widetilde{x}}) = j$ and $h(\ensuremath{\widetilde{w}}) = j-1$. Since
$\mathrm{TC}$ applies to $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$, there exists $\ensuremath{\widetilde{z}} \in \ensuremath{\widetilde{B}}_{j-1}$
such that $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{u}}, \ensuremath{\widetilde{x}}$. By $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$, $\ensuremath{\widetilde{z}}\sim \ensuremath{\widetilde{w}}$ and
thus $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ is a virtual 5-cycle with height
$h(\ensuremath{\widetilde{\pi}})-1$. By induction hypothesis, $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{v}}$, concluding the
analysis of Case 4.
\end{proof}
Now suppose that only $\mathrm{PC}$ applies to the edge $\ensuremath{\widetilde{e}}$. If
$\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$, then by $\mathrm{PC}$ there exist
$\ensuremath{\widetilde{w}}_1, \ensuremath{\widetilde{w}}_2 \in \ensuremath{\widetilde{B}}_{j-1}$, and $\ensuremath{\widetilde{z}}\in \ensuremath{\widetilde{B}}_{j-2}$ such that
$\ensuremath{\widetilde{\pi}}_1=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}_2\ensuremath{\widetilde{w}}$ is a pentagon of $\ensuremath{\widetilde{G}}_{i+1}$ (and thus
$\pi_1=xw_1zw_2w$ is a pentagon of $G$). If $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$, then
$h(\ensuremath{\widetilde{w}}) = j-1$ and by Lemma~\ref{lem-ind-PCdist2}, there exists
$\ensuremath{\widetilde{w}}_1 \in \ensuremath{\widetilde{B}}_{j-1}$ and $\ensuremath{\widetilde{z}}\in \ensuremath{\widetilde{B}}_{j-2}$ such that
$\ensuremath{\widetilde{\pi}}_2=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{u}}$ is a pentagon of $\ensuremath{\widetilde{G}}_{i+1}$ (and thus
$\pi_2=xw_1zwy$ is a pentagon of $G$). Note that in both cases, the
image $e$ of $\ensuremath{\widetilde{e}}$ is a common edge of $\pi$ and of the new pentagon
$\pi_1$ or $\pi_2$. Denote the union $\pi\cup \pi_1$ or
$\pi\cup \pi_2$ by $U$. Since the balls of radius 3 of $G$ are
convex, $U$ has diameter 2 or 3. Moreover, if $\ensuremath{\widetilde{e}} = \ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$, then
$\pi$ and $\pi_2$ share three vertices and in this case $U$ has
diameter $2$.
\begin{claim} \label{claim:better5virtcycle} Any virtual 5-cycle of the form $\ensuremath{\widetilde{\pi}}'=\ensuremath{\widetilde{p}}\ensuremath{\widetilde{a}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{b}}\ensuremath{\widetilde{q}}$ satisfies the induction hypothesis, thus $\ensuremath{\widetilde{p}}\sim \ensuremath{\widetilde{q}}$. \end{claim}
\begin{proof}
First notice that
$h(\ensuremath{\widetilde{\pi}}')=h(\ensuremath{\widetilde{p}})+h(\ensuremath{\widetilde{a}})+h(\ensuremath{\widetilde{z}})+h(\ensuremath{\widetilde{b}})+h(\ensuremath{\widetilde{q}})\le
j+j-1+j-2+j-1+j=5j-4$ and the equality holds only if
$h(\ensuremath{\widetilde{p}})=h(\ensuremath{\widetilde{q}})=j, h(\ensuremath{\widetilde{a}})=h(\ensuremath{\widetilde{b}})=j-1$, and $h(\ensuremath{\widetilde{z}})=j-2$. Therefore
$h(\ensuremath{\widetilde{\pi}}')=5j-4$ only if all edges of $\ensuremath{\widetilde{\pi}}'$ are vertical. In this
case, the result holds by Case 3. Now we will show that
$h(\ensuremath{\widetilde{\pi}})\ge 5j-4$.
First suppose that $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$. Then $h(\ensuremath{\widetilde{u}})\ge h(\ensuremath{\widetilde{x}})-1=j-1$,
$h(\ensuremath{\widetilde{y}})\ge h(\ensuremath{\widetilde{w}})-1=j-1$, and $h(\ensuremath{\widetilde{v}})\ge h(\ensuremath{\widetilde{y}})-1\ge j-2$, yielding
$h(\ensuremath{\widetilde{\pi}})\ge j-1+j+j+j-1+j-2=5j-4$.
Now suppose that $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$. Then $h(\ensuremath{\widetilde{u}})=h(\ensuremath{\widetilde{x}})=j$,
$h(\ensuremath{\widetilde{v}})\ge h(\ensuremath{\widetilde{u}})-1$, $h(\ensuremath{\widetilde{w}})=j-1$, and
$h(\ensuremath{\widetilde{y}})\ge h(\ensuremath{\widetilde{w}})-1=j-2$, we deduce that
$h(\ensuremath{\widetilde{\pi}})\ge j+j+j-1+j-2+j-1=5j-4$.
\end{proof}
\medskip\noindent
{\bf Case 5.} $\mathrm{diam}(U)=3$.
\begin{proof}
In this case, necessarily $\ensuremath{\widetilde{u}} = \ensuremath{\widetilde{w}}\ensuremath{\widetilde{x}}$ and the vertices of
$\pi \setminus e$ and $\pi_1\setminus e$ are pairwise distinct. By
Lemma \ref{lem: deuxpent}, $\pi$ or $\pi_1$ has a universal vertex
$t$. First suppose that $t$ is a universal for $\pi_1$. By ($R_i$)
there exists $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{z}}$ in $\ensuremath{\widetilde{B}}_{j-1}$. By ($R_{i+1}$), $\ensuremath{\widetilde{t}}$
is adjacent to all vertices of $\ensuremath{\widetilde{\pi}}_1=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}_2\ensuremath{\widetilde{w}}$. This
contradicts the assumption that only $\mathrm{PC}$ applies to $\ensuremath{\widetilde{e}}$.
Now suppose that $t$ is a universal vertex for $\pi=uvywx$. Since
$B_2(t)$ is convex in $G$ and $\mathrm{diam}(U)=3$, we have $d(t,z)=2$. Let
$s$ be a common neighbor of $t$ and $z$. By ($R_i$) there exists
$\ensuremath{\widetilde{s}} \in\ensuremath{\widetilde{B}}_{j-1}$ adjacent to $\ensuremath{\widetilde{z}}$. Applying ($R_i$) the second time
we conclude that there exists $\ensuremath{\widetilde{t}}\in \ensuremath{\widetilde{B}}_j$ such that
$\ensuremath{\widetilde{t}}\sim \ensuremath{\widetilde{s}}$. By Claim~\ref{claim:better5virtcycle} applied to the
virtual 5-cycle $\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}\ensuremath{\widetilde{s}}\ensuremath{\widetilde{t}}$, we have $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{x}}$.
Applying several times ($T_{i+1}$), we conclude that
$\ensuremath{\widetilde{t}}\sim \ensuremath{\widetilde{u}},\ensuremath{\widetilde{w}},\ensuremath{\widetilde{y}},\ensuremath{\widetilde{v}}$. Applying ($T_{i+1}$) once again to
$\ensuremath{\widetilde{t}}\sim \ensuremath{\widetilde{u}},\ensuremath{\widetilde{v}}$, we deduce that $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{v}}$.
\end{proof}
\medskip\noindent
{\bf Case 6.} $\mathrm{diam}(U)=2$.
\begin{proof}
We start with a special subcase, which is used to prove a useful claim.
\medskip\noindent
{\bf Subcase 6.1.} $h(\ensuremath{\widetilde{w}})=i-1$.
\medskip
Let $h(\ensuremath{\widetilde{v}})\ge h(\ensuremath{\widetilde{u}})$. Since $\max\{ h(\ensuremath{\widetilde{x}}),h(\ensuremath{\widetilde{y}})\}\le h(\ensuremath{\widetilde{w}})+1=i$ and $\ensuremath{\widetilde{\pi}}$ contains a vertex of height $i+1$, necessarily $h(\ensuremath{\widetilde{v}})=i+1$.
Since $h(\ensuremath{\widetilde{u}})\ge h(\ensuremath{\widetilde{v}})-1$ and $\ensuremath{\widetilde{\pi}}$ contains horizontal edges (by Cases 1-2), either $h(\ensuremath{\widetilde{u}})=i=h(\ensuremath{\widetilde{y}}), h(\ensuremath{\widetilde{x}})=h(\ensuremath{\widetilde{w}})=i-1, h(\ensuremath{\widetilde{v}})=i+1$ or
$h(\ensuremath{\widetilde{u}})=h(\ensuremath{\widetilde{x}})=i=h(\ensuremath{\widetilde{y}}),h(\ensuremath{\widetilde{w}})=i-1,h(\ensuremath{\widetilde{v}})=i+1$. Thus the edge $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$ or $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$ is the unique horizontal edge of $\ensuremath{\widetilde{\pi}}$.
First, let $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$. Since $\mathrm{diam}(U)=2$, $d(u,z)\le 2$. If $u=z$ or $u\sim z$, then Lemma \ref{lemma: virtC4} applied to the path $\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{x}}\ensuremath{\widetilde{u}}$ implies
$\ensuremath{\widetilde{z}}=\ensuremath{\widetilde{u}}$ or $\ensuremath{\widetilde{z}}\sim \ensuremath{\widetilde{u}}$, which is impossible because $h(\ensuremath{\widetilde{u}})=i$ and $h(\ensuremath{\widetilde{z}})=i-3$. Thus $d(u,z)=2$ and let $s$ be a common neighbor of $u$ and $z$.
By ($R_i$) there exists $\ensuremath{\widetilde{s}}\sim \ensuremath{\widetilde{z}}$ with $h(\ensuremath{\widetilde{s}})\le i-2$. The virtual 5-cycle $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}\ensuremath{\widetilde{s}}$ is included $\ensuremath{\widetilde{B}}_i$, thus $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{s}}$ by induction hypothesis, contrary
to $h(\ensuremath{\widetilde{u}})=i$ and $h(\ensuremath{\widetilde{s}})\le i-2$. Now, let $\ensuremath{\widetilde{e}}=\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$. Then $h(\ensuremath{\widetilde{z}})=i-2$. As $h(\ensuremath{\widetilde{u}})=i$, there exists $\ensuremath{\widetilde{v}}'\in \ensuremath{\widetilde{B}}_{i+1}$ in the preimage of $v$ such that $\ensuremath{\widetilde{v}}'\sim \ensuremath{\widetilde{u}}$. If $h(\ensuremath{\widetilde{v}}')\leq i$,
then setting $\pi'=\ensuremath{\widetilde{v}}'\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}$, we have $h(\ensuremath{\widetilde{\pi}}')<h(\ensuremath{\widetilde{\pi}})$, and by induction hypothesis we get $\ensuremath{\widetilde{v}}'\sim \ensuremath{\widetilde{y}}$. By ($R_{i+1}$) we get $\ensuremath{\widetilde{v}}'=\ensuremath{\widetilde{v}}$.
Assume now that $h(\ensuremath{\widetilde{v}}')=i+1$. If $d(v,z)<2$, we immediately get a contradiction with Lemma \ref{lemma: virtC4} applied to $\ensuremath{\widetilde{z}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{u}}\ensuremath{\widetilde{v}}'$. Thus $d(v,z)=2$
and there exists $t \sim v,z$. By ($R_i$) there is an associated $\ensuremath{\widetilde{t}}$ in the preimage of $t$ such that $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{z}}$. Applying Case 2 to $\ensuremath{\widetilde{v}}'\ensuremath{\widetilde{u}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}\ensuremath{\widetilde{t}}$, we get a
contradiction because $h(\ensuremath{\widetilde{v}}')-h(\ensuremath{\widetilde{t}})\geq 2$.
\begin{claim}\label{claim:virtC4} Let $\ensuremath{\widetilde{a}}\ensuremath{\widetilde{b}}\ensuremath{\widetilde{c}}\ensuremath{\widetilde{d}}$ be a path of $\ensuremath{\widetilde{G}}_{i+1}$ with $h(\ensuremath{\widetilde{d}})\le i-1$. Then $d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{a}},\ensuremath{\widetilde{d}})=d_G(a,d)$.
Moreover, if $d(a,d)=2$, for any $t\sim a,d$ in $G$, then there exists $\ensuremath{\widetilde{t}}\sim\ensuremath{\widetilde{a}},\ensuremath{\widetilde{d}}$ in $\ensuremath{\widetilde{G}}_{i+1}$ such that $f_{i+1}(\ensuremath{\widetilde{t}}) = t$.
\end{claim}
\begin{proof}
The cases $d(a,d)\in\{ 0,1\}$ follow from Lemma \ref{lemma: virtC4}. Now, let $d(a,d)=2$ and let $t \sim a,d$. Since $\ensuremath{\widetilde{d}}\in \ensuremath{\widetilde{B}}_{i-1}$, by ($R_i$) there exists $\ensuremath{\widetilde{t}} \in \ensuremath{\widetilde{B}}_i$ in the preimage of $t$ such that $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{d}}$. By ($R_{i+1}$) there exists $\ensuremath{\widetilde{a}}' \in \ensuremath{\widetilde{B}}_{i+1}$ in the preimage of $a$ such that $\ensuremath{\widetilde{a}}' \sim \ensuremath{\widetilde{t}}$.
Consider the virtual 5-cycle $\ensuremath{\widetilde{\pi}}'= \ensuremath{\widetilde{b}} \ensuremath{\widetilde{c}} \ensuremath{\widetilde{d}} \ensuremath{\widetilde{t}} \ensuremath{\widetilde{a}}'$. Since $h(\ensuremath{\widetilde{d}})\leq i-1$, to $\ensuremath{\widetilde{\pi}}'$ we can apply Subcase 6.1, whence $\ensuremath{\widetilde{a}}'\sim \ensuremath{\widetilde{b}}$.
By ($R_{i+1}$) we must have $\ensuremath{\widetilde{a}}=\ensuremath{\widetilde{a}}'$, hence $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{a}}, \ensuremath{\widetilde{d}}$.
\end{proof}
Since $\mathrm{diam}(U)=2$, in $G$ the vertex $z$ has distance at most 2 from all vertices of $\pi$.
Now we prove that in $\ensuremath{\widetilde{G}}_{i+1}$ the vertex $\ensuremath{\widetilde{z}}$ also has distance at most 2 from all vertices of $\ensuremath{\widetilde{\pi}}$.
Suppose first that $\ensuremath{\widetilde{e}} = \ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}$. Clearly,
$d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{x}})=d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{w}})=2$. Applying Claim
\ref{claim:virtC4} to the paths $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}_1\ensuremath{\widetilde{z}}$ and $\ensuremath{\widetilde{y}}\ensuremath{\widetilde{w}}\tw_2\ensuremath{\widetilde{z}}$,
we conclude that $d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{u}})=d_G(z,u)\le 2$ and
$d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{y}})=d_G(z,y)\le 2$. Finally, we show that
$d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{v}})\le 2.$ If this is not the case, then
necessarily $d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{y}})=d_G(z,y)=2$ and by Claim
\ref{claim:virtC4} there exists $\ensuremath{\widetilde{t}}$ such that $\ensuremath{\widetilde{t}} \sim
\ensuremath{\widetilde{z}},\ensuremath{\widetilde{y}}$. By Claim \ref{claim:virtC4} applied to the path
$\ensuremath{\widetilde{v}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{t}}\ensuremath{\widetilde{z}}$ we conclude that
$d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{v}})=d_G(z,v)\le 2$.
Assume now that $\ensuremath{\widetilde{e}} = \ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}$. Since in this case, $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{w}}$,
we have $d(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{y}}) \leq 2$ and Applying Claim \ref{claim:virtC4} to
the path $\ensuremath{\widetilde{v}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{w}} \ensuremath{\widetilde{z}}$, we have
$d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{v}})=d_G(z,v)\le 2$.
Consequently, in both cases $\ensuremath{\widetilde{z}}$ is at distance at most $2$ from $\ensuremath{\widetilde{u}}$ and $\ensuremath{\widetilde{v}}$.
If $d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{u}}) \leq 1$ or $d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}}, \ensuremath{\widetilde{v}})\leq 1$, then $\ensuremath{\widetilde{u}}$ and $\ensuremath{\widetilde{v}}$ can be connected by a path of length at most $3$ passing via $\ensuremath{\widetilde{z}}$.
Applying Lemma \ref{lemma: virtC4} to this path, we conclude that $\ensuremath{\widetilde{u}}\sim\ensuremath{\widetilde{v}}$. Assume now that $d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{u}})=d_{\ensuremath{\widetilde{G}}_{i+1}}(\ensuremath{\widetilde{z}},\ensuremath{\widetilde{v}})=2$.
Then there exist $\ensuremath{\widetilde{t}} \sim \ensuremath{\widetilde{z}},\ensuremath{\widetilde{u}}$ and $\ensuremath{\widetilde{s}} \sim \ensuremath{\widetilde{z}}, \ensuremath{\widetilde{v}}$. If $\ensuremath{\widetilde{t}}=\ensuremath{\widetilde{s}}$, then we get $\ensuremath{\widetilde{u}} \sim \ensuremath{\widetilde{v}}$ by ($R_{i+1}$) applied to $\ensuremath{\widetilde{t}}=\ensuremath{\widetilde{s}}$.
Otherwise, consider the virtual 5-cycle $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{t}}\ensuremath{\widetilde{z}}\ensuremath{\widetilde{s}}\ensuremath{\widetilde{v}}$. By Claim \ref{claim:better5virtcycle}, we conclude that $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{v}}$. This finishes the proof of Case 6.
\end{proof}
In all cases we proved that $\ensuremath{\widetilde{u}}\sim \ensuremath{\widetilde{v}}$, establishing ($S_{i+1}$) and concluding the proof of the proposition.
\end{proof}
\subsection{Universal Cover}
As for every $i\geq 0$, $\ensuremath{\widetilde{G}}_i$ satisfies ($Q_i$), the graph $\ensuremath{\widetilde{G}}$ satisfies $\mathrm{INC}(\ensuremath{\widetilde{v}}_0)$ and $\mathrm{TPC}(\ensuremath{\widetilde{v}}_0)$, so in particular by Lemma \ref{lem: triangle-pentagon-simply-connected}, the complex $\ensuremath{\widetilde{X}}= X_{\scaleobj{0.8}{\triangle}, \pentagon}(\ensuremath{\widetilde{G}})$ is simply connected.
We recall that to end the proof of Theorem \ref{thm: triangle-pentagon}, we need to show that $\ensuremath{\widetilde{X}}$ is a cover of $X$. Indeed, as there is a unique (up to isomorphism) simply connected cover of $X= X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$, this will imply that $\ensuremath{\widetilde{X}}=X$. As $\ensuremath{\widetilde{X}}$ can be constructed from any arbitrary basepoint $v_0\in V$, in particular $G$ satisfies $\mathrm{INC}(v)$ and $\mathrm{TPC}(v)$ for any $v\in V$, so it has convex balls.
To complete the proof that $f := \bigcup_{i\geq 0}f_i : \ensuremath{\widetilde{X}} \to X$ is
a covering map, we show that $f$ induces an isomorphism between the
(closed) star of a vertex of $\ensuremath{\widetilde{X}}$ and the (closed) star of its image in
$X$.
\begin{lemma}
\label{lem: etoiles}
For every $\widetilde w \in \ensuremath{\widetilde{X}}$, the map $f$ induces an isomorphism between the closed stars $\cSt(\ensuremath{\widetilde{w}},\widetilde{X})$ and $\cSt(w,X)$, where $w := f(\ensuremath{\widetilde{w}})$. Thus $f$ is a covering map.
\end{lemma}
\begin{proof} That $f$ defines a morphism from $\ensuremath{\widetilde{X}}$ to $X$ follows from Lemmas \ref{lem: RipTip} and \ref{lem: morphism}. It remains to show that $f$ is bijective on stars. To do so, we establish the stronger result that $f$ is bijective on closed stars. Let $i:=d(\ensuremath{\widetilde{v}}_0, \ensuremath{\widetilde{v}})$. Then the restriction of $f$ on the set $\cSt(\ensuremath{\widetilde{w}},\widetilde{X})$ equals to the restriction of $f_{i+2}$ on this set. By ($R_{i+1}$) , $f$ induces an isomorphism between $\ensuremath{\widetilde{G}}\left[B_1(\ensuremath{\widetilde{w}})\right]$ and $G\left[B_1(w)\right]$.
Let $u \in \cSt(w,X)\backslash B_1(w)$. Then there exists a pentagon $uxwyv$ in $G$. By iterated applications of ($R_{i+2}$), we find an associated virtual 5-cycle $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ in $\ensuremath{\widetilde{G}}$. By ($S_{i+2}$), $\ensuremath{\widetilde{u}}\ensuremath{\widetilde{x}}\ensuremath{\widetilde{w}}\ensuremath{\widetilde{y}}\ensuremath{\widetilde{v}}$ is a pentagon of $\ensuremath{\widetilde{G}}$, so it is a cell of $\cSt(\ensuremath{\widetilde{w}},\widetilde{X})$. Thus we proved that $f|_{\cSt(\ensuremath{\widetilde{w}},\widetilde{X})}$ is onto $\cSt(w,X)$.
It remains to show that $f|_{\cSt(\ensuremath{\widetilde{w}},\widetilde{X})}$ is injective. Assume that there exist $\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}' \in \cSt(\ensuremath{\widetilde{w}},\widetilde{X})$ such that $f(\ensuremath{\widetilde{u}}) = f(\ensuremath{\widetilde{u}}') = u$. By the previous remark, we cannot have $\ensuremath{\widetilde{u}}, \ensuremath{\widetilde{u}}' \in B_1(\ensuremath{\widetilde{w}})$, so we can assume that $\ensuremath{\widetilde{u}}\notin B_1(\ensuremath{\widetilde{w}})$. Let $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{u}}, \ensuremath{\widetilde{w}}$. Then by ($R_{i+2}$) we must have $u \neq w$ and $u \nsim w$, hence $\ensuremath{\widetilde{u}}' \notin B_1(\ensuremath{\widetilde{w}})$ and there exists $\ensuremath{\widetilde{z}} \sim \ensuremath{\widetilde{u}}', \ensuremath{\widetilde{w}}$. Set $x:= f(\ensuremath{\widetilde{x}})$ and $z:= f(\ensuremath{\widetilde{z}})$. By ($R_{i+2}$) we get $x\sim u, w$ and $z \sim u, w$.
If $x = z$, then by ($R_{i+2}$) applied two times we must have $\ensuremath{\widetilde{x}} = \ensuremath{\widetilde{z}}$ and $\ensuremath{\widetilde{u}} = \ensuremath{\widetilde{u}}'$. If $x\neq z$, then the 4-cycle $uxwz$ cannot be induced, so we must have $x\sim z$. Then by iterated applications of ($R_{i+2}$) we get $\ensuremath{\widetilde{x}} \sim \ensuremath{\widetilde{z}}$ and $\ensuremath{\widetilde{u}}' \sim \ensuremath{\widetilde{x}}$ and eventually $\ensuremath{\widetilde{u}} = \ensuremath{\widetilde{u}}'$. Hence we proved the injectivity of $f|_{\cSt(\ensuremath{\widetilde{w}},\widetilde{X})}$.
\end{proof}
\subsection{A structural characterization of CB-graphs}
The goal of this subsection is to derive from Theorem \ref{thm: triangle-pentagon} a characterization of CB-graphs in terms of forbidden substructures.
For this we will show that the local conditions of Theorem \ref{thm: triangle-pentagon} can be obtained as follows:
\begin{lemma}
\label{lem: INC3 TPC3}
Let $G$ be a graph without isometric $C_4, C_6, C_7$, and $\mathrm{PT}$ and such that any induced subgraph isomorphic to $\mathrm{PP}_2$ has diameter $2$. Then $G$ satisfies $\mathrm{INC}_{\leq 3}$ and $\mathrm{TPC}^0_{\leq 3}$.
\end{lemma}
\begin{proof}
We first show that $G$ satisfies $\mathrm{INC}_{\leq 3}$. $\mathrm{INC}_{\leq 2}$
trivially holds since $G$ does not contain induced $C_4$. Let
$u,v\in V$ with $d(u,v)=3$ and let $w,w'\in I(u,v)$ be two distinct
vertices both adjacent to $u$. Let $a \sim w,v$ and $b\sim w',v$. If
$a=b$ we are immediately done by $\mathrm{INC}_{\leq 2}(u,a)$. Thus assume
that $a\neq b$. We first establish that $w\nsim w'$. Suppose by
contradiction that $w\nsim w'$. The $6$-cycle $vawuw'b$ cannot be
isometric in $G$. As $d(u,v)=3$, either $d(a,w')\leq 2$ or
$d(b,w)\leq 2$. Assume without loss of generality that
$d(a,w')\leq 2$. If $a \sim w'$, then we are done when considering
the $4$-cycle $aw'uw$. Assume now that $d(a,w')=2$. Then there
exists $x\sim a, w'$. By hypothesis, $x\neq w$. If $x=b$, then as
$a\nsim w'$, $a\nsim u$, $w\nsim w'$, $b\nsim w$ and as there is no
isometric $\mathrm{PT}$, we must have $d(v,u)\leq 2$, which is a
contradiction. Thus we may assume that $x \neq b$ and $a\nsim
b$. Observe that if $x \sim v$, we can replace $b$ by $x$ and we get
a contradiction. Consequently, if $x \nsim b$, $abxw$ is an induced
square, which is impossible. Therefore, the $5$-cycle
$\pi_1:= vaxw'b$ is induced. If the $5$-cycle $\pi_2:= axw'uw$ is
also induced, then $\pi_1$ and $\pi_2$ induce an induced $\mathrm{PP}_2$
which must have diameter $2$, contradicting $d(u,v) =
3$. Consequently, $\pi_2$ has a chord and necessarily $x \sim
w,u$. Therefore, $\pi_1$ and $u$ induce an isometric $\mathrm{PT}$, which is
impossible. This shows that $w \sim w'$. Since $G$ does not contain
isometric $\mathrm{PT}$, the $5$-cycle $ww'avb$ cannot be induced and thus
necessarily either $a \sim w'$ or $b \sim w$. This shows that $G$
satisfies $\mathrm{INC}_{\leq 3}$.
We assume now that $\mathrm{INC}_{\leq 3}$ holds and prove $\mathrm{TPC}^0_{\leq 3}$. Let $v,x,y\in V$ such that $x\sim y$ and the edge $xy$ is at uniform distance $3$ from $v$. We let $w\in I(v,x)$, $w'\in I(v,y)$ such that $w\sim x$, $w'\sim y$, $a\sim v,w$, and $b\sim v,w'$. We may assume that $a\neq b$, $d(w,w')\geq 2$, $d(a,y)=3$, and $d(b,x)=3$ otherwise we are done. Moreover we may assume that $d(a,w')=3$, otherwise we get $w'\sim x$ by $\mathrm{INC}^0(y,a)$. Similarly we assume that $d(b,w)=3$. As the $7$-cycle $vawxyw'b$ is not isometric, this means that we must have $d(w,w')= 2$. Now by $\mathrm{INC}^0(b,w)$ we must have either $v\sim w'$ or $d(w,b)\leq 2$, contradicting in both cases our hypothesis.
\end{proof}
\begin{corollary}
For a graph $G$ the following conditions are equivalent:
\begin{itemize}
\item[(i)] $G$ is a CB-graph;
\item[(ii)] $G$ has no isometric $C_k$ for every $k\geq 4$ such that $k\neq 5$, no isometric $\mathrm{PT}$, every induced subgraph isomorphic to $\mathrm{PP}_1$ has diameter at most $3$ and every induced subgraph isomorphic to $\mathrm{PP}_2$ has diameter $2$.
\end{itemize}
Furthermore, if $G$ has no isometric $C_4, C_6, C_7, C_8$, no isometric $\mathrm{PT}$, every induced subgraph isomorphic to $\mathrm{PP}_1$ has diameter at most $3$ and every induced subgraph isomorphic to $\mathrm{PP}_2$ has diameter $2$, then the 1-skeleton $\widetilde{G}$ of the universal cover of
$X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is a graph with convex balls.
\end{corollary}
\begin{proof}
The implication (i)$\Rightarrow$(ii) is easy. To prove the converse, we would like to prove that item $(ii)$ from Theorem \ref{thm: triangle-pentagon} holds. For this, thanks to Lemma \ref{lem: INC3 TPC3} we only have to show that $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ is simply connected. We show by induction over $k\geq 3$ that every cycle $C_k$ is null-homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$. If $k\in \sg{3,5}$ the result is immediate.
Otherwise, $C_k$ is not isometric, so it has two vertices $a,b$ such that $d_G(a,b)<d_{C_k}(a,b)$. Taking a shortest path between $a$ and $b$ in $G$, we cut $C_k$ into two smaller cycles $C$ and $C'$. By induction hypothesis, both $C$ and $C'$ are null-homotopic in $X_{\scaleobj{0.8}{\triangle}, \pentagon}(G)$ and it is not hard to deduce that $C_k$ is also null-homotopic.
\end{proof}
\section{Contractibility of Rips complexes}
\label{sec: dism}
In this section, we prove that the square $G^2$ of any CB-graph $G$ is dismantlable and that the dismantling order can be obtained by BFS ordering the vertices of $G$.
This implies that if $G$ is locally-finite, then the clique complex $X(G^2)$ of $G^2$ is contractible. Consequently,
this shows that all Rips complexes $X_k(G)$, $k\ge 2$ of a locally-finite CB-graph $G$ are contractible.
\subsection{Dismantlability of squares}
For several subclasses of weakly modular graphs, BFS
(Breadth-First-Search) and its refinements turn out to provide
orderings with interesting and strong properties, which can be used, for instance,
to prove contractibility of associated clique complexes. First, it was shown
in~\cite{Ch_bridged} that for locally finite bridged graphs, any BFS ordering is a
dismantling ordering, showing in particular that the clique complexes of bridged
graphs are contractible. Polat \cite{Po_bridged1} proved that arbitrary
connected graphs (even if they are not locally finite) admit a BFS ordering and,
extending the result of~\cite{Ch_bridged}, he showed that BFS provides a dismantling
order for non-locally-finite bridged
graphs. For weakly bridged graphs the same kind of results has been obtained
for specific BFS orderings. Namely, any LexBFS ordering of a locally
finite weakly bridged graph provides a dismantling
ordering~\cite{ChepoiOsajda}. In the case of non-locally-finite graphs, it is
not always possible to define a LexBFS ordering. However, for graphs
without infinite cliques, it was shown in~\cite{Bresaretal2013} that it is always
possible to define an ordering, intermediate between BFS and LexBFS, and called SimpLexBFS,
and it was shown that for weakly bridged graphs, any SimpLexBFS ordering is a
dismantling ordering. Notice also that the contractibility of Kakimizu complexes was
established by defining a BFS-like orderings of their vertices; for details,
see Section 5 of \cite{PrzSch}.
A vertex $x$ of a graph $G$ is {\it dominated} by another vertex $y$ if the unit ball $B_1(y)$
includes $B_1(x).$ A graph $G$ is {\it dismantlable} if
its vertices can be well-ordered $\prec$ so that, for each $v$ there is a neighbor $w$ of $v$ with
$w\prec v$ which dominates $v$ in the subgraph of $G$ induced by the vertices $u\preceq v$. The order $\prec$ is then called a \emph{dismantling order} of $G$.
Following Polat \cite{Po_bridged1}, a well-order $\preceq$ on the vertex set $V(G)$ of a graph $G$ is called a {\it BFS order} if there exists a family
$\{ A_x: x\in V(G)\}$ of subsets of $V(G)$ such that, for
every vertex $x\in V(G)$,
\begin{enumerate}[{(S}1)]
\item $x\in A_x$;
\item if $x\preceq y$, then $(A_x,\preceq)$ is an initial segment of $(A_y,\preceq)$;
\item $A_x=A_{(x)}\cup N(x)$, where $A_{(x)}:=\{ x\}$ if $x$ is the least element of $(V(G),\preceq)$ and $A_{(x)}:=\bigcup_{y\prec x} A_y$ otherwise.
\end{enumerate}
\begin{lemma} \label{l:Polat_BFS} \cite{Po_bridged1}*{Lemma 3.6} There exists a BFS order on the vertex set of any connected graph.
\end{lemma}
The vertex $x$ will be called the {\it parent} of each vertex of $A_x\setminus A_{(x)}$. We will denote by $f$ the map from $V(G)$ to $V(G)$
such that $f(v)$ is the parent of $v$, for every $v\in V(G)$. The least element of $(V(G),\preceq)$ will be called the {\it base-point}
and will be denoted by $v_0$ (by convention, we set $f(v_0)=v_0)$). Notice that like in the case of finite graphs, for every vertices $x$ and $y$ of $G$,
$x\preceq y$ implies $d(v_0,x)\le d(v_0,y)$, and $d(v_0,x)<d(x_0,y)$ implies $x\prec y$. In particular, $d(v_0,x)=d(v_0,f(x))+1$.
For two distinct vertices $x$ and $y$ of $G$, we set $\max\{ x,y\}=x$ if $y\prec x$ and $\max\{ x,y\}=y$ if $x\prec y$.
The main goal is to prove the following result:
\begin{theorem} \label{dismantl} Any BFS order of the vertices of a CB-graph $G$ is a dismantling order of its square $G^2$.
\end{theorem}
\begin{proof} We start with several properties of BFS on all graphs.
Let $b$ be a basepoint of $G$ and let $\preceq$ be the basepoint (partial) order of the vertices of $G$:
for two vertices $u,v$ of $G$ we set $u\preceq v$ if and only if
$d(b,u)\le d(b,v)$. For a vertex $v$, let $F(v)=\{ u\in V: d(b,u)=d(b,v)+d(v,u)\}$ and call $F(v)$ the
{\it filter} of $v$ with respect to the basepoint order $\preceq$. Let $\prec$
be a BFS (total) order of the vertices of $G$ with basepoint $b$. Clearly, $\prec$ is a linear extension
of $\preceq$. For a vertex $v$ of $G$, let $f(v)$ be the \emph{parent} of $v$
defined by BFS, $f^2(v)=f(f(v))$ be the \emph{grandparent} of $v$, and so on,
let $f^i(v)=f(f^{i-1}(v))$; if $u = f^i(v)$ for some $i$, we say that $u$ is an \emph{ascendant} of $v$ and that $v$ is a \emph{descendant} of $u$.
\begin{lemma}If $x\prec y\prec z,$ $d(b,x)=d(b,z)=:k$, and $f^2(x)=f^2(z)=:s,$
then $f^2(y)=s$.
\end{lemma}
\begin{proof}
Suppose $f^2(y)=s'\ne s$. Since $x\prec y\prec z$ and
$d(b,x)=d(b,z)=k,$ necessarily $d(b,y)=k$ and
$f(x)\prec f(y)\prec f(z).$ Since $f^2(x)=f^2(z)=s$, all vertices
$t$ with $f(x)\prec t\prec f(z)$ will necessarily have $s$ as their
parent. In particular, $f(f(y))=s$.
\end{proof}
\begin{lemma}\label{BFS0}
Let $x,y,v$ be vertices of $G$ with $d(b,x)=d(b,y)$, $x=f^i(v)$, and
$v\in F(y)$. Then $x\prec y$.
\end{lemma}
\begin{proof} Suppose $y\prec x$. Since $v\in F(y)$ and $d(b,x)=d(b,y)$, we conclude that $d(y,v)=d(x,v)=i$.
Let $P=(y=z_{i},z_{i-1},\ldots,z_1,z_0=v)$ be a shortest path from $y$ to $v$ and $Q=(x=f^{i}(v),f^{i-1}(v),\ldots,f^1(v),v)$ be the shortest path
from $x$ to $v$ consisting of the ascendants of $v$. We can suppose that $P$ and $Q$ intersects only in $v$, otherwise
we can replace $v$ by a closest to $x$ and $y$ vertex from the intersection. Since $y\prec x$ and $x=f(f^{i-1}(v))$, we deduce that
$z_{i-1}\prec f^{i-1}(v)$. Continuing this way, we get $z_{i-j}\prec f^{i-j}(v)$ for any $j<i$. In particular,
$z_1\prec f^1(v)=f(v)$, contradicting that $f(v)$ is the parent of $v$.
\end{proof}
\begin{lemma} \label{BFS} Let $u,v$ be vertices of $G$ with $k:=d(b,v)-d(b,u)\ge 0$, $u\prec f^k(v)$, and $v\in F(f^i(u))$. Then $f^{k+i}(v)=f^i(u)$.
\end{lemma}
\begin{proof} Let $x=f^i(u)$ and $y=f^{i+k}(v)=f^i(f^k(v))$. Suppose $y\ne x$. Since $u\prec f^k(v)$, necessarily $x\prec y$.
Since $v\in F(x)$ and $d(v,y)=d(v,x)$, necessarily $d(b,x)=d(b,y)$. But this contradicts Lemma \ref{BFS0}.
\end{proof}
In the remaining results, we suppose that $G$ is a CB-graph.
\begin{lemma} \label{2-fellow-traveller} Let $u\prec v$.
If $d(b,u)=d(b,v)=k$ and $u\sim v$, then either $f(u)=f(v),$ or $f(v)\sim u,f(u),$ or $f(u)\nsim f(v)$ and $f^2(u)=f^2(v).$
\end{lemma}
\begin{proof}
Observe that if $f(u) \neq f(v)$, we have $f(u) \sim f(v)$ if and
only if $u \sim f(v)$. Indeed, if $u \sim f(v)$, then
$f(v),f(u)\in B_{k-1}(b)$ and $u\notin B_{k-1}(b),$ thus the
convexity of $B_{k-1}(b)$ implies that $f(v)\sim f(u)$. Conversely,
if $f(v)\sim f(u),$ since $G$ does not contain induced 4-cycles and
since $v\nsim f(u)$, as $f(v) \neq f(u)$ and $u \prec v$, we
conclude that $f(v)\sim u$.
We prove the lemma by induction on $k$, the cases $k=1,2$ being
trivial. Let $k\ge 3$. Suppose by way of contradiction that
$f(u)\ne f(v)$, that $f(v) \nsim u,f(u)$ and that
$f^2(u)\ne f^2(v)$.
By Lemma \ref{BFS}, since $f^2(v)\ne f^2(u)$ and $u\prec v$,
necessarily $d(v,f^2(u))=3.$ If $d(f^2(u),f(v))=2$, then since
$u,f(v)\in B_2(f^2(u))$ and $v\notin B_2(f^2(u))$, the convexity of
$B_2(f^2(u))$ implies that $f(v)\sim u$, which contradicts our
assumption. Thus, further we will suppose that
$d(f^2(u),f(v))=d(f^2(u),v)=3$. Since $f(u),f(v)\in B_{k-1}(b),$
$B_{k-1}(b)$ is convex, and $f(u)\nsim f(v)$, necessarily
$d(f(u),f(v))=2$. Let $w$ be a common neighbor of $f(u),f(v)$. If
$d(b,w)=k-2,$ then since $f^2(u),w\in B_{k-2}(b)$ and
$f(u)\notin B_{k-2}(b),$ the convexity of $B_{k-2}(b)$ implies that
$d(w,f^2(u)) \leq 1$, contrary to the assumption that
$d(f^2(u),f(v))=3$. Thus further we can assume that $d(b,w)=k-1$
and that $f^2(u)\nsim w$.
Now, we apply the induction assumption to $f(u)\sim w$ and
$w\sim f(v)$. We distinguish two cases.
\medskip\noindent
{\bf Case 1.} $f(w)\sim f(u),f^2(u)$.
\begin{proof}
Since $d(f^2(u),f(v))=3$, necessarily $f(w)\nsim f(v).$ Moreover,
as $f(v)\nsim u$, we can assume that $d(f(w),v)=2$, otherwise we
get a contradiction with the convexity of $B_2(f(w))$. First
suppose that $f^2(v)\sim w,f(w).$ Then $f(w)\prec f^2(v)$, so
$w\prec f(v)$. If $d(f(w),v)=2,$ then $v\in F(f(w))$ and, Lemma
\ref{BFS} yields $f^2(v)=f(w)$, a contradiction. Thus
$f^2(v)\nsim w,f(w)$ and by induction hypothesis
$f^3(v)=f^2(w)$. Since $f^2(v),f(w)\in B_2(v)$ and
$f^2(w)=f^3(v)\notin B_2(v),$ the convexity of $B_2(v)$ implies
that $f^2(v)\sim f(w),$ a contradiction.
\end{proof}
\medskip\noindent
{\bf Case 2.} $x:=f^2(w)=f^3(u)$ and $f(w)\nsim f^2(u)$.
\begin{proof}
Observe that $d(x,u) = 3$ and that $d(x,f(v)) \leq 3$. Since
$u \nsim f(v)$, the convexity of $B_3(x)$ implies that
$d(x,v) \leq 3$. Lemma \ref{BFS} implies that $x=f^3(v)$ and
consequently, $d(x,f(v) = 2$.
Note also that $d(f(v),f^2(u)) \geq 3$. Indeed, if
$d(f(v),f^2(u)) \leq 2$, then since $u,f(v) \in B_2(f^2(u))$ and
$u \nsim f(v)$, the convexity of $B_2(f^2(u))$ implies that
$v \in B_2(f^2(u))$, and by Lemma \ref{BFS}, we get that
$f^2(v) = f^2(u)$.
Therefore, we have $x,f(u) \in B_2(f(v))$ and
$f^2(u) \notin B_2(f(v))$ and thus $x \sim f(u)$, a
contradiction.
\end{proof}
This concludes the proof of the lemma.
\end{proof}
Now, we are ready to complete the proof of the theorem. By induction
on the labels of vertices of $G$, we will prove that the BFS ordering
of $G$ is a domination order of $G^2$. Let $u\prec v$ and
$d(u,v)\le 2$. We have to prove that $f(v)$ is adjacent to $u$ in
$G^2,$ i.e., that $d(f(v),u)\le 2$ in $G$. This is obviously true if
$u$ is adjacent to $v$ in $G$. Thus, further let $d(v,u)=2$ and let
$w$ be a common neighbor of $u$ and $v$. Since $u\prec v,$
$d(b,u)\le d(b,v)=k$. If $w\in B_{k+1}(b)$, then the convexity of
$B_k(b)$ implies that $u \sim v$, a contradiction. If
$w\in B_{k-1}(b)$, then since $v\sim f(v),w$ and $v\notin B_{k-1}(b),$
the convexity of $B_{k-1}(b)$ implies that $f(v)=w$ or $f(v)\sim
w$. Then obviously $d(f(v),u)\le 2$ and we are done. So, further let
$d(b,w)=k$. If $u\in B_{k-1}(b),$ again the convexity of $B_{k-1}(b)$
implies that $d(f(v),u)\le 2.$ So, we can suppose that $d(b,u)=k$. If
$f(v)=f(w),$ then again $d(f(v),u)\le 2.$ So, $f(v)\ne f(w)$ and
$f(v)\nsim w$.
We distinguish the following cases:
\medskip\noindent
{\bf Case 1:} $f(v)\sim f(w).$
\begin{proof}
Since $f(v)\nsim w,$ we conclude that $f(w)\sim v$ and thus $v\prec w$, $f(v)\prec f(w)$. Since $u\prec v\prec w,$ we obtain that $f(w)\ne f(u)$ and
$f(u)\prec f(w)$. If $f(u)\sim f(w),$ then $f(w)\sim u$ and thus $d(f(v),u)\le 2.$ So, assume that $f(u)\nsim f(w)$. Them Lemma \ref{2-fellow-traveller}
implies that $f^2(u)=f^2(w)=:x.$ Since $u\prec v\prec w$ and $f^2(u)=f^2(w),$ by BFS $f^2(v)=x$. Now, if $d(f(v),u)=3,$ then $v\sim x=f^2(v)$ and $w\sim f(u)$
by the convexity of the balls $B_2(f(v))$ and $B_2(u)$. Since this is impossible, $d(f(v),u)\le 2$.
\end{proof}
\medskip\noindent
{\bf Case 2:} $f(v)\nsim f(w).$
\begin{proof}
By Lemma \ref{2-fellow-traveller}, $f^2(v)=f^2(w)=:x$. First, let $f(w)\sim u$. If $d(f(v),u)=3,$ since $v,x\in B_2(u)$ and $f(v)\notin B_2(u),$
by convexity of $B_2(u)$ we conclude that $v\sim x=f^2(v),$ which is impossible. So $f(w)\nsim u$. In particular we have $f(w)\ne f(u).$ First suppose that $f(w)\sim f(u).$
Then $f(u)\sim w$ and thus $w\prec u$. Hence $w\prec u\prec v$. Since $f^2(v)=f^2(w),$ BFS implies that $f^2(u)=x$, i.e., $x\sim f(u)$. Again, if $d(f(v),u)=3,$ since
$v,x\in B_2(u)$ and $f(v)\notin B_2(u),$ we conclude that $v\sim x=f^2(v),$ a contradiction.
Finally, let $f(w)\nsim f(u)$. By Lemma \ref{2-fellow-traveller}, $f^2(u)=f^2(w)=x$. Again, if $d(f(v),u)=3,$ then the fact that $v\nsim x$ and $w\nsim f(u)$
leads to a contradiction with the convexity of the balls $B_2(f(v))$ and $B_2(u)$.
\end{proof}
This concludes the proof of the theorem.
\end{proof}
\begin{remark} The graphs $G_2, G_3, G_4, \ldots$ from Figure~\ref{fig: G2 nonwm} show that the powers of CB-graphs are not in general weakly modular graphs, and thus are not weakly bridged. Namely, for any $k$, the $k$th power $G_k^k$ of $G_k$ does not satisfy the quadrangle conditions for the vertices $u,v,w,y$ as indicated in the figure.
\end{remark}
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.65]
\begin{scope}[xshift=-6cm]
\def0.866{0.866}
\def4{2}
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\foreach \y in {1,..., 4}{
\foreach \x in {1,...,\y}{
\pgfmathsetmacro{\cx}{\x-(\y +1)*0.5}
\pgfmathsetmacro{\cy}{(\y-1)*0.866}
\ifnum \y=1
\node [fill=red, label=below: $u$] (x-\x-\y) at (\cx,\cy) {};
\else
\node (x-\x-\y) at (\cx,\cy) {};
\fi
\ifnum \x>1
\pgfmathtruncatemacro{\px}{\x -1}
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (x-\x-\y) -- (x-\px-\y);
\draw[thick] (x-\x-\y) -- (x-\px-0.866);
\fi
\ifnum \x<\y
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (x-\x-\y) -- (x-\x-0.866);
\fi
}
}
\pgfmathtruncatemacro{\nk}{4+1}
\foreach \y in {0,...,\nk}{
\pgfmathtruncatemacro{\hy}{(4 +1)-\y}
\foreach \x in {0,...,\hy}{
\pgfmathsetmacro{\cy}{(\y+4-1)*0.866}
\pgfmathsetmacro{\cx}{\x-(\hy)*0.5}
\ifnum \y=0
\ifnum \x = 0
\node [fill=red, label=left: $x$] (xx-\x-\y) at (\cx,\cy)
{};
\else
\ifnum \x = \hy
\node [fill=red, label=right: $y$] (xx-\x-\y) at (\cx,\cy)
{};
\else
\node (xx-\x-\y) at (\cx,\cy) {};
\fi
\fi
\else
\node (xx-\x-\y) at (\cx,\cy) {};
\fi
\ifnum \x>0
\pgfmathtruncatemacro{\px}{\x -1}
\draw[thick] (xx-\x-\y) -- (xx-\px-\y);
\fi
\ifnum \y>0
\pgfmathtruncatemacro{\nx}{\x +1}
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (xx-\x-\y) -- (xx-\x-0.866);
\draw[thick] (xx-\x-\y) -- (xx-\nx-0.866);
\fi
}
}
\pgfmathtruncatemacro{\pk}{4-1}
\foreach \y in {0,...,\pk}{
\pgfmathsetmacro{\cy}{(\y+2*4)*0.866}
\ifnum \y=\pk
\node [fill=red, label=above: $v$] (v-\y) at (0,\cy) {};
\else
\node (v-\y) at (0,\cy) {};
\fi
\ifnum \y>0
\pgfmathtruncatemacro{0.866}{\y-1}
\draw[thick] (v-0.866) -- (v-\y);
\fi
}
\node (lab) at (0, -1.5)[draw=white, fill=white, opacity =0, text opacity=1] {$G_2$};
\end{scope}
\begin{scope}
\def0.866{0.866}
\def4{3}
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\foreach \y in {1,..., 4}{
\foreach \x in {1,...,\y}{
\pgfmathsetmacro{\cx}{\x-(\y +1)*0.5}
\pgfmathsetmacro{\cy}{(\y-1)*0.866}
\ifnum \y=1
\node [fill=red, label=below: $u$] (x-\x-\y) at (\cx,\cy) {};
\else
\node (x-\x-\y) at (\cx,\cy) {};
\fi
\ifnum \x>1
\pgfmathtruncatemacro{\px}{\x -1}
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (x-\x-\y) -- (x-\px-\y);
\draw[thick] (x-\x-\y) -- (x-\px-0.866);
\fi
\ifnum \x<\y
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (x-\x-\y) -- (x-\x-0.866);
\fi
}
}
\pgfmathtruncatemacro{\nk}{4+1}
\foreach \y in {0,...,\nk}{
\pgfmathtruncatemacro{\hy}{(4 +1)-\y}
\foreach \x in {0,...,\hy}{
\pgfmathsetmacro{\cy}{(\y+4-1)*0.866}
\pgfmathsetmacro{\cx}{\x-(\hy)*0.5}
\ifnum \y=0
\ifnum \x = 0
\node [fill=red, label=left: $x$] (xx-\x-\y) at (\cx,\cy)
{};
\else
\ifnum \x = \hy
\node [fill=red, label=right: $y$] (xx-\x-\y) at (\cx,\cy)
{};
\else
\node (xx-\x-\y) at (\cx,\cy) {};
\fi
\fi
\else
\node (xx-\x-\y) at (\cx,\cy) {};
\fi
\ifnum \x>0
\pgfmathtruncatemacro{\px}{\x -1}
\draw[thick] (xx-\x-\y) -- (xx-\px-\y);
\fi
\ifnum \y>0
\pgfmathtruncatemacro{\nx}{\x +1}
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (xx-\x-\y) -- (xx-\x-0.866);
\draw[thick] (xx-\x-\y) -- (xx-\nx-0.866);
\fi
}
}
\pgfmathtruncatemacro{\pk}{4-1}
\foreach \y in {0,...,\pk}{
\pgfmathsetmacro{\cy}{(\y+2*4)*0.866}
\ifnum \y=\pk
\node [fill=red, label=above: $v$] (v-\y) at (0,\cy) {};
\else
\node (v-\y) at (0,\cy) {};
\fi
\ifnum \y>0
\pgfmathtruncatemacro{0.866}{\y-1}
\draw[thick] (v-0.866) -- (v-\y);
\fi
}
\node (lab) at (0, -1.5)[draw=white, fill=white, opacity =0, text opacity=1] {$G_3$};
\end{scope}
\begin{scope}[xshift=7cm]
\def0.866{0.866}
\def4{4}
\tikzstyle{every node}=[draw, circle,fill=black,minimum size=4pt,
inner sep=0pt]
\foreach \y in {1,..., 4}{
\foreach \x in {1,...,\y}{
\pgfmathsetmacro{\cx}{\x-(\y +1)*0.5}
\pgfmathsetmacro{\cy}{(\y-1)*0.866}
\ifnum \y=1
\node [fill=red, label=below: $u$] (x-\x-\y) at (\cx,\cy) {};
\else
\node (x-\x-\y) at (\cx,\cy) {};
\fi
\ifnum \x>1
\pgfmathtruncatemacro{\px}{\x -1}
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (x-\x-\y) -- (x-\px-\y);
\draw[thick] (x-\x-\y) -- (x-\px-0.866);
\fi
\ifnum \x<\y
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (x-\x-\y) -- (x-\x-0.866);
\fi
}
}
\pgfmathtruncatemacro{\nk}{4+1}
\foreach \y in {0,...,\nk}{
\pgfmathtruncatemacro{\hy}{(4 +1)-\y}
\foreach \x in {0,...,\hy}{
\pgfmathsetmacro{\cy}{(\y+4-1)*0.866}
\pgfmathsetmacro{\cx}{\x-(\hy)*0.5}
\ifnum \y=0
\ifnum \x = 0
\node [fill=red, label=left: $x$] (xx-\x-\y) at (\cx,\cy)
{};
\else
\ifnum \x = \hy
\node [fill=red, label=right: $y$] (xx-\x-\y) at (\cx,\cy)
{};
\else
\node (xx-\x-\y) at (\cx,\cy) {};
\fi
\fi
\else
\node (xx-\x-\y) at (\cx,\cy) {};
\fi
\ifnum \x>0
\pgfmathtruncatemacro{\px}{\x -1}
\draw[thick] (xx-\x-\y) -- (xx-\px-\y);
\fi
\ifnum \y>0
\pgfmathtruncatemacro{\nx}{\x +1}
\pgfmathtruncatemacro{0.866}{\y -1}
\draw[thick] (xx-\x-\y) -- (xx-\x-0.866);
\draw[thick] (xx-\x-\y) -- (xx-\nx-0.866);
\fi
}
}
\pgfmathtruncatemacro{\pk}{4-1}
\foreach \y in {0,...,\pk}{
\pgfmathsetmacro{\cy}{(\y+2*4)*0.866}
\ifnum \y=\pk
\node [fill=red, label=above: $v$] (v-\y) at (0,\cy) {};
\else
\node (v-\y) at (0,\cy) {};
\fi
\ifnum \y>0
\pgfmathtruncatemacro{0.866}{\y-1}
\draw[thick] (v-0.866) -- (v-\y);
\fi
}
\node (lab) at (0, -1.5)[draw=white, fill=white, opacity =0, text opacity=1] {$G_4$};
\end{scope}
\end{tikzpicture}
\caption{The graphs $G_2, G_3, G_4$ are CB-graphs whose powers are
not weakly modular. For any $k \geq 2$, in $G_k^k$, $u$ is
adjacent to $x, y$ and at distance $3$ from $v$, and the
vertices $v, x$, and $y$ are pairwise at distance $2$ but they
do not have a common neighbor.}\label{fig: G2 nonwm}
\end{figure}
\begin{remark}
General CB-graphs are not
dismantlable (even if the main result of this section shows that their squares are dismantlable). A finite graph $G$ is \emph{dismantlable to a subgraph} $H$, if $G$ and $H$ satisfy the following conditions:
(1) there exists a total ordering $v_k,\ldots,v_1$ of the vertices of $V(G)\setminus V(H)$ such that each vertex $v_i$ is dominated in the subgraph $G_i$ induced by $\{ v_i,v_{i-1},\ldots,v_1\}\cup V(H)$
either by a vertex $v_j$ with $j<i$ or by a vertex of $H$ and (2) $H$ does not contain any dominated vertex. It is well-known that all subgraphs $H$ of a graph $G$ to which $G$ is dismantlable are isomorphic \cite{Hell_Nesetril}
and any such subgraph is called the \emph{core} of $G$. One can easily show that if $G$ is a CB-graph, then all intermediate subgraphs $G_i$ are also CB-graphs, thus the core $H$ of $G$ is a CB-graph.
Since weakly systolic graphs are dismantlable \cite{ChepoiOsajda}, their cores are trivial (a single vertex).
On the other hand, the core of any triangle-free CB-graph $G$ in which each 2-connected component contains a cycle is the graph $G$ itself. By Proposition \ref{thm: Moore}, each 2-connected component of such a graph
is a Moore graph and thus has diameter 2. One can ask if this property extends to the cores of all finite CB-graphs:
\end{remark}
\begin{question} Is it true that the core of any finite CB-graph is a CB-graph in which all 2-connected components have diameter 1 or 2?
\end{question}
\subsection{Rips complexes and stabilized sets}
We continue with several consequences of Theorem \ref{dismantl}.
\begin{corollary} \label{Rips} For any $k\ge 2$, the Rips complex $X_{k}(G)$ of a locally-finite CB-graph is contractible. The maximal
simplices of $X_{k}(G)$ define convex sets of diameter $k$ of $G$.
\end{corollary}
\begin{proof} Note that $X_{k}(G)=X(G^k)$. Since $G^2$ is dismantlable, $G^k$ is also dismantlable for every $k \geq 2$. Indeed, pick any two vertices $u,v\in V$ such that $v$ dominates $u$ in $G^2$. Pick any vertex $x$ such that $d(u,x)\le k$ and let $y$ be a vertex at distance 2 from $u$ on a shortest $(u,x)$-path of $G$. Since $v$ dominates $u$ in $G^2$, $d(v,y)\le 2$, thus by triangle inequality we obtain that $d(v,x)\le k$. Consequently, $v$ dominates $u$ also in $G^k$. Since $G$ is locally-finite, all cliques of $G^k$ are finite. Thus $X_{k}(G)$ is contractible, as the flag complex of a dismantlable graph without infinite cliques is contractible.
Now, let $A$ be a maximal by inclusion simplex of $X_{k}(G)$. Then $\mathrm{diam}(A)\le k$. Since in CB-graphs the equality $\mathrm{diam}(\mathrm{conv}(A))=\mathrm{diam}(A)$ holds, the maximality of $A$ implies that $\mathrm{conv}(A)=A$.
\end{proof}
\begin{remark} The second assertion of Corollary \ref{Rips} implies that the Rips complex $X_k(G)$ of a CB-graph can be viewed as the thickening of
$G$ with respect to all convex sets of diameter at most $k$. Other similar thickening operations have been used in the context of Helly graphs in \cite{ChChGeHiOs}.
\end{remark}
It is known (see for example \cite{Hell_Nesetril}) that for every dismantlable graph $G$ and every graph homomorphism
$f:G\to G$, there exists a (nonempty) clique $C$ in $G$ stabilized by $f$, i.e. such that $f(C)=C$. Hence we immediately get the following corollary:
\begin{corollary}
\label{cor: fixpoint}
Let $G$ be a graph with convex balls and $f: G \to G$ be a graph homomorphism. Then there exists a convex set of diameter at most $2$ in $G$ which is stabilized by $f$.
\end{corollary}
Corollary \ref{cor: fixpoint} is tight in the sense that one cannot
hope to find a stabilized clique for any homomorphism. Indeed a
simple counter-example is the $5$-cycle, which is not dismantlable and
which have the cyclic permutation of order $5$ in its automorphism
group stabilizing no clique. The following example shows that the
5-cycle is not the only obstruction to get such a property.
\begin{example}\label{ex-circulant}
We describe now a graph $H$ with convex balls such that there exists
an automorphism $f:H\to H$ that does not stabilize a clique or a
5-cycle. See Figure \ref{fig: fixpointctre-ex} for an
illustration. Let
$V(H):=\sg{a_1, a_2, a_3, b_1, b_2, b_3, c_1, c_2, c_3}$ and define
the edge set of $H$ as follows\footnote{One may observe that $H$ is
isomorphic to the circulant graph
$\mathrm{Cay}(\mathbb{Z}_9, \sg{1,2})$.}:
\begin{gather*}
E(H):=\sg{a_ib_i, i \in \sg{1,2,3}}\cup \sg{a_{i+1}b_i, i \in \sg{1,2,3}}\cup \sg{a_ic_i, i \in \sg{1,2,3}}\\
\cup \sg{a_{i+1}c_i, i \in \sg{1,2,3}}\cup\sg{b_ic_i, i \in \sg{1,2,3}}\cup\sg{b_{i+1}c_i, i \in \sg{1,2,3}},
\end{gather*}
where the operations on the indices are done modulo $3$. Observe that
$H$ has diameter $2$ and that $H$ has no induced 4-cycle. This implies
that $H$ has convex balls. Now consider the automorphism $f$ defined
by $f(a_i):=a_{i+1}$, $f(b_i):= b_{i+1}$ and $f(c_i):= c_{i+1}$ for
every $i\in \sg{1,2,3}$ (again, every addition is done modulo $3$).
Note that $f$ cannot stabilize any clique as every vertex is nonadjacent to
its image. Moreover every 5-cycle of $H$ must have at least one of the
$a_i$'s, one of the $b_i$'s and one of the $c_i$'s in its vertex
set. By this observation, it follows that $f$ cannot stabilize a 5-cycle of
$H$.
\end{example}
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=2]
\tikzstyle{every node}=[draw,circle,fill=green,minimum size=4pt,
inner sep=0pt]
\node (1) at (0,1) [label=above: $a_1$] {};
\node (2) at (0.87,-0.5) [label=right: $a_2$] {};
\node (3) at (-0.87,-0.5) [label=left: $a_3$] {};
\tikzstyle{every node}=[draw,circle,fill=blue,minimum size=4pt,
inner sep=0pt]
\node (4) at (0.87,0.5) [label=right: $b_1$] {};
\node (5) at (0,-1) [label=below: $b_2$] {};
\node (6) at (-0.87,0.5) [label=left: $b_3$] {};
\tikzstyle{every node}=[draw,circle,fill=red,minimum size=4pt,
inner sep=0pt]
\node (7) [label=left: $c_1$] at (0.66,0) {};
\node (8) at (-0.33,-0.58) [label=above: $c_2$] {};
\node (9) at (-0.33,0.58) [label=below: $c_3$] {};
\draw (1) -- (4) -- (2) -- (5) -- (3) -- (6) -- (1) -- (7) -- (5) -- (8) -- (6) -- (9) -- (4) -- (7) -- (2) -- (8) -- (3) -- (9) -- (1);
\end{tikzpicture}
\caption{The graph $H$ of Example~\ref{ex-circulant}
} \label{fig:
fixpointctre-ex}
\end{figure}
\section{Biautomaticity}\label{sec: NCP}
In this section, we prove that CB-groups (i.e., groups acting geometrically on CB-graphs) are biautomatic.
For this, we construct in a canonical way clique-paths between all pairs of vertices of a graph with $2$-convex balls. In CB-graphs, we characterize those normal clique-paths locally. Then we prove that in CB-graphs, the normal clique-paths satisfy the 2-sided fellow traveller property. Using this property and the local definition of normal clique-paths, in the last subsection we prove biautomaticity of CB-groups.
\subsection{Clique-paths: definition and existence}
Following \cite{ChChGeHiOs}, for a set $S$ of vertices of a graph $G = (V,E)$ and an integer $k \ge 0$, let $B^*_k(S)=\bigcap_{s\in S} B_k(s)$.
If $S$ is a clique, then $B^*_1(S)$ is the union of $S$ and the set of vertices adjacent to all vertices in $S$. Notice also that if
$S \subseteq S'$, then $B^*_k(S')\subseteq B^*_k(S)$.
Let $u,v$ be two arbitrary vertices of a graph $G$ and let $d(u,v)=k$. For each $i$ running from $k$ to 0 we inductively define the sets $C_{(u,v)}^{(i)}\subseteq S_i(u)\cap I(u,v)$ by setting $C_{(u,v)}^{(k)}:=\sg{v}$ and
$$C_{(u,v)}^{(i-1)}:=B^*_1\left(C_{(u,v)}^{(i)}\right)\cap B_{i-1}(u)$$
for any $i<k$. We also set $C_{(u,v)}^{(i)}:= \sg{v}$ for any $i \geq k+1$. We denote this
clique-path by $\gamma(u,v):= (\sg{u} = C_{(u,v)}^{(0)}, C_{(u,v)}^{(1)}, \ldots, C_{(u,v)}^{(k)}=\sg{v})$.
Note that the resulting clique-paths are directed, since the cliques of $\gamma_{(u,v)}$ and $\gamma_{(v,u)}$
are not the same in general.
\begin{lemma}\label{lem: INCclique}
Let $G$ be a graph with $2$-convex balls and let $u,v\in V$ with $d(u,v)=k$. For any $i\in \sg{0,\ldots,k}$, the set $C_{(u,v)}^{(i)}$ is nonempty and defines a clique of $G$. Furthermore, $C_{(u,v)}^{(i)}\cup C_{(u,v)}^{(i+1)}$ is a clique for any $i=0,\ldots,k-1$.
\end{lemma}
\begin{proof}
We proceed by induction on decreasing values of $i\in \sg{0, \ldots, k}$. The cases $i=k,k-1$ are immediate since
$G$ is a graph with 2-convex balls, so by Theorem \ref{thm: INC} $\mathrm{INC}$ holds. Now, let $i < k-1$. By induction hypothesis,
$C_{(u,v)}^{(i+2)}$ and $C_{(u,v)}^{(i+1)}$ are nonempty cliques and their union is a clique.
Let $x\in C_{(u,v)}^{(i+2)}$.
Then $x$ is adjacent to all vertices of $C_{(u,v)}^{(i+1)}$ and $C_{(u,v)}^{(i+1)}\subset I(x,u)$. Applying
$\mathrm{INC}^+(x,u)$, there exists a vertex $z \in B_{i}(u)$ adjacent to every vertex of $S_1(x) \cap B_{i+1}(u)$.
In particular, $z$ is adjacent to all vertices of $C_{(u,v)}^{(i+1)}$, therefore $z$ is a vertex of
$C_{(u,v)}^{(i)}$. This shows that $C_{(u,v)}^{(i)}$ is nonempty. Since $C_{(u,v)}^{(i)}$ belongs to $S_1(y)\cap I(y,u)$
for any $y\in C_{(u,v)}^{(i+1)}$, by $\mathrm{INC}^0$ the set $C_{(u,v)}^{(i)}$ defines a clique.
From the definition of $C_{(u,v)}^{(i)}$, it immediately follows that $C_{(u,v)}^{(i)}\cup C_{(u,v)}^{(i+1)}$ is a clique as well.
\end{proof}
We denote the sequence of cliques $(\sg{u}=C_{(u,v)}^{(0)},C_{(u,v)}^{(1)},\ldots,C_{(u,v)}^{(k-1)},C_{(u,v)}^{(k)}=\sg{v})$
by $\gamma_{(u,v)}$ and call it a \emph{clique-path with source $u$ and sink $v$}.
For any graph $G$, any vertex $u$ of $G$, and any subset $K\subseteq S_{k}(u)$, one can define the sets $C_{(u,K)}^{(i)}$ in a similar way, by setting
$C_{(u,K)}^{(k)}:=K$ and
$C_{(u,K)}^{(i)}:=B_1^*\left(C_{(u,K)}^{(i)}\right)\cap B_{i}(u),$
for any $i\in \sg{0,\ldots, k-1}$. We also let $C_{(u,K)}^{(i)}:=K$ for any $i \geq k+1$. Again, we let $\gamma_{(u,C)}:=(\sg{u} = C_{(u,C)}^{(0)}, C_{(u,C)}^{(1)},\ldots, C_{(u,C)}^{(k-1)},C_{(u,C)}^{(k)}=C)$ denote such a clique-path.
However, Lemma \ref{lem: INCclique} does not hold anymore and the resulting sets may be empty. Nevertheless, the following simple observation
holds:
\begin{lemma}
\label{lem: inclalt}
Let $G$ be any graph, $u\in V$ and $k\geq 1$ be an integer. Then for any $L \subseteq K\subseteq S_k(u)$, the inclusion $C_{(u,K)}^{(k-1)} \subseteq C_{(u,L)}^{(k-1)}$ holds.
\end{lemma}
\begin{proof}
Pick $x\in C_{(u,K)}^{(k-1)}$. Then $x\in S_{k-1}(u)$ and clearly $x$ is adjacent to every vertex of $K$, so to every vertex of $L$ as well, which implies that $x \in C_{(u,L)}^{(k-1)}$.
\end{proof}
\tikzexternaldisable
\begin{figure}[h!]
\centering
\begin{tikzpicture}[scale=1]
\tikzstyle{every node}=[draw,circle,fill=red,minimum size=4pt,
inner sep=0pt]
\node (u) at (0,0) [label=left: $u$] {};
\node (11) at (-1,1) [] {};
\node (12) at (0,1) [] {};
\node (13) at (1,1) [] {};
\node (21) at (-1,2) [color = black] {};
\node (22) at (0,2) [] {};
\node (23) at (1,2) [color= black] {};
\node (31) at (-0.5,3) [] {};
\node (32) at (0.5,3) [] {};
\node (v) at (0,4) [label=left: $v$] [] {};
\draw[thick] (u) -- (11) -- (12) -- (13) -- (u) -- (12) -- (22) -- (21) -- (11);
\draw[thick] (11) to[bend right] (13);
\draw[thick] (11) -- (22) -- (13) -- (23) -- (22) -- (31) -- (v) -- (32) -- (23);
\draw[thick] (21) -- (31) -- (32) -- (22);
\end{tikzpicture}
\caption{In red, the vertices of the cliques $C_{(u,v)}^{(i)}$ of the clique-path $\gamma_{(u,v)}$ in a CB-graph.}
\end{figure}
\begin{comment}
\subsection{Normal clique paths and normal paths}
Let $G$ be a graph with convex balls, $u\in V$ some vertex and $C \subseteq V$ a clique of $G$. A sequence of cliques $(C_0, C_1, \ldots, C_k)$ is a \emph{geodesic clique-path} if bla. ( \textbf{\textcolor{red}{TODO}} mettre définition Jérémie ici).
A path $u = u_0, \ldots, u_k = v$ from $u$ to $v$ is called a \emph{normal path} if
\end{comment}
\subsection{Local characterization}
In this subsection, we provide a local characterization of the clique-paths $\gamma_{(u,v)}, (u,v)\in V\times V$ defined above. This will show that they
are canonically defined and are unique.
Let $u$ be a vertex and $C$ be a nonempty clique of $G$. A sequence of nonempty cliques $\eta_{(u,C)}=(\sg{u} = C_0, C_1, \ldots, C_k = C)$ is
called a \emph{normal clique-path} from $u$ to $C$ if the following local conditions hold:
\begin{itemize}
\item ($i$) for any $i=0,\ldots,k-1$, $C_i \cap C_{i+1} = \varnothing$ and $C_i \cup C_{i+1}$ form a clique;
\item ($ii$) for any $i=1,\ldots, k-1$, $C_{i-1}\cap C_{i+1}= \varnothing$ and there is no edge between $C_{i+1}$ and $C_{i-1}$;
\item ($iii$) for any $i=3, \ldots, k$, $C_i$ is at uniform distance $3$ from $C_{i-3}$.
\item ($iv$) for any $i=1,\ldots,k-1$, $C_i=B_1^*(C_{i+1})\cap B_1(C_{i-1}).$
\end{itemize}
In most of cases, condition ($ii$) is implied by condition ($iii$). However we keep ($ii$) as it allows to deal
with some special cases and it makes clearer that things ``behave well'' locally. Our next goal is to prove that the
clique-paths $\gamma_{(u,v)}$ are the unique normal clique-paths. To do induction, we prove this result in a more general
setting of clique-paths from vertices $u$ to the cliques $C$ at uniform distance from $u$.
\begin{lemma}
\label{lem: cliquepathsarenormal}
Let $G$ be a graph with $2$-convex balls, $u$ be a vertex, and $C$ be a nonempty clique of $G$ at uniform distance $k$ from $u$.
Additionally assume that all sets $C_{(u,C)}^{(i)}$ are nonempty. Then the sequence $\gamma_{(u,C)}=(\sg{u} = C_{(u,C)}^{(0)}, C_{(u,C)}^{(1)},\ldots, C_{(u,C)}^{(k-1)},C_{(u,C)}^{(k)}=C)$ is a normal clique-path. Consequently, for any pair of vertices $u,v$ of $G$, the clique-path
$\gamma_{(u,v)}$ is a normal clique-path.
\end{lemma}
\begin{proof}
To simplify the notation, we set $C_i := C_{(u,C)}^{(i)}$ for any $i \in \sg{0, \ldots, k}$. Since $G$ is a graph with 2-convex balls, each $C_i$ is a clique, thus
$\gamma_{(u,C)}=(\sg{u} = C_0,C_1,\ldots, C_{k-1},C_{k}=C)$ is a clique-path. From the definition of $\gamma_{(u,C)}$ it immediately follows that $\gamma_{(u,C)}$
satisfies ($i$), ($ii$) and ($iii$). It remains to establish ($iv$), i.e. that for any $i \in \sg{1, \ldots, k-1}$ we have
$C_i :=B_1^*(C_{i+1}) \cap B_i(u) =B_1^*(C_{i+1}) \cap B_1(C_{i-1})$. We proceed by induction on $k=d(u,C)$. This is direct if $k=2$. Hence assume that $k \geq 3$ and
that the required equality holds for every $i \in \sg{1, \ldots, k-2}$. Since $C_i = C_{(u,C_{k-1})}^{(i)}$ for every $i \in \sg{0, \ldots, k-1}$, it is enough
to prove the equality for $i = k-1$, i.e., that $B_1^*(C_{k}) \cap B_{k-1}(u) =B_1^*(C_{k}) \cap B_1(C_{k-2})$.
By definition, $C_{k-1} \subseteq B_1(C_{k-2})$, so the inclusion $B_1^*(C_{k}) \cap B_{k-1}(u) \subseteq B_1^*(C_{k}) \cap B_1(C_{k-2})$ is immediate.
Now we show the converse inclusion $B_1(C_{k-2})\subseteq B_{k-1}(u)$. This is also immediate because all vertices of $C_{k-2}$ have distance $k-2$ to $u$, thus
all vertices of $B_1(C_{k-2})$ have distance at most $k-1$ to $u$.
\end{proof}
Now we prove the converse direction for CB-graphs.
\begin{lemma}
\label{lem: normalcliquepathunicity} Let $G$ be a CB-graph, $u$ be a vertex, and $C$ be a nonempty clique of $G$ at uniform distance $k$ from $u$.
Let $\eta_{(u,C)}=(\sg{u} = C_0, C_1, \ldots, C_{\ell}= C)$ be a normal clique-path from $u$ to $C$. Then $\eta_{(u,C)}$ coincides with the clique-path $\gamma_{(u,C)}$ defined in Lemma \ref{lem: cliquepathsarenormal}. Consequently, for any pair of vertices $u,v$ of $G$ there exists a unique normal clique-path from $u$ to $v$ and this path is
$\gamma_{(u,v)}$.
\end{lemma}
\begin{proof}
We proceed by induction on the length $\ell$ of the normal clique-path
$\eta_{(u,C)}$. If $\ell\le 2$ there is nothing to prove. Thus, we assume that $\ell\ge 3$ and that the assertion of the lemma holds for any normal clique-path of length smaller than $\ell$. To prove the assertion for $\eta_{(u,C)}$, we will show that for any $i \in \sg{0,\ldots,\ell-1}$, (a) $C_i$ is at uniform distance $i$ from $u$ and (b) the equality $C_i = B^*_1(C_{i+1})\cap B_{i}(u)$ holds. If for some $j<\ell$, the clique $C_j$ is at uniform distance $j$ from $u$, then the subpath $\eta_{(u,C_j)}:=(\sg{u} = C_0, C_1, \ldots, C_{j})$ of $\eta_{(u,C)}$ is a normal clique-path from $u$ to $C_j$. Therefore, by induction assumption $\eta_{(u,C_j)}$ coincides with the clique-path $\gamma_{(u,C_j)}$, consequently, $\eta_{(u,C_j)}$ satisfies the properties (a) and (b).
By condition ($i$) of normal clique-paths, the union of two consecutive cliques of $\eta_{(u,C)}$ is a clique, therefore for any $i\le \ell-1$, all vertices of $C_{i}$ have distance at most $i$ from $u$. Now we prove that each $C_i$ is at uniform distance $i$ from the origin $u$, thus establishing (a). Suppose this is not true
and let $i$ be the smallest index such that $C_i$ is not at uniform distance $i$ from $u$.
By condition ($iii$) of normal clique-paths, we must have $i > 3$.
The minimality choice of $i$
implies that all cliques $C_j$ with $j<i\le \ell$ are at uniform distance $j$ from $u$. Consequently, for all such $C_j$, $\eta_{(u,C_j)}$ coincides with $\gamma_{(u,C_j)}$. From the choice of $i$ it also follows that there exists a vertex $x\in C_i$ with $d(u,x)<i$.
If $d(u,x)=i-1$, then we can apply $\mathrm{TPC}^+(u,xy)$, where $y\in C_{i-1}$. If $\mathrm{TC}$ applies, then there exists a vertex $z\sim x,y$ at distance $i-2$ from $u$. By $\mathrm{INC}^+(u)$, $z$ must be adjacent to every vertex of $C_{i-2}$ and there exists a vertex $z'$ at distance $i-3$ from $u$ which is adjacent to all vertices of the clique $C_{i-2}\cup\sg{z}$.
Since $\eta_{(u,C_{i-1})}=\gamma_{(u,C_{i-1})}$, we have $C_{i-3}=B^*_1(C_{i-2})\cap B_{i-3}(u)$, yielding $z' \in C_{i-3}$. This implies that $x$ is at distance at most $2$ from $z' \in C_{i-3}$, contradicting condition ($iii$). Otherwise, if $\mathrm{PC}^+$ applies, then
there exists a vertex $z$ at distance $i-3$ from $u$ and at distance 2 from $x,y$ and which is adjacent to every $w \in C_{i-2}$.
Again, since $C_{i-3}=B^*_1(C_{i-2})\cap B_{i-3}(u)$, we conclude that $z \in C_{i-3}$. This implies that $x$ is at distance $2$ from a vertex of $C_{i-3}$,
contradicting again ($iii$).
Now suppose that $d(u,x)\le i-2$ and that $C_i$ does not contain a vertex at distance $i-1$ from $u$. Since all vertices of $C_{i-1}$ are at distance $i-1$ from $u$ and $C_i\cup C_{i-1}$ is a clique, this implies that $C_i$ has uniform distance $i-2$ from $u$. Since $C_{i-2}$ is at uniform distance $i-2$ from $u$, all vertices of $C_i\cup C_{i-2}$ are adjacent to any vertex $w\in C_{i-1}$ and $C_i\cup C_{i-2}\subset I(w,u)$. Then $\mathrm{INC}^+(u)$ immediately implies that every vertex of $C_i$ is adjacent to every vertex of $C_{i-2}$, which contradicts condition ($ii$). This establishes that each clique $C_i$ of $\eta_{(u,C)}$ is at uniform distance $i$ from $u$. By induction hypothesis, $\eta_{(u,C_j)}=\gamma_{(u,C_j)}$ for any $j<\ell$.
Therefore it remains to prove that $C_{\ell-1}=B^*_1(C_{\ell}) \cap B_{\ell-1}(u)$. We use the following claim.
\begin{claim}
\label{clm: lemk-1}
Let $k\geq 2$, and $C,C',C''$ be three cliques at uniform distance respectively $k, k-1,$ and $k-2$ from $u$. Assume that $C'' = B^*_1(C')\cap B_{k-2}(u)$ and $C' = B^*_1(C)\cap B_1(C'')$. Then $C'=B^*_1(C)\cap B_{k-1}(u).$
\end{claim}
\begin{proof}
As $C'$ has uniform distance $k-1$ to $u$, the direct inclusion is trivial.
Since $C\neq \varnothing$, $B^*_1(C)\cap B_{k-1}(u)$ is a clique by $\mathrm{INC}^0(u)$. Hence by $\mathrm{INC}^+(u)$, there exists
$z \in B_{k-2}(u)$ such that $z$ is adjacent to any vertex of $B^*_1(C) \cap B_{k-1}(u)$. Necessarily, $z \in C''$ and we get the reverse inclusion $B^*_1(C) \cap B_{k-1}(u) \subseteq C'$.
\end{proof}
The desired equality $C_{\ell-1}=B^*_1(C_{\ell}) \cap B_{\ell-1}(u)$ follows by applying Claim \ref{clm: lemk-1} with $C:=C_{\ell}$, $C':= C_{\ell-1}$, and $C'':=C_{\ell-2}$.
\end{proof}
\subsection{Fellow traveler property}
Le $G=(V,E)$ be a CB-graph. Denote by $\Upsilon$ the set of normal clique-paths between all pairs of vertices of $G$: $\Upsilon=\{ \gamma_{(u,v)}: (u,v)\in V\times V\}$. A path $(u=u_0,u_1, \ldots, u_k = v)$ between two vertices $u,v$ of a graph $G$ is called a \emph{normal path from $u$ to $v$} if for every $i \in \sg{0, \ldots, k}$ we have $u_i \in C_{i},$ where $\gamma_{(u,v)}=(\sg{u} = C_0, C_1, \ldots, C_k = \sg{v})$ is the normal clique-path from $u$ to $v$. From the definition it follows that every normal path from $u$ to $v$ is a shortest $(u,v)$-path. Denote by $\Upsilon^+$ the set of all normal paths between all pairs of vertices of $G$.
Let ${\mathcal P}(G)$ be the set of all paths of $G$ and ${\mathcal CP}(G)$ be the set of all clique-paths of $G$. A \emph{(clique-)path system} $\mathcal P$ \cite{Swiat} is any subset of ${\mathcal P}(G)$ (respectively, of ${\mathcal CP}(G)$). A (clique-)path system $\mathcal P$ is \emph{complete} if any two vertices are endpoints of some (clique-)path in $\mathcal P$. Let $[0,k]^*$ denote the set of integer points
from the segment $[0,k].$ Given a (clique-)path $\gamma$ of length $k=|\gamma|$ in $G$, we can parametrize it and denote by $\gamma:[0,k]^*\rightarrow V(G)$. It will be convenient to extend $\gamma$ over $[0,\infty]$ by setting $\gamma(i)=\gamma(k)$ for any $i>k$. A (clique-)path system $\mathcal P$ of a graph $G$ satisfies the
\emph{2-sided fellow traveler property} if there are constants $C>0$ and $D\ge 0$ such that for any two paths $\gamma_1,\gamma_2\in \mathcal P$, the following inequality holds for all $i$:
$$d_G(\gamma_1(i),\gamma_2(i))\le C\cdot \max\{ d_G(\gamma_1(0),\gamma_2(0)),d_G(\gamma_1(\infty),\gamma_2(\infty))\}+D.$$
A \emph{bicombing} of a graph $G$ is a complete (clique-)path system $\mathcal P$ satisfying the $2$--sided fellow traveler property. If all (clique-)paths in
${\mathcal P}$ are shortest paths of $G$, then
${\mathcal P}$ is called a \emph{geodesic bicombing}.
Our goal in this subsection is to show that in CB-graphs, the systems of normal clique-paths and of normal paths enjoy the $2$-sided fellow traveler property and thus define a geodesic bicombing of $G$:
\begin{theorem}\label{thm: fellowtravdist1}
Let $G$ be a graph with convex balls. For arbitrary vertices $u,u',v,v'$ of $G$, let $\gamma_{(u,v)}=(\sg{u}=C_0,C_1,\ldots,C_k=\sg{v})$ and $\gamma_{(u',v')}=(\sg{u'}=C'_0,C'_1,\ldots,C'_{k}=\sg{v'})$ be the normal clique-paths from $u$ to $v$ and from $u'$ to $v'$, respectively.
Then for any $i \geq 0$ and for any pair $(x,y) \in C_{i}\times C'_{i}$, we have $d(x,y) \leq 7 \max(d(u,u'), d(v,v')).$
Consequently, the sets of normal clique-paths $\Upsilon$ and of normal paths $\Upsilon^+$ of $G$ enjoy the $2$-sided fellow
traveler property and thus define a geodesic bicombing of $G$.
\end{theorem}
\begin{proof}
We let $\ell:=d(u,u')$ and $m:=d(v,v')$. First we show by induction over $m\geq 0$ that if $\ell=0$, i.e $u=u'$, then for any $i \geq 0$ and $(x,y) \in C_{i}\times C_{i'}$ we have $d(x,y)\leq 3m$. Indeed, if $m=1$, this is a direct consequence of Lemma \ref{lem: cliquepath}. Assume now that $m\geq 2$ and let $(v=v_0, v_1, \ldots, v_m=v')$ be a shortest $(v,v')$-path, and set $v'':=v_{m-1}$.
By applying the induction hypothesis to $\gamma_{(u,v)}$ and $\gamma_{(u,v'')}$ and Lemma \ref{lem: cliquepath} to $\gamma_{(u,v'')}$ and $\gamma_{(u,v')}$, by triangle inequality we deduce that $d(x,y)\leq 3m$ for every $(x,y)\in C_i \times C_{i'}$ and $i \geq 0$. Similarly we can prove by Lemma \ref{lem: invcliquepath} and induction on $\ell \geq 0$ that if $m=0$, then for any $i \geq 0$ and $(x,y) \in C_{i}\times C_{i'}$ we have $d(x,y)\leq 4\ell$.
Now we prove Theorem \ref{thm: fellowtravdist1} by induction on $(\ell,m)$ together with the product order. The cases $\ell=0$ or $ m=0$ are covered by what we did. If
$\ell=1$ and $m=1$, we apply these two basis cases to the two pairs $\{ \gamma_{(u,v)},\gamma_{(u,v')}\}$,$\{ \gamma_{(u,v')},\gamma_{(v,v')}\}$ and, by
triangle inequality we deduce that $d(x,y)\le 7$. Hence we can assume that $\ell\geq 1$, $m\geq 1$, and $\ell+m \geq 3$. Let $i \geq 0$ and $(x,y) \in C_{i}\times C'_{i}$. Let $u'':=u_{\ell-1}$ and $v'':= v_{m-1}$ and let $z \in C''_i$, where $C''_i$ is the $i$-th clique of the normal clique-path $\gamma_{(u'',v'')}$. By induction hypothesis, $d(x, z) \leq 7\max\{ d(u,u''), d(v,v'')\}=7\max\{ \ell-1, m-1\}.$ On the other hand, since $u''\sim u'$ and $v''\sim v$, by the case $(\ell, m)=(1,1)$ we get $d(z,y)\le 7$. Consequently, $d(x,y)\le d(x,z)+d(z,y)\le 7\max\{ \ell-1, m-1\}+7=7\max \{ \ell, m\}$, concluding the proof of the theorem.
\end{proof}
In our next results, $G$ is a CB-graph and we follow the notations of Theorem \ref{thm: fellowtravdist1}.
We start with a structural property of normal clique-paths $\gamma_{(u,v)}=(\sg{u}=C_0,C_1,\ldots,C_{k-1},C_k=\sg{v})$ and $\gamma_{(u,v')}=(\sg{u}=C'_0,C'_1,\ldots,C'_{k-1},C'_{k}=\sg{v'})$
in the particular case when $v\sim v'$ and $d(u,v)=d(u,v')=k$.
\begin{lemma}\label{lem: clique2nvx} Let $u,v,v'$ be three vertices of $G$ such that $v\sim v'$ and $d(u,v)=d(u,v')=k$.
Then for any $1\le i\le k$ there exists $j\in \{ i-1,i\}$ for which either $C_j\cap C'_j \neq \varnothing$ or $C_{j}\cup C'_{j}$ is a clique. \end{lemma}
\begin{proof} We start with the particular case $i=k-1$.
\begin{claim} \label{claim:clique2nvx}
$C_{k-1}\cap C'_{k-1}\neq \varnothing$ or $C_{k-2}\cup C'_{k-2}$ is a clique.
\end{claim}
\begin{proof}
Notice that $C_{k-1}$ and $C'_{k-1}$ are cliques of $G$ coinciding with the intersections $B_{k-1}(u)\cap B_1(v)$ and $B_{k-1}(v)\cap B_1(v')$. Hence, $C_{k-1}\cap C'_{k-1}\neq \varnothing$ exactly when $\mathrm{TC}(u,vv')$ applies. Now assume that $C_{k-1}\cap C'_{k-1}=\varnothing$. This means that $\mathrm{PC}^2(u,vv')$ applies so $B_2(v)\cap B_{k-2}(u) = B_2(v')\cap B_{k-2}(u)$ and in particular $C_{k-2}\subseteq B_2(v')$ and $C'_{k-2}\subseteq B_2(v)$. Let $z\in C_{k-2}$ and $z'\in C'_{k-2}$. We assert that $z\sim z'$. Indeed there exists $x\sim v,z'$, and in particular we must have $x\in C_{k-1}$. Thus by $\mathrm{INC}^0(x,u)$ we get $z\sim z'$.
\end{proof}
Now, we prove the lemma by decreasing induction on $i$. The case $i=k$ is immediate since $C_k=\sg{v}, C'_k=\sg{v'}$, and $v\sim v'$. The case $i=k-1$ is covered by Claim \ref{claim:clique2nvx}. Hence, let $0<i\le k-2$. First suppose that there exists $x\in C_{i+1}\cap C'_{i+1}$. Then $C_i\cup C'_i\subset I(x,u)$. By $\mathrm{INC}^0(x,u)$, $C_i \cup C'_i$ forms a clique, thus we are in the second situation with $j=i$. Therefore, let $C_{i+1}\cap C'_{i+1} = \varnothing$. We can assume that $C_{i}\cap C'_{i}=\varnothing$ and
$C_{i-1}\cap C'_{i-1}=\varnothing$, otherwise we are immediately done.
By induction hypothesis, $C_{i+1}\cup C'_{i+1}$ forms a clique. We distinguish two cases.
\begin{claim}
\label{claim: cliqueTC}
If there exists $(x,x')\in C_{i+1}\times C'_{i+1}$ such that $\mathrm{TC}(u,xx')$ applies, then $C_{i}\cup C'_{i}$ is a clique.
\end{claim}
\begin{proof}
Pick $(x,x')\in C_{i+1}\times C'_{i+1}$ such that $\mathrm{TC}(u,xx')$ applies.
Then there exists $s\in B_{i}(u) \cap B_1(x)\cap B_1(x')$. If $s\in C_i \cap C'_i$ we are immediately done. Hence we can assume that $s\notin C'_i$, so there exists $b\in C'_{i+1}$ such that $s\nsim b$. Let $y'\in C'_i$. The $4$-cycle $sy'bx$ cannot be induced so we must have $y'\sim x$. Thus $y'$ is adjacent to any other vertex of $C_{i}$ by $\mathrm{INC}^0(u)$.
\end{proof}
\begin{claim}
\label{claim: cliquePC2}
If for any $(x,x')\in C_{i-1}\times C'_{i-1}$, $\mathrm{TC}(u,xx')$ does not apply, then $C_{i-1}\cup C'_{i-1}$ is a clique.
\end{claim}
\begin{proof}
Let $z\in C_{i-1}$ and $x\in C_{i+1}$. We will show that $z$ is adjacent to any vertex of $C'_{i-1}$. Note that $\mathrm{TPC}^2(u,xx')$ applies for any $x'\in C'_{i+1}$. In particular, we get that $z\in B_2(x)\cap B_{i-1}(u) = B_2(x')\cap B_{i-1}(u)$, so $z$ is at uniform distance $2$ from any vertex $x'$ of $C'_{i+1}$. As $i\leq k-2$, we can consider a vertex $t'\in C'_{i+2}$. By $\mathrm{INC}^+(t', z)$, there exists a vertex $s$ adjacent to $z$ and to every vertex of $C'_{i+1}$. In particular, we must have $d(u,s)=i$ so $s \in C'_i$. Thus for every $z'\in C_{i-1}$, we get $z\sim z'$ by $\mathrm{INC}^0(s,u)$ and we are done.
\end{proof}
We can now conclude applying Claim \ref{claim: cliqueTC} or Claim \ref{claim: cliquePC2}.
\end{proof}
\begin{lemma}\label{lem: cliquepath}
Let $u,v,v'$ be three vertices of $G$ such that $v\sim v'$ and $k:=d(u,v)\ge k':= d(u,v')$. Let $\gamma_{(u,v)}=(\sg{u}=C_0,C_1,\ldots,C_{k-1},C_k=\sg{v})$ and $\gamma_{(u,v')}=(\sg{u}=C'_0,C'_1,\ldots,C'_{k'-1},C'_{k'}=\sg{v'})$. Then for any $i \geq 0$ and any $(x,y) \in C_{i}\times C'_{i},$ we have $d(x,y) \leq 3.$
\end{lemma}
\begin{proof}
If $d(u,v')=k$, then the desired result is a consequence of Lemma \ref{lem: clique2nvx}. So, suppose that $d(u,v')=d(u,v)-1=k-1$. Then $v' \in C_{k-1}$, so by Lemma \ref{lem: inclalt} for every $i \in \sg{0, \ldots, k-1}$, either $C_{i}\subseteq C'_{i}$ or $C'_{i} \subseteq C_{i}$ according to the parity of $i$. Hence we are also done in this case.
\end{proof}
The next result has to be viewed as a symmetric version of the previous lemma, when we exchange the sources and sinks (i.e., $u\sim u'$ and $v=v'$).
\begin{lemma}
\label{lem: invcliquepath}
Let $u,u',v$ be three vertices of $G$ such that $u\sim u'$ and $k:=d(u,v)\ge k':=d(u',v)$. Let $\gamma_{(u,v)}=(\sg{u}=C_0,C_1,\ldots,C_{k-1},C_k=\sg{v})$ and $\gamma_{(u',v)}=(\sg{u'}=C'_0,C'_1,\ldots,C'_{k-1},C'_{k}=\sg{v})$ with the convention that $C'_i=\{ v\}$ if $i>k'$. Then for any
$0\le i\le k$ and any $(x,y) \in C_i \times C'_i$ we have:
\begin{itemize}
\item $d(x,y)=1$ if $d(u,v)>d(u',v)$;
\item $d(x,y) \leq 4$ if $d(u,v)=d(u',v)$.
\end{itemize}
\end{lemma}
\begin{proof}
Assume first that $k=d(u,v)>d(u',v)$. Then $d(u',v)=k-1$. Note that in this case for any $1\le i\le k$ the inclusion $C'_{i-1}\subseteq B_i(u)$ holds.
\begin{claim}
\label{clm: altinclusions}
For any $2\le i\le k-1$, the inclusion $C_{i-1}\subseteq C'_{i-2}$ implies the inclusion $C'_{i-3}\subseteq C_{i-2}.$
\end{claim}
\begin{proof}
Note that $C'_{i-3}\subseteq B_{i-3}(u')\cap I(u',v) \subseteq B_{i-2}(u)\cap I(u,v)$. If $C_{i-1}\subseteq C'_{i-2}$ holds, then any vertex of $C'_{i-3}$ is adjacent to any vertex of $C'_{i-2}$,
and therefore to any vertex of $C_{i-1}$, so we have $C'_{i-3}\subseteq C_{i-2}$.
\end{proof}
A particularly nice case occurs when the inclusions between the $C_i$'s and the $C'_{i-1}$'s alternate at each level, since in this case we are immediately done. However such an inclusion alternation does only hold in the sense of Claim \ref{clm: altinclusions}, i.e. the fact that $C'_{i-1}\subseteq C_i$ does not imply that $C_{i-2}\subseteq C'_{i-3}$. Hence we consider the maximal index $3\le i_0\le k-1$ (if it exists) such that $C'_{i_0-1}\subseteq C_{i_0}$ and $C_{i_0-1}\not\subseteq C'_{i_0-2}$. If $i_0$ does not exists, then we are done because by Claim \ref{clm: altinclusions} the inclusion alternation $C'_{k-2} \subseteq C_{k-1}$, $C_{k-2}\subseteq C'_{k-3}$, $C'_{k-4}\subseteq C_{k-3}$ hold and so on, until we reach the vertex $u'$. Otherwise, there is such an inclusion alternation until we reach $i_0$. This means
that for any $i \geq i_0-1$ and any $(x,y) \in C_{i}\times C'_i$ we have $d(x,y) \leq 1$. Now we show that this inequality still holds for
all $(x,y) \in C_{i}\times C'_i$ with $i<i_0$.
\begin{claim}
\label{clm: triv}
$C'_{i_0 - 2}\cup C_{i_0 - 1}$ is a clique.
\end{claim}
\begin{proof}
Pick any $a\in C'_{i_0 -1}\cap C_{i_0} \neq \varnothing$. By $\mathrm{INC}^0(a,u)$, any two vertices of $C'_{i_0 - 2}\cup C_{i_0 - 1}$ are adjacent. \end{proof}
Since $C_{i_0}\neq \varnothing$ and $C_{i_0-1}\not\subseteq C'_{i_0-2}$, there exists $p \in C_{i_0-1}\setminus C'_{i_0-2}$. Then $C'_{i_0-1}\subseteq C_{i_0}$ implies that
$p$ is adjacent to all vertices of $C_{i_0}$, hence the only reason why $p$ does not belong to $C'_{i_0 - 2}$ is that $d(u',p)=i_0-1$.
\begin{claim}\label{clm: disjointcliques}
For any $i \leq i_0-2$, there is no edge between $C'_{i-1}$ and $C_{i+1}$. In particular, $C'_{i-1} \cap C_i = \varnothing$.
\end{claim}
\begin{proof}
Suppose not, and let $ab$ be an edge with $a \in C'_{i-1}$ and $b \in C_{i+1}$. Then we can find a path of length $i_0-2$ from $u'$ to $p$ going via $ab$, with exactly one vertex in each clique $C'_j$ for $0\leq j\leq i-1$ and exactly one vertex in each clique $C_j$ for $ i+1 \leq j \leq i_0-1$, leading to a contradiction.
\end{proof}
\begin{claim}
\label{clm: zigzag}
For any $0\leq i\leq i_0-2$, the unions $C'_i\cup C_{i+1}$ and $C'_i\cup C_i$ are cliques.
\end{claim}
\begin{proof}
Pick any path $u = u_0, u_1, \ldots, u_{i_0-1} = p$ such that for every $i$, $u_i \in C_i$. Similarly consider any path $u' = u'_0, u'_1, \ldots, u'_{i_0-2}$ such that for every $i$, $u'_i \in C'_i$. By Claim \ref{clm: disjointcliques}, these paths are vertex-disjoint. By $\mathrm{INC}^0(v)$, $u' \sim u_1$. Because $d(u', p)=i_0-1$, the convexity of $B_{i_0-2}(u')$ implies that $u'_{i_0-2}\sim u_{i_0-2}$. By $\mathrm{INC}^0(u)$, we thus get $u_{i_0-2}\sim u'_{i_0 - 3}$. Note that this way we easily prove by induction that for every $1 \leq i\leq i_0 - 2$ we have the adjacencies $u_{i}\sim u'_i$ and $u_i \sim u'_{i-1}$. As we could have chosen any $u_i\in C_i$, $u'_i \in C'_i$ in this proof, this implies that $C'_i \cup C_{i+1}$ is a clique for every $i\leq i_0-3$, and that $C'_i \cup C_i$ is a clique for every $i \leq i_0 - 2$. Hence we are done, as the only remaining case was treated in Claim \ref{clm: triv}.
\end{proof}
By Claim \ref{clm: zigzag}, for any $i < i_0$ and any $(x,y) \in C'_{i}\times C_i$ we have $d(x,y) \leq 1$. This conclude the proof in the case $d(u,v) > d(u',v)$. Now, assume that $k = d(u,v) = d(u',v)$ and apply $\mathrm{TPC}^0(v,uu')$. If the triangle condition holds, then there exists $z \in B_{k-1}(v)$ with $z \sim u,u'$. Applying the
previous case to $v$ and the pairs $uz$ and $u'z$, by the triangle inequality we conclude that for any $(x,y)\in C_i\times C'_i$ we have $d(x,y) \leq 2.$
If the pentagon condition holds, then there exists $w_1, w_2, z$ with $d(w_1, v)=d(w_2, v) = k-1$, $d(z,v) = k-2$, $w_1 \sim u$, $w_2\sim u'$ and $z \sim w_1, w_2$. By the previous case applied to $v$ and the pairs $u w_1$, $w_1z$, $zw_2$, and $w_2u'$, the triangle inequality implies
that for any $(x,y)\in C_i\times C'_i$ we have $d(x,y) \leq 4.$
\end{proof}
\subsection{Biautomaticity}
\label{sec: biauto}
In this subsection, we apply the previous results to show that CB-groups are biautomatic.
We continue by recalling the definition of biautomatic group \cite{ECHLPT,BrHa}. Let $G=(V,E)$ be a graph and
suppose that $\Gamma$ is a group acting geometrically by automorphisms
on $G$. These assumptions imply that the graph $G$ is locally finite
and that the degrees of the vertices of $G$ are uniformly bounded.
The action of $\Gamma$ on $G$ induces the action
of $\Gamma$ on the set ${\mathcal P}(G)$ of all paths of $G$. A path
system ${\mathcal P}\subseteq {\mathcal P}(G)$ is called
$\Gamma$--\emph{invariant} if $g\cdot \gamma \in \mathcal P$, for all
$g\in \Gamma$ and $\gamma \in \mathcal P$.
Let $\Gamma$ be a group generated by a finite set $S$. A \emph{language} over $S$ is some set of words in $S\cup S^{-1}$ (in the
free monoid $(S\cup S^{-1})^{\ast}$).
A language over $S$ defines a $\Gamma$--invariant path system in the Cayley graph Cay$(\Gamma,S)$.
A language is \emph{regular} if it is accepted by some finite state automaton.
A \emph{biautomatic structure} is a pair $(S,\mathcal L)$, where $S$ is as above, $\mathcal L$ is a regular language over $S$, and
the associated path system in Cay$(\Gamma,S)$ is a bicombing. A group is
\emph{biautomatic} if it admits a biautomatic structure.
We use specific conditions implying biautomaticity for groups acting geometrically on graphs. The method, relying on the notion of locally recognized path system, was developed by {\'S}wi{\c{a}}tkowski \cite{Swiat}.
Let $G$ be a graph and let $\Gamma$ be a group acting geometrically on $G$.
Two paths $\gamma_1$ and $\gamma_2$ of $G$ are $\Gamma$-\emph{congruent} if there is $g\in \Gamma$ such that $g\cdot\gamma_1=\gamma_2$. Denote by ${\mathcal S}_k$ the set of $\Gamma$-congruence classes
of paths of length $k$ of $G$. Since $\Gamma$ acts cocompactly on $G$, the sets ${\mathcal S}_k$ are finite for any natural $k$. For any path $\gamma$ of $G$, denote by $[\gamma]$ its $\Gamma$-congruent class. For a subset $R\subset {\mathcal S}_k$, let ${\mathcal P}_R$ be the path system in $G$ consisting of all paths $\gamma$ satisfying the following two conditions:
\begin{itemize}
\item[(1)] if $|\gamma|\ge k$, then $[\eta]\in R$ for any subpath $\eta$ of length $k$ of $\gamma$;
\item[(2)] if $|\gamma|<k$, then $\gamma$ is a prefix of some path $\eta$ such that $[\eta]\in R$.
\end{itemize}
A path system $\mathcal P$ in $G$ is $k$--\emph{locally recognized} if for some $R\subset {\mathcal S}_k$, we have ${\mathcal P}={\mathcal P}_R$, and $\mathcal P$ is \emph{locally recognized}
if it is $k$--locally recognized for some $k$. The following result of {\'S}wi{\c{a}}tkowski \cite{Swiat} provides sufficient conditions of biautomaticity in terms of local recognition and bicombing.
\begin{theorem} \label{swiat} \cite[Corollary 7.2]{Swiat} \label{th:swiat} Let $\Gamma$ be a group acting geometrically on a graph $G$ and let $\mathcal P$ be a path system in
$G$ satisfying the following conditions:
\begin{itemize}
\item[(1)] $\mathcal P$ is locally recognized;
\item[(2)] there exists $v_0\in V(G)$ such that any two vertices from the orbit $\Gamma \cdot v_0$ are connected by a path from $\mathcal P$;
\item[(3)] $\mathcal P$ satisfies the $2$--sided fellow traveler property.
\end{itemize}
Then $\Gamma$ is biautomatic.
\end{theorem}
Based on Theorems \ref{thm: fellowtravdist1} and \ref{th:swiat}, we obtain the following result:
\begin{theorem}\label{thm: biauto}
Let $G$ be a CB-graph and $\Gamma$ be a group acting geometrically on $G$. Then $\Gamma$ is biautomatic.
\end{theorem}
\begin{proof}
As mentioned above, since $\Gamma$ acts geometrically on $G$, the graph $G$ has bounded degrees, so all cliques of $G$ are finite and we can use all previous results.
Recall that every group action of $\Gamma$ on $G$ induces a group action of $\Gamma$ on the barycentric subdivision $\beta(X(G))$
of the clique complex $X(G)$, and hence on the graph $\beta(G):= \beta(X(G))^{(1)}$. Moreover, if $\Gamma$ acts geometrically
on $G$, then $\Gamma$ also acts geometrically on $\beta(G)$. Now observe that every normal clique-path $\gamma_{(u,v)}(\sg{u}=C_0,C_1,\ldots,C_k=\sg{v})$
of length $k$ between two vertices $u,v$ of $G$ correspond to the path of $\gamma^*_{(u,v)}=(C_0, C_0\cup C_1, C_1, C_1\cup C_2, C_2, \ldots, C_{k-1}, C_{k-1}\cup C_k, C_k)$
of length $2k+1$ of the graph $\beta(G)$. Denote by $\Upsilon^*$ the set of all paths $\gamma^*_{(u,v)}$ of $\beta(G)$ derived from the set
$\Upsilon$ of normal clique-paths $\gamma^*_{(u,v)}$ of $G$. We assert that if $\Gamma$ acts geometrically on $G$, then $\Upsilon^*$
satisfies the conditions of Theorem \ref{th:swiat}.
First observe that if $u$ and $v$ are at distance $k>0$ in $G$, then for any two cliques $C,C'$ with $u\in C$ and $v\in C'$, the vertices of $\beta(G)$ corresponding to $C$ and $C'$ are at distance at most $5k$ in $\beta(G)$. Thus by Theorem \ref{thm: fellowtravdist1} the system of paths $\Upsilon^*$ of $\beta(G)$ also satisfies the $2$-sided fellow traveler property, with associated constants $C$ and $D$ five times higher that those for $\Upsilon$. A simple argument similar to the one used in \cite{JS,Swiat} implies that the system of paths $\Upsilon^*$ is $3$-locally recognizable, thanks to the local characterization of normal clique-paths provided by Lemma \ref{lem: normalcliquepathunicity}. Finally, observe that the action of $\Gamma$ on $\beta(G)$ preserve the size of the cliques, so vertices are sent to vertices, thus condition $(2)$ of Theorem \ref{th:swiat} also holds, by Lemma \ref{lem: normalcliquepathunicity}.
\end{proof}
\subsection{Falsification by fellow traveler property}
The \emph{falsification by fellow traveler property (FFTP)} was initially introduced for groups by Neumann and Shapiro \cite{NeSh} and further investigated by Elder \cite{Elder_thesis,Elder02}. It was show in \cite{NeSh} that in the groups satisfying FFTP the language of geodesics is regular. It was show in \cite{Elder02} that
the groups satisfying FFTP are almost convex (in the sense of Cannon \cite{Cannon}) and satisfy the quadratic isoperimetric inequality.
The definition of FFTP can be formulated in the language of graphs as follows.
We say that two paths $\gamma, \gamma'$ in $G$ \emph{asynchronously $K$-fellow travel} for some constant $K>0$ if there exists a proper nondecreasing function $f:\mathbb [0, \infty) \to \mathbb [0, \infty)$ such that for every $i\geq 0$,
$$d(\gamma(i), \gamma'(f(i)))\leq Kd(\gamma(0), \gamma'(0)).$$
A graph $G$ is said to have the \emph{falsification by fellow traveler property} (or FFTP for short) if there exists a constant $K>0$ such that for every path $\gamma$ of $G$ which is not a geodesic, there exists a path $\eta$ of $G$ such that $\eta(0)=\gamma(0)$, $\eta(\infty)=\gamma(\infty)$, $|\eta|<|\gamma|$ and $\gamma$ and $\eta$ asynchronously $K$-fellow travel.
\begin{proposition}
\label{lem: FFTP}
Let $G$ be a CB-graph or a weakly modular graph. Then $G$ enjoys FFTP. Consequently, any group $\Gamma$ whose Cayley graph is weakly modular or a CB-graph enjoys FFTP.
\end{proposition}
\begin{proof} We first prove the result for CB-graphs and start with a general claim about them.
\begin{claim}
\label{clm: path-homotopy}
Let $G$ be a CB-graph (respectively a weakly modular graph) and $u,v,w,w'$ be four vertices such that $k:=d(u,v)=d(u,w)+1=d(u,w')+1$ and $v\sim w,w'$. Then for every geodesic $\gamma$ from $u$ to $w$ there exists a geodesic $\gamma'$ from $u$ to $w'$ which $1$-fellow travels (respectively $2$-fellow travels)
with $\gamma$.
\end{claim}
\begin{proof}
Assume that $G$ is a CB-graph.
By $\mathrm{INC}^0(v,u)$ we have $w\sim w'$. Now we proceed by induction on $k\geq 2$. If $k =1$, then we are immediately done. If $k\geq 2$, then
by $\mathrm{INC}(v,u)$ there exists $z\sim w, w'$ such that $d(u,z)=k-2$. Now we can apply the induction hypothesis to the vertices $u,w,\gamma(k-2),z$ and find a geodesic from $u$ to $z$ which $1$-fellow travels with $(\gamma(0), \ldots, \gamma(k-2))$. Completing this path with $w'$ we are done. If $G$ is weakly modular, then we are done by using $\mathrm{QC}$ instead of $\mathrm{INC}$.
\end{proof}
We now show that FFTP holds with constant $K=2$ for CB-graphs and $K=3$ for weakly modular graphs. Let $G$ be a CB-graph or a weakly modular graph.
Let $\gamma$ be a non-geodesic path in $G$ and let $u:=\gamma(0)$ and $k:=|\gamma|$. Then $k\geq 2$.
As $\gamma$ is not a geodesic, we can define $i_0:=\min\sg{i\geq 1, d(u,\gamma(i))\geq d(u,\gamma(i+1))}= \min\sg{i\geq 1, d(u, \gamma(i+1))\leq i}$.
If $i_0=1$, then either $\gamma(2)=u$ or $\gamma(2)\sim u$, so we are done by removing $\gamma(0)$ and $\gamma(1)$ or only $\gamma(1)$ from $\gamma$.
Now, let $i_0\geq 2$. Let $w:=\gamma(i_0-1)$, $v:=\gamma(i_0)$, $v':=\gamma(i_0+1)$. If $d(u,v')=i_0-1$, then by Claim \ref{clm: path-homotopy} there exists a $(u,v')$-geodesic $\gamma'$ which either $1$-fellow or $2$-fellow travels with $(\gamma(0), \ldots, \gamma(i_0-1)=w)$. So we can choose $\eta := (\gamma'(0), \ldots, \gamma'(i_0-2), v',\gamma(i_0+2), \ldots, \gamma(k))$ and we are done.
Otherwise, $d(u,v')=i_0$ and we apply $\mathrm{TPC}^1(u, vv')$ or $\mathrm{TC}(u, vv')$ according to whether $G$ is a CB-graph or weakly modular. If $\mathrm{TC}(u,vv')$ applies, then there exists a vertex $w'\sim v,v'$ at distance $i_0-1$ from $u$. By Claim \ref{clm: path-homotopy}, we can find a geodesic $\gamma'$ from $u$ to $w'$ which $1$-fellow or $2$-fellow travels with $(\gamma(0), \ldots, \gamma(i_0-1)=w)$. We thus let $\eta:=(u=\gamma'(0), \ldots, \gamma'(i_0-2),w', v', \gamma(i_0+2), \ldots, \gamma(k))$.
If $\mathrm{PC}^1(u,vv')$ holds with respect to the neighbor $w$ of $v$, then there exists two vertices $w',z$ with $d(u,z)=i_0-2$, $d(u,w')=i_0-1$, $w'\sim v'$ and $z\sim w', w$. By Claim \ref{clm: path-homotopy} there exists a geodesic $\gamma'$ from $u$ to $z$ which $1$-fellow travels with $(u=\gamma(0), \ldots, \gamma(i_0-2))$. Thus we let
$\eta:=(u=\gamma'(0), \ldots, \gamma'(i_0-3), z, w', v', \gamma(i_0+2), \ldots, \gamma(k))$.
\end{proof}
In \cite{Cannon}, Cannon introduced the notion of \emph{almost convexity}. As for FFTP, the initial definition was for groups, and we extend it there in the language of graphs.
A graph $G$ is said to be \emph{$k$-almost convex} ($\mathrm{AC}(k)$ for short) for some $k\geq 2$ if there exists a constant $K_k>0$ such that for every $v\in V, n \geq 0$ and every pair of vertices $x, y\in S_n(v)$ with $d(x,y)\leq k$ there exists a path from $x$ to $y$ of length at most $K_k$ which is entirely included in $B_n(v)$. A graph $G$ is said to be \emph{almost convex} if there exists some $k\geq 3$ such that $G$ is $k$-almost convex.
Observe that the class almost convex graphs trivially generalizes the class of CB-graphs. It is easy to show (see for example \cite{Cannon}) that for every $k\geq 3$, the property $\mathrm{AC}(k)$ is equivalent to $\mathrm{AC}(3)$. The fact that CB-graphs are graphs with $3$-convex balls (item (ii) of Theorem \ref{thm: well-bridged}) holds for exactly the same kind of reason. It was shown in \cite[Proposition 1]{Elder02} that graphs satisfying FFTP are almost convex.
\begin{remark} In general, the Cayley graph of a weakly modular or a CB-group is not weakly modular or a CB-graph. Recently, the paper \cite{Soe} provides
sufficient conditions on a presentation of a group, which imply that its Cayley graph is systolic.
\end{remark}
In view of previous remark, one can ask the following questions:
\begin{question} Is it true that CB-groups and weakly modular groups satisfy FFTP? Are CB-groups and weakly modular groups almost convex?
\end{question}
\begin{remark} Example \ref{CB-pentagons-infinite} shows that there exist CB-graphs which are not weakly systolic and which have an arbitrary diameter and contain an arbitrary number of pentagons not included in 5-wheels.
Nevertheless, such graphs are quasi-isometric to weakly systolic graphs. In view of this, one can ask the following question: \end{remark}
\begin{question} Is it true that CB-groups are weakly systolic?
\end{question} \section{Metric triangles}
\label{sec: metriangle}
Recall that three vertices $u,v,w$ of a graph $G$ form a \emph{metric triangle} $uvw$ if $I(u,v)\cap I(u,w) = \sg{u}$, $I(u,v)\cap I(v,w) = \sg{v}$, and $I(u,w)\cap I(u,w) = \sg{w}$ \cite{Ch_metric}. The pairs $uv$, $vw$, and $wu$ are called the \emph{sides of $uvw$}. The metric triangle $uvw$ has \emph{type} $(k_1,k_2,k_3)$ if its sides have lengths $k_1, k_2,k_3$ and $k_1\ge k_2\ge k_3$. If $k_1=k_2=k_3=k$, then $uvw$ is called \emph{equilateral of size $k$}. If $k_3\ge k_1-1$, then we call $uvw$ \emph{almost equilateral}. A metric triangle $uvw$ is called \emph{strongly equilateral} if all vertices of $I(v,w)$ have the same
distance to the opposite vertex $u$. Clearly, strongly equilateral metric triangles are equilateral. A metric triangle $uvw$ is a \emph{quasi-median} of vertices $x,y,z$ if the following equalities hold:
$$d(x,y)=d(x,u)+d(u,v)+d(v,y),$$
$$d(y,z)=d(y,v)+d(v,w)+d(w,z),$$
$$d(z,x)=d(z,w)+d(w,u)+d(u,x).$$
If $uvw$ has size 0, then it is called the \emph{median} of $x,y,z$. Any triplet of vertices $x,y,z$ admits at least one quasi-median: it suffices to chose $u$ in $I(x,y)\cap I(x,z)$ as far as possible from $x$, $v$ in $I(y,u)\cap I(y,z)$ as far as possible from $y$, and $w$ in $I(z,u)\cap I(z,v)$ as far as possible from $z$; if $x,y,z$ form a metric triangle, then $xyz$ is the unique quasi-median of $x,y,z$.
Metric triangles and quasi-medians an important role and have interesting properties in classes of graphs defined by metric conditions \cite{BaCh_survey}. For example, it was shown in \cite{Ch_metric} that weakly modular graphs are exactly the graphs in which all metric triangles are strongly equilateral. Modular graphs are exactly the graphs in which all metric triangles have size $0$, i.e., all triplets of vertices have medians. Median graphs are exactly the graphs in which all triplets of vertices have unique medians.
In this section, we show that in CB-graphs metric triangles behave quite similarly to metric triangles in weakly modular graphs. Namely, our main result there is the following classification theorem:
\begin{theorem}
\label{thm: metrictriangle}
Let $uvw$ be a metric triangle of a CB-graph $G$ with
$d(u,v) \leq d(u,w) \leq d(v,w)$. Then $uvw$ is either strongly
equilateral or has type $(2,2,1)$ and is included in a pentagon of
the form
$uvxwy$. \end{theorem}
\begin{proof}
The proof is a consequence of Lemmas \ref{lem: almosteq}, \ref{lem: no2k-1}, and \ref{lem: no2k}. By Lemma \ref{lem: almosteq}, $uvw$ is almost equilateral. If $uvw$ is not equilateral, then Lemma \ref{lem: no2k-1} implies that $uvw$ has type $(k,k,k-1)$ for some $k \geq 2$. By Lemma \ref{lem: no2k}, we must have $k=2$, so $uvw$ lies on a pentagon as asserted. If $uvw$ is equilateral, then $uvw$ is strongly equilateral as a direct consequence of Lemma \ref{lem: almosteq}.
\end{proof}
We start with an auxiliary lemma:
\begin{lemma} \label{closestk-1} Let $uvw$ be a metric triangle in a graph with convex balls $G$ such that $k=d(u,v)\ge d(u,w)$ and $v'$ be a neighbor of $v$ in $I(v,u)$. Suppose that $I(v,w)$ contains a vertex $t$ such that $d(u,t)\le k-1$. Then $I(v,t)\subset I(v,w)$ contains a vertex $x$ such that $d(u,x)=k-1$ and $d(v,x)=d(v',x)=2$. Moreover, there exists a vertex $v''\in I(v,u)$ at distance 2 from $v$ and adjacent to $x$.
\end{lemma}
\begin{proof} Let $x$ be a closest to $v$ vertex of $I(v,t)$ such that $d(u,x)\le k-1$. First we prove that $d(v',x)=d(v,x)$. Indeed, since $v'\in I(v,u)$ and $uvw$ is a metric triangle, necessarily $v'\notin I(v,w)$, yielding $v'\notin I(v,x)\subset I(v,w)$. Thus $d(v',x)\ge d(v,x)$. If $d(v',x)>d(v,x)$, then $v\in I(v',x)$ and since $v',x\in B_{k-1}(u)$ and $v\notin B_{k-1}(u)$, we obtain a contradiction with the convexity of $B_{k-1}(u)$. This show that $d(v',x)=d(v,x)$.
Now we show that $d(v,x)=2$. Since $d(v',x)=d(v,x)$, we can apply $\mathrm{TPC}^1(x,vv')$. If $\mathrm{TC}(x,vv')$ applies, then there exists some $s \sim v,v'$ one step closer to $x$. In particular, $s \in I(v',x)$, so by convexity of $B_{k-1}(u)$ we have $d(u,s)\le k-1$. This implies that $s \in I(v,u)\cap I(v,w)$, which contradicts that $uvw$ is a metric triangle. Hence, only $\mathrm{PC}^1(x,vv')$ can be applied. By this property there exists a vertex $s\in I(v,x)\cap I(v',x)$ at distance 2 from $v$ and $v'$. Since $v',x\in B_{k-1}(u)$, the convexity of $B_{k-1}(u)$ implies that $d(u,s)\le k-1$. The minimality choice of $x$ implies that $x=s$, i.e., $d(v,x)=2$.
Let $z$ be a common neighbor of $v$ and $x$. Then $d(u,z)=k$ and we can apply $\mathrm{TPC}^1(u,vz)$. If $\mathrm{TC}(u,vz)$ applies, then there exists a vertex $v'\sim v,z$ at distance $k-1$ from $u$. Since $x,v'\in I(z,u)$ are neighbors of $z$, by $\mathrm{INC}^0(u)$ we conclude that $x\sim v'$, contrary to the conclusion that $d(x,v')=2$. Hence only $\mathrm{PC}^1(u,vz)$ applies and the neighbor $x\in I(z,u)$ of $z$. By this condition, there exists a vertex $v''\in I(v,u)\cap I(z,u)$ having distance 2 to $v$ and $z$ and adjacent to $x$.
\end{proof}
\begin{lemma}
\label{lem: almosteq}
Let $G$ be a graph with convex balls and $uvw$ be a metric triangle of
$G$. If $d(u,v)\geq d(u,w)$, then $d(u,v) \leq d(u,w) +1$ and for any
$x\in I(v,w)$ we have $d(u,w) \leq d(u,x) \leq d(u,v).$ In particular,
$uvw$ is almost equilateral and if $uvw$ is equilateral, then it is
strongly equilateral.
\end{lemma}
\begin{proof}
The first assertion immediately implies the remaining assertions. Indeed, the fact that any two sides of $uvw$ have length which differ by at most 1 implies that all metric triangles of $G$ are almost equilateral. If $uvw$ is equilateral of size $k$, then the fact that $d(u,w) \leq d(u,x) \leq d(u,v)$ implies that $d(u,x)=k$ for any $x\in I(v,w)$. Finally, if $k:=d(u,v)=d(u,w)+1$ and $\gamma(v,w)=(v=v_0, \ldots, v_m = w)$ is a shortest $(v,w)$-path, let $v_{i+1}$ be first vertex of $\gamma(v,w)$ (when moving from $v$ to $w$) such that $d(u,v_{i+1})=k-1$. Then the convexity of $B_{k-1}(u)$ implies that $d(u,v_j)\le k-1$ for any $i+1\le j\le m$.
So, it remains to show the first assertion of the lemma. Let $k:= d(u,v)\geq d(u,w)$. By convexity of the ball $B_k(u)$, $d(u,x)\le k$ for any $x\in I(v,w)$.
If the first assertion does not hold, then there exists $x\in I(v,w)$ such that $d(u,x)\le k-2$ if $d(u,v)>d(u,w)$ and such that $d(u,x)\le k-1$ if $d(u,v)=d(u,w)$.
We will select such a vertex $x$ as close as possible to the vertex $v$. We distinguish two cases.
\medskip\noindent
{\bf Case 1:} $k=d(u,v)>d(u,w)$.
\begin{proof}
Let $x$ be a closest to $v$ vertex of $I(v,w)$ such that $d(u,x)\le k-2$. From the choice of $x$ we conclude that $d(u,x)=k-2$ and $k-1\le d(u,x')\le k$ for any vertex $x'\in I(v,x)\setminus \{ x\}$. Let $z$ be a closest to $v$ vertex of $I(v,x)$ such that $d(u,z)\le k-1$. From Lemma \ref{closestk-1} we conclude that $d(v,z)=2$ and that there exists $v''\in I(v,u)$ having distance 2 to $v$ and adjacent to $z$. Since $d(u,v'')=k-2=d(u,x)$ and $d(u,z)=k-1$, from the convexity of the ball $B_{k-2}(u)$ we infer that $z\notin I(v'',x)$. This implies that $d(v'',x)\le d(z,x)$. Since $d(v,v'')=d(v,z)=2$, we conclude that $v''\in I(v,x)\cap I(v,u)\subset I(v,w)\cap I(v,u)$, contrary to the assumption that $I(v,w)\cap I(v,u)=\{ v\}$. This contradiction finishes the proof of Case 1.
\end{proof}
\medskip\noindent
{\bf Case 2:} $k=d(u,v)=d(u,w)$.
\begin{proof}
Let $x$ be a vertex of $I(v,w)$ with $d(u,x)\le k-1$. Let $x'$ be a closest to $v$ vertex of $I(v,x)$ at distance $\le k-1$ from $u$ and let $y'$ be a closest to $w$ vertex of $I(x,w)$ at distance $\le k-1$ from $u$. By Lemma \ref{closestk-1}, we get $d(u,x')=k-1=d(u,y')$, $d(v,x')=d(w,y')=2$ and for any neighbor $v'$ of $v$ in $I(v,u)$ and for any neighbor $w'$ of $w$ in $I(w,u)$ the equality $d(v,x')=d(v',x')=2=d(w,y')=d(w',y')$ holds. Moreover, there exists a vertex $v''\in I(v,u)$ at distance 2 from $v$ and adjacent to
$x'$ and a vertex $w''\in I(w,u)$ at distance 2 from $w$ and adjacent to $y'$. Let also $z$ be a common neighbor of $v$ and $x'$ and $v'$ be a common neighbor of $v$ and $v''$. If $z\sim v'$, to avoid an induced 4-cycle defined by the vertices $z,v',v'',$ and $x'$, either $v'\sim x'$ or $z\sim v''$. This implies that either $v'\in I(v,u)\cap I(v,w)$ or $z\in I(v,u)\cap I(v,w)$, contrary to the assumption that $I(v,u)\cap I(v,w)=\{ v\}$. Consequently, $v,v',v'',z,x$ induce a pentagon.
Since $v'',w''\in B_{k-2}(u)$ and $x',y'\notin B_{k-2}(u)$, the path consisting of the edges $v''x', y'w''$ and a shortest $(x',y')$-path cannot be a shortest $(v'',w'')$-path. If we suppose that $d(v,w)=m$, since $d(v,x')=d(y',w)=2$, this implies that $d(v'',w'')<m-4+2=m-2$. Hence $d(v'',w'')\le m-3$. On the other hand, if $d(v'',w'')\le m-4$, then $v''$ and $w''$ belong to a shortest $(v,w)$-path, contradicting the assumption that $I(v,u)\cap I(v,w)=\{ v\}$. This shows that $d(v'',w'')=m-3$. This implies that $d(w'',v')\le m-2$. Since $w''\sim y'$ and $d(y',z)=d(y',x')+1=m-3$, we also conclude that $d(w'',z)\le m-2$. Since $z\nsim v'$, we obtain that $v\in I(v',z)$. Since $v',z\in B_{m-2}(w'')$ and $B_{m-2}(w'')$ is convex, we conclude that $d(w'',v)\le m-2$. This implies that $w''\in I(w,v)$, contrary to the assumption that $I(w,v)\cap I(w,u)=\{ w\},$ concluding the analysis of Case 2.
\end{proof}
This finishes the proof of the lemma. \end{proof}
Using Lemma \ref{lem: almosteq} we can easily prove the following result:
\begin{lemma}
\label{lem: no2k-1}
A CB-graph $G$ does not contain metric triangles of the type $(k,k-1,k-1)$ for any $k\ge 2$.
\end{lemma}
\begin{proof}
Assume by way of contradiction that $G$ contains a metric triangle $uvw$ with $d(u,w)=d(v,w)=k-1$ and $d(u,v)=k$. Then obviously $k\geq 3$. By Lemma \ref{lem: almosteq}, for any vertex $y\in I(u,v)$ we have $k-1=d(w,v)\le d(w,y)\le d(w,u)=k-1$, i.e., $d(w,y)=k-1$. Let $x$ be a closest to $v$ vertex of $I(v,w)$ such that $d(u,x)\le k-1$ (such a vertex exists because $d(u,w)=k-1$). By Lemma \ref{closestk-1}, $d(v,x)=2$ and there exists a vertex $v''\in I(v,u)$ at distance 2 from $v$ and adjacent to $x$. Since $d(x,w)=k-3$, we conclude that $d(w,v'')\le k-2$, contrary to the fact that $d(w,y)=k-1$ for any $y\in I(u,v)$.
\end{proof}
To conclude the proof of Theorem \ref{thm: metrictriangle}, it only remains to show the following lemma:
\begin{lemma}
\label{lem: no2k}
A CB-graph $G$ does not contain metric triangles of type $(k,k,k-1)$ for any $k\ge 3$.
\end{lemma}
\begin{proof} We proceed by induction on $k$.
Suppose by way of contradiction that $G$ contains a metric triangle $uvw$ with $d(u,v)=d(u,w)=k \geq 3$ and $d(v,w)=k-1$. Let $x$ be a closest to $u$ vertex of $I(v,u)$ at distance $k-1$ from $w$ (such a vertex exists because $d(w,v)=k-1$).
By Lemma \ref{closestk-1}, $d(x,u)=2$ and there exists a vertex $y\in I(u,w)$ at distance 2 from $u$ and adjacent to $x$. Consider a quasi-median $x'v'w'$ of the triplet $x,v,w$. Since $I(v,x)\cap I(v,w)\subset I(v,u)\cap I(v,w)=\{ v\}$, we obtain that $v'=v$.
If $x'\ne x$, then $I(x,v)\cap I(x,w)$ contains a vertex $s$ adjacent to $x$. Since $s,y\in I(x,w)$ are neighbors of $x$, by $\mathrm{INC}^0(w)$ $s$ and $y$ are adjacent. Consequently, $y\in I(u,v)\cap I(u,w)$, contrary to the assumption that $I(u,v)\cap I(u,w)=\{ u\}$. Hence $x'=x$. Thus any quasi-median of $x,v,$ and $w$ has the form $xvw'$.
First, let $w'\ne w$. Since $d(x,w)=d(v,w)=k-1$ and $d(x,v)=k-2$, Lemma \ref{lem: no2k-1} implies that the metric triangle $xvw'$ has type $(k-2,k-2,k-2)$ and thus $w'$ is adjacent to $w$.
Notice that $y\notin I(x,w')$, otherwise $d(y,w')=k-3$, yielding $d(u,w')\le k-2$. This implies $w'\in I(w,u)\cap I(w,v)$, contrary to the assumption that $I(w,u)\cap I(w,v)=\{ w\}$. Thus $y\notin I(x,w')$. Pick any neighbor $y'$ of $x$ in $I(x,w')$.
Since $I(x,w')\subset I(x,w)$, we conclude that $y',y\in I(x,w)$. Hence by $\mathrm{INC}^0(w)$, $y'\sim y$. Let $s$ be a common neighbor of $x$ and $u$ and $t$ be a common neighbor of $y$ and $u$. If $s\sim t$, then $x,s,t,y$ define a 4-cycle which cannot be induced, thus $x\sim t$ or $y\sim s$. This implies that $t\in I(u,w)\cap I(u,v)$ or $s\in I(u,v)\cap I(u,w)$, which is impossible. Hence we have $s\nsim t$ and by convexity of $B_2(y')$ we get $d(y',u)\leq 2$. In particular this implies that $d(w', u)\leq k-1$, hence $w'\in I(u,w)\cap I(v,w)$, which is impossible.
Now, let $w'=w$. Then $xvw$ is a metric triangle of type $(k-1,k-1,k-2)$. By induction hypothesis, this is possible only if $k-1=2$, i.e., $k=3$. This implies that $x\sim v$, $y\sim w$, and $d(v,w)=2$. Since $uvw$ is a metric triangle, $x\nsim w$ and $y\nsim v$. Let $s$ be any common neighbor of $x$ and $u$ and $t$ be any common neighbor of $y$ and $u$. Then $xsuty$ is an induced 5-cycle, otherwise one of the vertices $s$ or $t$ belongs to $I(u,v)\cap I(u,w)$, which is impossible.
Let $z$ be a common neighbor of $v$ and $w$. By Lemma \ref{lem: almosteq}, $d(u,z)=d(u,v)=d(u,w)=3$. If $z$ is adjacent to one of the vertices $x$ and $y$, then $z$ is adjacent to both $x$ and $y$. In this case, $d(z,s)=d(z,t)=2$ and $u\in I(s,t)\setminus B_2(z)$, contrary to the convexity of $B_2(z)$. Hence $z\nsim x,y$. In particular $xywzv$ is an induced 5-cycle, and we can apply Lemma \ref{lem: deuxpent}. Thus there exists a vertex $a$ universal for one of the two pentagons $xsuty$ and $xywzv$. If $a$ is adjacent to all the vertices of $xsuty$, then $a\in I(u,v)\cap I(u,w)$, which is a contradiction as $uvw$ is a metric triangle. Thus $a$ is adjacent to all the vertices of $xywzv$. In particular, by convexity of $B_2(a)$ we get $d(a,u)=2$. Hence $a \in I(u,v)\cap I(u,w)\cap I(v,w)$, contradicting again the fact that $uvw$ is a metric triangle.
\end{proof}
We continue with a useful property of (strongly) equilateral metric triangles.
\begin{lemma}
\label{lem: equilateral1}
Let $uvw$ be an equilateral metric triangle of size $k\ge 3$ of a CB-graph $G$. If $u'\in I(v,u)$ and $w'\in I(v,w)$ are such that $d(v,u')=d(v,w')=2$ and $d(u',w')\le 2$, and $a,b,c$ are three vertices such that $a\sim u',v$, $b\sim w',v$, and $c\sim u',w'$, then $a,b,c$ are pairwise adjacent.
\end{lemma}
\begin{proof}
By Theorem \ref{thm: metrictriangle}, $uvw$ is strongly equilateral, thus $d(u',w')=2$ and $c\in I(u', w')\subset I(w,u')\cap I(u,w')$.
The convexity of $B_2(v)$ implies that $d(c,v)\leq 2$. As $uvw$ is a metric triangle and $d(u,c)=d(w,c)=k-1$, we must have $d(c,v)=2$.
For the same reason, $d(u',b)\ge 2 and d(a,w')\ge 2$.
Let $s$ be a common neighbor of $c$ and $v$. If $s$ coincides with one of the
vertices $a,b$ and is adjacent to the second, say $s=a$ and $s\sim b$, then $a$ and $b$ are adjacent and $a$ is adjacent to $c$.
Then we get the 4-cycle $acw'b$, which cannot be induced. If $a\sim w'$, then $a\in I(v,u)\cap I(v,w)$, which contradicts that $uvw$ is a
metric triangle. If $b\sim c$, then $a,b,c$ are pairwise adjacent and we are done. Thus, further we can suppose that either $s$ is
different from $a$ and $b$ or $s$ coincides with $a$ and $b$ but is not adjacent to the second vertex.
Let $w''$ be a neighbor of $w'$ in $I(w',w)$. Since $d(w,w'')=k-3$ and $d(w,u')=k$ because $uvw$ is strongly equilateral, we deduce that
$d(u',w'')=3$. This implies that $w''\ne c$ and $w''\nsim c$. Since $uvw$ is strongly equilateral, we have $d(u,w')=d(u,w'')=k$. Hence we can apply
$\mathrm{TPC}^1(u, w'w'')$ to the edge $w'w''$. First suppose that $\mathrm{PC}^1(u,w'w'')$ applies
with the neighbor $c\in I(u,w')$ of $w'$. By this condition there exists a pentagon $w'cz'zw''$ such that
$d(u,z')=k-2$ and $d(u,z)=d(u,c)=k-1$. Note that $z'\neq u'$, otherwise $d(u',w'')\leq 2$, contradicting the fact that $d(u',w'')=3$.
Since $z',u'\in I(c,u)$, by $\mathrm{INC}^0(u)$ we get $z'\sim u'$. Since $z,w'\in B_2(u')$, $w''\in I(z,w')$, and $w'\nsim z$, we obtain a
contradiction with the convexity of $B_2(u')$. This contradiction shows that only $\mathrm{TC}(u,w'w'')$ applies to edge $w'w''$. Consequently,
we can find $z\ne c$ adjacent to $w',w''$ and having distance $k-1$ to $u$. By $\mathrm{INC}(u)$, we get $z\sim c$ and there exists $z'\sim c,z$ at distance $k-2$ from $u$.
Again, since $d(u',w'')=3$, $z'\neq u'$. Since $z',u'\in I(u,c)$, by $\mathrm{INC}(u)$ we get $z'\sim u'$ and there exists $u''\sim z',u'$ at distance $k-3$ from $u$.
Since $d(z,u)=k-1$, we must have $d(v,z)=3$, otherwise $z\in I(v,w'')\subset I(v,w)$ and we obtain a contradiction with the fact that $uvw$ is strongly
equilateral. Analogously, one can show that $d(v,z')=3$. Summarizing, we obtain that $d(z',a)=d(z',s)=d(z,b)=d(z,s)=2$ and $d(v,z)=d(v,z')=3$.
The convexity of the balls $B_2(z)$ and $B_2(z')$ implies that $s$ either is different and adjacent to $a$ and $b$ or $s$
coincides with one of the vertices $a,b$ and is adjacent to the second. Since the second case is impossible, we deduce that $s\ne a,b$ and $s\sim a,b$.
Consequently, we obtain two $4$-cycles $u'csa$ and $w'csb$, which cannot be induced. If $s \sim u', w'$ we get $s\in I(u,v)\cap I(v,w)$,
contradicting that $uvw$ is a metric triangle. Thus we can assume without loss of generality that $c \sim b$. Therefore the vertices $a$
and $b$ belong to the ball $B_2(z')$. Since $d(z',v)=3$, the convexity of $B_2(z')$ implies that $a\sim b$. Then we also
have $a\sim c$ as the unique diagonal of the cycle $abcu'$, finishing the proof. \end{proof}
The following lemma will be useful in the next section. It shows that the conclusions of Lemma \ref{lem: equilateral1} hold under weaker
hypotheses.
\begin{lemma}
\label{lem: equilateral2}
Let $uvw$ be an equilateral metric triangle of size $k\ge 3$ of a CB-graph $G$. Pick three vertices $a,b,w'$ such that $a\in I(v,u)$, $w'\in I(v,w)$ and $a\sim v$,
$d(w',v)=2$, and $b\sim v,w'$. If $d(a,w')\le 2$, then for any common neighbor $c$ of $a$ and $w'$, either $c=b$ or the vertices $a,b,c$ are pairwise adjacent.
\end{lemma}
\begin{proof}
Since $uwv$ is strongly equilateral, we must have $d(w,a)=k$, whence $d(a,w')=2$. In particular, if $a \sim b$, then $b\sim c$ as the unique diagonal
of the 4-cycle $cabw'$. Hence, it suffices to show that $a\sim b$. Since $d(w,c)\le k-1$ and $uvw$ is strongly equilateral, necessarily $c\notin I(u,v)$, hence $d(u, c) \in \sg{k-1, k}$.
\medskip\noindent
{\bf Case 1:} $d(u,c)=k$.
\begin{proof}
Since $d(u,w')=k$, we can apply $\mathrm{TPC}^1(u,cw')$ to the edge $cw'$. If $\mathrm{TC}(u,cw')$ applies, then there exists a vertex $z\sim c,w'$ at distance $k-1$ from $u$.
Since $d(a,w')=2$, necessarily $z\ne a$. By $\mathrm{INC}(u)$ we have $a\sim z$ and there exists some $z'\sim a,z$ at distance $k-2$ from $u$. Hence $z'\in I(v,u)$ has distance 2 to
$v$ and $w'$. Analogously, if $\mathrm{PC}^1(u,cw')$ applies with respect to the neighbor $a\in I(c,u)$ of $c$, then we can find a pentagon
$caz'zw'$ with $z'$ at distance $k-2$ from $u$. In this case also $z'\in I(a,u)\subset I(v,u)$. In both cases we have $a\sim z',v$ and $b\sim w',v$ and $z\sim z',w'$, thus
we can apply Lemma \ref{lem: equilateral1} with $z'$ playing the role of $u'$ and $z$ playing the role of $c$ to deduce that $a,b,c$ are pairwise adjacent.
\end{proof}
\medskip\noindent
{\bf Case 2:} $d(u,c)=k-1$.
\begin{proof}
Since $d(u,a)=k-1$, we can apply $\mathrm{TPC}^0(u,ac)$. If $\mathrm{TC}(u,ac)$ applies, then there exists $z\sim a,c$ at distance $k-2$ from $u$. Since $z\in I(v,u)$ and $d(v,z)=d(z,w')=2$, we can apply Lemma \ref{lem: equilateral1} with $z$ playing the role of $u'$ to conclude that $a,b,c$ are pairwise adjacent. So suppose that $\mathrm{PC}^0(u,ac)$ applies and there exist a pentagon $au'z'zc$ with $d(u,u')=d(u,z)=k-2$ and $d(u,z')=k-3$. Then we are in position to apply Lemma \ref{lem: deuxpent} to the two pentagons $zz'u'ac$ and $acw'bv$. Since $d(v,z')=3$, one of these two pentagons has a universal vertex (i.e., a vertex adjacent to all vertices of the respective pentagon). Since we can suppose that $\mathrm{TC}(u,ac)$ does not apply, the pentagon $zz'u'ac$ cannot contain a universal vertex. Thus there exists a vertex $p$ adjacent to all vertices of the pentagon $acw'bv$. Since $u',z\in B_2(p)$ and $z'\in I(u',z)$, the convexity of $B_2(p)$ implies that $z'\in B_2(p)$. Consequently, $p\in I(v,z')\cap I(v,w')\subset I(v,u)\cap I(v,w)$, contrary to the assumption that $uvw$ is a metric triangle.
\end{proof}
In both cases we either obtained a contradiction or the desired conclusion that $a\sim b$.
\end{proof}
\section{Helly theorem}
\label{sec: helly}
In this section, we prove a Helly theorem for convex sets in CB-graphs. We follow the method from \cite{BaCh95} for weakly modular graphs
and we show how to adapt it to graphs with convex balls, thanks to the results of previous section. In what follows, $G$ will always be a graph
with convex balls.
A finite subset of vertices $A\subseteq V$ is \emph{Helly independent} (\emph{h-independent} for short) if
$$\bigcap_{a\in A}\mathrm{conv}\left(A\backslash \sg{a}\right) = \varnothing.$$
The \emph{Helly number} $h(G)$ of $G$ is the supremum of the size of an $h$-independent set of vertices of $G$ \cite{vdV}. Equivalently, $h(G)$ is the smallest integer $k$ such that any finite family $\mathcal F$ of convex sets of $G$ has a nonempty intersection if and only if any subfamily of $\mathcal F$ with $k+1$ members has a nonempty intersection. For a positive integer $k$, let $h_k(G)$ be the supremum of the size of an $h$-independent set of $G$ of diameter at most $k$. Clearly $h_1(G)$ is just the clique number $\omega(G)$ of $G$.
A subset of vertices $S\subseteq V$ of $G$ is called a \emph{simplex} \cite{BaCh95} if any three vertices form an equilateral metric triangle and
for any four vertices $u,v,w,x\in V$, we have $I(u, w)\cap I(v,x)=\varnothing$. Let $\sigma(G)$ be the supremum of the size of a simplex of $G$ and $\sigma_k(G)$ be the supremum of the size of a simplex of diameter at most $k$.
It is easy to see that $h(G)$ and $\sigma(G)$ are always lower bounded by the clique number $\omega(G)=h_1(G)$ of $G$ and upper
bounded by the Hadwiger number $\eta(G)$ (the size of the largest complete graph that can be obtained by contracting the edges of $G$) and that $h(G)$ and $\sigma(G)$ are not comparable in general. However, it is shown
in \cite{BaCh95} that $h(G)\leq \sigma(G)$ holds for graphs with equilateral metric triangles and that
$h(G)=\sigma(G)=\omega(G)=h_1(G)$ holds for weakly modular graphs. We adapt their proof scheme to establish that $h(G)=h_2(G)$
for graphs with convex balls. Notice that in graphs with convex balls, $h(G)$ may be much larger than the clique number. For example, the Petersen graph
has Helly number 4 (see Fig. \ref{fig: Petersen} for an $h$-independent set of size 4) and clique number 2.
\tikzexternaldisable
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1]
\tikzstyle{every node}=[draw,circle,fill=black,minimum size=4pt,
inner sep=0pt]
\node (1) at (90:2) [color=red] {};
\node (2) at (18:2) [color=red] {};
\node (3) at (306:2) {};
\node (4) at (234:2) {};
\node (5) at (162:2) {};
\node (6) at (90:1) {};
\node (7) at (18:1) {};
\node (8) at (306:1) [color=red] {};
\node (9) at (234:1) {};
\node (10) at (162:1) [color=red] {};
\draw[thick] (1) -- (2) -- (3) -- (4) -- (5) -- (1) -- (6) -- (8) -- (3);
\draw[thick] (2) -- (7) -- (9) -- (4);
\draw[thick] (8) -- (10) -- (5);
\draw[thick] (6) -- (9);
\draw[thick] (10) -- (7);
\end{tikzpicture}
\caption{The red vertices form an $h$-independent set of size $4$ of the Petersen graph.} \label{fig: Petersen}
\end{figure}
The main result of this section shows that the Helly number of a CB-graph can be defined locally:
\begin{theorem}
\label{thm: hellydiam2}
Let $G$ be a CB-graph. Then $h(G)=h_2(G)$.\end{theorem}
The proof of Theorem \ref{thm: hellydiam2} is based on three lemmas. First we recall the following general result:
\begin{lemma}\cite[Lemma 5]{BaCh95}
\label{lem: LBaCh}
Let $A\subseteq V$ be an $h$-independent set of a graph $G$. If $x \in I(u,v)\cap \mathrm{conv}\left(A\backslash \sg{u}\right)$ for some vertices $u,v\in A$,
then the set $B:=\left(A\backslash\sg{v} \right)\cup \sg{x}$ is $h$-independent with $|B|=|A|$. In particular, $B$ is $h$-independent when
$x$ is chosen from $I(u,v)\cap I(v,w)$ for distinct $u,v,w\in A$.
\end{lemma}
For the sake of simplicity, by a \emph{distance-minimal set} we mean any set $A\subseteq V$ such that for any two distinct $u, v \in A$ we have
$I(u,v)\cap \mathrm{conv}\left(A\backslash \sg{u}\right) = \sg{v}.$ We will use the operation of Lemma \ref{lem: LBaCh} to
transform $h$-independent sets into $h$-independent sets with ``nicer properties''.
\begin{lemma}\label{lem: hleqsigma}
Any finite $h$-independent set $A$ of a CB-graph $G$ can be
transformed to a distance-minimal $h$-independent simplex of $G$ or to an
$h$-independent set of diameter at most $2$. In particular,
$h(G)\leq \max\{ h_2(G), \sigma(G)\}$.
\end{lemma}
\begin{proof}
For any vertex $u$ of $A$, let $\Delta(u,A):= \sum_{v\in A\backslash \sg{u}} d(u,v).$ Let also
$\Delta(A):=\min_{z\in A}\Delta(z,A).$ Since $A$ is finite, these values are well defined. We call any vertex $z\in A$ such
that $\Delta(z,A)=\Delta(A)$ \emph{minimal}. Pick any minimal vertex $z\in A$ and we make any other vertex of $A$ as close as possible to $z$.
Namely, if there exists some $u\in A\backslash \sg{z}$ such that $I(u,z)\cap \mathrm{conv}(A\backslash \sg{z})\neq \sg{u}$, then we
pick any vertex $x\ne u$ from this intersection and set $B:= \left(A\backslash \sg{u}\right)\cup \sg{x}$. By Lemma \ref{lem: LBaCh},
the set $B$ is $h$-independent. Notice also that $\Delta(B)\leq \Delta(z, B)<\Delta(z,A)=\Delta(A)$. Hence, if we set $A:=B$ and
apply this transformation to minimal vertices of $A$ as long as we decrease $\Delta(A)$, after a finite number of steps we will
end up with an $h$-independent set $S$ of the same cardinality as the initial set $A$ and such that
$I(u,z)\cap \mathrm{conv}(A\backslash\sg{z})=\sg{u}$ holds for any minimal vertex $z$ of $S$. This
implies that $I(u,v)\cap I(u,z)=\{ u\}$ and $I(v,u)\cap I(v,z)=\{ v\}$ for any three distinct vertices $u,v,z$ of $S$
such that $z$ is minimal. We distinguish two cases.
\medskip\noindent
{\bf Case 1:} $z\sim u$ for a minimal vertex $z$ of $S$ and a vertex $u\in A\backslash \sg{z}$.
\begin{proof}
By construction, for any $v\in S\backslash\sg{z,u}$, we have $I(v,z)\cap I(v,u)=\sg{v}$ and $I(u,z)\cap I(u,v)=\sg{u}$. Moreover, since $S$ is $h$-independent, we have $u\notin I(v,z)$. Consequently, $zuv$ is a metric triangle of $G$. By Theorem \ref{thm: metrictriangle}, $zuv$ is either a graphic triangle or it lies on a pentagon, i.e., $d(z,uv)=d(u,v)=2$. This implies that $S\subseteq B_2(z)$ and that $B_1(z)\cap S$ form a clique. We assert that $\mathrm{diam}(S)\leq 2$. Pick any $v, w \in S\backslash \sg{z}$.
If $d(z,v)=2$ and $w\sim z$, then we saw that $d(v,w)=2$. Now, let $d(z,v)=d(z,w)=2$. If we assume that $d(v,w)\geq 3$, then by Theorem \ref{thm: metrictriangle} $zvw$ is not a metric triangle, by the property of the set $S$ there exists a vertex $y \in I(z, v)\cap I(z,w)$ different from $z$. Since $d(z,v)=d(z,w)=2$, $y$ must be adjacent to $v$ and $w$, contrary to the assumption $d(v,w)\ge 3$. Hence $d(v,w)\le 2$, establishing that $S$ has diameter at most 2. Thus in this case we have $|A|=|S|\leq h_2(G)$.
\end{proof}
\medskip\noindent
{\bf Case 2:} $z\nsim u$ for any minimal vertex $z$ of $S$ and any vertex $u\in A\backslash \sg{z}$.
\begin{proof}
We assert that in this case $S$ is a distance-minimal simplex. The second condition in the definition of a simplex is always true in $h$-independent sets, hence we only need to show that any three vertices of $S$ form an equilateral metric triangle. Pick any triplet $z,u,v$ of vertices of $S$ such that $z$ is minimal. We already know that $I(u,v)\cap I(u,z)=\{ u\}$ and $I(v,u)\cap I(v,z)=\{ v\}$. Let $y \in I(z,u)\cap I(z,v)$ such that $I(y,u)\cap I(y,v)=\sg{y}$. Then clearly $yuv$ is a metric triangle. We assert that $d(u,v)\le \min \{d(z,u), d(z,v)\}$ and that $d(u,v)<\min\{ d(z,u),d(z,v)\}$ if $z\ne y$ or $z=y$ and $d(z,u)=d(z,v)=2, d(u,v)=1$. By Theorem \ref{thm: metrictriangle}, the metric triangle $yuv$ is equilateral or has type $(2,2,1)$. If $yuv$ is equilateral, then $d(z,u)=d(z,v)=d(u,v)$ if $z=y$ and $d(z,u)=d(z,v)>d(u,v)$ if $z\ne y$. Now assume that $yuv$ has type $(2,2,1)$. If $d(u,v)=1$, then $d(z,u)=d(z,v)\ge d(y,u)=d(y,v)=2$ and we are done. Now suppose that $d(u,v)=2$. Then $y$ is adjacent to one of the vertices $u,v$, say $y\sim v$. Since $z$ is not adjacent
to any vertex of $S\setminus \{ z\}$, we must have $z\ne y$. We assert that $\Delta(v,S)<\Delta(z,S)$. First notice that $d(u,v)=2 < d(u,y)+d(y,z)=d(u,z)$.
For any $w \in S\backslash \sg{z,u,v}$, let $y'\in I(z,v)\cap I(z,w)$ be such that $y'vw$ is a metric triangle. By Theorem \ref{thm: metrictriangle}, either $y'vw$ is equilateral, in which case $d(v,w)=d(y', w)\leq d(z,w)$, or $y'vw$ has type $(2,2,1)$ and, since $z\notin B_1(S\backslash\sg{z})$, we must have $d(v,w)\leq 2 \leq d(z,w)$.
Consequently, $\Delta(v,S)< \Delta(z,S)$, contrary with the minimality choice of $z$. This shows that $d(u,v)\le \min\{ d(z,u), d(z,v)\}$ and that $d(u,v)<\min\{ d(z,u), d(z,v)\}$ holds if $z\ne y$ or $z=y$ and $zuv$ has type $(2,2,1)$. Since $\Delta(z)\le \Delta(u)$, all inequalities $d(u,v)\le d(z,u)$ are equalities. This proves that $z=y$ and that each metric triangle $zuv$ is equilateral. Consequently, all vertices of $S$ are minimal, yielding that all triplets $u,v,w$ of $S$ define equilateral metric triangles of size $k\ge 2$. In particular, $S$ is a simplex, whence $|S|\leq \sigma(G)$.
\end{proof}
Consequently, in Case 1 $S$ is an $h$-independent set of diameter at most 2 and in Case 2 $S$ is an $h$-independent distance-minimal simplex.
\end{proof}
Now we show that the simplex from the second case of the previous proof can be reduced to a clique of $G$ of the same size.
\begin{lemma}
\label{lem: h-indepdiam2}
Let $S\subseteq V$ be a distance-minimal $h$-independent simplex of a CB-graph $G$ and let $\mathrm{diam}(S)\ge 3$. Then $\mathrm{conv}(S)$
contains a clique $C$ of size $|S|$.
\end{lemma}
\begin{proof} The case $|S|=2$ is trivial. Thus, let $|S|\geq 3$. Let $k:=\mathrm{diam}(S) \geq 3$, which corresponds to the distance between any two distinct vertices of $S$. Pick $u,v\in S$ and let $x$ be a neighbor of $v$ in the interval $I(u,v)$. For every $w\in S\backslash \sg{u, v}$, we can apply $\mathrm{TPC}^0(w,vx)$. We partition $S\backslash \sg{u,v}$ into the sets $S_1$ and $S_2$: $S_1$ is the set of all vertices $w$ of $S\backslash \sg{u,v}$ such that $\mathrm{TC}(w,vx)$ applies and $S_2:= S\setminus S_1$ (notice that for all $w\in S_2$ the condition $\mathrm{PC}^0(w,vx)$ applies).
\begin{claim} \label{cl:empty}
The set $S_2$ is empty.
\end{claim}
\begin{proof}
Suppose by way of contradiction that $w\in S_2$. Let $z$ be a vertex at distance 2 from $v$ and $x$ such that $z\in I(v,w)\cap I(x,w)$.
Let $t$ be a common neighbor of $z$ and $v$. Then we can apply Lemma \ref{lem: equilateral2} with the vertices $u,v,w,t,z,x$ playing the role of $u,v,w,b,w',a$, respectively.
Thus we get that $t\sim x$. This implies that $\mathrm{TC}(w,vx)$ applies, contrary to the assumption that $w\in S_2$. Hence $S_2= \varnothing$.
\end{proof}
Let $A_1$ be the set of all common neighbors $z$ of $v$ and $x$ such that there exists a vertex $w\in S_1$ such that $z\in I(v,w)\cap I(x,w)$. We call $z$ an \emph{imprint} of $w$ on the edge $vx$.
\begin{claim}
\label{clm: S1'}
The set $C:=A_1\cup \{ v,x\}$ induces a clique of size at least $|S|$.
\end{claim}
\begin{proof}
Pick any two vertices $z_1, z_2$ in $A_1$ and suppose that $z_1,z_2$ are imprints of the vertices $w_1,w_2\in S_1$, respectively. If $w_1=w_2$, then $z_1\sim z_2$
by $\mathrm{INC}^0(w_1)$. Thus assume that $w_1\neq w_2$. Assume by way of contradiction that $z_1\nsim z_2$. This implies that $x \in I(z_1, z_2)$. Since $z_1, z_2\in I(w_1, v)\cup I(w_2, v)\subseteq \mathrm{conv}(S\backslash \sg{u})$, we obtain that $ x\in \mathrm{conv}(S\backslash \sg{u})\cap I(u, v).$
This contradicts the fact that the set $S$ is distance-minimal. This establishes that $A_1$ induces a clique. Since all vertices of $A_1$ are different from $v,x$ and are adjacent to $v$ and $x$, $C=A_1\cup \{ v,x\}$ also induces a clique.
From the definition of $S_1$ and since any three vertices of $S$ form a metric triangle it follows that any two vertices of $S_1$ have different imprints, thus $|A_1|\geq |S_1|$.
By Claim \ref{cl:empty} the set $S_2$ is empty, yielding $|A_1|\geq |S_1|=|S\backslash\sg{u,v}|=|S|-2.$ Consequently, $|C|\geq |S|$.
\end{proof}
The desired result is now a direct consequence of Claim \ref{clm: S1'}.
\end{proof}
We now prove the equality $h(G)=h_2(G)$. Obviously, $h(G)\ge
h_2(G)$. We now establish the converse inequalitiy. Given a finite
$h$-independent set $A$ of size $k$, by Lemma \ref{lem: hleqsigma},
$A$ can be transformed to an $h$-independent set $S$ of the same size
and which either has diameter at most $2$ or is a distance-minimal
simplex. In the second case, by Lemma \ref{lem: h-indepdiam2}, we
find a clique $C\subseteq\mathrm{conv}(S)$ of the same size as $S$ and
$A$. Since $C$ is an $h$-independent of diameter $1$, in both cases,
we have found an $h$-independent of diameter at most $2$ and of size
$k$. This establishes that $h(G) \leq h_2(G)$ when $h(G)$ is finite.
If $h(G)$ is infinite, then for each integer $k$ there exists an
$h$-independent set $A_k$ of size $k$. By the previous argument, each
$A_k$ can be transformed into an $h$-independent $S_k$ of size $k$ and
of diameter at most $2$. This implies that $h_2(G)$ is also infinite and thus $h(G) = h_2(G)$. This finishes the proof of Theorem \ref{thm: hellydiam2}.
\medskip
We were not able to prove that any $h$-independent set of diameter 2 can be transformed into an $h$-independent simplex of diameter 2 and of the same size.We formulate this as an open question:
\begin{question} Is it true that for any CB-graph $G$, the equality $h_2(G)=\sigma_2(G)$ holds?
\end{question}
\section*{Acknowledgements}
J.C.\ and V.C.\ were supported by ANR project DISTANCIA (ANR-17-CE40-0015). U.G. was supported by an internship grant of ``Pole Calcul" of LIS, Aix-Marseille Université and
by ANR project DISTANCIA (ANR-17-CE40-0015).
\bibliographystyle{plain}
|
\section{Introduction}
An atom-chip \cite{folman, folman1, den,fort, reichel, eckel, Reichel1} with micron sized wire structure patterns on a reflecting surface provides platform for trapping and manipulation of atoms on miniaturized scale. The manipulation of cold atoms on a miniaturized scale using atom-chip has been demonstrated in a number of ways including atom traps \cite{reichel}, atom guides\cite{muller, Dekker}, beam splitters \cite{casset, machluf} and Bose-Einstein condensations (BECs) \cite{hansel, Farkas, Ott}. Usually, a magneto-optical trap (MOT) on atom-chip is prepared by reflecting the MOT-beams (cooling and re-pumping laser beams) on atom-chip surface in presence of a quadrupole magnetic field produced by the bias coils (outside the chamber) and an U-shaped wire (kept behind the atom-chip). This MOT is usually called "U-MOT". After initial loading of the U-MOT, the quadrupole magnetic field and cooling laser beam detuning are increased which brings the atom cloud closer to chip surface but with reduced temperature and increased density. This MOT is called grey-MOT (G-MOT). Typically the cold $^{87}Rb$ atoms in a MOT cloud are equally distributed among different magnetic sub-states of hyperfine state F=2. For the magnetic trapping, however, the state $\ket{F=2, \, m_{F} = + 2}$ is a preferred state. Therefore, maximum number of atoms are to be prepared in this state to maximize the number of atoms in the magnetic trap \cite{mishra}. The presence of other states in the MOT atom cloud, particularly un-trappable states $\ket{F=2, \, m_{F} = -2,-1,0}$, results in a loss of the cold atoms from the magnetic trap. This loss can be as large as $\sim$ 60 $\%$ of the atoms from the MOT cloud \cite{david}. Thus, the preparation of atoms in the trappable state using optical pumping is an important step before the magnetic trapping.\\
The efficiency of optical pumping of atoms to the trappable state ($\ket{F=2, \, m_{F} = +2}$) can be quantified by applying Stern-Gerlach technique which can separate atoms in different Zeeman hyperfine states \cite{david}. In this work, we have used this SG technique to employ the Zeeman state dependent magnetic force (Stern-Gerlach force) on cold $^{87}Rb$ atoms in a MOT to spatially separate and detect them in different Zeeman hyperfine states. In our experiments, by varying the temperature of the cold atom cloud and duration of Stern-Gerlach magnetic field, the spatial separation of cold atoms in the different Zeeman states was optimized to accurately detect atoms in these states. This has allowed to estimate accurately the optical pumping efficiency in our experiments for pumping atoms to a trappable Zeeman hyperfine state ($\ket{F=2, \, m_{F} = +2}$). After optimizing the power and pulse duration of optical pumping pulse, $\sim 92 \% $ of atoms from a G-MOT on the atom-chip were transferred to trappable state $\ket{F=2, \, m_{F} = +2}$. The optically pumped atoms were trapped in external Ioffe-Pritchard (IP) magnetic trap near atom-chip surface formed by magnetic fields due to a current carrying Z-shaped copper wire and bias coils.\\
\section{Experimental setup}
\begin{figure}[ht]
\centering
\includegraphics[width=9.0 cm]{Fig1.png}
\caption{ Schematic of the atom-chip setup for MOT formation, with two beams in reflection geometry forming four MOT beams in yz-plane, two counter propagating MOT beams in $\pm$ x-axis (not shown in figure), a U-shaped copper wire and bias coils (y-direction bias coils shown and z-direction coils are not shown). The dotted curves show the magnetic field lines. The MOT is formed $\sim$ 6 mm below the atom-chip surface. }
\end{figure}
The schematic of the experimental setup used in this work is shown in Fig. 1. The details of the experimental setup can be found in our earlier published works \cite{vivek1, vivek2}. The experimental setup consists of an octagonal vacuum chamber which is connected to vacuum pumps through a six way cross. A turbo molecular pump (TMP) (77 l/s), a sputter ion pump (SIP) (300 l/s) and a titanium sublimation pump (TSP) are the part of vacuum system used for creating an ultrahigh vacuum (UHV) with a base pressure of $\sim$ 2.0 $\times 10^{-10}$ Torr in the chamber. The Rb-dispensers are fixed on the UHV side rods of a feed-through and this feed-through is inserted in the chamber such that dispensers are placed at a distance of $\sim$ 17 cm from the center of the octagonal chamber. The Rb vapor is generated in the chamber when a current of $\sim$ 3 A is passed through the feed-through rods. The Rb-vapor generation results in increase in the pressure in the UHV chamber to $\sim$ 3.5 $\times 10^{-10}$ Torr.\\
The atom-chip is mounted on a special mount and feed-through system. This is inserted inside the UHV chamber such that atom-chip active surface is facing vertically downwards (along z-axis in Fig.1). Four independent MOT beams (cooling laser beams mixed with re-pumper laser beams) are used to form U-MOT with quadrupole field produced by U-shaped external copper wire (cross-section: 1 mm $\times$ 10 mm ) and bias coils, as shown in Fig. 1. On the back surface of atom-chip, on the top of copper U-wire, a Z-shaped copper wire (cross section : 1 mm $\times$ 1 mm) is kept to produce Stern-Gerlach magnetic field. This wire is also used for magnetic trapping of atoms in Ioffe-Pritchard (IP) magnetic trap. The field due to this Z-wire and field due to the bias coils provide the net field required for an external Ioffe-Pritchard (IP) magnetic trap near the chip surface. The 2f- fluorescence imaging technique \cite{serre} has been used to image the cold atom cloud in x-z plane using a digital CCD camera.\\
\begin{figure}
\centering
\includegraphics[width=9.0 cm]{Fig2.png}
\caption{ CCD images of cold $^{87}Rb$ atoms in (a) U-Magneto-optical trap (U-MOT) and (b) Grey-magneto optical trap (G-MOT). }
\end{figure}
A FPGA based electronic controller is used to generate digital as well as analog pulses to control various equipment like power supplies, acousto-optic modulators, CCD camera, mechanical shutters etc to implement various experimental events, from MOT-loading to magnetic trapping and cloud imaging, in a sequential manner. The first step in this sequence is to load U-magneto-optical trap (U-MOT) for $^{87}Rb$ atoms. The duration of this satge is $\sim$ 20 s. A current of 60 A is flown in U-wire and bias fields of $\sim$ 10 G in y-direction and $\sim$ 1.5 G in z-direction is applied during U-MOT formation. The cooling laser frequency is kept - 14 MHz red-detuned from the cooling transition of $^{87}Rb$. The U-MOT cloud is formed $\sim$ 6 mm from the atom-chip surface, as shown in Fig. 2 (a). The number and temperature of cold atoms in U-MOT were $\sim 1 \times 10^{8}$ and $\sim 280 \mu K$ respectively. For magnetic trapping of atoms, the atom cloud from U-MOT needs to be further cooled and to be placed near the atom-chip surface. This can be achieved by performing G-MOT stage after U-MOT is loaded fully. To achieve G-MOT after U-MOT, the current in U-wire is ramped (in 100 ms) from 60 A to 80 A and current in Y-bias coil is ramped (in 100 ms) from from 15 A to 40 A. The cooling laser frequency is further detuned from -14 MHz to -22 MHz. In G-MOT, atom cloud is moved vertically in z-direction to a position $\sim$ 2 mm from the chip surface as shown in Fig. 2(b). The number and temperature of cold atoms in G-MOT are $\sim 7 \times 10^{7}$ and $\sim 510\mu K$ respectively.
\begin{figure}
\centering
\includegraphics[width=8.0 cm]{Fig3.png}
\caption{ The variation in number (squares) and temperature (circles) of cold atoms in the G-MOT with frequency detuning of the cooling laser beam.}
\end{figure}
The temperature of the G-MOT cloud was further reduced by varying the cooling laser frequency from - 22 MHz to - 74 MHz in time duration of 10 ms. The cooling laser beam power was observed to decrease linearly with detuning at the rate $\sim$ 0.25 mW/MHz due to spectral response of AOM. As a result, the temperature of the cold atom cloud reduced from $\sim 510 \mu K$ to $\sim 77 \mu K$. However, the number of cold atoms ($\sim 5.5 \times 10^{7}$) in G-MOT remains nearly unchanged as shown in Fig. 3.\\
\begin{figure}
\centering
\includegraphics[width=9.0 cm]{Fig5.png}
\caption{ Schematic of the sequence of various events from U-MOT formation to optical pumping and imaging of atom cloud in Stern-Gerlach (SG) experiments. }
\end{figure}
Fig. 4 shows the experimental time sequence used for performing optical pumping experiments in presence of SG magnetic field. At the end of G-MOT stage, all the magnetic fields and laser fields are switched-off for the duration 1.5 ms. Then optical pumping beam, Y-bias field, current in Copper Z-wire (i.e. SG field gradient) and imaging beam are switched-on in a particular sequence. The current in Copper Z-wire (90 A) is switched-on after the optical pumping of atoms, before imaging of atom cloud using fluorescence imaging technique.
\section{Stern-Gerlach Force}
In presence of a magnetic field gradient, the atoms in different Zeeman hyperfine state ($\ket{F, \, m_{F}}$) experience a force which is dependent on $m_{F}$. This force is called Stern-Gerlach (SG) force and is given as \cite{kafer} ,
\begin{equation}
F_{z}(m_{F})=\mu_{B} \, g_{F} \, m_{F} \, \dfrac{dB}{dz}
\end{equation}
where $\mu_{B}$ is Bohr magneton, $g_{F}$ is lande's g factor, $m_{F}$ is magnetic hyperfine angular momentum and dB/dz is magnetic field gradient (i.e. SG field gradient) along z-direction .\\
Owing to the above equation, because of state dependent acceleration, the atoms in different $m_{F}$ states in an atom cloud are expected to separate out according to their $m_{F}$ values, when they evolve in a SG magnetic field gradient \cite{stamper, marechal}. Thus the SG effect can be used to spatially separate (and detect) the population of atoms in different $m_{F}$ states. This effect has been utilized in our optical pumping experiments to identify the population in different $m_{F}$ states in an atom cloud of $^{87}Rb$ atoms cooled in F=2 hyperfine state. By accurately measuring population in different $m_{F}$ levels, the optical pumping of atoms was maximized to $m_{F}$=2 state.
\begin{figure}
\centering
\subfigure[]{
\resizebox*{3 cm}{!}{\includegraphics{Fig4a.png}}}\hspace{0pt}
\subfigure[]{
\resizebox*{5 cm}{!}{\includegraphics{Fig4b.png}}}
\caption{ (a) The simulated current density distribution in copper Z-wire is shown with color code. (b) The variation of magnetic field due to flow of current in copper Z-wire along z-direction (0,0,z). The Z-wire dimensions at Z-joint are : 1 mm width along y-direction, 1 mm thickness along z-direction, and 4.0 mm length along x-direction). The coordinates of mid point of Z-joint are (0,0,0). }
\label{sample-figure}
\end{figure}
In our experiments, the state dependent SG force on cold atoms from G-MOT was applied by flowing a dc current in a copper Z-wire placed behind the atom-chip. The current density variation in copper Z-wire due to flow of 90 A current is shown in Fig. 5(a). Using these current density values, the variation in magnetic field along z-direction, due to the Z-wire (carrying current of 90 A), is shown in Fig. 5(b). It can be noted here from the figure that field gradient at the G-MOT cloud position (which is $\sim$ 2.6 mm below the Copper Z-wire) is $(17.61 \pm 0.19)$ G/mm. \\
\section{Results and discussion}
We have investigated the effect of two experimental parameters, namely, SG magnetic field pulse duration and temperature of cold atom cloud, on the $m_{F}$ state dependent splitting of cold atom cloud. Both these parameters are important in SG experiment to separate and detect group of atoms with different $m_{F}$ states (i.e. Zeeman magnetic states) in the atom cloud. These results are discussed as follows.
\subsection{Effect of SG field pulse duration}
\begin{figure}[ht]
\centering
\includegraphics[width=9.0 cm]{Fig6.png}
\caption{ CCD image of the cold $^{87}Rb$ atom cloud after different values of duration $\Delta t_{SG}$ of SG field :(a) 4 ms, (b) 6 ms, and (c) 8 ms.}
\end{figure}
The pulse duration of SG field was varied in the experiment and the atom cloud was imaged after a time of flight of 5 ms after evolution in the magnetic field gradient. Fig. 6 shows the CCD images of the cold $^{87}Rb$ atoms for different values of SG magnetic field pulse duration used in the experiment. It is evident from the Fig. 6 that as pulse duration was increased from 4 ms to 8 ms, the relative separation between the group of cold atoms in different $m_{F}$ state was also increased. For small pulse duration (i.e. 4 ms), the separation between atoms in different $m_{F}$ states was not good enough to resolve the atom clouds clearly (Fig. 6(a). With increase in SG field pulse duration (from 4 ms to 6 ms and 8 ms), the separation between atom clouds in different $m_{F}$ states was increased and clouds in different states were clearly resolved (Fig. 6 (b) and (c)). The cold atoms in $m_{F}$ = +2 state experience maximum SG force along z-axis. Therefore, this group of atoms is found at the farthest position along z-direction in the CCD image. The cold atoms in $m_{F}$ = 0 state does not experience any SG force. Therefore, this group of atoms is not affected by SG field and its position remains nearly same in all the images ((a), (b) and (c)). Since atom clouds with negative $m_{F}$ values would move in negative direction of z-axis, these clouds may be lost after colliding with atom-chip surface. That is why we do not observe any cloud for state $m_{F}$ = -2 in the images in Fig.6. \\
\begin{figure}
\centering
\includegraphics[width=6.0 cm]{Fig7.png}
\caption{ The measured variation in peak to peak separation ($\Delta z$) between the cold atoms in $m_{F} = + 1$ and in $m_{F} = + 2$ states as a function of SG field pulse duration for a fixed value (90 A) of current in Copper Z-wire. }
\end{figure}
Fig. 7 shows the experimentally measured separation ($\Delta z$) between the cold atoms in $m_{F} = + 1$ and in $m_{F} = + 2$ states as a function of SG field pulse duration ($\Delta t_{SG}$), at a fixed value of magnetic field gradient ($\sim$ 17.6 G/mm) due to $\sim$ 90 A current in copper Z-wire. The separation between cold atom clouds in two states ($\Delta z$) increases with ($\Delta t_{SG}$). This relative separation ($\Delta z$) increases from $\sim$ 1.2 mm to $\sim$ 2.0 mm as SG pulse duration increases from 3 ms to 9 ms respectively as shown in Fig. 7. It was observed that for $\Delta t_{SG}$ = 6 ms, the separation of atoms in different $m_{F}$ states was clearly resolved. Therefore, this pulse duration was used in the subsequent experiments.
\subsection{Effect of temperature of atom cloud}
In SG experiments, we have also studied the effect of temperature of atom cloud in G-MOT on spatial splitting of atom cloud in different parts corresponding to different $m_{F}$ states. The temperature in G-MOT was changed by changing the cooling laser beam detuning. The SG pulse duration of 6 ms and a current of 90 A in copper Z-wire was used to apply the SG force the atom cloud. The images of atom cloud were taken after 5 ms duration in time of flight after the SG field pulse. Fig. 8 shows the CCD image of the $^{87}Rb$ atom cloud after splitting in different $m_{F}$ states for different initial temperature of cloud in G-MOT.
\begin{figure}[ht]
\centering
\includegraphics[width=8.0 cm]{Fig8.png}
\caption{ The images of split atom cloud of $^{87}Rb$ atoms after SG splitting. The temperature of cloud in G-MOT was kept at different values for these images : (a) $\sim$ for 510 $\mu K$, (b) $\sim$ 210 $\mu K$, (c) $\sim$ 130 $\mu K$, (d) $\sim$ 77 $\mu K$. The SG pulse duration was 6 ms at 90 A current in copper Z-wire, and atom cloud images were taken after 5 ms of time of flight after the SG field pulse.}
\end{figure}
It is evident from images in Fig.8 that lower temperature of initial atom cloud improves the spatial resolution of cloud split in different states. The higher initial temperature of cloud from G-MOT results in faster expansion of cloud in SG field, which keeps the split parts overlapped as shown in Fig. 8(a). The group of atoms in different $m_{F}$ states after SG splitting are better resolved at lower initial temperature of cloud (e.g. Fig. 8 (c) and (d)). \\
\subsection{Optimization of optical pumping process}
The SG state detection technique has been finally utilized to maximize the optical pumping efficiency for preparing the $^{87}Rb$ atom cloud in the trappable Zeeman hyperfine state ($\ket{ F=2, \, m_{F} = + 2}$ ). All magnetic fields and laser beams were switched-off after G-MOT stage for a duration of 1.5 ms. A right handed circularly polarized ($\sigma ^{+}$) optical pumping beam pulse resonant to $D_{2}$-line transition $F=2 \xrightarrow{} F'=2$ was directed along y-direction to pump atom cloud to Zeeman hyperfine state ($ \ket{F=2, m_{F} = + 2}$ ) in presence of an axial magnetic field of $\sim$ 3.6 G. A re-pumper laser beam pulse resonant to $D_{2}$-line transition $F=1 \xrightarrow{} F'=2$ was also overlapped with the optical pumping beam to avoid accumulation of atoms in F=1 state. Atoms are expected to be in ($\ket{ F=2, \, m_{F} = + 2}$ ) state after application of optical pumping pulse. After optical pumping, SG state detection was performed by applying a current pulse of 90 A (6 ms) to copper Z-wire to generate Stern-Gerlach (SG) force on atoms. Then after 5 ms of time of flight, the cloud was imaged by fluorescence technique to detect the population of atoms in different Zeeman states.\\
\begin{figure}[ht]
\centering
\includegraphics[width=9.0 cm]{Fig9.png}
\caption{ The CCD images of atom cloud under Stern-Gerlach (SG) force showing the distribution of population of $^{87}Rb$ atoms in different $m_{F}$ state: (a) without optical pumping, (b) with optical pumping pulse of duration $\sim$ 300 $\mu s$ , and (c) with optical pumping pulse of duration $\sim$ 500 $\mu s$. The atom clouds in different $m_{F}$ states ($m_{F} = -1, 0, +1, +2$) are shown in the images. The plots (d)-(f) show the line profiles along z-direction for images (a)-(c) respectively.}
\end{figure}
Fig. 9 shows the spatially separated cold atom clouds in different $m_{F}$ states and the corresponding line profiles along z-direction. In the absence of optical pumping beam (Fig.9 (a) and (d)), peaks correspond to different states ($\ket{F=2, m_{F} = -1, 0, +1, +2}$). The cold atoms in $(F=2, m_{F} = -2)$ states are not seen in the image as they are accelerated upwards and lost to the atom-chip surface. We noted that without optical pumping, in split cloud after SG pulse, $\sim$ 24 $\%$ of atoms were present in $\ket{F=2, m_{F} = + 2}$ state, with respect to number of atoms in G-MOT. With use of optical pumping, the population in $\ket{F=2, m_{F} = + 2}$ state was increased as shown in Fig.9 (b) and (c). After using an optimized optical pumping pulse ($\sim$ 500 $\mu s$, power = 300 $\mu W$) in presence of $\sim$ 3.6 G bias magnetic field, $\sim$ 92 $\%$ of G-MOT atoms were transferred to $\ket{F=2, m_{F} = + 2}$ state. Any further change in power or pulse duration of optical pumping beam did not help in improving optical pumping efficiency significantly.\\
\subsection{Magnetic trapping of atoms}
\begin{figure}[htp]
\centering
\includegraphics[width= 8.5 cm]{Fig10.png}
\caption{ The spatial variation in magnitude of the net magnetic field due to copper Z-wire and bias coils. The current in the copper Z-wire was $I_{Z} = 60 A$ and magnetic fields due to bias coils were $B_{y}$ = 38 G and $B_{x}$ = -25 G. The image in (d) is the fluorescence image of magnetically trapped atom cloud.}
\end{figure}
After optical pumping of atoms ( with pumping pulse of optimized duration of $\sim 500 \mu s$ ) to Zeeman hyperfine quantum state ($(\ket{F=2, \, m_{F} = + 2})$), the magnetic trapping of atoms was performed. For this, all lasers and magnetic fields were switched-off after optical pumping. Then trap fields were switching-on within 1 ms time. Nearly 60 A of dc current was switched-on in the copper Z-wire ($I_{Z} = 60 A$) along with the bias magnetic fields of strength $B_{y}$ = 38 G and $B_{x}$ = -25 G. Fig. 10 shows the calculated magnetic field strength due to Z-wire and bias fields for the above parameters. Fig. 10 (a), (b) and (c) show the calculated absolute magnetic field (or potential) profiles along x-, y- and z-direction respectively. The trap configuration is close to Ioffe-Pritchard configuration \cite{esslinger}. The trap frequencies at this current are $\omega_{x}/2\pi \sim 36$ Hz and $\omega_{y,z}/2\pi \sim 110$ Hz. Due to asymmetry in the trap frequencies, the trapped atom cloud is expected to be elongated along x-direction, as shown in Fig. 10(d). The CCD image in Fig. 10 (d) shows the fluorescence image of magnetically trapped atom cloud in X-Z plane. The total number of cold $^{87}Rb$ atoms in the magnetic trap was $\sim 1 \times 10^{7}$ out of $\sim 5 \times 10^{7}$ in G-MOT, leading to trapping efficiency $\sim 20 \% $. We believe that this trapping efficiency can be improved by improving several factors such as switching-on magnetic trap slowly, lowering the temperature of atom cloud in G-MOT, etc. \\
\section{Conclusion}
To conclude, Stern-Gerlach effect of magnetic field on cold $^{87}Rb$ atoms has been used to spatially separate and detect atoms in different Zeeman hyperfine quantum states. It is found that the spatial separation of cold atoms in different magnetic hyperfine states depends on SG field pulse duration and temperature of atom cloud temperature. Using this SG state detection technique, the optimization of optical pumping of cold $^{87}Rb$ atoms to a trappable Zeeman hyperfine state has been demonstrated. Nearly 92 $\%$ of cold atoms from grey magneto-optical trap (G-MOT) were optically pumped to $\ket{F=2, m_{F} = + 2}$ state using this SG technique.
\begin{acknowledgments}
We are grateful to R. Folman and Omer Amit, Ben-Gurion University, Israel for technical suggestions during experiments. We are also thankful to our colleagues S. P. Ram, S. Singh and K. Bhardwaj, LPAS, RRCAT for their help during the experiments. The technical assistance provided by Amit Chaudhary, LPAS, RRCAT during the experiments is also gratefully acknowledged.
\end{acknowledgments}
|
\section{Introduction}
Detecting gravitational waves (GWs) from the merging compact binaries consists of black holes and neutron stars has ushered in a new era of observational astronomy~\cite{gw150914, gw170817}. So far, the network of Advanced LIGO and Advanced Virgo Detectors has observed more than 90 of these events~\cite{ GWTC1, GWTC2, 2021arXiv211103606T, Venumadhav:2019tad, Nitz:2018imz, Nitz:2020oeq}. Observations of compact binary mergers allow us to design a range of novel tests of general relativity (GR) in the strong-field regime~\cite{TGR-GW150914, TGR-GWTC1, TGR-GWTC2}. A current limitation on tests of beyond-GR physics with compact binary coalescences is the lack of understanding of the strong-field merger regime in nearly all modified theories of gravity. Thereby, we generally perform the tests assuming GR to be the null hypothesis or evaluate the consistency of the data with predictions from the theory. At the same time, the latter test can reveal if any unknown binary parameter influences the observed signal.
The continuous wavelet transformation (CWT) based time-frequency analysis has played a crucial role in identifying the gravitational wave events and detecting the deviation from general relativity. The most potential gravitational wave sources, such as compact binary mergers, supernovae explosions, produce non-stationary signals. The CWT is a powerful analysis tool that allows us to obtain a time-frequency localized projection of a non-stationary signal. This procedure calculates the wavelet coefficients by performing a scalar product between the signal and wavelet basis. These coefficients are used not only for time-frequency analysis but also to reconstruct the original time-domain signal. The existing time-frequency-based methods such as coherent wave burst ($\textsc{cWB}$)~\cite{Klimenko_2004, Klimenko:2005xv, Klimenko_2008}, $\textsc{BayesWave}$~\cite{Cornish_2015, Cornish:2020dwh}, and X-pipeline~\cite{Sutton:2009gi, Chatterji:2006nh} are designed for searching unmodeled gravitational-wave bursts and reconstructing the signal from the coherent response of a network of detectors.
The previous studies reported satisfactory agreement between the reconstructed signal and the estimated waveform of binary black hole mergers for the louder events~\cite{GWTC1, Ghonge:2020suv, Salemi:2019uea}.
The conventional burst search method identifies a cluster of pixels from the time-frequency map of a detector's strain data if each of those pixels has more power than one expects from the noise alone, called triggers. If the triggers from a network of detectors are coincident, then the method claims a detection of a signal~\cite{Anderson:2000yy}. The $\textsc{cWB}$ method first constructs a multi-resolution time-frequency map of the strain data using Wilson-Daubechies-Meyer wavelets~\cite{Necula:2012zz} and uses the same strategy to detect the events. After that, the method employs a coherent framework of the constrained maximum likelihood approach to yield the event properties: sky location, wave polarization, and signal reconstruction~\cite{Klimenko:2005xv, Klimenko:2015ypf}. Whereas the $\textsc{BayesWave}$ method employs a framework of Bayesian statistics without relying on any prior assumption of waveform morphology~\cite{Cornish_2015}. This method reconstruct the gravitational wave signals and instrumental noise, where both of them are characterized as a superposition of Gabor-Morlet wavelets or chrirplets~\cite{Millhouse_2018}. The number of time-localized wavelets and their parameters are determined via Reversible Jump Markov Chain Monte Carlo algorithm. After that, whether the likelihood of the event is favoured to a true gravitational wave signal or instrumental glitch is determined by using Bayesian model selection strategy. Finally, the posterior samples are used to reconstruct the signal as a superposition of the wavelets.
A wavelet can be regarded as a time-frequency localized bandpass filter. The bandwidth of which is determined by the properties of that wavelet. The reconstructed signal has an inevitable noise contribution that passes through those wavelet filters. The match between the reconstructed signal and the estimated waveform increase with the signal-to-noise ratio (SNR). However, we would never achieve absolute agreement (unit match) even if our predicted theory is complete because of the additional noise that passes through the coherent wavelets.
In this paper, we propose a quasi model-dependent method for reconstruction of signal from noisy data. We identify the essential wavelets using the binary black holes merger waveforms estimated using the standard Bayesian analysis~\cite{Rover:2006ni, vanderSluys:2008qx, Veitch:2014wba, Skilling:2006gxv}. The unmodelled reconstruction methods identify the wavelets that are coherent across the network of detectors. The collection of those coherent wavelets in the time-frequency plane form a cluster. Similarly, the essential wavelets form a cluster in the time-frequency plane which can be collectively looked as a single patch with an irregular boundary.
The patch area determines how much noise persists in the reconstructed signal. The wavelets are discretely placed over the time-scale plane, where the patch area is determined by the placement method and sparseness. The dyadic grid placement is considered to be the most efficient method, leading to the construction of an orthonormal wavelet basis. This placement assumes an output of octave decomposition (power-of-two logarithmic) to construct the grid. The sparseness of the grid imposes a limit on the acceptable loss in signal characteristics. We target to achieve maximum sparseness and still have an adequate representation. A highly sparse grid might provide a sufficient signal representation, but the area covered by the essential wavelets would be larger than the case of a denser grid. On the contrary, an over-dense grid can not reduce the patch area beyond a specific limit and also increases the computational cost.
In this paper, we propose a log-uniform scale to place the wavelets over the time-scale domain.
Further, we derive an inverse wavelet transformation formula for log-uniform scale. The wavelets with such scales are highly redundant (i.e., over complete), but they can further reduce the area covered by the essential wavelets. As a result, We can reduce the noise contribution in the reconstructed signal, which leads to a more accurate signal reconstruction. Moreover, the essential wavelets with log-uniform scale can provide a more appropriate signal representation than the octave scale for the high-frequency linear chirp signals.
It implies that the log-uniform scale is more efficient for reconstructing the signals from low mass binary systems.
This paper is organized as follows: Section~\ref{sec:WaveletReconstruction} describes the reconstruction methods using the continuous wavelet transformation with octave and log-uniform scales;
Section~\ref{sec:performance} demonstrates the performance of the reconstruction methods for the linear chirp signals and gravitational wave signals in simulated Gaussian noise; Section~\ref{sec:Deviation} describes the efficiency of the wavelet reconstruction method when a signal in the data does not belong to the manifold of search template waveform by injecting an eccentric waveform; Section~\ref{sec:GWTC1Ananlysis} exhibits results of our analysis for binary black hole merger events observed by the Advanced LIGO and Virgo during the first and second observation runs. Finally, we conclude in Section~\ref{sec:Consclusion}.
\section{Reconstruction formulae for non-orthogonal wavelet}
\label{sec:WaveletReconstruction}
The CWT of a continuous signal $x(t)$ is a linear mapping onto a time-scale space of wavelets:
\begin{equation}
\label{eq:cwt}
\begin{split}
X( a, b) & = \int x(t) \, \Psi^\ast_{a, b}(t) \, dt \\
&=\frac{1}{\sqrt{a}} \int x(t) \, \psi^\ast\left(\frac{t-b}{a}\right) \: dt ,
\end{split}
\end{equation}
where, $\psi^\ast$ is the complex conjugate of the shifted and scaled version of the time-localized mother wavelet $\psi$, in which $a$ and $b$ are the scale and time-shift parameters, respectively. Therefore, the CWT provides a time-scale representation of the signal by performing a sliding cross-correlation with a continuous family of wavelets. The quantity $1/\sqrt{a}$ outside the integral is an energy normalized factor. It assures that each wavelet has the same energy, whatever the value of scaling and shift. The function $\psi(t)$ must satisfy a set of mathematical criteria to be a wavelet, it must have finite energy,
\begin{equation}
E_\psi = \int_\infty^\infty \abs{ \psi(t) }^2 \: dt < \infty,
\end{equation}
and must hold the admissibility condition,
\begin{equation}
\label{eq:admissibility_constant}
C_{\psi} = \int_\infty^\infty \frac{ \abs{\tilde{\psi}(\omega) }^2 }{\omega} \: d\omega < \infty ,
\end{equation}
where, $\tilde{\psi}(\omega)$ is the Fourier transform of $\psi(t)$, $\omega = 2\pi f$ is the angular frequency, and $C_{\psi}$ is called the admissibility constant. The above condition implies that $\tilde{\psi}(\omega)$ is endeavoured to zero faster than $\omega$ and must not have zero frequency component, $\tilde{\psi}(0)=0$. If the wavelet function satisfy both the criteria, then the inverse wavelet transform can be described as a superposition of the dual wavelets:
\begin{equation}
\label{eq:icwt1}
x(t) = \frac{1}{C_\psi} \int_0^\infty \int_{-\infty}^{\infty} X(a, b) \: \Psi_{a, b}(t) \: \frac{1}{a^2} \, db \, da
\end{equation}
An alternative approach of inversion formula was found by Morlet, we can even choose a a completely different wavelet function, Dirac $\delta$-function $\delta\left((b-t)/a\right)$ instead of the analysing wavelet, and that leads to a single integral inverse formula~\cite{Farge-1992},
\begin{equation}
\label{eq:icwt2}
x(t) = \frac{2}{C_\delta} \int_0^\infty \Re\left\{ X(a, t)\right\} \frac{1}{a^{3/2}} \, da ,
\end{equation}
where $C_\delta$ is the admissibility constant of the $\delta$-function that can be computed using the Eq.\eqref{eq:admissibility_constant}.
In this work, we use this single integral inversion formula for reconstructing the gravitational wave signals.
The wavelet function can be regarded as a impulse response of a bandpass filter. The associated frequency of the wavelet can be treated as frequency value in the time-frequency domain, which is known as pseudo-frequency ($f_p$). It depends on the central frequency ($f_c$) and the scale parameter of the wavelet,
\begin{equation}
f_p = f_c/a
\end{equation}
This equation can used to represent the CWT in time-frequency frame as like the short-time Fourier transform.
One of the most commonly used mother wavelet is the Morlet wavelet~\cite{Morlet-1984}, which is a complex non-orthogonal wavelet. As the complex wavelet can separate the phase and amplitude components associated with the signal, it is more suited to determine the instantaneous frequency. The Morlet wavelet consists of an harmonic oscillation with Gaussian window,
\begin{equation}
\psi(t) = A_{f_c} \pi^{-1/4} \left( e^{i 2\pi f_c t} - e^{-(2\pi f_c)^2/2} \right) e^{-t^2/2},
\end{equation}
where $A_{f_c} = (1 - e^{-4\pi^2 f_c^2} - 2e^{-3\pi^2 f_c^2})^{-1/2}$ the normalization constant. The second term in the bracket is a correction to preserve the admissibility condition. It ensures that the zero-frequency component vanishes. In practice, we ignore this term as it is approximately zero for the values of $f_c \gg 0$. Note that the above equation does not contain the time-shift and scale parameters. When we analyze a signal, we replace the variable $t$ with $(t-b)/a$.
The central frequency $f_c$ is a crucial parameter in the time-frequency analysis to regulate the tradeoff between temporal precession and spectral resolution.
It controls the number-of-cycles of the wavelet without modifying the shape of Gaussian window. A large value of central frequency leads to an increased temporal precision at the cost of decreased spectral precision and vice versa for a smaller value of central frequency. It is impossible to achieve simultaneously good precision in time as well as in frequency. A rule of thumb can be proposed for analyzing the gravitational waves from compact binary mergers.
The signals from low-mass systems spend several tens of seconds or more than a hundred seconds in the bandwidth of Advanced LIGO-like detectors, which look like a long-duration chirp signal. A larger value of central frequency is convenient to achieve an overall good time-frequency precision for those systems. On the other hand, a smaller value of central frequency is suitable for high mass systems as their signals are short-duration bursts with a tiny chirp.
\subsection{Choice of wavelet scales: octave}
The CWT generally suffers due to the redundancy at large scales, where the neighboring wavelets are highly correlated. Once a wavelet function is chosen, it is important to choose a tightest set of scales that forms an orthonormal wavelet set. The key mathematical criteria to choose a set of discrete wavelets is that every function $f\in L^2(\mathbb{R})$\footnote{$L^2(\mathbb{R})$ denotes the space of square integrable functions on the real line ($\mathbb{R}$). } must be fully expressed as a superposition of those wavelets i.e., the set is complete in $L^2(\mathbb{R})$. The conventional approach of choosing the scales is an output of octave decomposition~\cite{Grossmann-1989, APracticalGuidetoWaveletAnalysis},
\begin{equation}
\begin{split}
a_j & = a_0 \: 2^{j \delta j}, \ \ j = 0, 1, 2,..., J \\
J &= \delta j^{-1} \logtwo{\left(N \delta t/ a_0\right)},
\end{split}
\end{equation}
where $N$ is the total number of samples in $x(t)$, $\delta t$ is sampling interval, $a_0=2\delta t \, f_c$ is the smallest resolvable scale, and $\delta j$ is the spacing between the discrete scales. For Morlet wavelet, an adequate scale resolution can be achieved for the values of $\delta j \lessapprox 0.5$. The quantity $J$ that determines the highest value of the scale is associated with the Nyquist frequency. The octave scale leads to a simpler implementation of inverse discrete wavelet transformation as the quantity $da/a = a_0 \delta j \lnb{2} $ is a constant. Now, the inverse formula Eq.~\eqref{eq:icwt2} yields,
\begin{equation}
\label{eq:ReconstructionOctave}
x_n = \frac{\delta j \delta t^{1/2}}{C_{\delta} \psi_0(0)} \sum_{j=1}^{J} \frac{\Re{ \left\{ X_n (a_j) \right\} }}{ a_j^{1/2}}
\end{equation}
\subsection{Choice of log-uniform scale}
For non-orthogonal wavelet analysis, one can adopt an arbitrary set of scales to build up a more complete picture if that set satisfy the completeness criteria. We propose a log-uniform scale,
\begin{equation}
\label{eq:log-uniformSacle}
\begin{split}
\frac{1}{a_j} &= \frac{1}{a_{j-1}} - \delta\ell \\
&= \frac{1}{a_0} - j\delta\ell .
\end{split}
\end{equation}
The quantity $\delta\ell=\delta f_p/f_c$ stands for the inverse of spacing between two discrete wavelets and smallest resolvable scale $a_0 = f_c/\delta f_p$. In this work, we choose $\delta f_p = 1/T$, where $T$ is duration of the time-series.
The log-uniform scale leads to an uniformly spaced pseudo-frequency. In this formalism, Morlet wavelet transformation is equivalent to the scaled Gabor transformation as the pseudo-frequencies are uniformly spaced. For log-uniform scale $da/a^2 = \delta\ell$, thus, the discrete inverse formula Eq.~\eqref{eq:icwt2} yields,
\begin{equation}
\label{eq:Reconstructionlog-uniform}
x_n = \frac{\delta\ell \, \delta t^{1/2}}{C_{\delta} \, \psi_0(0)} \sum_{j=1}^{J} a_j^{1/2} \: \Re{ \left\{ X_n (a_j) \right\}}
\end{equation}
\subsection{Identifying the essential wavelets}
The conventional method for identifying essential wavelets is a straightforward approach, where one discards the low-magnitude wavelet coefficients by applying a threshold, which is known as wavelet thresholding.~\cite{Donoho-1994, Johnstone-1997}. This approach aims to remove the noise from data without affecting the basic features of the signal. In this work, we propose a slightly different approach. We employ the modelled waveforms that are obtained from the standard Bayesian parameter estimation analysis. We compute wavelet coefficients for a waveform and set a threshold on coefficient value to select the essential wavelets. The threshold is chosen such that the resultant power from the essential wavelets is equal to a percentage of total power of the spectrogram. We calculate the threshold ($E_{\ast}$) for a given value of fractional power loss ($ R_{\ast}$) by solving an equation,
\begin{equation}
\label{eq:ImportantWavelets}
R_{\ast} = 1 - \sum_{n, j} \abs{X(n, j)}^2 \, \left[ \abs{X(n, j)}^2 \geq E_{\ast} \right] \bigg/ \sum_{n, j} \abs{X(n, j)}^2.
\end{equation}
We shall call that $R_\ast$ is the spectral loss parameter. As the number of essential wavelets is considerably fewer than the total number of wavelets, a significant amount of noise can be removed while preserving the basic features of the signal.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{cluster_area_log}
\caption{ An illustration of the area covered by the cluster of essential wavelets is produced using a gravitational waveform of a nonspinning equal mass binary system with a chirp mass of $30 {M}_\odot$. The areas bounded by the log-uniform wavelets and octave scale wavelets are 13 and 15, respectively. For this computation, we have assumed $\aligo$~\cite{aLIGO_ZDHP} noise curve with a fixed lower cutoff frequency of $20\mathrm{Hz}$, the waveform is generated using $\mathrm{SEOBNRv4}$ model~\cite{Bohe:2016gbl}. We have considered $R_\ast=0.05$ for determining the essential wavelets.}
\label{fig:area_comparison}
\end{figure}
\subsection{Occupied area by the essential wavelets}
Reconstructing a signal using the essential wavelets trades with the sensitivity, by which we are able to remove the noise from data. However, the reconstructed signal would always contain a nominal amount of inseparable noise, which passes through the wavelet filters. As a wavelet is regarded as a time-localized bandpass filter, it can be seen as a patch on time-frequency plane. The area (time-frequency bandwidth) of that patch can be used to determine the amount noise released through that wavelet. Therefore, the total amount of inseparable noise can be estimated by calculating the total area covered by the essential wavelets.
The area on the time-frequency plane covered by a cluster of wavelets $\{ (b_i, a_i) \}_{i=1}^{n_0}$ is related to their placement. For octave scale wavelet with a central frequency of $f_c$, the covered area is:
\begin{equation}
\begin{split}
A_{\rm{Octave}} & = \delta t \: a_0 \delta j\lnb{2} \sum_{i=1}^{n_0} a_i \\
& = 2\lnb{2} \: \delta j \left( f_c \delta t\right)^2 \sum_{i=1}^{n_0} 1 / f_{p i},
\end{split}
\end{equation}
where, $f_{pi}$ is the $i\nth$ pseudo frequency, $f_{p i}=f_c/a_i$. Whereas, the time-frequency area governed by a cluster of wavelets $\{ (b_i, a_i) \}_{i=1}^{n_0}$ with log-uniform scale is:
\begin{equation}
A_{\rm{log-uniform}} = \sum_{i=1}^{n_0} \delta t /T .
\end{equation}
Fig.~\ref{fig:area_comparison} shows that log-uniform scale occupy a smaller area than the octave scale. The figure indicates that the log-uniform scale allows $\sim 13\%$ less noise in the reconstructed signal than the octave scale for a high mass system, can provide a nearly identical signal representation.
\section{Performance of the reconstruction procedures}
\label{sec:performance}
In this section, we demonstrate the performance of the wavelet based signal reconstruction and compared between the choice of Octave and log-uniform scale. First, we exhibit that the essential wavelets are adequate to represent a chirp signal containing a broad range of frequency. Second, we carry out injection analysis to evaluate the performance for real data analysis. The gravitational wave signals are drawn from the binary black merger and added to a simulated Gaussian noise.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{power_comparsion}
\caption{ Performance of octave and log-uniform scale wavelets reconstruction for the linear chirp signals. The signals are generated using Eq.~\eqref{eq:simple_chirp}. The y-axis represents the normalized minimum squared error in the reconstructed signal. The solid line and dashed line correspond to the log-uniform scale and octave scale, respectively. The bottom and top x-axes represent the starting and ending frequency of the chirp, respectively. The legend represents a set of spectral loss parameters ($R_{\ast}$) as described in Eq.~\eqref{eq:ImportantWavelets}. We used those values to choose the essential wavelets. }
\label{fig:power_comparison}
\end{figure}
\subsection{Reconstruction of chirp signal without noise }
\label{sec:RecChirpSignal}
We consider a simple chirp signal with a constant amplitude and phase is upto a quadratic order in time $(t)$,
\begin{equation}
\label{eq:simple_chirp}
x(t) = \sin\left( 2\pi f_0t + 2\pi f_0t^2 \right),
\end{equation}
where $f_0$ is the starting frequency of the signal
and the time range is chosen to be between 0 and $2\, \si{\second}$. Therefore, the ending frequency of the chirp is $3f_0$. This type of chirp signal is adequate to exhibit the performance of the reconstruction method over a broad range of frequencies. We consider five different cases to identify the essential wavelets.
For each case, the total power contained in the essential wavelets equals a fraction of signal power.
It is inevitable to have an overall amplitude loss since the number of essential wavelets is a small subset of the set of wavelets representing the whole time-scale space.
However, for a given value of $R_{\ast}$, we can set the overall amplitude by looking at the amplitude loss when determining the essential wavelets. An alternative approach is to define normalized minimum squared error (NMSE) for a reconstructed signal $x_{\rm{rec}}(t)$,
\begin{equation}
\label{eq:NMSE}
\mathcal{E}_{\rm{NMSE}} = \int \Big| x(t)/\norm{x} - x_{\rm{rec} }(t)/\norm{x_{\rm{rec}}} \Big|^2 \: dt ,
\end{equation}
where, the symbol $\norm{\cdot}$ denotes the norm. We use this above equation to quantify the adequateness of the essential wavelets to characterize a linear chirp signal.
In Fig.~\ref{fig:power_comparison}, we illustrate the accuracy of the wavelet reconstruction of the linear chirp signals. The solid and dashed lines correspond to the log-uniform scale and octave scale, respectively. We consider five different cases of fractional loss in spectrogram power to select the essential wavelets used for reconstruction. The reconstruction of high-frequency chirps using log-uniform scale wavelet is more accurate than the octave scale and vice versa for the low-frequency chirps.
\subsection{Reconstruction of gravitational wave signal in simulated noise }
\label{subsec:gwsimnoise}
We estimate the performance of the reconstruction methods for gravitational wave signals from compact binary mergers of equal mass nonspinning black holes. For each case, an identical signal is injected in many noise realizations of stationary Gaussian distribution weighted by Advanced LIGO zero-detuned high-power ($\aligo$) design sensitivity~\cite{aLIGO_ZDHP}. The waveforms are generated for an wide range of chirp mass between $10 {M}_\odot$ and $40 {M}_\odot$ using $\mathrm{SEOBNRv4}$ model~\cite{Bohe:2016gbl} with a fixed lower cutoff frequency of $20\mathrm{Hz}$. To determine the reconstruction accuracy with signal-to-noise ratio (SNR) of the injection (injected SNR), we choose a set of values between 5 and 50. The injected SNR ($\rho_{\rm{inj}} $) of a waveform $(h_{\rm{inj}})$ is defined as:
\begin{equation}
\rho_{\mathrm{inj}}^2 = 4 \int_{f_{\mathrm{low}}}^{f_{\mathrm{high}}} \frac{ \tilde{h}^{\ast}_{\mathrm{inj}}(f) \: \tilde{h}_{\mathrm{inj}}(f) }{S_n(f)} \, df ,
\end{equation}
where, $\tilde{h}_{\mathrm{inj}}(f)$ denotes the Fourier transform of $ h_{\mathrm{inj}} (t) $ and $S_{n} (f)$ denotes the one-sided detector noise power spectral density.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{spec_loss}
\caption{ The noise averaged match ( $ \langle \mathscr{M} \rangle_{\rm{noise}} $ ) as a function of spectral loss parameter ($R_\ast$) is produced using a gravitational waveform of a nonspinning equal mass binary system with a chirp mass of $30 {M}_\odot$. The solid dot on the curve indicates its maxima. The right y-axis shows the relative number of essential wavelets ($\bar{N}_W$) as a function of $R_\ast$. The unit value of $\bar{N}_W$ corresponds to a 99.9\% match between the original and reconstructed signal for the zero noise case. These results are produced assuming the $\aligo$ noise curve with a fixed lower cutoff frequency of $20\mathrm{Hz}$. The waveforms are generated using the $\mathrm{SEOBNRv4}$ model.}
\label{fig:spec_loss}
\end{figure}
We reconstruct the signal from noisy data using the log-uniform scale wavelet transform, where the essential wavelets are selected using the Eq.~\eqref{eq:ImportantWavelets}.
To quantify the signal reconstruction accuracy, we calculate the \textit{match} between the injected signal $ h_{\mathrm{inj}} $ and reconstructed signal $ h_{\mathrm{rec}} $, which is defined as a inner product between two normalized waveforms ($\hat{h}_\ast = {h_\ast}/\sqrt{ \inp{ h_\ast }{ h_\ast } }$):
\begin{equation}
\label{eq:overlap}
\mathscr{M} ( h_{\mathrm{inj}} , h_{\mathrm{rec}} ) := \inp{ \hat h_{ \mathrm{inj} } }{ \hat h_{ \mathrm{rec} } } ,
\end{equation}
where the term with angular brackets represents the following inner product,
\begin{equation}
\inp{ h_{ \mathrm{inj} } }{ h_{ \mathrm{rec} } } = 4 \Re \int_{f_{\mathrm{low}}}^{f_{\mathrm{high}}} df \, \frac{ \tilde{h}^{\ast}_{ \mathrm{inj} }(f) \, \tilde{h}_{ \mathrm{rec} }(f) }{S_n (f)} ,
\end{equation}
where $ \tilde{h}_{ \mathrm{inj} }(f) $ and $\tilde{h}_{ \mathrm{rec} }(f) $ denote the Fourier transform of $ h_{\mathrm{inj}} (t)$ and $ h_{\mathrm{rec}} (t)$, respectively.
The match between two waveforms varies between -1 and 1, depending on their correlation but not their overall amplitudes. A match value of 1 indicates a perfect positive correlation such that two waveforms change with equal proportion, and 0 means no correlation. The equal proportion changes with reverse direction indicate perfect negative correlation, for which match value is -1. To quantify the agreement between the wavelet reconstruction for a network of detectors, we compute the network match ($\mathscr{M} _{\rm{net}}$) as given in~\cite{Becsy:2016ofp},
\begin{equation}
\mathscr{M} _{\rm{net}} = \frac{ \sum_k \inp{ h^{k}_{ \mathrm{inj} } }{ h^{k}_{ \mathrm{rec} } } }{ \left( \sum_k\inp{ h^{k}_{ \mathrm{inj} } }{ h^{k}_{ \mathrm{inj} } } \cdot \sum_k\inp{ h^{k}_{ \mathrm{rec} } }{ h^{k}_{ \mathrm{rec} } } \right)^{1/2} } ,
\end{equation}
where $k$ represents the $k\nth$ detector. In this paper, we consider three detectors configuration of Hanford (H1), Livingston (L1), and Virgo (V1). We consider $\aligo$ design sensitivity for H1 and L1 detectors~\cite{LIGOScientific:2014pky, Harry:2010zz, aLIGO_ZDHP}, and advanced Virgo design sensitivity for V1~\cite{Accadia_2012, VIRGO:2014yos}. Please note that we demonstrate the injection analysis in this section assuming the single detector with the $\aligo$ design sensitivity.
A whiten Gaussian noise is considered to be distributed normally over time-scale domain. This implies that the noise energy is equally distributed over time-scale domain. The amount of noise retains in the reconstructed signal is determined by the spectral loss parameter $R_\ast$. A higher value of $R_\ast$ can further reduce the inseparable noise in the reconstructed signal. At the same time, the essential wavelets would not be able to represent the complete signal characteristics. It implies that the spectral loss parameter plays a role that imposes a limitation on achieving the maximum overlap between the injected and reconstructed signal. Therefore, we want to optimize this parameter. For a given source parameter and injected SNR, we perform the injection analysis with a set of $R_\ast$ values. The noise averaged match for a given injected signal $h_{\rm{inj}}( \vec{\lambda}, \rho_{\rm{inj} } )$ and $R_\ast$ can be defined as
\begin{equation}
\label{eq:SpectralLoss}
\langle \mathscr{M} ( R_\ast ) \rangle_{ \rm{noise} } = \frac{1}{N}\sum_{i=1}^N \mathscr{M} ( h_{\mathrm{inj}} , h_{\mathrm{rec}} [n_i] ),
\end{equation}
where, $ h_{\mathrm{rec}} [n_i] $ represents the reconstructed signal from the data of $i\nth$ noise realization $n_i$.
To determine the essential wavelets for a given value of $R_\ast$, we compute the wavelet coefficients of whitened waveform and plug in to Eq.~\eqref{eq:ImportantWavelets}.
Fig.~\ref{fig:spec_loss} shows the noise averaged match as a function of $R_\ast$ for a set of injected SNR values and the number of essential wavelets as a function of $R_\ast$. If we fix the spectral loss parameter and increase the injected SNR, the signal contribution to each essential wavelet increases. At the same time, the average noise contribution remains the same since the number of essential wavelets and their properties does not change. This leads to an improvement in the reconstruction accuracy. A fixed SNR curve in the Fig.~\ref{fig:spec_loss} indicates that a smaller value of $R_\ast$ allows many nonessential wavelets in the analysis. The signal contribution to those wavelets is trivial, but the inseparable noise in the reconstructed signal increases. On the other hand, a higher value of $R_\ast$ discards many moderate essential wavelets. Thus, the choice of $R_\ast$ is a tradeoff between the signal and noise contributions to the wavelets. To find an optimum value R, we maximize the quantity $\langle \mathscr{M} ( R_\ast ) \rangle_{\rm{noise}} $ over $R_\ast$
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{match_comparison_half}
\caption{ The figures shows the mismatch between the injected gravitational wave signal and reconstructed signal, where the filter wavelets were constructed using the log-uniform scale as defined in Eq.~\eqref{eq:log-uniformSacle}. For producing the injections, we have assumed the nonspinning equal mass binary black holes with a wide range of chirp mass as labelled in $x$-axis. The legend of the figures stand for the optimal SNR of the injected signal. We have injected an identical signal into many noise realization to produce the distribution of mismatch for each chirp mass. We have assumed the $\aligo$ noise curve with a fixed lower cutoff frequency of $20\mathrm{Hz}$. The waveforms are generated using the $\mathrm{SEOBNRv4}$ model.}
\label{fig:match_comparison}
\end{figure}
Fig.~\ref{fig:match_comparison} demonstrates the \textit{mismatch} ($1 - \mathscr{M} $) in reconstruction for nonspinning binary systems with equal component masses.
The solid curve represents the median of the mismatch distribution, and the shaded region shows the $\pm \sigma$ width of that distribution. For the case of a fixed optimal SNR, we can see that mismatch substantially decreases with an increase of chirp mass. It is intuitively expected: the number of essential wavelets for high chirp mass systems is fewer than low chirp mass systems.
The waveform of a binary system with high chirp mass can be characterized using a few wavelets as the waveform is short. In contrast, the waveform of a low chirp mass system is longer, for which one requires a large number of wavelets to represent the signal.
As the injection chirp mass increases, the number of essential wavelets decreases, and they cover a smaller area over the time-frequency plane. The signal contribution to the essential wavelet coefficients increases with the injection chirp mass for a fixed SNR case. At the same time, the noise contribution decreases due to the smaller area. That explains why the reconstruction signal for a high chirp mass system is more accurate than a low chirp mass.
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{recon_dev_LU_323_M60_v2}
\caption{ An illustration of wavelet reconstruction (solid red line) and $\textsc{LALInference}$ (cyan band) with $\mathrm{IMRPhenomXP}$ waveform model, obtained from an injection analysis using an NR eccentric waveform SXS:BBH:0323 picked from SXS catalog with a total mass of $60{M}_\odot$. The dashed black line represents the injected waveform. For this computation, we have used the three detector configuration of Advanced LIGO and Advanced Virgo. }
\label{fig:eccnReconstruction}
\end{figure*}
\section{Identifying the deviation}
\label{sec:Deviation}
As of yet, we discussed the performance of wavelet reconstruction where injected waveforms were used for determining the essential wavelets. That results tell us the efficiency of the semi-model dependent wavelet reconstruction method when a signal in the data belongs to the search template waveform manifold. However, a signal in real data may not belong to that manifold. The deviations could arise due to: the influence of unknown binary parameters (such as eccentricity if the waveform model only considers the circular binary), missing physics in the waveform model, deviation from GR, or noise artifacts.
We demonstrate one of such cases, an injection waveform simulation for an eccentric BBH merger. Since recent template-based analysis by LIGO-Virgo collaborations used quasi-circular waveform~\cite{GWTC2}, we consider a recently developed $\mathrm{IMRPhenomXP}$ model to generate the template waveform~\cite{Pratten:2020ceb}. This model includes the effect of orbital precession but does not consider orbital eccentricity.
The inclusion of eccentricity in the injected signal leads to a deviation from the search template waveform manifold. To demonstrate this, we consider a numerical-relativity simulation of eccentric IMR waveform picked from Simulating eXtreme Spacetimes (SXS) catalog~\cite{ian_hinder_2019_3326460}. This system is a nonspinning binary with mass ratio of 1.22. The numerical simulation is performed for an eccentric BBH system, where eccentricity evolves with time. The reference eccentricity is $e_{\rm{ref}} = 0.194$ as measured at a reference orbital frequency of $Mf_{0}=0.0137$. In our analysis, we scale the simulation for a total mass of $60 \, {M}_\odot$, which is suitable for ground based detectors.
We keep the system in the \textit{face-on} configuration where the inclination angle is $0^{\circ}$. As the component masses are nearly equal and zero inclination, the contribution of higher-order modes to the injected signal is negligible. This configuration assures us that the deviation enters only due to the orbital eccentricity. We inject the signal assuming a three detector configuration of Advanced LIGO and Advanced Virgo~\cite{LIGOScientific:2014pky, Harry:2010zz, aLIGO_ZDHP, Accadia_2012, VIRGO:2014yos}. We set the source's sky position and distance such that the injected SNR in H1, L1 and V1 are 25, 20, and 15, respectively. Since the average result of our wavelet reconstruction over many Gaussian noise realizations with an identical injected signal leads to the case of zero noise, we consider zero noise realization to construct the data stream. We analyze the data using the standard Bayesian parameter estimation library LALInferenceNest~\cite{Veitch:2014wba}, a Bayesian inference nested sampling code implemented in the LIGO Algorithm Library (LALSuite)~\cite{lalsuite}. We use the python-based package $\textsc{PESummary}$~\cite{PESummary} to process the data from parameter estimation analysis and generate the template waveform in the detector frame.
In order to reconstruct the signal from data stream, we follow these steps:
\begin{enumerate}
\label{enu:Reconstruction}
\item \label{en:step1} For each posterior sample, we generate the CBC template waveform in the detector frame. Since the GR allows only two polarization states, referred to as the plus ($h_+$) and cross ($h_\times$), the time-domain response $h_\mathcal{I}(t)$ of a given detector $\mathcal{I}$ is determined by the antenna response functions ($F^+_{\mathcal{I}}$ and $F^\times_{I}$) of those polarizations~\cite{LIGOScientific:2019hgc},
\begin{equation}
\begin{split}
h_{\mathcal{I}}(t) &= F^+_{\mathcal{I}}(\alpha, \delta, \psi, t) h_\times(t-\Delta t_{\mathcal{I}}; D_L, \iota, \vec{\lambda}) \\
& + F^\times_{\mathcal{I}} (\alpha, \delta, \psi, t) h_+(t-\Delta t_{\mathcal{I}}; D_L, \iota, \vec{\lambda}),
\end{split}
\end{equation}
where, $\alpha$ and $\delta$ refers to the source sky location in terms of right ascension and declination, $\psi$ is the polarization angle, $D_L$ is the luminosity distance to the source, $\iota$ is the inclination angle of the binary plane, $\vec{\lambda}$ represents the set of intrinsic parameters of the binary system, and $\Delta t_{\mathcal{I}} (\equiv \Delta t_{\mathcal{I}} (\alpha, \delta, t) )$ is the travel time of the signal from geocenter to the detector.
\item Whiten the waveform weighted by the noise amplitude spectral density such that the norm of the whitened waveform is equal to its optimal SNR.
\item Compute the wavelet coefficients for each whiten template waveform using the CWT as shown in Eq.~\eqref{eq:cwt},, where wavelets are constructed using the log-uniform scale.
\item Determine the essential wavelets using Eq.~\eqref{eq:ImportantWavelets}, where spectral loss parameter $R_\ast$ is obtained from Eq.~\eqref{eq:SpectralLoss}. We inject the best fit template (corresponds to maximum likelihood sample) waveform in many simulated noise realizations and estimate the value of $R_\ast$ for each detector’s data. We have found the value of $R_\ast$ for H1, L1, and V1 is 2.5\%, 4\%, and 5\%, respectively. Please note that we estimate the spectral loss parameter for the maximum likelihood sample and use that value for all the posterior samples.
\item Use those essential wavelets to reconstruct the signal from detector strain.
\end{enumerate}
\begin{table}[htbp!]
\centering
\begin{tabular}{l | c c c c c
\toprule[1pt]
\toprule[1pt]
Detector & \ $\rho_{\rm{inj}}$ \ & \ $ \widetilde{ \mathscr{M} }_{\rm{bif}}^{\rm{inj}}$ \ & \ $\widetilde{ \mathscr{M} }^{\rm{rec}}_{\rm{bif}} $ \ & \ $ \widetilde{\mathscr{M} }_{\rm{rec}}^{\rm{inj}} $ \ & \ $\widetilde{\rho}_{\rm{res}}$ \\
\midrule[1pt]
Hanford (H1) & 25 & 0.954 & 0.968 & 0.984 & 6.2 \\
Livingston (L1) & 20 & 0.95 & 0.959 & 0.978 & 5.5 \\
Virgo (V1) & 15 & 0.959 & 0.959 & 0.978 & 4.2 \\
Network & 35.6 & 0.949 & 0.958 & 0.981 & -- \\
\bottomrule[1pt]
\bottomrule[1pt]
\end{tabular}
\caption{ Comparison between the $\textsc{LALInference}$ template waveform and log-uniform wavelet reconstruction for an eccentric numerical relativity waveform SXS:BBH:0323 picked from SXS catalogue with a total mass of $60 {M}_\odot$. The numerical simulation is performed for a nonspinning, nearly equal-mass eccentric BBH system, where the measured reference eccentricity is $e_{\rm{ref}} = 0.194$ at a reference orbital frequency of $Mf_{0}=0.0137$. The quantity $\widetilde{ \mathscr{M} }$ denotes the median of the match values.}
\label{tab:matchComparison}
\end{table}
The above-described procedure is used to obtain a reconstructed signal for each posterior sample. These wavelet-based reconstructed signals are very similar, and their 90\% interval is very thin and looks like a line. Therefore, we illustrate the median of reconstructed signals at every time index.~\footnote{Note that the median of the reconstructed signals is used \emph{only} for the illustration in time-domain.}
Fig.~\ref{fig:eccnReconstruction} illustrates the results from wavelet reconstruction, 90\% credible region of $\textsc{LALInference}$ template waveform, and injected waveform. We can see the amplitude and phase of the reconstructed waveform are approximately consistent with the injected waveform. On the contrary, the $\textsc{LALInference}$ waveform is out of phase over a bit of the region, which indicates a significant deviation from the search template waveform manifold. We also compute three different matches using the reconstructed signal ($ h_{\mathrm{rec}} $), $\textsc{LALInference}$ template waveform ($ h_{\mathrm{bif}} $) obtained from posterior samples as described in step~\ref{en:step1}, and injected waveform ($ h_{\mathrm{inj}} $): $\mathscr{M} ( h_{\mathrm{inj}} , h_{\mathrm{bif}} )$, $\mathscr{M} ( h_{\mathrm{bif}} , h_{\mathrm{rec}} )$, and $\mathscr{M} ( h_{\mathrm{inj}} , h_{\mathrm{rec}} )$.
Table~\ref{tab:matchComparison} summarizes the match comparison. It signifies the wavelet-based reconstructed waveform is more faithful than the $\textsc{LALInference}$.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{gw150914_recon2}
\caption{ Log-uniform scale wavelet reconstruction results of GW150914 event, obtained using the on-source data of H1 and L1 detectors, where the times (in seconds) are shown relative to a reference time 1126259462.0. The solid black line in the top-left panel shows the median reconstructed signal obtained from H1 data and the solid blue line in the bottom-left panel for L1 data. The orange band in these panels is produced using the CBC template waveforms from parameter estimation samples. We demonstrate the reconstructed signals and CBC template waveforms in units of the standard deviation of the noise, which implies the norm is SNR. The right panel shows the histogram of the match between every posterior waveform and the corresponding wavelet-based reconstructed signal. }
\label{fig:gw150914Reconstruction}
\end{figure*}
However, we can not compute the $\mathscr{M} ( h_{\mathrm{rec}} , h_{\mathrm{inj}} )$ and $\mathscr{M} ( h_{\mathrm{bif}} , h_{\mathrm{inj}} )$ for an actual event case as it is infeasible to know the true signal in data. We propose to compute the residual SNR ($\rho_{\rm{res}}$) obtained by subtracting each template waveform of $\textsc{LALInference}$ posterior samples from the corresponding wavelet- based reconstructed waveform. LVC commonly uses this procedure in the test of GR with BBH events~\cite{TGR-GW150914, TGR-GWTC1, TGR-GWTC2, 2021arXiv211206861T}. We report the median of residual SNR in the last column of Table~\ref{tab:matchComparison}.
\section{Analysis of events in GWTC-1}
\label{sec:GWTC1Ananlysis}
We apply the proposed method to each binary black hole event in the first gravitational-wave transient catalog GWTC-1~\cite{GWTC1} to reconstruct the signals from individual detectors. In this analysis, we use the on-source data from the Gravitational Wave Open Science Center~\cite{GWOSC:catalog, LIGOScientific:2019lzm}, released for GWTC-1~\cite{GWTC1}. For determining the essential wavelets, we use the posterior samples of source properties obtained from $\textsc{Bilby}$'s~\cite{Ashton:2018jfp, Romero-Shaw:2020owr} reanalysis of GWTC-1. Parameter estimation analysis was performed using the $\mathrm{IMRPhenomPv2}$~\citep{Schmidt:2012rh, Hannam:2013oca} waveform model, and PSD was estimated using the $\textsc{BayesLine}$ algorithm~\cite{Littenberg:2014oda}.
In order to reconstruct the signal from data using the posterior samples, we follow the steps described in Section~\ref{sec:Deviation}. Fig.~\ref{fig:gw150914Reconstruction} shows the results of GW150914~\cite{gw150914}: CBC template waveform from posterior sample and wavelet reconstruction.
The agreement for H1 data is better than L1 data as the reconstruction accuracy increases with SNR. For H1 data, the most probable value (mode) of the match values is 0.975, and the maximum is 0.982. We also compute the network match values and its mode 0.962. It implies an excellent agreement between the GR template waveform and the observed data. We have seen that the reconstructed signal does remain almost identical even when the essential wavelets are selected using different posterior samples. Therefore, we consider the case of the maximum likelihood sample only to illustrate the signal is time-domain. Note that match values are computed between every posterior waveform and the corresponding wavelet-based reconstructed signal; we call this \emph{on-source} match.
Our reconstruction method trades with sensitivity for identifying the essential wavelets and removing the noise from on-source data. Consequently, the technique cannot discern the early inspiral or late ringdown part of the CBC waveforms where the signals are weaker. In the time-frequency domain, the early inspiral part spreads over time, whereas the late-ringdown part of the signal over frequency direction. In Fig.~\ref{fig:gw150914Reconstruction}, it is visible that the early inspiral part of the L1 signal (before 0.25 seconds) fades out. However, the H1 signal is still present and consistent with the template waveform because of the higher SNR in H1. At the same time, the late ringdown part of both the signals is disappeared. In a time-frequency frame, the ringdown part of a signal spreads out over frequency despite having a fixed frequency because of its exponential decay term, leading to a Lorentzian spread along the frequency direction.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{gwtc1_match}
\caption{ Network match between the CBC template waveform by Bayesian parameter estimation and wavelet-based reconstructed waveform. The vertical solid line represents the 90\% credible interval of the on-source match. The dot over each line is the mode of the on-source match distribution. The vertical dashed line indicates the 90\% interval of the expected match under the assumption of Gaussian noise, which is similar to Fig.~\ref{fig:match_comparison}.}
\label{fig:gwtc1_match}
\end{figure}
\begin{table}
\centering
\begin{tabular}{c | c c c c c c c c
\toprule[1pt]
\toprule[1pt]
Event \ & $\widetilde{\mathcal{M}}_c$ & \ $\widetilde{ \rho}_{\rm{bif}}^{\: \rm{H1}} $ & \ $\widetilde{ \rho}_{\rm{bif}}^{\: \rm{L1}} $ & \ $\widetilde{ \rho}_{\rm{bif}}^{\: \rm{V1}} $ \ & \ $ \widetilde{ \mathscr{M} }_{\rm{H1}} $ & \ $ \widetilde{ \mathscr{M} }_{\rm{L1}} $ & \ $ \widetilde{ \mathscr{M} }_{\rm{V1}} $ & \ $ \widetilde{ \mathscr{M} }_{\rm{net}} $ \\
\midrule[1pt]
GW150914 & 31.0 & 20.6 & 14.3 & -- & 0.96 & 0.94 & -- & 0.95\\
GW151012 & 18.3 & 6.5 & 5.9 & -- & 0.73 & 0.74 & -- & 0.73\\
GW151226 & 9.7 & 9.8 & 6.9 & -- & 0.80 & 0.66 & -- & 0.74 \\
GW170104 & 25.7 & 9.5 & 10.0 & -- & 0.86 & 0.89 & -- & 0.87 \\
GW170608 & 8.5 & 12.1 & 9.2 & -- & 0.78 & 0.49 & -- & 0.60 \\
GW170729 & 51.5 & 6.0 & 8.3 & 1.7 & 0.89 & 0.91 & -0.21 & 0.85 \\
GW170809 & 29.7 & 6.0 & 10.8 & 1.1 & 0.86 & 0.92 & 0.02 & 0.89 \\
GW170814 & 27.0 & 9.3 & 14.2 & 3.8 & 0.88 & 0.93 & 0.74 & 0.91 \\
GW170818 & 32.1 & 4.7 & 9.8 & 4.3 & 0.80 & 0.90 & 0.82 & 0.87 \\
GW170823 & 38.9 & 7.1 & 9.5 & -- & 0.88 & 0.93 & -- & 0.91 \\
\bottomrule[1pt]
\bottomrule[1pt]
\end{tabular}
\caption{ List of match values of GWTC-1 events obtained by computing the match between the reconstructed waveform and Bayesian inference template waveform. We report the chirp mass and the SNR to indicate the efficiency of the reconstruction as demonstrated in section~\ref{subsec:gwsimnoise}. We report the median (denoted by tilde) value of the distribution. Dashes (–) correspond to detector not included in the analysis. }
\label{tab:matchComparisonGWTC-1}
\end{table}
Further, we perform the reconstruction analysis on the remaining events of GWTC-1~\cite{LIGOScientific:2016sjg, LIGOScientific:2016dsl, LIGOScientific:2017bnn, LIGOScientific:2017vox, LIGOScientific:2017ycc} and report the results in Table~\ref{tab:matchComparisonGWTC-1}. In order to understand the reconstruction efficiency depending on the SNR and chirp mass, we also reported them in the same Table. These values are obtained from the posterior samples by $\textsc{Bilby}$~\cite{Romero-Shaw:2020owr}. The solid vertical line in Fig.~\ref{fig:gwtc1_match} shows the 90\% interval of the on-source network match values, and the solid circle marks their mode value. We found the best agreement with GR for the GW150914 event, the minimum for GW170608, albeit the observed SNR from the latter event was higher than the nominal threshold in both the detectors. The match value depends not only on SNR but also on the time-frequency area covered by the essential wavelets of the signal. In section~\ref{subsec:gwsimnoise}, we have seen that the match value increases with injection chirp mass while the injection SNR is kept at a fixed value. GW170608 event has the lowest chirp mass in the catalog, for which its essential wavelets occupy the largest area over the time-frequency plane.
The vertical dashed line in Fig.~\ref{fig:gwtc1_match} shows the 90\% interval of expected match that is determined by injecting the maximum likelihood $\textsc{Bilby}$ waveform ($ h_{\mathrm{bif}} ^\ast$) in many Gaussian noise realizations, similar study is shown in Fig.~\ref{fig:match_comparison}. We have reconstructed the signal from each realization and computed the match with $ h_{\mathrm{bif}} ^\ast$. We see that there is significant overlap between the distribution of expected match and on-source match, except for GW151226 and GW170608. In particular, the distributions are far from each other for GW170608, for which a further study is worthy. We will focus on it in future work.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{gw170729_recon}
\caption{Signal reconstruction results of GW170729 with a reference time 1185389807.3. We plot the 90\% credible interval of the CBC template waveform (orange band) and median reconstructed signal (blue). The blue dashed line in the bottom panel shows the projected signal of H1 and L1 over V1, where the sky location and polarization angle are taken from parameter estimation samples. }
\label{fig:gw170729_rec}
\end{figure}
For GW170729, the match V1 data is negative, which is unexpected. The reconstructed signal and template waveform are in nearly antiphase for most of the posterior samples as shown in Fig.~\ref{fig:gw170729_rec}, leading to a negative match value. It could be due to the noise artifacts. The antiphase could occur due to poor constraint of event time in the detector frame. The total width of event time distribution for H1, L1, and V1 are 19, 18, and 63 ms, respectively. The spread for V1 data is significantly larger than the other two. It is impossible to draw an appropriate conclusion about whether the deviation presents or not when the SNR is very low. Thereby, we also compute the network match excluding the contribution of V1. The median network match for two detectors is 0.88, and the mode is 0.89. An extensive study on agreement between the CBC template waveform and the $\textsc{BayesWave}$ reconstruction also found the similar results~\cite{Chatziioannou:2019dsz}, where the contribution of V1 data was not considered for this study. We excluded the contribution of V1 for GW170729 to produce Fig.~\ref{fig:gwtc1_match}. As a further investigation, we project the H1 and L1 signal over V1 based on the sky location and polarization angle of parameter estimation samples. We follow these steps to obtain the projected signal: (a) perform the inverse transformation of whiten procedure over H1 and L1 reconstructed signals and apply time-shift based on the sky location to obtain these signals in geocentric coordinate, (b) obtain the two GW polarizations based on the sky location and polarization angle, and (c) project on the frame of V1 detector as shown in the bottom panel (blue dashed) of Fig.~\ref{fig:gw170729_rec}. We can see that the projected signal and CBC template waveform are in phase, and the median match is 0.64. It implies that the noise probably plays a role in V1 reconstructed signal being antiphase.
\section{Conclusion}
\label{sec:Consclusion}
This paper presents a wavelet-based semi-model-dependent method for reconstructing the gravitational wave signals produced from compact binary mergers. We have employed the framework of continuous wavelet transformation, where Morlet wavelets represent the signals. The semi-model-dependent approach determines the essential wavelets using the posterior samples from parameter estimation to reconstruct the signals from the data.
In general, the wavelets are constructed using an output of octave scale, which provides the tightest set of wavelets. Such wavelets yield a nearly orthogonal wavelet basis. In this paper, we have proposed a log-uniform scale for constructing the wavelets. Such wavelets are highly redundant, i.e., non-orthogonal wavelets. However, this new scale is more efficient for representing the linear chirp signals at high frequencies than the octave scale. As the wavelets are regarded as time localized bandpass filters, a reconstructed signal always contains a nominal amount of noise that passes through the essential wavelets. The amount of noise depends on the area covered by the essential wavelets over the time-frequency plane. We have shown that the essential wavelets with the log-uniform scale occupy a smaller space than the octave scale, which enables us to reconstruct the weak signals better.
We have conducted injection analysis to evaluate the reconstruction efficiency for the signals from binary black hole mergers by computing the match between injected waveform and the reconstructed signal. The reconstruction accuracy increases with the SNR such that the mismatch is approximately proportional to SNR squared, $1 - \mathscr{M} \gtrsim 1/\mathrm{SNR}^2$. Also, the reconstruction accuracy strongly depends on chirp mass. We have seen that the mismatch at a fixed SNR decreases as chirp mass increases.
We have demonstrated the ability to detect the deviation where the injected waveform is outside the region of space enclosed by the search template waveform manifold. We have performed the parameter estimation analysis using $\textsc{LALInference}$ by injecting an electric BBH waveform, where the template waveforms are generated by a quasi-circular waveform model $\mathrm{IMRPhenomXP}$. After that, we studied the match between the injected waveform and the reconstructed waveform. As reported in Table~\ref{tab:matchComparison}, the $\textsc{LALInference}$ waveform agrees less with the injected waveform than the wavelet reconstruction. Also, the $\textsc{LALInference}$ waveform is out of phase over a bit of the region. In comparison, the amplitude and phase of the reconstructed waveform are nearly consistent with the injected waveform.
We have applied this wavelet-based reconstruction analysis to each binary black hole merger event in GWTC-1. We have seen a satisfactory agreement between the reconstructed signal and the estimated theoretical waveform. As the gravitational wave catalog grows, we expect the wavelet-based semi-model-dependent reconstruction method to provide a more precise view of the agreement between the observed data and the waveform model.
There are many avenues for future work and extensions of this method: combining the octave and log-uniform scales to have an optimal method for reconstructing the signals from compact binaries; applying this new method for reconstructing the signal from binary neutron star mergers; developing appropriate statistics for detecting the deviations and investigating the match study when instrumental glitch presents in on-source data; applying this method on GWTC-2 and GWTC-3~\cite{GWTC2, 2021arXiv211103606T}.
\begin{acknowledgments}
I gratefully acknowledge Ayatri Singha, Sudarshan Ghonge, M.K. Haris, Amit Reza, Khun Sang Phukon, Chinmay Kalaghatgi, Bhooshan Gadre, Melissa Lopez Portilla, Chris Van Den Broeck for helpful comments and suggestions. S.R. was supported by the research program of the Netherlands Organization for Scientific Research (NWO). The material of this paper is based upon work supported by NSF's LIGO Laboratory, which is a major facility fully funded by the National Science Foundation (NSF). I gratefully acknowledge computational resources provided by the LIGO Laboratory and supported by the NSF Grants No.~PHY-0757058 and No.~PHY-0823459. This research has made use of data, software and/or web tools obtained from the Gravitational Wave Open Science Center, a service of LIGO Laboratory~\cite{GWOSC:catalog}, the LIGO Scientific Collaboration and the Virgo Collaboration.
\end{acknowledgments}
|
\section{INTRODUCTION}
Advection Diffusion Reaction (ADR) Partial Differential Equations (PDEs) are widely used in a lot of disciplines for modeling of various phenomena. The PDEs found in these various application is however deviant from the standardized linear-constant-parameter versions for which analytical solutions exist. Consequently, numerical solution is commonly employed which is usually computationally very expensive and inadequate to real-time applications. Nevertheless, solving ADR PDEs in real time is critical in achieving monitoring and control goals for numerous applications \cite{Elkhashap.2019b,Elkhashap.2021b}. For example, one of these applications is water quality monitoring in water distribution networks which gained a lot of momentum in the past few decades. Open source as well as commercial tools propose several methods for water quality modeling all while considering the used models computational complexity \cite{WQ_Epanet,rossman1999epanet,RTP_Epanet_1}. However, such models are mostly constrained by the highly simplifying assumptions and can usually only handle a restricted set of operating conditions. For example until now the widely used EPANET tool developed by the U.S. Environmental Protection Agency \cite{rossman1999epanet} for hydraulic and water quality simulation of water distribution network uses a simplified ADR equation neglecting the diffusion phenomena in disinfectant propagation. Moreover, the classical junction-link \cite{rossman1999epanet} representation of the water network disregards an accurate representation of the spatial variability of the properties under consideration. Other methods exist, which consider various effects such as diffusion, e.g. \cite{SHang_ADR}, however such methods normally suffer from high computational complexity specially when a fine spatial discretization is adopted. Tools offering high spatial resolution solutions combined with real-time suitable computational effort are to the authors knowledge still not well investigated. In most of the present tools the focus is either laid upon accurate solutions considering the spatial variability, e.g. Computational Fluid Dynamics (CFD) and Finite Element Method (FEM) \cite{ABOKIFA2016107,matlabPDEtool} tools, or the real time capability of the solution method through crude approximation, e.g. as in crude spatial lumping in EPANET \cite{rossman1999epanet}. Recent efforts for the consideration of the water quality with chosen spatial resolution for real-time control purposes can be also found. For example in \cite{WQModels4MPC} state space models of water networks are constructed with direct spatio-temporal discretization of ADR PDEs and then directly used for model-based control. Moreover, efforts regarding the computational complexity reduction can be found. For example in \cite{WQMOR} the network state-space models are joined and reduced using adapted projection based techniques. However, the strategy of collectively modelling water networks with direct spatio-temporal discretization is unsuitable for a network size scale-up or fine discretization cases. Moreover, the employed Model Order Reduction (MOR) technique for such Network monolithic models causes the inability to fully preserve the structure, properties and parametric dependence of the Full Order Model (FOM) within the Reduced Order Model (ROM). Hence special adaptations in the reduction method are needed to guarantee the conservation of the essential properties, e.g. the ROM stability in \cite{WQMOR}. In this contribution, a general approach for the solution and model order reduction of ADR-PDEs with time-varying coefficients is proposed. The approach is based on using augmentations in the FOM formulation reaching a standard bilinear form on which system theoretic MOR methods for bilinear systems can be employed \cite{MORbilinH2,Elkhashap.2021a,Elkhashap2022realtime}. The proposed approach preserves the FOM structural interpretability and dependency on the time varying coefficients allowing for its use in real-time estimation and control purposes. Moreover, the proposed method inherits by construction the qualities of minimal reduction error, stability, and error bounds from the employed $\mathcal{H}_2$ norm MOR for bilinear systems \cite{redmann2021bilinear}. Furthermore, the approach is evaluated empirically in two examples including one for water quality modeling comparing the ROM solution to a FOM solution using the MATLAB's PDE FEM solver \cite{matlabPDEtool} as ground truth. In the water quality modeling example the EPANET tool solution is used in comparison highlighting the effect of the high spatial resolution and inclusion of the diffusion effects. The contribution is organized as following. First the general ADR PDE with time varying coefficients is introduced in section \ref{sec:1} highlighting the adopted solution method reaching a standard bilinear form. Second in section \ref{sec:2}, the MOR method is briefly introduced illustrating the produced ROM. Third in section \ref{sec:3}, the two evaluation scenarios of the ROM is are illustrated also introducing the water quality model utilized for the second scenario. Finally, the results are shown and discussed in section \ref{sec:4} followed by a brief conclusion.
\section{Methods}
\subsection{ADR PDE with time varying coefficients}\label{sec:1}
\begin{figure}[b!]
\centering
\vspace{-0.5cm}
\resizebox{\linewidth}{!}{
\input{Figures/FVDisc.tex}
}
\vspace{-0.8cm}
\caption{Schematic of ADR equation employed spatial discretization and boundary conditions handling}
\vspace{-0.5cm}
\label{fig:ADR_Disk}
\end{figure}
The main equation under consideration is the inhomogeneous ADR equation in one spatial dimension defined over a finite domain $x\in [0,L]$. The PDE governing a property $q(x,t),\quad t \in \mathbb{R}^+$ can be expressed as follows
\begin{equation}
\pdrev{q}{t}=-v(t)\pdrev{q}{x}+D(t)\pdrev{^2 q}{x^2}+r(t)q+s(t),\label{eq:ADR_PDE}
\end{equation}
with the propagation velocity coefficient $v(t)$, Diffusion coefficient $D(t)$, reaction rate $r(t)$, and source term $s(t)$. The property values at the boundaries are assumed to be directly known (Dirichlet boundary condition). This is as the junctions of the PDE domain (boundaries) are treated separately as illustrated in Fig. \ref{fig:ADR_Disk} using a mass conserving Robin-type boundary condition (cf. open boundary in \cite{Danckwerts.1995})
\begin{align}
\sum\limits_{j\in \mathcal{I}_{in}} \Phi_{j}(x_b,t) -\sum\limits_{j\in \mathcal{I}_o} \Phi_{j}( x_b,t)=0,\\
\Phi(x,t)=v(t)q(x,t)-D(t)\pdrev{q}{x}(x,t)\label{eq:FluxFcn},
\end{align}
with the flux $\Phi(x,t)$ (\ref{eq:FluxFcn}) crossing the respective boundary at $x_b$, the set of the property inflow $\mathcal{I}_{in}$, and set of the property outflow $\mathcal{I}_{o}$. The partial derivatives of the property in the flux functions are approximated using finite differences after introducing the function for the property at the respective boundary $g_b(t)$. Hence, for junctions with multiple in- and outflux algebraic conservation equations are to be solved (as shown in Fig. \ref{fig:ADR_Disk}). The boundary conditions are defined in this general way to allow constructing networks with arbitrary topology whose links are represented by ADR PDEs. The advantage of this formulation is that the diffusion across the junctions is considered, which is important for cases of networks with zero advection, e.g. dead ends in water distribution networks. Note that, the property state at the nodes of the network can also be represented with separate ODEs analogously if accumulation in junctions is to be considered. A finite difference based Method of Lines (semi-discretization) is applied for the PDE solution \cite{Abgrall.2017}. The spatial variable is discretized on a uniform grid of $N$ points with segment length $\Delta x=\frac{L}{N-1}$. The convective term is approximated using first order upwind scheme and the diffusive using central difference. The boundary conditions are handled using a ghost point technique. For illustrative simplicity, only the case of a single inflow flux $g(t)$ with positive flow velocity and an open right boundary is considered hereafter. Consequently, the following high dimensional nonlinear ODE system is produced
\begin{equation}
\dot{\vek{q}}=(v(t)\matr{Q}_1+D(t)\matr{Q}_2+r(t)\matr{Q}_3)\vek{q}+\matr{b}_1s(t) +\matr{b}_2{\varphi}(t),
\end{equation}
with the vector $\vek{q} \in \mathbb{R}^{N}:=[q(i\Delta x ,t)]^\mathrm{T},\quad i=\{1,\cdots,N\}$
\begin{subequations}\label{eq:FOMBilin}
\begin{equation}
\matr{Q}_1=\frac{1}{\Delta x}\small{
\begin{bmatrix}
-1 &0 &\ldots&0\\
1 & -1 &\ddots&\vdots\\
\vdots&\ddots&\ddots &0\\
0 &\ldots&1 &-1
\end{bmatrix}},
\matr{Q}_3=\matr{I},\,
\vek{b}_1=\begin{bmatrix} 1\\ 1\\ \vdots\\1\end{bmatrix},
\end{equation}
\begin{equation}
\vek{Q_2}=\frac{1}{\Delta x^2}\small{
\begin{bmatrix}
-2 &1 &\ldots&0\\
1 & -2 &\ddots&\vdots\\
\vdots&\ddots&\ddots &1\\
0 &\ldots&1 &-1
\end{bmatrix}},\,
\vek{b}_2=\frac{1}{\Delta x^2}\begin{bmatrix} 1 \\ 0\\ \vdots\\0\end{bmatrix},
\end{equation}
\end{subequations}
and the function $\varphi(v(t),g(t),D(t),q(\Delta x,t))$ representing the dependency due to the right boundary condition after discretization. This dependency is formulated according to the handling method, e.g. ghost point, Taylor-based expansion. A ghost point handling with partial derivative approximation at the right boundary point delivers $\varphi(t)=\Delta xv(t)g(t)+D(t)g(t)$. Now (\ref{eq:FOMBilin}) can be rearranged into a standard bilinear form by introducing the augmented input vector $\vek{u}:=[v(t),D(t),r(t)+1,s(t),\varphi(t)]^\mathrm{T}$
\begin{subequations}
\begin{equation}
\dot{\vek{q}}=\matr{A}\vek{q}+ \vek{\mathcal{Q}}^{[1]} \vek{u}\otimes\vek{q} +\matr{B}\vek{u},\label{eq:FOM}
\end{equation}
\begin{equation}
\vek{y}=\matr{C}\vek{q}\label{eq:FOM_out}
\end{equation}\label{eq:FOM_Tot}
\end{subequations}
with the system matrix $\matr{A} \in \mathbb{R}^{N \times N}=-\matr{I}$, input matrix $\matr{B}\in \mathbb{R}^{N\times5}=[\matr{0}_{N\times 3},\,\vek{b}_1,\vek{b}_2]$, and measurement matrix $\matr{C} \in \mathbb{R}^{N \timesN}$ assumed to be identity $\matr{I}$ indicating that a measurement is available at each point. The bilinear term is expressed using the Kronecker product notation $\otimes$, where $\matr{\mathcal{Q}}^{[1]}\in \mathbb{R}^{N\times 5N}$ is the mode-1 matricization of the $3^{rd}$ order tensor $\matr{\mathcal{Q}}\in\mathbb{R}^{N\timesN\times 5}$ having the matrices $\matr{Q}_i \in \mathbb{R}^{N\timesN}, \forall i\in\{1,..,5\}$ as frontal slices, i.e. $\matr{\mathcal{Q}}^{[1]}=[\matr{Q}_1,\matr{Q}_2,\matr{Q}_3,\matr{0}_{N\times3N}]$. The choice of the augmented input vector allowed a system matrix $\matr{A}$ with strictly negative eigenvalues of $-1$ with algebraic multiplicity $N$. Hence, \refeq{eq:FOM_Tot} represents a Bounded Input Bounded Output (BIBO) stable bilinear system \cite{MORbilinH2Zhang.2002,Redmann2019TheML} in standard form. The next step is to perform a projection based MOR method with the lowest possible reduction error.
\subsection{Model Order Reduction}\label{sec:2}
Classic projection-based reduction methods are mainly based on the time space separation assumption when considering the original PDE. Generally, A PDE solution satisfying the separation property can be expressed using an infinite series of the inner product between the spatial and temporal bases. Hence, the PDE solution can be approximated by an inner product between spatial and temporal modes. Namely, a truncation of such infinite series to only a finite number of terms with the selection of the most influential spatial and temporal modes delivers a good approximation of the solution. Empirical methods, e.g. POD-Galerkin, Trajectory Piece Wise Linear (TPWL) methods \cite{MORParamOverview}, rely on simulation data in finding the bases delivering the most accurate reconstruction of a full order dynamics. However, such methods (specially for nonlinear systems) have no guarantees nor bounds regarding the reduction error. Moreover, it is challenging in such methods to preserve the original system structure and parametric dependency. Hence these methods generally work when the ROM is used with the parameter and input profiles in the vicinity of the ones used in the reduction. In the context of water quality modeling and reduction special adaptations are needed to achieve a ROM inheriting the stability of the FOM \cite{WQMOR}. On the other hand, system theoretic methods introduce more rigor in the reduction procedure \cite{MORParamOverview}, where for some special cases notions of optimal reduction and ROM error bounds can be achieved. The main method of interest employed here is the $\mathcal{H}_{\text{2}}$ norm reduction method for bilinear systems. The main idea of the method is to apply a Petrov-Galerkin projection based reduction, i.e. double sided projection using the test and trial bases matrices $\matr{V},\, \matr{W} \in \mathbb{R}^{N\times n}$, on the FOM delivering the following ROM equivalent
\begin{subequations}
\begin{equation}
\vek{{\dot{\hat q}}}= \vek{\hat{A}} \vek{\hat q}+ \vek{\mathcal{\hat{Q}}}^{[1]} \vek{u}\otimes \vek{\hat q} + \vek{\hat{B}} \vek{u},\label{eq:ROM}
\end{equation}
\begin{equation}
\vek{y}= \vek{\hat{C}} \vek{\hat q}\label{eq:ROM_out}
\end{equation}
\end{subequations}
with the reduced state space vector $\vek{\hat q}\in \mathbb{R}^{n}$, reduced order system matrices $ \vek{\hat{A}} \in \mathbb{R}^{n\times n}$, $ \vek{\mathcal{\hat{Q}}}^{[1]} \in \mathbb{R}^{n \times 5n}$, $ \vek{\hat{B}} \in \mathbb{R}^{n \times 5}$, and $\vek{\hat{C}} \in \mathbb{R}^{n\times n}$ as follows
\begin{align}
& \vek{\hat{A}} =\vek{T}\vek{A}\vek{V},\quad \vek{\hat{B}} =\vek{T}\vek{B},\quad \vek{\hat{C}} =\vek{C}\vek{V},\\
& \vek{\mathcal{\hat{Q}}}^{[1]} = \vek{T} \vek{\mathcal{Q}}^{[1]} (\vek{I}_{m\times m} \otimes \vek{V}) ,\quad \vek{T}=(\vek{W}^\mathrm{T}\vek{V})^{-1}\vek{W}^\mathrm{T}.
\end{align}
The error between the FOM and ROM is formulated in terms of the $\mathcal{H}_{\text{2}}$ norm then used to find the projection bases leading to minimum error. Several methods are proposed for finding the bases, e.g, as constrained minimization in \cite{MORbilinH2Zhang.2002}, iterative algorithms in \cite{MORbilinH2}. Here a slightly modified version of the Sylvester equation based algorithm in \cite{MORbilinH2} is used. The modification is mainly in the progression criteria of the algorithm iteration, where instead of using the change in the eigenvalues of the ROM matrix $ \vek{\hat{A}} $, the representation of the $\mathcal{H}_2$ norm of the error system proposed in \cite{Redmann2019TheML} (error bound expression) is calculated explicitly and used as the algorithm progression criteria. This adaptation introduces significant increase in the algorithm computational effort as the explicit error representation requires the explicit calculation of the FOM, ROM, and cross system Gramians. The algorithm computational efficiency is however in our case irrelevant as it is used only in an offline step. The main advantage of the method is the preservation of the FOM structure and parametric dependency without relying in any way on empirical simulation data. Moreover, recent results on the relation between the $\mathcal{H}_{\text{2}}$ norm error and the explicit output error including bounds on the output error can be found in \cite{redmann2021bilinear,Redmann2019TheML}. The reduction method is applied to a time and space scaled version of the FOM (\ref{eq:FOMBilin}) with $N=500$ and $n=8$. The spatial modes emergent after the convergence of the reduction method are shown in Fig. \ref{fig:MORModes}.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{MOR_Modes_.pdf}
\vspace{-0.6cm}
\caption{Spatial Modes for the right (bottom) and left (top) projection matrices, i.e. columns of $\matr{V}$, rows of $\matr{T}$, resulting from the reduction method}
\vspace{-0.3cm}
\label{fig:MORModes}
\end{figure}
Besides the theoretical guarantees of the ROM accuracy, the ROM is also empirically evaluated in a set of scenarios presented in the next section.
\section{Case Studies}\label{sec:3}
The method elaborated in the previous sections is now evaluated in two scenarios including one of real world applications namely:
\begin{enumerate}
\item A general ADR equation with time-varying signals as its coefficients,
\item A water quality modeling and simulation test case.
\end{enumerate}
The simulation of the FOM for the different scenarios is performed using Matlab's FEM based PDE toolbox \cite{matlabPDEtool}. The tool requires the PDE to be solved in divergence form. It also requires an explicit time functional form of the inputs time varying profiles. Hence, for the time varying input profiles either a functional approximation is used or interpolation routines are used for the piece-wise changing functions, e.g. demand patterns from the EPANET software. For the implementation of the ROM time integration schemes CasADi \cite{Andersson2019} toolbox is employed utilizing the 3rd order legendre collocation scheme. In the following the two scenarios are elaborated and then the results are presented.
\subsection{Simulation-based Evaluations}
For the first scenario the ROM of the general ADR model is evaluated against a high fidelity simulation for a wide range of flow conditions with varying P\'{e}clet\ number. Moreover, a classical POD-Galerkin reduction (being the standard method for nonlinear systems) is compared against the proposed approach highlighting the efficacy of the proposed method. The time-varying coefficients are chosen to be either chirp signals of varying amplitude and frequency or a superposition of harmonics, steps, exponential growths, and decays.
\subsection{Water Quality Modeling Example}
There are a number of aspects which could be considered for the modeling of the water quality such as the disinfectant concentration or biological activity of bacteria within the water. Here the popular disinfectant concentration aspect is considered. The disinfectant concentration within the water flowing in the pipe is modelled by a standard ADR PDE (\ref{eq:ADR_PDE}) with time varying coefficients. However, the PDE coefficients in such case are interdependent. Namely, the axial diffusion coefficient $D(t)$ as well as the reaction rate $r(t)$ are functions of the flow regime, i.e. flow velocity $v(t)$. The exact dependency for the diffusion and reaction parameters $D(t),r(t)$ (and also effect of other variables as water temperature) is a separate research topic heavily relying on empirical studies \cite{DParamLee2004MASSDI,Rparam_kw,Dparam_Taylor_turb}. For the diffusion rate in laminar regime the relation proposed in \cite{DParamLee2004MASSDI} is used here. Moreover, for the turbulent regime the empirical formula achieved by the gene expression technique proposed in \cite{D_Gene_Exepr} is used due to its simplicity and acceptable accuracy. Hence, the total expression for the diffusion coefficient is as follows
\begin{equation*}
D(t)=\begin{cases}\frac{a^2v(t)^2}{48D_{0}}(1-\exp{-12.425\frac{\tau D_{0}}{a^2}}) & \Rey<2400 \\\frac{c_1(2v(t)+c_2)}{\Rey\,v(t)} & \Rey \geq 2400
\end{cases}
\label{eq:DispWQ}
\end{equation*}
with the pipe radius $a$, Reynolds number $\Rey$, molecular diffusion coefficient $D_0$, the gene expression empirical formula constants $c_1,c_2$, and the Lagrangian time $\tau$, which is calculated in the simulation by averaging the velocity and dividing by the pipe total length $L$. The reaction coefficient $r(t)$ is decomposed into its two main contributors, the bulk reaction rate $k_b$ and the wall reaction $R_w(t)$. The classical dimensionless analysis relations of mass transfer \cite{Rossman1} are used to express the wall reaction rate in dependence of $v(t)$:
\begin{subequations}
\begin{equation*}
r(t)=k_b+\underbrace{\frac{k_w k_f(t)}{a(k_w+k_f(t))}}_{R_w(t)},\, k_f=\Shear\frac{D_0}{2a},
\end{equation*}
\begin{equation*}
\Shear=\begin{cases} 3.65+\frac{0.0668(a/L)(\Rey\,\Schm)}{1+ [0.04(a/L)(\Rey\,\Schm)]^{2/3}} & \Rey <2400\\
0.023\,\Rey^{0.83}\Schm^{0.333}& \Rey\geq 2400
\end{cases}
\end{equation*}
\end{subequations}
with the wall decay constant $k_w$, mass transfer coefficient $k_f$, Sherwood number $\Shear$, and Schmidt number $\Schm$. After constructing the model and its inputs dependencies, a ROM is constructed and used for the prediction of the disinfectant concentration within an exemplary water distribution path.
As one of the purposes of this contribution is to offer an enhancement to the existing water quality modeling and simulation tools, EPANET open source tool is used to construct the example. Moreover, the quality analysis produced by EPANET is compared to the one produced by the ROM and the high fidelity solution produced by MATLAB PDE solver highlighting the effect of considering the diffusion phenomena. The exemplary network is an adapted version of the one used in \cite{SHang_ADR} with consideration of a random demand pattern, which also ensures that both turbulent and laminar regimes are included. The network considered for the example is shown in Fig. \ref{fig:EPA_net}.
\begin{figure}[h!]
\centering
\resizebox{\linewidth}{!}{
\input{Figures/EPA_ex_net}
}
\vspace{-0.8cm}
\caption{Exemplary Water Distribution Path considered for the test case}
\vspace{-0.2cm}
\label{fig:EPA_net}
\end{figure}
The network is composed of a source node, where chlorine concentration is injected, and 10 pipes with half meter diameter and length of 100 m (a total path of 1000 m). The chlorine concentration is initially set to zero then the profile shown in Fig. \ref{fig:Inprof} is injected for a 48 hour simulation time. The injected concentration profile includes a one hour pulse of 10 [mg/L] followed by a 16 hour zero injected concentration then finally a unit step superimposed by a random noise for the last 27 hours period. The case study includes periods for both laminar as well as turbulent flow regimes inducing the variability of the different regimes of the coefficients $D(t),r(t)$. Figure \ref{fig:Inprof} shows the profiles used for the 48 hour simulation, moreover the model parameters used for simulation are summarized in Table \ref{tb:Parmeter}.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{INProf_in.pdf}
\vspace{-0.7cm}
\caption{Injected chlorine concentration profile (top) along with the ADR coefficients profiles used in the water quality simulation}
\vspace{-0.6cm}
\label{fig:Inprof}
\end{figure}
\renewcommand{\arraystretch}{1}
\begin{table}[h!]
\vspace{-0.4cm}
\centering
\caption{Parameter values used for the case study}
\begin{tabular}{cccc}
Parameter & Value & Unit\\\hline
$a$&$0.5$&$\mathrm{m}$\\
$D_0$&$1.2\times10^{-9}$& $\mathrm{m^2/sec}$\\
$c_1$ & $4110$ & -\\
$c_2$ & $0.062$ & -\\
$k_b$ & $6.36\times 10^{-5}$ & $\mathrm{1/sec}$\\
$k_w$ & $8.33\times 10^{-5}$ & $\mathrm{m/sec}$\\
$N$ & $500$ &-\\
$n $ &$8$ &-\\
\hline
\end{tabular} \label{tb:Parmeter}
\vspace{-0.4cm}
\end{table}
\section{Results}\label{sec:4}
The results of the two evaluation cases are summarized hereafter. Due to the limited place only few result plots for the first case will be presented.
\subsection{Arbitrary Input Profiles Simulations}
The results for one selected case are shown in Fig. \ref{fig:TRsurf}. Moreover, the deviation between the ROM and FOM prediction is quantified using Normalised Mean Square Error (NMSE) in percentage and summarized for all scenarios in Table \ref{tab:Results} (with the shown result in Fig. \ref{fig:TRsurf} highlighted in the table). As expected, the POD-Galerkin based ROM showed a large steady state error or unstable behavior for a large number of the chosen scenarios. This is as the POD-Galerkin based ROM performance is generally limited to the behavior introduced in the snapshots matrix. And for such rich system represented by the ADR system with the time variable coefficients it was not possible to capture the dependency (dynamic and static) of the varying coefficients fully in the snapshots matrix. On the other hand the ROM with the proposed $\mathcal{H}_2$ method could produce stable and accurate prediction for almost all cases for P\'{e}clet\ number up to $10^{5}$. Oscillatory behavior within the ROM behavior is observed for high P\'{e}clet\ number, i.e. $\Pecl>10^2$ which is expected due to the general limitation of the global bases methods in combination with extremely high P\'{e}clet\ \footnote{P\'{e}clet\ number is a non dimensional quantity equal to the ratio between rate of advection and rate of diffusion, hence, represents a quantification for the dominating flow phenomena.} number and Eulerian frame of reference (see \cite{Elkhashap.2021b})
\newcommand{\vln}[1]{\multicolumn{1}{c|}{#1}}
\newcommand{\TabStack}[1]{{\multirow{2}{*}{\shortstack{#1}}}}
\definecolor{TabShade}{rgb}{0.00000,0.38039,0.39608}
\newcommand{\cellcolor{TabShade!20}}{\cellcolor{TabShade!20}}
\newcommand{\STAB}[1]{\begin{tabular}{@{}c@{}}#1\end{tabular}}
\begin{table}[h]
\caption{Results of the two ROMs accuracy compared to the FOM for all simulations at different P\'{e}clet\ number}
\vspace{-0.2cm}
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{0.65}
\centering
\begin{tabular}{cccccccc}
\toprule
& \multicolumn{1}{c}{\multirow{3}{*}{\begin{tabular}{@{}c@{}}$\Pecl$ \\ Range\end{tabular}}} & \multicolumn{3}{c}{\shortstack{$\mathcal{H}_2$ Norm Optimal ROM\\NMSE $\UnitBrackets{\percent}$}} & \multicolumn{3}{c}{\shortstack{POD-Galerkin ROM\\ NMSE $\UnitBrackets{\percent}$}} \\ \cmidrule(r){3-5} \cmidrule(r){6-8}
& \multicolumn{1}{c}{}
& Step & Pulse & Random & Step & Pulse & Random \\ \midrule
\vln{\multirow{6}{*}{\STAB{\rotatebox[origin=c]{90}{$n=8$}}}} &
\vln{$10^0$} & $0.00009$& $0.006$& $0.0003$& $207.6$& $80.1$& $30.2$\\ \vln{}&
\vln{$10^1$} & $0.01726$& $0.731$& $0.0025$& $167.9$& $13.4$& $5.70$\\ \vln{}&
\vln{$10^2$} &\cellcolor{TabShade!20}$0.24527$& $9.355$& $0.0219$&\cellcolor{TabShade!20}$75.39$& $12.6$& $5.90$\\ \vln{}&
\vln{$10^3$} & $2.06860$& $46.59$& $0.4077$& $64.57$& $40.7$& $6.42$\\ \vln{}&
\vln{$10^4$} & $3.11280$& $65.68$& $0.6161$& $63.84$& $60.3$& $6.71$\\ \vln{}&
\vln{$10^5$} & $3.25190$& $68.28$& $0.8092$& $63.79$& $63.1$& $6.77$\\
\bottomrule
\end{tabular}\label{tab:Results}
\vspace{-0.5cm}
\end{table}
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{TRsurf.pdf}
\vspace{-0.6cm}
\caption{The predicted concentration surface by the FOM (blue), $\mathcal{H}_2$ optimal ROM (green) and POD-Galerkin ROM (red) with the corresponding absolute error for a step change in injected concentration (middle left) and arbitrary changing profiles for $v(t),D(t),r(t),s(t)$ (bottom)}
\vspace{-0.5cm}
\label{fig:TRsurf}
\end{figure}
\subsection{Water Quality Modeling}
The FOM as well as ROM full spatial resolution prediction along with the error between both are shown in Fig. \ref{fig:WQsurf}. Moreover, the EPANET predicted quality plotted against both the FOM and ROM along the length of the water path (only nodes for EPANET) at certain time points are shown in Fig. \ref{fig:Zplots}.
It can be observed that the ROM maps the FOM prediction very well with minimal error mainly represented at the boundary (clf. Fig.\ref{fig:WQsurf}). Furthermore, the EPANET quality prediction can be seen as green spikes propagating through time and space in the left plot in Fig. \ref{fig:WQsurf}. This behavior is due to the combined effect of the sparse spatial resolution (10 nodes with 100 meter distance) and the neglection of the diffusive part of the flow. It can be also observed that the EPANET neglection of diffusion causes a significant deviation to the high fidelity simulation specially at periods of low P\'{e}clet\ number (dominating diffusion). For example underestimation of the concentration profile can be observed at the downstream nodes for several time instants, e.g. $t=2,3,19\, \mathrm{hr}$ (see Fig. \ref{fig:Zplots}), as the advective flow at these time instants has not yet propagated the concentration. However in reality as shown by the simulation including the diffusion phenomena, the concentration would have already diffused through these nodes at the specified times. On the other hand there are instants with overestimation of the concentration, e.g. $t=30\, \mathrm{hr}$ in Fig. \ref{fig:Zplots} (also observe peaks in Fig. \ref{fig:WQsurf}) as the neglected diffusion phenomena usually causes an evening out effect to sharp gradients contrary to pure advection wave propagation behavior. The deviation of the ROM from the high fidelity simulation is observed to be low for both regimes with very mild oscillations at certain time instants at abrupt changes near the left boundary (observe the error plot in Fig. \ref{fig:WQsurf}).
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{WQsurf.pdf}
\vspace{-0.7cm}
\caption{The predicted concentration surface for the 48hr simulation time (left) with the ROM (red), high fidelity simulation (blue) and EPANET (green), absolute error between FOM and ROM (right)}
\vspace{-0.2cm}
\label{fig:WQsurf}
\end{figure}
The ROM NMSE is below $2.3\,\%$ indicating a high prediction accuracy. A computational time reduction from $3487.5409$ sec needed by the FOM to $0.1507$ sec needed by the ROM could be achieved (on a Windows 10 development computer, Intel(R) Core(TM) i7-7700HQ CPU@2.8GHZ, 8GB RAM). This indicates computational speedup reaching a factor of four orders of magnitude compared to the MATLAB's PDE toolbox. However, it is not fair to compare a ROM with predetermined fixed steps calculations with high fidelity scheme with step size adaption FEM tool. However, the ROM computational complexity indicates the real-time potential of the ROM. The computational time needed for the EPANET water quality analysis including the hydraulic analysis using the matlab interface is $2.576$ sec.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{Zplot.pdf}
\vspace{-0.7cm}
\caption{The predicted concentration profiles with the ROM (dotted line), high fidelity simulation (solid line), and EPANET (asterisk Marker) at different time instants}
\vspace{-0.8cm}
\label{fig:Zplots}
\end{figure}
\section{CONCLUSIONS}
A model order reduction method for Advection Diffusion Reaction (ADR) PDEs with time varying coefficients is proposed. The method is based on formulating the PDE into a standard bilinear high dimensional ODE using semi-discretization and augmentation of the input vector. Then system theoretic projection based model order reduction method, namely $\mathcal{H}_2$ norm optimal method for bilinear system, is employed for reducing the high dimensional system. The method is evaluated for two simulation based test cases against a reference solution produced using MATLAB's Finite Element Method PDE solver. A real world water quality simulation scenario is constructed for the second evaluation scenario.The water quality prediction of the second scenario is also generated using EPANET tool and compared against the high spatial resolution prediction highlighting the advantage of considering the further effects of which the proposed method offers. Finally the ROM could achieve a significant speedup compared to the high fidelity simulation solving the 48 hour, 1000 m water quality prediction problem within $0.1507$ seconds with normalised mean square error below $2.3\,\%$. The shown results introduce the potential of combining selective model formulation with structure preserving model order reduction techniques in high spatial resolution real-time monitoring of water quality networks.
\addtolength{\textheight}{-12cm}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{section:introduction}
Pre-Training has revolutionized the way computational models are trained in the natural language processing (NLP) community~\cite{devlin2019bert,radford2019language,peters2018deep,clark2019electra}. For a long time, supervised learning has been the most successful natural language learning paradigm. The pioneers of the pre-training idea challenged this view by showing that a vast amount of general knowledge about language, including both linguistic and commonsense knowledge, can be acquired by (pre-)training a model in a {\em task-agnostic} manner using {\em self-supervised} learning tasks. Self-supervised learning tasks are NLP tasks for which the label associated with a training instance can be derived automatically from the text itself. Consider, for instance, one of the most well-known self-supervised learning tasks, Masked Language Modeling (MLM)~\cite{devlin2019bert}. Given a sequence of word tokens in which a certain percentage of tokens is {\em masked} randomly, the goal of MLM is to predict the masked tokens. As can be easily imagined, a model for MLM can therefore be trained on instances where each one is composed of a partially masked sequence of word tokens and the associated ``class'' value is the masked tokens themselves. Because no human annotation is needed, a model can be pre-trained on a very large amount of labeled data can be automatically generated, thereby acquiring a potentially vast amount of knowledge about language. A pre-trained model can then be optimized for a specific task by fine-tuning its parameters using task-specific labeled data in the standard supervised fashion.
Inspired by the successes of pre-trained models in NLP, a number of pre-trained models for source code have been proposed and applied to a variety of SE tasks including code summarization and code completion, with notable successes~\cite{buratti2020exploring,karampatsis2020scelmo,kanade2020learning,svyatkovskiy2020intellicode,feng2020codebert,guo2020graphcodebert,liu2020multi,mastropaolo2021studying,jiang2021treebert}. Despite these promising results, there are issues surrounding the application of pre-trained models for source code to SE tasks.
First, the majority of these pre-trained models focus on pre-training only the encoder of the Transformer~\cite{kanade2020learning,buratti2020exploring,feng2020codebert,guo2020graphcodebert}. This is not ideal, however. For instance, for generation tasks that are addressed using models with the encoder-decoder architecture, there is no reason why the decoder should be left out in the pre-training process. Second, these models have largely assumed as inputs the source code~\cite{kanade2020learning,karampatsis2020scelmo,svyatkovskiy2020intellicode,buratti2020exploring,liu2020multi} and the associated natural language description~\cite{feng2020codebert,mastropaolo2021studying}. In particular, code structure, which is also crucial to understanding source code, is largely missing from these models. The reason why code structure is left out is that SE researchers have for the most part simply reused the pre-training tasks designed for natural languages when pre-training models (by viewing source code as natural language), but none of these tasks are concerned with learning the structure of
natural language. Third, these pre-training tasks assume the availability of a bilingual corpus, where each method/function (henceforth collectively referred to as method) is ``labeled'' with the corresponding docstring, when pre-training a model on source code and the associated natural language description~\cite{feng2020codebert,guo2020graphcodebert}. However, such a bilingual corpus tends to be small in size compared to a monolingual (i.e., source code only) corpus, thus severely limiting the amount of data a model can be pre-trained on.
In general, we believe the reliance on a bilingual corpus would hinder the development of powerful pre-trained models for source code in the long run, as a key strength of the pre-trained models developed in the NLP community stems from their ability to be trained using self-supervised tasks for which \textit{very large} amounts of labeled data can be automatically generated.
Several attempts have been made to address the three problems mentioned above to different extents. To address the first problem, T5-learning~\cite{mastropaolo2021studying} and TreeBERT~\cite{jiang2021treebert} are proposed, which are sequence-to-sequence (i.e., seq2seq) pre-training models with the encoder-decoder architecture and enables both the encoder and the decoder to be jointly trained. To address the second problem, GraphCodeBERT~\cite{guo2020graphcodebert} employs two pre-training tasks specifically designed to acquire structural information. One task involves predicting the edges in the data flow while the other involves predicting the alignment between the nodes in the data flow and the code sequence, respectively. However, as the authors also pointed out, while the data flow captures information that is largely semantic in nature, it does not capture syntactic information (e.g., the syntactic structure encoded in an Abstract Syntax Trees, i.e., AST), which is arguably the most important type structural information about source code that is commonly exploited by SE researchers.
TreeBERT~\cite{jiang2021treebert} employs the set of constituent paths of ASTs as the input of its encoder. However, it only inputs code sequences at the decoder side during pre-training, trying to make the encoder learn lexical and semantic information (both of which can be easily obtained from the code sequences) from the AST (which contains mainly syntactic information) during the pre-training phase. But it is uncertain whether TreeBERT can extract the complete lexical and semantic information from the AST only by relying on pre-training, thus eliminating the need to input code tokens when fine-tuning.
To address the third problem, T5-learning treats code and natural language as two types of independent data instances. While this allows T5-learning to learn from a monolingual rather than bilingual corpus, the connection between a piece of code and the associated natural language description is no longer present in the corpus. Hence, it is no longer clear whether T5-learning can still learn to produce a natural language description of a piece of code. To our knowledge, there has been no attempt to address all three issues in a single model.
In light of the above discussion, we propose SPT-Code, a new pre-trained model for source code. Motivated by T5-learning, SPT-Code is a seq2seq pre-training model, enabling both the encoder and the decoder of Transformer to be jointly pre-trained. Each data instance for SPT-Code is composed of three types of information derived from a method, namely the code sequence, its AST, and the associated natural language description. Note that the incorporation of ASTs allows SPT-Code to exploit structural, specifically syntactic, information. In addition, to obviate the need to learn natural language descriptions from a bilingual corpus, we will simply use the name of the method and the names of the methods that are invoked in this method as a (very succinct) natural language description of the given source code.
We design three code-specific pre-training tasks for SPT-Code, each of which allows SPT-Code to acquire exactly one of the three types of information that comprise a data instance. The first task is a version the well-known Masked Sequence to Sequence (MASS)~\cite{song2019mass} pre-training task for natural language that we adapt to source code. Specifically, our modified MASS task seeks to acquire knowledge about source code via masking a random fragment of the code tokens. The second task, Code-AST prediction (CAP), is designed to enable the model to gain knowledge of the syntactic structure of a code fragment by predicting whether the given AST is the correct AST for the given piece of code. The final task, Method Name Generation (MNG), is a novel task that involves generating the subtokens of the method name, which we take to be an (extremely concise) natural language description of the method.
After SPT-Code is pre-trained on the CodeSearchNet dataset~\cite{husain2020codesearchnet}, we fine-tune and evaluate it on five downstream tasks, including code summarization, code completion, bug fixing, code translation, and code search. Experimental results show that SPT-Code achieves state-of-the-art results under virtually all circumstances.
In sum, we make the following contributions
\begin{enumerate}
\item Propose SPT-Code, a seq2seq pre-trained model for source code that is built with the encoder-decoder architecture and is applicable to both classification and generation tasks.
\item Extend the input representation of pre-trained models for source code with a simplified and linearized version of ASTs. To our knowledge, we are the first to use both natural language and AST as inputs in pre-training for source code.
\item Design special input representations and three code-specific seq2seq-based pre-training tasks enabling SPT-Code to be pre-trained without relying on any bilingual corpus or labeled data.
\item Pre-train SPT-Code on a large unlabeled monolingual (i.e., source code only) dataset across six programming languages, then fine-tune and evaluate it on five downstream code-related tasks, achieving state-of-the-art results on all tasks.
\end{enumerate}
\section{Related Work}
\label{section:related}
\subsection{Pre-Training Models in NLP and SE}
Table~\ref{table:pre_train_models}
\begin{table*}
\centering
\caption{Pre-training models in NLP and for source code.}
\label{table:pre_train_models}
\resizebox{\textwidth}{!}{
\begin{tabular}{llcccc|cccccccc|ccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Domain}} &
\multicolumn{1}{l}{\multirow{2}{*}{Models}} &
\multicolumn{4}{c|}{Modules} &
\multicolumn{8}{c|}{Objectives} &
\multicolumn{3}{c}{Inputs} \\
\cline{3-17}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{} &
LSTM &
Encoder &
Decoder &
Encoder-Decoder &
Forward LM &
Backward LM &
Masked LM &
NSP &
Permutation LM &
RTD &
- &
- &
NL &
Code &
Structure \\
\midrule
\multirow{8}{*}{NLP} & ELMo (2018) & \checkmark & & & & \checkmark & \checkmark & & & & & & & \checkmark & & \\ \cline{2-17}
& BERT (2019) & & \checkmark & & & & & \checkmark & \checkmark & & & & & \checkmark & & \\ \cline{2-17}
& XLNet (2019) & & \checkmark & & & & & & & \checkmark & & & & \checkmark & & \\ \cline{2-17}
& RoBERTa (2019) & & \checkmark & & & & & \checkmark & & & & & & \checkmark & & \\ \cline{2-17}
& ELECTRA (2019) & & \checkmark & & & & & \checkmark & & & \checkmark & & & \checkmark & & \\ \cline{2-17}
& GPT-2 (2019) & & & \checkmark & & \checkmark & & & & & & & & \checkmark & & \\ \cline{2-17}
& T5 (2020) & & & & \checkmark & & & \checkmark & & & & & & \checkmark & & \\ \cline{2-17}
& BART (2020) & & & & \checkmark & & & \checkmark & & & & & & \checkmark & & \\
\hline
\multirow{7}{*}{Code} & SCEMLo (2020) & \checkmark & & & & \checkmark & \checkmark & & & & & & & & \checkmark & \\ \cline{2-17}
& CuBERT (2019) & & \checkmark & & & & & \checkmark & \checkmark & & & & & & \checkmark & \\ \cline{2-17}
& C-BERT (2020) & & \checkmark & & & & & \checkmark & \checkmark & & & & & & \checkmark & \\ \cline{2-17}
& IntelliCode (2020) & & & \checkmark & & & \checkmark & & & & & & & & \checkmark & \\ \cline{2-17}
& CodeBERT (2020) & & \checkmark & & & & & \checkmark & & & \checkmark & & & \checkmark & \checkmark & \\ \cline{2-17}
& GraphCodeBERT (2020) & & \checkmark & & & & & \checkmark & & & & EP & NA & \checkmark & \checkmark & Data Flow \\ \cline{2-17}
& CugLM (2020) & & \checkmark & & & \checkmark & & \checkmark & \checkmark & & & & & & \checkmark & \\ \cline{2-17}
& T5-learning (2021) & & & & \checkmark & & & \checkmark & & & & & & \checkmark & \checkmark & \\ \cline{2-17}
& TreeBERT (2021) & & & & \checkmark & & & & & & & TMLM & NOP & & \checkmark & AST \\
\bottomrule
\end{tabular}
}
\end{table*}
presents an overview of the most prominent pre-trained models in NLP and for code. Each model is characterized along four dimensions: (1) Modules: what is being pre-trained (e.g., the encoder, the decoder, or both); (2) Objectives: the pre-training objectives\footnote{Forward LM~\cite{peters2018deep,radford2019language,karampatsis2020scelmo} aims to predict the next word given the preceding words in a sentence. Backward LM~\cite{peters2018deep,karampatsis2020scelmo} aims to predict the previous word given the words that appear after it in a sentence. Masked LM is the masked language modeling task described in Section~\ref{section:introduction}. NSP~\cite{devlin2019bert} aims to predict whether the second sentence in a sentence pair should immediately follow the first sentence in the pair. Permutation LM~\cite{yang2019xlnet} aims to predict a word using a set of context words randomly selected via the attention mask mechanism. RTD aims to predict which token in the input has been replaced. EP~\cite{guo2020graphcodebert} involves masking 20\% of the edges in a data flow and aims to predict the masked edges. NA~\cite{guo2020graphcodebert} involves masking a certain portion of edges of connecting data flow nodes and code tokens and aims to predict the masked edges. TMLM~\cite{jiang2021treebert} masks paths in the AST input on the encoder side and tokens in the code sequence input on the decoder side, then predicts the token of the masked code. NOP~\cite{jiang2021treebert} is to exchange the positions of some nodes in the path, and distinguish whether the order of nodes in the AST is correct or not.} and (3) Inputs: what information the model assumes as input (e.g., natural language, code and structural information of the code).
We can make a few observations from Table~\ref{table:pre_train_models}. First, while the majority of work has focused on pre-training the encoder, the newest models (T5~\cite{raffel2020exploring}, BART~\cite{lewis2020bart}, and T5-learning~\cite{mastropaolo2021studying}, which is modeled after T5) are all seq2seq pre-training models that allow the encoder and the decoder to be jointly trained. Second, except for GraphCodeBERT~\cite{guo2020graphcodebert} and TreeBERT~\cite{jiang2021treebert}, all pre-trained models in SE reuse training objectives designed for natural languages, with MLM being the most popular pre-training task. This indicates that the selection of pre-training tasks, which dictates what knowledge will be acquired and exploited by a model, is an area of research that is under-investigated in SE. Finally, while earlier pre-training models in SE assume only source code as input, the later ones all use both code and language.
In our experiments, we will use as our baselines the most recently developed (and also the state-of-the-art) pre-trained models for source code, namely CodeBERT~\cite{feng2020codebert}, GraphCodeBERT~\cite{guo2020graphcodebert}, CugLM~\cite{liu2020multi}, T5-learning~\cite{mastropaolo2021studying} and TreeBERT~\cite{jiang2021treebert}.
\subsection{Structural Information of Source Code}
\label{section:related_structure}
Structural information is very important for understanding source code. ASTs are widely used in code related tasks for representing structure information of code (e.g., ~\cite{zhang2020retrieval,wang2021code,svyatkovskiy2019pythia,hu2018deep,leclair2019neural}), which contains abundant syntactic structure information that cannot be expressed by code sequences. An AST should be flattened with linearization methods, e.g., pre-order traversal~\cite{zhang2020retrieval,wang2021code}, in-order traversal~\cite{svyatkovskiy2019pythia}, and Structure-based Traversal (SBT)~\cite{hu2018deep}, before being fed to an encoder. code2vec~\cite{alon2019code2vec}, code2seq~\cite{alon2018code2seq}, and SLM~\cite{alon2020structural} use a method that linearizes an AST as a series of ``path-contexts'' representing two terminal nodes and the path between them. Jiang et al.~\cite{jiang2021treebert} represent ASTs as the set of paths and then introduce node position embedding to obtain the position of the node in the AST. Besides, neural networks that take trees as input (e.g., tree-LSTMs~\cite{wan2018improving,wan2019multi,lin2021improving}, RvNNs~\cite{zhang2019novel} and GNNs/GCNs~\cite{yang2021multi,leclair2020improved,chen2021holistic}) utilize an AST directly instead of flattening it.
There are also approaches taking data flow and control flow extracted from code as structural information, e.g., ~\cite{guo2020graphcodebert} and~\cite{huo2020control}. However, these flows do not contain structural information as rich as ASTs~\cite{guo2020graphcodebert}. Their advantages over ASTs are that they have a lower demand on hardware and need less training time under the same conditions~\cite{guo2020graphcodebert}.
Taken together, we choose to use ASTs to represent the structural information in SPT-Code.
\section{SPT-Code}
In this section we first introduce the architecture of SPT-Code (Section~\ref{section:architechture}). We then describe the model input (Section~\ref{section:model_input}) and the pre-training tasks (Section~\ref{section:pre_training_tasks}), which are the key innovations of this paper. Finally, we illustrates how to fine-tune SPT-Code when it is applied to downstream tasks (Section~\ref{section:fine-tuning}).
\subsection{Model Architecture}
\label{section:architechture}
Architecturally, SPT-Code is essentially a multi-layer Transformer~\cite{vaswani2017attention}, which is widely used by pre-training models such as BART~\cite{lewis2020bart} and T5~\cite{raffel2020exploring}. As far as the parameter setting of the encoder and decoder is concerned, SPT-Code follows CodeBERT and GraphCodeBERT: (1) the number of layers (i.e., Transformer blocks) $L=12$, (2) the size of model $d_{model}=768$, (3) the dimension of feed forward $d_{ff}=3072$, (4) the number of self-attention heads $h=12$, and (4) the dropout rate $p_{dropout}=0.1$. The total number of parameters is 262M.
To pre-train both classification and generation tasks with an encoder-decoder structure, we adopt the strategy used in BART. In particular, note that the encoder and decoder will continue to learn jointly and collaboratively when pre-trained on classification tasks.
Figure~\ref{figure:model_cls_gen} shows how classification and generation tasks are pre-trained in SPT-Code. Specifically,
\begin{figure}[htbp]
\centering
\subfloat[When SPT-Code is used for classification, the inputs of the encoder and decoder are identical, and the output of the decoder at the last time step is used as the label for the classification.]{\includegraphics[width=0.4\textwidth]{figures/model_cls.pdf}\label{figure:model_cls}}\\
\subfloat[When SPT-Code is used for generation, the whole procedure is the same as the regular sequence-to-sequence model.]{\includegraphics[width=0.4\textwidth]{figures/model_gen.pdf}\label{figure:model_gen}}
\caption{SPT-Code for classification and generation.}
\Description{Input and output of SPT-Code for classification and generation.}
\label{figure:model_cls_gen}
\end{figure}
to use SPT-Code for classification tasks, the input of the decoder is the same as encoder, except that a special symbol ``[SOS]'' (indicating the start of the sequence) is added to the front, and another special symbol ``[EOS]'' (used here as a placeholder for the classification position) is added to the end. The output of the corresponding position of the ``[EOS]'' is then used for classification. For generation tasks, Figure~\ref{figure:model_gen} demonstrates the input and output of the model when translating ``A B C D E'' to ``\textalpha~\textbeta~\textgamma~\textdelta~\textepsilon''. The special symbol ``[EOS]'' denotes the end of the sequence, and the process of generation stops when the model outputs this symbol.
\subsection{Model Input}
\label{section:model_input}
The inputs of the model are three different types of components belonging to a complete \textit{Method} (i.e., it can be invoked by its name), namely, the code tokens, the linearized AST, and the natural language. In this subsection, we will demonstrate this with a real Java method\footnote{https://github.com/Unidata/thredds/blob/d2d68f9eee87f345625211324d71d5dc3e162\\ee1/cdm/src/main/java/thredds/client/catalog/Property.java\#L56-L63} shown in the bottom of Figure~\ref{figure:model_input}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.47\textwidth]{figures/input_formulation.pdf}
\caption{The input of a real world Java code snippet. Due to space constraints, we abbreviated the node names in the AST sequence. ``NL'' denotes the natural language input.}
\Description{Input of a Java method.}
\label{figure:model_input}
\end{figure}
\subsubsection{Code Tokens}
\label{section:code_tokens}
As we can see from Figure~\ref{figure:model_input}, the first part of the input is the code token sequence of a method. We use a lexical analyzer to tokenize the source code and then obtain the tokens $C=\{c_1,c_2,\dots,c_l\}$, where $l$ is the number of code tokens. Specifically, we use the Python standard library\footnote{https://docs.python.org/3.8/library/tokenize.html} to tokenize Python codes. For languages such as Java, JavaScript, PHP and Go, we use the Python binding\footnote{https://pypi.org/project/antlr4-python3-runtime/} of ANTLR 4\footnote{https://github.com/antlr/antlr4} to get the code tokens. The Ruby source code is tokenized by the calling of a Ruby binary program. The source codes of other programming languages are tokenized by the NLTK tokenizer\footnote{https://www.nltk.org/api/nltk.tokenize.html}.
\subsubsection{Linearized AST}
\label{section:linearized_ast}
To represent the second part of the input, i.e., the structural information of the code, we convert an AST into a specially formatted sequence by traversing it, and call the result of this converting a \textit{linearized} AST. We first employ an AST parser\footnote{https://tree-sitter.github.io/tree-sitter/} to get the corresponding AST, then use a traversal method to parse the AST into a sequence.
Instead of using the original SBT (please refer to~\cite{hu2018deep,hu2020deep} for more details), which has been shown to be more effective than classical traversal methods (i.e., pre-order traversal) but tend to produce excessively long sequences that are on average more than three times the length of the code, we propose a simplified version of SBT called X-SBT (XML-like SBT) to traverse ASTs. X-SBT can reduce the length of the resulting sequence of traversals by more than half, Figure~\ref{figure:xsbt}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figures/xsbt.pdf}
\caption{An example of the SBT and X-SBT, traversal sequences are formatted and indented for a better illustration.}
\Description{Comparison of SBT and X-SBT.}
\label{figure:xsbt}
\end{figure}
shows a comparison of SBT and X-SBT. It can be seen that when traversing AST, SBT takes two tokens, ``('' and the node name, as the starting flag of a certain AST node, and takes ``)'' and the name as the ending flag. We make one observation: for non-terminal nodes (i.e., nodes which are not leaves), we can replace the starting flag with one token in an XML-like form, similarly for the ending flag. For terminal nodes, i.e., leaf nodes, we can further merge the starting and ending tokens into one token. Therefore, it is easy to prove that X-SBT shortens the length by more than half.
However, X-SBT sequences are still long. To further shorten X-SBT sequences, we make X-SBT traverse only the nodes at or above the \textit{expression} level in the AST. This will also reduce redundancy. Commonly, the AST contains both lexical and syntactic information, where the lexical information is already represented by the first part of the input (i.e., code tokens). Since the lexical information in the AST is concentrated on the terminal nodes, we only keep the nodes at or above the \textit{expression} level, so that it contains only syntactic information. Figure~\ref{figure:xsbt_example}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.47\textwidth]{figures/xsbt_example.pdf}
\caption{The AST of the ``for'' statement block of the Java code snippet in Figure~\ref{figure:model_input}. Non-terminal nodes are ellipses and terminal nodes are represented as rectangles. The green nodes will be traversed by X-SBT at the expression level, while the red ones will not.}
\Description{Example of the AST.}
\label{figure:xsbt_example}
\end{figure}
shows the nodes in the AST that will be traversed by the X-SBT at the \textit{expression} level. Obviously, it traverses only one subtree of the AST, and consequently, it can further reduce the length of the sequence by ignoring some fine-grained (lexical) information that is already present in the code tokens. To conclude, we traverse the AST using X-SBT at the \textit{expression} level to obtain the tokens $A=\{a_1,a_2,\dots,a_m\}$, where $m$ is the length of the sequence.
\subsubsection{Natural Language}
\label{section:nl_input}
For extracting natural language information from the code only, we derive the method name and API call sequence of the code\footnote{The reason we do not use documentation in code, such as docstring and in-line comments, is that documentation is not always available as we mentioned in Section~\ref{section:introduction}.}. For example, the tokens extracted from the code snippet in Figure~\ref{figure:model_input} are \texttt{removeDups, size, contains, add}. We further split each token of the form \textit{CamelCase} and \textit{snake\_case}, so \texttt{removeDups} is split into \texttt{remove} and \texttt{dups}. Then we take the resulting linear sequence of tokens, $N=\{n_1,n_2,\dots,n_p\}$, as our natural language input, where $p$ is the number of tokens.
After completing the construction of the three input parts, we concatenate them and use a special symbol, i.e., ``[SEP]'', to separate the three inputs. Therefore, the input is represented as
\begin{equation}
Input=c_1,\dots,c_l,[\textup{SEP}],a_2,\dots,a_m,[\textup{SEP}],n_1,\dots,n_p
\end{equation}
\subsection{Pre-Training Tasks}
\label{section:pre_training_tasks}
In this section, we introduce the three tasks in the order they are used for (sequential) pre-training.
\subsubsection{Code-AST Prediction}
\label{section:cap}
The first pre-training task is Code-AST Prediction (CAP). Since we add structural information to the input, in CAP, we expect to have the model learn such information about the structure represented by X-SBT. So we introduce this binarized task that can be trivially generated from any code. Specifically, when we construct the input representation, namely $Input=C$,[SEP],$A$,[SEP],$N$, 50\% of the time $A$ is the actual AST sequence corresponding to $C$ (labeled as \texttt{IsAST}), and 50\% of the time it is a random AST sequence from the dataset (labeled as \texttt{NotAST}). As we show in Figure~\ref{figure:model_cls}, ``label'' is used for Code-AST Prediction. Note that CAP is closely similar to the next sentence prediction task~\cite{devlin2019bert}.
\subsubsection{MASS}
\label{section:mass}
Since SPT-Code is based on the encoder-decoder architecture, we expect to pre-train SPT-Code in a sequence-to-sequence style. Therefore, we then adopt MASS, which seeks to reconstruct a sentence fragment given the remaining part of the sentence in the encoder-decoder framework. We employ a modified version of MASS as one of our pre-training tasks, with the intention of training the model to understand, infer and generate code sequences.
Given an input $Input=C$,[SEP],$A$,[SEP],$N$, we denote $C^{\backslash u:v}_{\textup{origin}}$ as a modified version of $C$ where its fragment from position $u$ to $v$ is masked, that is,
\begin{equation}
C^{\backslash u:v}_{\textup{origin}}=\{c_0,\dots,c_{u-1},[MASK],\dots,[MASK],c_{v+1},\dots,c_l\}
\end{equation}
where $0<u<v<l$. In contrast, $C^{u:v}$ denotes the fragment of $C$ from $u$ to $v$. In our work, we merge a number of consecutive ``[MASK]'' symbols into one, then $C^{\backslash u:v}_{\textup{origin}}$ is replaced by $C^{\backslash u:v}$:
\begin{equation}
C^{\backslash u:v}=\{c_0,\dots,c_{u-1},[MASK],c_{v+1},\dots,c_l\}
\end{equation}
We utilize this modified version of MASS to pre-train our sequence to sequence model by predicting the fragment $C^{u:v}$ taking the input $Input=C^{\backslash u:v}$,[SEP],$A$,[SEP],$N$. Then, $k=u-v+1$ is the number of the masked consecutive tokens, and we follow Song et al.~\cite{song2019mass} and set $k$=50\% of $l$ to achieve the best performance.
\subsubsection{Method Name Generation}
\label{section:mng}
With the last pre-training task, we would expect the model to learn information such as the intent and functionality of the code. The method name, which is present in every method, can be seen as an extremely concise summary of the method. Xie et al.~\cite{xie2021exploiting} analyze the method name and code summary in a Java dataset built by Leclair and McMillan~\cite{leclair2019recommendations}, and find that on average 50.6\% of the words in method names appear in the corresponding summaries, and 21.3\% of the words in the summaries appear in the corresponding method names. For about 20\% of the methods, all the words in the method names appear in the corresponding summaries. Therefore, they improve code summarization performance successfully by predicting method names in advance.
Therefore, we exploit method names in this pre-training task, which we call Method Name Generation (MNG). Given the model input $Input=C$,[SEP],$A$,[SEP],$N$, we denote $name=c_i$ as the method name in $C$, where $i$ is the index of the method name token in $C$. From Section~\ref{section:nl_input}, we learn that the split $name=c_i$ also appears in $N$, so we remove the tokens from $N$ that are derived from the method name split. Assuming that the method name is split into $s$ subtokens, the final input of MNG will be as follows:
\begin{equation}
\begin{split}
Input_{MNG}=&c_0,\dots,c_{i-1},[MASK],c_{i+1},\dots,c_l,\\
&[SEP],a_1,a_2,\dots,a_m,\\
&[SEP],n_{s+1},n_{s+2},\dots,n_p
\end{split}
\end{equation}
Then we make the decoder try to generate the split method name, i.e., $n_1,n_2,\dots,n_s$.
\subsection{Fine-Tuning}
\label{section:fine-tuning}
We group all downstream tasks into two categories: classification and generation. For classification tasks, we use the setting shown in Figure~\ref{figure:model_cls}; otherwise, we use the setting in Figure~\ref{figure:model_gen}. For each task, we simply plug in the task-specific inputs and outputs into SPT-Code and fine-tune all the parameters end-to-end. For example, in code search, which we cast as a classification task, when obtaining a representation vector of a natural language query, we take only the natural language as input and do not consider the first two parts of the input, i.e., code and AST. For code summarization, which we cast as a generation task, we only include the natural language summary in the output. We will describe the task-specific details in the corresponding subsections of Section~\ref{section:experiment}.
\section{Experiment}
\label{section:experiment}
In this section, we first introduce the pre-training data and settings (Section~\ref{section:exp_pre_train}) and how we fine-tune on the downstream tasks (Section~\ref{section:exp_fine_tune}). Then, we answer four research questions (Section~\ref{section:exp_evaluation}). Finally, we manually evaluate SPT-Code through quantitative and qualitative analysis.
\subsection{Pre-Training}
\label{section:exp_pre_train}
\subsubsection{Dataset}
\label{section:pre_training_dataset}
The dataset we use for pre-training SPT-Code is CodeSearchNet~\cite{husain2020codesearchnet}, which has also been used to pre-train CodeBERT~\cite{feng2020codebert}, GraphCodeBERT~\cite{guo2020graphcodebert} and T5-learning~\cite{mastropaolo2021studying}. The CodeSearchNet corpus is programmatically obtained by scraping open-source repositories and pairing individual functions with their (processed) documentation as natural language annotation. It includes more than 6.4M codes from 6 programming language, i.e., Java, Python, JavaScript, PHP, Go and Ruby. The data statistics are shown in Table~\ref{table:csn_statistics}.
\begin{table}[htbp]
\centering
\caption{Pre-Training Dataset Size Statistics}
\label{table:csn_statistics}
\small
\begin{tabular}{lrr}
\toprule
\multirow{2}{*}{\textbf{Language}} & \multicolumn{2}{c}{\textbf{\# Methods/Function}} \\ \cline{2-3}
& \multicolumn{1}{r}{\textbf{w/ documentation}} & \multicolumn{1}{r}{\textbf{All}} \\
\midrule
Java & 542,991 & 1,569,889 \\
Python & 503,502 & 1,156,085 \\
JavaScript & 157,988 & 1,857,835 \\
PHP & 717,313 & 977,821 \\
GO & 347,789 & 726,768 \\
Ruby & 57,393 & 164,048 \\
\hline
All & 2,326,976 & 6,452,446 \\
\bottomrule
\end{tabular}
\end{table}
Since our input can be extracted from completely unlabeled data, we can make use of all the 6.4M data instances in CodeSearchNet. However, CodeBERT and GraphCodeBERT use labels (i.e., the documentation) in the input for training both code and natural language, so they are both pre-trained on all the data in the ``w/ documentation'' column of the ``All'' row, which we named \textit{CSN$_{w/ doc}$}. T5-learning, inputs the code and the documentation into the model separately when pre-training, so it resembles SPT-Code in the sense that it does not utilize the code-document correspondence. Nevertheless, the publicly available implementation of T5-learning uses only the data in the ``All'' column of ``Java'' row, named \textit{CSN$_{Java}$}.
\subsubsection{Settings}
\label{section:pre_train_settings}
The three pre-training tasks of SPT-Code are performed sequentially. We tried different orders and different numbers of epoch for these tasks and found that better results can be achieved by first pre-training CAP for 10 epochs, then MASS for 30 epochs, and eventually MNG for 30 epochs. All three tasks use cross-entropy as the loss function. We use AdamW~\cite{loshchilov2018decoupled} as our optimizer, the initial learning rate is 5e-5 and the warmup step is set as 2000. We pre-train SPT-Code on 4$\times$NIVDIA A100s\footnote{The GPUs are provided by Alibaba Group, which employ EFLOPS~\cite{eflops} and Alibaba Collective Communication Library (ACCL)~\cite{accl} techniques. EFLOPS is a high performance distributed training system that can achieve near-linear scalability of overall throughput, and ACCL brings the performant efficiency of EFLOPS architecture to general cluster systems and Cloud scenarios, which is able to achieve fully non-congested communication.} with a total batch size of 256.
We use Byte-Pair Encoding (BPE)~\cite{sennrich2016neural} to tokenize the code and the natural language (after CamelCase and snake\_case splitting), and use regular word tokenizer to tokenize X-SBT sequences. The tokenizer is built upon the whole pre-training data, and will be directly used on each downstream tasks without any modification.
\subsection{Fine-Tuning on Downstream Tasks}
\label{section:exp_fine_tune}
In this subsection, we detail the process of fine-tuning the pre-trained SPT-Code on the five downstream tasks. For each downstream task, we give a brief introduction, tell the datasets for fine-tuning, present the compared baselines, and eventually illustrate the metrics for evaluating.
\subsubsection{Code Summarization}
\label{section:fine_tune_summarization}
Code summarization, also known as the process of code summary or code comment generation, is the task of automatically generating a natural language description for a piece of source code that summarizes the overall actions of the code snippet accurately, adequately, and concisely~\cite{sridhara2010towards}. Work in this area of research has generally focused on generating a (typically short) natural language comment from a given method.
\textbf{Datasets.} In addition to CodeSearchNet, we use two widely used classical datasets that are collected from open source repositories in GitHub, i.e., the Java dataset (JCSD)~\cite{hu2018summarizing} and the Python dataset (PCSD)~\cite{miceli2017parallel}. As for JCSD, the comment of each method is the first sentence extracted from its Javadoc, and the comment of a method in PCSD is its docstring.
\textbf{Baselines.} We adopt CodeBERT~\cite{feng2020codebert}, GraphCodeBERT~\cite{guo2020graphcodebert}, CugLM~\cite{liu2020multi}, T5-learning~\cite{mastropaolo2021studying} and TreeBERT~\cite{jiang2021treebert} as baselines for all dataset
\footnote{We resize all the pre-trained models here and in all the following downstream tasks identical to ours.}. For JCSD, PCSD and CodeSearchNet's Java/Python datasets, we adopt NerualCodeSum~\cite{ahmad2020transformer}, which employs Relative Position~\cite{shaw2018self} and Copy Attention~\cite{see2017get} upon vanilla Transformer~\cite{vaswani2017attention}, as another baseline\footnote{We have tried to run it as Java for other languages in CodeSearchNet but got completely nonsensical output.}. Besides, we use three more baselines for JCSD and PCSD, namely, DeepCom~\cite{hu2018deep}, Rencos~\cite{zhang2020retrieval}, and SiT~\cite{wu2021code}, which are widely compared or recently proposed RNN or transformer-based models dedicated for code summarization.
\textbf{Metrics.} We use BLEU (B.)~\cite{papineni2002bleu}, METEOR (M.)~\cite{banerjee2005meteor} and ROUGE-L (R.L)~\cite{lin2002manual}, which are widely used in fields such as machine translation and text summarization, to measure the similarity between the sentences generated by the model and the goal sentences.
\subsubsection{Code completion}
\label{section:fine_tune_completion}
Code completion is to generate code given its surrounding code as context. We fine-tune SPT-Code on a task called \textit{any-code} completion~\cite{alon2020structural} here. Unlike \textit{restricted} completion where the target code contains only primitive types (e.g., \texttt{int} and \texttt{string}) and excludes user-defined functions, \textit{any-code} completion aims to generate code in a general-purpose programming language without any restriction on its vocabulary or structure. Specifically, consider a program $\mathcal{P}$ and some part of the program $p\in \mathcal{P}$, any-code completion makes the model to predict $p$ given the rest of the program $\mathcal{P}^-=\mathcal{P}/p$.
\textbf{Datasets.} The dataset we use for any-code completion is the Java dataset provided by Alon et al.~\cite{alon2020structural}. It is based on the Java-small dataset~\cite{alon2018code2seq}, which contains the least code duplication~\cite{allamanis2019adverse}. Alon et al. create any-code completion examples by selecting every expression larger than a single AST node as the target $p$, using the reminder of the method as the context $\mathcal{P}$. They also filter many methods to clean the dataset and make the task harder. The resulting dataset contains 1.3M/10k/20k train/dev/test examples.
\textbf{Baselines.} Besides CodeBERT, GraphCodeBERT, CugLM, T5-learning and TreeBERT, some specific methods for this task are also compared. Code2seq~\cite{alon2018code2seq} represents a code snippet as the set of compositional paths in its AST and uses attention to select the relevant paths while decoding. Transformer w/ copy uses the implementation of Open-NMT~\cite{klein2017opennmt} with a copy mechanism~\cite{gu2016incorporating}. Seq2tree w/ copy~\cite{aharoni2017towards} learns to generate the linearized, subtokenized target AST for code completion. SLM~\cite{alon2020structural} leverages joint modeling of an AST and its missing subtree using a structural language model, which estimates the probability of the program’s AST by decomposing it into a product of conditional probabilities over its nodes.
\textbf{Metrics.} Following Alon et al.~\cite{alon2020structural}, we use exact match accuracy at 1 (Acc@1) and 5 (Acc@5) for evaluation. An exact match is counted if and only if the sentence generated by the model is identical to the goal sentence (ignoring cases and whitespaces). If we let the model generate $k$ candidate sentences with the highest probability and if an exact match is counted while any one of the candidates matches the goal sentence exactly, the ultimate exact match accuracy is Acc@$k$ (e.g., if only one candidate, it is Acc@1).
\subsubsection{Bug fixing}
\label{section:fine_tune_bug_fix}
Bug fixing, aka. code repair or code refinement, aims to automatically fix bugs in the code. It can help software developers locate and fix bugs and thus save a lot of time~\cite{jorgensen2006systematic,tufano2019empirical}.
\textbf{Datasets.} We use the dataset collected by Tufano et al.~\cite{tufano2019empirical}, who extract method-level pairs of buggy and corresponding fixed code named BFPs (bug-fix pairs) from bug-fixing commits in thousands of GitHub Java repositories. Each BFP is composed of a tuple $BFP=<m_b, m_f>$, where $m_b$ represents a buggy code component, $m_f$ represents the corresponding fixed code. Based on the code size, Tufano et al. provide two datasets: BFP$_{\textup{small}}$ and BFP$_{\textup{medium}}$, with the former having a code length below 50 and the latter having a code length between 50 and 100.
\textbf{Baselines.} The original method proposed by the dataset creator~\cite{tufano2019empirical} is marked as \textit{Tufano et al.} S2S+COPYSAPN~\cite{panthaplackel2021copy} is an extension of seq2seq models which can copy entire spans of the input to the output in one step and reduce the number of decisions required during inference. CodeBERT, GraphCodeBERT, CugLM, T5-learning and TreeBERT are also used as baselines.
\textbf{Metrics.} We report Acc (=Acc@1) and BLEU on both datasets. Both $m_b$ and $m_f$ are abstracted (please refer to \cite{tufano2019empirical} for more details) before being fed to the model, and we do not translate the abstracted code predicted by the model back into source code before evaluation using the metrics because the result before and after translation are the same.
\subsubsection{Code Translation}
\label{section:fine_tune_translation}
Code translation is important for migrating legacy code in one language into an ecosystem built in a different language.
\textbf{Datasets.} Following Chen et al.~\cite{chen2018tree} and Guo et al.~\cite{guo2020graphcodebert}, we conduct our experiments on a dataset containing several open-source projects, which have both a Java and a C\# implementation.
\textbf{Baselines.} Baselines are Naive (i.e., directly copying the source code as the translation result), vanilla Transformer~\cite{vaswani2017attention}, CodeBERT, GraphCodeBERT, CugLM, T5-learning and TreeBERT.
\textbf{Metrics.} We use Acc and BLEU as metrics.
\subsubsection{Code Search}
\label{section:fine_tune_search}
Code search aims to find the code snippet that most closely resembles the semantics of the given natural language query statement from a set of codes, which is named \textit{codebase} in this task. It is a good choice for testing the performance of SPT-Code in classification mode. For each code-query pair $(c,q)$, we compute the representation vectors of $c$ and $q$, i.e., $V_c$ and $V_q$ respectively. Then, we randomly take another query statement from the dataset as a negative sample, denoted as $q-$, whose representation vector is $V_{q-}$. To ensure the effectiveness of the negative samples, we restrict the BLEU score~\cite{papineni2002bleu} between $q-$ and $q$ to be lower than 0.3. The training loss is:
\begin{gather}
\mathcal{L}_{search}=\sum_{(c,q,q-)\in D} \max (0, \varepsilon -\cos (V_c,V_q)+\cos (V_c,V_{q-}))\\
\cos (V_c,V_q)=\frac{V_c\cdot V_q}{\Vert V_c\Vert \Vert V_q\Vert }
\end{gather}
where $D$ is the dataset, $\varepsilon$ is a fixed value of 0.05 (following Gu et al.~\cite{gu2018deep}), and $\cos$ denotes the calculation of cosine similarity. In the evaluation, we first get the representation vectors of all codes in codebase as candidates. Then, for each vector of query statements, we select a few code representation vectors from the candidates that are closest to the vector of query statements. Finally, the metric is calculated.
\textbf{Datasets.} We use CodeSearchNet~\cite{husain2020codesearchnet}.
\textbf{Baselines.} We use CNN~\cite{kim2014convolutional} (i.e., 1D convolutional neural network), Bi-GRU~\cite{cho2014learning}, and Transformer (i.e., vanilla Transformer~\cite{vaswani2017attention}) as baselines in addition to CodeBERT and GraphCodeBERT.
\textbf{Metrics.} We use Mean Reciprocal Rank (MRR) for evaluation. MRR is a statistic measure for evaluating search algorithms. The reciprocal rank of a query response is the multiplicative inverse of the rank of the first correct answer: 1 for first place, 1/2 for second place. Therefore, the mean reciprocal rank is the average of the reciprocal ranks of results for a sample of queries Q:
\begin{equation}
\textup{MRR}=\frac{1}{|Q|}\sum_{i=1}^{|Q|}\frac{1}{\textup{rank}_i}
\end{equation}
\subsection{Evaluation}
\label{section:exp_evaluation}
We evaluate SPT-Code by answering four research questions.
\subsubsection*{RQ1: How effective is SPT-Code compared with the state-of-the-art baselines and other code pre-trained models on five downstream tasks?}
\label{section:rq_comparsion}
We conduct experiments on five downstream tasks introduced in the previous subsection. The results on classical and CodeSearchNet code summarization are shown in Tables~\ref{table:results_summarization}
\begin{table}[htbp]
\centering
\caption{Results on classical code summarization.}
\label{table:results_summarization}
\small
\begin{tabular}{lcccccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Methods}} &
\multicolumn{3}{c}{JCSD} &
\multicolumn{3}{c}{PCSD} \\ \cline{2-7}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{B.} &
\multicolumn{1}{c}{M.} &
\multicolumn{1}{c}{R.L} &
\multicolumn{1}{c}{B.} &
\multicolumn{1}{c}{M.} &
\multicolumn{1}{c}{R.L} \\
\midrule
DeepCom & 37.5 & 22.0 & 51.2 & 20.4 & 9.5 & 36.8 \\
Rencos & 36.9 & 26.5 & 51.2 & 28.9 & 20.4 & 45.0 \\
NerualCodeSum & 44.5 & 26.4 & 54.7 & 32.3 & 19.7 & 46.8 \\
SiT & 45.3 & 27.3 & 55.0 & 33.8 & 20.7 & 48.2 \\
\hline
CodeBERT & 45.2 & 26.2 & 54.7 & 33.3 & 21.6 & 49.2 \\
GraphCodeBERT & 46.0 & 26.5 & 56.5 & 33.9 & 22.1 & 50.4 \\
CugLM & 46.1 & 26.3 & 55.8 & 34.3 & 21.6 & 49.7 \\
T5-learning & 44.2 & 26.5 & 53.9 & 34.1 & 22.6 & 50.1 \\
TreeBERT & 47.9 & 27.2 & 56.6 & 34.7 & 23.0 & 50.5 \\
\hline
SPT-Code & \textbf{49.1} & \textbf{32.4} & \textbf{58.2} & \textbf{36.1} & \textbf{26.9} & \textbf{52.0} \\
\bottomrule
\end{tabular}
\end{table}
and Table~\ref{table:results_summarization_csn},
\begin{table*}[htbp]
\centering
\caption{Results on CodeSearchNet code summarization.}
\label{table:results_summarization_csn}
\small
\begin{tabular}{lcccccccccccccccccc}
\toprule
\multirow{2}{*}{Methods} &
\multicolumn{3}{c}{Java} &
\multicolumn{3}{c}{Python} &
\multicolumn{3}{c}{JavaScript} &
\multicolumn{3}{c}{PHP} &
\multicolumn{3}{c}{Go} &
\multicolumn{3}{c}{Ruby} \\
\cline{2-19} &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L \\
\midrule
NeuralCodeSum & 10.7 & 12.9 & 25.9 & 9.4 & 4.9 & 17.0 & - & - & - & - & - & - & - & - & - & - & - & - \\
CodeBERT & 13.6 & 16.4 & 32.2 & 10.7 & 6.0 & 21.0 & 10.0 & 10.5 & 22.2 & 20.1 & 19.3 & 42.7 & 16.0 & 11.7 & 30.4 & 8.3 & 7.5 & 17.2 \\
GraphCodeBERT & 14.5 & 17.8 & 33.5 & 11.0 & 7.4 & 22.1 & 10.9 & 11.4 & 24.9 & 20.0 & 20.2 & 42.8 & 16.5 & 12.5 & 30.1 & 8.3 & 8.1 & 18.0 \\
CugLM & 13.2 & 16.1 & 31.7 & 11.4 & 7.5 & 22.0 & 10.0 & 10.4 & 21.8 & 17.8 & 18.2 & 41.9 & 17.0 & 12.7 & 29.7 & 7.4 & 6.8 & 15.3 \\
T5-learning & 13.5 & 16.8 & 34.2 & 9.7 & 5.5 & 16.3 & 9.0 & 9.2 & 19.6 & 18.4 & 17.9 & 35.6 & 15.3 & 11.0 & 25.9 & 7.8 & 6.7 & 15.5 \\
TreeBERT & 13.8 & 17.1 & 34.3 & 11.1 & 7.2 & 21.6 & 10.3 & 10.5 & 22.0 & 18.0 & 19.1 & 42.4 & 17.1 & 12.8 & 30.1 & 7.4 & 6.9 & 15.5 \\
\hline
SPT-Code & \textbf{16.8} & \textbf{20.6} & \textbf{36.3} & \textbf{12.8} & \textbf{14.2} & \textbf{27.1} & \textbf{12.8} & \textbf{17.2} & \textbf{28.8} & \textbf{20.4} & \textbf{20.6} & \textbf{43.7} & \textbf{18.8} & \textbf{16.6} & \textbf{38.1} & \textbf{8.4} & \textbf{11.1} & \textbf{20.0} \\
\bottomrule
\end{tabular}
\end{table*}
respectively. Tables~\ref{table:results_completion}-\ref{table:results_search} show the results of code completion, bug fixing, code translation and code search, respectively.
\begin{table}[htbp]
\centering
\caption{Results on any-code completion.}
\label{table:results_completion}
\small
\begin{tabular}{lcc}
\toprule
Methods & Acc@1 & Acc@5 \\
\midrule
code2seq & 10.44 & 15.47 \\
Transformer w/ copy & 16.68 & 24.12 \\
seq2tree w/ copy & 16.46 & 22.89 \\
SLM & 18.00 & 24.77 \\
\hline
CodeBERT & 17.23 & 24.79 \\
GraphCodeBERT & 18.21 & 25.00 \\
CugLM & 18.43 & 25.51 \\
T5-learning & 16.03 & 23.78 \\
TreeBERT & 17.17 & 24.73 \\
\hline
SPT-Code & \textbf{19.09} & \textbf{26.57} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{Results on bug fixing.}
\label{table:results_bug_fix}
\small
\begin{tabular}{lcccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Methods}} &
\multicolumn{2}{c}{BFP$_{\textup{small}}$} &
\multicolumn{2}{c}{BFP$_{\textup{medium}}$} \\ \cline{2-5}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} \\
\midrule
Tufano et al. & 9.27 & - & 3.21 & - \\
S2S+COPYSPAN & \textbf{17.6} & - & 7.9 & - \\
\hline
CodeBERT & 6.23 & 61.47 & 2.43 & 68.54 \\
GraphCodeBERT & 7.49 & 63.40 & 3.16 & 67.33 \\
CugLM & 14.55 & 67.24 & 7.84 & 71.23 \\
T5-learning & 12.01 & 65.02 & 4.51 & 60.89 \\
TreeBERT & 13.15 & 65.74 & 7.61 & 70.67 \\
\hline
SPT-Code & 17.54 & \textbf{75.10} & \textbf{10.86} & \textbf{87.88} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{Results on code translation.}
\label{table:results_translation}
\small
\begin{tabular}{lcccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Methods}} &
\multicolumn{2}{c}{Java $\to$ C\#} &
\multicolumn{2}{c}{C\# $\to$ Java} \\ \cline{2-5}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} \\
\midrule
Naive & 00.0 & 18.52 & 00.0 & 18.66 \\
Transformer & 33.4 & 56.05 & 38.8 & 50.84 \\
\hline
CodeBERT & 58.8 & 79.34 & 57.3 & 71.10 \\
GraphCodeBERT & 59.3 & 80.33 & 58.1 & 72.39 \\
CugLM & 61.6 & 82.72 & 59.5 & 74.89 \\
T5-learning & 54.1 & 81.35 & 48.4 & 77.10 \\
TreeBERT & 60.0 & 80.92 & 57.8 & 77.15 \\
\hline
SPT-Code & \textbf{64.07} & \textbf{90.34} & \textbf{60.29} & \textbf{86.10} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[htbp]
\centering
\caption{Results on code search.}
\label{table:results_search}
\small
\begin{tabular}{lcccccc}
\toprule
Methods & Java & Python & JS & PHP & Go & Ruby \\
\midrule
CNN & 0.262 & 0.241 & 0.224 & 0.261 & 0.676 & 0.274 \\
Bi-GRU & 0.312 & 0.298 & 0.195 & 0.340 & 0.691 & 0.214 \\
Transformer & 0.404 & 0.399 & 0.289 & 0.430 & 0.729 & 0.278 \\
\hline
CodeBERT & 0.673 & 0.670 & 0.618 & 0.624 & 0.879 & 0.674 \\
GraphCodeBERT & 0.690 & 0.692 & \textbf{0.643} & 0.647 & \textbf{0.896} & \textbf{0.701} \\
\hline
SPT-Code & \textbf{0.700} & \textbf{0.699} & 0.641 & \textbf{0.651} & 0.895 & \textbf{0.701} \\
\bottomrule
\end{tabular}
\end{table}
First of all, we can see that whether the baselines are dedicated to a specific downstream task or pre-trained models, SPT-Code clearly outperforms them in the vast majority of cases, including code summarization, code completion, bug fixing on BFP$_{\textup{medium}}$, and code translation. This is because during the pre-training process, SPT-Code learns the representation of codes from a large amount of data, as well as the connection between codes and structural and natural language information. In addition, the model is enhanced to generate code and natural language sequences by pre-training.
As for bug fixing on BFP$_{\textup{small}}$, the accuracy of SPT-Code is slightly lower than that of S2S+COPYSPAN, which achieves the best accuracy on BFP$_{\textup{small}}$. The reason we believe is two-fold. The first is the dataset, where the lengths of the code are up to 50, which allows RNN-based S2S+COPYSPAN to adequately cope with inputs of these lengths. Another reason is that the mechanism proposed by S2S+COPYSPAN to copy the entire span of the input is well suited for a task like bug fixing, where only individual modifications are made on the input. However, in the next RQ, we will show that SPT-Code still outperforms S2S+COPYSPAN on BFP$_{\textup{small}}$ after removing MNG from the pre-training tasks.
Finally, we can learn that SPT-Code performs comparably to GraphCodeBERT on code search, which suggests that although we designed our model to perform better on generation tasks, it does not come at the cost of the performance on classification tasks.
\subsubsection*{RQ2: How do the three pre-training tasks, as well as the AST and the natural language input contribute to the model's performance on the different downstream tasks?}
\label{section:rq_component}
In order to find the impact of each component to the overall performance of SPT-Code, we conduct ablation study on all downstream tasks. We remove each/all pre-training task, and AST or/and natural language part from the input, respectively. Owing to space limitations, for code summarization, we only show results on Java and Python in CodeSearchNet\footnote{Java and Python represent static and dynamic languages, respectively.}; and for code search, we only show results on Java and Go\footnote{The results for both are above and below GraphCodeBERT, respectively.}. The results are shown in the first three groups of Table~\ref{table:ablation_mixed}.
\begin{table*}
\centering
\caption{Ablation study on downstream tasks.}
\label{table:ablation_mixed}
\resizebox{\textwidth}{!}{
\begin{tabular}{lcccccc|cc|cccc|cccc|cc}
\toprule
\multirow{3}{*}{Methods} &
\multicolumn{6}{c|}{Summarization} &
\multicolumn{2}{c|}{Completion} &
\multicolumn{4}{c|}{Bug fixing} &
\multicolumn{4}{c|}{Translation} &
\multicolumn{2}{c}{Search}\\ \cline{2-19}
&
\multicolumn{3}{c}{Java} &
\multicolumn{3}{c|}{Python} &
\multirow{2}{*}{Acc@1} &
\multirow{2}{*}{Acc@5} &
\multicolumn{2}{c}{BFP$_\textup{small}$} &
\multicolumn{2}{c|}{BFP$_\textup{medium}$} &
\multicolumn{2}{c}{Java $\to$ C\#} &
\multicolumn{2}{c|}{C\# $\to$ Java} &
\multirow{2}{*}{Java} &
\multirow{2}{*}{Go} \\ \cline{2-7} \cline{10-17}
& B. & M. & R.L & B. & M. & R.L & & & Acc & BLEU & Acc & BLEU & Acc & BLEU & Acc & BLEU & & \\
\midrule
SPT-Code & \textbf{16.79} & \textbf{20.55} & \textbf{36.34} & \textbf{12.77} & \textbf{14.16} & \textbf{27.10} & 19.09 & \textbf{26.57} & 17.54 & \textbf{75.10} & 10.86 & 87.77 & \textbf{64.07} & \textbf{90.34} & \textbf{60.29} & \textbf{86.10} & \textbf{0.700} & \textbf{0.895} \\
\hline
\ -w/o CAP & 15.85 & 20.33 & 36.30 & 12.26 & 14.16 & 26.57 & 18.42 & 25.63 & 17.20 & 73.54 & 10.06 & 87.26 & 57.25 & 88.80 & 54.40 & 84.66 & 0.668 & 0.872 \\
\ -w/o MASS & 15.75 & 20.19 & 35.70 & 12.16 & 13.48 & 26.00 & 17.42 & 23.71 & 16.30 & 72.71 & 9.65 & 86.10 & 53.93 & 82.09 & 50.55 & 82.32 & 0.686 & 0.889 \\
\ -w/o MNG & 15.45 & 19.86 & 35.65 & 12.01 & 13.81 & 25.98 & \textbf{19.71} & 26.10 & \textbf{18.11} & 73.09 & \textbf{11.96} & \textbf{87.79} & 58.80 & 88.79 & 57.16 & 85.19 & 0.673 & 0.875 \\
\ -w/o all & 13.54 & 17.72 & 33.31 & 11.90 & 12.74 & 25.23 & 13.49 & 19.31 & 14.28 & 72.89 & 8.27 & 85.30 & 49.19 & 79.28 & 47.49 & 78.69 & 0.656 & 0.855 \\
\hline
\ -w/o AST & 16.43 & 20.14 & 36.19 & 11.87 & 13.80 & 25.74 & 18.53 & 25.78 & 16.35 & 72.85 & 10.61 & 87.52 & 55.81 & 87.79 & 56.17 & 82.66 & 0.692 & 0.885 \\
\ -w/o NL & 15.68 & 20.15 & 35.86 & 11.96 & 13.56 & 25.55 & 18.47 & 26.09 & 16.12 & 72.70 & 10.50 & 87.57 & 56.29 & 87.83 & 56.92 & 83.56 & 0.679 & 0.865 \\
\ -only code & 15.35 & 19.97 & 35.63 & 10.25 & 12.74 & 23.90 & 18.33 & 25.88 & 16.01 & 72.55 & 10.19 & 86.96 & 54.13 & 87.34 & 54.59 & 83.81 & 0.669 & 0.864 \\
\hline
\ -CSN$_{\textup{w/ doc}}$ & 15.41 & 19.78 & 35.74 & 12.59 & 13.56 & 26.49 & 17.89 & 23.80 & 16.14 & 74.02 & 10.41 & 87.58 & 60.53 & 88.78 & 58.28 & 84.09 & 0.673 & 0.878 \\
\ -CSN$_{\textup{Java}}$ & 15.67 & 19.72 & 35.22 & 12.40 & 12.60 & 25.24 & 17.76 & 23.54 & 16.10 & 74.51 & 10.18 & 86.96 & 59.43 & 86.55 & 58.11 & 83.07 & 0.678 & 0.866 \\
\bottomrule
\end{tabular}
}
\end{table*}
The first thing we notice is that when we train from scratch (i.e., remove all pre-training tasks) or remove the AST and natural language from the input (i.e., only input code tokens), the performance of SPT-Code consistently drops considerably. Second, when we remove each of the three pre-training tasks, the results decreased in most cases, particularly in code summarization, code translation and code search. Of these, CAP is most useful for code search, MASS is most helpful for code translation, and removing MNG has the greatest impact on code summarization.
Interestingly, we find that for code completion and bug fixing, SPT-Code's performance w.r.t.\ accuracy improves instead when MNG is removed. This is understandable. On the one hand, code completion and bug fixing are tasks where the input and the output are both code, as is MASS, while the output of the MNG task is natural language, and thus the ability to generate natural language trained by MNG is not fully reflected in these two tasks. On the other hand, regarding code completion, MASS can be seen as totally unrestricted any-code completion. Both predict a piece of code based on its context, with the difference that in any-code completion, the piece of code is restricted to be an entire expression, while in MASS, the piece of code is selected completely at random. Therefore, fine-tuning code completion directly after MASS, i.e., removing the MNG, yields a higher result.
When we remove either the AST or the natural language from the input, the results of the model drop, indicating that they both help improve performance. In addition, for code summarization, bug fixing and code search, the results are lower when only natural language is removed, which indicates that in comparison, natural language helps these three tasks more than ASTs do. On the contrary, AST is more helpful for code completion and code translation.
Through ablation, we find that different pre-training tasks show different degrees of influence on the downstream tasks, As a result, appropriate trade-offs of pre-training tasks for different downstream tasks can help the model achieve better performance. ASTs and natural language both have a positive impact on the performance of the model, regardless of the downstream task.
\subsubsection*{RQ3: Is the ability of utilizing more unlabeled data an advantage of SPT-Code?}
\label{section:rq_pre_train_data}
As we know, CodeBERT, GraphCodeBERT and T5-learning are pre-trained on a subset of CodeSearchNet, whereas we can pre-train on the entire dataset. Therefore there is a question of whether the ability to utilize more data for pre-training also gives SPT-Code an unfair advantage. To ensure a fair comparison, we therefore pre-train SPT-Code only on the same data as CodeBERT and GraphCodeBERT (i.e., CSN$_{\textup{w/ doc}}$), and also on the same data as T5-learning (i.e., CSN$_{\textup{Java}}$). Notice, however, that here only the code from the labeled data is utilized and not the labels, even on the CSN$_{\textup{w/ doc}}$ where all samples have labels. So in this case, although we are using the data set of the same size, it actually utilize less information than they do. The results are displayed in the last group of Table~\ref{table:ablation_mixed}.
Comparing the results of SPT-Code with those of ``CSN$_{\textup{w/ doc}}$'' and ``CSN$_{\textup{Java}}$'' in Table~\ref{table:ablation_mixed}, we find that there is a significant performance decrease when the data used for pre-training is shrunk from 6.4M to 2.3M or 1.5M. However, considering the other pre-training models, it still maintains an advantage or is at a comparable level in performance. Therefore, we can conclude that SPT-Code is superior given the same amount of pre-trained data. In addition, the ability to use more unlabeled pre-training data can help SPT-Code achieve higher performance. This ensures that SPT-Code has a very advantageous scalability at the data level compared to the other pre-training models for source code.
\subsubsection*{RQ4: How would the size of the training data for fine-tuning SPT-Code influence its performance on downstream tasks?}
\label{section:rq_fine_tune_data}
To answer this question, we plot learning curves by varying the amount of (task-specific) training data used to fine-tune SPT-Code.
Owing to space limitations, we only report results of code summarization on Java in CodeSearchNet, and bug fixing on BFP$_{\textup{medium}}$. We select $k/10$ data from the training set each time for training, and then test on the same entire testing set. The results are shown in Figure~\ref{figure:data_ablation}.
\begin{figure}[htbp]
\centering
\subfloat[Results on Java CodeSearchNet code summarization.]
{\includegraphics[width=0.42\textwidth]{figures/summarization_subset.pdf}\label{figure:summarization_subset}}\\
\subfloat[Results on BFP$_{\textup{medium}}$ bug fixing.]
{\includegraphics[width=0.42\textwidth]{figures/bug_fix_subset.pdf}\label{figure:bug_fix_subset}}
\caption{Results when $k$ goes from 1 to 10.}
\Description{When the size of training set is different, results of the model.}
\label{figure:data_ablation}
\end{figure}
It is easy to see that the performance shows a decreasing trend in all evaluation metrics as $k$ decreases (i.e., the size of the training set decreases). It proves that even for pre-trained models, the size of the training set plays a key role in the performance as well. Meanwhile, considering Table~\ref{table:results_summarization_csn}, we see that for code summarization, the results of SPT-Code are comparable to that of CodeBERT when the training set size is reduced to 1/10, and the performance is similar to or even higher than GraphCodeBERT when it is reduced to 2/10. Further, recalling Table~\ref{table:results_bug_fix}, we can see that when the training set of BFP$_{\textup{medium}}$ is reduced to 1/10 of the original size, SPT-Code's results are still higher than other pre-trained models, and when it is reduced to half of the original size, SPT-Code's performance can reach the performance of the best baseline model, i.e., S2S+COPYSPAN.
The conclusion is that although there is an inevitable drop in SPT-Code's performance as the size of the training data decreases during fine-tuning, the performance of SPT-Code is comparable to that of the other models when the data used for fine-tuning SPT-Code is reduced to a very small size. This also implies the robustness of SPT-Code.
\subsection{Quantitative Analysis}
\label{section:quantitative_analysis}
We collect the output of SPT-Code and baselines for a wide range of test samples on each downstream task. Then we invited five participants, all of whom are graduate students in software engineering who are themselves not authors of this paper. For each downstream task, the five students are randomly assigned to the same number of samples. If there are multiple datasets for that downstream task, the samples of multiple datasets are the same. For downstream tasks that use Acc as a metric, such as code completion, we focus on the similarity between incorrect and correct answers, and the extent to which the incorrect answer can be helpful when all models fail. The results are shown in Table~\ref{table:results_quantitative}.
\begin{table}[htbp]
\centering
\caption{Results of quantitative analysis.}
\label{table:results_quantitative}
\footnotesize
\begin{tabular}{lrrrr}
\toprule
Dowstream Tasks & \# Sample & \# Better & \# Comparable & \# Worse \\
\midrule
Code Summarization & 800 & 328 & 368 & 104 \\
Code Completion & 600 & 312 & 233 & 55 \\
Bug Fixing & 200 & 89 & 75 & 36 \\
Code Translation & 200 & 83 & 65 & 52 \\
Code Search & 300 & 61 & 185 & 54 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Qualitative Analysis}
\label{section:qualitative_analysis}
By asking the participants in the previous subsection and browsing the output of each model by ourselves, we found that in comparison, SPT-Code can capture the semantic information of identifiers within code more accurately than other pre-trained models, and it can capture the semantic information of code segment globally instead of limiting to a certain region. The extracted code semantics are relatively more widely and evenly distributed in the code segment.
In the case of code summarization, for example, SPT-Code provides more complete and accurate descriptions of the method's overall functionality. Table~\ref{table:qualitative_example} shows summaries generated by different models for an example Java method in CodeSearchNet data.
\begin{table}
\centering
\caption{Qualitative example of SPT-Code and baselines.}
\footnotesize
\begin{tabular}{l}
\toprule
\begin{lstlisting}
void emitLoop() {
for ( ; ; ) {
AppendOnlyLinkedArrayList<Object> q;
q = queue;
if (q == null) {
emitting = false;
return; }
queue = null;
q.forEachWhile(this); } }
\end{lstlisting} \\
\midrule
{\color{teal} CodeBERT}: the queue is being destroyed \\
{\color{cyan} GraphCodeBERT}: emits the next loop in the queue \\
{\color{orange} T5-learning}: this method is called when the queue \\
\hline
{\color{purple} SPT-Code}: emits all of the elements in this queue \\
{\color{purple} \ -w/o AST}: emit the linked list \\
{\color{purple} \ -w/o NL}: emits all entries in the queue \\
\hline
{\color{red} Human Written}: loops until all notifications in the queue has been processed \\
\bottomrule
\end{tabular}
\label{table:qualitative_example}
\end{table}
We find that the ``queue'' object is recognized by all of these models. CodeBERT and T5-learning fail to capture the operation conducted on the queue, i.e., ``emits/emit'', but GraphCodeBERT and SPT-Code does. Compared to SPT-Code, although GraphCodeBERT also understand the relationship between the operation and the object, it fails to choose appropriate words for describing the operation. The reason may be that the data flow can indeed partially (but not fully) capture code structure information. Another interesting observation that may support this inference is that all models considering structure information generate ``in the queue'' correctly, i.e., GraphCodeBERT, SPT-Code, and SPT-Code-w/o NL, and models using AST all generate accurate, readable and smooth summaries. It also implies that data flow is less effectiveness than AST.
\section{Threats to Validity}
\label{section:threats}
\subsubsection*{Construct Validity} Like many existing code pre-training models, SPT-Code uses CodeSearchNet for pre-training. Since CodeSearchNet is also used in the evaluation of code summarization and code search, it is possible that the samples from the test sets for these two tasks have already been used for pre-training. This is not fair to methods that have not been pre-trained with CodeSearchNet, such as NeuralCodeSum in CodeSearchNet code summarization. We also recognize the impact on the results of not removing duplicates. However, our decision of not removing replicates is based on two considerations: (1) ensuring fairness of the comparison: the three code pre-training baselines were pre-trained with all data from CodeSearchNet without duplicate removal; (2) SPT-code is not affected by duplicate data: on one hand, downstream tasks that use pre-training dataset are CSN code summarization and code search, while the pre-training tasks designed for SPT-Code do not use the docstring, i.e., in testing, SPT-Code does not have the advantage of generating more accurate summaries because it has seen a piece of the tested code during pre-training, or has the advantage of better searching for code based on natural language. On the other hand, other downstream tasks that do not use CodeSearchNet have little overlap with CSN.
For rigorous consideration, we remove all test sets in CodeSearchNet and code duplicated with test sets of other downstream tasks\footnote{By using tools provided by Allamanis~\cite{allamanis2019adverse}, we find 9 in classical code summarization, 2 in $\textup{BFP}_\textup{medium}$ and code completion, respectively.}. Then re-pre-train and fine-tune the SPT-Code on three downstream tasks, i.e., JCSD, code completion and $\textup{BFP}_\textup{medium}$ bug fixing. It is found that the results of SPT-Code decrease very little after removing these duplicates\footnote{Results decrease 0.01 BLEU on JCSD, 0.03 Acc@1 on code completion and 0.07 Acc on $\textup{BFP}_\textup{medium}$}. Moreover, it is still not sure whether the change in results is due to the smaller pre-training dataset (shrunk by about 1/10) or the removal of duplicates.
\subsubsection*{Internal Validity} It is widely agreed that hyperparameters have a significant impact on the performance of deep learning models, but hyperparameters of SPT-Code are not tuned experimentally and are set empirically. Therefore, other hyperparameter settings may yield better results.
\subsubsection*{External Validity} Another threat posed by using CodeSearchNet as our pre-training dataset is that CodeSearchNet data is not balanced across six programming languages, which can be seen in Table~\ref{table:csn_statistics}, so our model may not perform the same on different languages, and we cannot guarantee the validity of SPT-Code for programming languages other than these six.
\section{Conclusion}
\label{section:conclusion}
We presented SPT-Code, a large model for source code based on an encoder-decoder architecture.
First, we design three code-specific pre-training tasks to pre-train SPT-Code. Secondly, we propose a new input representation whose is the first method that take into account both natural language and AST form of code, where we also propose a improved version of the AST traversal method, X-SBT. Both our pre-training tasks and input representation allow SPT-Code to be pre-trained on a completely unlabeled dataset.
SPT-Code was then fine-tuned on five code-related downstream tasks. Results indicate that fine-tuning SPT-Code enables it to achieve the state-of-the-art performance on five code-related downstream tasks. Ablation experiments reveal that the three pre-training tasks have different degrees of impact on different downstream tasks, and AST and natural language input also helped improve SPT-Code's performance. To facilitate future research, we also make our code and other artifacts publicly available at \url{https://github.com/NougatCA/SPT-Code}.
\begin{acks}
This work is supported by National Natural Science Foundation of China (61802167, 61802095), Natural Science Foundation of Jiangsu Province, China (BK20201250), Cooperation Fund of Huawei-NJU Creative Laboratory for the Next Programming, and NSF award 2034508. We thank Alibaba Cloud for its high-efficient AI computing service from EFlops Cluster. We also thank the reviewers for their helpful comments. Chuanyi Li and Jidong Ge are the corresponding authors.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
\label{section:introduction}
Pre-Training has revolutionized the way computational models are trained in the natural language processing (NLP) community~\cite{devlin2019bert,radford2019language,peters2018deep,clark2019electra}. For a long time, supervised learning has been the most successful natural language learning paradigm. The pioneers of the pre-training idea challenged this view by showing that a vast amount of general knowledge about language, including both linguistic and commonsense knowledge, can be acquired by (pre-)training a model in a {\em task-agnostic} manner using {\em self-supervised} learning tasks. Self-supervised learning tasks are NLP tasks for which the label associated with a training instance can be derived automatically from the text itself. Consider, for instance, one of the most well-known self-supervised learning tasks, Masked Language Modeling (MLM)~\cite{devlin2019bert}. Given a sequence of word tokens in which a certain percentage of tokens is {\em masked} randomly, the goal of MLM is to predict the masked tokens. As can be easily imagined, a model for MLM can therefore be trained on instances where each one is composed of a partially masked sequence of word tokens and the associated ``class'' value is the masked tokens themselves. Because no human annotation is needed, a model can be pre-trained on a very large amount of labeled data can be automatically generated, thereby acquiring a potentially vast amount of knowledge about language. A pre-trained model can then be optimized for a specific task by fine-tuning its parameters using task-specific labeled data in the standard supervised fashion.
Inspired by the successes of pre-trained models in NLP, a number of pre-trained models for source code have been proposed and applied to a variety of SE tasks including code summarization and code completion, with notable successes~\cite{buratti2020exploring,karampatsis2020scelmo,kanade2020learning,svyatkovskiy2020intellicode,feng2020codebert,guo2020graphcodebert,liu2020multi,mastropaolo2021studying,jiang2021treebert}. Despite these promising results, there are issues surrounding the application of pre-trained models for source code to SE tasks.
First, the majority of these pre-trained models focus on pre-training only the encoder of the Transformer~\cite{kanade2020learning,buratti2020exploring,feng2020codebert,guo2020graphcodebert}. This is not ideal, however. For instance, for generation tasks that are addressed using models with the encoder-decoder architecture, there is no reason why the decoder should be left out in the pre-training process. Second, these models have largely assumed as inputs the source code~\cite{kanade2020learning,karampatsis2020scelmo,svyatkovskiy2020intellicode,buratti2020exploring,liu2020multi} and the associated natural language description~\cite{feng2020codebert,mastropaolo2021studying}. In particular, code structure, which is also crucial to understanding source code, is largely missing from these models. The reason why code structure is left out is that SE researchers have for the most part simply reused the pre-training tasks designed for natural languages when pre-training models (by viewing source code as natural language), but none of these tasks are concerned with learning the structure of
natural language. Third, these pre-training tasks assume the availability of a bilingual corpus, where each method/function (henceforth collectively referred to as method) is ``labeled'' with the corresponding docstring, when pre-training a model on source code and the associated natural language description~\cite{feng2020codebert,guo2020graphcodebert}. However, such a bilingual corpus tends to be small in size compared to a monolingual (i.e., source code only) corpus, thus severely limiting the amount of data a model can be pre-trained on.
In general, we believe the reliance on a bilingual corpus would hinder the development of powerful pre-trained models for source code in the long run, as a key strength of the pre-trained models developed in the NLP community stems from their ability to be trained using self-supervised tasks for which \textit{very large} amounts of labeled data can be automatically generated.
Several attempts have been made to address the three problems mentioned above to different extents. To address the first problem, T5-learning~\cite{mastropaolo2021studying} and TreeBERT~\cite{jiang2021treebert} are proposed, which are sequence-to-sequence (i.e., seq2seq) pre-training models with the encoder-decoder architecture and enables both the encoder and the decoder to be jointly trained. To address the second problem, GraphCodeBERT~\cite{guo2020graphcodebert} employs two pre-training tasks specifically designed to acquire structural information. One task involves predicting the edges in the data flow while the other involves predicting the alignment between the nodes in the data flow and the code sequence, respectively. However, as the authors also pointed out, while the data flow captures information that is largely semantic in nature, it does not capture syntactic information (e.g., the syntactic structure encoded in an Abstract Syntax Trees, i.e., AST), which is arguably the most important type structural information about source code that is commonly exploited by SE researchers.
TreeBERT~\cite{jiang2021treebert} employs the set of constituent paths of ASTs as the input of its encoder. However, it only inputs code sequences at the decoder side during pre-training, trying to make the encoder learn lexical and semantic information (both of which can be easily obtained from the code sequences) from the AST (which contains mainly syntactic information) during the pre-training phase. But it is uncertain whether TreeBERT can extract the complete lexical and semantic information from the AST only by relying on pre-training, thus eliminating the need to input code tokens when fine-tuning.
To address the third problem, T5-learning treats code and natural language as two types of independent data instances. While this allows T5-learning to learn from a monolingual rather than bilingual corpus, the connection between a piece of code and the associated natural language description is no longer present in the corpus. Hence, it is no longer clear whether T5-learning can still learn to produce a natural language description of a piece of code. To our knowledge, there has been no attempt to address all three issues in a single model.
In light of the above discussion, we propose SPT-Code, a new pre-trained model for source code. Motivated by T5-learning, SPT-Code is a seq2seq pre-training model, enabling both the encoder and the decoder of Transformer to be jointly pre-trained. Each data instance for SPT-Code is composed of three types of information derived from a method, namely the code sequence, its AST, and the associated natural language description. Note that the incorporation of ASTs allows SPT-Code to exploit structural, specifically syntactic, information. In addition, to obviate the need to learn natural language descriptions from a bilingual corpus, we will simply use the name of the method and the names of the methods that are invoked in this method as a (very succinct) natural language description of the given source code.
We design three code-specific pre-training tasks for SPT-Code, each of which allows SPT-Code to acquire exactly one of the three types of information that comprise a data instance. The first task is a version the well-known Masked Sequence to Sequence (MASS)~\cite{song2019mass} pre-training task for natural language that we adapt to source code. Specifically, our modified MASS task seeks to acquire knowledge about source code via masking a random fragment of the code tokens. The second task, Code-AST prediction (CAP), is designed to enable the model to gain knowledge of the syntactic structure of a code fragment by predicting whether the given AST is the correct AST for the given piece of code. The final task, Method Name Generation (MNG), is a novel task that involves generating the subtokens of the method name, which we take to be an (extremely concise) natural language description of the method.
After SPT-Code is pre-trained on the CodeSearchNet dataset~\cite{husain2020codesearchnet}, we fine-tune and evaluate it on five downstream tasks, including code summarization, code completion, bug fixing, code translation, and code search. Experimental results show that SPT-Code achieves state-of-the-art results under virtually all circumstances.
In sum, we make the following contributions
\begin{enumerate}
\item Propose SPT-Code, a seq2seq pre-trained model for source code that is built with the encoder-decoder architecture and is applicable to both classification and generation tasks.
\item Extend the input representation of pre-trained models for source code with a simplified and linearized version of ASTs. To our knowledge, we are the first to use both natural language and AST as inputs in pre-training for source code.
\item Design special input representations and three code-specific seq2seq-based pre-training tasks enabling SPT-Code to be pre-trained without relying on any bilingual corpus or labeled data.
\item Pre-train SPT-Code on a large unlabeled monolingual (i.e., source code only) dataset across six programming languages, then fine-tune and evaluate it on five downstream code-related tasks, achieving state-of-the-art results on all tasks.
\end{enumerate}
\section{Related Work}
\label{section:related}
\subsection{Pre-Training Models in NLP and SE}
Table~\ref{table:pre_train_models}
\begin{table*}
\centering
\caption{Pre-training models in NLP and for source code.}
\label{table:pre_train_models}
\resizebox{\textwidth}{!}{
\begin{tabular}{llcccc|cccccccc|ccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Domain}} &
\multicolumn{1}{l}{\multirow{2}{*}{Models}} &
\multicolumn{4}{c|}{Modules} &
\multicolumn{8}{c|}{Objectives} &
\multicolumn{3}{c}{Inputs} \\
\cline{3-17}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{} &
LSTM &
Encoder &
Decoder &
Encoder-Decoder &
Forward LM &
Backward LM &
Masked LM &
NSP &
Permutation LM &
RTD &
- &
- &
NL &
Code &
Structure \\
\midrule
\multirow{8}{*}{NLP} & ELMo (2018) & \checkmark & & & & \checkmark & \checkmark & & & & & & & \checkmark & & \\ \cline{2-17}
& BERT (2019) & & \checkmark & & & & & \checkmark & \checkmark & & & & & \checkmark & & \\ \cline{2-17}
& XLNet (2019) & & \checkmark & & & & & & & \checkmark & & & & \checkmark & & \\ \cline{2-17}
& RoBERTa (2019) & & \checkmark & & & & & \checkmark & & & & & & \checkmark & & \\ \cline{2-17}
& ELECTRA (2019) & & \checkmark & & & & & \checkmark & & & \checkmark & & & \checkmark & & \\ \cline{2-17}
& GPT-2 (2019) & & & \checkmark & & \checkmark & & & & & & & & \checkmark & & \\ \cline{2-17}
& T5 (2020) & & & & \checkmark & & & \checkmark & & & & & & \checkmark & & \\ \cline{2-17}
& BART (2020) & & & & \checkmark & & & \checkmark & & & & & & \checkmark & & \\
\hline
\multirow{7}{*}{Code} & SCEMLo (2020) & \checkmark & & & & \checkmark & \checkmark & & & & & & & & \checkmark & \\ \cline{2-17}
& CuBERT (2019) & & \checkmark & & & & & \checkmark & \checkmark & & & & & & \checkmark & \\ \cline{2-17}
& C-BERT (2020) & & \checkmark & & & & & \checkmark & \checkmark & & & & & & \checkmark & \\ \cline{2-17}
& IntelliCode (2020) & & & \checkmark & & & \checkmark & & & & & & & & \checkmark & \\ \cline{2-17}
& CodeBERT (2020) & & \checkmark & & & & & \checkmark & & & \checkmark & & & \checkmark & \checkmark & \\ \cline{2-17}
& GraphCodeBERT (2020) & & \checkmark & & & & & \checkmark & & & & EP & NA & \checkmark & \checkmark & Data Flow \\ \cline{2-17}
& CugLM (2020) & & \checkmark & & & \checkmark & & \checkmark & \checkmark & & & & & & \checkmark & \\ \cline{2-17}
& T5-learning (2021) & & & & \checkmark & & & \checkmark & & & & & & \checkmark & \checkmark & \\ \cline{2-17}
& TreeBERT (2021) & & & & \checkmark & & & & & & & TMLM & NOP & & \checkmark & AST \\
\bottomrule
\end{tabular}
}
\end{table*}
presents an overview of the most prominent pre-trained models in NLP and for code. Each model is characterized along four dimensions: (1) Modules: what is being pre-trained (e.g., the encoder, the decoder, or both); (2) Objectives: the pre-training objectives\footnote{Forward LM~\cite{peters2018deep,radford2019language,karampatsis2020scelmo} aims to predict the next word given the preceding words in a sentence. Backward LM~\cite{peters2018deep,karampatsis2020scelmo} aims to predict the previous word given the words that appear after it in a sentence. Masked LM is the masked language modeling task described in Section~\ref{section:introduction}. NSP~\cite{devlin2019bert} aims to predict whether the second sentence in a sentence pair should immediately follow the first sentence in the pair. Permutation LM~\cite{yang2019xlnet} aims to predict a word using a set of context words randomly selected via the attention mask mechanism. RTD aims to predict which token in the input has been replaced. EP~\cite{guo2020graphcodebert} involves masking 20\% of the edges in a data flow and aims to predict the masked edges. NA~\cite{guo2020graphcodebert} involves masking a certain portion of edges of connecting data flow nodes and code tokens and aims to predict the masked edges. TMLM~\cite{jiang2021treebert} masks paths in the AST input on the encoder side and tokens in the code sequence input on the decoder side, then predicts the token of the masked code. NOP~\cite{jiang2021treebert} is to exchange the positions of some nodes in the path, and distinguish whether the order of nodes in the AST is correct or not.} and (3) Inputs: what information the model assumes as input (e.g., natural language, code and structural information of the code).
We can make a few observations from Table~\ref{table:pre_train_models}. First, while the majority of work has focused on pre-training the encoder, the newest models (T5~\cite{raffel2020exploring}, BART~\cite{lewis2020bart}, and T5-learning~\cite{mastropaolo2021studying}, which is modeled after T5) are all seq2seq pre-training models that allow the encoder and the decoder to be jointly trained. Second, except for GraphCodeBERT~\cite{guo2020graphcodebert} and TreeBERT~\cite{jiang2021treebert}, all pre-trained models in SE reuse training objectives designed for natural languages, with MLM being the most popular pre-training task. This indicates that the selection of pre-training tasks, which dictates what knowledge will be acquired and exploited by a model, is an area of research that is under-investigated in SE. Finally, while earlier pre-training models in SE assume only source code as input, the later ones all use both code and language.
In our experiments, we will use as our baselines the most recently developed (and also the state-of-the-art) pre-trained models for source code, namely CodeBERT~\cite{feng2020codebert}, GraphCodeBERT~\cite{guo2020graphcodebert}, CugLM~\cite{liu2020multi}, T5-learning~\cite{mastropaolo2021studying} and TreeBERT~\cite{jiang2021treebert}.
\subsection{Structural Information of Source Code}
\label{section:related_structure}
Structural information is very important for understanding source code. ASTs are widely used in code related tasks for representing structure information of code (e.g., ~\cite{zhang2020retrieval,wang2021code,svyatkovskiy2019pythia,hu2018deep,leclair2019neural}), which contains abundant syntactic structure information that cannot be expressed by code sequences. An AST should be flattened with linearization methods, e.g., pre-order traversal~\cite{zhang2020retrieval,wang2021code}, in-order traversal~\cite{svyatkovskiy2019pythia}, and Structure-based Traversal (SBT)~\cite{hu2018deep}, before being fed to an encoder. code2vec~\cite{alon2019code2vec}, code2seq~\cite{alon2018code2seq}, and SLM~\cite{alon2020structural} use a method that linearizes an AST as a series of ``path-contexts'' representing two terminal nodes and the path between them. Jiang et al.~\cite{jiang2021treebert} represent ASTs as the set of paths and then introduce node position embedding to obtain the position of the node in the AST. Besides, neural networks that take trees as input (e.g., tree-LSTMs~\cite{wan2018improving,wan2019multi,lin2021improving}, RvNNs~\cite{zhang2019novel} and GNNs/GCNs~\cite{yang2021multi,leclair2020improved,chen2021holistic}) utilize an AST directly instead of flattening it.
There are also approaches taking data flow and control flow extracted from code as structural information, e.g., ~\cite{guo2020graphcodebert} and~\cite{huo2020control}. However, these flows do not contain structural information as rich as ASTs~\cite{guo2020graphcodebert}. Their advantages over ASTs are that they have a lower demand on hardware and need less training time under the same conditions~\cite{guo2020graphcodebert}.
Taken together, we choose to use ASTs to represent the structural information in SPT-Code.
\section{SPT-Code}
In this section we first introduce the architecture of SPT-Code (Section~\ref{section:architechture}). We then describe the model input (Section~\ref{section:model_input}) and the pre-training tasks (Section~\ref{section:pre_training_tasks}), which are the key innovations of this paper. Finally, we illustrates how to fine-tune SPT-Code when it is applied to downstream tasks (Section~\ref{section:fine-tuning}).
\subsection{Model Architecture}
\label{section:architechture}
Architecturally, SPT-Code is essentially a multi-layer Transformer~\cite{vaswani2017attention}, which is widely used by pre-training models such as BART~\cite{lewis2020bart} and T5~\cite{raffel2020exploring}. As far as the parameter setting of the encoder and decoder is concerned, SPT-Code follows CodeBERT and GraphCodeBERT: (1) the number of layers (i.e., Transformer blocks) $L=12$, (2) the size of model $d_{model}=768$, (3) the dimension of feed forward $d_{ff}=3072$, (4) the number of self-attention heads $h=12$, and (4) the dropout rate $p_{dropout}=0.1$. The total number of parameters is 262M.
To pre-train both classification and generation tasks with an encoder-decoder structure, we adopt the strategy used in BART. In particular, note that the encoder and decoder will continue to learn jointly and collaboratively when pre-trained on classification tasks.
Figure~\ref{figure:model_cls_gen} shows how classification and generation tasks are pre-trained in SPT-Code. Specifically,
\begin{figure}[htbp]
\centering
\subfloat[When SPT-Code is used for classification, the inputs of the encoder and decoder are identical, and the output of the decoder at the last time step is used as the label for the classification.]{\includegraphics[width=0.4\textwidth]{figures/model_cls.pdf}\label{figure:model_cls}}\\
\subfloat[When SPT-Code is used for generation, the whole procedure is the same as the regular sequence-to-sequence model.]{\includegraphics[width=0.4\textwidth]{figures/model_gen.pdf}\label{figure:model_gen}}
\caption{SPT-Code for classification and generation.}
\Description{Input and output of SPT-Code for classification and generation.}
\label{figure:model_cls_gen}
\end{figure}
to use SPT-Code for classification tasks, the input of the decoder is the same as encoder, except that a special symbol ``[SOS]'' (indicating the start of the sequence) is added to the front, and another special symbol ``[EOS]'' (used here as a placeholder for the classification position) is added to the end. The output of the corresponding position of the ``[EOS]'' is then used for classification. For generation tasks, Figure~\ref{figure:model_gen} demonstrates the input and output of the model when translating ``A B C D E'' to ``\textalpha~\textbeta~\textgamma~\textdelta~\textepsilon''. The special symbol ``[EOS]'' denotes the end of the sequence, and the process of generation stops when the model outputs this symbol.
\subsection{Model Input}
\label{section:model_input}
The inputs of the model are three different types of components belonging to a complete \textit{Method} (i.e., it can be invoked by its name), namely, the code tokens, the linearized AST, and the natural language. In this subsection, we will demonstrate this with a real Java method\footnote{https://github.com/Unidata/thredds/blob/d2d68f9eee87f345625211324d71d5dc3e162\\ee1/cdm/src/main/java/thredds/client/catalog/Property.java\#L56-L63} shown in the bottom of Figure~\ref{figure:model_input}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.47\textwidth]{figures/input_formulation.pdf}
\caption{The input of a real world Java code snippet. Due to space constraints, we abbreviated the node names in the AST sequence. ``NL'' denotes the natural language input.}
\Description{Input of a Java method.}
\label{figure:model_input}
\end{figure}
\subsubsection{Code Tokens}
\label{section:code_tokens}
As we can see from Figure~\ref{figure:model_input}, the first part of the input is the code token sequence of a method. We use a lexical analyzer to tokenize the source code and then obtain the tokens $C=\{c_1,c_2,\dots,c_l\}$, where $l$ is the number of code tokens. Specifically, we use the Python standard library\footnote{https://docs.python.org/3.8/library/tokenize.html} to tokenize Python codes. For languages such as Java, JavaScript, PHP and Go, we use the Python binding\footnote{https://pypi.org/project/antlr4-python3-runtime/} of ANTLR 4\footnote{https://github.com/antlr/antlr4} to get the code tokens. The Ruby source code is tokenized by the calling of a Ruby binary program. The source codes of other programming languages are tokenized by the NLTK tokenizer\footnote{https://www.nltk.org/api/nltk.tokenize.html}.
\subsubsection{Linearized AST}
\label{section:linearized_ast}
To represent the second part of the input, i.e., the structural information of the code, we convert an AST into a specially formatted sequence by traversing it, and call the result of this converting a \textit{linearized} AST. We first employ an AST parser\footnote{https://tree-sitter.github.io/tree-sitter/} to get the corresponding AST, then use a traversal method to parse the AST into a sequence.
Instead of using the original SBT (please refer to~\cite{hu2018deep,hu2020deep} for more details), which has been shown to be more effective than classical traversal methods (i.e., pre-order traversal) but tend to produce excessively long sequences that are on average more than three times the length of the code, we propose a simplified version of SBT called X-SBT (XML-like SBT) to traverse ASTs. X-SBT can reduce the length of the resulting sequence of traversals by more than half, Figure~\ref{figure:xsbt}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{figures/xsbt.pdf}
\caption{An example of the SBT and X-SBT, traversal sequences are formatted and indented for a better illustration.}
\Description{Comparison of SBT and X-SBT.}
\label{figure:xsbt}
\end{figure}
shows a comparison of SBT and X-SBT. It can be seen that when traversing AST, SBT takes two tokens, ``('' and the node name, as the starting flag of a certain AST node, and takes ``)'' and the name as the ending flag. We make one observation: for non-terminal nodes (i.e., nodes which are not leaves), we can replace the starting flag with one token in an XML-like form, similarly for the ending flag. For terminal nodes, i.e., leaf nodes, we can further merge the starting and ending tokens into one token. Therefore, it is easy to prove that X-SBT shortens the length by more than half.
However, X-SBT sequences are still long. To further shorten X-SBT sequences, we make X-SBT traverse only the nodes at or above the \textit{expression} level in the AST. This will also reduce redundancy. Commonly, the AST contains both lexical and syntactic information, where the lexical information is already represented by the first part of the input (i.e., code tokens). Since the lexical information in the AST is concentrated on the terminal nodes, we only keep the nodes at or above the \textit{expression} level, so that it contains only syntactic information. Figure~\ref{figure:xsbt_example}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.47\textwidth]{figures/xsbt_example.pdf}
\caption{The AST of the ``for'' statement block of the Java code snippet in Figure~\ref{figure:model_input}. Non-terminal nodes are ellipses and terminal nodes are represented as rectangles. The green nodes will be traversed by X-SBT at the expression level, while the red ones will not.}
\Description{Example of the AST.}
\label{figure:xsbt_example}
\end{figure}
shows the nodes in the AST that will be traversed by the X-SBT at the \textit{expression} level. Obviously, it traverses only one subtree of the AST, and consequently, it can further reduce the length of the sequence by ignoring some fine-grained (lexical) information that is already present in the code tokens. To conclude, we traverse the AST using X-SBT at the \textit{expression} level to obtain the tokens $A=\{a_1,a_2,\dots,a_m\}$, where $m$ is the length of the sequence.
\subsubsection{Natural Language}
\label{section:nl_input}
For extracting natural language information from the code only, we derive the method name and API call sequence of the code\footnote{The reason we do not use documentation in code, such as docstring and in-line comments, is that documentation is not always available as we mentioned in Section~\ref{section:introduction}.}. For example, the tokens extracted from the code snippet in Figure~\ref{figure:model_input} are \texttt{removeDups, size, contains, add}. We further split each token of the form \textit{CamelCase} and \textit{snake\_case}, so \texttt{removeDups} is split into \texttt{remove} and \texttt{dups}. Then we take the resulting linear sequence of tokens, $N=\{n_1,n_2,\dots,n_p\}$, as our natural language input, where $p$ is the number of tokens.
After completing the construction of the three input parts, we concatenate them and use a special symbol, i.e., ``[SEP]'', to separate the three inputs. Therefore, the input is represented as
\begin{equation}
Input=c_1,\dots,c_l,[\textup{SEP}],a_2,\dots,a_m,[\textup{SEP}],n_1,\dots,n_p
\end{equation}
\subsection{Pre-Training Tasks}
\label{section:pre_training_tasks}
In this section, we introduce the three tasks in the order they are used for (sequential) pre-training.
\subsubsection{Code-AST Prediction}
\label{section:cap}
The first pre-training task is Code-AST Prediction (CAP). Since we add structural information to the input, in CAP, we expect to have the model learn such information about the structure represented by X-SBT. So we introduce this binarized task that can be trivially generated from any code. Specifically, when we construct the input representation, namely $Input=C$,[SEP],$A$,[SEP],$N$, 50\% of the time $A$ is the actual AST sequence corresponding to $C$ (labeled as \texttt{IsAST}), and 50\% of the time it is a random AST sequence from the dataset (labeled as \texttt{NotAST}). As we show in Figure~\ref{figure:model_cls}, ``label'' is used for Code-AST Prediction. Note that CAP is closely similar to the next sentence prediction task~\cite{devlin2019bert}.
\subsubsection{MASS}
\label{section:mass}
Since SPT-Code is based on the encoder-decoder architecture, we expect to pre-train SPT-Code in a sequence-to-sequence style. Therefore, we then adopt MASS, which seeks to reconstruct a sentence fragment given the remaining part of the sentence in the encoder-decoder framework. We employ a modified version of MASS as one of our pre-training tasks, with the intention of training the model to understand, infer and generate code sequences.
Given an input $Input=C$,[SEP],$A$,[SEP],$N$, we denote $C^{\backslash u:v}_{\textup{origin}}$ as a modified version of $C$ where its fragment from position $u$ to $v$ is masked, that is,
\begin{equation}
C^{\backslash u:v}_{\textup{origin}}=\{c_0,\dots,c_{u-1},[MASK],\dots,[MASK],c_{v+1},\dots,c_l\}
\end{equation}
where $0<u<v<l$. In contrast, $C^{u:v}$ denotes the fragment of $C$ from $u$ to $v$. In our work, we merge a number of consecutive ``[MASK]'' symbols into one, then $C^{\backslash u:v}_{\textup{origin}}$ is replaced by $C^{\backslash u:v}$:
\begin{equation}
C^{\backslash u:v}=\{c_0,\dots,c_{u-1},[MASK],c_{v+1},\dots,c_l\}
\end{equation}
We utilize this modified version of MASS to pre-train our sequence to sequence model by predicting the fragment $C^{u:v}$ taking the input $Input=C^{\backslash u:v}$,[SEP],$A$,[SEP],$N$. Then, $k=u-v+1$ is the number of the masked consecutive tokens, and we follow Song et al.~\cite{song2019mass} and set $k$=50\% of $l$ to achieve the best performance.
\subsubsection{Method Name Generation}
\label{section:mng}
With the last pre-training task, we would expect the model to learn information such as the intent and functionality of the code. The method name, which is present in every method, can be seen as an extremely concise summary of the method. Xie et al.~\cite{xie2021exploiting} analyze the method name and code summary in a Java dataset built by Leclair and McMillan~\cite{leclair2019recommendations}, and find that on average 50.6\% of the words in method names appear in the corresponding summaries, and 21.3\% of the words in the summaries appear in the corresponding method names. For about 20\% of the methods, all the words in the method names appear in the corresponding summaries. Therefore, they improve code summarization performance successfully by predicting method names in advance.
Therefore, we exploit method names in this pre-training task, which we call Method Name Generation (MNG). Given the model input $Input=C$,[SEP],$A$,[SEP],$N$, we denote $name=c_i$ as the method name in $C$, where $i$ is the index of the method name token in $C$. From Section~\ref{section:nl_input}, we learn that the split $name=c_i$ also appears in $N$, so we remove the tokens from $N$ that are derived from the method name split. Assuming that the method name is split into $s$ subtokens, the final input of MNG will be as follows:
\begin{equation}
\begin{split}
Input_{MNG}=&c_0,\dots,c_{i-1},[MASK],c_{i+1},\dots,c_l,\\
&[SEP],a_1,a_2,\dots,a_m,\\
&[SEP],n_{s+1},n_{s+2},\dots,n_p
\end{split}
\end{equation}
Then we make the decoder try to generate the split method name, i.e., $n_1,n_2,\dots,n_s$.
\subsection{Fine-Tuning}
\label{section:fine-tuning}
We group all downstream tasks into two categories: classification and generation. For classification tasks, we use the setting shown in Figure~\ref{figure:model_cls}; otherwise, we use the setting in Figure~\ref{figure:model_gen}. For each task, we simply plug in the task-specific inputs and outputs into SPT-Code and fine-tune all the parameters end-to-end. For example, in code search, which we cast as a classification task, when obtaining a representation vector of a natural language query, we take only the natural language as input and do not consider the first two parts of the input, i.e., code and AST. For code summarization, which we cast as a generation task, we only include the natural language summary in the output. We will describe the task-specific details in the corresponding subsections of Section~\ref{section:experiment}.
\section{Experiment}
\label{section:experiment}
In this section, we first introduce the pre-training data and settings (Section~\ref{section:exp_pre_train}) and how we fine-tune on the downstream tasks (Section~\ref{section:exp_fine_tune}). Then, we answer four research questions (Section~\ref{section:exp_evaluation}). Finally, we manually evaluate SPT-Code through quantitative and qualitative analysis.
\subsection{Pre-Training}
\label{section:exp_pre_train}
\subsubsection{Dataset}
\label{section:pre_training_dataset}
The dataset we use for pre-training SPT-Code is CodeSearchNet~\cite{husain2020codesearchnet}, which has also been used to pre-train CodeBERT~\cite{feng2020codebert}, GraphCodeBERT~\cite{guo2020graphcodebert} and T5-learning~\cite{mastropaolo2021studying}. The CodeSearchNet corpus is programmatically obtained by scraping open-source repositories and pairing individual functions with their (processed) documentation as natural language annotation. It includes more than 6.4M codes from 6 programming language, i.e., Java, Python, JavaScript, PHP, Go and Ruby. The data statistics are shown in Table~\ref{table:csn_statistics}.
\begin{table}[htbp]
\centering
\caption{Pre-Training Dataset Size Statistics}
\label{table:csn_statistics}
\small
\begin{tabular}{lrr}
\toprule
\multirow{2}{*}{\textbf{Language}} & \multicolumn{2}{c}{\textbf{\# Methods/Function}} \\ \cline{2-3}
& \multicolumn{1}{r}{\textbf{w/ documentation}} & \multicolumn{1}{r}{\textbf{All}} \\
\midrule
Java & 542,991 & 1,569,889 \\
Python & 503,502 & 1,156,085 \\
JavaScript & 157,988 & 1,857,835 \\
PHP & 717,313 & 977,821 \\
GO & 347,789 & 726,768 \\
Ruby & 57,393 & 164,048 \\
\hline
All & 2,326,976 & 6,452,446 \\
\bottomrule
\end{tabular}
\end{table}
Since our input can be extracted from completely unlabeled data, we can make use of all the 6.4M data instances in CodeSearchNet. However, CodeBERT and GraphCodeBERT use labels (i.e., the documentation) in the input for training both code and natural language, so they are both pre-trained on all the data in the ``w/ documentation'' column of the ``All'' row, which we named \textit{CSN$_{w/ doc}$}. T5-learning, inputs the code and the documentation into the model separately when pre-training, so it resembles SPT-Code in the sense that it does not utilize the code-document correspondence. Nevertheless, the publicly available implementation of T5-learning uses only the data in the ``All'' column of ``Java'' row, named \textit{CSN$_{Java}$}.
\subsubsection{Settings}
\label{section:pre_train_settings}
The three pre-training tasks of SPT-Code are performed sequentially. We tried different orders and different numbers of epoch for these tasks and found that better results can be achieved by first pre-training CAP for 10 epochs, then MASS for 30 epochs, and eventually MNG for 30 epochs. All three tasks use cross-entropy as the loss function. We use AdamW~\cite{loshchilov2018decoupled} as our optimizer, the initial learning rate is 5e-5 and the warmup step is set as 2000. We pre-train SPT-Code on 4$\times$NIVDIA A100s\footnote{The GPUs are provided by Alibaba Group, which employ EFLOPS~\cite{eflops} and Alibaba Collective Communication Library (ACCL)~\cite{accl} techniques. EFLOPS is a high performance distributed training system that can achieve near-linear scalability of overall throughput, and ACCL brings the performant efficiency of EFLOPS architecture to general cluster systems and Cloud scenarios, which is able to achieve fully non-congested communication.} with a total batch size of 256.
We use Byte-Pair Encoding (BPE)~\cite{sennrich2016neural} to tokenize the code and the natural language (after CamelCase and snake\_case splitting), and use regular word tokenizer to tokenize X-SBT sequences. The tokenizer is built upon the whole pre-training data, and will be directly used on each downstream tasks without any modification.
\subsection{Fine-Tuning on Downstream Tasks}
\label{section:exp_fine_tune}
In this subsection, we detail the process of fine-tuning the pre-trained SPT-Code on the five downstream tasks. For each downstream task, we give a brief introduction, tell the datasets for fine-tuning, present the compared baselines, and eventually illustrate the metrics for evaluating.
\subsubsection{Code Summarization}
\label{section:fine_tune_summarization}
Code summarization, also known as the process of code summary or code comment generation, is the task of automatically generating a natural language description for a piece of source code that summarizes the overall actions of the code snippet accurately, adequately, and concisely~\cite{sridhara2010towards}. Work in this area of research has generally focused on generating a (typically short) natural language comment from a given method.
\textbf{Datasets.} In addition to CodeSearchNet, we use two widely used classical datasets that are collected from open source repositories in GitHub, i.e., the Java dataset (JCSD)~\cite{hu2018summarizing} and the Python dataset (PCSD)~\cite{miceli2017parallel}. As for JCSD, the comment of each method is the first sentence extracted from its Javadoc, and the comment of a method in PCSD is its docstring.
\textbf{Baselines.} We adopt CodeBERT~\cite{feng2020codebert}, GraphCodeBERT~\cite{guo2020graphcodebert}, CugLM~\cite{liu2020multi}, T5-learning~\cite{mastropaolo2021studying} and TreeBERT~\cite{jiang2021treebert} as baselines for all dataset
\footnote{We resize all the pre-trained models here and in all the following downstream tasks identical to ours.}. For JCSD, PCSD and CodeSearchNet's Java/Python datasets, we adopt NerualCodeSum~\cite{ahmad2020transformer}, which employs Relative Position~\cite{shaw2018self} and Copy Attention~\cite{see2017get} upon vanilla Transformer~\cite{vaswani2017attention}, as another baseline\footnote{We have tried to run it as Java for other languages in CodeSearchNet but got completely nonsensical output.}. Besides, we use three more baselines for JCSD and PCSD, namely, DeepCom~\cite{hu2018deep}, Rencos~\cite{zhang2020retrieval}, and SiT~\cite{wu2021code}, which are widely compared or recently proposed RNN or transformer-based models dedicated for code summarization.
\textbf{Metrics.} We use BLEU (B.)~\cite{papineni2002bleu}, METEOR (M.)~\cite{banerjee2005meteor} and ROUGE-L (R.L)~\cite{lin2002manual}, which are widely used in fields such as machine translation and text summarization, to measure the similarity between the sentences generated by the model and the goal sentences.
\subsubsection{Code completion}
\label{section:fine_tune_completion}
Code completion is to generate code given its surrounding code as context. We fine-tune SPT-Code on a task called \textit{any-code} completion~\cite{alon2020structural} here. Unlike \textit{restricted} completion where the target code contains only primitive types (e.g., \texttt{int} and \texttt{string}) and excludes user-defined functions, \textit{any-code} completion aims to generate code in a general-purpose programming language without any restriction on its vocabulary or structure. Specifically, consider a program $\mathcal{P}$ and some part of the program $p\in \mathcal{P}$, any-code completion makes the model to predict $p$ given the rest of the program $\mathcal{P}^-=\mathcal{P}/p$.
\textbf{Datasets.} The dataset we use for any-code completion is the Java dataset provided by Alon et al.~\cite{alon2020structural}. It is based on the Java-small dataset~\cite{alon2018code2seq}, which contains the least code duplication~\cite{allamanis2019adverse}. Alon et al. create any-code completion examples by selecting every expression larger than a single AST node as the target $p$, using the reminder of the method as the context $\mathcal{P}$. They also filter many methods to clean the dataset and make the task harder. The resulting dataset contains 1.3M/10k/20k train/dev/test examples.
\textbf{Baselines.} Besides CodeBERT, GraphCodeBERT, CugLM, T5-learning and TreeBERT, some specific methods for this task are also compared. Code2seq~\cite{alon2018code2seq} represents a code snippet as the set of compositional paths in its AST and uses attention to select the relevant paths while decoding. Transformer w/ copy uses the implementation of Open-NMT~\cite{klein2017opennmt} with a copy mechanism~\cite{gu2016incorporating}. Seq2tree w/ copy~\cite{aharoni2017towards} learns to generate the linearized, subtokenized target AST for code completion. SLM~\cite{alon2020structural} leverages joint modeling of an AST and its missing subtree using a structural language model, which estimates the probability of the program’s AST by decomposing it into a product of conditional probabilities over its nodes.
\textbf{Metrics.} Following Alon et al.~\cite{alon2020structural}, we use exact match accuracy at 1 (Acc@1) and 5 (Acc@5) for evaluation. An exact match is counted if and only if the sentence generated by the model is identical to the goal sentence (ignoring cases and whitespaces). If we let the model generate $k$ candidate sentences with the highest probability and if an exact match is counted while any one of the candidates matches the goal sentence exactly, the ultimate exact match accuracy is Acc@$k$ (e.g., if only one candidate, it is Acc@1).
\subsubsection{Bug fixing}
\label{section:fine_tune_bug_fix}
Bug fixing, aka. code repair or code refinement, aims to automatically fix bugs in the code. It can help software developers locate and fix bugs and thus save a lot of time~\cite{jorgensen2006systematic,tufano2019empirical}.
\textbf{Datasets.} We use the dataset collected by Tufano et al.~\cite{tufano2019empirical}, who extract method-level pairs of buggy and corresponding fixed code named BFPs (bug-fix pairs) from bug-fixing commits in thousands of GitHub Java repositories. Each BFP is composed of a tuple $BFP=<m_b, m_f>$, where $m_b$ represents a buggy code component, $m_f$ represents the corresponding fixed code. Based on the code size, Tufano et al. provide two datasets: BFP$_{\textup{small}}$ and BFP$_{\textup{medium}}$, with the former having a code length below 50 and the latter having a code length between 50 and 100.
\textbf{Baselines.} The original method proposed by the dataset creator~\cite{tufano2019empirical} is marked as \textit{Tufano et al.} S2S+COPYSAPN~\cite{panthaplackel2021copy} is an extension of seq2seq models which can copy entire spans of the input to the output in one step and reduce the number of decisions required during inference. CodeBERT, GraphCodeBERT, CugLM, T5-learning and TreeBERT are also used as baselines.
\textbf{Metrics.} We report Acc (=Acc@1) and BLEU on both datasets. Both $m_b$ and $m_f$ are abstracted (please refer to \cite{tufano2019empirical} for more details) before being fed to the model, and we do not translate the abstracted code predicted by the model back into source code before evaluation using the metrics because the result before and after translation are the same.
\subsubsection{Code Translation}
\label{section:fine_tune_translation}
Code translation is important for migrating legacy code in one language into an ecosystem built in a different language.
\textbf{Datasets.} Following Chen et al.~\cite{chen2018tree} and Guo et al.~\cite{guo2020graphcodebert}, we conduct our experiments on a dataset containing several open-source projects, which have both a Java and a C\# implementation.
\textbf{Baselines.} Baselines are Naive (i.e., directly copying the source code as the translation result), vanilla Transformer~\cite{vaswani2017attention}, CodeBERT, GraphCodeBERT, CugLM, T5-learning and TreeBERT.
\textbf{Metrics.} We use Acc and BLEU as metrics.
\subsubsection{Code Search}
\label{section:fine_tune_search}
Code search aims to find the code snippet that most closely resembles the semantics of the given natural language query statement from a set of codes, which is named \textit{codebase} in this task. It is a good choice for testing the performance of SPT-Code in classification mode. For each code-query pair $(c,q)$, we compute the representation vectors of $c$ and $q$, i.e., $V_c$ and $V_q$ respectively. Then, we randomly take another query statement from the dataset as a negative sample, denoted as $q-$, whose representation vector is $V_{q-}$. To ensure the effectiveness of the negative samples, we restrict the BLEU score~\cite{papineni2002bleu} between $q-$ and $q$ to be lower than 0.3. The training loss is:
\begin{gather}
\mathcal{L}_{search}=\sum_{(c,q,q-)\in D} \max (0, \varepsilon -\cos (V_c,V_q)+\cos (V_c,V_{q-}))\\
\cos (V_c,V_q)=\frac{V_c\cdot V_q}{\Vert V_c\Vert \Vert V_q\Vert }
\end{gather}
where $D$ is the dataset, $\varepsilon$ is a fixed value of 0.05 (following Gu et al.~\cite{gu2018deep}), and $\cos$ denotes the calculation of cosine similarity. In the evaluation, we first get the representation vectors of all codes in codebase as candidates. Then, for each vector of query statements, we select a few code representation vectors from the candidates that are closest to the vector of query statements. Finally, the metric is calculated.
\textbf{Datasets.} We use CodeSearchNet~\cite{husain2020codesearchnet}.
\textbf{Baselines.} We use CNN~\cite{kim2014convolutional} (i.e., 1D convolutional neural network), Bi-GRU~\cite{cho2014learning}, and Transformer (i.e., vanilla Transformer~\cite{vaswani2017attention}) as baselines in addition to CodeBERT and GraphCodeBERT.
\textbf{Metrics.} We use Mean Reciprocal Rank (MRR) for evaluation. MRR is a statistic measure for evaluating search algorithms. The reciprocal rank of a query response is the multiplicative inverse of the rank of the first correct answer: 1 for first place, 1/2 for second place. Therefore, the mean reciprocal rank is the average of the reciprocal ranks of results for a sample of queries Q:
\begin{equation}
\textup{MRR}=\frac{1}{|Q|}\sum_{i=1}^{|Q|}\frac{1}{\textup{rank}_i}
\end{equation}
\subsection{Evaluation}
\label{section:exp_evaluation}
We evaluate SPT-Code by answering four research questions.
\subsubsection*{RQ1: How effective is SPT-Code compared with the state-of-the-art baselines and other code pre-trained models on five downstream tasks?}
\label{section:rq_comparsion}
We conduct experiments on five downstream tasks introduced in the previous subsection. The results on classical and CodeSearchNet code summarization are shown in Tables~\ref{table:results_summarization}
\begin{table}[htbp]
\centering
\caption{Results on classical code summarization.}
\label{table:results_summarization}
\small
\begin{tabular}{lcccccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Methods}} &
\multicolumn{3}{c}{JCSD} &
\multicolumn{3}{c}{PCSD} \\ \cline{2-7}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{B.} &
\multicolumn{1}{c}{M.} &
\multicolumn{1}{c}{R.L} &
\multicolumn{1}{c}{B.} &
\multicolumn{1}{c}{M.} &
\multicolumn{1}{c}{R.L} \\
\midrule
DeepCom & 37.5 & 22.0 & 51.2 & 20.4 & 9.5 & 36.8 \\
Rencos & 36.9 & 26.5 & 51.2 & 28.9 & 20.4 & 45.0 \\
NerualCodeSum & 44.5 & 26.4 & 54.7 & 32.3 & 19.7 & 46.8 \\
SiT & 45.3 & 27.3 & 55.0 & 33.8 & 20.7 & 48.2 \\
\hline
CodeBERT & 45.2 & 26.2 & 54.7 & 33.3 & 21.6 & 49.2 \\
GraphCodeBERT & 46.0 & 26.5 & 56.5 & 33.9 & 22.1 & 50.4 \\
CugLM & 46.1 & 26.3 & 55.8 & 34.3 & 21.6 & 49.7 \\
T5-learning & 44.2 & 26.5 & 53.9 & 34.1 & 22.6 & 50.1 \\
TreeBERT & 47.9 & 27.2 & 56.6 & 34.7 & 23.0 & 50.5 \\
\hline
SPT-Code & \textbf{49.1} & \textbf{32.4} & \textbf{58.2} & \textbf{36.1} & \textbf{26.9} & \textbf{52.0} \\
\bottomrule
\end{tabular}
\end{table}
and Table~\ref{table:results_summarization_csn},
\begin{table*}[htbp]
\centering
\caption{Results on CodeSearchNet code summarization.}
\label{table:results_summarization_csn}
\small
\begin{tabular}{lcccccccccccccccccc}
\toprule
\multirow{2}{*}{Methods} &
\multicolumn{3}{c}{Java} &
\multicolumn{3}{c}{Python} &
\multicolumn{3}{c}{JavaScript} &
\multicolumn{3}{c}{PHP} &
\multicolumn{3}{c}{Go} &
\multicolumn{3}{c}{Ruby} \\
\cline{2-19} &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L &
B. & M. & R.L \\
\midrule
NeuralCodeSum & 10.7 & 12.9 & 25.9 & 9.4 & 4.9 & 17.0 & - & - & - & - & - & - & - & - & - & - & - & - \\
CodeBERT & 13.6 & 16.4 & 32.2 & 10.7 & 6.0 & 21.0 & 10.0 & 10.5 & 22.2 & 20.1 & 19.3 & 42.7 & 16.0 & 11.7 & 30.4 & 8.3 & 7.5 & 17.2 \\
GraphCodeBERT & 14.5 & 17.8 & 33.5 & 11.0 & 7.4 & 22.1 & 10.9 & 11.4 & 24.9 & 20.0 & 20.2 & 42.8 & 16.5 & 12.5 & 30.1 & 8.3 & 8.1 & 18.0 \\
CugLM & 13.2 & 16.1 & 31.7 & 11.4 & 7.5 & 22.0 & 10.0 & 10.4 & 21.8 & 17.8 & 18.2 & 41.9 & 17.0 & 12.7 & 29.7 & 7.4 & 6.8 & 15.3 \\
T5-learning & 13.5 & 16.8 & 34.2 & 9.7 & 5.5 & 16.3 & 9.0 & 9.2 & 19.6 & 18.4 & 17.9 & 35.6 & 15.3 & 11.0 & 25.9 & 7.8 & 6.7 & 15.5 \\
TreeBERT & 13.8 & 17.1 & 34.3 & 11.1 & 7.2 & 21.6 & 10.3 & 10.5 & 22.0 & 18.0 & 19.1 & 42.4 & 17.1 & 12.8 & 30.1 & 7.4 & 6.9 & 15.5 \\
\hline
SPT-Code & \textbf{16.8} & \textbf{20.6} & \textbf{36.3} & \textbf{12.8} & \textbf{14.2} & \textbf{27.1} & \textbf{12.8} & \textbf{17.2} & \textbf{28.8} & \textbf{20.4} & \textbf{20.6} & \textbf{43.7} & \textbf{18.8} & \textbf{16.6} & \textbf{38.1} & \textbf{8.4} & \textbf{11.1} & \textbf{20.0} \\
\bottomrule
\end{tabular}
\end{table*}
respectively. Tables~\ref{table:results_completion}-\ref{table:results_search} show the results of code completion, bug fixing, code translation and code search, respectively.
\begin{table}[htbp]
\centering
\caption{Results on any-code completion.}
\label{table:results_completion}
\small
\begin{tabular}{lcc}
\toprule
Methods & Acc@1 & Acc@5 \\
\midrule
code2seq & 10.44 & 15.47 \\
Transformer w/ copy & 16.68 & 24.12 \\
seq2tree w/ copy & 16.46 & 22.89 \\
SLM & 18.00 & 24.77 \\
\hline
CodeBERT & 17.23 & 24.79 \\
GraphCodeBERT & 18.21 & 25.00 \\
CugLM & 18.43 & 25.51 \\
T5-learning & 16.03 & 23.78 \\
TreeBERT & 17.17 & 24.73 \\
\hline
SPT-Code & \textbf{19.09} & \textbf{26.57} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{Results on bug fixing.}
\label{table:results_bug_fix}
\small
\begin{tabular}{lcccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Methods}} &
\multicolumn{2}{c}{BFP$_{\textup{small}}$} &
\multicolumn{2}{c}{BFP$_{\textup{medium}}$} \\ \cline{2-5}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} \\
\midrule
Tufano et al. & 9.27 & - & 3.21 & - \\
S2S+COPYSPAN & \textbf{17.6} & - & 7.9 & - \\
\hline
CodeBERT & 6.23 & 61.47 & 2.43 & 68.54 \\
GraphCodeBERT & 7.49 & 63.40 & 3.16 & 67.33 \\
CugLM & 14.55 & 67.24 & 7.84 & 71.23 \\
T5-learning & 12.01 & 65.02 & 4.51 & 60.89 \\
TreeBERT & 13.15 & 65.74 & 7.61 & 70.67 \\
\hline
SPT-Code & 17.54 & \textbf{75.10} & \textbf{10.86} & \textbf{87.88} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{Results on code translation.}
\label{table:results_translation}
\small
\begin{tabular}{lcccc}
\toprule
\multicolumn{1}{l}{\multirow{2}{*}{Methods}} &
\multicolumn{2}{c}{Java $\to$ C\#} &
\multicolumn{2}{c}{C\# $\to$ Java} \\ \cline{2-5}
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} &
\multicolumn{1}{c}{Acc} &
\multicolumn{1}{c}{BLEU} \\
\midrule
Naive & 00.0 & 18.52 & 00.0 & 18.66 \\
Transformer & 33.4 & 56.05 & 38.8 & 50.84 \\
\hline
CodeBERT & 58.8 & 79.34 & 57.3 & 71.10 \\
GraphCodeBERT & 59.3 & 80.33 & 58.1 & 72.39 \\
CugLM & 61.6 & 82.72 & 59.5 & 74.89 \\
T5-learning & 54.1 & 81.35 & 48.4 & 77.10 \\
TreeBERT & 60.0 & 80.92 & 57.8 & 77.15 \\
\hline
SPT-Code & \textbf{64.07} & \textbf{90.34} & \textbf{60.29} & \textbf{86.10} \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[htbp]
\centering
\caption{Results on code search.}
\label{table:results_search}
\small
\begin{tabular}{lcccccc}
\toprule
Methods & Java & Python & JS & PHP & Go & Ruby \\
\midrule
CNN & 0.262 & 0.241 & 0.224 & 0.261 & 0.676 & 0.274 \\
Bi-GRU & 0.312 & 0.298 & 0.195 & 0.340 & 0.691 & 0.214 \\
Transformer & 0.404 & 0.399 & 0.289 & 0.430 & 0.729 & 0.278 \\
\hline
CodeBERT & 0.673 & 0.670 & 0.618 & 0.624 & 0.879 & 0.674 \\
GraphCodeBERT & 0.690 & 0.692 & \textbf{0.643} & 0.647 & \textbf{0.896} & \textbf{0.701} \\
\hline
SPT-Code & \textbf{0.700} & \textbf{0.699} & 0.641 & \textbf{0.651} & 0.895 & \textbf{0.701} \\
\bottomrule
\end{tabular}
\end{table}
First of all, we can see that whether the baselines are dedicated to a specific downstream task or pre-trained models, SPT-Code clearly outperforms them in the vast majority of cases, including code summarization, code completion, bug fixing on BFP$_{\textup{medium}}$, and code translation. This is because during the pre-training process, SPT-Code learns the representation of codes from a large amount of data, as well as the connection between codes and structural and natural language information. In addition, the model is enhanced to generate code and natural language sequences by pre-training.
As for bug fixing on BFP$_{\textup{small}}$, the accuracy of SPT-Code is slightly lower than that of S2S+COPYSPAN, which achieves the best accuracy on BFP$_{\textup{small}}$. The reason we believe is two-fold. The first is the dataset, where the lengths of the code are up to 50, which allows RNN-based S2S+COPYSPAN to adequately cope with inputs of these lengths. Another reason is that the mechanism proposed by S2S+COPYSPAN to copy the entire span of the input is well suited for a task like bug fixing, where only individual modifications are made on the input. However, in the next RQ, we will show that SPT-Code still outperforms S2S+COPYSPAN on BFP$_{\textup{small}}$ after removing MNG from the pre-training tasks.
Finally, we can learn that SPT-Code performs comparably to GraphCodeBERT on code search, which suggests that although we designed our model to perform better on generation tasks, it does not come at the cost of the performance on classification tasks.
\subsubsection*{RQ2: How do the three pre-training tasks, as well as the AST and the natural language input contribute to the model's performance on the different downstream tasks?}
\label{section:rq_component}
In order to find the impact of each component to the overall performance of SPT-Code, we conduct ablation study on all downstream tasks. We remove each/all pre-training task, and AST or/and natural language part from the input, respectively. Owing to space limitations, for code summarization, we only show results on Java and Python in CodeSearchNet\footnote{Java and Python represent static and dynamic languages, respectively.}; and for code search, we only show results on Java and Go\footnote{The results for both are above and below GraphCodeBERT, respectively.}. The results are shown in the first three groups of Table~\ref{table:ablation_mixed}.
\begin{table*}
\centering
\caption{Ablation study on downstream tasks.}
\label{table:ablation_mixed}
\resizebox{\textwidth}{!}{
\begin{tabular}{lcccccc|cc|cccc|cccc|cc}
\toprule
\multirow{3}{*}{Methods} &
\multicolumn{6}{c|}{Summarization} &
\multicolumn{2}{c|}{Completion} &
\multicolumn{4}{c|}{Bug fixing} &
\multicolumn{4}{c|}{Translation} &
\multicolumn{2}{c}{Search}\\ \cline{2-19}
&
\multicolumn{3}{c}{Java} &
\multicolumn{3}{c|}{Python} &
\multirow{2}{*}{Acc@1} &
\multirow{2}{*}{Acc@5} &
\multicolumn{2}{c}{BFP$_\textup{small}$} &
\multicolumn{2}{c|}{BFP$_\textup{medium}$} &
\multicolumn{2}{c}{Java $\to$ C\#} &
\multicolumn{2}{c|}{C\# $\to$ Java} &
\multirow{2}{*}{Java} &
\multirow{2}{*}{Go} \\ \cline{2-7} \cline{10-17}
& B. & M. & R.L & B. & M. & R.L & & & Acc & BLEU & Acc & BLEU & Acc & BLEU & Acc & BLEU & & \\
\midrule
SPT-Code & \textbf{16.79} & \textbf{20.55} & \textbf{36.34} & \textbf{12.77} & \textbf{14.16} & \textbf{27.10} & 19.09 & \textbf{26.57} & 17.54 & \textbf{75.10} & 10.86 & 87.77 & \textbf{64.07} & \textbf{90.34} & \textbf{60.29} & \textbf{86.10} & \textbf{0.700} & \textbf{0.895} \\
\hline
\ -w/o CAP & 15.85 & 20.33 & 36.30 & 12.26 & 14.16 & 26.57 & 18.42 & 25.63 & 17.20 & 73.54 & 10.06 & 87.26 & 57.25 & 88.80 & 54.40 & 84.66 & 0.668 & 0.872 \\
\ -w/o MASS & 15.75 & 20.19 & 35.70 & 12.16 & 13.48 & 26.00 & 17.42 & 23.71 & 16.30 & 72.71 & 9.65 & 86.10 & 53.93 & 82.09 & 50.55 & 82.32 & 0.686 & 0.889 \\
\ -w/o MNG & 15.45 & 19.86 & 35.65 & 12.01 & 13.81 & 25.98 & \textbf{19.71} & 26.10 & \textbf{18.11} & 73.09 & \textbf{11.96} & \textbf{87.79} & 58.80 & 88.79 & 57.16 & 85.19 & 0.673 & 0.875 \\
\ -w/o all & 13.54 & 17.72 & 33.31 & 11.90 & 12.74 & 25.23 & 13.49 & 19.31 & 14.28 & 72.89 & 8.27 & 85.30 & 49.19 & 79.28 & 47.49 & 78.69 & 0.656 & 0.855 \\
\hline
\ -w/o AST & 16.43 & 20.14 & 36.19 & 11.87 & 13.80 & 25.74 & 18.53 & 25.78 & 16.35 & 72.85 & 10.61 & 87.52 & 55.81 & 87.79 & 56.17 & 82.66 & 0.692 & 0.885 \\
\ -w/o NL & 15.68 & 20.15 & 35.86 & 11.96 & 13.56 & 25.55 & 18.47 & 26.09 & 16.12 & 72.70 & 10.50 & 87.57 & 56.29 & 87.83 & 56.92 & 83.56 & 0.679 & 0.865 \\
\ -only code & 15.35 & 19.97 & 35.63 & 10.25 & 12.74 & 23.90 & 18.33 & 25.88 & 16.01 & 72.55 & 10.19 & 86.96 & 54.13 & 87.34 & 54.59 & 83.81 & 0.669 & 0.864 \\
\hline
\ -CSN$_{\textup{w/ doc}}$ & 15.41 & 19.78 & 35.74 & 12.59 & 13.56 & 26.49 & 17.89 & 23.80 & 16.14 & 74.02 & 10.41 & 87.58 & 60.53 & 88.78 & 58.28 & 84.09 & 0.673 & 0.878 \\
\ -CSN$_{\textup{Java}}$ & 15.67 & 19.72 & 35.22 & 12.40 & 12.60 & 25.24 & 17.76 & 23.54 & 16.10 & 74.51 & 10.18 & 86.96 & 59.43 & 86.55 & 58.11 & 83.07 & 0.678 & 0.866 \\
\bottomrule
\end{tabular}
}
\end{table*}
The first thing we notice is that when we train from scratch (i.e., remove all pre-training tasks) or remove the AST and natural language from the input (i.e., only input code tokens), the performance of SPT-Code consistently drops considerably. Second, when we remove each of the three pre-training tasks, the results decreased in most cases, particularly in code summarization, code translation and code search. Of these, CAP is most useful for code search, MASS is most helpful for code translation, and removing MNG has the greatest impact on code summarization.
Interestingly, we find that for code completion and bug fixing, SPT-Code's performance w.r.t.\ accuracy improves instead when MNG is removed. This is understandable. On the one hand, code completion and bug fixing are tasks where the input and the output are both code, as is MASS, while the output of the MNG task is natural language, and thus the ability to generate natural language trained by MNG is not fully reflected in these two tasks. On the other hand, regarding code completion, MASS can be seen as totally unrestricted any-code completion. Both predict a piece of code based on its context, with the difference that in any-code completion, the piece of code is restricted to be an entire expression, while in MASS, the piece of code is selected completely at random. Therefore, fine-tuning code completion directly after MASS, i.e., removing the MNG, yields a higher result.
When we remove either the AST or the natural language from the input, the results of the model drop, indicating that they both help improve performance. In addition, for code summarization, bug fixing and code search, the results are lower when only natural language is removed, which indicates that in comparison, natural language helps these three tasks more than ASTs do. On the contrary, AST is more helpful for code completion and code translation.
Through ablation, we find that different pre-training tasks show different degrees of influence on the downstream tasks, As a result, appropriate trade-offs of pre-training tasks for different downstream tasks can help the model achieve better performance. ASTs and natural language both have a positive impact on the performance of the model, regardless of the downstream task.
\subsubsection*{RQ3: Is the ability of utilizing more unlabeled data an advantage of SPT-Code?}
\label{section:rq_pre_train_data}
As we know, CodeBERT, GraphCodeBERT and T5-learning are pre-trained on a subset of CodeSearchNet, whereas we can pre-train on the entire dataset. Therefore there is a question of whether the ability to utilize more data for pre-training also gives SPT-Code an unfair advantage. To ensure a fair comparison, we therefore pre-train SPT-Code only on the same data as CodeBERT and GraphCodeBERT (i.e., CSN$_{\textup{w/ doc}}$), and also on the same data as T5-learning (i.e., CSN$_{\textup{Java}}$). Notice, however, that here only the code from the labeled data is utilized and not the labels, even on the CSN$_{\textup{w/ doc}}$ where all samples have labels. So in this case, although we are using the data set of the same size, it actually utilize less information than they do. The results are displayed in the last group of Table~\ref{table:ablation_mixed}.
Comparing the results of SPT-Code with those of ``CSN$_{\textup{w/ doc}}$'' and ``CSN$_{\textup{Java}}$'' in Table~\ref{table:ablation_mixed}, we find that there is a significant performance decrease when the data used for pre-training is shrunk from 6.4M to 2.3M or 1.5M. However, considering the other pre-training models, it still maintains an advantage or is at a comparable level in performance. Therefore, we can conclude that SPT-Code is superior given the same amount of pre-trained data. In addition, the ability to use more unlabeled pre-training data can help SPT-Code achieve higher performance. This ensures that SPT-Code has a very advantageous scalability at the data level compared to the other pre-training models for source code.
\subsubsection*{RQ4: How would the size of the training data for fine-tuning SPT-Code influence its performance on downstream tasks?}
\label{section:rq_fine_tune_data}
To answer this question, we plot learning curves by varying the amount of (task-specific) training data used to fine-tune SPT-Code.
Owing to space limitations, we only report results of code summarization on Java in CodeSearchNet, and bug fixing on BFP$_{\textup{medium}}$. We select $k/10$ data from the training set each time for training, and then test on the same entire testing set. The results are shown in Figure~\ref{figure:data_ablation}.
\begin{figure}[htbp]
\centering
\subfloat[Results on Java CodeSearchNet code summarization.]
{\includegraphics[width=0.42\textwidth]{figures/summarization_subset.pdf}\label{figure:summarization_subset}}\\
\subfloat[Results on BFP$_{\textup{medium}}$ bug fixing.]
{\includegraphics[width=0.42\textwidth]{figures/bug_fix_subset.pdf}\label{figure:bug_fix_subset}}
\caption{Results when $k$ goes from 1 to 10.}
\Description{When the size of training set is different, results of the model.}
\label{figure:data_ablation}
\end{figure}
It is easy to see that the performance shows a decreasing trend in all evaluation metrics as $k$ decreases (i.e., the size of the training set decreases). It proves that even for pre-trained models, the size of the training set plays a key role in the performance as well. Meanwhile, considering Table~\ref{table:results_summarization_csn}, we see that for code summarization, the results of SPT-Code are comparable to that of CodeBERT when the training set size is reduced to 1/10, and the performance is similar to or even higher than GraphCodeBERT when it is reduced to 2/10. Further, recalling Table~\ref{table:results_bug_fix}, we can see that when the training set of BFP$_{\textup{medium}}$ is reduced to 1/10 of the original size, SPT-Code's results are still higher than other pre-trained models, and when it is reduced to half of the original size, SPT-Code's performance can reach the performance of the best baseline model, i.e., S2S+COPYSPAN.
The conclusion is that although there is an inevitable drop in SPT-Code's performance as the size of the training data decreases during fine-tuning, the performance of SPT-Code is comparable to that of the other models when the data used for fine-tuning SPT-Code is reduced to a very small size. This also implies the robustness of SPT-Code.
\subsection{Quantitative Analysis}
\label{section:quantitative_analysis}
We collect the output of SPT-Code and baselines for a wide range of test samples on each downstream task. Then we invited five participants, all of whom are graduate students in software engineering who are themselves not authors of this paper. For each downstream task, the five students are randomly assigned to the same number of samples. If there are multiple datasets for that downstream task, the samples of multiple datasets are the same. For downstream tasks that use Acc as a metric, such as code completion, we focus on the similarity between incorrect and correct answers, and the extent to which the incorrect answer can be helpful when all models fail. The results are shown in Table~\ref{table:results_quantitative}.
\begin{table}[htbp]
\centering
\caption{Results of quantitative analysis.}
\label{table:results_quantitative}
\footnotesize
\begin{tabular}{lrrrr}
\toprule
Dowstream Tasks & \# Sample & \# Better & \# Comparable & \# Worse \\
\midrule
Code Summarization & 800 & 328 & 368 & 104 \\
Code Completion & 600 & 312 & 233 & 55 \\
Bug Fixing & 200 & 89 & 75 & 36 \\
Code Translation & 200 & 83 & 65 & 52 \\
Code Search & 300 & 61 & 185 & 54 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Qualitative Analysis}
\label{section:qualitative_analysis}
By asking the participants in the previous subsection and browsing the output of each model by ourselves, we found that in comparison, SPT-Code can capture the semantic information of identifiers within code more accurately than other pre-trained models, and it can capture the semantic information of code segment globally instead of limiting to a certain region. The extracted code semantics are relatively more widely and evenly distributed in the code segment.
In the case of code summarization, for example, SPT-Code provides more complete and accurate descriptions of the method's overall functionality. Table~\ref{table:qualitative_example} shows summaries generated by different models for an example Java method in CodeSearchNet data.
\begin{table}
\centering
\caption{Qualitative example of SPT-Code and baselines.}
\footnotesize
\begin{tabular}{l}
\toprule
\begin{lstlisting}
void emitLoop() {
for ( ; ; ) {
AppendOnlyLinkedArrayList<Object> q;
q = queue;
if (q == null) {
emitting = false;
return; }
queue = null;
q.forEachWhile(this); } }
\end{lstlisting} \\
\midrule
{\color{teal} CodeBERT}: the queue is being destroyed \\
{\color{cyan} GraphCodeBERT}: emits the next loop in the queue \\
{\color{orange} T5-learning}: this method is called when the queue \\
\hline
{\color{purple} SPT-Code}: emits all of the elements in this queue \\
{\color{purple} \ -w/o AST}: emit the linked list \\
{\color{purple} \ -w/o NL}: emits all entries in the queue \\
\hline
{\color{red} Human Written}: loops until all notifications in the queue has been processed \\
\bottomrule
\end{tabular}
\label{table:qualitative_example}
\end{table}
We find that the ``queue'' object is recognized by all of these models. CodeBERT and T5-learning fail to capture the operation conducted on the queue, i.e., ``emits/emit'', but GraphCodeBERT and SPT-Code does. Compared to SPT-Code, although GraphCodeBERT also understand the relationship between the operation and the object, it fails to choose appropriate words for describing the operation. The reason may be that the data flow can indeed partially (but not fully) capture code structure information. Another interesting observation that may support this inference is that all models considering structure information generate ``in the queue'' correctly, i.e., GraphCodeBERT, SPT-Code, and SPT-Code-w/o NL, and models using AST all generate accurate, readable and smooth summaries. It also implies that data flow is less effectiveness than AST.
\section{Threats to Validity}
\label{section:threats}
\subsubsection*{Construct Validity} Like many existing code pre-training models, SPT-Code uses CodeSearchNet for pre-training. Since CodeSearchNet is also used in the evaluation of code summarization and code search, it is possible that the samples from the test sets for these two tasks have already been used for pre-training. This is not fair to methods that have not been pre-trained with CodeSearchNet, such as NeuralCodeSum in CodeSearchNet code summarization. We also recognize the impact on the results of not removing duplicates. However, our decision of not removing replicates is based on two considerations: (1) ensuring fairness of the comparison: the three code pre-training baselines were pre-trained with all data from CodeSearchNet without duplicate removal; (2) SPT-code is not affected by duplicate data: on one hand, downstream tasks that use pre-training dataset are CSN code summarization and code search, while the pre-training tasks designed for SPT-Code do not use the docstring, i.e., in testing, SPT-Code does not have the advantage of generating more accurate summaries because it has seen a piece of the tested code during pre-training, or has the advantage of better searching for code based on natural language. On the other hand, other downstream tasks that do not use CodeSearchNet have little overlap with CSN.
For rigorous consideration, we remove all test sets in CodeSearchNet and code duplicated with test sets of other downstream tasks\footnote{By using tools provided by Allamanis~\cite{allamanis2019adverse}, we find 9 in classical code summarization, 2 in $\textup{BFP}_\textup{medium}$ and code completion, respectively.}. Then re-pre-train and fine-tune the SPT-Code on three downstream tasks, i.e., JCSD, code completion and $\textup{BFP}_\textup{medium}$ bug fixing. It is found that the results of SPT-Code decrease very little after removing these duplicates\footnote{Results decrease 0.01 BLEU on JCSD, 0.03 Acc@1 on code completion and 0.07 Acc on $\textup{BFP}_\textup{medium}$}. Moreover, it is still not sure whether the change in results is due to the smaller pre-training dataset (shrunk by about 1/10) or the removal of duplicates.
\subsubsection*{Internal Validity} It is widely agreed that hyperparameters have a significant impact on the performance of deep learning models, but hyperparameters of SPT-Code are not tuned experimentally and are set empirically. Therefore, other hyperparameter settings may yield better results.
\subsubsection*{External Validity} Another threat posed by using CodeSearchNet as our pre-training dataset is that CodeSearchNet data is not balanced across six programming languages, which can be seen in Table~\ref{table:csn_statistics}, so our model may not perform the same on different languages, and we cannot guarantee the validity of SPT-Code for programming languages other than these six.
\section{Conclusion}
\label{section:conclusion}
We presented SPT-Code, a large model for source code based on an encoder-decoder architecture.
First, we design three code-specific pre-training tasks to pre-train SPT-Code. Secondly, we propose a new input representation whose is the first method that take into account both natural language and AST form of code, where we also propose a improved version of the AST traversal method, X-SBT. Both our pre-training tasks and input representation allow SPT-Code to be pre-trained on a completely unlabeled dataset.
SPT-Code was then fine-tuned on five code-related downstream tasks. Results indicate that fine-tuning SPT-Code enables it to achieve the state-of-the-art performance on five code-related downstream tasks. Ablation experiments reveal that the three pre-training tasks have different degrees of impact on different downstream tasks, and AST and natural language input also helped improve SPT-Code's performance. To facilitate future research, we also make our code and other artifacts publicly available at \url{https://github.com/NougatCA/SPT-Code}.
\begin{acks}
This work is supported by National Natural Science Foundation of China (61802167, 61802095), Natural Science Foundation of Jiangsu Province, China (BK20201250), Cooperation Fund of Huawei-NJU Creative Laboratory for the Next Programming, and NSF award 2034508. We thank Alibaba Cloud for its high-efficient AI computing service from EFlops Cluster. We also thank the reviewers for their helpful comments. Chuanyi Li and Jidong Ge are the corresponding authors.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
In this paper, we study the existence and the asymptotic behavior of the BGK model for multi-component gases suggested in \cite{mixmodel}:
\begin{align}\label{CCBGK}
\begin{aligned}
\partial_t F_1+v \cdot \nabla_xF_1&=n_1(\mathcal{M}_{11}-F_1)+n_2(\mathcal{M}_{12}-F_1), \cr
\partial_t F_2+v \cdot \nabla_xF_2&=n_2(\mathcal{M}_{22}-F_2)+n_1(\mathcal{M}_{21}-F_2), \cr
F_1(x,v,0)=F_{10}&(x,v), \qquad F_2(x,v,0)=F_{20}(x,v).
\end{aligned}
\end{align}
The distribution function $F_i(x,v,t)$ denotes the number density of $i$-th species particle at the phase point $(x,v) \in \mathbb{T}^3 \times \mathbb{R}^3$ at time $t\in\mathbb{R}^+$ for $i=1,2$. The intra-species Maxwell distributions in the BGK operator $\mathcal{M}_{ii}$ are defined as
\begin{align*}
\mathcal{M}_{ii} = \frac{n_i}{\sqrt{2\pi\frac{T_i}{m_i}}^3} \exp\left(-\frac{|v-U_i|^2}{2\frac{T_i}{m_i}}\right),\quad (i=1,2).
\end{align*}
Here $m_i$ $(i=1,2)$ denotes the mass of a molecule in the $i$-th component,
which we assume that $m_1 \geq m_2$ throughout the paper without loss of generality.
The number density $n_i$, the bulk velocity $U_i$, and the temperature $T_i$ of the $i$-th particle are defined by
\begin{align*}
n_i(x,t)&=\int_{\mathbb{R}^3}F_i(x,v,t)dv,\cr U_i(x,t)&=\frac{1}{n_i}\int_{\mathbb{R}^3}F_i(x,v,t)vdv, \cr T_i(x,t)&=\frac{1}{3n_i}\int_{\mathbb{R}^3}F_i(x,v,t)m_i|v-U_i|^2dv .
\end{align*}
The inter-species Maxwellian distributions are defined by
\begin{align*}
\mathcal{M}_{12} = \frac{n_1}{\sqrt{2\pi\frac{T_{12}}{m_1}}^3} \exp\left(-\frac{|v-U_{12}|^2}{2\frac{T_{12}}{m_1}}\right), \qquad \mathcal{M}_{21} = \frac{n_2}{\sqrt{2\pi\frac{T_{21}}{m_2}}^3} \exp\left(-\frac{|v-U_{21}|^2}{2\frac{T_{21}}{m_2}}\right),
\end{align*}
where the inter-species bulk velocities $ U_{12}, U_{21}$ and the inter-species temperatures $T_{12}, T_{21}$ are defined by
\begin{align*}
U_{12}&=\delta U_1 + (1-\delta)U_2, \cr
U_{21}&=\frac{m_1}{m_2}(1-\delta)U_1 + \left(1-\frac{m_1}{m_2}(1-\delta)\right)U_2,
\end{align*}
and
\begin{align*}
T_{12}&= \omega T_1 + (1-\omega)T_2 + \gamma |U_2-U_1|^2, \cr
T_{21}&= (1-\omega) T_1 + \omega T_2 +\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right) |U_2-U_1|^2.
\end{align*}
Here, the free parameter $\delta$ and $\omega$ denote the momentum interchange rate and the temperature interchange rate, respectively.
In \eqref{CCBGK}, $n_i(\mathcal{M}_{ii}-F_i)$ $(i=1,2)$ are the intra-species relaxation operators for $i$-th gas component, while
$n_j(\mathcal{M}_{ij}-F_i)$ $(i\neq j)$ are the inter-species relaxation operators between different components of the gas. We note that the inter-species relaxation operators describe the interchange of the macroscopic momentum and the temperature between two different species of gas.
These relaxation operators satisfy the following cancellation properties:
\begin{align*}
\begin{split}
&\int_{\mathbb{R}^3}(\mathcal{M}_{ii}-F_i) \left( 1 ,m_iv,m_i|v|^2\right)dv=0,\quad i=1,2 \cr
& \int_{\mathbb{R}^3}(\mathcal{M}_{12}-F_1) dv=0,\quad \int_{\mathbb{R}^3}(\mathcal{M}_{21}-F_2) dv=0, \cr
& \int_{\mathbb{R}^3}n_1(\mathcal{M}_{12}-F_1)m_1v dv+\int_{\mathbb{R}^3}n_2(\mathcal{M}_{21}-F_2)m_2v dv=0, \cr
& \int_{\mathbb{R}^3}n_1(\mathcal{M}_{12}-F_1)m_1|v|^2 dv+\int_{\mathbb{R}^3}n_2(\mathcal{M}_{21}-F_2)m_2|v|^2 dv=0,
\end{split}
\end{align*}
leading to the following conservation laws of the density, total momentum, and total energy:
\begin{align}\label{conservation}
\begin{split}
&\frac{d}{dt} \int_{\mathbb{T}^3 \times \mathbb{R}^3}F_1(x,v,t) dvdx = \frac{d}{dt} \int_{\mathbb{T}^3 \times \mathbb{R}^3}F_2(x,v,t) dvdx =0, \cr
&\frac{d}{dt} \int_{\mathbb{T}^3 \times \mathbb{R}^3}\left(F_1(x,v,t)m_1v + F_2(x,v,t)m_2v\right) dvdx =0, \cr
&\frac{d}{dt} \int_{\mathbb{T}^3 \times \mathbb{R}^3}\left(F_1(x,v,t)m_1|v|^2 + F_2(x,v,t)m_2|v|^2\right) dvdx =0.
\end{split}
\end{align}
To ensure the positivity of all temperatures, the free parameters $\omega$, $\delta$, and $\gamma$ are restricted to
\begin{align*}
\frac{ \frac{m_1}{m_2} - 1}{1+\frac{m_1}{m_2}} \leq \delta < 1, \qquad 0 \leq \omega < 1,
\end{align*}
and
\begin{align*}
0 \leq \gamma \leq \frac{m_1}{3}(1-\delta)\left[\left(1+\frac{m_1}{m_2} \right) \delta+1-\frac{m_1}{m_2} \right].
\end{align*}
For more details, see \cite{mixmodel}. \newline
The main goal of this paper is to establish the global-in-time classical solution of the mixture BGK model when the initial data is close to global equilibrium. For this, we consider the following global equilibrium for each particle distribution function:
\begin{align*}
\mu_1(v)=n_{10}\frac{\sqrt{m_1}^3}{\sqrt{2\pi}^3}e^{-\frac{m_1|v|^2}{2}},\qquad \mu_2(v) = n_{20}\frac{\sqrt{m_2}^3}{\sqrt{2\pi}^3}e^{-\frac{m_2|v|^2}{2}}.
\end{align*}
\iffalse
From the following computation,
\begin{align*}
\int_{\mathbb{R}^3}\mu_1 dv = n_{10} ,\qquad \int_{\mathbb{R}^3}m_1|v|^2\mu_1 dv = 3n_{10}, \cr
\int_{\mathbb{R}^3}\mu_2 dv = n_{20} ,\qquad \int_{\mathbb{R}^3}m_2|v|^2\mu_2 dv = 3n_{20},
\end{align*}
we note that the two equilibrium distributions have the same mean velocity $U_{10}=U_{20}=0$, and the same temperature $T_{10}=T_{20}=1$.
\fi
We then define the perturbations $f_k$ $(k=1,2)$ by $F_k=\mu_k+\sqrt{\mu_k}f_k$ and rewrite the mixture BGK model \eqref{CCBGK} in terms of $f_k$ as
\begin{align}\label{pertf12}
\begin{split}
\partial_t f_1+v\cdot \nabla_xf_1&=L_{11}(f_1)+L_{12}(f_1,f_2)+\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2), \cr
\partial_t f_2+v\cdot \nabla_xf_2&=L_{22}(f_2)+L_{21}(f_1,f_2)+\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2).
\end{split}
\end{align}
On the R.H.S, $L_{11}$ and $L_{22}$ denote the linearized part of the intra-species relaxation operators:
\[L_{kk}(f_k)=n_{k0}(P_kf_k-f_k), \quad (k=1,2),
\]
where $P_k$ is the $L^2$ projection onto the linear space spanned by
\[\left\{ \sqrt{\mu_k}, v\sqrt{\mu_k}, |v|^2\sqrt{\mu_k} \right\}.\]
The linearized operators for inter-species interactions $L_{12}$ and $L_{21}$ are given by
\begin{align*}
L_{12}(f_1,f_2) &=n_{20}(P_1f_1-f_1)\\
&\quad + n_{20} \bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i} \cr
&\quad+(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15} \bigg],
\end{align*}
\begin{align*}
L_{21}(f_1,f_2) &=n_{10}(P_2f_2-f_2)\\
&\quad+ n_{10} \bigg[\frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i} \cr
&\quad+ (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25} \bigg],
\end{align*}
for $0 \leq \delta,\omega<1$ and $\{e_{ki}\}_{1\leq i \leq 5}$ is an orthonormal basis spanned by $\left\{ \sqrt{\mu_k}, v\sqrt{\mu_k}, |v|^2\sqrt{\mu_k} \right\}$ for $k=1,2$.
Finally, $\Gamma_{11}$, $\Gamma_{22}$, $\Gamma_{12}$, and $\Gamma_{21}$ are nonlinear perturbations.
For detailed derivation of \eqref{pertf12}, see Sec. 2.
We introduce
\[
L(f_1,f_2)=(L_{11}(f_1)+L_{12}(f_1,f_2),L_{22}(f_2)+L_{21}(f_1,f_2)),
\]
and
\[
\Gamma(f_1,f_2)=(\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2),\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2)),
\]
to rewrite \eqref{pertf12} in the following succinct form:
\begin{align*}
(\partial_t+v\cdot\nabla_x)(f_1,f_2)=L(f_1,f_2)+\Gamma(f_1,f_2).
\end{align*}
To state our main result, we need to set up several notations.
\begin{itemize}
\item The constant $C$ in the estimates will be defined generically.
\iffalse
\item Throughout this paper,
\begin{align*}
\begin{split}
&E_1= \frac{1}{\sqrt{n_{10}}}(\sqrt{\mu_1},0), \quad E_2= \frac{1}{\sqrt{n_{20}}}(0,\sqrt{\mu_2}),\cr
&E_i= \frac{1}{\sqrt{m_1n_{10}+m_2n_{20}}}\left(m_1v_{i-2}\sqrt{\mu_1},m_2v_{i-2}\sqrt{\mu_2} \right),\qquad (i=3,4,5)\cr
&E_6 =\frac{1}{\sqrt{6n_{10}+6n_{10}}}\left((m_1|v|^2-3)\sqrt{\mu_1},(m_2|v|^2-3)\sqrt{\mu_2}\right),
\end{split}
\end{align*}
\fi
\item $\langle \cdot,\cdot\rangle_{L^2_{v}}$ and $\langle\cdot,\cdot\rangle_{L^2_{x,v}}$ denote the standard $L^2$ inner product on $\mathbb{R}^3_v$ and $\mathbb{T}^3_x \times \mathbb{R}^3_v$, respectively.
\begin{align*}
\langle f,g\rangle_{L^2_{v}}=\int_{\mathbb{R}^3}f(v)g(v)dv, \quad\langle f,g\rangle_{L^2_{x,v}}=\int_{\mathbb{T}^3\times\mathbb{R}^3}f(x,v)g(x,v)dvdx.
\end{align*}
\item $\|\cdot\|_{L^2_v}$ and $\|\cdot\|_{L^2_{x,v}}$ denote the standard $L^2$ norms in $\mathbb{R}^3_v$ and $\mathbb{T}^3_x \times \mathbb{R}^3_v$, respectively:
\begin{align*}
\|f\|_{L^2_v}\equiv \left(\int_{\mathbb{R}^3}|f(v)|^2 dv\right)^{\frac{1}{2}}, \quad\|f\|_{L^2_{x,v}}\equiv \left(\int_{\mathbb{T}^3\times\mathbb{R}^3}|f(x,v)|^2 dvdx\right)^{\frac{1}{2}}.
\end{align*}
\item We define an $L^2$ inner product between two vectors $(f_1,f_2)$ and $(g_1,g_2)$ as
\begin{align*}
&\langle(f_1,f_2),(g_1,g_2)\rangle_{L^2_{v}}=\int_{\mathbb{R}^3}f_1(v)g_1(v)+f_2(v)g_2(v)dv, \cr
&\langle(f_1,f_2),(g_1,g_2)\rangle_{L^2_{x,v}}=\int_{\mathbb{T}^3\times\mathbb{R}^3}f_1(x,v)g_1(x,v)+f_2(x,v)g_2(x,v)dvdx.
\end{align*}
\item The standard $L^2$ norm of a vector denotes
\begin{align*}
&\| \left(f(x,v),g(x,v)\right) \|_{L^2_v} = \left( \int_{\mathbb{R}^3}|f(v)|^2 +|g(v)|^2 dv\right)^{\frac{1}{2}}, \cr
&\| \left(f(x,v),g(x,v)\right) \|_{L^2_{x,v}} = \left( \int_{\mathbb{T}^3\times\mathbb{R}^3}|f(x,v)|^2 +|g(x,v)|^2 dvdx\right)^{\frac{1}{2}}.
\end{align*}
\item We use the following notations for multi-indices differential operators:
\begin{align*}
\alpha=[\alpha_0,\alpha_1,\alpha_2,\alpha_3], \quad \beta=[\beta_1,\beta_2,\beta_3],
\end{align*}
and
\begin{align*}
\partial^{\alpha}_{\beta}=\partial_t^{\alpha_0}\partial_{x_1}^{\alpha_1}\partial_{x_2}^{\alpha_2}\partial_{x_3}^{\alpha_3}\partial_{v_1}^{\beta_1}\partial_{v_2}^{\beta_2}\partial_{v_3}^{\beta_3}.
\end{align*}
\item We employ the following convention for simplicity.
\begin{align*}
\partial^{\alpha}_{\beta}(f_1,f_2)=\big(\partial^{\alpha}_{\beta}f_1,\partial^{\alpha}_{\beta}f_2\big).
\end{align*}
\item We define the high-order energy norm $\mathcal{E}_{N_1,N_2}(f_1(t), f_2(t))$:
\begin{align*}
\mathcal{E}_{N_1,N_2}(f_1(t), f_2(t))=\sum_{\substack{|\alpha|\leq N_1,~|\beta|\leq N_2 \cr N_1+N_2=N}} \|\partial^{\alpha}_{\beta}\big(f_1(t),f_2(t)\big)\|^2_{L^2_{x,v}}.
\end{align*}
For notational simplicity, we use $\mathcal{E}(t)$ to denote $\mathcal{E}_{N_1,N_2}(f_1(t), f_2(t))$ when the
dependency on $(N_1,N_2)$ is not relevant.
\end{itemize}
We are now ready to state our main result.
\begin{theorem} Let $N\geq 3$. We set the macroscopic quantities of the initial data to the same with that of the global equilibria:
\begin{align*}
\int_{\mathbb{T}^3 \times \mathbb{R}^3}F_{k0}(x,v) \left(\begin{array}{c} 1 \cr m_kv \cr m_k|v|^2 \end{array}\right) dvdx = \int_{\mathbb{T}^3 \times \mathbb{R}^3}\mu_k(v) \left(\begin{array}{c} 1 \cr m_kv \cr m_k|v|^2 \end{array}\right) dvdx,
\end{align*}
for $k=1,2$.
We define $f_{k0}$ as $F_{k0}=\mu_k+ \sqrt{\mu_k}f_{k0}$.
Then there exists $\epsilon>0$ such that if $\mathcal{E}_{N_1,N_2}(f_{10},f_{20}) < \epsilon$, then there exists a unique global-in-time classical solution of \eqref{CCBGK} satisfying
\begin{itemize}
\item The two distribution functions are non-negative:
\[F_k(x,v,t)=\mu_k + \sqrt{\mu_k} f_k \geq 0.\]
\item The conservation laws hold \eqref{conservation}.
\item The distribution functions converge exponentially to the global equilibrium:
\[ \mathcal{E}_{N_1,N_2}(f_1,f_2)(t) \leq Ce^{-\eta t} \mathcal{E}_{N_1,N_2}(f_{10},f_{20}). \]
In the case of $N_2=0$, that is, if $\mathcal{E}_{N_1,0}(f_{10},f_{20}) < \epsilon$, we have the following more detailed convergence estimate:
\[ \mathcal{E}_{N_1,0}(f_1,f_2)(t) \leq Ce^{-\eta \min\left\{(1-\delta),(1-\omega) \right\}t} \mathcal{E}_{N_1,0}(f_{10},f_{20}). \]
\item Let $(f_1,f_2)$ and $(\bar{f}_1,\bar{f}_2)$ be solutions corresponding to the initial data $(f_{10},f_{20})$ and $(\bar{f}_{10},\bar{f}_{20})$, respectively, then the system satisfies the following $L^2$ stability:
\[ \| (f_1 -\bar{f}_1 ,f_2 -\bar{f}_2) \|_{L^2_{x,v}} \leq C \|( f_{10} -\bar{f}_{10}, f_{20} -\bar{f}_{20}) \|_{L^2_{x,v}}. \]
\end{itemize}
\end{theorem}
\begin{remark} The convergence rate in the case of $N_2=0$ shows that the higher interchange rate ($\delta$ and $\omega$ close to $0$) gives the faster convergence rates.
\end{remark}
\iffalse
Note that
\begin{align*}
\int_{\mathbb{R}^3}m_1|v|^2\mathcal{M}_{11} = 3n_1T_1 + m_1n_1 |U_1|^2
\end{align*}
\begin{align*}
\int m_1|v|^2Q_{12} dv + \int m_2 |v|^2Q_{21} dv = 3(T_{12}-T_1) + 3(T_{21}-T_2) + m_1(|U_{12}|^2-|U_1|^2) + m_2(|U_{21}|^2-U_2^2)
\end{align*}
\begin{align*}
m_1(|U_{12}|^2-|U_1|^2) + m_2(|U_{21}|^2-U_2^2) = m_1 (1-\delta)\left(\frac{m_1}{m_2}(1-\delta)-(1+\delta)\right)|U_2-U_1|^2
\end{align*}
\begin{align*}
T_{21}&=(1-\omega) T_1 + \omega T_2 +\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right) |U_2-U_1|^2
\end{align*}
and $\delta,\omega \in [0,1]$ and $0 \leq \gamma \leq 2\delta(1-\delta)$
\fi
The most important step is the identification of the dissipation mechanism of the linearized multi-component relaxation operator. To investigate the dissipative property of $L$, we decompose the linearized inter-species relaxation operator $L_{ij} $ $(i\neq j)$ further into the mass interaction part $L^1_{ij}$ and the momentum-energy interaction part $L^2_{ij}$:
\begin{align*}
L_{12}^1(f_1)= n_{20}(P_1f_1-f_1), \quad L_{21}^1(f_2)= n_{10}(P_2f_2-f_2),
\end{align*}
and
\begin{align*}
\begin{split}
L_{12}^2(f_1,f_2) &= n_{20} \bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i} \cr
&\quad+(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15} \bigg], \cr
L_{21}^2(f_1,f_2) &= n_{10} \bigg[\frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i} \cr
&\quad+ (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25} \bigg],
\end{split}
\end{align*}
so that $L_{12}=L_{12}^1+L_{12}^2$ and $L_{21}=L_{21}^1+L_{21}^2$. We first derive from an explicit computation that the intra-species operator $L_{ii}$ and the mass interaction part of the inter-species operator $L^1_{12}$ and $L^1_{21}$ give rise to the following partial dissipative estimate:
\begin{align}\label{LtoP0}
\begin{split}
&\langle (L_{11}+ L_{12}^1)f_1 , f_1\rangle_{L^2_{x,v}}+\langle (L_{22}+L_{21})f_2 , f_2\rangle_{L^2_{x,v}} \cr
&\hspace{3cm}= -(n_{10}+n_{20})\| (I-P_1,I-P_2)(f_1,f_2) \|_{L^2_{x,v}}^2.
\end{split}
\end{align}
We note that the dissipation estimate above is too weak in that it involves 10-dimensional degeneracy,
which is 4-dimensional bigger than the 6-dimensional conservation laws in \eqref{conservation}.
It is the additional dissipation from the momentum-energy interaction parts $L_{12}^2$, $L_{21}^2$ of the inter-species operators $L_{12}$ and $L_{21}$ that make up for the deficiency:
\begin{align}\label{bring}
\begin{split}
\langle L_{12}^2,f_1 \rangle_{L^2_{x,v}}+\langle L_{21}^2,f_2 \rangle_{L^2_{x,v}} &\leq -\min\left\{(1-\delta),(1-\omega) \right\} (n_{10}+n_{20}) \cr
&\quad \times\left( \|(P_1,P_2)(f_1,f_2)\|_{L^2_{x,v}}^2-\|P(f_1,f_2) \|_{L^2_{x,v}}^2\right),
\end{split}
\end{align}
where $P$ is an orthonormal $L^2\times L^2$ projection on the space spanned by the following $6$-dimensional basis
\begin{align*}
\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}), (m_1v\sqrt{\mu_1},m_2v\sqrt{\mu_2}),\left((m_1|v|^2-3)\sqrt{\mu_1},(m_2|v|^2-3)\sqrt{\mu_2}\right)\}.
\end{align*}
Then partial dissipation estimates \eqref{LtoP0} and \eqref{bring} complement each other to give rise to the following multi-component dissipation estimate for $L$:
\begin{align}\label{Lm}
\begin{split}
\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}&\leq - (n_{10}+n_{20})\Big( \max\{\delta,\omega\}
\| (I-P_1,I-P_2)(f_1,f_2) \|_{L^2_{x,v}}^2 \cr
&\quad + \min\left\{(1-\delta),(1-\omega) \right\}\| (I-P)(f_1,f_2)\|_{L^2_{x,v}}^2 \Big).
\end{split}
\end{align}
The dissipation estimate \eqref{Lm}, together with further analysis on the degeneracy part through the standard micro-macro decomposition, provides the following full coercivity depending on the interchange rates:
\begin{align*}
\langle L(\partial^{\alpha}(f_1,f_2)),\partial^{\alpha}(f_1,f_2)\rangle_{L^2_{x,v}} \leq - \eta \min\left\{(1-\delta),(1-\omega) \right\}\sum_{|\alpha|\leq N} \| (\partial^{\alpha}(f_1,f_2) \|_{L^2_{x,v}}^2.
\end{align*}
Due to the presence of the momentum interchange rate $\delta$ and the energy interchange rate $\omega$ between different components in the dissipation estimate, we see that the larger interchange rate (when $\delta$ and $\omega$ are close to zero) leads to the stronger dissipation, and therefore, the faster convergence to the global equilibrium:
\begin{align*}
\sum_{|\alpha|\leq N}\|\partial^{\alpha} (f_1(t),f_2(t) )\|_{L^2_{x,v}}^2 \leq e^{-\eta \min\left\{(1-\delta),(1-\omega) \right\}t}\sum_{|\alpha|\leq N}\|\partial^{\alpha} (f_1(0), f_2(0))\|_{L^2_{x,v}}^2.
\end{align*}
\subsection{Literature review}
We start with a review of the mathematical results of the mono-species BGK model.
Perthame established the first result on global weak solutions for a general initial data in \cite{Perth}.
In \cite{PP1993}, the authors considered weighted-$L^{\infty}$ bounds to obtain the uniqueness.
Desvillettes considered the convergence to equilibrium in a weak sense \cite{Des}.
Ukai proved the existence of the stationary solution on a finite interval with inflow boundary condition in \cite{Ukai}. In \cite{ZH}, the $L^{\infty}$ work in \cite{Perth} is generalized to an weighted $L^p$ space.
Classical solutions near-global equilibrium is constructed in \cite{Bello} using the spectral analysis of Ukai \cite{Ukai spectral}, and by using the nonlinear energy method of Yan Guo \cite{Guo whole, Guo VMB, Guo VPB} in \cite{Yun1}.
The nonlinear energy method is then employed further to study several types of BGK models \cite{Yun1,Yun2,Yun3,HY,BY,Shakhov}.
Saint-Raymond considered the hydrodynamic limits of the BGK model in \cite{Saint,Saint2}.
For the numerical study of the BGK model, we refer to \cite{BCRY,MR3828279,Bennoune_2008,Crestetto_2012,Pirner4,CBRY1,CBRY2,BCRY,RSY,RY}.
Various BGK models to describe the dynamics of multi-component gases are proposed in the literature. Examples include the model of Gross and Krook \cite{gross_krook1956}, the model of Hamel \cite{hamel1965}, the model of Greene \cite{Greene}, the model of Garzo, Santos and Brey \cite{Garzo1989}, the model of Sofonea and Sekerka \cite{Sofonea2001}, the model by Andries, Aoki and Perthame \cite{AndriesAokiPerthame2002}, the model of Brull, Pavan and Schneider \cite{Brull_2012}, the model of Klingenberg, Pirner and Puppo \cite{mixmodel}, the model of Haack, Hauck, Murillo \cite{haack}, the model of Bobylev, Bisi, Groppi, Spiga \cite{Bobylev}.
The BGK model for gas mixtures has also been extended to the ES-BGK model, polyatomic molecules, chemical reactions, or the quantum case; See for example \cite{MR3828279, Groppi, MR3960644, Pirner6, Bisi, Bisi2, Quantum,Yun,Stru}.
For the applications of the mixture BGK models,
we refer to \cite{Puppo_2007, Jin_2010,Dimarco_2014, Bennoune_2008, Dimarco, Bernard_2015,BCRY,RSY}.
For the existence of the BGK model of gas mixtures, the mild solution was established in \cite{MR3720827}. In \cite{LiuPirner}, by constructing an entropy functional, the authors can prove exponential relaxation to equilibrium with explicit rates. The strategy is based on the entropy and spectral methods adapting Lyapunov’s direct method.
A review of the multi-species Boltzmann equation is in order.
In \cite{Guo VMB}, the author established the global existence for the mixture of a charged particle described by the Vlasov-Maxwell-Boltzmann equation.
The mild solution and uniform $L^1$ stability are obtained in \cite{HNYun}.
A mass diffusion problem of the mixture and the cross-species resonance is studied for a one-dimension case in \cite{SotiYu} based on the work in \cite{LiuYu}.
In \cite{Briant}, the author constructed the global-in-time mild solution near-global equilibrium for the mixture Boltzmann equation.
The Vlasov–Poisson–Boltzmann equation was considered in \cite{DuanLiu} about large time asymptotic profiles when the different-species gases tend to two distinct global Maxwellians.
In \cite{GambaP}, the existence and uniqueness are constructed in spatially homogeneous settings when an initial data has upper and lower bounds for some polynomial moments.
The authors in \cite{BGPS} obtained some energy estimates.
For physical or engineering references on the studies on multi-component gases at the kinetic level,
we refer \cite{ABT,YoAo,Valo,TakaAo,SotiYu,AS,MMM,TaAo,TaAoMu,Sofonea2001}. Some general reviews of the Boltzmann and the BGK model can be found in \cite{B.G.K,Cercignani,DL,Chap,Cerci2,CIP,GL,V}.
This paper is organized as follows:
In Sec. 2, we linearized the system \eqref{CCBGK} to obtain \eqref{pertf12}.
In Sec. 3, we derive the dissipation estimate of the linearized relaxation operator. The local-in-time classical solution is constructed in Sec 4.
In Section 5, The full coercivity of $L$ is recovered when the energy norm is sufficiently small.
Lastly, we establish the global-in-time classical solution in Sec 6.
\section{Linearization of the mixture BGK model}
\subsection{Linearization of the mixture Maxwellian}
In this part, we linearize the inter-species Maxwellian $\mathcal{M}_{12}$ and $\mathcal{M}_{21}$. We first define the macroscopic projection on $L^2_v$ and state the linearization result of the mono-species local Maxwellian $\mathcal{M}_{kk}$.
\begin{definition} We define the macroscopic projection operator $P_k$ in $L^2_v$ for $k=1,2$:
\begin{align*}
P_kf &= \frac{1}{n_{k0}} \int_{\mathbb{R}^3} f\sqrt{\mu_k} dv\sqrt{\mu_k} + \frac{m_k}{n_{k0}} \int_{\mathbb{R}^3} fv\sqrt{\mu_k} dv\cdot v\sqrt{\mu_k} \cr
&\quad + \frac{1}{6n_{k0}}\int_{\mathbb{R}^3} f(m_k|v|^2-3)\sqrt{\mu_k} dv (m_k|v|^2-3)\sqrt{\mu_k}.
\end{align*}
We denote $5$-dimensional basis as $(i=2,3,4)$
\begin{align}\label{basis}
\begin{split}
e_{k1}= \frac{1}{\sqrt{n_{k0}}}\sqrt{\mu_k},\qquad e_{ki}= \sqrt{\frac{m_k}{n_{k0}}}v_{i-1}\sqrt{\mu_k}, \qquad e_{k5} =\frac{m_k|v|^2-3}{ \sqrt{6n_{k0}} }\sqrt{\mu_k}.
\end{split}
\end{align}
\end{definition}
The $5$-dimensional basis set $\{e_{1i}\}_{i=1,\cdots,5}$ and $\{e_{2i}\}_{i=1,\cdots,5}$ construct an orthonormal basis in $L^2_v$, respectively. So, we can write
\begin{align*}
P_1f = \sum_{1\leq i \leq 5}\langle f, e_{1i} \rangle_{L^2_v}e_{1i}, \quad \textit{and} \quad
P_2f = \sum_{1\leq i \leq 5}\langle f, e_{2i} \rangle_{L^2_v}e_{2i}.
\end{align*}
\begin{lemma}\emph{\cite{Yun1}}\label{lin ii} The mono-species BGK Maxwellian $\mathcal{M}_{kk}$ is linearized as follows:
\begin{align*}
\mathcal{M}_{kk}(F_k) = \mu_k + \sqrt{\mu_k}P_kf_k + \sqrt{\mu_k}~\Gamma_{kk}(f_k,f_k),
\end{align*}
where the nonlinear term $\Gamma_{kk}(f_k,f_k)$ is given by
\begin{align*}
\Gamma_{kk}(f_k,f_k) &= \sum_{1 \leq i,j \leq 5} \frac{1}{\sqrt{\mu_k}}\int_0^1 \frac{P_{ij}(n_{k\theta},U_{k\theta},T_{k\theta},v-U_{k\theta},U_{k\theta})}{R_{ij}(n_{k\theta},T_{k\theta})}\mathcal{M}_{kk}(\theta)(1-\theta) d\theta \cr
&\quad \times \langle f_k,e_{ki} \rangle_{L^2_v}\langle f_k, e_{kj} \rangle_{L^2_v},
\end{align*}
for k=1,2. The function $P_{ij}(x_1,\cdots,x_5)$ denotes a generic polynomial depending on $(x_1,\cdots,x_5)$ and $R_{ij}(x,y)$ denotes a generic monomial $R_{ij}(x,y)=x^ny^m$, where $n,m\in \mathbb{N}\cup\{0\}$.
\end{lemma}
\begin{proof}
The linearization of the mono-species BGK Maxwellian $\mathcal{M}_{kk}$ is in \cite{Yun1} for the case $n_{k0}=1$ and $m_k=1$. For a general $n_{k0}$ and $m_k$, the linearization of $\mathcal{M}_{kk}$ is a special case of the linearization of $\mathcal{M}_{12}$ and $\mathcal{M}_{21}$
with the choice $\delta=\omega=1$ (See \eqref{lin M12} and \eqref{lin M21}, respectively).
\end{proof}
\begin{proposition}\label{linearize} The multi-species BGK Maxwellians $\mathcal{M}_{12}$ and $\mathcal{M}_{21}$ are linearized as follows:
\begin{align*}
\mathcal{M}_{12}(F)&= \mu_1 + P_1f_1\sqrt{\mu_1}
+ (1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i}\sqrt{\mu_1} \cr
&\quad + (1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15}\sqrt{\mu_1} +\sqrt{\mu_1}\Gamma_{12}(f_1,f_2),
\end{align*}
and
\begin{align*}
\mathcal{M}_{21}(F)&= \mu_2+ P_2f_2\sqrt{\mu_2} + \frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i}\sqrt{\mu_2} \cr
&\quad + (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25}\sqrt{\mu_2} +\sqrt{\mu_2}\Gamma_{21}(f_1,f_2)
\end{align*}
We give the precise definition of the nonlinear terms $\Gamma_{12}$ and $\Gamma_{21}$ in Section \ref{linMBGK}
\end{proposition}
\begin{proof}
We first define a transition of the macroscopic fields:
\begin{align}\label{transition}
n_{k\theta} = \theta n_k +(1-\theta)n_{k0}, \quad n_{k\theta}U_{k\theta} = \theta n_k U_k , \quad G_{k\theta} = \theta G_k,
\end{align}
where
\begin{align*}
G_k = \frac{3n_k T_k + m_k n_k |U_k|^2-3n_k}{\sqrt{6}},
\end{align*}
for $k=1,2$. We also denote multi-species macroscopic fields as
\begin{align}\label{12theta}
\begin{split}
U_{12\theta}&=\delta U_{1\theta} + (1-\delta)U_{2\theta}, \cr
U_{21\theta}&=\frac{m_1}{m_2}(1-\delta)U_{1\theta} + \left(1-\frac{m_1}{m_2}(1-\delta)\right)U_{2\theta}, \cr
T_{12\theta}&= \omega T_{1\theta} + (1-\omega)T_{2\theta} + \gamma |U_{2\theta}-U_{1\theta}|^2, \cr
T_{21\theta}&= (1-\omega) T_{1\theta} + \omega T_{2\theta} +\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right) |U_{2\theta}-U_{1\theta}|^2 .
\end{split}
\end{align}
Then we consider the multi-species BGK Maxwellians $\mathcal{M}_{12}$ and $\mathcal{M}_{21}$, which depend on $\theta$:
\begin{align*}
\mathcal{M}_{12}(\theta) = \frac{n_{1\theta}}{\sqrt{2\pi\frac{T_{12\theta}}{m_1}}^3} \exp\left(-\frac{|v-U_{12\theta}|^2}{2\frac{T_{12\theta}}{m_1}}\right), \quad \mathcal{M}_{21}(\theta) = \frac{n_{2\theta}}{\sqrt{2\pi\frac{T_{21\theta}}{m_2}}^3} \exp\left(-\frac{|v-U_{21\theta}|^2}{2\frac{T_{21\theta}}{m_2}}\right).
\end{align*}
The definition of $n_{k\theta},U_{k\theta},T_{k\theta}$ gives
\begin{align*}
\left(n_{k\theta},U_{k\theta},T_{k\theta} \right)|_{\theta=1} = (n_k,U_k,T_k), \quad \textit{and} \quad \left(n_{k\theta},U_{k\theta},T_{k\theta} \right)|_{\theta=0} = (n_{k0},0,1),
\end{align*}
so we have
\begin{align*}
\mathcal{M}_{12}(1) = \mathcal{M}_{12}, \quad \mathcal{M}_{12}(0) = \mu_1,
\end{align*}
and
\begin{align*}
\mathcal{M}_{21}(1) = \mathcal{M}_{21}, \quad \mathcal{M}_{21}(0) = \mu_2,
\end{align*}
where we used $U_{120} = U_{210}=0$ and $T_{120} = T_{210}=1$.
We apply the Taylor expansion to $\mathcal{M}_{12}(\theta)$ and $\mathcal{M}_{21}(\theta)$:
\begin{align*}
\mathcal{M}_{12}(1)=\mu_1+\mathcal{M}_{12}'(0)+\int_0^1\mathcal{M}_{12}''(\theta)(1-\theta)d\theta,
\end{align*}
and
\begin{align*}
\mathcal{M}_{21}(1)=\mu_2+\mathcal{M}_{21}'(0)+\int_0^1\mathcal{M}_{21}''(\theta)(1-\theta)d\theta.
\end{align*}
By the chain rule, we compute the linear term $\mathcal{M}_{ij}'(0)$:
\begin{multline}\label{M'(0)}
\mathcal{M}_{ij}'(0)=\left(\frac{d (n_{1\theta}, n_{1\theta}U_{1\theta}, G_{1\theta},n_{2\theta}, n_{2\theta}U_{2\theta}, G_{2\theta})}{d \theta}\right)^{T} \cr
\times \left( \frac{\partial(n_{1\theta}, n_{1\theta}U_{1\theta}, G_{1\theta},n_{2\theta}, n_{2\theta}U_{2\theta}, G_{2\theta})} {\partial(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})} \right)^{-1} \left(\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{ij}(\theta)\right)\bigg|_{\theta=0},
\end{multline}
for $(i,j)=(1,2)$ or $(2,1)$.
Although $\mathcal{M}_{12}$ does not depend on $n_2$, we use the above form for the convenience of the calculation. In this proposition, we focus on the linear term $\mathcal{M}_{12}'(0)$ and $\mathcal{M}_{21}'(0)$.
The exact form of the nonlinear terms will be presented in Section \ref{linMBGK}.
The remaining proof proceeds by stating some auxiliary lemmas below.
\end{proof}
\begin{lemma}\emph{\cite{Yun1}}\label{Jaco} Let us define
\[G = \frac{3n T + m n |U|^2-3n}{\sqrt{6}}. \]
Then we have
\begin{align*}
J = \frac{\partial(n,n U,G)} {\partial(n,U,T)} = \left[ {\begin{array}{cccccc}
1 & 0 & 0 & 0 & 0 \\
U_1 & n & 0 & 0 & 0 \\
U_2 & 0 & n & 0 & 0 \\
U_3 & 0 & 0 & n & 0 \\
\frac{3T+m|U|^2-3}{\sqrt{6}} & \frac{2n U_1m}{\sqrt{6}} & \frac{2n U_2m}{\sqrt{6}} & \frac{2n U_3m}{\sqrt{6}} & \frac{3n }{\sqrt{6}}
\end{array} } \right],
\end{align*}
and
\begin{align*}
J^{-1} = \left(\frac{\partial(n,n U,G)} {\partial(n,U,T)}\right)^{-1} = \left[ {\begin{array}{cccccc}
1 & 0 & 0 & 0 & 0 \\
-\frac{U_1}{n} & \frac{1}{n} & 0 & 0 & 0 \\
-\frac{U_2}{n} & 0 & \frac{1}{n} & 0 & 0 \\
-\frac{U_3}{n} & 0 & 0 & \frac{1}{n} & 0 \\
\frac{m|U|^2-3T+3}{3n} & -\frac{2m}{3}\frac{U_1}{n} & -\frac{2m}{3}\frac{U_2}{n} & -\frac{2m}{3}\frac{U_3}{n} & \sqrt{\frac{2}{3}}\frac{1}{n}
\end{array} } \right].
\end{align*}
\end{lemma}
\begin{proof}
In the case of $m_i=1$, it is proved in \cite{Yun1}, and by the same explicit calculation, we can extend the result for general $m_i$. We omit it.
\end{proof}
\subsubsection{Linearization of $\mathcal{M}_{12}$} We first consider the calculation of $\mathcal{M}'_{12}(0)$ in \eqref{M'(0)}.
\begin{lemma}\label{M_12 diff} We have
\begin{align*}
&(1) ~ \frac{\partial \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}}\bigg|_{\theta=0} =\frac{1}{n_{10}}\mu_1,
&(2) ~ \frac{\partial \mathcal{M}_{12}(\theta)}{\partial U_{1\theta}}\bigg|_{\theta=0} &= \delta m_1 v \mu_1, \cr
&(3) ~ \frac{\partial \mathcal{M}_{12}(\theta)}{\partial T_{1\theta}}\bigg|_{\theta=0} = \omega \frac{m_1|v|^2-3}{2}\mu_1,
&(4) ~\frac{\partial \mathcal{M}_{12}(\theta)}{\partial U_{2\theta}}\bigg|_{\theta=0} &= (1-\delta)m_1 v\mu_1, \cr
&(5) ~ \frac{\partial \mathcal{M}_{12}(\theta)}{\partial T_{2\theta}}\bigg|_{\theta=0} = (1-\omega)\frac{m_1|v|^2-3}{2}\mu_1. &
\end{align*}
\end{lemma}
\begin{proof}
For readability, we ignore the dependence on $\theta$
\newline
(1) By an explicit computation, we have
\begin{align*}
\frac{\partial \mathcal{M}_{12}}{\partial n_1} = \frac{1}{n_1}\mathcal{M}_{12}.
\end{align*}
(2) Note that both $U_{12}$ and $T_{12}$ depend on $U_1$. So that, the chain rule gives
\begin{align*}
\frac{\partial \mathcal{M}_{12}}{\partial U_1} &= \frac{\partial U_{12}}{\partial U_1}\frac{\partial \mathcal{M}_{12}}{\partial U_{12}} + \frac{\partial T_{12}}{\partial U_1}\frac{\partial \mathcal{M}_{12}}{\partial T_{12}} \cr
&= \delta m_1\frac{v-U_{12}}{T_{12}}\mathcal{M}_{12} -2\gamma(U_2-U_1)\left(-\frac{3}{2}\frac{1}{T_{12}}+\frac{m_1|v-U_{12}|^2}{2T_{12}^2}\right)\mathcal{M}_{12}.
\end{align*}
(3) An explicit calculation gives
\begin{align*}
\frac{\partial \mathcal{M}_{12}}{\partial T_1} = \frac{\partial T_{12}}{\partial T_1}\frac{\partial \mathcal{M}_{12}}{\partial T_{12}} = \omega \left(-\frac{3}{2}\frac{1}{T_{12}}+\frac{m_1|v-U_{12}|^2}{2T_{12}^2}\right)\mathcal{M}_{12}.
\end{align*}
(4) Similar to case (2), both $U_{12}$ and $T_{12}$ depend on $U_2$.
\begin{align*}
\frac{\partial \mathcal{M}_{12}}{\partial U_2} &= \frac{\partial U_{12}}{\partial U_2}\frac{\partial \mathcal{M}_{12}}{\partial U_{12}} + \frac{\partial T_{12}}{\partial U_2}\frac{\partial \mathcal{M}_{12}}{\partial T_{12}} \cr
&= (1-\delta)m_1\frac{v-U_{12}}{T_{12}}\mathcal{M}_{12} +2\gamma(U_2-U_1)\left(-\frac{3}{2}\frac{1}{T_{12}}+\frac{m_1|v-U_{12}|^2}{2T_{12}^2}\right)\mathcal{M}_{12}.
\end{align*}
(5) By an explicit computation, we have
\begin{align*}
\frac{\partial \mathcal{M}_{12}}{\partial T_2} = \frac{\partial T_{12}}{\partial T_2}\frac{\partial \mathcal{M}_{12}}{\partial T_{12}} = (1-\omega) \left(-\frac{3}{2}\frac{1}{T_{12}}+\frac{m_1|v-U_{12}|^2}{2T_{12}^2}\right)\mathcal{M}_{12}.
\end{align*}
Substituting
\begin{align}\label{1020}
(n_{1\theta},U_{1\theta},T_{1\theta},U_{2\theta},T_{2\theta})\big|_{\theta=0}&= (n_{10},U_{10},T_{10},U_{20},T_{20}) = (n_{10},0,1,0,1),
\end{align}
and
\begin{align}\label{12210}
U_{12\theta}|_{\theta=0}=U_{21\theta}|_{\theta=0} = 0, \quad \quad
T_{12\theta}|_{\theta=0}=T_{21\theta}|_{\theta=0} = 1,
\end{align}
on the above computations, we get the desired result. \newline
\end{proof}
Now we proceed with the proof of Proposition \ref{linearize} for $\mathcal{M}_{12}(F)$. By the definition of the transition of the macroscopic fields \eqref{transition} and the definition of the basis \eqref{basis}, we have
\begin{align}\label{lin1}
\begin{split}
\frac{d (n_{k\theta}, n_{k\theta}U_{k\theta}, G_{k\theta})}{d \theta}
&= \left( \int_{\mathbb{R}^3} f_k\sqrt{\mu_k} dv, \int_{\mathbb{R}^3} f_kv\sqrt{\mu_k}dv, \int _{\mathbb{R}^3}f_k\frac{m_k|v|^2-3}{\sqrt{6}}\sqrt{\mu_k} dv \right) \cr
&= \left( \langle f_k,e_{k1} \rangle_{L^2_v} , \langle f_k,e_{k2} \rangle_{L^2_v}, \langle f_k,e_{k3} \rangle_{L^2_v}, \langle f_k,e_{k4} \rangle_{L^2_v} , \langle f_k,e_{k5} \rangle_{L^2_v} \right),
\end{split}
\end{align}
for $k=1,2$. For notational brevity, we define
\[J_{k\theta} = \frac{\partial(n_{k\theta},n_{k\theta}U_{k\theta},G_{k\theta})} {\partial(n_{k\theta},U_{k\theta},T_{k\theta})}.\]
Then applying Lemma \ref{Jaco} gives
\begin{align*}
J^{-1}_{k\theta}\big|_{\theta=0} = diag\left(1,\frac{1}{n_{k0}},\frac{1}{n_{k0}},\frac{1}{n_{k0}},\sqrt{\frac{2}{3}}\frac{1}{n_{k0}}\right),
\end{align*}
and
\begin{align}\label{lin2}
\begin{split}
\left( \frac{\partial(n_{1\theta}, n_{1\theta}U_{1\theta}, G_{1\theta},n_{2\theta}, n_{2\theta}U_{2\theta}, G_{2\theta})} {\partial(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})} \right)^{-1}\bigg|_{\theta=0} = \left[ {\begin{array}{cccccc}
J^{-1}_{1\theta}\big|_{\theta=0} & 0 \\
0 & J^{-1}_{2\theta}\big|_{\theta=0}
\end{array} } \right] ,
\end{split}
\end{align}
where we used
\begin{align}\label{block inv}
\begin{split}
\left[ {\begin{array}{cccccc}
J_1 & 0 \\
0 & J_2
\end{array} } \right]^{-1}=
\left[ {\begin{array}{cccccc}
J^{-1}_1 & 0 \\
0 & J^{-1}_2
\end{array} } \right].
\end{split}
\end{align}
We substitute \eqref{lin1}, \eqref{lin2} and Lemma \ref{M_12 diff} into \eqref{M'(0)} to obtain
\begin{align*}
\mathcal{M}_{12}'(0)&=\frac{\mu_1}{n_{10}}\int_{\mathbb{R}^3} f_1\sqrt{\mu_1} dv + \frac{\delta m_1 v \mu_1}{n_{10}}\int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv \cr
&\quad+ \omega \frac{m_1|v|^2-3}{2}\mu_1\sqrt{\frac{2}{3}}\frac{1}{n_{10}} \int _{\mathbb{R}^3}f_1\frac{m_1|v|^2-3}{\sqrt{6}}\sqrt{\mu_1} dv \cr
&\quad+ \frac{(1-\delta)m_1v\mu_1}{n_{20}}\int_{\mathbb{R}^3} f_2v\sqrt{\mu_2}dv \cr
&\quad+ (1-\omega)\frac{m_1|v|^2-3}{2}\mu_1\sqrt{\frac{2}{3}}\frac{1}{n_{20}}\int _{\mathbb{R}^3}f_2\frac{m_2|v|^2-3}{\sqrt{6}}\sqrt{\mu_2} dv.
\end{align*}
Using the definition of the basis in \eqref{basis}, we simplify it as follows:
\begin{multline}\label{lin M12}
\mathcal{M}_{12}'(0)=\langle f_1, e_{11} \rangle_{L^2_v}e_{11}\sqrt{\mu_1} + \delta \sum_{2\leq i \leq 4}\langle f_1, e_{1i} \rangle_{L^2_v}e_{1i}\sqrt{\mu_1} + \omega \langle f_1, e_{15} \rangle_{L^2_v}e_{15}\sqrt{\mu_1} \cr
+ (1-\delta)\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\sum_{2\leq i \leq 4}\langle f_2, e_{2i} \rangle_{L^2_v}e_{1i}\sqrt{\mu_1} + (1-\omega) \sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}e_{15}\sqrt{\mu_1}.
\end{multline}
Adding and subtracting the following term
\begin{align*}
(1-\delta) \sum_{2\leq i \leq 4}\langle f_1, e_{1i} \rangle_{L^2_v}e_{1i}\sqrt{\mu_1} + (1-\omega) \langle f_1, e_{15} \rangle_{L^2_v}e_{15}\sqrt{\mu_1},
\end{align*}
gives
\begin{align*}
\mathcal{M}_{12}'(0)&= P_1f_1\sqrt{\mu_1}
+ (1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i}\sqrt{\mu_1} \cr
&\quad+ (1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15}\sqrt{\mu_1}.
\end{align*}
This completes the proof for the linearization of $\mathcal{M}_{12}$.
\iffalse
\begin{align*}
\mathcal{M}_{12}'(0)&=\int_{\mathbb{R}^3} f_1\sqrt{\mu_1} dv\mu_1 + \int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv~ v \mu_1 + \int _{\mathbb{R}^3}f_1\frac{|v|^2-3}{\sqrt{6}}\sqrt{\mu_1} dv ~ \frac{|v|^2-3}{\sqrt{6}}\mu_1 \cr
&+ \int_{\mathbb{R}^3} (f_2\sqrt{\mu_2}-f_1\sqrt{\mu_1})vdv(1-\delta)v\mu_1 + \int _{\mathbb{R}^3}(f_2\sqrt{\mu_2}-f_1\sqrt{\mu_1})\frac{|v|^2-3}{\sqrt{6}} dv (1-\omega)\frac{|v|^2-3}{\sqrt{6}}\mu_1 \cr
&= P_1f_1\sqrt{m} \cr
&+ \int_{\mathbb{R}^3} (f_2\sqrt{\mu_2}-f_1\sqrt{\mu_1})vdv(1-\delta)v\mu_1 + \int _{\mathbb{R}^3}(f_2\sqrt{\mu_2}-f_1\sqrt{\mu_1})\frac{|v|^2-3}{\sqrt{6}} dv (1-\omega)\frac{|v|^2-3}{\sqrt{6}}\mu_1
\end{align*}
\fi
\subsubsection{Linearization of $\mathcal{M}_{21}$} Now we consider the calculation of $\mathcal{M}_{21}$ in \eqref{M'(0)}.
\begin{lemma}\label{M_21 diff} We have
\begin{align*}
&(1) ~ \frac{\partial \mathcal{M}_{21\theta}}{\partial n_{2\theta}}\bigg|_{\theta=0} =\frac{1}{n_{20}}\mu_2,
&(2)~ \frac{\partial \mathcal{M}_{21\theta}}{\partial U_{2\theta}}\bigg|_{\theta=0} &=\left(1-\frac{m_1}{m_2}(1-\delta)\right) m_2 v \mu_2, \cr
&(3)~ \frac{\partial \mathcal{M}_{21\theta}}{\partial T_{2\theta}}\bigg|_{\theta=0} = \omega \frac{m_2|v|^2-3}{2}\mu_2 ,
&(4)~ \frac{\partial \mathcal{M}_{21\theta}}{\partial U_{1\theta}}\bigg|_{\theta=0}&= \frac{m_1}{m_2}(1-\delta)m_2 v\mu_2, \cr
&(5) ~ \frac{\partial \mathcal{M}_{21\theta}}{\partial T_{1\theta}}\bigg|_{\theta=0} = (1-\omega)\frac{m_2|v|^2-3}{2}\mu_2. &
\end{align*}
\end{lemma}
\begin{proof}
(1) By an explicit computation, we have
\begin{align*}
\frac{\partial \mathcal{M}_{21}}{\partial n_2} = \frac{1}{n_2}\mathcal{M}_{21}.
\end{align*}
(2) Note that $U_{21}$ and $T_{21}$ depend on $U_2$. The chain rule gives
\begin{align*}
\frac{\partial \mathcal{M}_{21}}{\partial U_2} &= \frac{\partial U_{21}}{\partial U_2}\frac{\partial \mathcal{M}_{21}}{\partial U_{21}} + \frac{\partial T_{21}}{\partial U_2}\frac{\partial \mathcal{M}_{21}}{\partial T_{21}}.
\end{align*}
So we differentiate
\begin{align*}
\frac{\partial U_{21}}{\partial U_2}\frac{\partial \mathcal{M}_{21}}{\partial U_{21}}&= (1-\frac{m_1}{m_2}(1-\delta)) m_2\frac{v-U_{21}}{T_{21}}\mathcal{M}_{21},
\end{align*}
and
\begin{align*}
\frac{\partial T_{21}}{\partial U_2}\frac{\partial \mathcal{M}_{21}}{\partial T_{21}}
&=2\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right)\cr
&\quad \times (U_2-U_1)\left(-\frac{3}{2}\frac{1}{T_{21}}+\frac{m_2|v-U_{21}|^2}{2T_{21}^2}\right)\mathcal{M}_{21}.
\end{align*}
(3) We have
\begin{align*}
\frac{\partial \mathcal{M}_{21}}{\partial T_2} = \frac{\partial T_{21}}{\partial T_2}\frac{\partial \mathcal{M}_{21}}{\partial T_{21}} = \omega \left(-\frac{3}{2}\frac{1}{T_{21}}+\frac{m_2|v-U_{21}|^2}{2T_{21}^2}\right)\mathcal{M}_{21}.
\end{align*}
(4) Since both $U_{21}$ and $T_{21}$ depend on $U_1$,
\begin{align*}
\frac{\partial \mathcal{M}_{21}}{\partial U_1} &= \frac{\partial U_{21}}{\partial U_1}\frac{\partial \mathcal{M}_{21}}{\partial U_{21}} + \frac{\partial T_{21}}{\partial U_1}\frac{\partial \mathcal{M}_{21}}{\partial T_{21}},
\end{align*}
we compute
\begin{align*}
\frac{\partial U_{21}}{\partial U_1}\frac{\partial \mathcal{M}_{21}}{\partial U_{21}} &= \frac{m_1}{m_2}(1-\delta)m_2\frac{v-U_{21}}{T_{21}}\mathcal{M}_{21},
\end{align*}
and
\begin{align*}
\frac{\partial T_{21}}{\partial U_1}\frac{\partial \mathcal{M}_{21}}{\partial T_{21}} &= -2\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right)\cr
&\quad \times (U_2-U_1)\left(-\frac{3}{2}\frac{1}{T_{21}}+\frac{m_2|v-U_{21}|^2}{2T_{21}^2}\right)\mathcal{M}_{21}.
\end{align*}
(5) We have
\begin{align*}
\frac{\partial \mathcal{M}_{21}}{\partial T_1} = \frac{\partial T_{21}}{\partial T_1}\frac{\partial \mathcal{M}_{21}}{\partial T_{21}} = (1-\omega) \left(-\frac{3}{2}\frac{1}{T_{21}}+\frac{m_2|v-U_{21}|^2}{2T_{21}^2}\right)\mathcal{M}_{21}.
\end{align*}
Similar to Lemma \ref{M_12 diff}, substituting \eqref{1020} and \eqref{12210} on the above calculations gives desired results.
\end{proof}
Substituting \eqref{lin1}, \eqref{lin2}, and Lemma \ref{M_21 diff} into \eqref{M'(0)} yields
\begin{align*}
\mathcal{M}_{21}'(0)&=\frac{\mu_2}{n_{20}}\int_{\mathbb{R}^3} f_2\sqrt{\mu_2} dv + \frac{\left(1-\frac{m_1}{m_2}(1-\delta)\right) m_2 v \mu_2}{n_{20}}\int_{\mathbb{R}^3} f_2v\sqrt{\mu_2}dv \cr
&\quad+ \omega \frac{m_2|v|^2-3}{2}\mu_2 \sqrt{\frac{2}{3}}\frac{1}{n_{20}} \int _{\mathbb{R}^3}f_2\frac{m_2|v|^2-3}{\sqrt{6}}\sqrt{\mu_2} dv \cr
&\quad+ \frac{\frac{m_1}{m_2}(1-\delta)m_2 v\mu_2}{n_{10}}\int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv \cr
&\quad+ (1-\omega)\frac{m_2|v|^2-3}{2}\mu_2\sqrt{\frac{2}{3}}\frac{1}{n_{10}}\int _{\mathbb{R}^3}f_1\frac{m_1|v|^2-3}{\sqrt{6}}\sqrt{\mu_1} dv .
\end{align*}
Using the notation of the basis in \eqref{basis}, it is equal to
\begin{align}\label{lin M21}
\begin{split}
\mathcal{M}_{21}'(0)&=\langle f_2, e_{21} \rangle_{L^2_v}e_{21}\sqrt{\mu_2} \cr
&\quad+ \left(1-\frac{m_1}{m_2}(1-\delta)\right) \sum_{2\leq i \leq 4}\langle f_2, e_{2i} \rangle_{L^2_v}e_{2i}\sqrt{\mu_2} + \omega \langle f_2, e_{25} \rangle_{L^2_v}e_{25}\sqrt{\mu_2} \cr
&\quad+ \frac{m_1}{m_2}(1-\delta)\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\sum_{2\leq i \leq 4}\langle f_1, e_{1i} \rangle_{L^2_v}e_{2i}\sqrt{\mu_2}\cr
&\quad+ (1-\omega)\sqrt{\frac{n_{20}}{n_{10}}} \langle f_1, e_{15} \rangle_{L^2_v}e_{25}\sqrt{\mu_2}.
\end{split}
\end{align}
Adding and subtracting the following term
\begin{align*}
\frac{m_1}{m_2}(1-\delta) \sum_{2\leq i \leq 4}\langle f_2, e_{2i} \rangle_{L^2_v}e_{2i}\sqrt{\mu_2} + (1-\omega) \langle f_2, e_{25} \rangle_{L^2_v}e_{25}\sqrt{\mu_2},
\end{align*}
gives
\begin{align*}
\mathcal{M}_{21}'(0)&= P_2f_2\sqrt{\mu_2} \cr
&\quad+ \frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i}\sqrt{\mu_2} \cr
&\quad+ (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25}\sqrt{\mu_2}.
\end{align*}
This completes the proof for the linearization of $\mathcal{M}_{21}$.
\subsection{Linearization of the mixture BGK model}\label{linMBGK} In this part, we linearize the mixture BGK model \eqref{CCBGK}. Applying the linearization of the BGK Maxwellian Lemma \ref{lin ii} and Proposition \ref{linearize}, we substitute $F_1=\mu_1+\sqrt{\mu_1}f_1$ on $(\ref{CCBGK})_1$ and divide it by $\sqrt{\mu_1}$ to have
\begin{align*}
\partial_t f_1+v\cdot \nabla_xf_1&=n_1(P_1f_1-f_1 +\frac{1}{\sqrt{\mu_1}}\int_0^1\mathcal{M}_{11}''(\theta)(1-\theta)d\theta)\cr
&\quad+n_2(P_1f_1-f_1 +\frac{1}{\sqrt{\mu_1}}\int_0^1\mathcal{M}_{12}''(\theta)(1-\theta)d\theta) \cr
&\quad+n_2 \bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i} \cr
&\quad+(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15} \bigg].
\end{align*}
Splitting $n_k$ by $n_k=(n_k-n_{k0})+n_{k0}$,
\begin{align}\label{rho decomp}
\begin{split}
n_k &= n_k-n_{k0} +n_{k0} = \int_{\mathbb{R}^3}f_k\sqrt{\mu_k} dv +n_{k0} =\sqrt{n_{k0}}\langle f_k, e_{k1} \rangle_{L^2_v}+n_{k0},
\end{split}
\end{align}
we can have the following linearized equation:
\begin{align}\label{pertf1}
\partial_t f_1+v\cdot \nabla_xf_1&=L_{11}(f_1)+L_{12}(f_1,f_2)+\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2),
\end{align}
where $L_{11}(f_1)=n_{10}(P_1f_1-f_1)$. The linear term $L_{12}$ is decomposed as $L_{12}=L_{12}^1+L_{12}^2$ with $L_{12}^1= n_{20}(P_1f_1-f_1)$. And $L_{12}^2$ denotes the linear term describing the interchange of momentum and temperature of each species as follows:
\begin{align}\label{def_AL1}
\begin{split}
L_{12}^2(f_1,f_2) &= n_{20} \bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i} \cr
&\quad+(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15} \bigg].
\end{split}
\end{align}
The nonlinear terms $\Gamma_{11}$ and $\Gamma_{12}$ denote
\begin{align*}
\Gamma_{11}(f_1) &= (n_1-n_{10})(P_1f_1-f_1) + n_1 \frac{1}{\sqrt{\mu_1}} \int_0^1\mathcal{M}_{11}''(\theta)(1-\theta)d\theta, \cr
\Gamma_{12}(f_1,f_2) &= (n_2-n_{20})(P_1f_1-f_1)+n_2 \frac{1}{\sqrt{\mu_1}} \int_0^1\mathcal{M}_{12}''(\theta)(1-\theta)d\theta \cr
&\quad+ (n_2-n_{20})\bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i} \cr
&\quad+(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15} \bigg].
\end{align*}
Similarly, we substitute $F_2=\mu_2+\sqrt{\mu_2}f_2$ on $(\ref{CCBGK})_2$ and divide it by $\sqrt{\mu_2}$ to have
\begin{align*}
\partial_t f_2+v\cdot \nabla_xf_2&=n_2(P_2f_2-f_2+\frac{1}{\sqrt{\mu_2}}\int_0^1\mathcal{M}_{22}''(\theta)(1-\theta)d\theta)\cr
&\quad +n_1(P_2f_2-f_2+\frac{1}{\sqrt{\mu_2}}\int_0^1\mathcal{M}_{21}''(\theta)(1-\theta)d\theta) \cr
&\quad+ n_1 \bigg[\frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i} \cr
&\quad+ (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25} \bigg],
\end{align*}
which yields
\begin{align}\label{pertf2}
\partial_t f_2+v\cdot \nabla_xf_2&=L_{22}(f_2)+L_{21}^2(f_1,f_2)+\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2),
\end{align}
where $L_{22}(f_2)=n_{20}(P_2f_2-f_2)$. The linear term $L_{21}$ also decomposed as $L_{21}=L_{21}^1+L_{21}^2$ with $L_{21}^1=n_{10}(P_2f_2-f_2)$. And $L_{21}^2$ denotes the interchange of the momentum and temperature between other species.
\begin{align*}
L_{21}^2(f_1,f_2) &=n_{10} \bigg[\frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i} \cr
&\quad+ (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25} \bigg].
\end{align*}
The nonlinear terms $\Gamma_{22}$ and $\Gamma_{21}$ denote
\begin{align*}
\Gamma_{22}(f_2) &= (n_2-n_{20})(P_2f_2-f_2) + n_2 \frac{1}{\sqrt{\mu_2}} \int_0^1\mathcal{M}_{22}''(\theta)(1-\theta)d\theta, \cr
\Gamma_{21}(f_1,f_2) &= (n_1-n_{10})(P_2f_2-f_2)+n_1 \frac{1}{\sqrt{\mu_2}} \int_0^1\mathcal{M}_{21}''(\theta)(1-\theta)d\theta \cr
&\quad+ (n_1-n_{10}) \bigg[\frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)e_{2i} \cr
&\quad+ (1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)e_{25} \bigg].
\end{align*}
Overall, we can write the linearized mixture BGK model \eqref{CCBGK} as
\begin{align}\label{linf}
\begin{split}
\partial_t f_1+v\cdot \nabla_xf_1&=L_{11}(f_1)+L_{12}(f_1,f_2)+\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2), \cr
\partial_t f_2+v\cdot \nabla_xf_2&=L_{22}(f_2)+L_{21}(f_1,f_2)+\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2), \cr
f_1(x,v,0)=f_{10}&(x,v), \qquad f_2(x,v,0)=f_{20}(x,v).
\end{split}
\end{align}
where $f_{10} =(F_{10}-\mu_1)/\sqrt{\mu_1}$, and $f_{20} = (F_{20}-\mu_2)/\sqrt{\mu_2}$. The linearized mixture BGK model \eqref{linf} satisfies the following conservation laws.
\begin{align}\label{conservf}
\begin{split}
&\int_{\mathbb{T}^3 \times \mathbb{R}^3}\sqrt{\mu_1}f_1(x,v,t) dvdx = \int_{\mathbb{T}^3 \times \mathbb{R}^3}\sqrt{\mu_2}f_2(x,v,t) dvdx =0, \cr
&\int_{\mathbb{T}^3 \times \mathbb{R}^3}\left(\sqrt{\mu_1}f_1(x,v,t)m_1v + \sqrt{\mu_2}f_2(x,v,t)m_2v\right) dvdx =0, \cr
&\int_{\mathbb{T}^3 \times \mathbb{R}^3}\left(\sqrt{\mu_1}f_1(x,v,t)m_1|v|^2 + \sqrt{\mu_2}f_2(x,v,t)m_2|v|^2\right) dvdx =0.
\end{split}
\end{align}
\section{Dissipative property of the linearized relaxation operator}
In this part, we investigate the dissipative property of the linearized multi-component relaxation operator. For simplicity of the notation, we denote the linear operator and the nonlinear perturbation as the vector forms:
\begin{align*}
L_1&=L_{11}(f_1)+L_{12}(f_1,f_2), \cr
L_2&=L_{22}(f_2)+L_{21}(f_1,f_2),
\end{align*}
and
\begin{align*}
\Gamma_1&=\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2), \cr
\Gamma_2&=\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2),
\end{align*}
then we can write \eqref{pertf1} and \eqref{pertf2} as
\begin{align}\label{pertff}
\begin{split}
(\partial_t +v\cdot \nabla_x)(f_1,f_2)&=L(f_1,f_2)+\Gamma(f_1,f_2),
\end{split}
\end{align}
where $L(f_1,f_2)=(L_1,L_2)$ and $\Gamma(f_1,f_2)=(\Gamma_1,\Gamma_2)$.
We also define the following $6$-dimensional orthonormal basis:
\begin{align*}
\begin{split}
&E_1= \frac{1}{\sqrt{n_{10}}}(\sqrt{\mu_1},0), \quad E_2= \frac{1}{\sqrt{n_{20}}}(0,\sqrt{\mu_2}),\cr
&E_i= \frac{1}{\sqrt{m_1n_{10}+m_2n_{20}}}\left(m_1v_{i-2}\sqrt{\mu_1},m_2v_{i-2}\sqrt{\mu_2} \right),\quad (i=3,4,5), \cr
&E_6 =\frac{1}{\sqrt{6n_{10}+6n_{10}}}\left((m_1|v|^2-3)\sqrt{\mu_1},(m_2|v|^2-3)\sqrt{\mu_2}\right).
\end{split}
\end{align*}
We also denote $E_i=(E_i^1,E_i^2)$ for $i=1,\cdots,6$. The macroscopic projection operator for mixture can be written as
\begin{align*}
P(f_1,f_2) = \sum_{1\leq i \leq 6}\langle (f_1,f_2), E_i \rangle_{L^2_v}E_i.
\end{align*}
The following is the main result of this section.
\begin{proposition}\label{dissipation} We have the following dissipation property for the linear operator $L$:
\begin{align*}
\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}&\leq - (n_{10}+n_{20})\Big( \max\{\delta,\omega\}\| (I-P_1,I-P_2)(f_1,f_2) \|_{L^2_{x,v}}^2 \cr
&\quad + \min\left\{(1-\delta),(1-\omega) \right\}\| (I-P)(f_1,f_2)\|_{L^2_{x,v}}^2 \Big).
\end{align*}
\end{proposition}
\begin{proof}
By an explicit computation, we have
\begin{align}\label{LtoP}
\begin{split}
\langle L(f_1,f_2) &,(f_1,f_2)\rangle_{L^2_{x,v}}=
\langle L_1f_1 , f_1\rangle_{L^2_{x,v}}+\langle L_2f_2 , f_2\rangle_{L^2_{x,v}} \cr
&= -(n_{10}+n_{20})\| (I-P_1,I-P_2)(f_1,f_2)\|_{L^2_{x,v}} +\langle L_{12}^2,f_1 \rangle_{L^2_{x,v}}+\langle L_{21}^2,f_2 \rangle_{L^2_{x,v}}.
\end{split}
\end{align}
We decompose the proof in the following 4-steps.\\
({\bf Step 1:}) We consider the dissipation from the momentum and temperature interchange part of the inter-species linearized relaxation operator. We claim that
\begin{align*}
\langle L_{12}^2,f_1 \rangle_{L^2_v}+\langle L_{21}^2,f_2 \rangle_{L^2_v} \leq 0,
\end{align*}
and the equality holds if and only if
\begin{align*}
\frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv=\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2v\sqrt{\mu_2}dv,
\end{align*}
and
\begin{align*}
\frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1(m_1|v|^2-3)\sqrt{\mu_1}dv=\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2(m_2|v|^2-3)\sqrt{\mu_2}dv.
\end{align*}
\noindent$\bullet$ Proof of the claim:
By the definition of $L_{12}^2$ in \eqref{def_AL1}, we have
\begin{align*}
\langle L_{12}^2, f_1 \rangle_{L^2_v}&= (1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)\langle f_1, e_{1i} \rangle_{L^2_v}n_{20} \cr
&\quad+ (1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)\langle f_1, e_{15} \rangle_{L^2_v} n_{20} \cr
&= I_1 + I_2.
\end{align*}
Similarly,
\begin{align*}
\langle L_{21}^2, f_2 \rangle_{L^2_v} &= \frac{m_1}{m_2}(1-\delta)\sum_{2\leq i \leq 4}\left(\sqrt{\frac{n_{20}}{n_{10}}}\sqrt{\frac{m_2}{m_1}}\langle f_1, e_{1i} \rangle_{L^2_v}-\langle f_2, e_{2i} \rangle_{L^2_v}\right)\langle f_2, e_{2i} \rangle_{L^2_v}n_{10} \cr
&\quad + \frac{m_1}{m_2}(1-\omega)\left(\sqrt{\frac{n_{20}}{n_{10}}}\langle f_1, e_{15} \rangle_{L^2_v}-\langle f_2, e_{25} \rangle_{L^2_v}\right)\langle f_2, e_{25} \rangle_{L^2_v}n_{10} \cr
&= I_3 + I_4.
\end{align*}
By an explicit computation, we have
\begin{align}\label{1+3}
\begin{split}
I_1+I_3
&= -(1-\delta) n_{20} \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)^2 \cr
&= -(1-\delta) m_1n_{10}n_{20} \left(\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2v\sqrt{\mu_2}dv-\frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv \right)^2 \leq 0 ,
\end{split}
\end{align}
and
\begin{align}\label{2+4}
\begin{split}
I_2&+I_4 = -(1-\omega)n_{20} \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)^2 \cr
&=-(1-\omega)\frac{n_{10}n_{20}}{6} \left(\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2(m_2|v|^2-3)\sqrt{\mu_2}dv-\frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1(m_1|v|^2-3)\sqrt{\mu_1}dv\right)^2 \cr
&\leq 0 .
\end{split}
\end{align}
which proves the claim of this step. \newline
({\bf Step 2:}) To estimate the gap of the macroscopic projection $(P_1,P_2)$ with $P$, we compute the following term:
\begin{align*}
\|(P_1,P_2)(f_1,f_2)-P(f_1,f_2) \|_{L^2_{x,v}}^2.
\end{align*}
We note that the element of $(P_1,P_2)(f_1,f_2)$ can be written as the linear combination of the following $10$-dimensional basis:
\begin{align*}
\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}),(v\sqrt{\mu_1},0),(0,v\sqrt{\mu_2}), \left(|v|^2\sqrt{\mu_1},0\right),\left(0,|v|^2\sqrt{\mu_2}\right)\}
\end{align*}
so that $(P_1,P_2)P = P$. Therefore,
\begin{align*}
\|(P_1,P_2)(f_1,f_2)-P(f_1,f_2) \|_{L^2_{x,v}}^2 = \|(P_1,P_2)(f_1,f_2)\|_{L^2_{x,v}}^2-\|P(f_1,f_2) \|_{L^2_{x,v}}^2.
\end{align*}
Since we have
\begin{align*}
\int_{\mathbb{R}^3}|P_kf_k|^2 dv &= \frac{1}{n_{k0}}\left(\int_{\mathbb{R}^3}f_k\sqrt{\mu_k} dv\right)^2 + \frac{m_k}{n_{k0}}\left(\int_{\mathbb{R}^3}f_kv\sqrt{\mu_k} dv\right)^2 \cr
& \quad + \frac{1}{6n_{k0}}\left(\int_{\mathbb{R}^3}f_k(m_k|v|^2-3)\sqrt{\mu_k} dv\right)^2,
\end{align*}
and
\begin{align*}
\int_{\mathbb{R}^3}|P(f_1,f_2)|^2 dv &= \frac{1}{n_{10}}\left(\int_{\mathbb{R}^3}f_1\sqrt{\mu_1} dv\right)^2 + \frac{1}{n_{20}}\left(\int_{\mathbb{R}^3}f_2\sqrt{\mu_2} dv\right)^2 \cr
&\quad + \frac{1}{m_1n_{10}+m_2n_{20}}\left(\int_{\mathbb{R}^3}f_1m_1v\sqrt{\mu_1} dv+\int_{\mathbb{R}^3}f_2m_2v\sqrt{\mu_2} dv\right)^2 \cr
&\quad + \frac{1}{6n_{10}+6n_{20}}\left(\int_{\mathbb{R}^3} f(m_1|v|^2-3)\sqrt{\mu_1} dv+\int_{\mathbb{R}^3} f(m_2|v|^2-3)\sqrt{\mu_2} dv\right)^2,
\end{align*}
which follows directly from explicit computations, we can write
\begin{align*}
\|(P_1f_1,P_2f_2)\|_{L^2_{x,v}}^2-\|P(f_1,f_2) \|_{L^2_{x,v}}^2= II_1+II_2,
\end{align*}
where
\begin{align}\label{II_1}
\begin{split}
II_1 &= \frac{1}{m_1n_{10}}\left(\int_{\mathbb{R}^3}f_1m_1v\sqrt{\mu_1} dv\right)^2+\frac{1}{m_2n_{20}}\left(\int_{\mathbb{R}^3}f_2m_2v\sqrt{\mu_2} dv\right)^2 \cr
&\quad -\frac{1}{m_1n_{10}+m_2n_{20}}\left(\left(\int_{\mathbb{R}^3}f_1m_1v\sqrt{\mu_1} dv+\int_{\mathbb{R}^3}f_2m_2v\sqrt{\mu_2} dv\right)^2\right)\\
&=\frac{1}{m_1n_{10}+m_2n_{20}} \left[ \sqrt{\frac{m_2n_{20}}{m_1n_{10}}}\int_{\mathbb{R}^3}f_1m_1v\sqrt{\mu_1} dv-\sqrt{\frac{m_1n_{10}}{m_2n_{20}}}\int_{\mathbb{R}^3}f_2m_2v\sqrt{\mu_2} dv \right]^2
\end{split}
\end{align}
and
\begin{align}\label{II_2}
\begin{split}
II_2&= \frac{1}{6n_{10}}\left(\int_{\mathbb{R}^3}f_1(m_1|v|^2-3)\sqrt{\mu_1} dv\right)^2+\frac{1}{6n_{20}}\left(\int_{\mathbb{R}^3}f_2(m_2|v|^2-3)\sqrt{\mu_2} dv\right)^2 \cr
&\quad - \frac{1}{6n_{10}+6n_{20}}\left(\left(\int_{\mathbb{R}^3} f(m_1|v|^2-3)\sqrt{\mu_1} dv+\int_{\mathbb{R}^3} f(m_2|v|^2-3)\sqrt{\mu_2} dv\right)^2\right)\\
&= \frac{1}{6n_{10}+6n_{20}}\left[\sqrt{\frac{n_{20}}{n_{10}}} \int_{\mathbb{R}^3}f_1(m_1|v|^2-3)\sqrt{\mu_1} dv-\sqrt{\frac{n_{10}}{n_{20}}}\int_{\mathbb{R}^3}f_2(m_2|v|^2-3)\sqrt{\mu_2} dv\right]^2.
\end{split}
\end{align}
\noindent({\bf Step 3:}) In this step, we compare $\langle L_{12}^2,f_1 \rangle_{L^2_{x,v}}+\langle L_{21}^2,f_2 \rangle_{L^2_{x,v}}$ with $\|(P_1,P_2)(f_1,f_2)-P(f_1,f_2) \|_{L^2_{x,v}}^2$ computed in (Step 1) and (Step 2), respectively. We claim that
\begin{align}\label{claimP}
\begin{split}
\langle L_{12}^2,f_1 \rangle_{L^2_{x,v}}+\langle L_{21}^2,f_2 \rangle_{L^2_{x,v}} &\leq -\min\left\{(1-\delta),(1-\omega) \right\} (n_{10}+n_{20}) \cr
&\quad \times\left( \|(P_1,P_2)(f_1,f_2)\|_{L^2_{x,v}}^2-\|P(f_1,f_2) \|_{L^2_{x,v}}^2\right).
\end{split}
\end{align}
which is equivalent to
\begin{align}\label{equivalent}
(n_{10}+n_{20})\left(II_1+II_2\right) \leq -\max\left\{\frac{1}{1-\delta},\frac{1}{1-\omega} \right\} \left[(I_1+I_3)+(I_2+I_4)\right],
\end{align}
where $I_i~(i=1,2,3,4)$ are defined in Step 1, and $II_i~(i=1,2)$ are defined in \eqref{II_1} and \eqref{II_2}.
We first compare $II_2$ with $I_2+I_4$. Multiplying $(n_{10}+n_{20})$ on \eqref{II_2} yields
\begin{align*}
(n_{10}+n_{20})II_2= \frac{1}{6}\left[\sqrt{\frac{n_{20}}{n_{10}}} \int_{\mathbb{R}^3}f_1(m_1|v|^2-3)\sqrt{\mu_1} dv-\sqrt{\frac{n_{10}}{n_{20}}}\int_{\mathbb{R}^3}f_2(m_2|v|^2-3)\sqrt{\mu_2} dv\right]^2,
\end{align*}
which is equal to $-\frac{1}{1-\omega}(I_2+I_4)$ by \eqref{2+4}:
\begin{align}\label{es1}
(n_{10}+n_{20})II_2 = -\frac{1}{1-\omega}(I_2+I_4).
\end{align}
Secondly, we compare $II_1$ with $I_1+I_3$. We multiply $(n_{10}+n_{20})$ on \eqref{II_1}:
\begin{align*}
(n_{10}+n_{20})II_1 &=\frac{n_{10}+n_{20}}{m_1n_{10}+m_2n_{20}} \left[ \sqrt{\frac{m_2n_{20}}{m_1n_{10}}}\int_{\mathbb{R}^3}f_1m_1v\sqrt{\mu_1} dv-\sqrt{\frac{m_1n_{10}}{m_2n_{20}}}\int_{\mathbb{R}^3}f_2m_2v\sqrt{\mu_2} dv \right]^2 \cr
&\leq \frac{1}{m_2}\left[ \sqrt{\frac{m_2n_{20}}{m_1n_{10}}}\int_{\mathbb{R}^3}f_1m_1v\sqrt{\mu_1} dv-\sqrt{\frac{m_1n_{10}}{m_2n_{20}}}\int_{\mathbb{R}^3}f_2m_2v\sqrt{\mu_2} dv \right]^2
\end{align*}
where we used the assumption $m_1\geq m_2$. From \eqref{1+3}, we compute
\begin{align*}
-m_2(I_1+I_3) &= (1-\delta) m_1m_2n_{10}n_{20} \left(\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2v\sqrt{\mu_2}dv-\frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv \right)^2
\end{align*}
which means that
\begin{align}\label{es2}
(n_{10}+n_{20})II_1 \leq -\frac{1}{1-\delta}(I_1+I_3).
\end{align}
Combining the estimates \eqref{es1} and \eqref{es2} yields the desired estimate \eqref{equivalent}.\\
({\bf Step 4:}) Finally, we go back to the estimate \eqref{LtoP}. Applying \eqref{claimP} on \eqref{LtoP} yields
\begin{multline*}
\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}\leq (n_{10}+n_{20})\left(\|(P_1,P_2)(f_1,f_2)\|_{L^2_{x,v}}^2-\|(f_1,f_2)\|_{L^2_{x,v}}^2\right) \cr
- \min\left\{(1-\delta),(1-\omega) \right\}(n_{10}+n_{20})\left( \|(P_1,P_2)(f_1,f_2)\|_{L^2_{x,v}}^2-\|P(f_1,f_2) \|_{L^2_{x,v}}^2\right).
\end{multline*}
So that,
\begin{align*}
\frac{\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}}{n_{10}+n_{20}}&\leq -\|(f_1,f_2)\|_{L^2_{x,v}}^2 + \max\{\delta,\omega\}\|(P_1,P_2)(f_1,f_2)\|_{L^2_{x,v}}^2 \cr
& \quad +\min\left\{(1-\delta),(1-\omega) \right\}\|P(f_1,f_2) \|_{L^2_{x,v}}^2.
\end{align*}
Finally, by splitting $1=\max\{\delta,\omega\}+\min\left\{(1-\delta),(1-\omega) \right\}$ on the coefficient of $\|(f_1,f_2)\|_{L^2}^2$, we conclude that
\begin{align*}
\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}&\leq - (n_{10}+n_{20})\Big( \max\{\delta,\omega\}\| (I-P_1,I-P_2)(f_1,f_2) \|_{L^2_{x,v}}^2 \cr
&\quad + \min\left\{(1-\delta),(1-\omega) \right\}\| (I-P)(f_1,f_2)\|_{L^2_{x,v}}^2 \Big).
\end{align*}
\end{proof}
\begin{lemma} The kernel of the linear operator $L$ satisfies
\begin{align*}
Ker L
&= span\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}),\cr
&\quad (m_1v\sqrt{\mu_1},m_2v\sqrt{\mu_2}),\left((m_1|v|^2-3)\sqrt{\mu_1},(m_2|v|^2-3)\sqrt{\mu_2}\right)\}.
\end{align*}
\end{lemma}
\begin{proof}
We prove the following equivalence condition.
\begin{align*}
\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}=0 \qquad \Leftrightarrow \qquad L(f_1,f_2)=0.
\end{align*}
($\Leftarrow$) This is trivial.\newline
($\Rightarrow$) By Proposition \ref{dissipation}, $\langle L(f_1,f_2),(f_1,f_2)\rangle_{L^2_{x,v}}=0$ implies $(f_1,f_2)=P(f_1,f_2)$. Now it is enough to show that $L(P(f_1,f_2))=0$.
By direct computation,
\begin{align*}
L(P(f_1,f_2)) &= (n_{10}+n_{20})((P_1,P_2)(P(f_1,f_2))-P(f_1,f_2))\cr
&\quad +(L_{12}^2(Pf)+L_{21}^2(Pf)).
\end{align*}
The first term is equal to $0$ since $(P_1,P_2)P = P$. From (Step 1) of Proposition \ref{dissipation}, we can observe that $A_1=A_2=0$ implies $L_{12}^2=L_{21}^2=0$ where
\begin{align*}
A_1&= \frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1v\sqrt{\mu_1}dv-\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2v\sqrt{\mu_2}dv,\cr
A_2&= \frac{1}{n_{10}}\int_{\mathbb{R}^3} f_1(m_1|v|^2-3)\sqrt{\mu_1}dv-\frac{1}{n_{20}}\int_{\mathbb{R}^3} f_2(m_2|v|^2-3)\sqrt{\mu_2}dv.
\end{align*}
Thus we want to prove that $A_1=A_2=0$ when $(f_1,f_2)= P(f_1,f_2) = \sum_{1\leq k \leq 6}\langle (f_1,f_2), E_k \rangle_{L^2_v}E_k$.
From the orthogonality of the basis $E_k^1$ with $v_i\sqrt{\mu_1}$,
\begin{align*}
A_1 &= \frac{1}{n_{10}}\int_{\mathbb{R}^3} \sum_{1\leq k \leq 6}\left[\langle (f_1,f_2), E_k \rangle_{L^2_v}E_k^1\right]v_i\sqrt{\mu_1}dv -\frac{1}{n_{20}}\int_{\mathbb{R}^3} \sum_{1\leq k \leq 6}\left[\langle (f_1,f_2), E_k \rangle_{L^2_v}E_k^2\right]v_i\sqrt{\mu_2}dv \cr
&= \langle (f_1,f_2), E_{i+2} \rangle_{L^2_v} \left(\frac{1}{n_{10}}\int_{\mathbb{R}^3}E_{i+2}^1v_i\sqrt{\mu_1}dv-\frac{1}{n_{20}}\int_{\mathbb{R}^3} E_{i+2}^2v_i\sqrt{\mu_2}dv\right),
\end{align*}
for $i=1,2,3$. By definition of $E_{i+2}$, we have
\begin{align*}
A_1 &=\frac{\langle (f_1,f_2), E_{i+2} \rangle_{L^2_v}}{\sqrt{m_1n_{10}+m_2n_{20}}} \left( \frac{1}{n_{10}}\int_{\mathbb{R}^3}m_1v_i^2\mu_1dv-\frac{1}{n_{20}}\int_{\mathbb{R}^3} m_2v_i^2\mu_2dv\right) =0 .
\end{align*}
Similarly, we compute
\begin{align*}
A_2 &= \frac{1}{n_{10}}\int_{\mathbb{R}^3} \sum_{1\leq k \leq 6}\left[\langle (f_1,f_2), E_k \rangle_{L^2_v}E_k^1\right](m_1|v|^2-3)\sqrt{\mu_1}dv\cr
&\quad -\frac{1}{n_{20}}\int_{\mathbb{R}^3} \sum_{1\leq k \leq 6}\left[\langle (f_1,f_2), E_k \rangle_{L^2_v}E_k^2\right](m_2|v|^2-3)\sqrt{\mu_2}dv \cr
&= \frac{\langle (f_1,f_2), E_6 \rangle_{L^2_v}}{\sqrt{6n_{10}+6n_{10}}}\left(\frac{1}{n_{10}}\int_{\mathbb{R}^3} (m_1|v|^2-3)^2\mu_1dv-\frac{1}{n_{20}}\int_{\mathbb{R}^3} (m_2|v|^2-3)^2\mu_2dv \right) \cr
&=0 ,
\end{align*}
where we used
\begin{align*}
\int_{\mathbb{R}^3}(m_i|v|^2-3)^2\mu_i dv &= \int_{\mathbb{R}^3}(m_i^2|v|^4-6m_i|v|^2+9)\mu_i dv
= 6n_{i0}.
\end{align*}
Thus, $L_{12}^2(Pf)=L_{21}^2(Pf)=0$. Therefore, we conclude that $L(P(f_1,f_2))=0$ and the kernel of $L$ is spanned by the basis of $P$. This completes the proof.
\end{proof}
\begin{remark} Note that in the extreme cases $\delta=1$ or $\omega=1$, we have
\begin{itemize}
\item For $\delta=1$ and $0\leq \omega <1$
\begin{align*}
Ker L
&= span\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}),(v\sqrt{\mu_1},0),(0,v\sqrt{\mu_2})\cr
&\quad \left((m_1|v|^2-3)\sqrt{\mu_1},(m_2|v|^2-3)\sqrt{\mu_2}\right)\}.
\end{align*}
\item For $0\leq \delta <1$ and $\omega=1$
\begin{align*}
Ker L
&= span\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}),(m_1v\sqrt{\mu_1},m_2v\sqrt{\mu_2}),\cr
&\quad \left(|v|^2\sqrt{\mu_1},0\right),\left(0,|v|^2\sqrt{\mu_2}\right)\}.
\end{align*}
\item For $\delta=\omega=1$
\begin{align*}
Ker L
&= span\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}),(v\sqrt{\mu_1},0),(0,v\sqrt{\mu_2})\cr
&\quad \left(|v|^2\sqrt{\mu_1},0\right),\left(0,|v|^2\sqrt{\mu_2}\right)\}.
\end{align*}
\end{itemize}
However, since $\delta=1$ or $\omega=1$ corresponds respectively to the cases where no interchange of momentum or temperature occurs. We exclude the cases in the paper sequel.
\end{remark}
\section{Local existence}
In this section, we prove the local-in-time existence of the mixture BGK model. We start with estimates of the macroscopic fields.
\subsection{Estimate of the macroscopic fields}
\begin{lemma}\label{macro esti} For sufficiently small $\mathcal{E}(t)$, there exists a positive constant $C>0$, such that
\begin{align*}
&(1)~ |n_{k\theta}(x,t)-n_{k0}|\leq C\sqrt{\mathcal{E}(t)},\cr
&(2)~ |U_{ij\theta}(x,t)|\leq C\sqrt{\mathcal{E}(t)}, \cr
&(3)~ |T_{ij\theta}(x,t)-1| \leq C\sqrt{\mathcal{E}(t)},
\end{align*}
for $k=1,2$ and $(i,j)=(1,2)$ or $(2,1)$.
\end{lemma}
\begin{proof}
We recall the estimates for the mono-species macroscopic fields in \cite{Yun1}:
\begin{align*}
|n_{k\theta}(x,t)-n_{k0}|,\quad|U_{k\theta}(x,t)|,\quad|T_{k\theta}(x,t)-1| \leq C\sqrt{\mathcal{E}(t)}.
\end{align*}
Therefore, from the definition of $U_{12\theta}$, $U_{21\theta}$, $T_{12\theta}$, and $T_{21\theta}$ in \eqref{12theta}, we have
\begin{align*}
|U_{12\theta}|&\leq\delta |U_{1\theta}| + (1-\delta)|U_{2\theta}| \leq C\sqrt{\mathcal{E}(t)}, \cr
|U_{21\theta}|&\leq\frac{m_1}{m_2}(1-\delta)|U_{1\theta}| + \left(1-\frac{m_1}{m_2}(1-\delta)\right)|U_{2\theta}| \leq C\sqrt{\mathcal{E}(t)}, \cr
|T_{12\theta}|&= \omega |T_{1\theta}| + (1-\omega)|T_{2\theta}| + \gamma |U_{2\theta}-U_{1\theta}|^2 \leq C\sqrt{\mathcal{E}(t)}+C\mathcal{E}(t), \cr
\end{align*}
and
\begin{align*}
|T_{21\theta}|&= (1-\omega) |T_{1\theta}| + \omega |T_{2\theta}| +\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right) |U_{2\theta}-U_{1\theta}|^2 \cr
& \leq C\sqrt{\mathcal{E}(t)}+C\mathcal{E}(t),
\end{align*}
for sufficiently small $\mathcal{E}(t)$.
\end{proof}
\begin{lemma}\label{macro diff} For $|\alpha|\geq 1$ and sufficiently small $\mathcal{E}(t)$, there exists a positive constant $C_{\alpha}>0$, such that
\begin{align*}
&(1)~ |\partial^{\alpha}n_{k\theta}(x,t)|\leq C_{\alpha}\| \partial^{\alpha}f_k \|_{L^2_v}, \cr
&(2)~ |\partial^{\alpha}U_{ij\theta}(x,t)|\leq C_{\alpha} \sum_{|\alpha_1|\leq|\alpha|}\| \partial^{\alpha_1}f_k \|_{L^2_v}, \cr
&(3)~ |\partial^{\alpha}T_{ij\theta}(x,t)| \leq C_{\alpha} \sum_{|\alpha_1|\leq|\alpha|}\| \partial^{\alpha_1}f_k \|_{L^2_v}+C_{\alpha}\sum_{|\alpha_1|\leq|\alpha|}\| \partial^{\alpha_1}(f_1,f_2) \|_{L^2_v}^2,
\end{align*}
for $k=1,2$ and $(i,j)=(1,2)$ or $(2,1)$.
\end{lemma}
\begin{proof}
We recall \eqref{12theta} and use the following estimates from \cite{Yun1}:
\begin{align}\label{preresult}
\begin{split}
&|\partial^{\alpha}n_{k\theta}(x,t)|,~|\partial^{\alpha}U_{k\theta}(x,t)|,
~|\partial^{\alpha}T_{k\theta}(x,t)| \leq C_{\alpha} \sum_{|\alpha_1|\leq|\alpha|}\| \partial^{\alpha_1}f_k \|_{L^2_v}. \quad (k=1,2),
\end{split}
\end{align}
to get
\begin{align*}
|\partial^{\alpha}U_{12\theta}|&\leq\delta |\partial^{\alpha}U_{1\theta}| + (1-\delta)|\partial^{\alpha}U_{2\theta}| \leq C_{\alpha} \sum_{|\alpha_1|\leq|\alpha|}\|\partial^{\alpha_1}(f_1,f_2) \|_{L^2_v}, \cr
|\partial^{\alpha}U_{21\theta}|&\leq\frac{m_1}{m_2}(1-\delta)|\partial^{\alpha}U_{1\theta}| + \left(1-\frac{m_1}{m_2}(1-\delta)\right)|\partial^{\alpha}U_{2\theta}| \leq C_{\alpha} \sum_{|\alpha_1|\leq|\alpha|}\|\partial^{\alpha_1}(f_1,f_2) \|_{L^2_v},
\end{align*}
and
\begin{align*}
|\partial^{\alpha}T_{12\theta}|&= \omega |\partial^{\alpha}T_{1\theta}| + (1-\omega)|\partial^{\alpha}T_{2\theta}| + \gamma \partial^{\alpha}|U_{2\theta}-U_{1\theta}|^2, \cr
|\partial^{\alpha}T_{21\theta}|&= (1-\omega) |\partial^{\alpha}T_{1\theta}| + \omega |\partial^{\alpha}T_{2\theta}| \cr
&\quad +\left(\frac{1}{3}m_1(1-\delta)\left(\frac{m_1}{m_2}(\delta-1)+1+\delta\right)-\gamma\right) \partial^{\alpha}|U_{2\theta}-U_{1\theta}|^2.
\end{align*}
Then by Young's inequality and using $(\ref{preresult})_2$, we have
\begin{align*}
\partial^{\alpha}|U_{2\theta}-U_{1\theta}|^2 &= \sum_{\alpha_1+\alpha_2=\alpha}2\partial^{\alpha_1}(U_{2\theta}-U_{1\theta})\cdot \partial^{\alpha_2}(U_{2\theta}-U_{1\theta}) \cr
&\leq C_{\alpha}\sum_{|\alpha_1|\leq|\alpha|}\|\partial^{\alpha_1}(f_1,f_2) \|_{L^2_v}^2,
\end{align*}
which gives desired result.
\end{proof}
\subsection{Estimate of the nonlinear term}
We now consider the estimates of nonlinear perturbation $\Gamma$.
\begin{lemma}\label{nonlin} There exist non-negative integer $\lambda$, $\nu$, $\xi$, and general polynomial $\mathcal{P}_{lm}$ satisfying
\begin{align*}
\{\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{ij}(\theta)\}_{l,m} =\frac{\mathcal{P}_{lm}(n_{1\theta},n_{2\theta},U_{1\theta},U_{2\theta},T_{1\theta},T_{2\theta},v-U_{ij\theta})}{n_{1\theta}^{\lambda}n_{2\theta}^{\nu}T_{ij\theta}^{\xi}}\mathcal{M}_{ij}(\theta),
\end{align*}
where $\mathcal{P}_{lm}(x_1,\cdots,x_n) = \sum_k a_k x_1^{k_1}\cdots x_n^{k_n}$ and the indices $k_1,\cdots,k_n$ are non-negative integer, $ij=12$ or $ij=21$, and $1\leq l,m \leq 10$.
\end{lemma}
\begin{proof} The estimates of $\mathcal{M}_{12}(\theta)$ and $\mathcal{M}_{21}(\theta)$ are similar. We only consider the former case. We compute
\begin{align*}
\nabla_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)&=\left( \frac{\partial(n_{1\theta}, n_{1\theta}U_{1\theta}, G_{1\theta},n_{2\theta}, n_{2\theta}U_{2\theta}, G_{2\theta})} {\partial(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})} \right)^{-1} \cr
&\quad\times \nabla_{(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta).
\end{align*}
Then, as in \eqref{block inv}, we have
\begin{align*}
\nabla_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)&=\left[ {\begin{array}{cccccc}
J^{-1}_{1\theta} & 0 \\
0 & J^{-1}_{2\theta}
\end{array} } \right] \times \nabla_{(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta) \cr
&= \left[ {\begin{array}{cccccc}
J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta) \\
J_{2\theta}^{-1}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta)
\end{array} } \right] .
\end{align*}
Applying the same process once more time, we get
\begin{align*}
\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)&= \left[ {\begin{array}{cccccc}
J^{-1}_{1\theta} & 0 \\
0 & J^{-1}_{2\theta}
\end{array} } \right] \cr
&\quad \times \nabla_{(n_{1\theta},U_{1\theta},T_{1\theta},n_{2\theta},U_{2\theta},T_{2\theta})}\left[ {\begin{array}{cccccc}
J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta) \\
J_{2\theta}^{-1}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta)
\end{array} } \right],
\end{align*}
where the second line on the R.H.S. is equal to
\begin{multline*}
\left[ {\begin{array}{cccccc}
\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\left(J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\right) && \nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\left(J_{2\theta}^{-1}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta)\right) \\
\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\left(J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\right) && \nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\left(J_{2\theta}^{-1}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta)\right)
\end{array} } \right].
\end{multline*}
Thus we get
\iffalse
\begin{multline*}
\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta) \cr
= \left[ {\begin{array}{cccccc}
J^{-1}_{1\theta}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\{J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\} & J^{-1}_{1\theta}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\left(J_{2\theta}^{-1}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta)\right) \\
J^{-1}_{2\theta}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\left(J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\right) & J^{-1}_{2\theta}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\{J_{2\theta}^{-1}\nabla_{(n_{2\theta},U_{2\theta},T_{2\theta})}\mathcal{M}_{12}(\theta)\}
\end{array} } \right]
\end{multline*}
\fi
\begin{align*}
\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)=\left[ {\begin{array}{cccccc}
T_{11} && T_{12} \\ T_{21} && T_{22}
\end{array} } \right],
\end{align*}
where
\begin{align*}
T_{ij}= J^{-1}_{i\theta}\nabla_{(n_{i\theta},U_{i\theta},T_{i\theta})}\left(J_{j\theta}^{-1}\nabla_{(n_{j\theta},U_{j\theta},T_{j\theta})}\mathcal{M}_{12}(\theta)\right),
\end{align*}
for $i,j=1,2$. Each $T_{ij}$ is a $5\times 5$ matrix. For simplicity, we only consider the $(1,1)$ and $(1,2)$ components of $\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)$. We can treat other components similarly. Recall that the first row of $J^{-1}_{1\theta}$ is $(1,0,0,0,0)$, so that
\begin{align*}
\{\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)\}_{11}=\frac{\partial }{\partial n_{1\theta}}\frac{\partial \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}} &= \frac{\partial }{\partial n_{1\theta}} \left(\frac{1}{n_{1\theta}}\mathcal{M}_{12}(\theta)\right) = 0.
\end{align*}
Now we consider the $(1,2)$ component of $\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)$ which is inner product of the first row of $J^{-1}_{1\theta}$ which is $(1,0,0,0,0)$, and the second column of $\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\{J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\}$. Thus,
we only need (1,2) component of $\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\{J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\}$:
\begin{align}\label{D2M12}
\begin{split}
\{\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)\}_{12}&= \left[\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\{J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)\}\right]_{12} \cr
&=\frac{\partial }{\partial n_{1\theta}} \left[J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta) \right]_2.
\end{split}
\end{align}
The second component of $\left[J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta) \right]$ is equal to the inner product of the second row of $J_{1\theta}^{-1}$ and $\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta)$:
\begin{align*}
\left[J_{1\theta}^{-1}\nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta) \right]_2 &= \left(-\frac{U_{1\theta}}{n_{1\theta}},\frac{1}{n_{1\theta}},0,0,0\right)\cdot \nabla_{(n_{1\theta},U_{1\theta},T_{1\theta})}\mathcal{M}_{12}(\theta) \cr
&=-\frac{U_{1\theta}}{n_{1\theta}} \frac{\partial \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}} + \frac{1}{n_{1\theta}} \frac{\partial \mathcal{M}_{12}(\theta)}{\partial U_{11\theta}}.
\end{align*}
Substituting this into \eqref{D2M12} gives
\begin{align*}
\{\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)\}_{12}
&=\frac{\partial }{\partial n_{1\theta}} \left(-\frac{U_{11\theta}}{n_{1\theta} }\frac{\partial \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}}+\frac{1}{n_{1\theta}}\frac{\partial \mathcal{M}_{12}(\theta)}{\partial U_{11\theta}}\right) \cr
&= \frac{U_{11\theta}}{n_{1\theta}^2 }\frac{\partial \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}}-\frac{U_{11\theta}}{n_{1\theta} }\frac{\partial^2 \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}^2}-\frac{1}{n_{1\theta}^2}\frac{\partial \mathcal{M}_{12}(\theta)}{\partial U_{11\theta}}+\frac{1}{n_{1\theta}}\frac{\partial^2 \mathcal{M}_{12}(\theta)}{\partial n_{1\theta}\partial U_{11\theta}}.
\end{align*}
Then, from Lemma \ref{M_12 diff} (1) and (2), we have
\begin{align*}
\{\nabla^2_{(H_{1\theta},H_{2\theta})}&\mathcal{M}_{12}(\theta)\}_{12}
= \frac{U_{11\theta}}{n_{1\theta}^3}\mathcal{M}_{12}(\theta)\cr
& + \frac{1}{n_{1\theta}}\left(\delta m_1\frac{v-U_{12\theta}}{T_{12\theta}} -2\gamma(U_{2\theta}-U_{1\theta})\left(-\frac{3}{2}\frac{1}{T_{12\theta}}+\frac{m_1|v-U_{12\theta}|^2}{2T_{12\theta}^2}\right)\right)\mathcal{M}_{12}(\theta).
\end{align*}
We observe that $(1,2)$ component of $\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)$ is expressed in the form presented in this lemma.
\end{proof}
We are now ready to estimate the nonlinear terms. The intra-species part is established in \cite{Yun1}:
\begin{lemma}\emph{\cite{Yun1}}\label{nonlin esti1} For sufficiently small $\mathcal{E}(t)$, we have the following inequality for $k=1,2$.
\begin{align*}
\langle \partial^{\alpha}_{\beta} \Gamma_{kk}(f_k), g\rangle_{L^2_v} \leq C\sum_{|\alpha_1|+|\alpha_2|\leq |\alpha|}\|\partial^{\alpha_1}f_k\|_{L^2_v}\|\partial^{\alpha_2}f_k\|_{L^2_v}\|g \|_{L^2_v}.
\end{align*}
\end{lemma}
So we focus on the inter-species part.
\begin{lemma}\label{nonlin esti} Let $N\geq3$ and $|\alpha|+|\beta|\leq N$. For sufficiently small $\mathcal{E}(t)$, we have
\begin{align*}
\langle \partial^{\alpha}_{\beta} \Gamma_{ij}, g\rangle_{L^2_v} \leq C\sum_{|\alpha_1|+|\alpha_2|\leq |\alpha|}\|\partial^{\alpha_1}(f_1,f_2)\|_{L^2_v}\|\partial^{\alpha_2}_{\beta}(f_1,f_2)\|_{L^2_v}\|g \|_{L^2_v},
\end{align*}
for $(i,j)=(1,2)$ or $(2,1)$.
\end{lemma}
\begin{proof}
We only consider the $\Gamma_{12}$ since the estimate of $\Gamma_{21}$ is similar.
Therefore, we focus on the estimates of the nonlinear terms $\Gamma_{12}$ and $\Gamma_{21}$. For convenience, we divide $\Gamma_{12}$ into three parts:
\begin{align*}
\Gamma_{12} = \Gamma_{12A}+\Gamma_{12B}+\Gamma_{12C},
\end{align*}
where
\begin{align*}
\Gamma_{12A}&= (n_2-n_{20})(P_1f_1-f_1), \cr
\Gamma_{12B}&= n_2 \frac{1}{\sqrt{\mu_1}} \int_0^1\mathcal{M}_{12}''(\theta)(1-\theta)d\theta, \cr
\Gamma_{12C}&= (n_2-n_{20})\bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle f_2, e_{2i} \rangle_{L^2_v}-\langle f_1, e_{1i} \rangle_{L^2_v}\right)e_{1i} \cr
&\quad +(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle f_2, e_{25} \rangle_{L^2_v}-\langle f_1, e_{15} \rangle_{L^2_v}\right)e_{15} \bigg].
\end{align*}
We first write $\Gamma_{12B}$ in a concise form before we delve into the estimate. For this, compute applying the chain rule twice on $\mathcal{M}_{ij}$:
\begin{align*}
&\mathcal{M}_{ij}''(\theta) \cr
&\quad= \frac{d}{d\theta} \bigg( \frac{d n_{\theta1}}{d \theta}\frac{d \mathcal{M}_{ij}}{d n_{\theta1}}+\frac{d (n_{\theta1}U_{\theta1})}{d \theta}\frac{d \mathcal{M}_{ij}}{d (n_{\theta1}U_{\theta1})}+\frac{d G_{\theta1}}{d \theta}\frac{d \mathcal{M}_{ij}}{d G_{\theta1}} \cr
&\qquad +\frac{d n_{\theta2}}{d \theta}\frac{d \mathcal{M}_{ij}}{d n_{\theta2}}+\frac{d (n_{\theta2}U_{\theta2})}{d \theta}\frac{d \mathcal{M}_{ij}}{d (n_{\theta2}U_{\theta2})}+\frac{d G_{\theta2}}{d \theta}\frac{d \mathcal{M}_{ij}}{d G_{\theta2}} \bigg) \cr
&\quad=(n_1-n_{10}, n_1 U_1, G_1,n_2-n_{20}, n_2 U_2, G_2)^T\left\{\nabla^2_{(n_{1\theta}, n_{1\theta}U_{1\theta}, G_{1\theta},n_{2\theta}, n_{2\theta}U_{2\theta}, G_{2\theta})}\mathcal{M}_{ij}(\theta)\right\} \cr
&\qquad\times (n_1-n_{10}, n_1 U_1, G_1,n_2-n_{20}, n_2 U_2, G_2).
\end{align*}
Therefore, if we define
\begin{align}\label{defH}
H_{k}=(n_{k}, n_{k}U_{k}, G_{k}) ,\quad \textit{and} \quad H_{k\theta}=(n_{k\theta}, n_{k\theta}U_{k\theta}, G_{k\theta}),
\end{align}
we can rewrite $\Gamma_{12B}$ as
\begin{align*}
\Gamma_{12B} &= \frac{n_2}{\sqrt{\mu_1}} (H_1-H_{10},H_2-H_{20})^T \cr
&\quad \times \int_0^1 \left\{\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{ij}(\theta)\right\} (1-\theta)d\theta (H_1-H_{10},H_2-H_{20}).
\end{align*}
Now we estimate each part of $\Gamma_{12}$. \\
$\bullet$ {\bf Estimate of $\Gamma_{12A}$:} We take a derivative $\partial^{\alpha}_{\beta}$ on $\Gamma_{12A}$:
\begin{align*}
\partial^{\alpha}_{\beta}\Gamma_{12A} &=\sum_{\alpha_1+\alpha_2=\alpha} C_{\alpha_1} \partial^{\alpha_1}(n_2-n_{20})\partial^{\alpha_2}_{\beta}(P_1f_1-f_1).
\end{align*}
From \eqref{rho decomp}, we have
\begin{align}\label{e1}
\partial^{\alpha}(n_2-n_{20}) \leq C\| \partial^{\alpha}f_2 \|_{L^2_v}.
\end{align}
For an estimate of the macroscopic projection $P_1f_1$, since $\partial_{\beta} e_{1i}$ has an exponential decay, we get
\begin{align*}
\| \partial^{\alpha}_{\beta}P_1f_1 \|_{L^2_v}= \| \partial_{\beta}P_1\partial^{\alpha}f_1 \|_{L^2_v} \leq C_{\beta}\| \partial^{\alpha}f_1 \|_{L^2_v}.
\end{align*}
Thus we have
\begin{align}\label{P1}
\langle \partial^{\alpha}_{\beta}(P_1f_1-f_1), g \rangle_{L^2_v
\leq C\left(\| \partial^{\alpha}f_1 \|_{L^2_v}+\| \partial^{\alpha}_{\beta}f_1 \|_{L^2_v}\right)\|g\|_{L^2_v}.
\end{align}
Combining \eqref{e1} and \eqref{P1}, we obtain
\begin{align*}
\langle \partial_{\beta}^{\alpha} \Gamma_{12A},g \rangle_{L^2_v} \leq C\sum_{|\alpha_1|+|\alpha_2|+|\beta|\leq N}\| \partial^{\alpha_1}f_2 \|_{L^2_v}\left(\| \partial^{\alpha_2}f_1 \|_{L^2_v}+\| \partial^{\alpha_2}_{\beta}f_1 \|_{L^2_v}\right)\|g\|_{L^2_v}.
\end{align*}
$\bullet$ {\bf Estimate of $\Gamma_{12C}$:} We take a derivative $\partial_{\beta}^{\alpha}$ on $\Gamma_{12C}$:
\begin{align*}
\partial^{\alpha}_{\beta}\Gamma_{12C}&= \sum_{\alpha_1+\alpha_2=\alpha}C_{\alpha_1} \partial^{\alpha_1}(n_2-n_{20})\cr
&\quad \times \bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle \partial^{\alpha_2}f_2, e_{2i} \rangle_{L^2_v}-\langle \partial^{\alpha_2}f_1, e_{1i} \rangle_{L^2_v}\right)\partial_{\beta}e_{1i} \cr
&\quad +(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle \partial^{\alpha_2}f_2, e_{25} \rangle_{L^2_v}-\langle \partial^{\alpha_2}f_1, e_{15} \rangle_{L^2_v}\right)\partial_{\beta}e_{15} \bigg].
\end{align*}
Since each $e_{1i}$ and $e_{2i}$ has exponential decay for $i=1,\cdots,5$, we can have
\begin{align}
\langle \partial^{\alpha}f_1, e_{1i} \rangle_{L^2_v} \leq C\| \partial^{\alpha}f_1 \|_{L^2_v}, \qquad \langle \partial^{\alpha}f_2, e_{2i} \rangle_{L^2_v} \leq C\| \partial^{\alpha}f_2 \|_{L^2_v},
\label{e2}
\end{align}
and
\begin{align}\label{e3}
\langle \partial_{\beta}e_{1i}, g \rangle_{L^2_v} \leq C\|g\|_{L^2_v} \qquad \langle \partial_{\beta}e_{2i}, g \rangle_{L^2_v} \leq C\|g\|_{L^2_v}.
\end{align}
Thus by using \eqref{e1}, \eqref{e2}, and \eqref{e3}, we get
\begin{align*}
\langle \partial^{\alpha}_{\beta} \Gamma_{12C}, g\rangle_{L^2_v} \leq C\sum_{|\alpha_1|+|\alpha_2|\leq |\alpha|}\|\partial^{\alpha_1}f_2\|_{L^2_v}\|\partial^{\alpha_2}(f_1,f_2)\|_{L^2_v}\|g \|_{L^2_v}.
\end{align*}
$\bullet$ {\bf Estimate of $\Gamma_{12B}$:} Taking $\partial^{\alpha}_{\beta}$ on $\Gamma_{12B}$ gives
\begin{align}\label{gamma12Bes}
\begin{split}
\partial^{\alpha}_{\beta}\Gamma_{12B} &= \sum_{\sum \alpha_i=\alpha}C_{\alpha_i}\partial^{\alpha_0}n_2 \partial^{\alpha_1}(H_1-H_{10},H_2-H_{20})^T \cr
&\times \int_0^1\partial^{\alpha_2}_{\beta}\left\{\frac{1}{\sqrt{\mu_1}}\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)\right\}(1-\theta)d\theta \partial^{\alpha_3}(H_1-H_{10},H_2-H_{20}) .
\end{split}
\end{align}
By the definition of $H_k$ in \eqref{defH}, applying \eqref{lin1} yields
\begin{align*}
\partial^{\alpha}(H_k-H_{k0})&=\partial^{\alpha}(n_k-n_{k0}, n_k U_k, G_k)= \left( \langle \partial^{\alpha}f_k, e_{k1} \rangle_{L^2_v}, \cdots, \langle \partial^{\alpha}f_k, e_{k5} \rangle_{L^2_v} \right),
\end{align*}
for $k=1,2$.
Thus we have
\begin{align}\label{HkH0}
|\partial^{\alpha}(H_k-H_{k0})| \leq C \| \partial^{\alpha}f_k \|_{L^2_v}.
\end{align}
For notational simplicity, we set
\begin{align*}
A_{lm}=\int_0^1\partial^{\alpha_2}_{\beta}\left\{\frac{1}{\sqrt{\mu_1}}\nabla^2_{(H_{1\theta},H_{2\theta})}\mathcal{M}_{12}(\theta)\right\}_{l,m}(1-\theta)d\theta.
\end{align*}
Then by Lemma \ref{nonlin}, we can write it as
\begin{align}\label{Alm}
A_{lm}=\int_0^1\partial^{\alpha_2}_{\beta}\left\{\frac{1}{\sqrt{\mu_1}} \frac{\mathcal{P}_{lm}(n_{1\theta},n_{2\theta},U_{1\theta},U_{2\theta},T_{1\theta},T_{2\theta},v-U_{12\theta})}{n_{1\theta}^{\lambda}n_{2\theta}^{\nu}T_{12\theta}^{\xi}}\mathcal{M}_{12}(\theta)\right\}(1-\theta)d\theta.
\end{align}
By the product rule, we have
\begin{align*}
&\partial^{\alpha}_{\beta}\left\{ \frac{\mathcal{P}_{lm}(n_{1\theta},n_{2\theta},U_{1\theta},U_{2\theta},T_{1\theta},T_{2\theta},v-U_{12\theta})}{n_{1\theta}^{\lambda}n_{2\theta}^{\nu}T_{12\theta}^{\xi}}\right\} \cr
&=C_{\alpha}\sum_{\sum \alpha_i=\alpha}\bigg\{\mathcal{P}_{lm}(\partial^{\alpha_1}n_{1\theta},\partial^{\alpha_2}n_{2\theta},\partial^{\alpha_3}U_{1\theta},\partial^{\alpha_4}U_{2\theta},\partial^{\alpha_5}T_{1\theta},\partial^{\alpha_6}T_{2\theta},\partial^{\alpha_7}_{\beta}(v-U_{12\theta})) \cr
&\quad \times \partial^{\alpha_8}\frac{1}{n_{1\theta}^{\lambda}n_{2\theta}^{\nu}T_{12\theta}^{\xi}}\bigg\}
\end{align*}
If $|\alpha_i|\leq N-2$, then by Sobolev embedding $H^2 \subset\subset L^{\infty}$ and Lemma \ref{macro diff}, we have
\begin{align*}
|\partial^{\alpha}n_{k\theta}(x,t)|+|\partial^{\alpha}U_{k\theta}(x,t)|+|\partial^{\alpha}T_{k\theta}(x,t)|\leq C\| \partial^{\alpha}f_k \|_{L^2_v}\leq \sqrt{\mathcal{E}(t)}.
\end{align*}
Since $N\geq 3$, there is at most one $\alpha_i$ that exceeds $N-2$. Thus, for sufficiently small $\mathcal{E}(t)$, we have
\begin{align*}
\partial^{\alpha}_{\beta}\left\{ \frac{\mathcal{P}_{lm}(n_{1\theta},n_{2\theta},U_{1\theta},U_{2\theta},T_{1\theta},T_{2\theta},v-U_{12\theta})}{n_{1\theta}^{\lambda}n_{2\theta}^{\nu}T_{12\theta}^{\xi}}\right\} \leq C\sqrt{\mathcal{E}(t)}\|\partial^{\alpha}f \|_{L^2_v} \mathcal{P}_{lm}(v).
\end{align*}
Substituting it in \eqref{Alm} yields
\begin{align*}
A_{lm} \leq C\sqrt{\mathcal{E}(t)}\|\partial^{\alpha}f \|_{L^2_v} \mathcal{P}_{lm}(v) \partial^{\alpha}_{\beta} \exp\left(-\frac{|v-U_{12\theta}|^2}{2\frac{T_{12\theta}}{m_1}}+\frac{m_1|v|^2}{4}\right).
\end{align*}
Similarly, the derivative of the exponential part can be bounded as follows:
\begin{multline*}
\partial^{\alpha}_{\beta} \exp\left(-\frac{|v-U_{12\theta}|^2}{2\frac{T_{12\theta}}{m_1}}+\frac{m_1|v|^2}{4}\right) \cr
\leq C\sqrt{\mathcal{E}(t)}\|\partial^{\alpha}f \|_{L^2_v} \mathcal{P}_{lm}(v) \exp\left(-\frac{|v-U_{12\theta}|^2}{2\frac{T_{12\theta}}{m_1}}+\frac{m_1|v|^2}{4}\right).
\end{multline*}
By Lemma \ref{macro esti} (3), a sufficiently small $\mathcal{E}(t)$ guarantees $T_{12\theta} \leq 3/2$, so that
\begin{align}\label{Almg}
\begin{split}
\langle A_{lm} , g \rangle_{L^2_v} &\leq C \left\| P(v) \exp\left(-\frac{2m_1|v-U_{12\theta}|^2}{3}+\frac{m_1|v|^2}{2}\right) \right\|_{L^2_v} \| g \|_{L^2_v} \cr
&\leq C \left\| P(v) \exp\left(-\frac{m_1|v-4U_{12\theta}|^2}{6}+2m_1|U_{12\theta}|^2\right) \right\|_{L^2_v} \| g \|_{L^2_v} \cr
&\leq C\| g \|_{L^2_v},
\end{split}
\end{align}
where we used $e^{2m_1|U_{12\theta}|^2}\leq C$ for sufficiently small $\mathcal{E}(t)$. Substituting \eqref{HkH0} and \eqref{Almg} on \eqref{gamma12Bes} gives the desired result.
\end{proof}
\subsection{Local existence}
In this part, we prove the existence of a local-in-time classical solution of the mixture BGK model \eqref{CCBGK}.
\begin{theorem}\label{theo_loc_ex}
Let $F_{10}=\mu_1 + \sqrt{\mu_1} f_{10}\geq 0$ and $F_{20}= \mu_2 + \sqrt{\mu_2} f_{20} \geq 0$. There exists $ T_*>0$ and $M_0>0$ such that if $\mathcal{E}(0) \leq \frac{M_0}{2}$, then there exists a unique local-in-time solution $(F_1,F_2)$ of \eqref{CCBGK} such that
\begin{enumerate}
\item The distribution functions $F_1(x,v,t)$ and $F_2(x,v,t)$ are non-negative.
\item The high-order energy $\mathcal{E}(t)$ is uniformly bounded:
$$ \sup_{0 \leq t \leq T_*} \mathcal{E}(t) \leq M_0.$$
\item The high-order energy is continuous in $t \in [0,T_*)$.
\item The conservation laws \eqref{conservf} hold for all $t \in [0,T_*)$.
\end{enumerate}
\end{theorem}
\begin{proof}
We define an iteration of the mixture BGK model \eqref{CCBGK} as follows:
\begin{align*}
\begin{aligned}
\partial_t F_1^{n+1}+v\cdot \nabla_xF_1^{n+1}&=n_1(F_1^n)(\mathcal{M}_{11}(F_1^n)-F_1^{n+1})\cr
&\quad +n_2(F_2^n)(\mathcal{M}_{12}(F_1^n,F_2^n)-F_1^{n+1}), \cr
\partial_t F_2^{n+1}+v\cdot \nabla_xF_2^{n+1}&=n_2(F_2^n)(\mathcal{M}_{22}(F_2^n)-F_2^{n+1})\cr
&\quad +n_1(F_1^n)(\mathcal{M}_{21}(F_1^n,F_2^n)-F_2^{n+1}),
\end{aligned}
\end{align*}
and $F_1^{n+1}(x,v,0)=F_{10}(x,v)$ and $F_2^{n+1}(x,v,0)=F_{20}(x,v)$ for all $n \geq 0$. We start the iteration with $F_1^0(x,v,t)=F_{10}(x,v)$ and $F_2^0(x,v,t)=F_{20}(x,v)$.
We split $F_1^n=\mu_1+\sqrt{\mu_1}f_1^n$, and $F_2^n=\mu_2+\sqrt{\mu_2}f_2^n$ for all $n \in \mathbb{N}$ and use the linearization of the Maxwellian given in Proposition \ref{linearize} and Lemma \ref{lin ii} to get
\begin{align*}
\partial_t f_1^{n+1}+v\cdot \nabla_xf_1^{n+1}&=(n_{10}+n_{20})(P_1f_1^n-f_1^{n+1})+L_{12}^2(f_1^n,f_2^n)+\Gamma_{11}(f_1^n)+\Gamma_{12}(f_1^n,f_2^n), \cr
\partial_t f_2^{n+1}+v\cdot \nabla_xf_2^{n+1}&=(n_{10}+n_{20})(P_2f_2^n-f_2^{n+1})+L_{21}^2(f_1^n,f_2^n)+\Gamma_{22}(f_2^n)+\Gamma_{21}(f_1^n,f_2^n).
\end{align*}
Then the local existence can be constructed by the standard argument as in \cite{Guo VMB}.
The key ingredient is the uniform control of the high-order energy norm in each iteration step. So we only prove the following auxiliary lemma below.
\end{proof}
\begin{lemma}
Let $\mathcal{E}(0)<\frac{M_0}{2}$. Then there exists $T_*>0$ and $M_0>0$ such that $\mathcal{E}(f^n(t))<M_0$ for all $n\geq 0$ and $t\in [0,T_*].$
\end{lemma}
\begin{proof} We take $\partial^{\alpha}_{\beta}$ on each side of \eqref{pertf1} and \eqref{pertf2}:
\begin{multline*}
\partial^{\alpha}_{\beta}\partial_t f_1^{n+1}+v\cdot \nabla_x\partial^{\alpha}_{\beta}f_1^{n+1}+\sum_{i=1}^3 \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1^{n+1}
=(n_{10}+n_{20})(\partial_{\beta}P_1\partial^{\alpha}f_1^n-\partial^{\alpha}_{\beta}f_1^{n+1})\cr
+\partial^{\alpha}_{\beta}L_{12}^2(f_1^n,f_2^n)+\partial^{\alpha}_{\beta}\Gamma_{11}(f_1^n)+\partial^{\alpha}_{\beta}\Gamma_{12}(f_1^n,f_2^n),
\end{multline*}
and
\begin{multline*}
\partial^{\alpha}_{\beta}\partial_t f_2^{n+1}+v\cdot \nabla_x\partial^{\alpha}_{\beta}f_2^{n+1}+\sum_{i=1}^3 \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_2^{n+1}=(n_{10}+n_{20})(\partial_{\beta}P_2\partial^{\alpha}f_2^n-\partial^{\alpha}_{\beta}f_2^{n+1})\cr
+\partial^{\alpha}_{\beta}L_{21}^2(f_1^n,f_2^n)+\partial^{\alpha}_{\beta}\Gamma_{22}(f_2^n)+\partial^{\alpha}_{\beta}\Gamma_{21}(f_1^n,f_2^n),
\end{multline*}
where $k_1=(1,0,0)$, $k_2=(0,1,0)$, $k_3=(0,0,1)$, and $\bar{k}_1=(0,1,0,0)$, $\bar{k}_2=(0,0,1,0)$, $\bar{k}_3=(0,0,0,1)$. We then take the inner product with $\partial^{\alpha}_{\beta}f_1^{n+1}$:
\begin{align}\label{eqtoest}
\begin{split}
\frac{1}{2}\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}^2&+(n_{10}+n_{20})\|\partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}^2 = - \sum_{i=1}^3 \langle \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1^{n+1}, \partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} \cr
&\quad +
\langle \partial_{\beta}P_1\partial^{\alpha}f_1^n,\partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}}
+\langle \partial^{\alpha}_{\beta}L_{12}^2(f_1^n,f_2^n), \partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} \cr
&\quad +\langle \partial^{\alpha}_{\beta}\Gamma_{11}(f_1^n),\partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}}+\langle \partial^{\alpha}_{\beta}\Gamma_{12}(f_1^n,f_2^n),\partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} \cr
&\quad = I_1+I_2+I_3+I_4+I_5.
\end{split}
\end{align}
Applying the H\"{o}lder inequality on $I_1$, we have
\begin{align*}
I_1=\sum_{i=1}^3 \langle \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1^{n+1}, \partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} &\leq \sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1^{n+1}\|_{L^2_{x,v}}\|\partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}} \cr
& \leq \sum_{|\alpha|+|\beta|\leq N } \|\partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}^2.
\end{align*}
Since $\partial_{\beta} e_{1i}$ and $\partial_{\beta} e_{2i}$ have exponential decay,
\begin{align*}
\| \partial_{\beta}P_1\partial^{\alpha}f_1^{n} \|_{L^2_{x,v}} \leq C_{\beta}\| \partial^{\alpha}f_1^{n} \|_{L^2_{x,v}}.
\end{align*}
Thus Young's inequality implies
\begin{align*}
I_2=\langle \partial_{\beta}P_1\partial^{\alpha}f_1^n,\partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} &\leq C_{\beta} \| \partial^{\alpha}f_1^n\|_{L^2_{x,v}}^2 + C \| \partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}^2.
\end{align*}
To estimate $I_3$, we take $\partial^{\alpha}_{\beta}$ on $L_{12}^2$:
\begin{multline*}
\partial^{\alpha}_{\beta}L_{12}^2(f_1,f_2) = n_{20} \bigg[(1-\delta) \sum_{2\leq i \leq 4} \left(\sqrt{\frac{n_{10}}{n_{20}}}\sqrt{\frac{m_1}{m_2}}\langle \partial^{\alpha}f_2, e_{2i} \rangle_{L^2_v}-\langle \partial^{\alpha}f_1, e_{1i} \rangle_{L^2_v}\right)\partial_{\beta}e_{1i} \cr
+(1-\omega) \left(\sqrt{\frac{n_{10}}{n_{20}}}\langle \partial^{\alpha}f_2, e_{25} \rangle_{L^2_v}-\langle \partial^{\alpha}f_1, e_{15} \rangle_{L^2_v}\right)\partial_{\beta}e_{15} \bigg],
\end{multline*}
and apply the H\"{o}lder inequality:
\begin{align*}
I_3=\langle \partial^{\alpha}_{\beta}L_{12}^2(f_1^n,f_2^n), \partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} &\leq C\int_{\mathbb{T}^3} \left(\| \partial^{\alpha}f_2^n\|_{L^2_v} + C\| \partial^{\alpha}f_1^n\|_{L^2_v}\right) \| \partial^{\alpha}_{\beta}f_1^{n+1}\|_{L^2_v} dx \cr
&\leq C\| \partial^{\alpha}(f_1^n,f_2^n)\|_{L^2_{x,v}} \| \partial^{\alpha}_{\beta}f_1^{n+1}\|_{L^2_{x,v}}.
\end{align*}
Since $I_4$ and $I_5$ are similar, we only consider $I_5$. Applying Lemma \ref{nonlin esti}, we have
\begin{align*}
I_5=\langle \partial^{\alpha}_{\beta}\Gamma_{12}(f_1^n,f_2^n),\partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} &\leq C\sum_{|\alpha_1|+|\alpha_2|\leq |\alpha|} \int_{\mathbb{T}^3}\|\partial^{\alpha_1}(f_1^n,f_2^n)\|_{L^2_v} \cr
&\quad\times \|\partial^{\alpha_2}(f_1^n,f_2^n)\|_{L^2_v}\| \partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_v}dx.
\end{align*}
Without loss of generality, we assume that $|\alpha_1|\leq |\alpha_2|$. Then the Sobolev embedding $H^2 \subset\subset L^{\infty}$ implies
\begin{align*}
I_5&=\langle \partial^{\alpha}_{\beta}\Gamma_{12}(f_1^n,f_2^n),\partial^{\alpha}_{\beta} f_1^{n+1} \rangle_{L^2_{x,v}} \cr
&\leq C\bigg(\sum_{|\alpha_1|\leq |\alpha|} \|\partial^{\alpha_1}(f_1^n,f_2^n)\|_{L^2_{x,v}}\bigg)^2 \| \partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}.
\end{align*}
Combining the estimate from $I_1$ to $I_5$, and taking $\sum_{|\alpha|+|\beta|\leq N}$ on \eqref{eqtoest}, we have
\begin{align}\label{est_dt1}
\begin{split}
\frac{1}{2}\sum_{|\alpha|+|\beta|\leq N }&\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}^2+(n_{10}+n_{20})\sum_{|\alpha|+|\beta|\leq N }\|\partial^{\alpha}_{\beta} f_1^{n+1} \|_{L^2_{x,v}}^2 \cr
&\leq C\mathcal{E}^n(t) + C \mathcal{E}^{n+1}(t) + C\sqrt{\mathcal{E}^n(t)}\sqrt{\mathcal{E}^{n+1}(t)} +C\mathcal{E}^n(t)\sqrt{\mathcal{E}^{n+1}(t)}.
\end{split}
\end{align}
Similarly,
\begin{align}\label{est_dt2}
\begin{split}
\frac{1}{2}\sum_{|\alpha|+|\beta|\leq N }&\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_2^{n+1} \|_{L^2_{x,v}}^2+\sum_{|\alpha|+|\beta|\leq N }(n_{10}+n_{20})\|\partial^{\alpha}_{\beta} f_2^{n+1} \|_{L^2_{x,v}}^2 \cr
&\leq C\mathcal{E}^n(t) + C \mathcal{E}^{n+1}(t) + C\sqrt{\mathcal{E}^n(t)}\sqrt{\mathcal{E}^{n+1}(t)} +C\mathcal{E}^n(t)\sqrt{\mathcal{E}^{n+1}(t)}.
\end{split}
\end{align}
Combining \eqref{est_dt1} and \eqref{est_dt2} yields
\begin{align*}
\frac{1}{2}\frac{d}{dt}\mathcal{E}^{n+1}(t)+(n_{10}+n_{20})\mathcal{E}^{n+1}(t) &\leq C
\mathcal{E}^n(t) + C \mathcal{E}^{n+1}(t) \cr
& \quad + C\sqrt{\mathcal{E}^n(t)}\sqrt{\mathcal{E}^{n+1}(t)} +C\mathcal{E}^n(t)\sqrt{\mathcal{E}^{n+1}(t)}.
\end{align*}
We integrate in time to get
\begin{align}\label{Eps_n+1}
\begin{split}
\mathcal{E}^{n+1}&(t)\leq \mathcal{E}^{n+1}(0)\cr
&+\int_0^t \left(C\mathcal{E}^n(s) +C \mathcal{E}^{n+1}(s)+ C\sqrt{\mathcal{E}^n(t)}\sqrt{\mathcal{E}^{n+1}(t)} +C\mathcal{E}^n(t)\sqrt{\mathcal{E}^{n+1}(t)}\right)ds.
\end{split}
\end{align}
We now apply an induction argument. We have $\mathcal{E}^0(0)<\frac{M_0}{2}$ from the assumption. Assume we have
\begin{align*}
\sup_{0\leq t \leq T_*}\mathcal{E}^{n}(t) \leq M_0, \quad \mathcal{E}^{n+1}(0)\leq M_0/2.
\end{align*}
Then, from \eqref{Eps_n+1}, we see that
\begin{align*}
\sup_{0\leq t \leq T_*}\mathcal{E}^{n+1}(t)&\leq \frac{M_0}{2}+CT_*M_0+CT_*\sup_{0\leq t \leq T_*}\mathcal{E}^{n+1}(t) \cr
&\quad + CT_*\sqrt{M_0}\sqrt{\sup_{0\leq t \leq T_*}\mathcal{E}^{n+1}(t)} +CT_*M_0\sqrt{\sup_{0\leq t \leq T_*}\mathcal{E}^{n+1}(t)}.
\end{align*}
By using Young's inequality, we have
\begin{align*}
(1-3CT_*)\sup_{0\leq t \leq T_*}\mathcal{E}_1^{n+1}(t)&\leq \frac{M_0}{2} +
2 C T_*M_0 +CT_*M_0^2 .
\end{align*}
Therefore, for sufficiently small $T_*$ and $M_0>0$, we can derive
\begin{align*}
\sup_{0\leq t \leq T_*}\mathcal{E}^{n+1}(t)&\leq M_0 .
\end{align*}
This completes the proof.
\end{proof}
\section{Coercivity estimate}
We write the macroscopic part $P(f_1,f_2)$ of the distribution function $(f_1,f_2)$ as
\begin{align*}
P(f_1,f_2) &= a_1(x,t)\left(\sqrt{\mu_1},0\right)+a_2(x,t)\left(0,\sqrt{\mu_2}\right) +b(x,t)\cdot v \left(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2} \right) \cr
&\quad + c(x,t)|v|^2\left(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2}\right),
\end{align*}
where
\begin{align}\label{abc}
\begin{split}
a_k(x,t) &= \frac{1}{n_{k0}} \int_{\mathbb{R}^3} f_k\sqrt{\mu_k} dv \cr
&\quad- \frac{1}{2n_{10}+2n_{20}}\left(\int_{\mathbb{R}^3} f_1(m_1|v|^2-3)\sqrt{\mu_1} dv+\int_{\mathbb{R}^3} f_2(m_2|v|^2-3)\sqrt{\mu_2} dv\right), \cr
b(x,t) &= \frac{1}{m_1n_{10}+m_2n_{20}}\left(\int_{\mathbb{R}^3} f_1m_1v\sqrt{\mu_1} dv +\int_{\mathbb{R}^3} f_2m_2v\sqrt{\mu_2} dv\right), \cr
c(x,t) &= \frac{1}{6n_{10}+6n_{20}}\left(\int_{\mathbb{R}^3} f_1(m_1|v|^2-3)\sqrt{\mu_1} dv+\int_{\mathbb{R}^3} f_2(m_2|v|^2-3)\sqrt{\mu_2} dv\right),
\end{split}
\end{align}
for $k=1,2$.
We substitute
\[
(f_1,f_2)=(I-P)(f_1,f_2)+P(f_1,f_2),
\]
into \eqref{pertff} to get
\begin{align}\label{split}
\begin{split}
\{\partial_t +v\cdot \nabla_x\} P(f_1,f_2)&=-\{\partial_t +v\cdot \nabla_x-L\} (I-P)(f_1,f_2) \cr
&\quad+(\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2),\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2)).
\end{split}
\end{align}
We write L.H.S. of \eqref{split} in the following form:
\begin{multline*}
\bigg\{\left(\partial_ta_1+v\cdot \nabla_x a_1\right)(\sqrt{\mu_1},0)+\left(\partial_ta_2+v\cdot \nabla_x a_2\right)(0,\sqrt{\mu_2}) \cr
+v\cdot\partial_t b(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2}) + \sum_{1\leq i<j \leq 3}v_iv_j(\partial_{x_i}b_j+\partial_{x_j}b_i)(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2}) \cr
+ \sum_{1\leq i \leq 3 }(\partial_{x_i}b_i+\partial_tc)v_i^2(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2}) + |v|^2v\cdot \nabla_x c (m_1\sqrt{\mu_1},m_2\sqrt{\mu_2}) \bigg\},
\end{multline*}
as a linear expansion with respect to the following $17$ basis:
\begin{align}\label{basis17}
\begin{split}
\{(\sqrt{\mu_1},0),(0,\sqrt{\mu_2}),v(\sqrt{\mu_1},0)&,v(0,\sqrt{\mu_2}),\cr
&v_iv_j(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2}),v|v|^2(m_1\sqrt{\mu_1},m_2\sqrt{\mu_2})\}.
\end{split}
\end{align}
Therefore, comparing both sides of \eqref{split}, we obtain the following system:
\begin{align*}
\partial_t a_1 &= l_{a1}+h_{a1}, \cr
\partial_t a_2 &= l_{a2}+h_{a2}, \cr
\partial_{x_i}a_1 + m_1\partial_t b_i &= l_{b1i} + h_{b1i}, \cr
\partial_{x_i}a_2 + m_2\partial_t b_i &= l_{b2i} + h_{b2i}, \cr
\partial_{x_i}b_j+\partial_{x_j}b_i&= l_{bbi}+ h_{bbi}, \quad (i\neq j ) \cr
\partial_{x_i}b_i+\partial_tc &= l_{bci}+ h_{bci}, \cr
\partial_{x_i}c &= l_{ci} + h_{ci},
\end{align*}
where $(l_{a1},l_{a2},l_{b1i},l_{b2i},l_{bbi},l_{bci},l_{ci})$, and $(h_{a1},h_{a2},h_{b1i},h_{b2i},h_{bbi},h_{bci},h_{ci})$ are the coefficients corresponding to the expansion of $l$ and $h$:
\begin{align*}
&l(f_1,f_2)= -\{\partial_t +v\cdot \nabla_x-L\} (I-P)(f_1,f_2), \cr
&h(f_1,f_2)=(\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2),\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2)),
\end{align*}
with respect to \eqref{basis17}. For brevity, we denote
\begin{align*}
\tilde{l} &= l_{a1}+l_{a2}+\sum_{i=1}^3\left(l_{b1i}+l_{b2i}+l_{bbi}+l_{bci}+l_{ci}\right) \cr
\tilde{h} &= h_{a1}+h_{a2}+\sum_{i=1}^3\left(h_{b1i}+h_{b2i}+h_{bbi}+h_{bci}+h_{ci}\right).
\end{align*}
\iffalse
\subsection{Coercivity estimate}
Our object is to prove
\begin{align*}
\|P_k\partial^{\alpha}f_k\|_{L^2_{x,v}} &\leq C \left(\|\partial^{\alpha}a_k\|_{L^2_x}+\|\partial^{\alpha}b\|_{L^2_x}+\|\partial^{\alpha}c\|_{L^2_x}\right)
\end{align*}
for $k=1,2$. In this section, we will show that
\begin{multline}
\|\partial^{\alpha}a_1\|_{L^2_x}+\|\partial^{\alpha}b_1\|_{L^2_x}+\|\partial^{\alpha}c_1\|_{L^2_x}+\|\partial^{\alpha}a_2\|_{L^2_x}+\|\partial^{\alpha}b_2\|_{L^2_x}+\|\partial^{\alpha}c_2\|_{L^2_x} \cr
\leq C \sum_{|\alpha|\leq N-1}\left( \|\partial^{\alpha}\tilde{l}^1\|_{L_{x}^2} + \|\partial^{\alpha}\tilde{h}^1\|_{L_{x}^2}\right) +C \sum_{|\alpha|\leq N-1}\left( \|\partial^{\alpha}\tilde{l}^2\|_{L_{x}^2} + \|\partial^{\alpha}\tilde{h}^2\|_{L_{x}^2}\right)
\end{multline}
for $|\alpha|=N$. Main point of this estimate is that $l$ and $h$ have one less derivative than $a$, $b$ and $c$. In addition, it include the estimate of $a$, $b$ and $c$ without derivative. When they have no derivative, the main method is the Poincar\'{e} inequality. The conservation laws of the total mass, momentum and energy
\begin{align*}
\frac{d}{dt}\int_{\mathbb{T}^3\times\mathbb{R}^3} \left(\begin{array}{c}1 \cr v \cr |v|^2 \end{array}\right) F_1 dvdx = 0
\end{align*}
can be changed to
\begin{align*}
\frac{d}{dt}\int_{\mathbb{T}^3\times\mathbb{R}^3} \left(\begin{array}{c}1 \cr v \cr |v|^2 \end{array}\right)\sqrt{\mu_1}f_1 dvdx = 0
\end{align*}
This is also equal to
\begin{align*}
\int_{\mathbb{T}^3} a(x,t) dx =\int_{\mathbb{T}^3} b(x,t) dx =\int_{\mathbb{T}^3} c(x,t) dx = 0 .
\end{align*}
This inequality contributes to the estimate of $a$, $b$, and $c$ when they have no derivative.
\newline
Before we state the Lemma, we first denote the following notation:
\[\alpha = (\alpha_0,\alpha_1,\alpha_2,\alpha_3) \]
\[\partial^\alpha= \partial_t^{\alpha_0}\partial_{x_1}^{\alpha_1}\partial_{x_2}^{\alpha_2}\partial_{x_3}^{\alpha_3}\]
Note that $\partial^{\alpha}$ include time derivative and spatial derivative.
\fi
\begin{lemma} We have
\begin{align*}
\int_{\mathbb{T}^3}a_1(x,t)dx=\int_{\mathbb{T}^3}a_2(x,t)dx=\int_{\mathbb{T}^3}b(x,t)dx=\int_{\mathbb{T}^3}c(x,t)dx=0.
\end{align*}
\end{lemma}
\begin{proof}
This follows from the conservation laws \eqref{conservf} and the definition of $a_1$, $a_2$, $b$, and $c$ in \eqref{abc}.
\end{proof}
\begin{lemma}\emph{\cite{Guo VMB}}\label{abc esti} Let $ 0\leq |\alpha| \leq N $ with $N\geq 3 $, then we have
\begin{multline*}
\| \partial^{\alpha}a_1 \|_{L^2_x}+\| \partial^{\alpha}a_2 \|_{L^2_x}+\|\partial^{\alpha}b \|_{L^2_x}+\|\partial^{\alpha}c \|_{L^2_x}
\leq \sum_{|\alpha|\leq N-1} \left(\|\partial^{\alpha}\tilde{l} \|_{L^2_x}+\|\partial^{\alpha}\tilde{h} \|_{L^2_x}\right).
\end{multline*}
\end{lemma}
\begin{proof}
The proof can be found in \cite[page 620, Proof of Theorem 3]{Guo VMB}. We omit it.
\end{proof}
\begin{lemma}\label{lh} For sufficiently small energy norm $\mathcal{E}(t)$, we have
\begin{align*}
&(1) \ \sum_{|\alpha|\leq N-1} \|\partial^{\alpha}\tilde{l}\|_{L^2_{x,v}} \leq C \sum_{|\alpha|\leq N}\|(I-P)\partial^{\alpha}(f_1,f_2)\|_{L^2_{x,v}}, \cr
&(2) \ \sum_{|\alpha|\leq N} \|\partial^{\alpha}\tilde{h}\|_{L^2_{x,v}} \leq C\sqrt{M} \sum_{|\alpha|\leq N}\|\partial^{\alpha}(f_1,f_2)\|_{L^2_{x,v}}.
\end{align*}
\end{lemma}
\begin{proof}
(1) The proof can be found in \cite[page 616, Lemma 7]{Guo VMB}. We omit it. \newline
(2) Let us define $\{e_i^*\}_{i=1}^{17}$ be the orthonormal basis corresponding to the basis \eqref{basis17}. Then we can write
\begin{align*}
e_i^* = \sum_{j=1}^{17} C_{ij}e_j, \qquad h(f_1,f_2) =\sum_{i=1}^{17} \langle h, e_i^* \rangle_{L^2_v} e_i^*,
\end{align*}
so that
\begin{align*}
\langle h, e_n^* \rangle_{L^2_v} = \sum_{1\leq i,j \leq 17}C_{ij}C_{ni}\langle h, e_i^* \rangle_{L^2_v},
\end{align*}
for $n=1,\cdots,17$. For the estimate of $h$, we compute
\begin{align*}
\bigg\| \int \partial^{\alpha}h(f_1,f_2) e_i^* dv \bigg\|_{L^2_x} &\leq \bigg\| \int \partial^{\alpha}\Gamma_{11}(f_1) (|v|^k\sqrt{\mu_1}) dv \bigg\|_{L^2_x} + \bigg\| \int \partial^{\alpha}\Gamma_{12}(f_1,f_2) (|v|^k\sqrt{\mu_1}) dv \bigg\|_{L^2_x} \cr
&+ \bigg\| \int \partial^{\alpha}\Gamma_{22}(f_2) (|v|^k\sqrt{\mu_2}) dv \bigg\|_{L^2_x} + \bigg\| \int \partial^{\alpha}\Gamma_{21}(f_1,f_2) (|v|^k\sqrt{\mu_2}) dv \bigg\|_{L^2_x},
\end{align*}
for $k=0,1,2,3$. For sufficiently small $\mathcal{E}(t)$, by Lemma \ref{nonlin esti1}, we have
\begin{align*}
\bigg\| \int \partial^{\alpha}\Gamma_{mm}(f_m) |v|^k\sqrt{\mu_m} dv \bigg\|_{L^2_x} \leq C\sum_{|\alpha_1|+|\alpha_2|\leq |\alpha|}\bigg\| \|\partial^{\alpha_1}f_m\|_{L^2_v}\|\partial^{\alpha_2}f_m\|_{L^2_v}\bigg\|_{L^2_x}.
\end{align*}
Similarly, we have from Lemma \ref{nonlin esti}
\begin{align*}
\bigg\| \int \partial^{\alpha}\Gamma_{lm}(f_l,f_m) |v|^k\sqrt{\mu_l} dv \bigg\|_{L^2_x}
\leq C\sum_{|\alpha_1|+|\alpha_2|\leq |\alpha|}\bigg\| \|\partial^{\alpha_1}(f_l,f_m)\|_{L^2_v}\|\partial^{\alpha_2}(f_l,f_m)\|_{L^2_v}\bigg\|_{L^2_x},
\end{align*}
for $l\neq m$. Without loss of generality, we assume that $|\alpha_1|\leq|\alpha_2|$ and apply the Sobolev embedding $H^2 \subset\subset L^{\infty}$ to obtain
\begin{align*}
\sum_{|\alpha|\leq N} \|\partial^{\alpha}\tilde{h}\|_{L^2_{x,v}} &\leq C \sum_{|\alpha_1|\leq |\alpha_2|}\sup_{x\in\mathbb{T}^3}\|\partial^{\alpha_1}(f_1,f_2)\|_{L^2_v} \sum_{|\alpha_2|\leq N} \|\partial^{\alpha_2}(f_1,f_2)\|_{L^2_{x,v}}\cr
&\leq C\sqrt{\mathcal{E}(t)}\sum_{|\alpha|\leq N}\|\partial^{\alpha}(f_1,f_2)\|_{L^2_{x,v}},
\end{align*}
which gives desired result.
\end{proof}
We are now ready to derive the full coercivity estimate. By Lemma \ref{abc esti}, we have
\begin{align*}
\sum_{|\alpha|\leq N} \| \partial^{\alpha} P(f_1,f_2) \|_{L^2_{x,v}}^2 &\leq \sum_{|\alpha|\leq N}\left( \| \partial^{\alpha}a_1 \|_{L^2_x}^2+\| \partial^{\alpha}a_2 \|_{L^2_x}^2+\|\partial^{\alpha}b \|_{L^2_x}^2+\|\partial^{\alpha}c \|_{L^2_x}^2\right) \cr
&\leq \sum_{|\alpha|\leq N-1} \left(\|\partial^{\alpha}\tilde{l} \|_{L^2_x}^2+\|\partial^{\alpha}\tilde{h} \|_{L^2_x}^2\right).
\end{align*}
We then apply Lemma \ref{lh} to get
\begin{align*}
\sum_{|\alpha|\leq N} \| &\partial^{\alpha} P(f_1,f_2) \|_{L^2_{x,v}}^2 \cr
&\leq C \sum_{|\alpha|\leq N}\|(I-P)\partial^{\alpha}(f_1,f_2)\|_{L^2_{x,v}}^2 + C\sqrt{M} \sum_{|\alpha|\leq N}\|\partial^{\alpha}(f_1,f_2)\|_{L^2_{x,v}}^2.
\end{align*}
Adding $\displaystyle{\sum_{|\alpha|\leq N}\|(I-P)\partial^{\alpha}(f_1,f_2)\|_{L^2_x}^2}$ on each side, we obtain
\begin{align*}
\sum_{|\alpha|\leq N} \| \partial^{\alpha}(f_1,f_2) \|_{L^2_{x,v}}^2 \leq \frac{C+1}{1-C\sqrt{M}} \sum_{|\alpha|\leq N}\|(I-P)(\partial^{\alpha}(f_1,f_2))\|_{L^2_{x,v}}^2.
\end{align*}
Combining it with the estimate in Proposition \ref{dissipation}, we derive the following full coercivity estimate
\begin{align}\label{full coer}
\langle L\partial^{\alpha}(f_1,f_2),\partial^{\alpha}(f_1,f_2)\rangle_{L^2_{x,v}} \leq - \eta \min\left\{(1-\delta),(1-\omega) \right\}\sum_{|\alpha|\leq N} \| \partial^{\alpha}(f_1,f_2) \|_{L^2_{x,v}}^2,
\end{align}
when $\mathcal{E}(t)$ is sufficiently small.
\section{Global existence}
In this section, we extend the local-in-time solution to the global one by establishing a uniform energy estimate. Let $(f_1,f_2)$ be the classical local-in-time solution constructed in Theorem \ref{theo_loc_ex}.
We take $\partial^{\alpha}$ on \eqref{pertf1} and take inner product with $\partial^{\alpha}f_1$ in $L^2_{x,v}$ to have
\begin{align}\label{f11}
\begin{split}
\frac{1}{2}\frac{d}{dt} \| \partial^{\alpha}f_1 \|_{L^2_{x,v}}^2&=\langle \partial^{\alpha}L_{11}(f_1),\partial^{\alpha}f_1 \rangle_{L^2_{x,v}}+ \langle \partial^{\alpha}L_{12}(f_1,f_2),\partial^{\alpha}f_1 \rangle_{L^2_{x,v}} \cr
&\quad +\langle \partial^{\alpha}f_1,\partial^{\alpha}(\Gamma_{11}+\Gamma_{12}) \rangle_{L^2_{x,v}} .
\end{split}
\end{align}
Similarly, we get from \eqref{pertf2} that
\begin{align}\label{f22}
\begin{split}
\frac{1}{2}\frac{d}{dt} \| \partial^{\alpha}f_2 \|_{L^2_{x,v}}^2&=\langle \partial^{\alpha}L_{22}(f_2),\partial^{\alpha}f_2 \rangle_{L^2_{x,v}}+ \langle \partial^{\alpha}L_{21}(f_1,f_2),\partial^{\alpha}f_2 \rangle_{L^2_{x,v}} \cr
&\quad +\langle \partial^{\alpha}f_2,\partial^{\alpha}(\Gamma_{22}+\Gamma_{21}) \rangle_{L^2_{x,v}} .
\end{split}
\end{align}
Combining \eqref{f11} and \eqref{f22} yields
\begin{align*}
\sum_{k=1,2} \frac{1}{2}\frac{d}{dt} \|\partial^{\alpha} f_k \|_{L^2_{x,v}}^2&\leq \langle L\partial^{\alpha}(f_1,f_2),\partial^{\alpha}(f_1,f_2)\rangle_{L^2_{x,v}} \cr
&\quad+\langle \partial^{\alpha} f_1,\partial^{\alpha} (\Gamma_{11}+ \Gamma_{12}) \rangle_{L^2_{x,v}} +\langle \partial^{\alpha} f_2,\partial^{\alpha}(\Gamma_{22}+\Gamma_{21}) \rangle_{L^2_{x,v}}.
\end{align*}
Then the first term of the R.H.S is controlled by the full coercivity estimate \eqref{full coer}, and the nonlinear terms on the second line are estimated by Lemma \ref{nonlin esti1} and Lemma \ref{nonlin esti}:
\begin{multline*}
\sum_{|\alpha|\leq N}\sum_{k=1,2}\left(\frac{1}{2}\frac{d}{dt} \|\partial^{\alpha} f_k \|_{L^2_{x,v}}^2+\eta \min\left\{(1-\delta),(1-\omega) \right\} \| \partial^{\alpha}f_k \|_{L^2_{x,v}}^2\right) \cr
\leq C_0\sqrt{\mathcal{E}_{N_1,0}(t)}\sum_{|\alpha|\leq N}\|\partial^{\alpha} (f_1,f_2) \|_{L^2_{x,v}}^2.
\end{multline*}
For $M_0$ satisfying Theorem \ref{theo_loc_ex} and \eqref{full coer}, we define
\begin{align*}
M = \left\{\frac{M_0}{2}, \frac{\eta^2 \min\left\{(1-\delta)^2,(1-\omega)^2 \right\}}{4C_0^2}\right\}, \qquad T= \sup_{t\in\mathbb{R}^+}\{t ~|~ \mathcal{E}_{N_1,0}(t) \leq 2M \} >0.
\end{align*}
We restrict our initial data to satisfy the following energy bound:
\begin{align*}
\mathcal{E}_{N_1,0}(0) \leq M \leq 2M_0.
\end{align*}
Once we define
\begin{align*}
y(t) =\sum_{|\alpha|\leq N}\sum_{k=1,2}\|\partial^{\alpha} f_k \|_{L^2_{x,v}}^2,
\end{align*}
then $y(t)$ satisfies
\begin{align*}
y'(t) + 2\eta \min\left\{(1-\delta),(1-\omega) \right\}y(t) &\leq 2C_0\sqrt{\mathcal{E}_{N_1,0}(t)} y(t) \cr
&\leq \eta \min\left\{(1-\delta),(1-\omega) \right\}y(t).
\end{align*}
Thus we obtain
\begin{align*}
y(t) \leq e^{- \eta \min\left\{(1-\delta),(1-\omega) \right\}t}y(0) \leq y(0) \leq M < 2M,
\end{align*}
and which is possible only when $T=\infty$. Note that this also gives
\begin{align*}
\sum_{|\alpha|\leq N}\|\partial^{\alpha} (f_1(t),f_2(t)) \|_{L^2_{x,v}}^2 \leq e^{-\eta \min\left\{(1-\delta),(1-\omega) \right\}t}\sum_{|\alpha|\leq N}\|\partial^{\alpha} (f_1(0),f_2(0)) \|_{L^2_{x,v}}^2.
\end{align*}
Now we consider the general case of $f$ having momentum derivatives. Taking $\partial^{\alpha}_{\beta}$ on \eqref{pertf1} and \eqref{pertf2} and applying an inner product with $\partial^{\alpha}_{\beta} f_1$ and $\partial^{\alpha}_{\beta} f_2$, respectively, we have
\begin{align}\label{d1}
\begin{split}
\frac{1}{2}\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2&+(n_{10}+n_{20})\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2 = - \sum_{i=1}^3 \langle \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1, \partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} \cr
&\quad + (n_{10}+n_{20})
\langle \partial_{\beta}P_1\partial^{\alpha}f_1,\partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}}
+\langle \partial^{\alpha}_{\beta}L_{12}^2(f_1,f_2), \partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} \cr
&\quad +\langle \partial^{\alpha}_{\beta}(\Gamma_{11}(f_1)+\Gamma_{12}(f_1,f_2)),\partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}},
\end{split}
\end{align}
and
\begin{align}\label{d2}
\begin{split}
\frac{1}{2}\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2&+(n_{10}+n_{20})\|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2 = - \sum_{i=1}^3 \langle \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_2, \partial^{\alpha}_{\beta} f_2 \rangle_{L^2_{x,v}} \cr
&\quad + (n_{10}+n_{20})
\langle \partial_{\beta}P_2\partial^{\alpha}f_2,\partial^{\alpha}_{\beta} f_2 \rangle_{L^2_{x,v}}
+\langle \partial^{\alpha}_{\beta}L_{21}^2(f_1,f_2), \partial^{\alpha}_{\beta} f_2 \rangle_{L^2_{x,v}} \cr
&\quad +\langle \partial^{\alpha}_{\beta}(\Gamma_{22}(f_2)+\Gamma_{21}(f_1,f_2)),\partial^{\alpha}_{\beta} f_2 \rangle_{L^2_{x,v}}.
\end{split}
\end{align}
\iffalse
Using the Young's inequality, we have
\begin{align*}
\sum_{i=1}^3 \langle \partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1, \partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} &\leq \sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1\|_{L^2_{x,v}}\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}} \cr
& \leq \frac{1}{2\epsilon}\sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1\|_{L^2_{x,v}}^2 +\frac{3 \epsilon }{2} \|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2,
\end{align*}
for an $\varepsilon >0$ which we will choose later.
Similarly we have
\begin{align*}
\langle \partial_{\beta}P_1\partial^{\alpha}f_1,\partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} &\leq \frac{C}{2\epsilon}\| \partial^{\alpha}f_1\|_{L^2_{x,v}}^2 + \frac{\epsilon}{2} \| \partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2 .
\end{align*}
Using \eqref{AL1 esti}, and the Young's inequality, we have
\begin{align*}
\langle \partial^{\alpha}_{\beta}L_{12}(f_1,f_2), \partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} &\leq \left(\| \partial^{\alpha}f_1\|_{L^2_{x,v}} + \| \partial^{\alpha}f_2\|_{L^2_{x,v}}\right) \| \partial^{\alpha}_{\beta}f_1\|_{L^2_{x,v}} \cr
&\leq \frac{1}{\epsilon}\left(\| \partial^{\alpha}f_1\|_{L^2_{x,v}}^2 + \| \partial^{\alpha}f_2\|_{L^2_{x,v}}^2\right) + \frac{\epsilon}{2}\| \partial^{\alpha}_{\beta}f_1\|_{L^2_{x,v}}^2.
\end{align*}
Appplying Lemma \ref{nonlin esti}, and \eqref{gamma12 esti} we have
\begin{align*}
\langle \partial^{\alpha}_{\beta}\Gamma_{12}(f_1,f_2),\partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} \leq C \mathcal{E}^{\frac{3}{2}}(t).
\end{align*}
Similarly, the first term of the third line can be estimated as follows:
\begin{align*}
\langle \partial^{\alpha}_{\beta}\Gamma_{11}(f_1),\partial^{\alpha}_{\beta} f_1 \rangle_{L^2_{x,v}} \leq C\bigg(\sum_{|\alpha_1|\leq |\alpha|} \|\partial^{\alpha_1}f_1\|_{L^2_{x,v}}\bigg)^2 \| \partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}} \leq C \mathcal{E}^{\frac{3}{2}}(t).
\end{align*}
Combining above estimates yields
\begin{align*}
\frac{1}{2}\frac{d}{dt}&\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2+(n_{10}+n_{20})\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2 \leq \frac{1}{2\epsilon}\sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1\|_{L^2_{x,v}}^2 +\frac{3 \epsilon }{2} \|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2 \cr
&\quad +
\frac{C}{2\epsilon}\| \partial^{\alpha}f_1\|_{L^2_{x,v}}^2 + \frac{\epsilon}{2} \| \partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2
+\frac{1}{\epsilon}\left(\| \partial^{\alpha}f_1\|_{L^2_{x,v}}^2 + \| \partial^{\alpha}f_2\|_{L^2_{x,v}}^2\right) + \frac{\epsilon}{2}\| \partial^{\alpha}_{\beta}f_1\|_{L^2_{x,v}}^2 \cr
&\quad +C \mathcal{E}^{\frac{3}{2}}(t).
\end{align*}
By the same way, we can have the following estimate for $f_2$:
\begin{align*}
\frac{1}{2}\frac{d}{dt}&\|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2+(n_{10}+n_{20})\|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2 \leq \frac{1}{2\epsilon}\sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_2\|_{L^2_{x,v}}^2 +\frac{3 \epsilon }{2} \|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2 \cr
&\quad +
\frac{C}{2\epsilon}\| \partial^{\alpha}f_2\|_{L^2_{x,v}}^2 + \frac{\epsilon}{2} \| \partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2
+\frac{1}{\epsilon}\left(\| \partial^{\alpha}f_1\|_{L^2_{x,v}}^2 + \| \partial^{\alpha}f_2\|_{L^2_{x,v}}^2\right) + \frac{\epsilon}{2}\| \partial^{\alpha}_{\beta}f_2\|_{L^2_{x,v}}^2 \cr
&\quad +C \mathcal{E}^{\frac{3}{2}}(t).
\end{align*}
Combining above two inequalities yields
\begin{align*}
\frac{1}{2}\frac{d}{dt}\left(\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2+\|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2\right)+(n_{10}+n_{20}-\frac{3 \epsilon}{2}-\frac{\epsilon}{2}-\frac{\epsilon}{2})\left(\|\partial^{\alpha}_{\beta} f_1 \|_{L^2_{x,v}}^2+\|\partial^{\alpha}_{\beta} f_2 \|_{L^2_{x,v}}^2\right) \cr
\leq \frac{1}{2\epsilon}\sum_{i=1}^3\left(\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_1\|_{L^2_{x,v}}^2+\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_2\|_{L^2_{x,v}}^2 \right) \cr
+\left(\frac{C}{2\epsilon}+\frac{1}{\epsilon}\right)\left(\| \partial^{\alpha}f_1\|_{L^2_{x,v}}^2 + \| \partial^{\alpha}f_2\|_{L^2_{x,v}}^2\right) +C \mathcal{E}^{\frac{3}{2}}(t).
\end{align*}
\fi
Combining \eqref{d1} and \eqref{d2}, and applying the H\"{o}lder inequality and Young's inequality, we can obtain
\begin{multline*}
\sum_{k=1,2}\left(\frac{1}{2}\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_k \|_{L^2_{x,v}}^2 + (n_{10}+n_{20}-2\epsilon)\|\partial^{\alpha}_{\beta} f_k \|_{L^2_{x,v}}^2\right) \cr
\leq \frac{1}{2\epsilon}\sum_{k=1,2}\sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_k\|_{L^2_{x,v}}^2 +\frac{C}{2\epsilon}\sum_{k=1,2}\| \partial^{\alpha}f_k\|_{L^2_{x,v}}^2 +C \mathcal{E}_{N_1,|\beta|}^{\frac{3}{2}}(t),
\end{multline*}
for some positive constant $\epsilon$ satisfying $(n_{10}+n_{20})/2>\epsilon>0$. We sum this over $|\beta|=m+1$ and multiply both sides with $\epsilon \eta_m$:
\begin{multline*}
\sum_{|\beta|=m+1}\left[\sum_{k=1,2}\left(\frac{\epsilon \eta_m}{2}\frac{d}{dt}\|\partial^{\alpha}_{\beta} f_k \|_{L^2_{x,v}}^2 + \epsilon \eta_m(n_{10}+n_{20}-2\epsilon)\|\partial^{\alpha}_{\beta} f_k \|_{L^2_{x,v}}^2\right)\right] \cr
\leq \sum_{|\beta|=m+1}\left[\frac{\eta_m}{2}\sum_{k=1,2}\sum_{i=1}^3\|\partial^{\alpha+\bar{k}_i}_{\beta-k_i}f_k\|_{L^2_{x,v}}^2 +\frac{C\eta_m}{2}\sum_{k=1,2}\| \partial^{\alpha}f_k\|_{L^2_{x,v}}^2 +C \mathcal{E}_{N_1,|\beta|}^{\frac{3}{2}}(t)\right].
\end{multline*}
Combining the previous cases $|\beta| \leq m$, the R.H.S of the inequality can be bounded by the energy $\mathcal{E}_{N_1,|\beta|}$ with $|\beta|\leq m$ and $\mathcal{E}_{N_1,0}$. Thus, we can conclude from induction that
\begin{align*}
\sum_{\substack{|\alpha|+|\beta|\leq N \cr |\beta|\leq m+1}}\sum_{k=1,2} \left( C_{m+1}\frac{d}{dt}\| \partial^{\alpha}_{\beta}f_k \|_{L^2_{x,v}}^2+ \eta_{m+1}\| \partial^{\alpha}_{\beta}f_k \|_{L^2_{x,v}}^2\right) \leq C_{m+1}^*\mathcal{E}_{N_1,|\beta|}^{\frac{3}{2}}(t).
\end{align*}
Applying the same continuity argument as to when $\beta=0$, we can construct the global-in-time classical solution. We mention that when $|\beta|=0$, the parameter $\eta_0$ depends on $1-\delta$ and $1-\omega$, and $C_0=1/2$. But when $|\beta|\geq1$, both $C_{m+1}$ and $\eta_{m+1}$ depend on the parameter $\eta_m$. That is why we cannot extract a decay rate depending explicitly on the parameter $\delta$ and $\omega$ when the velocity derivatives are involved.
For the uniqueness of the solution and $L^2$ stability, we can follow the standard arguments in \cite{Guo whole,Guo VMB,Guo VPB,Yun1}. This completes the proof.\newline\newline
\begin{center}
{\bf Acknowledgement:}
\end{center}
G.-C. Bae is supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. 2021R1C1C2094843).
C. Klingenberg is supported in part by the W\"urzburg University performance-orientated fund LOM 2021.
M. Pirner is supported by the Alexander von Humboldt Foundation
S.-B. Yun is supported by Samsung Science and Technology Foundation under Project Number SSTF-BA1801-02.
|
\section{Introduction}\label{sec:intro}
Entropic repulsion in low temperature $(2+1)$\textsc{d} crystals above a hard wall has been the subject of extensive study in statistical physics. Whereas in the absence of a wall, the surface of the crystal would typically be rigid at height $0$, in the presence of a wall, the surface is propelled in order to increase its entropy (i.e., to allow thermal fluctuations going downward), and becomes rigid at some height level which diverges with the side length $L$ of the box.
A rigorous study of this phenomenon in the $(2+1)$\textsc{d} Solid-On-Solid \abbr{(SOS)} model---a low temperature approximation of the 3\textsc{d} Ising model---dates back to Bricmont, El Mellouki and Fr\"{o}hlich~\cite{BEF86} in 1986, where it was shown that, in the presence of a hard wall at height $0$, the typical height of a site in the bulk is propelled to order $\log L$. Thereafter, a detailed description of the shape of this random surface was obtained by Caputo et al.~\cite{CLMST12,CLMST14,CLMST16}, showing that it typically becomes rigid at a height which is one of two consecutive (explicit) integers, through a sequence of nested level lines each encompassing a $(1-\epsilon)$-fraction of the sites (analogous behavior was later established~\cite{LMS16} for the more general family of $|\nabla\phi|^p$-random surface model, where the \abbr{SOS} model is the case $p=1$).
The level lines near the center sides of the box behave as random walks---a ubiquitous feature of interfaces in low temperature spin systems---albeit with cube-root fluctuations, as their laws are tilted by the entropic repulsion effect. The lower the level line, the higher the reward is for generating spikes going downward, and as such, the tilting effect of the level lines increases exponentially as the height decreases.
Whereas the $2$\textsc{d} Ising model with a pinning potential is known~\cite{IOSV21} to have an interface converging to a Ferrari--Spohn diffusion, the behavior in the \abbr{SOS} model---where there are $H\asymp\log L$ interacting level lines, each constrained not to cross its neighbors and inducing a tilt which is a function of the area it encompasses and its height---is far from being understood (see the review~\cite{IV18} for more information).
In this work, we investigate the limiting law of a line ensemble that was
studied by Caputo, Ioffe and Wachtel~\cite{CIW19a,CIW19b} to model the level lines of the \abbr{SOS} model in the presence of a hard wall: each level line, $X_1,X_2,\ldots$, where $X_1$ is the top one, is tilted by the area below it, with the coefficients of these area tilts increasing geometrically.
For more perspective on this model in the context of other models of Brownian polymers constrained above a barrier, starting from the influential work of Ferrari and Spohn~\cite{FS05} (the model there being equivalent by Girsanov's transformation---cf.~\cite{MZ16}---to a Brownian excursion with an area tilt), see, e.g.,~\cite{CIW19a,ISV15,IVW18} and the references therein.
Define
\[ \An = \{ \underline x\in \mathbb R^n \,:\; x_1 > x_2 > \ldots > x_n > 0\}\,,\]
its closure $\bAn$ and, for a designated interval
\[I=[\ell,r] \qquad (\ell<r\in \mathbb R)\,,\]
let
\[ \Omega_{n}^{I} = \left\{ X\in \mathsf C(I; \mathbb{R}^n)
\,:\; X(t)\in \An\mbox{ for all $t\in I$}\right\}\,.\]
(Here, for $T\subset \mathbb{R}$ and $\mathcal{X}$ a topological space, we denote by $\mathsf C(T;\mathcal{X})$ the space of continuous functions from $T$ to $\mathcal{X}$, equipped with the topology of uniform convergence on compact subsets of $\mathbb{R}$).
Further define the area tilt of $Y\in \mathsf C(I;\mathbb{R})$ to be
\[ {\mathcal A}_I(Y) = \int_I Y(t)\, \d t\,,\]
and, for given tilt parameters $\mathfrak a>0$ and $\mathfrak b>1$ and
endpoints $\underline x=(x_1,\ldots,x_n)\in \An$ and $\underline y=(y_1,\ldots,y_n)\in \An$, the partition function
\begin{equation}\label{eq:partition-func}
Z_{n}^{\underline x,\underline y,I} = \mathbf E_{n}^{\underline x, \underline y,I} \left[\mathbbm{1}_{\Omega_{n}^{I}}\, e^{-\mathfrak a \sum_{i=1}^n \mathfrak b^{i-1} {\mathcal A}_I(X_i(\cdot))}\right]\,,\end{equation}
in which $\mathbf E_n^{\underline x,\underline y,I} = \bigotimes_{i=1}^n \mathbf E_1^{x_i,y_i,I}$ and the expectation $\mathbf E_1^{x,y,I}$ for $I=[\ell,r]$ is w.r.t.\ the (unnormalized) path measures of the Brownian bridge which starts at $x$ at time $\ell$ and ends at $y$ at time~$r$;
that is, the total mass of $\mathbf E_1^{x_i,y_i,I}$ is $\phi_{r-\ell}(y_i-x_i)$,
where hereafter $\|\cdot\|$ stands for the relevant Euclidean norm, with
\begin{equation}\label{def:Gauss}
\phi_v(\underline x):=(2\pi v)^{-k/2} e^{-\|\underline x\|^2/(2v)}
\end{equation}
denoting the density of a centered Gaussian vector
of independent coordinates of variance $v$, whose dimension $k$ is
implicitly given by the argument we use.
(At no point in our analysis will we need to adjust the tilt parameters $(\mathfrak a,\mathfrak b)$, and as such we do not include them in the notation for brevity.)
Let ${\mathcal B}_n={\mathcal B}_{n,I}$ be the Borel $\sigma$-field on $\mathsf C(I,\mathbb R^n)$. (We omit $I$ from the notation when no confusion occurs.) For $\Gamma\in{\mathcal B}_{n,I}$ define
\begin{equation}\label{eq:P-ux-uy-def} \P_{n}^{\underline x,\underline y,I}(\Gamma) := \frac1{Z_{n}^{\underline x,\underline y,I}} \mathbf E_n^{\underline x,\underline y,I}\left[\mathbbm{1}_\Gamma \mathbbm{1}_{\Omega_{n}^I}\, e^{-\mathfrak a \sum_{i=1}^n \mathfrak b^{i-1} {\mathcal A}_I(X_i(\cdot))}\right]\,.
\end{equation}
Consider $I_T=[-T,T]$. Two classes of boundary conditions that are of interest are:
\begin{enumerate}[(a)]
\item \emph{Zero boundary conditions}: fixing both $\underline x$ and $\underline y$ to be zero:
\[ \mu^\mathfrak{0}_{n,T} = \P_{n}^{\underline 0,\underline 0, I_T} \;\]
(more precisely, this is the limit of
$\P_{n}^{\epsilon \underline x,\epsilon \underline y, I_T}$ as $\epsilon \downarrow 0$,
which by stochastic domination exists and is independent of the
fixed $\underline x$, $\underline y$ in $\An$ which one uses).
\item
\emph{Free boundary conditions} with respect to a $\sigma$-finite measure:
averaging $\mathbf E_n^{\underline x,\underline y,I_T}[\cdot]$
over $\underline x,\underline y$ according to a specified
$\sigma$-finite measure $\Theta_n$ on~$\mathbb R^n$:
\[ \mu^\mathfrak{f}_{n,T}(\Gamma) =
\frac1{{\mathcal Z}_{n,T}^{\mathfrak{f}}} \int_{\An} \int_{\An} \mathbf E_n^{\underline x,\underline y,I_T}\left[\mathbbm{1}_\Gamma \mathbbm{1}_{\Omega_n^{I_T}}\, e^{-\mathfrak a \sum_{i=1}^n \mathfrak b^{i-1} {\mathcal A}_{I_T}(X_i(\cdot))}\right] \,\Theta_n(\d \underline x) \Theta_n(\d \underline y) \,,
\]
where
\[ {\mathcal Z}_{n,T}^\mathfrak{f} := \int_{\An} \int_{\An}
\mathbf E_n^{\underline x,\underline y,I_T}\left[ \mathbbm{1}_{\Omega_n^{I_T}}\, e^{-\mathfrak a\sum_{i=1}^n \mathfrak b^{i-1} {\mathcal A}_{I_T}(X_i(\cdot))}\right] \,
\Theta_n(\d \underline x) \Theta_n(\d \underline y) \,.
\]
\end{enumerate}
We refer to such $\mu^{\mathfrak{f}}_{n,T}$ as $\Theta_n$-free boundary conditions, reserving
Leb-free for the special case of Lebesgue $\Theta_n$, considered in \cite{CIW19a,CIW19b}.
Caputo, Ioffe and Wachtel show in~\cite{CIW19a,CIW19b}
that $\mu_{n,T}^\mathfrak{0}$ converges to a limit $\mu_\infty^\mathfrak{0}$ as $n,T\to\infty$
(moreover, they proved that for any fixed $n$, the measures $\mu_{n,T}^\mathfrak{0}$ converge as $T\to\infty$ to a limit $\mu_n^\mathfrak{0}$, which then converges to $\mu_\infty^\mathfrak{0}$ as $n\to\infty$),
and that for any $c>0$, the family of Leb-free distributions
$\{\mu_{n,T}^\mathfrak{f}\}_{n\geq 1, T>c}$ is tight.
In this and subsequent statements, the measure $\mu_\infty^\mathfrak{0}$ is defined on
$\mathsf C(\mathbb{R},\mathbb{R}^{\mathbb N})$, and the convergence is in the sense that for any compact set $\mathcal K\subset \mathbb{R}$, integer $k\in \mathbb N$ and fixed bounded, continuous
function
$f: \mathsf C(\mathcal K;\mathbb{R}^k) \mapsto \mathbb{R}$, we have that
\begin{equation}
\label{def:conv}
\lim_{n,T\to\infty} \int f(X_1,\ldots,X_k) \mu_{n,T}^\mathfrak{0}(\d X)= \int f(X_1,\ldots,X_k)\mu_\infty^\mathfrak{0} (\d X)\,.
\end{equation}
For Leb-free boundary conditions, Caputo, Ioffe and Wachtel
conjectured that $\mu_{n,T}^\mathfrak{f}$ converges as well, and to the same limit
$\mu^\mathfrak{0}_\infty$ as $n,T\to\infty$.
Our main result confirms that when $T\to\infty$ followed by~$n\to\infty$, this holds
more generally, whenever for $n \ge 1$,
\begin{align}
\mathfrak c_n &:= \limsup_{r \to \infty} r^{-1} \log \Theta_n(\An \cap \{x_1 \le r\}) < \infty
\label{eq:exp-growth}
\end{align}
(in particular, note that $\mathfrak c_n = 0$ when $\Theta_n$ is Lebesgue measure).
\begin{theorem}\label{thm:1}
Assuming \eqref{eq:exp-growth}, for
any fixed tilt parameters $\mathfrak a>0$ and $\mathfrak b>1$ and any fixed integer $n$, the measures $\mu_{n,T}^\mathfrak{f}$ and $\mu_{n,T}^\mathfrak{0}$ have the same weak limit as $T\to\infty$. In particular, if we denote by $\mu_\infty^\mathfrak{0}$ the limit of $\mu_{n,T}^\mathfrak{0}$ as $n,T\to\infty$, then
\[ \exists\lim_{n\to\infty} \lim_{T\to\infty} \mu_{n,T}^\mathfrak{f} =\mu_\infty^\mathfrak{0}\,.
\]
\end{theorem}
\begin{remark} Our proof easily extends to allow in
$\mu_{n,T}^\mathfrak{f}$ different measures for $\underline x$ and for $\underline y$ (as long as
both satisfy \eqref{eq:exp-growth}). Note that Theorem \ref{thm:1} is optimal in terms of $\Theta_n$,
as merely having ${\mathcal Z}_{n,T}^\mathfrak{f}$ finite, requires that $\mathfrak a \, T \ge \mathfrak c_n$ (see \eqref{eq:mu-free-1}
and \eqref{eq:psi-bd} at $s=T$). Further, $\sup_n \{\mathfrak c_n\}$ must be finite if aiming to exchange
the order of limits in $n$ and $T$.
\end{remark}
Our proof of Theorem~\ref{thm:1} employs the Markovian structure of the problem. In a first step we introduce a (sub-)Markovian Kernel $K_t$, see~\eqref{eq-kernel}. The key part of the proof is Lemma~\ref{lem:K1-compact}, where we prove that $K_1$ is compact in the appropriate $L^2$ space; the proof of the lemma involves probabilistic arguments. With the lemma, standard contraction arguments, detailed in Section~\ref{subsec-2.1}, yield the exponential decay (in $T$) of the dependence in the boundary conditions. We note that some care is needed here due to the non-compactness of the set of possible boundary conditions, but that non-compactness was already handled in~\cite{CIW19a}.
Many interesting open questions remain, chief among which, perhaps, is describing the
limiting process $X_\infty(\cdot)$ (on, say, the interval $[0,1]$). We refer to~\cite{CIW19a,CIW19b} for a list of such problems
and note in passing that from \eqref{eq:mu-free-4} and
\abbr{PDE} theory, one can verify that
the
limit of $\mu^\mathfrak{f}_{n,T}$ when $T \to \infty$,
is the stationary solution of the
Langevin \abbr{SDE} for invariant probability density $\varphi_1^2$ on $\An$ (where
the Perron-Frobenius eigenvector $\varphi_1$ of $K_1$ is the positive $C^2$-solution of
the elliptic \abbr{PDE}
$\frac{1}{2} \Delta u = (c + \mathfrak a \langle \underline{\mathfrak b},\underline x\rangle) u$
with
Dirichlet boundary
conditions at $\partial \An$ and the largest possible $c
<0$).
\section{Proof of main result}
Fix the tilt parameters $\mathfrak a>0$ and $\mathfrak b>1$, and let $n\geq 1$ be an integer.
Throughout this proof, for $X\in \Omega_n^I$,
we use the abbreviated notation
\[ {\mathcal A}_I(X(\cdot)) := \mathfrak a \sum_{i=1}^n \mathfrak b^{i-1} {\mathcal A}_I(X_i(\cdot))\,.\]
Let $\Gamma\in{\mathcal B}_{n,[0,1]}$, and define
\[ K^\Gamma_1(\underline x,\underline y) = \mathbf E_n^{\underline x,\underline y,[0,1]}\left[\mathbbm{1}_\Gamma \mathbbm{1}_{\Omega_n^{[0,1]}}\, e^{-{\mathcal A}_{[0,1]}(X(\cdot))}\right]\,,
\]
which we view as a linear operator on $L^2(\An)=L^2(\An,{\rm Leb})$:
\[ (K^\Gamma_1 f)(\underline x) = \int_{\An} K^\Gamma_1(\underline x,\underline y) f(\underline y) \,\d \underline y\,.\]
With a slight abuse of notation,
we continue to write $K_1^\Gamma$ also when $\Gamma\in {\mathcal B}_{n,\mathbb{R}}$, in which case we understand that $\Gamma$ was replaced by its restriction to the interval $[0,1]$. With this convention in mind,
we will further be interested in the semigroup
\begin{equation}
\label{eq-kernel}
K^\Gamma_t(\underline x,\underline y) = \mathbf E_n^{\underline x,\underline y,[0,t]}\left[\mathbbm{1}_\Gamma \mathbbm{1}_{\Omega_n^{[0,t]}}\, e^{-{\mathcal A}_{[0,t]}(X(\cdot))}\right]\,.
\end{equation}
When referring to the case $\Gamma=\Omega_n^\mathbb R$ (i.e., the indicator $\mathbbm{1}_\Gamma$ within the expectation in the definition of $K_1^\Gamma$ is omitted), we simply write $K_t$ (with no superscript) in lieu of~$K_t^{\Omega_n^\mathbb R}$, noting that $K_t(\underline x,\underline y)$ is precisely the partition function $Z_n^{\underline x,\underline y,[0,t]}$ from~\eqref{eq:partition-func}.
Observe that $K_t$ is symmetric, in that $K_t(\underline x,\underline y)=K_t(\underline y,\underline x)$,
as well as positivity preserving:
\[ (K_t f)(\underline x) =\int K_t(\underline x,\underline y)f(\underline y)\d\underline y \geq 0\qquad\mbox{whenever}\qquad f \geq 0\,.\]
As $K_t$ is symmetric, and given by a continuous time Markov process with killing, it is positive definite (this follows, e.g., by~\cite[Theorems~1.3.1, Lemma~1.3.2 and Theorem~6.1.1]{Fukushima}, all applied to Example 1.2.3 there).
A key ingredient in the proof will be that $K_1$ is furthermore relatively compact:
\begin{lemma}\label{lem:K1-compact}
For every fixed $n$, the range of the symmetric positive definite operator
\[ (K_1 f)(\underline x) = \int_{\An} \mathbf E_n^{\underline x,\underline y,[0,1]}\left[ \mathbbm{1}_{\Omega_n^{[0,1]}}\, e^{-{\mathcal A}_{[0,1]}(X(\cdot))}\right]f(\underline y)\,\d\underline y
\]
consists of continuous functions on $\An$, and the operator $K_1$ is compact w.r.t.\ $L^2(\An)$.
\end{lemma}
\subsection{Proof of Theorem~\ref{thm:1} modulo Lemma~\ref{lem:K1-compact}}
\label{subsec-2.1}
We consider throughout the convergence over the interval $[0,1]$, the changes
needed for considering other compact sets (as the set $\mathcal K$ in \eqref{def:conv}) are minimal. Aiming
to express the measures $\mu_{n,T}^\mathfrak{0}$ and $\mu_{n,T}^\mathfrak{f}$ in terms of the operator $K_t$, we
define for $s >0$
\[
\psi_{s}(\underline u) = \int K_{s}(\underline u,\underline x)\, \Theta_n(\d \underline x) \,.
\]
Setting $s>0$ large enough so that $\psi_s \in L^2(\An)$,
in view of
the symmetry of $K_{s}$ and the semigroup property, our
goal is then to show that for every $\Gamma\in{\mathcal B}_{n,[0,1]}$, the limit of
\begin{equation}
\label{eq:mu-free-1}
\mu_{n,T}^\mathfrak{f}(\Gamma) = \frac{\iint \psi_s(\underline x) K_{T-s}(\underline x , \underline u)
K^\Gamma_1(\underline u, \underline v) K_{T-1-s} (\underline v, \underline y) \psi_s(\underline y) \, \d \underline u \d \underline v \, \d \underline x \d \underline y} {\iint \psi_s(\underline x) \, K_{2T-2s}(\underline x, \underline y)\, \psi_s(\underline y) \d\underline x \d\underline y}
\end{equation}
as $T\to\infty$ exists and coincides with that of
\begin{equation}
\label{eq:mu-zero-1}
\mu_{n,T}^\mathfrak{0}(\Gamma) = \lim_{\epsilon\downarrow 0}\frac{\iint K_{T}(\epsilon \underline x, \underline u) K_1^\Gamma(\underline u, \underline v) K_{T-1}(\epsilon \underline y, \underline v) \,\d \underline u \d \underline v }{K_{2T}(\epsilon \underline x,\epsilon \underline y)}\,,
\end{equation}
where by~\cite[Lemma~2.2]{CIW19b}, the limit \eqref{eq:mu-zero-1} exists and
is independent of
$\underline x,\underline y$ in $\An$.
By the spectral decomposition theorem, the compact positive definite operator~$K_1$ has a discrete spectrum (except for a possible accumulation point at $0$), with positive eigenvalues $\{\lambda_i\}$
and eigenvectors $\{\varphi_i\}$ that form a complete orthonormal basis of $L^2(\An)$
(see, e.g.,~\cite[Thms.~VI.15 and VI.16]{ReedSimonI}). In particular,
\[
K_1(\underline x,\underline y) = \sum_{i=1}^\infty \lambda_i \varphi_i(\underline x) \varphi_i(\underline y) \quad\mbox{for a complete basis $\{\varphi_i\}_{i\geq 1}$ with}\; \left<\varphi_i,\varphi_j\right>_{L^2(\An)}=\delta_{ij}\,.
\]
With $K_1(\underline x,\underline y)>0$ throughout $\An \times \An$ (e.g., due to parabolic regularity),
by the generalized Perron-Frobenius Theorem (see, e.g., the version of the
Krein-Rutman Theorem given in [9, Thm. XIII.43]), the
top eigenvalue $\lambda_1$ has a one-dimensional eigen-space and we may choose
the continuous function $\varphi_1$ to be strictly positive on $\An$. That is,
\[ \varphi_1 > 0\qquad\mbox{and}\qquad \lambda_1>\lambda_2\geq \lambda_3 \geq \ldots \geq 0\,.\]
Further, for
any $r \ge 1$ and $\underline x,\underline y\in \An$,
\begin{align}
\label{eq-K1K2}
K_{r}(\underline x,\underline y) = \mathbf E_n^{\underline x,\underline y,[0,r]}\left[ \mathbbm{1}_{\Omega_n^{[0,r]}}\, e^{-{\mathcal A}_{[0,r]}(X(\cdot))}\right]
&\leq \mathbf E_n^{\underline x,\underline y,[0,r]}\left[ \mathbbm{1}_{\Omega_n^{[0,r]}}\, e^{-\mathfrak a {\mathcal A}_{[0,r]}(X_1(\cdot))}\right] \nonumber \\
&\le
\mathbf E_1^{x_1,y_1,[0,r]}\left[
e^{-\mathfrak a{\mathcal A}_{[0,r]}(X_1(\cdot))}\right]\nonumber\\
& {\; = \;} \phi_{r}(y_1-x_1) e^{-\mathfrak a \, r \frac{x_1+y_1}2} \mathbb E \big[e^{-\mathfrak a \int_0^{r} B_s \,\d s}\big]
\nonumber\\
&
\le e^{C_r}
e^{-\mathfrak a \, r \frac{x_1+y_1}2},
\end{align}
where $\{B_s, s\in [0,r]\}$ is
the standard Brownian bridge over $[0,r]$ starting and ending at $0$
and $C_r =\frac{\mathfrak a^2}{2} \mathbb E(\int_0^r B_s \d s)^2$
(using in the second line that the total mass of
$\mathbf E_1^{x_i,y_i,[0,r]}$, $i \ge 2$, is at most one, while for the third line recall that
a Brownian bridge between fixed points has the law of the standard bridge
plus a straight line connecting these points). Since $K_{r}(\underline x,\underline y)$ vanishes if either
$\underline x\not\in \An$ or $\underline y\not\in \An$, it follows that
\begin{align}
\label{eq-K2bound}
\iint K_{r}(\underline x,\underline y) \,\d \underline x \d\underline y &\leq e^{C_r} \Big[
\int_0^\infty x_1^{n-1}
e^{-\mathfrak a \, r x_1/2} \,\d x_1 \Big]^2 < \infty \,,
\end{align}
and
\begin{align} \label{eq-K1bound}
\int K_{r} (\underline x,\underline x) \,\d \underline x &\leq e^{C_r} \int_0^\infty x_1^{n-1} e^{-\mathfrak a \, r x_1}\,\d x_1
< \infty \,.
\end{align}
Similarly,
by the symmetry of $K_s$, the semigroup property and \eqref{eq-K1K2},
\begin{equation}\label{eq:psi-bd}
\int \psi_{s}(\underline u)^2\,\d\underline u =
\iint K_{2s}(\underline x,\underline y) \Theta_n(\d \underline x) \Theta_n(\d \underline y) \le
e^{C_{2s}} \Big[ \int_{\An} e^{-\mathfrak a s x_1} \Theta_n(\d \underline x) \Big]^2 <\infty \,,
\end{equation}
provided that $\mathfrak a \, s > \mathfrak c_n$ of \eqref{eq:exp-growth}, in which case we can
decompose
\begin{equation}\label{eq:l2-dec}
\psi_{s} = \sum_{i=1}^\infty \alpha_{i,s} \varphi_i\quad\mbox{where}\quad \alpha_{i,s} := \left<\psi_{s},\varphi_i\right>_{L^2(\An)}\,, \quad \sum_{i=1}^\infty \alpha_{i,s}^2 = \|\psi_{s}\|_2^2
\end{equation}
(hereafter $\|\cdot\|_2$ denotes the
$L^2(\mathbb{R}^n,{\rm Leb})$-norm, using $\|\cdot\|_{L^2(\An)}$ when restricting the domain
to $\An$). Fixing an integer $\ell > \mathfrak c_n/\mathfrak a$, we have from \eqref{eq:psi-bd} that
$\{\psi_s\}_{s \in [\ell,\ell+1]}$ is bounded in $L^2(\An)$, and we split any
$T \ge \ell+1$ as $T=t+s$ for $s
\in [\ell,\ell+1)$ and integer $t \ge 1$,
to get from the decomposition~\eqref{eq:l2-dec} that
\begin{align*}
\int K_{t-1}(\underline u,\underline y) \psi_{s} (\underline y) \,\d\underline y
= \sum_{i=1}^\infty \lambda_i^{t-1}\alpha_{i,s} \varphi_i(\underline u) \,.
\end{align*}
Similarly, for $t\geq 1$,
\begin{align*}
\iint K_{t}(\underline x,\underline y) \psi_{s}(\underline x)\psi_{s}(\underline y)\,\d\underline x\d\underline y
= \sum_{i=1}^\infty \lambda_i^{t}\alpha_{i,s}^2 := c_{t,s}
\,.\end{align*}
Hence,~\eqref{eq:mu-free-1} translates for $s=\ell+\{T\}$ and $t=T-s$, into
\begin{align} \mu_{n,T}^\mathfrak{f}(\Gamma) &= \frac{1}{c_{2t,s}} \iint \sum_{i=1}^\infty\sum_{j=1}^\infty
\alpha_{i,s} \alpha_{j,s} \lambda_i^{t} \lambda_j^{t-1} \varphi_i(\underline u) \, K^\Gamma_1(\underline u,\underline v)\, \varphi_j(\underline v) \,\d\underline u\d\underline v \,.
\label{eq:mu-free-2}
\end{align}
Looking at $K^\Gamma_1$ and arguing as we did for $K_1$, we see that for any $\underline u\in\mathbb R^n$,
\[ \int K^\Gamma_1(\underline u,\underline v)^2\, \d\underline v \leq \int K_1(\underline u,\underline v)^2\,\d\underline v = K_2(\underline u,\underline u) < \infty \,,\]
where the equality holds by the symmetry of $K_1$ and the definition of $K_t$ and the last inequality by~\eqref{eq-K1K2}.
In other words, $K_1^\Gamma(\underline u,\cdot)\in L^2(\An)$ for every $\underline u\in\mathbb R^n$.
Moreover, by
\eqref{eq-K1bound}
we have that
\[ \iint K^\Gamma_1(\underline u,\underline v)^2 \,\d\underline u\d\underline v \leq \iint K_1(\underline u,\underline v)^2\,\d\underline u\d\underline v =\int K_2(\underline u,\underline u) \d\underline u<\infty\]
and it follows that
\[ K_1^\Gamma \in L^2(\red{\An \times \An}) \,.\]
A complete orthonormal system $\{\varphi_i\}$ w.r.t.\ $L^2(\An)$ induces a complete orthonormal system $\{\varphi_i\otimes\varphi_j\}_{i,j\geq 1}$ w.r.t.\ $L^2(\red{\An \times \An})$; hence, we may decompose $K^\Gamma_1$ into
\[ K^\Gamma_1(\underline u,\underline v) = \sum_{i,j}\gamma_{i,j}\varphi_i(\underline u)\varphi_j(\underline v)\]
where
\[ \gamma_{i,j} := \iint K^\Gamma_1(\underline u,\underline v) \varphi_i(\underline u) \varphi_j(\underline v)\,
\d\underline u\d\underline v \,, \qquad
\sum_{i,j \geq 1} \gamma_{i,j}^2 =\| K^\Gamma_1 \|^2_{L^2(\red{\An \times \An})}
< \infty\,.
\]
This reduces~\eqref{eq:mu-free-2} into $\mu_{n,T}^\mathfrak{f}(\Gamma) = \Xi^{(1)}_{n,T}/\Xi^{(2)}_{n,T}$ where
\begin{equation}
\label{eq:mu-free-3}
\Xi^{(1)}_{n,T} := \sum_{i,j \ge 1} \gamma_{i,j} \alpha_{i,s} \alpha_{j,s} \widehat\lambda_i^{t} \widehat\lambda_j^{t-1} \,,
\qquad
\Xi^{(2)}_{n,T} :=
\lambda_1\sum_{i=1}^\infty \widehat\lambda_i^{2t}\alpha_{i,s}^2 \,,
\end{equation}
and the rescaled eigenvalues $\widehat\lambda_i := \lambda_i / \lambda_1 \in [0,1]$ $(i=1,2,\ldots)$ satisfy
\[ \widehat\lambda_i=1\qquad\mbox{and}\qquad\sup_{i>1}\widehat\lambda_i \leq 1-\delta\quad\mbox{for $\delta=(\lambda_1-\lambda_2)/\lambda_1>0$}\,.
\]
We immediately see that $\Xi_{n,T}^{(2)}$ of ~\eqref{eq:mu-free-3} satisfies
\begin{equation}\label{eq:err-Xi2}
\lambda_1\alpha_{1,s}^2 \leq \Xi_{n,T}^{(2)} \leq
\lambda_1\alpha_{1,s}^2 + \lambda_1 (1-\delta)^{2t}\|\psi_{s}\|_2^2\,.
\end{equation}
Further, with $K_s \varphi_1 = \lambda_1^s \varphi_1$ continuous and positive on $\An$,
for any non-zero
$\Theta_n$,
\begin{align*}
\alpha_{1,s}
=\int_{\An} (K_s \varphi_1)(\underline x) \Theta_n(\d \underline x)
= \lambda_1^s \int_{\An} \varphi_1 (\underline x) \Theta_n(\d \underline x)
\end{align*}
is bounded away from zero, uniformly over $s \le \ell+1$. Consequently,
\[ \lim_{T\to\infty} \,
\frac{\Xi_{n,T}^{(2)}}{\alpha_{1,s}^2} = \lambda_1 \,.
\]
To treat $\Xi_{n,T}^{(1)}$ of~\eqref{eq:mu-free-3}, note that
by Cauchy--Schwarz and having $\sup_{i \ge 2}
|\widehat \lambda_i| \le 1-\delta$,
\begin{align}\label{eq:err-Xi1}
\bigg|\sum_{\substack{i,j\geq 1 \\ i+j > 2}} \gamma_{i,j}\alpha_{i,s} \alpha_{j,s}
\widehat\lambda_i^{t}\widehat\lambda_j^{t-1} \bigg|
&\leq (1-\delta)^{t-1} \sum_{i,j}
\left|\gamma_{i,j}\alpha_{i,s} \alpha_{j,s} \right| \nonumber \\
& \leq (1-\delta)^{t-1} \sqrt{\sum_{i,j} \gamma_{i,j}^2}
\sqrt{\sum_{i,j} \alpha_{i,s}^2 \alpha_{j,s}^2} \nonumber \\
&= (1-\delta)^{t-1} \| K^\Gamma_1 \|_{L^2(\red{\An \times \An})} \, \|\psi_{s}\|_2^2 \,.
\end{align}
Taking $T\to\infty$, we see that
\[
\lim_{T\to\infty} \frac{\Xi_{n,T}^{(1)}}{\alpha_{1,s}^2} = \gamma_{1,1} \,.
\]
Altogether, we have thus established that
\begin{equation}
\label{eq:mu-free-4}
\lim_{T\to\infty}\mu_{n,T}^\mathfrak{f}(\Gamma) = \frac{\gamma_{1,1}}{\lambda_1}\,.
\end{equation}
We now repeat the same analysis for $\mu_{n,T}^\mathfrak{0}$, where
since the limit as $T \to \infty$ exists, we assume hereafter that
$T$ is integer (and set $s=\ell=1$). Further, for
simplicity we opt to take $\underline y=\underline x$ and let
$\psi^{(\epsilon)} (\underline u) := K_1(\epsilon \underline x,\underline u)$.
Inferring that $\psi^{(\epsilon)} \in L^2(\An)$ (because
$K_2(\epsilon \underline x,\epsilon \underline x) < \infty$), we can write
\[
\psi^{(\epsilon)} = \sum_{i=1}^\infty \alpha_i^{(\epsilon)} \varphi_i \,,
\]
where
\[\alpha_i^{(\epsilon)} := \big<\psi^{(\epsilon)},\varphi_i\big>_{L^2(\An)}\,, \quad
\|\psi^{(\epsilon)}\|_{L^2(\An)}^2 = K_2(\epsilon \underline x,\epsilon \underline x) =
\sum_{i=1}^\infty (\alpha^{(\epsilon)}_i)^2<\infty \,.
\]
The exact same argument then shows that
$\mu_{n,T}^\mathfrak{0}(\Gamma)$ is the limit at $\epsilon \to 0$ of
$\Xi^{(1,\epsilon)}_{n,T}/\Xi^{(2,\epsilon)}_{n,T}$ where
\begin{equation}
\label{eq:mu-zero-3}
\Xi^{(1,\epsilon)}_{n,T} := \sum_{i,j \ge 1} \gamma_{i,j} \alpha^{(\epsilon)}_i
\alpha^{(\epsilon)}_j \widehat\lambda_i^{T-1} \widehat\lambda_j^{T-2} \,,
\qquad
\Xi^{(2,\epsilon)}_{n,T} :=
\lambda_1\sum_{i=1}^\infty \widehat\lambda_i^{2T-2} (\alpha^{(\epsilon)}_i)^2 \,.
\end{equation}
With $\psi^{(\epsilon)}>0$ and $\varphi_1>0$, we have as before that
$\alpha^{(\epsilon)}_1 > 0$. Moreover, setting
\[ \kappa_\epsilon := \frac{\|\psi^{(\epsilon)}\|^2_{L^2(\An)}}{(\alpha^{(\epsilon)}_1)^2}\,,\]
we have
analogously to~\eqref{eq:err-Xi2} and~\eqref{eq:err-Xi1} that
\begin{align*}
0 \le \frac{\Xi_{n,T}^{(2,\epsilon)}}{(\alpha^{(\epsilon)}_1)^2}
- \lambda_1 & \le \lambda_1 (1-\delta)^{2T-2}
\kappa_\epsilon \,,
\\
\Big|\frac{\Xi_{n,T}^{(1,\epsilon)}}{(\alpha^{(\epsilon)}_1)^2} - \gamma_{1,1} \Big|
& \le (1-\delta)^{T-2}
\| K^\Gamma_1 \|_{L^2(\red{\An \times \An})} \, \kappa_\epsilon \,.
\end{align*}
We shall employ the following asymptotic as
$\epsilon \to 0$, the proof of which we defer to Section \ref{subsec-2.3}.
\begin{lemma}\label{lem:ker-asymptot}
Setting $\underline n:=(2n-1,2n-3,\ldots,1)$, we have that
\begin{equation}\label{eq:bd-kap-eps}
\limsup_{\epsilon \to 0} \frac{K_2(\epsilon \underline n,\epsilon \underline n)}{
\Big( \int_{u_1 \le 1} K_1(\epsilon \underline n, \underline u) \varphi_1(\underline u) d\underline u \Big)^2} < \infty \,.
\end{equation}
\end{lemma}
Since $K_1$ and $\varphi_1$ are both positive, \eqref{eq:bd-kap-eps} applies also
without the restriction to $u_1 \le 1$, with Lemma \ref{lem:ker-asymptot} yielding that
$\kappa_\epsilon$ is uniformly bounded (as $\epsilon \to 0$), when $\underline x=\underline n$. Hence, thanks
to our freedom to choose the boundary, we have that
\begin{equation}
\label{eq:mu-zero-2}
\lim_{T\to\infty}\mu_{n,T}^\mathfrak{0}(\Gamma) = \frac{\gamma_{1,1}}{\lambda_1}\,,
\end{equation}
which in light of~\eqref{eq:mu-free-4} concludes our proof.
\qed
\subsection{Proof of Lemma~\ref{lem:K1-compact}}
Letting
\begin{align*}
\mathfrak B_0 &= \big\{ f \,:\; \|f\|_{L^2(\An)}\leq 1\big\}\qquad\mbox{and} \\
\mathfrak B_1 &= \big\{ (K_1 f)\,:\; f\in \mathfrak B_0\big\}\,,
\end{align*}
we will establish compactness by verifying the Fr\'echet--Kolmogorov criteria
(see~\cite[p.~275]{Yosida}, as well as~\cite{Sudakov}).
First, with $\P$ denoting the law of Brownian motion $\{W(t)\}_{t \in [0,1]}$ in $\mathbb R^n$ started at the origin and $\mathbb E$ its corresponding expectation, note that
\begin{equation}\label{eq:K1f-alt}
(K_1 f)(\underline x) = \mathbb E \left[ \mathbbm{1}_{\Omega_n^{[0,1]}}(\underline x + W(\cdot))\, e^{-{\mathcal A}_{[0,1]}(\underline x+W(\cdot))} f(\underline x+W(1))\right] \,.
\end{equation}
Now, setting for $f$ supported on $\An$,
\begin{align}\label{eq:M(f)-def}
M(f) & := \sup_{\underline x\in\An} \mathbb E [|f(\underline x+W(1))|]
\,,
\end{align}
note that by Cauchy--Schwarz,
\begin{equation}\label{eq:M(f)-bound}
M(f)^2 \leq \sup_{\underline x\in\An} \mathbb E [f(\underline x+W(1))^2] \le
\|f \|^2_{L^2(\An)} \,\sup_{\underline x,\underline y \in\An} \{\phi_1(\underline y-\underline x)\}
\leq 1\,,
\end{equation}
where $\phi_v(\cdot)$ denotes the
density in~\eqref{def:Gauss} and
the last inequality holds for all $f\in\mathfrak B_0$.
This readily implies the following uniform bound on $g=K_1f \in \mathfrak B_1$,
where by a computation similar to the third line of \eqref{eq-K1K2},
for any $\underline x
\in \An$,
\begin{align}\label{eq:g(x)^2-unif-bnd}
\left|g(\underline x)\right| &\leq \int_{\An} \mathbf E_n^{\underline x,\underline y,[0,1]}\big[\,
e^{-\mathfrak a \int_0^1 X_1(s) ds} \big]|f(\underline y)|\,\d\underline y
\leq c e^{-\frac{\mathfrak a}{2} x_1} M(f) \leq
c e^{-\frac{\mathfrak a}{2} x_1}\,,
\end{align}
for some finite $c=c(\mathfrak a)$,
independent of $\underline x$ and $f \in \mathfrak B_0$.
We deduce in particular that
\begin{equation}\label{eq:equi-tight}
\limsup_{R\to\infty} \sup_{g\in \mathfrak B_1} \int_{\substack{\underline x\in\An \\ x_1>R}}\left|g(\underline x)\right|^2 d\underline x =0\,,
\end{equation}
establishing equitightness (and, due to \eqref{eq:g(x)^2-unif-bnd}, also
uniform boundedness, although it is not needed in view of~\cite{Sudakov}).
It remains to establish equicontinuity for $\mathfrak B_1$, where in view of
\eqref{eq:equi-tight} and the compactness of $\bAn \cap \{x_1 \le R \}$
it suffices to bound, in terms of $\|\underline h\|$,
the value of
\[
\sup_{g \in \mathfrak B_1, \underline x \in \An, \underline x+\underline h \in \An}
\{ |g(\underline x+\underline h) - g(\underline x)|
\} \,.
\]
Using the representation \eqref{eq:K1f-alt} for $g=K_1 f$,
we start by reducing to $\tilde g(\cdot)$ in which we extracted out the explicit
dependence of the area tilt on $\underline x$. Specifically, let
\[ \tilde g(\underline x) := \mathbb E \left[ \mathbbm{1}_{\Omega_n^{[0,1]}}(\underline x + W(\cdot))\, e^{-{\mathcal A}_{[0,1]}(W(\cdot))} f(\underline x+W(1))\right]\,. \]
By a slight abuse of notation, letting ${\mathcal A}_{[0,1]}(\underline x)$ denote ${\mathcal A}_{[0,1]}(X(\cdot))$ for $X\equiv \underline x$, which is nothing but $\mathfrak a\left< \underline\mathfrak b,\underline x\right>$ for $\underline\mathfrak b:=(1,\mathfrak b,\ldots,\mathfrak b^{n-1})$, we see that
\[ g(\underline x) = e^{-{\mathcal A}_{[0,1]}(\underline x)} \, \tilde g(\underline x)\,,
\]
and therefore,
\begin{align*} | g(\underline x)-g(\underline x+\underline h)|
&= | e^{-{\mathcal A}_{[0,1]}(\underline x)}\big(\tilde g(\underline x)- e^{-{\mathcal A}_{[0,1]}(\underline h)}\tilde g(\underline x+\underline h)\big)| \\
&\leq \big| e^{{\mathcal A}_{[0,1]}(\underline h)} - 1\big|\, | g(\underline x+\underline h)| +
e^{-{\mathcal A}_{[0,1]}(\underline x)} | \tilde g(\underline x)- \tilde g(\underline x+\underline h) | \,.
\end{align*}
For the first term note that
$|{\mathcal A}_{[0,1]}(\underline h)| = |\mathfrak a \left<\underline\mathfrak b,\underline h\right>| \leq \mathfrak a \|\underline{\mathfrak b}\| \|\underline h\|$ and though $\underline h$ may be outside $\An$,
by Taylor expansion and~\eqref{eq:g(x)^2-unif-bnd} we have that for any $\|\underline h\| \le 1$,
\begin{align*}
\sup_{g\in\mathfrak B_1,\underline x+\underline h \in\An} \big| e^{{\mathcal A}_{[0,1]}(\underline h)}-1\big| \,
| g(\underline x+\underline h)| \leq C(\mathfrak a,\mathfrak b,n) \|\underline h\|\,.
\end{align*}
Further, with ${\mathcal A}_{[0,1]}(\underline x)\geq 0$ for all $\underline x\in\An$,
it remains only to bound $|\tilde g(\underline x)-\tilde g(\underline y)|$
uniformly over $g \in \mathfrak B_1$, $\underline x\in\An$ and $\underline y\in \An$
such that $\|\underline y-\underline x\| \le \delta$.
To this end, let
\[
\tau_{\underline x} := \inf\left\{ t \geq 0 \,:\; \underline x+W(t) \notin \An \right\}\,,\quad\mbox{so that }\quad \mathbbm{1}_{\Omega_n^{[0,1]}}(\underline x+W(\cdot)) = \mathbbm{1}_{\left\{ \tau_{\underline x} > 1\right\}}\,.
\]
We then have in terms of
\[ \Delta(\underline x,\underline y) := \mathbbm{1}_{\left\{\tau_{\underline y}> 1\right\}} f(W(1)+\underline y) - \mathbbm{1}_{\left\{\tau_{\underline x}> 1\right\}} f(W(1)+\underline x)
\]
\and $\eta \in (0,1)$, that
\begin{align*}
| \tilde g(\underline y)-\tilde g(\underline x)| &=
| \mathbb E \big[ e^{-{\mathcal A}_{[0,1]}(W(\cdot))} \Delta(\underline x,\underline y) \big]
| \le \mathbb E\left[ |\Psi_1|\right] + | \mathbb E\left[\Psi_2\right]|\,,
\end{align*}
where
\begin{align*}
\Psi_1 &:= e^{-{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))}
\left(
e^{-{\mathcal A}_{[1-\eta,1]}(W(\cdot)-W(1-\eta))}-1
\right)
\Delta(\underline x,\underline y)\,, \\
\Psi_2 &:= e^{-{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))}\Delta(\underline x,\underline y)\,,
\end{align*}
and
\[ {\mathcal A}_{[0,1-\eta]}^*(W(\cdot)) = {\mathcal A}_{[0,1-\eta]}(W(\cdot)) + {\mathcal A}_{[1-\eta,1]} (W(1-\eta)) \,.
\]
To bound $\mathbb E|\Psi_1|$, use the fact that $|\Delta(\underline x,\underline y)| \leq |f(W(1)+\underline y)| + |f(W(1)+\underline x)|$ together with H\"older's inequality to infer that
$\mathbb E |\Psi_1|$ is at most
\begin{align*}
\mathbb E\bigg[ e^{-4{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))} \bigg]^{\frac14}
\mathbb E\bigg[\left|e^{-{\mathcal A}_{[1-\eta,1]}(W(\cdot)-W(1-\eta))}-1\right|^4\bigg]^{\frac14} \bigg(2\sup_{\underline x\in \An}\mathbb E\left[f(W(1)+\underline x)^2\right]\bigg)^{\frac12}.
\end{align*}
Noting that the variance of the centered Gaussian ${\mathcal A}^*_{[0,1-\eta]}(W(\cdot))$ is at most some $v=v(\mathfrak a,\mathfrak b,n)$ finite,
the first expectation above is uniformly bounded (namely, by $e^{8v}$).
Similarly, by \eqref{eq:M(f)-bound}, the third term is at most $\sqrt{2}$,
uniformly over $f \in \mathfrak B_0$. Finally, with
${\mathcal A}_{[1-\eta,1]}(W(\cdot)-W(1-\eta))$
a centered Gaussian of variance $c(\mathfrak a,\mathfrak b,n) \eta^2$ for some
finite $c(\mathfrak a,\mathfrak b,n)$,
the expectation in the second term is at most $\epsilon_0(\eta) \to 0$
as $\eta\to 0$. Overall, we conclude that
\begin{equation}\label{eq:eps1}
\mathbb E |\Psi_1| \leq \epsilon_{1}(\eta)\downarrow 0 \quad \hbox{as} \quad
\eta\downarrow 0\,, \quad \hbox{uniformly over} \quad g \in \mathfrak B_1, \underline x \in \An\,.
\end{equation}
Turning to $\Psi_2=e^{-{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))}\Delta(\underline x,\underline y)$, utilizing
the identity
\[
\mathbbm{1}_{\{\tau_{\underline y}>1\}} = 1 - \mathbbm{1}_{\{\tau_{\underline y}\leq 1-\eta \;,\; \tau_{\underline x}\leq 1-\eta\}}
-\mathbbm{1}_{\{1-\eta <\tau_{\underline y} \leq 1 \}} - \mathbbm{1}_{\{\tau_{\underline y} \le 1-\eta\;,\; 1-\eta < \tau_{\underline x}\leq 1\}} - \mathbbm{1}_{\{\tau_{\underline y}\leq 1-\eta\:,\: \tau_{\underline x}>1\}} \,,
\]
and its dual where the roles of $\tau_{\underline y}$ and $\tau_{\underline x}$ have been exchanged,
yields the decomposition
\begin{align*}
\Delta(\underline x,\underline y) &= \Upsilon_1 - \Upsilon_2(\underline y) -\Upsilon_3(\underline y,\underline x)-\Upsilon_4(\underline y,\underline x)
+ \Upsilon_2(\underline x) + \Upsilon_3(\underline x,\underline y)+ \Upsilon_4(\underline x,\underline y) \,,
\end{align*}
where
\begin{align*}
\Upsilon_1 &:= [f(W(1)+\underline y) - f(W(1)+\underline x)](1-\mathbbm{1}_{\{\tau_{\underline x}\leq 1-\eta \;,\; \tau_{\underline y}\leq 1-\eta\}})\,,\\
\Upsilon_2(\underline y) &:= f(W(1)+\underline y)\mathbbm{1}_{\{1-\eta < \tau_{\underline y}\leq 1 \}} \,, \\
\Upsilon_3(\underline y,\underline x) &:= f(W(1)+\underline y)\mathbbm{1}_{\{\tau_{\underline y} \le 1-\eta\;,\; 1-\eta < \tau_{\underline x}\leq 1\}} \,, \\
\Upsilon_4(\underline y,\underline x) &:= f(W(1)+\underline y) \mathbbm{1}_{\{\tau_{\underline y}\leq 1-\eta\:,\: \tau_{\underline x}>1\}} \,.
\end{align*}
For the contribution to $|\mathbb E[\Psi_2]|$
due to $\Upsilon_1$, condition on
${\mathcal F}_{1-\eta} = \sigma(\{W(s)\}_{s\leq 1-\eta})$,
on which the indicator in $\Upsilon_1$ is measurable, to get
\begin{align*}
\left|\mathbb E\big[ \Upsilon_1 e^{-{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))}\big]\right| &\leq \mathbb E\left[e^{-{\mathcal A}^*_{[0,1-\eta]}W(\cdot))}\right] \\
&\cdot \sup_{\underline z}
\Big| \mathbb E\big[f(W(1)+\underline y)- f(W(1)+\underline x) \;\big|\; W(1-\eta)=\underline z\big]\Big|\,.
\end{align*}
While treating $\mathbb E|\Psi_1|$, we saw that the first term on the right-hand
is some finite $C(\mathfrak a,\mathfrak b,n)$, independently of $\underline x,\underline h$.
For the second term, extending $f\in\mathfrak B_0$ from $\An$ to $\mathbb R^n$
via $f(\underline x)=0$ for $\underline x\notin \An$, yields that
$\|f\|_2 = \|f\|_{L^2(\An)} \leq 1$. Thus, performing
a change of variable
$\underline v := W(1)+\underline y$ in $\mathbb E[f(W(1)+\underline y) \mid W(1-\eta)=\underline z]$ and
$\underline v := W(1)+\underline x$ in $ \mathbb E[f(W(1)+\underline x)\mid W(1-\eta)=\underline z]$, we
get that the absolute difference between these expectations is
\begin{align*}
& \Big| \int \big[\phi_\eta(\underline v-\underline y-\underline z) -\phi_\eta(\underline v-\underline x-\underline z)\big] f(\underline v)\,\d\underline v \Big| \\
& \leq \|f\|_{L^2(\An)} \eta^{-n/4}
\big\| \phi_1(\underline w-\eta^{-1/2} (\underline y-\underline x))-\phi_1(\underline w)
\big\|_2
\le C(n) \, \eta^{-n/4-1/2} \delta \,,
\end{align*}
where the first inequality is obtained by Cauchy--Schwarz and an
additional change of variable $\underline w = \eta^{-1/2} (\underline v-\underline z-\underline x)$, and
the second inequality by an easy computation (utilizing that $1-e^{-r} \le r$).
Thus, choosing
\begin{equation}\label{eq:eta-def}
\delta \le \eta^{n/4+1},
\end{equation}
makes the contribution of $\Upsilon_1$ negligible.
To deal with the contribution of $\Upsilon_2(\underline y)$ to $|\mathbb E[\Psi_2]|$, observe that
by H\"older's inequality,
\begin{align}\label{eq:basic-holder}
\mathbb E&\left[
e^{-{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))}
\mathbbm{1}_{\{1-\eta < \tau_{\underline y} \leq 1\}}
|f(W(1)+\underline y)|
\right] \nonumber \\
& \leq
\mathbb E\Big[ e^{-4{\mathcal A}^*_{[0,1-\eta]}(W(\cdot))} \Big]^{\frac14}
\P\Big(1-\eta < \tau_{\underline y} \leq 1\Big)^{\frac14}
\Big(\sup_{\underline y\in \An}\mathbb E\left[f(W(1)+\underline y)^2\right]\Big)^{\frac12}.
\end{align}
While bounding $\mathbb E|\Psi_1|$ we have seen that the first and
third terms are at most some $c(\mathfrak a,\mathfrak b,n)$ finite, uniformly over $\mathfrak B_0$, so
it suffices to show that
\begin{equation}\label{eq:eps2}
\epsilon_2(\eta):=\sup_{\underline y \in \An} \{
\P(1-\eta < \tau_{\underline y}\leq 1) \} \to 0 \qquad \mbox{as \;\; $\eta\to 0$}\,.
\end{equation}
Indeed, taking a union bound over the $n$ different boundaries of $\An$ that are considered in $\tau_{\underline y}$, reduces, up to the factor $n$, to the bound in case $n=1$, namely for the first hitting time $T_b$
of level $-b<0$ by a standard Brownian motion $B_t$. The corresponding probability density
$f_{T_b}(t)=be^{-b^2/(2t)}/\sqrt{2\pi t^3}$
is bounded, uniformly over $b$ and $t \ge 1/2$,
thereby yielding \eqref{eq:eps2}.
The same analysis applies to the contributions
from the $\Upsilon_3$ terms.
Analogously to \eqref{eq:basic-holder} the contribution of
$\Upsilon_4(\underline y,\underline x)$ to $|\mathbb E[\Psi_2]|$ is bounded above by
\begin{align*}
& \mathbb E\left[e^{-{\mathcal A}_{[0,1-\eta]}^*(W(\cdot))}
\mathbbm{1}_{\{\tau_{\underline y}\leq 1-\eta \;,\; \tau_{\underline x}> 1\}} |f(W(1)+\underline y)|
\right] \\
\quad&\leq C \, \|f\|_{L^2(\An)}
\P(\tau_{\underline y}\leq 1-\eta\;, \;\tau_{\underline x} > 1)^{1/4} \le
C \, \epsilon_3(\delta,\eta)^{1/4} \,,
\end{align*}
for some $C=C(\mathfrak a,\mathfrak b,n)$, any $f \in \mathfrak B_0$ and
\[
\epsilon_3(\delta,\eta) := \sup_{\underline x,\underline y \in \An, \|\underline x-\underline y\| \le \delta}
\P(\tau_{\underline y}\leq 1-\eta\;, \;\tau_{\underline x} > 1) \,.
\]
With the same bound applying for $\Upsilon_4(\underline x,\underline y)$, it remains only
to show that $\epsilon_3(\delta,\eta) \to 0$ as $\delta \to 0$ (for any
fixed $\eta>0$). To this end, by a union bound over the $n$ different boundaries of $\An$, as done for proving \eqref{eq:eps2},
the probability in question is at most~$n$ times the probability
that standard Brownian motion $B(t):=\frac{1}{\sqrt{2}}(W_i(t)-W_{i+1}(t))$
reach level $-b$ by time $1-\eta$ (here $b=(y_i-y_{i+1})/\sqrt{2}$),
while remaining above $-(b+\delta)$
up till time $1$. With Brownian motion a strong Markov process
of independent increments, we thus deduce by the reflection principle
that
\begin{align*}
n^{-1} \epsilon_3(\delta,\eta) \le \P(\inf_{s\leq \eta} \{B(s)\} > -\delta)
= 1 - 2\P(B(\eta) \ge \delta) = \P(|B(\eta)| <\delta)\,,
\end{align*}
which goes to zero as $\delta \to 0$ (for any fixed $\eta>0$).
\qed
\subsection{Proof of Lemma \ref{lem:ker-asymptot}}\label{subsec-2.3}
Setting $\widehat K_t$ for the operator $K_t$ in the case $\mathfrak a=0$ (no area tilt),
we first establish~\eqref{eq:bd-kap-eps} for $\widehat K_t$. Namely, we show
that,
\begin{equation}\label{eq:bd-kap-a}
\limsup_{\epsilon \to 0} \frac{\widehat K_2(\epsilon \underline n,\epsilon \underline n)}{
\Big( \int_{u_1 \le 1} \widehat K_1(\epsilon \underline n, \underline u) \varphi_1(\underline u) d\underline u \Big)^2} < \infty \,.
\end{equation}
Our starting point for \eqref{eq:bd-kap-a} is the following explicit formula,
valid for any $\underline y \in \An$ and any $t,\epsilon>0$,
\begin{equation}\label{eq:exact}
\widehat K_t(\epsilon \underline n,\underline y) = 2^{n^2} \phi_t(\underline y)
e^{-\epsilon^2\|\underline n\|^2/(2t)} \prod_{i} \sinh\big(\frac{\epsilon y_i}{t}\big)
\prod_{j<k} \Big[
\sinh^2\big(\frac{\epsilon y_j}{t}\big) - \sinh^2\big(\frac{\epsilon y_k}{t}\big) \Big] \,.
\end{equation}
Indeed, for $\epsilon=1$ this is the explicit evaluation in
\cite[Display below (24)]{Grabiner99} of the Karlin--McGregor
determinantal formula~\cite{KarlinMcGregor59} for the transition kernel,
\[
q_t(x,y)=\phi_t(y-x)-\phi_t(y+x)=2 \phi_t(y) e^{-x^2/(2t)} \sinh(xy/t) \,,
\]
of a scalar Brownian motion absorbed at level zero, when starting at the
distinguished point $\underline n$. We thus get \eqref{eq:exact} by noting
that the non-trivial factors $\sinh(x_i y_j/t)$ are invariant to
changing from $(\epsilon \underline n,\underline y)$ to $(\underline n,\epsilon \underline y)$.
\begin{comment}
For $t>0$, $\mathfrak a=0$ and $\underline x,\underline y$ in the interior of $\An$, note that
$\widehat K_t(\underline x,\underline y) = \mathbf E_n^{\underline x,\underline y,[0,t]}\big[\mathbbm{1}_{\Omega_n^{[0,t]}} \big]$,
is precisely the sub-probability density at time $t$ and point $\underline y$, for
$n$ independent scalar Brownian motions absorbed upon hitting zero, or
whenever two of them coincide. Thus, by the Karlin--McGregor formula~\cite{KarlinMcGregor59},
\[
\widehat K_t(\underline x,\underline y) = \det(\{q_t(x_i,y_j)\}_{i,j \le n}) \,,
\]
for the transition probability density
\[
q_t(x,y)=\phi_t(y-x)-\phi_t(y+x)=2 \phi_t(y) e^{-x^2/(2t)} \sinh(xy/t) \,,
\]
of a scalar Brownian motion absorbed at level zero. Since all factors which depend
only on the row or column index scale out of a determinant, we get that
\[
\widehat K_t(\epsilon \underline x,\underline y) = 2^n \phi_t(\underline y) e^{-\epsilon^2 \|\underline x|^2/(2t)}
\det(\{\sinh((\epsilon/t) x_i y_j) \}_{i,j \le n}) \,.
\]
Now, from the multi-linearity of the determinant and the Taylor expansion
\[
\sinh(u)=\sum_{r=1}^\infty \frac{u^{2r-1}}{(2r-1)!} \,,
\]
it follows that
\begin{equation}\label{eq:det-exp}
\det(\{\sinh((\epsilon/t) x_i y_j) \}) = \sum_{r:[n] \to \mathbb N}
(\epsilon/t)^{\ell(r)} f(r) \prod_{i=1}^n x_i^{2r(i)-1}
\det(\{y_j^{2r(i)-1}\}) \,.
\end{equation}
The two-dimensional central minor of the matrix $\{y_j^{2r(i)-1}\}$
at rows and columns $k \ne k'$, has rank one when $r(k)=r(k')$,
hence a zero determinant for any such matrix. Consequently, it suffices
to consider in \eqref{eq:det-exp} only invertible $r:[n] \to \mathbb N$.
Further, for any permutation $\pi$ of $[n]$ we have that
$\ell(\cdot)$ and $f(\cdot)$ are invariant to replacing $r$ by
$r \circ \pi^{-1}$, whereas such permutation of the rows of
$\{y_j^{2r(i)-1}\}$ merely multiplies its determinant by $(-1)^{{\rm sign}(\pi)}$. Thus, combining together
the contributions from all
permutations of each strictly decreasing
map $r:[n] \downarrow \mathbb N$, we can rewrite the \abbr{rhs}
of \eqref{eq:det-exp} as
\[
\sum_{r:[n] \downarrow \mathbb N}
(\epsilon/t)^{\ell(r)} f(r) \det(\{x_j^{2r(i)-1}\}) \det(\{y_j^{2r(i)-1}\}) \,.
\]
The minimal value $n^2$ of $\ell(r)$ over all strictly increasing
$r :[n] \downarrow \mathbb N$, is uniquely attained at $r_\star$.
The contribution of this map to the
preceding sum is $2^{-n} (\epsilon/t)^{n^2} c_n V_n(\underline x) V_n(\underline y)$,
\green{whereas the contribution of all other strictly decreasing maps
is $O(\epsilon^{n^2+1})$,} which is relatively negligible when
$\epsilon \to 0$.
\end{comment}
In particular, with $g(x):=\sinh(x/2)$ being zero at $x=0$
and globally Lipschitz($L$) on $[0,2n]$, we get from \eqref{eq:exact} that
for some $c_n$, $C_n$ finite and any $\epsilon \in [0,1]$,
\begin{align}\label{eq:ubd-wK2}
\widehat K_2(\epsilon \underline n,\epsilon \underline n) &\le c_n
\prod_i g(\epsilon^2 n_i) \prod_{j<k} [g^2(\epsilon^2 n_j) - g^2(\epsilon^2 n_k)]
\nonumber \\
&\le c_n L^{n^2} \prod_i (\epsilon^2 n_i) \prod_{j<k} \big[(\epsilon^2 n_j)^2
- (\epsilon^2 n_k)^2 \big] = C_n \epsilon^{2 n^2} \,.
\end{align}
Next, noting that on $\mathbb R_+$ both $\sinh(x) \ge x$ and $\sinh^2(x)-x^2$ are
non-decreasing, we deduce from \eqref{eq:exact} that for any $\underline u \in \An$
and $\epsilon \in [0,1]$,
\[
\widehat K_1(\epsilon \underline n,\underline u) \ge 2^{n^2} e^{-\|\underline n\|^2/2} \epsilon^{n^2}
\hat{\phi}(\underline u) \,, \quad \hbox{where} \quad
\hat{\phi}(\underline u) := \phi_1(\underline u) \prod_{i} u_i \prod_{j<k} (u_j^2 - u_k^2) \,.
\]
With $\hat{\phi}(\cdot)$ and $\varphi_1(\cdot)$ positive
on $\An$, we get from the latter bound that
\[
\inf_{\epsilon \in [0,1]} \epsilon^{-n^2}
\int_{u_1 \le 1} \widehat K_1(\epsilon \underline n, \underline u) \varphi_1(\underline u) d\underline u > 0 \,,
\]
which in combination with \eqref{eq:ubd-wK2} establishes~\eqref{eq:bd-kap-a}.
Next, recall that $K_t(\underline x,\underline y)$ is point-wise decreasing in $\mathfrak a$ and
in particular bounded from above by $\widehat K_t(\underline x,\underline y)$;
thus, the sought bound \eqref{eq:bd-kap-eps} for $K_t$ follows from~\eqref{eq:bd-kap-a} once we show that for some finite $C=C(\mathfrak a,\mathfrak b,n)$
and any $\underline u \in \An$ with $u_1\leq 1$,
\begin{equation}\label{eq:unif-bd}
\sup_{\epsilon\in (0,1]} \Big\{ \frac{\widehat K_1(\epsilon\underline n,\underline u)}{K_1(\epsilon\underline n,\underline u)} \Big\} \le C \,.
\end{equation}
Turning to the latter bound, we define for finite $M$ the event
\[ \Gamma_M := \left\{ \max_{t \in [0,1]} \{X_1(t)\} \le M \right\}\,,\]
noting that for $c := \mathfrak a \langle \underline{\mathfrak b}, \underline 1 \rangle$,
any $u_1 \le 1$ and $\epsilon \le 1$,
\begin{align*}
K_1(\epsilon\underline n,\underline u) &\geq e^{-c M} \mathbf E_n^{\epsilon\underline n,\underline u,[0,1]}
\left[ \mathbbm{1}_{\Gamma_M} \mathbbm{1}_{\Omega_n^{[0,1]}}\right]
= e^{-c M} \widehat K_1(\epsilon\underline n,\underline u) \widehat\P^{\epsilon\underline n,\underline u,[0,1]}_n (\Gamma_M) \\
&\geq e^{-c M} \widehat K_1(\epsilon\underline n,\underline u) \widehat\P^{\underline n,\underline n,[0,1]}_n (\Gamma_M) \,,
\end{align*}
where $\widehat\P^{\underline x,\underline y,[0,1]}_n$ is the measure $\P^{\underline x,\underline y,[0,1]}_n$ from~\eqref{eq:P-ux-uy-def} corresponding to $\mathfrak a=0$, and with the second inequality due to~\cite[Lemma~2.7]{CorwinHammond14} (taking there $A=[0,1]$,
$f \equiv 0$,
noting that $\underline n > \underline u$ and $\underline n > \epsilon \underline n$
whenever $u_1 \leq 1$ and $\epsilon \le 1$
and that the event $\Gamma_M$ is decreasing).
Finally, moving to the unconditional space of $n$ independent bridges rooted at $\underline n,\underline n$ via a multiplicative cost of at most $1/\widehat K_1(\underline n,\underline n)$, we see that
$\widehat\P^{\underline n,\underline n,[0,1]}_n (\Gamma_M^c)$ is at most $\P(\sup_{s \in [0,1]}
\{B(s)\} > M-2n) /\widehat K_1(\underline n,\underline n)$
for a one dimensional Brownian bridge from $(0,1)$ to $(1,1)$. By the tightness of the maximum of the latter bridge (and recalling that
$\widehat{K}_1(\underline n,\underline n)>0$), one thus has for $M$ large,
depending only on $n$, that
\[
\widehat\P_n^{\underline n,\underline n,[0,1]}(\Gamma_M) \ge \tfrac12\,.
\]
Combining the last two displays yields \eqref{eq:unif-bd},
thereby completing the proof.
\qed
\subsection*{Acknowledgment}
We thank Ivan Corwin for bringing to our attention the paper~\cite{Grabiner99}.
We thank the referees for a careful reading of the manuscript and their comments.
A.D.\ was supported in part by NSF grant DMS-1954337.
E.L.\ was supported by NSF grants DMS-1812095 and DMS-2054833.
O.Z.\ was partially supported by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No.~692452).
This research was further supported in part by BSF grant 2018088.
\bibliographystyle{abbrv}
|
\subsection{Related work}
Dedicated machines with single-axis servo mechanisms are omnipresent in industrial production processes, especially given the tendency to evolve towards dedicated actuators for each machine movement \cite{Berselli2016}. In this context, several approaches have been presented in the past literature to optimize the motion profiles of these one DOF systems.
In \cite{Richiedei2016,Lee2020,Park1996,Carabin2021,Botan2010}, optimal motion profiles are obtained for mechanical systems with constant load parameters such as the inertia $J$. Nevertheless, as indicated in literature \cite{Pellicciari2015,Berselli2016,Hsu2014}, it is essential to consider varying loads to cover the majority of machine applications.
However, in those cases obtaining an analytic description of the position-dependent system properties is a challenging task for industrial machine designers. For instance, \cite{Hsu2014,Hsu2016,Huang2012,Ha2006} use Hamilton's principle and Lagrange multipliers to obtain the differential-algebraic equations of the system. In \cite{Sollmann2010}, a method of virtual work is described to obtain the system matrix while \cite{Vanbecelaere2020} determines the inertia profile using the method of kinetic energy. Such approaches are cumbersome, complex and error-prone, and are not easily applicable. Especially given the trend indicated in \cite{Walsch2014} that there is a demand for methods that take into account the ease of implementation.
Fortunately, machine builders often already design their machines in 3D CAD multibody software, which can be used to extract crucial information. Hence, in \cite{VanOosterwyck2019} and \cite{Berselli2016}, the authors of this paper describe a technique to derive the position dependency of critical parameters inertia $J(\theta)$ and load torque $\tau_l(\theta)$, based on only three CAD motion simulations. If the load model is known, optimization algorithms can iterate on it to minimize energy usage.
For what concerns these optimization algorithms, several approaches are possible. On the one hand, \cite{Park1996} and \cite{Shiller1996} use an indirect approach such as Pontryagin's Maximum Principle to obtain the best possible control. However, this method tends to be abandoned recently due to the small convergence area and difficulties incorporating constraints \cite{Chettibi2004}.
On the other hand, direct approaches recast the optimization into a nonlinear optimization problem, which can be solved with different numerical methods. In particular, \cite{VanOosterwyck2020} and \cite{Pellicciari2015} use gradient-based methods such as Sequential Quadratic Programming (SQP) or quasi-Newton methods that are known to have very low solve times and good scalability. However, these algorithms can only deliver local optimal solutions and are not suited for problems with multiple minima. As indicated in \cite{Huang2017}, the optimum obtained with gradient-based methods is greatly influenced by the selected starting points, which are to be chosen arbitrarily.
To avoid this problem, heuristic optimization algorithms \cite{VanOosterwyck2019,Huang2012,Hsu2014,Berselli2016} such as generalized pattern search (GPS) or genetic algorithms (GA) are of interest. In contrast to gradient-based algorithms that do not search the entire design space, derivative-free algorithms like GA often sample a wide part of the design space in order to be successful \cite{Wenzhong2005}. Nevertheless, because these heuristic solvers do not exploit gradient information, they are not computationally competitive with gradient-based methods \cite{Betts1998}.
Regarding the motion profile function, several papers rely on piecewise position functions (\cite{Pellicciari2015,Berselli2016}), where either cubic, quintic, or trigonometric splines are used. However, the objective functions in these works, are characterized by many local minima, causing the risk of getting stuck in a suboptimal solution. For instance, in \cite{Piazzi1998}, the usage of cubic splines resulted in a savings difference of $18\%$ between the global and local optimum.
On the other hand, continuous motion profile functions such as classic polynomials (\cite{VanOosterwyck2019,Huang2012,Hsu2014,Carabin2021,Lee2020}) are also popular because they do not introduce high jerk peaks into the system, which increases the wear of the components. However, the resulting optimization problem is known to be badly conditioned. For example in \cite{VanOosterwyck2019}, the coefficients reached values up to $1.8 \, 10^{20}$. Therefore, the authors of this paper proposed Chebyshev polynomials and rescalings in \cite{VanOosterwyck2020} to obtain a more numerically stable problem formulation and to increase the robustness against getting stuck in local minima. Moreover, in contrast to the classical polynomial descriptions, the design variables of Chebyshev polynomials can be bounded as is shown in this paper.
\subsection{Method}
In light of the considerations mentioned above, a CAD-based method for computing energy-optimal PTP (Point-to-Point) motion profiles of single DOF mechanisms using Chebyshev polynomials has been previously presented by the authors in \cite{VanOosterwyck2020}. However, due to the numerous symbolic calculations involved in constructing the objective function, solve times of almost 2 hours were reported. In addition, the solutions in \cite{VanOosterwyck2020} were obtained using gradient-based solvers which have a high risk of getting stuck in local minima. Finally, only theoretical reductions were reported, thus, leaving the feasibility of the proposed motion profiles undetermined. This paper builds upon these previous results by providing five critical improvements:
\begin{itemize}
\item In order to reduce the computational burden, a discrete approach is presented which eliminates the use of symbolic operations. To do so, the discrete system property data which originates from the CAD motion simulations have to be properly rescaled and interpolated.
\item As an accurate model of the system dynamics is crucial for a correct optimization, the dynamics of the mechanism are extended by including damping and friction into the optimization routine. In addition, a new identification procedure is described which is validated on an industrial case.
\item A derivation for exact bounds of the Chebyshev polynomial coefficients is introduced. This allows limiting the feasible design space. The latter is essential for heuristic optimizers to reduce their computation time and the chances that the global optimum remains unidentified.
\item To check the robustness of the proposed method against getting stuck in local optima, the resulting optimization problem is solved with both a fast gradient-based and a global heuristic solver (i.e. GA).
\item Experimental tests have been carried out on an industrial pick-and-place unit to quantify the actual measured energy savings.
\end{itemize}
\section{System Modeling}
The complete mechatronic system can be divided into two subsystems (Fig. \ref{fig:Simplified_Model}). On the one hand, there is the \textit{mechanical subsystem} which describes the dynamics of a generic single-axis system. For high dynamical applications, these systems usually consist of slider-crank mechanisms and four-bar linkages \cite{Berselli2016}. Nevertheless, the approach is valid to any position-controlled system where the mechanism is driven by a single actuator.
On the other hand, there is the \textit{actuation subsystem} which converts the electrical energy into mechanical energy and drives the mechanism. For the envisaged position-controlled systems, PMSMs are becoming the industry standard for rotary applications, whereas linear motors are used for fast and precise linear movements \cite{Kiel2008}. In Fig. \ref{fig:Simplified_Model}, the PMSM actuator is represented by an equivalent DC model.
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Simplified_Model_Motor.pdf}
\caption{Schematic of the q-axis of a single axis mechanism.}
\label{fig:Simplified_Model}
\end{figure}
\subsection{Mechanical subsystem}
The dynamics of a single axis DOF mechanism can be described by means of the torque equation \cite{Pellicciari2015,Dresig2010} :
\begin{equation} \label{eq:torque_equation}
\tau_m(t) = \tau_l(\theta) + \underbrace{J(\theta)\ddot{\theta}}_{\tau_a} + \underbrace{\frac{1}{2}\frac{\mathrm{d}J(\theta)}{\mathrm{d}\theta}(\dot{\theta})^2}_{\tau_{v}} + \tau_f(\dot{\theta}) \, .
\end{equation}
With reference to Fig. \ref{fig:Simplified_Model} and equation (\ref{eq:torque_equation}), let us define $\theta = \theta(t)$ as the Lagrangian coordinate which describes the angular position of the main driving axis as a function of time $t$. The motor torque $\tau_m(\theta)$ is defined as the driving torque generated by the motor. The load torque $\tau_l(\theta)$ contains both gravitational forces as well as external process powers that act on the mechanism.
Furthermore, all inertias of the mechanism's components are related to the main driving axis resorting to the concept of reduced moment of inertia. Therefore, the reduced inertia of the complete system $J(\theta)$ is defined as a combination of the reduced load inertia $J_l(\theta)$ and inertia of the motor shaft itself $J_m$. Note that the position-dependent inertia of the system $J(\theta)$, results in two torque components when it is reduced to the motor side. The acceleration torque $\tau_a$ represents the part of the motor torque responsible for the motor acceleration forces that arise during the movement, while the variation torque $\tau_v$ compensates for the variation of inertia in the system.
Finally, the frictional torque $\tau_f(\dot{\theta})$ is defined as the result of frictional forces such as, for instance, viscous brush friction or dry bearing friction in the motor bearings and mechanical system. A commonly used model of friction shows three components of force: Coulomb (sliding) friction, viscous damping, and static friction \cite{Ellis2012}. Regarding the PMSM, as indicated in \cite[p.~175]{Westphal2001}, the only appreciable friction effect in operation is viscous friction. Thus, coulomb and static frictions can be neglected in the PMSM model. For what concerns the mechanical model, only the viscous damping is modeled since the other friction components are constant and will not have an effect on the optimal motion profile:
\begin{equation} \label{eq:driction_torque}
\tau_f(\dot{\theta}) = \mu_v\dot{\theta} \, ,
\end{equation}
with $\mu_v$ the equivalent viscous friction coefficient.
The key benefit of the formulation in \eqref{eq:torque_equation} is that it permits to model every possible mechanism with a known geometry and allows to define a generic optimization approach.
\subsection{Actuation subsystem}
Concerning the dynamics of the PMSM as depicted in Fig. \ref{fig:Simplified_Model} (represented by an equivalent DC model), the electromechanical behavior can be easily described by the following basic laws \cite[p.~843]{Rizzoni2003}:
\begin{equation} \label{eq:basis_torque}
\tau_m = k_t i \, ,
\end{equation}
\begin{equation} \label{eq:voltage}
u = R i + L\frac{\mathrm{d}i}{\mathrm{d}t} + e = R i + L\frac{\mathrm{d}i}{\mathrm{d}t} + p k_v \dot{\theta} \, ,
\end{equation}
with electric back emf $e$, resistance $R$, back emf constant $k_v$, motor torque constant $k_t$, and number of pole pairs $p$, which can be found in the motor data sheet.
In equation (\ref{eq:voltage}), the voltage drop $L\frac{\mathrm{d}i}{\mathrm{d}t}$ due to the armature inductance is omitted as the mean value of its reactive power will be zero and therefore does not contribute to the system's energy need \cite{Pellicciari2015}.
Depending on whether the electric power flows from the drive unit to the PMSM's or vice versa, the PMSM operates in respectively motor or generator mode. In this latter condition, depending on the capabilities of the drive unit, the generated electric power can be either stored in a capacitor, dissipated as heat on a braking resistance, or transferred back to the energy source. Recent commercial PMSM drives are sized so that no electric power is actually dissipated during normal functioning so that the braking resistance is actually activated only under emergency conditions \cite{Berselli2016}. Therefore, in what follows, it is assumed that all the generated energy is returned to the grid and no losses occur in the process.
For a correct model of the actuation subsystem and prediction of the energy usage, it is important to model other losses such as cooling fans and drive circuitry as well. Nevertheless, the power consumption of these devices is generally considered constant and is therefore not affected by the motion profile \cite{Gadaleta2019}.
In order to minimize the total energy need $E$ of the application, it is crucial to quantify the input energy of the complete system. Therefore, similar to \cite{Berselli2016}, a formulation of the input electrical energy $E$ is derived and a torque-based design objective is obtained which allows minimizing the energy solely based on the mechanical parameters.
Starting from equations \eqref{eq:basis_torque} and \eqref{eq:voltage}, the instantaneous power $P_e$ is defined as
\begin{equation}
P_{e} = u \, i = \frac{R}{k_t^2}\tau_m^{2} + \frac{p k_v}{k_t} \tau_m \dot{\theta} \, .
\end{equation}
The motion profile is defined on the time interval $t \in [t_A,t_B]$ and must have zero initial and final speed and acceleration, i.e. $\dot{\theta}(t_A) = \dot{\theta}(t_B) = \ddot{\theta}(t_A) = \ddot{\theta}(t_B) = 0$. The total energy can be expressed as
\begin{equation}
\begin{aligned}
E= \int_{t_A}^{t_B} P_e \, \mathrm{d} t &= \int_{t_A}^{t_B} \left[ \frac{R}{k_t^2}\tau_m^{2} + \frac{p k_v}{k_t} \tau_m \dot{\theta} \right] \, \mathrm{d} t \, .
\end{aligned}
\end{equation}
Then, by incorporating the torque equation from \eqref{eq:torque_equation}, the total energy of the motion is given by:
\begin{equation} \label{eq:energy}
\begin{aligned}
E=\frac{p k_v}{k_t} \underbrace{\int_{t_A}^{t_B} (\tau_a + \tau_v) \dot{\theta} \, \mathrm{d} t}_{E_{k}} + \frac{p k_v}{k_t} \underbrace{\int_{t_A}^{t_B} \tau_l \, \dot{\theta} \, \mathrm{d} t}_{E_{p}} \\
+\underbrace{\int_{t_A}^{t_B}\left[\frac{R}{k_t^2} \tau_m^{2}+ \frac{p k_v}{k_t} \tau_f \dot{\theta}\right] \mathrm{d} t}_{E_{l}} \, .
\end{aligned}
\end{equation}
Here, the first term $E_k$ represents the kinetic energy of the moving masses in the system. Due to the rest-to-rest motion of the envisaged applications, this term reduces to zero. Further, the term $E_p$ represents the potential energy stored in the system. As this term $E_p$ only depends on the fixed start $\theta_A$ and end position $\theta_B$, it is disregarded in the optimization routine \cite{Berselli2016}. The final term $E_l$ represents the energy that is lost due to the coil resistance and frictional forces and is the only term that is affected by optimizing the motion profile $\theta(t)$. Nevertheless, in many industrial applications, the frictional forces $\tau_f$ are negligible \cite{Park1996}, especially if the inertial loads are predominant. Thus, the energy losses $E_l$ can be expressed as:
\begin{equation}
E_l = \int_{t_A}^{t_B} \frac{R}{k_t^2} \tau_m^{2} \mathrm{d} t = \frac{R \Delta t}{k_t^2}\tau_{rms}^2 \, ,
\end{equation}
where $\tau_{rms}$ is the RMS value of the motor Torque $\tau_{m}$. This proves that the RMS torque $\tau_{rms}$ can be effectively used as an optimization objective to minimize the total energy usage of the system. This is very useful in situations where the motor coil properties are unknown or where parameters are missing \cite{Berselli2016}.
\section{Identification}
\subsection{Inertia and Load Torque}
Identification of all the position varying parameters in the highly nonlinear differential torque equation (\ref{eq:torque_equation}) is not straightforward. Fortunately, machine builders design their machines in 3D CAD multibody software. For this reason, \cite{Berselli2016} and \cite{VanOosterwyck2019} describe a technique to derive the position dependency of critical parameters inertia $J(\theta)$ and load torque $\tau_l(\theta)$, based on three CAD motion simulations. (Fig. \ref{fig:Property_Extraction}).
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Property_Extraction.pdf}
\caption{Schematic overview of the procedure for extracting position-dependent properties $J(\theta)$ and $\tau_l(\theta)$ based on three different CAD motion simulations \cite{VanOosterwyck2019}.}
\label{fig:Property_Extraction}
\end{figure}
In this paper, the identification routine is illustrated by applying it to an industrial pick-and-place unit (Fig. \ref{fig:Experimental_Setup}) that performs repetitive movements between start point $A$ with angular position $\theta_A = 0$ and endpoint $B$ with angular position $\theta_B = 173.6^\circ$. The resulting inertia $J(\theta)$ and load torque $\tau_l(\theta)$ profiles are presented in Fig. \ref{fig:sytem_properties}. Because of the machine position limits $\theta_A$ and $\theta_B$, only the green shaded part of the system properties is relevant during operation.
\begin{figure}
\centering
\includegraphics[width=0.47\columnwidth]{Images/Nedschroef_Real2.png}
\includegraphics[width=0.51\columnwidth]{Images/Nedschroef_Schematic_Simple.pdf}
\caption{Experimental set-up (left) and schematic overview (right) of the pick- and place unit.}
\label{fig:Experimental_Setup}
\end{figure}
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/System_Properties.pdf}
\caption{Values of system properties inertia $J(\theta)$ and load torque $\tau_l(\theta)$.}
\label{fig:sytem_properties}
\end{figure}
\subsection{Viscous Friction Coefficient}
Once the system properties $J(\theta)$ and $\tau_l(\theta)$ are determined, the only indefinite term in Eq. \eqref{eq:torque_equation} is the friction torque $\tau_f$, and more specifically $\mu_k$. In the previous description of the energy flows, the friction torque $\tau_f$ was neglected, leading to a simple objective (i.e. $\tau_{rms}$) to quantify the energy consumption. However, it is important to verify this statement for the intended setup. Therefore this section describes a method to quantify the frictional forces $\tau_f$.
Since the viscous friction coefficient $\mu_k$ parameter is highly dependent on the practical setup, it is often only possible to determine this parameter experimentally. Therefore, a first measurement is carried out by using an arbitrary motion profile $\theta^{*}(t)$ as a set point and recording the resulting actual motor torque $\tau_{m}^{e}(t)$ and position $\theta^{e}(t)$. The arbitrary motion profile $\theta^{*}(t)$ can be determined by using a default motion law such as a trapezoidal or s-curve profile.
After this measurement, a least squares fit can be used to determine the experimental value of $\mu_k$, by fitting the torque model $\tau_{m}(\theta^{e}(t),\mu_k)$ with the measured torque $\tau_{m}^{e}$.
However, using the measured position $\theta^{e}$ and its time derivatives $\dot{\theta}^{e}$, $\ddot{\theta}^{e}$ in the torque equation \eqref{eq:torque_equation} leads to unfeasible results since the derivatives amplify any noise that is present in the measurement. Therefore, the measured position $\theta^{e}(t)$ is fitted with an $n$-th degree polynomial $\theta^p(t) = \sum_{i=1}^{n} a_it^i$ and differentiated symbolically to smooth out any noise.
The friction parameter $\mu_v$ is thus determined by comparing the measured torque $\tau_{m}^{e}(t)$ with virtual model and fitted motion profile $\tau_{m}(\theta^{p}(t))$:
\begin{equation}
\begin{aligned}
& \underset{\mu_v \, \in \, \mathbb{R}}{\text{minimize}}
& & ||\tau_{m}^{e}(t) - \tau_{m}(\theta^{p}(t),\mu_v)||_2 \, .
\end{aligned}
\end{equation}
For the pick-and-place unit, a viscous damping coefficient of $0.0157Nms/rad$ was found. In Fig. \ref{fig:Torque_Validation}, a comparison of the measured $\tau_{m}^{e}(t)$ and virtual $\tau_{m}(\theta^{p}(t))$ torque is presented. The difference between the virtual torques with and without friction is minimal, which indicates that the friction can be neglected for the present case. The graph also shows a close correlation between the virtual and measured torque, which indicates that the virtual model can be effectively used to minimize the RMS torque $\tau_{rms}$ and, by extension, the energy consumption $E$.
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Torque_Validation.pdf}
\caption{Comparison of the virtual $\tau_{m}(\theta^{p}(t))$ and measured torque $\tau_{m}^{e}(t)$ (with and without friction).}
\label{fig:Torque_Validation}
\end{figure}
\section{Optimization Approach}
\subsection{Motion Profile Definition \& Rescaling}
In this paper, a Chebyshev polynomial $\sum_{i=0}^{n} p_iT_i(x)$ is used to define the position profile $\theta(t)$, where $t \in [t_A,t_B]$, in between the start-$\theta(t_A) = \theta_A$ and endpoint $\theta(t_B) = \theta_B$ of the motion task. The sequence of orthogonal Chebyshev polynomials $T_n(x) = T_n(\cos(\vartheta))$, defined on the interval $x \in [-1,1]$, is obtained from the recurrence relation:
\begin{equation} \label{eq:def_cheb}
\begin{aligned}
T_0(x) &= 1, \quad T_1(x) = x, \\
T_{k+1}(x) &= 2xT_k(x)-T_{k-1}(x),
\end{aligned}
\end{equation}
This gives exactly the same results as from the trigonometric definition:
\begin{equation} \label{eq:def_cheb_tri}
T_k(\cos(\vartheta)) = \cos(k\vartheta).
\end{equation}
To use $T_n(x)$ as a representation for the position profile, a linear transformation from $t$ into the range $[-1,1]$ of $x$ is required \cite{Thompson2013}:
\begin{equation} \label{eq:rescale_tx}
\begin{aligned}
t&=\frac{1}{2}(t_B-t_A)x+\frac{1}{2}(t_B+t_A) =a x+ b,
\end{aligned}
\end{equation}
where scale factors $a$ and $b$ are defined for the purpose of the following paragraphs. In addition, the position $\theta$ is also rescaled to the interval $[-1,1]$, which makes it possible to obtain strict bounds on the design space in \eqref{boundsDS}. Thus, the rescaled motion profile description $\phi(x)$ of degree $n$ with optimizable coefficients $\mathbf{p} = [p_0, p_1, \ldots ,p_n]^T$ is obtained.
\begin{equation} \label{eq:position_function_cheb}
\phi(x)=\sum_{i=0}^{n} p_iT_i(x), \quad x\in [-1,1].
\end{equation}
The output of the motion simulations in the previous section deliver $n_s$ samples of inertia $\mathbf{J}= [J_1, \ldots ,J_{n_s}]^T$, load torque $\mathbf{T}_l= [T_{l,1}, \ldots ,T_{l,n_s}]^T$ and corresponding angle query points $\boldsymbol{\uptheta} = [\theta_1, \ldots ,\theta_{n_s}]^T$. Due to the position rescaling of the motion profile $\phi(x)$, the angle query points $\boldsymbol{\uptheta}$ of every system property vector $\mathbf{Y}$ have to be rescaled accordingly:
\begin{equation} \label{eq:rescale_prop}
\begin{aligned}
\boldsymbol{\upphi} &= \frac{2}{(\theta_B -\theta_A)} \, \boldsymbol{\uptheta} - \frac{(\theta_B +\theta_A)}{(\theta_B -\theta_A)} = c \, \boldsymbol{\uptheta} + d.
\end{aligned}
\end{equation}
Moreover, as the property description $Y(\phi)$ is now defined on the rescaled interval $\phi \in [-1,1]$, the following relationship holds with regard to the derivative properties:
\begin{equation} \label{eq:rescale_pder}
\frac{\mathrm{d}J(\phi)}{\mathrm{d}\phi} = \frac{1}{2}(\theta_B-\theta_A) \frac{\mathrm{d}J(\theta)}{\mathrm{d}\theta} = e \, \frac{\mathrm{d}J(\theta)}{\mathrm{d}\theta}.
\end{equation}
When using the rescaled position profile $\phi(x)$, it is important to rescale the torque equation \eqref{eq:torque_equation} as well. Otherwise, the resulting values of the torque profile $T(x)$ are distorted which results in different objective values (i.e. $T_{rms}$) and solutions. To preserve the motor torque's absolute values, the following rescaled torque equation is used:
\begin{equation} \label{eq:torque_equation_rescaled}
T_m(x) = T_l(\phi) + \frac{1}{2}\frac{\mathrm{d}J(\phi)}{\mathrm{d}\phi}\frac{1}{e}\left(\frac{\dot{\phi}}{a.c}\right)^2 + J(\theta)\frac{\ddot{\phi}}{a^2.c} + \mu_k\frac{\dot{\phi}}{a}.
\end{equation}
An overview of the position and torque rescalings is presented in Fig. \ref{fig:rescaling}. The new system equation \eqref{eq:torque_equation_rescaled} ensures the system dynamics are equally scaled and the minima are not altered.
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Rescaling2.pdf}
\caption{Original $\theta(t)$ and rescaled position profiles $\theta(x)$, $\phi(x)$ with their corresponding torque equations.}
\label{fig:rescaling}
\end{figure}
For what concerns the constraints, the rest-to-rest motion required zero speed $\dot{\phi}$ and acceleration $\ddot{\phi}$ in the start and endpoint:
\begin{equation}
\begin{array}{ccccc}
\phi(-1)=-1 & , & \dot{\phi}(-1)=0 & , & \ddot{\phi}(-1)=0,\\
\phi(1)=1 & , & \dot{\phi}(1)=0 & , & \ddot{\phi}(1)=0.
\end{array}{}
\label{eq:constraints}
\end{equation}
Referring to \eqref{eq:position_function_cheb}, and by incorporating the motion profile constraints \eqref{eq:constraints}, the lower degree coefficients $[p_0,... , p_5]^T$ can be written as a function of the remaining coefficients $[p_6,... , p_n]^T$, such that $n-5$ degrees of freedom (DOF) are kept available for the optimization algorithm \cite{Hsu2014}. Thus, the energy optimal motion profile problem is formulated as the following minimization problem with design variable vector $\mathbf{o}=[p_6,... , p_n]^T$:
\begin{equation}
\begin{aligned}
& \underset{\mathbf{o} \, \in \, \mathbb{R}^{n-5}}{\text{minimize}}
& & T_{m;rms}(x) = \sqrt{\frac{1}{2}\int_{-1}^{1} {T_m(\phi(x,\mathbf{o}))}^2 \, \mathrm{d}x} .
\end{aligned}
\end{equation}
In some applications, an additional constraint of zero jerk in the begin and endpoint can be imposed to limit the vibrations:
\begin{equation} \label{eq:constraint_jerk}
\dddot{\phi}(-1)=0 \quad;\quad \dddot{\phi}(1)=0.
\end{equation}
Because of these two extra equations, the DOF is reduced to $n-7$ and the design variable vector can be expressed as $\mathbf{o}=[p_8,... , p_n]^T$.
\subsection{Initialization \& Design Space}
In this paper, the resulting optimization problem is solved with both a fast \textit{gradient-based} solver, the BFGS (Broyden–Fletcher–Goldfarb–Shanno) quasi-Newton method \cite{Nocedal2006}, and a global \textit{heuristic} solver, the genetic algorithm \cite{Holland1992}.
For gradient-based optimization, a starting point needs to be defined. The use of the Chebyshev basis $T_i(x)$ in representation \eqref{eq:position_function_cheb} allows initializing the optimization parameter vector at zero, since the coefficients in a convergent Chebyshev series development of the motion profile function $\theta(x)$ would converge to zero \cite{Majidian2017}. Here, we can safely assume some similar behavior for the coefficients $p_i$ in \eqref{eq:position_function_cheb}.
For what concerns the genetic algorithm, a similar approach is used for the initialization of the population. However, because a GA often samples a wide part of the design space \cite{Wenzhong2005}, it is beneficial to determine the exact bounds on the design vector $\mathbf{o}$. By doing so, the solver can cover a large part of the design space and reveal the global optimal solution. In the following paragraphs, thanks to the rescaled Chebyshev motion profile, strict bounds on the design vector $\mathbf{o}$ can be derived.
To define these bounds, we first introduce the integral $I$:
\begin{equation}\label{int}
I = \int\limits_{0}^{2\pi} T_k(\cos\theta) \cos(\ell\theta) \, \mathrm{d}\theta.
\end{equation}
With respect to Eq. \eqref{eq:def_cheb_tri} and by using the inverse Simpson rule, the integral can be written as:
\begin{equation}
\begin{aligned}
I &= \int\limits_{0}^{2\pi} \cos(k\theta) \cos(\ell\theta) \, \mathrm{d}\theta \\
&= \frac{1}{2} \int\limits_{0}^{2\pi} \cos\big((k+\ell)\theta\big)\,\mathrm{d}\theta
+ \frac{1}{2} \int\limits_{0}^{2\pi} \cos\big((k-\ell)\theta\big)\,\mathrm{d}\theta.
\end{aligned}
\end{equation}
The problem is divided into three cases of which the results will be used in the following paragraphs:
\begin{enumerate}
\item \underline{$k+\ell = 0$} \\
Since $k$ and $\ell$ are both positive, this is only possible for $k=\ell=0$. In this case, we find
\begin{equation}\label{eq:case1}
I =
2\pi.
\end{equation}
\item \underline{$k-\ell = 0$} \\
In this case, we find that $k = \ell$, which means that
\begin{equation}\label{eq:case2}
\begin{aligned}
I
= \pi.
\end{aligned}
\end{equation}
\item \underline{$k+\ell \neq 0$ and $k-\ell \neq 0$} \\
In this case, $k \neq \ell$ and both are not 0 at the same time. For the integral we then find
\begin{equation}\label{eq:case3}
\begin{aligned}
I
= 0.
\end{aligned}
\end{equation}
\end{enumerate}
Further, a second integral $F$ is introduced which is used later on to obtain the strict bounds.
\begin{equation} \label{eq:integralF}
\begin{aligned}
F &= \int\limits_{0}^{2\pi} \phi(\cos\theta)\cos(\ell\theta)\,\mathrm{d}\theta.
\end{aligned}
\end{equation}
By taking into account \eqref{eq:position_function_cheb}, we find the following result:
\begin{equation}
\begin{aligned}
F &= \int\limits_{0}^{2\pi} \left(\sum\limits_{k=0}^n p_k T_k(\cos\theta)\right)\cos(\ell\theta)\,\mathrm{d}\theta \\
&= \sum\limits_{k=0}^n p_k \int\limits_{0}^{2\pi} T_k(\cos\theta) \cos(\ell\theta) \, \mathrm{d}\theta.
\end{aligned}
\end{equation}
Then, by taking into account \eqref{eq:case3}, only the term for which $k=l$ remains in the summation:
\begin{equation}
\begin{aligned}
F = p_\ell \int\limits_{0}^{2\pi} T_\ell(\cos\theta) \cos(\ell\theta) \, \mathrm{d}\theta.
\end{aligned}
\end{equation}
This can be split into two cases. For $\ell = 0$ and with respect to \eqref{eq:case1} and \eqref{eq:integralF} we find:
\begin{equation}\label{c0}
p_0 = \dfrac{1}{2\pi} \int\limits_{0}^{2\pi} \phi(\cos\theta)\,\mathrm{d}\theta,
\end{equation}
and for $\ell > 0$ with respect to \eqref{eq:case2} and \eqref{eq:integralF}:
\begin{equation}\label{cl}
p_\ell = \dfrac{1}{\pi} \int\limits_{0}^{2\pi} \phi(\cos\theta)\cos(\ell\theta)\,\mathrm{d}\theta.
\end{equation}
For $\theta \in [0,2\pi]$, $\cos\theta$ lies in interval $[-1,1]$, so the image of $\phi(\cos\theta)$ also lies in the interval $[-1,1]$. Thus, we find:
\begin{equation}
|p_0| \leq \dfrac{1}{2\pi}\int\limits_{0}^{2\pi} |\phi(\cos\theta)|\,\mathrm{d}\theta \leq \dfrac{1}{2\pi}\int\limits_{0}^{2\pi} \,\mathrm{d}\theta = 1.
\end{equation}
and
\begin{equation}
|p_\ell| \leq \dfrac{1}{\pi}\int\limits_{0}^{2\pi} |\phi(\cos\theta)||\cos(\ell\theta)|\,\mathrm{d}\theta \leq \dfrac{1}{\pi}\int\limits_{0}^{2\pi}|\cos(\ell\theta)| \,\mathrm{d}\theta.
\end{equation}
To calculate this last integral, we use the periodicity of the function $\cos(\ell\theta)$. This function has a period of $2\pi / \ell$, so goes $\ell$ times up and down on the interval $[0,2\pi]$. So, after taking the absolute value of this function, we find $2\ell$ times the integral over the positive part of a period, for example, the interval $[-\pi/2\ell, \pi/2\ell]$:
\begin{equation}
\begin{aligned}
\dfrac{1}{\pi}\int\limits_{0}^{2\pi}|\cos(\ell\theta)| \,\mathrm{d}\theta = \dfrac{2\ell}{\pi} \int\limits_{-\pi/2\ell}^{\pi/2\ell} \cos(\ell\theta)\,\mathrm{d}\theta = \dfrac{4}{\pi}.
\end{aligned}
\end{equation}
Thus, the following bounds were obtained:
\begin{equation} \label{boundsDS}
|p_0| \leq 1 \hspace{0.5cm} \text{ and } \hspace{0.5cm} |p_\ell| \leq \dfrac{4}{\pi}, \hspace{0.3cm} \ell =1,\dots,n.
\end{equation}
These constraints on the design space are exploited in the subsequent optimization.
\section{Results}
\subsection{Motion Profile Optimization}
In order to assess the performance of the proposed method, a set of optimizations has been performed on the industrial pick-and-place unit depicted in Fig. \ref{fig:Experimental_Setup}. The mechanism is required to move between its start position $\theta_A$ of $0^\circ$ and end position $\theta_B$ of $173.6^\circ$ and has a motion time $\Delta t$ of $73.5ms$. As for the constraint, two different cases are considered, namely
\begin{itemize}
\item \textit{Jerk Free (JF)}: Only the boundary constraints of \eqref{eq:constraints} are taken into account. The corresponding rescaled Chebyshev position profile $\phi(x)$ of degree $n$ is hereafter referred to as \textit{cheb"n"}. A 5th-degree polynomial, hereafter indicated as \textit{poly5}, is taken as the reference motion profile for comparison purposes. This is the smallest degree polynomial that satisfies the constraints.
\item \textit{Jerk Zero (J0)}: In addition to the constraint of a jerk-free optimization, a zero-jerk constraint is added in the start, and endpoint \eqref{eq:constraint_jerk} is added to the motion profile definition. The resulting $n$-th degree position profile $\phi(x)$ is referred to as \textit{cheb"n"J0}. The reference motion profile is in this case a 7th-degree polynomial, hereafter referred to as \textit{poly7J0}.
\end{itemize}
For every case, the resulting optimization problem is solved in a MATLAB environment for degrees $n= 7, 9, 11,$ and $13$. The results are presented in Fig. \ref{fig:Optimization_Results} and Tables \ref{tab:results_jerkfree} \& \ref{tab:results_jerkzero} where for every motion profile, the corresponding RMS torque $\tau_{rms}$ and solve time $t_{sol}$ are displayed. Savings up to $54.4\%$ are obtained in under $0.77$ s. The results clearly converge towards a minimal value for increasing degree $n$. In general, the motion profiles which include the jerk constraint \eqref{eq:constraint_jerk}, have slightly bigger $\tau_{rms}$ values, which is to be expected due to the fact that this extra constraint limits the acceleration near the endpoints while it is desirable to have high accelerations here since the inertia is low.
In Table \ref{tab:results_jerkfree}, the $\tau_{rms}$ values of a conventional trapezoidal 1/3 motion profile are presented as well, which accelerates during 1/3rd of the time, moves at a constant speed during 1/3rd, and decelerates at the last 1/3rd \cite{Park1996}. What is interesting in this table is that the torque demand can already be significantly reduced by selecting an adequate default motion law. Notwithstanding that the greatest savings are realized after optimization.
It is worth noting that for the jerk-free motion profiles, the same solution was found for both the genetic algorithm and gradient-based solver. However, the calculation times with GA are considerably higher. When including the jerk constraint, the GA comes close but does not completely reveal the full optimization potential. Therefore, for what concerns the present study, gradient-based optimizations algorithms are preferable. Since the GA did not obtain a better solution for any motion profile in the bounded search space, we can expect that the results obtained with the gradient-based method are global optimal solutions.
Although only the forward motion is considered here, similar results can be obtained for the return motion by simply changing the position constraints.
\begin{table}
\caption{Results of the motion profile optimization (Jerk Free).}
\label{tab:results_jerkfree}
\centering
\begin{tabular}{lcccc}
& \multicolumn{2}{c}{\textbf{Gradient-Based }} & \multicolumn{2}{c}{\textbf{Genetic Algorithm}} \\
\textbf{JF} & \textbf{$\tau_{rms}\,[Nm]$} & \textbf{$t_{sol}\, [s]$} & \textbf{\textbf{$\tau_{rms}\,[Nm]$}} & \textbf{$t_{sol}\,[s]$} \\
\hline\hline
poly5 (ref.) & \begin{tabular}[c]{@{}c@{}}22.48\end{tabular} & - & \begin{tabular}[c]{@{}c@{}}22.48\end{tabular} & - \\
\cmidrule(lr){1-5}
trap & \begin{tabular}[c]{@{}c@{}}17.16 \\-23.7\%\end{tabular} & - & \begin{tabular}[c]{@{}c@{}}17.16\\-23.7\%\end{tabular} & - \\
\cmidrule(lr){1-5}
cheb7 & \begin{tabular}[c]{@{}c@{}}13.78 \\-38.7\%\end{tabular} & 0.21 & \begin{tabular}[c]{@{}c@{}}13.78\\-38.7\%\end{tabular} & 3.28 \\
\cmidrule(lr){1-5}
cheb9 & \begin{tabular}[c]{@{}c@{}}12.47 \\-44.5\%\end{tabular} & 0.32 & \begin{tabular}[c]{@{}c@{}}12.47\\-44.5\%\end{tabular} & 40.33 \\
\cmidrule(lr){1-5}
cheb11 & \begin{tabular}[c]{@{}c@{}}12.33 \\-45.2\%\end{tabular} & 0.51 & \begin{tabular}[c]{@{}c@{}}12.33\\-45.2\%\end{tabular} & 67.05 \\
\cmidrule(lr){1-5}
cheb13 & \begin{tabular}[c]{@{}c@{}}12.29 \\-45.4\%\end{tabular} & 1.06 & \begin{tabular}[c]{@{}c@{}}12.29\\-45.4\%\end{tabular} & 142.34 \\
\cmidrule(lr){1-5}
\end{tabular}
\end{table}
\begin{table}
\caption{Results of the motion profile optimization (Jerk 0).}
\label{tab:results_jerkzero}
\centering
\begin{tabular}{lcccc}
& \multicolumn{2}{c}{\textbf{Gradient-Based }} & \multicolumn{2}{c}{\textbf{Genetic Algorithm}} \\
\textbf{J0} & \textbf{$\tau_{rms}\,[Nm]$} & \textbf{$t_{sol}\, [s]$} & \textbf{\textbf{$\tau_{rms}\,[Nm]$}} & \textbf{$t_{sol}\,[s]$} \\
\hline\hline
poly7J0 (ref.) & \begin{tabular}[c]{@{}c@{}}28.44\end{tabular} & - & \begin{tabular}[c]{@{}c@{}}28.44\end{tabular} & - \\
\cmidrule(r){1-5}
cheb9J0 & \begin{tabular}[c]{@{}c@{}}16.12 \\-43.3\%\end{tabular} & 0.27 & \begin{tabular}[c]{@{}c@{}}16.12\\-43.3\%\end{tabular} & 6.15 \\
\cmidrule(r){1-5}
cheb11J0 & \begin{tabular}[c]{@{}c@{}}13.61 \\-52.2\%\end{tabular} & 0.38 & \begin{tabular}[c]{@{}c@{}}14.11\\-50.4\%\end{tabular} & 175.23 \\
\cmidrule(lr){1-5}
cheb13J0 & \begin{tabular}[c]{@{}c@{}}12.98\\-54.4\%\end{tabular} & 0.77 & \begin{tabular}[c]{@{}c@{}}13.15\\-53.8\%\end{tabular} & 195.02 \\
\cmidrule(lr){1-5}
\end{tabular}
\end{table}
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Optimization_Results.pdf}
\caption{Results of the motion profile optimization for different degrees $n$.}
\label{fig:Optimization_Results}
\end{figure}
\subsection{Measurements}
The theoretical results are validated against experimental measurements on the pick-and-place unit (Fig. \ref{fig:Experimental_Setup}). The setup comprises a Beckhoff CX5140 PLC, a Beckhoff AX5901 motor drive, and a Beckhoff AM3064 PMSM, which is connected to the shaft of the mechanism. In order to measure the input electrical energy, a Tektron PA4000 power analyzer is used to analyze the power supply (Fig. \ref{fig:Schematic_Experimental_Setup}).
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Experimental_Setup.pdf}
\caption{Schematic overview of the experimental setup.}
\label{fig:Schematic_Experimental_Setup}
\end{figure}
The theoretical savings potential of the motion profile optimization is only fulfilled when the motor is capable of following the optimized position setpoint. Therefore, a performant motion controller needs to be designed in order to keep the tracking error as low as possible. Here, similar to \cite{VanOosterwyck2019}, a cascade controller with torque and speed feedforward is employed as it has proven to be successful for high dynamic systems. The look-up table for the feedforward torque is determined using the torque equation \eqref{eq:torque_equation_rescaled}.
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Motion_Controller.pdf}
\caption{Schematic overview of the cascade motion controller with feedforward \cite{VanOosterwyck2019}.}
\label{fig:Motion_Controllers}
\end{figure}
In Tables \ref{tab:measurement_jerkfree} and \ref{tab:measurement_jerkzero}, the results of both the measured RMS torque $\tau_{rms}$ and measured input electrical energy $E$ for different motion profiles are presented. As expected from the simulations, the lowest absolute energy consumption is obtained when using jerk-free motion profiles. When the jerk constraint is active, a decrease of 62.9\% in energy consumption can be achieved by optimizing the motion profile, while a relative saving of 52.5\% is possible if no extra constraint on the jerk is imposed.
The measured $\tau_{rms;meas}$ and calculated RMS motor torque $\tau_{rms}$ show a very high similarity, which confirms that the present system model is valid.
\begin{table}
\caption{Experimental results with energy measurement (Jerk Free).}
\label{tab:measurement_jerkfree}
\centering
\begin{tabular}{cccc}
\textbf{JF} & \textbf{$\tau_{rms}\,[Nm]$} & \textbf{$\tau_{rms;meas}\,[Nm]$} & \textbf{$E_{meas} \, [Wh] $} \\
\hline\hline
poly5 & 22.48 & 19.59 & 312.2 \\
\cmidrule(r){1-4}
trap & \begin{tabular}[c]{@{}c@{}}17.16\\-23.7\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}15.88\\-18.98\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}215.1\\-31.1\%\end{tabular} \\
\cmidrule(r){1-4}
cheb7 & \begin{tabular}[c]{@{}c@{}}13.78\\-38.7\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.40\\-31.6\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}181.7\\-41.8\%\end{tabular} \\
\cmidrule(lr){1-4}
cheb9 & \begin{tabular}[c]{@{}c@{}}12.47\\-44.5\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.07\\-38.4\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}152.3\\-51.2\%\end{tabular} \\
\cmidrule(lr){1-4}
cheb11 & \begin{tabular}[c]{@{}c@{}}12.33\\-45.2\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.93\\-39.1\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}150.1\\-51.9\%\end{tabular} \\
\cmidrule(lr){1-4}
cheb13 & \begin{tabular}[c]{@{}c@{}}12.29\\-45.4\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}11.83\\-39.6\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}148.2\\-52.5\%\end{tabular} \\
\cmidrule(lr){1-4}
\end{tabular}
\end{table}
\begin{table}
\caption{Experimental results with energy measurement (Jerk Zero).}
\label{tab:measurement_jerkzero}
\centering
\begin{tabular}{cccc}
\textbf{J0} & \textbf{$\tau_{rms}\,[Nm]$} & \textbf{$\tau_{rms;meas}\,[Nm]$} & \textbf{$E_{meas} \, [Wh] $} \\
\hline\hline
poly7J0 & 28.44 & 25.30 & 458.5 \\
\cmidrule(lr){1-4}
cheb9J0 & \begin{tabular}[c]{@{}c@{}}16.12\\-43.3\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}15.81\\-37.5\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}222.9\\-51.4\%\end{tabular} \\
\cmidrule(lr){1-4}
cheb11J0 & \begin{tabular}[c]{@{}c@{}}13.61\\-52.2\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}13.08\\-48.3\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}170.3\\-62.9\%\end{tabular} \\
\cmidrule(lr){1-4}
cheb13J0 & \begin{tabular}[c]{@{}c@{}}12.98\\-54.4\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}12.72\\-49.7\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}170.8\\-62.7\%\end{tabular} \\
\cmidrule(lr){1-4}
\end{tabular}
\end{table}
\section{Conclusion}
This study proposes a novel approach for motion profile optimization of PTP motions with Chebyshev polynomials. At first, system properties have been extracted from both CAD motion simulations and measurements to obtain an accurate virtual twin of the system. A Chebyshev motion profile with scaling laws is presented. Especially novel in this paper is the derivation of the boundary conditions of this profile which enables to define bounds for the design variables. The latter allows to use an optimizer that is designed to obtain globally optimal solutions, i.e. Genetic Algorithm. In addition, the solutions are validated with fast gradient-based optimization algorithms. Finally, experimental optimization results have been considered to verify the feasibility of the proposed solutions.
The numerical results, achieved on an exemplary model, clearly show that large $\tau_{rms}$ savings of up to 53.8\% can be achieved. In addition, it is shown that by employing Chebyshev polynomials for the motion profile, a fast gradient-based optimization can be effectively employed with solve times under $0.8$s. At last, the validation measurements show that similar savings are obtained on the real machine with a maximum energy reduction of $62.9 \%$.
Due to the straightforward implementation of both the optimization itself and integration of the resulting motion profiles in the motor drive, the proposed method can be easily adopted in any existing configuration where the CAD is data available. Therefore, the proposed method is expected to have a beneficial impact on the energy usage of the envisaged PTP applications.
\section{Optimization Approach}
\subsection{Motion Profile Definition \& Rescaling}
In this paper, a Chebyshev polynomial $\sum_{i=0}^{n} p_iT_i(x)$ is used to define the position profile $\theta(t)$, where $t \in [t_A,t_B]$, in between the start- ($\theta(t_A) = \theta_A$) and endpoint ($\theta(t_B) = \theta_B$) of the motion task. The sequence of orthogonal Chebyshev polynomials $T_k(x) = T_k(\cos(\vartheta))$, defined on the interval $x \in [-1,1]$, is obtained from the recurrence relation:
\begin{equation} \label{eq:def_cheb}
\begin{aligned}
T_0(x) &= 1, \quad T_1(x) = x, \\
T_{k+1}(x) &= 2xT_k(x)-T_{k-1}(x),
\end{aligned}
\end{equation}
Alternatively, the polynomials can be derived from the trigonometric definition, which gives exactly the same results:
\begin{equation} \label{eq:def_cheb_tri}
T_k (x) = T_k(\cos(\vartheta)) = \cos(k\vartheta).
\end{equation}
To use $T_n(x)$ as a representation for the position profile, a linear transformation from $t$ into the range $[-1,1]$ of $x$ is required \cite{Thompson2013}:
\begin{equation} \label{eq:rescale_tx}
\begin{aligned}
t&=\frac{1}{2}(t_B-t_A)x+\frac{1}{2}(t_B+t_A) =a x+ b,
\end{aligned}
\end{equation}
where scale factors $a$ and $b$ are defined for the purpose of the following paragraphs. In addition, the position $\theta \in [\theta_A, \theta_B]$ is also rescaled to the interval $\phi \in [-1,1]$, which makes it possible to obtain strict bounds on the design space in \eqref{boundsDS}. Thus, the rescaled motion profile description $\phi(x)$ of degree $n$ with optimizable coefficients $\mathbf{p} = [p_0, p_1, \ldots ,p_n]^T$ is obtained.
\begin{equation} \label{eq:position_function_cheb}
\phi(x)=\sum_{i=0}^{n} p_iT_i(x), \quad x\in [-1,1].
\end{equation}
The output of the motion simulations in the previous section deliver $n_s$ samples of inertia $\mathbf{J}= [J_1, \ldots ,J_{n_s}]^T$, load torque $\boldsymbol{\uptau}_l= [\tau_{l,1}, \ldots ,\tau_{l,n_s}]^T$ and corresponding angle query points $\boldsymbol{\uptheta} = [\theta_1, \ldots ,\theta_{n_s}]^T$. Due to the position rescaling of the motion profile $\phi(x)$, the angle query points $\boldsymbol{\uptheta}$ have to be rescaled accordingly:
\begin{equation} \label{eq:rescale_prop}
\begin{aligned}
\boldsymbol{\upphi} &= \frac{2}{(\theta_B -\theta_A)} \, \boldsymbol{\uptheta} - \frac{(\theta_B +\theta_A)}{(\theta_B -\theta_A)} = c \, \boldsymbol{\uptheta} + d.
\end{aligned}
\end{equation}
Moreover, as the property description is now defined on the rescaled interval $\phi \in [-1,1]$, the following relationship holds with regard to the derivative properties such inertia variation $\frac{\mathrm{d}J(\phi)}{\mathrm{d}\phi}$:
\begin{equation} \label{eq:rescale_pder}
\frac{\mathrm{d}J(\phi)}{\mathrm{d}\phi} = \frac{1}{2}(\theta_B-\theta_A) \frac{\mathrm{d}J(\theta)}{\mathrm{d}\theta} = e \, \frac{\mathrm{d}J(\theta)}{\mathrm{d}\theta}.
\end{equation}
When using the rescaled position profile $\phi(x)$, it is important to rescale the torque equation \eqref{eq:torque_equation} as well. Otherwise, the resulting values of the torque profile $\tau(x)$ are distorted which results in different objective values (i.e. $\tau_{rms}$) and solutions. To preserve the motor torque's absolute values, the following rescaled torque equation is introduced:
\begin{equation} \label{eq:torque_equation_rescaled}
\tau_m(x) = \tau_l(\phi) + \frac{1}{2}\frac{\mathrm{d}J(\phi)}{\mathrm{d}\phi}\frac{1}{e}\left(\frac{\dot{\phi}}{a.c}\right)^2 + J(\theta)\frac{\ddot{\phi}}{a^2.c} + \mu_k\frac{\dot{\phi}}{a.c}.
\end{equation}
An overview of the position and torque rescalings is presented in Fig. \ref{fig:rescaling}. The new system equation \eqref{eq:torque_equation_rescaled} ensures the system dynamics are equally scaled and the minima are not altered.
\begin{figure}[thpb]
\centering
\includegraphics[width=\columnwidth]{Images/Rescaling2.pdf}
\caption{Original $\theta(t)$ and rescaled position profiles $\theta(x)$, $\phi(x)$ with their corresponding torque equations.}
\label{fig:rescaling}
\end{figure}
For what concerns the constraints, the rest-to-rest motion requires zero speed $\dot{\phi}$ and acceleration $\ddot{\phi}$ in the start and endpoint:
\begin{equation}
\begin{array}{ccccc}
\phi(-1)=-1 & , & \dot{\phi}(-1)=0 & , & \ddot{\phi}(-1)=0,\\
\phi(1)=1 & , & \dot{\phi}(1)=0 & , & \ddot{\phi}(1)=0.
\end{array}{}
\label{eq:constraints}
\end{equation}
Referring to \eqref{eq:position_function_cheb}, and by incorporating the motion profile constraints \eqref{eq:constraints}, the lower degree coefficients $[p_0,... , p_5]^T$ can be written as a function of the remaining coefficients $[p_6,... , p_n]^T$, such that $n-5$ degrees of freedom (DOF) are kept available for the optimization algorithm \cite{Hsu2014}. Thus, the energy optimal motion profile problem is formulated as the following minimization problem with design variable vector $\mathbf{o}=[p_6,... , p_n]^T$:
\begin{equation}
\begin{aligned}
& \underset{\mathbf{o} \, \in \, \mathbb{R}^{n-5}}{\text{minimize}}
& & \tau_{rms} = \sqrt{\frac{1}{2}\int_{-1}^{1} {\tau_m(\phi(x,\mathbf{o}))}^2 \, \mathrm{d}x} .
\end{aligned}
\end{equation}
In some applications, an additional constraint of zero jerk in the begin and endpoint can be imposed to limit the vibrations:
\begin{equation} \label{eq:constraint_jerk}
\dddot{\phi}(-1)=0 \quad;\quad \dddot{\phi}(1)=0.
\end{equation}
Because of these two extra equations, the DOF is reduced to $n-7$ and the design variable vector can be expressed as $\mathbf{o}=[p_8,... , p_n]^T$.
\subsection{Initialization \& Design Space}
In this paper, the resulting optimization problem is solved with both a fast \textit{gradient-based} solver, the BFGS (Broyden–Fletcher–Goldfarb–Shanno) quasi-Newton method \cite{Nocedal2006}, and a global \textit{heuristic} solver, the genetic algorithm \cite{Holland1992}.
For gradient-based optimization, a starting point needs to be defined. The use of the Chebyshev basis $T_i(x)$ in representation \eqref{eq:position_function_cheb} allows initializing the optimization parameter vector at zero since the coefficients in a convergent Chebyshev series development of the motion profile function $\phi(x)$ would converge to zero \cite{Majidian2017}. Here, we can safely assume some similar behavior for the coefficients $p_i$ in \eqref{eq:position_function_cheb}.
For what concerns the genetic algorithm, a similar approach is used for the initialization of the population. However, because a GA often samples a wide part of the design space \cite{Wenzhong2005}, it is beneficial to determine the exact bounds on the design vector $\mathbf{o}$. By doing so, the solver can cover a large part of the design space and reveal the global optimal solution. In the following paragraphs, thanks to the rescaled Chebyshev motion profile $\phi(x)$, strict bounds on the design vector $\mathbf{o}$ can be derived.
To define these bounds, we take a look at the projection of the position profile $\phi(x)$ onto the orthogonal Chebyshev polynomial basis $T_l(x)$. Given that $x =\cos(\theta)$, we introduce the inner product $F$:
\begin{equation} \label{eq:integralF}
\begin{aligned}
F = \langle\phi(x),T_l(x)\rangle &= \int\limits_{-1}^{1} \frac{\phi(x)T_l(x)}{\sqrt{1-x^2}}\,\mathrm{d}x\\
&= \int\limits_{0}^{2\pi} \phi(\cos\theta)T_l(\cos\theta)\,\mathrm{d}\theta.
\end{aligned}
\end{equation}
Then, by taking into account the position function definition \eqref{eq:position_function_cheb}, we find the following result:
\begin{equation}
\begin{aligned}
F &= \int\limits_{0}^{2\pi} \left(\sum\limits_{k=0}^n p_k T_k(\cos\theta)\right)T_l(\cos\theta)\,\mathrm{d}\theta \\
&= \sum\limits_{k=0}^n p_k \int\limits_{0}^{2\pi} T_k(\cos\theta) T_l(\cos\theta) \, \mathrm{d}\theta.
\end{aligned}
\end{equation}
Here, the integral $I= \int\limits_{0}^{2\pi} T_k(\cos\theta) T_l(\cos\theta) \, \mathrm{d}\theta$ can be further simplified by using the Chebyshev polynomial orthogonality properties, which are rederived here for the sake of readability. Because of Eq. \eqref{eq:def_cheb_tri} and by using the inverse Simpson rule of trigonometry, the integral $I$ can be written as:
\begin{equation}
\begin{aligned}
I &= \int\limits_{0}^{2\pi} \cos(k\theta) \cos(\ell\theta) \, \mathrm{d}\theta \\
&= \frac{1}{2} \int\limits_{0}^{2\pi} \cos\big((k+\ell)\theta\big)\,\mathrm{d}\theta
+ \frac{1}{2} \int\limits_{0}^{2\pi} \cos\big((k-\ell)\theta\big)\,\mathrm{d}\theta.
\end{aligned}
\end{equation}
This integral can be split into three cases:
\begin{enumerate}
\item \underline{$k =\ell = 0$} \\
\begin{equation}\label{eq:case1}
I =
2\pi,
\end{equation}
\item \underline{$k =\ell \neq 0$} \\
\begin{equation}\label{eq:case2}
\begin{aligned}
I
= \pi,
\end{aligned}
\end{equation}
\item \underline{$k \neq \ell$} \\
\begin{equation}\label{eq:case3}
\begin{aligned}
I
= 0.
\end{aligned}
\end{equation}
\end{enumerate}
Thus, by taking into account \eqref{eq:case3}, only the term for which $k=l$ remains in the summation $F$:
\begin{equation}
\begin{aligned}
F = p_\ell \int\limits_{0}^{2\pi} \cos^2 (\ell\theta) \, \mathrm{d}\theta.
\end{aligned}
\end{equation}
This can be split into two cases. For $\ell = 0$ and by making use of \eqref{eq:case1} and \eqref{eq:integralF} we find:
\begin{equation}\label{c0}
p_0 = \dfrac{1}{2\pi} \int\limits_{0}^{2\pi} \phi(\cos\theta)\,\mathrm{d}\theta,
\end{equation}
and for $\ell > 0$, by making use of \eqref{eq:case2} and \eqref{eq:integralF}:
\begin{equation}\label{cl}
p_\ell = \dfrac{1}{\pi} \int\limits_{0}^{2\pi} \phi(\cos\theta)\cos(\ell\theta)\,\mathrm{d}\theta.
\end{equation}
For $\theta \in [0,2\pi]$, $\cos\theta$ lies in interval $[-1,1]$. Because of the position rescalings of the motion profile $\phi(x)$, the image $\phi(\cos\theta)$ also lies in the interval $[-1,1]$. Thus, we find:
\begin{equation}
|p_0| \leq \dfrac{1}{2\pi}\int\limits_{0}^{2\pi} |\phi(\cos\theta)|\,\mathrm{d}\theta \leq \dfrac{1}{2\pi}\int\limits_{0}^{2\pi} \,\mathrm{d}\theta = 1.
\end{equation}
and
\begin{equation}
|p_\ell| \leq \dfrac{1}{\pi}\int\limits_{0}^{2\pi} |\phi(\cos\theta)||\cos(\ell\theta)|\,\mathrm{d}\theta \leq \dfrac{1}{\pi}\int\limits_{0}^{2\pi}|\cos(\ell\theta)| \,\mathrm{d}\theta.
\end{equation}
To calculate this last integral, we use the periodicity of the function $\cos(\ell\theta)$. This function has a period of $2\pi / \ell$, so goes $\ell$ times up and down on the interval $[0,2\pi]$. So, after taking the absolute value of this function, we find $2\ell$ times the integral over the positive part of a period, for example, the interval $[-\pi/2\ell, \pi/2\ell]$:
\begin{equation}
\begin{aligned}
\dfrac{1}{\pi}\int\limits_{0}^{2\pi}|\cos(\ell\theta)| \,\mathrm{d}\theta = \dfrac{2\ell}{\pi} \int\limits_{-\pi/2\ell}^{\pi/2\ell} \cos(\ell\theta)\,\mathrm{d}\theta = \dfrac{4}{\pi}.
\end{aligned}
\end{equation}
Thus, the following bounds for the coefficients $p_i$ are obtained:
\begin{equation}\label{boundsDS}
|p_0| \leq 1 \hspace{0.5cm} \text{ and } \hspace{0.5cm} |p_\ell| \leq \dfrac{4}{\pi}, \hspace{0.3cm} \ell =1,\dots,n.
\end{equation}
These constraints on the design space simplify the subsequent optimization.
\section{Introduction}
\input{Sections/01_Introduction}
\input{Sections/02_SystemModelling}
\input{Sections/03_Identification}
\input{Sections/04_Optimization}
\input{Sections/05_Results}
\input{Sections/06_Conclusion}
\section*{Acknowledgements}
Research funded by a PhD grant of the Research Foundation Flanders (FWO) [1S88120N].
|
\section{Video link: Gunwale bobbing}
\url{https://youtu.be/XitE82v-mdY}
\section{Derivation of the wave-field for cruising}
In this section, we derive expressions for the wave-field height and wave resistance for a cruising canoe, which are based on the linear theory first developed by Havelock \cite{S_havelock1919wave}. To avoid following these derivations from first principles (since they are very long), we begin by referencing the main equations for the wave-field (in general form) and resistance taken from other literature, and continue thereon. Nevertheless, a full derivation of the theory of wave resistance can be found in \cite{S_havelock1919wave}.
We start with the expression for the wave-field, as formulated by Benzaquen {\it et al.} \cite{S_benzaquen2014wake}, such that
\begin{equation}
h_\mathcal{C}=- \frac{1}{4\pi^2}\lim_{\epsilon\rightarrow 0} \iint\limits_{0}^{\infty} \frac{\kappa \left( \mathcal{F} p_\mathcal{C}\right) }{\kappa-\mathrm{Fr}^2k_X^2+2i\epsilon k_X}\exp\left({-i(k_X X+k_y y)}\right)\, \mathrm{d}k_X \,\mathrm{d} k_y,\label{hfield22}
\end{equation}
where $k_X,k_y$, are the wavenumbers in the $X$ and $y$ directions, $\kappa=(k_X^2+k_y^2)^{1/2}$ is the wavenumber magnitude, $\epsilon$ is a dummy asymptotic variable and $\mathcal{F} p_\mathcal{C}$ is the Fourier transform of the cruising pressure disturbance, which is
\begin{equation}
\mathcal{F} p_\mathcal{C}=\frac{\delta}{\alpha\beta} \exp\left({-(k_X^2 +k_y^2 /\alpha^2)/4\pi^2}\right).
\end{equation}
We write \eqref{hfield22} in a form which is amenable to the Sokhotski-Plemelj theorem from complex analysis, such that
\begin{equation}
h_\mathcal{C}=- \frac{\delta}{4\pi^2 \alpha \beta }\lim_{{\epsilon}\rightarrow 0} \iint\limits_{0}^{\infty} \frac{({\kappa}/2{k}_X) }{({\kappa}-\mathrm{Fr}^2{k}_X^2)/2{k}_X+i{\epsilon}}\exp\left({-i({k}_X {X}+{k}_y {y})-({k}_X^2 +{k}_y^2/\alpha^2)/4\pi^2}\right)\, \mathrm{d}{k}_X \,\mathrm{d} {k}_y.\label{hfield2}
\end{equation}
By defining the functions
\begin{align}
A({k}_X,{k}_y)&=({\kappa}/2{k}_X) \exp\left({-i({k}_X {X}+{k}_y {y})-({k}_X^2 +{k}_y^2/\alpha^2)/4\pi^2}\right),\\
B({k}_X,{k}_y)&=({\kappa}-\mathrm{Fr}^2{k}_X^2)/2{k}_X,\label{Beqn}
\end{align}
the integral \eqref{hfield2} can be written (via a change of variables) as
\begin{equation}
h_\mathcal{C}=- \frac{\delta}{4\pi^2 \alpha \beta }\lim_{{\epsilon}\rightarrow 0} \int_0^\infty \int_{\infty}^{-\infty} \frac{A \left(\partial B/\partial {k}_X\right)^{-1} }{B+i{\epsilon}}\, \mathrm{d}B \,\mathrm{d} {k}_y.\label{hfield3}
\end{equation}
Hence, upon application of the Sokhotski-Plemelj theorem, this reduces to
\begin{equation}
h_\mathcal{C}\approx - \frac{\delta}{4\pi^2 \alpha \beta } \int_0^\infty -i\pi A\left[{k}^*_X({k}_y),{k}_y\right]\left\{\partial B/\partial {k}_X\left[{k}^*_X({k}_y),{k}_y\right]\right\}^{-1} \,\mathrm{d} {k}_y,\label{hfield4}
\end{equation}
where ${k}^*_X$ is the critical wave-number for which $B=0$, which is
\begin{equation}
{k}^*_X=\mathrm{Fr}^{-2}\left[{ 1/2+\left({ {1}/{4}+\mathrm{Fr}^4{k}_y^2}\right)^{1/2}} \right]^{1/2}.
\end{equation}
Note that \eqref{hfield4} ignores a second term in the Sokhotski-Plemelj theorem (proportional to the Cauchy principal value), since it is asymptotically small. The partial derivative of $B$ with respect to ${k}_X$ in \eqref{hfield4} is evaluated as \begin{equation}
\partial B/\partial {k}_X=\frac{1}{2{\kappa}}-\frac{{\kappa}}{2{k}_X^2}-\frac{\mathrm{Fr}^2}{2},
\end{equation}
which, evaluated at the critical wave-number is
\begin{equation}
\partial B/\partial {k}_X\left[{k}^*_X({k}_y),{k}_y\right]=\frac{1}{2\mathrm{Fr}^2\left[{k}^*_X({k}_y)\right]^2}-\mathrm{Fr}^2.
\end{equation}
Hence, the wave-field is ultimately written as
\begin{equation}
h_\mathcal{C}=\frac{-i \mathrm{Fr}^{2}\delta}{4\pi\alpha\beta}\int_0^\infty {F\left[{k}_X^*({k}_y),{k}_y,{X},{y}\right]}\,\mathrm{d}{k}_y,
\end{equation}
where the function
\begin{equation}
F= \frac{{\kappa}}{{k}_X(2\mathrm{Fr}^{4}-{k}_X^{-2})}\exp\left({{-\frac{1}{4\pi^2\alpha^2}\left({k}_X^2\alpha^{2}+{k}_y^2\right)-i\left( {k}_{X} {X}+{k}_{y} {y}\right)}}\right).
\end{equation}
The corresponding wave resistance is written (following Benham {\it et al.} \cite{S_benham2020hysteretic}, except using an inertial scaling $\rho U^2L^2$ as the normalisation instead of $mg=\rho g L^3/\alpha\beta $) as
\begin{equation}
R_{\mathcal{C},\mathcal{C}}=\frac{\delta^2}{4 \mathrm{Fr}^4 \alpha^2 \beta^2} \int_0^\infty C\left[{k}^*_X({k}_y),{k}_y\right]\left\{2\mathrm{Fr}^{-2}\,\partial B/\partial {k}_X\left[{k}^*_X({k}_y),{k}_y\right]\right\}^{-1} \,\mathrm{d} {k}_y,
\end{equation}
where the function
\begin{equation}
C({k}_X,{k}_y)={\kappa} \exp\left({-({k}_X^2 +{k}_y^2/\alpha^2)/2\pi^2}\right).
\end{equation}
Hence, the wave resistance simplifies to
\begin{equation}
R_{\mathcal{C},\mathcal{C}}=-\frac{\delta^2}{4 \alpha^2\beta^2}\int_0^\infty { G\left[{k}_X^*({k}_y),{k}_y\right]}\, \mathrm{d}{k}_y,
\end{equation}
where the function
\begin{align}
G= \frac{{\kappa} }{2\mathrm{Fr}^{4}-{k}_X^{-2}}\exp\left({{-\frac{1}{2\pi^2\alpha^2}({k}_X^2\alpha^2+{k}_y^2)}}\right).
\end{align}
\section{Derivation of the wave-field for oscillating heave or pitch}
Consider the linear wave equation
\begin{equation}
\mathrm{Fr}_\omega^{-2}\frac{\partial^2 h_n}{\partial t^2}-\nabla^2 h_n = \nabla^2 p_n,\quad n=\mathcal{H},\mathcal{P},\label{waveeq1}
\end{equation}
with source terms $p_n$ given by \eqref{heave_p}-\eqref{pitch_p} in the manuscript. We apply a Lorentz transformation
\begin{align}
{X'}&=\gamma (x-\mathrm{Fr}\,t)=\gamma X,\label{lorentz1}\\
{y'}&=y,\\
{t'}&=\gamma(t-\mathrm{Fr}\mathrm{Fr}_\omega^{-2}x)=t/\gamma-\gamma \mathrm{Fr}\mathrm{Fr}_\omega^{-2} X,\label{lorentz3}
\end{align}
where $\gamma=(1-\mathrm{Fr}^2/\mathrm{Fr}_\omega^2)^{-1/2}=(1-\mathrm{Ma}^2)^{-1/2}$.
Under the transformation \eqref{lorentz1}-\eqref{lorentz3}, the left hand side of the wave equation \eqref{waveeq1} is invariant. Hence, \eqref{waveeq1} is rewritten as
\begin{equation}
\mathrm{Fr}_\omega^{-2}\frac{\partial^2 h_n}{\partial {t'}^2}-{\nabla'}^2 h_n = \left( \gamma^2\frac{\partial^2}{\partial {X'}^2}+\frac{\partial^2}{\partial {y'}^2} \right) p_n\left[{{X'}}/{\gamma},{y'},\gamma\left({t'}+{\mathrm{Fr}\mathrm{Fr}_\omega^{-2}{X'}}\right)\right],\label{waveeqq}
\end{equation}
where the original time variable $t$ is replaced using the inverse transformation identity $t=\gamma({t'}+\mathrm{Fr}\mathrm{Fr}_\omega^{-2}{X'})$.
By inserting the expressions
\begin{align}
h_n&=\mathrm{Im}\left\{\gamma^{-1} \mathrm{Fr}_\omega^{2}h^*_ne^{i\gamma \mathrm{Fr}_\omega^{-1} t'}\right\},\label{hexpr}\\
p_n&= \mathrm{Im}\left\{ \gamma \mathrm{Fr}_\omega^{-2} p^*_ne^{i\gamma\mathrm{Fr}_\omega^{-1} t'}\right\},\label{pexpr}
\end{align}
into \eqref{waveeqq}, and by re-scaling the variables $(\tilde{X},\tilde{y})=\gamma\mathrm{Fr}_\omega^{-2} (X',y')$, we arrive at the Helmholtz equation
\begin{equation}
h^*_n+\tilde{\nabla}^2 h^*_n
=-\gamma^2\mathrm{Fr}_\omega^{-4}\left(\gamma^2\frac{\partial^2}{\partial \tilde{X}^2}+\frac{\partial^2}{\partial \tilde{y}^2}\right) p^*_n,\label{helm}
\end{equation}
where the complex source terms are
\begin{align}
p^*_\mathcal{H}&=\frac{\delta\mathrm{Fr}_\omega^{2}}{\beta \gamma} \exp\left({-{\mathrm{Fr}_\omega^{4}(\tilde{X}^2+(\alpha\gamma)^2\tilde{y}^2)}/{2\gamma^{4}}+i \mathrm{Ma} \tilde{X}}\right),\\
p^*_\mathcal{P}&=\frac{\delta\mathrm{Fr}_\omega^{4}}{\beta \gamma^3} {\tilde{X}}\exp\left({i\theta-{\mathrm{Fr}_\omega^{4}(\tilde{X}^2+(\alpha\gamma)^2\tilde{y}^2)}/{2\gamma^{4}}+i \mathrm{Ma} \tilde{X}}\right).
\end{align}
As described by Devauchelle {\it et al.} \cite{S_devauchelle2020walkers}, the solution to \eqref{helm} is
\begin{equation}
h^*_n=\frac{i}{4}\iint\limits_{-\infty}^{+\infty} H_0^{(1)}(|\tilde{\mathbf{X}}-\tilde{\boldsymbol{\mathcal{X}}}|) \gamma^2\mathrm{Fr}_\omega^{-4}\left(\gamma^2\frac{\partial^2}{\partial \tilde{\mathcal{X}}^2}+\frac{\partial^2}{\partial \tilde{\mathcal{Y}}^2}\right) p^*_n(\tilde{\boldsymbol{\mathcal{X}}})\,\mathrm{d}\tilde{\boldsymbol{\mathcal{X}}},
\end{equation}
written in terms of the Hankel function of the first kind and zeroth order $H_0^{(1)}$ (which is the Green's function for the Helmholtz equation) and integrated over dummy variables $\tilde{\boldsymbol{\mathcal{X}}}=(\tilde{\mathcal{X}},\tilde{\mathcal{Y}})$.
To acquire the final solution for the wave-field, we re-write expressions \eqref{hexpr} and \eqref{pexpr} in terms of the original time variable $t$, noting that
\begin{equation}
i\mathrm{Fr}_\omega^{-1} \gamma t'=i\mathrm{Fr}_\omega^{-1} t-i\tilde{X}\mathrm{Fr}\mathrm{Fr}_\omega^{-1}=i\mathrm{Fr}_\omega^{-1} t-i\mathrm{Ma}\tilde{X}.
\end{equation}
Hence, the wave-field and pressure source are given by
\begin{align}
h_n&=\mathrm{Im}\left\{\gamma^{-1}\mathrm{Fr}_\omega^{2}\bar{h}_n(\tilde{X},\tilde{y})\,e^{i\mathrm{Fr}_\omega^{-1} t}\right\},\\
p_n&=\mathrm{Im}\left\{\gamma \mathrm{Fr}_\omega^{-2}\bar{p}_n(\tilde{X},\tilde{y})\, e^{i\mathrm{Fr}_\omega^{-1} t}\right\},
\end{align}
where
\begin{align}
\bar{h}_n&=h^*_n e^{-i \mathrm{Ma} \tilde{X}}=\left[\frac{i}{4}\iint\limits_{-\infty}^{+\infty} H_0^{(1)}(|\tilde{\mathbf{X}}-\tilde{\boldsymbol{\mathcal{X}}}|)\mathcal{L}\bar{p}_n(\tilde{\boldsymbol{\mathcal{X}}}) e^{i\mathrm{Ma}\tilde{{\mathcal{X}}}} \,\mathrm{d}\tilde{\boldsymbol{\mathcal{X}}}\right]e^{-i\mathrm{Ma}\tilde{X}},\\
\bar{p}_n&=p^*_n e^{-i \mathrm{Ma} \tilde{X}}=\begin{cases}
\frac{\delta\mathrm{Fr}_\omega^{2}}{\beta \gamma} \exp\left({-{\mathrm{Fr}_\omega^{4}(\tilde{X}^2+(\alpha\gamma)^2\tilde{y}^2)}/{2\gamma^{4}}}\right),\quad & n=\mathcal{H},\\
& \\
\frac{\delta\mathrm{Fr}_\omega^{4}}{\beta \gamma^3} {\tilde{X}}\exp\left({i\theta-{\mathrm{Fr}_\omega^{4}(\tilde{X}^2+(\alpha\gamma)^2\tilde{y}^2)}/{2\gamma^{4}}}\right),\quad & n=\mathcal{P},
\end{cases}
\end{align}
where the operator $\mathcal{L}=\gamma^2\mathrm{Fr}_\omega^{-4}(\gamma^2\partial^2/\partial \tilde{\mathcal{X}}^2+\partial^2/\partial \tilde{\mathcal{Y}}^2)$.
Now to calculate the force $R_{n,m}$, we start with the expression for the horizontally resolved pressure force
\begin{equation}
R_{n,m}=\frac{1}{\mathrm{Fr}^2T}\int_0^T\left[\iint\limits_{-\infty}^{+\infty} p_n\frac{\partial h_m}{\partial X}\,\mathrm{d}X\,\mathrm{d}y\right]\,\mathrm{d}t.\label{bouncefor2}
\end{equation}
Noting that
\begin{equation}
\frac{\partial h_n}{\partial X}=\mathrm{Im}\left\{\gamma \frac{\partial \bar{h}_n}{\partial \tilde{X}} \,e^{i\mathrm{Fr}_\omega^{-1} t}\right\},
\end{equation}
the force \eqref{bouncefor2} is given in dimensionless terms as
\begin{equation}
R_{n,m}=
\frac{\mathrm{Fr}_\omega^{2}}{\mathrm{Fr}^2\gamma T}\int_0^T\iint\limits_{-\infty}^{+\infty}\mathrm{Im}\left\{ \bar{p}_ne^{i\mathrm{Fr}_\omega^{-1} t}\right\} \mathrm{Im}\left\{ \frac{\partial \bar{h}_m}{\partial \tilde{X}}e^{i\mathrm{Fr}_\omega^{-1} t}\right\} \,\mathrm{d}\tilde{X}\,\mathrm{d}\tilde{y}\,\mathrm{d}t.
\end{equation}
However, by noticing that the time average of any product of the form
\begin{equation}
\frac{1}{T}\int_0^T \mathrm{Im}\left\{ A e^{i\mathrm{Fr}_\omega^{-1} t}\right\} \mathrm{Im}\left\{ B e^{i\mathrm{Fr}_\omega^{-1} t}\right\} \,\mathrm{d}t=\frac{1}{2T}\int_0^T \mathrm{Im}\left\{ A\right\} \mathrm{Im}\left\{ B \right\} + \mathrm{Re}\left\{ A\right\} \mathrm{Re}\left\{ B \right\} \,\mathrm{d}t,
\end{equation}
then we arrive at the final expression
\begin{equation}
R_{n,m}=\frac{\mathrm{Fr}_\omega^2}{2\mathrm{Fr}^2\gamma}\iint \limits_{-\infty}^{+\infty}\mathrm{Re} \,\bar{p}_n \,\mathrm{Re} \frac{\partial \bar{h}_m}{\partial \tilde{X}}+\mathrm{Im}\, \bar{p}_n \,\mathrm{Im} \frac{\partial \bar{h}_m}{\partial \tilde{X}} \,\mathrm{d}\tilde{X}\,\mathrm{d}\tilde{y}.
\end{equation}
\section{Canoe surface area and natural frequencies}
For the sake of simplicity we approximate the surface area of the canoe by considering two tetrahedrons stuck together with the same total dimensions, as illustrated in Fig. \ref{canoe}. The wetted area of one of the exposed triangles of the tetrahedron surface can be calculated through trigonometry, giving
\begin{equation}
A=\frac{1}{8}\left[ L^2 W^2 + 4 D^2 (L^2 + W^2) \right]^{1/2}.
\end{equation}
Hence, in dimensionless coordinates, the total surface area (four triangles) is
\begin{equation}
\mathcal{S}=\left[\frac{1}{4\alpha^2}+ \frac{1}{\beta^2} \left( 1 + \frac{1}{\alpha^2}\right) \right]^{1/2}.
\end{equation}
Next, we use this approximate canoe shape to estimate the natural frequencies for heaving and pitching. This is done by considering the simple harmonic motion of small perturbations to the vertical position of the centre of mass and pitch angle.
To calculate these natural frequencies it is assumed that the canoe is not cruising, such that $\mathrm{Fr}=0$, $U=0$.
Whilst all of the analysis in this study has so far remained dimensionless, we keep the subsequent analysis dimensional for convenience. To make this clear we have capitalised many variables so as not to be confused with their dimensionless counterparts in the manuscript. The two variables for which this doesn't apply are $x$ and $t$, whose capitals already have prescribed definitions in the manuscript, so we replace these with hatted capitals, $\hat{X},\hat{T}$, for clarity.
\begin{figure}
\centering
\begin{tikzpicture}[scale=1.55]
\node at (7.3,-8.5) {\includegraphics[width=0.5\textwidth]{canoe.png}};
\node[blue] at (9.3,-8.2) {\large$\boldsymbol{L}$};
\node[blue] at (7.5,-7.8) {\large$\boldsymbol{W}$};
\node[blue] at (3.9,-8.3) {\large$\boldsymbol{D}$};
\draw[line width=1.5,<->,blue] (4.55,-7.9) -- (9.85,-8.5);
\draw[line width=1.5,<->,blue] (4.25,-8.7) -- (4.25,-7.95);
\draw[line width=1.5,<->,blue] (7.35,-8) -- (6.9,-8.35);
\end{tikzpicture}
\caption{Canoe shape approximated by two tetrahedrons with total length, width and draft, $L$, $W$ and $D$.\label{canoe}}
\end{figure}
Let the vertical position of the centre of mass (measured from the waterline) and the pitch angle of the canoe be denoted by $Z(\hat{T})$ and $\varphi(\hat{T})$ respectively. The equations of motion for each of these are given by conservation of linear and angular momentum in the heave and pitch directions, such that
\begin{align}
(m+m_a)\ddot{Z}&=-\iint_S (P-P_a) (\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}) \,\mathrm{d}S-m g,\label{vertf}\\
(I+I_a)\ddot{\varphi}&=\iint_S (P-P_a) ( \mathbf{r}\times\hat{\mathbf{n}})\cdot \hat{\boldsymbol{\jmath}} \, \mathrm{d}S,\label{angf}
\end{align}
where $P_a$ is atmospheric pressure, $\mathbf{r}$ is the position vector measured from the origin, $\hat{\mathbf{n}}$ is the outward pointing unit normal vector to the hull surface (where $S=L^2\mathcal{S}$), $m$ is the mass of the canoe, $m_a$ and $I_a$ are the linear and angular added masses in the heave and pitch directions \cite{S_ursell1949heaving}, and $I$ is the moment of inertia which is defined as
\begin{equation}
I=\rho_0\iiint_V |\mathbf{r}|^2 \,\mathrm{d}V,\label{momi}
\end{equation}
in terms of the hull material density $\rho_0$ (assumed constant), and volume $V$.
For simplicity we have ignored the presence of a gunwale bobber in \eqref{momi} and instead we have assumed that the mass of the canoe is distributed uniformly over its hull.
We first consider the case of pure heaving ($\varphi=0,Z=Z(\hat{T})$) and attempt to find an approximate expression for the natural frequency. In this case, the centre of mass is perturbed vertically according to
\begin{equation}
Z=Z_0+\zeta(\hat{T}),
\end{equation}
where $Z_0$ is a constant and $|\zeta/Z_0|\ll 1$. Likewise, if $Z=-\Gamma(\hat{X},Y)$ is the shape function of the canoe composed of two tetrahedrons, then this is also perturbed by $Z=-\Gamma(\hat{X},Y)+ \zeta(\hat{T})$.
Since the oscillations are of small amplitude, the pressure on the hull surface is approximately hydrostatic
\begin{equation}
P\approx P_a-\rho g (-\Gamma+ \zeta) .
\end{equation}
Hence, the leading order and first order terms of \eqref{vertf} reduce to
\begin{align}
0&=-\rho g \iint_S (\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}) \Gamma \,\mathrm{d}S-mg,\label{masseq}\\
(m+m_a)\ddot{\zeta}&=\left[\rho g\iint_S (\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}) \, \mathrm{d}S\right] \zeta,
\end{align}
the first of which sets the mass of the boat and the second of which determines the perturbation dynamics. These are neatly rearranged to give a simple harmonic oscillator equation for the small perturbation, such that
\begin{equation}
\ddot{\zeta}+ \omega^2 \zeta=0,
\end{equation}
where the natural frequency is
\begin{equation}
\omega^2=\frac{ g\iint_S (\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}})\,\mathrm{d}S}{ \iint_S (\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}) \Gamma \,\mathrm{d}S(1+m_a/m)}.\label{nat1}
\end{equation}
The added mass ratio $m_a/m$ for the tetrahedron shape is unknown, so instead (as a very approximate estimate) we use the value for a heaving ellipse with the same length-depth aspect ratio (see Ref. \cite{S_newman2018marine}), which is
\begin{equation}
\frac{m_a}{m}=\frac{\pi\rho L^2/4}{\pi \rho LD/2}=\frac{\beta}{2}.
\end{equation}
Due to the symmetries of the tetrahedra, it suffices to consider just one of the outer triangular faces to calculate \eqref{nat1}. Taking $\Gamma$ as the planar outer face
\begin{equation}
\Gamma=D\left[1+\frac{2\hat{X}}{L}+\frac{2Y}{W}\right],\label{quarter}
\end{equation}
which is defined in the range $Y\in[-W(1/2 + \hat{X}/L),0]$, $\hat{X}\in[-L/2,0]$, we calculate
\begin{equation}
\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}=-\left[1+\frac{4}{\beta^2}(1+\alpha^2)\right]^{-1/2}=-\frac{1}{2\alpha \mathcal{S}(\alpha,\beta)},
\end{equation}
which is a constant. Now the natural heaving frequency $\omega$ may be calculated by evaluating the integrals in \eqref{nat1}. After simplification and re-writing in terms of the oscillating Froude number, this provides the result
\begin{equation}
\mathrm{Fr}_\omega=\frac{1}{\omega}\left(\frac{g}{L}\right)^{1/2}=\left(\frac{1+\beta/2}{3\beta}\right)^{1/2}.\label{heavenat}
\end{equation}
Next we repeat the above analysis for the case of pure pitching ($\zeta=0,\varphi=\varphi(\hat{T}) $), centred about the origin. A small perturbation $\varphi\ll1$ is applied to the pitch angle of the boat, causing the vertical position of the hull to be moved to $Z\approx -\Gamma+\varphi \hat{X}$.
To evaluate the right hand side of \eqref{angf}, we need to calculate the cross product of the radial vector with the pressure force vector. Each of these vectors is given by
\begin{align}
\mathbf{r}&=\left\{\hat{X},Y,-\Gamma+\varphi \hat{X}\right\},\\
(P-P_a)\hat{\mathbf{n}}&=\frac{-\rho g (-\Gamma+\varphi \hat{X})}{\left[ \left( -\frac{\partial \Gamma}{\partial \hat{X}}+\varphi\right)^2+\frac{\partial \Gamma}{\partial Y}^2+1 \right]^{1/2}}\left\{ -\frac{\partial \Gamma}{\partial \hat{X}}+\varphi,-\frac{\partial \Gamma}{\partial Y},-1\right\}.
\end{align}
Next we perform the cross product, expand out the variables in powers of $\varphi$, and integrate over the surface area of the hull (note all anti-symmetric terms vanish upon integration), ignoring terms of order $\mathcal{O}(\varphi^2)$. This gives the result
\begin{equation}
\begin{split}
&\iint_S \left( P-P_a\right) \left( \mathbf{r}\times \hat{\mathbf{n}}\right)\cdot \hat{\boldsymbol{\jmath}} \, \mathrm{d}S\approx \\
&-\left[\rho g \iint_S \frac{ \hat{X}^2\left( 1+\frac{\partial \Gamma}{\partial \hat{X}}^2+\frac{\partial \Gamma}{\partial Y}^2\right)+\Gamma^2\left( 1+\frac{\partial \Gamma}{\partial Y}^2\right)+\hat{X}\Gamma\frac{\partial \Gamma}{\partial \hat{X}}\left( 1+2\frac{\partial \Gamma}{\partial \hat{X}}^2+2\frac{\partial \Gamma}{\partial Y}^2\right)}{\left[ \frac{\partial \Gamma}{\partial \hat{X}}^2+\frac{\partial \Gamma}{\partial Y}^2+1 \right]^{3/2}} \, \mathrm{d}S\right] \varphi .\label{angf2}
\end{split}
\end{equation}
The next step is to calculate the moment of inertia in \eqref{angf}. To do so we assume that the hull mass is distributed over a thin solid shell of vertical thickness $H\ll \Gamma$, such that the volume element is approximately $dV\approx H dS$ and the hull density $\rho_0 ={m}/{H S}$. Hence, the moment of inertia \eqref{momi} is given by
\begin{equation}
I\approx \frac{m}{S}\iint_S (\hat{X}^2+Y^2+\Gamma^2) \,\mathrm{d}S.\label{momi1}
\end{equation}
Note that only leading order terms are kept in \eqref{momi1} since the left hand side of \eqref{angf} is already of order $\mathcal{O}(\varphi)$.
Using \eqref{masseq} to replace $m$ (at leading order) with an expression for the mass, then \eqref{momi1} becomes
\begin{equation}
I\approx\frac{\rho}{S}\iint_S (\hat{X}^2+Y^2+\Gamma^2) \,\mathrm{d}S\iint_S -(\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}) \Gamma \,\mathrm{d}S.
\end{equation}
Hence, the equation of motion \eqref{angf} reduces to a simple harmonic oscillator of the form
\begin{equation}
\ddot{\varphi}+ \omega^2\varphi = 0,
\end{equation}
where
\begin{equation}
\begin{split}
\omega^2=\frac{gS \iint_S { \left[\hat{X}^2\left( 1+\frac{\partial \Gamma}{\partial \hat{X}}^2+\frac{\partial \Gamma}{\partial Y}^2\right)+\Gamma^2\left( 1+\frac{\partial \Gamma}{\partial Y}^2\right) + \hat{X}\Gamma\frac{\partial \Gamma}{\partial \hat{X}}\left( 1+2\frac{\partial \Gamma}{\partial \hat{X}}^2+2\frac{\partial \Gamma}{\partial Y}^2\right)\right]}{\left[ \frac{\partial \Gamma}{\partial \hat{X}}^2+\frac{\partial \Gamma}{\partial Y}^2+1 \right]^{-3/2}} \, \mathrm{d}S}{
{\iint_S (\hat{X}^2+Y^2+\Gamma^2) \,\mathrm{d}S\iint_S -(\hat{\mathbf{n}}\cdot \hat{\boldsymbol{k}}) \Gamma\,\mathrm{d}S}(1+I_a/I)}.\label{compint}
\end{split}
\end{equation}
Since the angular added mass ratio $I_a/I$ is unknown for the tetrahedron shape, we use the value for an ellipse with the same length-depth aspect ratio as a very approximate estimate, which is
\begin{equation}
\frac{I_a}{I}=\frac{1/8\pi\rho (L^2/4-D^2)^2}{1/8\pi \rho LD (L^2/4+D^2)}=\frac{ (\beta^2-4)^2}{ 4\beta (\beta^2+4)}.
\end{equation}
As before, we exploit the symmetry of the tetrahedra to evaluate the integrals in \eqref{compint} by considering only one of the outer triangular faces of the surface \eqref{quarter}. In this way the the natural frequency for pitching may be calculated analytically, and we write this (after simplification) in terms of the oscillating Froude number, which is
\begin{equation}
\mathrm{Fr}_\omega=\left[\frac{
\left( 4 + 4 \alpha^2 + \beta^2\right)^{1/2} \left( \beta^2 + \alpha^2 \left( 4 + \beta^2\right)\rb}{3 \alpha^2 \left( -16 + \left( 6 + 4 \alpha^2\right) \beta^2 + \beta^4\right)}\left( 1+ \frac{ (\beta^2-4)^2}{ 4\beta (\beta^2+4)}\right) \right]^{1/2}.\label{pitchnat}
\end{equation}
Hence, inserting $\alpha=5$, $\beta=31$, into the (very approximate) expressions for the natural oscillating Froude number \eqref{heavenat},\eqref{pitchnat}, we calculate $\mathrm{Fr}_\omega=0.42$ in the case of heaving and $\mathrm{Fr}_\omega=0.30$ in the case of pitching.
\section{Accelerometer data}
\begin{figure}
\centering
\begin{tikzpicture}[scale=1]
\node at (0,0) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-122}};
\node at (4,0) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-123}};
\node at (8,0) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-124}};
\node at (12,0) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-125}};
\node at (0,-4) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-126}};
\node at (4,-4) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-127}};
\node at (8,-4) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-128}};
\node at (12,-4) {\includegraphics[width=0.2\textwidth,trim={10cm 0 0 0},clip]{image-129}};
\end{tikzpicture}
\caption{Photographs of Jerome A. Neufeld gunwale bobbing on his canoe of dimensions $L,W,D=4.7,0.94,0.15\un{m}$, taken at intervals of 0.1 s. \label{jeromepics}}
\end{figure}
The gunwale bobbing experiments were performed by Jerome A. Neufeld and Miles Neufeld on Muldrew Lake in Ontario, Canada in August 2021. The accelerometer data used in our study was taken by Jerome A. Neufeld using the \textit{Accelerometer} app on an \textit{iPhone 7} whilst bobbing on a canoe of dimensions $L,W,D=4.7,0.94,0.15\un{m}$. The photographs in Figs. \ref{pitch}a and \ref{thewaves}e of the main text corresponded to gunwale bobbing performed by Miles Neufeld on a paddle board of dimensions $L,W,D=3.05,0.76,0.10\un{m}$. Photographs of Jerome A. Neufeld on his canoe are illustrated here in Fig. \ref{jeromepics} and a video is uploaded as a separate file.
Data for the vertical acceleration during gunwale bobbing over five different trials are plotted in Fig. \ref{expdata}. Aside from the trials corresponding to the displayed accelerometer data, a total eight further gunwale bobbing trials were performed by Jerome A. Neufeld over a measured distance of $24.73$ m between two jetties on the lake to measure the speed of the canoe. The times to complete this journey were $T=21.5,21.1,24.66,32.98,29.15,47.53,21.41,20.4$ s. Hence, the average speed plus or minus one standard deviation was $U=0.97\pm0.24$ m/s.
\begin{figure}
\centering
\begin{tikzpicture}[scale=1]
\node at (0,0) {\includegraphics[width=0.9\textwidth]{exp_extra_1}};
\node at (0,-2.5) {\includegraphics[width=0.9\textwidth]{exp_extra_2}};
\node at (0,-5) {\includegraphics[width=0.9\textwidth]{exp_extra_3}};
\node at (0,-7.5) {\includegraphics[width=0.9\textwidth]{exp_extra_4}};
\node at (0,-10.3) {\includegraphics[width=0.9\textwidth]{exp_extra_5}};
\end{tikzpicture}
\caption{Experimental data for the vertical acceleration (normalised by
$g$) measured using an accelerometer over five separate trials. \label{expdata}}
\end{figure}
\section{Modelling profile drag}
In this section we briefly describe how we model profile drag, which is given by \eqref{Rskin} in the manuscript.
As described by \cite{S_boucher2018thin}, there is a significant contribution to the drag on a canoe from viscous friction at the wetted surface, and from the form drag due to vortex shedding. The skin and form drag are summed together and modelled with a combined profile drag term $R_d$, given in terms of the dimensionless wetted surface area $\mathcal{S}$ and a drag coefficient $C_d$ (see \eqref{Rskin} in the manuscript).
Following \cite{S_boucher2018thin}, the drag coefficient is approximated by the empirical relationship
\begin{equation}
C_d=C_f(1+2/\alpha+60/\alpha^4),\label{dragapprox}
\end{equation}
where $C_f$ is the skin friction coefficient for a flat plate \cite{S_hoerner1965practical}. This varies weakly with the Reynolds number $\mathrm{Re}=UL/\nu$, where $\nu$ is the kinematic viscosity, and is approximated for turbulent flows \cite{S_hadler1958coefficients} as
\begin{equation}
C_f={0.075}(\log\mathrm{Re}-2)^{-2}.
\end{equation}
For example, a $4.7\un{m}$ canoe cruising at 1 m/s in water corresponds to a Reynolds number of $4.7\times10^6$, producing a skin friction value $C_f=3.4\times 10^{-3}$.
|
\section{Introduction and main results}\label{Intro}
Let $\Omega \subset \re^N$\tcr{, $N \ge 2$,} be a domain with $a \in \Omega$ and $1 < p <N$.
The Hardy inequality
\begin{equation}
\label{H_p}
\( \frac{N-p}{p} \)^p \int_{\Omega} \frac{|u|^p}{|x -a|^p} dx \le \int_{\Omega} | \nabla u |^p dx,
\end{equation}
holds for all $u \in \dot{W}^{1,p}_0(\Omega)$, where $\dot{W}_0^{1,p}(\Omega)$ is \tcr{the homogeneous Sobolev space defined as} the completion of $C_c^{\infty}(\Omega)$
with respect to the \tcr{(semi-)}norm $\| \nabla (\cdot )\|_{L^p(\Omega)}$.
It is well-known that $(\frac{N-p}{p})^p$ is the best constant and is not attained.
Hardy's best constant $\( \frac{N-p}{p} \)^p$ plays an important role \tcr{in investigating qualitative} properties of solutions to elliptic \tcr{or} parabolic partial differential equations,
\tcr{such as, stability, instantaneous blow-up, and global-in-time asymptotics,} see \tcr{for example} \cite{BG, BV}.
On the other hand, in the limiting case where $p=N$, the Hardy inequality (\ref{H_p}) looks degenerate as the best constant vanishes.
However \tcr{in this case,} we can obtain the critical Hardy inequality on bounded domains:
\begin{align}\label{H_N}
\( \frac{N-1}{N} \)^N \int_{\Omega} \frac{|u|^N}{|x-a|^N \( \log \frac{R}{|x-a|} \)^N} dx
\le \int_{\Omega} | \nabla u |^p dx \quad \( u \in \dot{W}_0^{1,N}(\Omega)\)
\end{align}
as a limit of the Hardy inequality (\ref{H_p}) as $p \nearrow N$, where $R=\sup_{x \in \Omega} |x-a|$,
see \cite{I} or I.-(ii) in \S \ref{S Harmonic}.
It is also known that $(\frac{N-1}{N})^N$ is the best constant and is not attained, see e.g. \cite{AS,II, B-TF}.
In the present paper, we introduce \tcr{a} critical Hardy inequality \tcr{similar to} (\ref{H_N}) when $\Omega$ is unbounded, especially the half-space
$\re^N_+ =\{ (x,y) \,| \, x \in \re^{N-1}, y>0 \}$.
Note that if $\Omega = \re^N$, this kind of inequality does not hold even if we restrict functions to radially symmetric \tcr{ones}, see Proposition \ref{Prop re^N} in \S \ref{S App}.
Our first result is the following.
\begin{theorem}\label{Thm CH}(\tcr{Critical} Hardy inequality on the half-space)
Let $N \ge 2$. Then the inequality
\begin{align}\label{H_N half}
\( \frac{N-1}{N} \)^N \int_{\re^N_+} &\frac{|u(x,y)|^N}{\( |x|^2 +(1-y)^2 \)^{\frac{N}{2}} \( \frac{|x|^2 +(1+y)^2}{4} \)^{\frac{N}{2}} \( \log \sqrt{ \frac{|x|^2 +(1+y)^2}{|x|^2 +(1-y)^2}} \)^N } \,dxdy \notag \\
&\le \int_{\re^N_+} |\nabla u(x,y)|^N \,dxdy
\end{align}
holds for any $u \in \dot{W}_0^{1,N}(\re^N_+)$. Furthermore, $( \frac{N-1}{N})^N$ is the best constant and is not attained.
\end{theorem}
\begin{remark}(Asymptotic behavior of the potential function)
Set
\begin{equation}
\label{V_N}
V_N (x, y) := \frac{1}{\( |x|^2 +(1-y)^2 \) \( \frac{|x|^2 +(1+y)^2}{4} \) \( \log \sqrt{ \frac{|x|^2 +(1+y)^2}{|x|^2 +(1-y)^2}} \)^2 }.
\end{equation}
\tcr{Then the inequality (\ref{H_N half}) is of the form
\begin{align*}
\( \frac{N-1}{N} \)^N \int_{\re^N_+} V_N(x,y)^{\frac{N}{2}} |u(x,y)|^N \, dxdy \le \int_{\re^N_+} |\nabla u(x,y)|^N \,dxdy.
\end{align*}
The inequality (\ref{H_N half}) has two aspects:
one is the critical Hardy inequality on bounded domains and the other is the geometric Hardy inequality on $\re^N_+$,
which involves \tcb{the distance from the boundary} $\pd \re^N_{+}$.
\tcb{Indeed,} the potential function $V_N(x,y)^{\frac{N}{2}}$ behaves like $\( |x|^2 +(1-y)^2 \)^{-\frac{N}{2}} \( \log \frac{2}{\sqrt{|x|^2 +(1-y)^2}} \)^{-N}$ when $(x, y)$ is near to $e_N = (0,1) \in \re^N_{+}$,
which is similar to the critical Hardy potential on the ball $B_2(e_N)$ with radius $2$ and center $e_N$.
Also $V_N(x,y)^{\frac{N}{2}}$ behaves like $y^{-N} = {\rm dist }((x,y), \pd\re^N_{+})^{-N}$ near the boundary $\pd \re^N_+$ or $\infty$, which is similar to the geometric Hardy potential on $\re^N_+$.
In fact, since $Y:=\frac{|x|^2 +(1+y)^2}{|x|^2 +(1-y)^2}=1 + o(1)$ as $|x|^2 +(y-1)^2 \to \infty$ or $y \to 0$ and $\log Y = Y-1 + o(1)$ as $Y \to 1$, we have
\begin{align*}
V_N(x,y)^{\frac{N}{2}}
&= \frac{1}{\( |x|^2 +(1-y)^2 \)^{\frac{N}{2}} \( \frac{|x|^2 +(1+y)^2}{4} \)^{\frac{N}{2}} \( \log \sqrt{Y} \)^N } \\
&= \frac{1}{\( |x|^2 +(1-y)^2 \)^{\frac{N}{2}} \( \frac{|x|^2 +(1+y)^2}{4} \)^{\frac{N}{2}} \( \frac{Y-1}{2} \)^N } + o(1) \\
&= \frac{\( |x|^2 +(1-y)^2 \)^{\frac{N}{2}}}{ y^N \( |x|^2 +(1+y)^2 \)^{\frac{N}{2}} } + o(1) \\
& = O(y^{-N})
\end{align*}
}
as $|x|^2 +(y-1)^2 \to \infty$ or $y \to 0$.
\end{remark}
Next, we give an improvement of (\ref{H_p}) which \tcr{yields} (\ref{H_N half}) as $p \nearrow N$.
Improvements of the Hardy and the Hardy-Sobolev inequalities on balls are studied for radially symmetric functions in \cite{FKR,I,S(ArXiv)}.
However, on the half-space $\re^N_+$, we cannot consider \tcr{radial} symmetry
since radial functions which are zero on the boundary $\pd \re^N_+$ \tcr{must be identically zero.}
Instead of \tcr{radial} symmetry, we introduce the following \tcr{new} symmetry for functions $u=u(x,y)$ on $\re^N_+$:
\tcr{Put}
\begin{align}\label{def U}
U_p (x,y)=
\begin{cases}
\frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}}
\left[ \( |x|^2 +(1-y)^2 \)^{-\frac{N-p}{2(p-1)}} - \( |x|^2 +(1+y)^2 \)^{-\frac{N-p}{2(p-1)}} \right]
\, &\text{if} \, p \in (1,N), \\
\w_{N-1}^{-\frac{1}{N-1}} \log \sqrt{ \frac{ |x|^2 +(1+y)^2}{ |x|^2 +(1-y)^2}}
&\text{if}\, p=N,
\end{cases}
\end{align}
\tcr{where} $\omega_{N-1}$ is the area of the unit sphere $\mathbb{S}^{N-1}$ in $\re^N$.
\tcr{Note that the function $U_p(x,y)$ is obtained from the fundamental solution of the $p$-Laplacian on $\re^N$ with singularity $e_N = (0, 1)$, by ``reflecting the singularity" with respect to the boundary $\pd \re^N_{+}$.
So $U_p \equiv 0$ on $\pd \re^N_{+}$.
However, it is different from $p$-Green's function $G_{\re^N_+, e_N}(x,y)$ on $\re^N_{+}$ when $p \not= 2$ and $p \not= N$, see \S \ref{S Green}.}
\tcr{We consider functions on $\re^N_{+}$ of the form}
\begin{align}\label{*'}
u(x,y) = \tilde{u}(s),\,\text{where}\,\, s= U_p(x,y), \quad \tcr{(x, y) \in \re^N_{+},}
\end{align}
\tcr{for some function $\tilde{u}$ on $\re$ with the property $\tilde{u}(0) = 0$.}
\tcr{In the following, with some ambiguity, we identify $\tilde{u}$ as $u$ and write, for example, $u(x, y) = u(s)$, $s = U_p(x,y)$ for $(x, y) \in \re^N_{+}$.}
\tcr{Thus a function of the form \eqref{*'} has the same value on each level set of $U_p$ and vanishes on $\pd \re^N_{+}$.}
Our second result is an improvement of the Hardy inequality (\ref{H_p}) on $\re^N_+$ for functions with the symmetry (\ref{*'}).
\begin{theorem}(\tcr{Improved} Hardy inequality for $p \ge 2$)
\label{Thm IH}
Let $2 \le p < N$.
Then the inequality
\begin{align}
\label{IH}
\( \frac{N-p}{p} \)^p &\int_{\re^N_+} \frac{V_p(x,y)^{\frac{p}{2}}}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} |u(x,y)|^p \,dxdy
\le \int_{\re^N_+} |\nabla u(x,y)|^p \,dxdy
\end{align}
holds for any $u \in \dot{W}_0^{1,p} (\re^N_+)$ \tcr{of the form} (\ref{*'}),
where
\tcr{
\begin{align}
\label{V_p}
\begin{cases}
&V_p(x,y) = \frac{1+ X^{\frac{N-1}{p-1}} -2X^{\frac{N-p}{2(p-1)}} \( |x|^2 +(1+y)^2 \)^{\tcrr{-1}} (|x|^2 +y^2 -1)}
{\left[ 1-X^{\frac{N-p}{2(p-1)}} \right]^{2}}, \\
&X = \frac{|x|^2 +(1-y)^2}{|x|^2 + (1+y)^2} \in [0, 1).
\end{cases}
\end{align}
}
Furthermore, $(\frac{N-p}{p})^p$ is the best constant and is not attained.
\end{theorem}
\begin{remark}\label{Rem without sym}
\tcrr{Actually, the inequality (\ref{IH}) holds for functions without the symmetry (\ref{*'}) by combining Proposition \ref{Prop sol U} and a result in \cite{DD}, see Theorem \ref{Thm IH without sym} in \S \ref{S Proof}. Our method is based on the harmonic transplantation.}
\end{remark}
\begin{remark}($V_p \ge 1$)
\tcr{We remark here} that (\ref{IH}) is an improvement of the Hardy inequality (\ref{H_p}) \tcr{with $a = e_N$} since $V_p (x, y) \ge 1$.
In fact, for any $(x,y) \in \re^N_+ \cap \overline{B_1^N}$, we have
$V_p(x,y) \ge 1+ X^{\frac{N-1}{p-1}} \ge1$.
Also, for any $(x,y) \in \re^N_+ \setminus \overline{B_1^N}$,
we have
\begin{align*}
V_p(x,y) &=
1+ \frac{X^{\frac{N-p}{p-1}}}{(1-X^{\frac{N-p}{2(p-1)}} )^2}
\left[ X-1+ X^{- \frac{N-p}{2(p-1)}}\frac{4(y+1)}{|x|^2 + (1+y)^2} \right] \\
&= 1+ \frac{X^{\frac{N-p}{p-1}}}{(1-X^{\frac{N-p}{2(p-1)}} )^2}
\left[ \left\{ X^{- \frac{N-p}{2(p-1)}} -1 \right\} \frac{4y}{|x|^2 + (1+y)^2} + X^{- \frac{N-p}{2(p-1)}}\frac{4}{|x|^2 + (1+y)^2} \right] \\
&\ge 1
\end{align*}
\tcr{since $X \in [0,1)$.}
\end{remark}
\begin{remark}
Unlike (\ref{H_p}), it is possible to take the limit \tcr{$p \nearrow N$ in} the improved Hardy inequality (\ref{IH}).
\tcr{Thus we obtain Theorem \ref{Thm CH} (for functions with symmetry \eqref{*'}) from Theorem \ref{Thm IH} in this way.}
In fact, since $1-X^s = s \log \frac{1}{X} + \tcr{o(s)}$ as $s \to 0$,
\tcr{taking $s = \frac{N-p}{2(p-1)}$,} we see that \tcr{$V_p$ in \eqref{V_p} satisfies}
\begin{align*}
&\( \frac{N-p}{p} \)^p \frac{V_p(x,y)^{\frac{p}{2}}}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}} } \\
&= \( \frac{N-p}{p} \)^p
\frac{\left\{ 1+X -2 \( |x|^2 +(y+1)^2 \)^{-1} (|x|^2+ y^2 -1) \right\}^{\frac{N}{2}}}{ \( |x|^2 +(1-y)^2 \)^{\frac{N}{2}}
\left[ \frac{N-p}{2(p-1)} \log \frac{1}{X} \right]^{p}} + o(1) \\
&= \( \frac{N-1}{N} \)^N V_N (x,y)^{\frac{N}{2}} + o(1) \quad (p \nearrow N)
\end{align*}
\tcr{where $V_N$ is defined in \eqref{V_N}.}
Therefore, we obtain (\ref{H_N half}) as a limit of (\ref{IH}) as $p \nearrow N$.
\tcr{However, note that, Theorem \ref{Thm CH} is proved by another method in \S \ref{S Proof} and valid for functions without any symmetry.}
\end{remark}
This paper is organized as follows:
In \S \ref{S Green}, we show propositions about the function $U_p(x,y)$ in \eqref{def U}, which coincides with \tcr{$p$-}Green's function \tcr{$G_{\re^N_+, e_N}(x,y)$} when $p=2$ or $N$.
Although $U_p$ is \tcr{different from $G_{\re^N_+, e_N}$} for $p \in (2, N)$, \tcr{we can prove that} $U_p$ is superharmonic for $p \in (2, N)$ on \tcr{$\re^N_+ \setminus \{ e_N \}$.}
This is a key point of the proof of Theorem \ref{Thm IH}.
In \S \ref{S Transformation}, we recall \tcr{the} M\"obius transformation and \tcr{the} harmonic transplantation proposed by Hersch \cite{H}.
We point out that various transformations so far \tcr{appeared in references} can be understood as a special or a general case of harmonic transplantation.
Also, we explain the difference between \tcr{two transformations.}
In \S \ref{S Proof}, we show Theorem \ref{Thm CH} by \tcr{exploiting} the M\"obius transformation.
Due to the lack of the explicit form of \tcr{$p$-}Green's function \tcr{$G_{\re^N_+, e_N}$} for $p \in (2,N)$,
it seems difficult to apply the original harmonic transplantation which exploits the $p$-Green's functions,
to obtain an improvement of the Hardy inequality on the half-space $\re^N_+$, see Theorem \ref{Thm psi} in \S \ref{S Transformation}.
We use $U_p$ in \eqref{def U} instead of $p$-Green's function $G_{\re^N_{+}, e_N}$ to define a modified version of the harmonic transplantation.
By usng this new transformation, we show Theorem \ref{Thm IH}.
In the last of \S \ref{S Proof}, we mention that these transformations can be also applicable to Sobolev type inequalities.
In \S \ref{S App}, we show several propositions related to main theorems and give \tcr{an} application of \tcr{a special type} of harmonic transplantation.
We fix several notations:
$B_R$ or $B_R^N$ denotes the $N$-dimensional ball centered $0$ with radius $R$. As a matter of convenience, we set $B_\infty^N = \re^N$ and $\frac{1}{\infty} = 0$.
$\omega_{N-1}$ denotes the area of the unit sphere $\mathbb{S}^{N-1}$ in $\re^N$.
$[f > \ep]$ denotes the set $\{ (x,y) \in \re^N_+ \,|\, f(x,y) >\ep \}$.
$|A|$ denotes the Lebesgue measure of a set $A \subset \re^N$.
\section{Green's function on the half-space}\label{S Green}
Let $G_{\Omega, a} = G_{\Omega, a} (z): \Omega \setminus \{ a\} \to \re$ be the \tcr{$p$-}Green function
with singularity at $a \in \Omega$ associated with $p$-Laplace operator $\lap_p (\cdot)= {\rm div }(|\nabla (\cdot ) |^{p-2} \nabla (\cdot ))$.
Namely, $G_{\Omega, a} (z)$ satisfies
\begin{align}
\label{lap_p}
\begin{cases}
-\lap_p G_{\Omega, a}(z) = \delta_{a}(z), &\quad z \in \Omega,\\
\quad \,\,\,\, G_{\Omega, a} (z) = 0 \,\,&\quad z \in \pd \Omega,
\end{cases}
\end{align}
where $\delta_a$ is the Dirac measure giving unit mass to a point $a \in \Omega$.
When $\Omega = \re^N_+$ and \tcr{$a= e_N = (0,1)$}, we have
\begin{align}
\label{G half}
\tcr{G_{\re^N_+, e_N}(x,y)} =
\begin{cases}
\frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \( |x|^2 +(1-y)^2 \)^{-\frac{N-p}{2(p-1)}} - \psi_p (x, y) \right] \, &\text{if} \, p \in (1,N), \\
\w_{N-1}^{-\frac{1}{N-1}} \log \sqrt{ \frac{ |x|^2 +(1+y)^2}{ |x|^2 +(1-y)^2}} &\text{if}\, p=N,
\end{cases}
\end{align}
where \tcr{$\psi_p$ is a function with} $\psi_p \in L^\infty_{\rm loc} (\re^N_+)$,
$\lim_{|x|^2 + (y-1)^2 \to 0} \( |x|^2 +(1-y)^2 \)^{\frac{N-1}{2(p-1)}} \nabla \psi_p (x, y) = 0$,
and $\psi_2 (x,y)= \( |x|^2 +(1+y)^2 \)^{-\frac{N-2}{2}}$
(see \cite{KV}).
\tcr{Note} that $U_p$ \tcr{in \eqref{def U} coincides with} $G_{\re^N_+, e_N}$ for $p=2$ or $p=N$.
To the best of our knowledge, we do not know the explicit form of $\psi_p$ when $p \not= 2$ and $p \not= N$.
\tcr{This fact causes some difficulty in the application of harmonic transplantation in \S \ref{S Transformation} on $\re^N_{+}$,
since we need the explicit form of Green's function in the use of harmonic transplantation.}
However, fortunately, we see that $U_p$ is a super (or sub) solution of (\ref{lap_p}) in the distributional sense \tcr{according to the range of $p$} as follows.
This fact enables us to use $U_p$ instead of $G_{\re^N_{+}, e_N}$ in the proof of Theorem \ref{Thm IH}.
\begin{prop}\label{Prop sol U}
Let $1 < p \le N$ \tcr{and let $U_p$ be as in \eqref{def U}.}
Then for any $\phi \in C_c^\infty (\re^N_+)$ with $\phi \ge 0$,
\begin{align}
\label{eq:sol U}
\int_{\re^N_+} | \nabla U_p|^{p-2} \nabla U_p \cdot \nabla \phi \,dxdy
&= \phi (0,1) + \int_{\re^N_+ } (-\lap_p U)\, \phi \, dxdy \\
&\begin{cases}
\le \phi (0, 1) \quad &\text{if} \,\, p \in (1, 2], \\
\ge \phi (0, 1) &\text{if} \,\, p \in [2, N), \\
= \phi (0, 1) &\text{if} \,\, p = N.
\end{cases} \notag
\end{align}
\end{prop}
Proposition \ref{Prop sol U} follows from Proposition \ref{Prop cal U}.
\begin{prop}
\label{Prop cal U}
\tcr{Let $1 < p \le N$ and let $U_p$ be as in \eqref{def U}.}
Then for $(x,y) \in \re^N_+ \setminus \{ (0,1) \}$, we have the followings:
(I) Let $1<p<N$.
Then
\begin{align*}
-\lap_p U_p
&= \frac{(N-p)(p-2)}{(p-1)^2 \, \w_{N-1}^{\frac{2}{p-1}}} \, |\nabla U_p|^{p-4} U_p
\left[ |x|^2 + (y-1)^2 \right]^{-\frac{N-p}{2(p-1)} -1} \left[ |x|^2 + (y+1)^2 \right]^{-\frac{N-p}{2(p-1)} -1} \\
&\hspace{3em}\times \left[ N-p + (N+p-2) \frac{(|x|^2+ y^2 -1)^2}{\{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} } \right].
\end{align*}
(II) $-\lap_N U_N = 0$.
\tcr{Especially, we see that the pointwise estimates
$$
\begin{cases}
-\Delta_p U_p \le 0 &\quad \text{on} \ \re^N_{+} \setminus \{ e_N \}, \quad (1 < p \le 2), \\
-\Delta_p U_p \ge 0 &\quad \text{on} \ \re^N_{+} \setminus \{ e_N \}, \quad (2 \le p < N), \\
-\Delta_p U_p = 0 &\quad \text{on} \ \re^N_{+} \setminus \{ e_N \}, \quad (p = N)
\end{cases}
$$
hold.}
\end{prop}
\begin{proof}(Proof of Proposition \ref{Prop cal U})
\noindent
(I) For $(x,y) \in \re^N_+ \setminus \{ (0,1) \}$, we have
\begin{align*}
\w_{N-1}^{\frac{1}{p-1}} \nabla U_p = -\left[ |x|^2 + (y-1)^2 \right]^{-\frac{N-p}{2(p-1)}-1}
\begin{pmatrix}
x \\
y-1
\end{pmatrix}
+ \left[ |x|^2 + (y+1)^2 \right]^{-\frac{N-p}{2(p-1)} -1}
\begin{pmatrix}
x \\
y+1
\end{pmatrix}
\end{align*}
which implies that
\begin{align*}
|\nabla U_p|^2
&= \w_{N-1}^{-\frac{2}{p-1}} \Biggl[ \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1}-1}
+ \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1}-1} \\
&\hspace{1em}- 2 \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\{ |x|^2 +y^2 -1 \} \Biggl].
\end{align*}
\tcr{We put $V = |\nabla U_p|^2$.}
Then we have
\begin{align*}
&{\rm div}(|\nabla U_p |^{p-2} \nabla U_p)
= {\rm div}(V^{\frac{p-2}{2}} \nabla U_p)
= V^{\frac{p-4}{2}} \left[ V \lap U_p + \frac{p-2}{2} \nabla V \cdot \nabla U_p \right], \\
&\lap U_p = {\rm div} (\nabla U_p)
= -\frac{(N-1)(p-2)}{(p-1)\, \w_{N-1}^{\frac{1}{p-1}}}
\left[ \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} - \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \right].
\end{align*}
Also, we have
\begin{align*}
&\w_{N-1}^{\frac{2}{p-1}} \nabla V \\
&=-2 \,\frac{N-1}{p-1} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1}-2}
\begin{pmatrix}
x \\
y-1
\end{pmatrix}
- 2 \, \frac{N-1}{p-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1}-2}
\begin{pmatrix}
x \\
y+1
\end{pmatrix} \\
&-4 \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\begin{pmatrix}
x \\
y
\end{pmatrix} \\
&+2 \, \( \frac{N-p}{p-1}+2 \) \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-2}
\left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} (|x|^2 +y^2 -1)
\begin{pmatrix}
x \\
y-1
\end{pmatrix} \\
&+2 \, \( \frac{N-p}{p-1}+2 \) \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-2} (|x|^2 +y^2-1)
\begin{pmatrix}
x \\
y+1
\end{pmatrix}
\end{align*}
which implies that
\begin{align*}
&\frac{p-2}{2} \nabla V \cdot \nabla U_p \\
&= \frac{p-2}{2} \w_{N-1}^{-\frac{3}{p-1}} \Biggl[\left[ |x|^2 + (y-1)^2 \right]^{-\frac{N-p}{2(p-1)}-1}
\begin{pmatrix}
x \\
y-1
\end{pmatrix}
- \left[ |x|^2 + (y+1)^2 \right]^{-\frac{N-p}{2(p-1)} -1}
\begin{pmatrix}
x \\
y+1
\end{pmatrix} \Biggl] \cdot \\
&\Biggl[2 \,\frac{N-1}{p-1} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1} -2}
\begin{pmatrix}
x \\
y-1
\end{pmatrix}
+ 2 \, \frac{N-1}{p-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1}-2}
\begin{pmatrix}
x \\
y+1
\end{pmatrix} \\
&+4 \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \left\{ x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\begin{pmatrix}
x \\
y
\end{pmatrix} \\
&-2 \, \( \frac{N-p}{p-1} +2 \) \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-2}
\left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)} -1} (|x|^2 +y^2 -1)
\begin{pmatrix}
x \\
y-1
\end{pmatrix} \\
&-2 \, \( \frac{N-p}{p-1} +2 \) \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-2} (|x|^2 +y^2 -1)
\begin{pmatrix}
x \\
y+1
\end{pmatrix} \Biggl].
\end{align*}
Therefore, we have
\begin{align*}
&\frac{p-2}{2} \nabla V \cdot \nabla U_p \\
&=\frac{p-2}{2} \w_{N-1}^{-\frac{3}{p-1}} \Biggl[ 2 \,\frac{N-1}{p-1}
\left[ |x|^2 + (y-1)^2 \right]^{-\frac{3(N-p)}{2(p-1)} -2} \\
&+2 \,\frac{N-1}{p-1}
\left[ |x|^2 + (y-1)^2 \right]^{-\frac{N-p}{2(p-1)} -1} \left[ |x|^2 + (y+1)^2 \right]^{-\frac{N-p}{p-1} -2} (|x|^2 +y^2 -1) \\
&+4 \left[ |x|^2 + (y-1)^2 \right]^{-\frac{N-p}{p-1} -2} \left[ |x|^2 + (y+1)^2 \right]^{-\frac{N-p}{2(p-1)} -1} (|x|^2 +y^2 -y) \\
&-2 \, \( \frac{N-1}{p-1} +1 \) \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1} -2} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)} -1} (|x|^2 +y^2 -1) \\
&-2 \, \( \frac{N-1}{p-1} +1 \) \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1} -2} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)} -2} (|x|^2 +y^2 -1)^2 \\
&-2 \, \frac{N-1}{p-1} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1} -2} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)} -1} (|x|^2 +y^2 -1) \\
&-2 \,\frac{N-1}{p-1}
\left[ |x|^2 + (y+1)^2 \right]^{-\frac{3(N-p)}{2(p-1)} -2}
+4 \left[ |x|^2 + (y-1)^2 \right]^{-\frac{N-p}{2(p-1)} -1} \left[ |x|^2 + (y+1)^2 \right]^{-\frac{N-p}{2(p-1)} -2} (|x|^2 +y^2 +y) \\
&+2 \, \( \frac{N-1}{p-1} +1 \) \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)} -2} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1} -2} (|x|^2 +y^2 -1)^2 \\
&+2 \, \( \frac{N-1}{p-1} +1 \) \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)} -1} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1} -2} (|x|^2 +y^2 -1)
\Biggl].
\end{align*}
Since
\begin{align*}
V \lap U_p = - \w_{N-1}^{-\frac{3}{p-1}} \frac{p-2}{2} \,
&\Biggl[ 2 \,\frac{N-1}{p-1} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{3(N-p)}{2(p-1)}-2} \\
&- 2 \,\frac{N-1}{p-1} \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1}-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \\
&+ 2 \,\frac{N-1}{p-1} \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1}-1} \\
&- 2 \,\frac{N-1}{p-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{3(N-p)}{2(p-1)}-2} \\
&-4 \,\frac{N-1}{p-1} \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1} -2}
\left\{ x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} (|x|^2 +y^2 -1) \\
&+4 \,\frac{N-1}{p-1} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1}-2} (|x|^2 + y^2 - 1) \Biggl],
\end{align*}
we have
\begin{align*}
&{\rm div}(|\nabla U_p |^{p-2} \nabla U_p) V^{-\frac{p-4}{2}} \w_{N-1}^{\frac{3}{p-1}} \frac{2}{p-2}
= \left[ V \lap U_p + \frac{p-2}{2} \nabla V \cdot \nabla U_p \right] \w_{N-1}^{\frac{3}{p-1}} \frac{2}{p-2} \\
&= -2 \,\frac{N-1}{p-1} \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} U_p \frac{N-p}{p-1} \, \w_{N-1}^{\frac{1}{p-1}} \\
&+2 \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}
\Biggl[ \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} (|x|^2 +y^2 -1) \\
&- \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} (|x|^2 +y^2 -1)
+2 \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} (|x|^2 +y^2 -y) \\
&- 2 \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} (|x|^2 +y^2 +y)
\Biggl] \\
&- 2 \( \frac{N-1}{p-1} +1 \) \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)} -2}
\left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)} -2} (|x|^2 +y^2 -1)^2 U_p\\
&=- \frac{2(N-p) \, \w_{N-1}^{\frac{1}{p-1}}}{(p-1)^2} U_p
\left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)} -1} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)} -1} \\
&\left[ N-p + (N+p-2) \frac{(|x|^2+ y^2 -1)^2}{\{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} } \right]
\end{align*}
\tcr{which implies Proposition \ref{Prop cal U} (I).}
\noindent
(II) The proof is done by direct calculation in the same way as (I). We omit the proof here.
\begin{comment}
The proof of (II):
For $(x,y) \in \re^N_+ \setminus \{ (0, \cdots , 0, 1) \}$, we have
\begin{align*}
\nabla U_N = \frac{\w_{N-1}^{-\frac{1}{N-1}} }{|x|^2 + (y+1)^2} \(
\begin{array}{c}
x \\
y+1
\end{array}
\) -\frac{\w_{N-1}^{-\frac{1}{N-1}}}{|x|^2 + (y-1)^2} \(
\begin{array}{c}
x \\
y-1
\end{array}
\)
\end{align*}
which implies that
\begin{align*}
|\nabla U_N |^2
&= \w_{N-1}^{-\frac{2}{N-1}} \left[ \frac{1}{|x|^2 + (y+1)^2} + \frac{1}{|x|^2 + (y-1)^2} - \frac{2(|x|^2 +y^2 -1)}{ \{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} } \right] \\
&= \frac{4\w_{N-1}^{-\frac{2}{N-1}} }{ \{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} } =: V(x,y).
\end{align*}
Then we have
\begin{align*}
{\rm div}(|\nabla U_N |^{N-2} \nabla U_N)
&= {\rm div}(V^{\frac{N-2}{2}} \nabla U_N)
= V^{\frac{N-4}{2}} \left[ V \lap U_N + \frac{N-2}{2} \nabla V \cdot \nabla U_N \right], \\
\lap U_N &= {\rm div} (\nabla U_N) = \frac{-4y \, (N-2) \, \w_{N-1}^{-\frac{1}{N-1}} }{ \{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} }.
\end{align*}
Also, we have
\begin{align*}
\nabla V =
\frac{-8\, \w_{N-1}^{-\frac{2}{N-1}}}{ \{ |x|^2 + (y-1)^2 \}^2 \{ |x|^2 + (y+1)^2 \} } \(
\begin{array}{c}
x \\
y-1
\end{array}
\)- \frac{8\, \w_{N-1}^{-\frac{2}{N-1}}}{ \{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \}^2 } \(
\begin{array}{c}
x \\
y+1
\end{array}
\)
\end{align*}
which implies that
\begin{align*}
&\frac{N-2}{2} \nabla V \cdot \nabla U_N \\
&= \frac{-4 (N-2) \w_{N-1}^{-\frac{3}{N-1}}}{ \{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} } \left[ \frac{1}{|x|^2 + (y+1)^2} \(
\begin{array}{c}
x \\
y+1
\end{array}
\) -\frac{1}{|x|^2 + (y-1)^2} \(
\begin{array}{c}
x \\
y-1
\end{array}
\) \right] \cdot \\
&\hspace{1em} \left[ \frac{1}{|x|^2 + (y+1)^2} \(
\begin{array}{c}
x \\
y+1
\end{array}
\) +\frac{1}{|x|^2 + (y-1)^2} \(
\begin{array}{c}
x \\
y-1
\end{array}
\) \right] \\
&= \frac{16y \, (N-2)\w_{N-1}^{-\frac{3}{N-1}} }{ \{ |x|^2 + (y-1)^2 \}^2 \{ |x|^2 + (y+1)^2 \}^2 }.
\end{align*}
Hence, we have
\begin{align*}
{\rm div}(|\nabla U_N |^{N-2} \nabla v)
&= \( \frac{4 \, \w_{N-1}^{-\frac{2}{N-1}}}{ \{ |x|^2 + (y-1)^2 \} \{ |x|^2 + (y+1)^2 \} } \)^{\frac{N-4}{2}} \\
&\left[ \frac{-16y \, (N-2) \, \w_{N-1}^{-\frac{3}{N-1}} }{ \{ |x|^2 + (y-1)^2 \}^2 \{ |x|^2 + (y+1)^2 \}^2 }
+ \frac{16y (N-2) \, \w_{N-1}^{-\frac{3}{N-1}} }{ \{ |x|^2 + (y-1)^2 \}^2 \{ |x|^2 + (y+1)^2 \}^2 } \right] =0
\end{align*}
\end{comment}
\end{proof}
\begin{proof}(Proof of Proposition \ref{Prop sol U})
\tcr{Let $B_\ep(e_N)$ be the ball with center $e_N$ and radius $\ep$.}
For any $\phi \in C_c^{\infty} (\re^N_+)$, we have
\begin{align}
&\int_{\re^N_+} | \nabla U_p |^{p-2} \nabla U_p \cdot \nabla \phi \, dxdy \notag \\
\label{PU0}
&= \int_{\pd B_\ep(e_N)} | \nabla U_p |^{p-2} ( \nabla U_p \cdot \nu )\,\phi \, dS
+ \int_{\re^N_+ \setminus B_\ep(e_N)} (-\lap_p U_p)\, \phi \, dxdy
+ \int_{B_\ep(e_N)} | \nabla U_p |^{p-2} \nabla U_p \cdot \nabla \phi \, dxdy.
\end{align}
where $\nu = -(x, y-1)^{T} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{1}{2}}$.
\tcr{We claim}
\begin{align}\label{1st term}
\int_{\pd B_\ep(e_N)} | \nabla U_p |^{p-2} ( \nabla U_p \cdot \nu )\,\phi \, dS
&= \phi (0,1) + o(1)\quad (\ep \to 0), \\
\label{3rd term}
\int_{B_\ep(e_N)} | \nabla U_p |^{p-2} \nabla U_p \cdot \nabla \phi \, dxdy
&= o(1) \quad (\ep \to 0),
\end{align}
Indeed, a direct calculation shows that
\begin{align*}
&\int_{\pd B_\ep(e_N)} | \nabla U_p |^{p-2} ( \nabla U_p \cdot \nu )\,\phi \, dS\\
&= \w_{N-1}^{-1} \int_{\pd B_\ep} \Biggl[ \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{p-1}-1}
+ \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{p-1}-1}
- 2 \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \\
&\left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \{ |x|^2 +y^2 -1 \} \Biggl]^{\frac{p-2}{2}}
\Biggl[ \left\{|x|^2 + (y-1)^2 \right\}^{-\frac{N-p}{2(p-1)} } - \left\{|x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1} \\
&\{ |x|^2 +y^2 -1 \} \Biggl] \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{1}{2}} \,\phi \, dS\\
&= \w_{N-1}^{-1} \ep^{-1-\frac{N-p + (N-1)(p-2)}{p-1}}
\int_{\pd B_\ep(e_N)} \Biggl[1 + \left\{ \frac{\ep^2}{|x|^2 + (y+1)^2} \right\}^{\frac{N-1}{p-1}}
- 2 \ep^{-\frac{N-1}{p-1}-1} \left\{ |x|^2 + (y+1)^2 \right\}^{-\frac{N-p}{2(p-1)}-1}\\
&\{ |x|^2 +y^2 -1 \} \Biggl]^{\frac{p-2}{2}}
\Biggl[1 - \left\{ \frac{\ep^2}{|x|^2 + (y+1)^2} \right\}^{\frac{N-p}{2(p-1)}} \frac{|x|^2 +y^2 -1}{|x|^2 +(y+1)^2} \Biggl] \,\phi \, dS\\
&= \phi (0,1) + o(1) \quad (\ep \to 0)
\end{align*}
which implies (\ref{1st term}).
On the other hand, we \tcr{see}
\begin{align*}
\left| \, \int_{B_\ep(e_N)} | \nabla U_p |^{p-2} \nabla U_p \cdot \nabla \phi \, dxdy\, \right|
&\le C \, \| \nabla \phi \|_\infty \int_{B_\ep(e_N)} \left\{ |x|^2 + (y-1)^2 \right\}^{-\frac{N-1}{2}} \,dxdy \\
&= C \, \| \nabla \phi \|_\infty \, \w_{N-1} \int_0^\ep \, dr = o(1) \quad (\ep \to 0),
\end{align*}
\tcr{which proves \eqref{3rd term}.}
\tcr{
Finally, we check that the second term in \eqref{PU0} satisfies
\begin{equation}
\label{2nd term}
\int_{\re^N_+ \setminus B_\ep(e_N)} (-\lap_p U_p)\, \phi \, dxdy \to \int_{\re^N_+ } (-\lap_p U_p)\, \phi \, dxdy
\end{equation}
as $\ep \to 0$.
Actually, we have
\begin{align*}
|\Delta_p U_p| &= O\( |\nabla U_p|^{p-4} U_p (|x|^2 + (y-1)^2 )^{-\frac{N-p}{2(p-1)}-1} \), \\
|\nabla U_p|^{p-4} &= O\( (|x|^2 + (y-1)^2 )^{(-\frac{N-p}{p-1}-1)(\frac{p-4}{2})} \), \\
|U_p| &= O\( (|x|^2 + (y-1)^2 )^{-\frac{N-p}{2(p-1)}} \)
\end{align*}
near $(x, y) = (0, 1)$ by Proposition \ref{Prop cal U}.
Thus we have
\begin{align*}
|\Delta_p U_p| &= O\( (|x|^2 + (y-1)^2 )^{(-\frac{N-p}{p-1}-1)(\frac{p-4}{2}) - \frac{N-p}{2(p-1)} - \frac{N-p}{2(p-1)}-1} \) \\
&= O\( \(\sqrt{|x|^2 + (y-1)^2}\)^{(-\frac{N-p}{p-1}-1)(p-4) - \frac{N-p}{(p-1)} - \frac{N-p}{(p-1)}-2} \) \\
&= O\( \(\sqrt{|x|^2 + (y-1)^2}\)^{-\frac{(N-1)(p-2)}{p-1}} \)
\end{align*}
near $(x, y) = (0, 1)$.
This is locally integrable if
$$
\frac{(N-1)(p-2)}{p-1} < N
$$
which always holds for $p \in (1, N]$.
Thus \eqref{2nd term} follows from Lebesgue's dominated convergence theorem.
}
\tcr{Returning to \eqref{PU0} with \eqref{1st term}, \eqref{3rd term}, and \eqref{2nd term},
we obtain \eqref{eq:sol U}.}
\end{proof}
\section{M\"obius transformation and harmonic transplantation}\label{S Transformation}
In this section, we recall M\"obius transformation and harmonic transplantation.
Both transformations preserve the norm $\| \nabla (\cdot) \|_p$ and coincide in the critical case $p=N$.
However in the subcritical case $p<N$, these transformations are different \tcr{from} each other.
\subsection{M\"obius transformation}\label{S Mobius}
First, we recall \tcr{the definition of} M\"obius transformation and its properties.
\begin{definition}\label{def M}(M\"obius transformation)
For $b \in \re^N$, $\la >0$, $R \in O(N)$, where $O(N)$ is the orthogonal group in \tcr{$\re^N$}, set
\begin{align*}
T_b (z) &= z+b\quad (\text{translation}),\\
S_\la (z) &= \la z \quad (\text{scaling}),\\
R(z) &= Rz \quad (\text{rotation}),\\
J(z) &= z^* = \frac{z}{|z|^2} \quad (\text{reflection}).
\end{align*}
A {\it M\"obius transformation} $M: \re^N \to \re^N$ is a finite composition of $T_b, S_\la, R$ and $J$.
Also, the group of M\"obius transformations is denoted by $M(\re^N)$.
\end{definition}
\begin{remark}\label{Jacobian}
Set $u \otimes v = (u_i v_j)_{1 \le i, j \le N}$ for $u=(u_1, \cdots, u_N)^T$ and $v=(v_1, \cdots, v_N)^T$.
The \tcr{differential} and the Jacobian of each transformation are as follows.
\begin{align*}
(T_b)' (z) &= I, \,\,\det (T_b)'(z) =1\\
(S_\la)' (z) &= \la I, \,\, \det (S_\la)'(z) =\la^N\\
R'(z) &= R, \,\, \det R' (z) = \det R = \pm 1\\
J'(z) &= \frac{1}{|z|^2}\( I - 2 \frac{z}{|z|} \otimes \frac{z}{|z|} \), \,\, \det J'(z) = \frac{(-1)}{|z|^{2N}}
\end{align*}
\tcr{where $I$ is the identity matrix on $\re^N$.}
For the proof of the last one, see Proposition \ref{Prop detJ}.
\end{remark}
\tcr{For a function $f: \re^N \to \re$, we set}
\begin{align}\label{M trans}
(M^{\#} f ) (z) = |\det M' (z)|^{\frac{N-p}{Np}} f(M(z))
\end{align}
for $z \in \re^N$.
\tcr{We call $M^{\sharp} f$ is also the M\"obius transformation of the function $f$.}
Then we see that the transformation $M^{\#}$ preserves several quantities as follows.
\begin{prop}\label{Prop Mobius}
Let $1\le p \le N$ and $p^* = \frac{Np}{N-p}$ for $p < N$.
If $M \in M(\re^N)$, then
\begin{align}\label{nabla M}
\int_{\re^N} |\nabla (M^{\#} f) (z) |^p\,dz &= \int_{\re^N} |\nabla f (w)|^p \,dw \quad \text{for} \,\, p=2 \,\text{or} \,N,\\
\label{p^* M}
\int_{\re^N} |(M^{\#} f) (z) |^{p^*}\,dz &= \int_{\re^N} |f (w)|^{p^*} \,dw \quad \text{for} \,\, p <N, \\
\label{Hardy M}
\int_{\re^N} \frac{|(M^{\#} f) (z) |^{p}}{|z|^p}\,dz &= \int_{\re^N} \frac{|f (w)|^{p}}{|w|^p} \,dw
\end{align}
hold for any $f \in C_c^1 (\re^N \setminus \{ 0 \})$.
\end{prop}
\begin{remark}\label{Kelvin}
From Remark \ref{Jacobian}, we have
\begin{align*}
(T_b^{\#} \,f) (z) &= f(z+b),\\
(S_\la^{\#}\, f) (z) &= \la^{\frac{N-p}{p}} f(\la z),\\
(R^{\#} f) (z) &= f(Rz),\\
(J^{\#} f )(z) &= |z|^{\frac{2}{p} (p-N)} f \( \frac{z}{|z|^2} \).
\end{align*}
The last transformation is called \tcr{the} Kelvin transformation when $p=2$ or $N$.
In the case where $p \not=2$ and $p \not= N$, there is no radial function $\rho= \rho (|z|) \not\equiv 0$ such that
\begin{align*}
\int_{\re^N} |\nabla g(z) |^p\,dz = \int_{\re^N} |\nabla f (w)|^p \,dw \,\,\text{holds for}\,\, g(z) = \rho (|z|) f \( \frac{z}{|z|^2} \),
\end{align*}
see the proof below. \tcb{Furthermore, all transformations except for $J^{\#}$ above preserve the $p$-harmonicity of functions: $\Delta_p f = 0$ implies $\Delta_p (M^{\#} f) = 0$, where $M$ is one of $T_b$, $S_{\la}$, and $R$. Also $J^{\#}$ preserves the $p-$harmonicity of functions when $p=2$ or $p=N$. }
\tcr{When $p \ne 2$ and $p \ne N$, it is shown in \cite{Lind} that there is no radial function $\rho$ such that $\Delta_p g = 0$, $g$ as above, for any function $f$ satisfying $\Delta_p f = 0$.}
\end{remark}
\begin{proof} \tcr{(Proof of Proposition \ref{Prop Mobius})}
We can easily show (\ref{p^* M}) and (\ref{Hardy M}). We show (\ref{nabla M}) only.
First, we claim that \tcr{\eqref{nabla M} holds for each transformation $T_b, S_\la, R, J$.}
We shall show (\ref{nabla M}) only \tcr{for} $M=J$.
We use the polar coordinate $z= r\w, r=|z|, \w \in \mathbb{S}^{N-1}$.
Then we have $w := Jz = s \omega$, $|w| = s = r^{-1}$, and
\begin{align*}
(J^{\#}f )(r\w) =\rho (r) f (s \w), \quad \text{where}
\,\, \rho (r)= r^{\frac{2}{p} (p-N)}.
\end{align*}
Therefore, we have
\begin{align*}
&\int_{\re^N} |\nabla (J^{\#} f) (z) |^p\,dz \\
&= \int_0^\infty \int_{\mathbb{S}^{N-1} } \left[ \left| \,\frac{\pd (J^{\#}f )}{\pd r} \,\right|^2 + \frac{1}{r^2} |\nabla_{\mathbb{S}^{N-1}} (J^{\#}f) |^2 \right]^{\frac{p}{2}} r^{N-1} \,dr dS_{\w} \\
&= \iint \left[ \left| -\frac{\pd f}{\pd s} \rho (r) r^{-2} + \rho' (r) f \,\right|^2 + \tcr{\rho (r)^2} r^{-2} |\nabla_{\mathbb{S}^{N-1}} f |^2 \right]^{\frac{p}{2}} r^{N-1} \,dr dS_{\w} \\
&= \iint \left[ \left| \frac{\pd f}{\pd s} - \frac{\rho' (r) r^2}{\rho (r)} f \,\right|^2 + r^2 |\nabla_{\mathbb{S}^{N-1}} f |^2 \right]^{\frac{p}{2}} \rho(r)^{p} r^{-2p + N-1} \,dr dS_{\w} \\
&= \iint \left[ \left| \,\frac{\pd f}{\pd s} \, \right|^2 -\frac{\rho'(r)}{\rho (r) s^2} \frac{\pd}{\pd s} (f^2) + \frac{|\rho'(r)|^2}{\rho (r)^2 s^4} f^2 + \frac{1}{s^2} |\nabla_{\mathbb{S}^{N-1}} f |^2 \right]^{\frac{p}{2}} \rho(r)^p r^{2(N-p)} s^{N-1}\,ds dS_{\w}
\end{align*}
\tcr{where $r = s^{-1}$.}
Since $\rho (r)^p r^{2(N-p)} =1$ and $\rho (r) =1$ for $p=N$, we obtain (\ref{nabla M}) for $p=N$.
In the case where $p=2$, by the integration by parts, we have
\begin{align*}
\int_{\re^N} |\nabla (J^{\#} f) (z) |^2\,dz
&= \int_{\re^N} |\nabla f (w) |^2 \,dw + \tcr{(N-2)^2} \iint \( \frac{\pd}{\pd s} \( \frac{1}{s} \) f^2 + \frac{1}{s^2} f^2 \) s^{N-1}\,ds dS_{\w}\\
&= \int_{\re^N} |\nabla f (w) |^2 \,dw.
\end{align*}
Therefore, we obtain the claim. Let $A, B \in \{ T_b, S_\la, R, J\}$.
Since
\begin{align*}
(A \circ B)^{\#} (z)
&=|\det (A \circ B)' (z)|^{\frac{N-p}{Np}} f((A \circ B)(z)) \\
&=|\det A' ( B (z) ) \cdot \det B'(z)|^{\frac{N-p}{Np}} f(A (B(z) ) ) \\
&=|\det B'(z)|^{\frac{N-p}{Np}} (A^{\#}f )(B(z) ) \\
&= [(B^{\#} \circ A^{\#}) f] (z),
\end{align*}
we have
\begin{align*}
(A \circ B)^{\#} = B^{\#} \circ A^{\#}.
\end{align*}
From this and the claim, we have
\begin{align*}
\int_{\re^N} |\nabla [(A \circ B)^{\#} f ](z) |^p\,dz
&=\int_{\re^N} |\nabla [ (B^{\#} \circ A^{\#} ) f ](z) |^p\,dz \\
&=\int_{\re^N} |\nabla (A^{\#} f )(w) |^p\,dw \\
&= \int_{\re^N} |\nabla f (\xi)|^p \,d\xi
\end{align*}
for $p=2$ or $N$, where $w=B(z)$ and $\xi = A(w) = (A \circ B) (z)$.
Since M\"obius transformation is a finite composition of $T_b, S_\la, R, J$, we obtain (\ref{nabla M}) for any $M \in M(\re^N)$ by induction.
\end{proof}
For more information \tcr{on} M\"obius transformation, see e.g. \cite{A,B}.
\subsection{An example of M\"obius transformation: Cayley type transformation}\label{S Cayley}
Let $N \ge 2$ and $p=2$ or $N$.
Consider the transformation ${\bf B}$
\tcr{from $\re^N$ to $\re^N$ as follows} (Ref. \cite{BFL}).
\begin{align}
\label{Cayley type}
\( \tx, \ty \) = {\bf B} (x, y)= \( \frac{2x, \,\, 1-|x|^2 -y^2}{(1+y)^2 + |x|^2 } \) \quad (x, y) \in \re^N, \,\, (\tx, \ty) \in \re^N.
\end{align}
\tcr{We see
\begin{equation}
\label{B maps to ball}
|{\bf B}(x,y)|^2 = \frac{|x|^2 + (y-1)^2}{|x|^2 + (y +1)^2},
\end{equation}
thus if we restrict ${\bf B}$ on $\re^N_{+}$, then ${\bf B}$ maps $\re^N_{+}$ to the unit ball $B_1^N \subset \re^N$.
Also $|{\bf B}(x,y)| = 1$ if and only if $y = 0$, thus ${\bf B} (\pd \re^N_{+}) = \pd B_1^N$.}
We can check that the inverse function ${\bf B}^{-1}$ is the same as ${\bf B}$, that is
$$
\( x, y \) = \tcr{{\bf B}^{-1}} (\tx, \ty)= \( \frac{2\tx, \,\, 1-|\tx|^2 -\ty^2}{(1+\ty)^2 + |\tx|^2 } \).
$$
Note that the transformation (\ref{Cayley type}) is a M\"obius transformation\tcr{:${\bf B} \in M (\re^N)$.}
In fact, we see that
\begin{align}\label{Cayley is Mobius}
{\bf B} (z) = R \circ J \circ T_{e_N} \circ S_2 \circ J \circ T_{-e_N} (z), \text{where}\,\,
R=
\begin{pmatrix}
1 & & & \\
& \ddots & & \\
& & 1 & \\
& & & -1
\end{pmatrix},\,
e_N = \begin{pmatrix}
0 \\
\vdots \\
0 \\
1
\end{pmatrix},
\end{align}
see \cite{A} p.34 or Proposition \ref{Prop Cayley is Mobius} in \S \ref{S App}.
Therefore, we obtain the following.
\begin{prop}\label{Prop Cayley}
Let ${\bf B}: \re^N \to B_1^N$ be given by (\ref{Cayley type}) and let $z=(x,y)$. Then
\begin{align*}
\det \tcr{{\bf B'}} (z) = -\left\{ \, \frac{2}{(1+y)^2 + |x|^2} \,\right\}^N.
\end{align*}
\end{prop}
\begin{proof}
From (\ref{Cayley is Mobius}) and Remark \ref{Jacobian}, we have
\begin{align*}
\det {\bf B}'(z) &= \underbrace{\det R}_{=-1} \cdot \det J'\( T_{e_N} \circ S_2 \circ J \circ T_{-e_N} (z) \) \cdot \underbrace{\det (T_{e_N})' \( S_2 \circ J \circ T_{-e_N} (z) \)}_{=1} \\
&\cdot \underbrace{\det S_2'\(J \circ T_{-e_N}(z) \)}_{=2^N} \cdot \det J'(T_{-e_N}(z)) \cdot \det \underbrace{(T_{-e_N})'(z)}_{=1} \\
&= (-1) \cdot \frac{(-1)}{|T_{e_N} \circ S_2 \circ J \circ T_{-e_N}(z)|^{2N}} \cdot 2^N \cdot \frac{(-1)}{|T_{-e_N} (z)|^{2N}} \\
&= (-1) \frac{1}{|e_N + 2(z-e_N)^*|^{2N}} \cdot \frac{1}{|z-e_N|^{2N}} \cdot 2^N \\
&\tcr{= (-1) \frac{1}{|e_N + 2\frac{(z-e_N)}{|z-e_N|^2}|^{2N}} \cdot \frac{1}{|z-e_N|^{2N}} \cdot 2^N}
= (-1) \frac{2^N}{\left\{ |x|^2 + (1 + y)^2 \right\}^N}.
\end{align*}
\end{proof}
\subsection{Harmonic transplantation}\label{S Harmonic}
Harmonic transplantation \tcr{was first} proposed by \tcr{J.} Hersch \cite{H}
\tcr{in the attempt to extend several isoperimetric problems on two-dimensional simply-connected domains to higher connectivity and higher dimensions,}
see also \cite{F book,BBF}.
Here, we recall the original harmonic transplantation from $B_1^N$ to $\Omega \subset \re^N$.
For $v \in \dot{W}_{0, {\rm rad}}^{1,p}(B_1^N)$ and $a \in \Omega$,
define $\tcr{H_a (v)} =u : \Omega \setminus \{ a\} \to \re$ by
\begin{align}\label{Harmonic trans}
u(y) = \tcr{H_a (v)}(y)
= v \( \, \( G_{B_1^N, O} \)^{-1} \( G_{\Omega, a} (y) \) \,\),
\end{align}
\tcr{where $G_{B_1^N, O}$ and $G_{\Omega, a}$ are $p$-Green's functions on the ball $B_1^N$ with the pole $O$
and on $\Omega$ with the pole $a \in \Omega$, respectively.}
In the case $p \in (1, N)$, we have
\begin{equation}
\label{GB0}
G_{B_1^N, O} (z) = \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \, |z|^{-\frac{N-p}{p-1}} - 1 \, \right]
\end{equation}
which implies that
\begin{align*}
\( G_{B_1^N, O} \)^{-1} \( G_{\Omega, a} (y) \)
= \left[ \, \frac{N-p}{p-1} \w_{N-1}^{\frac{1}{p-1}} G_{\Omega, a} (y) +1 \, \right]^{-\frac{p-1}{N-p}}.
\end{align*}
Also, we can rewrite the transformation (\ref{Harmonic trans}) to
\begin{align}\label{Omega trans}
u(y) = v(z), \,\text{where}\,\, G_{\Omega, a} (y) = G_{B_1^N, O} (z).
\end{align}
\tcr{Hereafter, we call the transformed function $u = H_a(v)$ on $\Omega \setminus \{ a \}$ via (\ref{Omega trans})
{\it the harmonic transplantation} of a function $v \in \dot{W}^{1,p}_{0, rad}(B^N_1)$.}
\tcr{We see} that harmonic transplantation (\ref{Omega trans}) preserves $\| \nabla (\cdot)\|_{L^p}$ for \tcr{$p \in (1, N]$}.
\tcr{A proof of this fact is shown for the sake of reader's convenience.}
\begin{lemma}\label{Lemma Omega}(\cite{F book} Theorem 10.3, \cite{CRN} Lemma 22)
Let $v \in \dot{W}_{0, {\rm rad}}^{1,p}(B_1^N)$ and $1<p \le N$.
Then $\tcr{H_a(v)} \in \dot{W}_0^{1,p}(\Omega)$ and $\| \nabla \( \tcr{H_a(v)} \) \|_{L^p (\Omega)} = \| \nabla v \|_{L^p (B_1^N)}$.
\end{lemma}
\begin{proof}
In the case where $p=N$, see \cite{CRN} Lemma 22.
Let $1<p<N$.
We write $G=G_{\Omega, a}$. Let $h$ be defined by
$$
h(y) = \tcr{\( G_{B^N_1, O} \)^{-1} \( G(y) \)} = \left[ \, \frac{N-p}{p-1} \w_{N-1}^{\frac{1}{p-1}} G (y) +1 \, \right]^{-\frac{p-1}{N-p}},
\quad \tcr{y \in \Omega,}
$$
and hence $u(y)= v(h(y))$.
In particular, $\nabla u(y)= v' (h(y)) \nabla h(y)$.
Note that since $G \ge 0$ in $\Omega$, we get that \tcr{$0 < h(y) \le 1$ on $\ol{\Omega}$ and if} $y \in h^{-1}(\{ t\} ) \cap \Omega$,
then $t \in [0,1]$.
Thus the coarea formula gives that
\begin{align*}
\int_{\Omega} |\nabla u |^p
&= \int_{\Omega} |v'(h(y))|^p |\nabla h (y)|^{p-1} |\nabla h (y)| \,dy\\
&=\int_0^1 \left[ \, \int_{h^{-1}(\{ t\} ) \cap \Omega} |v'(h(y))|^p |\nabla h (y)|^{p-1} \,d \mathcal{H}^{N-1}(y) \,\right] \,dt.
\end{align*}
Using $|\nabla h| = \w_{N-1}^{\frac{1}{p-1}} h(y)^{\frac{N-1}{p-1}} |\nabla G (y)|$,
\tcr{we have}
\begin{align*}
\int_{\Omega} |\nabla u |^p
=\int_0^1 \w_{N-1} t^{N-1} |v'(t)|^p\left[ \, \int_{h^{-1}(\{ t\} ) \cap \Omega} |\nabla G (y)|^{p-1} \,d \mathcal{H}^{N-1}(y) \,\right] \,dt.
\end{align*}
Note that $h^{-1}(\{ t\} ) \cap \Omega$ is \tcr{also} a level set of $G$.
Since
\begin{align*}
\int_{\{ \tcr{G} < t \}} |\nabla \tcr{G}(y) |^{p} \,dy = t, \quad
\int_{\{ \tcr{G} = t \}} |\nabla \tcr{G}(y) |^{p-1} \,d\mathcal{H}^{N-1}(y) = 1.
\end{align*}
for any $t \in [0, \infty )$
(Ref. \tcr{\cite{F book} Lemma 9.1, or} \cite{CRN} Proposition 4),
we obtain
\begin{align*}
\int_{h^{-1}(\{ t\} ) \cap \Omega} |\nabla G (y)|^{p-1} \,d \mathcal{H}^{N-1}(y)
= 1 \quad ({}^{\forall}t \in (0,1)),
\end{align*}
which implies that
\begin{align*}
\int_{\Omega} |\nabla u |^p =\int_0^1 \w_{N-1} t^{N-1} |v'(t)|^p \,dt = \int_{B_1^N} |\nabla v|^p.
\end{align*}
\end{proof}
\tcr{Up to now,} various transformations are \tcr{found in literature} so far.
These transformations can be understood as a \tcr{variant} of harmonic transplantation.
Here, we classify these transformations into the \tcr{following} three types:\\
\noindent
{\it I. \,\,\,\,Domains of two \tcr{Green's} functions in (\ref{Omega trans}) are different \tcr{from} each other. \\
II. \,\,Operators of two \tcr{Green's} functions in (\ref{Omega trans}) are different \tcr{from} each other. \\
III. Dimensions of two \tcr{Green's} functions in (\ref{Omega trans}) are different \tcr{from} each other.}\\
\noindent
Original harmonic transplantation (\ref{Omega trans}) is type I.
For reader's convenience, we unify these transformations in the form of (\ref{Omega trans}) and summarize their \tcr{properties} briefly.
In the present paper, we use harmonic transplantation (\ref{Omega trans}) in I.-(ii) \tcr{below}.
\\
\noindent
{\it I.-(i): Critical case: $1< p = N$}
\noindent
If $\Omega = B_R^N, a=O$, then \tcr{the} harmonic transplantation \tcr{$u = H_a(v)$ in} (\ref{Omega trans}) becomes
\begin{align*}
u(y)= v(z), \,\text{where}\,\, G_{B_1^N, O}(y)= \w_{N-1}^{-\frac{1}{N-1}} \log \frac{1}{|y|} = \w_{N-1}^{-\frac{1}{N-1}} \log \frac{R}{|z|} = G_{B_R^N, O}(z)
\end{align*}
which coincides with the scaling $z= S_R (y) = Ry$.
On the other hand, if $\Omega = \re_+^N$, $\tcr{a=(0,1)}$,
then \tcr{the} harmonic transplantation \tcr{$u = H_a(v)$ in} (\ref{Omega trans}) coincides with \tcr{the function ${\bf B}^{\sharp} v$ by}
the Cayley type transformation \tcr{${\bf B}$} in (\ref{Cayley type}), see \S \ref{S Cayley}.
\tcr{Similar to the various rearrangement techniques,
harmonic transplantation (\ref{Omega trans}) enables us to construct appropriate test functions for various minimization or maximization problems:
we refer the readers to the application of harmonic transplantation to the study of} the Trudinger-Moser maximization problem on general bounded domain $\Omega$ (Ref. \cite{F, CR, L, CRN}). \\
\noindent
{\it I.-(ii): Subcritical case: $1<p<N$}
\noindent
If $\Omega = B_R^N$ (let $B_\infty^N = \re^N$ and $\frac{1}{\infty} = 0$), $a=O$,
then \tcr{the} harmonic transplantation \tcr{$u = H_a(v)$ in} (\ref{Omega trans}) becomes
\begin{align*}
u(y)= v(z), \,\text{where}\,\, G_{B_1^N, O}(y)
&= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \, |y|^{-\frac{N-p}{p-1}} -1 \, \right] \\
&= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \, |z|^{-\frac{N-p}{p-1}} -R^{-\frac{N-p}{p-1}} \, \right]
= G_{B_R^N, O}(z)
\end{align*}
which does not coincide with \tcr{$S_R^{\sharp}(v)$, here $S_R$ is the dilation} $z= S_R (y) = Ry \,(R<\infty)$, unlike I.-(i).
We can obtain improved Hardy-Sobolev inequalities on $B_R^N$ via (\ref{Omega trans}),
which are equivalent to the Hardy-Sobolev inequalities on $\re^N$ (Ref. \cite{I}. See also \cite{S(NA),S(ArXiv)}).
Not only \tcr{the improvement of the inequalities},
but also a limit of the improved Hardy-Sobolev inequalities on $B_R^N$ as $p \nearrow N$ can be \tcr{considered},
unlike the classical \tcr{cases}.
For the subcritical Rellich inequality, a part of this argument still holds, see \S 2 in \cite{S(Rellich)}.
A limit of \tcr{the} Hardy-Sobolev and the Poincar\'e \tcr{inequalities} (in some sense) can be considered,
\tcr{see \cite{I} and \cite{BP} for taking a limit $p \nearrow N$ or $N \nearrow \infty$ in the Sobolev inequality respectively,
\cite{SS} for $p \nearrow N$ in the Hardy inequality, and $|\Omega| \searrow 0$ in the Poincar\'e inequality.
Also see \cite{S(RIMS)} for a survey.}
In the present paper, we consider \tcr{the} harmonic transplantation \tcr{$u = H_a(v)$, $v \in \dot{W}^{1,p}_{0, rad}(B^N_1)$, in} (\ref{Omega trans})
\tcr{for} $\Omega = \re_+^N$, \tcr{$a = e_N = (0,1)$, and $p \in (1,N)$.}
Namely,
\begin{align}\label{half-space trans}
u(x,y)&= v(\tx, \ty), \,\text{where} \notag \\
G_{\re^N_+, (0,1)} (x, y)
&= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \, \( |x|^2 +(1-y)^2 \)^{-\frac{N-p}{2(p-1)}} - \psi_p (x,y) \,\right] \notag \\
&= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \, \( |\tx|^2 + |\ty|^2 \)^{-\frac{N-p}{2(p-1)}} -1 \, \right]= G_{B_1^N, O}(\tx, \ty)
\end{align}
\tcr{where $\psi_p$ is as in \eqref{G half}.}
\begin{remark}\label{Rem two scaling}
\tcr{
We point out that, in the case $2=p<N$,
there are at least two transformations \tcr{$u$ of $v \in \dot{W}^{1,2}_{0,rad}(B^N_1)$},
by which $\| \nabla u \|_{L^2(\Omega)} = \| \nabla v \|_{L^2(B_1^N)}$ holds.
Indeed, when $\Omega = B_R^N$ for $R < \infty$, the harmonic transplantation $u = H_0(v)$ and the M\"obius transformation $u = S_R^{\#}(v)$
from $\dot{W}^{1,2}_{0,rad}(B_1^N)$ to $\dot{W}^{1,2}_0(B_R^N)$ preserve the $L^2$ norm of the gradient.
Also when $\Omega = \re^N_{+}$, the harmonic transplantation $u = H_{e_N}(v)$ and the M\"obius transformation $u = {\bf B}^{\#}(v)$
via the Cayley type transformation (\ref{Cayley type}) from $\dot{W}^{1,2}_{0,rad}(B_1^N)$ to $\dot{W}^{1,2}_0(\re^N_+)$ have the same property,
see \S \ref{S Cayley}.
}
\end{remark}
\noindent
{\it II.}: {\it From weighted problem to \tcr{unweighted} problem}
\noindent
Let $1<p<N$ and $\tilde{G}_{B_1^N, O}$ be Green's function with singularity at $O$ associated with the weighted $p$-Laplace operator
${\rm div }(|y|^{p-N} |\nabla (\cdot)|^{p-2} \nabla (\cdot))$.
Define \tcr{$u: B^N_1 \to \re$ by}
\begin{align*}
u(y)= v(z), \,&\text{where}\,\, \tilde{G}_{B_1^N, O}(y)
= \w_{N-1}^{-\frac{1}{p-1}} \log \frac{1}{|y|}
= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} |z|^{-\frac{N-p}{p-1}}
= G_{\re^N, O}(z), \\
&\tcr{y \in B_1^N, \, z \in \re^N.}
\end{align*}
Then we have $\| \nabla u \|_{L^p(B_1^N ;\, |y|^{p-N} \,dy)} = \| \nabla v \|_{L^p (\re^N)}$ (Ref. \cite{Z, HK}).
We can remove the weight $|y|^{p-N}$ thanks to the above transformation. \\
\noindent
{\it III.-(i): From higher dimensions to one dimension}
\noindent
Consider \tcr{the} Moser transformation
\begin{align*}
u(y)= v(z), \,\text{where}\,\, G_{\tcr{B_R^N,} O}(y)
= \w_{N-1}^{-\frac{1}{N-1}} \log \frac{R}{|y|}
= z.\,
\quad \tcr{y \in B_R^N, z \in \re_{+}}.
\end{align*}
Then we have $\| \nabla u \|_{L^N(B_R^N)} = \| v' \|_{L^N (\re_+)}$.
\tcr{The} Moser transformation is used to reduce the Trudinger-Moser maximization problem on $\dot{W}_{0, \text{rad}}^{1,N}(B_R^N)$
to the one-dimensional problem (Ref. \cite{M}).
On the other hand, if we consider the Moser transformation on the subcritical Sobolev spaces $\dot{W}_{0, {\rm rad}}^{1,p}(B_R^N)\,(p< N)$,
then we have
\begin{align*}
u(y)= v(z), \,\text{where}\,\, G_{\tcr{B_R^N}, O}(y)
= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} \left[ \, |y|^{-\frac{N-p}{p-1}} -R^{-\frac{N-p}{p-1}} \,\right]
= z.\,
\end{align*}
Then \tcr{again} we have $\| \nabla u \|_{L^p(B_R^N)} = \| v' \|_{L^p (\re_+)}$.
For an application of these transformations, see Proposition \ref{Prop Bliss} in \S \ref{S App}.
\\
\noindent
{\it III.-(ii): Relation between the critical and the subcritical Sobolev spaces}
\noindent
Let $p=N < m$. If we consider \tcr{the relation}
\begin{align}\label{trans dim}
u(y)= v(z), \,\text{where}\,\, G_{B_R^N, O}(y) &= \w_{N-1}^{-\frac{1}{N-1}} \log \frac{R}{|y|} \notag \\
&= \frac{N-1}{m-N} \w_{m-1}^{-\frac{1}{N-1}} \left[ \, |z|^{-\frac{m-N}{N-1}} -R^{-\frac{m-N}{N-1}} \,\right] = G_{B_R^m, O}(z),
\end{align}
\tcr{for $u \in \dot{W}^{1,N}_{0, rad}(B^N_R)$, $v \in \dot{W}^{1,N}_{0, rad}(B^m_R)$.}
then we have $\| \nabla u \|_{L^N (B_R^N)} = \| \nabla v \|_{L^N (B_R^m)}$.
Namely, we obtain the equality between two norms of the critical Sobolev spaces $\dot{W}_{0, \text{rad}}^{1,N} (B_R^N)$
and the higher dimensional subcritical Sobolev spaces $\dot{W}_{0, \text{rad}}^{1,p} (B_R^m)$ (Ref. \cite{ST}).
This transformation (\ref{trans dim}) gives a direct relation between the subcritical Sobolev embeddings
\begin{align*}
\dot{W}_{0, \text{rad}}^{1,p} \hookrightarrow L^{p^*, p} \hookrightarrow L^{p^*, q} \hookrightarrow L^{p^*, \infty}
\end{align*}
\tcr{where $p < q$,}
and the critical Sobolev embeddings
\begin{align*}
\dot{W}_{0, \text{rad}}^{1,N} \hookrightarrow L^{\infty, N}(\log L)^{-1} \hookrightarrow L^{\infty, q}(\log L)^{-1+ \frac{1}{N} - \frac{1}{q}} \hookrightarrow L^{\infty, \infty}(\log L)^{-1+\frac{1}{N}} = {\rm Exp L}^{\frac{N}{N-1}}.
\end{align*}
For the subcritical and the critical Sobolev embeddings, see e.g. \cite{SS} \S 1.
\\
\noindent
{\it III.-(iii): An infinite dimensional form of the Sobolev inequality}
\noindent
Let $p< N < m$. If we consider
\begin{align*}
u(y)= v(z), \,&\text{where}\,\, G_{\re^N, O}(y)
= \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} |y|^{-\frac{N-p}{p-1}}
= \frac{p-1}{m-p} \w_{m-1}^{-\frac{1}{p-1}} |z|^{-\frac{m-p}{p-1}}
= G_{\re^m, O}(z), \\
&\tcr{y \in \re^N, \, z \in \re^m}
\end{align*}
then we have $\| \nabla u \|_{L^p (\re^N)} = \| \nabla v \|_{L^p (\re^m)}$.
Namely, we can reduce the $m$-dimensional Sobolev inequality:
$S_{m,p} \| v \|^p_{L^{p^*}(\re^m)} \le \| \nabla v\|_{L^p(\re^m)}^p$
to \tcr{an} $N$-dimensional inequality \tcr{for $u$, which involves $m$ as a parameter and $m$ can be arbitrarily large}.
Therefore, we can take a limit of the $m$-dimensional Sobolev inequality as $m \nearrow \infty$ \tcr{in this sense.}
As a consequence, we can obtain the $N$-dimensional Hardy inequality:
$\( \frac{N-p}{p} \)^p \int_{\re^N} \frac{|u|^p}{|y|^p} \, dy \le \int_{\re^N} |\nabla u|^p \,dy$
as an infinite dimensional form of the Sobolev inequality (Ref. \cite{S(NA)}).
\section{Proof of Theorems}\label{S Proof}
First, we show Theorem \ref{Thm CH}.
\begin{proof}(Proof of Theorem \ref{Thm CH})
Let $p=N$.
\tcr{We will ``transplant" the critical Hardy inequality on $B_1^N$ to $\re^N_{+}$ by the Cayley type transformation
${\bf B}$ \eqref{Cayley type} in \S \ref{S Cayley}.}
\tcr{By \eqref{B maps to ball}, we consider ${\bf B}$ maps $\re^N_{+}$ to $B^N_1$.}
\tcr{Let $u \in C_c^1(\re^N_{+})$ and put $v(\tx, \ty) = ({\bf B}^{-1})^{\#}(u)(\tx, \ty)$ for $(\tx, \ty) \in B^N_1$.
Then we see $u(x,y) = {\bf B}^{\#} (v)(x,y) = v({\bf B}(x,y))$ for $(x,y) \in \re^N_{+}$.}
\tcr{From the fact that ${\bf B} \in M(\re^N)$} and Proposition \ref{Prop Mobius},
we have
\begin{align*}
\int_{\re^N_+} |\nabla u\tcr{(x,y)}|^N \,dx d y = \int_{B_1^N} |\nabla v\tcr{(\tx, \ty)}|^N \,d\,\tx \, d\, \ty
\end{align*}
\tcr{where $(\tx, \ty) = {\bf B}(x, y)$.}
Since $|\tx|^2 + \ty^2 = \frac{|x|^2 + (y-1)^2}{|x|^2 + (y +1)^2}$ \tcr{by \eqref{B maps to ball}},
we have
\begin{align*}
&\int_{B_1^N} \frac{|\tcr{v(\tx, \ty)}|^N}{\{ | \tx|^2 + |\ty|^2 \}^{\frac{N}{2}}
\( \log \frac{1}{\sqrt{ |\tx|^2 + |\ty|^2}} \)^N }\,d \,\tx \,d \,\ty \\
&=\tcr{\int_{\re^N_+} \frac{|u(x,y)|^N}{|{\bf B}(x,y)|^{\frac{N}{2}} \( \log \frac{1}{|{\bf B}(x,y)|} \)^N} | \det {\bf B'}(x,y) |
\,\,dx \,dy} \\
&=\int_{\re^N_+} \frac{|\tcr{u(x,y)}|^N }{\{ |x|^2 + (1-y)^2 \}^{\frac{N}{2}} \left\{ \frac{|x|^2 + (1+y)^2}{4} \right\}^{\frac{N}{2}} \( \log \sqrt{ \frac{|x|^2+ (1+y)^2}{|x|^2 + (1-y)^2}} \)^N } \,dx dy\tcr{.}
\end{align*}
\tcr{Thus, we obtain \eqref{H_N half} by the critical Hardy inequality on the unit ball for $v$} (Ref. \cite{AS,II, TF}):
\begin{align*}
\( \frac{N-1}{N} \)^N \int_{B_1^N} \frac{|\tcr{v(\tx, \ty)}|^N}{\{ |\tx|^2 + |\ty|^2 \}^{\frac{N}{2}}
\( \log \frac{1}{\sqrt{ |\tx|^2 + |\ty|^2}} \)^N }\,d \,\tx \,d \,\ty \le \int_{B_1^N} |\nabla \tcr{v(\tx, \ty)}|^N \,d\,\tx \, d\, \ty.
\end{align*}
\tcr{Optimality} and the non-attainability of the constant $\( \frac{N-1}{N} \)^N$ \tcr{in \eqref{H_N half}} follows
from results for the critical Hardy inequality on the unit ball also.
\end{proof}
In the same way as above, we also obtain \tcr{a} Trudinger-Moser type inequality on the half-space from the result on balls
\begin{theorem}\label{Thm TM}
Let $N \ge 2$. Then
\begin{align*}
\sup \left\{ \int_{\re^N_+} e^{\alpha \,|u(x,y)|^{\frac{N}{N-1}}} \,\frac{2^N \,dxdy}{\left\{ |x|^2 + (y+1)^2 \right\}^N} \,\,\middle|
\,\, \|\nabla u\|_{\tcr{L^N}(\re^N_+)} \le 1, \, u \in \dot{W}_{0}^{1,N}(\re^N_+) \,\, \right\}
\end{align*}
is finite if and only if $\alpha \le N \w_{N-1}^{\frac{1}{N-1}}$.
Moreover, the above maximization problem is attained for any $\alpha \le N \w_{N-1}^{\frac{1}{N-1}}$.
\end{theorem}
\tcr{
\begin{proof}
Again, we use the transformation $v = ({\bf B}^{-1})^{\#}(u)$, $u \in \dot{W}^{1,N}_0(\re^N_{+})$.
Since $|{\rm det} \, {\bf B}'(x,y)| = \frac{2^N}{\left\{|x|^2 + (y+1)^2 \right\}^N}$,
the theorem follows from the Trudinger-Moser inequality on $B^N_1$ and its attainability: see \cite{CC}.
\end{proof}
}
Next, we show Theorem \ref{Thm IH}.
\tcr{Before that, we claim the following Theorem:}
\begin{theorem}\label{Thm psi}
Let $1<p<N$. Then the Hardy type inequality
\begin{align*}
\( \frac{N-p}{p} \)^p \int_{\re^N_+} \frac{W_p(x,y)^{\frac{p}{2}}}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} |u(x,y)|^{p} \,dxdy
\le \int_{\re^N_+} |\nabla u(x,y)|^p \,dxdy,
\end{align*}
holds for any $u \in C_{c}^{1}(\re^N_+)$ \tcr{of the form $u(x,y) = u\(G_{\re^N_+, \tcr{e_N}}(x,y)\)$},
where $\psi_p$ is given in \S \ref{S Green} and for $(x,y) \in \re^N_+$,
\begin{align*}
W_p(x,y) &= \frac{1+ \( |x|^2 +(1-y)^2 \)^{\frac{N-1}{p-1}} |\nabla \psi_p|^2 -2 \( |x|^2 +(1-y)^2 \)^{\frac{N-p}{2(p-1)}+1} \nabla \psi_p \cdot
\begin{pmatrix}
x \\
y-1
\end{pmatrix}
}
{\left[ 1-\tcr{\tilde{X}}^{\frac{N-p}{2(p-1)}} \right]^2}, \\
\tcr{\tilde{X}} &= \psi_p (x,y)^{\frac{2(p-1)}{N-p}} \left\{ |x|^2 +(1-y)^2\right\}.
\end{align*}
\end{theorem}
\tcr{The proof of Theorem \ref{Thm psi} consists of the use of the harmonic transplantation
between $\re^N_{+}$ and $\re^N$, the Hardy inequality (\ref{H_p}), and Lemma \ref{Lemma Omega}.
Since the proof of Theorem \ref{Thm psi} is almost the same as that of Theorem \ref{Thm IH} below,}
we omit the proof.
As we mentioned in \S \ref{S Green},
we do not know the explicit form of $\psi_p$ when $2 \not= p < N$.
Due to the lack of the explicit form of $\psi_p$,
we cannot check that the inequality in Theorem \ref{Thm psi} is improved, i.e., $W_p(x,y) \ge 1$.
Therefore, we consider a modification of harmonic transplantation by using $U_p$ \tcr{in \eqref{def U}}
instead of $G_{\re^N_+, \tcr{e_N}}$ in (\ref{half-space trans}).
\tcr{This is the main idea we have invented in the proof of Theorem \ref{Thm IH}.}
Consider the following modified transformation for radial functions
$v = v(z) = v(t)$, $(t= |z| \tcr{\in [0,1]})$ on $B_1^N$, or $w = w(\tilde{z})= w(r)$, $(r= |\tilde{z}| \tcr{\in [0, +\infty)})$ on $\re^N$:
\begin{align}\label{U trans}
\tcr{
\begin{cases}
&u(x,y) = v(t), \,\text{where}\,\,U_p (x,y) = G_{B_1^N, O}(t), \quad (x, y) \in \re^N_{+}, \\
&u(x,y) = w(r), \,\text{where}\,\,U_p (x,y) = G_{\re^N, O}(r), \quad (x, y) \in \re^N_{+}.
\end{cases}
}
\end{align}
\tcr{We call the function $u$ on $\re^N_{+}$ in \eqref{U trans} {\it the generalized harmonic transplantation} of $v$ (or $w$).}
We obtain the following Lemma instead of Lemma \ref{Lemma Omega}.
\begin{lemma}\label{Lemma Omega gene}
Let $1< p < N, v, w$ be radial functions on $B_1^N, \re^N$, and $u$ be given by (\ref{U trans}). Then we have
\begin{align*}
\int_{\re^N_+} |\nabla u |^p
&= \int_{B_1^N} |\nabla v|^p + \int_0^1 \w_{N-1} t^{N-1} |v'(t)|^p F_p \( G_{B_1^N, O} (t) \) \,dt\\
&=\int_{\re^N} |\nabla w|^p + \int_0^\infty \w_{N-1} r^{N-1} |w'(r)|^p F_p \( G_{\re^N, O}(r) \) \,dr
\end{align*}
where
\begin{equation}
\label{F_p}
F_p(s)= \int_{[U_p > s] } (-\lap_p U_p) \, dxdy.
\end{equation}
Especially,
\tcr{
if $p \ge 2$ (resp. $p \le 2$), then $F_p (s) \ge 0$ (resp. $F_p(s) \le 0$)
and $\| \nabla u \|_p \ge \| \nabla v \|_p$, $\| \nabla u \|_p \ge \| \nabla w \|_p$
(resp. $\| \nabla u \|_p \le \| \nabla v \|_p$, $\| \nabla u \|_p \le \| \nabla w \|_p$)
holds.}
\end{lemma}
\begin{proof}
\tcr{We prove the first equality only, since the proof of the second equality is similar.}
\tcr{Also we note that the proof below is an analogue to that of Lemma \ref{Lemma Omega}.}
Let $h$ be defined by
$$
\tcr{h(x,y) = \(G_{B^N_1, O}\)^{-1} \( U_p(x,y) \)} = \left[ \, \frac{N-p}{p-1} \w_{N-1}^{\frac{1}{p-1}} U_p (x,y) +1 \, \right]^{-\frac{p-1}{N-p}},
\quad \tcr{(x, y) \in \re^N_{+}}.
$$
\tcr{Thus $t = h(x,y)$ is equivalent to $U_p(x,y) = G_{B^N_1, O}(t)$}
and $u(x,y)= v(h(x,y))$.
In particular, $\nabla u(x, y)= v'(h(x,y)) \nabla h(x,y)$.
\tcr{Note that since $U_p \ge 0$ in $\re^N_+$,
we get that $0 < h(x,y) \le 1$ for $(x, y) \in \ol{\re^N_{+}}$ and} if $(x,y) \in h^{-1}(\{ t\} ) \cap \re^N_{+}$,
Thus, the coarea formula gives that
\begin{align*}
\int_{\re^N_+} |\nabla u |^p
&= \int_{\re^N_+} |v'(h(x,y))|^p |\nabla h (x, y)|^{p-1} |\nabla h (x,y)| \,dxdy\\
&=\int_0^1 \left[ \, \int_{h^{-1}(\{ t\} ) \cap \re^N_+} |v'(h(x,y))|^p |\nabla h (x,y)|^{p-1} \,d \mathcal{H}^{N-1}(x,y) \,\right] \,dt.
\end{align*}
\tcr{Inserting} $|\nabla h| = \w_{N-1}^{\frac{1}{p-1}} h(y)^{\frac{N-1}{p-1}} |\nabla U_p (x,y)|$,
\tcr{we have}
\begin{align}
\label{LL1}
\int_{\re^N_+} |\nabla u |^p
=\int_0^1 \w_{N-1} t^{N-1} |v'(t)|^p
\left[ \, \int_{h^{-1}(\{ t\} ) \cap \re^N_+} |\nabla U_p (x, y)|^{p-1} \,d \mathcal{H}^{N-1}(x,y) \,\right] \,dt.
\end{align}
Note that $h^{-1}(\{ t\} ) \cap \re^N_+$ is a level set of $U_p$.
Applying $\phi_t = {\rm min} \{ t, U_p \}$, $(t >0)$ as a test function \tcr{of \eqref{eq:sol U}} in Proposition \ref{Prop sol U},
\tcr{we have}
\begin{align*}
\int_{[U_p < t]} | \nabla U_p |^{p} \, dxdy
= t + \int_{[\tcr{U_p} < t] } (-\lap_p U_p)\, U_p\, dxdy + \int_{[U_p > t] } (-\lap_p U_p)\,t \, dxdy.
\end{align*}
If we differentiate the above with respect to $t$, then we have
\begin{align*}
&\int_{[U_p = t]} | \nabla U_p |^{p-1} \, d \mathcal{H}^{N-1}(x,y) \\
&= 1 + \int_{[U_p = t] } (-\lap_p U_p)\, \frac{U_p}{|\nabla U_p|}\, d \mathcal{H}^{N-1}(x,y)
- \int_{[U_p = t] } (-\lap_p U_p)\,\frac{t}{|\nabla U_p|} \, d \mathcal{H}^{N-1}(x,y) \\
&\hspace{2em} + \int_{[U_p > t] } (-\lap_p U_p) \, dxdy \\
&= 1 + \int_{[U_p > t] } (-\lap_p U_p) \, dxdy
\end{align*}
thanks to the coarea formula.
Therefore, \tcr{replacing $t$ by $G_{B^N_1, O}(t)$} for any $t \in (0,1)$,
we have
\begin{align}
\label{LL2}
\int_{h^{-1}(\{ t\} ) \cap \re^N_+} |\nabla U_p (x,y)|^{p-1} \,d \mathcal{H}^{N-1}(x,y)
= 1 + \int_{ \left[ U_p > G_{B^N_1, O}(t) \right] } (-\lap_p U_p) \, dxdy.
\end{align}
\tcr{Inserting \eqref{LL2} into \eqref{LL1}, we obtain}
\begin{align*}
&\int_{\re^N_+} |\nabla u |^p
= \int_0^1 \w_{N-1} t^{N-1} |v'(t)|^p \( 1+ \int_{[U_p > t] } (-\lap_p U_p) \, dxdy\) \,dt \\
&= \int_{B_1^N} |\nabla v|^p + \int_0^1 \w_{N-1} t^{N-1} |v'(t)|^p
F_p \( \w_{N-1}^{-\frac{1}{p-1}} \frac{p-1}{N-p} \( t^{-\frac{N-p}{p-1}} -1 \) \) \,dt.
\end{align*}
\end{proof}
\begin{lemma}\label{Lemma IH gene}
Let $1< p < N$, $0 \le s \le p$, \tcr{$v$ and $w$} be radial functions on $B_1^N$ and $\re^N$ \tcr{respectively},
and \tcr{let} $u$ be given by (\ref{U trans}).
Then we have
\begin{align*}
&\int_{\re^N_+} \frac{V_p(x,y)^{\frac{p}{2}} }{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} |u(x,y)|^{p} \,dxdy \\
&=\int_{B_1^N} \frac{|v|^{p}}{|z|^p \left[ 1-|z|^{\frac{N-p}{p-1}} \right]^p} \,dz
+ \w_{N-1} \int_0^1 |v|^{p} t^{N-1-p} \left[ 1-t^{\frac{N-p}{p-1}} \right]^{-p} F_p \( G_{B_1^N, O}(t) \) \,dr \\
&= \int_{\re^N} \frac{|w|^{p}}{|\tilde{z}|^p}\,d\tilde{z}
+ \w_{N-1} \int_0^\infty |w|^{p} r^{N-1-p} F_p \( G_{\re^N, O}(r) \) \,dr.
\end{align*}
where \tcr{$F_p(s)$ is defined in \eqref{F_p},}
and for $(x,y) \in \re^N_+$, \tcbb{$V_p$ and $X$ are defined in \eqref{V_p}.}
\end{lemma}
\begin{proof
\tcr{We prove the second equality for $v$ only. The first equality is similar.}
\tcr{For $(x, y) \in \re^N_{+}$, define $h(x,y)$ by the relation}
\begin{align*}
\tcr{U_p (x, y) = G_{\re^N, O}(h(x,y))} = \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} h(x,y)^{-\frac{N-p}{p-1}},
\end{align*}
\tcr{that is,}
\begin{align*}
h(x,y) = \left[ \,\( |x|^2 +(1-y)^2 \)^{-\frac{N-p}{2(p-1)}} - \( |x|^2 +(1+y)^2 \)^{-\frac{N-p}{2(p-1)}} \,\right]^{\frac{p-1}{p-N}}.
\end{align*}
\tcr{As in \eqref{LL2}, we can obtain}
\begin{align}
\label{LL3}
\tcr{\int_{h^{-1}(\{ t\} ) \cap \re^N_+} |\nabla U_p (x,y)|^{p-1} \,d \mathcal{H}^{N-1}(x,y)
= 1 + \int_{ \left[ U_p > G_{\re^N, O}(t) \right] } (-\lap_p U_p) \, dxdy.}
\end{align}
\tcr{Thus by the coarea formula and (\ref{LL3}),} we have
\begin{align}
\label{LL4}
&\int_{\re^N_+} |u|^{p} \frac{|\nabla h(x,y)|^p}{h(x,y)^p} \,dx dy
= \int_0^\infty \int_{h^{-1}(\{ r\} ) \cap \re^N_+} |w|^{p} r^{-p} |\nabla h (x,y)|^{p-1} \,d \mathcal{H}^{N-1}(x,y) \,dr \\
&= \,\w_{N-1} \int_0^\infty |w|^{p} r^{N-1-p} \int_{h^{-1}(\{ r \} ) \cap \re^N_+}
|\nabla U_p (\tcr{x,y})|^{p-1} \,d \mathcal{H}^{N-1}(x,y) \,dr \notag \\
&\overset{\eqref{LL3}}{=} \w_{N-1} \int_0^\infty |w|^{p} r^{N-1-p} \,dr \tcr{+} \w_{N-1} \int_0^\infty |w|^{p} r^{N-1-p}
F_p \( \tcr{G_{\re^N, O} (r)} \) \,dr \notag \\
&=\int_{\re^N} \frac{|w|^{p}}{|\tilde{z}|^p}\,d\tilde{z}
+ \w_{N-1} \int_0^\infty |w|^{p} r^{N-1-p} F_p \( G_{\re^N, O} (r) \) \,dr, \notag
\end{align}
where $|\tilde{z}| = r=h(x,y)$.
On the other hand, we have
\begin{align*}
\nabla h(x,y) = h(x,y)^{\frac{N-1}{p-1}} \left[ \, \( |x|^2 +(1-y)^2 \)^{\frac{p-N}{2(p-1)}-1}
\begin{pmatrix}
x \quad \\
y-1
\end{pmatrix}
- \( |x|^2 +(1+y)^2 \)^{\frac{p-N}{2(p-1)}-1}
\begin{pmatrix}
x \quad \\
y+1
\end{pmatrix}
\,\right]
\end{align*}
which implies that
\begin{align*}
&\frac{|\nabla h(x,y) |^p}{h(x,y)^p}
= h(x,y)^{\frac{N-p}{p-1}p}
\left| \, \( |x|^2 +(1-y)^2 \)^{\frac{p-N}{2(p-1)}-1}
\begin{pmatrix}
x \quad \\
y-1
\end{pmatrix}
- \( |x|^2 +(1+y)^2 \)^{\frac{p-N}{2(p-1)}-1}
\begin{pmatrix}
x \quad \\
y+1
\end{pmatrix}
\,\right|^{p}\\
&=\frac{
\left[ \, \( |x|^2 +(1-y)^2 \)^{\frac{1-N}{p-1}}
+ \( |x|^2 +(1+y)^2 \)^{\frac{1-N}{p-1}} -2 \( |x|^2 +(1-y)^2 \)^{\frac{2-p-N}{2(p-1)}}
\( |x|^2 +(1+y)^2 \)^{\frac{2-p-N}{2(p-1)}} (|x|^2 +y^2 -1) \,\right]^{\frac{p}{2}}
}{
\left[ \, \( |x|^2 +(1-y)^2 \)^{-\frac{N-p}{2(p-1)}} - \( |x|^2 +(1+y)^2 \)^{-\frac{N-p}{2(p-1)}} \,\right]^{p} }\\
&= \frac{\left[ 1+ X^{\frac{N-1}{p-1}} -2X^{\frac{N-p}{2(p-1)}} \( |x|^2 +(1+y)^2 \)^{\tcrr{-1}}
(|x|^2 +y^2 -1) \right]^{\frac{p}{2}}}{ \left[ |x|^2 +(1-y)^2 \right]^{\frac{p}{2}} \left[ 1-X^{\frac{N-p}{2(p-1)}} \right]^p} \\
&= \frac{V_p(x,y)^{\frac{p}{2}}}{ \left[ |x|^2 +(1-y)^2 \right]^{\frac{p}{2}} }.
\end{align*}
\tcr{Inserting this in the left hand-side of \eqref{LL4}, we obtain the second equality of} Lemma \ref{Lemma IH gene}.
\end{proof}
\tcr{Now, we prove Theorem \ref{Thm IH}.}
\tcr{First, we claim the next lemma.}
\begin{lemma}
\label{Lemma:claim 2}
Let $2 \le p \le N$ and let $F_p$ be defined in \eqref{F_p}.
Then
\begin{align}\label{claim 2}
\( \frac{N-p}{p} \)^p \int_0^\infty |w|^{p} r^{N-1-p} F_p \( G_{\re^N, O} (r) \) \,dr
\le \int_0^\infty |w'|^{p} r^{N-1} F_p \( G_{\re^N, O} (r) \) \,dr
\end{align}
holds for any radial function $w = w(r) \in C_c^1(\re^N)$.
\end{lemma}
\begin{proof}
Since for $p \ge 2$,
\begin{align*}
F'_p (s) = - \int_{[U_p=s]} \frac{-\lap_p U_p}{ |\nabla U_p|} d\mathcal{H}^{N-1} (x,y) \le 0
\end{align*}
\tcr{by Proposition \ref{Prop cal U}} and $G^{'}_{\re^N, O} (r) \le 0$,
we have
\begin{align*}
&\int_0^\infty |w|^{p} r^{N-1-p} F_p \( G_{\re^N, O} (r) \) \,dr \\
&= -\frac{p}{N-p} \int_0^\infty |w|^{p-2} w w' r^{N-p} F_p \( G_{\re^N, O} (r) \) \,dr
- \frac{1}{N-p} \int_0^\infty |w|^{p} r^{N-p} F^{'}_p \( G_{\re^N, O} (r) \) G^{'}_{\re^N, O} (r) \,dr \\
&\le -\frac{p}{N-p} \int_0^\infty |w|^{p-2} w w' r^{N-p} F_p \( G_{\re^N, O} (r) \) \,dr \\
&\le \frac{p}{N-p} \( \int_0^\infty |w'|^{p} r^{N-1} F_p \( G_{\re^N, O} (r) \) \,dr \)^{\frac{1}{p}}
\( \int_0^\infty |w|^p r^{N-1-p} F_p \( G_{\re^N, O} (r) \) \,dr \)^{\frac{p-1}{p}}.
\end{align*}
Therefore, we obtain (\ref{claim 2}).
\end{proof}
\begin{proof}(Proof of Theorem \ref{Thm IH})
From Lemma \ref{Lemma IH gene}, \eqref{claim 2}, the classical Hardy inequality (\ref{H_p}) on $\re^N$, and Lemma \ref{Lemma Omega gene},
we have
\begin{align*}
&\( \frac{N-p}{p} \)^p \int_{\re^N_+} \frac{|u(x,y)|^{p}}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} V_p(x,y)^{\frac{p}{2}} \,dxdy \\
&\overset{{\rm Lemma} \, \ref{Lemma IH gene}}{=} \( \frac{N-p}{p} \)^p\int_{\re^N} \frac{|w|^{p}}{|\tilde{z}|^p}\,d\tilde{z}
+ \( \frac{N-p}{p} \)^p\w_{N-1} \int_0^\infty |w|^{p} r^{N-1-p} F_p \( G_{\re^N, O}(r) \) \,dr \\
&\overset{\eqref{claim 2}}{\le} \( \frac{N-p}{p} \)^p\int_{\re^N} \frac{|w|^{p}}{|\tilde{z}|^p}\,d\tilde{z}
+ \int_0^\infty |w'|^{p} r^{N-1} F_p \( G_{\re^N, O} (r) \) \,dr \\
&\overset{\eqref{H_p}}{<} \int_{\re^N} |\nabla w|^p\,d\tilde{z}
+ \int_0^\infty |w'|^{p} r^{N-1} F_p \( G_{\re^N, O} (r) \) \,dr \\
&\overset{{\rm Lemma} \, \ref{Lemma Omega gene}}{=} \int_{\re^N_+} |\nabla u (x,y) |^p \, dxdy.
\end{align*}
\tcr{Thus the inequality \eqref{IH}
\begin{align*}
&\( \frac{N-p}{p} \)^p\int_{\re^N_+} \frac{|u(x,y)|^p}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} V_p(x,y)^{\frac{p}{2}} \,dxdy
< \int_{\re^N_+} |\nabla u (x,y)|^p \, dxdy.
\end{align*}
is proven.}
The remaining is to show the optimality of the constant $\( \frac{N-p}{p} \)^p$ in the above inequality.
For large $M >0$ and small $\ep >0$, consider the following test function:
\begin{align*}
u_{\ep, M} (x,y)= U_p (x,y)^{\frac{p-1}{p} - \frac{p-1}{N-p} \ep} \psi_M \( U_p (x,y) \)
\end{align*}
where $U_p (x,y)$ \tcr{is in \eqref{def U}.}
\tcr{Put}
$$
C(N,p) = \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}}
$$
and define $\psi_M \in C^\infty (0,\infty)$, $0\le \psi_M \le 1$, $\psi_M (s) =1$ for $s \ge M$, $\psi_M (s) =0$ for $s \le \frac{M}{2}$.
Let $\delta_1 = \delta_1 (M)$, $\delta_2 =\delta_2(M) >0$ satisfy
\begin{align*}
\delta_1^{-\frac{N-p}{p-1}} - (2-\delta_1)^{-\frac{N-p}{p-1}} = C(N,p)^{-1} M, \quad
\delta_2 = \( \frac{M}{C(N,p)} \)^{\frac{p-1}{N-p}}.
\end{align*}
Then we have $B_{\delta_1} = B_{\delta_1}(\tcr{0,1}) \subset [U_p \ge M] \subset B_{\delta_2}$.
Then we have
\begin{align}\label{deno}
&\int_{\re^N_+} \frac{|u_{\ep, M}(x,y)|^{p} V_p(x,y)^{\frac{p}{2}}}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} \,dxdy
\ge \int_{[U_p \ge M]} \frac{|U_p|^{p-1-\frac{p-1}{N-p} p\ep} }{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} \,dxdy \notag \\
&\ge C(N,p)^{p-1-\frac{p-1}{N-p} p\ep} \int_{B_{\delta_1}} \( |x|^2 +(1-y)^2 \)^{-\frac{N}{2} + \frac{p\ep}{2} }
\left[ 1-\( \frac{|x|^2 +(y-1)^2}{|x|^2 + (y+1)^2} \)^{\frac{N-p}{2(p-1)}} \right]^{p-1} \, dxdy \notag \\
&\ge C(N,p)^{p-1-\frac{p-1}{N-p} p\ep} \int_{B_{\delta_1}} \( |x|^2 +(1-y)^2 \)^{-\frac{N}{2} + \frac{p\ep}{2} }
\left[ 1- (p-1) \( \frac{|x|^2 +(y-1)^2}{|x|^2 + (y+1)^2} \)^{\frac{N-p}{2(p-1)}} \right] \, dxdy \notag \\
&\ge C(N,p)^{p-1-\frac{p-1}{N-p} p\ep} \w_{N-1} \left[ \int_0^{\delta_1} r^{-1+ p\ep} \,dr - \frac{p-1}{(2 -\delta_1)^{\frac{N-p}{p-1}}}
\int_0^{\delta_1} r^{-1 + p\ep + \frac{N-p}{p-1}} \,dr \right] \notag \\
&= \frac{C(N,p)^{p-1}}{p} \w_{N-1} \ep^{-1} + o(\ep^{-1}) \quad (\ep \to 0).
\end{align}
Since
\begin{align*}
\nabla u_{\ep, M} = \( \frac{p-1}{p} -\frac{p-1}{N-p} \ep \) U_p^{-\frac{1}{p} -\frac{p-1}{N-p} \ep} \, (\nabla U_p ) \, \psi_M
+ \psi^{'}_M \, U_p^{1-\frac{1}{p} -\frac{p-1}{N-p} \ep} \, (\nabla U_p )
\end{align*}
\tcr{and $(a + b)^p \le a^p + p a^{p-1}b$ for $a, b \ge 0$},
we have
\begin{align*}
&|\nabla u_{\ep, M} |^p \\
&\le \( \frac{p-1}{p} -\frac{p-1}{N-p} \ep \)^p U_p^{-1 -\frac{p-1}{N-p} p \ep} \, |\nabla U_p |^p
+ p \( \frac{p-1}{p} \)^{p-1} \psi^{'}_M \, U_p^{-\frac{p-1}{N-p} \ep -\frac{(p-1)^2}{N-p} \ep} \, |\nabla U_p |^p.
\end{align*}
Then we have
\begin{align}\label{nume}
&\int_{\re^N_+} |\nabla u_{\ep, M}(x,y)|^{p} \,dxdy \notag \\
&\le \int_{[U_p \ge M]} \( \frac{p-1}{p} -\frac{p-1}{N-p} \ep \)^p U_p^{-1 -\frac{p-1}{N-p} p \ep} \, |\nabla U_p |^p \,dxdy
+ \int_{[\frac{M}{2} \le U_p \le M]} |\nabla u_{\ep, M}|^{p} \,dxdy \notag \\
&\le \( \frac{p-1}{p} \)^p C(N,p)^{-1 -\frac{p-1}{N-p}p\ep} \w_{N-1}^{-\frac{p}{p-1}}
\int_{B_{\delta_2}} \( |x|^2 +(1-y)^2 \)^{-\frac{N}{2} + \frac{p\ep}{2} } \tcr{\times} \notag \\
&\hspace{1em} \Biggl[1 + \left\{ \frac{|x|^2 + (y-1)^2}{(2-\delta_2)^2} \right\}^{\frac{N-p}{p-1} +1}
+ 2 \left\{ \frac{|x|^2 + (y-1)^2}{(2-\delta_2)^2} \right\}^{\frac{N-p}{2(p-1)} }
\frac{ \( |x|^2 +y^2 -1 \)_-}{(2-\delta_2)^2} \Biggl] \,dxdy +o(\ep^{-1}) \notag \\
&\le \( \frac{p-1}{p} \)^p C(N,p)^{-1 -\frac{p-1}{N-p}p\ep} \w_{N-1}^{-\frac{p}{p-1}+1}
\int_0^{\delta_2} r^{-1 + p\ep} \, dr +o(\ep^{-1}) \notag \\
&= \( \frac{p-1}{p} \)^p \frac{C(N,p)^{-1}}{p} \w_{N-1}^{-\frac{1}{p-1}} \ep^{-1} + o(\ep^{-1}) \quad (\ep \to 0),
\end{align}
where $(\,f(x) \,)_- := \max \{ 0, \, -f(x) \}$.
From (\ref{deno}) and (\ref{nume}), we have
\begin{align*}
\frac{\int_{\re^N_+} |\nabla u_{\ep, M}(x,y)|^{p} \,dxdy}{\int_{\re^N_+} \frac{|u_{\ep, M}(x,y)|^{p} V_p(x,y)^{\frac{p}{2}}}
{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} \,dxdy }
&\le \frac{\( \frac{p-1}{p} \)^p \frac{C(N,p)^{-1}}{p} \w_{N-1}^{-\frac{1}{p-1}} \ep^{-1} + o(\ep^{-1})}
{\frac{C(N,p)^{p-1}}{p} \w_{N-1} \ep^{-1} + o(\ep^{-1})} \\
&= \( \frac{p-1}{p\,C(N,p)} \)^p \w_{N-1}^{-\frac{p}{p-1}}+ o(1) \\
&= \( \frac{N-p}{p} \)^{p} + o(1) \quad (\ep \to 0).
\end{align*}
Therefore, the constant $\( \frac{N-p}{p} \)^p$ in the inequality (\ref{IH}) is optimal.
\end{proof}
\tcrr{As we mention in Remark \ref{Rem without sym}, the improved inequality (\ref{IH}) is valid for functions without any symmetry by using Proposition \ref{Prop sol U} and a result in \cite{DD}. }
\begin{theorem}\label{Thm IH without sym}
Let $2 \le p < N$.
Then the inequality
\begin{align}
\label{IH}
\( \frac{N-p}{p} \)^p \int_{\re^N_+} \frac{V_p(x,y)^{\frac{p}{2}}}{\( |x|^2 +(1-y)^2 \)^{\frac{p}{2}}} |u(x,y)|^p \,dxdy
&\le \int_{\re^N_+} \left| \nabla u(x,y) \right|^p \,dxdy
\end{align}
holds for any $u \in \dot{W}_0^{1,p} (\re^N_+)$,
where \tcbb{$V_p$ and $X$ are defined in \eqref{V_p}.}
Furthermore, $(\frac{N-p}{p})^p$ is the best constant and is not attained.
\end{theorem}
\begin{proof}
\tcrr{
$U_p$ is a nonnegative function and Proposition \ref{Prop sol U} implies that $-\lap_p U_p \ge 0$ in weak sense for $p \in [2, N)$. Substituting $U_p$ for $\rho$ in \cite{DD}:Theorem 2.1, we have the inequality (\ref{IH}) for any functions $\dot{W}_0^{1,p} (\re^N_+)$, since
\begin{align*}
\( \frac{p-1}{p} \)^p \frac{|\nabla \rho |^p}{\rho^p}
&=\( \frac{p-1}{p} \)^p \frac{|\nabla U_p |^p}{U_p^p} \\
&= \( \frac{p-1}{p} \)^p \( \frac{N-p}{p-1} \)^p \frac{|\nabla h(x,y) |^p}{h(x,y)^p} \\
&= \( \frac{N-p}{p} \)^p \frac{V_p(x,y)^{\frac{p}{2}}}{ \left[ |x|^2 +(1-y)^2 \right]^{\frac{p}{2}} },
\end{align*}
where $h(x,y)$ is given by the proof of Lemma \ref{Lemma IH gene}.
The optimality of the constant $(\frac{N-p}{p})^p$ in the inequality (\ref{IH}) follows from Theorem \ref{Thm IH} and the} \tcbb{non-attainability} \tcrr{follows from \cite{DD}:Theorem 4.1.}
\end{proof}
In the last of this section, we give \tcr{an} improved Hardy-Sobolev \tcr{inequality} on the half-space for $p=2$.
The proof is simpler than the \tcr{that} of Theorem \ref{Thm IH}.
We omit \tcr{it} here.
\begin{theorem}\label{Thm IHS}(Improved Hardy-Sobolev \tcr{inequality} for $p =2$)
Let \tcr{$p=2 < N$}, $0 \le s <2$, and $2^*(s) = \frac{2 (N-s)}{N-2}$.
Then the \tcr{inequality}
\begin{align}\label{IHS}
S_{N,2,s} &\( \int_{\re^N_+} \frac{|u(x,y)|^{2^*(s)}}{\( |x|^2 +(1-y)^2 \)^{\frac{s}{2}}}
\frac{V_2(x,y)}{ \left[ 1-X^{\frac{N-2}{2}} \right]^{\frac{2-s}{N-2}}} \,dxdy \)^{\frac{2}{2^* (s)}}
\le \int_{\re^N_+} |\nabla u(x,y)|^2 \,dxdy,
\end{align}
\tcr{holds} for any $u \in \dot{W}_0^{1,2}(\re^N_+)$
\tcr{of the form $u(x,y)= \tilde{u}(G_{\re^N_+, \tcr{e_N}}(x,y))$ for some function $\tilde{u}$ on $[0, +\infty)$,}
where $V_2(x,y)$ and $X$ is given by \eqref{V_p} in Theorem \ref{Thm IH},
and $S_{N,2,s}$ is the Hardy-Sobolev best constant, i.e.,
\begin{align*}
S_{N,2,s}
= \inf_{u \in C_c^\infty(\re^N) \setminus \{ 0\}} \frac{\int_{\re^N} |\nabla u|^2 \,dx}{\( \int_{\re^N} \frac{|u|^{2^*(s)}}{|x|^s} \,dx \)^{\frac{2}{2^*(s)}}}.
\end{align*}
\end{theorem}
\begin{remark}
\tcrr{The inequality (\ref{IHS}) does not hold \tcr{for functions} without the symmetry (\ref{*'}),
see Proposition \ref{Prop all fct. zero}.}
\end{remark}
\section{Appendix}\label{S App}
First, we show that
the radial critical Sobolev \tcr{space} $\dot{W}_{0, {\rm rad}}^{1,N} (\re^N)$ cannot be embedded
to \tcr{any} weighted Lebesgue \tcr{space} $L^q (\re^N; g(x) \,dx)$ for $q \in [1, \infty)$ and for $g >0$.
\begin{prop}\label{Prop re^N}
There is no weight function $g >0$ such that the inequality
\begin{align*}
C \( \int_{\re^N} |u|^q g(x) \,dx \)^{\frac{N}{q}} \le \int_{\re^N} |\nabla u|^N \,dx
\end{align*}
holds for any $u \in C_{c, {\rm rad}}^1(\re^N)$ for some $C>0$.
\end{prop}
\begin{proof}
Consider the radial test function
\begin{align*}
\phi_R(|x|) =
\begin{cases}
1 \quad&\text{if} \,\, |x| \le 1,\\
\frac{\log \frac{R}{|x|}}{\log R} &\text{if} \,\, 1< |x| < R,\\
0 &\text{if} \,\, |x| \ge R.
\end{cases}
\end{align*}
Direct calculation shows that
\begin{align*}
&\int_{\re^N} |\phi_R |^q g(x) \,dx \ge \int_{B_1} g(x) \,dx >0, \\
&\int_{\re^N} |\nabla \phi_R|^N \,dx = \w_{N-1} \( \log R \)^{1-N} \to 0 \quad (R \to \infty).
\end{align*}
\tcr{Though $\phi_R$ is not $C^1$, we can mollify it as in \cite{HK}:Lemma 8.1, to obtain a $C^1_{rad}(\re^N)$ function with the same property.}
Therefore, we obtain Proposition \ref{Prop re^N}.
\end{proof}
\tcr{Next}, we show \tcr{that} the improved inequalities (\ref{IH}), (\ref{IHS}) in Theorem \ref{Thm IH} \tcr{and Theorem} \ref{Thm IHS}
do not hold without the symmetry (\ref{*'}).
\begin{prop}\label{Prop all fct. zero}
Let $1<p<N$, $0 \le s \,\tcrr{<} \, p$, $p^*(s )= \frac{p(N-s)}{N-p}$,
and $V_p$, \tcr{$X$ }be given in \tcr{\eqref{V_p}}.
Then
\begin{align*}
S:= \inf_{u \in C_c^1(\re^N_+) \setminus \{ 0\}}
\frac{\int_{\re^N_+} |\nabla u(x,y)|^p \,dxdy}{\( \int_{\re^N_+} \frac{|u(x,y)|^{p^*(s)}}
{\( |x|^2 +(1-y)^2 \)^{\frac{s}{2}}} \frac{V_p(x,y)^{\frac{p}{2}}}{\left[ 1-X^{\frac{N-p}{2(p-1)}} \right]^{\frac{(p-1)(p-s)}{N-p}}} \,dxdy \)^{\frac{p}{p^* (s)}}}
= 0\tcr{.}
\end{align*}
\end{prop}
\tcrr{
\begin{proof}
We use the same test function as it in \cite{S(NA)}:Propositon 2.
Let $z=(x,y) \in \re^N_+$ and $z_\ep = (0, \ep)$.
Note that $X = \frac{|x|^2 + (1-y)^2}{|x|^2 + (1+y)^2} \to 1$ and
\begin{align*}
V_p (x, y) &= \( \frac{4(p-1)}{N-p} \)^2 (1-X)^{-2} + o\( (1-X)^{-2} \) \\
&=\( \frac{p-1}{N-p} \)^2 y^{-2} + o\( y^{-2} \)
\end{align*}
as $|z| = \sqrt{ |x|^2 + y^2} \to 0$.
For small $\ep >0$, we define $u_\ep$ as follows:
\begin{align*}
u_{\ep}(z) =
\begin{cases}
v\( \frac{|z-z_{\ep}|}{\ep} \) \,\,\,&\text{if} \,\,\, z \in B_{\ep}(z_{\ep}), \\
0 &\text{if} \,\,\, z \in \re^N_+ \setminus B_{\ep}(z_{\ep}),
\end{cases}
\,\, \text{where}\,\, v(t)=
\begin{cases}
1 \,\,\,&\text{if} \,\,\,0\le t \le \frac{1}{2}, \\
2(1-t) &\text{if} \,\,\, \frac{1}{2} < t \le 1.
\end{cases}
\end{align*}
Then we have
\begin{align*}
&\int_{\re^N_+} | \nabla u_{\ep} (z)|^p \,dz = \ep^{N-p} \int_{B_1} |\nabla v (|z|)|^p \,dz = C \ep^{N-p}, \\
&\int_{\re^N_+} \frac{|u_\ep (x,y)|^{p^*(s)}}{\( |x|^2 +(1-y)^2 \)^{\frac{s}{2}}}
\frac{V_p(x,y)^{\frac{p}{2}}}{ \left[ 1-X^{\frac{N-p}{2(p-1)}} \right]^{\frac{(p-1)(p-s)}{N-p}}} \,dxdy \\
&\ge C \, \ep^{-p} \int_{B_{\ep/2}(z_\ep)} \ep^{-\frac{(p-1)(p-s)}{N-p} } dz
= O \( \ep^{N-p -\frac{(p-1)(p-s)}{N-p}} \) \quad (\ep \to 0).
\end{align*}
Applying $u_\ep$ \tcr{as a test function for} $S$, \tcr{we see}
\begin{align*}
S \le C \, \ep^{(N-p) \(1- \frac{p}{p^*(s)} \) + \frac{(p-1) (p-s) p}{(N-p) p^*(s)}} = O\( \ep^{\frac{N-1}{N-s} (p-s)} \) \to 0\,\,\text{as}\,\, \ep \to 0.
\end{align*}
\end{proof}
}
\tcr{Next proposition is a fact from linear algebra.}
\begin{prop}\label{Prop detJ}
Let $v \in \re^N$, $|v|=1$, $t \in \re$, $A= I + t \,v \otimes v$, \tcr{where $I$ is the identity matrix on $\re^N$.}
Then $A$ has two eigenvalues $1$ and $1+t$.
The multiplicity of $1$ is $N-1$, and the multiplicity of $1+t$ is $1$.
Especially,
\begin{align*}
\det A = 1+t.
\end{align*}
If $t \not=-1$, then there exists the inverse matrix
\begin{align}\label{A^{-1}}
A^{-1} = I - \frac{t}{t+1} \,v \otimes v.
\end{align}
\end{prop}
\begin{proof}
Let $u =(u_1, \dots, u_N)^T$ satisfy $u \cdot v =0$.
Then we have
\begin{align*}
\( (v \otimes v ) u \)_i = \sum_{j=1}^N (v \otimes v)_{i,j} u_j = \sum_{j=1}^N v_i v_j u_j
= \( \sum_{j=1}^N v_j u_j \) v_i = 0.
\end{align*}
Therefore, $A u = (I + t v \otimes v) u = u$ which means that $u$ is the \tcr{eigenvector} of the eigenvalue $1$ of $A$.
\tcr{Note that there are $N-1$ such linearly independent $u$, thus the multiplicity of the eigenvalue $1$ is $N-1$.}
\tcr{Also} since $|v|=1$, for any $i = 1, \dots, N$, we have
\begin{align*}
\( (v \otimes v ) v \)_i & = \sum_{j=1}^N v_i v_j v_j = \( \sum_{j=1}^N v_j v_j \) v_i = v_i
\end{align*}
which implies that $(v \otimes v) v = v$.
Therefore, $A v = (I + t v \otimes v) v = (1 + t) v$ which means that $v$ is the \tcr{eigenvector} of the eigenvalue $1+t$ of $A$.
Hence,
the multiplicity of $1+t$ is $1$ and ${\rm det} (\tcr{I} + t v \otimes v) = 1 + t$.
Next, we show (\ref{A^{-1}}).
Since
\begin{align*}
\tcr{\((v \otimes v)^2\)_{ik}} &= \sum_{j=1}^N (v \otimes v)_{i,j} (v \otimes v)_{jk} = \sum_{j=1}^N v_i v_j v_j v_k \\
&= \( \sum_{j=1}^N v_j^2 \) v_i v_k = (v \otimes v)_{ik},
\end{align*}
we have $(v \otimes v)^2 = (v \otimes v)$. Therefore, we have
$$
(I + t v \otimes v)(\tcr{I} - s v \otimes v) = I + (t-s-st) v \otimes v.
$$
If $t-s-st = 0$, then the right-hand side is $I$.
Therefore, $A^{-1} = I - \frac{t}{t+1} v \otimes v$.
\end{proof}
\begin{prop}\label{Prop Cayley is Mobius}
Let $J, T_b, S_\la$ be given by Definition \ref{def M} and ${\bf B}$ be given by (\ref{Cayley type}).
Then
\begin{align*}
{\bf B} (z) = R \circ J \circ T_{e_N} \circ S_2 \circ J \circ T_{-e_N} (z), \text{where}\,\,R
=
\begin{pmatrix}
1 & & & \\
& \ddots & & \\
& & 1 & \\
& & & -1
\end{pmatrix},\,
e_N = \begin{pmatrix}
0 \\
\vdots \\
0 \\
1
\end{pmatrix}
\end{align*}
for any $z \in \re^N$.
\end{prop}
\begin{proof}
Direct calculation shows that
\begin{align*}
J \circ T_{e_N} \circ S_2 \circ J \circ T_{-e_N} (z)
&=[e_N + 2(z - e_N)^* ]^* = \frac{e_N + 2(z - e_N)^*}{|e_N + 2(z - e_N)^*|^2} \\
&= \frac{e_N + \frac{2(z - e_N)}{|z - e_N|^2}}{\left| e_N + \frac{2(z - e_N)}{|z - e_N |^2} \right|^2} \\
&= \frac{e_N + \frac{2(z - e_N)}{|z - e_N|^2}}{1 + \frac{4}{|z - e_N |^2} + \frac{4 e_N \cdot (z - e_N)}{|z - e_N |^2}} \\
&= \frac{e_N |z-e_N|^2 + 2(z - e_N)}{|z - e_N|^2 + 4 z_N} \\
&= \frac{e_N (|z|^2 -2 z_N + 1) + 2(z - e_N)}{|z|^2 + 2 z_N + 1} \\
&= \frac{(2x, |z|^2 -1)}{|x|^2 + (1 + y)^2} \\
&= \frac{(2x, |x|^2 + y^2 -1)}{|x|^2 + (1 + y)^2}.
\end{align*}
Therefore, we have
\begin{align*}
R \circ J \circ T_{e_N} \circ S_2 \circ J \circ T_{-e_N} (z)
= \frac{(2x, 1- |x|^2 - y^2 )}{|x|^2 + (1 + y)^2}
= {\bf B} (z).
\end{align*}
\end{proof}
Finally, we describe an application of the transformations in III.-(i) in \S \ref{S Harmonic}.
For more general case, see \cite{HK}.
It is well-known that the Sobolev inequality
\begin{align}\label{Sobolev}
\tcr{S_{N,p}} \( \int_{\re^N} |u|^{p^*}\,dx \)^{\frac{p}{p^*}} \le \int_{\re^N} |\nabla u|^p\,dx
\end{align}
for any $u \in \dot{W}_0^{1,p}(\re^N)$, $1 < p < N$, with the best constant
$$
\tcr{S_{N,p}} = \pi^{\frac{p}{2}} N \( \frac{N-p}{p-1} \)^{p-1}
\( \frac{\Gamma(\frac{N}{p})\Gamma(1 + N - \frac{N}{p})}{\Gamma(1 + \frac{N}{2}) \Gamma(N)} \)^{\frac{p}{N}}\tcr{,}
$$
follows from a one-dimensional inequality obtained by Bliss \cite{Bliss}:
Let $v: [0, +\infty) \to \re$ be an absolutely continuous function on $(0, +\infty)$ such that $v' \in L^p(0,+\infty)$, $v(0) = 0$.
Put $q > p > 1$.
Then the inequality
\begin{equation}
\label{Bliss}
C(p,q) \( \int_0^{\infty} \frac{|v(t)|^q}{t^{1 + q(\frac{p-1}{p})}} dt \)^{1/q} \le \( \int_0^{\infty} |v'(t)|^p dt \)^{1/p}
\end{equation}
holds where
$$
C(p,q) = \( \frac{\Gamma\(\frac{q}{q-p}\)\Gamma\(\frac{p(q-1)}{q-p}\)}{\Gamma\(\frac{pq}{q-p}\)} \)^{1/p-1/q} \( \frac{q(p-1)}{p} \)^{1/q}.
$$
See Maz'ya \cite{Mazya}, pp. 274, the equation (4.6.4).
Also see \cite{Os} for a new proof of this classical inequality.
In the following, we consider radial functions only.
To obtain the best Sobolev inequality (\ref{Sobolev}) for radial functions $u(r)$, $r = |x|$, we change the variables
$$
v(t) = u(r), \,\text{where}\,\,t = \frac{p-1}{N-p} \w_{N-1}^{-\frac{1}{p-1}} r^{-\frac{N-p}{p-1}}
$$
and apply the Bliss inequality (\ref{Bliss}) for $q = p^* > p > 1$.
Note that the condition $v(0) = 0$ is satisfied if $u(r) \to 0$ as $r \to \infty$.
\vspace{1em}
Instead, let us change the variables
$$
v(t) = u(r), \,\text{where}\,\, t = \w_{N-1}^{-\frac{1}{N-1}} \log \frac{1}{r}
$$
and put $q > p = N$.
Then the usual computation shows that
\begin{align*}
&\omega_{N-1}^{1-\frac{q}{N}} \int_0^{\infty} \frac{|v(t)|^q}{t^{1 + q(\frac{N-1}{N})}} dt = \omega_{N-1} \int_0^1 \frac{|u(r)|^q}{r \( \log \frac{1}{r} \)^{1 + q(\frac{N-1}{N})}} \, dr
= \int_{B_1^N} \frac{|u(x)|^q}{|x|^N \( \log \frac{1}{|x|} \)^{1 + q(\frac{N-1}{N})}} dx, \\
&\int_0^{\infty} |v'(t)|^N dt = \int_{B_1^N} |\nabla u|^N dx.
\end{align*}
Also in this case, the condition $v(0) = 0$ is equivalent to $u(1) = 0$.
Inserting these identities into \eqref{Bliss}, we obtain
\begin{equation}
\label{Bliss2}
C(q) \( \int_{B_1^N} \frac{|u(x)|^q}{|x|^N \( \log \frac{1}{|x|} \)^{1 + q(\frac{N-1}{N})}} dx \)^{N/q} \le \int_{B_1^N} |\nabla u|^N dx
\end{equation}
for any $u \in \dot{W}^{1,N}_{0, rad}(B_1^N)$ where
\begin{equation}
\label{C_q}
C(q) = \omega_{N-1}^{1-N/q} C(N, q)^{-N}
= \omega_{N-1}^{1-N/q} \( \frac{\Gamma\(\frac{q}{q-N}\)\Gamma\(\frac{N(q-1)}{q-N}\)}{\Gamma\(\frac{Nq}{q-N}\)} \)^{1-N/q}
\( \frac{q(N-1)}{N} \)^{N/q}.
\end{equation}
For the inequality (\ref{Bliss2}), see e.g. \cite{S(JDE)}.
Now, we check that $\lim_{q \to N + 0} C(q) = \( \frac{N-1}{N} \)^N$.
Recall the Stirling formula
$$
\Gamma(s) = \sqrt{2\pi} s^{s-1/2} e^{-s} + o(1), \quad \text{as} \quad s \to +\infty
$$
and put $s = \frac{q}{q-N}$.
Then, we see
\begin{align*}
&\frac{\Gamma\(\frac{q}{q-N}\)\Gamma\(\frac{N(q-1)}{q-N}\)}{\Gamma\(\frac{Nq}{q-N}\)} = \frac{\Gamma(s)\Gamma(\frac{(q-1)N}{q} s)}{\Gamma(Ns)}
\sim \frac{\Gamma(s)\Gamma((N-1)s)}{\Gamma(Ns)} \\
&\sim \sqrt{2\pi} \frac{s^{s-1/2} e^{-s} ((N-1)s))^{(N-1)s-1/2} e^{-(N-1)s}}{(Ns)^{Ns-1/2} e^{-Ns}}
= \sqrt{2\pi} \frac{(N-1)^{(N-1)s-1/2}}{N^{Ns-1/2}} s^{-1/2}
\end{align*}
\tcr{as $q \to N+0$ (which is equivalent to $s= \frac{q}{q-N} \to \infty$),}
and for $C(q)$ in \eqref{C_q}, we have
\begin{align*}
C(q) &\sim \( \frac{\Gamma(s)\Gamma(\frac{N(q-1)}{q} s)}{\Gamma(Ns)} \)^{1/s} \( \frac{q(N-1)}{N} \)^{N/q} \\
&\sim \( \sqrt{2\pi} \frac{(N-1)^{(N-1)s-1/2}}{N^{Ns-1/2}} s^{-1/2} \)^{1/s} (N-1) \\
&\sim \frac{(N-1)^{N-1}}{N^N} (N-1) s^{-1/2s} \to \( \frac{N-1}{N} \)^N \quad \text{as} \quad \tcr{s \to +\infty.}
\end{align*}
Thus if we take a limit $q \to N + 0$ in the inequality \eqref{Bliss2}, we have
the critical Hardy inequality
$$
\( \frac{N-1}{N} \)^N \int_{B_1^N} \frac{|u(x)|^N}{|x|^N \( \log \frac{1}{|x|} \)^N} dx \le \int_{B_1^N} |\nabla u|^N dx
$$
on a unit ball for any $u \in \dot{W}^{1,N}_{0, {\rm rad}}(B_1^N)$.
In conclusion, we obtain the following.
\begin{prop}\label{Prop Bliss}
The Bliss inequality (\ref{Bliss}) \tcr{yields} both the best Sobolev inequality (\ref{Sobolev})
and the generalized critical Hardy \tcr{inequality} (\ref{Bliss2}) for radially symmetric functions.
\end{prop}
\section*{Acknowledgment}
The first author (M.S.) was supported by JSPS KAKENHI Early-Career Scientists, No. JP19K14568.
The second author (F.T.) was supported by JSPS Grant-in-Aid for Scientific Research (B), No. JP19136384.
This work was partly supported by Osaka City University Advanced
Mathematical Institute (MEXT Joint Usage/Research Center on Mathematics
and Theoretical Physics) \tcbb{JPMXP0619217849}.
|
\subsection{Threat Model}
\subsubsection{Attacker Knowledge}
We presume a white-box setting where the attacker is well aware of the parameters of the victim classifier and has direct access to the model gradient. This information is used by the attacker to construct adversarial examples.
\subsubsection{Adversarial Goal}
The objective of the attacker is to compromise the integrity of the victim model. The attack success Rate is defined as (1 - Classification Accuracy) and represents the proportion of total perturbed images in a dataset for which the adversarial noise forces the model to output a wrong label. A lower classification accuracy corresponds to higher attack success rate.
Adversarial goals can be divided into two categories:
\noindent\textbf{Untargeted adversarial attack}: The goal of a non-targeted attack is to slightly modify the source image so that it is classified incorrectly by the target model, without special preference towards any particular output.
\begin{equation}
f(x) \neq f(x^*)
\end{equation}
\noindent\textbf{Targeted adversarial attack}: The goal of a targeted attack is to slightly modify the source image so that it is classified incorrectly into a specified target class $t$ by the target model.
\begin{equation}
f(x^*) = t
\end{equation}
In this work, we consider the targeted adversarial attack setting.
\subsubsection{Noise Budget}
The adversarial examples should be imperceptible by humans, and hence are constrained in amplitude. The noise budget is generally expressed in terms $L_p$-norm distance, mainly $L_0$, $L_2$, and $L_ \infty$:
\begin{equation}
\left\|x\right\|_p = \left( \sum^{n}_{i = 1} \left |x_i \right | ^{p} \right)^{\frac{1}{p}}
\end{equation}
\subsection{Generating Adversarial Examples}
\noindent\textbf{Problem definition}
An adversary, using information learnt about the structure of the classifier, tries to craft perturbations added to the input to cause incorrect classification. For illustration purposes, consider a CNN used for image classification. Given an original input image $x$ and a target classification model $ f(.) $, the problem of generating an adversarial example $x^*$ can be formulated as a constrained optimization \cite{pbform}:
\begin{equation}
\label{eq:adv}
\begin{array}{rlclcl}
x^* = \displaystyle \argmin_{x^* } \mathcal{D}(x,x^* ),\\
s.t. ~ f(x^* ) = l^* , ~ l \neq l^*
\end{array}
\end{equation}
Where $\mathcal{D}$ is a distance metric used to quantify similarity between two images and the goal of the optimization is to minimize the added noise, typically to avoid detection of the adversarial perturbations. $l$ and $l^*$ are the two labels of $x$ and $x^*$, respectively: $x^*$ is considered as an adversarial example if and only if the label of the two images are different ($ f(x) \neq f(x^*) $) and the added noise is bounded ($\mathcal{D}(x,x^*) < \epsilon $ where $\epsilon \geqslant 0 $).
To solve this optimization problem, several approaches have been proposed in the literature from which present the most widely used attacks:
\noindent \textbf{Fast Gradient Sign Method (FGSM).}
FGSM \cite{fgsm} is a single-step, gradient-based, attack. An adversarial example is generated by performing a one step gradient update along the direction of the sign of gradient at each pixel as follows:
\begin{equation}
x^* = x + \epsilon sign (\nabla_{x}J_{\theta}(x,y))
\end{equation}
Where $\nabla J()$ computes the gradient of the loss function $J$ and $\theta$ is the set of model parameters. The $sign()$ denotes the sign function and $\epsilon$ is the perturbation magnitude.
\noindent \textbf{Projected gradient descent (PGD).}
PGD \cite{pgd} is a stronger iterative variant of the FGSM where the adversarial example is generated as follows:
\begin{equation}
x^{t+1} = \mathcal{P}_{\mathcal{S}_x}(x^t + \alpha \cdot sign (\nabla_{x}\mathcal{L}_{\theta}(x^t,y)) )
\end{equation}
Where $\mathcal{P}_{\mathcal{S}_x}()$ is a projection operator projecting the input into the feasible region $\mathcal{S}_x$ and $\alpha$ is the added noise at each iteration.
The PGD attack tries to find the perturbation that maximizes the loss of a model on a particular input while keeping the size of the perturbation smaller than a specified amount.
\noindent \textbf{Carlini \& Wagner (C\&W).}
This attack \cite{C&W} is one of the state-of-the-art attacks. This latter has 3 forms based on different distortion measures ($l_0, l_2, l_{\infty}$). In this work we only consider the $l_2$ form as it has the best performance. It generates adversarial examples by solving the following optimization problem: \vspace{-5pt}
\begin{equation}
\begin{array}{rrclcl}
\displaystyle \minimize_{\delta} & \multicolumn{3}{l}{\left\Vert \delta \right\Vert_2 + c\cdot l(x+\delta)}\\
\textrm{s.t.} & x + \delta \in [0,1]^n
\end{array}
\end{equation}
Where $\left \Vert \delta \right \Vert_2$ is the smallest perturbation measured by the $l_2$ norm that makes the model misclassify into another/target class. $l(\cdot)$ is the loss function reflecting the distance between the current situation and the objective of the attack defined as:
\begin{equation}
l(x) = max(max_{i \neq t } \{Z(x)_i\}-Z(x)_t - \kappa )
\end{equation}
Where $Z(x)$ is the output of the layer before the softmax called \textit{logits}. $t$ is the target label, and $\kappa$ is called the confidence, a hyper-parameter used to enhance the transferability of the output.
An adversarial example is considered as successful if $max_{i \neq t}\{Z(x)_i\}-Z(x)_t \leq 0$. In the C\& W attack, the box constrained optimization problem $x + \delta \in [0,1]^n$ is turned to an unconstrained problem by replacing $\delta$ with $\frac{1}{2}(tanh(w)+1)-x$, where $w$ is a new optimizer ranging in $(-\infty , +\infty)$.
\subsection{Setup}
Our experiments include implementations of a CNN architecture (Four convolutional layers and three fully connected layers) trained with MNIST \cite{mnist} for handwritten digit recognition. MNIST is composed of $60,000$ images, with $10$ classes and is composed of grey-scaled images of size $28\times28$ pixels.
We also use Wide ResNet-34 CNN trained on CIFAR-10 database \cite{CIFAR} for object recognition in the evaluation. This database consists of 60,000 $32 \times 32$ RGB images in $10$ classes, with $6,000$ images per class. Finally, we consider VGG-19 CNN trained on ImageNet \cite{imagenet}, which contains over $14$ labeled million images of $224 \times 224$ pixels each.
To evaluate the attacks, we used two commonly used adversarial attack generation algorithms, namely:
PGD \cite{pgd} and C\&W \cite{C&W}.
The PGD attack is currently the strongest known attack for the $l_\infty$ metric. The C\&W attack is considered to be one of the strongest $l_2$ attacks.
Our implementations are built using the open source machine learning framework PyTorch \cite{PyTorch}. We modified the FoolBox Library \cite{foolbox} to support our approach and to evaluate the attacks.
Experiments were taken on NVIDIA Tesla K80 GPU.
\subsection{Evaluation of the offline adversarial pattern exploration}
In this section, we evaluate the efficiency of the adversarial noise pattern extraction proposed in Section \ref{sec:off}.
For this reason, we use different patterns with different similarity score levels as the exploration initial starting point and monitor the classification accuracy of the model under attack. Specifically, we chose patterns with three similarity score levels: the minimum, the median, and the maximum among the exploration set. Recall that this latter corresponds to our choice of noise pattern when using the offline exploration.
Figures \ref{scores_mnist} and \ref{scores_cifar} show that using patterns with higher similarity scores leads to more powerful attacks. In fact, with a perturbation magnitude of $0.2$, we report a model classification accuracy degraded to less than $20\%$ when using the pattern with the highest score, while it remains up to $90\%$ when using the pattern with the lowest score. Changing the size of the exploration set ($m$) used to produce distinct adversarial patterns also reveals that the larger the employed set, the more likely it is to uncover a pattern with a higher similarity score and thereby a higher attack success rate. Notice that the $m$ samples of the considered sets are randomly chosen.
\begin{figure}[!htp]
\centering
\includegraphics[width=0.8\columnwidth]{figures/pcc_different_scores.PNG}
\caption{Model classification accuracy under PGD attack when using patterns generated with max, median and min similarity scores for MNIST (m is the exploration Set size)}
\label{scores_mnist}
\end{figure}
\begin{figure}[!htp]
\centering
\includegraphics[width=0.8\columnwidth]{figures/pcc_different_scores_cifar.png}
\caption{Model classification accuracy under PGD attack when using patterns generated with max, median and min similarity scores for CIFAR-10 (m is the exploration Set size)}
\label{scores_cifar}
\end{figure}
\subsection{Evaluation under noise and time constraint}
In these experiments, we set a time budget for the online processing while varying the noise budget and comparing the performance of the state-of-the-art attacks to ROOM-enhanced version of the attacks.
\begin{figure*}[!htp]
\centering
\includegraphics[width=2\columnwidth]{figures/pgd_room_mnist.png}
\caption{Model classification accuracy under the PGD (blue) and the PGD-ROOM (red) attacks for different time budgets (MNIST).}
\label{pgd_mnist_acc}
\end{figure*}
\noindent\textbf{Attacks on MNIST.} To assess the effectiveness of ROOM, we randomly choose four different pairs of (source, target) classes from MNIST test set and generate the optimal adversarial pattern for each to be used as initialization when generating the AE.
In Figure \ref{pgd_mnist_acc}, we show the model accuracy under ROOM and conventional PGD attack comparatively for different noise budgets $\varepsilon$, and for different time budgets.
We can observe that PGD-ROOM efficiently generates adversarial examples at a maximum throughput of $1250$ FPS under a noise budget of $0.25$. More specifically, ROOM can \emph{totally} jeopardize the classification accuracy of the victim model processing data streaming in real-time with a speed of $1250$ FPS. However, at the same pace, the conventional online PGD attack was unable to have any impact on model accuracy even for a higher noise budget of $0.4$.
Moreover, for a throughput of $28$ FPS, the victim model \emph{misclassifies the totality of samples} under ROOM attack, while the baseline attack's maximum success rate is $75\%$ even for a higher noise budget.
\begin{figure*}[!htp]
\centering
\includegraphics[width=2\columnwidth]{figures/cw_room_mnist.png}
\caption{Model classification accuracy under the C\&W (blue) and the C\&W-ROOM (red) attacks for different time budgets at various confidence levels (MNIST).}
\label{cw_mnist_acc}
\end{figure*}
In Figure \ref{cw_mnist_acc}, we evaluate the performance of the conventional C\&W comparatively to ROOM-C\&W. We measure the model accuracy under both attacks while varying the time budget for different confidence levels. Notice that \textbf{confidence} parameter for C\&W is the target confidence of the misclassification for an adversarial example. Thus, unlike the PGD attack that crafts adversarial examples within a given perturbation level, C\&W finds the smallest perturbation needed to cause misclassification with a given target confidence level.
Furthermore, higher confidence requires more time to be reached since it controls the gap between the generated AE and the decision boundary.
For zero-confidence, and with a throughput of $2$ FPS, C\&W-ROOM reaches $67\%$ success rate, while the model accuracy remains intact with the conventional attack. The conventional C\&W attack requires $5 \times$ more time to reach the same performance, which illustrates the effectiveness of our approach.
\begin{figure*}[!htp]
\centering
\includegraphics[width=1.6\columnwidth]{figures/pgd_room_cifar10.png}
\caption{Model classification accuracy under the PGD (blue) and PGD-ROOM (red) attacks for different time budgets (CIFAR-10).}
\label{pgd_cifar_acc}
\end{figure*}
\noindent\textbf{Attacks on CIFAR-10.} We generate adversarial examples for images from the CIFAR-10 dataset. Attacking deeper models and larger size and more complex images (RGB) requires a higher time budget. As shown in Figure \ref{pgd_cifar_acc}, ROOM outperforms the conventional PGD attack.
For the same target attack success rate of $60\%$ under a noise budget $\varepsilon = 0.03$, PGD-ROOM offers a throughput of $200$ FPS, while the maximum throughput of the conventional PGD is only $1.4$ FPS.
The same trend was observed for C\&W-based attacks: Figure \ref{cw_cifar_acc} shows clearly that ROOM outperforms the conventional attack. For instance, for a time budget of $10$ seconds, ROOM is $30\%$ more efficient than the conventional C\&W. For a confidence equal to $5$, the conventional attack takes $2 \times$ more time to reach the same success rate.
\begin{figure*}[!htp]
\centering
\includegraphics[width=2\columnwidth]{figures/cw_room_cifar10.png}
\caption{Model classification accuracy under the C\&W (blue) and the C\&W-ROOM (red) attacks for different time budgets at various confidence levels (CIFAR-10).}
\label{cw_cifar_acc}
\end{figure*}
\noindent\textbf{Attacks on ImageNet.}
In this section, we evaluate our approach on the ImageNet dataset. As shown in Figure \ref{pgd_imagenet_acc}, our proposed PGD-ROOM attack is found to be more effective.
PGD-ROOM delivers the same attack success rate of $60\%$ for a throughput of $16$ FPS when the noise budget $\varepsilon = 0.025$. However, the conventional attack achieves the same attack effectiveness for a throughput of $0.625$ FPS.
The same trend has been observed for C\&W-based attacks, as shown in Figure \ref{c&w_imagenet_acc}. C\&W-ROOM is $15\%$ more effective than the conventional C\&W attack for the same allocated time.
\begin{figure*}[!htp]
\centering
\includegraphics[width=1.6\columnwidth]{figures/pgd_room_imagenet.png}
\vspace{-8pt}
\caption{Model classification accuracy under the PGD (blue) and PGD-ROOM (red) attacks for different time budgets (ImageNet).}
\label{pgd_imagenet_acc}
\end{figure*}
\begin{figure*}[!htp]
\centering
\includegraphics[width=1.3\columnwidth]{figures/cw_room_imagenet.png}
\vspace{-8pt}
\caption{Model classification accuracy under the C\&W (blue) and C\&W-ROOM (red) attacks for different time budgets at various confidence levels (ImageNet).}
\label{c&w_imagenet_acc}
\end{figure*}
\subsection{Does ROOM have inherent impact on attack generation?}
In this section, our objective is to investigate the impact of the offline exploration from a \emph{pure attack efficiency perspective} for a given noise budget. We want to answer whether ROOM's impact is inherent or can be achieved by any random initialization?
For a conclusive evaluation, we explore ROOM comparatively with both zero and random initialization of the adversarial noise. We use PGD with different numbers of steps.\\
\noindent\textbf{MNIST.} We set the size of perturbation as $\varepsilon = 0.3$ with a step size of $\frac{8}{255}$. We use PGD with different numbers of iterations and we report the classification accuracy in Table \ref{initial_MNIST}. We notice that ROOM outperforms all other approaches. For instance, with PGD-10 (PGD with $10$ iterations), ROOM is $3\times$ more powerful than random noise-based attack and more than $6\times$ more powerful than zero noise-based attack.
\begin{table}[!htp]
\centering
\caption{Model classification accuracy under PGD attack with different initialization methods on MNIST.}
\label{initial_MNIST}
\begin{tabular}{cccccc}
\toprule
\textbf{Initialization} & \textbf{PGD-40} & \textbf{PGD-10} & \textbf{PGD-5} & \textbf{PGD-3} & \textbf{PGD-2} \\
\midrule
Zero & 1.6\% & 84.9\% & 96.8\% & 98.5\% &99.5\%\\
Random & 0.2\% & 47\% & 89.5\% &95.8\%&97.1\%\\
\textbf{ROOM} & \textbf{0\%} & \textbf{13.7\%} & \textbf{47.7\%} & \textbf{69.4\%} & \textbf{76.3\%} \\
\bottomrule
\end{tabular}
\end{table}
\noindent\textbf{CIFAR-10.} We set the size of perturbation as $\varepsilon = 0.03$ with a step size of $\frac{2}{255}$. As shown in Table \ref{initial_CIFAR}, using ROOM made the attack more powerful, $1.5\times$ and $1.3\times$ more effective than zero and random noise-based attacks, respectively.
\begin{table}[!htp]
\centering
\caption{Model classification accuracy under PGD attack with different initialization methods on CIFAR-10.}
\label{initial_CIFAR}
\begin{tabular}{cccccc}
\toprule
\textbf{Initialization} & \textbf{PGD-20} & \textbf{PGD-4} & \textbf{PGD-3} & \textbf{PGD-2} & \textbf{PGD-1} \\
\midrule
Zero & 0\% & 9.56\% & 19\% &40.87\% & 71.66\%\\
Random & 0\% & 8.8\% & 15.2\%&32.28\% & 62.1\%\\
\textbf{ROOM} & \textbf{0\%} & \textbf{6.74\%} & \textbf{13.1\%} & \textbf{26.53\%} & \textbf{48.81\%} \\
\bottomrule
\end{tabular}
\end{table}
\noindent\textbf{ImageNet.} We use VGG-19 as the classifier for testing ImageNet. The noise magnitude is set to $\varepsilon = 0.03$ with step size $\frac{2}{255}$. As illustrated in Table \ref{initial_imagenet}, using ROOM with PGD attack made the attack more powerful, $1.6\times$ and $1.3\times$ more effective than zero and random noise-based attacks, respectively for PGD-3.
Those results confirm that our offline exploration significantly impacts an attack generation perspective even without considering time constraints. ROOM helps generate more effective adversarial attacks when compared to other initialization methods and for a limited number of attack iterations.
\begin{table}[!htp]
\centering
\caption{Model classification accuracy under PGD attack with different initialization methods on ImageNet.}
\label{initial_imagenet}
\begin{tabular}{ccccc}
\toprule
\textbf{Initialization} & \textbf{PGD-20} & \textbf{PGD-5} & \textbf{PGD-3} & \textbf{PGD-1} \\
\midrule
Zero & 0\% & 15.62\% & 34.37\% & 75\%\\
Random & 0\% & 12.5\% & 28.12\% & 71.87\%\\
\textbf{ROOM} & \textbf{0\%} & \textbf{8.37\%} & \textbf{20.87\%} & \textbf{55.62\%} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{ROOM vs YOPO}
In the quest of accelerating the adversarial training process, You Only Propagate Once (YOPO) \cite{yopo} has been recently proposed. YOPO is based on reducing the total number of full forward and backward propagation to only one for each group of adversary updates by restricting most of the forward and back propagation within the first layer of the network, taking advantage of the baseline training gradient backpropagation.
To evaluate our approach comparatively with YOPO, we run experiments where we set a time budget to $0.04$s and $0.08$s for adversarial example generation (corresponding respectively to $25$ and $12$ FPS) and we compare the effectiveness of each attack. Since ROOM is orthogonal to YOPO (YOPO focus on accelerating gradient-based noise generation, while ROOM focuses on noise patch initialization), we also explore the results of combining both techniques.
\noindent\textbf{MNIST.} As illustrated in Table \ref{yopo_mnist}, PGD-ROOM is more efficient in generating adversarial examples than YOPO initialized with zero and random noise for a throughput of 25 FPS. PGD-ROOM is nearly $6.5\times$ and $2.5\times$ more efficient than YOPO-Zero and YOPO-Random, respectively. Interestingly, we also noticed that combining YOPO with ROOM resulted in more gain in attack success rate.
\begin{table}[!htp]
\centering
\caption{Model classification accuracy under ROOM vs YOPO on MNIST for a throughput of 25 FPS.}
\label{yopo_mnist}
\begin{tabular}{ccc}
\toprule
\textbf{Attacks} & \textbf{Eps = 0.2} & \textbf{Eps = 0.3} \\
\midrule
PGD-Zero & 52.5\% & 51.2\% \\
PGD-Random & 41.9\% & 17.6\% \\
PGD-ROOM & 23.2\% & 6.8\% \\
YOPO-Zero & 44.8\% & 44.6\% \\
YOPO-Random & 43\% & 16\% \\
YOPO-ROOM & 20.2\% & 3.7\% \\
\bottomrule
\end{tabular}
\end{table}
\noindent\textbf{CIFAR-10.} For a larger model, and a throughput of 12 FPS, we noticed the same trend; ROOM with PGD outperforms YOPO, and combining ROOM and YOPO yields to even better performance. For instance, YOPO-ROOM is more than $2\times$ more powerful than YOPO-random, and PGD-ROOM is $1.4\times$ more successful than YOPO-random.
\begin{table}[!htp]
\centering
\caption{Model classification accuracy under ROOM vs YOPO on CIFAR-10 for a throughput of 12 FPS.}
\label{yopo_cifar}
\begin{tabular}{ccc}
\toprule
\textbf{Attacks} & \textbf{Eps = 0.02} & \textbf{Eps = 0.03} \\
\midrule
PGD-Zero & 37\% & 30\% \\
PGD-Random & 28\% & 21\% \\
PGD-ROOM & 15\% & 12\% \\
YOPO-Zero & 23\% & 19\% \\
YOPO-Random & 20\% & 17\% \\
YOPO-ROOM & 16\% & 8\% \\
\bottomrule
\end{tabular}
\end{table}
In conclusion, we notice that ROOM generates more effective adversarial attacks than YOPO under the same time constraint, which is a critical metric in scenarios like adversarial training when adversarial examples need to be generated for the whole training set.
Moreover, YOPO can be applied only to gradient-based attacks, while ROOM can be integrated into any attack generation method.
Interestingly, since ROOM is orthogonal to YOPO, we noticed that combining both techniques results in an even more efficient adversarial example generation. We believe this is a promising property of ROOM to help provide more optimized AML protection.
\subsection{Proposed Approach: Offline-Online Attack Model }
\begin{figure*}[!htp]
\centering
\includegraphics[width=2\columnwidth]{figures/Real_time_adv_ROOM.pdf}
\caption{Architecture of the proposed approach: Offline and Online exploration.
\label{process}
\end{figure*}
In contrast with the existing work on adversarial machine learning, this paper suggests including time as an analysis perspective of adversarial noise generation. More specifically, in addition to the noise budget used in the state-of-the-art, we consider \textbf{time budget} as an orthogonal constraint taken into account in the adversarial noise generation towards more practical threat models and defenses.
To include time as a constraint in the adversarial noise generation process, we first distinguish the time budget from the two design spaces, i.e., online and offline, as follows:
\begin{itemize}
\item \textbf{Online time budget:} which we note $T_{on}$, is defined as the time required for the online exploration, i.e., after the acquisition of the victim sample to target.
\item \textbf{Offline time budget:} which we note $T_{off}$, is defined as the time required for the offline exploration to generate the adversarial perturbation. During this time, we assume that the attacker has no access to the victim data sample to target.
\end{itemize}
Hence, we consider the adversarial noise generation as a continuous process combining offline and online processing.
We distinguish the two corner cases mentioned earlier in Section \ref{sec:intro}, where: (i) $(T_{off} = 0, T_{on}= \infty)$ corresponds to the conventional digital attack where all the computations are performed online without time limit,
and (ii) $(T_{off}= \infty , T_{on} = 0)$ corresponds to the universal adversarial perturbation referred to as Offline attack where all the computations are performed offline without time limit.
In this section, we propose real-time Offline-Online Model (ROOM), a new methodology for adversarial noise generation by combining both offline and online exploration under a time constraint. The proposed offline exploration is based on an analysis detailed later in Section \ref{sec:motiv}, where we explore the opportunities provided by patterns in adversarial examples that we exploit to warm-up the online exploration instead of starting from random initialization. Figure \ref{process} shows an overview on the proposed methodology and the two phases (offline and online) are detailed in Sections \ref{sec:off} and \ref{sec:on}.
\subsection{Offline Exploration}\label{sec:off}
The main objective of the offline exploration is to identify the most efficient adversarial noise pattern that corresponds to a static adversarial component on which we can later (during the online exploration) build upon to quickly converge to an adversarial example. Algorithm \ref{offline_exp_algo} gives a detailed description of the offline exploration mechanism. To efficiently identify a potential noise pattern, the exploration process implements the following steps: First, we select a set of images for which we generate the corresponding adversarial examples (AEs) and collect adversarial perturbations $P_i$ (Line $4$ -- $7$ in Algorithm \ref{offline_exp_algo}). Next, we calculate the correlation between the resulting noise distributions (Line 9 -- 13).
Subsequently, the aim is to identify the perturbation that has the highest correlation with the other samples, which correspond to the highest \emph{intra-class similarity} (detailed analysis in Section \ref{sec:motiv}). For this reason, we define the similarity score of a given noise as the sum of its PCCs with the remaining noise candidates:
\begin{align*}
\mathcal{S}^{i}_{score} = \sum_{k \neq i} PCC_{i,k}
\end{align*}
Notice that the higher the similarity score, the closer the noise sample is to the potential static component (adversarial pattern) of the given intra-class setting. In fact, the noise candidate with the maximum $\mathcal{S}_{score}$ represents the highest static component that is redundant within most of the set's noise samples. Therefore, we finally identify the noise pattern as the noise candidate with the highest similarity score (Line 19--20).
\begin{algorithm}
\caption{Offline Exploration (Optimal noise pattern identification) }
\label{offline_exp_algo}
\begin{algorithmic}[1]
\State \textbf{Input:} a classifier $f$, $x$ input image, $y$ true class, $y_{target}$ target class, noise budget $\varepsilon$, $m$ size of the used Set
\State \textbf{Output:} $pattern_{adv}$
\State \textit{// Generate AEs and collect $P_i$}
\For {i = 0 \dots m-1}
\State $x^*_i = Attack(f, x_i, y_i, y_{target}, \varepsilon)$
\State $P_i = {x^*_i - x_i}$
\EndFor
\State \textit{// Calculate the correlation between patterns $P_i$ }
\For {s = 0 \dots m-1}
\For {k = 0 \dots m-1}
\State $PCC_{s,k} = corr(P_s, P_k)$
\EndFor
\EndFor
\State \textit{// Set a similarity score for each candidate pattern $P_i$ as the sum of PCCs }
\For {i = 0 \dots m-1}
\State $\mathcal{S}^{i}_{score} = \sum_{k \neq i}PCC_{i,k}$
\EndFor
\State \textit{// Identify the $P_i$ with the highest correlation score}
\State $j = argmax(\mathcal{S}_{score}) $
\State $pattern_{adv} = P_j$
\end{algorithmic}
\end{algorithm}
\subsection{Online Exploration}\label{sec:on}
After the offline exploration, an adversarial noise pattern is identified as a potential static component that better characterizes the path "Source--Target" given a decision boundary defined by the trained victim classifier.
Our goal is to propose a novel approach for attacking trained models while considering a time budget. Hence, in the quest for a rapid adversarial example generation, the idea is to take advantage of the offline exploration to enhance the online generation efficiency. More specifically, we accelerate the conventional adversarial noise generation approaches with the offline-identified \emph{adversarial pattern}.
Essentially, the perturbation identified from the previous analysis is used as the initial starting point for a new adversarial attack targeting the same class. In fact, the noise pattern is identified such that it represents a static noise component that brings a given input from a source class $l$ closer to a target class $k$. Therefore, instead of starting the exploration of the adversarial noise from a zero or a random matrix, we start the online space exploration from an intermediate point that has a higher chance to be close to the decision boundary, and hence easier to flip the data sample classification to the target label. More importantly, the online exploration becomes faster in producing adversarial examples, allowing it to meet real-time constraints.
Algorithm \ref{proposed_pgd} details an illustration of the online exploration where we use the proposed technique to build a Projected Gradient Descent (PGD) based attack where the adversarial example is initialized using the previously generated pattern.
\begin{algorithm}
\caption{Proposed PGD-ROOM attack}
\label{proposed_pgd}
\begin{algorithmic}[1]
\State \textbf{Input:} a classifier $f$ with loss $J$, noise budget $\varepsilon$,step size $\alpha$, $x$ input image, $y_{target}$ targeted class, $pattern_{adv}$ adversarial pattern (identified offline), $m$ number of iterations, $T_{on}$ online time budget
\State \textbf{Output:} $x_{adv}$
\State $begin\_time = time.time()$
\State Initialize~ $x_{adv} = pattern_{adv}$
\For {i= 0...m-1}
\If {$time.time()-begin\_time > T_{on}$}
\State break
\EndIf
\State prediction = $argmax(f(x_{adv})) $
\State $x_{adv} = Clip \{x - \varepsilon sign (\nabla_{x_{adv}}J_{\theta}(f(x_{adv}),y_{target})) \}$
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Introduction}\label{sec:intro}
\input{Introduction}
\section{Background}\label{sec:bckg}
\input{Background}
\section{Time-Aware Analysis of Adversarial Noise}
\input{Methodology}
\section{Experiments}
\input{Experiments}
\label{exp}
\section{How does the offline exploration accelerate adversarial noise generation?}\label{sec:motiv}
\input{motivation}
\section{Discussion}
\input{Discussion}
\section{Related work}\label{sec:rw}
\input{related_work}
\section{Conclusions}
\input{Conclusions}
\bibliographystyle{IEEEtran}
\subsection{Pearson Correlation Coefficient}
PCC coefficient is defined as follows:
\begin{equation}
PCC_{X,Y} = \frac{cov(X,Y)}{\sigma_X \sigma_Y}
\end{equation}
where $cov$ indicates the covariance and $\sigma_X$ and $\sigma_Y$ are the
standard deviations of matrices $X$ and $Y$, respectively, and the $PCC$ values range from $-1$ to $1$. The absolute value indicates the extent to which the two variables are linearly correlated, with $1$ indicating perfect linear correlation, $0$ indicating zero linear correlation, and the sign indicates whether they are positively or negatively correlated.
The targeted adversarial noise is iteratively constructed to transform a sample from a source class (correct label) towards an adversarial sample $x^*_i$ with a target class (target misclassification label). This transformation path has naturally higher similarity for noise generated on samples from the same source targeting the same class, compared to noise generated on samples belonging to different pair (source,target) classes. Therefore, our intuition is that a static noise component could be present in high-dimensional space among the samples from a given class in the path towards a specific target class given a decision boundary.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{figures/pcc_pgd_mnist_intravsinter.png}
\caption{Gaussian distribution of Pearson correlation coefficients PGD MNIST intra-class vs inter-class.}
\label{pgd_mnist_pcc}
\end{figure}
We analyze the similarities between noise distributions within a given setting (source,target), which we call intra-class similarity, comparatively with the distribution of noise similarities between two different settings (source,target), which we call inter-class similarity.
Let $x_i$ be an input sample and $f$ a given classifier such that $f(x_i) = l$. Let be a targeted adversarial attack that generates an adversarial noise defined as:
\begin{equation}
\varepsilon^{lk}_i ~. s.t. ~ f(x_i + \varepsilon^{lk}_i) = k \neq l
\end{equation}
Let ${S}_{ij}$ be the similarity between two noises $\varepsilon^{lk}_i $ and $\varepsilon^{pq}_j$ for two samples $x_i$ and $x_j$, respectively, defined as follows:
\begin{equation}
\mathcal{S}_{ij} = PCC(\varepsilon^{lk}_i , \varepsilon^{pq}_j)
\end{equation}
Where PCC is the Pearson correlation coefficient \cite{pcc}, used as a similarity metric. We call $\mathcal{S}_{ij}$ an \textbf{inter-class} similarity when $(p,q) \neq (l,k)$, and we call $\mathcal{S}_{ij}$ an \textbf{intra-class} similarity when $(p,q) = (l,k)$.
In Figures \ref{pgd_mnist_pcc}, \ref{pgd_cifar_pcc} and \ref{imagenet_pcc_}, we represent the distribution of different measured similarity $\mathcal{S}_{ij}$ values for MNIST, CIFAR-10 and ImageNet datasets, respectively. We notice that the intra-class similarity distribution is clearly higher than the inter-class similarity distribution for the three datasets. This observation supports our intuition of the presence of patterns for the same (source, target) setting of adversarial noise. It, therefore, explains the mechanism by which an offline exploration allows to converge more quickly to an adversarial example.
\begin{figure}[!htp]
\centering
\includegraphics[width=\columnwidth]{figures/pcc_pgd_cifar_intravsinter.png}
\caption{Gaussian distribution of Pearson correlation coefficients PGD cifar-10 intra-class vs inter-class.}
\label{pgd_cifar_pcc}
\end{figure}
\begin{figure}[!htp]
\centering
\includegraphics[width=\columnwidth]{figures/pcc_pgd_imagenet_intravsinter.png}
\caption{Gaussian distribution of Pearson correlation coefficients PGD ImageNet intra-class vs inter-class.}
\label{imagenet_pcc_}
\end{figure}
|
\section*{Conflict of interest}
\bibliographystyle{spmpsci}
\section{Method}
\subsection{Ranking as a Markov Decision Process for Online Learning to Rank}
Ranking can be formalized as an MDP problem, where the search engine (agent) has to decide which document to place (action) at rank $i$, given the current candidate document set (state)~\cite{wei2017reinforcement}. In the offline LTR setting, i.e., when relevance labels are provided, any IR evaluation measure can be used as a reward function for the ranking created by the search engine. In the online LTR setting, however, rewards are typically inferred via users‘ implicit feedback, such as clicks. Following Wei et al.~\cite{wei2017reinforcement}, who have limited their attention to offline LTR only, we define each component of the MDP ranking for the OLTR as follows:
\textbf{States} $\boldsymbol {S}$ indicates the set of states that the agent (search engine) will observe from the environment. In the MDP ranking, at a time step $t$, a state $s_t\in{S}$ is the candidate document set $D_t$ which contains the documents that need to be ranked.
\textbf{Actions} $\boldsymbol {A}$ is the set of possible actions that the agent can take when at a state $s_t$. An action at a time step $t$, denoted as $a_t\in{A(s_t)}$, consists of selecting a document $d_{m(a_t)}\in{D_t}$ to place at rank $i$, where rank $k = t+1$\footnote{Note that the time step $t$ starts from 0 and the rank position $k$ starts from 1.}, $m(a_t)$ is the function that maps the action to the document index, and $d_{m(a_t)}$ is represented by a feature vector. We note the feature elements in $d_{m(a_t)}$ are not only obtained based on the document but also on the query.
\textbf{Transition} $\boldsymbol {T(s,a)}$ is the function $T(s_t,a_t) = s_{t+1}$, which maps a state $s_t$ to its next state $s_{t+1}$ based on the action $a_t$, where $s_{t+1}$ is the candidate document set $D_t$ without the selected document $d_{m(a_t)}$:
\begin{equation} \label{transaction}
s_{t+1} = s_t \setminus d_{m(a_t)} =D_t \setminus d_{m(a_t)} = D_{t+1}
\end{equation}
\textbf{Policy} $\boldsymbol {\pi(a|s, \theta)}$ is a probability distribution over all possible actions that the agent can take when in the state $s$, given the current model parameters $\theta$ (where $\theta$ is a vector containing the model's parameters). We compute this distribution using the softmax function over document scores:
\begin{equation}\label{policy}
\pi(a_t|s_t, \theta) =
\frac{exp \left( f_\theta(d_{m(a_t)} \right) }
{\sum_{a\in A(s_t)}exp \left( f_\theta(d_{m(a_t)} \right) }
\end{equation}
$f_\theta(d_{m(a_t)})$ is the document relevance score estimated by the ranking model. In our experiments, we use a linear ranking model, this means the dimensionality of $\theta$, i.e., the number of parameters in the ranking model, is equal to the dimensionality of the feature vector. However, we note that the ranking model can be extended to any dimensionality, e.g., a neural ranking model. For simplicity, we use $a_t \sim \pi(s_t)$ to denote the action $a_t$ selected according to the policy $\pi$ at state $s_t$.
\textbf{Reward} $\boldsymbol {R(s,a)}$ is the reward function that models the immediate reward given by the environment and its value represents the reward assigned to the action $a$ that has been selected by the agent in state $s$.
When full information is provided (i.e., the true relevance labels are given), it is straightforward to define the reward function on the basis of IR evaluation measures such as DCG. However, in the OLTR setting, relevance labels are not provided -- in place of these, the ranker observes implicit, noisy and biased user feedback such as clicks. Thus, in such a partial information setting, the reward function is defined based on the implicit feedback, e.g., on click labels. We define the value of the reward function next.
\subsection{Unbiased reward shaping}\label{unbiasedness}
In RL, reward shaping is used to reshape the original reward function to better guide the direction of the gradient update~\cite{ng1999policy}. Prior knowledge about the environment is needed to formalize a reliable reward shaping function to avoid otherwise to bias learning~\cite{popov2017data}.
\subsubsection{Na\"ive reward function}
In the (offline) LTR approach by Wei et al.~\cite{wei2017reinforcement}, DCG scores have been used as the original reward function to instruct the search engine to learn a ranker, when true relevance labels are supplied. This reward function $R_{DCG}(s_t,a_t)$ is defined as:
\begin{equation} \label{eq:3}
R_{DCG}(s_t,a_t) = \lambda(t)\cdot y_{m(a_t)}
\end{equation}
\noindent where $\lambda(t)=\frac{1}{log_2(t+2)}$ is the DCG weight term, and $y_{m(a_t)}\in\{0,1\}$ is the binary relevance label of the document at rank $t+1$. This is a reasonable reward function, provided that relevance labels are known. This reward function directly corresponds to the DCG evaluation metric (because of the DCG weight term), and thus the agent will attempt to learn a policy $\pi$ that directly maximizes the cumulative rewards $\Delta$ (DCG scores) of the ranking episode, for any given initial state $s_0$:
\begin{equation} \label{eq:5}
\Delta_{R_{DCG}}( s_0, \pi, y) = \sum_{t=0}R_{DCG}(s_t,\pi(s_t))
= \sum_{y_{m(a_t)}=1}\lambda(t)
\end{equation}
However, in the \textit{online} LTR setting, relevance labels are unknown (i.e., this is a partial information setting); instead, users' clicks are used as learning signal. A na\"ive reward function in this context is then to treat clicks as binary relevance labels and thus use the DCG reward function of equation~\ref{eq:3}:
\begin{equation} \label{eq:6}
R_{NAIVE^+} (s_t,a_t) = \lambda(t)\cdot c_{m(a_t)}
\end{equation}
\noindent where $c_{m(a_t)}=1$ represents that a user has clicked on the document at rank $t+1$, while $c_{m(a_t)}=0$ represents no click on the document. If we assume that the probability of a document $d_{m(a_t)}$ to be observed by a user (known as propensity) only depends on the rank position $t+1$ (note that $t$ starts from 0) and no click noise is present, i.e., $P(o_{t+1}=1 | t+1)$, then the expectation of the final ranking rewards is biased to the users' propensity~\cite{joachims2017unbiased,jagerman2019model}:
\begin{equation} \label{eq:7}
\begin{aligned}
\mathbb{E}_o[\Delta_{R_{NAIVE^+}}(s_0, \pi, c) ] &=\sum_{t=0}R_{NAIVE^+}(s_t,\pi(s_t)) \\
&= \sum_{y_{m(a_t)}=1}P(o_{t+1}=1 | t+1) \cdot \lambda(t)
\end{aligned}
\end{equation}
This is the so-called position bias effect: top-ranked documents usually have a larger chance of being observed, and thus the rewards assigned to them are biasedly higher.
\subsubsection{IPS reshaping}
Recent work in CLTR has attempted to account for this position bias effect~\cite{ai2018unbiased,joachims2017unbiased}. One of the most commonly used approaches to mitigate position bias is \textit{Inverse Propensity Scoring} (IPS). We follow this direction to unbias the click signal in our RL framework, and define the IPS reward shaping function as $F_{IPS^+}$ and the reshaped reward function as $R_{IPS^+}$:
\begin{equation}
\begin{aligned}
R_{IPS^+} (s_t,a_t)& = F_{IPS} (R_{NAIVE^+} (s_t,a_t))\\
&= \frac{R_{NAIVE^+}(s_t,a_t)}{P(o_{t+1}=1 | t+1)}= \frac{\lambda(t)}{P(o_{t+1}=1 | t+1)} \cdot c_{m(a_t)}
\end{aligned}
\end{equation}
It can be proven that the cumulative na\"ive rewards reshaped by $F_{IPS^+}$ are an unbiased estimate of the ranking rewards with true relevance labels~\cite{joachims2017unbiased}: $\mathbb{E}_o[\Delta_{R_{IPS^+}}(s_0, \pi, c) ] =\Delta_{R_{DCG}}(s_0, \pi, y)$.
Thus, $F_{IPS}$ can be used to obtain unbiased cumulative rewards with respect to position bias, likely providing more reliable gradient estimations than the na\"ive reward function.
\subsubsection{Negative rewards}
The reward functions above only provide positive rewards to the clicked/relevant documents, while they assign a zero reward to the unclicked/irrelevant documents. However, in other RL problems, it is often found that negative rewards help the agent to avoid selecting poor actions~\cite{sutton2018reinforcement}: for a ranker, negative rewards can help avoid selecting irrelevant documents from the document set, thus leading to better user experience. Based on this, we introduce a negative DCG reward function for online LTR (a similar one could be defined for offline LTR):
\begin{equation}
R_{NAIVE^-} (s_t,a_t) = \lambda(t)\cdot (c_{m(a_t)}-1)
\end{equation}
Hence, the negative cumulative ranking score is calculated by:
\begin{equation}
\Delta_{R_{NAIVE^-}} (s_0, \pi, c) = \sum_{t=0}R_{NAIVE^-}(s_t,\pi(s_t)) = \sum_{c_{m(a_t)}=0}-\lambda(t)
\end{equation}
This means highly ranked unclicked documents will be penalised more and thus drag down the final ranking reward. It can be proven that this na\"ive negative reward function is a biased estimate of the negative DCG score of the ranking ($\Delta_{R_{DCG^-}} ( s_0, \pi, y)$):
\begin{align*}
&\mathbb{E}_o[\Delta_{R_{NAIVE^-}} (s_0, \pi, c)] \\
&= \mathbb{E}_o\left[\sum_{t=0}R_{NAIVE^-} (s_t,\pi(s_t))\right]
=\mathbb{E}_o\left[\sum_{c_{m(a_t)}=0}-\lambda(t)\right] \\
&=\mathbb{E}_o\left[\sum_{ y_{m(a_t)}=0}-\lambda(t)\right] + \mathbb{E}_o\left[\sum_{o_{t+1}=0\wedge y_{m(a_t)}=1}-\lambda(t)\right] \\
&= \sum_{y_{m(a_t)}=0}-\lambda(t) - \sum_{y_{m(a_t)}=1}(1-P(o_{t+1}=1|t+1))\cdot\lambda(t)\\
&=\Delta_{R_{DCG^-}} ( s_0, \pi, y) - \sum_{y_{m(a_t)}=1}(1-P(o_{t+1}=1|t+1))\cdot\lambda(t) \numberthis \label{eqn}
\end{align*}
That is, the na\"ive negative reward overestimates the true negative DCG reward by counting the relevant but not observed documents as irrelevant.
\subsubsection{Negative IPS reward shaping function}
In order to get unbiased negative DCG rewards, we propose the negative IPS reward shaping function $F_{IPS^-}$ and denote the reshaped negative DCG reward as $R_{IPS^-}$:
\begin{align*}
&R_{IPS^-} (s_t,a_t)=F_{IPS^-}(R_{NAIVE^-} (s_t,a_t)) \numberthis \label{eqn}\\
& = \lambda(t)\cdot (c_{m(a_t)} - 1) + \left (\frac{1-P(o_{t+1}=1|t+1)}{P(o_{t+1}=1|t+1)} \right )\lambda(t)\cdot c_{m(a_t)}
\end{align*}
We mathematically prove that $\mathbb{E}_o[\Delta_{R_{IPS^-}} (s_0, \pi, c)]$ is an unbiased estimate of $\Delta_{R_{DCG^-}} (s_0, \pi, y)$:
\begin{align*}
&\mathbb{E}_o[\Delta_{R_{IPS^-}} (s_0, \pi, c)] = \mathbb{E}_o\left[\sum_{t=0}F_{IPS^-}(R_{NAIVE^-}(s_t,\pi(s_t)))\right] \\
&=\mathbb{E}_o\left[\sum_{t=0}\lambda(t)\cdot (c_{m(a_t)} - 1) + \left (\frac{1-P(o_{t+1}=1|t+1)}{P(o_{t+1}=1|t+1)} \right )\lambda(t)\cdot c_{m(a_t)}\right] \\
&=\mathbb{E}_o\left[ \sum_{t=0}\lambda(t)\cdot (c_{m(a_t)} - 1)\right] +\\
&+\mathbb{E}_o\left[ \sum_{t=0}\left(\frac{1-P(o_{t+1}=1|t+1)}{P(o_{t+1}=1|t+1)} \right )\lambda(t)\cdot c_{m(a_t)}\right] \\
&=\mathbb{E}_o[\Delta_{R_{NAIVE^-}} (s_0, \pi, c)] + \\
&+\mathbb{E}_o\left[\sum_{o_{t+1}=1\wedge y_{m(a_t)}=1}\left(\frac{1-P(o_{t+1}=1|t+1)}{P(o_{t+1}=1|t+1)} \right )\lambda(t)\right] \\
&=\Delta_{R_{DCG^-}} (s_0, \pi, y) - \sum_{y_{m(a_t)}=1}(1-P(o_{t+1}=1|t+1))\cdot\lambda(t) \\
& +\sum_{y_{m(a_t)}=1}(1-P(o_{t+1}=1 | t+1)) \cdot \lambda(t)\\
&=\Delta_{R_{DCG^-}} (s_0, \pi, y)
\numberthis \label{eqn}\\
\end{align*}
This reward shaping function allows to assign unbiased negative rewards to unclicked documents so as to obtain an unbiased cumulative negative DCG ranking score.
\subsubsection{Prior knowledge of propensity}
All the unbiased reward shaping functions above require to know a priori the users' propensities $P(o_{t+1}=1|t+1)$. Many recent works have considered estimating such propensities from historical click-logs~\cite{agarwal2017effective,fang2019intervention,agarwal2019estimating} and during training~\cite{ai2018unbiased,wang2018position}. We regard propensity estimation as being beyond the scope of this article and in our experiments we assume the propensities to be known. Nevertheless, for completeness, we also test how sensitive our method is to propensity mismatch (Section~\ref{sec_propensity_missmatch}).
\subsection{Learning with policy gradient} \label{policy_gradient}
Following previous work~\cite{wei2017reinforcement,yao2020rlper}, we learn the policy model parameters $\theta$ with REINFORCE~\cite{williams1992simple,sutton2018reinforcement,sutton2000policy}, a widely used policy gradient algorithm. In REINFORCE, the objective is to find an optimal policy that can maximize the expectation of cumulative reward from the beginning of each episode, $J(\theta) = \mathbb{E}_\pi\left[ G_t\right]$, where $t=0$.
Here $G_t$ is the discounted future cumulative reward starting from time step $t$, $G_t = \sum_{m=t}^{M}\gamma^{m-t}\cdot R(s_{m},a_{m})$, and
$M$ is the maximum depth of the ranking episode, and $\gamma \in [0, 1]$ is the discount factor. In previous offline LTR work~\cite{wei2017reinforcement,jun2020ppg}, $\gamma$ has been set to $1$ (maximum value), which results in $G_0 = \Delta_{R_{DCG}}$. This value results in the learning algorithm directly maximizing DCG. However, it is well known in RL that larger $\gamma$ values will lead the agent to care more about future rewards but at the same time to produce gradient estimations with significantly high variance, thus slowing down the learning speed~\cite{jun2020ppg,sutton2018reinforcement}. In OLTR settings, however, learning speed is very important as it is highly entangled with user experience, as measured by the online performance of the ranker.
Hence, in order to reduce variance and speed up learning, we simplify MDP to Contextual Bandits~\cite{jagerman2020safe,adomavicius2005incorporating,hofmann2011contextual} by setting $\gamma=0$. This setting makes REINFORCE to choose $a_t$ so as to maximize only the expectation of immediate reward $R(s_t, a_t)$:
\begin{equation}
J(\theta) = \mathbb{E}_\pi\left[ R(s_t, a_t)\right]
\end{equation}
This setting actually does not change the fact that the objective is to optimize DCG. This is because maximizing the expected immediate reward at a time step $t$ is equivalent to selecting the most likely relevant document from the candidate set $D_t$ for the state $s_t$: this is guaranteed to have maximum expected cumulative reward for that ranking episode. Thus, it is safe to ignore future rewards in this case. In Section~\ref{sec_discount}, we empirically show that $\gamma=0$ enjoys a much faster learning speed than other settings, without any loss in final convergence. It is important to note that the above is not true for tasks such as search result diversification, since greedily choosing the most relevant document at each rank position may cause lower final ranking scores~\cite{feng2018greedy}. Thus, we use the full MDP algorithm for this special situation.
Following the standard policy gradient practice~\cite{wei2017reinforcement,sutton2018reinforcement}, we estimate gradients with \textit{Monte Carlo sampling} and the gradient $\nabla_\theta J(\theta)$ can be calculated as:
\begin{equation}\label{PG}
\nabla_\theta J(\theta) = R(s_t, a_t)\nabla_\theta log(\pi(a_t|s_t, \theta) )
\end{equation}
Intuitively, the gradients will update the ranker parameters towards the actions that yield the highest immediate reward $R(s_t, a_t)$.
The complete procedure of ROLTR is described in Algorithm~\ref{algo:ReOLT}. At each ranking episode $i$, i.e., at each round of user interaction (line \ref{lst:line:2} for-loop), the search engine receives a query $q_i$ and the initial candidate document set $D_0$ is generated (line \ref{lst:line:3}). Then the algorithm first draws an action (document) from the distribution created by the current policy (line \ref{lst:line:6}). Then, the selected action, the state information, and the corresponding document are recorded (line \ref{lst:line:7}) and the environment moves to the next state (line \ref{lst:line:8}). The same procedure is repeated for the next rank positions, until the algorithm reaches the maximum depth for the ranking (line \ref{lst:line:5} for-loop). After finishing ranking, the final result list is shown to the user, who provides feedback to the search engine in the form of click labels (line \ref{lst:line:10}). Next, for each recorded state-action pair (line \ref{lst:line:11} for-loop), the gradient is calculated (line \ref{lst:line:14}) and the current policy is updated at the end of each ranking episode (line \ref{lst:line:16}).
\begin{algorithm}[t]
\caption{ Reinforcement Online Learning to Rank (ROLTR).}
\begin{algorithmic}[1]
\STATE {\bf Input}: Initial weights $\theta$, learning rate $\alpha$, reshaped reward function $R$, transition function $T$, number of docs on SERPs $M$.
\FOR{$i$ $\leftarrow$ 0....$\infty$} \label{lst:line:2}
\STATE $q_{i} \leftarrow receive\_query(i)$, $D_{0}\leftarrow canditate\_set(q_i)$ \label{lst:line:3}
\STATE $s_0 \leftarrow D_0$, $L_{i}\leftarrow [$ \space $]$, $\Delta \theta \leftarrow 0$
\FOR{$t$ $\leftarrow$ 0....$M$} \label{lst:line:5}
\STATE $a_t \leftarrow \pi(s_t, \theta)$\hfill// Eq.\ref{policy} \label{lst:line:6}
\STATE $L_i.append(s_t, a_t, d_{m(a_t)})$ \label{lst:line:7}
\STATE $s_{t+1} \leftarrow T(s_t,a_t)$\hfill// Eq.\ref{transaction} \label{lst:line:8}
\ENDFOR
\STATE $C_i\leftarrow receive\_clicks(L_{i})$ \label{lst:line:10}
\FOR{$t$ $\leftarrow$ 0....$M$} \label{lst:line:11}
\STATE $s_t, a_t \leftarrow L_i.get(t)$, $c_{m(a_t)}\leftarrow C_i.get(t)$
\STATE $\Delta \theta \leftarrow \Delta \theta+ R(s_t, a_t)\nabla_\theta log(\pi(a_t|s_t, \theta) )$ \hfill// Eq.\ref{PG} \label{lst:line:14}
\ENDFOR
\STATE $\theta \leftarrow \theta + \alpha\Delta \theta$ \label{lst:line:16}
\ENDFOR
\end{algorithmic}
\label{algo:ReOLT}
\end{algorithm}
\subsection{Guarantee of Unbiasedness for ROLTR}
ROLTR has two main parts: the learning part (Section~\ref{policy_gradient}), which uses the REINFORCE algorithm, and the unbiased reward shaping part (Section~\ref{unbiasedness}). The theoretical guarantees of REINFORCE are well studied in reinforcement learning, including the guarantee that the gradient estimate is unbiased with respect to maximizing the expected rewards~\cite{sutton2018reinforcement}. However, if the rewards are biased (which are when the naive reward function is used), then the learning algorithm is biased. This is where the second part of ROLTR comes into play: the unbiased rewards shaping functions guarantee the reward signals assigned to REINFORCE are unbiased with respect to position bias (mathematical proof in Eq.~\ref{eqn}). Thus, the update gradients estimated by ROLTR are guaranteed to maximize unbiased expected rewards, and thus the gradient estimation is unbiased with respect to position bias.
\section{Conclusion}
We proposed a novel OLTR algorithm, ROLTR, which formalises OLTR as an MDP ranking problem and uses reinforcement learning with unbiased reward shaping functions to directly optimize an IR metric. In contrast to traditional online evaluation based methods and the current state-of-the-art PDGD method, ROLTR does not fully rely on online interventions to overcome user position bias. Instead, ROLTR directly uses the IPS reward shaping functions to de-bias rewards given by the environment which can be further used to guide gradient estimation. As a result, the gradient calculated by ROLTR is unbiased with respect to position bias. In order to accelerate the learning speed and obtain better user experience, we first simplified the MDP ranking by setting the reward discount factor $\gamma$ to 0, which we empirically confirmed it provides a lower gradient variance without hurting final convergence. Furthermore, to fully leverage user click feedback in each training episode, we introduced a negative IPS reward shaping function for ROLTR which provides additional unbiased reward learning signal from unclicked documents. We have proven that the rewards reshaped by the negative IPS reward shaping function are also unbiased with respect to position bias. Our experimental results show that ROLTR achieves state-of-the-art offline performance requiring less user interactions, which results in considerably better user experience (online performance) over other OLTR methods.
Future work will be directed towards considering estimating user observation propensity during training since our method requires propensity to be known a priori. Furthermore, other biases such as selection bias and other type of user behaviour signals, such as dwell time, mouse move, etc, could be modelled within ROLTR, potentially improving its performance. This could be achieved by simply changing the reward function while maintaining other parts of the method unchanged. This makes ROLTR flexible and with plenty of room for further extension.
\section{Results}
\subsection{RQ1: Impact of the reward discount factor}
\label{sec_discount}
To answer RQ1, we study the reward discount factor $\gamma$ (set to 1 in previous work~\cite{wei2017reinforcement}). As discussed in Section~\ref{policy_gradient}, we expect that setting $\gamma=0$ will lead to smaller variance in gradient estimation and a consequent speed up of training, without loss in terms of final convergence.
\begin{figure}
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{gammas_biased.png}
\caption{Reward function: $R_{NAIVE^+}$}
\label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{gammas_unbiased.png}
\caption{Reward function: $R_{IPS^+}$}
\label{fig:sub2}
\end{subfigure}
\caption{Offline nDCG@10 score of ROLTR with different reward discount factor $\gamma$. (MSLR10K dataset)}
\label{fig:1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{gammas_unbiased_longterm.png}
\caption{Long-term offline nDCG@10 score of ROLTR with different reward discount factor $\gamma$. (MSLR10K dataset, reward function: $R_{IPS^+}$).}
\label{fig:7}
\end{figure}
To confirm this assumption, in this section we empirically investigate the convergence of the learned rankers with different values of $\gamma$. In this study, we use the reward function $R_{NAIVE^+}$ and $R_{IPS^+}$, and simulate the user propensity with $\eta=1$. We report results on MSLR10k. Figure~\ref{fig:1} illustrates the offline nDCG@10 learning curves obtained throughout the training process. The plots clearly show that the learning curves obtained when $\gamma=0$ (blue curves) is always above that associated with other $\gamma$ values, for both reward functions and both click settings. For naive reward and noisy click setting, rankers trained with lower value of $\gamma$ achieved higher offline nDCG@10 score at the point of 100,000 impressions. This indicates that ROLTR learns fastest when $\gamma=0$. On the other hand, ROLTR converges slower when $\gamma$ increases, and when $\gamma=1$ (red curves) the convergence speed is the slowest. Interestingly, unbiased reward function $R_{IPS^+}$ seems more sensitive to $\gamma$ than the naive reward function $R_{NAIVE^+}$. Although $R_{IPS^+}$ has faster learning speed than $R_{NAIVE^+}$ when $\gamma$ is small (e.g., $\gamma = 0, 0.3, 0.5$), the learning speed decreases dramatically when $\gamma$ is large, and $R_{IPS^+}$ even fails to converge at 100,000 impressions under perfect clicks. We also studied the convergences of ROLTR with the $R_{IPS^+}$ reward function and $\gamma = 0, 0.5, 1$ for long-term impressions (1,000,000 impressions). These results are presented in Figure~\ref{fig:7}. Similar to the trends in Figure~\ref{fig:1}, larger $\gamma$ values exhibit slower convergence and fail to converge to the same nDCG value as when $\gamma=0$, even after performing one million impressions. This is due to the IPS significantly enlarging the click signal at lower ranks and this effect being cumulated to the rewards assigned to the clicks at higher ranks with large $\gamma$, thus resulting in considerably larger variance.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{variance_bar_new.png}
\caption{Average variance of the gradient vector for each sampled episode. Error bars correspond to the 95\% confidence intervals. (MSLR10K dataset)}
\label{fig:2}
\vspace{-10pt}
\end{figure}
Following Xu et al.~\cite{jun2020ppg}, we also directly compare the variance of the estimated gradients when $\gamma = 0$, $\gamma = 0.5$ and $\gamma = 1$ for both naive and unbiased reward functions. For this, we calculate the variance of the gradient vectors at each training episode using the \textit{trace of the covariance matrix}. We report the results for all runs on the MSLR10K in Figure~\ref{fig:2} (along to those for $\gamma=0.5$ to show the overall trend for varying values between the two extremes).
The variance from the noisy click runs is higher than that from the perfect click runs for all values of $\gamma$ and both reward functions. This indicates that strong click noise causes high variance in gradient estimations. Moreover, the variance recorded for $\gamma=0$ is almost half that recorded for $\gamma=1$, for both noisy and perfect clicks and both reward functions. This finding agrees with the reinforcement learning practice: smaller discount factors always result in lower gradient variance.
When comparing $R_{NAIVE^+}$ and $R_{IPS^+}$, we notice that $R_{IPS^+}$ yields almost 10 times larger variance than that of the same setting for $R_{NAIVE^+}$. This explains why $R_{IPS^+}$ is more sensitive to $\gamma$ and larger $\gamma$ significantly slows down the learning speed, as shown in Figure~\ref{fig:1}.
In reinforcement learning, it is well known that $\gamma$ controls the trade-off between variance and final convergence. Larger values of $\gamma$ result in higher final convergence at the expense of higher variance (thus slower learning speed), and vice versa: smaller values result in lower variance (faster learning speed) but suboptimal convergence. However, in OLTR, our experiments show that optimizing the expected immediate reward (i.e. $\gamma=0$) will not hurt the final convergence at least after 100,000 impressions: the optimal policy learned with $\gamma=0$ is equal to the optimal policy learned with $\gamma=1$, which is maximizing the DCG score of the ranking. We further note that previous work that formalizing LTR as an MDP shows high variance in gradient estimation~\cite{wei2017reinforcement} and that recent work has attempted to fix this issue for offline LTR~\cite{jun2020ppg}: setting $\gamma = 0$ can be considered a simpler way to fix this problem. We believe this insight may have positive uptake for offline LTR.
With respect to RQ1, then, setting $\gamma=0$ would not hurt the offline performance (convergence) of the learnt ranker, while delivering lower gradient variance than when $\gamma=1$. Hence, we set $\gamma=0$ in the remaining experiments.
\subsection{RQ2: Impact of reward shaping functions}
To answer RQ2, we consider the impact of different reward shaping in ROLTR. Figure~\ref{fig:3} reports the \textit{offline} performance of ROLTR with different reward functions over $100,000$ impressions on the MLSR-WEB10K dataset. For both noisy and perfect click settings, all IPS reward functions outperform their naive versions in terms of learning speed. This means that it is important to de-bias the otherwise biased reward shaping function. In addition, we observe that ROLTR with $R_{IPS^{+}} + R_{IPS^{-}}$ and $R_{NAIVE^{+}} + R_{NAIVE^{-}}$ converges faster than when only using positive or negative reward functions. This is under both click settings, and it suggests that leveraging both clicked and unclicked signals is beneficial. This phenomenon is however less obvious with noisy clicks due to noisy clicks decreasing all the rankers' performance, thus resulting in worse and closer nDCG@10 scores. On the other hand, when noisy clicks are considered, $R_{IPS^{+}}$ and $R_{IPS^{-}}$ have a very similar learning curve; while, in the perfect click settings, $R_{IPS^{+}}$ converges much faster than $R_{IPS^{-}}$.
We explain this difference of behaviour between perfect and noisy click settings as follows. In the perfect click setting, $R_{IPS+}$ only uses signal from the clicked documents, which are relevant documents only in this click setting. $R_{IPS-}$ instead also exploits signal from not-clicked documents (it down-weights this signal), which are non-relevant documents only in this click settings -- this however is a worse signal than that from relevant documents only ($R_{IPS+}$). Thus, it is logical to see $R_{IPS+}$ performing better (faster learning curve) than $R_{IPS-}$.
In the noisy click setting, however, the signals exploited by $R_{IPS+}$ and $R_{IPS-}$ are similar: while still $R_{IPS+}$ only uses clicked documents (while $R_{IPS-}$ also uses not-clicked), in the noisy click setting both relevant and non-relevant documents are clicked. This therefore reduces the differences in performance observed between $R_{IPS+}$ and $R_{IPS-}$.
\begin{figure}[th]
\includegraphics[width=\linewidth]{unbiased_large.png}
\caption{Offline nDCG for ROLTR when different reward shaping functions are used (MSLR10K dataset). }
\label{fig:3}
\vspace{-10pt}
\end{figure}
When studying the final convergence at $100,000$ impressions, we do not record significant differences (i.e., $p>0.05$) except for $R_{NAIVE^{-}}$. This exception can be explained as $R_{NAIVE^{-}}$ only assigns biased negative rewards to unclicked documents and as result, in this case ROLTR just aims to avoid selecting documents that give low rewards, but it ignores valuable documents.
With respect to RQ2, then, we found that using unbiased reward shaping functions to reshape naive rewards make ROLTR converge faster (i.e. it requires less impressions). Specifically, the use of both unbiased positive and negative reward functions make ROLTR converge the fastest. Hence, in the remaining experiments, we set $R_{IPS^{+}} + R_{IPS^{-}}$ as reward function.
The fact that biased rewards do not change the final performance but slow down convergence may seem counter-intuitive at first.
In fact, in CLTR studies, the biased learning objective results in a local optimal for the final convergence~\cite{joachims2017unbiased}. However, this may not be the case for OLTR. This is because CLTR experiments use a logging ranker to collect clicks and then train a new ranker with the click log: thus the rank of a document is decided by the logging ranker and has been fixed in the log. In OLTR settings like those used here, instead, the deployed ranker is interactively updated after each session: this means every document has a chance to be ranked at top of the displayed rankings hence be observed: these online interventions can eliminate position bias. However, this does not come for free, as it requires a large number of interactions to eventually obtain a good convergence. This is exactly what ROLTR attempts to solve: reduce the number of interactions to reach convergence. ROLTR in fact is the first OLTR algorithm that directly de-biases user position bias, with the results above showing that it can significantly speed up learning (this in turn will translate in better user experience, see Section~\ref{rq4}).
\subsection{RQ3: Final convergence (offline nDCG)}
\begin{figure}[p!]
\centering
\includegraphics[width=0.9\columnwidth]{mslr10k_new.png}
\includegraphics[width=0.9\columnwidth]{yahoo_new.png}
\includegraphics[width=0.9\columnwidth]{istella_new.png}
\caption{Offline nDCG@10 for OLTR methods and MDPRank under noisy and perfect clicks.}
\label{fig:4}
\end{figure}
To answer RQ3, we compare the \textit{offline} performance of the learned rankers over $100,000$ impressions.
Figure~\ref{fig:4} reports the learning curves for different methods. COLTR is clearly better than PMGD and DBGD in terms of learning speed and final convergence across almost all datasets: this is similar to previous findings~\cite{oosterhuis2016probabilistic,zhuang2020counterfactual}. The gradient descent based methods (ROLTR and PDGD) significantly outperform candidate ranker sampling based methods (DBDG, PMGD and COLTR), rendering DBGD based methods outdated.
On the other hand, ROLTR and PDGD have similar learning curves across all three datasets. This is especially the case when perfect clicks are used. The final convergence of the two methods across all datasets and click settings is not statistically significantly different ($p>0.05$). However, when noisy clicks are considered, we find that ROLTR has a faster learning speed at the beginning of training. Figure~\ref{fig:5} considers the learning curves for the early impressions (first $10,000$) when noisy clicks are used. This figure clearly shows that the learning curves of ROLTR are almost always above other OLTR baselines at early impressions (except for the first $3,000$ impressions on MSLR10k, for which PMGD is best).
With respect to RQ3, then, we conclude that, when clicks are noisy, the learning speed of ROLTR is faster than PDGD in the early stage of training.
Finally, comparing the final convergence of ROLTR with the effectiveness of the offline MDPRank (trained with relevance labels and thus indicating skyline effectiveness), we can clearly observe that ROLTR's effectiveness is at par to that of MDPRank, when the perfect click setting is used (Figure~\ref{fig:4}). Recall that the perfect click setting still exhibits position bias. This findings empirically demonstrates the unbiasedness of ROLTR. However, when noisy clicks are used for ROLTR, then it's performance is sensibly lower to that of the offline MDPRank, showing how much noise in the click signal hurts performance, compared to perfect relevance labels.
\newcommand*\rot{\rotatebox{90}}
\newcommand{\bfseries}{\bfseries}
\subsection{RQ4: User experience during training (online nDCG)} \label{rq4}
\begin{figure}[p!]
\includegraphics[width=\columnwidth]{mslr10k_10k_new.png}
\includegraphics[width=\columnwidth]{yahoo_10k_new.png}
\includegraphics[width=\columnwidth]{istella_10k_new.png}
\caption{Offline nDCG@10 for the first 10,000 impressions (noisy clicks).}
\label{fig:5}
\end{figure}
To answer RQ4, we consider the quality of the experience users withstand while rankers are trained (online nDCG). Table~\ref{online} reports the \textit{online} performance of each method. As expected, online NDCG@10 scores obtained when learning from perfect clicks are higher than for noisy clicks, suggesting the latter hurt user experience.
Although COLTR has better offline performance than online evaluation based methods, it does lead to a worse user experience during the learning phase. This indicates that using counterfactual evaluation for candidate ranker comparison requires a lot more exploration. These findings are in agreement with our early findings~\cite{zhuang2020counterfactual}.
We also find that online evaluation based methods are significantly worse than ROLTR and PDGD ($p<0.01$) for both perfect and noisy clicks, agreeing with the offline nDCG learning curves (Figure~\ref{fig:4}). This is because ROLTR and PDGD assemble result lists by sampling documents from a softmax probability distribution, while online evaluation based methods use interleaving or multileaving to create these lists. This suggests that online evaluation based methods perform more exploration, thus hurting user experience more.
When comparing ROLTR and PDGD, we find that ROLTR obtains the best average online performance across all datasets under noisy clicks (agreeing with the learning curves in Figure~\ref{fig:5} for offline nDCG). Even for perfect clicks, for which the offline learning curves of ROLTR and PDGD are similar, ROLTR does statistically significantly outperform PDGD on MSLR10K and Istella ($p<0.05, p<0.01$ respectively). There are no statistically significant differences between ROLTR and PDGD on Yahoo! ($p>0.05$). This may be because the average proportion of irrelevant documents for each query is small (21.92\%), and thus it may be less likely the methods perform a ranking error, making them hard to distinguish. To conclude, with respect to RQ4, we found that ROLTR delivers the best user experience among the investigated methods.
\subsection{RQ5: Sensitivity to propensity mismatch}
\label{sec_propensity_missmatch}
\begin{small}
\begin{table}[t!]
\centering
\caption[centre]{Online nDCG@10. Bold values indicate the highest average performance. Significant gains and losses of ROLTR over PDGD are marked by ${\vartriangle}$ ($p<0.05$) and ${\blacktriangle}$ ($p<0.01$). }\label{online}
\begin{tabularx}{\linewidth}{@{}p{1.2cm}XXXX@{}}
\toprule
Clicks&Algorithm& MSLR10K & Yahoo!& Istella\\
\midrule
&DBGD & 519.99& 1165.11&510.34\\
\multirow{ 2}{*}{\textit{Perfect}}
&PMGD& 545.22 & 1191.26 &564.58\\
&COLTR& 448.65 & 1121.92 &348.15\\
&PDGD & 579.22& \bfseries 1310.99&741.45\\\cline{2-5}
&ROLTR & \bfseries 587.46 ${\vartriangle}$& 1302.11& \bfseries 808.66${\blacktriangle}$\\
\midrule
&DBGD & 477.29&1116.89&408.09\\
\multirow{ 2}{*}{\textit{Noisy}}
&PMGD & 535.42 & 1137.39&426.77\\
&COLTR& 431.27 & 1105.05 &238.30\\
&PDGD & 516.77& 1227.28&481.85\\\cline{2-5}
&ROLTR & \bfseries 543.28 ${\blacktriangle}$& \bfseries 1238.10&\bfseries 654.04${\blacktriangle}$\\
\bottomrule
\end{tabularx}
\end{table}
\end{small}
To answer RQ5, we investigate the sensitivity of ROLTR to mismatched user propensity. Our previous experiments, in fact, assumed that the user propensity is known a priori and used the true propensity to obtain the unbiased rewards. However, this assumption does not always hold true in practice, as the user propensity could be overestimated or underestimated for various reasons~\cite{joachims2017unbiased}.
\begin{figure}
\includegraphics[width=\linewidth]{propensities_large.png}
\vspace{-6pt}
\caption{Offline nDCG@10 for ROLTR when different prior propensities are used (MSLR10K dataset).}
\label{fig:6}
\end{figure}
Figure \ref{fig:6} shows the \textit{offline} nDCG@10 of ROLTR under different propensity values $\eta$ and for different click settings. When $\eta=0$, ROLTR assumes no user position bias and the unbiased reward function $R_{IPS}$ reduces to $R_{NAIVE}$. On the other hand, $\eta=1$ is the same propensity used to simulate clicks, and thus the propensity assumed by the ranker and that seen in the clicks are the same. Note that when simulating the user, we always assume the user has propensity $\eta=1$ (while instead we change ROLTR's propensity value).
From Figure \ref{fig:6} we can observe that, in the noisy click setting, an underestimation of propensity ($\eta<1$) leads to a similar final convergence as that obtained when the true propensity is considered. In fact, when $\eta=0.5$ we record a higher final convergence. However, when the propensity is overestimated ($\eta>1$), both learning speed and final convergence become much worse. A similar behaviour has been found in CLTR and relates to the large variance introduced by the extreme IPS weights~\cite{jagerman2020accelerated}: Because of the overestimated propensities, the IPS weights will be much bigger than those of the true propensity, resulting in a large amount of variance in the gradient estimation, thus hurting the final performance. There are many ways to prevent this from happening, e.g., `propensity clipping' which trades-off bias against variance.
On the other hand, under perfect clicks, the final convergence of all propensity priors is similar, except for the extremely overestimated propensity ($\eta=2.0$). In fact, in the perfect setting the main difference is with respect to the learning speed: overestimated propensities converge faster than underestimated propensities.
With respect to RQ5, then, we conclude that, when clicks are perfect, ROLTR is not sensitive to propensity mismatch in terms of final convergence; but when clicks are noisy, only underestimated propensities are not sensitive while overestimated propensities will hurt both the final convergence and learning speed.
\section{Related work}
\subsection{Counterfactual Learning to Rank}
Unlike traditional LTR where rankers are learned from explicitly labelled datasets~\cite{liu2011learning}, counterfactual LTR~\cite{agarwal2018counterfactual,ai2018unbiased,joachims2017unbiased} uses historical interaction data, typically click logs, to learn a ranker. However, clicks are a biased signal. The most prominent bias in the click signal is the \textit{position bias}: assuming that users examine search engine result pages (SERPs) from top to bottom, then the results that are ranked higher are more likely to be observed by the users~\cite{joachims2017unbiased}. Joachims et al. refer to the probability of a search result at a rank $i$ to be observed as its \textit{propensity}~\cite{joachims2017unbiased}. They then define the \textit{inverse propensity scoring} (IPS) method to re-weight user clicks: when IPS is used, the estimated ranking score is unbiased with respect to position bias. One crucial requirement of IPS-based CLTR methods is the prior knowledge of user propensity. This is usually estimated by conducting online result randomization which can negatively affect the user experience. To address this issue, Ai et al.~\cite{ai2018unbiased} proposed the Dual
Learning Algorithm (DLA) to jointly learn an unbiased ranker
and an unbiased propensity model, thus avoiding the preprocessing of propensity estimations. Unlike these CLTR works, our method also uses IPS but in the context of reinforcement learning for OLTR and we further propose a new IPS method for unclicked documents so as to gain an unbiased training signal from unclicked documents. This accelerates the learning process because both click and unclick information is used for training, thus increasing the number of implicit signals obtained from each search and used for training.
In addition to position bias, recent work in counterfactual LTR has considered correcting the \textit{selection bias}~\cite{ovaisi2020correcting,oosterhuis2020policy}, in which some documents have zero probability of being observed by the users (thus never having a chance to be clicked and contribute a training signal). We note that the selection bias is fundamentally different from the position bias in terms of that this bias is introduced by the system itself: some documents will never be included by the system in the top SERPs. In contrast, the position bias instead comes from the users. This often occurs in web search where SERPs only show a small subset of documents in the first page (e.g., 10) and users do not proceed beyond the first SERP. In this circumstance, documents ranked beyond the first SERP have no chance of being observed, hence never get identified as positive training examples. To debias the selection bias in the click signals, Oosterhuis and Rijke~\cite{oosterhuis2020policy} proposed a policy-aware counterfactual estimator for CLTR to directly account for the selection bias introduced by a stochastic logging policy. On the other hand, Ovaisi et al.~\cite{ovaisi2020correcting} adapted Heckman’s two-stage method
to account for selection and position bias in LTR systems. In our work, we follow the standard OLTR and CLTR experimental setup and simulate typical, real-world circumstances; we also include selection bias in our experiments by only placing 10 documents in the SERPs. Although in this paper we directly focus on correcting position bias, selection bias is also partially corrected by the proposed method via online intervention. Nevertheless, recently proposed methods~\cite{ovaisi2020correcting,oosterhuis2020policy} could be used to further modify our reward shaping functions: this would have the effect of further reducing the selection bias of the gradient estimations.
\subsection{Online Learning to Rank}
Similarly to CLTR, online LTR (OLTR) also considers implicit user feedback to learn a ranker. Unlike CLTR, however, this is done online, by directly interacting with users. This online training process allows to control data acquisition and handle biases and noise through online interventions with regards to which documents to display. Figure~\ref{OLTR-process} provides a schematic representation of the OLTR process.
\begin{figure}[t]
\includegraphics[width=\linewidth]{OLTR-process.png}
\caption{A schematic view of the online learning to rank process. Users pose queries to the ranking system. This uses a online LTR algorithm to identify the candidate documents to display to the user. Depending on the OLTR algorithm used, candidate documents will be used to form a displayed list. The user examines the displayed search results (SERP) and clicks on items of interest. The click feedback is used by the OLTR algorithm to perform ranker updates.}
\label{OLTR-process}
\end{figure}
The \textit{Dueling Bandit Gradient Descent} (DBGD)~\cite{yue2009interactively} models OLTR as a dueling bandit problem and uses \textit{online evaluation} with users' clicks on interleaved or multileaved SERPs to indicate user preference among a pool of candidate rankers~\cite{schuth2015probabilistic,oosterhuis2016probabilistic,hofmann2011balancing,hofmann2013reusing,schuth2016multileave,hofmann2011probabilistic}. More recent work has studied methods to reduce the variance of the gradients estimated by online evaluation~\cite{wang2019variance,wang2018efficient}.
However, because gradients are updated towards the winning rankers in the candidate pool, the quality of the gradient estimation is influenced by the number of candidate rankers. When the number of candidate rankers is large, the performance of DBGD is limited in terms of both effectiveness and efficiency~\cite{schuth2015probabilistic,li2020mergedts}. The recently proposed \textit{Counterfactual Online Learning to Rank} (COLTR)~\cite{zhuang2020counterfactual} attempts to overcome the issues associated with online evaluation. This is done by replacing online evaluation with counterfactual evaluation for DBGD. Instead of interleaving, COLTR uses clicks collected by the current ranker to evaluate candidate rankers thus providing high efficiency. However, empirical results have shown COLTR requires more exploration, thus possibly hurting user experience (online evaluation metrics).
Unlike DBGD and COLTR, the \textit{Pairwise Differentiable Gradient Descent} (PDGD) \cite{oosterhuis2018differentiable} does not require to sample candidate rankers for online evaluation. Instead, PDGD directly estimates gradients based on pairwise preferences between documents in the SERP, inferred by users' clicks; then, stochastic gradient descent is used to update the ranker. The gradient estimation of PDGD is unbiased with respect to user document pair preferences~\cite{oosterhuis2018differentiable}. PDGD is empirically found to be significantly better than DBGD in terms of final convergence, learning speed and user experience during optimization, making PDGD the current state-of-the-art method for OLTR~\cite{jagerman2019model,zhuang2020counterfactual,oosterhuis2019optimizing,wang2021federated}. PDGD has also been adapted to the federated OLTR context~\cite{wang2021effective}, exhibiting again state-of-the-art performance. However, both PDGD and DBGD based methods do not directly use IPS to handle position bias, which has been proven to be very important in previous CLTR results~\cite{joachims2017unbiased}.
As mentioned above, CLTR and OLTR differ from the setting considered for learning the ranker: use an historical log of implicit interactions for counterfactual, versus use direct observation and interaction with rankings in the case of online LTR. We note that Jagerman et al.~\cite{jagerman2019model} have specifically studied the similarity and differences between counterfactual and online LTR. They suggested that, from a theoretical standpoint, counterfactual LTR exploits position bias better, however they also indicated that empirical results have shown that OLTR (and in particular PDGD) is more reliable. On the other hand, recent works have been focusing on adapting the offline counterfactual learning to the online setting~\cite{ai2021unbiased,zhuang2020counterfactual,oosterhuis2020taking,oosterhuis2021unifying}. These works have suggested that OLTR algorithms can benefit from the counterfactual learning framework. Based on this context, our proposed method can also be thought of as embracing this direction.
\subsection{Reinforcement Learning to Rank}
Reinforcement learning (RL) has previously been applied to offline LTR problems, but not to online LTR. Wei et al.~\cite{wei2017reinforcement} have formalized ranking as a Markov Decision Process (MDP) problem and introduced the MDPRank algorithm to optimize a linear ranker. Specifically, the document ranking is modelled as a sequential decision making process where each time step corresponds to a position in the ranking and the action taken at each time step corresponds to the selection of a document for the corresponding position. The rewards given by the environment are generated according to the relevance label of the documents, and the classic policy gradient algorithm of REINFORCE~\cite{sutton2000policy} is used to maximize the expectation of cumulative rewards received by the ranker. An attractive property of MDPRank is that it can use gradient descent to directly optimize non-differentiable ranking metrics such as DCG. However, Xu et al.~\cite{jun2020ppg} have recently shown that the gradient estimation of the original MDPRank exhibits high variance, and thus this method requires more training episodes to learn an effective ranker. A similar reinforcement learning framework has also been used for search result diversification~\cite{feng2018greedy,xia2017adapting}, multi-page search~\cite{zeng2018multi} and recommendation system~\cite{zhao2018deep,zhao2018recommendations,zhao2021dear}.
In this paper, we also apply a MDPRank-like algorithm, although in the context of OLTR and with two important adaptions. First, the original MDPRank is designed for traditional offline LTR, i.e., the relevance labels of query-document pairs are provided. In the OLTR setting, however, only biased user interaction data such as clicks is available: thus the rewards given by the online environment are biased. In order to obtain unbiased rewards, we introduce unbiased reward shaping functions for MDPRank to discount position bias. Secondly, the high variance of the gradient estimation of MDPRank makes it converge slowly. However, for OLTR to be viable, we need the algorithm to converge fast so that it does not hurt the user experience too much. To reduce variance, we simplify the objective function of the policy gradient used in the original MDPRank to achieve an objective function with much lower variance of gradient estimation, without changing its optimization target.
\section{Introduction} \label{intro}
Learning to rank (LTR) is a supervised machine learning technique that has been widely used in modern search engines to learn rankers. \textit{Explicit feedback}, consisting of assessors manually judging the relevance of query-document pairs, is required for LTR~\cite{liu2011learning,li2011learning}.
This labelled dataset requirement poses obvious limitations: these datasets take substantial effort and cost to compile~\cite{qin2010letor,qin2013introducing,chapelle2011yahoo}, labelling personal documents is unethical and often impossible~\cite{wang2016learning}, static datasets cannot model user intent change over time~\cite{lefortier2014online,zhuang2021how}, and user preferences may not agree with that of annotators~\cite{sanderson2010test}.
In order to overcome these limitations, \textit{implicit feedback}, such as clicks, has been leveraged.
This type of training signal is not affected by the above limitations and has been an attractive alternative to annotated datasets~\cite{joachims2002optimizing}. However, training rankers with implicit feedback has its own drawbacks and challenges. For example, clicks are a weak relevance signal because they often are affected by a number of biases and noise. One of the most prominent bias in web search is the \textit{position bias}, where higher ranked documents have a higher chance to be observed and thus gain more clicks, even if they may be not relevant~\cite{guan2007eye,pan2007google,hofmann2016online,joachims2017unbiased}. Therefore, it is important to consider the influence of such biases.
Two main families of approaches have emerged that attempt to learn effective rankers from users' implicit feedback~\cite{jagerman2019model}: \textit{counterfactual learning to rank} (CLTR) \cite{joachims2017unbiased} and \textit{online learning to rank} (OLTR) \cite{yue2009interactively}.
In CLTR, given a historical click through log, clicks are treated as pure binary relevance labels and \textit{inverse propensity scoring} (IPS) is used to re-weight clicks in order to discount the effect of biases. Rankers are trained in an offline manner and deployed online after training. This offline batch updates pipeline can avoid the risk of exposing users to low-quality results since it only displays the best search engine results that are possible for a given CLTR algorithm and training data.
On the other hand, OLTR algorithms interactively update rankers after each user interaction has taken place, thus being more responsive to a non-stationary user environment~\cite{zhuang2021how}. In contrast to CLTR, current OLTR methods do not directly model position or selection bias, and only assume relevant documents are more likely to be clicked than non-relevant documents~\cite{joachims2002optimizing}. The biases and noise of users clicks are handled by \textit{online interventions}~\cite{jagerman2019model}, i.e., slightly perturbated result lists are displayed and preferences towards rankers are informed by users clicks. This is one of the key aspects of OLTR, but also one of its biggest disadvantages that limits OLTR's uptake in practice: such online interventions carry the risk of displaying a ``sub-optimal'' ranking list directly to the user, thus hurting user experience (as measured by online evaluation metrics). Hence, it is a requirement for OLTR methods to efficiently leverage the click feedback so that a good ranker is learnt as fast as possible to avoid displaying low quality search results to a large number of users.
To move beyond the limitations of existing OLTR methods, in this article, we propose a novel OLTR algorithm called \textit{Reinforcement Online Learning to Rank} (ROLTR), which exploits the reinforcement learning (RL) approach adapting it to OLTR. Our motivation for using RL in OLTR is based on the following observations:
\begin{enumerate}
\item RL suits OLTR setting very well: RL is powerful for modeling interactive environments and maximizing the long-term rewards yield from the environment. In the OLTR setting, the interactive environment is composed by the users and the search system, and the rewards are the users’ satisfaction.
\item RL with carefully designed reward functions allows OLTR algorithms to directly remove the biases present in the user clicks: this is currently hard for other OLTR algorithms to achieve.
\end{enumerate}
As we show in this article, our proposed ROLTR can directly remove position bias and thus effectively and, importantly, efficiently (i.e. within less impressions) update the ranker. To achieve this, we formalize OLTR as a \textit{Markov Decision Process} (MDP) problem and use \textit{policy gradient} with rewards assigned on clicked and unclicked documents to estimate the update gradients. In order to de-bias users' clicks, we further introduce unbiased reward shaping functions that re-weight the rewards for both clicked and unclicked documents. We mathematically prove that the gradient estimation of ROLTR is unbiased with respect to position bias, and it can directly optimize IR metrics such as discounted cumulative gain (DCG). The idea of leveraging unclicked data has been recently explored in offline counterfactual learning studies~\cite{hu2019unbiased,wang2021non}, however, our method is currently the only OLTR approach that can gain unbiased learning signals from unclicked documents, thus speeding up convergence. Empirical results further show that ROLTR significantly outperforms traditional OLTR methods and is at par with current state-of-the-art methods (offline performance), although requiring fewer user interactions. As a result, our method delivers considerably better user experience (online performance).
\section{Empirical Evaluation}
To study the effectiveness of ROLTR, we designed a number of empirical experiments aimed to answer the following research questions:
\begin{itemize}[leftmargin=40pt]
\item[\bf RQ1:] How does the reward discount factor in ROLTR affect gradient variance and final convergence (i.e., the offline nDCG score on the test dataset)?
\item[\bf RQ2:] How do the unbiased reward shaping functions of ROLTR impact performance?
\item[\bf RQ3:] How does ROLTR compare in terms of convergence and learning speed against current representative OLTR methods?
\item[\bf RQ4:] Does ROLTR deliver better user experience than current OLTR methods, i.e., higher online nDCG?
\item[\bf RQ5:] How sensitive is ROLTR to propensity mismatch?
\end{itemize}
\subsection{Datasets and synthetic data generation}
We consider three benchmark datasets that are commonly used to evaluate OLTR~\cite{zhuang2020counterfactual, schuth2015probabilistic,oosterhuis2016probabilistic,oosterhuis2018differentiable,wang2019variance} and CLTR~\cite{jagerman2019model,ai2018unbiased,joachims2017unbiased}: MSLR-WEB10K~\cite{qin2013introducing}, Yahoo! Webscope~\cite{chapelle2011yahoo}, and Istella~\cite{dato2016fast}. MSLR-WEB10K contains 10,000 queries and 125 retrieved documents on average; documents are represented by 136 features. Yahoo! is a bigger dataset, with 29,921 queries and an average of 23.7 documents per query, represented using 700 features. Istella is the largest dataset we consider, with 33,118 queries and an average of 315 documents per query, represented by 220 features. Query-document pair relevance labels for all datasets are recorded on a five-point scale from not relevant (0) to perfectly relevant (4) and have been split into training, validation and test sets (according to the standard splits in the datasets). Queries in the three sets are disjoint. We use the training set to train the rankers, the validation set to tune the hyper-parameters, and the test set to evaluate the rankers' performance.
To avoid hurting user experience, it is common for research in OLTR and CLTR to simulate users' clicks by relying on the relevance labels recorded in the datasets~\cite{zhuang2020counterfactual, schuth2015probabilistic,oosterhuis2016probabilistic,oosterhuis2018differentiable,jagerman2019model,joachims2017unbiased,vardasbi2020cascade}. This also allows to fully control users' biases and noise so that algorithms can be tested under different, controllable conditions. Queries are uniformly sampled from the dataset (sampling query IDs). The candidate document set associated with the query ID is provided by each dataset. Then OLTR algorithms generate a result list of documents to display. Clicks are simulated based on two fixed variables: the click probability and the position bias.
The click probability is the probability of a user clicking on a document after observing it. This probability is conditioned on the document's relevance label. Following previous OLTR work~\cite{oosterhuis2018differentiable,zhuang2020counterfactual,hofmann2011balancing,hofmann2013reusing,oosterhuis2016probabilistic}, we set two types of click behaviour: \textit{perfect} and \textit{noisy}. The click probability of the \textit{perfect} click behaviour is proportional to the relevance level of the documents, and has $0$ probability for non-relevant documents. This simulates an ideal user that is able to always determine the relevance of a document in the SERP. The \textit{noisy} click behaviour mimics instead a realist behaviour on SERPs by assigning a small click probability to non-relevant documents and a small skip probability to relevant documents. Table~\ref{click_model} provides the click probabilities for the two user models.
Position bias is modelled by the document observation probabilities; we assume the observation probabilities only depend on the rank position of the document and set these probabilities to:
\begin{equation}
P(o_{k}=1 | k) = \left( \frac{1}{k} \right)^\eta
\end{equation}
where $k$ is the rank position and $\eta$ is a parameter that determines the level of position bias. Following Joachims et al.~\cite{joachims2017unbiased} and Jagerman et al.~\cite{jagerman2019model}, we set $\eta=1$.
Thus, the probability of a click occurring on a document at rank $k$ in the result list is:
\begin{equation}
P(c_k=1) = P(c_k=1|o_k=1, rel(d_k))P(o_k=1 | k)
\end{equation}
\newcolumntype{b}{X}
\newcolumntype{s}{>{\hsize=.5\hsize}X}
\newcommand{\tc}[1]{\multicolumn{1}{c}{#1}}
\setlength{\tabcolsep}{3mm}
\begin{table}
\centering
\caption[centre]{Click probabilities for different user behaviours.}\label{click_model}
\begin{tabularx}{\linewidth}{bsssss}
\toprule
& & \multicolumn{4}{c}{$P(c=1|o=1, rel(d))$} \\
\midrule
$rel(d)$ & 0& 1 & 2 & 3 & 4\\
\midrule
\textit{perfect} & 0.0 & 0.2 & 0.4 & 0.8 & 1.0\\
\textit{noisy} & 0.4 & 0.6 & 0.7 & 0.8 & 0.9\\
\bottomrule
\end{tabularx}
\end{table}
\subsection{Evaluation measures}
We measure effectiveness using standard OLTR evaluation practice, which considers two aspects: \textit{offline} and \textit{online} performance.
\textit{Offline} performance is the final convergence of the learned ranker. We evaluate this using the average nDCG@10 of the ranker over the queries in the held-out test-set across $100,000$ impressions, as all rankers would have reached convergence at this point.
\textit{Online} performance measures user experience during training. This is quantified by the nDCG@10 obtained by the rank list $L_i$ that the user observes in the training episode $i$, times a discount rate $\tau < 1$:
\begin{equation}
online\_performance = \sum_{i}\tau^{i}\cdot NDCG(L_i)
\end{equation}
The discount rate assigns less weight to the later impressions to reward OLTR algorithms that learn an effective ranker fast, so to limit the amount of low-quality user experience. As in previous work~\cite{oosterhuis2018differentiable,wang2019variance,zhuang2020counterfactual,hofmann2013reusing}, we set $\tau=0.9995$: this means that impressions that occur after $10,000$ iterations have less than a $1\%$ impact. We further note that, while this online performance measure aims to quantify the user experience during training, it only does so partially: the relevance of the results in SERP is in fact only one of the many aspects influencing user experience~\cite{al2010review,maxwell2017study}. These other aspects however are not measurable in the context of the typical simulated experiments performed in OLTR research.
\subsection{Experimental runs}
We compare ROLTR to four OLTR baselines. First, the \textit{Dual Bandit Gradient Descent (DBGD)}~\cite{yue2009interactively} method is used as it is one of the standard and most influential algorithms for OLTR. This method uses interleaving for online evaluation, where only one candidate rank is compared to the production ranker at each update step.
The second baseline we consider is the \textit{Probabilistic Multileaving Gradient Descent (PMGD)}~\cite{oosterhuis2016probabilistic}, which has been reported to be the best traditional OLTR method that uses online evaluation with multileaving comparison~\cite{oosterhuis2018differentiable,zhuang2020counterfactual}. For this baseline, we use the same hyper-parameters settings reported in previous work~\cite{zhuang2020counterfactual}, where the number of candidate rankers is $n = 49$, the step size is $\delta =1$ and learning rate is $\alpha = 0.01$.
The third baseline we consider is \textit{Counterfactual Online Learning to Rank (COLTR)}~\cite{zhuang2020counterfactual}, which uses DBGD to update candidate rankers but is combined with counterfactual evaluation. This method has been reported to be more effective than online evaluation based methods in terms of final convergence, but it does present an overall deterioration of user experience during the learning cycle. For this baseline, we use the best hyper-parameters from the original paper~\cite{zhuang2020counterfactual}.
The last baseline we consider is the \textit{Pairwise Differentiable Gradient Descent (PDGD)}~\cite{oosterhuis2018differentiable}, which represents the current state-of-the-art in OLTR. This method does not require online evaluation; instead, it directly optimizes a ranker using gradient descent, and its gradient estimations are unbiased with respect to document pair preferences inferred from user clicks. The comparison between ROLTR and PDGD is interesting because both methods use gradient descent but they differ in the way the unbiased gradient estimation is computed. Following the original PDGD paper~\cite{oosterhuis2018differentiable}, we set its learning rate to $\alpha = 0.1$.
ROLTR has two hyper-parameters: the reward function and the learning rate $\alpha$. We study six reward functions in total: $R_{IPS^+}$, $R_{IPS^-}$, $R_{IPS^+} + R_{IPS^-}$ (i.e. using both clicked and unclicked signal) and their corresponding unshaped naive reward functions. We set the learning rate $\alpha=0.01$ for MSLR-WEB10K and Istella, and $\alpha=0.005$ for Yahoo! (values tuned based on validation set).
To simulate selection bias, we set $M=10$ to only display 10 documents in the result lists for all experimental runs. For fair comparison, all methods are used to optimize a linear ranker. In order to measure statistically significant differences between methods, all runs are repeated 15 times spread evenly over the available dataset folds with different random seeds. The results are reported and compared using averaged performance and the two-tailed t-test.
In addition to comparing ROLTR with the other OLTR baselines mentioned above, we also compare our method against the MDPRank trained using full information, i.e., the offline LTR settings with actual relevance labels. This is to be thought of as the skyline for ROLTR as this method is built on MDPRank but training in ROLTR occurs with partial, noisy information (user clicks). For MDPRank, we use the experiment settings from the original paper~\cite{wei2017reinforcement}. |
\section{Introduction}
\IEEEPARstart{I}{n} the past few years, distributed coordination of multiagent systems (MASs) has attracted much attention due to its broad application prospects in civil, military and other fields \cite{vicsek1995novel,moreau2005stability,fax2004information}, such as distributed sensor networks \cite{olfati2002distributed}, formation control \cite{su2009flocking} and distributed optimization \cite{olfati2007distributed}. One of its fundamental problems is consensus, which requires that agents achieve agreement about certain quantities of interest that depends on all agents’ states. Many scientific problems of consensus have emerged, and lots of control protocols are proposed in this area, such as consensus tracking \cite{hong2006tracking}, average consensus \cite{olfati2004consensus} and robust consensus \cite{shi2013robust,wang2008robust}.
In most of the above consensus problems, agents' states are not constrained. However, there are various state constraints on agents in many real-world scenarios, such as restricted actuators and limited communication distance. Imposing state constraints on agents has notable significance and great research value. The constrained consensus problems have been studied from different perspectives. For example, the constrained consensus and optimization problems have been studied, where agents' states are constrained in closed convex sets \cite{nedic2010constrained}. Under positivity constraints, the necessary and sufficient conditions for MASs to achieve consensus are obtained in \cite{valcher2017consensus}. A novel state-constrained consensus (named interval consensus) problem has been proposed and studied in \cite{fontan2019interval}. Moreover, alternative approaches for imposing state constraints have been proposed in \cite{meng2016consensus,sun2013consensus}.
Information sharing is a necessary condition for MASs to admit a consensus solution. In the process of information transmitting, there are various unfavorable factors such as environmental interference, noise, and attenuation. Therefore, it is impractical and ideal to assume that agents can receive neighbors' information without distortions and noise disturbances. Previous researches on the imperfect information transmission often focused on switching topology, communication delay, communication link fault, packet loss, etc. For example, average consensus problem is studied in \cite{wu2012average}, where time-varying delay and packet loss are considered under the undirected communication network. A synchronization protocol is proposed to solve the communication link faults and improve MAS's synchronization resilience \cite{chen2020adaptive}.
In the above constrained consensus problems, the constraints are imposed on agents' states or inputs, directly.
In some multiagent systems, the inputs of agents are the combination of neighbours' states. Hence, on one hand, the constraints imposed on agents' states may impact the sending information to theirs neighbours. On the other hand, the constraints imposed on agents' inputs can be caused by changing the received neighbours' states. Therefore, the above analysis inspires us that the constrained consensus can be achieved by changing the transmitted information between agents.
To distinguish our problem from other constrained consensus problems, we call it \textbf{transmission-constrained consensus}.
In this problem, the deformed transmitted information is depicted by heterogeneous functions (named transmission constraint function). And a variety of functions can be chosen as transmission constraint function,
hence this study is so universal that it can be applied in many cases. Some applications and motivating examples are given in subsection \ref{subsc:applications}.
The first part of this work can be regarded as a study on the consensus conditions for the MAS with distorted transmitted state information. The necessary and sufficient conditions for the ranges of transmission constraint functions are obtained.
A non-empty intersection of constraints is an important condition for MASs to achieve consensus. In many pieces of literature that study the consensus problem for MASs with state constraints, it is assumed that the constraints have a non-empty intersection. However, in our model, we do not need the above assumption. Therefore, the multiagent systems may not achieve consensus, but an equilibrium. In the second part of this study, it is proved that when the transmission constraint functions are distributed in a specific range, the system will reach an asymptotically stable equilibrium even though that intersection of constraints is empty.
Compared with the existing works about the consensus of networked systems with constraints, the contributions of this work can be obtained as follows.
\begin{enumerate}
\item This work is the first time to study the transmission-constrained consensus of multiagent networks with heterogeneous constraint functions. The transmission-constrained consensus model studied in this paper does not have a definite form, so that it can be regarded as a paradigm. The multiagent system that can be translated to this model are able to make the system achieve consensus under the necessary conditions, such as interval consensus \cite{fontan2019interval}.
\item Unlike traditional constrained consensus problems, transmission-constrained consensus problem has the following features:
\begin{enumerate}
\item each link in the interaction network is limited by an individual constraint function, which is more general in reality;
\item constraint functions do not have a uniform type, and they can be various functions, such as trigonometric function, saturation function and Sigmoid function, etc.
\end{enumerate}
Those features make the transmission-constrained model have a wide range of application, but also bring heterogeneity into the dynamics, which increases the difficulty of analysis.
\item For the transmission-constrained consensus problem, we obtain some consensus conditions, in which a necessary and sufficient condition is related to the range of constraint functions. As a more general case than the consensus case, equilibrium of MAS is seldom studied. We investigate this phenomenon and obtain conditions of equilibrium’s existence, uniqueness and stability.
\item Due to the novel model, where the unknown transmission constraints make the dynamics nonlinear, the analysis of MAS's stability is quite a challenge.
We design some linear boundaries and propose the corresponding lemmas to analyze the monotonicity of dynamics. Then, we construct a novel Lyapunov function to analyze the convergence of MASs. Applying theorems in algebraic topology, we prove the existence of equilibrium. By coordinate transformations, another Lyapunov function is constructed to study the stability and uniqueness of equilibrium.
\end{enumerate}
The paper is organized as follows. Preliminaries and problem statement are given in Section II. Main results are provided in Section III. Supports of numerical examples are provided in Section IV, and Section V concludes this work. Finally, we put all proofs in appendixes.
\section{PRELIMINARIES}
\subsection{Graph Theory}
This paper studies the problem of transmission-constrained consensus of multiagent networks. Consider a MAS with $n$ agents, and denote $\mathbf{N}=\{1,2,\dots,n\}$. The finite vertex set is denoted by $\mathcal{V}=\{v_1,\dots,v_n\}$, and $\mathcal{E}\subseteq \mathcal{V}\times\mathcal{V}$ denotes edge set where $(v_j,v_i)\in \mathcal{E}$ means that there exists a communication link from agent $j$ to agent $i$. The adjacency weight matrix $\mathcal{A}\in\mathbb{R}^{n\times n}$ is defined as $a_{ij}>0$ if and only if $(v_j,v_i)\in \mathcal{E}$, and $a_{ij}=0$, otherwise. Then, the underlying interaction network of MAS is described by a (weighted) graph $\mathcal{G}=\{\mathcal{V},\mathcal{E},\mathcal{A}\}$ which is a triple. $(v_j,v_i)$ is defined as the directed edge from agent $j$ to agent $i$, and $\mathcal{N}_i=\{v_j\in \mathcal{V}:(v_j,v_i)\in\mathcal{E}\}$ denotes the neighbor set of agent $i$.
Denote $\alpha_i=\sum_{j=1}^na_{ij}$ as the row sum of $\mathcal{A}$, and $\bar{a}=\max \alpha_i$.
\subsection{Problem Statement}
For any $i\in\mathbf{N}$, denote the state of $v_i$ by $x_i(t)\in\mathbb{R}$. Then consider the continuous-time dynamics of single-integrator MAS with $n$ agents:
\begin{equation*}
\dot{x}_i(t)=u_i(t),\quad i\in\mathbf{N},
\end{equation*}
where $u_i(t)\in\mathbb{R}$ is the control input.
The problem studied in this work is different from the general MAS dynamics. In this problem, the information transmissions between agents are disturbed by interference functions (or attenuation functions), i.e., the transmission of state $x_i(t)$ is replaced by a transmission constrain function $f_{ij}\big(x_i(t)\big)$.
Assume the transmission constrain function $f(x)$ is piecewise continuous, which is a standing assumption throughout the rest of this paper.
Then, the transmission-constrained consensus algorithm of $x_i(t)$ is
\begin{equation}
\dot{x}_i(t)=u_i(t)=\sum_{j\in \mathcal{N}_i}a_{ij}\Big[f_{ji}\big(x_j(t)\big)-x_i(t)\Big].
\label{MAS_1}
\end{equation}
\begin{remark}
Since $f(x)$ is piecewise continuous, the autonomous system \eqref{MAS_1} may have multiple solutions. However, the following theorems and corollaries apply to both unique and multiple solutions.
\end{remark}
This paper aims to find which transmission constraints could make MAS stable and obtain the consensus conditions for MAS \eqref{MAS_1}.
\subsection{Applications and Motivating Examples}\label{subsc:applications}
The distortion (attenuation or saturation) in information transmission or detection is an actual embodiment of transmission constraints. Those transmission constraints may be caused by objective physical constraints, or those constraints are added on purpose.
\subsubsection{Objective Constraints}
There are three kinds of objective constraints to show those transmission constraints are common in real world scenarios.
\begin{itemize}
\item \emph{Information distortion caused by transmission.}
Energy loss exists during the signal transmission, which may cause information distortion. The voltage drop on wires is an inevitable
phenomenon during signal transmission, which causes signal attenuation. If state of agent (or device) are represented by voltage of a signal, and this signal is transmitted on wires, then we can get an information distortion
\begin{equation*}
f_{ij}(x_i)=\frac{R_r}{R_L+R_r}x_i,
\end{equation*}
where $R_L$ is the resistance of wires, and $R_r$ is the equivalent resistance of the port.
\item \emph{Information distortion caused by detection.}
In real world scenarios, agents use sensors to get themselves or neighbors’ states. However, except for noise interference, state information cannot be obtained precisely due to the inherent characteristics of sensors. For example, temperature offset leads to signal fluctuation in ultrasonic distance measurements \cite{carullo2001ultrasonic}.
Likewise, the saturation characteristic of hall sensor may cause information distortion, i.e., $f_{ij}(x_i)=\text{sat}(x_i)$ \cite{ramsden2011hall}.
\item \emph{Information distortion caused by privacy protection.}
In social networks, individuals may express an opinion that is different from his/her private opinion, probably due to the pressure of conforming to a group standard or norm \cite{YE2019371,HOU2021125968}. Hence, $x_i$ could represent the private opinion, and $f_{ij}(x_i)$ is the expressed opinion.
\end{itemize}
\subsubsection{Subjective Constraints}
As a class of constraints, transmission constraints could make agent’s states converge into the expected set (see Theorem \ref{th:consensus} and Remark \ref{re:constrained state}). Especially, Since we do not specify the formula of transmission constraints, different transmission constraints can be designed to suit different scenarios, such as interval consensus \cite{fontan2019interval}, quantized consensus \cite{kashyap2007quantized,FRASCA2012273} and discarded consensus \cite{liu2012discarded}, etc. Related discussions are in remarks \ref{re:interval consensus} and \ref{re:discarded consensus}.
Those above examples show that information distortion during transmission is a common phenomenon in the real world. Hence, study consensus under transmission constraints is necessary.
\subsection{Notations and Some Definitions}
\textbf{Notations:}
The set of positive integers is denoted by $\mathbb{N}^+$.
Consider a matrix $B=[b_{ij}]\in M_{m,n}$ and denote $|B|=[|b_{ij}|]$ (i.e., element-wise absolute value of matrix $B$).
Consider a set $N=\{1,2,\dots,n\}$, then denote the cardinality of set $N$ by $\textnormal{Card}(N)=n$.
$d^+Z(t)$ denotes the upper right Dini derivative of $Z(t)$.
The arrow `$\implies$' means `implies', and the arrow `$\iff$' means `if and only if'.
Denote sign function
\begin{equation*}
\text{sign}(x)=\begin{cases}
1, &\text{if } x>0,\\
0, &\text{if } x=0,\\
-1, &\text{if } x<0.
\end{cases}
\end{equation*}
The distance between set $[ \partial_m, \partial_M]^n$ and $x(t)$ is denoted by
\begin{align*}
\textit{distance}\big([ \partial_m, \partial_M]^n,x(t)\big)&=\|x(t)\|_{[ \partial_m, \partial_M]^n}\\
&=\min\limits_{c\in[ \partial_m, \partial_M]^n}\|x(t)-c\|.
\end{align*}
Denote $\mathbf{e}=\{\mathbf{e}_1,\dots,\mathbf{e}_n\}^T$ to be an equilibrium of MAS \eqref{MAS_1}, then it can be concluded that for all $i\in\mathbf{N}$,
\begin{align*}
&\dot{x}_i(t)\big|_{\mathbf{e}_i}=\sum\limits_{j\in\mathcal{N}_i}a_{ij}\big(f_{ji}(\mathbf{e}_j)-\mathbf{e}_i\big)=0.
\end{align*}
Denote the error between the state $x(t)$ and equilibrium $\mathbf{e}$ by $\varepsilon_i(t)=x_i(t)-\mathbf{e}_i$, $\forall\, i\in\mathbf{N}$.
Denote $\Theta_{ij}=\{x:f_{ij}(x)=x\}$. Then, introduce the definition of consensus zone, and this work can be divided into two parts: the part of non-empty consensus zone and the part of empty consensus zone.
\begin{definition}
\textbf{consensus zone} $\Phi=\bigcap\limits_{(v_j,v_i)\in \mathcal{E}}\Theta_{ij}$.
\end{definition}
\begin{remark}
For any time, if an agent's state is in consensus zone, then the information it transmits to its neighbors is without transmission constraints. If all agents' initial states are in consensus zone, then the MAS becomes a standard consensus dynamics. Moreover, under a strongly connected digraph (or the digraph has a spanning tree), the MAS will reach consensus and the consensus value is in the consensus zone. That is why we name it consensus zone.
\end{remark}
\section{Main Results}
In this section, the transmission-constrained consensus problems are studied.
Initially, we analyze the convergence of MAS in subsection \ref{Part:CT_A}.
Secondly, for the non-empty consensus zone, we get the consensus conditions in subsection \ref{Part:CT_B}.
Thirdly, for the empty consensus zone, the system's states may achieve an equilibrium, and the existence, stability and uniqueness of equilibrium are studied in subsection \ref{Part:CT_C}.
\subsection{Convergence analysis}\label{Part:CT_A}
The following theorem states the conditions where the states of the multiagent system are bounded, and gives the boundary.
Furthermore, the conclusion of Theorem \ref{th:distance=0} plays an important role in the proofs of Theorems \ref{th:consensus} and \ref{th:exist_equilibrium}.
\begin{Theorem}\label{th:distance=0}
Along the system \eqref{MAS_1}, suppose following conditions hold:
\begin{enumerate}
\item the directed graph $\mathcal{G}$ is strongly connected;
\item\label{it:th_distance=0_2} there exists a set $[\partial_m,\partial_M]$, a value $\partial\in[\partial_m,\partial_M]$ and two rays
\begin{align*}
L_{1}(x)&=k_{1}(x-\partial)+\partial,\quad x\in(-\infty,\partial];\\
L_{2}(x)&=k_{2}(x-\partial)+\partial,\quad x\in[\partial,+\infty),
\end{align*}
where $k_{1},k_{2}<0$ and $k_{1}k_{2}=1$,
such that $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\begin{alignedat}{2}
&x\le f_{ij}(x)< L_{1}(x),\quad & &x\in(-\infty,\partial_m);\\
&L_{2}(x)< f_{ij}(x)\le x, & &x\in(\partial_M,+\infty).
\end{alignedat}
\end{equation*}
\item\label{it:th_distance=0_3} for any $x'\in(-\infty,m)\cup(M,+\infty)$, there eixst $j\in \mathbf{N}$ and $i\in\mathcal{N}_j$, such that $f_{ij}(x)$ is continuous on $x'$ and $f_{ij}(x')\ne x'$.
\end{enumerate}
Then for any initial state $x_*\in\mathbb{R}^n$, \textbf{if and only if} $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\partial_m\le f_{ij}(x)\le\partial_M,\enspace x\in[\partial_m,\partial_M],
\end{equation*}
MAS \eqref{MAS_1} satisfies that
\begin{equation*}
\lim\limits_{t\to\infty}\textnormal{distance}\big([ \partial_m, \partial_M]^n,x(t)\big)=0.
\end{equation*}
\end{Theorem}
\subsection{Nonempty Consensus Zone: Consensus}\label{Part:CT_B}
Then, we introduce conditions of transmission-constrained consensus.
The following theorem states the consensus conditions for MAS, and it is worth noting that the condition about range of constraint functions is necessary and sufficient.
\begin{Theorem}\label{th:consensus}
Along the system \eqref{MAS_1}, suppose following conditions hold:
\begin{enumerate}
\item the directed graph $\mathcal{G}$ is strongly connected;
\item\label{item:th_consensus_C2} there exist a non-empty set $[\partial_m,\partial_M]$ and a constant value $\partial\in[\partial_m,\partial_M]$ and two rays
\begin{align*}
L_1(x)&=k_1x+(1-k_1)\partial,\quad x\in(-\infty,\partial];\\
L_2(x)&=k_2x+(1-k_2)\partial,\quad x\in[\partial,+\infty),
\end{align*}
such that for all $j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\begin{alignedat}{2}
&x\le f_{ij}(x)< L_{1}(x),\quad & &x\in(-\infty,\partial_m);\\
&f_{ij}(x)=x, & &x\in[\partial_m,\partial_M];\\
&L_{2}(x)< f_{ij}(x)\le x, & &x\in(\partial_M,+\infty).
\end{alignedat}
\end{equation*}
\item\label{item:th_consensus_C3} for any $x'\in(-\infty,m)\cup(M,+\infty)$, there eixst $j\in \mathbf{N}$ and $i\in\mathcal{N}_j$, such that $f_{ij}(x)$ is continuous on $x'$ and $f_{ij}(x')\ne x'$.
\end{enumerate}
Then, $\lim\limits_{t\to\infty}x_i(t)=v^*$, $v^*\in[\partial_m,\partial_M]$ \textbf{if and only if} $k_1k_2\le1$.
\end{Theorem}
\begin{remark}
In this dynamic model, convergence happens in a fully distributed manner, and without the need to know any information about any constraint functions. Conditions \ref{item:th_consensus_C2} and \ref{item:th_consensus_C3} in Theorem \ref{th:consensus} require $[\partial_m,\partial_M]^n$ to be the largest set in consensus zone and the consensus zone is connected.
\end{remark}
\begin{remark}\label{re:constrained state}
Theorem \ref{th:consensus} shows that although the initial state is not in the consensus zone, the transmission constraints will limit the final state of the system. The fact inspires us that limiting agents' states can be done indirectly by imposing constraints on links of the interaction network.
\end{remark}
\begin{remark}
Under the conditions in Theorem \ref{th:consensus}, the constraint functions can be various functions. For example, constraint functions can be Sigmoid function or tanh function, which have many applications such as activation function in artificial neural networks, logistic function in biology, etc. More candidate constraint functions are given in the numeral example section.
\end{remark}
\begin{remark}\label{re:interval consensus}
In \cite{fontan2019interval}, the interval consensus problem is studied.
We propose a smooth interval consensus model:
\begin{equation*}
\dot{x}_i(t)=\sum_{j\in \mathcal{N}_i}a_{ij}\big[T_j\big(x_j(t)\big)-x_i(t)\big],\enspace i\in\mathbf{N},
\end{equation*}
\begin{equation*}
T(x)=\begin{cases}
\rho x+(1-\rho)q, &\text{if } x>q,\\
x, &\text{if } p\le x\ge q,\\
\rho x+(1-\rho)p, &\text{if } x<p,
\end{cases}
\end{equation*}
where $\rho\in(0,1)$ is a constant. By Theorem \ref{th:consensus}, it can be concluded that the system will reach interval consensus.
\end{remark}
\begin{remark}\label{re:discarded consensus}
The discarded consensus problem is studied in \cite{liu2012discarded}, but the initial states of system must be in the constraint set. Another discarded consensus model can be proposed:
\begin{equation}
\dot{x}_i(t)=\sum_{j\in \mathcal{N}_i,x_j(t)\in\Omega_{c_i}}a_{ij}x_j(t)-\sum_{j\in \mathcal{N}_i}a_{ij}x_i(t).
\label{eq:discarded consensus}
\end{equation}
By Theorem \ref{th:consensus}, we can get that the MAS \eqref{eq:discarded consensus} will reach discarded consensus with arbitrarily initial states.
\end{remark}
\begin{remark}
Consider the following multiagent system:
\begin{equation}\label{MAS_sin}
\dot{x}_i(t)=\sum_{j\in \mathcal{N}_i}a_{ij}\big[\sin\big(x_j(t)+\pi\big)-x_i(t)\big],\enspace i\in\mathbf{N}.
\end{equation}
Due to the consensus zone of MAS \eqref{MAS_sin} $\Phi=\{0\}$, by Theorem \ref{th:consensus}, it shows that the agents' states will converge to $0$, when the underlying directed graph $\mathcal{G}$ is strongly connected.
\end{remark}
\begin{Corollary}
Along the system \eqref{MAS_1}, suppose following conditions hold:
\begin{enumerate}
\item the connected graph $\mathcal{G}$ is strongly connected;
\item the consensus zone $\Phi\ne\emptyset$;
\item for any $j\in\mathbf{N},\,i\in\mathcal{N}_j$ and $\omega\ne0$,
\begin{equation*}
-1<\frac{f_{ij}(x+\omega)-f_{ij}(x)}{\omega}\le1,\quad x\in\mathbb{R}.
\end{equation*}
\end{enumerate}
Then, $\lim\limits_{t\to\infty}x_i(t)=v^*$, $v^*\in\Phi$.
\end{Corollary}
\subsection{Empty Consensus Zone: Existence, Stability and Uniqueness of Equilibria}\label{Part:CT_C}
In this part, the existence, stability and uniqueness of equilibria are discussed and proved.
Theorem \ref{th:exist_equilibrium} gives the existence conditions of equilibrium, which is a prerequisite for the Theorem \ref{th:single equilibrium}.
\begin{Theorem}\label{th:exist_equilibrium}
Suppose $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$, $f_{ij}(x)$ is a continuous function.
If there exists a set $[\partial_m,\partial_M]$ and $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\partial_m\le f_{ij}(x)\le\partial_M,\quad x\in[\partial_m,\partial_M],
\end{equation*}
then the system \eqref{MAS_1} exists at least one equilibrium. In fact, all equilibria of the system lie within $[ \partial_m, \partial_M]^n$, if the following conditions hold:
\begin{enumerate}
\item the directed graph $\mathcal{G}$ is strongly connected;
\item there exists a value $\partial\in[\partial_m,\partial_M]$ and two rays
\begin{align*}
L_{1}(x)&=k_{1}(x-\partial)+\partial,\quad x\in(-\infty,\partial];\\
L_{2}(x)&=k_{2}(x-\partial)+\partial,\quad x\in[\partial,+\infty),
\end{align*}
where $k_{1}k_{2}=1$,
such that $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\begin{alignedat}{2}
&x\le f_{ij}(x)< L_{1}(x),\quad & &x\in(-\infty,\partial_m);\\
&\partial_m\le f_{ij}(x)\le\partial_M, & &x\in[\partial_m,\partial_M];\\
&L_{2}(x)< f_{ij}(x)\le x, & &x\in(\partial_M,+\infty).
\end{alignedat}
\end{equation*}
\item $f_{ij}(x)$ is a continuous function, $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$.
\item for any $x'\in(-\infty,\partial_m)\cup(\partial_M,+\infty)$, there exist $j\in\mathbf{N},i\in\mathcal{N}_j$, such that $f_{ij}(x')\ne x'$.
\end{enumerate}
\end{Theorem}
Theorem \ref{th:exist_equilibrium} establishes the existence of equilibrium and gives the region where all equilibria exist. But it does not illustrate whether the MAS will reach equilibria, not to mention the stability of equilibria.
The following theorem indicates that the system will converge to an asymptotically stable equilibrium, if some conditions hold.
\begin{Theorem}\label{th:converge_to_equilibrium}
Along the system \eqref{MAS_1}, suppose following conditions hold:
\begin{enumerate}
\item the directed graph $\mathcal{G}$ is strongly connected;
\item there exists a equilibrium $\mathbf{e}=\{\mathbf{e}_1,\dots,\mathbf{e}_n\}^T$, i.e.,
\begin{equation*}
\dot{x}_i(t)\big|_{\mathbf{e}_i}=\sum\limits_{j\in\mathcal{N}_i}a_{ij}\big(f_{ji}(\mathbf{e}_j)-\mathbf{e}_i\big)=0,\enspace\forall\,i\in\mathbf{N}.
\end{equation*}
\item\label{item:th_converge_to_equilibrium_C3} there exist two rays
\begin{align*}
L_{e1}(\varepsilon)&=k_{e1}\varepsilon,\quad \varepsilon\in(-\infty,0];\\
L_{e2}(\varepsilon)&=k_{e2}\varepsilon,\quad \varepsilon\in[0,+\infty),
\end{align*}
where $k_{e1},k_{e2}<0$ and $k_{e1}k_{e2}=1$,
such that for all $j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{align*}
\varepsilon\le f_{ij}(\mathbf{e}_i+\varepsilon)-f_{ij}(\mathbf{e}_i)< L_{e1}(\varepsilon),\quad \varepsilon\in(-\infty,0);\\
L_{e2}(\varepsilon)< f_{ij}(\mathbf{e}_i+\varepsilon)-f_{ij}(\mathbf{e}_i)\le \varepsilon,\quad\varepsilon\in(0,+\infty).
\end{align*}
\item for any $\varepsilon'\ne0$, there exist $j\in\mathbf{N},i\in\mathcal{N}_j$, such that $f_{ij}(\mathbf{e}_i+\varepsilon')$ is continuous on $\varepsilon'$ and $f_{ij}(\mathbf{e}_i+\varepsilon')-f_{ij}(\mathbf{e}_i)\ne \varepsilon'$.
\end{enumerate}
Then, the equilibrium $\mathbf{e}$ is a unique, asymptotically stable equilibrium, i.e., $\lim\limits_{t\to\infty}x_i(t)=\mathbf{e}_i$.
\end{Theorem}
\begin{remark}
Unlike Theorem \ref{th:consensus}, the boundary rays in Theorem \ref{th:converge_to_equilibrium} are two clusters of parallel lines with the same slopes, but the endpoints may be different.
\end{remark}
\begin{remark}
The unique equilibrium's values are only decided by the network structure and transmission constraint functions, but not related to the initial states of MAS \eqref{MAS_1}.
\end{remark}
The following theorem can be regarded as a combination of theorems \ref{th:exist_equilibrium} and \ref{th:converge_to_equilibrium}, which gives the conditions for the system to converge to an asymptotically stable equilibrium.
\begin{Theorem}\label{th:single equilibrium}
Along the system \eqref{MAS_1}, suppose following conditions hold:
\begin{enumerate}
\item the directed graph $\mathcal{G}$ is strongly connected;
\item the consensus zone $\Phi=\emptyset$;
\item\label{item:th_single equilibrium_C3} for any $j\in\mathbf{N},\,i\in\mathcal{N}_j$ and $\omega\ne0$,
\begin{equation*}
-1<\frac{f_{ij}(x+\omega)-f_{ij}(x)}{\omega}<1,\quad x\notin\Theta_{ij}.
\end{equation*}
\end{enumerate}
Then the MAS \eqref{MAS_1} exits a unique, asymptotically stable equilibrium.
\end{Theorem}
\begin{remark}
Theorem \ref{th:converge_to_equilibrium} requires a known equilibrium of MAS.
In contrast, Theorem \ref{th:single equilibrium} relaxes the condition that the equilibrium is known, i.e., we just need to know the constraints functions and the connectivity of interaction networks, then we can predict the trajectory of MAS.
In conclusion, Theorem \ref{th:single equilibrium} states the existence, stability and uniqueness of Equilibria.
\end{remark}
\begin{Corollary}\label{co:single equilibrium}
Suppose the directed graph $\mathcal{G}$ is strongly connected and the consensus zone $\Phi=\emptyset$.
If for all $j\in\mathbf{N},\,i\in\mathcal{N}_j$, $f_{ij}(x_i)=k_{ij}(x_i)x_i+m_{ij}(x_i)$ is a continuous and piecewise linear function with its slopes $k_{ij}\in(-1,1]$ and $m_{ij}=0$ when $k_{ij}=1$, then the MAS \eqref{MAS_1} has a unique, asymptotically stable equilibrium point.
\end{Corollary}
\section{Numeral Example}
In this section, we present four numeral examples to illustrate the theorems and corollaries proposed in this paper.
Examples \ref{ex1} and \ref{ex2} illustrate the consensus theorem.
Additionally, example \ref{ex3} illustrates the theorem for stability, uniqueness of equilibrium, i.e., Theorem \ref{th:single equilibrium}.
Finally, to demonstrate the theorem for existence and region of equilibrium, i.e., Theorem \ref{th:exist_equilibrium}, we design a special example \ref{ex4} where the equilibria are affected by the initial states of agents.
In all following examples, the interaction networks are strongly connected and the number of agents $n=5$.
For simplicity, in examples \ref{ex2}, \ref{ex3} and \ref{ex4}, let $f_{ij}(x)=f_i(x)$, $\forall\,j\in\mathbf{N}$.
\begin{example}\label{ex1}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex01_1_new.eps}
\caption{The constraint functions in Example \ref{ex1}.}
\label{fig 0101}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex01_2.eps}
\caption{The trajectories of $\mathbf{x}(t)$ in Example \ref{ex1}.}
\label{fig 0102}
\end{figure}
The adjacency matrix in this example is
\begin{equation*}
\mathcal{A} = \begin{bmatrix}
0 & 0 & 3.6 & 0 & 0 \\
0 & 0 & 4.6 & 1.3 & 6.5 \\
3.6 & 0 & 0 & 0 & 7.6 \\
0.5 & 1.4 & 2.1 & 0 & 0 \\
2.9 & 6.5 & 0 & 0 & 0
\end{bmatrix},
\end{equation*}
Fig. \ref{fig 0101} shows candidates for the constraint function imposed on the information transmissions, and the configuration of constraint functions is shown in Table \ref{table2}. It is easy to know that the consensus zone $\Phi=\{0\}$. Fig. \ref{fig 0102} shows that MAS achieves transmission-constrained consensus with $\lim\limits_{t\to\infty}x(t)=0$.
In this example, the constraint function $f_5(x)$ is a continuous function similar to a sawtooth wave.
And the constraint function $f_4(x)$ can be chosen as the boundary rays since it satisfies the Condition (ii) of Theorem \ref{th:consensus} and $k_1k_2=0.8<1$.
\renewcommand\arraystretch{1.3}
\begin{table}[t]
\caption{Configuration of transmission constraints in Example \ref{ex1}}
\label{table2}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\multicolumn{6}{|c|}{Transmission Constraints $f_{ij}$} \\
\hline
\diagbox{$i$}{$f_{ij}$}{$j$} & 1 & 2 & 3 & 4 & 5\\
\hline
1 & \text{---} & \text{---} & $(f_b+f_d)/2$ & \text{---} & \text{---} \\
\hline
2 & \text{---} & \text{---} & $f_a$ & $f_b$ & $(f_e+f_a)/2$ \\
\hline
3 & $f_d$ & \text{---} & \text{---} & \text{---} & $f_c$ \\
\hline
4 & $f_a$ & $f_b$ & $(f_c+f_d)/2$ & \text{---} & \text{---} \\
\hline
5 & $f_c$ & $f_e$ & \text{---} & \text{---} & \text{---} \\
\hline
\end{tabular}
\end{table}
\renewcommand\arraystretch{1}
\end{example}
\begin{example}\label{ex2}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex02_1.eps}
\caption{The constraint functions in Example \ref{ex2}.}
\label{fig 0201}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex02_2ss.eps}
\caption{The trajectories of $\mathbf{x}(t)$ in Example \ref{ex2}.}
\label{fig 0202}
\end{figure}
The adjacency matrix in the example is $\mathcal{A} = \mathbf{1}_{5\times5}-\mathbf{I}_5$.
In Fig. \ref{fig 0201}, we can get that the consensus zone $\Phi=[-1,1]^5$.
The auxiliary line in Fig. \ref{fig 0201} represents the boundary rays with $k_1k_2=0.64<1$.
Fig. \ref{fig 0202} shows that $x(t)\to\Phi$ as $t\to\infty$.
\end{example}
\begin{example}\label{ex3}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex03_1.eps}
\caption{The constraint functions in Example \ref{ex3}.}
\label{fig 0301}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex03_2.eps}
\caption{The trajectories of $\mathbf{x}(t)$ in Example \ref{ex3}.}
\label{fig 0302}
\end{figure}
The adjacency matrix in this continuous-time example is $\mathcal{A} = \mathbf{1}_{5\times5}-\mathbf{I}_5$.
Fig. \ref{fig 0302} shows that no matter which initial states of agents are, MAS will reach the same equilibrium, i.e., unique equilibrium (see the points $\big(\mathbf{e}_i,f_i(\mathbf{e}_i)\big)$ shown by circles in Fig. \ref{fig 0301}).
It is easy to know that for any equilibrium $\mathbf{e}$, there exist two clusters of rays (see auxiliary lines in Fig. \ref{fig 0301}) satisfying the Condition (iii) of Theorem \ref{th:converge_to_equilibrium} or the Condition (iii) of Theorem \ref{th:single equilibrium}, which means that the system will converge to a unique, asymptotically stable equilibrium, even though we do not know the value of equilibrium.
\end{example}
\begin{example}\label{ex4}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex04_1.eps}
\caption{The constraint functions in Example \ref{ex4}.}
\label{fig 0401}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{ex04_2.eps}
\caption{The trajectories of $\mathbf{x}(t)$ in Example \ref{ex4}.}
\label{fig 0402}
\end{figure}
The adjacency matrix $\mathcal{A} = \mathbf{1}_{5\times5}-\mathbf{I}_5$.
Fig. \ref{fig 0402} shows that the initial state of MAS will affect the equilibrium of MAS, and all equilibria lie in the largest positively invariant set (see the square surrounded by dotted lines).
Incidentally, this example shows that the condition `$m_{ij}=0$ when $k_{ij}=1$' in Corollary \ref{co:single equilibrium} is necessary. Otherwise, the equilibria may not be unique without this condition.
\end{example}
\section{Conclusion}
This paper focuses on the transmission-constrained consensus problem of multiagent networks, where information transmissions between agents are affected by irregular constraint functions. We obtain the necessary and sufficient conditions about the range of transmission constraint functions where agents' states can converge to consensus.
Due to the piecewise continuous constraint functions, the LaSalle invariance principle is not applicable in those proofs. We construct a sophisticated Lyapunov function and discuss the boundaries of multiple limit points of MAS states to facilitate the convergence analysis.
Meanwhile, in some cases where the system cannot achieve consensus, there is an asymptotically stable equilibrium independent of the initial values of agents' states. The existence, uniqueness and stability of the system equilibrium point are proved by Brouwer fixed-point theorem and Lyapunov's direct method. Finally, the numerical simulations are presented to verify the effectiveness of the proposed theoretical results.
\appendices
\section{Proof of Theorem \ref{th:distance=0}}
\subsection{Technical lemmas}
First of all, we introduce some technical lemmas.
\begin{Lemma}\label{le:der_to_max}
\emph{\bf{(Lemma 2.2 in \cite{lin2007state})}} If for all $i\in\mathbf{N}$, $Z_i(\mathbf{x}):\mathbb{R}^d\to\mathbb{R}$ is of class $C^1$, and denote $Z(\mathbf{y})=\max\limits_{i\in\mathbf{N}}Z_i(\mathbf{y})$. Denote $\mathsf{N}_m(t)=\{i\in\mathbf{N}:Z(\mathbf{y})=Z_i(\mathbf{y})\}$ the indices set in which the maximum is reached at time $t$. Then it turns out that
\begin{equation*}
d^+Z\big(\mathbf{y}(t)\big)=\max\limits_{i\in\mathsf{N}_m(t)}\dot{Z}_i\big(\mathbf{y}(t)\big).
\end{equation*}
\end{Lemma}
\begin{Lemma}\label{le:one_line_crosses_boundary}
If $\partial\in[\partial_m,\partial_M]$ and $k_1k_2=1$, then
\begin{equation*}
\partial_M-\partial_m\ge\min\{(1-k_2)(\partial_M-\partial),(1-k_1)(\partial-\partial_m)\}.
\end{equation*}
\end{Lemma}
\begin{proof}
When $\partial_M-\partial_m=0$ or $(\partial_M-\partial)(\partial-\partial_m)=0$, the conclusion is obvious.
When $\partial_M-\partial_m>0$ and $(\partial_M-\partial)(\partial-\partial_m)>0$, we use a contradiction argument to prove it.
Let $\partial_M-\partial_m<\min\{(1-k_2)(\partial_M-\partial),(1-k_1)(\partial-\partial_m)\}$, i.e.,
\begin{equation}\label{eq:le_one_line_crosses_boundary_1}
\left\{
\begin{aligned}
\partial_M-\partial_m&<(1-k_2)(\partial_M-\partial);\\
\partial_M-\partial_m&<(1-k_1)(\partial-\partial_m).
\end{aligned}
\right.
\end{equation}
Since $\partial\in(\partial_m,\partial_M)$, we let $\partial=\rho\partial_M+(1-\rho)\partial_m$ with $\rho\in(0,1)$. Then, \eqref{eq:le_one_line_crosses_boundary_1} can be rewritten as
\begin{equation*}
\left\{
\begin{aligned}
\partial_M-\partial_m&<(1-k_2)(1-\rho)(\partial_M-\partial_m);\\
\partial_M-\partial_m&<(1-k_1)\rho(\partial_M-\partial_m).
\end{aligned}
\right.
\end{equation*}
Then, we can get that
\begin{equation*}
\left\{
\begin{aligned}
-k_2>\frac{\rho}{1-\rho}>0;\\
-k_1>\frac{1-\rho}{\rho}>0,
\end{aligned}
\right.
\end{equation*}
which implies that $k_1k_2>1$. We get a contradiction, and prove the Lemma \ref{le:one_line_crosses_boundary}.
\end{proof}
The following lemma shows the implicit inequality from the given condition, and it helps us discuss the monotonicity of transmission-constrained consensus dynamics.
\begin{Lemma}\label{le:x_M and x_m}
Denote $x_m(t)=\min\limits_{i\in\mathbf{N}}x_i(t)$, $x_M(t)=\max\limits_{i\in\mathbf{N}}x_i(t)$, $L_1\big(x_i(t)\big)=k_1x_i(t)+(1-k_1)\partial$, $L_2\big(x_i(t)\big)= k_2x_i(t)+(1-k_2)\partial$ and
\begin{align*}
Y(t)=\max\Big\{(1-k_1)\big(\partial-x_m(t)\big),\;(1-k_2) \big(x_M(t)-\partial\big)\Big\},
\end{align*}
where $\partial$ is a constant value.
If $k_1,k_2<0$ and $k_1k_2=1$, then $\forall\,i\in\mathbf{N}$,
\begin{enumerate}
\item\label{item:con1} $Y(t)=(1-k_2) \big(x_M(t)-\partial\big)\implies$$x_M(t)\ge L_1\big(x_i(t)\big)$;
\item\label{item:con2} $Y(t)=(1-k_1)\big(\partial-x_m(t)\big)\implies$$x_m(t)\le L_2\big(x_i(t)\big)$.
\end{enumerate}
\end{Lemma}
\begin{proof}
We first discuss Item \ref{item:con1}, i.e., the case where $Y(t)=(1-k_2) \big(x_M(t)-\partial\big)$, which implies that $x_M(t)\ge\partial$.
Note that if $x_i(t)\ge\partial$, it is easy to know that $\forall\,i\in\mathbf{N}$,
\begin{equation*}
x_M(t)>\partial\ge\partial+(1-k_1)\big(\partial-x_i(t)\big)=L_1(x_i(t)).
\end{equation*}
Hence, we assume $x_i(t)<\partial$, and continue this discussion.
\begin{align*}
&Y(t)=(1-k_2) \big(x_M(t)-\partial\big)\\
\iff& (1-k_2) \big(x_M(t)-\partial\big)\ge (1-k_1)\big(\partial-x_i(t)\big)\\
\iff& x_M(t)-k_1x_i(t)-(1-k_1)\partial\\
&\ge k_2x_M(t)-x_i(t)+(1-k_2)\partial\\
&=\big(\partial-x_i(t)\big)+k_2\big(x_M(t)-\partial\big).
\end{align*}
Since $\partial-x_i(t)>0$ and $k_1k_2=1$, it turns out that
\begin{align*}
&x_M(t)-k_1x_i(t)-(1-k_1)\partial\\
&\ge k_2x_M(t)+k_1k_2\big(\partial-x_i(t)\big)-k_2\partial\\
\iff& (1-k_2)\big(x_M(t)-k_1x_i(t)\big)\ge(1-k_2)(1-k_1)\partial\\
\implies& x_M(t)-k_1x_i(t)-(1-k_1)\partial\ge0.
\end{align*}
Then it concludes that $x_M(t)-k_1x_i(t)-(1-k_1)\partial\ge0$, which implies
\begin{equation*}
x_M(t)\ge L_1\big(x_i(t)\big)=k_1x_i(t)+(1-k_1)\partial,\enspace \forall\,i\in\mathbf{N}.
\end{equation*}
Therefore, the proof of Item \ref{item:con1} is completed.
The proof method of Item \ref{item:con2} is similar to that of Item \ref{item:con1}, and hence is omitted here.
Hence, Lemma \ref{le:x_M and x_m} is proved.
\end{proof}
\begin{Lemma}\label{le:positive_invariant}
For the MAS \eqref{MAS_1}, if there exists a set $[ \partial_m, \partial_M]^n$ such that for all $j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\partial_m\le f_{ij}(x)\le\partial_M,\enspace x\in[\partial_m,\partial_M],
\end{equation*}
then $[ \partial_m, \partial_M]^n$ is a positively invariant set.
\end{Lemma}
\begin{proof}
The dynamics of MAS \ref{MAS_1} can be rewritten as
\begin{equation}\label{eq:vector field}
\dot{x}(t)=h\big(x(t)\big)=\Big(h_1\big(x(t)\big),\dots, h_n\big(x(t)\big)\Big)^T,
\end{equation}
where $h_i\big(x(t)\big)=\sum_{j\in \mathcal{N}_i}a_{ij}\Big[f_{ji}\big(x_j(t)\big)-x_i(t)\Big]$.
The initial states of \eqref{eq:vector field} is $x_0=x(t_0)$. Assume that $x(t_0)\in[ \partial_m, \partial_M]^n$. Since the vector field $h$ is pointing inwards $[ \partial_m, \partial_M]^n$ that is an $n$-dimensional cube, it concludes that
\begin{equation*}
x(t)\in[ \partial_m, \partial_M]^n,\enspace\forall\,t\ge t_0.
\end{equation*}
It shows that $[ \partial_m, \partial_M]^n$ is a positively invariant set and the proof is completed.
\end{proof}
\begin{Lemma}\label{le:non-increasing function}
Along the system \eqref{MAS_1}, suppose there exists a set $[\partial_m,\partial_M]$, a value $\partial\in[\partial_m,\partial_M]$ and two rays
\begin{align*}
L_{1}(x)&=k_{1}(x-\partial)+\partial,\quad x\in(-\infty,\partial];\\
L_{2}(x)&=k_{2}(x-\partial)+\partial,\quad x\in[\partial,+\infty),
\end{align*}
where $k_{1}k_{2}=1$,
such that $\forall\,j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{equation*}
\begin{alignedat}{2}
&x\le f_{ij}(x)< L_{1}(x),\quad & &x\in(-\infty,\partial_m);\\
&\partial_m\le f_{ij}(x)\le\partial_M,\quad & &x\in[\partial_m,\partial_M];\\
&L_{2}(x)<f_{ij}(x)\le x, & &x\in(\partial_M,+\infty).
\end{alignedat}
\end{equation*}
Denote $x_m(t)=\min\limits_{i\in\mathbf{N}}x_i(t)$, $x_M(t)=\max\limits_{i\in\mathbf{N}}x_i(t)$ and
\begin{align*}
Y(t)=&\max\big\{\partial_M-\partial_m,x_M(t)-\partial_m,\partial_M-x_m(t),\\
&(1-k_2)\big(x_M(t)-\partial\big),(1-k_1)\big(\partial-x_m(t)\big)\big\}.
\end{align*}
If $\partial_M-\partial_m\ge\max\{(1-k_2)(\partial_M-\partial),(1-k_1)(\partial-\partial_m)\}$, then $Y(t)$ is a non-increasing function for any initial state $x_*\in\mathbb{R}^n$.
\end{Lemma}
\begin{proof}
By the structure of $Y(t)$, there exists five cases:
\begin{enumerate}
\item\label{item:th_distance=0_partA_case1} $Y(t)=Y_1(t)=(1-k_2)\big(x_M(t)-\partial\big)$;
\item\label{item:th_distance=0_partA_case2} $Y(t)=Y_2(t)=x_M(t)-\partial_m$;
\item\label{item:th_distance=0_partA_case3} $Y(t)=Y_3(t)=(1-k_1)\big(\partial-x_m(t)\big)$;
\item\label{item:th_distance=0_partA_case4} $Y(t)=Y_4(t)=\partial_M-x_m(t)$;
\item\label{item:th_distance=0_partA_case5} $Y(t)=Y_5(t)=\partial_M-\partial_m$.
\end{enumerate}
At first, we analyze the Case \ref{item:th_distance=0_partA_case1}.
Since
\begin{equation*}
(1-k_2)\big(x_M(t)-\partial\big)\ge\partial_M-\partial_m\ge(1-k_2)(\partial_M-\partial),
\end{equation*}
it turns out that $x_M(t)\ge\partial_M$.
Denote $\mathcal{I}_M(t)=\big\{k:x_k(t)=\max\limits_{l\in\mathbf{N}}x_l(t)\big\}$. By Lemma \ref{le:der_to_max}, we have
\begin{align*}
&\phantom{=}d^+Y_1(t)=d^+\max\limits_{i\in\mathbf{N}}\Big\{(1-k_2)\big(x_i(t)-\partial\big)\Big\}\\
&=\max\limits_{i\in\mathcal{I}_M(t)}\Big\{\big(1-k_2\big)\dot{x}_i(t)\Big\}\\
&=\max\limits_{i\in\mathcal{I}_M(t)}\Big\{(1-k_2)\sum\limits_{j\in\mathcal{N}_{i}(t)}a_{ij}\Big(f_{ji}\big(x_j(t)\big)-x_{i}(t)\Big)\Big\}.
\end{align*}
For all $i\in\mathcal{I}_M(t)$, which implies that $x_i(t)=x_M(t)$, and we conduct the following analysis:
\begin{equation*}
\begin{cases}
x_M\ge L_{1}(x_j)> f_{ji}(x_j) &\text{if } x_j\le\partial_m;\\
x_M\ge\partial_M\ge f_{ji}(x_j) &\text{if } x_j\in[\partial_m, \partial_M];\\
x_M\ge x_j\ge f_{ji}(x_j) &\text{if } x_j>\partial_M,
\end{cases}
\end{equation*}
where the first inequality follows from Lemma \ref{le:x_M and x_m} and the fact that $(1-k_2)\big(x_M(t)-\partial\big)\ge(1-k_1)\big(\partial-x_m(t)\big)$.
Then, it can be concluded that $d^+Y_1(t)\le0$ when $Y(t)=Y_1(t)=(1-k_2)\big(x_M(t)-\partial\big)$.
Secondly, we discuss the Case \ref{item:th_distance=0_partA_case2}. Since $x_M(t)-\partial_m\ge\partial_M-\partial_m$, it turns out that $x_M(t)\ge\partial_M$.
Since $x_M(t)-\partial_m\ge(1-k_1)\big(\partial-x_m(t)\big)$, we can get that
\begin{align*}
x_M(t)\ge& k_1x_m(t)+(1-k_1)\partial+\partial_m-x_m(t)\\
=&L_{1}\big(x_m(t)\big)+\partial_m-x_m(t)\\
\ge&L_{1}\big(x_i(t)\big)+\partial_m-x_m(t),\enspace\forall\,i\in\mathbf{N}.
\end{align*}
It turns out that
\begin{equation*}
\begin{cases}
x_M\ge L_{1}(x_j)> f_{ji}(x_j) &\text{if } x_j\le\partial_m;\\
x_M\ge\partial_M\ge f_{ji}(x_j) &\text{if } x_j\in[\partial_m, \partial_M];\\
x_M\ge x_j\ge f_{ji}(x_j) &\text{if } x_j>\partial_M,
\end{cases}
\end{equation*}
where the first inequality follows from $x_M\ge L_{1}(x_i)+\partial_m-x_m\ge L_{1}(x_i)$ when $x_m\le\partial_m$.
Then, it shows that
\begin{align*}
&\phantom{=}d^+Y_2(t)=\max\limits_{i\in\mathcal{I}_M(t)}\dot{x}_i(t)\\
&=\max\limits_{i\in\mathcal{I}_M(t)}\Big\{\sum\limits_{j\in\mathcal{N}_{i}(t)}a_{ij}\Big(f_{ji}\big(x_j(t)\big)-x_{i}(t)\Big)\Big\}\le0.
\end{align*}
The analyses of cases \ref{item:th_distance=0_partA_case3} and \ref{item:th_distance=0_partA_case4} are symmetric to those of cases \ref{item:th_distance=0_partA_case1} and \ref{item:th_distance=0_partA_case2}, hence they are omitted. As for Case \ref{item:th_distance=0_partA_case5}, the conclusion is obvious.
Therefore, by the above five cases, it can be concluded that $Y(t)$ is a non-increasing function.
\end{proof}
\begin{Lemma}\label{le:the boundary of x_M and x_m}
Suppose the MAS \eqref{MAS_1} satisfies the conditions in Lemma \ref{le:non-increasing function}. Denote the initial time $t_0$. Then for any $t\ge t_0$, we have
\begin{enumerate}
\item $Y(t_0)=\partial_M-\partial_m\implies \partial_m\le x_i(t)\le\partial_M,\,\forall\,i\in\mathbf{N}$;
\item $Y(t_0)=(1-k_2)\big(x_M(t_0)-\partial\big)$$\implies$$ x_m(t)\ge L_{2}\big(x_M(t_0)\big)$;
\item $Y(t_0)=(1-k_1)\big(\partial-x_m(t_0)\big)$$\implies$$ x_M(t)\le L_{1}\big(x_m(t_0)\big)$;
\item $Y(t_0)=x_M(t_0)-\partial_m$$\implies$$ x_m(t)\ge\min\{x_m(t_0),\partial_m\}$;
\item $Y(t_0)=\partial_M-x_m(t_0)$$\implies$$ x_M(t)\le\max\{x_M(t_0),\partial_M\}$.
\end{enumerate}
\end{Lemma}
\begin{proof}
$Y(t_0)=\partial_M-\partial_m$ means that $\mathbf{x}(t_0)\in[\partial_m,\partial_M]^n$. By Lemma \ref{le:positive_invariant}, we can get that $[\partial_m,\partial_M]^n$ is a positively invariant set, and this case is proven trivially.
When $Y(t_0)=(1-k_2)\big(x_M(t_0)-\partial\big)$, if $ x_m(t)< L_{2}\big(x_M(t_0)\big)$, we have $Y(t)\ge (1-k_1)\big(\partial-x_m(t)\big)>Y(t_0)$, which contradicts the Lemma \ref{le:non-increasing function}. By symmetry, the case where $Y(t_0)=(1-k_1)\big(\partial-x_m(t_0)\big)$ is also proven.
When $Y(t_0)=x_M(t_0)-\partial_m$, it turns out that $L_{2}\big(x_M(t_0)\big)\ge \partial_m$. Since $x_M(t)\le x_M(t_0)$, $\forall\,t\ge t_0$, we have $L_{2}\big(x_i(t)\big)\ge L_{2}\big(x_M(t_0)\big)$, $\forall\,i\in\mathbf{N},\,t\ge t_0$. Hence, it can be concluded that $f_{ji}\big(x_j(t)\big)\ge\min\{x_j(t),\partial_m,L_{2}\big(x_j(t)\big)\}\ge\min\{x_m(t),\partial_m\}$, $\forall\,i,j\in\mathbf{N}$. Therefore, we can get that $ x_m(t)\ge\min\{x_m(t_0),\partial_m\}$. By symmetry, the case where $Y(t_0)=\partial_M-x_m(t_0)$ is also proven.
This proof is completed.
\end{proof}
\subsection{Proof of Theorem \ref{th:distance=0}}
\begin{proof}
\subsubsection{Necessity}
We use a contradiction argument.
For simplicity, we assume that there are only two agents in MAS \eqref{MAS_1}, i.e., $\mathbf{N}=\{1,2\}$.
Since $\mathcal{G}$ is strongly connected, it turns out that $\mathcal{N}_1=\{2\}$ and $\mathcal{N}_2=\{1\}$.
Denote the initial time $t_0\ge0$. Suppose there exist $j\in\mathbf{N},i\in\mathcal{N}_j$ and $x_j(t_0)\in[\partial_m, \partial_M]$ such that
\begin{equation*}
f_{ji}\big(x_j(t_0)\big)=\partial_M+\omega,\enspace\omega>0.
\end{equation*}
Without loss of generality, assume that $j=1$ and $i=2$, i.e., $f_{12}\big(x_1(t_0)\big)=\partial_M+\omega$ in which $x_1(t_0)\in[\partial_m, \partial_M]$.
Let $x_2(t_0)=f_{12}\big(x_1(t_0)\big)$ and $f_{21}\big(x_2(t_0)\big)=x_1(t_0)$, then it can be concluded that for all $t\ge t_0$,
\begin{align*}
\dot{x}_1(t)&=a_{12}\Big(f_{21}\big(x_2(t)\big)-x_1(t)\Big)=0,\\
\dot{x}_2(t)&=a_{21}\Big(f_{12}\big(x_1(t)\big)-x_2(t)\Big)=0.
\end{align*}
which implies that $x_1(t)=x_1(t_0)$, $x_2(t)=x_2(t_0)$, $\forall\,t\ge t_0$.
Moreover, because
\begin{equation*}
\left\{
\begin{aligned}
&x_1(t_0)\in[\partial_m, \partial_M],\\
&f_{12}\big(x_1(t_0)\big)=\partial_M+\omega>\partial_M,\\
&x_2(t_0)=f_{12}\big(x_1(t_0)\big)=\partial_M+\omega>\partial_M,\\
&f_{21}\big(x_2(t_0)\big)=x_1(t_0)\le\partial_M,
\end{aligned}
\right.
\end{equation*}
it is easy to find two rays $L_{1}$ and $L_{2}$ satisfying the Condition \ref{it:th_distance=0_2}, and the Condition \ref{it:th_distance=0_3} is also satisfied.
Since $\forall\,t\ge t_0$, $x_2(t)=\partial_M+\omega>\partial_M$, it shows that
\begin{equation*}
\lim\limits_{t\to\infty}\textnormal{distance}\big([ \partial_m, \partial_M]^n,x(t)\big)\ne0.
\end{equation*}
Hence, we get a contradiction and the proof for the necessity statement of Theorem \ref{th:distance=0} is proved.
\subsubsection{Sufficiency}
We prove it in three steps.
\begin{step1}\label{th_distance=0_partA_step1}
Since $\partial\in[\partial_m,\partial_M]$ and $k_1k_2<1$, by Lemma \ref{le:one_line_crosses_boundary}, there exist two possibilities:
\begin{enumerate}
\item\label{item:th_distance=0_P1} $\partial_M-\partial_m\ge\max\{(1-k_2)(\partial_M-\partial),(1-k_1)(\partial-\partial_m)\}$;
\item\label{item:th_distance=0_P2} $\partial_M-\partial_m<(1-k_2)(\partial_M-\partial)$ or $\partial_M-\partial_m<(1-k_1)(\partial-\partial_m)$.
\end{enumerate}
We discuss the Possibility \ref{item:th_distance=0_P1}) in the rest of Step \ref{th_distance=0_partA_step1}, and the Possibility \ref{item:th_distance=0_P2}) is analyzed in Step \ref{th_distance=0_partB_step1}.
Assume that $\partial_M-\partial_m\ge\max\{(1-k_2)(\partial_M-\partial),(1-k_1)(\partial-\partial_m)\}$. Form Lemma \ref{le:non-increasing function}, we have that $Y(t)$ is a non-increasing function.
Let $n$ be the number of agents. Since
\begin{align*}
Y(t)=&\max\big\{\partial_M-\partial_m,x_M(t)-\partial_m,\partial_M-x_m(t),\\
&(1-k_2)\big(x_M(t)-\partial\big),(1-k_1)\big(\partial-x_m(t)\big)\big\},
\end{align*}
we continue this proof case by case.
\begin{case}\label{case:Y_1}
$Y(t_0)=Y_1(t_0)=(1-k_2)\big(x_M(t_0)-\partial\big)$.
By Lemma \ref{le:the boundary of x_M and x_m}, we have for any $t\ge t_0$,
\begin{align*}
&L_{1}\big(x_m(t)\big)=k_{1}\big(x_m(t)-\partial\big)+\partial\\
\le& k_{1}k_{2}x_M(t_0)+k_{1}(1-k_{2})\partial+(1-k_{1})\partial=x_M(t_0).
\end{align*}
Choose $i_0\in\mathcal{I}_0:=\{i:x_i(t_0)=x_m(t_0)\}$. For any $j\in\mathcal{N}_{i_0}$, we can get that
\begin{equation*}
f_{ji_0}\big(x(t)\big)\le\max\big\{x_M(t),L_{1}\big(x_m(t)\big),\partial_M\big\}\le x_M(t_0).
\end{equation*}
It turns out that
\begin{align*}
\dot{x}_{i_0}(t)&=\sum_{j\in\mathcal{N}_{i_0}}a_{i_0j}\big[f_{ji_0}\big(x_j(t)\big)-x_{i_0}(t)\big]\\
&\le \alpha_{i_0}[x_M(t_0)-x_{i_0}(t)],
\end{align*}
which implies that
\begin{equation*}
x_{i_0}(t)\le e^{-\alpha_{i_0}(t-t_0)}x_m(t_0)+[1-e^{-\alpha_{i_0}(t-t_0)}]x_M(t_0).
\end{equation*}
If $t\in[t_0,t_0+\tau]$, then we have for any $i_0\in\mathcal{I}_0$,
\begin{equation}\label{eq:the boundary of x_i_0}
x_{i_0}(t)\le \gamma_0x_m(t_0)+(1-\gamma_0)x_M(t_0),
\end{equation}
where $\gamma_0=e^{-\tau\bar{a}}$.
Choose $i_1\in\mathcal{I}_1:=\{i:\exists\,j\in\mathcal{I}_0$, $j\in\mathcal{N}_i\}$. By the conditions of Theorem \ref{th:distance=0} and the equation \eqref{eq:the boundary of x_i_0}, it is trivial to get that $f_{i_0i_1}\big(x_{i_0}(t)\big)<x_M(t_0)$. Hence, for any $t\in[t_0,t_0+\tau/n]$, there exists a constant $\gamma_0'\in(0,1)$ such that
\begin{equation*}
f_{i_0i_1}\big(x_{i_0}(t)\big)\le\gamma_0'x_m(t_0)+(1-\gamma_0')x_M(t_0).
\end{equation*}
Then, we can get that
\begin{align*}
&x_{i_1}(t_0+\frac{\tau}{n})\\
\le& e^{-\alpha_{i_1}\frac{\tau}{n}}x_{i_1}(t_0)+\big[a_{i_1i_0}\gamma_0'\big(x_m(t_0)-x_M(t_0)\big)\\
&+\alpha_{i_1}x_M(t_0)\big]\int_{t_0}^{t_0+\frac{\tau}{n}}e^{-\alpha_{i_1}(t_0+\frac{\tau}{n}-s)}ds\\
\le& e^{-\alpha_{i_1}\frac{\tau}{n}}x_M(t_0)+(1-e^{-\alpha_{i_1}\frac{\tau}{n}})x_M(t_0)\\
&+a_{i_1i_0}\gamma_0'\big(x_m(t_0)-x_M(t_0)\big)\int_{t_0}^{t_0+\frac{\tau}{n}}e^{-\alpha_{i_1}(t_0+\frac{\tau}{n}-s)}ds\\
=& x_M(t_0)+\frac{a_{i_1i_0}}{\alpha_{i_1}}(1-e^{-\alpha_{i_1}\frac{\tau}{n}})\gamma_0'\big(x_m(t_0)-x_M(t_0)\big).
\end{align*}
Since there exists a constant $\rho_1>0$ such that for any $i_1\in\mathcal{I}_1$, $i_0\in\mathcal{N}_{i_1}$, $\rho_1\le \frac{a_{i_1i_0}}{\alpha_{i_1}}(1-e^{-\alpha_{i_1}\frac{\tau}{n}})$. Therefore, it turns out that
\begin{equation*}
x_{i_1}(t_0+\frac{\tau}{n})\le\rho_1\gamma_0'x_m(t_0)+(1-\rho_1\gamma_0')x_M(t_0).
\end{equation*}
Similar to \eqref{eq:the boundary of x_i_0}, we can get that for any $i_1\in\mathcal{I}_1$, $t\in[t_0+\frac{\tau}{n},t_0+\tau]$,
\begin{equation}\label{eq:the boundary of x_i_1}
x_{i_1}(t)\le \gamma_1 x_m(t_0)+(1-\gamma_1)x_M(t_0),
\end{equation}
where $\gamma_1=\rho_1\gamma_0'\gamma_0$. Continuing the above analysis over $[t_0+\frac{\tau}{n}m,t_0+\tau]$, $\forall\,m=1,2,\dots,n-1$, it can be concluded that for all $i\in\mathbf{N}$,
\begin{equation*}
x_i(t_0+\tau)\le\gamma_{n-1}x_m(t_0)+(1-\gamma_{n-1})x_M(t_0),
\end{equation*}
where $\gamma_{n-1}=\rho_{n-1}\gamma_{n-2}'\gamma_{0}$.
If $x_m(t_0)<x_M(t_0)$, then there exists a constant $\omega\in(0,1]$ such that $x_m(t_0)\le\omega\partial+(1-\omega)x_M(t_0)$. Here, we have
\begin{equation*}
Y_1(t_0+\tau)=(1-k_2)\big(x_M(t_0+\tau)-\partial\big)\le(1-\omega\gamma_{n-1})Y_1(t_0).
\end{equation*}
If $x_m(t_0)=x_M(t_0)>\partial_M$, we use the Condition \ref{it:th_distance=0_3}) of Theorem \ref{th:distance=0} to get the convergence. Since there exists $i_0$, $i_1$ such that $f_{i_0i_1}\big(x_{i_0}(t_0)\big)$ is continuous on $x_{i_0}(t_0)=x_M(t_0)$ and $f_{i_0i_1}\big(x_{i_0}(t_0)\big)<x_M(t_0)$. Hence, there exists $T(\omega')$ such that $\forall\,t\in[t_0,t_0+T(\omega')]$, $f_{i_0i_1}\big(x_{i_0}(t)\big)\le \omega'\partial+(1-\omega')x_M(t_0)$, where the constant $\omega'\in(0,1]$. Similar to \eqref{eq:the boundary of x_i_1}, we have $\forall\,t\in[t_0+T(\omega'),t_0+nT(\omega')]$,
\begin{equation*}
x_{i_1}(t)\le \omega_1 \partial+(1-\omega_1)x_M(t_0),
\end{equation*}
where $\omega_1=\rho_1\omega'\gamma_0$. Furthermore, it shows that
\begin{equation*}
Y_1\big(t_0+nT(\omega')\big)\le(1-\omega_{n-1})Y_1(t_0),
\end{equation*}
where $\omega_{n-1}=\rho_{n-1}\omega_{n-2}'\gamma_{0}$.
The analysis of Case \ref{case:Y_1} is completed. The case $Y(t_0)=Y_3(t_0)=(1-k_1)\big(\partial-x_m(t_0)\big)$ is symmetric to Case \ref{case:Y_1}, so we omit its analysis.
\end{case}
\begin{case}\label{case:Y_2}
$Y(t_0)=Y_2(t_0)=x_M(t_0)-\partial_m$.
By Lemma \ref{le:the boundary of x_M and x_m}, we can get that $\forall\,i\in\mathbf{N}$, $t\ge t_0$, $x_M(t_0)\ge L_1\big(x_i(t)\big)$. Hence, it it trivially to get that
\begin{equation*}
x_i(t_0+\tau)\le\gamma_{n-1}x_m(t_0)+(1-\gamma_{n-1})x_M(t_0),\enspace \forall\,i\in\mathbf{N}.
\end{equation*}
Furthermore, we can get that $Y_2(t_0+\tau)\le(1-\omega\gamma_{n-1})Y_2(t_0)$ or $Y_2\big(t_0+nT(\omega')\big)\le(1-\omega_{n-1})Y_2(t_0)$. The case $Y(t_0)=Y_4(t_0)=\partial_M-x_m(t_0)$ is symmetric to this case, so we omit its analysis.
Finally, we can get that $x(t)\to[\partial_m,\partial_M]^n$ as $t\to\infty$, and the proof of the Possibility \ref{item:th_distance=0_P1} is completed.
\end{case}
\end{step1}
\begin{step1}\label{th_distance=0_partB_step1}
In this step, we will complete the proof of Possibility \ref{item:th_distance=0_P2}, i.e., $\partial_M-\partial_m<(1-k_2)(\partial_M-\partial)$ or $\partial_M-\partial_m<(1-k_1)(\partial-\partial_m)$. By symmetry, let $\partial_M-\partial_m<(1-k_1)(\partial-\partial_m)$, which implies that $\exists\,\partial'_M>\partial_M$, such that $\partial'_M-\partial_m=(1-k_1)(\partial-\partial_m)$. Denote
\begin{align*}
Y'(t)=&\max\big\{\partial_M'-\partial_m,x_M(t)-\partial_m,\partial_M'-x_m(t),\\
&(1-k_2)\big(x_M(t)-\partial\big),(1-k_1)\big(\partial-x_m(t)\big)\big\}.
\end{align*}
Similar to the proof in Step \ref{th_distance=0_partA_step1}, it concludes that $x(t)\to[\partial_m,\partial_M']^n$ as $t\to\infty$.
Then, we use a contradiction argument to prove that for any solution $x(t)\to[\partial_m,\partial_M]^n$ as $t\to\infty$.
Assume that there exist a solution $\hat{x}(t)$ and $i^*\in\mathbf{N}$, such that $\hat{x}_{i^*}(t)\to(\partial_M,\partial_M']$ as $t\to\infty$. Then, there is a $T^*$ such that for any $t>T^*$ and $j\in\mathbf{N}$, $f_{i^*j}\big(\hat{x}_{i^*}(t)\big)>\partial_m$.
Since the directed graph $\mathcal{G}$ is strongly connected, we can get that there exists a $T'>T^*$ such that $\forall\,t>T'$, $i\in\mathbf{N}$, $\hat{x}_i(t)>\partial_m$.
Denote $Z(t)=\max\{\hat{x}_M(t),\partial_M\}$. Repeat the analysis of Step \ref{th_distance=0_partA_step1}, it shows that $\lim\limits_{t\to\infty}Z(t)=\partial_M$, which implies that
\begin{equation*}
\limsup\limits_{t\to\infty}\hat{x}_{i^*}(t)\le\partial_M.
\end{equation*}
Therefore, the trajectory of $\hat{x}_{i^*}(t)$ cannot converge to $(\partial_M,\partial_M']$ as $t\to\infty$.
Here we prove that for any solution $x(t)\to[\partial_m,\partial_M]^n$ as $t\to\infty$.
\end{step1}
The proof is completed.
\end{proof}
\section{Proof of Theorem \ref{th:consensus}}
\subsection{Technical lemma}
\begin{Lemma}\label{le:rc}
\emph{\bf{(Proposition 4.10 in \cite{shi2013robust})}}
Let graph $\mathcal{G}$ has a directed spanning tree, and consider the dynamics of MAS defined over $\mathcal{G}$:
\begin{equation*}
\dot{x}_i(t)=\sum^N_{j=1}a_{ij}\big(x_j(t)-x_i(t)\big)+\theta_i(t),\quad i=\mathbf{N},
\end{equation*}
in which $\theta_i(t)$ is piecewise continuous on $[t_0,\infty)$ and is finite. If $\lim\limits_{t\to\infty}\theta_i(t)=0$, $\forall\,i\in\mathbf{N}$, then $\lim\limits_{t\to\infty}x_i(t)-x_j(t)=0$, $\forall\,i,j\in\mathbf{N}$.
\end{Lemma}
\subsection{Proof of Theorem \ref{th:consensus}}
\begin{proof}
\subsubsection{Necessity}
We use a contradiction argument.
Consider the MAS with antagonistic interactions
\begin{equation}\label{MAS_3}
\dot{x}_i(t)=\sum^N_{j=1}|a^s_{ij}|\big(\text{sign}(a^s_{ij})\cdot x_j(t)-x_i(t)\big),
\end{equation}
in which the signed graph $\mathcal{G}_s=\{\mathcal{V}_s,\mathcal{E}_s,\mathcal{A}_s=[a^s_{ij}]\}$,
and suppose the signed graph $\mathcal{G}_s$ is strongly connected and structurally balanced. By the bipartite consensus theorem (Theorem 2 in \cite{altafini2012consensus}), it shows that the system \eqref{MAS_3} reaches bipartite consensus but not consensus.
Let $\mathcal{A}'=|\mathcal{A}_s|=[|a^s_{ij}|]$ and $\mathcal{G}'=\{\mathcal{V}_s,\mathcal{E}_s,\mathcal{A}'\}$, i.e., $\mathcal{G}'$ is a strongly connected graph with only cooperative interactions.
Assume the MAS \eqref{MAS_1} is under the graph $\mathcal{G}'$, and for all $i\in\mathbf{N},\,j\in\mathcal{N}_i$, $f_{ji}\big(x_j(t)\big)=\text{sgn}(a^s_{ij})\cdot x_j(t)$, which implies that $k_1k_2>1$. With the above assumption, the dynamics of MAS \eqref{MAS_1} is equivalent to the dynamics of MAS \eqref{MAS_3}. Therefore, it is turns out that MAS \eqref{MAS_1} cannot achieves consensus.
On the other hand, it is obviously that under the above assumption, the system \eqref{MAS_1} satisfies all conditions in Theorem \ref{th:consensus}. By Theorem \ref{th:consensus}, the states of agents will converge to a consensus value. Hence, we get a contradiction and the proof for the necessity statement of Theorem \ref{th:consensus} is proved.
\subsubsection{Sufficiency}
Applying Theorem \ref{th:distance=0}, we have:
\begin{equation}\label{eq:th_consensus_tending_x}
x(t)\to[\partial_m,\partial_M]^n,\enspace \text{as }t\to\infty.
\end{equation}
Notice that if $\partial_m=\partial_M$, then it turns out that
\begin{equation*}
\lim\limits_{t\to+\infty}x_i(t)=\partial_m=\partial_M=v^*,\quad \forall\, i\in\mathbf{N},
\end{equation*}
and therefore the sufficiency statement of Theorem \ref{th:consensus} is proved.
Hence, we continue our proof in the condition that $\partial_m<\partial_M$. Assume $\partial_m<\partial_M$ in the following.
\begin{step2}\label{step:th_consensus_error=0}
In this step, it shows that the states of agents tend to achieve consensus.
Denote $\theta_i(t)=\sum\limits_{j\in\mathcal{N}_i}a_{ij}\Big(f_{ji}\big(x_j(t)\big)-x_j(t)\Big)$, and the dynamics of MAS \eqref{MAS_1} can be rewritten as
\begin{equation}\label{eq:th_consensus_robust_system}
\frac{d}{dt}x_i(t)=\sum\limits_{j\in\mathcal{N}_i}a_{ij}\big(x_j(t)-x_i(t)\big)+\theta_i(t).
\end{equation}
Then, we use a contradiction argument to prove that $\lim\limits_{t\to\infty}\theta_i(t)=0$, $\forall\,i\in\mathbf{N}$.
Without loss of generality, assume that there exists a solution $x(t)$, such that $\forall\,i\in\mathbf{N}$, $\liminf\limits_{t\to\infty}\theta_i(t)\ge0$ and $\exists\,i^*\in\mathbf{N}$, $\liminf\limits_{t\to\infty}\theta_{i^*}(t)>0$.
$\mathcal{P}=\{\chi_1,\chi_2,\dots\}$ denotes the set of all limit points of $x_{i^*}(t)$ as $t\to\infty$, i.e., there are time sequences $\{t_n'\}$ with $\lim\limits_{n\to\infty}t_n'=\infty$ and $\lim\limits_{n\to\infty}x_{i^*}(t_n')=\chi_1$, and $\{t_n''\}$ with $\lim\limits_{n\to\infty}t_n''=\infty$ and $\lim\limits_{n\to\infty}x_{i^*}(t_n'')=\chi_2$, etc.
Since $\liminf\limits_{t\to\infty}\theta_{i^*}(t)>0$, it turns out that $\exists\,i':i^*\in\mathcal{N}_{i'}$, such that
\begin{equation*}
\sup\limits_{\chi_i\in\mathcal{P}}\Big\{\limsup\limits_{x\to\chi_i}\big(f_{i^*i'}(x)-x\big)\Big\}>0.
\end{equation*}
Furthermore, since $x_{i^*}(t)\to[\partial_m,\partial_M]$ as $t\to\infty$, it can be concluded that $\limsup\limits_{x\to\partial_m}f_{i^*i'}(x)>\partial_m$.
Therefore, there is a time sequence $\{t_n^*\}\to\infty$ with $\{x_{i^*}(t_n^*)\}\to\partial_m$ and $\lim\limits_{n\to\infty}f_{i^*i'}\big(x_{i^*}(t_n^*)\big)>\partial_m$. It is clear that $\partial_m\in\mathcal{P}$.
Since $f_{i^*i'}(x)$ is continuous in $[\partial_m,\partial_M]$ and have finite breaks in $(-\infty,\partial_m)$, there exists a time sequence $\{t_n^\sim\}\subseteq\{t_n^*\}$ and $f_{i^*i'}(x)$ is continuous on $x_{i^*}(t)$, $\forall\,t\in\{t_n^\sim\}$, which implies that $f_{i^*i'}\big(x_{i^*}(t)\big)$ is continuous on $\{t_n^\sim\}$.
Further, there exist $\epsilon>0$ and $T(\epsilon)$, such that $f_{i^*i'}\big(x_{i^*}(t)\big)>\partial_m$ and $f_{i^*i'}\big(x_{i^*}(t)\big)$ is continuous in $(t-\epsilon,t+\epsilon)$, for all $t>T(\epsilon)$ and $t\in\{t_n^\sim\}$.
Denote closed and connected interval $\mathbf{I}_k=[t_k-\frac{\epsilon}{2},t_k+\frac{\epsilon}{2}]$, where $t_k>T(\epsilon)$ and $t_k\in\{t_n^\sim\}$, $k\in\mathbb{N}^+$.
Since $\liminf\limits_{t\to\infty}\theta_i(t)\ge0$, $\forall\,i\in\mathbf{N}$, it can be concluded that $\liminf\limits_{t\to\infty}f_{ij}\big(x_i(t)\big)\ge\partial_m$, $\forall\,i,j\in\mathbf{N}$.
Notice that $\int_{\mathbf{I}_k}f_{i^*i'}\big(x_{i^*}(t)\big)>\partial_m$ and repeat the analysis of Theorem \ref{th:distance=0}, we can get that
\begin{equation*}
\liminf\limits_{t\to\infty}x_{i^*}(t)>\partial_m.
\end{equation*}
Then, it turns out that $\partial_m\notin\mathcal{P}$. Here, we find a contradiction.
Hence, we have proven that $\lim\limits_{t\to\infty}\theta_i(t)=0$, $\forall\,i\in\mathbf{N}$.
Applying Lemma \ref{le:rc}, we can get that
\begin{equation}\label{eq:th_consensus_error=0}
\lim\limits_{t\to+\infty}x_i(t)-x_j(t)=0,\quad\forall\, i,j\in\mathbf{N}.
\end{equation}
\end{step2}
\begin{step2}
In Step \ref{step:th_consensus_error=0}, it shows that the states of agents will converge to consensus. In this step, by the fact that $x(t)\to[\partial_m,\partial_M]^n$, we prove that for any $i\in\mathbf{N}$, $\lim\limits_{t\to\infty}x_i(t)=v^*$ and $v^*\in[\partial_m,\partial_M]$.
By \eqref{eq:th_consensus_tending_x}, it turns out that for any $\omega_1>0$, there exists a finite $T_1>0$, which holds the following inequation:
\begin{equation}\label{eq:omega_1}
\partial_m-\omega_1\le x_i(t)\le \partial_M+\omega_1,\quad\forall\, t\ge T_1,\,i\in\mathbf{N}.
\end{equation}
Without loss of generality, assume that $\frac{\partial_m+\partial_M}{2}\le x_k(T_1)\le \partial_M+\omega_1$, where $k$ is a fixed node.
Similarly, by \eqref{eq:th_consensus_error=0}, $\exists\,T_2>0$ which is finite, there holds
\begin{equation}\label{eq:omega_2}
|x_i(t)-x_k(t)|\le\omega_2,\quad\forall\,t\ge T_2,\,i\in\mathbf{N}.
\end{equation}
According to \eqref{eq:omega_1} and \eqref{eq:omega_2}, let $\omega_1$ and $\omega_2$ be sufficiently small, and we get that
\begin{equation*}
\partial_m<x_i(T_*)<\partial_M+\omega_1+\omega_2,\quad\forall\,i\in\mathbf{N}.
\end{equation*}
where $T_*>\max\{T_1,T_2\}$.
Depending on whether $\exists\,l\in\mathbf{N}$, $x_l(T_*)\ge \partial_M$ or not, there are two cases in the following proof.
\begin{enumerate}
\item $\exists\,l\in\mathbf{N}$, $x_l(T_*)\ge \partial_M$.
Repeating the analysis in Step \ref{th_distance=0_partA_step1} in the proof of Theorem \ref{th:distance=0}, we can get that
\begin{equation*}
\bar{Y}(t)=\max\limits_{i\in\mathbf{N}}\Big\{(1-k_2)\big(x_i(t)-\partial\big)\Big\}
\end{equation*}
is non-increasing for $t\ge T_*$. Since $1-k_2>0$, it turns out that $\max\limits_{i\in\mathbf{N}}x_i(t)$ is non-increasing for $t\ge T_*$.
\item $\forall\,i\in\mathbf{N}$, $\partial_m<x_i(T_*)< \partial_M$.
It is easy to get $f_{ij}\big(x_i(T_*)\big)=x_i(T_*)$, $\forall\,i,j\in\mathbf{N}$, and the system degenerates into a standard multiagent system at time $T_*$. Therefore, it is easy to know that $\max\limits_{i\in\mathbf{N}}x_i(t)$ is non-increasing for $t\ge T_*$.
\end{enumerate}
Combining the above analyses, we can conclude that $\max\limits_{i\in\mathbf{N}}x_i(t)$ is non-increasing for $t\ge T_*$. Furthermore, $\max\limits_{i\in\mathbf{N}}x_i(t)$ converges to a finite limit value (denote the value by $\overline{v}$). According to \eqref{eq:th_consensus_error=0}, $\min\limits_{i\in\mathbf{N}}x_i(t)$ must converge to the same limit value $\overline{v}$. Since $\min\limits_{i\in\mathbf{N}}x_i(t)\le x_j(t)\le \max\limits_{i\in\mathbf{N}}x_i(t)$, $\forall\, j\in\mathbf{N}$, it is trivial to get that
\begin{equation*}
\lim\limits_{t\to\infty}x_i(t)=\bar{v}=v^*,\quad \forall\, i\in\mathbf{N}.
\end{equation*}
Using \eqref{eq:th_consensus_tending_x}, we can conclude that $v^*\in[\partial_m,\partial_M]$.
Based on the above analysis, it is shown that all $x_i(t)$ will converge to a finite limit $v^*$ and $v^*\in[\partial_m,\partial_M]$.
\end{step2}
Hence, the proof of Theorem \ref{th:consensus} is completed.
\end{proof}
\section{Proof of Theorem \ref{th:exist_equilibrium}}
\begin{proof}
By Lemma \ref{le:positive_invariant}, it shows $[ \partial_m, \partial_M]^n$ is a positively invariant set.
By the Brouwer fixed point Theorem extended to dynamical systems \cite{basener2006brouwer}, we can conclude that there exists an equilibrium in $[ \partial_m, \partial_M]^n$.
Hence, along the system \eqref{MAS_1}, the existence of equilibria is proven.
By Theorem \ref{th:distance=0}, it shows that
\begin{equation*}
\lim\limits_{t\to\infty}\textnormal{distance}\big([ \partial_m, \partial_M]^n,x(t)\big)=0,
\end{equation*}
which implies that every equilibrium $\mathbf{e}\in[\partial_m,\partial_M]$.
We complete the proof of Theorem \ref{th:exist_equilibrium}.
\end{proof}
\section{Proof of Theorem \ref{th:converge_to_equilibrium}}
\begin{proof}
Denote
\begin{align*}
V(t)=&\max\limits_{i\in\mathbf{N}}\big\{(1-k_{e1})\big(\mathbf{e}_i-x_i(t)\big),(1-k_{e2})\big(x_i(t)-\mathbf{e}_i\big)\big\}\\
=&\max\limits_{i\in\mathbf{N}}\big\{(1-k_{e1})\big(-\varepsilon_i(t)\big),(1-k_{e2})\varepsilon_i(t)\big\},
\end{align*}
and clearly $V$ is Lipschitz continuous. In this step, we will prove that $V(t)$ is a non-increasing function.
Denote $\varepsilon_m(t)=\min\limits_{i\in\mathbf{N}}\varepsilon_i(t)$, $\varepsilon_M(t)=\max\limits_{i\in\mathbf{N}}\varepsilon_i(t)$.
By the structure of $V(t)$, there exits two cases:
\begin{enumerate}
\item\label{it:th_converge_to_equilibrium_case1} $V(t)=(1-k_{e2})\varepsilon_M(t)$;
\item\label{it:th_converge_to_equilibrium_case2} $V(t)=(1-k_{e1})\big(-\varepsilon_m(t)\big)$.
\end{enumerate}
We first consider the Case \ref{it:th_converge_to_equilibrium_case1}, which implies that $V(t)=\max\limits_{i\in\mathbf{N}}\big\{(1-k_{e2})\varepsilon_i(t)\big\}$ and $\varepsilon_M(t)>0$.
Denote $\mathcal{I}_e(t)=\big\{k:\varepsilon_k(t)=\max\limits_{i\in\mathbf{N}}\varepsilon_i(t)\big\}$. By Lemma \ref{le:der_to_max}, we can get that
\begin{align}
&d^+V(t)=d^+\max\limits_{i\in\mathbf{N}}\Big\{\big(1-k_{e2}\big)\varepsilon_i(t)\Big\}\notag\\
=&\max\limits_{i\in\mathcal{I}_e(t)}\Big\{\big(1-k_{e2}\big)\dot{\varepsilon_i}(t)\Big\}=\max\limits_{i\in\mathcal{I}_e(t)}\Big\{(1-k_{e2})\dot{x}_{i}(t)\Big\}\notag\\
=&\max\limits_{i\in\mathcal{I}_e(t)}\Big\{(1-k_{e2})\sum\limits_{j\in\mathcal{N}_{i}}a_{ij}\Big(f_{ji}\big(x_j(t)\big)-x_{i}(t)\Big)\Big\}\notag\\
\label{case_i}=&\max\limits_{i\in\mathcal{I}_e(t)}\Big\{(1-k_{e2})\sum\limits_{j\in\mathcal{N}_{i}}a_{ij}\Big(f_{ji}\big(\mathbf{e}_j+\varepsilon_j(t)\big)-\mathbf{e}_{i}-\varepsilon_{i}(t)\Big)\Big\}.
\end{align}
Furthermore, noticing that $\sum\limits_{j\in\mathcal{N}_{i}}a_{ij}\big(f_{ji}(\mathbf{e}_j)-\mathbf{e}_{i}\big)=0$,
we can conclude from \eqref{case_i} that
\begin{align}\label{eq:asymptotically stabel th_case_i}
\dot{\varepsilon_{i}}=\sum\limits_{j\in\mathcal{N}_{i}}a_{ij}\big(f_{ji}\big(\mathbf{e}_j+\varepsilon_j(t)\big)-f_{ji}(\mathbf{e}_j)-\varepsilon_{i}(t)\big).
\end{align}
Let $\varepsilon_{i'}=\varepsilon_M$, which implies $i'\in\mathcal{I}_e(t)$.
Applying Lemma \ref{le:x_M and x_m} on \eqref{eq:asymptotically stabel th_case_i}, we can get that
\begin{equation*}
\begin{cases}
\varepsilon_{i'} \ge \varepsilon_j\ge f_{ji'}\big(\mathbf{e}_j+\varepsilon_j\big)-f_{ji'}(\mathbf{e}_j) &\text{if } \varepsilon_j\ge0;\\
\varepsilon_{i'} \ge L_{e1}(\varepsilon_j)> f_{ji'}\big(\mathbf{e}_j+\varepsilon_j\big)-f_{ji'}(\mathbf{e}_j) &\text{if } \varepsilon_j<0.
\end{cases}
\end{equation*}
By the above two analysis, it can be concluded that $d^+V(t)\le0$ when $(1-k_{e2})\varepsilon_M(t)>(1-k_{e1})\big(-\varepsilon_m(t)\big)$.
The proof of Case \ref{it:th_converge_to_equilibrium_case2} is similar to the above proof, and hence are omitted here.
Combining the two cases, we have proved that $d^+V(t)\le0$ for all $t\ge t_0$.
Repeating the analysis of Theorem \ref{th:distance=0}, we have $\mathbf{x}(t)\to\mathbf{e}$, as $t\to\infty$.
Since the equilibrium $\mathbf{e}$ is asymptotically stable and the initial states $\mathbf{x}(t_0)$ can be arbitrary, $\mathbf{e}$ is a unique equilibrium.
The proof is completed.
\end{proof}
\section{Proofs of Theorem \ref{th:single equilibrium} and Corollary \ref{co:single equilibrium}}
\subsection{Proof of Theorem \ref{th:single equilibrium}}
\begin{proof}
We first prove the existence of equilibria.
By conditions 2 and 3, we can get that for any $j\in\mathbf{N},\,i\in\mathcal{N}_j$ and $\omega\ne0$,
\begin{equation}\label{eq:f_ij is continuous}
-1<k^*\le\frac{f_{ij}(x+\omega)-f_{ij}(x)}{\omega}\le1,\quad x\in\mathbb{R},
\end{equation}
which implies that $f_{ij}$ is a continuous function. For simplicity, let $k^*\in(-1,0)$.
By \eqref{eq:f_ij is continuous}, it can be concluded that there exists at least one intersection between functions $f_{ij}$ and $f(x)=x$, i.e., $\Theta_{ij}\ne\emptyset$. Denote $X_M=\max\big\{x:x\in\bigcup\Theta_{ij}\big\}$ and $X_m=\min\big\{x:x\in\bigcup\Theta_{ij}\big\}$.
Since $\Theta_{ij}\ne\emptyset$, let $x_{ij}^*\in\Theta_{ij}$, i.e., $f_{ij}(x_{ij}^*)=x_{ij}^*$. By \eqref{eq:f_ij is continuous}, it shows that
\begin{equation*}
\left\{
\begin{aligned}
f_{ij}(x_{ij}^*+\omega)\le f_{ij}(x_{ij}^*)+\omega=x_{ij}^*+\omega,\quad \omega>0;\\
f_{ij}(x_{ij}^*+\omega)\ge f_{ij}(x_{ij}^*)+\omega=x_{ij}^*+\omega,\quad \omega<0,
\end{aligned}
\right.
\end{equation*}
which implies that
\begin{equation}\label{eq:f_ij range1}
\left\{
\begin{aligned}
f_{ij}(x)\le x,\quad x\ge X_M;\\
f_{ij}(x)\ge x,\quad x\le X_m.
\end{aligned}
\right.
\end{equation}
There are two parallel lines with slope $k^*\in(-1,0)$:
\begin{equation*}
\left\{
\begin{aligned}
L_m(x)&=k^*x+(1-k^*)X_m,\\
L_M(x)&=k^*x+(1-k^*)X_M,
\end{aligned}
\right.
\end{equation*}
and it can be concluded that
\begin{equation*}
\left\{
\begin{aligned}
f_{ij}(x)\le L_M(x),\quad x\le X_M;\\
f_{ij}(x)\ge L_m(x),\quad x\ge X_m.
\end{aligned}
\right.
\end{equation*}
Let $L^*(x)=-x+X_M+X_m$ with slope $k=-1$.
Since $-1<k^*<0$, it is easy to know that $L^*$ and the parallel lines $L_M,L_m$ exist two intersections $(y_m,y_M)$ and $(y_M,y_m)$ where $y_M\ge y_m$, which implies that $L_M(y_m)=y_M$ and $L_m(y_M)=y_m$. And it is easy to know that $y_m<X_m\le X_M<y_M$.
Then, it can be concluded that
\begin{equation}\label{eq:f_ij range2}
\left\{
\begin{aligned}
f_{ij}(x)\le L_M(x)\le y_M,\quad y_m\le x\le X_M;\\
f_{ij}(x)\ge L_m(x)\ge y_m,\quad X_m\le x\le y_M.
\end{aligned}
\right.
\end{equation}
Combine \eqref{eq:f_ij range1} and \eqref{eq:f_ij range2}, we can get that
\begin{equation*}
y_m\le f_{ij}(x)\le y_M,\quad y_m\le x\le y_M.
\end{equation*}
By Theorem \ref{th:exist_equilibrium}, it shows that the system \eqref{MAS_1} has at least one equilibrium.
Then, we will show that the MAS \eqref{MAS_1} has only one equilibrium and this equilibrium is asymptotically stable.
Assume one of equilibria is $\mathbf{e^*}=\{\mathbf{e}_1^*,\dots,\mathbf{e}_n^*\}^T$, and denote the error between the state $x(t)$ and equilibrium $\mathbf{e}$ by $\varepsilon_i(t)^*=x_i(t)-\mathbf{e}_i^*$. By \eqref{eq:f_ij is continuous}, it turns out that for all $j\in\mathbf{N},\,i\in\mathcal{N}_j$,
\begin{align*}
\varepsilon_i^*\le f_{ij}(\mathbf{e}_i+\varepsilon_i^*)-f_{ij}(\mathbf{e}_i)\le k^*\varepsilon_i^*,\quad\varepsilon_i^*<0;\\
k^*\varepsilon_i^*\le f_{ij}(\mathbf{e}_i+\varepsilon_i^*)-f_{ij}(\mathbf{e}_i)\le \varepsilon_i^*,\quad\varepsilon_i^*>0.
\end{align*}
Since $k^*\in(-1,0)$, it shows that $k^*k^*<1$. Hence, the Condition 3 of Theorem \ref{th:converge_to_equilibrium} holds.
Because $\bigcap\limits_{(v_j,v_i)\in \mathcal{E}}\Theta_{ij}=\emptyset$ and for any $j\in\mathbf{N},\,i\in\mathcal{N}_j$ and $\omega\ne0$,
\begin{equation*}
-1<\frac{f_{ij}(x+\omega)-f_{ij}(x)}{\omega}<1,\quad x\notin\Theta_{ij},
\end{equation*}
we can conclude that for any $\varepsilon^*\ne0$, there exist $j\in\mathbf{N},i\in\mathcal{N}_j$ and $\delta>0$, such that
\begin{equation*}
f_{ij}(\mathbf{e}_i+\varepsilon')-f_{ij}(\mathbf{e}_i)\ne \varepsilon',\quad \varepsilon'\in(\varepsilon^*-\delta,\varepsilon^*+\delta).
\end{equation*}
Hence, the Condition 4 of Theorem \ref{th:converge_to_equilibrium} holds.
Apply Theorem \ref{th:converge_to_equilibrium}, it shows that the equilibrium $\mathbf{e}$ is a unique, asymptotically stable equilibrium.
\end{proof}
\subsection{Proof of Corollary \ref{co:single equilibrium}}
\begin{proof}
Because $f_{ij}(x_i)=k_{ij}(x_i)x_i+m_{ij}(x_i)$ is a continuous and piecewise linear function with its slopes $k_{ij}\in(-1,1]$ and $m_{ij}=0$ when $k_{ij}=1$, it turns out that for any $\omega\ne0$,
\begin{equation*}
-1<\frac{f_{ij}(x+\omega)-f_{ij}(x)}{\omega}<1,\quad x\notin\Theta_{ij}.
\end{equation*}
Apply Theorem \ref{th:single equilibrium}, this corollary is proved.
\end{proof}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Comets are relatively well-preserved remnant building blocks of our planets. Their interiors may provide us with clues about planetesimal formation and the composition of the outer solar nebula. One of the key quantities relevant in this context is the relative abundance of refractories and volatiles inside the cometary nucleus, often referred to as the refractory-to-ice (mass) ratio.
This ratio cannot be measured directly with current spacecraft or remote observation techniques. An alternative is to determine it indirectly by measuring the dust-to-gas (mass) ratio of the material that was released from the nucleus into interplanetary space. This technique however comes with caveats. For one, estimating the lost dust mass relies on models that require knowledge of or assumptions about the dust size distribution and either optical properties (for remote sensing data) or spatial distribution (for in situ data).
Then, translating the dust-to-gas to the refractory-to-ice ratio is not straightforward. Part of the refractory material is contained in blocks that are too heavy to be accelerated past the comet's escape speed, and so they never leave the nucleus or fall back onto its surface \citep{choukroun-altwegg2020}. The blocks' volatiles on the other hand may have escaped entirely. Based on the coma dust-to-gas ratio, the refractory-to-ice ratio would thus be underestimated. Meanwhile, investigation of the surface would tend to overestimate it because of the refractory deposits. Such deposits may also quench the nucleus' outgassing \citep[e.g.,][]{gundlach-blum2016}, while ejected material may be outgassing too \citep{reach-vaubaillion2009}, making the translation to the refractory-to-ice ratio more complicated. Both issues would be better understood with a firmer grasp of the material's dynamics.
The European Space Agency's Rosetta mission to comet \object{67P/Churyumov-Gerasimenko} has revealed that fall-back of refractory material is a common phenomenon, as wide parts of \object{67P}'s surface are uniformly covered in loose material \citep{thomas-davidsson2015}. Images obtained during the final descents of both the lander Philae \citep{mottola-arnold2015, pajola-mottola2016} and the Rosetta orbiter \citep{pajola-lucchetti2017} show the ground coated in a loose assembly of irregularly shaped blocks with typical sizes down to the centimeter-scale resolution limit of the images. Compared to the consolidated terrains thought to be more representative of the comet's "bedrock", from further out, these areas look relatively smooth.
Smooth terrains are predominantly found in the northern hemisphere \citep{thomas-davidsson2015}. This regional distribution of fall-back material is likely related to the asymmetric seasons on \object{67P}, with short ($\sim$\,1 year), hot perihelion summers in the southern hemisphere and long ($\sim$\,5.5 years), yet colder aphelion summers in the north \citep{keller-mottola2015_erosion}. \citet{pajola-lucchetti2017} have plausibly modeled the inter-region transport of fall-back material driven by the differences in local gas pressure due to varying solar irradiation. They showed that debris should be carried from regions of high gas pressure to regions where gas pressure is too low to keep the material afloat. \citet{lai-ip2016} have followed a more global approach studying the trajectories of dust particles embedded in a 3D Direct Simulation Monte Carlo model, and found that regional change in dust mantle thickness can be on the meter-scale.
For such models, it is mandatory to have good knowledge of the debris' source distribution, (i.e., its production rate as a function of time and surface region), and the constituents' initial velocities and accelerations. But accelerations can also provide information about the ice content of larger chunks, because sublimating ice may manifest as an acceleration component toward the antisolar direction \citep{kelley-lindler2013, kelley-lindler2015, 10.1093/mnras/stw2179}.
To learn more about these factors, \cite{10.1093/mnras/stw2179} manually tracked 238 decimeter-sized particles in an image sequence obtained by the Optical, Spectroscopic, and Infrared Remote Imaging System (OSIRIS) on January 6, 2016. Of the particles whose projected velocities were measured, at least 10\,\% were faster than the local escape speed of the nucleus; hence they likely reached interplanetary space, contributing to the comet's debris trail \citep[e.g.,][]{sykes-walker1992a}. \citet{keller-mottola2017} estimated that of the remaining 90\,\%, at least 20\,\% fell back to the surface within several hours, possibly accumulating in a regional layer of debris that still contains some water ice.
Other studies that looked for ejected debris in OSIRIS images also did so manually, or mostly by making use of the elongated particle trails in long-exposure images: \citet{bertiniSearchSatellitesComet2015} searched for satellites near \object{67P} using the SExtractor software \citep{1996A&AS..117..393B}, but found no unambiguous candidates. A moon (dubbed "churymoon") was later discovered visually by \citet{landru79MaybeLittleCoorbital2019}, and in the following tracked by \citet{marin-yaselidelaparraAnalysisParticleSurroundings2020} using TrackMate \citep{tinevezTrackMateOpenExtensible2017}. \citet{rotundiDustMeasurementsComa2015} and \citet{fulleEvolutionDustSize2016} manually identified $\sim$\,400 and 204 particles respectively using image differencing. \citet{davidssonOrbitalElementsMaterial2015} detected, manually tracked, and determined the orbital elements of four particles. \citet{ottPaDeParticleDetection2016} developed an algorithm to detect elongated particle trails based on Canny edge detection \citep{cannyComputationalApproachEdge1986} and Hough transformation \citep{houghMachineAnalysisBubble1959}, and used it to measure 262 particles \citep{drolshagenDistanceDeterminationMethod2017, ottDustMassDistribution2017}. \citet{guttlerCharacterizationDustAggregates2017} used the blurriness of defocussed particles--instead of the parallax effect used in most of the previous studies--to derive the properties of 109 particles semi-automatically. Finally, \citet{frattinPostperihelionPhotometryDust2017} developed an automated detection method based on line-shaped matching functions to detect elongated particle trails, and identified 1925 tracks (and again 1916 tracks in \citealp{frattinObservationalConstraintsDynamics2021}). No algorithm however has been developed for OSIRIS images to automatically track point-source-like particles.
During the post-perihelion phase of the mission, OSIRIS has regularly obtained image sequences like the one analyzed by \citet{10.1093/mnras/stw2179}. Many of them show fountains of debris that seem to stem from locally confined sources. Exploiting the sequences' very specific properties, we have created a tool that can automatically detect and track the motion of the point-source-like debris.
Instead of tracing particle tracks manually on a stacked image sequence as performed by \citet{10.1093/mnras/stw2179}, our algorithm first examines each image individually before recovering tracks from the gathered data. Because the particles scatter sunlight but are not spatially resolved, they appear as point sources in the images, which already distinguishes them from most other features. Nevertheless we further clean the images to improve their signal-to-noise ratio and then use the SExtractor software \citep{1996A&AS..117..393B} to detect them. Once located, their positions are passed on to the core of our project, the tracking algorithm. Here, we exploit the dataset's pair-nature to reconstruct the particle motions.
Our work presents a new approach in the large field of particle tracking \citep[for other disciplines see e.g.,][]{westerweelParticleImageVelocimetry2013, chenouardObjectiveComparisonParticle2014, ulmanObjectiveComparisonCelltracking2017, roseParticleTrackingNanoparticles2020}. In astronomy, tracking algorithms were developed to discover small Solar System bodies (SSSBs) in large-scale sky surveys, such as: Spacewatch \citep[MODP][]{1991AJ....101.1518R}; CFHT, SKADS, OSSOS \citep[MOP][]{10.1111/j.1365-2966.2004.07217.x, GLADMAN2009104, Bannister2016}; Pan-STARRS, LSST \citep[MOPS][]{KUBICA2007151, Denneau2013, JONES2018181}; (NEO)WISE \citep[WMOPS][]{Mainzer_2011}; PTF \citep[PTF MOPS][]{10.1093/mnras/stt951}; and ZTF \citep[ZMODE][]{Masci2018}. Recently, also more general-purpose SSSB discovery engines have been developed, such as HelioLinC \citep{Holman2018}, THOR \citep{Moeyens2021}, or tracee \citep{ohsawaDevelopmentTrackletExtraction2021}. All these algorithms were however mostly designed for Earth-based observations of SSSBs, where the object density is low, the apparent speed small, and the motion near linear (as pointed out by \citealp{https://doi.org/10.1029/2019EA000843}). They are therefore not very well suited for the dense and more dynamic dust environment in the coma of \object{67P}.
Much closer related to our project are the methods for particle tracking around asteroid \object{(101955) Bennu}, which were developed in parallel to our own algorithm and recently published by the OSIRIS-REx team (see \citealp{https://doi.org/10.1029/2020JE006549} for an overview of the special issue). Following the discovery of \object{Bennu}'s activity \citep{theosiris-rexteamOperationalEnvironmentRotational2019, doi:10.1126/science.aay3544}, \citet{https://doi.org/10.1029/2019EA000843} developed a dedicated algorithm to detect and track the ejected material. \citet{https://doi.org/10.1029/2019EA000938}, \citet{https://doi.org/10.1029/2019EA000937}, and \citet{https://doi.org/10.1029/2019JE006363} then estimated the trajectories and orbits of the identified particles and traced them back onto the surface of \object{Bennu} to reconstruct the ejection events.
In this paper we describe our methodology in detail and apply it to the same image sequence analyzed by \citet{10.1093/mnras/stw2179}. Because our algorithm can detect much fainter tracks, we find more than three times as many tracks than the manual procedure and hence significantly improve the statistics.
\begin{figure}[!b]
\centering
\includegraphics[width=\linewidth]{./figures/data/timeline}
\caption{Timeline of image sequence STP090. The sequence was constructed from the two subsequences "JETS\_MOVIE" (20 images, principal sequence) and "DUST\_JET" (24 images, extended sequence). The whole sequence spans almost two hours. Due to the alternating time intervals between recordings, the images come in pairs.}
\label{fig:timeline}
\end{figure}
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/data/source}
\caption{One of the source images of STP090. It showcases the appearance of dust particles as small point sources, as well as the bright surface of the irradiated nucleus and its radiant features. Full image on the left, close-up of central region on the right (contrast enhanced for visibility).}
\label{fig:source}
\end{figure*}
The data and image processing are described in Sect.~\ref{sec:data}. The tracking algorithm is described in Sect.~\ref{sec:tracking}, and the parameter optimization in Sect.~\ref{sec:optimization}. In Sect.~\ref{sec:results}, we evaluate the performance of the algorithm and present initial scientific results. Our findings are summarized in Sect.~\ref{sec:conclusions}. The entire project is written in the Python programming language version 3.7.7 \citep{van1995python} and we are happy to provide access to the code on request.
\section{Data}
\label{sec:data}
The source material for our tracking algorithm are the previously discussed image sequences recorded by OSIRIS' Narrow Angle Camera (NAC) on board the Rosetta spacecraft (for NAC specifications see Table~\ref{tab:mission_details} and \citealp{2007SSRv..128..433K}). The sequences typically show (parts of) \object{67P}'s nucleus and coma from $\sim$\,20--400\,km distance, and share a characteristic that is essential for our tracking algorithm: their images were recorded in pairs with the time interval between pairs being much longer than the intra-pair cadence (see Sect.~\ref{sect:pair_tracking}).
We refer to the image sequence that we exemplarily analyze in the following as STP090. It was obtained with NAC on January 6, 2016, starting from UT 07:01:03 when Rosetta was $\sim$\,86\,km away from the nucleus, and \object{67P} was at a heliocentric distance of $\sim$\,2.06\,AU post-perihelion. We constructed STP090 from two sub-sequences, a short and a long one. The short sequence (OSIRIS activity tag "JETS\_MOVIE") consists of 20 images and covers roughly six minutes, while the long sequence (OSIRIS activity tag "DUST\_JET") contains 24 images and spans almost two hours, starting roughly two minutes before the short one (see Fig.~\ref{fig:timeline}). While the exposure time for the short sequence is constant at 0.24\,s, it alternates between 0.24 and 6\,s for the long one. In the following we refer to the short and long one as the principal and extended sequences respectively, a distinction that becomes clearer in Sect.~\ref{sect:extended_tracking}. The relevant mission details are summarized in Table~\ref{tab:mission_details}.
\ctable[
caption = Mission details for sequence STP090.,
label = tab:mission_details,
pos = b,
width = \linewidth,
notespar
]{lc}{
}{ \FL
Date of recording & January 6, 2016 \NN
Time of recording & UT 07:01:03--UT 08:51:15 \NN
Total duration & 1\,h 50\,min 11\,s \NN
Heliocentric distance & $\sim 2.06$\,AU \NN
Nucleocentric distance & $\sim 86$\,km \ML
Camera & OSIRIS NAC \NN
Field of view (FOV) & $2.208\degree \times 2.208\degree$ \NN
CCD resolution & $2048 \times 2048$\,px \hspace{6pt}|\hspace{6pt} $\sim 3.3 \times 3.3$\,km \NN
Pixel resolution & $18.6 \times 18.6$\,\textmu rd \hspace{6pt}|\hspace{6pt} $\sim 1.6 \times 1.6$\,m \NN
Filter (NAC F22) & center: 649.2\,nm, bandwidth: 85\,nm \LL
}
We use OSIRIS images of calibration level 3E (Committee on Data Management, Archiving, and Computing, CODMAC, level 4), which includes solar and in-field stray-light correction, radiometric calibration and geometric distortion correction\footnote{The data are available at the Planetary Science Archive of the European Space Agency under \url{https://www.cosmos.esa.int/web/psa/rosetta}.} \citep{2015A&A...583A..46T}. The pixel values of this level are provided in radiance units (W/m$^2$/sr/nm).
At this "raw" stage, it is already possible to see some of the brighter particles (see Fig.~\ref{fig:source}). To also detect fainter particles however and track their motion, the images are first cleaned before the point source coordinates are extracted (see Sect.~\ref{sec:detection}).
\subsection{Image cleaning}
\label{sec:cleaning}
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/cleaning/cleaning}
\caption{Illustration of the cleaning pipeline: (1) the unaltered source image (OSIRIS level 3E); (2) the masked-out nucleus; (3) the estimated background level (a) and the corresponding RMS map of the background-subtracted image (b); (4) the background- and nucleus-subtracted image predominated by dust particles. All images show the same central region indicated in Fig.~\ref{fig:source} and are brightness-inverted for better reading.}
\label{fig:cleaning}
\end{figure}
To optimize particle detection, we aim to minimize signals not associated with point sources. Of those, we identify three types: (1) ambient background noise that stems from the diffuse coma and bright, roughly cone-shaped dust streams radiating from the nucleus (in the following called radiant features, Fig.~\ref{fig:source}); (2) prints of cosmic ray hits; and (3) the nucleus itself. While cosmic ray hits may confuse the point source detector occasionally, we found that due to their small number, they do not significantly affect the tracking results. The background noise and the nucleus however need to be removed.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/data/stack}
\caption{Stacked image of sequence STP090. It was created by selecting the maximum value for each pixel across the image sequence. Full (brightness-inverted) image on the left, close-up of central region on the right. The stacked image is only used to check the tracking results and identify sidereal objects.}
\label{fig:stack}
\end{figure*}
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/detection/detections}
\caption{Sample detection set (red ellipses) from one of the processed images of sequence STP090. Full image on the left, close-up of central region on the right.}
\label{fig:detections}
\end{figure*}
The diffuse coma signal is determined by the background estimator from the library for Source Extraction and Photometry \citep[SEP,][]{Barbary2016, 1996A&AS..117..393B, 1987PASP...99..191S}. It subdivides an image into a grid of rectangular sections, calculates the background locally for each (with the help of iterative $\kappa$-$\sigma$-clipping and mode estimation), and merges the resulting background patches smoothly back together (via natural bicubic spline interpolation) to form the global background map. This approach has the advantage that it can account for medium-scale changes in the background level--such as radiant features--and is therefore generally well-suited for our datasets (Fig.~\ref{fig:cleaning}).
The bright nucleus on the other hand poses an issue for the background estimation. For sections at its limb that include both nucleus and coma the background level would be overestimated. To prevent this, we mask out the entire nucleus using its approximate shape retrieved from the OSIRIS level 4S (CODMAC level 5) georeferencing layers, and subsequently refine the mask with the help of edge-detection algorithms. The shape is then passed on to the background estimator which ignores the masked area during processing. Because particle detection is not possible in front of the illuminated surface, by removing the nucleus we only lose information of particles that appear in front shadowed regions.
The background estimation also renders a root mean square (RMS) map. It is calculated in a similar fashion as the background signal, where the RMS values are first determined locally, before being smoothed out to form the global map. Since the RMS map is calculated from the background-subtracted image, it gives us an idea about the remaining random noise. This information is used during the point source detection.
With the nucleus and background removed, the images are predominated by the signal of dust particles. We call the remaining area that still contains data the dust field.
Lastly, the processed images are stacked by selecting the maximum value that each pixel assumed over the sequence (see Fig.~\ref{fig:stack}). Unlike \cite{10.1093/mnras/stw2179} however, we do not use this stacked image to track the particles, but instead only as a visual aid to check the tracking results and identify sidereal objects.
\subsection{Point source detection}
\label{sec:detection}
The particles are detected with the help of the SEP software \citep{Barbary2016, 1996A&AS..117..393B}. It employs a thresholding approach based on \citet{1980CompJ..23..262L}'s one-pass algorithm that can be used to identify point sources. Only pixels whose values are above the local RMS level (see Fig.~\ref{fig:cleaning}) multiplied by some user-defined detection threshold are considered during the detection process.
The algorithm then extracts sources based on the number of contiguous pixels, which are later-on deblended (using their brightness topology, \citealp{1990MNRAS.247..311B}) to separate neighboring point sources that have been extracted together. The resulting dataset can additionally be "cleaned", meaning it is checked whether each source would have also been detected without its neighbors being present. In the following we call identified sources detections, and the entirety of all sources detected in a single image a detection set. Figure~\ref{fig:detections} shows a sample of such a set.
\section{Tracking algorithm}
\label{sec:tracking}
In the following we assume an image sequence comprised of $N$ images recorded at times $t_n$ ($n \in \{0, 1, ..., N-1\}$), thus containing $N/2$ image pairs. We start by briefly defining key concepts:
\begin{description}
\item[Track and candidate track:] Track refers to a collection of detections that are all of the same object and thus depict the object's path through the recorded scene. Candidate track refers to any collection of detections, independent of whether or not they belong to the same object. They are only accepted as tracks once they pass a quality check.
\item[Pursuit and tracking run:] Pursuit refers to the tracking of a single object throughout a dataset, while tracking run describes the exhaustive analysis of an entire dataset, encompassing every possible pursuit for a fixed set of tracking parameters.
\item[Tracking parameters:] Tracking parameters are the parameters that govern the execution of a tracking run and each of its pursuits. They influence for example which detections and detection pairs are considered during a pursuit and define how many of them candidate tracks must contain to be accepted.
\end{description}
\subsection{Pair-tracking}
\label{sect:pair_tracking}
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/tracking/data_reduction}
\caption{Illustration of the relation between image sequence, detection sets, and pair groups. Together, the pair groups comprise the pool of available pairs.}
\label{fig:data_reduction}
\end{figure}
A central aspect of our tracking algorithm is the exploit of the image sequences' pair-nature. We assume that during the time interval of an image pair, the particles only travel a short distance (typically no more than a few pixels). This allows us to pair neighboring detections, one from each of the images, and analyze them as a unit: close detections likely belong to the same object. Consequently, our tracking algorithm predominantly operates pairwise, reverting to search for single detections only when there are no suitable pairs. We call this process pair-tracking.
To create the detection pairs, our algorithm iterates over the detection sets of each image pair. For every detection in the first set, the algorithm looks for detections in the second set within a predefined search radius we call the initial search radius. Each secondary detection found this way then forms a detection pair with the primary detection (see Fig.~\hyperref[fig:static]{\ref*{fig:static}a}). Thus, any detection can be part of multiple detection pairs.
During a tracking run, we treat each pair as a singular unit with its own location in time \mbox{$t_{i, i+1} = (t_i+t_{i+1})/2$} and space \mbox{$\vec{p}_\text{pair} = (\vec{p}_{\text{det}, i}+\vec{p}_{\text{det}, i+1})/2$}, where $\vec{p}_{\text{det}, i}$ and $t_i$ are the positions and recording times of its two detections, \mbox{$i \in \{0, 2, 4, ..., N-2\}$} (for simplicity we refer to any time-step as $t_i$ following Eq.~\ref{eq:velocity}). What discriminates pairs from single detections however, is that we additionally attribute each pair with a velocity vector:
\begin{equation}
\label{eq:velocity} \vec{v}_{\text{pair}} = \frac{\vec{p}_{\text{det}, i+1} - \vec{p}_{\text{det}, i}}{t_{i+1} - t_i}.
\end{equation}
Once all the pairs are created and their properties computed, they make up the initial pool of available pairs. Because pairs that stem from the same two images all share the same point in time, the pool of available pairs is quantized into $N/2$ pair groups (see Fig.~\ref{fig:data_reduction}).
Each pair from this pool is considered as part of a candidate track at least once: either to establish a new one, or to become part of another. We allow detections and pairs to be associated with only one track however, thus as soon as a candidate track is accepted, its components (and any other unrelated pairs its detections were part of) become unavailable throughout the rest of the tracking run.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/tracking/tracking_run}
\caption{Flowchart and pseudo-code illustrating the structure of an entire tracking run. The algorithm iterates over the pool of available pairs (I), using each pair as the starting point for a new candidate track (II, III). If a candidate track is accepted (IV), its pairs and detections are removed from the respective sources (as well as any other available pair that shares detections with the track) and cannot be used to create future tracks.}
\label{fig:tracking_run}
\end{figure*}
\begin{figure*}[!ht]
\centering
\includegraphics[width=\textwidth]{./figures/tracking/pursuit}
\caption{Typical track demonstrating the pursuit process. The algorithm operates pair-step-wise, going from one pair group to the next. In this case, it starts with a pair from the second group. Because the origin lies in the first half of the image sequence, the candidate track is first tracked forward, then backward in time (indicated by the circled numbers). If no suitable pair is found at a given step, the algorithm switches to searching for suitable single detections (single-step) instead, starting with the detection set that is closer in time to the previous step. The algorithm only searches the second set as well if it finds no suitable detection in the first. Afterwards, the pair-tracking continues. The track's color gradient from red to yellow indicates the direction of time (and therefore the object's motion through space). While the red ellipses mark the detections that make up the track, the black, dashed circles indicate where detections are missing. The background image is part of the stacked image similar to Fig.~\ref{fig:stack}.}
\label{fig:pursuit}
\end{figure*}
Accordingly, a complete tracking run consists of a series of individual pursuits of one candidate track at a time (see Fig.~\ref{fig:tracking_run}). The algorithm walks forward in time through the pool of available pairs, and starts a new candidate track with each (though successful pursuits lower the number of remaining available pairs). We call this initial pair of a candidate track its origin, and track from it forward and backward in time.
Each candidate track is pursued from one pair group to the next, an operation quantized in what we call pair-steps (see Fig.~\ref{fig:pursuit}). If at any pair-step no suitable pair could be found, the algorithm switches for that instance to single-steps, looking for a single suitable detection instead. Afterwards, the algorithm switches back to pair-tracking, searching the next pair group in line.
In this manner, each candidate track is pursued throughout the whole dataset, independent of how many pairs or detections may have been missed along the way. The pursuit only stops prematurely if, after no suitable pair or detection were found at a given step, it is determined that the center of the search area lies outside the dust field. The pursuit at the other end of the candidate track remains unaffected by this. Any pursuit concludes by checking if the candidate track qualifies as a track (see below). Only then does the algorithm move on to pursue a new candidate track.
\subsection{Tracking parameters}
Whether a candidate track qualifies as a track and which criteria single detections and pairs need to satisfy to become part of one is governed by a set of tracking parameters. While some of them are static and do not change during the whole tracking run, others are dynamic and adjust as the candidate track in pursuit evolves. The static tracking parameters only play a role at the beginning and end of a pursuit. They are (see Fig.~\ref{fig:static}):
\begin{description}
\item[The initial search radius $R_\text{init}$,] which is used to create the detection pairs (Fig.~\hyperref[fig:static]{\ref*{fig:static}a}). It limits the maximum velocity any pair can have and sets the stage for individual pursuits of candidate tracks, as the properties of the origin are decisive in what the algorithm is looking for.
\item[The residual offset $R_\text{off}$,] which is the final tracking parameter that affects the candidate track itself (Fig.~\hyperref[fig:static]{\ref*{fig:static}b}). Once the pursuit of a candidate track is over, a final curve is fitted to its detections, and the distances $d_\text{off}$ between them ($\vec{p}_{\text{det}, i}$) and the locations where they should lie according to the fit ($\vec{p}_\text{fit} (t_i)$) are calculated. Any detection where $d_\text{off} > R_\text{off}$ is removed from the candidate track.
\item[The] minimum number of detections $N_\text{det}$ and detection pairs $N_\text{pair}$, which define the acceptance thresholds (Fig.~\hyperref[fig:static]{\ref*{fig:static}c}). After the residual offsets have been checked, any candidate track must have at least that many detections and detection pairs to be accepted.
\end{description}
Even though exceeding the acceptance thresholds does not guarantee that a group of detections all belong to the same object, it does increase our confidence in the tracking results. The more detections a candidate track contains, the less likely it is that they are unrelated (i.e., stem from different particles or sources). Thus, for the remainder of the tracking run, we treat any candidate track that passes these thresholds as a valid track.
Avoiding to add unrelated detections is also helped by the dynamic tracking parameters: They repeatedly adjust to the properties of the candidate track during its pursuit and therefore narrow the track-specific parameter space that the algorithm searches for suitable detections and detection pairs. With the exception of the first pair-step, where the properties of the origin are used, these parameters depend on the properties of a curve that is fitted to the candidate track at every step. We refer to the detections or detection pairs that satisfy the criteria derived from these parameters as candidate pairs or detections. The dynamic tracking parameters are (see Fig.~\ref{fig:dynamic}):
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/tracking/static_parameters}
\caption{Diagrams illustrating how the static tracking parameters operate: a) the initial search radius $R_\text{init}$ around a primary detection (orange circle), used to create pairs with secondary detections (violet circles); b) the residual offset $R_\text{off}$, which defines the maximum distance $d_\text{off}$ any detection $\vec{p}_{\text{det}, i}$ (red circle) can have from the corresponding location $\vec{p}_\text{fit} (t_i)$ of the curve (black line) fitted to the candidate track (orange path); and c) the minimum number of detections $N_\text{det}$ and detection pairs $N_\text{pair}$, which any candidate track ($n_\text{det}$, $n_\text{pair}$) must have to be accepted as a track. The candidate track is shown as a gradient line from red to yellow, the present and missing detections as red, and gray dashed circles, respectively.}
\label{fig:static}
\end{figure}
\begin{description}
\item[The dynamic search radius $R_\text{dyn}$,] which defines the area within which the algorithm looks for candidate pairs and detections (Fig.~\hyperref[fig:dynamic]{\ref*{fig:dynamic}a}). The size of the area depends on the distance $d$: the spatial distance between the candidate track's pair $\vec{p}_\text{pair}$ that is closest in time to the investigated step (in case of single-tracking the closest detection), and the predicted position $\vec{p}_\text{fit} (t_i)$ of where the next pair (or detection) is expected to lie according to the curve fitted to the candidate track. We chose the relation between the search radius and distance to be that of an arctangent, whose free parameters $R_\text{min}$, $R_\text{max}$, $\hat{d}$ (shift), and $\bar{d}$ (stretch) we can control at the beginning of the tracking run:
\begin{equation}
\label{eq:dmax} R_\text{dyn}(d) := R_\text{min} + \left(\frac{R_\text{max} - R_\text{min}}{2}\right)\left(1 + \frac{2}{\pi}\arctan\left(\frac{d - \hat{d}}{\bar{d}}\right)\right).
\end{equation}
Increasing with distance, this function still allows for meaningful search radii at the smallest distances, while being capped at larger distances, as to not include candidate pairs or detections that are too far out.
\item[The (maximum) offset angle $\Omega$,] which defines a circular sector within which the candidate pairs or detections must lie (Fig.~\hyperref[fig:dynamic]{\ref*{fig:dynamic}b}). It measures from the vector that points in the same direction as the fitted curve at the time of the investigated step ($\vec{v}_\text{fit} (t_i)$), and, like the dynamic search radius, it depends on the distance $d$. The sector originates from the center of the candidate track's closest pair (or closest detection in case of single-tracking), and opens up in tracking direction. Its arc spans twice the offset angle. The relation between $\Omega$ and $d$ is otherwise identical to that of Eq.~\ref{eq:dmax}, but with the arctangent flipped:
\begin{align}
\label{eq:amax} \nonumber &\Omega(d) := \\ &\Omega_\text{min} + \left(\frac{\Omega_\text{max} - \Omega_\text{min}}{2}\right)\left(1 - \frac{2}{\pi}\arctan\left(\frac{d - \hat{d}_\Omega}{\bar{d}_\Omega}\right)\right),
\end{align}
where we again have control over the free parameters $\Omega_\text{min}$, $\Omega_\text{max}$, $\hat{d}_\Omega$, and $\bar{d}_\Omega$. In this case however, we allow the largest deviations for the smallest distances, since even small positional changes perpendicular to the candidate track can mean large angular ones. The opposite is true for large distances.
\item[The (maximum) inclination angle $I$,] which is also measured with respect to the candidate track's direction at the investigated time-step ($\vec{v}_\text{fit} (t_i)$, Fig.~\hyperref[fig:dynamic]{\ref*{fig:dynamic}c}). It shares the same value as $\Omega$, but instead limits the inclination that pairs can have toward the reference vector. Because single detections do not have an inclination, this parameter is only relevant during pair-tracking.
\item[The relative difference in speed $\Delta V$,] which restricts how much the speed of candidate pairs can stray from that of the fitted curve at the investigated time-step ($|\vec{v}_\text{fit} (t_i)|$, Fig.~\hyperref[fig:dynamic]{\ref*{fig:dynamic}d}). It is calculated as the relative deviation from $|\vec{v}_\text{fit} (t_i)|$ in percent, from a relation that has the same shape as Eq.~\ref{eq:amax}:
\begin{align}
\label{eq:vmax} \nonumber &\Delta V(|\vec{v}_\text{fit}|) := \\ &\Delta V_\text{min} + \left(\frac{\Delta V_\text{max} - \Delta V_\text{min}}{2}\right)\left(1 - \frac{2}{\pi}\arctan\left(\frac{|\vec{v}_\text{fit}| - \hat{v}}{\bar{v}}\right)\right),
\end{align}
where we also have control over the free parameters $\Delta V_\text{min}$, $\Delta V_\text{max}$, $\hat{v}$, and $\bar{v}$. Analogous to the offset and inclination angle, we allow the largest relative deviation for the smallest speeds, because in this regime, pixelization and uncertainties in the pointing of the camera and source detection can have a significant effect. For high speeds on the other hand, we only expect small deviations, for example due to a curved flight path. Because single detections cannot be assigned a velocity, this parameter is also only relevant during pair-tracking.
\end{description}
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/tracking/dynamic_parameters}
\caption{Diagrams illustrating how the dynamic tracking parameters operate (pairs that satisfy the respective criteria are shown in violet, the ones that do not in gray): a) the dynamic search radius $R_\text{dyn}$, which defines the area the algorithm searches for candidate pairs or detections. It depends on the distance $d$ between the candidate track's pair $\vec{p}_\text{pair}$ closest in time to the investigated step, and the predicted position $\vec{p}_\text{fit} (t_i)$ where the next pair is expected to lie according the curve (black, partly dashed line) fitted to the track (orange path). The relation between $R_\text{dyn}$ and $d$ is that of an arctangent (see Eq.~\ref{eq:dmax}) shown by the graph on the right. b) the (maximum) offset angle $\Omega$, which defines a circular sector within which candidate pairs or detections must lie. The sector originates from the candidate track's closest pair and opens up in the same direction as the fitted curve at the investigated time-step ($\vec{v}_\text{fit} (t_i)$, black arrow). The offset angle also depends on $d$ in the form of an arctangent, although reversed, as shown by the graph on the right (see Eq.~\ref{eq:amax}). c) the (maximum) inclination angle $I$, whose value is equal to that of $\Omega$. It defines the maximum inclination candidate pairs can have with respect to $\vec{v}_\text{fit} (t_i)$. d) the relative difference in speed $\Delta V$, which determines how much the speed of a candidate pair can relatively deviate from $|\vec{v}_\text{fit} (t_i)|$. The relation between $\Delta V$ and $|\vec{v}_\text{fit} (t_i)|$ is also that of a reversed arctangent as shown by the graph on the right (see Eq.~\ref{eq:vmax}).}
\label{fig:dynamic}
\end{figure}
For each candidate pair or detection that satisfies all criteria set up by the dynamic tracking parameters, we compute a match-factor $M$ as a proxy for the candidate's validity. The match-factor is used in two ways: to decide between different candidate pairs or detections, and to weigh the contribution of the selected one on the curve fitted to the candidate track:
\begin{equation}
\label{eq:pair_match} M_\text{cand} := 1 - \frac{1}{4} \left(\frac{r_\text{cand}}{R_\text{dyn}} + \frac{\omega_\text{cand} + I_\text{cand}}{\Omega} + \frac{\Delta v_\text{cand}}{\Delta V}\right),
\end{equation}
for pair-tracking, or
\begin{equation}
\label{eq:single_match} M_\text{cand} := 1 - \frac{1}{2} \left(\frac{r_\text{cand}}{R_\text{dyn}} + \frac{\omega_\text{cand}}{\Omega}\right),
\end{equation}
for single-tracking, where $r_\text{cand}$, $\omega_\text{cand}$, $I_\text{cand}$ and $\Delta v_\text{cand}$ are the dynamic parameter values of the candidate pair or detection, which are normalized by the respective maximum values as determined by Eqs.~\ref{eq:dmax}--\ref{eq:vmax}. We then choose the pair or detection with the highest match-factor to become part of the candidate track.
\subsection{Principal and extended tracking}
\label{sect:extended_tracking}
To address the different time-steps of the two sub-sequences (see Fig.~\ref{fig:timeline}), our tracking algorithm has two operating modes on the pursuit level: principal, and extended tracking. Due to the shorter intervals of the principal sequence, particle tracks are generally easier to identify during principal tracking--both visually and by the tracking algorithm. Thus, candidate tracks are only pursued during extended tracking if they passed the acceptance thresholds after principal tracking. For the same reason, any track becomes part of the final tracking results independent of how many detections were missed during this second stage.
Both modes have their own set of predefined tracking parameters. Extended tracking however has an additional parameter we call life. The lives of a track define how many detection pairs are allowed to be missed during extended tracking. If no pair and no single detection is found at a given step, then the life counter is reduced by 1. Lives also cannot be replenished: should the counter fall to zero, the pursuit is stopped. This prevents adding unrelated detections to a track, something that becomes increasingly more likely the further the search area is away from the established part of the track.
Once the algorithm checked the residual offset again after extended tracking, another control mechanism executes. Because detections may have been removed during the residual offset check, the extended part of the track is inspected for larger gaps (i.e., missing pairs). Should any individual gap or the sum of all gaps be larger than the granted extra lives, then all detections that come after the critical gap (i.e., the gap that let the sum of all gaps exceed the number of extra lives) are removed as well.
Finally, principal and extended tracking differ by the kind of curve that is fitted to the candidate track during its pursuit. Because the particles only travel relatively short distances during the principal sequence, we fit their tracks with straight lines. This is more robust than a parabola for example, since we found that the parabola's extra degree of freedom often causes the tracking algorithm to trail off in the wrong direction when the detections of a candidate track are not perfectly aligned. During the extended sequence however, we expect tracks to curve significantly because it covers a much longer time period. Thus at this stage, we fit parabolas to the tracks, which is also less likely to fail now that the tracks already consist of a considerable amount of detections.
The curve fitted to a candidate track to determine the residual offsets on the other hand is always a parabola. And no matter the circumstance under which a curve is fitted, the detections are always weighted by the match-factor (Eqs.~\ref{eq:pair_match} \& \ref{eq:single_match}) they were assigned when added to the track.
\subsection{Sidereal-motion-based attitude correction}
\begin{figure}[!b]
\centering
\includegraphics[width=\linewidth]{./figures/correction/correction}
\caption{Two sample tracks (1, 2) from the principal sequence, once without (a) and once with (b) the pointing correction applied (the image shown in the background is left unchanged). Tracks are shown as colored lines from red to yellow, detections indicated by red ellipses.}
\label{fig:correction}
\end{figure}
This process precedes any particle tracking, but as it uses the same tracking algorithm, it is described only now. While studying the stacked image of sequence STP090, we noticed a pointing fluctuation with a typical amplitude of a few pixels that occurred during the principal sequence and which can be observed in every track (see Fig.~\ref{fig:correction}). It compromises the tracking results in several ways: (1) A significant spread of detections from their expected positions can quickly lead the algorithm to go off trail. (2) To account for a higher variance in location, velocity, and orientation of candidate pairs and detections, we need to chose more lenient tracking parameters. Inevitably, this further increases the chances of adding unrelated detections and going off trail. (3) Drastic changes in velocity also translate into incorrectly computed accelerations. This makes extended tracking based on parabolas virtually impossible, as predictions over the long time intervals between image pairs require accurate accelerations; otherwise, the search areas are too far off, again leading the candidate track to go astray.
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/correction/stars}
\caption{Sidereal objects identified in the dust field of sequence STP090. On the left: results from searching the SIMBAD database (lines colored violet to aquamarine). On the right: sidereal tracks obtained from our tracking algorithm that matched some of those results (lines colored dark blue to green). The objects move from right to left.}
\label{fig:stars}
\end{figure*}
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/correction/star_track}
\caption{Sample of a sidereal track (top line from dark blue to green, detections indicated by red ellipses) and the expected motion of a sidereal object it was matched with (bottom line from violet to aquamarine, expected positions indicated by blue circles). The top panel shows the whole track, the bottom one a close-up of the first 24 detections, including the entire principal sequence. The pointing fluctuation is mainly acting along the direction of motion from right to left.}
\label{fig:star_track}
\end{figure}
\begin{figure*}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/correction/wobble}
\caption{Pointing fluctuation derived from the apparent motion of 21 sidereal objects in the dust field of sequence STP090 (see Eqs.~\ref{eq:wobble1}--\ref{eq:wobble3}). The reference point indicates which image and therefore which detections we used to calculate the relative distances. The black circles and dashed line mark the zero line (no pointing fluctuation). The filled and the open gray circles show the measured offsets from sidereal tracks: while the filled ones were used to calculate the mean values that represent the pointing fluctuation (orange circles with errorbars), the open ones are outliers that were excluded from the calculations, as they are assumed to result from unrelated detections.}
\label{fig:wobble}
\end{figure*}
Since the deviations are systematic, we attribute them to unexpected changes in spacecraft attitude. During STP090 (and other sequences like it), Rosetta's orbit around \object{67P} is noticeable. But to keep the camera's reference frame fixed to the comet's center of mass, the spacecraft's attitude was constantly adjusted. Sidereal objects therefore describe an apparent linear motion across the dust field. By comparing the pointing data of the image headers--which represent the commanded pointing--with the actual motion derived from tracks, we can thus reconstruct the pointing fluctuation.
To identify sidereal objects in a sequence, we query the SIMBAD Astronomical Database \citep{2000A&AS..143....9W} via the astroquery library \citep{2019AJ....157...98G}. Objects such as binaries that are spaced too close to each other to be distinguishable in the images are recorded only once. We then use gnomonic projection to transform the objects' equatorial coordinates back to image coordinates and generate the expected motions (see Fig.~\ref{fig:stars}).
Next, we run the tracking algorithm in a local area around each of the identified objects, visually compare the tracking results to the expected motions, and match them manually. Figure~\ref{fig:star_track} shows an example of such a track we call sidereal track, and its companion, the previously estimated motion. By choosing a reference image, calculating the relative distances of the commanded positions to that reference image, doing the same for the detections of the sidereal tracks, and subtracting the former distances from the latter, we can calculate the relative offsets induced by the pointing fluctuation:
\begin{align}
\label{eq:wobble1} \vec{\delta}_i &= (\vec{p}_{\text{det}, i} - \vec{p}_{\text{det, ref}}) - (\vec{p}_{\text{com}, i} - \vec{p}_{\text{com, ref}}),
\end{align}
where $\vec{\delta}_i$ is the relative offset of a sidereal track at the $i$th image, $\vec{p}_{\text{det, ref}}$ the position of the track's detection in the reference image, and $\vec{p}_{\text{com}, i}$ the commanded position of the $i$th image relative to the position of the reference image $\vec{p}_{\text{com, ref}}$.
However, because we need to choose particularly liberal tracking parameters during the pursuit of sidereal tracks to account for the still present pointing fluctuation, there is an increased chance to pick up unrelated detections. Thus before we estimate the pointing fluctuation, we calculate the mean absolute offset values:
\begin{align}
\label{eq:wobble2} \bar{|\vec{\delta}_i|} = \frac{1}{M_i} \sum_{j=0}^{M_i} |\vec{\delta}_{i, j}|,
\end{align}
where $M_i$ is the number of sidereal tracks that have a detection in the $i$th image, and exclude any data point from the signal estimation whose absolute offset lies outside a certain range around its mean ($\pm 1.7\,\sigma_i$ in case of sequence STP090). Only then do we calculate the mean offsets (see Fig.~\ref{fig:wobble}) and use them to correct our detection sets:
\begin{align}
\label{eq:wobble3} \bar{\vec{\delta}}_i = \frac{1}{\Tilde{M}_i} \sum_{j=0}^{\Tilde{M}_i} \vec{\delta}_{i, j},
\end{align}
where $\Tilde{M}_i$ is $M_i$ minus the excluded data points. Figure~\ref{fig:correction} shows two sample tracks from the principal sequence with and without the pointing correction.
The choice of the reference point that is used to calculate the relative distances is crucial in this, since a) sidereal tracks that are missing the respective detection cannot be considered for the signature estimation, and b) sidereal tracks that have an unrelated detection as the reference end up with shifted offsets. For sequence STP090, we decided to use the first image of the principal sequence as the reference, as we found that detections from this image are usually not only included in all sidereal tracks but also the most reliable.
Finally, identifying the sidereal tracks also allows us to remove their constituents from the detection sets prior to the actual tracking run, ridding the tracking results of a statistical bias.
\section{Parameter optimization}
\label{sec:optimization}
The whole tracking process--including image cleaning, point source detection, attitude correction and the tracking run itself--involves far too many parameters ($>74$) for a systematic grid search. However for most parameters, preliminary tests indicate that their exact value is (within some range) secondary to achieving good results. We therefore focused on optimizing only the detection threshold (see Sect.~\ref{sec:detection}) and 15 dynamic tracking parameters of the principal and extended tracking (in the following referred to as principal and extended parameters, cf. Table~\ref{tab:parameters}), which we found to be more influential. In the following, we analyze their effect using a single quality index: the miss-rate $\Gamma$. It measures the percentage of detections that were missed during the pursuit of a track (i.e., whenever no suitable pair or detection was found, or when detections were later-on removed during offset checks):
\begin{align}
\label{eq:miss_rate} \Gamma := 100 \cdot \frac{\tilde{N}_\text{det} - n_\text{det}}{\tilde{N}_\text{det}},
\end{align}
where $\tilde{N}_\text{det} \leq N$ is the maximum possible number of detections a specific track can have, which depends on whether and when the track supposedly left the dust field (e.g., if it lies close to the edge, the detections expected outside the field do not count toward the total).
We then estimated the quality of tracking results by looking not only at the total number of tracks, but more importantly at the numbers of tracks with \mbox{$\Gamma=0\,\%$} and \mbox{$\Gamma<30\,\%$}. By visual inspection we found 30\,\% to be a reasonable threshold where most tracks still belong to real particles and only occasionally incorporate unrelated detections.
Because the principal parameters directly affect the total number of tracks (as the acceptance criteria are applied only once after the principal tracking), we optimized them first. Each of the twelve free parameters from Eqs.~\ref{eq:dmax}--\ref{eq:vmax} was varied at least ten times around an initial guess. Since testing all value combinations would still take $10^{12}$ individual tracking runs, we decided on a different strategy: First, we reduced the tracking runs to principal tracking only; and second, we tested each parameter value only once, keeping all other parameters constant. After the full value range for a given parameter was explored, we chose the value that produced the best results and used it as the parameter's new fixed value for the remaining runs. The results of this process are listed in Table~\ref{tab:parameters} in the appendix.
Next, we optimized the extended parameters. Since we decoupled principal and extended tracking, and on its own the latter runs much faster than the former, we adapted our approach. Instead of testing the whole set, we only varied the three parameters that we deemed the most influential ($R_\text{max}$, $\Omega_\text{max}$, and $\Delta V_\text{max}$), and used the optimized principal values for the rest. We again chose ten different values for each of the three variable parameters, but this time, we explored all of the 1000 corresponding value combinations. The parameter set that produced the best results according to our miss-rate criteria is also shown in Table~\ref{tab:parameters}.
Lastly, we estimated the optimal detection threshold (in units of signal-to-noise $S/N$). This dimensionless parameter determines the sensitivity of the detection algorithm toward weaker sources and is therefore directly linked to the number of detections per image. While being able to detect weaker sources can be beneficial in case of fainter particles or oblate rotators (i.e., particles that strongly vary in brightness), it also means to pick up more noise. Hence the detection threshold can neither be too high, as a significant portion of signal would be ignored, nor too low, as the signal would be overwhelmed by noise.
So to optimize the detection threshold, we again chose ten values around an initial guess and measured the average detection density (within the dust field). The detection density however also strongly depends on exposure time ($T_\text{exp}$). In case of sequence STP090, the average detection density for images with \mbox{$T_\text{exp}=6$}\,s was roughly twice of that for images with \mbox{$T_\text{exp}=0.24$}\,s. Thus to keep the detection densities roughly constant, we adopted two separate detection thresholds, one for each exposure time. The one for \mbox{$T_\text{exp}=6$}\,s was then adjusted so that its corresponding detection densities would approximately match those of the \mbox{$T_\text{exp}=0.24$}\,s one. The detection sets produced by each of the ten threshold pairs then underwent their own attitude correction before their tracking runs were started (using the parameters listed in Table~\ref{tab:parameters}). As with the previous optimization processes, we surveyed the total number of tracks, and the numbers of tracks with \mbox{$\Gamma=0\,\%$} and \mbox{$\Gamma<30\,\%$}. After inspecting the most promising results more closely, we found that the best were produced by detection thresholds of $S/N=2.7$ (\mbox{$T_\text{exp}=6$}\,s) and \mbox{$S/N=3.6$} (\mbox{$T_\text{exp}=0.24$}\,s). They roughly correspond to an average detection density of $27.12 \times 10^{-4}$ detections per pixel, or about 7000 detections per image.
Compared to the results produced by our initial set of tracking parameters (see Fig.~\ref{fig:optimization}), the optimization increased the total number of tracks by $\sim$\,18\,\% (from 1922 to 2268), the number tracks with \mbox{$\Gamma<30\,\%$} by $\sim$\,21\,\% (from 642 to 775), and the number of tracks without missing detections by $\sim$\,46\,\% (from 96 to 140)\footnote{While Fig.~\ref{fig:optimization} and the numbers discussed here are based on data produced by the latest version of the tracking algorithm, the optimization was unfortunately run on a previous version where the pointing fluctuation was slightly miscalculated due to a bug. However, because the error in the pointing fluctuation was small (\mbox{$<0.5$}\,px), and because the selected parameter values are only estimates of the optimal values that also work well with the correct pointing fluctuation, we decided against rerunning the optimization procedure.}.
\begin{figure}[!b]
\centering
\includegraphics[width=\linewidth]{./figures/optimization/optimized_tracks}
\caption{Effect of parameter optimization on the miss-rate distribution. Gray, hatched bars show the results produced by the initial set of tracking parameters, orange ones the optimized set.}
\label{fig:optimization}
\end{figure}
\section{Results and discussion}
\label{sec:results}
The goal of this study was to develop a robust algorithm to track dust particles of \object{67P} in image sequences recorded by OSIRIS NAC. As proof of concept, we applied the algorithm to sequence STP090 and optimized the tracking parameters. In the following, we first assess the general reliability of the tracking algorithm, and then give examples of how the tracking results can be evaluated to answer scientific questions.
\subsection{Algorithm assessment}
\subsubsection{Simulation}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figures/results/simulation}
\caption{Results from simulated data: (1) one of the ten simulations with a detection density of $27.59 \times 10^{-4}$\,det./px. The identified tracks are shown as gradient lines from red to yellow. The background image serves merely as a visual aid, showing the rough locations of detections (it is not a stack of images created to run the detection algorithm on; instead the detections were simulated first and the background image was created retroactively). (2) the combined miss-rate distribution from the ten simulations with a detection density of $27.59 \times 10^{-4}$\,det./px. (3) the velocity distribution of the same track population, showing a clear tendency of the algorithm to create more fast spurious tracks, especially when compared to velocity distributions from real data (Fig.~\hyperref[fig:track_stats]{\ref*{fig:track_stats}.2}). The probability density functions were created with Gaussian kernel density estimation.}
\label{fig:simulation}
\end{figure}
To test our algorithm's tendency to create spurious tracks we simulated datasets which consisted entirely of random noise (i.e., "detections"), with detection densities ranging from $27.59 \times 10^{-4}$ to $39.42 \times 10^{-4}$\,det./px. Although the algorithm identified a few hundred to more than two thousand spurious tracks in the simulations depending on their detection density, it found few to none in the critical miss-rate regime below 30\,\%. In particular, we simulated ten different datasets for the detection density closest to that of the optimal detection thresholds ($27.59 \times 10^{-4}$\,det./px, Fig.~\ref{fig:simulation}). In those cases, only 260 tracks were found on average, and in total only 6 with \mbox{$\Gamma<30\,\%$}.
\subsubsection{Manual assessment}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figures/results/track_stats}
\caption{Miss-rate (1), velocity (2), and brightness (3) distributions of tracks identified as either genuine (orange) or ambiguous (gray hatched).}
\label{fig:track_stats}
\end{figure}
To further assess the reliability of our algorithm, we inspected and manually flagged each track found in sequence STP090 according to the following system: if they are (a) genuine, that is, whether we believe that they belong to actual particles and contain few to no unrelated detections, or if they are (b) ambiguous, that is, whether we believe that (the majority of) their detections do not belong to the same particle, stem from noise, or when it is impossible to tell.
Of the 2268 tracks, we flagged 1081 ($\sim$\,48\,\%) as ambiguous, leaving 1187 ($\sim$\,52\,\%) as genuine. Figure~\hyperref[fig:track_stats]{\ref*{fig:track_stats}.1} shows that the miss-rate distributions of the ambiguous and genuine tracks have distinct shapes. In particular, only very few (4) of the ambiguous tracks have miss-rates less than 30\,\%. This is a good sign that our decision to base the parameter optimization on the number of tracks with \mbox{$\Gamma<30\,\%$} was appropriate. This is also further supported by the fact that the miss-rate distribution of the spurious tracks (Fig.~\hyperref[fig:simulation]{\ref*{fig:simulation}.2}) is very similar in shape to that of the ambiguous ones.
Because manually judging the validity of tracks becomes increasingly difficult with the spread of their detections, we expect a bias against faster particles in the flagged tracks. Figure~\hyperref[fig:track_stats]{\ref*{fig:track_stats}.2} shows that such a trend seems to exist in our data, though only slightly. Figure~\hyperref[fig:simulation]{\ref*{fig:simulation}.3} on the other hand shows that our algorithm tends to create more fast than slow spurious tracks. Both effects probably contribute to the excess of fast ambiguous tracks.
We also expect a bias toward flagging faint tracks more often as ambiguous. Figure~\hyperref[fig:track_stats]{\ref*{fig:track_stats}.3} however shows that the opposite was the case. This is likely caused by the overabundance of detections in the bright active area in the center of sequence STP090 (e.g., see Fig.~\ref{fig:source}): while of the genuine tracks only $\sim$\,15\,\% originate from here, of the ambiguous ones it is $\sim$\,22\,\% (the section indicated in Fig.~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.1} was used to calculate those numbers).
\subsection{First results}
In the following, we present examples of how our tracking results can be used and interpreted. Since they mainly serve as a technical demonstration, we do not perform detailed analyses. Nevertheless, because \mbox{$\Gamma<30\,\%$} proved to be a good criterion to identify genuine tracks, we only consider the 775 tracks that satisfy it--more than three times as many tracks than were identified by \cite{10.1093/mnras/stw2179}.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figures/results/genuine_angles}
\caption{Angle distributions of the projected velocity (1) and acceleration (2) of all 775 selected tracks. The orientations of the diagrams coincide with how the images of sequence STP090 are displayed (i.e., $0\degree$ corresponds to the right direction, $90\degree$ to the up direction, etc.). While the projected velocity of most tracks seems to be pointing away from the nucleus, the acceleration of a similar number of tracks seems to be pointing toward it.}
\label{fig:angles}
\end{figure}
Figure~\ref{fig:angles} shows the velocity- and acceleration-angle distributions of all 775 tracks. The projected velocity components of most tracks point upward, seemingly away from the nucleus and the central active area. This aligns well with what would be expected and \cite{10.1093/mnras/stw2179}'s findings. The projected acceleration components on the other hand mostly point downward and seem to be dominated by the nucleus' gravity.
\begin{figure}[p]
\centering
\subfloat{\includegraphics[width=\linewidth]{./figures/results/selection_stats_section1}}\vspace{6.5pt}
\subfloat{\includegraphics[width=\linewidth]{./figures/results/selection_stats_section2}}\vspace{6.5pt}
\subfloat{\includegraphics[width=\linewidth]{./figures/results/selection_stats_section3}}\vspace{6.5pt}
\subfloat{\includegraphics[width=\linewidth]{./figures/results/selection_stats_section4}}
\caption{Selection process and statistics of particles that likely originated from the central active area: (1) the starting points of all 775 tracks (i.e., their earliest confirmed locations) and the tracks we selected (orange circles) that start near the active area. (2) a further reduction of the tracks selected in (1) by choosing only the ones directed upward \mbox{$\pm 45\degree$} (orange). (3) the acceleration angle distribution of the tracks selected in (2), which is further divided into tracks that are accelerated upward (green) and downward (gray hashed). (4, 5, 6) the projected velocity, magnitude of acceleration and radius distributions for the two track populations defined in (3). Escape speed and gravitational acceleration based on \cite{2016Natur.530...63P}.}
\label{fig:selection_stats}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=\linewidth]{./figures/results/selection_tracks}
\caption{The 89 selected tracks near the central active area.}
\label{fig:selection_tracks}
\end{figure}
To derive particle radii and convert particle velocities and accelerations to physical units (e.g., from px/s to m/s), we need to know the particle distances to the spacecraft. Since the only accurate distance measurement we have is of the nucleus ($\sim$86\,km), we focus on particles which were seemingly just ejected from the active area in the center of the images (at that distance \mbox{1\,px $\equiv$ $\sim$\,1.6\,m}). In the following example, we isolated this group in two steps. First, we chose a region around the active area and selected the tracks that originate within it (106 tracks, Fig.~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.1}); then, we further reduced the group by selecting only tracks whose projected velocities are pointing upward within \mbox{$\pm 45\degree$} (89 tracks, Fig.~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.2}). Figure~\ref{fig:selection_tracks} shows the selected tracks as they appear in front of the stacked image of sequence STP090.
Figure~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.3} shows that roughly 47\,\% of the selected tracks have a projected acceleration that points away from the nucleus. The velocity distribution (Fig.~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.4}) shows that they are on average faster than the particles that are accelerated downward. Most particles show a net acceleration less negative than gravity (Fig.~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.5}). Assuming that on a first order the gravitational acceleration is comparable for all particles they must be experiencing an upward directed acceleration of variable strength that partially compensates or even exceeds gravity. A likely candidate for this upward force is gas drag.
Figure~\hyperref[fig:selection_stats]{\ref*{fig:selection_stats}.6} shows the distribution of the particle radii, which were calculated as:
\begin{align}
\label{eq:radius} r = \sqrt{J \frac{r_\text{h}^2 \Delta^2}{R I_\odot}},
\end{align}
where $r$ is the radius in m, $J$ the average particle flux in W/m$^2$/nm, $r_\text{h}$ the dimensionless heliocentric distance measured in units of AU, $\Delta$ the observer-particle distance in m, \mbox{$R=0.0021$} the particle reflectance (computed for decimeter-sized particles using the model in \citealp{8663e156a2a14932bb66452963913cec}), and \mbox{$I_\odot=1.565$\,W/m$^2$/nm} the solar flux in the NAC F22 filter at 1\,AU. The distribution agrees with \cite{10.1093/mnras/stw2179}'s findings when considering that their calculation is affected by a numerical error that leads them to systematically underestimate the radii by a factor of 4.4 (Agarwal et al., in preparation). It furthermore shows no clear trend between the upward- and downward-accelerated particles, which is remarkable because the gas drag we deem responsible for the upward-acceleration should be stronger for smaller particles.
If we assume that the particles have the same bulk density as the nucleus (533\,kg/m$^3$, \citealp{2016Natur.530...63P}), then the upward-accelerated particles contain about 1300\,kg, while the downward-accelerated ones contain roughly 3000\,kg. The largest boulder alone contains more than half of the mass of the upward-accelerated particles.
A similar extended analysis would be interesting to compare to typical models of cometary dust size distributions \citep[e.g.,][and references therein]{10.1093/mnras/stx2741}, as they predict that the majority of mass lost due to refractory material is likely contained in the largest specimen. Hence knowing the size limit and emission rate of the largest chunks is crucial to estimate a comet's contribution to the interplanetary dust environment and the zodiacal cloud \citep{nesvorny-janches2011}.
Lastly, we can also extrapolate our tracks back in time to find out when and where the particles were likely ejected. As the process behind lifting decimeter-sized debris from the surface is not entirely understood (although it appears now to be more straightforward to explain than the lifting of smaller, micron-sized dust, \citealp{gundlach-blum2015}), this can provide us with possible clues about the lifting mechanism or its conditions.
\section{Summary and outlook}
\label{sec:conclusions}
In this paper we present our algorithm for tracking the motion of debris near the nucleus of comet \object{67P}. The algorithm operates on image sequences recorded by Rosetta's camera system OSIRIS. The sequences typically show part of \object{67P}'s surface that ideally has at least one clearly discernible active area which is ejecting particles that appear as point sources against the dark backdrop of interplanetary space.
As an example and to assess the algorithm's reliability as well as presenting tentative first results, we applied our algorithm to image sequence STP090. The evaluation not only showed that our algorithm can find a large number of tracks, but also revealed a robust criterion--having a miss rate \mbox{$\Gamma<30\,\%$}--to separate genuine from ambiguous tracks. Our first results from a group of particles that satisfied the criterion and likely originated from the central area in sequence STP090 demonstrate one way of how our tracking results can be used. And finally, knowing the projected particle velocities and accelerations can help us estimate the fall-back fraction and the refractory-to-ice ratio--which are key to understanding more about cometary interiors and the role comets play in planetesimal formation.
\begin{acknowledgements}
We thank Eberhard Bodenschatz, Ulrich Christensen, and Pablo Lemos for our fruitful discussions; Carsten Güttler, Michael Mommert, and Jakob Deller for their early support; Kyle Barbary, Benne Holwerda, Peter Stetson, Gábor Kovács, Cecilia Tubiana, Guus Bertens, Jan Molacek, and Maurizio Berti for their technical support; Asmus Freytag for proofreading; and Steve Chesley for reviewing our paper and providing constructive comments.
\newline
We acknowledge the operation and calibration team at MPS and the Principal Investigator Holger Sierks on behalf of the OSIRIS Team for providing the OSIRIS images and related datasets. OSIRIS was built by a consortium of the Max-Planck-Institut für Sonnensystemforschung, Göttingen, Germany; the CISAS University of Padova, Italy; the Laboratoire d'Astrophysique de Marseille, France; the Instituto de Astrofísica de Andalucia, CSIC, Granada, Spain; the Research and Scientific Support Department of the European Space Agency, Noordwijk, The Netherlands; the Instituto Nacional de Técnica Aeroespacial, Madrid, Spain; the Universidad Politéchnica de Madrid, Spain; the Department of Physics and Astronomy of Uppsala University, Sweden; and the Institut für Datentechnik und Kommunikationsnetze der Technischen Universität Braunschweig, Germany. The support of the national funding agencies of Germany (DLR), France (CNES), Italy (ASI), Spain (MEC), Sweden (SNSB), and the ESA Technical Directorate is gratefully acknowledged. We thank the Rosetta Science Ground Segment at ESAC, the Rosetta Missions Operations Centre at ESOC and the Rosetta Project at ESTEC for their outstanding work enabling the science return of the Rosetta Mission.
\newline
MP and JA acknowledge funding by the ERC Starting Grant No. 757390 Comet and Asteroid Re-Shaping through Activity (CAstRA). JA acknowledges funding by the Volkswagen Foundation.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction} \label{sec:intro}
In this paper we are interested in solving the one-dimensional parabolic equation
\begin{equation} \label{eq:parabolic}
u_t = b(u)_{xx},
\end{equation}
where $u = u(x,t)$ is a scalar quantity and $b'(u) \geqslant 0$.
One example of such nonlinear equation is the porous medium equation (PME) of the degenerate parabolic type,
\begin{equation} \label{eq:pme}
u_t = (u^m)_{xx},~~m>1,
\end{equation}
which describes the flow of an isentropic gas through a porous medium \cite{Aronson, Muskat}, the heat radiation in plasmas \cite{ZeldovichRaizer}, and various physical processes.
The classical linear heat equation can be considered to be the limit of PME \eqref{eq:pme} as $m \to 1$.
Assuming $u \geqslant 0$, the PME could be written in the form
$$
u_t = \left( m u^{m-1} u_x \right)_x.
$$
Then the PME is parabolic only at those points where $u \ne 0$, while it degenerates as the vanishing of the term $m u^{m-1}$ wherever $u = 0$.
In other words, the PME is a degenerate parabolic equation.
One important property of the PME is the finite propagation, which is different from the infinite speed of propagation in the classical heat equation.
This property implies the appearance of free boundaries that separate the regions where the solution is positive from those where $u = 0$, giving rise to the sharp interfaces \cite{Vazquez}.
Since the free boundaries move with respect to time, their behavior looks similar to the behavior of shocks in the hyperbolic conservation laws.
So it is reasonable to apply the weighted essentially non-oscillatory (WENO) philosophy to the PME, enabling the free boundaries to be well resolved.
Before we discuss the WENO schemes, we would like to mention several different schemes that specialize in nonlinear degenerate parabolic equations in the literature.
For example, the explicit diffusive kinetic schemes have been designed in \cite{Aregba}.
Also, the high-order relaxation scheme has been introduced in \cite{Cavali}.
A local discontinuous Galerkin finite element method for the PME was studied in \cite{ZhangWu}.
Other approaches based on the finite volume method were investigated in \cite{Bess, Arbogast}.
In the more general nonlinear degenerate convection-diffusion case, the entropy stable finite difference schemes were proposed in \cite{Jerez}.
Our focus of this paper is on the WENO schemes.
In \cite{Liu}, Liu et al. constructed the finite difference WENO (WENO-LSZ) schemes for the equation \eqref{eq:parabolic}, which approximate the second derivative term directly by a conservative flux difference.
However, unlike the positive linear weights of WENO schemes for hyperbolic conservation laws \cite{JiangShu, Shu}, the negative linear weights exist so that some special care, such as the technique in \cite{Shi}, was applied to guarantee the non-oscillatory performance in regions of sharp interfaces.
Following the definition of the smoothness indicators in \cite{JiangShu, Shu} and invoking the mapped function in \cite{Henrick}, the resulting nonlinear weights meet the requirement of sixth order accuracy.
In \cite{Hajipour}, Hajipour and Malek proposed the modified WENO (MWENO) scheme with Z-type nonlinear weights \cite{Borges} and nonstandard Runge–Kutta (NRK) schemes.
Further, the hybrid scheme, based on the spatial MWENO and the temporal NRK schemes, was employed to solve the equation \eqref{eq:parabolic} numerically.
Recently, Abedian et al. \cite{AbedianAdibiDehghan, Abedian} aimed at avoiding negative linear weights and presented some modifications to the numerical flux.
In \cite{Rathan}, Rathan et al. showed a new type of local and global smoothness indicators in $L^1$ norm via undivided differences and subsequently constructed the new Z-type nonlinear weights.
Christelieb et al. employed a kernel based approach with the philosophy of the method of lines transpose, giving a high-order WENO method with a nonlinear filter in \cite{Christlieb}.
In \cite{Jiang}, Jiang designed an alternative formulation to approximate the second derivatives in a conservative form, where the odd order derivatives at half points were used to construct the numerical flux.
In this paper we present the central WENO (CWENO) scheme based on the point values for the diffusion term, following the notion of compact CWENO schemes based on the cell avarages for the convection term proposed by Levy et al. in \cite{Levy}.
The negative linear weights in \cite{Liu}, which require some special care, can be circumvented in our scheme.
We further devise the Z-type nonlinear weights \cite{Borges}, which are dependent on the smoothness indicators in \cite{Liu} of $L^2$ norms.
The global smoothness indicator can be designed to attain higher order so that we do not need the power to maintain the order of accuracy.
In our scheme, not only is the computational cost reduced without estimating the mapped function in WENO-LSZ or applying the splitting technique to treat the negative weights in both WENO-LSZ and MWENO, but the non-oscillatory performance is improved since there exist small-scale oscillations around the sharp interfaces for WENO-LSZ in some cases as the time advances whereas those oscillations are largely damped by our scheme.
We also provide the sufficient conditions for sixth order accuracy in smooth regions and an analysis of nonlinear weights shows that the proposed WENO scheme is in compliance with those criteria.
The implementation of WENO schemes for the equation \eqref{eq:parabolic} could be extended to the convection–diffusion equations with the WENO schemes for convection terms \cite{Borges, Henrick, JiangShu, Shu, Gu} and to multi-space dimensions in a dimension-by-dimension approach.
The paper is organized as follows.
In Section \ref{sec:weno}, the sixth-order WENO scheme for the parabolic equation \eqref{eq:parabolic} and some of the relevant analytical results are reviewed.
Section \ref{sec:cweno} presents the CWENO approximation for the diffusion term with the new Z-type nonlinear weights based on $L^2$ norms, and the sufficient conditions for sixth order accuracy in smooth regions, which the devised nonlinear weights satisfy.
In Section \ref{sec:nr}, the proposed CWENO scheme and the WENO-LSZ and MWENO schemes are compared with the simulation of one- and two-dimensional numerical experiments, including 1D and 2D heat equations for the sixth-order verification; 1D and 2D PMEs with various initial conditions; 1D and 2D Buckley–Leverett equations; 1D and 2D strongly degenerate parabolic convection-diffusion equations.
A brief concluding remark is presented in Section \ref{sec:conclusion}.
\section{WENO approximation to the second derivative} \label{sec:weno}
In this section, we review the direct WENO discretization to the second derivative in the conservation form \cite{Liu}.
Consider a uniform grid defined by the points $x_0 < x_1 < \cdots < x_{N-1} < x_N$ with $x_{i \pm 1/2} = x_i \pm \Delta x/2,~i = 0, \dots N$, where $\Delta x = x_{i+1} - x_i$ is the uniform grid spacing.
Then the spatial domain is discretized by this uniform grid.
The semi-discrete form of Equation \eqref{eq:parabolic} with respect to $t$, yields
\begin{equation} \label{eq:parabolic_discrete}
\frac{du_{i}(t)}{dt} = \left. \frac{\partial^2 b}{\partial x^2} \right|_{x=x_i},
\end{equation}
where $u_i(t)$ is the numerical approximation to the point value $u(x_i,t)$.
Define the function $h(x)$ implicitly by
\begin{equation} \label{eq:primitive}
b(u(x)) = \frac{1}{\Delta x^2} \int^{x+\Delta x/2}_{x-\Delta x/2} \left( \, \int^{\eta+\Delta x/2}_{\eta-\Delta x/2} h(\xi) d\xi \right) d\eta.
\end{equation}
Differentiating both sides twice with respect to $x$, we obtain
$$
b(u)_{xx} = \frac{h(x+\Delta x) - 2 h(x) + h(x-\Delta x)}{\Delta x^2}.
$$
Setting $g(x) = h(x+\Delta x/2) - h(x-\Delta x/2)$ gives the equation
$$
\left. \frac{\partial^2 b}{\partial x^2} \right|_{x=x_i} = \frac{g_{i+1/2} - g_{i-1/2}}{\Delta x^2},
$$
where $g_{i \pm 1/2} = g(x_{i \pm 1/2})$.
Then the equation \eqref{eq:parabolic_discrete} becomes
\begin{equation} \label{eq:parabolic_discrete_h}
\frac{du_{i}(t)}{dt} = \frac{g_{i+1/2} - g_{i-1/2}}{\Delta x^2}.
\end{equation}
In order to approximate $g_{i+1/2}$, a polynomial approximation $q(x)$ to $h(x)$ of degree at most 5,
$$
q(x) = a_0 + a_1 x + a_2 x^2 + a_3 x^3 + a_4 x^4 + a_5 x^5,
$$
can be constructed on the 6-point stencil $S^6$, as shown in Figure \ref{fig:stencil}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{stencil.eps}
\vspace{-4.5cm}
\captionof{figure}{The numerical flux $\hat{g}_{i+1/2}$ is constructed on the stencil $S^6 = \{ x_{i-2}, \cdots, x_{i+3} \}$ with six uniform points, as well as three 4-point substencils $S_0, S_1, S_2$.}
\label{fig:stencil}
\end{figure}
The polynomial $q(x)$ interpolates $b_{i+j} = b(u(x_{i+j},t)),~j = -2,\cdots,3$, in the sense of \eqref{eq:primitive}, which gives
\begin{align*}
a_0 &= \frac{2 b_{i-2} - 23 b_{i-1} + 222 b_i - 23 b_{i+1} + 2 b_{i+2}}{180}, \\
a_1 &= \frac{8 b_{i-2} - 55 b_{i-1} - 70 b_i + 160 b_{i+1} - 50 b_{i+2} + 7 b_{i+3}}{120 \Delta x}, \\
a_2 &= - \frac{b_{i-2} - 10 b_{i-1} + 18 b_i - 10 b_{i+1} + b_{i+2}}{120 \Delta x^2}, \\
a_3 &= - \frac{b_{i-2} + 4 b_{i-1} - 20 b_i + 26 b_{i+1} - 13 b_{i+2} + 2 b_{i+3}}{36 \Delta x^3}, \\
a_4 &= \frac{b_{i-2} - 4 b_{i-1} + 6 b_i - 4 b_{i+1} + b_{i+2}}{24 \Delta x^4}, \\
a_5 &= - \frac{b_{i-2} - 5 b_{i-1} + 10 b_i - 10 b_{i+1} + 5 b_{i+2} - b_{i+3}}{120 \Delta x^5}.
\end{align*}
The polynomial $p(x)$ of degree at most 4 approximating $g(x)$ is obtained by taking the difference of $q(x+\Delta x/2)$ and $q(x-\Delta x/2)$,
$$
g(x) = h(x+\Delta x/2) - h(x-\Delta x/2) \approx q(x+\Delta x/2) - q(x-\Delta x/2) = p(x).
$$
Then we have
\begin{equation} \label{eq:polynomial}
p(x) = c_0 + c_1 x + c_2 x^2 + c_3 x^3 + c_4 x^4,
\end{equation}
where
\begin{align*}
c_0 &= \frac{341 b_{i-2} - 2785 b_{i-1} - 2590 b_i + 6670 b_{i+1} - 1895 b_{i+2} + 259 b_{i+3}}{5760}, \\
c_1 &= - \frac{b_{i-2} - 12 b_{i-1} + 22 b_i - 12 b_{i+1} + b_{i+2}}{8 \Delta x}, \\
c_2 &= - \frac{5 b_{i-2} + 11 b_{i-1} - 70 b_i + 94 b_{i+1} - 47 b_{i+2} + 7 b_{i+3}}{48 \Delta x^2}, \\
c_3 &= \frac{b_{i-2} - 4 b_{i-1} + 6 b_i - 4 b_{i+1} + b_{i+2}}{6 \Delta x^3}, \\
c_4 &= - \frac{b_{i-2} - 5 b_{i-1} + 10 b_i - 10 b_{i+1} + 5 b_{i+2} - b_{i+3}}{24 \Delta x^4}.
\end{align*}
Evaluating $p(x)$ at $x = x_{i+1/2}$ yields the finite difference numerical flux
\begin{equation} \label{eq:numerical_flux_plus}
\hat{g}_{i+1/2}^{\text{FD}} = p(x_{i+1/2}) = - \frac{1}{90} b_{i-2} + \frac{5}{36} b_{i-1} - \frac{49}{36} b_i + \frac{49}{36} b_{i+1} - \frac{5}{36} b_{i+2} + \frac{1}{90} b_{i+3}.
\end{equation}
The numerical flux $\hat{g}_{i-1/2}^{\text{FD}}$ is obtained directly by shifting one grid to the left,
\begin{equation} \label{eq:numerical_flux_minus}
\hat{g}_{i-1/2}^{\text{FD}} = - \frac{1}{90} b_{i-3} + \frac{5}{36} b_{i-2} - \frac{49}{36} b_{i-1} + \frac{49}{36} b_i - \frac{5}{36} b_{i+1} + \frac{1}{90} b_{i+2}.
\end{equation}
Applying the Taylor expansions to $\hat{g}_{i \pm 1/2}^{\text{FD}}$ \eqref{eq:numerical_flux_plus} and \eqref{eq:numerical_flux_minus} would give
\begin{align}
\hat{g}_{i+1/2}^{\text{FD}} &= g_{i+1/2} + \frac{1}{560} h_i^{(7)} \Delta x^7 + O(\Delta x^8), \label{eq:numerical_flux_plus_Taylor} \\
\hat{g}_{i-1/2}^{\text{FD}} &= g_{i-1/2} + \frac{1}{560} h_i^{(7)} \Delta x^7 + O(\Delta x^8). \label{eq:numerical_flux_minus_Taylor}
\end{align}
Replacing $g_{i \pm 1/2}$ in \eqref{eq:parabolic_discrete_h} by \eqref{eq:numerical_flux_plus_Taylor} and \eqref{eq:numerical_flux_minus_Taylor}, respectively, we have the sixth-order approximation
\begin{equation} \label{eq:FD_scheme}
\frac{du_{i}(t)}{dt} = \frac{\hat{g}_{i+1/2}^{\text{FD}} - \hat{g}_{i-1/2}^{\text{FD}}}{\Delta x^2} + O(\Delta x^6).
\end{equation}
A similar approach can be used to obtain a polynomial $p_k(x)$ of degree at most 2 on each 4-point substentil $S_k = \{ x_{i-2+k}, \cdots, x_{i+1+k} \}$ with $k = 0,1,2$,
where
\begin{equation} \label{eq:subpolynomials}
\begin{aligned}
p_0(x) &= \frac{5 b_{i-2} - 27 b_{i-1} + 15 b_i + 7 b_{i+1}}{24} + \frac{b_{i-1} - 2 b_i + b_{i+1}}{\Delta x} x - \frac{b_{i-2} - 3 b_{i-1} + 3 b_i - b_{i+1}}{2 \Delta x^2} x^2, \\
p_1(x) &= - \frac{7 b_{i-1} + 15 b_i - 27 b_{i+1} + 5 b_{i+2}}{24} + \frac{b_{i-1} - 2 b_i + b_{i+1}}{\Delta x} x - \frac{b_{i-1} - 3 b_i + 3 b_{i+1} - b_{i+2}}{2 \Delta x^2} x^2, \\
p_2(x) &= - \frac{43 b_i - 69 b_{i+1} + 33 b_{i+2} - 7 b_{i+3}}{24} + \frac{2 b_i - 5 b_{i+1} + 4 b_{i+2} - b_{i+3}}{\Delta x} x - \frac{b_i - 3 b_{i+1}+ 3 b_{i+2} - b_{i+3}}{2 \Delta x^2} x^2.
\end{aligned}
\end{equation}
This results in the numerical fluxes $\hat{g}^k_{i+1/2}$,
\begin{equation} \label{eq:numerical_flux_substencil}
\begin{aligned}
\hat{g}^0_{i+1/2} &= \frac{1}{12} b_{i-2} - \frac{1}{4} b_{i-1} - \frac{3}{4} b_i + \frac{11}{12} b_{i+1}, \\
\hat{g}^1_{i+1/2} &= \frac{1}{12} b_{i-1} - \frac{5}{4} b_i + \frac{5}{4} b_{i+1} - \frac{1}{12} b_{i+2}, \\
\hat{g}^2_{i+1/2} &= - \frac{11}{12} b_i + \frac{3}{4} b_{i+1} + \frac{1}{4} b_{i+2} - \frac{1}{12} b_{i+3}.
\end{aligned}
\end{equation}
We could obtain the numerical fluxes $\hat{g}^k_{i-1/2}$ after shifting each index by -1.
Hence the Taylor series expansion gives
\begin{equation} \label{eq:numerical_flux_substencil_Taylor}
\begin{aligned}
\hat{g}^0_{i \pm 1/2} &= g_{i \pm 1/2} + \frac{1}{12} h_i^{(4)} \Delta x^4 + O(\Delta x^5), \\
\hat{g}^1_{i \pm 1/2} &= g_{i \pm 1/2} - \frac{1}{90} h_i^{(5)} \Delta x^5 + O(\Delta x^6), \\
\hat{g}^2_{i \pm 1/2} &= g_{i \pm 1/2} - \frac{1}{12} h_i^{(4)} \Delta x^4 + O(\Delta x^5).
\end{aligned}
\end{equation}
It is clear that the linear combination of all $\hat{g}^k_{i+1/2}$ can produce $\hat{g}_{i+1/2}^{\text{FD}}$ that approximates the flux $g_{i+1/2}$ in \eqref{eq:parabolic_discrete_h}, that is, there are linear weights $d_0 = d_2 = - \frac{2}{15}$ and $d_1 = \frac{19}{15}$ such that
\begin{equation} \label{eq:numerical_flux_linear_weights}
\hat{g}_{i+1/2}^{\text{FD}} = \sum_{k=0}^2 d_k \hat{g}^k_{i+1/2}.
\end{equation}
Similarly, an index shift by -1 returns the corresponding relation between $\hat{g}_{i-1/2}^{\text{FD}}$ and $\hat{g}^k_{i-1/2}$.
Since \eqref{eq:numerical_flux_linear_weights} is not a convex combination of \eqref{eq:numerical_flux_substencil} as the linear weights $d_0$ and $d_2$ are negative, the WENO procedure cannot be applied directly to obtain a stable scheme.
The test cases in \cite{Shi} showed that WENO schemes without special treatment to the negative weights may lead to the blow-up of the numerical solution.
Thus the splitting technique in \cite{Shi} could be utilized to treat the negative weights $d_0$ and $d_2$.
The linear weights are split into positive and negative parts,
$$
\tilde{\gamma}^+_k = \frac{1}{2} \left( d_k + 3 | d_k | \right),~~\tilde{\gamma}^-_k = \tilde{\gamma}^+_k - d_k,~~k=0,1,2.
$$
Then $d_k = \tilde{\gamma}^+_k - \tilde{\gamma}^-_k$ and
\begin{equation*}
\begin{aligned}
\tilde{\gamma}^+_0 &= \frac{2}{15}, & \tilde{\gamma}^+_1 &= \frac{38}{15}, & \tilde{\gamma}^+_2 &= \frac{2}{15}; \\
\tilde{\gamma}^-_0 &= \frac{4}{15}, & \tilde{\gamma}^-_1 &= \frac{19}{15}, & \tilde{\gamma}^+_2 &= \frac{4}{15}.
\end{aligned}
\end{equation*}
We scale them by
$$
\sigma^{\pm} = \sum_{k=0}^2 \tilde{\gamma}_k^{\pm},~~\gamma_k^{\pm} = \tilde{\gamma}_k^{\pm} / \sigma^{\pm},~~k=0,1,2.
$$
Then the linear positive and negative weights $\gamma_k^{\pm}$ are given by
\begin{equation} \label{eq:gamma}
\begin{aligned}
\gamma^+_0 &= \frac{1}{21}, & \gamma^+_1 &= \frac{19}{21}, & \gamma^+_2 &= \frac{1}{21}; \\
\gamma^-_0 &= \frac{4}{27}, & \gamma^-_1 &= \frac{19}{27}, & \gamma^+_2 &= \frac{4}{27},
\end{aligned}
\end{equation}
which satisfy
\begin{equation} \label{eq:linear_weight_relation}
d_k = \sigma^+ \gamma_k^+ - \sigma^- \gamma_k^-.
\end{equation}
Following the definition of the smoothness indicators in \cite{JiangShu, Shu}, which measure the regularity of the polynomial approximation $p_k(x)$ over some interval, the smoothness indicators are defined as
$$
\beta_k = \sum_{l=1}^2 \Delta x^{2l-1} \int_{x_i}^{x_{i+1}} \left( \frac{d^l}{d x^l} p_k(x) \right)^2 dx,
$$
which gives
\begin{equation} \label{eq:smooth_indicator_Liu}
\begin{aligned}
\beta_0 &= \frac{13}{12} \left( b_{i-2} - 3 b_{i-1} + 3 b_i - b_{i+1} \right)^2 + \frac{1}{4} \left( b_{i-2} - 5 b_{i-1} + 7 b_i - 3 b_{i+1} \right)^2, \\
\beta_1 &= \frac{13}{12} \left( b_{i-1} - 3 b_i + 3 b_{i+1} - b_{i+2} \right)^2 + \frac{1}{4} \left( b_{i-1} - b_i - b_{i+1} + b_{i+2} \right)^2,\\
\beta_2 &= \frac{13}{12} \left( b_i - 3 b_{i+1} + 3 b_{i+2} - b_{i+3} \right)^2 + \frac{1}{4} \left( -3 b_i + 7 b_{i+1} - 5 b_{i+2} + b_{i+3} \right)^2.
\end{aligned}
\end{equation}
The integration over the interval $[x_i, x_{i+1}]$ is performed to satisfy the symmetry property of the parabolic equation, and the factor $\Delta x^{2l-1}$ is introduced to remove any $\Delta x$ dependency in the derivatives.
Depending on the linear weights \eqref{eq:gamma} and the smoothness indicators \eqref{eq:smooth_indicator_Liu}, the nonlinear weights could be defined for the WENO approximation.
In \cite{Liu}, Liu et al. derived the sufficient conditions to attain sixth order accuracy in smooth regions,
\begin{gather}
\omega_0 - \omega_2 = O(\Delta x^4), \label{eq:nonlinear_weights_condition} \\
\omega_k - d_k = O(\Delta x^3). \label{eq:linear_nonlinear_condition}
\end{gather}
In \cite{Liu}, the nonlinear positive and negative weights $\omega^{\pm}_k$ are defined as
\begin{equation} \label{eq:weights_Liu_pm}
\omega^{\pm}_k = \frac{\alpha^{\pm}_k}{\sum^2_{l=0} \alpha^{\pm}_l},~~\alpha^{\pm}_k = \frac{\gamma^{\pm}_k}{(\beta_k + \epsilon)^2},~~k=0,1,2,
\end{equation}
where $\epsilon>0$ is known to avoid the denominator becoming zero.
Based on the relation \eqref{eq:linear_weight_relation} for the linear weights, the nonlinear weights are defined by
\begin{equation} \label{eq:weights}
\omega_k = \sigma^+ \omega^+_k - \sigma^- \omega^-_k.
\end{equation}
However, the nonlinear weights defined in \eqref{eq:weights_Liu_pm} and \eqref{eq:weights} give
$$
\omega_k - d_k = O(\Delta x),
$$
where the condition \eqref{eq:linear_nonlinear_condition} is not satisfied.
To increase the accuracy of the nonlinear weights, the mapped function in \cite{Henrick} is employed:
$$
g_k(\omega) = \frac{\omega (d_k + d_k^2 - 3 d_k \omega + \omega^2)}{d_k^2 + \omega (1 - 2d_k)},~~k = 0,1,2.
$$
The final nonlinear weights are formulated as
$$
\omega^{\text{\tiny LSZ}}_k = \frac{\alpha_k}{\sum^2_{l=0} \alpha_l},~~\alpha_k = g_k(\omega_k),~~k=0,1,2.
$$
It is shown in \cite{Liu} with Taylor expansion that
\begin{gather*}
\omega^{\text{\tiny LSZ}}_0 - \omega^{\text{\tiny LSZ}}_2 = O(\Delta x^4), \\
\omega^{\text{\tiny LSZ}}_k - d_k = O(\Delta x^3).
\end{gather*}
So both conditions \eqref{eq:nonlinear_weights_condition} and \eqref{eq:linear_nonlinear_condition} are satisfied.
\begin{remark} \label{rmk:epsilon_LSZ}
{\rm As $\epsilon = 10^{-6}$ is used in \cite{Liu}, it causes some small-scale oscillations around the sharp interfaces, e.g., in Examples \ref{ex:Barenblatt}, \ref{ex:two_box} and \ref{ex:PME_2d}, Section \ref{sec:nr}, and the NAN values in some computer systems, e.g., in Example \ref{ex:PME_2d}, Section \ref{sec:nr}.
The value of $\epsilon$ is thus replaced by $10^{-10}$ in the WENO-LSZ scheme for all numerical experiments in Section \ref{sec:nr}, so that some oscillations are smoothed and there is none NAN value for all tested computer systems.}
\end{remark}
In \cite{Hajipour}, the MWENO scheme were proposed with Z-type nonlinear weights \cite{Borges}, where the global smoothness indicator is supposed to give higher order, which implies that the lower order terms happen to cancel out if the function is smooth in the stencil.
The global smoothness indicator $\tau$ is simply the absolute difference between $\beta_0$ and $\beta_2$,
$$
\tau = |\beta_0 - \beta_2|,
$$
and the nonlinear positive and negative weights are defined as
\begin{equation} \label{eq:weights_MWENO_pm}
\omega^{\pm}_k = \frac{\alpha^{\pm}_k}{\sum^2_{l=0} \alpha^{\pm}_l},~~\alpha^{\pm}_k = \gamma^{\pm}_k \left( 1 + \left( \frac{\tau}{\beta_k + \epsilon} \right)^2 \right),~~k=0,1,2,
\end{equation}
with $\gamma^{\pm}_k$ in \eqref{eq:gamma} and $\epsilon = 10^{-30}$.
As defined in \eqref{eq:weights}, the MWENO nonlinear weights are
$$
\omega^{\text{\tiny MWENO}}_k = \sigma^+ \omega^+_k - \sigma^- \omega^-_k,
$$
which satisfy the sufficient conditions for sixth order accuracy in smooth regions as shown in \cite{Hajipour}.
Hence the WENO numerical flux is
$$
\hat{g}_{i+1/2} = \sum_{k=0}^2 \omega_k \hat{g}^k_{i+1/2},
$$
where $\hat{g}^k_{i+1/2},~k=0,1,2$ are given by \eqref{eq:numerical_flux_substencil}.
Then the semi-discrete finite difference WENO scheme of the conservation form is
\begin{equation} \label{eq:FD_WENO_scheme}
\frac{du_{i}(t)}{dt} = \frac{\hat{g}_{i+1/2} - \hat{g}_{i-1/2}}{\Delta x^2}.
\end{equation}
\begin{remark}
{\rm A major difference between the WENO approximations to the first and second derivatives is that in one stencil, two numerical fluxes $\hat{f}^+_{i-1/2}$ and $\hat{f}^-_{i+1/2}$ at the respective points $x_{i-1/2}$ and $x_{i+1/2}$ need evaluating for the first derivative, while only one numerical flux $\hat{g}_{i+1/2}$ at $x_{i+1/2}$ is estimated for the second derivative.}
\end{remark}
\section{Central WENO approximation to the second derivative} \label{sec:cweno}
Our goal is to obtain a convex combination of the numerical fluxes $\hat{g}^k_{i+1/2}$ in \eqref{eq:numerical_flux_substencil} as a new approximation to $g_{i+1/2}$ in \eqref{eq:parabolic_discrete_h}.
Motivated by the compact central WENO schemes for hyperbolic conservation laws \cite{Levy}, a centered polynomial $p_C(x)$ is introduced for the WENO approximation, such that all linear weights are positive without the concern about dealing with negative linear weights.
To conform to the notation in \cite{Levy}, we set $p_{\text{\fontsize{5pt}{6pt}\selectfont OPT}}(x) = p(x),~p_\text{\tiny L} (x) = p_0(x),~p_\text{\tiny M} (x) = p_1(x),~p_\text{\tiny R} (x) = p_2(x)$ with $p(x)$ \eqref{eq:polynomial} and $p_0(x),~p_1(x),~p_2(x)$ \eqref{eq:subpolynomials}, and apply this setting to every related term in the previous section accordingly.
The centered polynomial $p_C(x)$ is constructed by the following relation
\begin{equation} \label{eq:cpolynomial_relation}
p_{\text{\fontsize{5pt}{6pt}\selectfont OPT}}(x) = C_\text{\tiny L} p_\text{\tiny L}(x) + C_\text{\tiny M} p_\text{\tiny M}(x) + C_\text{\tiny R} p_\text{\tiny R}(x) + C_\text{\tiny C} p_\text{\tiny C}(x),~~\sum_k C_k = 1,~~k \in \{ \text{\tiny L}, \text{\tiny M}, \text{\tiny R}, \text{\tiny C} \},
\end{equation}
where $C_\text{\tiny L},~C_\text{\tiny M},~C_\text{\tiny R}$ and $C_\text{\tiny C}$ are positive constants.
It is required that $C_\text{\tiny L}=C_\text{\tiny R}$, as discussed further in the following subsection.
With different combinations of $(C_\text{\tiny L},~C_\text{\tiny M},~C_\text{\tiny R})$ attempted, we pick out $(C_\text{\tiny L},~C_\text{\tiny M},~C_\text{\tiny R}) = \left( \frac{1}{6},~\frac{1}{3},~\frac{1}{6} \right)$.
Then
$$
p_\text{\tiny C}(x) = 3 p_{\text{\fontsize{5pt}{6pt}\selectfont OPT}}(x) - \frac{1}{2} p_\text{\tiny L}(x) - p_\text{\tiny M}(x) - \frac{1}{2} p_\text{\tiny R}(x).
$$
The evaluation of the polynomial at $x = x_{i+1/2}$ gives rise to the central numerical flux $\hat{g}^\text{\tiny C}_{i+1/2}$
\begin{equation} \label{eq:numerical_flux_central}
\hat{g}^\text{\tiny C}_{i+1/2} = p_\text{\tiny C}(x_{i+1/2}) = - \frac{3}{40} b_{i-2} + \frac{11}{24} b_{i-1} - 2b_i + 2b_{i+1} - \frac{11}{24} b_{i+2} + \frac{3}{40} b_{i+3},
\end{equation}
and the Taylor series expansion shows that
\begin{equation} \label{eq:numerical_flux_central_plus_Taylor}
\hat{g}^\text{\tiny C}_{i+1/2} = g_{i+1/2} + \frac{23}{360} h_i^{(5)} \Delta x^5 + O(\Delta x^6).
\end{equation}
From \eqref{eq:cpolynomial_relation}, we have
\begin{equation} \label{eq:numerical_flux_central_linear_weights}
\hat{g}_{i+1/2}^{\text{FD}} = \sum_k C_k \hat{g}^k_{i+1/2},~~k \in \{ \text{\tiny L}, \text{\tiny M}, \text{\tiny R}, \text{\tiny C} \}.
\end{equation}
It is straightforward to obtain the corresponding $\hat{g}^\text{\tiny C}_{i-1/2}$ and the relation between $\hat{g}_{i-1/2}^{\text{FD}}$ and $\hat{g}^k_{i-1/2}$ with every index shifted by -1.
Similarly, expanding in Taylor series gives
\begin{equation} \label{eq:numerical_flux_central_minus_Taylor}
\hat{g}^\text{\tiny C}_{i-1/2} = g_{i-1/2} + \frac{23}{360} h_i^{(5)} \Delta x^5 + O(\Delta x^6).
\end{equation}
We define the central smoothness indicator as
$$
\beta_\text{\tiny C} = \sum_{l=1}^4 \Delta x^{2l-1} \int_{x_i}^{x_{i+1}} \left( \frac{d^l}{d x^l} p_{\text{\fontsize{5pt}{6pt}\selectfont OPT}}(x) \right)^2 dx,
$$
where the polynomial $p_\text{\tiny C} (x)$ is not used but replaced by $p_{\text{\fontsize{5pt}{6pt}\selectfont OPT}}(x)$, and the order of the derivative is up to 4.
After some algebra, the central smoothness indicator could be written as
\begin{equation} \label{eq:smooth_indicator_central}
\begin{aligned}
\beta_\text{\tiny C} = & \frac{4273}{20160} \left( b_{i-2} -5 b_{i-1} + 10 b_i - 10 b_{i+1} + 5 b_{i+2} - b_{i+3} \right)^2 + \\
& \frac{29}{345600} \left( 5 b_{i-2} + 11 b_{i-1} - 70 b_i + 94 b_{i+1} - 47 b_{i+2} + 7 b_{i+3} \right)^2 + \\
& \frac{1}{3600} \left( 35 b_{i-2} - 139 b_{i-1} + 230 b_i - 206 b_{i+1} + 103 b_{i+2} - 23 b_{i+3} \right)^2 + \\
& \frac{1}{576} \left( 7 b_{i-2} - 51 b_{i-1} + 134 b_i - 166 b_{i+1} + 99 b_{i+2} - 23 b_{i+3} \right)^2 + \\
& \frac{1}{2304} \left( 7 b_{i-2} - 56 b_{i-1} + 106 b_i - 76 b_{i+1} + 23 b_{i+2} - 4 b_{i+3} \right)^2 + \\
& \frac{1}{9216} \left( 65 b_{i-2} - 353 b_{i-1} + 690 b_i - 602 b_{i+1} + 221 b_{i+2} - 21 b_{i+3} \right)^2 + \\
& \frac{1}{9216} \left( 23 b_{i-2} - 63 b_{i-1} - 34 b_i + 186 b_{i+1} - 133 b_{i+2} + 21 b_{i+3} \right)^2 + \\
& \frac{1}{2304} \left( 13 b_{i-2} - 28 b_{i-1} + 30 b_i - 28 b_{i+1} + 13 b_{i+2} \right)^2 + \\
& \frac{2}{15} \left( b_{i-2} - 4 b_{i-1} + 6 b_i - 4 b_{i+1} + b_{i+2} \right)^2 +
\frac{1}{1152} \left( b_{i-2} - 12 b_{i-1} + 22 b_i - 12 b_{i+1} + b_{i+2} \right)^2.
\end{aligned}
\end{equation}
We set the new global smoothness indicator $\tau_6$ on the stencil $S^6$ as
$$
\tau_6 = \left| \beta_\text{\tiny C} - \frac{1}{24} ( 5 \beta_\text{\tiny L} + 14 \beta_\text{\tiny M} + 5 \beta_\text{\tiny R}) \right|.
$$
The nonlinear weights are defined by
\begin{equation} \label{eq:weights_CWENO}
\omega^{\text{\tiny CWENO}}_k = \frac{\alpha_k}{\sum_l \alpha_l},~~\alpha_k = C_k \left( 1 + \left( \frac{\tau_6}{\beta_k + \epsilon} \right)^p \right),~~k,~l \in \{ \text{\tiny L}, \text{\tiny M}, \text{\tiny R}, \text{\tiny C} \},
\end{equation}
with $C_\text{\tiny L} = C_\text{\tiny R} = \frac{1}{6},~C_\text{\tiny M} = C_\text{\tiny C}= \frac{1}{3}$ and $\epsilon = 10^{-40}$.
The free parameter $p$ is important to achieve sixth order accuracy in smooth regions, as well as control the amount of numerical dissipation.
The choice of $p$ will be discussed below.
We end up with the CWENO numerical flux
\begin{equation} \label{eq:CWENO_numerical_flux}
\hat{g}_{i+1/2} = \sum_k \omega^{\text{\tiny CWENO}}_k \hat{g}^k_{i+1/2},~~k \in \{ \text{\tiny L}, \text{\tiny M}, \text{\tiny R}, \text{\tiny C} \}.
\end{equation}
Note that the central numerical flux $\hat{g}^\text{\tiny C}_{i-1/2}$ is used in smooth regions.
Otherwise its contribution vanishes and the WENO numerical flux is determined by the nonlinear weight(s) corresponding to the smoothness indicators of smaller magnitude.
\subsection{Spatial sixth order accuracy in smooth regions}
We next consider the sufficient conditions of the finite difference WENO scheme \eqref{eq:FD_WENO_scheme} with the new numerical flux \eqref{eq:CWENO_numerical_flux} so as to maintain sixth order accuracy in smooth regions.
Let
$$
\hat{g}_{i \pm 1/2} = \sum_k \omega^{\pm}_k \hat{g}^k_{i \pm 1/2},~~k \in \{ \text{\tiny L}, \text{\tiny M}, \text{\tiny R}, \text{\tiny C} \}.
$$
Here we drop the superscript CWENO in \eqref{eq:CWENO_numerical_flux} to simplify the notation.
The superscripts $\pm$ in the nonlinear weights $\omega^{\pm}_k$ represent two different stencils, with $+$ for $\{ x_{i-2}, \cdots, x_{i+3} \}$ and $-$ for $\{ x_{i-1}, \cdots, x_{i+2} \}$.
The nonlinear weights $\omega^{\pm}_k$ in this subsection are not the nonlinear positive and negative weights $\omega^{\pm}_k$ \eqref{eq:weights_Liu_pm} and \eqref{eq:weights_MWENO_pm} in Section \ref{sec:weno}.
From the relation \eqref{eq:numerical_flux_central_linear_weights}, the numerical flux \eqref{eq:CWENO_numerical_flux} can be rewritten as
$$
\hat{g}_{i+1/2} = \sum_k C_k \hat{g}^k_{i+1/2} + \sum_k ( \omega^+_k - C_k ) \hat{g}^k_{i+1/2}
= \hat{g}^{\text{FD}}_{i+1/2} + \sum_k ( \omega^+_k - C_k ) \hat{g}^k_{i+1/2}.
$$
We expand the last term by using \eqref{eq:numerical_flux_substencil_Taylor} and \eqref{eq:numerical_flux_central_plus_Taylor},
\begin{align*}
\sum_k ( \omega^+_k - C_k ) \hat{g}^k_{i+1/2} =
{} & ( \omega^+_\text{\tiny L} - C_\text{\tiny L} ) \left[ g_{i+1/2} + \frac{1}{12} h_i^{(4)} \Delta x^4 + O(\Delta x^5) \right] + \\
{} & ( \omega^+_\text{\tiny M} - C_\text{\tiny M} ) \left[ g_{i+1/2} - \frac{1}{90} h_i^{(5)} \Delta x^5 + O(\Delta x^6) \right] + \\
{} & ( \omega^+_\text{\tiny R} - C_\text{\tiny R} ) \left[ g_{i+1/2} - \frac{1}{12} h_i^{(4)} \Delta x^4 + O(\Delta x^5) \right] + \\
{} & ( \omega^+_\text{\tiny C} - C_\text{\tiny C} ) \left[ g_{i+1/2} + \frac{23}{360} h_i^{(5)} \Delta x^5 + O(\Delta x^6) \right] \\
= {} & g_{i+1/2} \sum_k(\omega^+_k - C_k)+\frac{1}{12} h_i^{(4)} \Delta x^4 (\omega^+_\text{\tiny L} - \omega^+_\text{\tiny R})+\sum_k (\omega^+_k - C_k) O(\Delta x^5)\\
= {} & \frac{1}{12} h_i^{(4)} \Delta x^4 (\omega^+_\text{\tiny L} - \omega^+_\text{\tiny R}) + \sum_k ( \omega^+_k - C_k ) O(\Delta x^5).
\end{align*}
Then
$$
\hat{g}_{i+1/2} = \hat{g}^{\text{FD}}_{i+1/2}+ \frac{1}{12} h_i^{(4)} \Delta x^4 (\omega^+_\text{\tiny L} - \omega^+_\text{\tiny R})+ \sum_k ( \omega^+_k - C_k ) O(\Delta x^5).
$$
Similarly, with the help of \eqref{eq:numerical_flux_substencil_Taylor} and \eqref{eq:numerical_flux_central_minus_Taylor}, we find that
$$
\hat{g}_{i-1/2} = \hat{g}^{\text{FD}}_{i-1/2}+ \frac{1}{12} h_i^{(4)} \Delta x^4 (\omega^-_\text{\tiny L} - \omega^-_\text{\tiny R})+ \sum_k ( \omega^-_k - C_k ) O(\Delta x^5).
$$
By \eqref{eq:numerical_flux_plus_Taylor} and \eqref{eq:numerical_flux_minus_Taylor}, we have
\begin{align*}
\frac{\hat{g}_{i+1/2} - \hat{g}_{i-1/2}}{\Delta x^2} = {} & \frac{g_{i+1/2} - g_{i-1/2}}{\Delta x^2} + O(\Delta x^6) + \frac{1}{12} h_i^{(4)} \Delta x^2 (\omega^+_\text{\tiny L} - \omega^+_\text{\tiny R}) - \frac{1}{12} h_i^{(4)} \Delta x^2 (\omega^-_\text{\tiny L} - \omega^-_\text{\tiny R}) \\
{} & + \sum_k ( \omega^+_k - C_k ) O(\Delta x^3) - \sum_k ( \omega^-_k - C_k ) O(\Delta x^3).
\end{align*}
Thus the sufficient conditions for sixth order accuracy are given by
\begin{gather}
\omega_\text{\tiny L} - \omega_\text{\tiny R} = O(\Delta x^4), \label{eq:CWENO_nonlinear_weights_condition} \\
\omega_k - C_k = O(\Delta x^3), \label{eq:CWENO_linear_nonlinear_condition}
\end{gather}
where the superscripts are dropped, meaning that the nonlinear weights $\omega_k$ for each stencil $S^6$ are supposed to satisfy both conditions in smooth regions for sixth order accuracy.
Expanding the smoothness indicators $\beta_k,~k \in \{ \text{\tiny L}, \text{\tiny M}, \text{\tiny R}, \text{\tiny C} \}$ \eqref{eq:smooth_indicator_Liu} and \eqref{eq:smooth_indicator_central} in Taylor series at $x = x_{i+1/2}$, we obtain
\begin{align*}
\beta_\text{\tiny L} = b''^2_{i+1/2} \Delta x^4 &+ \left( \frac{13}{12} b'''^2_{i+1/2} - \frac{7}{12} b''_{i+1/2} b^{(4)}_{i+1/2} \right) \Delta x^6 \\
&+ \left( - \frac{13}{6} b'''_{i+1/2} b^{(4)}_{i+1/2} + \frac{1}{2} b''_{i+1/2} b^{(5)}_{i+1/2} \right) \Delta x^7 + O (\Delta x^8), \\
\beta_\text{\tiny M} = b''^2_{i+1/2} \Delta x^4 &+ \left( \frac{13}{12} b'''^2_{i+1/2} + \frac{5}{12} b''_{i+1/2} b^{(4)}_{i+1/2} \right) \Delta x^6 + O (\Delta x^8), \\
\beta_\text{\tiny R} = b''^2_{i+1/2} \Delta x^4 &+ \left( \frac{13}{12} b'''^2_{i+1/2} - \frac{7}{12} b''_{i+1/2} b^{(4)}_{i+1/2} \right) \Delta x^6 \\
&+ \left( \frac{13}{6} b'''_{i+1/2} b^{(4)}_{i+1/2} - \frac{1}{2} b''_{i+1/2} b^{(5)}_{i+1/2} \right) \Delta x^7 + O (\Delta x^8), \\
\beta_\text{\tiny C} = b''^2_{i+1/2} \Delta x^4 &+ \frac{13}{12} b'''^2_{i+1/2} \Delta x^6 + O (\Delta x^8).
\end{align*}
If there is no inflection (or undulation) point at $x_{i+1/2}$, i.e., the second derivative is nonzero, then $\tau_6 = O (\Delta x^8)$.
Since $\tau_6$ is of order $O (\Delta x^8)$ and each $\beta_k$ is of order $O (\Delta x^4)$, one can find that
$$
\left( \frac{\tau_6}{\beta_k} \right)^p = O (\Delta x^{4p}),
$$
by setting $\epsilon = 0$ in the Taylor expansion analysis.
From the definitions \eqref{eq:weights_CWENO},
\begin{equation} \label{eq:linear_nonlinear_relation}
\omega^{\text{\tiny CWENO}}_k = C_k + O (\Delta x^{4p}).
\end{equation}
The minimum value $p$ to satisfy both conditions \eqref{eq:CWENO_nonlinear_weights_condition} and \eqref{eq:CWENO_linear_nonlinear_condition} is $p = 1$.
Note that the condition \eqref{eq:CWENO_nonlinear_weights_condition} combined with \eqref{eq:linear_nonlinear_relation} explains the requirement $C_\text{\tiny L} = C_\text{\tiny R}$.
Now we consider the convergence behavior of the nonlinear weights when there exists an inflection point at $x_{i+1/2}$, that is, the second derivative is zero but the third derivative is nonzero.
Then it can be verified through the Taylor expansion analysis above that
$$
\omega^{\text{\tiny CWENO}}_k = C_k + O (\Delta x^{2p}),
$$
and it is clear that $p=2$ is the minimum value to maintain sixth order accuracy.
As pointed out by Borges et al. \cite{Borges}, increasing the value of $p$ amplifies the numerical dissipation around the discontinuities.
We then choose $p = 1$ in this paper even if it does not satisfy the sufficient conditions \eqref{eq:CWENO_nonlinear_weights_condition} and \eqref{eq:CWENO_linear_nonlinear_condition} at the inflection points.
However, our numerical experiments in the next section show that it still provide sixth order accuracy overall.
\section{Numerical results} \label{sec:nr}
This section presents some numerical experiments to demonstrate the performance of the proposed central WENO scheme and compare with the WENO-LSZ and MWENO schemes.
We examine the accuracy of the WENO schemes for one- and two-dimensional heat equations in terms of $L^1,~L^2$ and $L^{\infty}$ error norms:
\begin{align*}
& L^1 = \frac{1}{N+1} \sum_{i=0}^N \left| u_i(T) - u(x_i, T) \right|, \\
& L^2 = \sqrt{\frac{1}{N+1} \sum_{i=0}^N \left( u_i(T) - u(x_i, T) \right)^2}, \\
& L^{\infty} = \max_{0 \leqslant i \leqslant N} \left| u_i(T) - u(x_i, T) \right|,
\end{align*}
where $u(x_i, T)$ denotes the exact solution and $u_i(T)$ is the numerical approximation at the final time $t = T$.
The rest numerical experiments show the resolution of the numerical solutions with the WENO-LSZ, MWENO and central WENO schemes.
For time discretization, we use the explicit third-order total variation diminishing Runge-Kutta method \cite{ShuOsherI}
\begin{align*}
u^{(1)} &= u^n + \Delta t L(u^n), \\
u^{(2)} &= \frac{3}{4} u^n + \frac{1}{4} u^{(1)} + \frac{1}{4} \Delta t L \left( u^{(1)} \right), \\
u^{n+1} &= \frac{1}{3} u^n + \frac{2}{3} u^{(2)} + \frac{2}{3} \Delta t L \left( u^{(2)} \right),
\end{align*}
where $L$ is the spatial operator.
We follow the CFL condition in \cite{Liu} to set $\text{CFL} = 0.4$ unless otherwise stated.
The central WENO scheme in Section \ref{sec:cweno} is termed as CWENO-DZ with $p=1$.
We choose $\epsilon = 10^{-40}$ for the CWENO-DZ scheme whereas $\epsilon = 10^{-15}$ is set for WENO-LSZ as explained in Remark \ref{rmk:epsilon_LSZ} and $\epsilon = 10^{-30}$ for MWENO as in \cite{Hajipour}.
\subsection{One-dimensional numerical examples}
\begin{example} \label{ex:heat_1d}
We test the accuracy of those WENO schemes for the one-dimensional heat equation
$$
u_t = u_{xx},~~-\pi \leqslant x \leqslant \pi,~~t>0
$$
with the following initial data
$$
u(x,0) = \sin(x),
$$
and the periodic boundary condition.
The exact solution is given by
$$
u(x,t)= e^{-t} \sin(x).
$$
The numerical solution is computed up to the time $T=2$ with the time step $\Delta t = \text{CFL} \cdot \Delta x^2$.
We present the $L_1, L_2$ and $L_{\infty}$ errors versus $N$, as well as the order of accuracy, for the WENO-LSZ, MWENO, CWENO-DZ schemes in Tables \ref{tab:heat_1d_L1}, \ref{tab:heat_1d_L2} and \ref{tab:heat_1d_Linf}, respectively.
It is clear that the expected order of accuracy is achieved for all schemes.
Although the errors of the CWENO-DZ scheme are larger than WENO-LSZ for $N=10$, CWENO-DZ yields the most accurate results as $N$ increases.
\end{example}
\begin{table}[h!]
\centering
\caption{$L_1$ error and order of accuracy for Example \ref{ex:heat_1d}.}
\begin{tabular}{clcrlcrlc}
\hline
N & \multicolumn{2}{l}{WENO-LSZ} & & \multicolumn{2}{l}{MWENO} & & \multicolumn{2}{l}{CWENO-DZ} \\
\cline{2-3} \cline{5-6} \cline{8-9}
& Error & Order & & Error & Order & & Error & Order \\
\hline
10 & 6.31E-6 & -- & & 3.17E-5 & -- & & 4.15E-5 & -- \\
20 & 1.41E-7 & 5.4883 & & 2.16E-7 & 7.1985 & & 1.77E-8 & 11.1951 \\
40 & 2.27E-9 & 5.9514 & & 2.36E-9 & 6.5124 & & 1.94E-9 & 3.1896 \\
80 & 3.54E-11 & 6.0028 & & 3.55E-11 & 6.0562 & & 3.47E-11 & 5.8050 \\
160 & 5.70E-13 & 5.9582 & & 5.70E-13 & 5.9613 & & 5.69E-13 & 5.9304 \\
\hline
\end{tabular}
\label{tab:heat_1d_L1}
\end{table}
\begin{table}[h!]
\centering
\caption{$L_2$ error and order of accuracy for Example \ref{ex:heat_1d}.}
\begin{tabular}{clcrlcrlc}
\hline
N & \multicolumn{2}{l}{WENO-LSZ} & & \multicolumn{2}{l}{MWENO} & & \multicolumn{2}{l}{CWENO-DZ} \\
\cline{2-3} \cline{5-6} \cline{8-9}
& Error & Order & & Error & Order & & Error & Order \\
\hline
10 & 7.50E-6 & -- & & 3.79E-5 & -- & & 4.91E-5 & -- \\
20 & 1.61E-7 & 5.5422 & & 2.47E-7 & 7.2580 & & 2.11E-8 & 11.1843 \\
40 & 2.56E-9 & 5.9742 & & 2.66E-9 & 6.5387 & & 2.21E-9 & 3.2551 \\
80 & 3.96E-11 & 6.0136 & & 3.97E-11 & 6.0664 & & 3.89E-11 & 5.8281 \\
160 & 6.35E-13 & 5.9633 & & 6.35E-13 & 5.9663 & & 6.34E-13 & 5.9391 \\
\hline
\end{tabular}
\label{tab:heat_1d_L2}
\end{table}
\begin{table}[h!]
\centering
\caption{$L_\infty$ error and order of accuracy for Example \ref{ex:heat_1d}.}
\begin{tabular}{clcrlcrlc}
\hline
N & \multicolumn{2}{l}{WENO-LSZ} & & \multicolumn{2}{l}{MWENO} & & \multicolumn{2}{l}{CWENO-DZ} \\
\cline{2-3} \cline{5-6} \cline{8-9}
& Error & Order & & Error & Order & & Error & Order \\
\hline
10 & 1.01E-5 & -- & & 5.22E-5 & -- & & 6.43E-5 & -- \\
20 & 2.31E-7 & 5.4501 & & 3.54E-7 & 7.2038 & & 3.74E-8 & 10.7476 \\
40 & 3.66E-9 & 5.9780 & & 3.80E-9 & 6.5398 & & 3.21E-9 & 3.5424 \\
80 & 5.64E-11 & 6.0217 & & 5.65E-11 & 6.0722 & & 5.54E-11 & 5.8565 \\
160 & 9.01E-13 & 5.9677 & & 9.02E-13 & 5.9702 & & 8.99E-13 & 5.9454 \\
\hline
\end{tabular}
\label{tab:heat_1d_Linf}
\end{table}
\begin{example} \label{ex:Barenblatt}
Consider the PME \eqref{eq:pme}.
If the initial condition is set as the Dirac delta, the Barenblatt solution $B_m(x,t)$ \cite{Barenblatt, ZeldovichKompaneetz}, representing the heat release from a point source, takes the explicit formula
\begin{equation} \label{eq:Barenblatt}
B_m(x,t) = t^{-q} \left[ \left( 1 - \frac{q(m-1)}{2m} \frac{x^2}{t^{2q}} \right)_+ \right]^{1/(m-1)},~~m>1
\end{equation}
where $s_+ = \max (s,0)$ and $q = (m + 1)^{-1}$.
For $t > 0$, the solution has a compact support $[ -a_m(t), a_m(t) ]$, where
$$
\alpha_m(t) = \sqrt{\frac{2m}{k(m-1)}}~t^k,
$$
and the interfaces $|x| = a_m(t)$ move outward at a finite speed.
Moreover, the larger the value of $p$, the sharper the interfaces that separate the compact support and the zero solution.
We simulate the Barenblatt solution \eqref{eq:Barenblatt} of the PME \eqref{eq:pme} with the initial condition as the Barenblatt solution at $t=1$, $u(x,0) = B_m(x,1)$, and the boundary conditions $u(\pm 6,t)=0$ for $t>0$.
The final time is $T = 2$ and the time step is $\Delta t = \text{CFL} \cdot \Delta x^2/m$.
We take $N=160$ and plot the numerical solutions at the final time for $m = 5, \, 7$ and $9$, in Figures \ref{fig:Barenblatt_m5}, \ref{fig:Barenblatt_m7} and \ref{fig:Barenblatt_m9}, respectively.
We can see that the solution of the proposed CWENO-DZ almost overlaps the one of MWENO but both give more accurate solution profiles around the interfaces than WENO-LSZ.
This is also demonstrated by Table \ref{tab:Barenblatt}, which provides the $L_1, L_2$ and $L_{\infty}$ errors for the WENO-LSZ, MWENO and CWENO-DZ schemes.
\end{example}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{Barenblatt_m5.eps}
\includegraphics[width=0.32\textwidth]{Barenblatt_m5_1.eps}
\includegraphics[width=0.32\textwidth]{Barenblatt_m5_2.eps}
\captionof{figure}{Barenblatt solution profiles for Example \ref{ex:Barenblatt} with $m=5$ at $T=2$ (left), close-up view of the solutions in the boxes on the left/right (middle/right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N=160$. The dashed black lines are the exact solution.}
\label{fig:Barenblatt_m5}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{Barenblatt_m7.eps}
\includegraphics[width=0.32\textwidth]{Barenblatt_m7_1.eps}
\includegraphics[width=0.32\textwidth]{Barenblatt_m7_2.eps}
\captionof{figure}{Barenblatt solution profiles for Example \ref{ex:Barenblatt} with $m=7$ at $T=2$ (left), close-up view of the solutions in the boxes on the left/right (middle/right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N=160$. The dashed black lines are the exact solution.}
\label{fig:Barenblatt_m7}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{Barenblatt_m9.eps}
\includegraphics[width=0.32\textwidth]{Barenblatt_m9_1.eps}
\includegraphics[width=0.32\textwidth]{Barenblatt_m9_2.eps}
\captionof{figure}{Barenblatt solution profiles for Example \ref{ex:Barenblatt} with $m=9$ at $T=2$ (left), close-up view of the solutions in the boxes on the left/right (middle/right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N=160$. The dashed black lines are the exact solution.}
\label{fig:Barenblatt_m9}
\end{figure}
\begin{table}[h!]
\centering
\caption{$L_1,~L_2$ and $L_\infty$ errors for Example \ref{ex:Barenblatt}.}
\begin{tabular}{ccccc}
\hline
m & error & WENO-LSZ & MWENO & CWENO-DZ \\
\hline
\multirow{3}{*}{5} & $L_1$ & 2.81E-3 & 1.47E-3 & 1.45E-3 \\
& $L_2$ & 1.82E-2 & 1.15E-2 & 1.14E-2 \\
& $L_\infty$ & 1.77E-1 & 1.03E-1 & 1.02E-1 \\
\hline
\multirow{3}{*}{7} & $L_1$ & 2.77E-3 & 1.39E-3 & 1.37E-3 \\
& $L_2$ & 1.74E-2 & 1.05E-2 & 1.04E-2 \\
& $L_\infty$ & 1.73E-1 & 9.38E-2 & 9.31E-2 \\
\hline
\multirow{3}{*}{9} & $L_1$ & 3.25E-3 & 3.19E-3 & 3.19E-3 \\
& $L_2$ & 2.48E-2 & 2.16E-2 & 2.15E-2 \\
& $L_\infty$ & 2.45E-1 & 1.92E-1 & 1.91E-1 \\
\hline
\end{tabular}
\label{tab:Barenblatt}
\end{table}
\begin{example} \label{ex:two_box}
We continue to consider the PME \eqref{eq:pme}, where the shape of the initial condition is two separate boxes.
If the solution $u$ represents the temperature, the PME models the variations in temperature when two hot spots are situated in the domain.
We first consider the PME with $m=5$, where the initial condition is given by
\begin{equation} \label{eq:two_box_init_sh}
u(x,0) = \begin{cases}
1, & x \in (-3.7,-0.7) \cup (0.7, 3.7), \\
0, & \mbox{otherwise},
\end{cases}
\end{equation}
in which the two boxes have the same height, and the boundary conditions are $u(\pm 5.5,t)=0$ for $t>0$.
We divide the computational domain $[-5.5,~5.5]$ into $N = 220$ uniform cells.
The final time is $T = 1.5$ and the time step is $\Delta t = \text{CFL} \cdot \Delta x^2/m$.
We present the numerical solutions at $t = 0.5, \; 1.0, \; 1.5$, as shown in Figure \ref{fig:two_box_sh}.
The numerical solution, computed by MWENO with a high resolution of $N = 11000$ points, will be referred to as the ``exact'' solution.
Now we consider the PME with $m=6$.
The initial condition in this case is
\begin{equation} \label{eq:two_box_init_dh}
u(x,0) = \begin{cases}
1, & -4 < x < -1, \\
2, & 0 < x < 3, \\
0, & \mbox{otherwise},
\end{cases}
\end{equation}
and the boundary conditions are $u(\pm 6,t) = 0$ for $t>0$.
We select $N = 240$ for the computational domain $[-6,~6]$.
Figure \ref{fig:two_box_dh} shows the approximate results obtained when solving PME up to the final time $T = 0.15$ with the time step $\Delta t = \text{CFL} \cdot \left. \Delta x^2 \right/ \left( m 2^{m-1} \right)$.
We still take the solution computed by MWENO with $N = 6000$ points as the ``exact'' solution.
As seen in Figures \ref{fig:two_box_sh} and \ref{fig:two_box_dh}, all schemes are able to capture the sharp interfaces, and MWENO and CWENO-DZ yield very similar solution profiles.
\end{example}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{two_box_sh_1.eps}
\includegraphics[width=0.32\textwidth]{two_box_sh_2.eps}
\includegraphics[width=0.32\textwidth]{two_box_sh_3.eps}
\captionof{figure}{Solution profiles for PME \eqref{eq:pme} ($m=5$) with the initial condition \eqref{eq:two_box_init_sh} at $t = 0.5$ (left), $1.0$ (middle) and $1.5$ (right) approximated by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N = 220$.
The black lines are generated by MWENO with $N = 11000$.}
\label{fig:two_box_sh}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{two_box_dh_1.eps}
\includegraphics[width=0.32\textwidth]{two_box_dh_2.eps}
\includegraphics[width=0.32\textwidth]{two_box_dh_3.eps}
\captionof{figure}{Solution profiles for PME \eqref{eq:pme} ($m=6$) with the initial condition \eqref{eq:two_box_init_dh} at $t = 0.05$ (left), $0.1$ (middle) and $0.15$ (right) approximated by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N = 240$.
The black lines are generated by MWENO with $N = 6000$.}
\label{fig:two_box_dh}
\end{figure}
Next, we solve the one-dimensional scalar convection-diffusion equation of the form
$$
u_t + f(u)_x = g(u)_{xx}.
$$
For the convection term, the fifth-order finite difference Lax–Friedrichs flux splitting WENO scheme, WENO-JS \cite{JiangShu, Shu}, is employed as we want to see how those WENO schemes for the diffusion term affect the numerical solutions.
The numerical solution, computed by WENO-M \cite{Henrick} and MWENO for the respective convection and diffusion terms with a high resolution, will be referred to as the ``exact'' solution.
\begin{example} \label{ex:Buckley_Leverett}
The Buckley-Leverett equation \cite{Buckley} is of the form
\begin{equation} \label{eq:Buck}
u_t + f(u)_x = \epsilon \left( \nu(u) u_x \right)_x,~~\epsilon \nu(u) \geqslant 0,
\end{equation}
which is a prototype model for oil reservoir simulation.
This is an example of degenerate parabolic equations since $\nu(u)$ vanishes at some values of $u$.
Following \cite{Kurganov}, the convection flux $f(u)$ is of the s-shaped form
\begin{equation} \label{eq:nongravitaional_flux}
f(u) = \frac{u^2}{u^2+(1-u)^2},
\end{equation}
$\epsilon=0.01$, and
\begin{equation} \label{eq:diffusion_coeff_BLE}
\nu(u) = \begin{cases}
4u(1-u), & 0 \leqslant u \leqslant 1, \\
0, & \mbox{otherwise}.
\end{cases}
\end{equation}
The diffusion term $\epsilon \left( \nu(u) u_x \right)_x$ can be written in the form of $g(u)_{xx}$, where
$$
g(u) = \begin{cases}
0, & u < 0, \\
\epsilon \left( - \frac{4}{3} u^3 + 2 u^2 \right), & 0 \leqslant u \leqslant 1, \\
\frac{2}{3} \epsilon, & u > 1.
\end{cases}
$$
The initial condition is given by
$$
u(x,0) = \begin{cases}
1-3x, & 0 \leqslant x \leqslant 1/3, \\
0, & 1/3 < x \leqslant 1.
\end{cases}
$$
and the Dirichlet boundary condition is $u(0,t)=1$.
The computational domain $[0,~1]$ is divided into $N = 100$ uniform cells and the time step is $\Delta t = \text{CFL} \cdot \Delta x^2$.
The numerical solution computed by CWENO-DZ at $T=0.2$ is very close to those by WENO-LSZ and MWENO.
This results in the overlapping in Figure \ref{fig:BLE}.
\end{example}
\begin{figure}[h]
\centering
\includegraphics[width=0.32\textwidth]{BLE.eps}
\includegraphics[width=0.32\textwidth]{BLE_1.eps}
\includegraphics[width=0.32\textwidth]{BLE_2.eps}
\captionof{figure}{Solution profiles for Buckley-Leverett equation in Example \ref{ex:Buckley_Leverett} at $T = 0.2$ (left), close-up view of the solutions in the boxes on the left/right (middle/right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N = 100$.
The dashed black lines are generated by WENO-M and MWENO with $N = 1000$.}
\label{fig:BLE}
\end{figure}
\begin{example} \label{ex:Buckley_Leverett_gravitation}
We continue to consider the Buckley-Leverett equation \eqref{eq:Buck} with the same $\epsilon=0.01$ and $\nu(u)$ \eqref{eq:diffusion_coeff_BLE} as in Example \ref{ex:Buckley_Leverett}.
The flux function $f(u)$ with gravitational effects is
\begin{equation} \label{eq:gravitaional_flux}
f(u)=\frac{u^2}{u^2+(1-u)^2}(1-5(1-u)^2),
\end{equation}
where the sign of $f'(u)$ changes in $[0,~1]$.
The Riemann initial condition is
$$
u(x,0) = \begin{cases}
0, & 0 \leqslant x < 1 - 1/\sqrt{2}, \\
1, & 1 - 1/\sqrt{2} \leqslant x \leqslant 1.
\end{cases}
$$
We divide the computational domain $[0,~1]$ into $N=100$ uniform cells and the time step is $\Delta t = \text{CFL} \cdot \Delta x^2$.
Figure \ref{fig:BLE_grav} shows the numerical solutions at $T=0.2$ for the convection flux $f(u)$ \eqref{eq:gravitaional_flux} with gravitational effects while Figure \ref{fig:BLE_nograv} presents the ones for $f(u)$ \eqref{eq:nongravitaional_flux} without gravitational effects.
In Figure \ref{fig:BLE_grav}, all WENO schemes yield comparable results, while in Figure \ref{fig:BLE_nograv}, CWENO-DZ produces the numerical solution slightly closer to the reference solution than WENO-LSZ and MWENO around the shock.
\end{example}
\begin{figure}[h]
\centering
\includegraphics[width=0.32\textwidth]{BLE_grav.eps}
\includegraphics[width=0.32\textwidth]{BLE_grav_1.eps}
\includegraphics[width=0.32\textwidth]{BLE_grav_2.eps}
\captionof{figure}{Solution profiles for Buckley-Leverett equation \eqref{eq:Buck} with gravitation in Example \ref{ex:Buckley_Leverett_gravitation} at $T=0.2$ (left), close-up view of the solutions in the boxes on the left/right (middle/right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N = 100$. The dashed black lines are generated by WENO-M and MWENO with $N = 1000$.}
\label{fig:BLE_grav}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.32\textwidth]{BLE_nograv.eps}
\includegraphics[width=0.32\textwidth]{BLE_nograv_1.eps}
\includegraphics[width=0.32\textwidth]{BLE_nograv_2.eps}
\captionof{figure}{Solution profiles for Buckley-Leverett equation \eqref{eq:Buck} without gravitation in Example \ref{ex:Buckley_Leverett_gravitation} at $T=0.2$ (left), close-up view of the solutions in the boxes on the left/right (middle/right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N = 100$. The dashed black lines are generated by WENO-M and WENO-LSZ with $N = 1000$.}
\label{fig:BLE_nograv}
\end{figure}
\begin{example} \label{ex:strongly_degenerate_cd_1d}
In this example, we consider the strongly degenerate parabolic convection-diffusion equation
\begin{equation} \label{eq:strongly_degenerate_cd_1d}
u_t + f(u)_x = \epsilon \left( \nu(u) u_x \right)_x,~~\epsilon \nu(u) \geqslant 0.
\end{equation}
We take $\epsilon=0.1,~f(u)=u^2$, and
\begin{equation} \label{eq:diffusion_coeff_sdp}
\nu(u) = \begin{cases}
0, & |u| \leqslant 0.25, \\
1, & |u| > 0.25.
\end{cases}
\end{equation}
If $|u| \leqslant 0.25$, the equation \eqref{eq:strongly_degenerate_cd_1d} returns to the hyperbolic equation.
The diffusion term $\epsilon \left( \nu(u) u_x \right)_x$ can be written in the form of $g(u)_{xx}$, where
$$
g(u) = \begin{cases}
\epsilon (u+0.25), & u < -0.25, \\
\epsilon (u-0.25), & u > 0.25, \\
0, & \mbox{otherwise}.
\end{cases}
$$
The initial condition is given by
$$
u(x,0) = \begin{cases}
1, & -1/\sqrt{2}-0.4 < x < -1/\sqrt{2}+0.4, \\
-1, & 1/\sqrt{2}-0.4 < x < 1/\sqrt{2}+0.4, \\
0, & \mbox{otherwise}.
\end{cases}
$$
We divide the computational domain $[-2,~2]$ into $N=200$ uniform cells and the time step is $\Delta t = \text{CFL} \cdot \Delta x^2$.
The simulations at $T=0.7$ are presented in Figure \ref{fig:SDP}, where the numerical result with CWENO-DZ is comparable to those with WENO-LSZ and MWENO.
\end{example}
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{SDP.eps}
\includegraphics[width=0.45\textwidth]{SDP_1.eps}
\includegraphics[width=0.45\textwidth]{SDP_2.eps}
\includegraphics[width=0.45\textwidth]{SDP_3.eps}
\captionof{figure}{Solution profiles for Example \ref{ex:strongly_degenerate_cd_1d} at $T=0.7$ (top left), close-up view of the solutions in the boxes on the left/middle/right (top right/bottom left/bottom right) computed by WENO-LSZ (red), MWENO (green) and CWENO-DZ (blue) with $N = 200$. The dashed black lines are generated by WENO-M and WENO-LSZ with $N = 2000$.}
\label{fig:SDP}
\end{figure}
\subsection{Two-dimensional numerical examples}
\begin{example} \label{ex:heat_2d}
We test the accuracy of those WENO schemes for the two-dimensional heat equation
$$
u_t = u_{xx} + u_{yy},~~-\pi \leqslant x, y \leqslant \pi,~~t>0
$$
subject to the initial data
$$
u(x,y,0) = \sin(x+y),
$$
and the periodic boundary conditions in both directions.
The exact solution is
$$
u(x,t)= e^{-2t} \sin(x+y).
$$
The numerical solutions are computed at the final time $T=2$ with the time step $\Delta t = 0.2 \cdot \min (\Delta x, \Delta y)^2$.
The $L_1, L_2$, and $L_{\infty}$ errors, along with the orders of accuracy, are provided in Tables \ref{tab:heat_2d_L1}, \ref{tab:heat_2d_L2} and \ref{tab:heat_2d_Linf}, respectively.
All WENO schemes exhibit sixth order accuracy overall.
As in Examples \ref{ex:heat_1d}, the errors produced by CWENO-DZ are larger than WENO-LSZ for $N=10$, but we see that the proposed CWENO-DZ scheme performs the best in terms of accuracy subsequently.
\end{example}
\begin{table}[h!]
\centering
\caption{$L_1$ errors and order of convergence for Example \ref{ex:heat_2d}.}
\begin{tabular}{clcrlcrlc}
\hline
N & \multicolumn{2}{l}{WENO-LSZ} & & \multicolumn{2}{l}{MWENO} & & \multicolumn{2}{l}{CWENO-DZ} \\
\cline{2-3} \cline{5-6} \cline{8-9}
& Error & Order & & Error & Order & & Error & Order \\
\hline
10 $\times$ 10 & 1.83E-6 & -- & & 9.20E-6 & -- & & 1.20E-5 & -- \\
20 $\times$ 20 & 3.97E-8 & 5.5240 & & 6.10E-8 & 7.2381 & & 3.18E-9 & 11.8756 \\
40 $\times$ 40 & 6.30E-10 & 5.9785 & & 6.55E-10 & 6.5412 & & 5.40E-10 & 2.5608 \\
80 $\times$ 80 & 9.71E-12 & 6.0189 & & 9.73E-12 & 6.0719 & & 9.51E-12 & 5.8271 \\
160 $\times$ 160 & 1.55E-13 & 5.9665 & & 1.55E-13 & 5.9696 & & 1.55E-13 & 5.9402 \\
\hline
\end{tabular}
\label{tab:heat_2d_L1}
\end{table}
\begin{table}[h!]
\centering
\caption{$L_2$ errors and order of convergence for Example \ref{ex:heat_2d}.}
\begin{tabular}{clcrlcrlc}
\hline
N & \multicolumn{2}{l}{WENO-LSZ} & & \multicolumn{2}{l}{MWENO} & & \multicolumn{2}{l}{CWENO-DZ} \\
\cline{2-3} \cline{5-6} \cline{8-9}
& Error & Order & & Error & Order & & Error & Order \\
\hline
10 $\times$ 10 & 2.09E-6 & -- & & 1.06E-5 & -- & & 1.37E-5 & -- \\
20 $\times$ 20 & 4.44E-8 & 5.5563 & & 6.83E-8 & 7.2737 & & 4.16E-9 & 11.6802 \\
40 $\times$ 40 & 7.01E-10 & 5.9869 & & 7.29E-10 & 6.5513 & & 6.04E-10 & 2.7848 \\
80 $\times$ 80 & 1.08E-11 & 6.0212 & & 1.08E-11 & 6.0740 & & 1.06E-11 & 5.8355 \\
160 $\times$ 160 & 1.72E-13 & 5.9672 & & 1.72E-13 & 5.9702 & & 1.72E-13 & 5.9419 \\
\hline
\end{tabular}
\label{tab:heat_2d_L2}
\end{table}
\begin{table}[h!]
\centering
\caption{$L_\infty$ errors and order of convergence for Example \ref{ex:heat_2d}.}
\begin{tabular}{clcrlcrlc}
\hline
N & \multicolumn{2}{l}{WENO-LSZ} & & \multicolumn{2}{l}{MWENO} & & \multicolumn{2}{l}{CWENO-DZ} \\
\cline{2-3} \cline{5-6} \cline{8-9}
& Error & Order & & Error & Order & & Error & Order \\
\hline
10 $\times$ 10 & 2.76E-6 & -- & & 1.41E-5 & -- & & 1.78E-5 & -- \\
20 $\times$ 20 & 6.26E-8 & 5.4641 & & 9.61E-8 & 7.1989 & & 7.46E-9 & 11.2224 \\
40 $\times$ 40 & 9.91E-10 & 5.9814 & & 1.03E-9 & 6.5445 & & 8.61E-10 & 3.1154 \\
80 $\times$ 80 & 1.53E-11 & 6.0212 & & 1.53E-11 & 6.0728 & & 1.50E-11 & 5.8460 \\
160 $\times$ 160 & 2.44E-13 & 5.9673 & & 2.44E-13 & 5.9701 & & 2.43E-13 & 5.9433 \\
\hline
\end{tabular}
\label{tab:heat_2d_Linf}
\end{table}
\begin{example} \label{ex:PME_2d}
Consider the two-dimensional PME given by
$$
u_t = \left( u^2 \right)_{xx} + \left( u^2 \right)_{yy},
$$
with the initial condition
$$
u(x,y,0) = \begin{cases}
\exp \left( - \frac{1}{6 - (x-2)^2 - (y+2)^2} \right), & (x-2)^2 + (y+2)^2 < 6, \\
\exp \left( - \frac{1}{6 - (x+2)^2 - (y-2)^2} \right), & (x+2)^2 + (y-2)^2 < 6, \\
0, & \mbox{otherwise},
\end{cases}
$$
and the periodic boundary condition in each direction.
We divide the square computational domain $[-10,~10] \times [-10,~10]$ into $N_x \times N_y = 80 \times 80$ uniform cells and the time step $\Delta t = \text{CFL} \cdot \min (\Delta x, \Delta y)^4/2$.
The numerical solutions at $t=1$ and $t=4$ are shown in Figures \ref{fig:PME_2d_T1} and \ref{fig:PME_2d_T4}, respectively.
At the time $t=1$, there are some small-scale oscillations around the free boundaries in the solution by WENO-LSZ, which are implied by the white spots in the surface plot on the top left of Figure \ref{fig:PME_2d_T1}.
The oscillations are largely damped by MWENO and CWENO-DZ as there is no obvious white spot in the surface plot on the top middle and right, respectively.
However, at the time $t=4$, all WENO schemes are able to capture the free boundaries without noticeable oscillation, as shown in Figure \ref{fig:PME_2d_T4}.
Table \ref{tab:PME_2d} shows the minimum value of every numerical solution, which agrees with our observation above.
\end{example}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{PME_2d_T1_LSZ_surface.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T1_MWENO_surface.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T1_DC_surface.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T1_LSZ_contour.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T1_MWENO_contour.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T1_DC_contour.eps}
\captionof{figure}{Solutions in the surface (top) and contour (bottom) plots for Example \ref{ex:PME_2d} at $t = 1$ by WENO-LSZ (left), MWENO (middle) and CWENO-DZ (right) with $N_x \times N_y = 80 \times 80$. Each contour plot includes 18 contours of $u$.}
\label{fig:PME_2d_T1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{PME_2d_T4_LSZ_surface.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T4_MWENO_surface.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T4_DC_surface.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T4_LSZ_contour.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T4_MWENO_contour.eps}
\includegraphics[width=0.32\textwidth]{PME_2d_T4_DC_contour.eps}
\captionof{figure}{Solutions in the surface (top) and contour (bottom) plots for Example \ref{ex:PME_2d} at $t = 4$ by WENO-LSZ (left), MWENO (middle) and CWENO-DZ (right) with $N_x \times N_y = 80 \times 80$. Each contour plot includes 18 contours of $u$.}
\label{fig:PME_2d_T4}
\end{figure}
\begin{table}[h!]
\centering
\caption{Minimum values of the numerical solutions at $t=1$ and $t=4$ for Example \ref{ex:PME_2d}.}
\begin{tabular}{cccc}
\hline
t & WENO-LSZ & MWENO & CWENO-DZ \\
\hline
1 & -9.0127E-2 & -1.1547e-16 & -4.5836e-22 \\
\hline
4 & -2.0504E-8 & -2.3381e-16 & -9.6261e-22 \\
\hline
\end{tabular}
\label{tab:PME_2d}
\end{table}
Finally, we use WENO schemes to solve the two-dimensional scalar convection-diffusion equations.
The WENO-JS scheme for the convection term is combined with WENO-LSZ for the diffusion term, while WENO-ZR \cite{Gu}, which gives sharper approximations around the shocks, is applied with both MWENO and CWENO-DZ.
\begin{example} \label{ex:Buckley_Leverett_2d}
We consider the two-dimensional Buckley-Leverett equation of the form
$$
u_t + f_1(u)_x + f_2(u)_y = \epsilon \left( u_{xx} + u_{yy} \right),
$$
with $\epsilon = 0.01$ and the flux functions given by
$$
f_1(u) = \frac{u^2}{u^2+(1-u)^2},~\quad f_2(u) = \left( 1 - 5(1-u)^2 \right) f_1(u).
$$
Then the equation includes gravitational effects only in the y-direction.
The initial condition is
$$
u(x,y,0) = \begin{cases}
1, & x^2 + y^2 < 0.5, \\
0, & \mbox{otherwise}.
\end{cases}
$$
The square computational domain $[-1.5,~1.5] \times [-1.5,~1.5]$ is divided into $N_x \times N_y = 120 \times 120$ uniform cells and the time step is $\Delta t = \text{CFL} \cdot \min (\Delta x, \Delta y)^2$.
The solutions at $T = 0.5$ are plotted in Figure \ref{fig:BLE_2d}.
The white spot in the surface plot on the top left indicates the small-scale oscillations around the discontinuities in the solution by WENO-JS and WENO-LSZ.
Those oscillations are smoothed by WENO-ZR with both MWENO and CWENO-ZR, corresponding to the surface plot on the top middle and right, respectively.
We also provide Table \ref{tab:BLE_2d} showing the minimum value of each solution.
\end{example}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{BLE_2d_LSZ_surface.eps}
\includegraphics[width=0.32\textwidth]{BLE_2d_MWENO_surface.eps}
\includegraphics[width=0.32\textwidth]{BLE_2d_DC_surface.eps}
\includegraphics[width=0.32\textwidth]{BLE_2d_LSZ_contour.eps}
\includegraphics[width=0.32\textwidth]{BLE_2d_MWENO_contour.eps}
\includegraphics[width=0.32\textwidth]{BLE_2d_DC_contour.eps}
\captionof{figure}{Solutions in the surface (top) and contour (bottom) plots for Example \ref{ex:Buckley_Leverett_2d} at $T = 0.5$ by WENO-JS/WENO-LSZ (left), WENO-ZR/MWENO (middle) and WENO-ZR/CWENO-DZ (right) with $N_x \times N_y = 120 \times 120$. Each contour plot includes 18 contours of $u$.}
\label{fig:BLE_2d}
\end{figure}
\begin{table}[h!]
\centering
\caption{Minimum values of the numerical solutions at $T = 0.5$ for Example \ref{ex:Buckley_Leverett_2d}.}
\begin{tabular}{cccc}
\hline
T & WENO-LSZ & MWENO & CWENO-DZ \\
\hline
0.5 & -6.2550E-3 & 1.5645E-39 & 8.0662E-39 \\
\hline
\end{tabular}
\label{tab:BLE_2d}
\end{table}
\begin{example} \label{ex:strongly_degenerate_cd_2d}
We conclude this section with the two-dimensional strongly degenerate parabolic convection-diffusion equation
$$
u_t + f(u)_x + f(u)_y = \epsilon \left( \nu(u) u_x \right)_x + \epsilon \left( \nu(u) u_y \right)_y,
$$
where $\epsilon=0.1,~f(u)=u^2$, and $\nu(u)$ \eqref{eq:diffusion_coeff_sdp} are the same as in Example \ref{ex:strongly_degenerate_cd_1d}.
The initial condition is
$$
u(x,y,0) = \begin{cases}
1, & (x+0.5)^2+(y+0.5)^2<0.16, \\
-1, & (x-0.5)^2+(y-0.5)^2<0.16, \\
0, & \mbox{otherwise}.
\end{cases}
$$
We divide the computational domain $[-1.5,~1.5] \times [-1.5,~1.5]$ into $N_x \times N_y = 120 \times 120$ uniform cells and the time step is $\Delta t = \text{CFL} \cdot \min (\Delta x, \Delta y)^2$.
The numerical solutions at $T = 0.5$, generated by those WENO schemes, look similar in Figure \ref{fig:SDP_2d}.
\end{example}
\begin{figure}[h!]
\centering
\includegraphics[width=0.32\textwidth]{SDP_2d_LSZ_surface.eps}
\includegraphics[width=0.32\textwidth]{SDP_2d_MWENO_surface.eps}
\includegraphics[width=0.32\textwidth]{SDP_2d_DC_surface.eps}
\includegraphics[width=0.32\textwidth]{SDP_2d_LSZ_contour.eps}
\includegraphics[width=0.32\textwidth]{SDP_2d_MWENO_contour.eps}
\includegraphics[width=0.32\textwidth]{SDP_2d_DC_contour.eps}
\captionof{figure}{Solutions in the surface (top) and contour (bottom) plots for Example \ref{ex:strongly_degenerate_cd_2d} at $T = 0.5$ by WENO-JS/WENO-LSZ (left), WENO-ZR/MWENO (middle) and WENO-ZR/CWENO-DZ (right) with $N_x \times N_y = 120 \times 120$. Each contour plot includes 18 contours of $u$.}
\label{fig:SDP_2d}
\end{figure}
\section{Conclusion} \label{sec:conclusion}
In this paper, we proposed a six-order finite difference CWENO scheme to solve nonlinear degenerate parabolic equations.
The key idea is to introduce a centered polynomial such that the positivity of linear weights is guaranteed.
Numerical examples show that the proposed CWENO scheme achieves sixth order accuracy with smaller errors than WENO-LSZ and MWENO, and inhibits the small-scale oscillations introduced by WENO-LSZ.
\section*{Acknowledgments}
The first author is supported by IIPE, Visakhapatnam, India, under the IRG grant number $\text{IIPE/DORD/IRG/001}$ and NBHM, DAE, India (Ref. No. $\text{02011/46/2021 NBHM(R.P.)/R \& D II/14874}$).
The second author is supported by POSTECH Basic Science Research Institute under the NRF grant number $\text{NRF2021R1A6A1A1004294412}$.
\input{reference}
\end{document}
|
\section{Introduction}
Large lifetimes of chiral magnetic structures like skyrmions, antiskyrmions, spin helices, or chiral domain walls resulted in numerous experimental and theoretical studies of their application as bits of information in standard \cite{parkin2008magnetic,tomasello2014strategy,hag:NC2015,schaffer2020rotating} and probabilistic computing~\cite{pinna2018skyrmion,zazvorka2019thermal}. The vast majority of these publications relies on the stability of the chiral magnetic objects. Recently, however, it has been shown theoretically \cite{schaffer2019stochastic} and experimentally \cite{song2021commensurability,lindner2020temperature} that skyrmions in various confined geometries can be structurally stable while performing a diffusional motion. Characteristics of this Brownian-like motion are of high importance for memory applications that use confined geometries such as race-tracks or as reservoirs. The dynamics and experimental appearance of skyrmions in this regime of stochastic motion depends on the geometry of the confined regions or islands, temperature, and energy parameters. Importantly, the measured spatial distribution of skyrmions depends on the time resolution of observation. Taking an example of three skyrmions on a triangular nanoisland, one sees a chaotic skyrmion motion in real-time videos, but three immobile skyrmions in time-averaged static images \cite{schaffer2019stochastic}. Even more peculiar, one sees three immobile skyrmionic features in the time-averaged signal if only two skyrmions are diffusing on a triangular island, but stripes and skyrmion if the island has a round shape. This phenomenon appears particularly important in view of the fact that most experimental techniques resolving chiral magnetic structures are time-integrating. Thus, to rely on chiral magnetic bits of information, it is necessary to distinguish between the spatially and structurally stable magnetic entities from their structurally stable but diffusing counterparts.
This distinction is, however, not an easy task. Indeed, the required classification of structures cannot be reliably done with the unaided eye, and the direct way to solve it is to increase the time resolution, which is a technically tremendous assignment. Meanwhile, many complex problems have been solved using machine learning techniques. Particularly, machine learning algorithms have been used for finding \cite{Kwon:PRB2019}, recognizing \cite{Iak:PRB2018}, and classifying \cite{Salcedo:JMMM2020} ground states of systems with Dzyaloshinskii-Moriya interactions (DMI) and for reconstructing the spin configuration from data obtained in reciprocal space \cite{Kwon:SciRep2019}. Machine learning has also been used to define phase transitions in Ising-like spin systems \cite{Carras:NatPhys2017} and to estimate the DMI parameters from images of magnetic domains \cite{Kawa:njp2021,kwon2020magnetic}. In contrast to these investigations of statical properties of skyrmionic systems, a recent study \cite{Wang:PRA2021} applied neural networks for deep learning of skyrmionic dynamical phases from videos. Those videos have been created using numerical simulations representing skyrmions as interacting rigid-point particles. Changes in topological charges of chiral structures as well as different time-integrating schemes were disregarded in \cite{Wang:PRA2021}.
\\
Here, we show how machine learning categorically outperforms standard advanced techniques for distinguishing diffusing and spatially stable skyrmions and extracting the total topological charge from time-integrated data of the $z$-direction of the magnetic moments. To achieve this goal, we train a neural network (NN) using blurred, time-integrated data in various confined geometries in real space and then calculate topological charges of a complementary set of images using the trained NN. Because the topological charge directly corresponds to the number of topologically non-trivial magnetic objects (skyrmions) in our sample, this procedure delivers complete information about the dynamical system from time-averaged, statical images. Our results yield an outstanding degree of recognition and are important for experimental investigations of stability of chiral magnets.
\section{Skyrmion dynamics in confined geometries}
We study metastable skyrmions and their motion at finite temperatures on magnetic nanoislands. The simulations are performed with generic parameters close to those of a Pd/Fe bilayer on an Ir(111) surface, known for hosting nanometer-sized skyrmions at moderate magnetic bias fields. A detailed description of the phase diagram for extended systems of ultrathin biatomic layers of Pd/Fe on an Ir(111) substrate hosting magnetic skyrmions can be found in Refs.~\cite{rozsa2016complex,schaffer2019stochastic}.
Here, simulations are performed for the same system in the atomistic spin dynamics framework, by solving the Landau-Lifshitz-Gilbert equation (LLG)\cite{gilbert2004phenomenological}
\begin{align}
\dot{\vec{m}}_i=&-\frac{\gamma}{1+\alpha^2}\vec{m}_i\times\left(\vec{B}_{i}^{\mathrm{eff}}+\alpha\vec{m}_i\times\vec{B}_{i}^{\mathrm{eff}}\right)\ ,
\end{align}
on a discrete square lattice. The gyromagnetic ratio of an electron is $\gamma=1.76\times 10^{11}($T$^{-1}$s$^{-1})$, $\alpha$ is the Gilbert damping parameter, $\vec{m}_i$ is the normalized magnetization vector at lattice site $i$ and $\vec{B}_{i}^{\mathrm{eff}}$ is the effective magnetic field.
The classical Heisenberg Hamiltonian reads
\begin{align}
\begin{split}
\mathcal{H} &= -\frac{1}{2}\sum_{\langle ij \rangle}\left[ J\ind{ex} \vec{m}_i\cdot \vec{m}_j -\vec{D}_{ij}\cdot(\vec{m}_i\times \vec{m}_j)\right]\\
& - \sum_i \left[\mu\ind{s} \vec{B}\ind{ext}\cdot\vec{m}_i - K\ind{u}(m_i^z)^2\right] \, ,
\end{split}
\end{align}
and includes the Heisenberg exchange, interfacial Dzyaloshinskii-Moriya (DMI) interaction, Zeeman coupling to an external field $\vec{B}\ind{ext}$, and the uniaxial anisotropy in $z$-direction. The material parameters (magnetic moment $\mu_\ind{s}=1.037\times10^{-4}~$eV\,T$^{-1}$, interfacial DMI constant $D=1.52~$meV, exchange constant $J\ind{ex}=5.72~$meV, uniaxial anisotropy constant $K\ind{u}=0.4~$meV and Gilbert damping parameter $\alpha = 0.1$) are adopted from Ref.~\cite{Rozsa:PRB2018} while the strength of the magnetic moment is chosen to be twice lower than that in \cite{Rozsa:PRB2018}.
The effective magnetic field is calculated as the functional derivative of the Hamiltonian with respect to the normalized magnetization
\begin{align}
\vec{B}_{i}^{\mathrm{eff}}= -\frac{1}{\mu_s}\frac{\delta \mathcal{H}}{\delta\vec{m}_i} + \vec{B}_{i}^{\mathrm{th}}\, ,
\end{align}
in addition to the stochastic thermal field contribution $\vec{B}_{i}^{\mathrm{th}}=\vec{\eta}_i(t)(2\alpha k\ind{B} T / \mu\ind{s}\gamma\Delta t)^{1/2}$. The randomly oriented unit vector $\vec{\eta}$ is reevaluated after each simulation step $\Delta t = 8 \times 10^{-15}\,$s and involves the thermal energy $k\ind{B} T$ (Boltzmann's constant $k\ind{B}$, temperature $T$), hence emulating a stochastic thermal field. If not denoted otherwise, the external parameters are set to $\vec{B}\ind{ext}=1.5\,$T\,$\vec{e}_z$, $T=15\,$K, for which metastable isolated skyrmions are expected to form and move diffusively \cite{schaffer2019stochastic,lindner2020temperature}.
In Ref.~\cite{schaffer2019stochastic} we estimated the time for which a pattern formation in time-integrated measurement techniques stabilizes to be on the scale of $10\,$ns. Hence our simulations are performed over $20\,$ns and the $z-$component of the space-dependent magnetization is averaged over $2.5\times 10^4$ snapshots (cf. Fig.~\ref{fig:simImagTimeEvo}).
Exemplary resulting patterns with different total topological charges can be seen in Fig. \ref{fig:simImagsBeforeNorm}. As expected, the blurred magnetic contrast prohibits an immediate identification of the skyrmion number with the unaided eye, although the true topological charge can be calculated from the spin structure as
\begin{align}
\label{eq:SkyrmionNumber}
Q=\frac{1}{4\pi}\iint \vec{m}(x,y)\cdot\left[\partial_x \vec{m}(x,y)\times\partial_y\vec{m}(x,y) \right] \mathrm{d} x\mathrm{d} y\, ,
\end{align}
for each time-step. This condition is ideal for training an artificial neural network in a supervised learning method. The goal is to overcome the limitations in interpreting time-integrated experimental or simulation data. In the next step, we will elaborate on the design, training and evaluation of the neural network.
\begin{figure}[htb]
\centering
\includegraphics[width=1.0\linewidth]{Fig1.eps}
\caption{Snapshots of the simulated system at different times and the resulting time-averaged image: In three snapshots the number of skyrmions can be identified to be 5 at all times. Determining this number in the time-averaged image is more difficult.}
\label{fig:simImagTimeEvo}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=1.0\linewidth]{Fig2.eps}
\caption{Representative examples of used simulated images with metastable skyrmions for triangular (a), circular (b) and irregular islands (c). The blurred contrast shows the difficulty of predicting the number of skyrmions in a given image. The number of skyrmions is given above the respective figure. The islands are simulated in a rectangle of the size $100\times 100$ atomic sites and the skyrmion size is roughly $15$ atomic sites. Notably, the skyrmion number is assigned correctly by the NN, while counting by hand mostly results in a wrong estimate.}
\label{fig:simImagsBeforeNorm}
\end{figure}
\section{Neural network-driven characterization of skyrmion ensembles}
\subsection{Neural Network}
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth]{Fig3.eps}
\caption{(a) Flowchart showing the network architecture for the used convoluted neural network and scatterplots of the sum over all pixels in a given image: The data set is divided with respect to the number of skyrmions for each configuration. (b) shows the raw data. (c) shows plots where the images are normalized, randomly rescaled and a random noise is added. The size of an individual scatter point scales linearly with the number samples, which exist in its surroundings. Each dot corresponds to an interval with a length of 25 values for (b) and 50 values for (c).}
\label{fig:NNArchAndScatterAfterNorm}
\end{figure*}
The schematics of our machine learning algorithm is depicted in Fig.~\ref{fig:NNArchAndScatterAfterNorm}(a). The target is the topological number $Q$ corresponding to the vertical magnetization $m_z(x,y)$. In the approach of \cite{BERG1981412}, $Q$ is calculated using all components of the simulated atomic spins.
However, this calculation can not be performed here by the NN, because only the $z$-components of the spins are known from the experimental measurements. Here, the NN hence has to predict $Q$ solely on the information given by the time-averaged $z$-projection of the spins $\bar{m}_z(x,y)=\langle m_z(x,y)\rangle_t$.
The $\bar{m}_z(x,y)$ map is used as the input layer. The specifics of the used data augmentation and preprocessing on the input will be explaind later.
Interestingly, this reduced information is insufficient to reconstruct the skyrmion number by mathematical means, i.e., without knowledge about the in-plane rotational sense of the magnetic moments, the skyrmion number in Eq.~(\ref{eq:SkyrmionNumber}) can not be calculated per se. It is hence astonishing that the neural network is capable of correctly assigning the topological number from the out-of-plane magnetization alone. The reason for this success lies in the specifics of the system, which exclusively hosts non-overlapping localized skyrmions with a fixed sense of rotation. By identifying these structures, the network is able to effectively reconstruct the missing directional information and assign the skyrmion number correctly.
The output layer corresponds to a category vector with $8$ entries. Each entry corresponds to a topological number between $1$ and $8$. Since the case $Q=0$ is analogous to an empty island with no magnetic structure after time-integration, the classification is trivial. Therefore, $Q=0$ is omitted from the data set. For the systems examined here, the topological number $Q$ is identical to the number of localized skyrmions in the system, because with the used simulation parameters the skyrmions are well separated.
The maximum value of $Q$ is 8, because when placing more than 8 skyrmions in the island, our simulations show that skyrmions would escape at the edge of the island dynamically. In general, the range of $Q$ can be approximated in advance by dividing the area of the island by the equilibrium skyrmion area and the characteristic spacing between the skyrmions in the system.
The key point is to train the NN to statistically learn the relationship between the time-averaged magnetic contrast along the $z$-axis (Input) and the topological number $Q$ (Output), which is time-independent for the samples evaluated here. The network architecture consists of a convolutional neural network, with three convolutional layers, each followed by a maximum pooling layer and a fully connected layer at the end. The number of filters for the three convolutional layers are 32, 64, and 128, respectively. After each convolution step and after the fully connected layer the ReLU activation function is applied. The output layer consists of 8 neurons, each corresponding to a different amount of skyrmions in the configuration. Categorical cross-entropy is chosen for the loss function
\begin{equation}
\mathcal{L}(\mathbf{y},\mathbf{\hat{y}})=- \frac{1}{N} \sum_{i=0}^N \sum_{j=1}^8 y_{ij}\, ln(\hat{y}_{ij})\ ,
\end{equation}
where $N$ denotes the batch size (amount of data over which one training iteration is performed), $\mathbf{y}$ is the category vector of the labels, and $\mathbf{\hat{y}}$ is the prediction. At the end of the NN the softmax function is applied to the output yielding a probability distribution for $\mathbf{\hat{y}}$
\begin{equation}
\hat{y}_{ij} = \frac{e^{z_{ij}}}{\sum_{k=1}^{8}e^{z_{ik}}}\ .
\end{equation}
$\mathbf{z}$ is the vector corresponding to the last layer. As the final prediction of the network, we choose the Q corresponding to the neuron with the highest output signal. The batch size is set to 32. The Adam optimizer~\cite{kingma2015ICLR} with a learning rate of 0.001, is used to update the weights and biases of the NN. The custom architecture used here is similar to that of LeNet-5~\cite{lecun1998leNet}. A comparable custom architecture has been used for other analyses of magnetic systems~\cite{Kawa:njp2021,Salcedo:JMMM2020}. The network structure was chosen after varying the number of convolutional and fully connected layers, the filter and kernel size, the number of neurons, and the activation function. This architecture resulted in a stable and fast training process.
The NN was implemented in TensorFlow~\cite{tensorflow2015-whitepaper}. The training was performed on Google Colaboratory~\footnote{on Google Colaboratory, available at: https://colab.research.google.com/}.
\subsection{Training}
The simulations are performed for three different kinds of island geometries: Circular, triangular and irregular islands. A sample of the generated data is shown in Fig.~\ref{fig:simImagsBeforeNorm}. The generated data is then split into training and test data sets with 40\% of the data used for testing as shown in table ~\ref{tab:trainingTestQuantities}.
\begin{table}[h]
\caption{\label{tab:trainingTestQuantities}Size of the data sets for three classes of islands without data augmentation}
\centering
\begin{ruledtabular}
\begin{tabular}{ccc}
Island & Size of training set & Size of testing set\\ \hline
Circular & 3622 & 2415\\
Triangular & 3030 & 2035\\
Irregular & 3584 & 2404\\
\end{tabular}
\end{ruledtabular}
\end{table}
Through sample flipping and four-fold rotation, the size of the training data set is increased by a factor of 8.
Since each skyrmion contributes a constant amount of magnetization in the $z$-direction, the simulated data can be classified by only looking at the net magnetization. This is shown in Fig.~\ref{fig:NNArchAndScatterAfterNorm}(b). For each image in the data set, the brightness value of each pixel is summed up. This quantity, which is proportional to the net magnetization, is plotted as a function of the number of skyrmions in the system. The gaps between the data points enable the classification without the NN. To ensure that the data represents experimentally obtained data more closely, different steps of image processing are applied:
\begin{equation}
\Bar{x}_{ij}= \frac{x_{ij} + 0.2 \epsilon_{ij}}{\max({x_{ij}+ 0.2 \epsilon_{ij}})}\cdot (0.8 \eta + 0.2)\ .
\end{equation}
Here, $\mathbf{x}$ and $\mathbf{\Bar{x}}$ are the pixel values before and after the image processing respectively, $\mathbf{\eta}$ is a uniformly distributed random variable with $\eta \in [0,1]$, and $\mathbf{\epsilon}$ is uniformly distributed noise with $\epsilon_{ij} \in [0,1]$.
This forces the NN to utilize the patterns which form in the configurations and impedes the classification on net magnetization. The example images after each step are shown in Fig.~\ref{fig:ImgProcessing}.
After the training process, 100\% of the testing data for the triangular islands are correctly identified. For that, the network is trained for $10$ epochs. One epoch is a training iteration over the whole training data set. For the irregular and the circular islands the network is capable of predicting 99.5\% of the testing data set correctly, as can be seen in the confusion matrix in Fig.~\ref{fig:ConfMatIreg}. The matrix shows the predicted skyrmion number as a function of skyrmions in the system. Due to the regularity of the patterns formed in the triangular and circular islands, conventional machine learning methods such as k-nearest neighbors (k-NN) and support vector machines (SVM) can achieve similar results. For the irregular islands, these methods fall significantly behind our approach based on artificial neural networks. Here, we find a maximal accuracy for k-NN of 32\% and a maximal accuracy for SVM of 61\% when applied to irregularly shaped islands.
\begin{figure}[h!]
\centering
\includegraphics[width=1.0\linewidth]{Fig4.eps}
\caption{Image processing of the data: The images are normalized, randomly rescaled and injected with noise. This is done to prevent a classification solely on net magnetization.}
\label{fig:ImgProcessing}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=1.0\linewidth]{Fig5.eps}
\caption{Confusion matrix of the convolutional neural network for irregular islands: This matrix shows the distribution of the predictions of the network. Each row is normalized, giving a probability distribution for the prediction over all samples with the same number of skyrmions. In total $99.5\%$ of the samples are classified correctly. }
\label{fig:ConfMatIreg}
\end{figure}
\section{Discussion}
In this paper we showed the successful training of an artificial neural network to correctly identify stochastically moving ensembles of skyrmions in confined geometries. The data sets are obtained from classical spin dynamics simulations, mimicking measurement techniques that are sensitive to the out-of-plane component of the magnetization exclusively. Additionally, the time resolution is below the characteristic skyrmion dynamics, which leads to mixtures of blurred out skyrmion traces and localized skyrmions. Though, these patterns do not allow an identification of the true skyrmion number with the unaided eye, nor by analytical means, as this would demand time-resolved data of the magnetization for all three directions. The NN, trained with supervised learning, grants a topological charge identification with vast reliability.
Future experiments on the dynamics of magnetic configurations can benefit from these findings immediately, which help to identify the magnetic configuration at hand, even for highly mobile quasiparticles whose velocity exceeds the experimentally limited time resolution.
Our results show a further possibility of benefiting from state-of-the-art machine learning techniques. Possible next steps include the generalization of the pattern recognition to different non-collinear magnetic textures. Especially coexisting localized and delocalized skyrmions and spin spirals are relevant cases that are hard to characterize with conventional methods.
It should be noted that we train the network for the physical parameters that describes the physical system best. To account for different physical systems, e.g., different materials or strongly altered system size and geometry, the training dataset must be adjusted accordingly. For example, if we doubled the island size and thus decrease the relative skyrmion size, the network trained with the original data fails to predict the correct topological charge Q. The patterns that form in an island twice the size are too different.
Finally, the correct identification of ensembles of skyrmions at finite temperatures lays the basis for realizing skyrmion-based applications in conventional or unconventional computing applications like reservoir, stochastic or neuromorphic computing.
\section*{Acknowledgments}
A.F.S. acknowledges the Universität Hamburg's funding line Next Generation Partnerships funded under the Excellence Strategy of the Federal Government and the Länder, as well as financial support from the German research foundation (DFG) through the collaborative research center CRC/TRR 227. T.P. and R.W. acknowledge funding by the DFG via the Cluster of Excellence ''Advanced Imaging of Matter'' (EXC 2056, Project ID 390715994).
|
\section{Introduction}
In many engineering applications, it is desired to infer the parameters of a filtering system by observing its output. This \textit{inverse filtering} is useful in applications such as system identification, fault detection, image deblurring, and signal deconvolution \cite{idier2013bayesian,gustafsson2007statistical}.
Conventional inverse filtering is limited to non-dynamic systems. However, applications such as cognitive and counter-adversarial systems \cite{haykin2006cognitive,kuptel2017counter,mattila2020inverse}
have recently been shown to require designing the inverse of classical stochastic filters such as hidden Markov model (HMM) filter \cite{elliott2008hidden} and Kalman filter (KF) \cite{kalman1960new}. The cognitive systems are intelligent units that sense the environment, learn relevant information about it, and then adapt themselves in real-time to optimally enhance their performance. For example, a cognitive radar \cite{mishra2020toward} adapts both transmitter and receiver processing in order to achieve desired goals such as improved target detection \cite{mishra2017performance} and tracking \cite{sharaga2015optimal}.
In this context, \cite{krishnamurthy2019how} recently introduced \textit{inverse cognition}, in the form of inverse stochastic filters, to detect cognitive sensor and further estimate the information that the same sensor may have learnt. In this two-part paper, we focus on inverse stochastic filtering for such inverse cognition applications.
At the heart of inverse cognition are two agents: `defender' (e.g., an intelligent target) and an `adversary' (e.g., a sensor or radar) equipped with a Bayesian tracker. The adversary infers an estimate of the defender's kinematic state and cognitively adapts its actions based on this estimate. The defender observes adversary's actions with the goal to predict its future actions in a Bayesian sense.
In particular, \cite{krishnamurthy2020identifying} developed stochastic revealed preferences-based algorithms to ascertain if the adversary's actions are consistent with optimizing a utility function; and if so, estimate that function.
If the defender aims to guard against the adversary's future actions, it requires an estimate of the adversary's inference. This is precisely the objective of inverse Bayesian filtering. In (forward) Bayesian filtering, given noisy observations, a posterior distribution of the underlying state is obtained. An example is the KF, which provides optimal estimates of the underlying state in linear system dynamics with Gaussian measurement and process noises. The inverse filtering problem, on the other hand, is concerned with estimating this posterior distribution of a Bayesian filter given the noisy measurements of the posterior. An example of such a system is the recently introduced inverse Kalman filter (I-KF) \cite{krishnamurthy2019how}. Note that, historically, the Wiener filter -- a special case of KF when the process is stationary -- has long been used for frequency-domain inverse filtering for deblurring in image processing \cite{biemond1990iterative}. Further, some early works \cite{kalman1964linear} have investigated the inverse problem of finding cost criterion for a control policy.
Although KF and its continuous-time variant Kalman-Bucy filter \cite{kalman1961new} are highly effective in many practical applications, they are optimal for only linear and Gaussian models. In practice, many engineering problems involve non-linear processes \cite{haykin2004kalman,simon2006optimal}. In these cases, a \textit{linearized KF} is used, wherein the states of a linear system represent the deviations from a nominal trajectory of a non-linear system. The KF estimates the deviations from the nominal trajectory and obtains an estimate of the states of the non-linear system. The linearized KF is extended to directly estimate the states of a non-linear system in the extended KF (EKF) \cite{schmidt1966application}. The linearization is
locally at the state estimates through Taylor series expansion. This is very similar to the Volterra series filters \cite{zaknich2005principles} that are non-linear counterparts of adaptive linear filters.
While inverse non-linear filters have been studied for adaptive systems in some previous works \cite{broomhead1996nonlinear,shen2001robust}, the inverse of non-linear stochastic filters such as EKF remain unexamined so far. To address the aforementioned non-linear inverse cognition scenarios, contrary to prior works which focus on only linear I-KF \cite{krishnamurthy2019how}, our goal is to derive and analyze inverse EKF (I-EKF). Note that the I-EKF is different from the \textit{inversion of EKF} \cite{zhengyu2021iterated}, which may not take the same form as EKF, is employed on the adversary's side, and is unrelated to our inverse cognition problem. Similarly, the non-linear extended information filter (EIF) proposed in \cite{mutambara1999information} used inverse of covariance matrix and was compared with KF for estimation of the same states. Our inverse EKF is a different formulation that is focused on estimating the inference of an adversary who is also using an EKF to estimate the defender's state.
Preliminary results of this work appeared in our conference publication \cite{singh2022inverse}, where only I-EKF without any unknown inputs was formulated. In this paper, we present inverses of many other EKF formulations for systems with unknown inputs and provide their stability analyses.
The companion paper (Part II) \cite{singh2022inverse_part2} further develops the I-EKF theory to highly non-linear systems where first-order EKF does not sufficiently address the linear approximation. Our main contributions in this paper (Part I) are:\\
\textbf{1) I-KF and I-EKF with unknown inputs.} In the inverse cognition scenario, the target may introduce additional motion or jamming that is known to the target but not to the adversarial cognitive sensor. In this context, while deriving I-EKF, we consider a more general non-linear system model with unknown input. Unknown inputs refer to exogenous excitations to the system which affect the state transition and observations but are not known to the agent employing the stochastic filter.
In the process, we also obtain I-KF-with-unknown-input that was not examined in the I-KF developed in \cite{krishnamurthy2019how}.
Here, similar to the inverse cognition frameworks investigated in \cite{krishnamurthy2019how,mattila2020inverse}, we assume that the adversary's filter is known to the defender.
In the companion paper (Part II) \cite{singh2022inverse_part2}, we consider the case when no prior information about the adversary's filter is available
\\
\textbf{2) Augmented states for I-EKF.} For systems with unknown inputs, the adversary's state estimate depends on its estimate of the unknown input. As a result, the adversary's forward filters vary with system models. We overcome this challenge by considering augmented states in the inverse filter so that the unknown input estimation is performed jointly with state estimation, including for KF with direct feed-through. For different inverse filters, separate augmented states are considered depending on the state transitions for the inverse filter.
\\
\textbf{3) Stability of I-EKF.} The treatment of linear filters includes filter stability and model error sensitivity. But, in general, stability and convergence results for non-linear KFs, and more so for their inverses, are difficult to obtain. In this work, we show the stability of I-EKF using two techniques.
The first approach is based on bounded non-linearities, which has been earlier employed for proving stochastic stability of discrete-time \cite{reif1999stochastic} and continuous-time \cite{reif2000continuousEKFstability} EKFs.
Here, the estimation error was shown to be exponentially bounded in the mean-squared sense.
The second method relaxes the bound on the initial estimation error by introducing unknown matrices to model the linearization errors \cite{xiong2006performance_ukf}. Besides providing the sufficient conditions for error boundedness, this approach also rigorously justifies the enlarging of the noise covariance matrices to stabilize the filter
\cite{wu2007comments}.
Since the I-EKF's error dynamics depends on the forward filter's recursive updates, the derivations of these theoretical guarantees are not straightforward. We validate the estimation errors of all inverse filters through extensive numerical experiments with recursive Cram\'{e}r-Rao lower bound (RCRLB) \cite{tichavsky1998posterior} as the performance metric.
The rest of the paper is organized as follows. In the next section, we provide the background of inverse cognition model. The inverse EKF with unknown input is then derived in Section~\ref{sec:ekfunknown} for the case of the forward EKF with and without direct feed-through. Here, we also obtain the standard I-EKF in the absence of unknown input. Then, similar cases are considered for inverse KF with unknown input in Section~\ref{sec:kfunknown}. We then derive the stability conditions in Section~\ref{sec:stability}. In Section~\ref{sec:simulations}, we corroborate our results with numerical experiments before concluding in Section~\ref{sec:summary}.
Throughout the paper, we reserve boldface lowercase and uppercase letters for vectors (column vectors) and matrices, respectively. The transpose operation and $l_{2}$ norm (for a vector) are denoted by $(\cdot)^T$ and $||\cdot||_{2}$, respectively. The notation $\textrm{Tr}(\mathbf{A})$, $\textrm{rank}(\mathbf{A})$, and $||\mathbf{A}||$, respectively, denote the trace, rank, and spectral norm of $\mathbf{A}$. For matrices $\mathbf{A}$ and $\mathbf{B}$, the inequality $\mathbf{A}\preceq\mathbf{B}$ means that $\mathbf{B}-\mathbf{A}$ is a positive semidefinite (p.s.d.) matrix. For a function $f:\mathbb{R}^{n}\rightarrow\mathbb{R}^{m}$, $\nabla f$ denotes the $\mathbb{R}^{m \times n}$ Jacobian matrix. Similarly, for a function $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$, $\nabla f$ denote the gradient vector ($\mathbb{R}^{n\times 1}$). A $n\times n$ identity matrix is denoted by $\mathbf{I}_{n}$ and a $n\times m$ all zero matrix is denoted by $\mathbf{0}_{n\times m}$. The notation $\lbrace a_{i}\rbrace_{i_{1}\leq i\leq i_{2}}$ denotes a set of elements indexed by integer $i$.
The notation $\mathbf{x} \sim \mathcal{N}(\boldsymbol{\mu},\mathbf{Q})$ and $x \sim \mathcal{U}[u_l,u_u]$, respectively, represent a random variable drawn from a normal distribution with mean $\boldsymbol{\mu}$ and covariance matrix $\mathbf{Q}$, and the uniform distribution over $[u_l,u_u]$.
\begin{comment}
The other direction of designing optimal NLF is the particle filter (PF) \cite{arulampalam2002tutorial,bain2008fundamentals}, which is developed from sequential Monte Carlo method and is applicable to non-linear, non-Gaussian state update and observation equations. The PF can become asymptotically optimal as the number of particles goes to infinity. However, it is difficult to implement PF in real-time (while the observation data keep coming in) because of the computationally expensive Monte-Carlo simulations \cite{ristic2003beyond}. The derivative-free Gaussian filters called sigma-point KFs (SPKFs) form another popular NLF family. They are based on a weighted sum of function evaluations at a finite number of deterministic ``sigma'' points within the domain of integration, as opposite to the stochastic sampling performed by particle filtering methods. These algorithms involve different sigma-points and weights, which are used to approximate the transformation of means and covariances to approximate the posterior probability density in a non-linear system. For example, the unscented KF (UKF) \cite{julier2004unscented} draws points using the Unscented Transform, delivering estimates that are exact in mean for monomials up to third degree, while covariance computation is exact only for linear functions. It assumes the posterior distribution of the state is Gaussian and is, therefore, limited in its applications. The cubature KF (CKF) \cite{arasaratnam2009cubature} uses the third-degree spherical-radial rules, which are exact in mean for all real polynomials up to the third degree and in covariance for linear functions, to numerically approximate the multidimensional integral involved in Bayesian filtering. The quadrature KF (QKF) \cite{arasaratnam2008square,closas2012multiple} is based on Gauss-Hermite quadrature rules. The ensemble KF (EnKF) \cite{houtekamer1998data} tackles weak non-linearities (that is almost linear) and integrates the data assimilation into ensemble generation problem.
\end{comment}
\vspace{-8pt}
\section{Desiderata for Inverse Cognition}
\label{sec:background}
Consider a discrete-time stochastic dynamical system as the defender's state evolution process $\{\mathbf{x}_k\}_{k \geq 0}$, where $\mathbf{x}_k \in \mathbb{R}^{n \times 1}$ is the state at the $k$-th time instant. The defender perfectly knows its current state $\mathbf{x}_{k}$. The control input $\mathbf{u}_k \in \mathbb{R}^{m \times 1}$ is known to the defender but not to the adversary.
In a linear state-space model, we denote the state-transition and control input matrices by $\mathbf{F} \in \mathbb{R}^{n\times n}$ and $\mathbf{B} \in \mathbb{R}^{n\times m}$, respectively. The defender's state evolves as
\par\noindent\small
\begin{align}
\mathbf{x}_{k+1}=\mathbf{Fx}_{k}+\mathbf{Bu}_{k}+\mathbf{w}_{k},\label{eqn: linear x with input}
\end{align}
\normalsize
where $\mathbf{w}_{k}\sim\mathcal{N}(\mathbf{0}_{n\times 1},\mathbf{Q})$ is the process noise with covariance matrix $\mathbf{Q} \in \mathbb{R}^{n\times n}$. At the adversary, the observation and control input matrices are given by $\mathbf{H} \in \mathbb{R}^{p\times n}$ and $\mathbf{D} \in \mathbb{R}^{p\times m}$, respectively. The adversary makes a noisy observation $\mathbf{y}_k \in \mathbb{R}^{p \times 1}$ at time $k$ as
\par\noindent\small
\begin{align}
\mathbf{y}_{k}=\mathbf{Hx}_{k}+\mathbf{Du}_{k}+\mathbf{v}_{k},
\label{eqn: linear y withdf}
\end{align}
\normalsize
where $\mathbf{v}_{k}\sim\mathcal{N}(\mathbf{0}_{p\times 1},\mathbf{R})$ is the adversary's measurement noise with covariance matrix $\mathbf{R} \in \mathbb{R}^{p\times p}$.
The adversary uses $\{\mathbf{y}_j\}_{1 \leq j \leq k}$ to compute the estimate $\hat{\mathbf{x}}_k$ of the defender's state $\mathbf{x}_{k}$ using a (forward) stochastic filter. The adversary then uses this estimate to administer an action matrix $\mathbf{G} \in \mathbb{R}^{n_a \times n}$ on $\hat{\mathbf{x}}_{k}$. The defender makes noisy observations of this action as
\par\noindent\small
\begin{align}
\mathbf{a}_{k}=\mathbf{G}\hat{\mathbf{x}}_{k}+\bm{\epsilon}_{k}\;\; \in \mathbb{R}^{n_{a} \times 1},
\label{eqn: linear a}
\end{align}
\normalsize
where $\bm{\epsilon}_{k}\sim \mathcal{N}(\mathbf{0}_{n_{a}\times 1},\bm{\Sigma_{\epsilon}})$ is the defender's measurement noise with covariance matrix $\bm{\Sigma}_{\epsilon} \in \mathbb{R}^{n_{a}\times n_{a}}$. Finally, the defender uses $\{\mathbf{a}_j, \mathbf{x}_j,\mathbf{u}_{j}\}_{1 \leq j \leq k}$ to compute the estimate $\doublehat{\mathbf{x}}_k \in \mathbb{R}^{n \times 1}$ of $\hat{\mathbf{x}}_k $ in the (inverse) stochastic filter. Define $\hat{\mathbf{u}}_{k}$ to be the estimate of $\mathbf{u}_k$ as computed in the adversary's forward filter, while $\doublehat{\mathbf{u}}_{k}$ is an estimate of $\hat{\mathbf{u}}_{k}$ as computed by the defender's inverse filter. The noise processes $\{\mathbf{w}_{k}\}_{k \geq 0}$, $\{\mathbf{v}_{k}\}_{k \geq 1}$ and $\{\bm{\epsilon}_{k}\}_{k \geq 1}$ are mutually independent and i.i.d. across time. These noise distributions are known to the defender as well as the adversary. When the unknown input is absent, either $\mathbf{B}=\mathbf{0}_{n \times m}$ or $\mathbf{D}=\mathbf{0}_{p \times m}$ or both vanish. Throughout the paper, we assume that both parties (adversary and defender) have perfect knowledge of the system model and parameters.
The companion paper (Part II) \cite{singh2022inverse_part2} considers the case when the perfect knowledge is not available.
When the system dynamics are non-linear, then the matrix pairs $\{\mathbf{F, B}\}$, $\{\mathbf{H, D}\}$, and the matrix $\mathbf{G}$ are replaced by non-linear functions
$f(\cdot, \cdot)$, $h(\cdot,\cdot)$, and $g(\cdot)$, respectively, as
\par\noindent\small
\begin{align}
\mathbf{x}_{k+1}&=f(\mathbf{x}_{k},\mathbf{u}_{k})+\mathbf{w}_{k},\label{eqn: non x with input}\\
\mathbf{y}_{k}&=h(\mathbf{x}_{k},\mathbf{u}_{k})+\mathbf{v}_{k},\label{eqn: non y withdf}\\
\mathbf{a}_{k}&=g(\hat{\mathbf{x}}_{k})+\bm{\epsilon}_{k}.\label{eqn: non a}
\end{align}
\normalsize
This is a \textit{direct feed-through} (DF) model, wherein $\mathbf{y}_{k}$ depends on the unknown input. Without DF, observations \eqref{eqn: non y withdf} becomes
\par\noindent\small
\begin{align}
\mathbf{y}_{k}=h(\mathbf{x}_{k})+\mathbf{v}_{k}.\label{eqn: non y withoutdf}
\end{align}
\normalsize
We show in the following Section~\ref{sec:ekfunknown}, the presence or absence of the unknown input leads to different solution approaches towards forward and inverse filters. For simplicity, the presence of known exogenous inputs is also ignored in state evolution and observations. However, it is trivial to extend the inverse filters developed in this paper for these modifications in the system model. Throughout the paper, we focus on discrete-time models
\vspace{-8pt}
\section{I-EKF with Unknown Input}
\label{sec:ekfunknown}
One of the earliest approach to treat the unknown input was to model the inputs as a stochastic process with known evolution dynamics and jointly estimate the state and inputs. Relaxing the known input dynamics assumption, \cite{kitanidis1987unbiased,gillijns2007unknownkf,gillijns2007kfb,zhang2022boundedness} developed and analyzed unbiased minimum variance linear filters with unknown inputs. Recently, \cite{marco2022regularized,kong2021kalman} have also considered non-persistent and norm-constrained unknown input estimation in linear systems. Various EKF variants to handle unknown inputs in non-linear systems have also been proposed\cite{yang2007adaptive,pan2010applying,xiao2018adaptive,meyer2020unknown,kim2020simultaneous}. We consider a more general EKF with unknown inputs in case of both without \cite{pan2010applying} and with \cite{yang2007adaptive} DF. We do not make any other assumption on the inputs.
The EKF linearizes the model about the nominal values of the state vector and control input. It is similar to the iterated least squares (ILS) method except that the former is for dynamical systems and the latter is not \cite{mendel1995lessons}.
Note that the optimal forward EKFs with and without DF are conceptually different. In the latter case, while the observation $\mathbf{y}_{k}$ is unaffected by the unknown input $\mathbf{u}_{k}$, it is still dependent on $\mathbf{u}_{k-1}$ through $\mathbf{x}_k$; this induces a one-step delay in the adversary's estimate of $\mathbf{u}_k$. On the other hand, with DF,
there is no such delay in estimating $\mathbf{u}_k$.
We now show that this difference results in different inverse filters for these two cases.
\subsection{I-EKF-without-DF unknown input}
\label{subsec:ekfwithoutdf}
Consider the non-linear system without DF given by \eqref{eqn: non x with input} and \eqref{eqn: non y withoutdf}. Linearize the model functions as $\mathbf{F}_{k}\doteq\nabla_{\mathbf{x}}f(\mathbf{x},\hat{\mathbf{u}}_{k-1})|_{\mathbf{x}=\hat{\mathbf{x}}_{k}}$, $\mathbf{B}_{k} \doteq\nabla_{\mathbf{u}}f(\hat{\mathbf{x}}_{k},\mathbf{u})|_{\mathbf{u}=\hat{\mathbf{u}}_{k-1}}$ and $\mathbf{H}_{k+1}\doteq\nabla_{\mathbf{x}}h(\mathbf{x})|_{\mathbf{x}=\hat{\mathbf{x}}_{k+1|k}}$.
\subsubsection{Forward filter}
The forward filter's recursive state estimation procedure first obtains the prediction $\hat{\mathbf{x}}_{k+1|k}$ of the current state using the previous state and input estimates, with $\bm{\Sigma}^{x}_{k+1|k}$ as the associated state prediction error covariance matrix of $\hat{\mathbf{x}}_{k+1|k}$.
Then, the state and input gain matrices $\mathbf{K}^{x}_{k+1}$ and $\mathbf{K}^{u}_{k}$, respectively, are computed along with the input estimation (with delay) covariance matrix $\bm{\Sigma}^{u}_{k}$. Finally, the state $\hat{\mathbf{x}}_{k+1}$, input $\hat{\mathbf{u}}_{k}$, and covariance matrix $\bm{\Sigma}^{x}_{k+1}$ are updated using current observation $\mathbf{y}_{k+1}$, and gain matrices $\mathbf{K}^{x}_{k+1}$ and $\mathbf{K}^{u}_{k}$. Note that the current observation $\mathbf{y}_{k+1}$ provides an estimate $\hat{\mathbf{u}}_{k}$ of the input $\mathbf{u}_{k}$ at the previous time step. The adversary's forward EKF's recursions are\cite{pan2010applying}:
\par\noindent\small
\begin{align}
&\textit{Prediction:}\;\hat{\mathbf{x}}_{k+1|k}=f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k-1}),\label{eqn: ekfwithoutdf predict}\\
&\textit{Gain computation:}\;\bm{\Sigma}^{x}_{k+1|k}=\mathbf{F}_{k}\bm{\Sigma}^{x}_{k}\mathbf{F}_{k}^{T}+\mathbf{Q},\nonumber\\
&\mathbf{K}^{x}_{k+1}=\bm{\Sigma}^{x}_{k+1|k}\mathbf{H}_{k+1}^{T}\left(\mathbf{H}_{k+1}\bm{\Sigma}^{x}_{k+1|k}\mathbf{H}_{k+1}^{T}+\mathbf{R}\right)^{-1},\nonumber\\
&\bm{\Sigma}^{u}_{k}=\left(\mathbf{B}_{k}^{T}\mathbf{H}_{k+1}^{T}\mathbf{R}^{-1}(\mathbf{I}_{p\times p}-\mathbf{H}_{k+1}\mathbf{K}^{x}_{k+1})\mathbf{H}_{k+1}\mathbf{B}_{k}\right)^{-1},\nonumber\\
&\mathbf{K}^{u}_{k}=\bm{\Sigma}^{u}_{k}\mathbf{B}_{k}^{T}\mathbf{H}_{k+1}^{T}\mathbf{R}^{-1}(\mathbf{I}_{p\times p}-\mathbf{H}_{k+1}\mathbf{K}^{x}_{k+1}),\nonumber\\
&\textit{Update:}\;\hat{\mathbf{x}}_{k+1}=\hat{\mathbf{x}}_{k+1|k}+\mathbf{K}^{x}_{k+1}(\mathbf{y}_{k+1}-h(\hat{\mathbf{x}}_{k+1|k})),\label{eqn: ekfwithoutdf update x}\\
&\hat{\mathbf{u}}_{k}=\mathbf{K}^{u}_{k}(\mathbf{y}_{k+1}-h(\hat{\mathbf{x}}_{k+1|k})+\mathbf{H}_{k+1}\mathbf{B}_{k}\hat{\mathbf{u}}_{k-1}),\label{eqn: ekfwithoutdf update u}\\
&\textit{Covariance matrix update:}\;\bm{\Sigma}^{x}_{k+1}\nonumber\\
&=(\mathbf{I}_{n\times n}-\mathbf{K}^{x}_{k+1}\mathbf{H}_{k+1})\left(\bm{\Sigma}^{x}_{k+1|k}+\mathbf{B}_{k}\bm{\Sigma}^{u}_{k}\mathbf{B}_{k}^{T}(\mathbf{I}_{n\times n}-\mathbf{K}^{x}_{k+1}\mathbf{H}_{k+1})^{T}\right).\nonumber
\end{align}
\normalsize
Forward filter exists if $\textrm{rank}(\bm{\Sigma}^{u}_{k})=m$, for all $k\geq 0$, and $p \geq m$ \cite{pan2010applying}.
\subsubsection{Inverse filter}
Consider an augmented state vector $\mathbf{z}_{k}=\begin{bmatrix}
\hat{\mathbf{x}}_{k}^{T} & \hat{\mathbf{u}}_{k-2}^{T}
\end{bmatrix}^{T}$. The defender's observation $\mathbf{a}_{k}$ in \eqref{eqn: non a} is the first observation that contains the information about unknown input estimate $\hat{\mathbf{u}}_{k-2}$, because of the delay in forward filter input estimate. Hence, the delayed estimate $\hat{\mathbf{u}}_{k-2}$ is considered in the augmented state $\mathbf{z}_{k}$. Define $\widetilde{\phi}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k-1},\mathbf{x}_{k+1},\mathbf{v}_{k+1})=f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k-1})-\mathbf{K}^{x}_{k+1}h(f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k-1}))+\mathbf{K}^{x}_{k+1}h(\mathbf{x}_{k+1})+\mathbf{K}^{x}_{k+1}\mathbf{v}_{k+1}$.
From \eqref{eqn: non y withoutdf}
\eqref{eqn: ekfwithoutdf update u}, state transition equations of augmented state vector are
$\hat{\mathbf{x}}_{k+1}=\widetilde{f}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k-2},\hat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{x}_{k+1},\mathbf{v}_{k},\mathbf{v}_{k+1})$ and
$\hat{\mathbf{u}}_{k-1}=\widetilde{h}_{k}(\hat{\mathbf{u}}_{k-2},\hat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{v}_{k})$,
where
\par\noindent\small
\begin{align}
&\widetilde{h}_{k}(\hat{\mathbf{u}}_{k-2},\hat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{v}_{k})\nonumber\\
&\;\;\;\;=\mathbf{K}^{u}_{k-1}(\mathbf{H}_{k}\mathbf{B}_{k-1}\hat{\mathbf{u}}_{k-2}-h(f(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{u}}_{k-2}))+h(\mathbf{x}_{k})+\mathbf{v}_{k}),\label{eqn: state transition ekf without df input}\\
&\widetilde{f}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k-2},\hat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{x}_{k+1},\mathbf{v}_{k},\mathbf{v}_{k+1})\nonumber\\
&\;\;\;\;=\widetilde{\phi}_{k}(\hat{\mathbf{x}}_{k},\widetilde{h}_{k}(\hat{\mathbf{u}}_{k-2},\hat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{v}_{k}),\mathbf{x}_{k+1},\mathbf{v}_{k+1}).\label{eqn: state transition ekf without df state}
\end{align}
\normalsize
In these state transition equations, the actual states $\mathbf{x}_{k}$ and $\mathbf{x}_{k+1}$ are perfectly known to the defender and henceforth treated as known exogenous inputs. Note that, unlike the forward filter, the process noise terms $\mathbf{v}_{k}$ and $\mathbf{v}_{k+1}$ are non-additive because the filter gains $\mathbf{K}^{x}_{k+1}$ and $\mathbf{K}^{u}_{k-1}$ depend on the previous estimates (through the Jacobians).
Denote $\hat{\mathbf{z}}_{k+1} \doteq \begin{bmatrix}
\doublehat{\mathbf{x}}_{k+1}^{T} & \doublehat{\mathbf{u}}_{k-1}^{T}
\end{bmatrix}^{T}$. The state transition of the augmented state $\mathbf{z}_{k+1}$ depends on the estimate $\hat{\mathbf{x}}_{k-1}$ which the defender approximates by its previous estimate $\doublehat{\mathbf{x}}_{k-1}$. With this approximation, $\hat{\mathbf{x}}_{k-1}$ is treated as a known exogenous input for the inverse filter while the augmented process noise vector is $\begin{bmatrix}
\mathbf{v}_{k}^{T} & \mathbf{v}_{k+1}^{T}
\end{bmatrix}^{T}$. Define the Jacobians $\widetilde{\mathbf{F}}^{z}_{k}\doteq \begin{bmatrix}
\nabla_{\doublehat{\mathbf{x}}_{k}}\widetilde{f}_{k} & \nabla_{\doublehat{\mathbf{u}}_{k-2}}\widetilde{f}_{k}\\
\mathbf{0}_{m\times n} & \nabla_{\doublehat{\mathbf{u}}_{k-2}}\widetilde{h}_{k}
\end{bmatrix}$, and $\mathbf{G}_{k+1}\doteq\begin{bmatrix}
\nabla_{\doublehat{\mathbf{x}}_{k+1|k}}g & \mathbf{0}_{n_{a}\times m}\end{bmatrix}$ with respect to the augmented state; Jacobian $\widetilde{\mathbf{F}}^{v}_{k}\doteq\begin{bmatrix}
\nabla_{\mathbf{v}_{k}}\widetilde{f}_{k} & \nabla_{\mathbf{v}_{k+1}}\widetilde{f}_{k}\\
\nabla_{\mathbf{v}_{k}}\widetilde{h}_{k} & \mathbf{0}_{m\times p}
\end{bmatrix}$ with respect to the augmented process noise vector; and $\overline{\mathbf{Q}}_{k}=\widetilde{\mathbf{F}}^{v}_{k}\begin{bmatrix}
\mathbf{R} & \mathbf{0}_{p\times p}\\ \mathbf{0}_{p\times p} & \mathbf{R}
\end{bmatrix}(\widetilde{\mathbf{F}}^{v}_{k})^{T}$. Then, the I-EKF-without-DF's recursions yield the estimate $\hat{\mathbf{z}}_{k}$ of the augmented state and the associated covariance matrix $\overline{\bm{\Sigma}}_{k}$
as:
\par\noindent\small
\begin{align}
&\textit{Prediction:}\;\doublehat{\mathbf{x}}_{k+1|k}=\widetilde{f}_{k}(\doublehat{\mathbf{x}}_{k},\doublehat{\mathbf{u}}_{k-2},\doublehat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{x}_{k+1},\mathbf{0}_{p\times 1},\mathbf{0}_{p\times 1}),\nonumber\\
&\doublehat{\mathbf{u}}_{k-1|k}=\widetilde{h}_{k}(\doublehat{\mathbf{u}}_{k-2},\doublehat{\mathbf{x}}_{k-1},\mathbf{x}_{k},\mathbf{0}_{p\times 1}),\nonumber\\
&\hat{\mathbf{z}}_{k+1|k}=\begin{bmatrix}
\doublehat{\mathbf{x}}_{k+1|k}^{T} & \doublehat{\mathbf{u}}_{k-1|k}^{T}
\end{bmatrix}^{T},\nonumber\\
&\overline{\bm{\Sigma}}_{k+1|k}=\widetilde{\mathbf{F}}^{z}_{k}\overline{\bm{\Sigma}}_{k}(\widetilde{\mathbf{F}}^{z}_{k})^{T}+\overline{\mathbf{Q}}_{k},\label{eqn: I-EKF without DF covariance predict}\\
&\textit{Update:}\;\overline{\mathbf{S}}_{k+1}=\mathbf{G}_{k+1}\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}+\bm{\Sigma}_{\epsilon},\label{eqn: I-EKF without DF S compute}\\
&\hat{\mathbf{z}}_{k+1}=\hat{\mathbf{z}}_{k+1|k}+\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\left(\mathbf{a}_{k+1}-g(\doublehat{\mathbf{x}}_{k+1|k})\right),\label{eqn: I-EKF without DF a predict}\\
&\overline{\bm{\Sigma}}_{k+1}=\overline{\bm{\Sigma}}_{k+1|k}-\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\mathbf{G}_{k+1}\overline{\bm{\Sigma}}_{k+1|k}.\label{eqn: I-EKF without DF covariance update}
\end{align}
\normalsize
The I-EKF-without-DF's recursions take the same form as that of the standard EKF \cite{anderson2012optimal} but with modified system matrices. In particular, the former employs an augmented state such that the Jacobian of the state transition function with respect to the state is computed as $\widetilde{\mathbf{F}}^{z}_{k}$ while for the latter, it is simply $\mathbf{F}_{k}\doteq\nabla_{\mathbf{x}}f(\mathbf{x})\vert_{\mathbf{x}=\hat{\mathbf{x}}_{k}}$. Further, unlike standard KF or EKF, the noise terms, i.e., {$\mathbf{v}_{k}$ and $\mathbf{v}_{k+1}$ in \eqref{eqn: state transition ekf without df input} and \eqref{eqn: state transition ekf without df state} are non-additive such that linearization $\widetilde{\mathbf{F}}^{v}_{k}$ of the state transition function with respect to the noise terms yields the process noise covariance matrix approximation $\overline{\mathbf{Q}}_{k}$.
The forward filter gains $\mathbf{K}^{x}_{k+1}$ and $\mathbf{K}^{u}_{k-1}$ are treated as time-varying parameters of the state transition equation and not as a function of the state and input estimates ($\hat{\mathbf{x}}_{k}$ and $\hat{\mathbf{u}}_{k-1}$) in the inverse filter. The inverse filter approximates them by evaluating their values at its own estimates ($\doublehat{\mathbf{x}}_{k}$ and $\doublehat{\mathbf{u}}_{k-1}$) recursively in the similar manner as the forward filter evaluates them using its own estimates. On the contrary, in I-KF formulation introduced in \cite{krishnamurthy2019how}, the forward Kalman gain $\mathbf{K}_{k+1}$ is deterministic, fully determined by the model parameters for a given initial covariance estimate $\bm{\Sigma}_{0}$, and computed offline independent of the current I-KF's estimate.
\subsection{I-EKF-with-DF unknown input}
\label{subsec:ekfwithdf}
Consider the non-linear system with DF given by \eqref{eqn: non x with input} and \eqref{eqn: non y withdf}. Linearize the functions as $\mathbf{F}_{k}\doteq\nabla_{\mathbf{x}}f(\mathbf{x},\hat{\mathbf{u}}_{k})|_{\mathbf{x}=\hat{\mathbf{x}}_{k}}$, $\mathbf{H}_{k+1}\doteq\nabla_{\mathbf{x}}h(\mathbf{x},\hat{\mathbf{u}}_{k})|_{\mathbf{x}=\hat{\mathbf{x}}_{k+1|k}}$ and $\mathbf{D}_{k}\doteq\nabla_{\mathbf{u}}h(\hat{\mathbf{x}}_{k+1|k},\mathbf{u})|_{\mathbf{u}=\hat{\mathbf{u}}_{k}}$.
\subsubsection{Forward filter}
Denote the state and input estimation covariance and gain matrices identical to Section~\ref{subsec:ekfwithoutdf}. Here, the current observation $\mathbf{y}_{k+1}$ depends on the current unknown input $\mathbf{u}_{k+1}$ such that the forward filter infers $\hat{\mathbf{u}}_{k+1}$ without any delay. For input estimation covariance without delay, we use $\bm{\Sigma}^{u}_{k+1}$.
Then, the forward EKF-with-DF's recursions
are \cite{yang2007adaptive}
\par\noindent\small
\begin{align}
&\textit{Prediction:}\;\hat{\mathbf{x}}_{k+1|k}=f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k}),\;\bm{\Sigma}^{x}_{k+1|k}=\mathbf{F}_{k}\bm{\Sigma}^{x}_{k}\mathbf{F}_{k}^{T}+\mathbf{Q},\label{eqn: ekfwithdf predict}\\
&\mathbf{K}^{x}_{k+1}=\bm{\Sigma}^{x}_{k+1|k}\mathbf{H}_{k+1}^{T}(\mathbf{H}_{k+1}\bm{\Sigma}^{x}_{k+1|k}\mathbf{H}_{k+1}^{T}+\mathbf{R})^{-1},\nonumber\\
&\bm{\Sigma}^{u}_{k+1}=\left(\mathbf{D}_{k}^{T}\mathbf{R}^{-1}(\mathbf{I}_{p\times p}-\mathbf{H}_{k+1}\mathbf{K}^{x}_{k+1})\mathbf{D}_{k}\right)^{-1},\nonumber\\
&\mathbf{K}^{u}_{k+1}=\bm{\Sigma}^{u}_{k+1}\mathbf{D}_{k}^{T}\mathbf{R}^{-1}(\mathbf{I}_{p\times p}-\mathbf{H}_{k+1}\mathbf{K}^{x}_{k+1}),\nonumber\\
&\textit{Update:}\;\hat{\mathbf{u}}_{k+1}=\mathbf{K}^{u}_{k+1}\left(\mathbf{y}_{k+1}-h(\hat{\mathbf{x}}_{k+1|k},\hat{\mathbf{u}}_{k})+\mathbf{D}_{k}\hat{\mathbf{u}}_{k}\right),\label{eqn: ekfwithdf update u}\\
&\hat{\mathbf{x}}_{k+1}=\hat{\mathbf{x}}_{k+1|k}+\mathbf{K}^{x}_{k+1}\left(\mathbf{y}_{k+1}-h(\hat{\mathbf{x}}_{k+1|k},\hat{\mathbf{u}}_{k})-\mathbf{D}_{k}(\hat{\mathbf{u}}_{k+1}-\hat{\mathbf{u}}_{k})\right),\label{eqn: ekfwithdf update x}\\
&\textit{Covariance matrix update:}\;\bm{\Sigma}^{x}_{k+1}\nonumber\\
&\hspace{-0.4mm}=(\mathbf{I}_{n\times n}+\mathbf{K}^{x}_{k+1}\mathbf{D}_{k}\bm{\Sigma}^{u}_{k+1}\mathbf{D}_{k}^{T}\mathbf{R}^{-1}\mathbf{H}_{k+1})(\mathbf{I}_{n\times n}-\mathbf{K}^{x}_{k+1}\mathbf{H}_{k+1})\bm{\Sigma}^{x}_{k+1|k}.\nonumber
\end{align}
\normalsize
The forward filter exists if $\textrm{rank}(\mathbf{D}_{k})=m$ for all $k\geq 0$, which implies $p\geq m$\cite{yang2007adaptive}.
\subsubsection{Inverse filter}
Consider an augmented state vector $\mathbf{z}_{k}=\begin{bmatrix}
\hat{\mathbf{x}}_{k}^{T} & \hat{\mathbf{u}}_{k}^{T}
\end{bmatrix}^{T}$ (note the absence of delay in the input estimate). Define $\widetilde{\phi}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\hat{\mathbf{u}}_{k+1},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1})=f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k})-\mathbf{K}^{x}_{k+1}h(f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k}),\hat{\mathbf{u}}_{k})-\mathbf{K}^{x}_{k+1}\mathbf{D}_{k}(\hat{\mathbf{u}}_{k+1}-\hat{\mathbf{u}}_{k})+\mathbf{K}^{x}_{k+1}h(\mathbf{x}_{k+1},\mathbf{u}_{k+1})+\mathbf{K}^{x}_{k+1}\mathbf{v}_{k+1}$.
From \eqref{eqn: non y withdf} and \eqref{eqn: ekfwithdf predict}-\eqref{eqn: ekfwithdf update x}, state transitions for inverse filter are
$\hat{\mathbf{x}}_{k+1}=\widetilde{f}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1})$ and
$\hat{\mathbf{u}}_{k+1}=\widetilde{h}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1})$,
where
\par\noindent\small
\begin{align}
&\widetilde{h}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1})\nonumber\\
&=\mathbf{K}^{u}_{k+1}(h(\mathbf{x}_{k+1},\mathbf{u}_{k+1})+\mathbf{v}_{k+1}-h(f(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k}),\hat{\mathbf{u}}_{k})+\mathbf{D}_{k}\hat{\mathbf{u}}_{k})\nonumber\\
&\widetilde{f}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1})\nonumber\\
&=\widetilde{\phi}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\widetilde{h}_{k}(\hat{\mathbf{x}}_{k},\hat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1}),\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{v}_{k+1}).\label{eqn:state transition ekf with df}
\end{align}
\normalsize
Then, \textit{ceteris paribus}, following similar steps as in I-EKF-without-DF,
the I-EKF-with-DF estimate $\hat{\mathbf{z}}_{k}=\begin{bmatrix}
\doublehat{\mathbf{x}}_{k}^{T} & \doublehat{\mathbf{u}}_{k}^{T}
\end{bmatrix}^{T}$ from observations \eqref{eqn: non a} is computed recursively. The predicted augmented state is $\hat{\mathbf{z}}_{k+1|k}=\begin{bmatrix}
\doublehat{\mathbf{x}}_{k+1|k}^{T} & \doublehat{\mathbf{u}}_{k+1|k}^{T}
\end{bmatrix}^{T}$, where
$\doublehat{\mathbf{x}}_{k+1|k}=\widetilde{f}_{k}(\doublehat{\mathbf{x}}_{k},\doublehat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{0}_{p\times 1})$ and
$\doublehat{\mathbf{u}}_{k+1|k}=\widetilde{h}_{k}(\doublehat{\mathbf{x}}_{k},\doublehat{\mathbf{u}}_{k},\mathbf{x}_{k+1},\mathbf{u}_{k+1},\mathbf{0}_{p\times 1})$.
Hereafter, the remaining steps are as in \eqref{eqn: I-EKF without DF covariance predict}-\eqref{eqn: I-EKF without DF covariance update}. For I-EKF-with-DF, the Jacobians with respect to the augmented state are $\widetilde{\mathbf{F}}^{z}_{k}\doteq\begin{bmatrix}
\nabla_{\doublehat{\mathbf{x}}_{k}}\widetilde{f}_{k} & \nabla_{\doublehat{\mathbf{u}}_{k}}\widetilde{f}_{k} \\
\nabla_{\doublehat{\mathbf{x}}_{k}}\widetilde{h}_{k} & \nabla_{\doublehat{\mathbf{u}}_{k}}\widetilde{h}_{k}
\end{bmatrix}$ and $\mathbf{G}_{k+1}\doteq\begin{bmatrix}
\nabla_{\doublehat{\mathbf{x}}_{k+1|k}}g & \mathbf{0}_{n_{a}\times m}\end{bmatrix}$; the Jacobian with respect to the process noise term is $\widetilde{\mathbf{F}}^{v}_{k}\doteq\begin{bmatrix}
\nabla_{\mathbf{v}_{k+1}}\widetilde{f}_{k}\\
\nabla_{\mathbf{v}_{k+1}}\widetilde{h}_{k}
\end{bmatrix}$; and $\overline{\mathbf{Q}}_{k}=\widetilde{\mathbf{F}}^{v}_{k}\mathbf{R}(\widetilde{\mathbf{F}}^{v}_{k})^{T}$. Here, unlike I-EKF-without-DF, the inverse filter's prediction dispenses with any approximation of $\hat{\mathbf{x}}_{k-1}$. The absence of delay in input estimation also results in a simplified process noise term $\mathbf{v}_{k+1}$, in place of I-EKF-without-DF's augmented noise vector.
Examples of EKF with unknown inputs include fault detection with unknown excitations \cite{yang2007adaptive}
and missile-target interception with unknown target acceleration \cite{pan2010applying}. The inverse cognition in these applications would then resort to the I-EKFs described until now.
\begin{comment}
\begin{align*}
&\textit{Prediction:}\\
\\
&\overline{\bm{\Sigma}}_{k+1|k}=\widetilde{\mathbf{F}}^{z}_{k}\overline{\bm{\Sigma}}_{k}(\widetilde{\mathbf{F}}^{z}_{k})^{T}+\overline{\mathbf{Q}}_{k},\\
&\textit{Update:}\\
&\overline{\mathbf{S}}_{k+1}=\mathbf{G}_{k+1}\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}+\bm{\Sigma}_{\epsilon},\\
&\hat{\mathbf{z}}_{k+1}=\hat{\mathbf{z}}_{k+1|k}+\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\left(\mathbf{a}_{k+1}-g(\doublehat{\mathbf{x}}_{k+1|k})\right),\\
&\overline{\bm{\Sigma}}_{k+1}=\overline{\bm{\Sigma}}_{k+1|k}-\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\mathbf{G}_{k+1}\overline{\bm{\Sigma}}_{k+1|k},
\end{align*}
\end{comment}
\subsection{I-EKF without any unknown inputs}
\label{subsec:ekf}
Consider a non-linear system model without unknown inputs in the system equations \eqref{eqn: non x with input} and \eqref{eqn: non y withoutdf}, i.e.
\par\noindent\small
\begin{align}
\mathbf{x}_{k+1}=f(\mathbf{x}_{k})+\mathbf{w}_{k}.\label{eqn: ekf x}
\end{align}
\normalsize
Linearize the functions as $\mathbf{F}_{k}\doteq\nabla_{\mathbf{x}}f(\mathbf{x})\vert_{\mathbf{x}=\hat{\mathbf{x}}_{k}}$ and
$\mathbf{H}_{k+1}\doteq\nabla_{\mathbf{x}}h(\mathbf{x})\vert_{\mathbf{x}=\hat{\mathbf{x}}_{k+1|k}}$.
Then, \textit{ceteris paribus}, setting $\mathbf{B}_{k}=\mathbf{0}_{n\times p}$ and neglecting computation of $\bm{\Sigma}^{u}_{k}$, $\mathbf{K}^{u}_{k}$ and $\hat{\mathbf{u}}_{k}$ in forward EKF-without-DF yields forward EKF-without-unknown-input whose state prediction and updates are
\par\noindent\small
\begin{align}
&\hat{\mathbf{x}}_{k+1|k}=f(\hat{\mathbf{x}}_{k}),\label{eqn: ekf predict}\\
&\hat{\mathbf{x}}_{k+1}=\hat{\mathbf{x}}_{k+1|k}+\mathbf{K}_{k+1}(\mathbf{y}_{k+1}-h(\hat{\mathbf{x}}_{k+1|k})),\label{eqn: ekf update}
\end{align}
\normalsize
with $\mathbf{K}_{k+1}=\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\left(\mathbf{H}_{k+1}\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}+\mathbf{R}\right)^{-1}$. Here, we have dropped the superscript in the covariance matrix $\bm{\Sigma}^{x}_{k+1|k}$ and gain $\mathbf{K}^{x}_{k+1}$ to replace with $\bm{\Sigma}_{k+1|k}$ and $\mathbf{K}_{k+1}$, respectively (because only the state estimation covariances and gains are computed here). Thence, the I-EKF-without-DF's state transition equations and recursions yield I-EKF-without-unknown-input. Dropping the input estimate term in the augmented state $\mathbf{z}_{k}$, the state transition equations become
\par\noindent\small
\begin{align}
&\hat{\mathbf{x}}_{k+1}=\widetilde{f}_{k}(\hat{\mathbf{x}}_{k},\mathbf{x}_{k+1},\mathbf{v}_{k+1})\nonumber\\
&=f(\hat{\mathbf{x}}_{k})-\mathbf{K}_{k+1}h(f(\hat{\mathbf{x}}_{k}))+\mathbf{K}_{k+1}h(\mathbf{x}_{k+1})+\mathbf{K}_{k+1}\mathbf{v}_{k+1}.\label{eqn: inverse ekf state transition}
\end{align}
\normalsize
Denote $\widetilde{\mathbf{F}}^{x}_{k}\doteq\nabla_{\mathbf{x}}\widetilde{f}_{k}(\mathbf{x},\mathbf{x}_{k+1},\mathbf{0}_{p\times 1})|_{\mathbf{x}=\doublehat{\mathbf{x}}_{k}}$, $\mathbf{G}_{k+1}\doteq\nabla_{\mathbf{x}}g(\mathbf{x})_{\mathbf{x}=\doublehat{\mathbf{x}}_{k+1|k}}$, $\widetilde{\mathbf{F}}^{v}_{k}\doteq\nabla_{\mathbf{v}}\widetilde{f}_{k}(\doublehat{\mathbf{x}}_{k},\mathbf{x}_{k+1},\mathbf{v})|_{\mathbf{v}=\mathbf{0}_{p\times 1}}$, and $\overline{\mathbf{Q}}_{k}=\widetilde{\mathbf{F}}^{v}_{k}\mathbf{R}(\widetilde{\mathbf{F}}^{v}_{k})^{T}$. Then, the I-EKF's recursions are similar to I-EKF-without-DF except that the I-EKF's predicted state estimate and the associated prediction covariance matrix are computed, respectively, as
$\doublehat{\mathbf{x}}_{k+1|k}=\widetilde{f}_{k}(\doublehat{\mathbf{x}}_{k},\mathbf{x}_{k+1},\mathbf{0}_{p\times 1})$ and
$\overline{\bm{\Sigma}}_{k+1|k}=\widetilde{\mathbf{F}}^{x}_{k}\overline{\bm{\Sigma}}_{k}(\widetilde{\mathbf{F}}^{x}_{k})^{T}+\overline{\mathbf{Q}}_{k}$,
followed by the update procedure in \eqref{eqn: I-EKF without DF S compute}-\eqref{eqn: I-EKF without DF covariance update}.
Unlike I-KF \cite{krishnamurthy2019how}, the I-EKF approximates the forward gain $\mathbf{K}_{k+1}$ online at its own estimates recursively and is sensitive to the initial estimate of forward EKF's initial covariance matrix. I-EKF could be applied in various non-linear target tracking applications, where EKF is a popular forward filter\cite{ristic2003beyond}.
The two-step prediction-update formulation (as discussed for EKF and I-EKF so far) infers an estimate of the current state. However, often for stability analyses, the one-step prediction formulation is analytically more useful. In this formulation, the estimate $\hat{\mathbf{x}}_{k}$ is the one-step prediction estimate, i.e., an estimate of state $\mathbf{x}_{k}$ at $k$-th instant given the observations $\lbrace\mathbf{y}_{j}\rbrace_{1\leq j\leq k-1}$ up to time instant $k-1$ with $\bm{\Sigma}_{k}$ as the corresponding prediction covariance matrix. The forward one-step prediction EKF formulation\cite{reif1999stochastic} for the same system but with $\mathbf{F}_{k}\doteq\nabla_{\mathbf{x}}f(\mathbf{x})\vert_{\mathbf{x}=\hat{\mathbf{x}}_{k}}$ and $\mathbf{H}_{k}\doteq\nabla_{\mathbf{x}}h(\mathbf{x})\vert_{\mathbf{x}=\hat{\mathbf{x}}_{k}}$ is
\par\noindent\small
\begin{align}
&\mathbf{K}_{k}=\mathbf{F}_{k}\bm{\Sigma}_{k}\mathbf{H}_{k}^{T}(\mathbf{H}_{k}\bm{\Sigma}_{k}\mathbf{H}_{k}^{T}+\mathbf{R})^{-1},\label{eqn: one step forward ekf gain}\\
&\hat{\mathbf{x}}_{k+1}=f(\hat{\mathbf{x}}_{k})+\mathbf{K}_{k}(\mathbf{y}_{k}-h(\hat{\mathbf{x}}_{k})),\label{eqn: one step forward ekf update}\\
&\bm{\Sigma}_{k+1}=\mathbf{F}_{k}\bm{\Sigma}_{k}\mathbf{F}_{k}^{T}+\mathbf{Q}-\mathbf{K}_{k}(\mathbf{H}_{k}\bm{\Sigma}_{k}\mathbf{H}_{k}^{T}+\mathbf{R})\mathbf{K}_{k}^{T}.\label{eqn: one step forward ekf covariance}
\end{align}
\normalsize
From \eqref{eqn: non y withoutdf} and \eqref{eqn: one step forward ekf update}, the state transition equation for one-step formulation of I-EKF is $\hat{\mathbf{x}}_{k+1}=\widetilde{f}_{k}(\hat{\mathbf{x}}_{k},\mathbf{x}_{k},\mathbf{v}_{k})\doteq f(\hat{\mathbf{x}}_{k})-\mathbf{K}_{k}h(\hat{\mathbf{x}}_{k})+\mathbf{K}_{k}h(\mathbf{x}_{k})+\mathbf{K}_{k}\mathbf{v}_{k}$.
With this state transition, the I-EKF one-step prediction formulation follows directly from EKF's one-step prediction formulation treating $\mathbf{a}_{k}$ as the observation with the Jacobians with respect to state estimate $\widetilde{\mathbf{F}}^{x}_{k}=\nabla_{\mathbf{x}}\widetilde{f}_{k}(\mathbf{x},\mathbf{x}_{k},\mathbf{0})\vert_{\mathbf{x}=\doublehat{\mathbf{x}}_{k}}=\mathbf{F}_{k}-\mathbf{K}_{k}\mathbf{H}_{k}$ and $\mathbf{G}_{k}=\nabla_{\mathbf{x}}g(\mathbf{x})\vert_{\mathbf{x}=\doublehat{\mathbf{x}}_{k}}$, and the process noise covariance matrix $\overline{\mathbf{Q}}_{k}=\mathbf{K}_{k}\mathbf{R}\mathbf{K}_{k}^{T}$.
\begin{comment}
\subsubsection{Forward filter}
IEKF one step
represented by the following recursive equations.
\par\noindent\small
\begin{align*}
&\overline{\mathbf{K}}_{k}=\widetilde{\mathbf{F}}_{k}\overline{\bm{\Sigma}}_{k}\mathbf{G}_{k}^{T}\left(\mathbf{G}_{k}\overline{\bm{\Sigma}}_{k}\mathbf{G}_{k}^{T}+\overline{\mathbf{R}}_{k}\right)^{-1},\\
&\doublehat{\mathbf{x}}_{k+1}=\widetilde{f}_{k}(\doublehat{\mathbf{x}}_{k},\mathbf{x}_{k},\mathbf{0})+\overline{\mathbf{K}}_{k}(\mathbf{a}_{k}-g(\doublehat{\mathbf{x}}_{k})),\\
&\overline{\bm{\Sigma}}_{k+1}=\widetilde{\mathbf{F}}_{k}\overline{\bm{\Sigma}}_{k}\widetilde{\mathbf{F}}_{k}^{T}+\overline{\mathbf{Q}}_{k}-\overline{\mathbf{K}}_{k}(\mathbf{G}_{k}\overline{\bm{\Sigma}}_{k}\mathbf{G}_{k}^{T}+\overline{\mathbf{R}}_{k})\overline{\mathbf{K}}_{k}^{T},
\end{align*}
\normalsize
where , and .
\label{subsubsec:forward ekf}
The adversary employs EKF as the forward filter to estimate our state a
\par\noindent\small
\begin{align}
&\textit{Prediction:}\nonumber\\
&\hat{\mathbf{x}}_{k+1|k}=f(\hat{\mathbf{x}}_{k}),\label{eqn: ekf predict}\\
&\bm{\Sigma}_{k+1|k}=\mathbf{F}_{k}\bm{\Sigma}_{k}\mathbf{F}_{k}^{T}+\mathbf{Q},\nonumber\\
&\textit{Update:}\nonumber\\
&\mathbf{S}_{k+1}=\mathbf{H}_{k+1}\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}+\mathbf{R},\nonumber\\
&\hat{\mathbf{x}}_{k+1}=\hat{\mathbf{x}}_{k+1|k}+\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{S}_{k+1}^{-1}(\mathbf{y}_{k+1}-h(\hat{\mathbf{x}}_{k+1|k})),\label{eqn: ekf update}\\
&\bm{\Sigma}_{k+1}=\bm{\Sigma}_{k+1|k}-\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{S}_{k+1}^{-1}\mathbf{H}_{k+1}\bm{\Sigma}_{k+1|k},\nonumber
\end{align}
\normalsize
where .
\subsubsection{Inverse filter}
\label{subsubsec:inverse ekf}
The state transition equation for the I-EKF is obtained from \eqref{eqn: ekf y}-\eqref{eqn: ekf update} as
This yields the recursive form of I-EKF as
\par\noindent\small
\begin{align*}
&\textit{Prediction:}\\
&\textit{Update:}\\
&\overline{\mathbf{S}}_{k+1}=\mathbf{G}_{k+1}\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}+\bm{\Sigma}_{\epsilon},\\
&\doublehat{\mathbf{x}}_{k+1}=\doublehat{\mathbf{x}}_{k+1|k}+\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}\overline{\mathbf{S}}_{k+1}^{-1}(\mathbf{a}_{k+1}-g(\doublehat{\mathbf{x}}_{k+1|k})),\\
&\overline{\bm{\Sigma}}_{k+1}=\overline{\bm{\Sigma}}_{k+1|k}-\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}_{k+1}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\mathbf{G}_{k+1}\overline{\bm{\Sigma}}_{k+1|k},
\end{align*}
\normalsize
where .
\end{comment}
\vspace{-8pt}
\section{Inverse KF with unknown input}
\label{sec:kfunknown}
For linear Gaussian state-space models, our methods developed in the previous section are useful in extending the I-KF mentioned in \cite{krishnamurthy2019how} to unknown input. Again, the forward KFs employed by the adversary with and without DF are conceptually different \cite{gillijns2007kfb} because of the delay involved in input estimation. The forward KFs with unknown input provide unbiased minimum variance state and input estimates.
\subsection{I-KF-without-DF}
\label{subsec:kfwithoutdf}
Consider the
system in \eqref{eqn: linear x with input} and \eqref{eqn: linear y withdf} with $\mathbf{D}=\mathbf{0}_{p\times m}$.
\subsubsection{Forward filter}
Unlike EKF-without-DF, the forward KF-without-DF considers an intermediate state update step using the estimated unknown input before the final state updates. In this step, the unknown input is first estimated (with one-step delay) using the current observation $\mathbf{y}_{k+1}$ and input estimation gain matrix $\mathbf{M}_{k+1}$.
In the update step, the current state estimate $\hat{\mathbf{x}}_{k+1}$ is computed by again considering the current observation $\mathbf{y}_{k+1}$ as\cite{gillijns2007unknownkf}
\par\noindent\small
\begin{align}
&\textit{Prediction:}\;\hat{\mathbf{x}}_{k+1|k}=\mathbf{F}\hat{\mathbf{x}}_{k},\;\bm{\Sigma}_{k+1|k}=\mathbf{F}\bm{\Sigma}_{k}\mathbf{F}^{T}+\mathbf{Q},\label{eqn: kfwithoutdf predict}\\
&\textit{Unknown input estimation:}\;\mathbf{S}_{k+1}=\mathbf{H}\bm{\Sigma}_{k+1|k}\mathbf{H}^{T}+\mathbf{R},\\
&\bm{M}_{k+1}=(\mathbf{B}^{T}\mathbf{H}^{T}\mathbf{S}_{k+1}^{-1}\mathbf{HB})^{-1}\mathbf{B}^{T}\mathbf{H}^{T}\mathbf{S}_{k+1}^{-1},\\
&\hat{\mathbf{u}}_{k}=\mathbf{M}_{k+1}(\mathbf{y}_{k+1}-\mathbf{H}\hat{\mathbf{x}}_{k+1|k}),\label{eqn: kfwithoutdf update u}\\
&\widetilde{\mathbf{x}}_{k+1|k+1}=\hat{\mathbf{x}}_{k+1|k}+\mathbf{B}\hat{\mathbf{u}}_{k},\label{eqn: kfwithoutdf update x with u}\\
&\widetilde{\bm{\Sigma}}_{k+1|k+1}=(\mathbf{I}_{n\times n}-\mathbf{BM}_{k+1}\mathbf{H})\bm{\Sigma}_{k+1|k}(\mathbf{I}_{n\times n}-\mathbf{BM}_{k+1}\mathbf{H})^{T}\nonumber\\
&\hspace{1.5cm}+\mathbf{BM}_{k+1}\mathbf{RM}_{k+1}^{T}\mathbf{B}^{T},\\
&\textit{Update:}\;\mathbf{K}_{k+1}=\bm{\Sigma}_{k+1|k}\mathbf{H}^{T}\mathbf{S}_{k+1}^{-1},\\
&\hat{\mathbf{x}}_{k+1}=\widetilde{\mathbf{x}}_{k+1|k+1}+\mathbf{K}_{k+1}(\mathbf{y}_{k+1}-\mathbf{H}\widetilde{\mathbf{x}}_{k+1|k+1}),\label{eqn: kfwithoutdf update x}\\
&\bm{\Sigma}_{k+1}=\widetilde{\bm{\Sigma}}_{k+1|k+1}-\mathbf{K}_{k+1}(\widetilde{\bm{\Sigma}}_{k+1|k+1}\mathbf{H}^{T}-\mathbf{BM}_{k+1}\mathbf{R})^{T}.\label{eqn: kfwithoutdf sigma update}
\end{align}
\normalsize
The forward filter exists if $\textrm{rank}(\mathbf{HB})=\textrm{rank}(\mathbf{B})=m$ which implies $n\geq m$ and $p\geq m$\cite{gillijns2007unknownkf}. Here, unlike I-EKFs, the gain matrices $\mathbf{K}_{k+1}$ and $\mathbf{M}_{k+1}$, are deterministic and completely determined by the model parameters and the initial covariance matrix similar to I-KF\cite{krishnamurthy2019how}.
\subsubsection{Inverse filter}
Denote $\widetilde{\mathbf{F}}_{k}=(\mathbf{I}_{n\times n}-\mathbf{K}_{k+1}\mathbf{H})(\mathbf{I}_{n\times n}-\mathbf{BM}_{k+1}\mathbf{H})\mathbf{F}$ and $\mathbf{E}_{k}=\mathbf{BM}_{k+1}-\mathbf{K}_{k+1}\mathbf{HBM}_{k+1}+\mathbf{K}_{k+1}$. From \eqref{eqn: linear y withdf} with $\mathbf{D}=\mathbf{0}_{p\times m}$, and \eqref{eqn: kfwithoutdf predict}-\eqref{eqn: kfwithoutdf update x}, the state transition equation for I-KF-without-DF is
\par\noindent\small
\begin{align}
\label{eqn: state for kfwithoutdf}
\hat{\mathbf{x}}_{k+1}=\widetilde{\mathbf{F}}_{k}\hat{\mathbf{x}}_{k}+\mathbf{E}_{k}\mathbf{Hx}_{k+1}+\mathbf{E}_{k}\mathbf{v}_{k+1}.
\end{align}
\normalsize
Unlike the state transition \eqref{eqn: state transition ekf without df input} and \eqref{eqn: state transition ekf without df state} of I-EKF-without-DF, the state transition for I-KF-without-DF is not an explicit function of the forward filter input estimate and hence, an augmented state is not needed. The difference arises from the forward EKF-without-DF, where the current input estimate explicitly depends on the previous input estimates as observed in \eqref{eqn: ekfwithoutdf update u}, which is not the case in KF-without-DF. The I-KF-without-DF's recursions with observation \eqref{eqn: linear a}
are:
\par\noindent\small
\begin{align}
&\textit{Prediction:} \;
\doublehat{\mathbf{x}}_{k+1|k}=\widetilde{\mathbf{F}}_{k}\doublehat{\mathbf{x}}_{k}+\mathbf{E}_{k}\mathbf{Hx}_{k+1},\label{eqn: inverse kfwithoutdf state predict}\\
&\overline{\bm{\Sigma}}_{k+1|k}=\widetilde{\mathbf{F}}_{k}\overline{\bm{\Sigma}}_{k}\widetilde{\mathbf{F}}_{k}^{T}+\overline{\mathbf{Q}}_{k},\label{eqn: inverse kfwithoutdf covariance predict}\\
&\textit{Update:}\;\overline{\mathbf{S}}_{k+1}=\mathbf{G}\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}^{T}+\bm{\Sigma}_{\epsilon},\label{eqn: inverse kfwithoutdf gain}\\
&\doublehat{\mathbf{x}}_{k+1}=\doublehat{\mathbf{x}}_{k+1|k}+\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}^{T}\overline{\mathbf{S}}_{k+1}^{-1}(\mathbf{a}_{k+1}-\mathbf{G}\doublehat{\mathbf{x}}_{k+1|k}),\label{eqn: inverse kfwithoutdf state update}\\
&\overline{\bm{\Sigma}}_{k+1}=\overline{\bm{\Sigma}}_{k+1|k}-\overline{\bm{\Sigma}}_{k+1|k}\mathbf{G}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\mathbf{G}\overline{\bm{\Sigma}}_{k+1|k},\label{eqn: inverse kfwithoutdf covariance update}
\end{align}
\normalsize
where (inverse) process noise covariance matrix $\overline{\mathbf{Q}}_{k}=\mathbf{E}_{k}\mathbf{R}\mathbf{E}_{k}^{T}$.
\begin{comment}
The inverse filter seeks to find an estimate $\doublehat{\mathbf{x}}_{k}$ of the adversary's estimate $\hat{\mathbf{x}}_{k}$ given the observation . From , \eqref{eqn: kfwithoutdf update u}, \eqref{eqn: kfwithoutdf update x with u}, and , we obtain
\par\noindent\small
\begin{align*}
\hat{\mathbf{x}}_{k+1}=\widetilde{\mathbf{F}}_{k}\hat{\mathbf{x}}_{k}+\mathbf{E}_{k}\mathbf{y}_{k+1},
\end{align*}
\normalsize
where . Substituting for $\mathbf{y}_{k+1}$ using , we obtain
There is a significant difference in the above inverse filter and the one for non-linear systems without DF developed in Section \ref{subsec:ekfwithoutdf}.
\end{comment}
\subsection{I-KF-with-DF}
\label{subsec:kfwithdf}
Consider the linear system model with DF given by \eqref{eqn: linear x with input} and \eqref{eqn: linear y withdf}.
\subsubsection{Forward filter}
Denote the state estimation covariance, input estimation (without delay) covariance, and cross-covariance of state and input estimates by $\bm{\Sigma}^{x}_{k}$, $\bm{\Sigma}^{u}_{k}$ and $\bm{\Sigma}^{xu}_{k}$, respectively. The forward KF-with-DF is \cite{gillijns2007kfb}:
\par\noindent\small
\begin{align}
&\textit{Prediction:}\;\hat{\mathbf{x}}_{k+1|k}=\mathbf{F}\hat{\mathbf{x}}_{k}+\mathbf{B}\hat{\mathbf{u}}_{k},\label{eqn: kfwithdf predict}\\
&\bm{\Sigma}^{x}_{k+1|k}=\begin{bmatrix}
\mathbf{F} & \mathbf{B}
\end{bmatrix}\begin{bmatrix}
\bm{\Sigma}^{x}_{k} & \bm{\Sigma}^{xu}_{k}\\
\bm{\Sigma}^{ux}_{k} & \bm{\Sigma}^{u}_{k}
\end{bmatrix}\begin{bmatrix}
\mathbf{F}^{T}\\
\mathbf{B}^{T}
\end{bmatrix}+\mathbf{Q},\nonumber\\
&\textit{Gain computation:}\;\mathbf{S}_{k+1}=\mathbf{H}\bm{\Sigma}^{x}_{k+1|k}\mathbf{H}^{T}+\mathbf{R},\nonumber\\
&\mathbf{M}_{k+1}=(\mathbf{D}^{T}\mathbf{S}_{k+1}^{-1}\mathbf{D})^{-1}\mathbf{D}^{T}\mathbf{S}_{k+1}^{-1},\;\;\;\mathbf{K}_{k+1}=\bm{\Sigma}^{x}_{k+1|k}\mathbf{H}^{T}\mathbf{S}_{k+1}^{-1},\nonumber\\
&\textit{Update:}\;\hat{\mathbf{u}}_{k+1}=\mathbf{M}_{k+1}(\mathbf{y}_{k+1}-\mathbf{H}\hat{\mathbf{x}}_{k+1|k}),\label{eqn: kfwithdf update u}\\
&\hat{\mathbf{x}}_{k+1}=\hat{\mathbf{x}}_{k+1|k}+\mathbf{K}_{k+1}(\mathbf{y}_{k+1}-\mathbf{H}\hat{\mathbf{x}}_{k+1|k}-\mathbf{D}\hat{\mathbf{u}}_{k+1}),\label{eqn: kfwithdf update x}\\
&\textit{Covariance updates:}\;\bm{\Sigma}^{u}_{k+1}=(\mathbf{D}^{T}\mathbf{S}_{k+1}^{-1}\mathbf{D})^{-1},\nonumber\\
&\bm{\Sigma}^{x}_{k+1}=\bm{\Sigma}^{x}_{k+1|k}-\mathbf{K}_{k+1}(\mathbf{S}_{k+1}-\mathbf{D}\bm{\Sigma}^{u}_{k+1}\mathbf{D}^{T})\mathbf{K}_{k+1}^{T},\nonumber\\
&\bm{\Sigma}^{xu}_{k+1}=(\bm{\Sigma}^{ux}_{k+1})^{T}=-\mathbf{K}_{k+1}\mathbf{D}\bm{\Sigma}^{u}_{k+1}.\nonumber
\end{align}
\normalsize
The forward filter exists if $\textrm{rank}(\mathbf{D})=m$ (which implies $p\geq m$).
\subsubsection{Inverse filter}
Consider an augmented state vector $\mathbf{z}_{k}=\begin{bmatrix}
\hat{\mathbf{x}}_{k}^{T} & \hat{\mathbf{u}}_{k}^{T}
\end{bmatrix}^{T}$. Denote $\widetilde{\mathbf{F}}_{k}=(\mathbf{I}_{n\times n}-\mathbf{K}_{k+1}\mathbf{H}+\mathbf{K}_{k+1}\mathbf{DM}_{k+1}\mathbf{H})\mathbf{F}$, $\widetilde{\mathbf{B}}_{k}=(\mathbf{I}_{n\times n}-\mathbf{K}_{k+1}\mathbf{H}+\mathbf{K}_{k+1}\mathbf{DM}_{k+1}\mathbf{H})\mathbf{B}$, $\mathbf{E}_{k}=\mathbf{K}_{k+1}(\mathbf{I}_{p\times p}-\mathbf{DM}_{k+1})$, $\widetilde{\mathbf{H}}_{k}=-\mathbf{M}_{k+1}\mathbf{HF}$ and $\widetilde{\mathbf{D}}_{k}=-\mathbf{M}_{k+1}\mathbf{HB}$. From \eqref{eqn: linear y withdf}, and \eqref{eqn: kfwithdf predict}
\eqref{eqn: kfwithdf update x}, the state transition equations for I-KF-with-DF are
\par\noindent\small
\begin{align*}
&\hat{\mathbf{x}}_{k+1}=\widetilde{\mathbf{F}}_{k}\hat{\mathbf{x}}_{k}+\widetilde{\mathbf{B}}_{k}\hat{\mathbf{u}}_{k}+\mathbf{E}_{k}\mathbf{H}\mathbf{x}_{k+1}+\mathbf{E}_{k}\mathbf{D}\mathbf{u}_{k+1}+\mathbf{E}_{k}\mathbf{v}_{k+1},
\end{align*}\normalsize
and\par\noindent\small
\begin{align*}
&\hat{\mathbf{u}}_{k+1}\nonumber\\
&=\widetilde{\mathbf{H}}_{k}\hat{\mathbf{x}}_{k}+\widetilde{\mathbf{D}}_{k}\hat{\mathbf{u}}_{k}+\mathbf{M}_{k+1}\mathbf{H}\mathbf{x}_{k+1}+\mathbf{M}_{k+1}\mathbf{D}\mathbf{u}_{k+1}+\mathbf{M}_{k+1}\mathbf{v}_{k+1}.
\end{align*}
\normalsize
\color{black}
Also, $\begin{bmatrix}
(\mathbf{E}_{k}\mathbf{v}_{k+1})^{T} & (\mathbf{M}_{k+1}\mathbf{v}_{k+1})^{T}
\end{bmatrix}^{T}$ is the augmented noise vector involved in this state transition with noise covariance matrix $\overline{\mathbf{Q}}_{k}=\begin{bmatrix}
\mathbf{E}_{k}\mathbf{R}\mathbf{E}_{k}^{T} & \mathbf{E}_{k}\mathbf{R}\mathbf{M}_{k+1}^{T}\\
\mathbf{M}_{k+1}\mathbf{R}\mathbf{E}_{k}^{T} & \mathbf{M}_{k+1}\mathbf{R}\mathbf{M}_{k+1}^{T}
\end{bmatrix}$. Then, \textit{ceteris paribus}, following similar steps as in I-KF-without-DF, the I-KF-with-DF computes the estimate $\hat{\mathbf{z}}_{k}=\begin{bmatrix}
\doublehat{\mathbf{x}}_{k}^{T} & \doublehat{\mathbf{u}}_{k}^{T}
\end{bmatrix}^{T}$ of the augmented state vector using the observation $\mathbf{a}_{k}$ given by \eqref{eqn: linear a}. The system matrices for the augmented state are $\widetilde{\mathbf{F}}^{z}_{k}=\begin{bmatrix}
\widetilde{\mathbf{F}}_{k} & \widetilde{\mathbf{B}}_{k}\\
\widetilde{\mathbf{H}}_{k} & \widetilde{\mathbf{D}}_{k}
\end{bmatrix}$ and $\overline{\mathbf{G}}=\begin{bmatrix}
\mathbf{G} & \mathbf{0}_{n_{a}\times m}
\end{bmatrix}$. The I-KF-with-DF predicts the augmented state as
\par\noindent\small
\begin{align*}
&\doublehat{\mathbf{x}}_{k+1|k}=\widetilde{\mathbf{F}}_{k}\doublehat{\mathbf{x}}_{k}+\widetilde{\mathbf{B}}_{k}\doublehat{\mathbf{u}}_{k}+\mathbf{E}_{k}\mathbf{Hx}_{k+1}+\mathbf{E}_{k}\mathbf{Du}_{k+1},\\
&\doublehat{\mathbf{u}}_{k+1|k}=\widetilde{\mathbf{H}}_{k}\doublehat{\mathbf{x}}_{k}+\widetilde{\mathbf{D}}_{k}\doublehat{\mathbf{u}}_{k}+\mathbf{M}_{k+1}\mathbf{Hx}_{k+1}+\mathbf{M}_{k+1}\mathbf{Du}_{k+1},\\
&\hat{\mathbf{z}}_{k+1|k}=\begin{bmatrix}
\doublehat{\mathbf{x}}_{k+1|k}^{T} & \doublehat{\mathbf{u}}_{k+1|k}^{T}
\end{bmatrix}^{T},\;
\overline{\bm{\Sigma}}_{k+1|k}=\widetilde{\mathbf{F}}^{z}_{k}\overline{\bm{\Sigma}}_{k}(\widetilde{\mathbf{F}}^{z}_{k})^{T}+\overline{\mathbf{Q}}_{k},
\end{align*}
\normalsize
followed by the update procedure \eqref{eqn: inverse kfwithoutdf gain}-\eqref{eqn: inverse kfwithoutdf covariance update} with $\mathbf{G}$ and $\doublehat{x}_{k+1}$ replaced by $\overline{\mathbf{G}}$ and $\hat{\mathbf{z}}_{k+1}$, respectively.
Since the observation $\mathbf{y}_{k}$ explicitly depends on the unknown input $\mathbf{u}_{k}$ for a system with DF, I-KF-with-DF and I-EKF-with-DF require perfect knowledge of the current input $\mathbf{u}_{k}$ as a known exogenous input to obtain their state and input estimates, which is not the case in I-KF-without-DF and I-EKF-without-DF.
\begin{comment}
Using the KF with known exogenous inputs formulation, the recursive form of I-KF-with-DF is
\par\noindent\small
\begin{align*}
&\textit{Prediction:}\\
&\textit{Update:}\\
&\overline{\mathbf{S}}_{k+1}=\overline{\mathbf{G}}\overline{\bm{\Sigma}}_{k+1|k}\overline{\mathbf{G}}^{T}+\bm{\Sigma}_{\epsilon},\\
&\hat{\mathbf{z}}_{k+1}=\hat{\mathbf{z}}_{k+1|k}+\overline{\bm{\Sigma}}_{k+1|k}\overline{\mathbf{G}}^{T}\overline{\mathbf{S}}_{k+1}^{-1}(\mathbf{a}_{k+1}-\overline{\mathbf{G}}\hat{\mathbf{z}}_{k+1|k}),\\
&\overline{\bm{\Sigma}}_{k+1}=\overline{\bm{\Sigma}}_{k+1|k}-\overline{\bm{\Sigma}}_{k+1|k}\overline{\mathbf{G}}^{T}\overline{\mathbf{S}}_{k+1}^{-1}\overline{\mathbf{G}}\overline{\bm{\Sigma}}_{k+1|k},
\end{align*}
\normalsize
where , , and . Here, $\hat{\mathbf{z}}_{k}$ is the estimate of augmented state and $\overline{\bm{\Sigma}}_{k}$ is the associated covariance of estimate.
Note that, unlike the I-KF-without-DF, the inverse filter above needs to estimate an augmented state vector with both state and input estimates. Further, I-KF-without-DF does not require information about the actual input $\mathbf{u}_{k}$ while this information is needed as a known exogenous input to the inverse filter in the system with DF.
\end{comment}
\vspace{-8pt}
\section{Stability Analyses}
\label{sec:stability}
For continuous-time non-linear Kalman filtering, some convergence results were mentioned in \cite{krener2003convergenceOfEKF}. In case of EKF, sufficient conditions for stability of non-linear systems with linear output map were described in \cite{la1995conditionsforEKFforfreq}. Recently, the stability of deterministic EKF was studied based on contraction theory in \cite{bonnabel2014contraction}. The asymptotic convergence of EKF for a special class of systems, where EKF is applied for joint state and parameter estimation of linear stochastic systems, was studied in \cite{ljung1979asymptotic,ursin1980asymptotic}. If the non-linearities have known bounds, then the Riccati equation is slightly modified
to guarantee stability for the continuous-time EKF \cite{reif1998ekf}.
To derive the sufficient conditions for stochastic stability of non-linear filters,
one of the common approaches is to introduce unknown instrumental matrices to account for the linearization errors \cite{xiong2006performance_ukf}. It does not assume any bound on the estimation error, but its sufficient conditions for stability, especially the bounds assumed on the unknown matrices, are difficult to verify for practical systems.
Alternatively, \cite{reif1999stochastic} considers the one-step prediction formulation of the filter and provides sufficient conditions under which the state prediction error is \textit{exponentially bounded in mean-squared} sense. We restate some definitions and a useful Lemma from \cite{reif1999stochastic}.
\begin{definition}[Exponential mean-squared boundedness \cite{reif1999stochastic}] A stochastic process $\{\bm{\zeta}_{k} \}_{k \geq 0}$ is defined to be exponentially bounded in mean-squared sense if there are real numbers $\eta,\nu>0$ and $0<\lambda<1$ such that
$\mathbb{E}\left[\|\bm{\zeta}_{k}\|_{2}^{2}\right]\leq \eta\mathbb{E}\left[\|\bm{\zeta}_{0}\|_{2}^{2}\right]\lambda^{k}+\nu$
holds for every $k\geq 0$.
\end{definition}
\begin{definition}[Boundedness with probability one \cite{reif1999stochastic}] A stochastic process $\{\bm{\zeta}_{k} \}_{k \geq 0}$ is defined to be bounded with probability one if
$\sup_{k\geq 0}\|\bm{\zeta}_{k}\|_{2} < \infty$
holds with probability one.
\end{definition}
\begin{lemma}[Boundedness of stochastic process {\cite[Lemma 2.1]{reif1999stochastic}}]
\label{lemma:exponential boundedness}
Consider a function $V_{k}(\bm{\zeta}_{k})$ of the stochastic process $\bm{\zeta}_{k}$ and real numbers $v_{\textrm{min}}$, $v_{\textrm{max}}$, $\mu>0$, and $0<\lambda\leq 1$ such that for all $k\geq 0$
\par\noindent\small
\begin{align*}
v_{\textrm{min}}\|\bm{\zeta}_{k}\|_{2}^{2}\leq V_{k}(\bm{\zeta}_{k})\leq v_{\textrm{max}}\|\bm{\zeta}_{k}\|_{2}^{2},
\end{align*}
and
\begin{align*}
\mathbb{E}\left[ V_{k+1}(\bm{\zeta}_{k+1})|\bm{\zeta}_{k}\right]-V_{k}(\bm{\zeta}_{k})\leq\mu-\lambda V_{k}(\bm{\zeta}_{k}).
\end{align*}
\normalsize
Then, the stochastic process $\{\bm{\zeta}_{k}\}_{k \geq 0}$ is exponentially bounded in mean-squared sense, i.e.,
\par\noindent\small
\begin{align*}
\mathbb{E}\left[\|\bm{\zeta}_{k}\|_{2}^{2}\right]\leq\frac{v_{\textrm{max}}}{v_{\textrm{min}}}\mathbb{E}\left[\|\bm{\zeta}_{0}\|_{2}^{2}\right](1-\lambda)^{k}+\frac{\mu}{v_{\textrm{min}}}\sum_{i=1}^{k-1}(1-\lambda)^{i},
\end{align*}
\normalsize
for every $k\geq 0$. Further, $\{\bm{\zeta}_{k}\}_{k \geq 0}$ is also bounded with probability one.
\end{lemma}
In the bounded mean-squared sense, \cite[Sec. III]{reif1999stochastic} showed that, while the two-step prediction and update recursion (described in previous sections) and one-step formulation of (forward) filters may differ in their performance and transient behaviour, they have similar convergence properties.
However, the conditions of Lemma \ref{lemma:exponential boundedness} were proved to hold when the error remained within suitable bounds; the guarantees fail if the error exceeds this bound at any instant. However, it was numerically shown \cite[Sec. V]{reif1999stochastic} that the bound on the error was only of theoretical interest and, in practice, the filter remained stable for much larger estimation errors.
In the following, we first derive stability conditions for I-KF-without-DF in which we rely on the stability of the forward KF-without-DF as proved in \cite{fang2012on}. The procedure is similar for the stability of I-KF-with-DF and I-KF-without-unknown-input \cite{krishnamurthy2019how} and hence, we omit the details for these filters. For I-EKF stability, we employ both unknown matrix and bounded non-linearity approaches. In the process, we also derive the forward EKF stability conditions using unknown matrix approach; note that the same was obtained using bounded non-linearity method in \cite{reif1999stochastic}.
\subsection{I-KF-with-unknown-input}
\label{subsec:kfstable}
Consider I-KF-without-DF of Section \ref{subsec:kfwithoutdf}, where the forward filter is asymptotically stable under the sufficient conditions provided by \cite{fang2012on}. The following Theorem~\ref{theorem: inverse kf without DF} states conditions for stability of the inverse filter.
\begin{theorem} [Stability of I-KF-without-DF]
\label{theorem: inverse kf without DF}
Consider an asymptotically stable forward KF-without-DF \eqref{eqn: kfwithoutdf predict}-\eqref{eqn: kfwithoutdf sigma update} such that the gain matrices $\mathbf{M}_{k}$ and $\mathbf{K}_{k}$ asymptotically approach to limiting gain matrices $\overline{\mathbf{M}}$ and $\overline{\mathbf{K}}$, respectively. The measurement noise covariance matrix $\bm{\Sigma}_{\epsilon}$ is positive definite (p.d.). Denote the limiting matrices
$\overline{\mathbf{F}}=(\mathbf{I}-\overline{\mathbf{K}}\mathbf{H})(\mathbf{I}-\mathbf{B}\overline{\mathbf{M}}\mathbf{H})\mathbf{F}$ and $\overline{\mathbf{Q}}=\overline{\mathbf{E}}\mathbf{R}\overline{\mathbf{E}}^{T}$, where $\overline{\mathbf{E}}=\mathbf{B}\overline{\mathbf{M}}-\overline{\mathbf{K}}\mathbf{HB}\overline{\mathbf{M}}+\overline{\mathbf{K}}$. Then, the I-KF-without-DF \eqref{eqn: inverse kfwithoutdf state predict}-\eqref{eqn: inverse kfwithoutdf covariance update} is asymptotically stable under the assumption that pair ($\overline{\mathbf{F}}$,$\mathbf{G}$) is observable and the pair ($\overline{\mathbf{F}}$,$\mathbf{C}$) is controllable for the system given by \eqref{eqn: linear a} and \eqref{eqn: state for kfwithoutdf}, where $\mathbf{C}$ is such that $\overline{\mathbf{Q}}=\mathbf{C}^{T}\mathbf{C}$.
\end{theorem}
\begin{proof}
See Appendix~\ref{App-thm-inverse kf without DF}.
\end{proof}
Note that, for I-KF-with-DF's stability, the stability conditions of basic KF need to hold for the augmented state considered in inverse filter formulation of Section \ref{subsec:kfwithdf}. For forward KF-with-DF's stability conditions, we refer the reader to \cite{fang2012on}.
\subsection{I-EKF-without-unknown-input: Unknown matrix approach}
\label{subsec:ekf stable unknown}
Consider the I-EKF's two-step prediction and update formulation of Section~\ref{subsec:ekf}, with forward filter as EKF-without-unknown-input.
\subsubsection{Forward EKF stability}
Denote the forward EKF's state prediction, state estimation and measurement prediction errors by $\widetilde{\mathbf{x}}_{k+1|k}\doteq\mathbf{x}_{k+1}-\hat{\mathbf{x}}_{k+1|k}$, $\widetilde{\mathbf{x}}_{k}\doteq\mathbf{x}_{k}-\hat{\mathbf{x}}_{k}$ and $\widetilde{\mathbf{y}}_{k}\doteq\mathbf{y}_{k}-\hat{\mathbf{y}}_{k}$, with $\hat{\mathbf{y}}_{k}=h(\hat{\mathbf{x}}_{k|k-1})$, respectively. Using \eqref{eqn: ekf x}, \eqref{eqn: ekf predict} and the Taylor series expansion of $f(\cdot)$ at $\hat{\mathbf{x}}_{k}$, we get
\par\noindent\small
\begin{align*}
\widetilde{\mathbf{x}}_{k+1|k}=\mathbf{F}_{k}(\mathbf{x}_{k}-\hat{\mathbf{x}}_{k})+\mathbf{w}_{k}+\mathcal{O}(\|\mathbf{x}_{k}-\hat{\mathbf{x}}_{k}\|_{2}^{2})\approx\mathbf{F}_{k}\widetilde{\mathbf{x}}_{k}+\mathbf{w}_{k}.
\end{align*}
\normalsize
We consider the general case of time-varying process and measurement noise covariances and denote $\mathbf{Q}$, $\mathbf{R}$ and $\bm{\Sigma}_{\epsilon}$ by $\mathbf{Q}_{k}$, $\mathbf{R}_{k}$ and $\overline{\mathbf{R}}_{k}$, respectively. To account for the residuals and obtain an exact equality, we introduce an unknown instrumental diagonal matrix $\mathbf{U}^{x}_{k} \in\mathbb{R}^{n\times n}$\cite{xiong2006performance_ukf,li2012stochastic_ukf} as
\par\noindent\small
\begin{align}
\widetilde{\mathbf{x}}_{k+1|k}=\mathbf{U}^{x}_{k}\mathbf{F}_{k}\widetilde{\mathbf{x}}_{k}+\mathbf{w}_{k}\label{eqn:forward EKF predict error with alpha}.
\end{align}
\normalsize
However, using \eqref{eqn: ekf update}, we have $\widetilde{\mathbf{x}}_{k}=\widetilde{\mathbf{x}}_{k|k-1}-\mathbf{K}_{k}\widetilde{\mathbf{y}}_{k}$, which when substituted in \eqref{eqn:forward EKF predict error with alpha} yields $\widetilde{\mathbf{x}}_{k+1|k}=\mathbf{U}^{x}_{k}\mathbf{F}_{k}\widetilde{\mathbf{x}}_{k|k-1}-\mathbf{U}^{x}_{k}\mathbf{F}_{k}\mathbf{K}_{k}\widetilde{\mathbf{y}}_{k}+\mathbf{w}_{k}$.
Similarly, using Taylor series expansion of $h(\cdot)$ at $\hat{\mathbf{x}}_{k+1|k}$ in \eqref{eqn: non y withoutdf} and introducing an unknown diagonal matrix $\mathbf{U}^{y}_{k+1}\in\mathbb{R}^{p\times p}$ gives $\widetilde{\mathbf{y}}_{k+1}=\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}\widetilde{\mathbf{x}}_{k+1|k}+\mathbf{v}_{k+1}$.
The prediction error dynamics of the forward EKF becomes
\par\noindent\small
\begin{align}
\widetilde{\mathbf{x}}_{k+1|k}=\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})\widetilde{\mathbf{x}}_{k|k-1}-\mathbf{U}^{x}_{k}\mathbf{F}_{k}\mathbf{K}_{k}\mathbf{v}_{k}+\mathbf{w}_{k}.\label{eqn:forward EKF prediction error dynamics}
\end{align}
\normalsize
Denote the true prediction covariance by $\mathbf{P}_{k+1|k}=\mathbb{E}\left[\widetilde{\mathbf{x}}_{k+1|k}\widetilde{\mathbf{x}}_{k+1|k}^{T}\right]$. Define $\delta\mathbf{P}_{k+1|k}$ as the difference of estimated prediction covariance $\bm{\Sigma}_{k+1|k}$ and the true prediction covariance $\mathbf{P}_{k+1|k}$ while $\Delta\mathbf{P}_{k+1|k}$ as the error in the approximation of the expectation
\small
$\mathbb{E}\left[\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})\widetilde{\mathbf{x}}_{k|k-1}\widetilde{\mathbf{x}}_{k|k-1}^{T}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})^{T}\mathbf{F}_{k}^{T}\mathbf{U}^{x}_{k}\right]$
\normalsize
by $\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})\bm{\Sigma}_{k|k-1}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})^{T}\mathbf{F}_{k}^{T}\mathbf{U}^{x}_{k}$. Denoting $\hat{\mathbf{Q}}_{k}=\mathbf{Q}_{k}+\mathbf{U}^{x}_{k}\mathbf{F}_{k}\mathbf{K}_{k}\mathbf{R}_{k}\mathbf{K}_{k}^{T}\mathbf{F}_{k}^{T}\mathbf{U}^{x}_{k}+\delta\mathbf{P}_{k+1|k}+\Delta\mathbf{P}_{k+1|k}$ and following similar steps as in \cite{xiong2006performance_ukf, li2012stochastic_ukf}, we have
\par\noindent\small
\begin{align*}
&\bm{\Sigma}_{k+1|k}=\\
&\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})\bm{\Sigma}_{k|k-1}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})^{T}\mathbf{F}_{k}^{T}\mathbf{U}^{x}_{k}+\hat{\mathbf{Q}}_{k}.
\end{align*}
\normalsize
Similarly, denoting the true measurement prediction covariance and true cross-covariance by $\mathbf{P}^{yy}_{k+1}$ and $\mathbf{P}^{xy}_{k+1}$, respectively, we obtain
\par\noindent\small
\begin{align*}
\mathbf{S}_{k+1}&=\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}+\hat{\mathbf{R}}_{k+1},\\
\bm{\Sigma}^{xy}_{k+1}&=\begin{cases}\bm{\Sigma}_{k+1|k}\mathbf{U}^{xy}_{k+1}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}, & n\geq p\\
\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}\mathbf{U}^{xy}_{k+1}, & n<p\end{cases},
\end{align*}
\normalsize
where $\hat{\mathbf{R}}_{k+1}=\mathbf{R}_{k+1}+\Delta\mathbf{P}^{yy}_{k+1}+\delta\mathbf{P}^{yy}_{k+1}$ and $\mathbf{U}^{xy}_{k+1}$ is an unknown instrumental matrix introduced to account for errors in the estimated cross-covariance $\bm{\Sigma}_{k+1}^{xy}$\cite{xiong2007authorreply}.
The following Theorem~\ref{theorem:Forward ekf stable unknown matrix} provides stability conditions for the forward EKF using the unknown matrices $\mathbf{U}^{x}_{k}$, $\mathbf{U}^{y}_{k}$ and $\mathbf{U}^{xy}_{k}$.
\begin{theorem}[Stochastic stability of forward EKF]
\label{theorem:Forward ekf stable unknown matrix}
Consider the non-linear stochastic system in \eqref{eqn: ekf x} and \eqref{eqn: non y withoutdf}. The two-step forward EKF formulation is as in Section~\ref{subsec:ekf}. Let the following assumptions hold true:
\begin{enumerate}
\item There exist positive real numbers $\overline{f}$, $\overline{h}$, $\overline{\alpha}$, $\overline{\beta}$, $\overline{\gamma}$, $\underline{\sigma}$, $\overline{\sigma}$, $\overline{q}$, $\overline{r}$, $\hat{q}$ and $\hat{r}$ such that the following bounds are fulfilled for all $k\geq 0$.
\par\noindent\small
\begin{align*}
\|\mathbf{F}_{k}\|\leq\overline{f},&\hspace{0.4cm}\|\mathbf{H}_{k}\|\leq\overline{h},\hspace{0.4cm}\|\mathbf{U}^{x}_{k}\|\leq\overline{\alpha}, \hspace{0.4cm}\|\mathbf{U}^{y}_{k}\|\leq\overline{\beta},\\
\|\mathbf{U}^{xy}_{k}\|\leq\overline{\gamma},&\hspace{0.4cm}\mathbf{Q}_{k}\preceq\overline{q}\mathbf{I},\hspace{0.5cm} \mathbf{R}_{k}\preceq\overline{r}\mathbf{I},\hspace{0.6cm}\hat{q}\mathbf{I}\preceq\hat{\mathbf{Q}}_{k},\\
\hat{r}\mathbf{I}\preceq\hat{\mathbf{R}}_{k},&\hspace{0.4cm}
\underline{\sigma}\mathbf{I}\preceq\bm{\Sigma}_{k|k-1}\preceq\overline{\sigma}\mathbf{I}.
\end{align*}
\normalsize
\item $\mathbf{U}^{x}_{k}$ and $\mathbf{F}_{k}$ are non-singular for every $k\geq 0$.
\end{enumerate}
Then, the prediction error $\widetilde{\mathbf{x}}_{k|k-1}$ and the estimation error $\widetilde{\mathbf{x}}_{k}$ of the forward EKF are exponentially bounded in mean-squared sense and bounded with probability one provided that the constants satisfy the inequality
\par\noindent\small
\begin{align}
\overline{\sigma}\overline{\gamma}\overline{h}^{2}\overline{\beta}^{2}<\hat{r}.\label{eqn:inequality on constants}
\end{align}
\normalsize
\end{theorem}
\begin{proof}
See Appendix~\ref{App-thm-Forward ekf stable unknown matrix}.
\end{proof}
\subsubsection{Inverse EKF stability}
For a stable forward EKF in the previous subsection, we prove the stochastic stability of the I-EKF as an extension of Theorem \ref{theorem:Forward ekf stable unknown matrix}. Similar to the forward EKF, we introduce unknown matrices $\overline{\mathbf{U}}^{x}_{k}$ and $\overline{\mathbf{U}}^{a}_{k}$ to account for the errors in the linearization of functions $\widetilde{f}_{k}(\cdot)$ and $g(\cdot)$, respectively, and $\overline{\mathbf{U}}^{xa}_{k}$ for the errors in cross-covariance matrix estimation. Similarly, denote $\hat{\overline{\mathbf{Q}}}_{k}$ and $\hat{\overline{\mathbf{R}}}_{k}$ as the counterparts of $\hat{\mathbf{Q}}_{k}$ and $\hat{\mathbf{R}}_{k}$, respectively, in the I-EKF dynamics. The following Theorem~\ref{theorem: inverse EKF stable unknown matrix} states the stability criteria for I-EKF. Note that, when compared to Theorem~\ref{theorem:Forward ekf stable unknown matrix}, the following result requires an additional condition $\underline{r}\mathbf{I}\preceq\mathbf{R}_{k}$ for all $k\geq 0$ for some $\underline{r}>0$.
\begin{theorem}[Stochastic stability of I-EKF]
\label{theorem: inverse EKF stable unknown matrix}
Consider the adversary's forward EKF that is stable as per Theorem \ref{theorem:Forward ekf stable unknown matrix}. Additionally, assume that the following hold true for all $k\geq 0$.
\par\noindent\small
\begin{align*}
\underline{r}\mathbf{I}&\preceq\mathbf{R}_{k},&\|\mathbf{G}_{k}\|&\leq\overline{g},&\|\overline{\mathbf{U}}^{a}_{k}\|&\leq\overline{c},&
\|\overline{\mathbf{U}}^{xa}_{k}\|&\leq\overline{d},\\
\overline{\mathbf{R}}_{k}&\preceq\overline{\epsilon}\mathbf{I},&\hat{c}\mathbf{I}&\preceq\hat{\overline{\mathbf{Q}}}_{k},& \hat{d}\mathbf{I}&\preceq\hat{\overline{\mathbf{R}}}_{k},&
\underline{p}\mathbf{I}&\preceq\overline{\bm{\Sigma}}_{k|k-1}\preceq\overline{p}\mathbf{I},
\end{align*}
\normalsize
for some real positive constants $\underline{r}, \overline{g}, \overline{c}, \overline{d}, \overline{\epsilon}, \hat{c}, \hat{d}, \underline{p}, \overline{p}$. Then, the state estimation error of I-EKF is exponentially bounded in mean-squared sense and bounded with probability one provided that the constants satisfy the inequality
$\overline{p}\overline{d}\overline{g}^{2}\overline{c}^{2}<{\hat{d}}$.
\end{theorem}
\begin{proof}
See Appendix~\ref{App-thm-inverse EKF stable unknown matrix}.
\end{proof}
Note that Theorem \ref{theorem:Forward ekf stable unknown matrix} requires both $\hat{\mathbf{Q}}_{k}$ and $\hat{\mathbf{R}}_{k}$ to be p.d. In general, the difference matrices $\Delta\mathbf{P}_{k+1|k}$, $\delta\mathbf{P}_{k+1|k}$, $\Delta\mathbf{P}^{yy}_{k+1}$ and $\delta\mathbf{P}^{yy}_{k+1}$ may not be p.d. One could enhance the stability of EKF by enlarging the noise covariance matrices by adding sufficiently large $\Delta\mathbf{Q}_{k}$ and $\Delta\mathbf{R}_{k}$ to $\mathbf{Q}_{k}$ and $\mathbf{R}_{k}$, respectively \cite{xiong2006performance_ukf,xiong2007authorreply}. The same argument also holds true for I-EKF noise covariance matrices.
\subsection{I-EKF-without-unknown-input: Bounded non-linearity method}
\label{subsec:ekf stable Reif}
Consider the forward EKF's one step prediction formulation \eqref{eqn: one step forward ekf gain}-\eqref{eqn: one step forward ekf covariance}. Using Taylor series expansion around the estimate $\hat{\mathbf{x}}_{k}$, we have
\par\noindent\small
\begin{align*}
&f(\mathbf{x}_{k})-f(\hat{\mathbf{x}}_{k})=\mathbf{F}_{k}(\mathbf{x}_{k}-\hat{\mathbf{x}}_{k})+\phi(\mathbf{x}_{k},\hat{\mathbf{x}}_{k}),\\
&h(\mathbf{x}_{k})-h(\hat{\mathbf{x}}_{k})=\mathbf{H}_{k}(\mathbf{x}_{k}-\hat{\mathbf{x}}_{k})+\chi(\mathbf{x}_{k},\hat{\mathbf{x}}_{k}),
\end{align*}
\normalsize
where $\phi(\cdot)$ and $\chi(\cdot)$ are suitable non-linear functions to account for the higher-order terms of the expansions. Denoting the estimation error by $\mathbf{e}_{k}\doteq\mathbf{x}_{k}-\hat{\mathbf{x}}_{k}$, the error dynamics of the forward filter is
\par\noindent\small
\begin{align}
\mathbf{e}_{k+1}=(\mathbf{F}_{k}-\mathbf{K}_{k}\mathbf{H}_{k})\mathbf{e}_{k}+\mathbf{r}_{k}+\mathbf{s}_{k},\label{eqn: forward ekf error}
\end{align}
\normalsize
where $\mathbf{r}_{k}=\phi(\mathbf{x}_{k},\hat{\mathbf{x}}_{k})-\mathbf{K}_{k}\chi(\mathbf{x}_{k},\hat{\mathbf{x}}_{k})$ and $\mathbf{s}_{k}=\mathbf{w}_{k}-\mathbf{K}_{k}\mathbf{v}_{k}$.
The following Theorem~\ref{theorem: ekf stable Reif} (reproduced from \cite{reif1999stochastic}) provides sufficient conditions for forward EKF's stochastic stability.
\begin{theorem}[Exponential boundedness of forward EKF's error \cite{reif1999stochastic}]\label{theorem: ekf stable Reif} Consider a non-linear stochastic system defined by \eqref{eqn: ekf x} and \eqref{eqn: non y withoutdf}, and the one-step prediction formulation of forward EKF \eqref{eqn: one step forward ekf gain}-\eqref{eqn: one step forward ekf covariance}. Let the following assumptions hold true.
\begin{enumerate}
\item There exist positive real numbers $\overline{f}$,$\overline{h}$,$\underline{\sigma}$,$\overline{\sigma}$,$\underline{q}$,$\underline{r}$, $\delta$ such that the following bounds are fulfilled for all $k\geq 0$.
\par\noindent\small
\begin{align*}
\underline{\sigma}\mathbf{I}&\preceq\bm{\Sigma}_{k}\preceq\overline{\sigma}\mathbf{I},&\underline{q}\mathbf{I}&\preceq \mathbf{Q}_{k}\preceq\delta\mathbf{I},&\\
\underline{r}\mathbf{I}&\preceq \mathbf{R}_{k}\preceq\delta\mathbf{I},&\|\mathbf{F}_{k}\|&\leq\overline{f},\hspace{0.5cm}\|\mathbf{H}_{k}\|\leq\overline{h}.&
\end{align*}
\normalsize
\item $\mathbf{F}_{k}$ is non singular for every $k\geq 0$.
\item There exist positive real numbers $\kappa_{\phi}$, $\epsilon_{\phi}$, $\kappa_{\chi}$, $\epsilon_{\chi}$ such that the non-linear functions $\phi(\cdot)$ and $\chi(\cdot)$ satisfy
\par\noindent\small
\begin{align*}
&\|\phi(\mathbf{x},\hat{\mathbf{x}})\|_{2}\leq \kappa_{\phi}\|\mathbf{x}-\hat{\mathbf{x}}\|_{2}^{2}\hspace{0.2cm}\text{for}\hspace{0.2cm} \|\mathbf{x}-\hat{\mathbf{x}}\|_{2}\leq\epsilon_{\phi},\\
&\|\chi(\mathbf{x},\hat{\mathbf{x}})\|_{2}\leq \kappa_{\chi}\|\mathbf{x}-\hat{\mathbf{x}}\|_{2}^{2}\hspace{0.2cm}\text{for}\hspace{0.2cm} \|\mathbf{x}-\hat{\mathbf{x}}\|_{2}\leq\epsilon_{\chi}.
\end{align*}
\normalsize
\end{enumerate}
Then the estimation error given by \eqref{eqn: forward ekf error} is exponentially bounded in mean-squared sense and bounded with probability one provided that the estimation error is bounded by suitable constant $\epsilon>0$.
\end{theorem}
Theorem~\ref{theorem: ekf stable Reif} guarantees that the estimation error remains exponentially bounded in mean-squared sense as long as the error is within suitable $\epsilon$ bounds. Further, the mean drift $\mathbb{E}[V_{k+1}(\mathbf{e}_{k+1})|\mathbf{e}_{k}]-V_{k}(\mathbf{e}_{k})$ for a suitably defined $V_{k}(\cdot)$ (for application of Lemma \ref{lemma:exponential boundedness}) is negative when $\widetilde{\epsilon}\leq\|\mathbf{e}_{k}\|_{2}\leq\epsilon$, which drives the system towards zero error in an expected sense. However, with some finite probability, the estimation error at some time-steps may be outside the $\epsilon$ bound. In this case, we cannot guarantee with probability one that the error will be within $\epsilon$ bound again at some future time-steps. As mentioned earlier, bounded non-linearity approach may not provide theoretical guarantees for the filter to be stable for all time-steps but, practically, the filter remains stable even if the estimation error is outside the $\epsilon$ bound provided that the assumed bounds on the system model are satisfied.
For the inverse filter observations \eqref{eqn: non a}, the Taylor series expansion of $g(\cdot)$ at estimate $\doublehat{\mathbf{x}}_{k}$ of I-EKF's one step prediction formulation of Section~\ref{subsec:ekf}, considering suitable non-linear function $\overline{\chi}(\cdot)$ is
\par\noindent\small
\begin{align*}
&g(\hat{\mathbf{x}}_{k})-g(\doublehat{\mathbf{x}}_{k})=\mathbf{G}_{k}(\hat{\mathbf{x}}_{k}-\doublehat{\mathbf{x}}_{k})+\overline{\chi}(\hat{\mathbf{x}}_{k},\doublehat{\mathbf{x}}_{k}).
\end{align*}
\normalsize
Finally, the error dynamics of the inverse filter, with the estimation error denoted by $\overline{\mathbf{e}}_{k}\doteq\hat{\mathbf{x}}_{k}-\doublehat{\mathbf{x}}_{k}$ and the inverse filter's Kalman gain and estimation error covariance matrix by $\overline{\mathbf{K}}_{k}$ and $\overline{\bm{\Sigma}}_{k}$, respectively, is
\par\noindent\small
\begin{align}
\overline{\mathbf{e}}_{k+1}=(\widetilde{\mathbf{F}}^{x}_{k}-\overline{\mathbf{K}}_{k}\mathbf{G}_{k})\overline{\mathbf{e}}_{k}+\overline{\mathbf{r}}_{k}+\overline{\mathbf{s}}_{k},\label{eqn: inverse ekf error}
\end{align}
\normalsize
where $\overline{\mathbf{r}}_{k}=\overline{\phi}_{k}(\hat{x}_{k},\doublehat{\mathbf{x}}_{k})-\overline{\mathbf{K}}_{k}\overline{\chi}(\hat{\mathbf{x}}_{k},\doublehat{\mathbf{x}}_{k})$ and $\overline{\mathbf{s}}_{k}=\mathbf{K}_{k}\mathbf{v}_{k}-\overline{\mathbf{K}}_{k}\bm{\epsilon}_{k}$ with $\overline{\phi}_{k}(\hat{\mathbf{x}}_{k},\doublehat{\mathbf{x}}_{k})=\phi(\hat{\mathbf{x}}_{k},\doublehat{\mathbf{x}}_{k})-\mathbf{K}_{k}\chi(\hat{\mathbf{x}}_{k},\doublehat{\mathbf{x}}_{k})$.
The following Theorem~\ref{theorem: inverse ekf stable Reif} guarantees the stability of I-EKF. Note the additional assumption of $\mathbf{H}_{k}$ to be full column rank for all $k\geq 0$, which implies $p\geq n$.
\begin{theorem}[Exponential boundedness of I-EKF's error]
\label{theorem: inverse ekf stable Reif}
Consider the adversary's forward one-step prediction EKF that is stable as per Theorem \ref{theorem: ekf stable Reif}. Additionally, assume that the following hold true.
\begin{enumerate}
\item There exist positive real numbers $\overline{g}$,
\underline{$m$},
$\overline{m}$, $\underline{\epsilon}$, $\underline{\epsilon}$, $\overline{\delta}$ such that the following bounds are fulfilled for all $k\geq 0$.
\par\noindent\small
\begin{align*}
&\|\mathbf{G}_{k}\|\le\overline{g},\;\;\underline{m}\mathbf{I}\preceq\overline{\bm{\Sigma}}_{k}\preceq\overline{m}\mathbf{I},\;\;\underline{\epsilon}\mathbf{I}\preceq \overline{\mathbf{R}}_{k}\preceq\overline{\delta}\mathbf{I}.
\end{align*}
\normalsize
\item $\mathbf{H}_{k}$ is full column rank for every $k\geq 0$.
\item There exist positive real numbers $\kappa_{\bar{\chi}}$ and $\epsilon_{\bar{\chi}}$ such that the non-linear function $\overline{\chi}(\cdot)$ satisfies
\par\noindent\small
\begin{align*}
\|\overline{\chi}(\hat{\mathbf{x}},\doublehat{\mathbf{x}})\|_{2}\leq \kappa_{\bar{\chi}}\|\hat{\mathbf{x}}-\doublehat{\mathbf{x}}\|_{2}^{2}\hspace{0.2cm}\text{for}\hspace{0.2cm} \|\hat{\mathbf{x}}-\doublehat{\mathbf{x}}\|_{2}\leq\epsilon_{\bar{\chi}}.
\end{align*}
\normalsize
\end{enumerate}
Then, the estimation error for I-EKF given by \eqref{eqn: inverse ekf error} is exponentially bounded in mean-squared sense and bounded with probability one provided that the estimation error is bounded by suitable constant $\overline{\epsilon}>0$.
\end{theorem}
\begin{proof}
See Appendix~\ref{App-thm-inverse ekf stable Reif}.
\end{proof}
\vspace{-8pt}
\section{Numerical Experiments}\label{sec:simulations}
We illustrate the performance of the proposed inverse filters for different example systems. The efficacy of the inverse filters is demonstrated by comparing the estimation error with RCRLB. The CRLB provides a lower bound on mean-squared error (MSE) and is widely used to assess the performance of an estimator. For the discrete-time non-linear filtering, we employ the RCRLB as $\mathbb{E}\left[(\mathbf{x}_{k}-\hat{\mathbf{x}}_{k})(\mathbf{x}_{k}-\hat{\mathbf{x}}_{k})^{T}\right]\succeq\mathbf{J}_{k}^{-1}$
where $\mathbf{J}_{k}=\mathbb{E}\left[-\frac{\partial^{2}\ln{p(Y^{k},X^{k})}}{\partial\mathbf{x}_{k}^{2}}\right]$ is the Fisher information matrix\cite{tichavsky1998posterior}.
Here, $X^{k}=\lbrace\mathbf{x}_{0},\mathbf{x}_{1},\hdots,\mathbf{x}_{k}\rbrace$ is the state vector series while $Y^{k}=\lbrace\mathbf{y}_{0},\mathbf{y}_{1},\hdots,\mathbf{y}_{k}\rbrace$ are the noisy observations. Also, $p(Y^{k},X^{k})$ is the joint probability density of pair $(Y^{k},X^{k})$ and $\hat{\mathbf{x}}_{k}$ (a function of $Y^{k}$) is an estimate of $\mathbf{x}_{k}$ with $\frac{\partial^{2}(\cdot)}{\partial\mathbf{x}^{2}}$ denoting the Hessian with second order partial derivatives. The information matrix $\mathbf{J}_{k}$ can be computed recursively as \cite{tichavsky1998posterior}
\par\noindent\small
\begin{align}
\mathbf{J}_{k}&=\mathbf{D}_{k}^{22}-\mathbf{D}_{k}^{21}(\mathbf{J}_{k-1}+\mathbf{D}_{k}^{11})^{-1}\mathbf{D}_{k}^{12},\label{eqn: general Jk recursions}\\
\text{where}\hspace{0.25cm}\mathbf{D}_{k}^{11}&=\mathbb{E}\left[-\frac{\partial^{2}\ln{p(\mathbf{x}_{k}\vert\mathbf{x}_{k-1})}}{\partial\mathbf{x}_{k-1}^{2}}\right],\nonumber\\
\mathbf{D}_{k}^{12}&=\mathbb{E}\left[-\frac{\partial^{2}\ln{p(\mathbf{x}_{k}\vert\mathbf{x}_{k-1})}}{\partial\mathbf{x}_{k}\partial\mathbf{x}_{k-1}}\right]=(\mathbf{D}_{k}^{21})^{T},\nonumber\\
\mathbf{D}_{k}^{22}&=\mathbb{E}\left[-\frac{\partial^{2}\ln{p(\mathbf{x}_{k}\vert\mathbf{x}_{k-1})}}{\partial\mathbf{x}_{k}^{2}}\right]+\mathbb{E}\left[-\frac{\partial^{2}\ln{p(\mathbf{y}_{k}\vert\mathbf{x}_{k})}}{\partial\mathbf{x}_{k}^{2}}\right].\nonumber
\end{align}
\normalsize
For the non-linear system given by \eqref{eqn: ekf x} and \eqref{eqn: non y withoutdf}, the forward information matrices $\lbrace\mathbf{J}_{k}\rbrace$ recursions reduces to \cite{xiong2006performance_ukf}
\par\noindent\small
\begin{align}
&\mathbf{J}_{k+1}=\mathbf{Q}_{k}^{-1}\nonumber\\
&\;\;+\mathbf{H}_{k+1}^{T}\mathbf{R}_{k+1}^{-1}\mathbf{H}_{k+1}-\mathbf{Q}_{k}^{-1}\mathbf{F}_{k}(\mathbf{J}_{k}+\mathbf{F}_{k}^{T}\mathbf{Q}_{k}^{-1}\mathbf{F}_{k})^{-1}\mathbf{F}_{k}^{T}\mathbf{Q}_{k}^{-1},\label{eqn: additive Jk recursions}
\end{align}
\normalsize
where $\mathbf{F}_{k}=\nabla_{\mathbf{x}}f(\mathbf{x})\vert_{\mathbf{x}=\mathbf{x}_{k}}$ and $\mathbf{H}_{k}=\nabla_{\mathbf{x}}h(\mathbf{x})\vert_{\mathbf{x}=\mathbf{x}_{k}}$. Note that, for the information matrices recursion, the Jacobians $\mathbf{F}_{k}$ and $\mathbf{H}_{k}$ are evaluated at the true state $\mathbf{x}_{k}$ while for forward EKF recursions, these are evaluated at the estimates of the state. These recursions can be trivially extended to other system models considered in this paper and to compute the information matrix $\overline{\mathbf{J}}_{k}$ for inverse filter's estimate $\doublehat{\mathbf{x}}_{k}$.
Throughout all experiments, $100$ time-steps (indexed by $k$) were considered. The initial information matrices $\mathbf{J}_{0}$ and $\overline{\mathbf{J}}_{0}$ were set to $\bm{\Sigma}_{0}^{-1}$ and $\overline{\bm{\Sigma}}_{0}^{-1}$, respectively,
unless mentioned otherwise. Note that these initial estimates only affect the RCRLB in the transient phase. The steady state RCRLB is independent of the initialization.
\subsection{Inverse KF with unknown inputs}\label{subsec:sim KF with unknown inputs}
Consider a discrete-time linear system without DF\cite{hsieh2000robust},
\par\noindent\small
\begin{align*}
&\mathbf{x}_{k+1}=\begin{bmatrix}0.1 & 0.5 & 0.08\\ 0.6 & 0.01 & 0.04\\ 0.1 & 0.7 & 0.05\end{bmatrix}\mathbf{x}_{k}+\begin{bmatrix}0\\ 2\\ 1\end{bmatrix}u_{k}+\mathbf{w}_{k},\\
&\mathbf{y}_{k}=\begin{bmatrix}1 & 1 & 0\\ 0 & 1 & 1\end{bmatrix}\mathbf{x}_{k}+\mathbf{v}_{k},\;\;\;a_{k}=\begin{bmatrix}1 & 1 & 1\end{bmatrix}\hat{\mathbf{x}}_{k}+\epsilon_{k},
\end{align*}
\normalsize
with $\mathbf{w}_{k}\sim\mathcal{N}(\mathbf{0},\mathbf{I}_{3})$, $\mathbf{v}_{k}\sim\mathcal{N}(\mathbf{0},2\mathbf{I}_{2})$ and $\epsilon_{k}\sim\mathcal{N}(0,5)$. The
unknown input $u_{k}$ was set to $50$ for $1\leq k \leq 50$ and $-50$ thereafter.
The initial state was $\mathbf{x}_{0}=[1,1,1]^{T}$. For the forward filter, the initial state estimate was set to $[0,0,0]^{T}$ with initial covariance $\bm{\Sigma}_{0}=\mathbf{I}_{3}$. For the inverse filter, the initial state estimate was set to $\mathbf{x}_{0}$ (known to the defender) itself with initial covariance $\overline{\bm{\Sigma}}_{0}=5\mathbf{I}_{3}$.
For KF-with-DF, we modify the forward filter's observations as\cite{pan2011study}:
\par\noindent\small
\begin{align*}
\mathbf{y}_{k}=\begin{bmatrix}1 & 1 & 0\\ 0 & 1 & 1\end{bmatrix}\mathbf{x}_{k}+\begin{bmatrix}0\\1\end{bmatrix}u_{k}+\mathbf{v}_{k}.
\end{align*}
\normalsize
Here, the initial input estimate was set to $10$ with initial input estimate covariance $\bm{\Sigma}^{u}_{0}=10$ and initial cross-covariance $\bm{\Sigma}^{xu}_{0}=[0,0,0]^{T}$. The inverse filter's initial augmented state estimate $\mathbf{z}_{0}$ was set to $[1,1,1,50]^{T}$ with initial covariance $\overline{\bm{\Sigma}}_{0}=5\mathbf{I}_{4}$.
Fig. \ref{fig:KF unknown input} shows the time-averaged RMSE (AMSE) $=\sqrt{(\sum_{i=1}^{k}\|\mathbf{x}_{i}-\hat{\mathbf{x}}_{i}\|_{2}^{2})/nk}$ at $k$-th time step for $n$-dimensional actual state $\mathbf{x}_{i}$ and its estimate $\hat{\mathbf{x}}_{i}$, and RCRLB for state estimation for both forward and inverse filters in the two cases, respectively, averaged over 200 runs. For KF-without-DF, we plot the root MSE (RMSE) $=\sqrt{(\|\mathbf{x}_{k}-\hat{\mathbf{x}}_{k}\|^{2}_{2})/n}$ for comparison here but omit it for later plots for clarity.
Note that in Fig.~\ref{fig:KF unknown input}a, the I-KF-without-DF's RMSE fluctuates about the RCRLB because of a finite number of sample paths; see also similar phenomena in \cite{xiong2006performance_ukf,djuric2008target,vsimandl2001filtering}. The RCRLB value for state estimation is $\sqrt{\textrm{Tr}(\mathbf{J}^{-1})}$ with $\mathbf{J}$ denoting the associated information matrix.
Fig. \ref{fig:KF unknown input} shows that the effect of change in unknown input after 50 time-steps is negligible for KF-without-DF in both forward and inverse filters. However, for KF-with-DF, the sudden change in unknown input leads to an increase in state estimation error of the forward filter and, consequently, of the inverse filter. The estimation error of I-KF-without-DF is less than the corresponding forward filter while for KF-with-DF, the inverse filter has a higher estimation error than the forward filter.
Only I-KF-without-DF efficiently achieves the RCRLB bound on the estimation error. Note that in this and the following numerical experiments, the forward and inverse filters are compared only to highlight the relative estimation accuracy.
\begin{figure}
\centering
\includegraphics[width = 1.0\columnwidth]{kf_unknown2.png}
\caption{RMSE, AMSE and RCRLB for forward and inverse filters (a) KF-without-DF; (b) KF-with-DF.}
\label{fig:KF unknown input}
\end{figure}
\subsection{Inverse EKF without unknown inputs}\label{subsec:sim EKF}
Consider the discrete-time non-linear system model of FM demodulator without unknown inputs \cite[Sec. 8.2]{anderson2012optimal}
\par\noindent\small
\begin{align*}
&\mathbf{x}_{k+1}\doteq\begin{bmatrix}\lambda_{k+1}\\\theta_{k+1}\end{bmatrix}=\begin{bmatrix}\exp{(-T/\beta)}&0\\-\beta \exp{(-T/\beta)}-1&1\end{bmatrix}\begin{bmatrix}\lambda_{k}\\\theta_{k}\end{bmatrix}+\begin{bmatrix}1\\-\beta\end{bmatrix}w_{k},\\
&\mathbf{y}_{k}=\sqrt{2}\begin{bmatrix}\sin{\theta_{k}}\\\cos{\theta_{k}}\end{bmatrix}+\mathbf{v}_{k},\;\;
a_{k}=\hat{\lambda}_{k}^{2}+\epsilon_{k},
\end{align*}
\normalsize
with $w_{k}\sim\mathcal{N}(0,0.01)$, $\mathbf{v}_{k}\sim\mathcal{N}(\mathbf{0},\mathbf{I}_{2})$, $\epsilon_{k}\sim\mathcal{N}(0,5)$, $T=2\pi/16$ and $\beta=100$. Here, the observation function $g(\cdot)$ for the inverse filter is quadratic. Also, $\hat{\lambda}_{k}$ is the forward EKF's estimate of $\lambda_{k}$.
The initial state $\mathbf{x}_{0}\doteq[\lambda_{0},\theta_{0}]^{T}$ was set randomly with $\lambda_{0}\sim\mathcal{N}(0,1)$ and $\theta_{0}\sim\mathcal{U}[-\pi,\pi]$. The initial state estimates of forward and inverse EKF were also similarly drawn at random. The initial covariances were set to $\bm{\Sigma}_{0}=10\mathbf{I}_{2}$ and $\overline{\bm{\Sigma}}_{0}=5\mathbf{I}_{2}$ for forward and inverse EKF, respectively. The phase term of the state $\theta$ and its estimates $\hat{\theta}$ and $\doublehat{\theta}$ (for both prediction and measurement updates) were considered to be modulo $2\pi$ \cite{anderson2012optimal}.
Note that the process covariance $\mathbf{Q}$ is a singular matrix. For numerical stability and to facilitate computation of $\mathbf{Q}^{-1}$ for evaluating information matrices $\mathbf{J}_{k}$, we used an enlarged covariance matrix by adding $10^{-10}\mathbf{I}_{2}$ to $\mathbf{Q}$ in the forward filters. Similarly, we added $10^{-10}\mathbf{I}_{2}$ to $\overline{\mathbf{Q}}_{k}$ in the inverse filter because $\overline{\mathbf{Q}}_{k}$ is time-varying and may be ill-conditioned. The initial $\overline{\mathbf{J}}_{0}$ was taken close to the inverse of the steady state estimation covariance matrix of the forward filter. The initial $\overline{\mathbf{J}}_{0}$ only affects the RCRLB calculated for initial few time-steps. The RCRLB after these initial time-steps (around 20 for the considered system) shows same behaviour irrespective of the initial $\overline{\mathbf{J}}_{0}$.
Fig. \ref{fig:EKF and EKF with unknown inputs}a shows the AMSE and RCRLB for forward and inverse EKF averaged over 200 runs. The I-EKF's estimation error is comparable to that of forward EKF with I-EKF's average error being slightly higher than that of forward EKF. However, the difference between AMSE and RCRLB for I-EKF is less than that for forward EKF. Hence, we conclude that I-EKF is more efficient here. The I-EKF assumes initial covariance $\bm{\Sigma}_{0}$ as $5\mathbf{I}_{2}$ (the true $\bm{\Sigma}_{0}$ of forward EKF is $10\mathbf{I}_{2}$) and a random initial state for these recursions. In spite of this difference in the initial estimates, I-EKF's error performance is comparable to that of the forward EKF.
\subsection{Inverse EKF with unknown inputs}\label{subsec:sim EKF with unknown inputs}
\begin{figure}
\centering
\includegraphics[width = 1.0\columnwidth]{ekf_and_ekf_unknown.png}
\caption{(a) AMSE and RCRLB for forward and inverse EKF; (b) Time-averaged RMSE for forward and inverse EKF with and without DF, averaged over 200 runs.}
\label{fig:EKF and EKF with unknown inputs}
\end{figure}
For inverse EKF with unknown input, we modified the non-linear system model of Section \ref{subsec:sim EKF} to include an unknown input $u_{k}$ as
\par\noindent\small
\begin{align*}
\mathbf{x}_{k+1}=\begin{bmatrix}\exp{(-T/\beta)}&0\\-\beta \exp{(-T/\beta)}-1&1\end{bmatrix}\begin{bmatrix}\lambda_{k}\\\theta_{k}\end{bmatrix}+\begin{bmatrix}0.001\\1\end{bmatrix}u_{k}+\begin{bmatrix}1\\-\beta\end{bmatrix}w_{k},
\end{align*}
\normalsize
where $u_{k}$ was set to $\pi/4$ for $1\leq k \leq 50$ and $-\pi/4$ thereafter.
The observation $\mathbf{y}_{k}$ of the forward EKF-without-DF was same as in Section \ref{subsec:sim EKF}.
Consider a linear measurement $a_{k}$ for the inverse filter as
$a_{k}=\hat{\lambda}_{k}+\epsilon_{k}$.
For the forward filter, the initial input estimate was set to $0$ while the inverse filter initial augmented state estimate consisted of the true state $\mathbf{x}_{0}$ and true input $u_{0}$ (known to the defender) with initial covariance estimate $\overline{\bm{\Sigma}}_{0}=15\mathbf{I}_{3}$.
Similarly, for system with DF, we again considered the same non-linear system (without any unknown input in $\mathbf{x}_{k}$ state transition) but with a modified forward filter's observation
$\mathbf{y}_{k}=\sqrt{2}\begin{bmatrix}\sin{(\theta_{k}+u_{k})}\\\cos{(\theta_{k}+u_{k})}\end{bmatrix}+\mathbf{v}_{k}$.
The input estimates $\hat{u}$ and $\doublehat{u}$ were also, as before, modulo $2\pi$. The Gaussian noise terms in the inverse filter state transitions (\eqref{eqn: state transition ekf without df state} and \eqref{eqn:state transition ekf with df}) are transformed through non-linear functions such that \eqref{eqn: additive Jk recursions} is not applicable. The RCRLB in this case is derived using the general $\mathbf{J}_{k}$ recursions given by \eqref{eqn: general Jk recursions}, which is omitted here.
Fig. \ref{fig:EKF and EKF with unknown inputs}b shows that for both EKF with and without DF, the change in unknown input after $50$ time-steps does not increase the estimation error (as for KF-with-DF in Fig. \ref{fig:KF unknown input}b).
The estimation error of I-EKF-without-DF (I-EKF-with-DF) is higher (lower) than that of the corresponding forward filter.
Any change in unknown input affects the inverse filter's performance only when a significant change occurs in the forward filter's performance.
\vspace{-8pt}
\section{Summary}\label{sec:summary}
We studied the inverse filtering problem for non-linear systems with and without unknown inputs in the context of counter-adversarial applications.
For systems with unknown inputs, the adversary's observations may or may not be affected by the unknown input known to the defender but not the adversary.
The stochastic stability of a forward filter with certain additional system assumptions is also sufficient for the stability of the inverse filter.
Our experiments suggested that the impact of the unknown input on inverse filter's performance strongly depends on its impact on the forward filter. For certain systems, the inverse filter may perform more efficiently than the forward filter. In the companion paper (Part II) \cite{singh2022inverse_part2}, we develop I-EKF for second-order, Gaussian sum, and dithered EKFs and consider the case of uncertain information about the forward filter.
\vspace{-8pt}
\appendices
\section{Proof of Theorem \ref{theorem: inverse kf without DF}}
\label{App-thm-inverse kf without DF}
Under the stability assumption of the forward filter, $\widetilde{\mathbf{F}}_{k}$ and $\mathbf{E}_{k}$ converge to $\overline{\mathbf{F}}$ and $\overline{\mathbf{E}}$, respectively, where $\overline{\mathbf{F}}=(\mathbf{I}-\overline{\mathbf{K}}\mathbf{H})(\mathbf{I}-\mathbf{B}\overline{\mathbf{M}}\mathbf{H})\mathbf{F}$ and $\overline{\mathbf{E}}=\mathbf{B}\overline{\mathbf{M}}-\overline{\mathbf{K}}\mathbf{HB}\overline{\mathbf{M}}+\overline{\mathbf{K}}$, obtained by replacing $\mathbf{K}_{k+1}$ and $\mathbf{M}_{k+1}$ by the limiting matrices $\overline{\mathbf{K}}$ and $\overline{\mathbf{M}}$, respectively, in $\widetilde{\mathbf{F}}_{k}$ and $\mathbf{E}_{k}$. In this limiting case, the state transition equation \eqref{eqn: state for kfwithoutdf} becomes $\hat{\mathbf{x}}_{k+1}=\overline{\mathbf{F}}\hat{\mathbf{x}}_{k}+\overline{\mathbf{E}}\mathbf{Hx}_{k+1}+\overline{\mathbf{E}}\mathbf{v}_{k+1}$.
From \eqref{eqn: inverse kfwithoutdf covariance predict}, \eqref{eqn: inverse kfwithoutdf gain}, and \eqref{eqn: inverse kfwithoutdf covariance update} and substituting the limiting matrices, the Riccati equation $\overline{\bm{\Sigma}}_{k+1|k}=\overline{\mathbf{F}}\left[\overline{\bm{\Sigma}}_{k|k-1}-\overline{\bm{\Sigma}}_{k|k-1}\mathbf{G}^{T}(\mathbf{G}\overline{\bm{\Sigma}}_{k|k-1}\mathbf{G}^{T}+\overline{\mathbf{R}})^{-1}\mathbf{G}\overline{\bm{\Sigma}}_{k|k-1}\right]\overline{\mathbf{F}}^{T}+\overline{\bm{Q}}$ is obtained,
where $\overline{\mathbf{Q}}=\overline{\mathbf{E}}\mathbf{R}\overline{\mathbf{E}}^{T}$. For the forward filter to be stable, covariance $\mathbf{R}$ needs to be p.d.\cite{fang2012on} and hence, $\overline{\mathbf{Q}}$ is a p.s.d. matrix. With $\overline{\mathbf{R}}$ being p.d. and the observability and controllability assumptions, $\overline{\bm{\Sigma}}_{k|k-1}$ tends to a unique p.d. matrix $\overline{\bm{\Sigma}}$ satisfying $\overline{\bm{\Sigma}}=\overline{\mathbf{F}}[\overline{\bm{\Sigma}}-\overline{\bm{\Sigma}}\mathbf{G}^{T}\left(\mathbf{G}\overline{\bm{\Sigma}}\mathbf{G}^{T}+\overline{\mathbf{R}}\right)^{-1}\mathbf{G}\overline{\bm{\Sigma}}]\overline{\mathbf{F}}^{T}+\overline{\mathbf{Q}}$,
and $\overline{\mathbf{F}}-\overline{\mathbf{F}}\overline{\bm{\Sigma}}\mathbf{G}^{T}(\mathbf{G}\overline{\bm{\Sigma}}\mathbf{G}^{T}+\overline{\mathbf{R}})^{-1}\mathbf{G}$ has eigenvalues strictly within the unit circle. These results follow directly from the application of \cite[Proposition 4.1, Sec. 4.1]{bertsekas1995dynamic} similar to the stability and convergence results for the standard KF for linear systems \cite[Appendix E.4]{bertsekas1995dynamic}.
In this limiting case, the inverse filter prediction and update equations take the following asymptotic form
\par\noindent\small
\begin{align*}
&\doublehat{\mathbf{x}}_{k+1|k}=\overline{\mathbf{F}}\doublehat{\mathbf{x}}_{k}+\overline{\mathbf{E}}\mathbf{Hx}_{k+1},\\
&\doublehat{\mathbf{x}}_{k+1}=\doublehat{\mathbf{x}}_{k+1|k}+\overline{\bm{\Sigma}}\mathbf{G}^{T}(\mathbf{G}\overline{\bm{\Sigma}}\mathbf{G}^{T}+\overline{\mathbf{R}})^{-1}(\mathbf{a}_{k+1}-\mathbf{G}\doublehat{\mathbf{x}}_{k+1|k}).
\end{align*}
\normalsize
Denoting the inverse filter's one-step prediction error as $\overline{\mathbf{e}}_{k+1|k}\doteq\hat{\mathbf{x}}_{k+1}-\doublehat{\mathbf{x}}_{k+1|k}$, the error dynamics for the inverse filter is obtained from this asymptotic form using \eqref{eqn: linear a} as
\par\noindent\small
\begin{align*}
\overline{\mathbf{e}}_{k+1|k}&=\left(\overline{\mathbf{F}}-\overline{\mathbf{F}}\overline{\bm{\Sigma}}\mathbf{G}^{T}(\mathbf{G}\overline{\bm{\Sigma}}\mathbf{G}^{T}+\overline{\mathbf{R}})^{-1}\mathbf{G}\right)\overline{\mathbf{e}}_{k|k-1}\nonumber\\ &\hspace{4mm}-\overline{\mathbf{F}}\overline{\bm{\Sigma}}\mathbf{G}^{T}(\mathbf{G}\overline{\bm{\Sigma}}\mathbf{G}^{T}+\overline{\mathbf{R}})^{-1}\bm{\epsilon}_{k}+\overline{\mathbf{E}}\mathbf{v}_{k+1}.
\end{align*}
\normalsize
Since $\overline{\mathbf{F}}-\overline{\mathbf{F}}\overline{\bm{\Sigma}}\mathbf{G}^{T}(\mathbf{G}\overline{\bm{\Sigma}}\mathbf{G}^{T}+\overline{\mathbf{R}})^{-1}\mathbf{G}$ has eigenvalues strictly within the unit circle, this error dynamics is asymptotically stable.
\vspace{-8pt}
\section{Proof of Theorem~\ref{theorem:Forward ekf stable unknown matrix}}
\label{App-thm-Forward ekf stable unknown matrix}
For simplicity, we consider the case of $n\geq p$ with $\mathbf{U}^{xy}_{k+1}\in\mathbb{R}^{n\times n}$. It is trivial to show that the proof remains valid for $n<p$ as well. Using the expressions for $\bm{\Sigma}^{xy}_{k+1}$ and $\mathbf{S}_{k+1}$, we have
\par\noindent\small
\begin{align*}
\mathbf{K}_{k+1}&=\bm{\Sigma}_{k+1|k}\mathbf{U}^{xy}_{k+1}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}\\
&\hspace{1.0cm}\times\left(\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}+\hat{\mathbf{R}}_{k+1}\right)^{-1},\\
\bm{\Sigma}_{k+1}&=\bm{\Sigma}_{k+1|k}-\bm{\Sigma}_{k+1|k}\mathbf{U}^{xy}_{k+1}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}\\
&\times\left(\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{U}^{y}_{k+1}+\hat{\mathbf{R}}_{k+1}\right)^{-1}\\
&\times\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}(\mathbf{U}^{xy}_{k+1})^{T}\bm{\Sigma}_{k+1|k}.
\end{align*}
\normalsize
Define $V_{k}(\widetilde{\mathbf{x}}_{k|k-1})=\widetilde{\mathbf{x}}_{k|k-1}^{T}\bm{\Sigma}_{k|k-1}^{-1}\widetilde{\mathbf{x}}_{k|k-1}$. Using the bounds assumed on $\bm{\Sigma}_{k|k-1}$, we have for all $k\geq 0$
\par\noindent\small
\begin{align*}
\frac{1}{\overline{\sigma}}\|\widetilde{\mathbf{x}}_{k|k-1}\|_{2}^{2}\leq V_{k}(\widetilde{\mathbf{x}}_{k|k-1})\leq\frac{1}{\underline{\sigma}}\|\widetilde{\mathbf{x}}_{k|k-1}\|_{2}^{2}.
\end{align*}
\normalsize
Hence, the first condition of Lemma \ref{lemma:exponential boundedness} is satisfied with $v_{\textrm{min}}=1/\overline{\sigma}$ and $v_{\textrm{max}}=1/\underline{\sigma}$.
Using \eqref{eqn:forward EKF prediction error dynamics} and the independence of noise terms, we have
\par\noindent\smal
\begin{align}
&\mathbb{E}\left[V_{k+1}(\widetilde{\mathbf{x}}_{k+1|k})\vert\widetilde{\mathbf{x}}_{k|k-1}\right]\nonumber\\
&= \widetilde{\mathbf{x}}_{k|k-1}^{T}(\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))^{T}\nonumber\\
&\hspace{3mm} \times \bm{\Sigma}_{k+1|k}^{-1}(\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))\widetilde{\mathbf{x}}_{k|k-1}\nonumber\\
& +\mathbb{E}\left[\mathbf{v}_{k}^{T}(\mathbf{U}^{x}_{k}\mathbf{F}_{k}\mathbf{K}_{k})^{T}\bm{\Sigma}_{k+1|k}^{-1}(\mathbf{U}^{x}_{k}\mathbf{F}_{k}\mathbf{K}_{k})\mathbf{v}_{k}\vert\widetilde{\mathbf{x}}_{k|k-1}\right]\nonumber\\
& +\mathbb{E}\left[\mathbf{w}_{k}^{T}\bm{\Sigma}_{k+1|k}^{-1}\mathbf{w}_{k}\vert\widetilde{\mathbf{x}}_{k|k-1}\right].\label{eqn: Vk expectation}
\end{align}
\normalsize
The difference of two matrices $\mathbf{A}-\mathbf{B}$ is invertible if maximum singular value of $\mathbf{B}$ is strictly less than the minimum singular value of $\mathbf{A}$. Using the assumed bounds, we have $\|\mathbf{K}_{k}\|\leq\overline{k}=(\overline{\sigma}\overline{\gamma}\overline{h}\overline{\beta})/\hat{r}$. Hence, maximum singular value of $\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}$ is upper-bounded by $(\overline{\sigma}\overline{\gamma}\overline{h}^{2}\overline{\beta}^{2})/\hat{r}$ and the inequality \eqref{eqn:inequality on constants} guarantees that $\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}$ is invertible (singular value of $\mathbf{I}$ is 1) such that
\par\noindent\smal
\begin{align*}
&\bm{\Sigma}_{k+1|k} \nonumber\\
&=\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}) (\bm{\Sigma}_{k|k-1}+(\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))^{-1}\\
&\hspace{3mm} \times\hat{\mathbf{Q}}_{k}((\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))^{-1})^{T}) (\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})^{T}\mathbf{F}_{k}^{T}\mathbf{U}^{x}_{k},
\end{align*}
\normalsize
because $\mathbf{U}^{x}_{k}$ and $\mathbf{F}_{k}$ are also assumed to be invertible. Again with the assumed bounds, we have $\|\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k})\|\leq\overline{\alpha}\overline{f}(1+\overline{k}\overline{\beta}\overline{h})$ which implies
\par\noindent\small
\begin{align*}
(\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))^{-1}\hat{\mathbf{Q}}_{k}((\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))^{-1})^{T}\\
\succeq\frac{\hat{q}}{(\overline{\alpha}\overline{f}(1+\overline{k}\overline{\beta}\overline{h}))^{2}}\mathbf{I}.
\end{align*}
\normalsize
Using this bound in the expression of $\bm{\Sigma}_{k+1|k}$ as in \cite{li2012stochastic_ukf}, we have
\par\noindent\small
\begin{align*}
(\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))^{T}\bm{\Sigma}_{k+1|k}^{-1}(\mathbf{U}^{x}_{k}\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k}\mathbf{U}^{y}_{k}\mathbf{H}_{k}))\\
\preceq(1-\lambda)\bm{\Sigma}_{k|k-1}^{-1},
\end{align*}
\normalsize
where $1-\lambda=\left(1+\frac{\hat{q}}{\overline{\sigma}(\overline{\alpha}\overline{f}(1+\overline{k}\overline{\beta}\overline{h}))^{2}}\right)^{-1}$ with $0<\lambda<1$. The last two expectation terms in \eqref{eqn: Vk expectation} can be bounded by $\mu=(\overline{r}p\overline{\alpha}^{2}\overline{f}^{2}\overline{k}^{2}/\underline{\sigma})+(\overline{q}n/\underline{\sigma})>0$ following similar steps as in \cite{li2012stochastic_ukf} such that
\par\noindent\small
\begin{align*}
\mathbb{E}\left[ V_{k+1}(\widetilde{\mathbf{x}}_{k+1|k})|\widetilde{\mathbf{x}}_{k|k-1}\right]-V_{k}(\widetilde{\mathbf{x}}_{k|k-1})\leq-\lambda V_{k}(\widetilde{\mathbf{x}}_{k|k-1})+\mu.
\end{align*}
\normalsize
Hence, the second condition of Lemma \ref{lemma:exponential boundedness} is also satisfied and the prediction error $\widetilde{\mathbf{x}}_{k|k-1}$ is exponentially bounded in mean-squared sense and bounded with probability one.
Furthermore, with the bounds assumed on various matrices, it is straightforward to show that
\par\noindent\small
\begin{align*}
\mathbb{E}\left[\|\widetilde{\mathbf{x}}_{k}\|^{2}_{2}\right]\leq(1+\overline{k}\overline{\beta}\overline{h})^{2}\mathbb{E}\left[\|\widetilde{\mathbf{x}}_{k|k-1}\|^{2}_{2}\right]+\overline{k}^{2}\overline{r}p.
\end{align*}
\normalsize
Finally, the exponential boundedness of $\widetilde{\mathbf{x}}_{k|k-1}$ leads to $\widetilde{\mathbf{x}}_{k}$ also being exponentially bounded in mean-squared sense as well as bounded with probability one.
\vspace{-8pt}
\section{Proof of Theorem~\ref{theorem: inverse EKF stable unknown matrix}}
\label{App-thm-inverse EKF stable unknown matrix}
We will show that the I-EKF's dynamics also satisfies the assumptions of Theorem \ref{theorem:Forward ekf stable unknown matrix}. For this, the following conditions \textbf{C1}-\textbf{C13} need to hold true for all $k\geq 0$ for some real positive constants $\overline{a},\overline{g},\overline{b},\overline{c},\overline{d},\hat{q},\overline{\epsilon},\hat{c},\hat{d},\underline{p},\overline{p}$.
\begin{description}
\item [C1]
$\|\widetilde{\mathbf{F}}^{x}_{k}\|\leq\overline{a}$;
\item [C2]
$\|\overline{\mathbf{U}}^{x}_{k}\|\leq\overline{b}$;
\item [C3]
$\overline{\mathbf{U}}^{x}_{k}$ is non-singular;
\item [C4]
$\widetilde{\mathbf{F}}^{x}_{k}$ is non-singular;
\item [C5]
$\overline{\mathbf{Q}}_{k}\preceq\widetilde{q}\mathbf{I}$;
\item [C6]
$\|\mathbf{G}_{k}\|\leq\overline{g}$;
\item [C7]
$\|\overline{\mathbf{U}}^{a}_{k}\|\leq\overline{c}$;
\item [C8]
$\|\overline{\mathbf{U}}^{xa}_{k}\|\leq\overline{d}$;
\item [C9]
$\overline{\mathbf{R}}_{k}\preceq\overline{\epsilon}\mathbf{I}$;
\item [C10]
$\hat{c}\mathbf{I}\preceq\hat{\overline{\mathbf{Q}}}_{k}$;
\item [C11]
$\hat{d}\mathbf{I}\preceq\hat{\overline{\mathbf{R}}}_{k}$;
\item [C12]
$\underline{p}\mathbf{I}\preceq\overline{\bm{\Sigma}}_{k|k-1}\preceq\overline{p}\mathbf{I}$; and
\item [C13]
the constants satisfy the inequality $\overline{p}\overline{d}\overline{g}^{2}\overline{c}^{2}<\hat{d}$.
\end{description}
Next, we prove that under the assumptions of Theorem \ref{theorem: inverse EKF stable unknown matrix}, \textbf{C1}-\textbf{C13} are satisfied. From the I-EKF's state transition \eqref{eqn: inverse ekf state transition}, the Jacobians $\widetilde{\mathbf{F}}^{x}_{k}=\mathbf{F}_{k}-\mathbf{K}_{k+1}\mathbf{H}_{k+1}\mathbf{F}_{k}$ and $\widetilde{\mathbf{F}}^{v}_{k}=\mathbf{K}_{k+1}$ such that $\overline{\mathbf{Q}}_{k}=\mathbf{K}_{k+1}\mathbf{R}_{k+1}\mathbf{K}_{k+1}^{T}$.
For \textbf{C1}, using $\|\mathbf{K}_{k+1}\|\leq\overline{k}$ (as proved in Theorem \ref{theorem:Forward ekf stable unknown matrix}) and the bounds on $\mathbf{F}_{k}$ and $\mathbf{H}_{k+1}$ from the assumptions of Theorem \ref{theorem:Forward ekf stable unknown matrix}, it is trivial to show that $\|\widetilde{\mathbf{F}}^{x}_{k}\|=\|\mathbf{F}_{k}-\mathbf{K}_{k+1}\mathbf{H}_{k+1}\mathbf{F}_{k}\|\leq\overline{f}+\overline{k}\overline{h}\overline{f}$.
Hence, \textbf{C1} is satisfied with $\overline{a}=\overline{f}+\overline{k}\overline{h}\overline{f}$.
For \textbf{C2}-\textbf{C4}, consider the unknown matrix $\overline{\mathbf{U}}^{x}_{k}$ introduced to account for the residuals in linearization of $\widetilde{f}_{k}(\cdot)$. Let $\hat{\widetilde{\mathbf{x}}}_{k+1|k}$ and $\hat{\widetilde{\mathbf{x}}}_{k}$ denote the state prediction error and state estimation error of I-EKF. Similar to forward EKF with the introduction of the unknown matrix, we have
\par\noindent\small
\begin{align}
\hat{\widetilde{\mathbf{x}}}_{k+1|k}=\overline{\mathbf{U}}^{x}_{k}(\mathbf{F}_{k}-\mathbf{K}_{k+1}\mathbf{H}_{k+1}\mathbf{F}_{k})\hat{\widetilde{\mathbf{x}}}_{k}+\mathbf{K}_{k+1}\mathbf{v}_{k+1}\label{eqn:inverse EKF error unknown alpha}.
\end{align}
\normalsize
Also, $\hat{\widetilde{\mathbf{x}}}_{k+1|k}=f(\hat{\mathbf{x}}_{k})-f(\doublehat{\mathbf{x}}_{k})-\mathbf{K}_{k+1}(h(f(\hat{\mathbf{x}}_{k}))-h(f(\doublehat{\mathbf{x}}_{k})))+\mathbf{K}_{k+1}\mathbf{v}_{k+1}$.
Using the unknown matrices $\mathbf{U}^{x}_{k}$ and $\mathbf{U}^{y}_{k}$ introduced in the linearization of $f(\cdot)$ and $h(\cdot)$, respectively, we have
\par\noindent\small
\begin{align*}
\hat{\widetilde{\mathbf{x}}}_{k+1|k}=(\mathbf{U}^{x}_{k}\mathbf{F}_{k}-\mathbf{K}_{k+1}\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}\mathbf{U}^{x}_{k}\mathbf{F}_{k})\hat{\widetilde{\mathbf{x}}}_{k}+\mathbf{K}_{k+1}\mathbf{v}_{k+1}.
\end{align*}
\normalsize
Comparing with \eqref{eqn:inverse EKF error unknown alpha}, we have
\par\noindent\small
\begin{align}
\overline{\mathbf{U}}^{x}_{k}(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1})\mathbf{F}_{k}=(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1})\mathbf{U}^{x}_{k}\mathbf{F}_{k}\label{eqn: unknown alpha for inverse EKF}.
\end{align}
\normalsize
With the additional assumption of $\underline{r}\mathbf{I}\preceq\mathbf{R}_{k}$ and using matrix inversion lemma as in proof of \cite[Lemma 3.1]{reif1999stochastic}, we have
\par\noindent\small
\begin{align*}
(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1})\bm{\Sigma}_{k+1|k}=\left(\bm{\Sigma}_{k+1|k}^{-1}+\mathbf{H}_{k+1}^{T}\mathbf{R}_{k+1}^{-1}\mathbf{H}_{k+1}\right)^{-1}.
\end{align*}
\normalsize
Since $\bm{\Sigma}_{k+1|k}$ is invertible by the assumptions of Theorem \ref{theorem:Forward ekf stable unknown matrix}, $\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1}$ is invertible for all $k\geq 0$ and
\par\noindent\small
\begin{align*}
(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1})^{-1}=\mathbf{I}+\bm{\Sigma}_{k+1|k}\mathbf{H}_{k+1}^{T}\mathbf{R}_{k+1}^{-1}\mathbf{H}_{k+1}.
\end{align*}
\normalsize
With the bounds assumed on various matrices, we have $ \|(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1})^{-1}\|\leq 1+\frac{\overline{\sigma}\overline{h}^{2}}{\underline{r}}$.
Furthermore, using this bound and the invertibility of $\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1}$ in \eqref{eqn: unknown alpha for inverse EKF}, it is straightforward to show that $\overline{\mathbf{U}}^{x}_{k}=(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1})\mathbf{U}^{x}_{k}(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1})^{-1}$ is non-singular (both $\mathbf{U}^{x}_{k}$ and $\mathbf{I}-\mathbf{K}_{k+1}\mathbf{U}^{y}_{k+1}\mathbf{H}_{k+1}$ are invertible under the assumptions of Theorem \ref{theorem:Forward ekf stable unknown matrix}) and satisfies $\|\overline{\mathbf{U}}^{x}_{k}\|\leq\overline{\alpha}(1+\overline{k}\overline{\beta}\overline{h})(1+(\overline{\sigma}\overline{h}^{2})/\underline{r})$.
Also, since both $\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1}$ and $\mathbf{F}_{k}$ are invertible, $\widetilde{\mathbf{F}}^{x}_{k}=\mathbf{F}_{k}(\mathbf{I}-\mathbf{K}_{k+1}\mathbf{H}_{k+1})$ is non-singular. Hence, \textbf{C2}-\textbf{C4} are also satisfied with $\overline{b}=\overline{\alpha}(1+\overline{k}\overline{\beta}\overline{h})(1+(\overline{\sigma}\overline{h}^{2})/\underline{r})$.
For \textbf{C5}, using the upper bound on $\mathbf{R}_{k}$ from assumptions of Theorem \ref{theorem:Forward ekf stable unknown matrix}, we have $\overline{\mathbf{Q}}_{k}\preceq\overline{r}\mathbf{K}_{k+1}\mathbf{K}_{k+1}^{T}$. Since, $\|\mathbf{K}_{k+1}\|\leq\overline{k}$, the maximum eigenvalue of $\mathbf{K}_{k+1}\mathbf{K}_{k+1}^{T}$ is bounded by $\overline{k}^{2}$ such that $\overline{\mathbf{Q}}_{k}\preceq\overline{k}^{2}\overline{r}\mathbf{I}$. Hence, \textbf{C5} is satisfied with $\title{q}=\overline{k}^{2}\overline{r}$.
The conditions \textbf{C6-C13} are assumed to hold true in Theorem \ref{theorem: inverse EKF stable unknown matrix}. Hence, all the conditions hold true for the I-EKF's error dynamics and Theorem \ref{theorem:Forward ekf stable unknown matrix} is applicable for the I-EKF as well, i.e., the estimation error is exponentially bounded in mean-squared sense and bounded with probability one.
\vspace{-8pt}
\section{Proof of Theorem~\ref{theorem: inverse ekf stable Reif}}
\label{App-thm-inverse ekf stable Reif}
We will show that the error dynamics of the I-EKF given by \eqref{eqn: inverse ekf error} satisfies the following conditions for all $k\geq 0$ for some real positive constants $\underline{c},\kappa_{\bar{\phi}}, \epsilon_{\bar{\phi}}$.
\begin{description}
\item [C1]
$\underline{c}\mathbf{I}\preceq\overline{\mathbf{Q}}_{k}$.
\item [C2]
$\widetilde{\mathbf{F}}^{x}_{k}$ is non-singular matrix for all $k\geq 0$.
\item [C3]
$\|\overline{\phi}_{k}(\hat{\mathbf{x}},\doublehat{\mathbf{x}})\|_{2}\leq\kappa_{\bar{\phi}}\|\hat{\mathbf{x}}-\doublehat{\mathbf{x}}\|^{2}_{2}$ for all $\|\hat{\mathbf{x}}-\doublehat{\mathbf{x}}\|_{2}\leq\epsilon_{\bar{\phi}}$ for some $\kappa_{\bar{\phi}}>0$ and $\epsilon_{\bar{\phi}}>0$.
\end{description}
All other conditions of Theorem \ref{theorem: ekf stable Reif} can be proved to hold true for the I-EKF's error dynamics under the assumptions of Theorem \ref{theorem: inverse ekf stable Reif} following similar approach as in proof of Theorem \ref{theorem: inverse EKF stable unknown matrix}, such that the estimation error given by \eqref{eqn: inverse ekf error} is exponentially bounded in mean-squared sense and bounded with probability one provided that the estimation error is bounded with $\overline{\epsilon}>0$ where $\overline{\epsilon}$ depends on the various bounds in the same manner as $\epsilon$ depends in the forward filter case.
For \textbf{C1}, using the bound on $\mathbf{R}_{k}$ from one of the assumptions of Theorem \ref{theorem: ekf stable Reif}, we have $\overline{\mathbf{Q}}_{k}=\mathbf{K}_{k}\mathbf{R}_{k}\mathbf{K}_{k}^{T}\succeq\underline{r}\mathbf{K}_{k}\mathbf{K}_{k}^{T}$.
Substituting for $\mathbf{K}_{k}$, we have
\par\noindent\small
\begin{align*}
\mathbf{K}_{k}\mathbf{K}_{k}^{T}=\mathbf{F}_{k}\bm{\Sigma}_{k}\mathbf{H}_{k}^{T}(\mathbf{H}_{k}\bm{\Sigma}_{k}\mathbf{H}_{k}^{T}+\mathbf{R}_{k})^{-2}\mathbf{H}_{k}\bm{\Sigma}_{k}\mathbf{F}_{k}^{T}.
\end{align*}
\normalsize
With the assumption that $\mathbf{H}_{k}$ is full column rank, $\mathbf{K}_{k}\mathbf{K}_{k}^{T}$ is p.d. as $\mathbf{F}_{k}$ is assumed to be non-singular in Theorem \ref{theorem: ekf stable Reif}. Hence, there exists a constant $\widetilde{q}>0$ which is the minimum eigenvalue of $\mathbf{K}_{k}\mathbf{K}_{k}^{T}$ such that $\mathbf{K}_{k}\mathbf{K}_{k}^{T}\succeq\widetilde{q}\mathbf{I}$ and $\overline{\mathbf{Q}}_{k}\succeq\underline{r}\widetilde{q}\mathbf{I}$. Hence, \textbf{C1} is satisfied with $\underline{c}=\underline{r}\widetilde{q}$.
For \textbf{C2}, $\widetilde{\mathbf{F}}^{x}_{k}=\mathbf{F}_{k}-\mathbf{K}_{k}\mathbf{H}_{k}$ is proved to be invertible for all $k\geq 0$ as an intermediate result in the proof of Theorem \ref{theorem: ekf stable Reif} in \cite[Lemma 3.1]{reif1999stochastic}.
For \textbf{C3}, using $\|\mathbf{K}_{k}\|\leq(\overline{f}\overline{\sigma}\overline{h}/\underline{r})$ (proved in \cite[Lemma 3.1]{reif1999stochastic}) and the bounds on functions $\phi(\cdot)$ and $\chi(\cdot)$ from the assumptions of Theorem \ref{theorem: ekf stable Reif}, we have $\|\overline{\phi}_{k}(\hat{\mathbf{x}},\doublehat{\mathbf{x}})\|_{2}\leq\|\phi(\hat{\mathbf{x}},\doublehat{\mathbf{x}})\|_{2}+\frac{\overline{f}\overline{\sigma}\overline{h}}{\underline{r}}\|\chi(\hat{\mathbf{x}},\doublehat{\mathbf{x}})\|_{2}\leq\left(\kappa_{\phi}+\frac{\overline{f}\overline{\sigma}\overline{h}}{\underline{r}}\kappa_{\chi}\right)\|\hat{\mathbf{x}}-\doublehat{\mathbf{x}}\|_{2}^{2}$,
for $\|\hat{\mathbf{x}}-\doublehat{\mathbf{x}}\|_{2}\leq \textrm{min}(\epsilon_{\phi},\epsilon_{\chi})$. Hence, \textbf{C3} is satisfied with $\kappa_{\bar{\phi}}=\kappa_{\phi}+(\overline{f}\overline{\sigma}\overline{h}/\underline{r})\kappa_{\chi}$ and $\epsilon_{\bar{\phi}}=\textrm{min}(\epsilon_{\phi},\epsilon_{\chi})$.
\bibliographystyle{IEEEtran}
|
\section{INTRODUCTION}
\label{sec:intro}
Non-redundant aperture-masking (NRM) interferometry introduces a mask with well selected holes into the pupil plane of a telescope to enable interferometric capabilities. Major ground-based observatories have successfully implemented NRM interferometry using existing instruments and observing procedures. Landmark discoveries such as dusty disks imaged around young stellar objects, mass-loss shells of evolved stars and the fascinating time-varying spiral plumes surrounding dusty Wolf-Rayet systems have been reported among the 50-odd peer-reviewed papers describing results produced by this technique\cite{1999Natur.398..487T, 2008ApJ...675..698T, 2006ApJ...650L.131L, 2007ApJ...661..496M, 2008ApJ...678L..59I}. The modern era of extreme adaptive optics (XAO) pushes this technique to its next level by dramatically reducing atmospherically-induced wavefront error. The last five years have seen the resurgence of this technique, which is now offered on major observatories around the world---the Very Large Telescope (VLT) with SPHERE and recently VISIR, on Gemini South with GPI, and on the Keck with NIRC2. Recent studies have revealed the full potential of this technique on various astrophysical domains such as planetary systems, protoplanetary disks, brown dwarfs, low-mass stars, and more\cite{2015Natur.527..342S, 2016A&A...588A.117S, 2018MNRAS.480.1006L, 2019A&A...628A.101H, 2019A&A...621A...7W, 2019A&A...622A..96C, 2019AJ....157..249G}.
However, the atmosphere still places limitations on NRM performance. Aperture masking suffers from rapid temporal instabilities due to atmospheric scintillation and transparency variations, as well as from differential atmospheric refraction. Space will provide an exceptionally stable environment that translates to improved performance. Factors that were once secondary will now limit space-based NRM contrast performance. Detector-related effects such as flat-fielding, intra-pixel response (IPR), inter-pixel coupling and charge diffusion, as well as operational considerations such as target acquisition repeatability, guiding jitter and pupil wander, will limit contrasts. Such flagship space missions as JWST has developed thorough knowledge of these diverse noise sources, with decades of modeling as well as plans to characterize them in-orbit. This should result in space-based NRM delivering an order of magnitude better performance than its existing ground-based precursors.
The Aperture Masking Interferometry (AMI\cite{2009SPIE.7440E..0YS}) mode on JWST's NIRISS\cite{2012SPIE.8442E..2RD} instrument will offer three medium bandpass filters (3.8, 4.3 and 4.8\,$\mu m$). AMI promises both high sensitivity and high angular resolution---beyond the Rayleigh criteria---down to $0.5\lambda/D$ (60\,mas at 3.8\,$\mu m$). These filters preserve the non-redundancy of the Fourier space sampling (i.e.: $u-v$ coverage) and provide image plane pixel scales comparable to or finer than Nyquist-sampled.
\begin{figure}[htbp!]
\centering
\includegraphics[width=.65\textwidth]{figures/spectra_phillips_niriss.pdf}
\caption[]{\label{fig:spectra_phillips_niriss}Theoretical spectrum of a cool T-Y brown dwarf\cite{2020A&A...637A..38P} ($\mathrm{T_{eff}}=900$ K, log(g) = 4.5). The NIRISS filters bandwidth available in AMI mode are overplotted. The relevant spectral features of CH$_4$, CO$_2$, CO and H$_2$0 are represented.}
\end{figure}
A faint companion detected at sub-Nyquist distance from a host star cannot be observed with NIRSpec or MIRI, and the 3 AMI filters will therefore provide a unique means to constrain its properties, hence the importance of its filter choice. The 3 main AMI filters (Fig.~\ref{fig:spectra_phillips_niriss}) sample the brighter part of the spectral-energy distribution of sub-stellar objects between very strong water features at $<3.5$\,$\mu$m and $>5$\,$\mu$m. The set of filters provides the best constraints on the properties of ultra-cool companions to young stars, being particularly sensitive to the CH4$_2$, CO$_2$ and CO features (3.8, 4.3, 4.8 $\mu m$ respectively \cite{2014SPIE.9143E..40A}). A fourth filter, operating at 2.77 $\mu m$, samples water absorption features although its performance is degraded since NIRISS' pixel scale is coarser than Nyquist at that wavelength. Fig.~\ref{fig:color-plot} shows the evolution of objects in the AMI color-color diagram, as a function of companion age temperature and surface gravity and illustrates how brown dwarfs gradually move through the color-color diagram depending on their bulk properties.
\begin{figure}[!htbp]
\centering
\includegraphics[width=.75\textwidth]{figures/cmd1.pdf}
\caption[]{\label{fig:color-plot} Color-color plot (top-left) in AMI bandpasses for brown dwarfs ranging in mass between 20 and 70\,M$_{\rm Jup}$, and ages of 30\,Myr to 2.4\,Gyr. Curves of a given color correspond in all plots to a given mass, while ages are given for the 20\,$_{\rm Jup}$ tracks. For objects colder than $\sim$1700\,K, the F380M-F430M color provides an excellent proxy of temperature (bottom-left) as it probes the onset of CH$_4$ absorption, a hallmark feature of colder brown dwarfs. The F430M-F480M color is affected by both temperature and surface gravity (top-right), informing on the radius and mass of an object of a given age.}
\end{figure}
The primary advantage of the AMI mode is its ability to probe the very core of the point spread function (PSF), where coronagraphic techniques on JWST are blind. Thus AMI complements other high contrast JWST imaging on NIRCAM and MIRI\cite{2010PASP..122..162B} in the range of inner working angles between 0.5$\lambda$/D and 4$\lambda$/D which are masked by the JWST coronagraphic spots. This opens the search for companions down to roughly 1~AU for the closest star systems ($\leq50$~pc).
The main scientific application of AMI is for high-contrast detection of point sources, suitable for exoplanets and brown dwarf characterization. With a contrast limit of $10^{-4}$ (Sect. \ref{sec:limits}) within the Rayleigh criterion ($\lambda/D$), AMI will be able to detect a few Jupiter mass exoplanets around nearby young (1-100\,Myrs) stars. For comparison, the current facilities offering such masking modes are limited to a contrast of $10^{-3}$, restraining the scientific cases to the faint binaries\cite{2019A&A...622A..96C, 2019AJ....157..249G}. AMI has also shown its ability to probe the inner structure of nearby Active Galactic Nuclei (AGN), thanks to its relatively good Fourier (or $u-v$) plane coverage\cite{2012SPIE.8442E..2SS}. AMI can benefit from the JWST orientation between two visit to extend its Fourier coverage, filling gaps in the Fourier plane. Such $u-v$ coverage allows image reconstruction using the existing algorithms (\texttt{IRBIS}, \texttt{Squeeze}, \texttt{Mira}, etc.) to resolve the core of circumstellar environments (protoplanetary disks, Wolf-Rayet stars, arc structures, etc.).
AMI will be particularly suited for follow-up observations of the major giant planet imagers such as SPHERE on the Very Large Telescope (VLT) and GPI on Gemini, providing a complementary wavelength coverage and better overall sampling of their spectral energy distribution (SED) and hence effective temperature and cooling history. The number of exoplanets directly imaged by such instruments continues to rise (51 planets in 2020), indicating a rich field for NIRISS' AMI mode to exploit.
The key advantage of space-based aperture masking will be the dramatically more stable optical system compared to any on the ground. In particular, atmospheric instability typically makes ground-based visibility amplitude calibration laborious, and often scientifically unusable. Space-based observations will add this new dimension to AMI data reduction, and also yield more accurate closure phases than are obtainable from the ground. NIRISS' AMI will deliver interferometric observables that have no ground-based peer.
In this work, we present the advanced simulation tool \texttt{ami\_sim} (Sect. \ref{sec:ami-sim}). We review the two existing methods of extracting interferometric observables (Sect. \ref{sec:extraction}) from image plane interferograms. We apply these tools in an end-to-end analysis to quantify the expected contrast limits of the first space-based aperture mask data (Sect. \ref{sec:limits}). We discuss our candidate calibrator star vetting program which was deemed essential to support AMI's expected performance improvement and present our ground-based measurements of ten calibrators using both adaptive optics instrument and long baseline interferometry (Sect. \ref{sec:vetting}).
\section{ADVANCED SIMULATION TOOLS}
\label{sec:ami-sim}
In preparation for the start of JWST operations, planned for six months after launch, a few different simulation packages were developed to provide data sets compatible with that of the future facility. \texttt{MIRAGE}\footnote{Available on \url{https://mirage-data-simulator.readthedocs.io/en/latest/}.} is the STscI-supported data simulation tool. It produces raw data cubes (including instrument artifacts), and is used to test STScI's Data Management System (DMS), i.e. JWST's supported data pipeline. But \texttt{MIRAGE} simulations are limited to native pixel size sampling which currently prevents seeding close binaries in a robust manner. In this work, we use a more tailored software, \texttt{ami\_sim}\footnote{Available on \url{https://github.com/anand0xff/ami_sim}.}, which is dedicated to the simulation of the aperture masking mode of NIRISS. The main advantage of \texttt{ami\_sim} is its ability to provide bad-pixel-free and reduced (\i.e., post-pipeline) data, coupled with a versatile interface, appropriate for rapidly testing different combinations of astronomical scenes (binary, planetary systems, disks, etc.) and different levels of telescope- and instrument-dependent noise. We also quantify and reduce errors in numerically-generated NRM PSFs that are an input to both data simulations. Our \texttt{ami\_sim} runs utilize these improved PSFs.
\subsection{Non-Redundant Mask Image}
\label{sec:newPSF}
Simulated AMI observations use a set of point-spread functions (PSFs), typically produced using the \texttt{WebbPSF} Python package\cite{2012SPIE.8442E..3DP, 2015ascl.soft04007P}. \texttt{WebbPSF} allows details of the filter bandpass, detector, and input spectral source to be specified, and provides considerable flexibility in the choice of telescope wavefront errors used to calculate the PSF. The pupil mask is defined by a numerical input FITS file. \texttt{WebbPSF} utilizes a matrix Fourier transform which enables a user-specified image pixel scale\cite{2007OExpr..1515935S}. However, there were discrepancies between the numerical \texttt{WebbPSF} simulations made with the previous version of the non-redundant mask image and the PSFs calculated analytically using the ImPlaneIA package\cite{2015ApJ...798...68G, 2018ascl.soft08004G}. What caused this inaccuracy is the input NRM pupil mask image \cite{JWST-STScI-005724}. By constructing the pupil mask on a larger array and binning down to the size required, we create a pupil mask image that produces simulated PSFs that agreed more closely with comparable analytical PSFs.
In order to be able to reliably detect features with contrast as low as $10^{-4}$ in AMI observations, we limited the maximum difference between comparable numerical and analytical PSFs to be of the order of $10^{-5}$. Using the locations of the centers of the primary mirror segments corresponding to the mask holes and the hexagonal sub-apertures’ flat-to-flat distance of 0.82\,m, we created a square ``parent” pupil mask array of side-length mag~$\times$~1024 (where mag is a positive integer). We binned the parent array down to \texttt{WebbPSF}’s commonly-used 1024$\times$1024 32-bit floating point array. This greyscales or ``anti-aliases” the pupil array we provide to \texttt{WebbPSF}. The geometry of the pupil mask and the difference in the treatment of the pixels at the sub-aperture edges in the improved mask compared to the previous version can be seen in Figure \ref{fig:mask}.
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.85\textwidth]{figures/maskfig_3.png}
\caption[example]
{\label{fig:mask} \textbf{Left:} Improved NRM pupil image. \textbf{Center:} Close-up of the edge pixels of a single sub-aperture in the previous mask version. \textbf{Right:} Close-up of the same region in the improved mask, showing subtle differences in sub-aperture edge pixel values.}
\end{figure}
Agreement between numerical and analytical PSFs improved monotonically with the parent array magnification ``mag”. At a magnification of 100 we achieved our desired agreement between the two PSFs. The largest parent array we could create (with a 335$\times$ magnification, given the limitations of our computing setup) produced numerical PSFs with maximum differences from the analytical ones on the order of $10^{-5}$ and median differences of the order of $10^{-7}$ between analytical and numerical PSFs when the peak pixel of each PSF is normalized to unity. The trend of median difference with increasing parent array size can be seen in Figure \ref{fig:nrm_improvement}. The improved agreement of the numerical PSFs created using the new mask to the analytical PSFs is supported by the agreement of the central pixel fractions of the numerical PSFs. While the PSFs made using the default pupil mask displayed a 2.8\% difference in the central pixel fraction (CPF) from the analytical ones, the new-mask’s PSF's CPFs differ by 0.17\%.
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.52\textwidth]{figures/parentarraytrends_mediandiff.pdf}
\caption[example]
{\label{fig:nrm_improvement} Median difference of analytically computed PSFs minus numerically computed (\texttt{WebbPSF}) PSFs: There is a sharp decrease with increased parent array size until a magnification factor of $\sim100$, remaining approximately constant over the rest of the range of magnifications tested.}
\end{figure}
\subsection{Simulation with \texttt{ami\_sim}}
The new pupil mask presented above was used to create a spatially dependent NRM PSF using \texttt{WebbPSF}\cite{2014SPIE.9143E..3XP}. In turn, that PSF is used by our simulation code (\texttt{ami\_sim}) to generate simulated binary point source image data.
We use a simulation framework that takes two-dimensional input sky scenes from the user and generates simulated observations at a single dither pointing near the center of an 80$\times$80 array with the non-redundant mask (NRM). The dither step is modeled with a 15 mas (rms, single-axis) Gaussian-distributed error. We do not model the pixel response function that is described by Hardy et al., 2014\cite{2014SPIE.9154E..2DH}, or inter-pixel capacitance (IPC) effects. We may implement IPC in a future release. We simulate uncorrelated, normally distributed, 0.1\% standard deviation pixel-to-pixel flat field error, correlated double sampling (CDS) readnoise of 21\,e$^-$, dark current of 0.04\,e$^-$/sec and an assumed background of 0.125\,e$^-$/sec in a pixel.
The AMI mode typically uses the NISRAPID readout pattern, which retains all frames read out up-the-ramp, i.e. one frame per group (NFRAME=1) and no gaps between groups (GROUP\_GAP=0). This means that after the initial reset that starts the integration, the subarray is continuously read out until the final read out of the integration is followed by a reset. An exposure at each dither position consists of NINT integrations (a.k.a. up-the-ramp) that are placed in a FITS data cube with dimensions FOV$\,\times\,$FOV$\,\times\,$NINT, where FOV is the size of the input target scene (in detector pixels). Each integration is made of NGROUPS up-the-ramp frames, each of TFRAME=0.07544$\,$s duration (which is the frame read time for the SUB80 AMI subarray). The number of groups in an integration (NGROUPS) is chosen to keep the peak pixel below 30$\,$000 e$^{–}$ in the brightest pixel. Integrations are accumulated to reach the required exposure depth (for simplicity we ignore the possibility that we reach the operational maximum number of integrations in the observation in our simulations).
A pointing jitter of 7 mas rms (one-axis) is introduced between each integration. We start with a noiseless count rate input image that is 11 times oversampled and convolve it with the appropriate NRM PSF. The size of the oversampled images is (FOV\,$\times$\,11)\,$\times$\,(FOV\,$\times$\,11) pixels. Thus, we can center our images at 121 different locations (1 pixel/(11$\times$11)) within a detector pixel. For each of the NINT realizations we create a simulated ramp with non-destructively read frames, and add Poisson noise, read noise, dark current, and background realizations to our frames. We then fit a slope to the simulated ramp for each individual pixel using a linear least-squares regression routine and create a ``slope image''. This slope image is divided by a flat field error array to introduce a 0.1\% flat field error standard deviation, and then divided by TFRAME to convert the slope image from counts (\i.e., detected photoelectrons) per frame to counts per second.
When used this way \texttt{ami\_sim}'s output is a datacube ($80$\,$\times$\,80\,$\times$\,$\mathrm{NINT}$). This corresponds to a JWST post-pipeline dataset, including the different noise sources described above (jitter, flat field, dark current, background, and readnoise). \texttt{ami\_sim} can simulate bad-pixel-free simulated images with realistic noise and flux fluctuation, optimally centered on the detector sub-array. Treatment of missing data from bad pixels prior to extracting interferometric observables is still under development, so we do not generate a realistic bad pixel map in \texttt{ami\_sim}.
\section{EXTRACTING INTERFEROMETRIC OBSERVABLES}
\label{sec:extraction}
There are two different approaches to extracting the relevant quantities (observables) from NRM data: SAMP, which relies on fitting an analytical fringe model in the image plane \cite{2011A&A...532A..72L}, and a second making measurements in the Fourier plane (\texttt{AMICAL}).
\subsection{From the image plane: \texttt{IMPLANEIA} }
\label{sec:implaneia}
The analytical SAMP fringe fitting methods were extended to include hexagonal sub-aperture shapes in the \texttt{IMPLANEIA} package\cite{2015ApJ...798...68G , 2018ascl.soft08004G}. The advantages of fitting data in the image plane include the ability to ignore bad pixel data, to avoid effects of windowing and tapering in the image plane inherent in numerical Fourier approaches, and to address image plane noise in the image plane itself. \texttt{IMPLANEIA} finds the centroid of the interferogram to sub-pixel accuracy, and generates fringes with a primary beam envelope, which are fit to the data after being binned down to detector pixels.
The outputs include the total flux and the pedestal level of the image, the fringe (or Fourier) phases, and the fringe amplitudes of the interferogram. \texttt{IMPLANEIA} implements a user-selectable image plane oversampling to use when creating its models, and has access to JWST's filter throughput files, as well as STScI’s \texttt{Pysynphot} used to reckon synthetic stellar spectra (e.g.:\, Phoenix model). It accepts three-dimensional cubes of integrations, each of which has the same exposure time, and outputs measurements for each cube in an OIFITS file, following the standard convention to store the interferometric observables\cite{2017A&A...597A...8D}. This data format is broadly use by the interferometric community and compatible with publicly available specialized software developed by JMMC\footnote{Jean-Marie Mariotti Center, \url{http://jmmc.fr}.}.
The package also provides routines to calibrate one OIFITS file with another, to create calibrated OIFITS input data for interferometric image reconstruction or other analysis software.
\subsection{From the Fourier plane: \texttt{AMICAL}}
\label{sec:amical}
The Aperture Masking Interferometry Calibration and Analysis Library (\texttt{AMICAL}\footnote{Available at \url{https://github.com/SydneyAstrophotonicInstrumentationLab/AMICAL}.}) is a modern Python version of the widely used Sydney aperture masking pipeline developed and improved during the last two decades\cite{1998SPIE.3350..839T, 2000SPIE.4006..491T, 2008ApJ...679..762K}. The code was developed with the main objective to be compatible with existing instruments having an AMI mode. The main advantages of the Fourier sampling approach lie in (1) its flexibility combining different masks, detectors or wavelengths and (2) its speed of execution (few minutes vs. hours).
We focused our efforts to propose a user-friendly interface, though different sub-classes allowing to (1) Clean the reduced datacube from the standard instrument pipelines, (2) Extract the interferometrical quantities (visibilities and closure phases) using a Fourier sampling approach and (3) Calibrate those quantities to remove the instrumental biases.
\\[12pt]
\noindent\textbf{Data processing} The standard calibration procedures followed by the usual pipelines are not sufficient to produce datasets suitable for our interferometric purposes. In the following, we present the extra steps required to calculate the observables.
The initial step consists of cropping the frames to reduce the numerical pressure on the Fast Fourier Transform (FFT), and retrieve a centered image. The centering aspect is essential to avoid numerical artifacts in the Fourier domain. Secondly, several latent issues are caused by the presence of bad pixels. In some cases, the data pipeline does not correct them and they need to be processed carefully during the cleaning step. In \texttt{AMICAL}, we propose a solution to remove these undesirable pixels applying a standard linear interpolation method. The replacement values are computed using the 2D Gaussian kernel convolution available within \texttt{Astropy}\cite{astropy:2018}. In practice, the bad pixel map -- formally retrieved by the JWST pipeline --- is used to generate a list of bad pixel coordinates. These bad pixels are then replaced by the values of the raw image convolved by the 2D Gaussian kernel function at the same native coordinates. This method allows us to interpolate the bad pixels by the values interpolated from a few pixels around, including the same image' noise levels. The use of \texttt{Astropy} package is computationally optimized, allowing to replace a potentially large quantity of pixels in a minimum of time.
Once cleared of bad pixels and centered, we perform background subtraction to ensure a zero pedestal. Then, individual frames are windowed with a super-Gaussian function of the form $e^{-ar^4}$, where $r$ is the radius in pixels and $a=1$ the amplitude of the Gaussian function. The windowing is used to create an interpixel correlation in the Fourier space and ensures to correctly sample the information on each Fourier peaks (so-called splodges). This approach also maintains a smooth transition to zero at the edge of the frames and limits the sensitivity to readout noise.
\begin{figure}[htbp!]
\centering
\includegraphics[width=.42\textwidth]{figures/cleaning_example.pdf}
\caption[example]
{\label{fig:example_cleaning} Example of reduced post-pipeline dataset from \texttt{MIRAGE} (top) versus cleaned data (bottom) and the associated power spectrum (right).}
\end{figure}
Finally, we perform a first data selection across the cube by maximizing the total flux in each frame (lucky imaging). This selection allows rejecting any bad integrations due to the adaptive optics system, seeing conditions or instrumental issues. In Figure \ref{fig:example_cleaning}, we present an illustration of raw and cleaned subframes and the dramatic impact on the power spectrum.
\paragraph{Extracting the observables}
As any interferometric approach, the NRM technique consists of extracting complex quantities encoded within each of the Fourier splodges (Fig. \ref{fig:method_sampling}) where the amplitude of the Fourier peak is identified with the corresponding baseline's fringe visibility (the absolute value of the Fourier transform) and its argument with the fringe phase (imaginary part of the Fourier transform). Although the Fourier phases are corrupted by atmospheric fluctuations and internal optical paths, the observable known as the closure phase (CP) -- the sum of baseline phases around closed triangles -- has been developed by radio astronomers to overcome this problem\cite{1958MNRAS.118..276J}.
Following established practice for NRM data\cite{Tuthill2000a}, we sampled Fourier spectra for each windowed frame depending on the mask coordinates, the wavelength and the pixel size. For a 7-holes mask ($N=7$), the corresponding $u-v$ coverage offers 21 ($N(N-1)/2$) baselines and 35 ($N(N-1)(N-2)/6$) closure phases.
\noindent For each baseline, we compute the $u$ and $v$ Fourier coordinates and convert them to centered detector units with:
\begin{align}
\label{eq:uvcoord}
u &= \left(\frac{x_{h1} - x_{h2}}{\lambda} \times \mathrm{FOV}\right) + N_{pix} \Mod{N_{pix}},\\
v &= \left(\frac{y_{h1} - y_{h2}}{\lambda} \times \mathrm{FOV}\right) + N_{pix} \Mod{N_{pix}},
\end{align}
where $x_{hi}$, $y_{hi}$ are the mask coordinates (m), FOV is the field of view (rad) and $N_{pix}$ the image dimension in detector pixels.
The finite mask hole size causes information from each baseline to be spread over multiple pixels in the Fourier transform (FT). To exploit information spread beyond just the $u$, $v$ position of the hole centers (Eq. \ref{eq:uvcoord}), we developed four different methods to sample the overall splodges or just a few pixels. The first naive method consists of rounding the $u$, $v$ position on the closest true pixel (``unique'', Fig. \ref{fig:method_sampling}). The second uses the fractional part of the $u$, $v$ position to weight 4 pixels around the expected positions (``square'', Fig. \ref{fig:method_sampling}). The third samples the splodge using a 2D Gaussian weight centered on the rounded $u$, $v$ position (``gauss'', Fig. \ref{fig:method_sampling}). And finally, we apply a ``Fourier optics'' approach using a series of FT applied to the aperture pairs. Then, we threshold and normalize the result to ensure a sum equal to 1 for each splodge (``fft'', Fig. \ref{fig:method_sampling}). This last method takes advantage of computing pixel weights on the ``true'' Fourier grid by the use of FFTs. We recommend this method because of the close-to-Nyquist sampling of the NIRISS detector, which leads to a lack of resolution in Fourier space.
\begin{figure}[htbp!]
\centering
\includegraphics[width=.75\textwidth]{figures/method_sampling.pdf}
\caption[example]
{\label{fig:method_sampling} Different sampling methods implemented in \texttt{AMICAL}. Zoom on one of the 43 splodges (symmetric 21 baselines + central peak, Fig. \ref{fig:example_cleaning}) present in the amplitude map of the Fourier plane.}
\end{figure}
For a given sampling method, we generate the pixel coordinates and the associated weights for each baseline which are used to yield complex visibility data (per frame).
We then accumulate the robust observables comprising the squared visibilities (V$_{ij}^2$) and closure phases (CP$_{ijk}$) over the data cube. In practice, we compute the squared complex visibility (power spectrum, Eq. \ref{eq:v2}) and the complex triple product (bispectrum, Eq. \ref{eq:bispectrum}) formed by a subset of closing triangles (e.g., apertures 1, 2 and 3). The closure phase is then computed as the argument of this bispectrum (Eq. \ref{eq:cp}). For the V$_{ij}^2$, we measure any bias by taking the average power for regions in Fourier space without any signal. We finally subtract this bias and then normalize by the power at zero baseline (equivalent to normalizing by total flux in the image).
\begin{align}
\label{eq:v2}
V^2_{12} &= |C_{12}|^2,\\
\label{eq:bispectrum}
BS_{123} &= C_{12}C_{23}C_{13}^*,\\
\label{eq:cp}
CP_{123} &= arg(BS_{123}) = \atantwo(BS_{imag}, BS_{real}),
\end{align}
We also implement an alternative method to compute the closure phases using the so-called ``Monnier'' method \cite{1999PhDT........19M}. It consists of averaging many bispectra in each individual image for each triangle of baselines. Once more, the splodges are spread over many pixels which can represent a variety of independent closing triangles. For a given splodge, we define a diameter and account for the different combinations existing to compute the averaged closure phases. This method is computationally more expensive than the Fourier sampling, since hundreds of pixel triangles could exist for each bispectra. However, for most data tested with \texttt{AMICAL}, this method did not present any advantages in terms of scatter or accuracy. We therefore recommend to use the standard sampling approach for NIRISS (or other adaptive optics assisted instruments).
Following the formalism presented by Gordon et al. 2012\cite{2012A&A...541A..46G}, we use their bias-free estimator to compute the closure phase errors (Eq. 40). This method is preferred from the naive standard deviation because it is insensitive to wrapping uncertainty in the complex plane when the variance is large. For the squared visibilities, we used the diagonal of the covariance matrix as a proxy for the errors.
\paragraph{Calibration procedure} Closure phases and square visibilities suffer from systematic terms, caused by the wavefront fluctuations (temporal, polychromatic sources, non-zero size mask, etc.). To calibrate aperture masking data, these quantities are measured on identified point source calibrator stars. In practice, we subtract the calibrator signal from the raw closure phases and normalize the target visibilities by the calibrator's visibilities.
If several calibrators are available, the calibration factors are computed using a weighted average to account for variations between sources. The extra errors induced are then quadratically added to the calibrated uncertainties. During the calibration procedure, a second data selection is performed to reject bad calibrator-source pairs using a sigma-clipping approach. The calibrated observables are finally stored in \texttt{OIFITS} files.
\section{PLANET AND BINARY DETECTION LIMITS}
\label{sec:limits}
In this section, we present the updated detection limit estimates achievable with the AMI mode of NIRISS. Previous determinations were performed on cryogenic laboratory data and were focused on the reachable closure phases uncertainties\cite{2015ApJ...798...68G}. In this work, we use the extracted interferometric quantities themselves in an observational approach with the use of standard interferometric tools. Ireland et al. 2013\cite{2013MNRAS.433.1718I} demonstrated that the CP uncertainty is proportional to the number of apertures ($N_{holes}$) and photons ($N_{photons}$) following:
\begin{equation}
\label{eq:ireland2013}
\sigma_{CP} = \sqrt{1.5}\times\frac{N_{holes}}{\sqrt{N_{photons}}}
\end{equation}
However, a companion signal in the closure phase is directly proportional to the contrast ratio relative to its host star. For two unresolved stars (or star and planet), the ultimate detection limit can be approximated by the CP uncertainty in radians. According to Eq. \ref{eq:ireland2013}, a contrast of $10^{-4}$ (10 mags) can be reached with $\approx 10^{10}$ photons. This limit represents the goal of AMI-NIRISS and will be investigated using our best understanding of the instrument.
To compute our simulated data set, we used the improved simulated PSF generated based on our revised mask treatment (Sect. \ref{sec:newPSF}). We use \texttt{ami\_sim} (Sect. \ref{sec:ami-sim}) to include the different noise sources (flat-field, background, photon noise, jitter, etc.) and produce the post-pipeline data cube. As for the standard observational procedure, CDS is applied on individual groups (or frames) to reduce them to a single integration, which is accumulated to reach the desired number of photons. The number of groups and integrations are determined using the official Exposure Time Calculator (ETC\footnote{Available on \url{https://jwst.etc.stsci.edu}.}). We considered a sun-like star of spectral star G2V ($\mathrm{T_{eff}=5450\,K}$, $\mathrm{log\,g}=4.5$), represented by the Phoenix spectrum model. We normalize the continuum in the F380M filter of NIRISS to represent a star of magnitude 6 (Vega). This number was chosen to be far enough below the saturation limit, allowing to integrate more groups in a realistic amount of time (Tab. \ref{tab:simu}). Our astrophysical scene is represented by two unresolved components, with a contrast of 20 mags (too faint for NIRISS), a separation of 150 mas and a position angle of 20$^\circ$.
\begin{table}[htbp!]
\begin{center}
\caption{Operational parameters used for the AMI-NIRISS simulation (G2V star of 6 mag in F380M filter). The different number of integrations represent the collected number of photons of 10$^7$, 10$^8$, 10$^9$ and 10$^{10}$ respectively.}
\vspace{.1cm}
\label{tab:simu}
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{c c c c}
\hline
\hline
& F380M & F430M & F480M\\
\hline
Countrate [e$^{-}$/sec] & 1999651 & 1279391 & 1137034\\
N$_{grp}$ & 7 & 14 & 19\\
N$_{int}$ & 11, 97, 960, 9590 & 8, 76, 750, 7495 & 7, 63, 622, 6214 \\ \hline
\end{tabular}
\end{center}
\end{table}
The interferometric quantities (visibility and CP) are then extracted using the \texttt{AMICAL} software. For NIRISS, we apply the standard procedure consisting of cleaning the data (centering, windowing, background subtraction) and extracting the observables using the FFT method sampling (Sect. \ref{sec:amical}). The contrast limit performances are computed using the CANDID package\cite{2015A&A...579A..68G} between 50 mas ($\approx 0.5\lambda/D$) and 400 mas ($3.5\lambda/D$), which is the inner working angle domain best probed by NIRISS-AMI. Closer companions should be reached using long baseline interferometry or radial velocity techniques whereas outer companions are best observed using coronagraphy techniques.
\begin{figure}[htbp!]
\centering
\floatbox[{\capbeside\thisfloatsetup{capbesideposition={right, top},capbesidewidth=5cm}}]{figure}[\FBwidth]
{\caption{\label{fig:NIRISSlimits} Expected performance of NIRISS AMI: 3--$\sigma$ contrast as a function exposure depth. The photon noise limited cases are represented as the horizontal dashed-dotted lines (see Eq. \ref{eq:ireland2013}). Colors stand for the number of collected photons and the line styles represent the AMI filters (full, dashed and dotted lines for 3.8 µm, 4.3 µm and 4.8 µm respectively).}}
{\includegraphics[width=.55\textwidth]{figures/contrast_niriss_candid_mag6_v3.pdf}}
\end{figure}
We varied exposure depth by simulating data with varying number of photons collected (10$^7$, 10$^8$, 10$^9$ and 10$^{10}$), and derived NIRISS AMI's contrast detection limit curves as a function of separation (Figure \ref{fig:NIRISSlimits}). The given performance correspond to the 3-$\sigma$ confidence interval ($99.73\,\%$) for the three filters of NIRISS (F380M, F430M and F480M). A maximum contrast detection limit of 10 magnitudes is reachable with 10$^{10}$ collected photons (Tab. \ref{tab:limits}), remarkably close to the photon noise limited prediction (Eq. \ref{eq:ireland2013}). We notice that on brighter sources, the total number of photons collectible in a reasonable amount of time ($<$ few hours) and under the technical limitation of 10000 individual integrations per exposure, can be almost doubled allowing to gain up to 0.2 magnitudes in contrast.
\begin{table}[htbp!]
\begin{center}
\caption{Contrast performance of AMI-NIRISS between 50 and 400 mas depending on the exposure depths (10$^7$, 10$^8$, 10$^9$ and 10$^{10}$ photons). The given magnitude ranges represent the minimum and maximum values showed in Fig. \ref{fig:NIRISSlimits}.}
\vspace{.1cm}
\label{tab:limits}
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{c c c c}
\hline
\hline
N$_{phot}$ & F380M [mag] & F430M [mag] & F480M [mag]\\
\hline
10$^{7}$ & 5.4--6.3 & 4.5--6.2 & 4.1--5.9\\
10$^{8}$ & 6.8--7.7 & 5.9--7.7 & 5.4--7.1\\
10$^{9}$ & 7.8--8.9 & 6.9--8.5 & 6.7--8.6\\
10$^{10}$ & 8.9--10.1 & 8.2--9.9 & 8.1--10.0\\
\end{tabular}
\end{center}
\end{table}
The additional source of error ($\approx 15\%$ compared to the theoretical photon noise limited calculation) are probably due a combination of instrumental uncertainties (blurring caused by the jitter, correlated pixel-to-pixel flat field error, etc.) and data processing effects (imperfect Fourier sampling due to the limited field of view, incomplete background subtraction, etc.). We defer the investigation of further sources of noise such as the effects of bad pixel correction, the intra-pixel response (IPR\cite{2014SPIE.9154E..2DH}) or the inter-pixel capacitance (IPC) to a future paper.
\section{CALIBRATOR STAR VETTING PROGRAM}
\label{sec:vetting}
The vastly improved levels of performance that the NRM interferometric mode of JWST will deliver should contribute significantly in several science topics. With no turbulent atmosphere to disturb the wavefronts, no varying gravity loading to distort the optics, and an exquisitely cold and stable thermal environment, the levels of optical stability provided by the instrument will be unprecedented\cite{2010SPIE.7731E..0FD, 2012SPIE.8442E..2SS}. Nevertheless, in order to benefit from these intrinsic gains in calibration fidelity, it is also necessary to ensure a perfectly known instrumental PSF. To do so, we observed a set of PSF reference stars chosen by the accepted AMI-NIRISS commissioning and observing programs, to verify their suitability. We used among the most advanced ground-base facilities available at the Very Large Telescope (VLT) with SPHERE\cite{2019A&A...631A.155B} and the Very Large Telescope Interferometer (VLTI) with GRAVITY\cite{2017A&A...602A..94G}.
\subsection{Observation and data reduction}
\subsubsection{Aperture Masking Interferometry with SPHERE}
SPHERE is a high-performance imaging instrument equipped with an Extreme Adaptive Optics system (XAO)\cite{2019A&A...631A.155B}. We observed HD15633, the selected calibrator for the AGN program of NIRISS (PI: K. Ford), and two additional point sources HD15720 and HD16261. Our observations were obtained between October and November 2017 with the IRDIS infrared camera \cite{2008SPIE.7018E..59D, 2014SPIE.9147E..1RL}. We used SPHERE in its interferometric mode using the 7-holes mask available since 2017 \cite{2016SPIE.9907E..2TC}, using the dual-band imaging with K1K2 filter pair (2.110$\pm0.102$ and 2.251$\pm0.109$ $\mu m$).
Data reduction was performed following the procedures described in Zurlo et al. 2016\cite{2016A&A...587A..57Z} with the official ESO-pipeline\footnote{Version 0.4, \url{https://www.eso.org/sci/software/pipelines/index.html}.}. The raw images were reduced by performing background subtraction, bad-pixel correction and flat fielding. We extracted the interferometric observables using the Python package \texttt{AMICAL} (Sect. \ref{sec:amical}). The post-pipeline cubes were cropped, windowed and background subtracted. Once cleaned, we used the FFT method to sample the Fourier space. We calibrated each target by the ones observed closest in time. We finally generated a total of 12 standards calibrated OIFITS2\cite{2017A&A...597A...8D} files for each target.
\subsubsection{Long baseline interferometry with GRAVITY}
The second dataset were observed with VLTI/GRAVITY\cite{2017A&A...602A..94G} in the $K$ band (1.95–2.5$\mu$m) between 2017 and 2018 using the four 1.8\,m Auxiliary Telescopes (ATs). Two targets (HD36805, HD37093) were observed using the small baseline configuration (A0-B2-C1-D0, 11-34m) and five ($\delta$ Crv, HD4981, HD93372, HD93649 and HD101531) using the large configuration (A0-G1-J2-J3, 58-132m). Our sample includes seven different targets, covering four JWST/NIRISS guaranteed time observation proposals (program ID: 1242 (PI: Johnstone), 1200 (PI: Rameau), 1260 (PI: Ford)). HD36805 and HD37093 will be observed during the calibration plan of NIRISS to compute the phase reference. For each target, we observed 2 calibrators (before and after) to retrieve the atmospheric transfer function.
The data were reduced using the GRAVITY pipeline\cite{2014SPIE.9146E..2DL}, through reduction recipes made available by the
GRAVITY consortium in their python toolkit\footnote{Available at \url{https://version-lesia.obspm.fr/repos/DRS_gravity/python_tools/}.}. Six squared visibilities and four closure phases were obtained with each observation, for both the p and s polarization directions. This yielded two data sets at medium spectral resolution (R=500), and low resolution for the fringe tracker. As there was no significant difference between the two polarisations, we averaged them. We only used the science camera data in our analysis.
\subsection{Companion search and sensitivity limits}
According to our analysis, there is no evidence of multiplicity in our sample. An example of interferometric observables obtained with AMI-SPHERE on HD15633 is presented in Fig. \ref{fig:example_irdis_data} showing both closure phases and squared visibilities compatible with a point source model. In this case, the visibilities seem to be systematically above one (within the uncertainties), a sign of marginally resolved calibrator.
\begin{figure}[htbp!]
\centering
\includegraphics[width=.9\textwidth]{figures/example_uv_hd15633}
\caption[example]
{\label{fig:example_irdis_data} Example of interferometric observables obtained on HD15633. \textbf{Left:} $u-v$ coverage, \textbf{Top-right}: Squared visibilities, \textbf{Bottom-right}: Closure phases. Both visibilities and closure phases are compatible with an unresolved point source (CP $=0$ deg, V$^2=1$).}
\end{figure}
As for NIRISS, we use \texttt{CANDID}\footnote{Available at \url{https://github.com/amerand/CANDID}} to perform a grid search to probe the immediate surroundings of our reference stars (20-250 mas). At each starting position in the grid, the companion position, its flux ratio, and the angular diameters of the primary are fitted. The \texttt{CANDID} code also includes a tool to estimate the companion detection significance level (in units of sigma). We perform this systematic search using both sets of available observables at once (CP+V$^2$) or separately (CP, V$^2$) without any evidence of multiplicity.
For the SPHERE dataset, we utilize \texttt{AMICAL} and the package \texttt{pymask}\footnote{Available at \url{https://github.com/AnthonyCheetham/pymask}.} to compute the sensitivity limits (Fig. \ref{fig:contrast_limit_sphere}). \texttt{pymask} only uses the closure phases which are more reliable measurements and drops the often biased visibilities of the NRM technique. Our best contrast limit was obtained during the first epoch and reached a maximum contrast of 6.8 mags ($\approx520$ in flux ratio) at $\approx\lambda/D$ separation. The contrast ratio achieved with our SPHERE data set is relatively constant between 50 and 250 mas for all calibrators, being between 6.5 to 7 at 2.11 $\mu$m. These contrast limits are notably close to the best contrast ratio reported from the ground with the NRM techniques \cite{2011ApJ...731....8K, 2019JATIS...5a8001S}.
\begin{figure}[htbp!]
\centering
\includegraphics[width=.6\textwidth]{figures/contrast_SPHERE_epoch1.pdf}
\caption[example]
{\label{fig:contrast_limit_sphere}Contrast curves as a function of separation obtained with SPHERE in $K$ band (3--$\sigma$ confidence).}
\end{figure}
For the GRAVITY sample, the data quality was unequal depending on the baseline configurations and targets. Some targets appeared to be smaller than the associated calibrator and so present visibilities above one. For these stars (HD36805, HD93372, HD93649), we applied the same procedure as for SAM-SPHERE dataset and used only the closure phases to compute their contrast limits (Left, Fig. \ref{fig:contrast_gravity}).
\begin{figure}[htbp!]
\centering
\includegraphics[width=.8\textwidth]{figures/contrast_GRAVITY}
\caption[example]
{\label{fig:contrast_gravity} Contrast curves as a function of separation obtained with GRAVITY in K-band (3--$\sigma$ confidence). \textbf{Left:} Limits for the targets using only the visibilities \textbf{Right:} Limits for the targets resolved by GRAVITY (V$^2$+CP).}
\end{figure}
The other stars ($\delta$ Crv, HD37093, HD101531 and HR4981) were barely resolved by the longest baselines. For these targets, we used \texttt{CANDID} to estimate the contrast limits and include the resolved primary component (Fig. \ref{fig:contrast_gravity}, right panel). The contrast detection limits obtained with GRAVITY are relatively constant across the separation, between 5 and 6 magnitudes (100-250 in flux). We carefully checked the region around the detected peak found for the shortest baseline configuration (around 30 and 50 mas), but without any clear companion detection.
With this study, we confirmed that the 10 reported calibrator stars are trustworthy point sources (in the range of performance of VLT/SPHERE and VLTI/GRAVITY). They are suitable to be observed by the NIRISS consortium during the commissioning and across the GTO programs.
\section{Conclusion}
With the launch planned for October 2021, the James Webb Space Telescope will represent the flagship infrared observatory, surpassing all its precursors in both sensitivity and angular resolution for the next decades. The brand new interferometric mode offered by NIRISS, Aperture Masking Interferometry (AMI), will push back these limits allowing to probe inside the Rayleigh criterion in an unprecedented way for this spectral-domain (3--5 $\mu$m). As we have seen, the predicted contrast 10$^{-4}$ with AMI-NIRISS is much better than that achieved from the ground (e.g., SPHERE, Fig. \ref{fig:contrast_limit_sphere}). Free of atmospheric constraints, the AMI technique will reach its full potential, opening studies of planetary systems as much as ten magnitudes fainter than their host star. Such contrasts allow characterizing planets of a few Jupiter mass and their parent protoplanetary disk. The promising performance predictions of AMI-NIRISS should inspire further development of masking techniques for the next generation of space telescopes (e.g., NASA/LUVOIR), enabling the full potential of such facilities for a minimal investment.
\acknowledgments
We thanks the JWST/NIRISS consortium both at the IREX institute of Montréal, Canada and at the Space Telescope Science Institute of Baltimore, USA. We acknowledge support from the Australian Research Council (DP 180103408) that funded this work. We also thank the ESO Paranal staff for support for conducting the observations reported in this paper. We thanks the support of the French VLTI center, the ``\textit{Service Utilisateur du VLTI}" (SUV, PI A. Matter OCA-LAGRANGE), especially the help of K. Perraut (IPAG) for the GRAVITY data reduction. We thanks all \texttt{AMICAL} friends for their involvement and work and thus acknowledge C. Robert, A. Cheetham, D. Johnstone, D. Blakely for their assistance. We thanks the major developers involved in the IDL version of the Sydney code J. Monnier, M. Ireland, A. Cheetham among others, and A. Greenbaum, L. Pueyo, and S. Lacour for their leadership developing analytical approach to extracting observables implemented in \texttt{ami\_sim}.
\newpage
|
\section{Introduction}
Nuclear spins in material such as diamond associated with single defects represent a promising platform for quantum information registers \cite{Taminiau2014UniversalCA,2018Pulse,Waldherr2014QuantumEC,2011High} and sensing purposes \cite{Taminiau2012DetectionAC,Kolkowitz2012SensingDN,Zhao2012SensingSR,Ermakova2013DetectionOA,Mamin2013NanoscaleNM,Zaiser2016EnhancingQS,
Staudacher2013NuclearMR,Wu2016DiamondQD} due to their long coherence times and the potentially large number of available spins. Nuclear spins can be initialized, controlled, and read out through the electron spin of the nitrogen vacancy (NV) center \cite{Doherty2013TheNC} driven by optical fields and microwave radiation \cite{2016Optomechanical,2014Optically}. Recent progress manifests in a number of works, such as electron-nuclear \cite{Casanova2016NoiseResilientQC,Bermdez2011ElectronmediatedNI,zimmermann2020selective,rong2014implementation,hegde2020efficient,
abobeih2019atomic,Bradley2019ATS,Wang2017DelayedEE,Tratzmiller2021ParallelSN,Casanova2017ArbitraryNG}, electron-electron \cite{degen2021entanglement} and nuclear-nuclear quantum gates \cite{Bian2017UniversalQC}. Many schemes are proposed to realize electron-nuclear quantum gate operations, both with \cite{Bradley2019ATS,Wang2017DelayedEE} and without \cite{Tratzmiller2021ParallelSN,Casanova2017ArbitraryNG} additional radio-frequency (RF) control on the nuclear spins themselves.
Nuclear-nuclear quantum gate is implemented by a sequence of the NV-nuclear quantum operations \cite{Taminiau2014UniversalCA,2018Pulse}, and it is a delicate issue to have a complete set of quantum gates on specific nuclei in samples \cite{Bradley2019ATS}. Current related experiments are operated at low temperature, because of the relatively short lifetime of the NV center. Several outstanding challenges are caused by the relaxation and decoherence processes of the electron spin as these limit quantum gate fidelities on nuclear registers as well as spectral resolution and selectivity. A scheme was proposed to have highly selective and high fidelity quantum gates between nuclear spins under ambient conditions \cite{Chen2017DissipativelySQ}. However, it has strict requirements of resonance conditions and is applicable for special cases, i.e., for nuclear spins in the same species with the parallel coupling components between the NV and nuclear spins far smaller than the vertical components.
Here we extend the scheme in Ref. \cite{Chen2017DissipativelySQ} to a heteronuclear case by using radio-frequency fields to control different nuclear spin species individually. The effective substantial second-order coupling between the nuclear spins obtained through a MW driving NV center which is periodically reinitialized by a dissipative process \cite{Chen2017DissipativelySQ}. Similarly, the periodical reinitialization of the NV center decouple it from the dynamics and its effect on system is an effective weak dissipation process. Thus high selectivity and fidelity of nuclear-spin quantum gate for different nuclear spin species could be possible even at ambient condition, which is an extension for quantum computation and simulation applications by using different nuclear species controlled by NV centers. Additionally, one can use a nuclear spin as a quantum sensor to detect nuclear spins in another species as well as analysis of complex spin structures.
This paper is structured as follows. We start with the derivation of effective Hamiltonian of the system by using Schrieffer-Wolf transformation\cite{Kessler2012GeneralizedSF,Bravyi2011SchriefferWolffTF}, which demonstrates the validity of the indirect interaction between nuclear spins by the application of suitably tuned RF fields. We take two relevant species, i.e., carbon-13 and silicon-29 spins as examples to show our protocol efficiency. Coherent evolution between a carbon-13 and a silicon-29 spin is limited by the NV lifetime, this limitation is overcome through periodical resets of NV spin. The feasibility of high selectivity and fidelity nuclear-nuclear quantum gates by using RF controls is discussed as well as the sensing application. Finally, we compare our approach to the previous scheme in Ref. \cite{Chen2017DissipativelySQ}.
\section{Indirect interaction between nuclear spins}
We consider two nuclear spins in different species are coupled to an electron spin of single NV center, and microwave (MW) field and radio-frequency (RF) field are used for external control over the electron and nuclear spins as well as for achieving selective internuclear interactions, see Fig. \ref{NoReset}(a). The magnetic field $B$ is applied along the NV axis (the $\hat{z}$ axis), which is large enough to split the degenerate states of $|m_s=|\pm1\rangle$. The Hamiltonian of the whole system is given by
\begin{eqnarray}
H&=&DS_{z}^{2}+\gamma_{e}BS_z+\sum_{i}\gamma_{n_i}BI_i^z \\
& &+S_z\sum_{i}\overrightarrow{A_i}\cdot\overrightarrow{I_i}+H_{w}+H_{r}. \notag
\label{eq1}
\end{eqnarray}
Note that we hereafter omit the Dirac constant $\hbar$ for simplicity. Here $D=(2\pi)2.87$ GHz denotes zero-field splitting of the electronic ground state, $\gamma_{e}$ and $\gamma_{n_i}$ are the gyromagenetic ratio of the electron spin and nuclear spins, respectively. The interaction between the NV center
and the $i$th nucleus is mediated by the hyperfine vector $\overrightarrow{A_i}$, $A_{i}=(a_{\parallel i},a_{\perp i})$ with $a_{\parallel i}$ and $a_{\perp i}$ denotes the related coupling parallel and perpendicular components to the nuclear spin quantization axes $a_{\parallel i}=\overrightarrow{A}_{i}\cdot\hat{z}$ and $a_{\perp i}=\sqrt{|\overrightarrow{A}_{i}|^{2}-a^{2}_{\parallel i}}$. Hamiltonians $H_{w}$ and $H_{r}$ describe the action of the MW driving and RF fields, respectively, which reads
\begin{eqnarray}
H_{w}&=&\sqrt{2}\Omega\cos\omega tS_{x}, \\
H_{r}&=&\sum_{i}2\Omega_{rfi}\cos\omega_{rfi}tI_{i}^{x},
\label{driving}
\end{eqnarray}
where $\Omega$ ($\Omega_{rfi}$) is the Rabi frequency of the MW (RF) driving field with the corresponding frequency $\omega$ ($\omega_{rfi}$).
The MW driving field is applied to be on resonance of the transition $|-1\rangle\leftrightarrow|0\rangle$ as $\omega=D-\gamma_eB$, we can rewrite the effective Hamiltonian of the NV center as $H_{NV}=\Omega\sigma_{z}$. Then in the rotating frame with $H_0=(D-\gamma_eB)|-1\rangle\langle-1|$, the Hamiltonian could be rewritten as
\begin{eqnarray}
H_{t}&=&\Omega\sigma_{z}+\sum_{i}(\gamma_{ni}B+\frac{a_{\parallel i}}{2})I^{z}_{i} \notag\\
& &+a_{\parallel i}\sigma_{x}I^{z}_{i}+a_{\perp i}\sigma_{x}I^{x}_{i}+H_{r},
\label{Ht}
\end{eqnarray}
where $\sigma_{z}=\frac{1}{2}(|+\rangle_e\langle+|-|-\rangle_{e}\langle-|)$ with $|+\rangle_{e}=\frac{1}{\sqrt{2}}(|0\rangle+|-1\rangle)$ and $|-\rangle_{e}=\frac{1}{\sqrt{2}}(|0\rangle-|-1\rangle)$.
\begin{figure*}[t]
\center
\includegraphics[width=6.8in]{Fig1.pdf}
\caption{(a) The NV center mediates the coupling between a silicon-29 spin with $\gamma_{n_1}B=(2\pi)4$ MHz and a carbon-13 spin $\gamma_{n_2}B=(2\pi)5.06$ MHz to achieve a quantum gate, while itself is decoupled from the dynamics. RF fields are applied to control two nuclear spin species individually, when we use continual MW field to drive the NV spin. (b) Population evolutions are shown by considering the ideal case in which no dissipation is included and the nuclear spins is initialized in state $|+_{1}-_{2}\rangle$. The Rabi frequencies of RF fields are given by $\Omega_{rf1}=\Omega_{rf2}=(2\pi) 1$ kHz, microwave driving field $\Omega=(2\pi) 400$ kHz, the parallel coupling components $[a_{\parallel 1},a_{\parallel 2}]=(2\pi) [9,11]$ kHz and the effective coupling coefficient $g_{e}=(2\pi) 0.12$ kHz. These exact numerical simulations by using Eq. (\ref{H't}) fit well to the theoretical derivation of the effective (eff) dynamics under Hamiltonian Eq. (\ref{He}). (c) By considering the practical case, we use the same parameters in (b) and the NV dissipation is included as $T_{1\rho}=200$ $\mu s$, exact numerical simulation of the master equation Eq. (\ref{master}) shows that there is no coherent evolution between the nuclear spins due to the NV life time limitation. (d) The periodical NV resets makes it possible to extend the coherent evolution of the heteronuclear spins well beyond the NV spin lifetime. The exact numerical calculation is based on Eq. (\ref{Evolution}), when we simulate the effective master equation Eq. (\ref{masterN}) of nuclear spins. The parameters are as the same as in (b) except that the NV center is reinitialied to state $|-\rangle_e$ every $t_{re}=20$ $\mu s$. (e) Process infidelity for different detuning and errors of the MW driving of the NV center. }
\label{NoReset}
\end{figure*}
Two weak RF fields are applied to individually control two different nuclear spins with $\omega_{rfi}=\gamma_{n_i}B+\frac{a_{\parallel i}}{2}$. Working in the rotating frame with $H'_{0}=\sum_i(\gamma_{ni}B+\frac{a_{\parallel i}}{2})I^{z}_{i}$, by assuming $a_{\parallel i},a_{\perp i}\ll\Omega\ll\omega_{rfi}$ and $a_{\parallel i}$ is comparable or larger than $a_{\perp i}$, and the total Hamiltonian of the system can be approximated as
\begin{eqnarray}\label{eq3}
H'_{t}&=&\Omega\sigma_{z}+\sum_{i}\Omega_{rfi}I^{x}_{i}+a_{\parallel i}\sigma_{x}I^{z}_{i}.
\label{H't}
\end{eqnarray}
Consider the Hamiltonian is of the form $H'_{t}= H'_{NV}+V$, where $H'_{NV}=\Omega\sigma_{z}$ denotes the Hamiltonian of the MW driving and $V=\sum_{i}\Omega_{rfi}I^{x}_{i}+a_{\parallel i}\sigma_{x}I^{z}_{i}$ is the weak perturbation with $a_{\parallel i},\Omega_{rfi}\ll\Omega$. By using the Schrieffer-Wolff transformation \cite{Kessler2012GeneralizedSF,Bravyi2011SchriefferWolffTF} in condensed matter, the second order expansion due to perturbation terms $V$ can be obtained as
\begin{eqnarray}
\langle\alpha|H_{e}|\beta\rangle_e=\frac{1}{2}\sum_{i}(\frac{\langle\alpha|V|i\rangle\langle i|V|\beta\rangle_e}{E_\alpha-E_i}-\frac{\langle\alpha|V|i\rangle\langle i|V|\beta\rangle_e}{E_i-E_{\beta}}),
\label{SW}
\end{eqnarray}
in which $\alpha,\beta,i=\{+,-\}$, $H_e$ is defined as the effective Hamiltonian of the system and $E_k$ ($k=\{+,-\}$) is the energy corresponding to the state $|k\rangle_e$. By simple calculations, we find $\langle+|H_{e}|+\rangle_e=g_{e}I^{z}_{1}I^{z}_{2}$ in which $|i\rangle_e=|-\rangle_e$ is the channel state of virtual electron spin flip, when $\langle-|H_{e}|-\rangle_e=- \langle+|H_{e}|+\rangle_e=g_{e}I^{z}_{1}I^{z}_{2}$ with the channel state $|i\rangle_e=|+\rangle_e$, here $g_{e}\approx\frac{a_{\parallel 1}a_{\parallel 2}}{2\Omega}$. Thus we adiabatically eliminate the fast electronic degrees of freedom from the slow nuclear dynamics by a Schrieffer-Wolff (SW) transformation, and derive the effective Hamiltonian
\begin{eqnarray}\label{He}
H_{e}\approx\sum_{i=1,2}\Omega_{rfi}I^{x}_{i}+g_{e}I^{z}_{1}I^{z}_{2}\otimes(|+\rangle_e\langle+|-|-\rangle_e\langle-|).
\end{eqnarray}
To investigate the dynamics of the whole system described by density matrix $\rho$, the system is governed by the master equation
\begin{eqnarray}
\label{masterT}
\frac{d\rho}{dt}=-i[H'_{t},\rho]+ L_e\rho L_e^{\dagger}-\frac{1}{2}(L_eL_e^{\dagger} \rho+\rho L_eL_e^{\dagger})
\end{eqnarray}
in which $L_e=\sqrt{\gamma_e}|-\rangle_e\langle+|$ with $\gamma_{e}=1/T_{1\rho}$. We apply this method and evaluate the performance of the gate between the first spin (carbon-13) with $\gamma_{n_1}B=(2\pi)4$ MHz and the second spin (silicon-29) $\gamma_{n_2}B=(2\pi)5.06$ MHz coupled to an NV center with $[a_{\parallel 1},a_{\parallel 2}]=(2\pi)[9,11]$ kHz, [see Fig. \ref{NoReset}(a)]. Here the driving fields are given as RF Rabi frequencies $\Omega_{rf1}=\Omega_{rf2}=(2\pi)1$ kHz, and the MW Rabi frequency $\Omega=(2\pi)400$ kHz. The NV center is initialized to state $|-\rangle_{e}$, when the nuclear spins are initialized to $|+_{1}-_{2}\rangle$ with $|+_{i}\rangle=\frac{1}{2}(|\uparrow_{i}\rangle+|\downarrow_{i}\rangle)$. Perfect state transfer between nuclear spins silicon-29 and carbon-13 $|+_{1}-_{2}\rangle\rightarrow|-_{1}+_{2}\rangle$ is shown in Fig. \ref{NoReset}(b), in which the population evolutions by using effective second order Hamiltonian Eq. (\ref{He}) fit well with the exact total Hamiltonian of the system Eq. (\ref{H't}). The effective Hamiltonian Eq. (\ref{He}) includes a three-body interaction and there are two channels that can mediate internuclear interaction via virtual electron spin flips through the NV microwave dressed states $|+\rangle_{e}$ and $|-\rangle_{e}$. However, the two channels can be mixed up if the nuclear-nuclear interaction time is longer than the NV relaxation time. As shown in Fig. \ref{NoReset}(c), in the case when the life time of NV center $T_{1\rho}=200$ $\mu s$, there is no coherent evolution of the nuclear spins. Therefore, the system is limited by the NV relaxation at room temperature.
\begin{figure*}[t]
\center
\includegraphics[width=6.6in]{Operation.pdf}
\caption{Nuclear-nuclear quantum gate of two nuclear spins mediated by an NV center in diamond. The basic logic-operations contain the initialization, gate operation based on second-order coupling and readout of nuclear spins. The electron spin works as a quantum bus and is stabilized in a quasi-steady state by periodically resetting in $|-\rangle_{e}$, which is shown in the yellow dashed box. Here P and M stand for the electron (nuclear) spin polarization, $Y_{\theta}$ presents operation with rotation angles $\theta$ about the $Y$ axes and $N$ is the total number of resets.}
\label{Operation}
\end{figure*}
\section{The effect of electron spin resets}
In our scheme, in order to overcome the limitation of NV relaxation, we intend to periodical reset the electron spin to state $|-\rangle_{e}$ within its relaxation time $T_{1\rho}$. It is reasonable to consider the steady-state polarization of the NV spin in a quasi-steady state as
\begin{eqnarray}
\langle2\sigma_z\rangle=e^{-t_{re}/T_{1\rho}}=p.
\end{eqnarray}
The three-body interaction in the effective Hamiltonian Eq. (\ref{He}) can be estimated as $g_{e}I^{z}_{1}I^{z}_{2}\otimes(|+\rangle_e\langle+|-|-\rangle_e\langle-|)\approx g_{e}I^{z}_{1}I^{z}_{2}\langle2\sigma_z\rangle=pg_{e}I^{z}_{1}I^{z}_{2}$. Therefore, the NV spin is not involved in the nuclear-nuclear interaction Hamiltonian. Resets also bring another dissipation item in the master equation and it is reasonable to have the effective relaxation of the NV spin as $\gamma_{r}=1/T_{1\rho}+1/t_{re}$. Thus, we can rewrite the master equation of the system as
\begin{eqnarray}
\label{master}
\frac{d\rho}{dt}=-i[H'_{t},\rho]+L_r \rho L_r^{\dagger}-\frac{1}{2}(L_r ^{\dagger}L_r \rho+\rho L_r ^{\dagger}L_r)
\end{eqnarray}
in which $L_{r}=\sqrt{\gamma_r}|-\rangle_e\langle+|$. The non-Hermitian Hamiltonian of the quantum jump formalism \cite{Reiter2012EffectiveOF} of the NV spin is given by
\begin{eqnarray}
H_H=\Omega\sigma_z-\frac{1}{2}L_r^{\dagger}L_r
\end{eqnarray}
According to the second expansion of SW transformation, we have energy difference between $|+\rangle_e$ and $|-\rangle_e$ as a complex energy $E_+-E_-=\Omega-\frac{i}{2}\gamma_r$, one can derive effective Hamiltonian
\begin{eqnarray}
H'_{e}\approx\sum_{i}\Omega_{rfi}I^{x}_{i}+g'_{e}I^{z}_{1}I^{z}_{2}\otimes(|+\rangle_e\langle+|-|-\rangle_e\langle-|),\\
L_n\approx\sum_{k}\frac{\sqrt{\gamma_r}a_{\parallel k}}{\Omega-\frac{i}{2}\gamma_r}I_k^z\otimes(|+\rangle_{e}\langle+|-|-\rangle_{e}\langle-|),
\end{eqnarray}
in which the effective dissipation item is given by $\langle \alpha |L_e|\beta\rangle=\sum_{i}\frac{\langle\alpha| L_r|i\rangle\langle i|V|\beta\rangle}{E_\alpha-E_i}$, and the effective coupling is $$g'_e\approx\frac{\Omega a_{\parallel 1}a_{\parallel 2}}{2(\Omega^2+\frac{\gamma_r^2}{4})}.$$ By considering the electron spin is periodically reset in state $|-\rangle_e$ with the definition $\langle 2\sigma_z\rangle=p$, we have the effective master equation of the nuclear spins as
\begin{eqnarray}\label{masterN}
\frac{d}{dt}\rho_n&=&-i[H_{N},\rho_{n}]+L_{N} \rho_n L_{N}^{\dagger} \notag \\
& &-\frac{1}{2}(L_{N} ^{\dagger}L_{N} \rho_n+\rho_n L_{N} ^{\dagger}L_{N}),
\end{eqnarray}
with the effective Hamiltonian and Lindblad operators
\begin{eqnarray}\label{HLN}
H_{N}&\approx&\sum_{i}\Omega_{rfi}I^{x}_{i}+pg'_{e}I^{z}_{1}I^{z}_{2},\\
L_{N}&\approx&\sum_{k}\frac{\sqrt{p\gamma_r}a_{\parallel k}}{\Omega-\frac{i}{2}\gamma_r}I_z^k.
\end{eqnarray}
The effective dissipation rate is given by
$$\gamma_N=\sum_{i,j}\frac{p\gamma_ra_{\parallel i}a_{\parallel j}}{\Omega^2+\frac{\gamma_r^2}{4}}.$$ Therefore, in principle, if resonant condition $\Omega_{rf1}=\Omega_{rf2}$ and $pg'_{e}\gg\gamma_N$ are satisfied, perfect coherent state transition between the nuclear spins can be possible.
\begin{figure}
\center
\includegraphics[width=3.1in]{infidelity.pdf}
\caption{Numerical simulations are show with the same parameters in the Fig. \ref{NoReset}, i.e., parallel couplings are given as $[a_{\parallel 1},a_{\parallel 2}]=(2\pi)[9,11]$ kHz. (a) The detection bandwidth of the swept RF frequency based on different MW Rabi frequencies. The RF field is applied to the carbon-13 sensor resonantly when the frequency of the other RF field is swept. The evolution time is fixed as $T=[8.8,4.4]$ ms for $\Omega=(2\pi)[400,200]$ kHz. (b) The infidelity of quantum gate between carbon-13 and silicon-29 spins when there is another silicon-29 spin. The infidelity varies with the detuning $\delta_{3}$, $\delta_{3}=(a_{\parallel 3}-a_{\parallel 2})/2$. }
\label{Robust}
\end{figure}
For comparison, it is necessary to employ exact numerical simulations as well to show the efficiency of our theory. Exact state evolutions of the system are simulated as follows. The NV center is reinitialized periodically to the state $|-\rangle_e$ of the dressed state basis, namely, $\rho(Nt_{re})\rightarrow[\Tr_{e}\rho(Nt_{re})]\otimes|-\rangle_e\langle-|$, where $\Tr_{e}$ denotes the partial trace over the electron spin and N is an integer and reset of the NV spin every $t_{re}$ introduces an effective interaction time in each cycle. The density matrix of the system evolves according to
\begin{eqnarray}\label{Evolution}
\rho_{n}\rightarrow\cdot\cdot\cdot U_{t}\Tr_{e}[U_{t}(\rho_{n}\otimes|-\rangle_e\langle-|)U^{\dagger}_{t}]\otimes|-\rangle_e\langle-|U^{\dagger}_{t},
\end{eqnarray}
in which $U_{t}$ is the time evolution operator according master equation Eq. (\ref{master}). Both of the state evolutions by using exact numerical simulation and effective second-order master equation based on Eq. (\ref{Evolution}) and Eq. (\ref{masterN}) are shown in Fig. \ref{NoReset}(d), and the importance for periodic reinitializing of electron spin is illustrated.
We consider the same case of a silicon-29 and a carbon-13 spin coupled to an NV spin which is presented in Section II except that the NV center is reinitialized to state $|-\rangle_{e}$ every $t_{re}=20$ $\mu s$. The resonant condition is matched as $\Omega_{rf1}=\Omega_{rf2}=(2\pi)1$ kHz, both of the simulations of exact numerical calculations and effective master equation show perfect state transfer between nuclear spins silicon-29 and carbon-13 $|+_{1}-_{2}\rangle\rightarrow|-_{1}+_{2}\rangle$, [see Fig. \ref{NoReset}(d)]. The state evolutions based on effective master equation are slightly different from the exact simulations. The possible reason is high orders of SW transformation expansions are not included.
In the absence of periodic reinitialization of the electron spin, channel mixing up leads to no coherent evolution of the nuclear spins [see Fig. \ref{NoReset}(c)], while periodic reinitialization of the NV center to state $|-\rangle_{e}$ provides a priority channel during the operations and coherent evolution between the nuclear spins is possible to extend beyond the life time of the NV spin. High fidelity of the nuclear-nuclear quantum gate in one-step is possible ($>$0.99), which could be significantly higher than the fidelity ($<$0.66) of nuclear-nuclear gates achieved so far with NV centers by using 4 electron-nuclear spin quantum gates (if each fidelity $<$0.90 \cite{Taminiau2014UniversalCA}). Additionally, because of the second-order coupling, the scheme is not sensitive to the detunings from the resonance and Rabi frequency errors of the MW driving of the NV center [see Fig. \ref{NoReset}(e)]. The process fidelity of two nuclear spins is given by the overlap of the states corresponding to the implemented evolution $|\psi\rangle=U^{zz}\otimes\sum_{i,j}|i_1\rangle|j_2\rangle/4$ (i,j=+,-) to the target states.
\section{Quantum gate implementation}
In the Fig. \ref{Operation}, we include all the related operations of the nuclear ZZ gate implementation for experiments. The initialization consists of the electron (nuclear) spin polarization P (M). The electron spin of NV center can be optically initialized and read out by using laser illumination. Here, P is obtained by the optical pumping cycle available for NV centers \cite{Jelezko2002SingleSS,Jelezko2004ObservationOC}, whereas M is based on the techniques developed for the nuclear single-shot measurement \cite{Neumann2010SingleShotRO}, followed by the electron state-dependent fluorescence \cite{Jelezko2002SingleSS,Jelezko2004ObservationOC}. Gate operations includes 3 steps: (i) Polarize the carbon-13 and silicon-29 nuclear spins by using the NV center, which leads to the initial state $|-\rangle_{e}\otimes|+_{1}-_{2}\rangle$ of the system. (ii) Implement the gate operation, which is governed by the effective master equation of Eq. (\ref{masterN}), $U^{ZZ}(t)=\exp(-iH_{N}t)$. In this process, we periodically reset the NV center and we control the heteronuclear spins via RF fields individually. (iii) Use the NV center to read out the states of nuclear spins. Here we suppose that the carbon-13 and silicon-29 spin can be well controlled by a SWAP gate between NV and nuclear spin; i.e., it can be polarized and its position can be detected precisely.
Assuming that the RF field is applied to the carbon-13 spin resonantly and $\delta_2=\gamma_{n2}B+\frac{a_{\parallel2}}{2}-\omega_{rf2}$ is a detuning from the resonance of RF driving to the silicon-29 spin, the effective Hamiltonian is given by
\begin{eqnarray}\label{Detuning}
H'_{N}\approx\bar{\Omega}_{rf2}I^{x'}_{2}+\Omega_{rf1}I^{x}_{1}+pg'_{e}I^{z}_{1}I^{z}_{2},
\end{eqnarray}
in which $\bar{\Omega}_{rf2}=\sqrt{\delta_2^2+\Omega_{rf2}^2}$. The initial state of the system is $|+_1-_2\rangle$, the population ($P_+$) of the initial state $|+_1\rangle$ of carbon-13 spin is given by
\begin{eqnarray}\label{P}
P_+=1-\frac{(pg'_{e}\cos\theta)^{2}\sin^{2}[t
\sqrt{(pg'_{e}\cos\theta)^{2}+\Delta^{2}}/2]}{(pg'_{e}\cos\theta)^{2}+\Delta^{2}},
\end{eqnarray}
$\cos\theta=\frac{\Omega_{rf2}}{\sqrt{\delta_2^2+\Omega_{rf2}^2}}$ and $\Delta=\bar{\Omega}_{rf2}-\Omega_{rf1}$. The frequency of the RF driving to silicon-29 spin is swept to show the detection bandwidth of our method [see Fig. \ref{Robust}]. The perfect population transfer could be possible if $|\bar{\Omega}_{rf2}-\Omega_{rf1}|\ll pg'_{e}\cos\theta$. Namely, the dip position with $\omega_{rf2}=\gamma_{n2}B+\frac{a_{\parallel2}}{2}=(2\pi)5065.5$kHz indicates $\delta_2=0$ and $\Omega_{rf2}=\Omega_{rf1}$. The dip height depends on the longitudinal dipolar coupling with $P_+=\cos^{2}(pg'_{e}T/2)$ with $T$ the total evolution time. Increasing the Rabi frequency of the MW driving induces a smaller effective coupling between the nuclear spins, decreasing the frequency bandwidth [see Fig. \ref{Robust}], which also gives a tunable frequency filter. We investigate the case when there is a third silicon-29 nuclear spin with coupling $a_{\parallel_3}$ to NV center and show the efficiency of the selectivity of our scheme. In Fig. \ref{Robust}(b), we initialize the nuclei in state $|+_{1}-_{2}-_{3}\rangle$. We calculate the infidelity of nuclear ZZ gate between two nuclear spins affected by the third silicon-29 spin. The first carbon-13 and second silicon spins are coupled to an NV center with $[a_{\parallel 1},a_{\parallel 2}]=(2\pi)[11,9]$ kHz. The coupling of the third spin matches $\delta_{3}=(a_{\parallel 2}-a_{\parallel 3})/2>(2\pi)0.5$ kHz gives high fidelity of the gate operation of target nuclear spins ($>0.95$). Thus, our method can address two heteronuclear spins by using the RF fields individually, and implement a near perfect quantum gate at room temperature with high selectivity.
\section{Sensing application}
Another important application of our scheme is to detect nuclear spins in another species outside of the diamond. For example, one can use a carbon-13 spin which is well controlled by the NV center to detect hydrogen-1 spins outside of the diamond. Thus carbon-13 spin is initialized in state $|+_{1}\rangle$ and hydrogen-1 spins are in maximally mixed states, one can detect the signal by measuring the probability of that the sensor carbon-13 spin remains in the state $|+_{1}\rangle$. The related operations are shown in Fig. \ref{Operation}, except of the initialization and readout of hydrogen-1 spins. To evaluate the effectiveness of our sensing application, we initialize the carbon-13 spin to $|+_{1}\rangle$ state to detect the other three nuclear (hydrogen-1) spins, see Fig. \ref{Third}. Based on Eq. (\ref{P}), we can adjust the parallel components and control frequency of RF to match the resonance condition for target (hydrogen-1) nuclear spins. Once the resonance condition is achieved, the signal dips mark the presence of the nuclear spins. Similar to the case in the previous scheme \cite{Chen2017DissipativelySQ}, the frequency resolution is not limited by the NV relaxation $T_{1\rho}$, it requires $|\bar{\Omega}_{rfi}-\Omega_{rf1}|>pg'_{e}\cos\theta$ which is limited by the target spin decoherence time $T_2=200$ ms, the sensitivity per unit time of our scheme is proportional to $pg'_{e}/\sqrt{1/\Gamma_N}\sim(a_{\parallel i}/4)/\sqrt{T_{1\rho}}.$
\begin{figure}
\center
\includegraphics[width=3.0in]{sensing.pdf}
\caption{Assuming a carbon-13 and three hydrogen-1 spins coupled to the NV center with parallel couplings $[a_{\parallel 1},a_{\parallel 2},a_{\parallel 3},a_{\parallel 4}]=(2\pi)[11,4,9,11]$ kHz. The carbon-13 spin is the sensor with the initial state $|+_{1}\rangle$ and the other three hydrogen-1 are the targets which are in maximum mixed states. The population of the initial state $|+_{1}\rangle$ of the carbon-13 sensor is calculated by exact simulation of Eq. (\ref{Evolution}) with evolution time $T=8$ ms for $\Omega=(2\pi)400$ kHz and the target spins decoherence time $T_2=200$ ms. }
\label{Third}
\end{figure}
\section{Conclusion}
Before summarizing, we would like to compare our new method with the previous scheme \cite{Chen2017DissipativelySQ}. Both of schemes realize the effective coupling of target nuclear spins by periodical resetting of electron spin, and the impact of NV decoherence and relaxation processes on these nuclei are suppressed. The difference mainly comes from that we employ two weak RF fields to control the heteronuclear spins individually. In the previous approach, nuclear spins in the same species are considered, coarse tuning of the direction of the magnetic field in advance is necessary for matching the resonant condition. Our method allows for effective interaction of nuclear spins in different species which loosen the requirement of the stringent resonant condition of the previous scheme. The signal accumulation and resonant condition depend on longitudinal dipolar coupling component $a_{\parallel_i}$, which leads to no information of the transverse coupling component $a_{\perp_i}$ to be detected in this new scheme. Therefore, less information is detected when one uses this new method for sensing nuclear spins near the NV center. But it also bring advantage that the quantum gate implementation is not limited by the transverse coupling component.
In this work we presented an extension of dissipatively stabilized NV center to a case, in which one achieves a tunable second-order effective coupling between distant nuclear spins in different species. We employ RF fields and MW driving to match the resonant condition, and control electron and nuclear spins individually. The coupling is mediated by a dissipatively decoupled electron spin of a NV center, when the impact of NV decoherence and relaxation processes on these nuclei are suppressed. Thus the effective indirect interaction enables selectively initialization and coherent control of nuclear spins as well as analysis of complex spin structures at ambient condition.
\emph{Acknowledgements ---} The authors thank Natural Science Foundation of Hunan Province, China (2019JJ10002), Hunan Provincial Hundred People Plan (2019), Huxiang High-level Talent Gathering Project (2019RS1043).
|
\section{ Introduction}
\subsection{The model}
An increasing interest in studying the kinetic-fluid models, which describe the evolution of dispersed particles in the fluid, is stimulated by extensive applications, for example, in the modeling of reaction flows of sprays, sedimentation phenomenon, wastewater treatment, sedimentation-consolidation process and rainfall formation, see \cite{BBPS,BBKT,BBE,BWC,FFS,S,SG}.
In this paper, we consider the coupling of the Vlasov-Poisson-Fokker-Planck equations and compressible Navier-Stokes system in three dimensions:
\begin{equation}\label{VNS}
\left\{
\begin{aligned}
&\partial_tf+v\cdot\nabla_xf+{\rm{div}}_v((u-v)f)-\nabla_x\Phi\cdot\nabla_vf-\Delta_vf=0,\\
&-\Delta_x\Phi=\int_{\mathbb{R}^3}f\,dv-c(x),\\
&\partial_t\rho+{\rm{div}}_x(\rho u)=0,\\
&\partial_t(\rho u)+{\rm{div}}_x(\rho u\otimes u)+\nabla_x\rho^\gamma
-{\rm{div}}_x\mathbb{S}(\nabla u)=-\int_{\mathbb{R}^3}(u-v)f\,dv,
\end{aligned}
\right.
\end{equation}
where $f = f(t, x, v)\geq 0$ stands for the distribution of particles at time $t>0$ with respect to position $x$
and velocity $v$,
$\Phi(t,x)$ is the internal electric potential of the kinetic system,
and $\rho(t,x)$ and $u(t,x)$ represent the density and velocity of the fluid, respectively.
$\gamma>3/2$ is the adiabatic coefficient,
$c(x)$ is a given background function,
and $\mathbb{S}$ is given by
$$\mathbb{S}(\nabla u)=\mu_1(\nabla_x u+\nabla_x^\top u)+\mu_2{\rm{div}}u\,\mathbb{I},$$
where $\mu_1 $ and $\mu_2$ are coefficients of viscosity satisfying $\mu_1 > 0$ and $2\mu_1 + 3\mu_2 \geq 0$,
and $\mathbb{I}$ is the $3\times3$ identity matrix.
The kinetic equation in \eqref{VNS} describes an interacting large system with $M$ particles whose evolution is driven by the velocity $u$ of the compressible fluid system, which can be understood in the following microscopic mean field structure
\begin{align*}
dX_{M,i}(t)&=V_{M,i}(t)dt, \quad 1\leq i\leq M,\\
dV_{M,i}(t)&=\Big(-u(X_{M,i}(t))+V_{M,i}(t)-\nabla U*c(X_{M,i}(t))+\frac{1}{M}\sum_{j\neq i}^{M} \nabla U(X_{M,i}(t)-X_{M,j}(t))\Big)dt + dW_t^i,
\end{align*}
where $(X_{M,i}(t),V_{M,i}(t))_{1\leq i\leq M}$ are the positions and velocities of a group of $M$ particles, $(W^i_t)_{1\leq i\leq N}$ is a set of independent Brownian motions, and $U$ is the Newtonian potential in three dimensions.
In this particle formulation, the driven force $-u(X_{M,i}(t))+V_{M,i}(t)$ for the $i$-th particle is coupled with the fluid system through the fluid velocity $u$. The rigorous derivation of Vlasov-Poisson-Fokker-Planck through mean field limit was recently shown in \cite{CCS, HLP}. In the last decades, there have been many notable works in the derivation of one particle effective equation through the mean field limit of first (or second) order particle systems (in both deterministic or stochastic setting), for example, \cite{CCH,CDJ,CDHJ,CHJZ,G,J,JW,LP}, to name a few. However, due to the singular coupling of the particle system on the right hand side of the momentum equation, a rigorous derivation of the whole Vlasov-Poisson-Fokker-Planck-Navier-Stokes system is still missing.
We point out that a similar incompressible version of the system \eqref{VNS} was introduced in \cite{AIS14} by Anoshchenko, Iegorov, and Khruslov. And the existence of global weak solutions to it with homogeneous Dirichlet boundary conditions was obtained by mainly using the modified Galerkin method and Schauder's fixed point
theorem.
The main goal of this paper is to prove existence of weak solutions to the system \eqref{VNS} with given initial and boundary conditions. More precisely, the setting of the problem is the following. Let $\Omega\subset\mathbb{R}^3$ be a given bounded domain with smooth boundary $\partial \Omega$.
We impose the following initial conditions
\begin{equation}\label{1}
\big(f(0,x,v),\rho(0,x),u(0,x) \big)=\big(f_0(x,v),\rho_0(x),u_0(x) \big).
\end{equation}
For the fluid equations \eqref{VNS}$_3$-\eqref{VNS}$_4$ we consider the nonhomogeneous Dirichlet boundary conditions:
\begin{align}
&\rho(t,x)=\rho_B(x),\quad (t,x)\in (0,T)\times\Gamma_{\rm{in}},\\
&u(t,x)=u_B(x),\quad (t,x)\in (0,T)\times\partial\Omega,
\end{align}
and for the kinetic equation\eqref{VNS}$_1$ and the Poisson equation \eqref{VNS}$_2$, we consider the following inflow boundary conditions:
\begin{align}
&\gamma^- f(t,x,v)=g(t,x,v),\quad (t,x,v)\in (0,T)\times \Sigma^-,\\
&\Phi(t,x)=0,\quad (t,x)\in (0,T)\times\partial\Omega,\label{2}
\end{align}
where $\Gamma_{\rm{in}}:=\{x\in\partial\Omega\,\big|\,u_B\cdot \nu(x)<0\}$ and $\gamma ^{-}f(t,x,v)$ is the trace of $f$, $\Sigma^{-}:=\{(x,v)\in\partial\Omega\times\mathbb{R}^3| v\cdot \nu(x)<0\}$.
Here $\nu(x)$ denotes the outward unit normal vector to $x\in\partial\Omega$.
Furthermore, we use $\gamma^+f$ to denote the trace of $f$ on $(0,T)\times\Sigma^+$ and
$\Sigma^{+}:=\{(x,v)\in\partial\Omega\times\mathbb{R}^3| v\cdot \nu(x)>0\}$.
\subsection{Some previous results}
Kinetic-fluid models have been studied in different contexts depending on, for example, the compressibility of the fluid, the types of friction force and the magnitude of initial data.
In the following, we summarize some results in two categories: compressible kinetic-fluid model and incompressible kinetic-fluid model.
We will restrict ourselves to the kinetic-fluid model with non-density dependent viscosity.
The typical compressible Vlasov-Fokker-Planck-Navier-Stokes system reads:
\begin{equation}\label{s1}
\left\{
\begin{aligned}
&\partial_tf+v\cdot\nabla_xf+{\rm{div}}_v((u-v)f)-\Delta_vf=0,\\
&\partial_t\rho+{\rm{div}}_x(\rho u)=0,\\
&\partial_t(\rho u)+{\rm{div}}_x(\rho u\otimes u)+\nabla_x\rho^\gamma- \Delta_xu=-\int_{\mathbb{R}^3}(u-v)f\,dv.
\end{aligned}
\right.
\end{equation}
Chae, Kang and Lee \cite{CKL'} showed that the system \eqref{s1} possesses a global classical solution close to an equilibrium
\big(\bar f, \bar \rho, \bar u\big)=\big(e^{- {|v|^2}/{2}}, 1,0\big)
$ and proved that the solution converges to the equilibrium exponentially fast in $\mathbb{T}^3$.
Mellet and Vasseur \cite{MV} established the global existence of weak solutions to the system \eqref{s1} for
large initial data, furthermore, they \cite{MV'} studied the asymptotic analysis of the solutions.
In the case of nonhomogeneous Dirichlet boundary condition for the velocity of the fluid, the global existence of weak solutions to the system \eqref{s1} was given by the second author \cite{Li}.
Moreover, this result was extended to the system \eqref{s1} with a local alignment force in \cite{LL2}.
For the case that the friction force is dependent on the density $\rho$, Li, Mu and Wang \cite{LMW} established the global well-posedness of a strong solution when the initial data is a small perturbation of some given equilibrium.
Now we focus on the incompressible kinetic-fluid system.
Chae, Kang and Lee \cite{CKL} derived the global existence of weak and classical solutions to the Vlasov-Fokker-Planck-Navier-Stokes system.
Boudin, Desvillettes, Grandmont and Moussa \cite{BDGM} proved that the Vlasov-Navier-Stokes system possesses global weak solutions in three-dimensional
periodic case.
Furthermore, a similar result in three-dimensional bounded domain was obtained in \cite{Yu}.
When the friction force depends on the density,
Wang and Yu \cite{WY} derived the global well-posedness of weak solutions to the Vlasov-Navier-Stokes equations.
Anoshchenko, Khruslov and Stephan \cite{AKS} proved that the Vlasov-Poisson-Navier-Stokes system possesses global weak solution.
Choi and Jung \cite{CJ} studied the hydrodynamic limit of an incompressible Vlasov-Poisson-Fokker-Planck-Navier-Stokes system in a bounded domain by the relative entropy method.
Note that when taking the distribution function $f$ equal to zero, then the system \eqref{VNS} becomes the classical compressible isentropic Navier-Stokes equations. Lions \cite{L} proved the existence of global weak solutions of the three-dimensional Navier-Stokes equations for the adiabatic constant $\gamma\geq 9/5$ with the boundary condition $u|_{\partial\Omega}=0$. Later, Feireisl, Novotn\'{y} and Petzeltov\'{a} \cite{FNP} improved the above result to $\gamma>3/2$. In \cite{JZ}, Jiang and Zhang established the global existence of axisymmetric weak solutions for $\gamma>1$ with axisymmetric initial data. Bresch and Jabin \cite{BJ} proved the global existence of appropriate weak solutions to compressible Navier-Stokes equations with general viscous stress tensor. In the case of nonhomogeneous Dirichlet boundary condition, the global existence of weak solutions was given by Plotnikov and Sokolowski \cite{PS} by means of Young measure and other tools. They first showed the Navier-Stokes equations possess weak solution when $\gamma$ is big enough, then this result was extended to the case of $\gamma>3/2$ with the help of kinetic theory. Further, Chang, Jin and Novotn\'{y} \cite{CJN} gave an other proof of this result thanks to the effective viscous flux identity, oscillations defect measure and renormalization techniques for the continuity equation in the spirt of \cite{L,FNP}.
\subsection{Our results}
Weak solutions to the problem \eqref{VNS}-\eqref{2} is defined as follows:
\begin{defn}\label{defn}
Let $T>0$ arbitrary. The function $(f,\Phi, \rho, u) : (0,T)\times\Omega \to \mathbb{R}_+\times\mathbb{R}\times\mathbb{R}_+\times\mathbb{R}^3$
is called a {\em weak solution} to the problem \eqref{VNS}-\eqref{2} on $[0,T]$ if
it possesses the regularity
\begin{gather*}
f\in L^\infty(0,T;L^1\cap L^\infty(\Omega\times\mathbb{R}^3)),\quad |v|^2f\in L^\infty(0,T;L^1(\Omega\times\mathbb{R}^3)),\\
\Phi\in L^\infty\big(0,T; W^{2,\frac{5}{3}}(\Omega)\big),\quad
\rho\in L^\infty(0,T;L^\gamma(\Omega)),\\
u\in L^2(0,T;H^1(\Omega)),\quad
\rho u\in L^\infty\big(0,T;L^\frac{2\gamma}{\gamma+1}(\Omega)\big),
\end{gather*}
and fulfills the following relations:\medskip\\
1.~{\em Weak formulation of the Vlasov-Fokker-Planck equation:} for any $\varphi\in C^\infty_c([0,T)\times\bar\Omega\times\mathbb{R}^3)$ such that $\varphi=0$ on $(0,T)\times\Sigma^+$, it holds
\begin{align*}
&\int^{T}_0\int_{\Omega\times\mathbb{R}^3}f\big(\partial_t\varphi+v\cdot\nabla_x\varphi+(u-v)\cdot\nabla_v\varphi
-\nabla_x\Phi\cdot\nabla_v\varphi
+\Delta_v\varphi\big)\,dxdvdt\\
&\qquad+\int_{\Omega\times\mathbb{R}^3}f_0\varphi(0,x,v)\,dxdv
=\int_0^T\int_{\Sigma^-}(v\cdot \nu(x))g\varphi\,d\sigma(x)dvdt.
\end{align*}
2.~{\em Weak formulation of the Poisson equation:} for any $\Psi\in C^\infty_c((0,T)\times\Omega)$, it holds
\begin{align*}
\int_0^T\int_\Omega\nabla_x\Phi\cdot\nabla_x\Psi\,dxdt=
\int_0^T\int_\Omega (n-c(x))\Psi\,dxdt,
\end{align*}
where $n:=\int_{\mathbb{R}^3}f\,dv$;\\
3.~{\em Weak formulation of the continuity equation:} for any $\psi\in C^\infty_c([0,T)\times(\Omega\cup\Gamma_{\rm{in}}))$, it holds
\begin{align*}
\int^{T}_0\int_{\Omega}(\rho\partial_t\psi+\rho u\cdot\nabla_x\psi)\,dxdt+\int_{\Omega}\rho_0\psi(0,x)\,dx=
\int_0^T\int_{\Gamma_{\rm{in}}}\rho_B u_B\cdot \nu(x)\varphi\,d\sigma(x)dt;
\end{align*}
4.~{\em Weak formulation of the momentum balance equation:} for any $\phi\in C^\infty_c((0,T)\times\Omega;\mathbb{R}^3)$, it holds
\begin{align*}
\int^{T}_{0}\int_{\Omega}\big(&\rho u\cdot\partial_t\phi+(\rho u\otimes u):\nabla_x\phi+\rho^\gamma{\rm{div}}_x\phi
-\mathbb{S}(\nabla_x u):\nabla_x\phi\\
&+(j-nu)\cdot\phi\big)\,dxdt
+\int_{\Omega}\rho_0u_0\cdot \phi(0,x)\,dx=0,
\end{align*}
where $j:=\int_{\mathbb{R}^3}vf\,dv$;\\
5.~{\em Energy balance inequality:}
for any $\tau\in (0,T)$,
\begin{align}\label{energy main}
\int_{\Omega}&\Big(\frac{1}{2}\rho|u-u_{\infty}|^2+\frac{1}{\gamma-1}\rho^{\gamma}
+\frac{1}{2}|\nabla\Phi|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f\,dv\Big)(\tau)\,dx\nonumber\\
&+\int^{\tau}_0\int_{\Omega}\mathbb{S}(\nabla(u-u_\infty)):\nabla(u-u_\infty)\,dxdt
+\int^{\tau}_0\int_{\Sigma^-}(v\cdot \nu(x))\frac{|v|^2}{2}g\,d\sigma(x)dvdt\nonumber\\
&+\int_0^\tau\int_{\Gamma_{\rm{in}}}\rho^\gamma|u_B\cdot \nu(x)|\,d\sigma(x)dt
+\int_0^\tau\int_{\Gamma_{\rm{out}}}\frac{1}{\gamma-1}\rho^\gamma|u_B\cdot\nu(x)|\,d\sigma(x)dt\nonumber\\
\leq&\int_{\Omega}\Big(\frac{1}{2}\rho_0|u_0-u_{\infty}|^2+\frac{1}{\gamma-1}\rho_0^{\gamma}+\frac{1}{2}|\nabla\Phi_0|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_0\,dv\Big)\,dx\nonumber\\
&+\int^{\tau}_0\int_{\Omega}\big(-\rho^{\gamma}{\rm{div}}u_{\infty}-\mathbb{S}(\nabla u_{\infty}):\nabla(u-u_\infty)
-\rho u\cdot\nabla u_\infty\cdot(u-u_\infty)\big)\,dxdt\nonumber\\
&+\int^{\tau}_0\int_{\Gamma_{\rm{in}}}\frac{\gamma}{\gamma-1}\rho^{\gamma-1}\rho_{B}|u_B\cdot \nu(x)|\,d\sigma(x)dt
+3\int_{\Omega\times\mathbb{R}^3}f\,dxdv
-\int^{\tau}_0\int_{\Omega}(j-nu)\cdot u_\infty\,dxdt,
\end{align}
where $\Phi_0$ is determined by the equation $-\Delta\Phi_0=\int_{\mathbb{R}^3}f_0\,dv-c(x)$, and
$u_\infty(x)\in W^{1,\infty}(\Omega;\mathbb{R}^3)$ is an extension of $u_B(x)$, satisfying
\begin{gather}
{\rm{div}}u_\infty\geq 0\;\;\textrm{a.e.}\;\;{\rm{in}}\;\;U^-_h\equiv\{x\in\Omega\big|{\rm{dist}}(x,\partial\Omega)<h\},\;\;h>0 \;\;{\rm{is\,\,small}}.\label{extension2}
\end{gather}
\end{defn}
\begin{remark}
The existence of the extension $u_\infty$ of $u_B$ was given by Girinon \cite{Gir}.
\end{remark}
Now we state our main result.
\begin{theorem}\label{main}
Let $\Omega\subset\mathbb{R}^3$ be a bounded domain with smooth boundary, $c(x)\in L^p(\Omega)$ $(1\leq p<+\infty)$ be a given function and $\gamma >3/2$. Suppose that the boundary data
$u_B\in C^2(\partial\Omega;\,\mathbb{R}^3)$, $\rho_B\in C(\partial\Omega)$, and $\mathop{{\rm{min}}}\limits_{\Gamma_{\rm{in}}}\rho_B=\underline{\rho}_B>0$.
Assume that the initial and boundary data are of finite energy:
\begin{align}
&0\leq f_0\in L^1\cap L^\infty(\Omega\times\mathbb{R}^3),\;\;0\leq g\in L^1\cap L^\infty((0,T)\times\Sigma^-), \label{fv}\\
&\int_{\Omega}\rho_0\,dx>0,\;\;\;\;\int^T_0\int_{\Sigma^-}|v|^2g(t,x,v)|v\cdot \nu(x)|\,d\sigma(x)dvdt<\infty, \\
&\int_{\Omega}\Big(\frac{1}{2}\rho_0|u_0-u_\infty|^2+\frac{1}{\gamma-1}\rho_0^\gamma
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_0\,dv\Big)\,dx
<\infty,\label{mv}
\end{align}
then for any $T>0$, the problem \eqref{VNS}-\eqref{2} possesses at least one global weak solution $(f,\Phi,\rho,u)$ on $[0,T]$.
\end{theorem}
Compared with the weak existence result obtained in \cite{LL2}, where the
Vlasov-Fokker-Planck-Navier-Stokes equations with alignment force term is considered, one can clearly see that the main difficulty comes from the coupling with the Poisson equation \eqref{VNS}$_2$. In fact, the weak solution theory for Vlasov-Poisson-Fokker-Planck system has been established in a classical work \cite{C}. However, the $k$-th moment estimates for the iteration scheme can only be obtained with additional assumption on the solution. If we perform the same iteration scheme for the kinetic part of our system, due to the complicated structure of this system, we can not arrive at the similar moment estimates. Therefore, instead of using the method developed in \cite{C}, we construct a new approximation scheme to decouple the Vlasov-Poisson-Fokker-Planck system and apply a special case of Leray-Schauder fixed point theorem to complete a sequence of well-defined approximate solutions. Consequently, the global weak solution is obtained by entropy estimate and compactness argument.
Next we give a brief description of the method in handling with the kinetic equation. For a given $\Phi^*\in L^\infty(0,T;W^{1,\infty}(\Omega))$, we can construct a unique solution $f$ to the Vlasov-Fokker-Planck equation with the help of
a main result in \cite{C}. More precisely, there exists a unique weak solution to the following linear equation:
\begin{align*}
\partial_t f +v\cdot \nabla_x f+{\rm{div}}_v((\nabla\Phi^*+\tilde u\chi_{\{|\tilde u|\leq N\}}-v)f)-\Delta_vf=0,
\end{align*}
where $\chi$ is a cut-off function for the fluid velocity. For more detailed technics for the fluid part, we refer to Section \ref{app}.
The next thing to do is to find a unique solution $\Phi$ to the Poisson equation $-\Delta \Phi=\int_{\mathbb{R}^3} f dv-c(x)$.
However, a $W^{1,\infty}$ bound for $\Phi$ can not be obtained by strong solution theory. Therefore, in order to close the definition of the map, we need to introduce the regularization of the Poisson equation. Furthermore, since the moment estimates are only valid for fixed point of this map, instead of using the iteration scheme as in \cite{C}, we use a special type of Leray-Schauder fixed point theorem to establish the existence of $\nabla\Phi$.
More details can be found in Section \ref{app}.
The rest of this paper is organized as follows.
Section \ref{app} is devoted to constructing the approximate solutions to the problem \eqref{VNS}-\eqref{2}.
In Section \ref{l}, we pass to the limits on regularized parameters in turn and complete the proof of Theorem \ref{main}.
\section{Approximate solutions}\label{app}
Our first goal of this section is to solve the regularized system:
\begin{align}
&\partial_tf+v\cdot \nabla_xf+{\rm{div}}_v\big((u-v)f\big)-\nabla_x\Phi\cdot\nabla_vf-\Delta_vf=0,\label{3} \\
&-\Delta\Phi-\varepsilon\Delta^{2m+1}\Phi=\int_{\mathbb{R}^3}f\,dv-c(x),\label{4}\\
&\partial_t\rho+{\rm{div}}(\rho u)=\varepsilon\Delta\rho,\label{5} \\
&\partial_t(\rho u)+{\rm{div}}(\rho u\otimes u)+\nabla\rho^{\gamma}+\delta\nabla\rho^\beta
+\varepsilon\nabla\rho\cdot\nabla u
\nonumber\\
&\qquad\quad ={\rm{div}}\mathbb{S}(\nabla u)
+\varepsilon{\rm{div}}(|\nabla(u-u_\infty)|^2\nabla(u-u_\infty))
-\int_{\mathbb{R}^3}(u-v)f\,dv,\label{6}
\end{align}
where $m\in \mathbb{N}$ is fixed large as necessary, $\varepsilon>0$, $\delta>0$ and $\beta>{\rm{max}}\{\gamma,9/2\}$.
The system \eqref{3}-\eqref{6} is equipped with the following initial and boundary data:
\begin{align}\label{7}
\big(f(0,x,v),\rho(0,x),u(0,x)\big)=v(f_{0\varepsilon}(x,v),\rho_0(x),u_0(x)\big),
\end{align}
and
\begin{align}
&\gamma^-f(t,x,v)\big|_{(0,T)\times\Sigma^-}=g_\varepsilon(t,x,v),\label{8}\\
&\Phi|_{(0,T)\times\partial\Omega}=\Delta\Phi|_{(0,T)\times\partial\Omega}=\cdot \cdot \cdot
=\Delta^{2m}\Phi|_{(0,T)\times\partial\Omega}=0,\label{8'}\\
&(-\varepsilon\nabla\rho+\rho u)\cdot \nu(x)\big|_{(0,T)\times\partial\Omega}=
\left\{
\begin{array}{ll}
\rho_Bu_B\cdot \nu(x) & \;\;\;{\rm{on}}\;\;\;\Gamma_{\rm{in}},\\
\rho u_B\cdot \nu(x)&\;\;\;{\rm{on}}\;\;\;\partial\Omega\backslash\Gamma_{\rm{in}},
\end{array}
\right.\label{9}\\
&u(t,x)\big|_{(0,T)\times\partial \Omega}=u_B(x),\label{10}
\end{align}
where $f_{0\varepsilon}$ and $g_\varepsilon$
are approximate sequences of $f_0$ and $g$, respectively. They satisfy \eqref{fv}-\eqref{mv} uniformly with respect to $\varepsilon$ and
\begin{gather*}
\int_{\Omega\times\mathbb{R}^3}|v|^\kappa f_{0\varepsilon}(x,v)\,dxdv<\infty,\quad \forall\; \kappa\in[0,\kappa_0]\;\; {\rm{with}}\;\; \kappa_0\geq 5,\\
\int_0^T\int_{\Sigma^-}|v|^\kappa g_{\varepsilon}|v\cdot\nu(x)|\,d\sigma(x)dvdt<\infty,\quad \forall\; \kappa\in[0,\kappa_0]\;\; {\rm{with}}\;\; \kappa_0\geq 5.
\end{gather*}
Here the initial and boundary data verify the assumptions of Theorem \ref{main} and
\begin{equation}\label{11}
\left\{
\begin{aligned}
&u_0\in L^2(\Omega),\quad \rho_0\in W^{1,2}(\Omega),\\
&0<\underline{\rho}\leq\rho_0(x)\leq\bar{\rho}<\infty,\quad x\in\Omega,\\
&0<\underline{\rho}\leq\rho_B(x)\leq\bar{\rho}<\infty,\quad x\in\Gamma_{\rm{in}}.
\end{aligned}
\right.
\end{equation}
We still use $(\rho_0,u_0,\rho_B)$ to denote the initial and boundary data to the system \eqref{3}-\eqref{6} and will emphasize the dependence
of the parameter if it is needed.
In the rest of this section, we focus on the solvability of the problem \eqref{3}-\eqref{10}.
\subsection{Decoupled system}\label{2.1.1}
To establish the global existence of weak solutions to the problem \eqref{3}-\eqref{10}, Galerkin method is a classical one.
We introduce a finite dimensional space $X={\rm{span}}\{\Psi_i\}_{i=1}^N$, where the smooth functions
$\Psi_i(x)\,(1\leq i\leq N)$ are orthonormal in $L^2(\Omega)$.
We are now in a position to construct approximate solutions to the following system:
\begin{align}
&\partial_tf+v\cdot \nabla_xf+{\rm{div}}_v\big((\tilde u\chi_{\{|\tilde u|\leq N\}}-v)f \big)
-\nabla_x\Phi^*\cdot \nabla_v f-\Delta_vf=0,\label{12}\\
&f|_{t=0}=f_{0\varepsilon}(x,v),\quad f|_{(0,T)\times\Sigma^-}=g_\varepsilon(t,x,v),\label{v213}\\
&-\Delta\Phi-\varepsilon\Delta^{2m+1}\Phi=\int_{\mathbb{R}^3}f\,dv-c(x),\label{13}\\
&\Phi|_{(0,T)\times\partial\Omega}=\Delta\Phi|_{(0,T)\times\partial\Omega}=\cdot\cdot\cdot
=\Delta^{2m}\Phi|_{(0,T)\times\partial\Omega}=0,\label{14}\\
&\partial_t\rho+{\rm{div}}_x(\rho u)=\varepsilon\Delta_x\rho, \label{26}\\
&\rho|_{t=0}=\rho_0(x),\quad
(-\varepsilon\nabla\rho+\rho u)\cdot \nu(x)\big|_{(0,T)\times\partial\Omega}=
\left\{
\begin{array}{llr}
\rho_Bu_B\cdot \nu(x)& \;\;{\rm{on}}\;\;\Gamma_{\rm{in}},\\
\rho u_B\cdot \nu(x)&\;\;{\rm{on}}\;\;\partial\Omega\backslash\Gamma_{\rm{in}},
\end{array}
\right.\label{27}\\
&\partial_t(\rho u)+{\rm{div}}(\rho u\otimes u)+\nabla\rho^{\gamma}+\delta\nabla\rho^{\beta}+\varepsilon\nabla\rho\cdot\nabla u\nonumber\\
&\qquad ={\rm{div}}\mathbb{S}(\nabla u)+\varepsilon{\rm{div}}(|\nabla(u-u_\infty)|^{2}\nabla(u-u_\infty))
-\chi_{\{|\tilde{u}|\leq N\}}\int_{\mathbb{R}^3}(\tilde u-v)f\,dv, \label{28}\\
&u|_{t=0}=u_0(x),\quad u(t,x)\big|_{(0,T)\times\partial\Omega}=u_B(x),\label{29}
\end{align}
where $(\tilde{u},\nabla\Phi^*)\in L^2(0,T;L^2(\Omega))\times L^\infty(0,T;W^{4m,\frac{5}{3}}(\Omega))$ is a given pair and $\chi$ is continuous.
For this purpose, we commence by recalling a known result due to \cite{MV}.
\begin{lemma}\label{1'}
If one has
\begin{align*}
&\|f\|_{L^\infty((0,T)\times\Omega\times\mathbb{R}^3)}\leq M, \\
&\int_{\Omega\times \mathbb{R}^3}|v|^\kappa f(t,x,v)\,dxdv\leq M,\;\; t\in[0,T],\;\; \kappa\in[0,\kappa_0],
\end{align*}
where $\kappa_0>0$,
then there exists a constant $C(M)$ such that
\begin{align*}
&\Big\|\int_{\mathbb{R}^3}f(t,x,v)\,dv\Big\|_{L^p(\Omega)}\leq C(M), \quad p\in\Big[1,\frac{\kappa_0+3}{3}\Big],\\
&\Big\|\int_{\mathbb{R}^3}vf(t,x,v)\,dv\Big\|_{L^p(\Omega)}\leq C(M), \quad p\in\Big[1,\frac{\kappa_0+3}{4}\Big],
\end{align*}
for any $t\in[0,T]$.
\end{lemma}
It is a direct consequence of \cite{C} that there exists a unique solution $f\in L^\infty(0,T;L^1\cap L^\infty(\Omega\times\mathbb{R}^3))$
to the problem \eqref{12}-\eqref{v213}.
Moreover, the solution $f$ satisfies
\begin{align}
&\partial_t\int_{\mathbb{R}^3}f\,dv=-{\rm{div}}_x\int_{\mathbb{R}^3}vf\,dv,\label{v24}\\
&\|f\|_{L^\infty(0,T;L^p(\Omega\times\mathbb{R}^3))}
\leq e^{\frac{3T}{p'}}(\|f_{0\varepsilon}\|_{L^p(\Omega\times\mathbb{R}^3)}
+\|g_\varepsilon\|_{L^p((0,T)\times\Sigma^-)}),\quad 1\leq p\leq \infty,\label{v25}\\
&\|\gamma^+f\|_{L^p((0,T)\times\Sigma^+)}
\leq e^{\frac{3T}{p'}}(\|f_{0\varepsilon}\|_{L^p(\Omega\times\mathbb{R}^3)}
+\|g_\varepsilon\|_{L^p((0,T)\times\Sigma^-)}),\quad 1\leq p\leq \infty,\label{v26}
\end{align}
where $p'$ verifies $\frac{1}{p}+\frac{1}{p'}=1$, and for any $1\leq l<\infty$,
\begin{align}\label{v23}
\frac{d}{dt}\int_{\Omega\times\mathbb{R}^3}|v|^lf\,dxdv
&=-l\int_{\Omega\times\mathbb{R}^3}|v|^lf\,dxdv+l(l+1)\int_{\Omega\times\mathbb{R}^3}|v|^{l-2}f\,dxdv\nonumber\\
&-\int_{\Sigma^+}|v\cdot\nu(x)||v|^l\gamma^+f\,d\sigma(x)dv+\int_{\Sigma^-}|v\cdot\nu(x)||v|^lg_\varepsilon\,d\sigma(x)dv\nonumber\\
&+l\int_{\Omega\times\mathbb{R}^3}\tilde{u}\chi_{\{|\tilde{u}|\leq N \}}\cdot v|v|^{l-2}f\,dxdv
-l\int_{\Omega\times\mathbb{R}^3}\nabla\Phi^*\cdot v|v|^{l-2}f\,dxdv.
\end{align}
This together with the imbedding $W^{4m,\frac{5}{3}}(\Omega)\hookrightarrow L^\infty(\Omega)$ (m is big enough) ensure us to derive that
\begin{align*}
\big\||v|^lf\big\|_{L^\infty(0,T;L^1(\Omega\times\mathbb{R}^3))}\leq C\Big(N,\varepsilon,T,\|\nabla\Phi^*\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}\Big).
\end{align*}
With this observation at hand, and combining \eqref{v25} and Lemma \ref{1'}, we have
\begin{align}
&\Big\|\int_{\mathbb{R}^3}f\,dv \Big\|_{L^\infty\big(0,T;L^\frac{l+3}{3}(\Omega)\big)}\leq
C\Big(N,\varepsilon,T,\|\nabla\Phi^*\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}\Big),\label{v214}\\
&\Big\|\int_{\mathbb{R}^3}vf\,dv \Big\|_{L^\infty\big(0,T;L^\frac{l+3}{4}(\Omega)\big)}\leq
C\Big(N,\varepsilon,T,\|\nabla\Phi^*\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}\Big)\label{v215}.
\end{align}
Based on \eqref{v214} and the properties of the elliptic problem (\cite{Ab}), we infer that there exists a unique solution $\Phi$ to the following problem
\begin{equation}\label{15}
\left\{
\begin{aligned}
&-\Delta\Phi-\varepsilon\Delta^{2m+1}\Phi=\int_{\mathbb{R}^3}f\,dv-c(x),\\
&\Phi|_{(0,T)\times\partial\Omega}=\Delta\Phi|_{(0,T)\times\partial\Omega}=\cdot\cdot\cdot=
\Delta^{2m}\Phi|_{(0,T)\times\partial\Omega}=0,
\end{aligned}
\right.
\end{equation}
where $f$ is the weak solution to system \eqref{12}-\eqref{v213}. Furthermore,
\begin{align*}
&\|\Phi\|_{L^\infty\big(0,T;W^{4m+2,\frac{5}{3}}(\Omega)\big)}\leq C(\varepsilon)
\Big(\Big\|\int_{\mathbb{R}^3}f\,dv \Big\|_{L^\infty\big(0,T;L^\frac{5}{3}(\Omega)\big)}+\|c(x)\|_{L^\frac{5}{3}(\Omega)} \Big),\\
&\|\Phi\|_{L^\infty\big(0,T;W^{2,\frac{5}{3}}(\Omega)\big)}\leq C
\Big(\Big\|\int_{\mathbb{R}^3}f\,dv \Big\|_{L^\infty\big(0,T;L^\frac{5}{3}(\Omega)\big)}+\|c(x)\|_{L^\frac{5}{3}(\Omega)} \Big).
\end{align*}
It is time to find an approximate solution to the Navier-Stokes system \eqref{26}-\eqref{29} when $f$
is the weak solution to the problem \eqref{12}-\eqref{v213}.
Notice that $u_N\in C([0,T];X)$ can be written as
$$u_N(t,x)=\sum^N_{i=1}\zeta_i(t)\Psi_i(x),$$
where $\zeta_i(t)$ $(i=1,2,\cdots N)$ are functions of $t$.
Thanks to \eqref{v214}-\eqref{v215}, we derive that
\begin{align*}
\Big\|\chi_{\{|\tilde u| \leq N\}}\int_{\mathbb{R}^3}(\tilde u-v)f\,dv \Big\|_{L^\infty(0,T;L^2(\Omega))}\leq
C\Big(N,\varepsilon,T,\|\nabla\Phi^*\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}\Big).
\end{align*}
An argument similar to the one used in \cite{CJN} shows that the problem \eqref{26}-\eqref{29}
possesses a unique approximate solution $(\rho_{N},u_{N})$
on the whole time interval $[0,T]$.
In addition, the couple $(\rho_{N},u_{N})$ satisfies, for any $(\tau,x)\in (0,T)\times\Omega$,
\begin{equation}\label{30}
\inf_{x\in\Omega}\rho_0(x)e^{-\int_0^T\|{\rm{div}}u_N\|_{L^\infty(\Omega)}\,dt}\leq \rho_N(\tau,x)\leq
\sup_{x\in \Omega}\rho_0(x) e^{\int_0^T\|{\rm{div}}u_N\|_{L^\infty(\Omega)}\,dt},
\end{equation}
and,
\begin{align}\label{31}
&\int_{\Omega}\Big(\frac{1}{2}\rho_{N}|u_{N}-u_\infty|^2+\frac{1}{\gamma-1}\rho_{N}^{\gamma}
+\frac{\delta}{\beta-1}\rho_{N}^{\beta}+\frac{1}{2}\rho_{N}^{2}\Big)\,dx
+\varepsilon\int_0^{\tau}\int_{\Omega}|\nabla\rho_{N}|^2\,dxdt\nonumber\\
&+\frac{1}{2}\int^\tau_0\int_{\partial\Omega}\rho_{N}^2|u_B\cdot \nu(x)|\,d\sigma(x)dt
+\varepsilon\int^{\tau}_0\int_{\Omega}(\gamma\rho_{N}^{\gamma-2}+\delta\beta\rho_{N}^{\beta-2})|\nabla\rho_{N}|^2\,dxdt\nonumber\\
&+\int_0^{\tau}\int_{\Omega}\mathbb{S}(\nabla(u-u_\infty)):\nabla(u-u_\infty)\,dxdt
+\int_0^\tau\int_{\Gamma_{\rm{out}}}\Big(\frac{1}{\gamma-1}\rho_{N}^\gamma+\frac{\delta}{\beta-1}\rho_{N}^\beta \Big)|u_B\cdot \nu(x)|\,d\sigma(x)dt
\nonumber\\
&+\varepsilon\int_0^\tau\int_{\Omega}|\nabla(u_{N}-u_\infty)|^4\,dxdt
+\int_0^\tau\int_{\Gamma_{\rm{in}}}(\rho_{N}^\gamma+\delta\rho_{N}^\beta)|u_B\cdot\nu(x)|\,d\sigma(x)dt
\nonumber\\
\leq\;&\int_{\Omega}\Big(\frac{1}{2}\rho_0|u_0-u_\infty|^2+\frac{1}{\gamma-1}\rho^{\gamma}_0
+\frac{\delta}{\beta-1}\rho_0^{\beta}+\frac{1}{2}\rho_0^2\Big)\,dx
+\int^\tau_0\int_{\Gamma_{\rm{in}}}\rho_{N}\rho_B|u_B\cdot \nu(x)|\,d\sigma(x)dt \nonumber\\
&+\int^{\tau}_0\int_{\Gamma_{\rm{in}}}\Big(\frac{\gamma}{\gamma-1}\rho_{N}^{\gamma-1}
+\frac{\delta\beta}{\beta-1}\rho_{N}^{\beta-1}\Big)\rho_B|u_B\cdot \nu(x)|\,d\sigma(x)dt
-\frac{1}{2}\int^{\tau}_0\int_{\Omega}\rho_{N}^2{\rm{div}}u_{N}\,dxdt\nonumber\\
&+\int^{\tau}_0\int_{\Omega}\big(-(\rho_{N}^{\gamma}+\delta\rho_{N}^{\beta}){\rm{div}}u_\infty
-\mathbb{S}(\nabla u_\infty):\nabla(u_{N}-u_\infty)
-\rho_{N} u_{N}\cdot\nabla u_\infty\cdot(u_{N}-u_\infty)\nonumber\\
&\qquad+\varepsilon\nabla\rho_{N}\cdot\nabla(u_{N}-u_\infty)\cdot u_\infty \big)\,dxdt
+\int^{\tau}_0\int_{\Omega\times\mathbb{R}^3}f(v-\tilde{u})\chi_{\{|\tilde u|\leq N\}}\cdot (u_{N}-u_\infty)\,dxdvdt.
\end{align}
\subsection{An argument for $(\tilde u,\nabla\Phi^*)=(u_N,\nabla\Phi_N)$}\label{fix}
Before proceeding further, we recall a special case of Leray-Schauder fixed point theorem \cite{GT}:
\begin{lemma}\label{LS}
Let $\mathscr{T}$ be a compact mapping of a Banach space $\mathscr{B}$ into itself, and suppose there exists a constant $M$ such that
\begin{align*}
\|w\|_{\mathscr{B}}<M,
\end{align*}
for all $w\in\mathscr{B}$ and $\sigma\in [0,1]$ satisfying $w=\sigma\mathscr{T}w$. Then $\mathscr{T}$ has a fixed point.
\end{lemma}
To establish the existence of approximate solution to the problem \eqref{12}-\eqref{29} with $(\tilde u,\nabla\Phi^*)$ replaced by $(u_{N},\nabla\Phi_N)$, we define an operator $\mathscr{T}$:
\begin{align*}
\mathscr{T}: L^2((0,T)\times\Omega)\times L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big) &\rightarrow L^2((0,T)\times\Omega)\times L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)\\
(\tilde{u},\nabla\Phi^*) &\mapsto (u_{N},\nabla\Phi_N).
\end{align*}
Let $\{(\tilde{u_i},\nabla\Phi^*_i) \}_i$ be a uniform bounded sequence in $L^2((0,T)\times\Omega)\times L^\infty(0,T;W^{4m,\frac{5}{3}}(\Omega))$ and
$(f_{Ni},\rho_{Ni},u_{Ni})$ be the corresponding sequence solutions constructed in Subsection \ref{2.1.1}.
It is easy to see
\begin{align*}
\|\nabla\Phi_{i}^*\|_{L^\infty((0,T)\times\Omega)}\leq
C\|\nabla\Phi_i^*\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}\leq C,
\end{align*}
where $C$ appeared in this subsection is independent of $i$.
With this observation at hand, and combining the inequality \eqref{v23} with $l=2$, we get
\begin{align*}
\left\||v|^2f_{Ni}\right\|_{L^\infty(0,T;L^1(\Omega\times\mathbb{R}^3))}\leq C.
\end{align*}
And in accordance with \eqref{v25} and Lemma \ref{1'}, we derive that
\begin{align}\label{v27}
\Big\|\int_{\mathbb{R}^3}f_{Ni}\,dv \Big\|_{L^\infty\big(0,T;L^\frac{5}{3}(\Omega)\big)}
+\Big\|\int_{\mathbb{R}^3}vf_{Ni}\,dv \Big\|_{L^\infty\big(0,T;L^\frac{5}{4}(\Omega)\big)}\leq C.
\end{align}
From \eqref{15}, we obtain that
\begin{align}\label{v28}
\|\Phi_{Ni}\|_{L^\infty\big(0,T;W^{4m+2,\frac{5}{3}}(\Omega)\big)}\leq C.
\end{align}
Similarly, we can deduce that
\begin{align}\label{v29}
\|\partial_t\Phi_{Ni}\|_{L^\infty\big(0,T;W^{4m+1,\frac{5}{4}}(\Omega)\big)}\leq C,
\end{align}
where we have used \eqref{v24}, \eqref{15} and \eqref{v27}.
The inequalities \eqref{v28} and \eqref{v29} allow us to use Aubin-Lions lemma to get
\begin{align}
\Phi_{Ni}\rightarrow\Phi_N \quad {\rm{in}}\quad C\big([0,T];W^{4m+1,\frac{5}{3}}(\Omega)\big).
\end{align}
For the Navier-Stokes equations, the last term on the right-hand side of the inequality \eqref{31} written with $(f_{Ni},\rho_{Ni},u_{Ni})$ can be calculated as
\begin{align*}
&\int^{\tau}_0\int_{\Omega\times\mathbb{R}^3}f_{Ni}(v-\tilde{u_i})\chi_{\{|\tilde u_i|\leq N\}}\cdot (u_{Ni}-u_\infty)\,dxdvdt\\
\leq\; &\|u_{Ni}-u_\infty\|_{L^5((0,T)\times\Omega)}\Big(\Big\|\int_{\mathbb{R}^3}vf_{Ni}\,dv \Big\|_{L^\frac{5}{4}((0,T)\times\Omega)}
+N \Big\|\int_{\mathbb{R}^3}f_{Ni}\,dv \Big\|_{L^\frac{5}{4}((0,T)\times\Omega)}\Big)\\
\leq \;& C\big(\|u_{Ni}-u_\infty\|_{L^2((0,T)\times\Omega)}+\|\nabla(u_{Ni}-u_\infty)\|_{L^2((0,T)\times\Omega)}\big )\\
\leq \;& \alpha\int_0^\tau\int_\Omega |\nabla(u_{Ni}-u_\infty)|^2\,dxdt+C,
\end{align*}
where $\alpha$ is a small constant. In addition, we notice that (\cite{CJN})
\begin{align*}
\|u_{Ni}-u_\infty\|_{H^1(\Omega)}^2\leq C\|\mathbb{S}(\nabla(u_{Ni}-u_\infty)):\nabla(u_{Ni}-u_\infty)\|_{L^1(\Omega)}.
\end{align*}
Following the same path as in \cite{CJN}, we can deal with the other terms on the right-side hand of the inequality \eqref{31} to deduce that
\begin{align}\label{v210}
\|u_{Ni}\|_{L^2(0,T;H^1(\Omega))}\leq C.
\end{align}
An argument similar to the one used in \cite{LL2} shows that there exists a constant $p\in (1,\infty)$ such that
\begin{align}\label{v211}
\|\partial_t u_{Ni}\|_{L^p(0,T;W^{-1,p}(\Omega))}\leq C,
\end{align}
where we have used \eqref{28}, \eqref{30} and \eqref{31}.
By means of \eqref{v210}, \eqref{v211} and Aubin-Lions lemma, we infer
\begin{align*}
u_{Ni}\rightarrow u_{N}\quad {\rm{in}}\quad L^2((0,T)\times\Omega).
\end{align*}
For any
$(\tilde u,\nabla\Phi^*)\in L^2((0,T)\times\Omega)\times L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)$ satisfying
$(\tilde u,\nabla\Phi^*)=\sigma\mathscr{T}(\tilde u,\nabla\Phi^*)=\sigma(u_N,\nabla\Phi_N)$, $\sigma\in [0,1]$,
let $(f_N,\rho_N,u_N)$ be the associated solution.
Thanks to \eqref{v24} and \eqref{15}, we can rewrite the last term on the right-hand side of \eqref{v23} with $l=2$ as
\begin{align*}
-2\int_{\Omega\times\mathbb{R}^3}\nabla\Phi^*\cdot vf_N\,dxdv&=-2\sigma\int_\Omega\nabla\Phi_N\cdot\int_{\mathbb{R}^3}vf_N\,dvdx\\
&=-2\sigma\int_\Omega\Phi_N\,\partial_t\int_{\mathbb{R}^3}f_N\,dvdx\\
&=-\sigma\frac{d}{dt}\int_\Omega(\varepsilon|\nabla^{2m+1}\Phi_N |^2+|\nabla\Phi_N|^2)\,dx.
\end{align*}
As a consequence, we have
\begin{align*}
&\frac{d}{dt}\int_{\Omega}\Big(\frac{\sigma}{2}\varepsilon|\nabla^{2m+1}\Phi_N|^2
+\frac{\sigma}{2}|\nabla\Phi_N|^2+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_N\,dv\Big) \,dx
+\int_{\Sigma^\pm}(v\cdot\nu(x))\frac{|v|^2}{2}\gamma^\pm f_N\,d\sigma(x)dv\\
\leq\; & 3\int_{\Omega\times\mathbb{R}^3}f_N\,dxdv
-\int_{\Omega\times\mathbb{R}^3}|v|^2f_N\,dxdv+\int_{\Omega\times\mathbb{R}^3}\tilde u\chi_{\{|\tilde u|\leq \lambda \}}\cdot vf_N\,dxdv.
\end{align*}
It is apparent from this inequality and Gronwall inequality that
\begin{align}\label{v212}
\left\||v|^2f_N\right\|_{L^\infty(0,T;L^1(\Omega\times\mathbb{R}^3))}\leq \tilde C,
\end{align}
where $\tilde C$ is a positive constant independent of $\sigma$.
From \eqref{v25}, \eqref{15}, \eqref{v212} and Lemma \ref{1'}, we get
\begin{align}
\|\nabla\Phi_N\|_{L^\infty\big(0,T;W^{4m+1,\frac{5}{3}}(\Omega)\big)}\leq \tilde C.
\end{align}
Therefore, it holds
\begin{align*}
\|\nabla\Phi^*\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}
\leq \sigma\|\nabla\Phi_N\|_{L^\infty\big(0,T;W^{4m,\frac{5}{3}}(\Omega)\big)}\leq \tilde C.
\end{align*}
We perform the same reasoning as that in the proof of \eqref{v210} to deduce that
\begin{align*}
\|\tilde u\|_{L^2((0,T)\times\Omega)}\leq \sigma\|u_N\|_{L^2((0,T)\times\Omega)}\leq \tilde C.
\end{align*}
It is possible to use Lemma \ref{LS} to yield that there exist $(u_N,\nabla\Phi_N)$ such that
$(u_N,\nabla\Phi_N)=\mathscr{T}(u_N,\nabla\Phi_N)$.
We are led to the conclusion that $(f_N,\Phi_{N},\rho_{N},u_{N})$ satisfies\\
1. for any $\varphi\in C^\infty_c([0,T)\times\bar\Omega\times\mathbb{R}^3)$ such that $\varphi=0$ on $(0,T)\times\Sigma^+$, it holds
\begin{align}\label{32}
&\int^{T}_0\int_{\Omega\times\mathbb{R}^3}f_N\big(\partial_t\varphi+v\cdot\nabla_x\varphi
+(u_{N}\chi_{\{|u_{N}|\leq N\}}-v)\cdot\nabla_v\varphi
-\nabla_x\Phi_{N}\cdot\nabla_v\varphi
+\Delta_v\varphi\big)\,dxdvdt\nonumber\\
&=-\int_{\Omega\times\mathbb{R}^3}f_{0\varepsilon}\varphi(0,x,v)\,dxdv
+\int_0^T\int_{\Sigma^-}(v\cdot \nu(x))g_\varepsilon\varphi\,d\sigma(x)dvdt;
\end{align}
2. for any $\Psi\in C_c^\infty((0,T)\times\Omega)$, it holds
\begin{align}\label{33}
&\int_0^T\int_{\Omega}\nabla\Phi_{N}\cdot \nabla\Psi\,dxdt
+\varepsilon\int_0^T\int_\Omega\nabla^{2m+1}\Phi_{N}\cdot\nabla^{2m+1}\Psi\,dxdt\nonumber\\
&=\int_0^T\int_\Omega\big(n_{N}-c(x) \big)\Psi\,dxdt,
\end{align}
where $n_{N}:=\int_{\mathbb{R}^3}f_{N}\,dv$;\\
3. for any $\psi\in C^\infty_c([0,T)\times(\Omega\cup\Gamma_{\rm{in}}))$, it holds
\begin{align}\label{34}
&\int^{T}_0\int_{\Omega}(\rho_{N}\partial_t\psi+\rho_{N} u_{N}\cdot\nabla\psi-\varepsilon\nabla\rho_{N}\cdot \nabla\psi)\,dxdt
+\int_{\Omega}\rho_0\psi(0,x)\,dx\nonumber\\
&\qquad\qquad\qquad=\int_0^T\int_{\Gamma_{\rm{in}}}\rho_B u_B\cdot \nu(x)\varphi\,d\sigma(x)dt;
\end{align}
4. for any $\phi\in C^\infty_c((0,T)\times\Omega;\mathbb{R}^3)$, it holds
\begin{align}\label{35}
\int^{T}_{0}&\int_{\Omega}\big(\rho_{N} u_{N}\cdot\partial_t\phi+(\rho_{N} u_{N}\otimes u_{N}):\nabla\phi+\rho_{N}^\gamma{\rm{div}}\phi
+\delta\rho_{N}^\beta{\rm{div}}\phi
-\varepsilon\nabla\rho_{N}\cdot\nabla u_{N}\phi\nonumber\\
&-\varepsilon|\nabla(u_{N}-u_\infty)|^2\nabla(u_{N}-u_\infty):\nabla\phi
-\mathbb{S}(\nabla u_{N}):\nabla\phi
+(j_N-n_N u_{N})\chi_{\{|u_{N}|\leq N\}}\cdot\phi\big)\,dxdt\nonumber\\
&+\int_{\Omega}\rho_0u_0\cdot \phi(0,x)\,dx=0,
\end{align}
where $j_N:=\int_{\mathbb{R}^3}vf_N\,dv$.
Moreover, the energy inequalities \eqref{v23} and \eqref{31} with $(\tilde u,\nabla\Phi^*)$ replaced by $(u_{N},\nabla\Phi_N)$ hold.
\subsection{Uniform estimates independent of $N$}\label{uN}
In order to take the limit on the regularized paremeter $N$,
we derive the uniform estimates satisfied by $(f_N,\Phi_{N},\rho_{N},u_{N})$ in this subsection.
Summing \eqref{v23} with $l=2$ and \eqref{31} with $(\tilde u,\nabla\Phi^*)=(u_{N},\nabla\Phi_N)$ up, we have
\begin{align}\label{e2}
&\int_{\Omega}\Big(\frac{1}{2}\rho_{N}|u_{N}-u_\infty|^2+\frac{1}{\gamma-1}\rho_{N}^{\gamma}
+\frac{\delta}{\beta-1}\rho_{N}^{\beta}+\frac{1}{2}\rho_{N}^{2}
+\frac{\varepsilon}{2}|\nabla^{2m+1}\Phi|^2+\frac{1}{2}|\nabla\Phi|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_N\,dv\Big)\,dx\nonumber\\
&+\varepsilon\int_0^{\tau}\int_{\Omega}|\nabla\rho_{N}|^2\,dxdt
+\frac{1}{2}\int^\tau_0\int_{\partial\Omega}\rho_{N}^2|u_B\cdot \nu(x)|\,d\sigma(x)dt\nonumber\\
&+\int_0^{\tau}\int_{\Omega}\mathbb{S}(\nabla(u-u_\infty)):\nabla(u-u_\infty)\,dxdt
+\int_0^\tau\int_{\Gamma_{\rm{out}}}\Big(\frac{1}{\gamma-1}\rho_{N}^\gamma+\frac{\delta}{\beta-1}\rho_{N}^\beta\Big)|u_B\cdot\nu(x)|\,d\sigma(x)dt
\nonumber\\
&+\varepsilon\int^{\tau}_0\int_{\Omega}(\gamma\rho_{N}^{\gamma-2}+\delta\beta\rho_{N}^{\beta-2})|\nabla\rho_{N}|^2\,dxdt
+\int^{\tau}_0\int_{\Sigma^-}(v\cdot \nu(x))\frac{|v|^2}{2}g_\varepsilon\,d\sigma(x)dvdt\nonumber\\
&+\varepsilon\int_0^{\tau}\int_{\Omega}|\nabla(u_{N}-u_\infty)|^4\,dxdt
+\int_0^\tau\int_{\Gamma_{\rm{in}}}(\rho_{N}^\gamma+\delta\rho_{N}^\beta)|u_B\cdot\nu(x)|\,d\sigma(x)dt\nonumber\\
\leq\;&\int_{\Omega}\Big(\frac{1}{2}\rho_0|u_0-u_\infty|^2+\frac{1}{\gamma-1}\rho^{\gamma}_0
+\frac{\delta}{\beta-1}\rho_0^{\beta}+\frac{1}{2}\rho_0^2
+\frac{\varepsilon}{2}|\nabla^{2m+1}\Phi_0|^2+\frac{1}{2}|\nabla\Phi_0|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_{0\varepsilon}\,dv\Big)\,dx \nonumber\\
&+\int^{\tau}_0\int_{\Gamma_{\rm{in}}}\Big(\frac{\gamma}{\gamma-1}\rho_{N}^{\gamma-1}
+\frac{\delta\beta}{\beta-1}\rho_{N}^{\beta-1}\Big)\rho_B|u_B\cdot \nu(x)|\,d\sigma(x)dt\,d\sigma(x)dt
-\frac{1}{2}\int^{\tau}_0\int_{\Omega}\rho_{N}^2{\rm{div}}u_{N}\,dxdt\nonumber\\
&+\int^{\tau}_0\int_{\Omega}\big(-(\rho_{N}^{\gamma}+\delta\rho_{N}^{\beta}){\rm{div}}u_\infty
-\mathbb{S}(\nabla u_\infty):\nabla(u_{N}-u_\infty)
-\rho_{N} u_{N}\cdot\nabla u_\infty\cdot(u_{N}-u_\infty)\nonumber\\
&\qquad +\varepsilon\nabla\rho_{N}\cdot\nabla(u_{N}-u_\infty)\cdot u_\infty \big)\,dxdt
-\int^{\tau}_0\int_{\Omega}(j_N-n_Nu_{N})\chi_{\{|u_{N}|\leq N\}}\cdot u_\infty\,dxdt\nonumber\\
&+3\int^{\tau}_0\int_{\Omega\times\mathbb{R}^3}f_N\,dxdvdt
+\int_0^\tau\int_{\Gamma_{\rm{in}}}\rho_N\rho_B|u_B\cdot \nu(x)|\,d\sigma(x)dt.
\end{align}
Using the same arguments as that in \cite{LL2}, we can carry out that the external force term of the energy inequality \eqref{e2} can be calculated as
\begin{align}\label{100}
&\Big|\int_0^\tau\int_\Omega(j_N-n_Nu_{N})\chi_{\{|u_{N}|\leq N\}}\cdot u_\infty\,dxdt \Big|\nonumber\\
\leq &\alpha\int_0^\tau\int_{\Omega}|\nabla(u_{N}-u_\infty)|^2\,dxdt
+\tilde C\int_0^\tau\int_{\Omega\times\mathbb{R}^3}|v|^2f_N\,dxdvdt+\tilde C,
\end{align}
where $\tilde C>0$ is independent of $N$ and $\varepsilon$.
Inserting \eqref{100} into the energy inequality \eqref{e2}, we obtain that
\begin{align}
&\|f_N\|_{L^\infty(0,T;L^1(\Omega\times\mathbb{R}^3))}+\|f_N\|_{L^\infty(0,T;L^\infty(\Omega\times\mathbb{R}^3))}
+\||v|^2f_N\|_{L^\infty(0,T;L^1(\Omega\times\mathbb{R}^3))}\leq C,\label{40}\\
&\|\rho_{N}|u_{N}-u_\infty|^2\|_{L^\infty(0,T;L^1(\Omega))}
+\|u_{N}-u_\infty\|_{L^2(0,T;H^1(\Omega))}\leq C,\label{41}\\
&\|\rho_{N}\|_{L^\infty(0,T;L^\beta(\Omega))}
+\varepsilon\|\nabla(u_{N}-u_\infty)\|^4_{L^4(0,T;L^4(\Omega))}\leq C,\label{42}\\
&\varepsilon\|\nabla\rho_{N}\|^2_{L^2(0,T;L^2(\Omega))}
+\varepsilon\|\nabla(\rho_{N}^{\frac{\beta}{2}})\|^2_{L^2(0,T;L^2(\Omega))}\leq C,\label{43}\\
&\Big\|\int_{\mathbb{R}^3}f_N\,dv\Big\|_{L^\infty(0,T;L^p(\Omega))}\leq C, \quad p\in\Big[1,\frac{5}{3}\Big],\label{113}\\
&\Big\|\int_{\mathbb{R}^3}vf_N\,dv\Big\|_{L^\infty(0,T;L^p(\Omega))}\leq C, \quad p\in\Big[1,\frac{5}{4}\Big],\label{114}
\end{align}
where $C>0$ is independent of $N$, but depends on $\varepsilon$.
\subsection{Taking the limit $N\rightarrow +\infty$}\label{lN}
Based on the uniform bounds obtained in Subsection \ref{uN}, we take the limit $N\rightarrow +\infty$ in the equalities \eqref{32}-\eqref{35}.
The procedure of passing to the limit in the fluid equations \eqref{34}-\eqref{35} can be found in \cite{CJN} and so is omitted.
Here we only list the key points of handling with the kinetic system for brevity.
\begin{prop}\label{p2}
There exists a subsequence $(f_N,\Phi_{N})$ (not relabeled) satisfying
\begin{align}
&f_N\stackrel{*}{\rightharpoonup}f \quad {\rm{in}}\quad L^\infty(0,T;L^p(\Omega\times\mathbb{R}^3)),\quad 1<p\leq\infty,\label{20}\\
&\int_{\mathbb{R}^3}f_N\,dv \stackrel{*}{\rightharpoonup}\int_{\mathbb{R}^3}f\,dv \quad {\rm{in}} \quad L^\infty(0,T;L^p(\Omega)),\quad
p\in \Big[1,\frac{5}{3} \Big],\label{21}\\
&\int_{\mathbb{R}^3}vf_N\,dv \stackrel{*}{\rightharpoonup}\int_{\mathbb{R}^3}vf\,dv \quad {\rm{in}} \quad L^\infty(0,T;L^p(\Omega)),\quad
p\in \Big[1,\frac{5}{4} \Big],\label{22}\\
&\Phi_{N}\rightarrow \Phi \quad {\rm{in}}\quad C([0,T];W^{1,q}(\Omega)),\quad q\in\Big[1,\frac{15}{4} \Big),\label{23}\\
&\nabla^{2m+1}\Phi_{N}\rightarrow \nabla^{2m+1}\Phi \quad {\rm{in}}\quad C([0,T];W^{1,q}(\Omega)),\quad q\in\Big[1,\frac{15}{4} \Big).\label{101}
\end{align}
\end{prop}
\begin{proof}
It is apparent from \eqref{40}, \eqref{113} and \eqref{114} that \eqref{20}-\eqref{22} hold. It only remains to prove \eqref{23} and \eqref{101}.
Using \eqref{15}, we have
\begin{align*}
-\Delta \partial_t\Phi_{N}-\varepsilon\Delta^{2m+1}\partial_t\Phi_{N}=-{\rm{div}}_x\int_{\mathbb{R}^3}vf_{N}\,dv.
\end{align*}
Meanwhile, by means of the inequality \eqref{114}, it holds
\begin{align}\label{116}
\|\partial_t\Phi_{N}\|_{L^\infty(0,T;W^{1,p}(\Omega))}\leq C,\quad p\in \Big(1,\frac{5}{4} \Big].
\end{align}
Proceeding as that in the proof of \eqref{116}, we obtain
\begin{align}\label{117}
\|\Phi_{N}\|_{L^\infty(0,T;W^{2,p}(\Omega))}\leq C,\quad p\in \Big(1,\frac{5}{3} \Big].
\end{align}
The estimates \eqref{116} and \eqref{117} allow us to use Aubin-Lions lemma to get \eqref{23}.
Similarly, we can obtain \eqref{101}.
\end{proof}
The convergence stated in Proposition \ref{p2} makes it possible to perform the limit $N\rightarrow +\infty$
in the equalites \eqref{32}-\eqref{33}.
Thanks to the weak lower semicontinuity of convex functions, we are able to pass to the limit $N\rightarrow+\infty$ in the energy inequality \eqref{e2}. In conclusion, we have proved the following proposition.
\begin{prop}\label{p1}
The problem \eqref{3}-\eqref{10} has a global weak solution $(f,\Phi,\rho,u)$ satisfying, for any $\tau\in(0,T)$,
\begin{align}\label{e1}
&\int_{\Omega}\Big(\frac{1}{2}\rho|u-u_\infty|^2+\frac{1}{\gamma-1}\rho^{\gamma}
+\frac{\delta}{\beta-1}\rho^{\beta}+\frac{1}{2}\rho^{2}
+\frac{1}{2}|\nabla\Phi|^2+\frac{\varepsilon}{2}|\nabla^{2m+1}\Phi|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f\,dv\Big)\,dx\nonumber\\
&+\varepsilon\int_0^{\tau}\int_{\Omega}|\nabla\rho|^2\,dxdt
+\varepsilon\int^{\tau}_0\int_{\Omega}(\gamma\rho^{\gamma-2}+\delta\beta\rho^{\beta-2})|\nabla\rho|^2\,dxdt
+\frac{1}{2}\int^\tau_0\int_{\partial\Omega}\rho^2|u_B\cdot \nu(x)|\,d\sigma(x)dt \nonumber\\
&+\int_0^{\tau}\int_{\Omega}\mathbb{S}(\nabla(u-u_\infty)):\nabla(u-u_\infty)\,dxdt
+\varepsilon\int_0^{\tau}\int_{\Omega}|\nabla(u-u_\infty)|^4\,dxdt
\nonumber\\
&+\int_0^\tau\int_{\Gamma_{\rm{in}}}(\rho^\gamma+\delta\rho^\beta)|u_B\cdot\nu(x)|\,d\sigma(x)dt
+\int_0^\tau\int_{\Gamma_{\rm{out}}}\Big(\frac{1}{\gamma-1}\rho^\gamma+\frac{\delta}{\beta-1}\rho^\beta \Big)|u_B\cdot\nu(x)|\,d\sigma(x)dt\nonumber\\
&+\int^{\tau}_0\int_{\Sigma^-}(v\cdot \nu(x))\frac{|v|^2}{2}g_\varepsilon\,d\sigma(x)dvdt
-\int^\tau_0\int_{\Gamma_{\rm{in}}}\rho\rho_B|u_B\cdot \nu(x)|\,d\sigma(x)dt\nonumber\\
\leq\;&\int_{\Omega}\Big(\frac{1}{2}\rho_0|u_0-u_\infty|^2+\frac{1}{\gamma-1}\rho^{\gamma}_0
+\frac{\delta}{\beta-1}\rho_0^{\beta}+\frac{1}{2}\rho_0^2
+\frac{1}{2}|\nabla\Phi_0|^2+\frac{\varepsilon}{2}|\nabla^{2m+1}\Phi_0|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_{0\varepsilon}\,dv\Big)\,dx \nonumber\\
&-\frac{1}{2}\int^{\tau}_0\int_{\Omega}\rho^2{\rm{div}}u\,dxdt
-\int^{\tau}_0\int_{\Omega}(j-nu)\cdot u_\infty\,dxdt
+3\int_{0}^\tau\int_{\Omega\times\mathbb{R}^3}f\,dxdvdt\nonumber\\
&+\int^{\tau}_0\int_{\Omega}\big(-(\rho^{\gamma}+\delta\rho^{\beta}){\rm{div}}u_\infty
-\mathbb{S}(\nabla u_\infty):\nabla(u-u_\infty)
-\rho u\cdot\nabla u_\infty\cdot(u-u_\infty)\nonumber\\
&\qquad +\varepsilon\nabla\rho\cdot\nabla(u-u_\infty)\cdot u_\infty \big)\,dxdt
+\int_0^\tau\int_{\Gamma_{\rm{in}}}\Big(\frac{\gamma}{\gamma-1}\rho^{\gamma-1}
+\frac{\delta\beta}{\beta-1}\rho^{\beta-1} \Big)\rho_B|u_B\cdot\nu(x)|\,d\sigma(x)dt.
\end{align}
\end{prop}
\begin{remark}
The weak formulation of momentum equation \eqref{6} reads as
\begin{align*}
\int^{T}_{0}&\int_{\Omega}\big(\rho u\cdot\partial_t\phi+(\rho u\otimes u):\nabla\phi+\rho^\gamma{\rm{div}}\phi
+\delta\rho^\beta{\rm{div}}\phi
-\varepsilon\nabla\rho\cdot\nabla u\cdot\phi\nonumber\\
&-Z_\varepsilon:\nabla\phi
-\mathbb{S}(\nabla u):\nabla\phi
+(j-nu)\cdot\phi\big)\,dxdt
+\int_{\Omega}\rho_0u_0\cdot \phi(0,x)\,dx=0,
\end{align*}
for any $\phi\in C_c^\infty((0,T)\times\Omega;\mathbb{R}^3)$.
Here $Z_\varepsilon$ is the weak limit of the term $\varepsilon|\nabla(u_{N}-u_\infty)|^2\nabla(u_{N}-u_\infty)$ in the equality \eqref{35}, that is,
\begin{align*}
\varepsilon|\nabla(u_{N}-u_\infty)|^2\nabla(u_{N}-u_\infty)\rightharpoonup Z_\varepsilon \quad
{\rm{in}} \quad L^\frac{4}{3}((0,T)\times\Omega),
\end{align*}
and
\begin{align*}
\|Z_\varepsilon\|_{L^\frac{4}{3}((0,T)\times\Omega)}\rightarrow 0 \quad {\rm{as}}\quad \varepsilon\rightarrow 0,
\end{align*}
where we have used the inequality \eqref{42}.
\end{remark}
\section{Vanishing limits and the proof of Theorem \ref{main}}\label{l}
Our ultimate goal is to take limits on $\varepsilon$ and $\delta$ in turn to get the desired existence result.
\subsection{Taking the limit $\varepsilon\rightarrow 0$}\label{le}
We use $(f_\varepsilon,\Phi_\varepsilon,\rho_\varepsilon,u_\varepsilon)$ to denote the approximate solutions of this level.
With the help of regularized technique of the continuity equation, we can take limit in the weak formulation of fluid system
\eqref{5}-\eqref{7}, \eqref{9} and \eqref{10}. For a rigorous proof the reader is referred to \cite{CJN}.
Using the same arguments as that in Subsection \ref{lN}, we can easily
take limit $\varepsilon\rightarrow 0$ in the weak formulation of the problem
\eqref{3}, \eqref{4}, \eqref{7}-\eqref{8'}. Here we only list the difference.
Owing to the inequality \eqref{e1} written with $(f_\varepsilon,\Phi_\varepsilon,\rho_\varepsilon,u_\varepsilon)$
and Gronwall inequality, we derive that
\begin{align*}
\|\nabla\Phi_\varepsilon\|_{L^\infty(0,T;L^2(\Omega))}
+\sqrt{\varepsilon}\|\nabla^{2m+1}\Phi_\varepsilon\|_{L^\infty(0,T;L^2(\Omega))}\leq C,
\end{align*}
where $C>0$ appeared in this subsection is independent of $\varepsilon$. With this observation at hand, we infer that
\begin{align*}
\varepsilon\int_0^T\int_{\Omega}\nabla^{2m+1}\Phi_{\varepsilon}\cdot\nabla^{2m+1}\Psi\,dxdt
&\leq \varepsilon\|\nabla^{2m+1}\Phi_{\varepsilon}\|_{L^2((0,T)\times\Omega)}\|\nabla^{2m+1}\Psi\|_{L^2((0,T)\times\Omega)}\\
&\leq C\sqrt{\varepsilon}\rightarrow 0\quad {\rm{as}}\quad \varepsilon\rightarrow 0.
\end{align*}
Thus we can summarize what we have proved as the following proposition.
\begin{prop}
There exists a global weak solution $(f,\Phi,\rho,u)$ to the following system:
\begin{align}
&\partial_t f+v\cdot \nabla_x f
+{\rm{div}}_v((u-v)f)-\nabla_x\Phi\cdot\nabla_vf-\Delta_vf=0, \label{54}\\
&-\Delta\Phi=\int_{\mathbb{R}^3}f\,dv-c(x),\label{55}\\
&\partial_t\rho+{\rm{div}}_x(\rho u)=0, \label{56}\\
&\partial_t(\rho u)+{\rm{div}}(\rho u\otimes u)+\nabla\rho^{\gamma}+\delta\nabla\rho^{\beta}={\rm{div}}\mathbb{S}(\nabla u)
-\int_{\mathbb{R}^3}(u-v)f\,dv, \label{57}
\end{align}
with
\begin{align*}
&f(0,x,v)=f_{0}(x,v),\quad \rho(0,x)=\rho_0(x),\quad u(0,x)=u_0(x),\\
&\gamma^-f(t,x,v)\big|_{(0,T)\times\Sigma^-}=g(t,x,v),
\quad \Phi(t,x)\big|_{(0,T)\times\partial\Omega}=0,\\
&\rho(t,x)\big|_{(0,T)\times\Gamma_{\rm{in}}}=\rho_B(x),\quad
u(t,x)\big|_{(0,T)\times\partial\Omega}=u_B(x),
\end{align*}
where $(f_0,\rho_0,u_0)$ and $(g,\rho_B,u_B)$ satisfy \eqref{fv}-\eqref{mv} and \eqref{11}.
In addition, $(f,\Phi,\rho,u)$ verifies
\begin{align}\label{e3}
&\int_{\Omega}\Big(\frac{1}{2}\rho|u-u_\infty|^2+\frac{1}{\gamma-1}\rho^{\gamma}
+\frac{\delta}{\beta-1}\rho^{\beta}
+\frac{1}{2}|\nabla\Phi|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f\,dv\Big)\,dx\nonumber\\
&+\int_0^{\tau}\int_{\Omega}\mathbb{S}(\nabla(u-u_\infty)):\nabla(u-u_\infty)\,dxdt
+\int^{\tau}_0\int_{\Sigma^-}(v\cdot \nu(x))\frac{|v|^2}{2}g\,d\sigma(x)dvdt\nonumber\\
&+\int_0^\tau\int_{\Gamma_{\rm{in}}}(\rho^\gamma+\delta\rho^\beta)|u_B\cdot\nu(x)|\,d\sigma(x)dt
+\int_0^\tau\int_{\Gamma_{\rm{out}}}\Big(\frac{1}{\gamma-1}\rho^\gamma+\frac{\delta}{\beta-1}\rho^\beta \Big)|u_B\cdot\nu(x)|\,d\sigma(x)dt\nonumber\\
\leq\;&\int_{\Omega}\Big(\frac{1}{2}\rho_0|u_0-u_\infty|^2+\frac{1}{\gamma-1}\rho^{\gamma}_0
+\frac{\delta}{\beta-1}\rho_0^{\beta}
+\frac{1}{2}|\nabla\Phi_0|^2
+\int_{\mathbb{R}^3}\frac{|v|^2}{2}f_{0}\,dv\Big)\,dx \nonumber\\
&+\int^{\tau}_0\int_{\Gamma_{\rm{in}}}\Big(\frac{\gamma}{\gamma-1}\rho^{\gamma-1}
+\frac{\delta\beta}{\beta-1}\rho^{\beta-1}\Big)\rho_B|u_B\cdot \nu(x)|\,d\sigma(x)dt
+3\int^{\tau}_0\int_{\Omega\times\mathbb{R}^3}f\,dxdvdt
\nonumber\\
&+\int^{\tau}_0\int_{\Omega}\big(-(\rho^{\gamma}+\delta\rho^{\beta}){\rm{div}}u_\infty
-\mathbb{S}(\nabla u_\infty):\nabla(u-u_\infty)
-\rho u\cdot\nabla u_\infty\cdot(u-u_\infty) \big)\,dxdt\nonumber\\
&-\int^{\tau}_0\int_{\Omega}(j-nu)\cdot u_\infty\,dxdt.
\end{align}
\end{prop}
\subsection{Passing the limit as $\delta\rightarrow 0$}
In this section, we take the limit $\delta\rightarrow 0$ and relax our hypotheses on the initial and boundary data.
Performing the same arguments as those in Subsection \ref{le}, we can take the limit on $\delta$ in the weak formulation of kinetic system \eqref{54}-\eqref{55}.
And the details of taking limit in the weak formulation of fluid system \eqref{56}-\eqref{57} can be found in \cite{CJN}. We omit it for brevity.
Hence the proof of Theorem \ref{main} is completed.
\hfill\qedsymbol
\medskip
\indent
{\bf Acknowledgements:}
F. Li and Y. Li are supported by NSFC (Grant No. 12071212). And F. Li is also supported by a project funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions. NZ acknowledges support from the Alexander von Humboldt Foundation (AvH), from the Austrian Science Foundation (FWF) grant P30000, and from the bilateral Croatian-Austrian Project of the Austrian Agency for International Cooperation in Education and Research (ÖAD) grant HR 19/2020.
|
\section{Introduction}
Including wave function relaxation in state-specific approximations can provide an accurate
representation of excited states where there is significant electron density rearrangement relative to the
electronic ground state.
This relaxation is particularly important for the description of charge
transfer,\cite{Barca2018,Liu2012,Jensen2018,Shea2018,Hait2021}
core electron excitations,\cite{Hait2020a,Oosterbaan2018,Oosterbaan2020,Garner2020}
or Rydberg states with diffuse orbitals,\cite{CarterFenk2020,Shea2018,Clune2020}
\edit{and can be visualised using the eigenvectors of the difference density matrix for
the excitation.\cite{Plasser2014a,Plasser2014b}}
In contrast, techniques based on linear response theory
--- including time-dependent Hartree--Fock\cite{McLachlan1964} (TD-HF),
time-dependent density functional theory\cite{Runge1984,Dreuw2005,Burke2005} (TD-DFT),
configuration interaction singles\cite{Foresman1992,Dreuw2005} (CIS),
and equation of motion coupled cluster theory\cite{Stanton1993,Krylov2008} (EOM-CC) ---
are evaluated using the ground-state orbitals, making a balanced treatment of the ground
and excited states more difficult. \cite{Burke2005}
Furthermore, linear response methods are generally applied under the adiabatic approximation and are
limited to single excitations.\cite{Maitra2004,Burke2005}
In principle, state-specific approaches can approximate both single and double
excitations,\cite{Gilbert2008,Barca2018,Otis2020} although the open-shell character of single excitations requires a
multi-configurational approach.\cite{Shea2018,Hardikar2020,Zhao2020a,Zhao2020b,Ye2017}
Underpinning excited state-specific methods is the fundamental idea that ground-state wave functions
can also be used to describe an electronic excited state.
This philosophy relies on the existence of additional higher-energy mathematical solutions, which have been found
in Hartree--Fock (HF),\cite{Slater1951,Stanton1968,Fukutome1971,Fukutome1974a,Fukutome1974,Fukutome1975,Fukutome1973,Mestechkin1978,Mestechkin1979,Mestechkin1988,Davidson1983,Kowalski1998,Gilbert2008,Thom2008,Li2009a,Jimenez-Hoyos2011,Jimenez-Hoyos2014,Toth2016,Burton2018,Huynh2019,Lee2019,Burton2021}
density functional theory (DFT),\cite{Theophilou1979,Perdew1985,Zarotiadis2020,Hait2020}
multiconfigurational self-consistent field (MC-SCF),%
\cite{Olsen1982,Golab1983,Olsen1983,Golab1985,Angeli2003,Guihery1997,Evenhuis2011,Tran2019,Tran2020}
and coupled cluster (CC) theory.\cite{PiecuchBook,Mayhall2010,Lee2019b,Marie2021a,Kossoski2021}
These multiple solutions correspond to higher-energy stationary points of a parametrised approximate energy function,
including local energy minima, saddle points, or maxima.
\edit{It has long been known that the exact $k$-th excited state forms a saddle point of the energy
with $k$ negative Hessian eigenvalues (where $k=0$ is the ground state).
These stationary properties have been identified using the exponential parametrisation of MC-SCF calculation
\cite{Werner1981,Olsen1982,Golab1985,Olsen1983,Golab1983}
and can also be derived using local expansions around an exact eigenstate.\cite{Bacalis2016,Bacalis2020}
However, questions remain about the global structure of the exact energy landscape and the
connections between exact excited states.
}
Multiple self-consistent field (SCF) solutions in HF or Kohn--Sham DFT are the most
widely understood state-specific approximations.
Their existence was first identified by Slater,\cite{Slater1951} and later characterised in detail by
Fukutome.\cite{Fukutome1971,Fukutome1974a,Fukutome1974,Fukutome1975,Fukutome1973}
Physically, these solutions appear to represent single-determinant approximations for
excited states,\cite{Gilbert2008,Besley2009,Barca2014,Barca2018,Barca2018a}
or mean-field quasi-diabatic states.\cite{Thom2009,Jensen2018}
In the presence of strong electron correlation, multiple SCF solutions often break symmetries of the exact Hamiltonian
\cite{Davidson1983,Li2009a,Huynh2019,Jimenez-Hoyos2011,Jimenez-Hoyos2014,Lee2019}
and can disappear as the molecular geometry changes.\cite{Mestechkin1978,Mestechkin1979,Mestechkin1988,Burton2018}
The stability analysis pioneered by \v{C}i\v{z}ek and Paldus\cite{Cizek1967,Cizek1970,Thouless1960,Paldus1970}
allows SCF solutions to be classified according to their Hessian index (the number of downhill orbital rotations).
There are usually only a handful of low-energy SCF minima, connected by index-1 saddle points, while
symmetry-broken solutions form several degenerate minima that are connected by higher-symmetry saddle points.\cite{Burton2021}
At higher energies, stationary points representing excited states generally become higher-index saddle points of
the energy.\cite{Perdew1985,Dardenne2000,Burton2021}
Recent interest in locating higher-energy SCF solutions has led to several new approaches including:
modifying the iterative SCF approach with orbital occupation constraints\cite{Gilbert2008,Barca2018} or level-shifting;\cite{CarterFenk2020}
second-order direct optimisation of higher-energy stationary points;\cite{Levi2020,Levi2020a}
minimising an alternative functional such as the variance\cite{Ye2017,Ye2019,Shea2017,Shea2018,Cuzzocrea2020}
or the square-magnitude of the energy gradient.\cite{Hait2020}
\edit{The success of these algorithms depends on the structure of the approximate energy landscape, the
stationary properties of excited states, and the quality of the initial guess.
In principle, the approximate energy landscape is determined by the relationship between an approximate wave function and the
exact energy landscape.
However, the nature of this connection has not been widely investigated.}
Beyond single-determinant methods, state-specific approximations using multi-configurational wave functions
have been developed to describe open-shell or statically correlated excited states, including
MC-SCF,\cite{Olsen1982,Golab1985,Olsen1983,Golab1983,Tran2019,Tran2020}
excited-state mean-field (ESMF) theory,\cite{Shea2018,Hardikar2020,Zhao2020a,Zhao2020b}
half-projected HF,\cite{Ye2019} or multi-Slater-Jastrow functions.\cite{Dash2019,Cuzzocrea2020,Dash2021}
\edit{The additional complexity of these wave functions compared to a single determinant has led to the use
of direct second-order optimisation algorithms\cite{Olsen1982,Golab1985,Olsen1983,Golab1983,Werner1981}
or, more recently,
methods based on variance optimisation.\cite{Messmer1969,Shea2017,PinedaFlores2019}}
Variance optimisation exploits the fact that both ground and excited states form minima of the
Hamiltonian variance $\mel*{\Psi}{(\mathcal{H} - E)^2}{\Psi}$,\cite{MacDonald1934}
and thus excited-states can be identified using downhill minimisation techniques.
\edit{Alternatively, the folded-spectrum method uses an objective function with the form $\mel*{\Psi}{(\mathcal{H} - \omega)^2}{\Psi}$ to
target the state with energy closest to $\omega$.\cite{Wang1994}}
These approaches are particularly easy to combine with stochastic methods such as
variational Monte--Carlo\cite{Umrigar2005,PinedaFlores2019,Cuzzocrea2020} (VMC) and have
been proposed as an excited-state extension of variational quantum eigensolvers.\cite{Zhang2020,Zhang2021}
However, variance optimisation is prone to convergence issues that include
drifting away from the intended target state,\cite{Cuzzocrea2020,Otis2020} and very little is known about the
properties of the variance optimisation landscape, or its stationary points.
In my opinion, our limited understanding about the relationship between exact and approximate
state-specific solutions arises because exact electronic structure is traditionally
viewed as a matrix eigenvalue problem, while state-specific approximations are considered as higher-energy stationary
points of an energy landscape.
The energy landscape concept is more familiar to theoretical chemists in the context
of a molecular potential energy surface,\cite{Born1927} where
local minima correspond to stable atomic arrangements and index-1 saddles can be interpreted as
reactive transition states.\cite{Murrell1968,WalesBook}
To bridge these concepts, this article introduces a fully geometric perspective on
exact electronic structure within a finite Hilbert space.
In this representation, ground and excited states form stationary points of an energy landscape constrained
to the surface of a unit hypersphere.
Analysing the differential geometry of this landscape reveals the stationary properties of ground and excited states
and the pathways that connect them.
Furthermore, the square-magnitude of the exact gradient is shown to be directly proportional to the Hamiltonian
variance, allowing the structure of the exact variance optimisation landscape to be derived.
Finally, the relationship between approximate wave functions and the exact energy or variance is
explored, revealing how the stationary properties of state-specific solutions are controlled
by the structure of the exact energy landscape.
Throughout this work, key concepts are illustrated using the electronic singlet states
of \ce{H2}.
While this minimal model is used to allow visualisation of the exact energy landscape,
the key conclusions are mathematically general and can be applied to any number of electrons or basis functions.
Unless otherwise stated, all results are obtained with the STO\nobreakdash-3G basis set\cite{Hehre1969}
using Mathematica 12.0\cite{Mathematica} and are available in an accompanying notebook available for
download from \href{https://doi.org/10.5281/zenodo.5615978}{https://doi.org/10.5281/zenodo.5615978}.
Atomic units are used throughout.
\section{Exact Electronic Energy Landscape}
\label{sec:ExactLandscape}
\subsection{Traditional eigenvalue representation}
For a finite $N$-dimensional Hilbert space, the exact electronic wave function can
be represented using a full configuration interaction (FCI) expansion
constructed from a linear expansion of orthogonal Slater determinants as
\begin{equation}
\ket{\Psi} = \sum_{I=1}^{N} c_{I} \ket{\Phi_I}.
\end{equation}
This expansion is invariant to the particular choice of orthogonal Slater determinants $\ket{\Phi_I}$, but the
set of all excited configurations from a self-consistent HF determinant is most commonly used.\cite{SzaboBook}
Normalisation of the wave function introduces a constraint on the expansion coefficients
\begin{equation}
\sum_{I=1}^{N} \abs{c_{I}}^2 = 1,
\label{eq:ExactNorm}
\end{equation}
while the electronic energy is given by the Hamiltonian expectation value
\begin{equation}
E = \mel{\Psi}{\mathcal{H}}{\Psi} = \sum_{I,J=1}^{N} c_{I}^{*} \mel{\Phi_I}{\mathcal{H}}{\Phi_J} c_{J}^{\vphantom{*}}.
\end{equation}
The optimal coefficients are conventionally identified by solving the secular equation
\begin{equation}
\sum_{J=1}^{N} \mel{\Phi_I}{\mathcal{H}}{\Phi_J} c_{J}^{\vphantom{*}} = E\, c_{I},
\end{equation}
giving the exact ground- and excited-state energies as the eigenvalues of the Hamiltonian matrix $H_{IJ} = \mel{\Phi_I}{\mathcal{H}}{\Phi_J}$.
\subsection{Differential geometry of electronic structure theory}
In what follows, the wave function and Hamiltonian are assumed to be real-valued, although the key conclusions
can be extended to complex wave functions.
A geometric energy landscape for exact electronic structure can be constructed by representing
the wave function $\ket{\Psi}$ as an $N$-dimensional vector $\bm{c} \in \mathbb{R}^N$ with coefficients
\begin{equation}
\bm{c} =
\begin{pmatrix}
c_{1}, \cdots, c_{N}
\end{pmatrix}^{\intercal}.
\label{eq:VectorForm}
\end{equation}
The energy is then defined by the quadratic expression
\begin{equation}
E(\bm{c}) = \bm{c}^\intercal \mathbold{H}} #1 \right) \bm{c}^{\vphantom{\intercal}},
\end{equation}
where $\mathbold{H}} #1 \right)$ represents the Hamiltonian matrix in the orthogonal basis of
Slater determinants with elements $H_{IJ}~=~\mel{\Phi_I}{\mathcal{H}}{\Phi_J}$.\cite{SzaboBook}
Normalisation of the wave function is geometrically represented as
\begin{equation}
\bm{c}^{\intercal} \bm{c}^{\vphantom{\intercal}} = 1
\end{equation}
and requires the coefficient vector $\bm{c}$ to be constrained to a unit hypersphere of dimension $(N-1)$
embedded in the full $N$-dimensional space (see Fig.~\ref{fig:localGradient}).
In this representation, optimal ground and excited states are stationary points of the
electronic energy constrained to the surface of this unit hypersphere.
The stationary conditions are obtained by applying the framework
of differential geometry under orthogonality constraints, as described in Ref.~\onlinecite{Edelman1998}.
In particular, a stationary point requires that the global gradient of the energy in the full Hilbert space,
given by the vector
\begin{equation}
\frac{\partial E}{\partial \bm{c}} = 2\, \mathbold{H}} #1 \right) \bm{c},
\label{eq:GlobalGradient}
\end{equation}
has no component in the tangent space to the hypersphere.
At a point $\bm{c}$ on the surface of the hypersphere, tangent vectors $\boldsymbol{\Delta}$ satisfy
the condition\cite{Edelman1998}
\begin{equation}
\boldsymbol{\Delta}^{\intercal} \bm{c} + \bm{c}^{\intercal} \boldsymbol{\Delta} = 0.
\end{equation}
The orthogonal basis vectors that span this $(N-1)$-dimensional tangent space form the columns of
a projector into the local tangent basis,\cite{Savas2010}
denoted $\bm{c}_{\bot} \in \mathbb{R}^{N \times (N-1)}$ .
\edit{Note that $\bm{c}_{\bot}$ forms a \textit{matrix} whose columns span the tangent space, while $\bm{c}$ is
a \textit{vector} representing the current position.}
The corresponding projectors satisfy the completeness condition
\begin{equation}
\bm{c} \bm{c}^{\intercal} + \bm{c}_{\bot}^{\vphantom{\intercal}} \bm{c}_{\bot}^{\intercal} = \bm{I}_{N},
\label{eq:ProjectorExpansion}
\end{equation}
\edit{where $\bm{I}_{N}$ is the $N$-dimensional identity matrix},
and span disjoint vector spaces such that
\begin{equation}
\bm{c} \bm{c}_{\bot}^{\intercal} = \bm{c}_{\bot} \bm{c} ^{\intercal} = \bm{0}.
\label{eq:ProjectorExpansion2}
\end{equation}
The constrained energy gradient is then obtained by
projecting the global gradient Eq.~\eqref{eq:GlobalGradient} into the tangent space to give
\begin{equation}
\grad E = \bm{c}_{\bot}^\intercal \frac{\partial E}{\partial \bm{c}} = 2\, \bm{c}_{\bot}^\intercal \mathbold{H}} #1 \right) \bm{c},
\label{eq:LocalGradient}
\end{equation}
with constrained stationary points satisfying $\grad E = \bm{0}$.
Figure~\ref{fig:localGradient} illustrates this geometric relationship between the unit hypersphere,
the exact tangent space, the global gradient in the full Hilbert space, and the local gradient in the tangent space.
\begin{figure}[t]
\includegraphics[width=\linewidth]{figure1.pdf}
\caption{Geometric relationship between the position vector $\bm{c}$ (black) for a wave function constrained to the
unit hypersphere, the global energy gradient in the full Hilbert space [red; Eq.~\eqref{eq:GlobalGradient}],
and the local gradient in the tangent space [blue; Eq.~\eqref{eq:LocalGradient}]} \label{fig:localGradient}
\end{figure}
The stationary condition $\grad E = \bm{0}$ requires that the global gradient Eq.~\eqref{eq:GlobalGradient} has no
component in the tangent space.
Therefore, it can only be satisfied if $\mathbold{H}} #1 \right) \bm{c}$ is (anti)parallel to the position vector $\bm{c}$.
This condition immediately recovers the expected eigenvector expression $\mathbold{H}} #1 \right) \bm{c}_k = E_k \bm{c}_k$, where the
eigenvalue $E_k$ is the exact energy of the $k$-th excited state with coefficient vector $\bm{c}_k$.
As a result, each exact eigenstate is represented by two stationary points on the hypersphere that are related
by a sign-change in the wave function (i.e., $\pm \bm{c}_k$ or, equivalently, $\pm \ket{\Psi_k}$).
There are no other stationary points on the exact landscape.
The exact hypersphere can be compared to the constraint surface in HF theory,
where the occupied orbitals represent the current position on a Grassmann manifold and occupied-virtual orbital rotations
define the tangent basis vectors.\cite{Voorhis2002,Edelman1998}
In HF theory, the global gradient is given by $2 \bm{F}\bm{C}_\text{occ}$,
where $\bm{F}$ is the Fock matrix and
$\bm{C}_\text{occ(vir)}$ are the occupied (virtual) orbital coefficients.
Projection into the Grassmann tangent space then yields the local gradient as
$2 \bm{C}_\text{vir}^{\intercal} \bm{F}\bm{C}_\text{occ}^{\vphantom{\intercal}}$, which
corresponds to the virtual--occupied block of the Fock matrix in the molecular orbital (MO)
basis.\cite{Douady1980,Voorhis2002,Chaban1997,Burton2021}
Therefore, in both HF theory and the exact formalism, optimisation of the energy requires the off-diagonal
blocks of an effective Hamiltonian matrix to become zero such that the ``occupied--virtual'' coupling between orbitals
or many-particle states vanishes.
\subsection{Properties of exact stationary points}
\label{subsec:exactProperties}
\begin{figure*}
\includegraphics[width=\linewidth]{figure2.pdf}
\caption{\textit{Left:} Exact singlet electronic energy for \ce{H2} (STO-3G) at a bond length of $\SI{2}{\bohr}$ constrained
to the unit hypersphere.
\textit{Right:} Stereographic projection of the exact singlet energy using Eq.~\eqref{eq:stereo}.
Ground and excited states correspond to stationary points of the energy (black dots).
The $-\ket{\Psi_1}$ state is at infinity in this representation.
Each pair of stationary points is directly connected by a gradient extremal (black line) where the gradient is an eigenvector of the Hessian.
}
\label{fig:fciTopology}
\end{figure*}
Stationary points on an energy landscape can be characterised as either minima,
index-$k$ saddles, or maxima,
depending on the number of downhill directions (or negative eigenvalues of the Hessian matrix).
Following Ref.~\onlinecite{Edelman1998}, the analytic Hessian $\mathbf{Q}$ of the exact energy constrained to
the hypersphere is given by
\begin{equation}
\mathbf{Q} = \frac{\partial^2 E}{\partial \bm{c}^2} - \qty(\frac{\partial E}{\partial \bm{c}})^\intercal \bm{c}.
\end{equation}
Taking the global second derivative in the full Hilbert space
\begin{equation}
\frac{\partial^2 E}{\partial \bm{c}^2} = 2 \mathbold{H}} #1 \right)
\end{equation}
and exploiting the relationship
\begin{equation}
\qty(\frac{\partial E}{\partial \bm{c}})^\intercal \bm{c} = 2\, \bm{c}^\intercal \mathbold{H}} #1 \right) \bm{c} = 2 E
\end{equation}
allows the Hessian to be expressed as a shifted and rescaled Hamiltonian matrix
\begin{equation}
\mathbf{Q} = 2\qty(\mathbold{H}} #1 \right) - E\,\bm{I}_N).
\label{eq:GlobalHessian}
\end{equation}
Projecting into the space spanned by the tangent vectors then gives the constrained
local Hessian as an $(N-1)\times(N-1)$ matrix defined as
\begin{equation}
\widetilde{\mathbf{Q}} = 2\,\bm{c}_{\bot}^{\intercal} \qty(\mathbold{H}} #1 \right) - E\,\bm{I}_{N})\,\bm{c}_{\bot}^{\vphantom{\intercal}} ,
\label{eq:LocalHessian}
\end{equation}
\edit{where it should be remembered that $\bm{c}_\bot$ is an $N\times(N-1)$ matrix.}
This expression for the Hessian of a CI wave function is also obtained with the exponential
transformation used in second-order MC-SCF approaches.\cite{Olsen1983,RoosBook}
The constrained Hessian allows the properties of exact stationary points to be recovered.
Consider a constrained stationary point $\bm{c}_k$ corresponding to an exact
eigenstate $\ket{\Psi_k}$ with energy $E_k$.
To satisfy the completeness condition Eq.~\eqref{eq:ProjectorExpansion}, the tangent basis vectors
at this point must correspond to the remaining $(N-1)$ exact eigenstates.
The local Hessian $\widetilde{\mathbf{Q}}$ is then simply the full Hamiltonian shifted by $E_k$ and projected into
the basis of these $(N-1)$ exact eigenstates.
As a result, the Hessian eigenvalues $\lambda_i$ are directly proportional to the excitation energies, giving
\begin{equation}
\lambda_i =2(E_{i} - E_{k}) = 2\Delta E_{ik},
\end{equation}
where $E_i$ are the exact energies with $i\neq k$.
Furthermore, the corresponding eigenvectors coincide with the position vectors $\bm{c}_i$ representing the
remaining exact eigenstates $\ket{\Psi_i}$.
Remarkably, this means that only the energy and second derivatives
at an exact stationary point are required to deduce the electronic energies of the \textit{entire} system.
In other words, the full electronic energy spectrum is encoded in the local structure of the
energy landscape around a single stationary point.
Now, at the stationary point $\bm{c}_k$ with energy $E_k$, the number of exact eigenstates
that are lower in energy is equal to the excitation level $k$.
The number of negative eigenvalues $\lambda_i = 2 \Delta E_{ik}$ is then equivalent to the excitation level.
Significantly, there are only two minima on the exact energy landscape
corresponding to positive and negative sign-permutations of the exact ground state, i.e.\ $\pm \ket{\Psi_0}$,
\edit{and no higher-energy local minima.}
The $k$-th excited state forms a pair of index-$k$ saddle points that are also related by
a sign-change in the wave function.
\edit{The saddle-point nature of exact excited states was previously derived in the context of MC-SCF
theory,\cite{Olsen1982,Olsen1983,Golab1985,Golab1983} and has been described by Bacalis using local expansions
around an excited state.\cite{Bacalis2020}
In fact, the Hessian index has been suggested as a means of targeting and characterising a
particular MC-SCF excited state.\cite{Golab1983,Olsen1983}
In contrast, here the stationary properties of exact excited states have been derived using only the
differential geometry of functions under orthogonality constraints.
As will be shown later, this differential geometry also reveals the global structure of the energy
landscape and the connections between exact eigenstates.
}
These properties also apply within a particular symmetry subspace.
For example, the first excited state of a given symmetry is an index-1 saddle on the energy landscape
projected into the corresponding symmetry subspace, but may be a higher-index saddle on the full energy landscape.
\edit{For a pair of degenerate eigenstates, the corresponding stationary points will have a zero Hessian
eigenvalue and the corresponding eigenvector will interconvert the two states.
Therefore, degenerate eigenstates form a flat continuum of stationary points on the exact energy landscape
and any linear combination of the two states must also be a stationary point of the energy.}
In Fig.~\ref{fig:fciTopology}, the structure of the exact energy landscape is illustrated for the
singlet states of \ce{H2} at a bond length of $\SI{2}{\bohr}$ using the STO-3G basis set.\cite{Hehre1969}
An arbitrary spin-pure singlet wave function can be constructed as a linear combination
of singlet configuration state functions to give
\begin{equation}
\ket{\Psi} = c_1 \abs*{\sig_g \bar{\sig_g}} + \frac{c_2}{\sqrt 2} \qty(\abs*{\sig_g \bar{\sig_u}} + \abs*{\sig_u \bar{\sig_g}}) + c_3 \abs{\sig_u \bar{\sig_u}}.
\label{eq:genH2wfn}
\end{equation}
Here, $\sig_g$ and $\sig_u$ are the symmetry-adapted MOs,
and the absence (presence) of an overbar indicates an occupied high-spin (low-spin) orbital.
A stereographic projection centred on $(c_1,c_2,c_3) = (0,1,0)$ is used to
highlight the topology of the energy landscape (Fig.~\ref{fig:fciTopology}: right panel),
with new coordinates $X$ and $Y$ defined as
\begin{align}
X = \frac{c_3}{1+c_2}
\quad\text{and}\quad
Y = \frac{c_1}{1+c_2}.
\label{eq:stereo}
\end{align}
The ground and excited states (black dots) form stationary points constrained to the hypersphere
(Fig.~\ref{fig:fciTopology}: left panel) with the global minima
representing the ground state, index-1 saddles representing the first excited state, and the global maxima
representing the second excited state (Fig.~\ref{fig:fciTopology}: right panel).
At the index-1 saddle, the downhill directions connect the two sign-permutations of
the ground-state wave function, while
the two uphill directions connect sign-permutations of the second excited singlet state.
\subsection{Gradient extremals on the electronic energy surface}
\label{subsec:GradExtremal}
Energy landscapes can also be characterised by the pathways that connect stationary points.
For molecular potential energy surfaces, pathways can be interpreted as
reaction trajectories between stable molecular structures, with saddle points representing
reactive transition states.\cite{WalesBook}
However, unlike stationary points, these pathways do not have a unique mathematical definition.
On the exact electronic energy landscape, gradient extremals represent the most
obvious pathways between stationary points.
A gradient extremal is defined as a set of points where the gradient is either maximal or minimal
along successive energy-constant contour lines.\cite{Hoffman1986}
For the contour line with energy $E_\text{c}$, these points can be identified by the
constrained optimisation\cite{SzaboBook}
\begin{equation}
\begin{split}
\frac{\partial }{\partial \bm{c}}\qty[ \abs{\grad E}^2 - 2 \lambda \qty(E - E_c ) ]
=2 \qty[ \widetilde{\mathbf{Q}}\cdot \grad E - \lambda \grad E ] = 0.
\end{split}
\label{eq:SqGradCondition1}
\end{equation}
Therefore, gradient extremals are pathways where the local gradient is an eigenvector of the
local Hessian, i.e.
\begin{equation}
\widetilde{\mathbf{Q}}(\bm{c})\, \grad E(\bm{c}) = \lambda(\bm{c}) \grad E(\bm{c}).
\label{eq:SqGradCondition2}
\end{equation}
These pathways propagate away from each stationary point along the ``normal mode'' eigenvectors of the Hessian
and provide the softest or steepest ascents from a minimum.\cite{Hoffman1986}
On the exact FCI landscape, gradient extremals directly connect each pair of stationary points,
as illustrated by the black lines in Fig.~\ref{fig:fciTopology}.
Each extremal corresponds to the geodesic connecting the two stationary points along the
surface of the hypersphere.
The wave function along these pathways is only a linear combination of the
two eigenstates at each end of the path.
Therefore, gradient extremals provide a well-defined route along which a ground-state wave function can
be continuously evolved into an excited-state wave function, or vice-versa.
Furthermore, as a gradient extremal moves from the lower-energy to the higher-energy stationary point,
the corresponding Hessian eigenvalue $\lambda(\bm{c})$ changes from positive to negative.
This leads to an inflection point with $\lambda(\bm{c}) = 0$ exactly halfway along each gradient extremal
where the wave function is an equal combination of the two exact eigenstates.
These inflection points are essential for understanding the structure of the variance optimisation
landscape in Section~\ref{subsec:SquaredGradientVariance}.
\subsection{Structure of the exact variance landscape}
\label{subsec:SquaredGradientVariance}
\begin{figure*}
\includegraphics[width=\linewidth]{figure3}
\caption{Energy square-gradient for the singlet states of \ce{H2} (STO-3G) at a bond length of $\SI{2}{\bohr}$.
(a) Energy along the gradient extremal connecting the ground and first excited singlet state.
(b) Square-gradient of the energy along the gradient extremal connecting the ground and first excited singlet state.
(c) Square-gradient landscape for singlet wave functions in \ce{H2} (STO-3G), with gradient extremals (black) connecting the physical minima. The gradient extremal plotted in (b) is highlighted in red.
}
\label{fig:fciVariance}
\end{figure*}
The accuracy of a general point on the exact energy landscape can
be assessed using the square-magnitude of the local gradient, defined using Eq.~\eqref{eq:LocalGradient} as
\begin{equation}
\abs{\grad E(\bm{c})}^2 = 4\, \bm{c}^\intercal \mathbold{H}} #1 \right) \bm{c}_{\bot}^{\vphantom{\intercal}} \bm{c}_{\bot}^\intercal \mathbold{H}} #1 \right) \bm{c}.
\label{eq:GradMag1}
\end{equation}
Since all stationary points are minima of the squared-gradient
with $\abs{\grad E(\bm{c})}^2=0$, minimising this objective function has been proposed
as a way of locating higher-index saddle points in various contexts.\cite{Broderix2000,Angelani2000,Hait2020}
For the electronic structure problem, exploiting the relationship between the tangent-
and normal-space projectors [Eq.~\eqref{eq:ProjectorExpansion}] allows Eq.~\eqref{eq:GradMag1} to be
expressed using only the position vector $\bm{c}$ as
\begin{equation}
\abs{\grad E}^2 = 4\, \bm{c}^\intercal \mathbold{H}} #1 \right) \qty(\bm{I}_{N} - \bm{c} \bm{c}^\intercal) \mathbold{H}} #1 \right) \bm{c}.
\label{eq:GradMag2}
\end{equation}
Further expanding this expression gives
\begin{equation}
\abs{\grad E}^2 = 4\qty( \bm{c}^\intercal \mathbold{H}} #1 \right)^2 \bm{c} - \qty(\bm{c}^\intercal \mathbold{H}} #1 \right) \bm{c})^2),
\label{eq:GradMag3}
\end{equation}
which can be recognised as four times the Hamiltonian variance, i.e.\
\begin{equation}
\abs{\grad E}^2 = 4\mel*{\Psi}{(\mathcal{H} - E)^2}{\Psi}.
\end{equation}
While variance optimisation has previously inspired the development of excited-state variational
principles,\cite{Messmer1969,Shea2017,PinedaFlores2019,Cuzzocrea2020,Ye2017,Ye2019} these approaches
have generally been motivated by the fact that exact eigenstates of $\mathcal{H}$ also have zero variance.
In contrast, the relationship between the variance and $\abs{\grad E}^2$ provides a purely geometric
motivation behind searching for excited states in this way, derived from the structure of the exact energy landscape.
Hait and Head-Gordon alluded to a relationship of this type by noticing similarities
between the equations for SCF square-gradient minimisation and optimising the SCF variance.\cite{Hait2020}
By connecting the squared-gradient of the energy to the variance, the exact energy
landscape can be used to deduce the structure of the variance optimisation landscape.
Exact eigenstates form minima on the square-gradient landscape with $\abs{\grad E}^2 = 0$.
However, the square-gradient can have additional non-zero stationary points
corresponding to local minima, higher-index saddle points, or local maxima.\cite{Doye2002,Doye2003}
These ``non-stationary'' points do not represent stationary points of the energy, but
they provide important information about the structure of the square-gradient landscape away from exact
eigenstates.
In particular,
a stationary point of
$\abs{\grad E}^2$ with $\grad E \neq \bm{0}$ can only occur when the local gradient is an eigenvector
of the Hessian with a zero eigenvalue, $\widetilde{\mathbf{Q}}(\bm{c})\, \grad E(\bm{c}) = \bm{0}.$\cite{Doye2002,Doye2003}
Non-stationary points therefore occur on the gradient extremals
described in Section~\ref{subsec:GradExtremal} and correspond to the inflection points
exactly halfway between each pair of eigenstates.
Since gradient extremals only connect two eigenstates, the value of $\abs{\grad E}^2$ at these
non-stationary points can be obtained by parametrising the wave function as
\begin{equation}
\ket{\Psi(\theta)} = \cos \theta \ket*{\Psi_i} + \sin \theta \ket*{\Psi_j}.
\end{equation}
The energy and square-gradient are then given by
\begin{subequations}
\begin{align}
E(\theta) &= E_i + \Delta E_{ji}\, \sin^2 \theta,
\\
\abs{\grad E(\theta)}^2 &= \Delta E_{ji}^2\, \sin^2 2\theta,
\end{align}
\end{subequations}
where $0 \le \theta \le \pi / 2$ and $\Delta E_{ji} = E_j - E_i$.
There are only two stationary points of the energy along each pathway, corresponding to
$\abs{\grad E}^2 = 0$ at $\theta = 0$ and $\pi / 2$, as illustrated for the gradient extremal
connecting the ground and first excited singlet states of \ce{H2} (STO-3G) in Fig.~\ref{fig:fciVariance}a.
In contrast, the square-gradient has an additional stationary point at the inflection point $\theta = \pi/4$ with
$\abs{\grad E}^2 = \Delta E_{ji}^2$ (see Fig.~\ref{fig:fciVariance}b).
This point corresponds to an unphysical maximum of
$\abs{\grad E}^2 $ along the gradient extremal and the height of this barrier depends on the square of the energy
difference between the two states.
Therefore, the exact square-gradient (or variance) landscape
contains exact minima separated by higher-variance stationary points that form barriers at the inflection points of the energy,
with the height of each barrier directly proportional to
the square of the energy difference between the connected eigenstates.
The lowest square-gradient barrier always connects states that are adjacent in energy
to form an index-1 saddle point, while barriers connecting states that are not adjacent in energy
form higher-index saddle points of $\abs{\grad E}^2$. %
This structure of the exact square-gradient landscape is illustrated for the singlet states of
\ce{H2} in Fig.~\ref{fig:fciVariance}c.
Connecting the variance to the exact energy square-gradient reveals that the general structure of the
variance optimisation landscape is universal and completely determined by the energy difference between
exact eigenstates.
Systems with very small energy gaps will have low barriers between exact variance minima,
while systems with well-separated energies will have high-variance barriers.
\edit{This structure may also play a role in explaining the convergence behaviour of variance minimisation
approaches, as discussed in Section~\ref{sec:Discussion}.}
\section{Understanding Approximate Wave Functions}
\label{sec:ApproximateMethods}
\subsection{Differential geometry on the exact energy landscape}
\label{subsec:RelationshipToExactEnergy}
\begin{figure*}
\includegraphics[width=\linewidth]{figure4}
\caption{The space of possible RHF wave functions forms a one-dimensional submanifold
(right panel; red curve) on the exact singlet energy surface for \ce{H2} (STO-3G) with a bond length of $\SI{3}{\bohr}$.
Multiple RHF solutions (left panel) correspond to constrained stationary points on the RHF manifold
(right panel; red dots).
\edit{The sign-permuted RHF wave functions are denoted by a dashed red curve (red panel).}
}
\label{fig:fci_rhf_Topology}
\end{figure*}
Any normalised wave function approximation $\ket*{\tilde{\Psi}(\bm{t})}$ with variational parameters $\bm{t}$
can be represented as a point on the exact hypersphere using a linear expansion in the many-particle basis
\begin{equation}
\ket*{\tilde{\Psi}(\bm{t})} = \sum_{I=1}^{N} \tilde{c}_I(\bm{t}) \ket{\Phi_I}.
\end{equation}
Like the exact wave function, this expansion is invariant to the particular choice of orthogonal basis determinants.
Since the number of parameters $t_i$ is generally smaller than the Hilbert space size,
the approximate wave functions form a constrained submanifold of the exact hypersphere.
The structure of this approximate submanifold is implicitly defined by the mathematical form of the parametrisation.
Geometrically, the approximate energy is given as
\begin{equation}
E(\bm{t}) = \tilde{\bm{c}}(\bm{t})^\intercal\, \mathbold{H}} #1 \right)\, \tilde{\bm{c}}(\bm{t}),
\end{equation}
and the constrained local gradient is defined as
\begin{equation}
\qty(\widetilde{\grad} E)_i
=
\frac{\partial E(\bm{t})}{\partial t_i} = 2\, \qty( \frac{\partial\tilde{\bm{c}}(\bm{t})}{\partial t_i})^\intercal \mathbold{H}} #1 \right) \, \tilde{\bm{c}}(\bm{t}).
\label{eq:localGradientTerms}
\end{equation}
Here, the partial derivatives of the coefficient vector define the local
tangent basis of the approximate submanifold, representing the ket vectors
\begin{equation}
\ket*{\eta_i} = \pdv{t_i}\ket*{\tilde{\Psi}(\bm{t})}.
\label{eq:TangetVectors}
\end{equation}
In analogy with the exact wave function, the approximate local gradient
corresponds to the global gradient [Eq.~\eqref{eq:GlobalGradient}] projected into the space
spanned by the approximate tangent vectors [Eq.~\eqref{eq:TangetVectors}].
Optimal stationary points of the approximate energy then occur
when $\widetilde{\grad} E = \bm{0}$.
The HF approach illustrates how well-known approximate stationary
conditions can be recovered with this geometric perspective.
Although HF theory is usually presented as an iterative self-consistent approach,\cite{Roothaan1951,Hall1951}
the HF wave function can also be parametrised using an exponential transformation of a reference determinant to give\cite{Thouless1960,Douady1980}
\begin{equation}
\ket{\Phi(\bm{\kappa})} = \exp(\hat{\kappa})\ket{\Phi_0}.
\end{equation}
Here, $\hat{\kappa}$ is a unitary operator constructed from closed-shell single excitations and de-excitations,
represented in second-quantisation as\cite{HelgakerBook}
\begin{equation}
\hat{\kappa} = \sum_{ai} \kappa_{ai} \qty(
a_{a}^{\dagger} a_{i}^{\vphantom{\dagger}} - a_{i}^{\dagger} a_{a}^{\vphantom{\dagger}}
)
\label{eq:singlesExp}
\end{equation}
where $\kappa_{ai}$ are the variable parameters.
The tangent vectors at $\bm{\kappa} = \bm{0}$ correspond to the singly-excited configurations, i.e.
\begin{equation}
\ket{\eta_{ai}}
= \left. \frac{\partial\ket{\Phi(\bm{\kappa})}}{\kappa_{ai}} \right\rvert_{\bm{\kappa}=\bm{0}}
= \ket{\Phi_i^a}.
\end{equation}
These approximate tangent vectors span a subspace of the exact tangent space on the full hypersphere.
Using Eq.~\eqref{eq:localGradientTerms}, the local HF gradient is then given by\cite{Douady1980,Voorhis2002}
\begin{equation}
\qty(\widetilde{\grad} E)_{ai}
= 2 \mel*{\Phi_i^a}{\mathcal{H}}{\Phi_0} = 2 F_{ai},
\end{equation}
which corresponds to twice the virtual-occupied components of the Fock matrix in the MO basis.
As expected, the stationary condition $\widetilde{\grad} E =\bm{0}$ recovers Brillouin's theorem for HF
convergence.\cite{SzaboBook}
\subsection{Multiple Hartree--Fock solutions}
\label{subsec:MultipleSolutions}
\begin{figure}[b!]
\includegraphics[width=\linewidth]{figure5.pdf}
\caption{Mapping between the RHF submanifold (solid red line) and the exact
singlet energy surface for \ce{H2} (STO-3G) at (a) the equilibrium bond length $\SI{1.437707}{\bohr}$ and
(b) a stretched geometry $\SI{3}{\bohr}$.
\edit{Sign-permuted RHF wave functions are denoted by dashed red curves.}
}
\label{fig:fci_rhf_topo_comparison}
\end{figure}
Although the HF wave function has fewer parameters than the exact wave function,
there can be more HF stationary points than exact eigenstates.\cite{Stanton1968,Burton2018}
For example, in dissociated \ce{H2} with a minimal basis set, there are four closed-shell restricted HF (RHF) solutions
and only three exact singlet states.\cite{Burton2018}
In contrast to the exact eigenstates, there can also be multiple HF solutions with the same Hessian index,
although this index generally increases with energy.\cite{Burton2018}
Furthermore, HF solutions do not necessarily exist for all molecular geometries and can disappear at so-called
``Coulson--Fischer'' points.\cite{Coulson1949,Slater1951,Thom2008,Huynh2019,Burton2018,Burton2021}
These phenomena can all be understood through the geometric mapping between the approximate HF
submanifold and the exact energy landscape.
Consider the RHF approximation for the singlet states of \ce{H2} (STO-3G).
Only two RHF solutions exist at the equilibrium geometry, while an additional higher-energy
pair of degenerate solutions emerge in the dissociation limit, as shown in the left panel of Fig.~\ref{fig:fci_rhf_Topology}
(see Ref.~\onlinecite{Burton2018} for further details).
In this system, the RHF submanifold forms a continuous one-dimensional subspace of the exact
energy surface, illustrated by the red curve in Fig.~\ref{fig:fci_rhf_Topology} (right panel).
This submanifold includes \emph{all possible} closed-shell Slater determinants for the system
and is fixed by the wave function parametrisation.
Approximate solutions then correspond to constrained stationary points of the energy along
the RHF submanifold, which occur when the global energy gradient has no component
parallel to the red curve.
The existence and properties of these solutions is completely determined by the mapping between
the RHF submanifold and the exact energy landscape.
At bond lengths near the equilibrium structure of \ce{H2} (STO\nobreakdash-3G), the RHF submanifold extends relatively close to the exact
global minimum and maximum, as illustrated in Fig.~\ref{fig:fci_rhf_topo_comparison}a.
This mapping results in only two constrained stationary
points, the global minimum and maximum of the RHF energy, which correspond to
the symmetry-pure $\sig_g^2$ and $\sig_u^2$ configurations respectively.
Notably, the RHF $\sig_u^2$ global maximum represents a doubly-excited
state that cannot be accurately described by TD-HF or CIS.\cite{Burke2005}
However, the RHF submanifold cannot get sufficiently close to the exact open-shell
singlet state to provide a good approximation, and there is no stationary point representing this single excitation.
\begin{figure}[b!]
\includegraphics[width=\linewidth]{figure6}
\caption{Comparison of the energy and SCF Hessian index computed for the orbital-optimised excited configurations
of \ce{H2} (cc-pVDZ\cite{Dunning1989}) at $R=\SI{1.437707}{\bohr}$ using unrestricted HF.}
\label{fig:HessCompare}
\end{figure}
As the \ce{H-H} bond is stretched towards dissociation, the exact energy landscape on the
hypersphere changes while the RHF submanifold remains fixed by the functional form of the approximate wave function.
Therefore, the exact energy evolves \textit{underneath} the RHF submanifold, creating changes in the approximate energy
that alter the properties of the constrained stationary points.
For example, at a sufficiently large bond length, the RHF submanifold no longer
provides an accurate approximation to the exact global maximum and instead encircles it to give two local
maxima and a higher-energy local minimum of the RHF energy (Fig.~\ref{fig:fci_rhf_topo_comparison}b).
These local maxima represent the spatially-symmetry-broken RHF solutions that tend towards to the ionic
dissociation limit (left panel in Fig.~\ref{fig:fci_rhf_Topology}), while the higher-energy local minimum represents the $\sig_u^2$ configuration.
Combined with the global minimum, this gives a total of four RHF solutions,
in contrast to only three exact singlet states.
Consequently, we find that the number of RHF solutions can exceed the number of exact eigenstates
because the RHF submanifold is a highly constrained non-linear subspace of the exact energy landscape.
Furthermore, it is the structure of this constrained subspace that creates a high-energy local minimum
at dissociation, while the exact energy landscape has no local minima at any geometry.
\begin{figure*}[htb!]
\includegraphics[width=0.8\linewidth]{figure7}
\caption{
\textit{Left:} RHF (solid and dashed red curves) and ESMF (blue mesh) constrained submanifolds superimposed on a
stereographic projection [see Eq.~\eqref{eq:stereo}]
of the exact singlet energy landscape for \ce{H2} (STO-3G) at a bond length
of $R=\SI{1.437707}{\bohr}$.
Exact stationary points and RHF solutions occur at the black and red dots respectively.
\textit{Right:}
The ESMF energy for singlet \ce{H2} (STO-3G) as a function of the wave function parameters
[see Eq.~\eqref{eq:OOcisH2}].
Black and grey symbols indicate stationary points of the ESMF or RHF energy respectively, and the RHF submanifold corresponds to
$\theta = 0$ (dashed black line).}
\label{fig:ooCIS}
\end{figure*}
Finally,
for real-valued HF wave functions, the Hessian of the approximate energy is computed using the second derivatives\cite{Seeger1977}
\begin{equation}
\frac{\partial^2 E(\bm{t}) }{\partial t_{ia} \partial t_{jb}} = 2\, \Big(\mel*{\Phi_{i}^{a}}{\mathcal{H} - E}{\Phi_{j}^{b}} + \mel*{\Phi}{\mathcal{H} - E}{\Phi_{ij}^{ab}} \Big),
\label{eq:OrbHess}
\end{equation}
where now open-shell orbital rotations are now allowed.
To investigate how the HF Hessian index changes with energy,
the unrestricted excited configurations obtained from the ground-state RHF orbitals of
\ce{H2} (cc-pVDZ\cite{Dunning1989}) at
$R~=~\SI{1.437707}{\angstrom}$ were optimised using the initial maximum overlap method\cite{Barca2018}
in Q-Chem 5.4.\cite{QChem54}
The corresponding Hessian indices are plotted against the optimised energy in Fig.~\ref{fig:HessCompare}.
Similarly to the exact eigenstates, there is a general increase in the Hessian index at higher energies.
However, unlike the exact eigenstates, the approximate Hessian index does not increase monotonically with the energy.
These results strengthen the conclusion of Ref.~\onlinecite{Burton2021} that approximate HF excited states are generally
higher-index saddle points of the energy.
\subsection{Excited-state mean-field theory}
\label{subsec:CISandPCCD}
While multiple HF solutions are relatively well understood, the energy landscape of orbital-optimised post-HF
wave functions remains less explored.
The simplest excited-state extension of HF theory is the CIS wave function,\cite{Foresman1992,Dreuw2005}
constructed as a linear combination of singly-excited determinants.
\edit{CIS generally provides a qualitatively correct description of singly excited states,
but it is less reliable for multiconfigurational or charge transfer excitations.\cite{Dreuw2005}
The systematic overestimate of CIS for charge transfer excitations can be attributed to the absence of
orbital relaxation effects.\cite{Subotnik2011}}
Therefore, to improve this description, the orbitals and CI coefficients can be
simultaneously optimised to give the state-specific excited-state mean-field (ESMF) wave function, defined for singlet
states as\cite{Shea2018}
\begin{equation}
\ket{\Psi_{\text{ESMF}}} = \exp(\hat{\kappa}) \qty[c_0
+ \frac{1}{\sqrt{2}} \sum_{ia} c_{ia}
\qty( a^{\dagger}_{a} a^{\vphantom{\dagger}}_{i}
+ a^{\dagger}_{\bar{a}} a^{\vphantom{\dagger}}_{\bar{i}} ) ] \ket{\Phi_0} .
\end{equation}
Here, the reference determinant is retained in the expansion and orbital rotations
are parametrised by the unitary rotation defined in
Eq.~\eqref{eq:singlesExp}.
In recent years, efficient optimisation of this \textit{ansatz} has been
successfully applied to charge transfer and core excitations.\cite{Shea2018,Zhao2020a,Hardikar2020,Garner2020}
Alternative approaches to include orbital relaxation effects in CIS using perturbative corrections
have also been investigated.\cite{Liu2012,Liu2013,Liu2014}
Geometrically, the single excitations for any closed-shell reference determinant
define the tangent vectors to the RHF submanifold.
Therefore, the orbital-optimised ESMF submanifold contains the RHF wave functions and
all points that lie in the combined tangent spaces of the RHF submanifold,
as illustrated for the singlet states of \ce{H2} (STO-3G) in Fig.~\ref{fig:ooCIS} (left panel).
Like multiple RHF solutions, state-specific ESMF solutions correspond to stationary
points of the energy constrained to the ESMF wave function manifold.
The ESMF wave function for singlet \ce{H2} (STO-3G) can be constructed by parametrising the occupied
and virtual molecular orbitals with a single rotation angle $\phi$ as
\begin{subequations}
\begin{align}
\psi_1(\bm{r}) &= \cos \phi\, \sig_g(\bm{r}) + \sin \phi\, \sig_u(\bm{r}),
\\
\psi_2(\bm{r}) &= \cos \phi\, \sig_u(\bm{r}) -\sin \phi\, \sig_g(\bm{r}).
\end{align}
\end{subequations}
and defining the normalised singlet CI expansion with a second rotation angle $\theta$ to give
\begin{equation}
\ket{\Psi_{\text{ESMF}}} = \cos \theta\, \abs*{\psi_1 \bar{\psi}_1 } + \sin \theta \, \frac{\abs*{\psi_1 \bar{\psi}_2 } + \abs*{\psi_2 \bar{\psi}_1 } }{\sqrt{2}} .
\label{eq:OOcisH2}
\end{equation}
The corresponding energy landscape at the equilibrium bond length $R=\SI{1.437707}{\bohr}$ is
shown in Fig.~\ref{fig:ooCIS} (right panel) as a function of $\theta$ and $\phi$,
with the RHF approximation indicated by the dashed black line at $\theta=0$.
Stationary points representing the exact open-shell singlet state occur at $(\theta, \phi) = (\pm\frac{\pi}{2},0)$ and
$(\pm \frac{\pi}{2},\pm \frac{\pi}{2})$, denoted by \edit{black stars}, with optimal orbitals that correspond to $\sig_g(\bm{r})$ and $\sig_u(\bm{r})$.
In common with the exact energy landscape, these open-shell singlet solutions form index-1
saddle points of the singlet energy.
On the other hand, the local maximum representing the double excitation has no contribution from the single excitations and
reduces to the closed-shell $\sig_u^2$ RHF solution \edit{(grey diamonds)}.
This lack of improvement beyond RHF can be understood from the structure of the singlet ESMF submanifold:
the exact double excitation is encircled by the RHF submanifold and therefore cannot be reached by any of the
tangent spaces to the RHF wave function (Fig.~\ref{fig:ooCIS}: left panel).
However, the most surprising observation from Fig.~\ref{fig:ooCIS} is not the higher-energy
stationary points of the ESMF energy, but the location of the global minimum.
Counter-intuitively, the RHF ground state becomes an index-1 saddle point of the ESMF energy \edit{(grey circles)} and there is a lower-energy
solution at $(\theta, \phi) = (\pm 0.5026,\mp 0.3304)$ that corresponds to the exact ground state \edit{(black squares)}.
This lower-energy solution occurs because rotating the orbitals away from an optimal HF solution breaks
Brillioun's condition and introduces new coupling terms between the reference and singly-excited configurations
that allow the energy to be lowered below the RHF minimum.
Since the RHF ground state is stationary with respect to both orbital rotations and the introduction of single excitations,
this cooperative effect can only occur when the orbital and CI coefficients are optimised simultaneously.
Furthermore, the combined orbital and CI Hessian is required to diagnose the RHF ground state as a saddle point of the ESMF energy,
highlighting the importance of considering the full parametrised energy landscape.
The existence of an ESMF global minimum below the RHF ground state challenges
the idea that CIS-based wave functions are only useful for approximating excited states.
From a practical perspective, current ESMF calculations generally underestimate excitation energies because the
multiconfigurational wave function used for the excited state can capture some electron correlation,
while the single-determinant RHF ground state remains completely uncorrelated.\cite{Shea2018}
\edit{Although CIS is often described as an uncorrelated excited-state theory, the wave function
is inherently multi-configurational and becomes correlated
when the first-order density matrix is not idempotent.\cite{Surjan2007}
These circumstances generally correspond to excitations with more than one dominant
natural transition orbital.\cite{Plasser2016}
Therefore, it is not too surprising that the ESMF global minimum can provide a
correlated representation of the ground state.}
As a result, using state-specific ESMF wave functions for both the ground and excited states
should provide a more balanced description of an electronic excitation.
Since the global minimum is exact across all \ce{H2} bond lengths,
orbital-optimised ESMF may also provide an alternative reference wave function for capturing
static correlation in single-bond breaking processes.
\edit{The success of this \textit{ansatz} can be compared to the spin-flip CIS (SF-CIS) approach, where the
CI expansion is constructed using spin-flipping excitations from a high-spin reference determinant.\cite{Krylov2001}
SF-CIS gives an accurate description of the \ce{H2} ground state because single spin-flip excitations
from an open-shell $\sig_g \sig_u$ reference produce both the $\sig_g^2$ and $\sig_u^2$ configurations.
In contrast, the ESMF global minimum contains a closed-shell reference determinant
with symmetry-broken orbitals, as shown in Fig.~\ref{fig:esmf_orbs}.
These orbitals resemble the $\sig_g$ and $\sig_u$ MOs at short geometries [Fig.~\ref{fig:esmf_orbs}(a)--(b)] where the reference determinant $|\psi_1 \bar{\psi}_1|$
dominates the ESMF wave function.
As the bond is stretched, the optimised orbitals localise on opposite H atoms to give an ionic reference
state [Fig.~\ref{fig:esmf_orbs}(e)--(f)] and the single excitations correspond to the localised configurations required for the diradical ground state.
}
\begin{figure}[b!]
\setlength\fboxrule{1pt}
\includegraphics[width=\linewidth]{figure8}
\caption{%
\edit{%
Spatial orbitals for the optimised ESMF ground state of \ce{H2} (STO-3G) at bond lengths of
(a)--(b) $\SI{1.437707}{\bohr}$,
(c)--(d) $\SI{3.0}{\bohr}$, and
(e)--(f) $\SI{6.0}{\bohr}$,
plotted with an isosurface value of $\pm0.05$.
The reference configuration corresponds to $|\psi_1 \bar{\psi}_1|$ with the ESMF wave function defined in Eq.~\eqref{eq:OOcisH2}.}
}
\label{fig:esmf_orbs}
\end{figure}
\begin{figure*}
\setlength\fboxrule{1pt}
\includegraphics[width=\linewidth]{figure9}
\caption{(a) Exact square-gradient landscape for singlet \ce{H2} (STO-3G) at a bond length of $\SI{3}{\bohr}$.
Restricted $\sigma$-SCF solutions identified using variance optimisation are equivalent to
constrained stationary points (blue dots) on the single-determinant subspace (blue curve) and its sign-related
copy (dashed blue curve).
(b) RHF energy as a function of the single orbital rotation angle $\phi$ [see Eq.~\eqref{eq:orbParam}].
(c) Exact square-gradient of RHF wave functions with stationary points corresponding to $\sigma$-SCF solutions.
(d) Square-magnitude of the constrained RHF energy gradient, with minima corresponding to RHF energy stationary points.
}
\label{fig:fci_rhf_variance}
\end{figure*}
\subsection{Unphysical solutions in variance optimisation}
\label{subsec:VarianceChallenges}
Approximate state-specific variance minimisation can also be considered as an optimisation of
the exact variance constrained to the approximate wave function submanifold.
Variance minimisation is increasingly being applied to target excited states because it
turns the optimisation of higher-energy stationary points into a minimisation
problem\cite{Messmer1969,Shea2017,PinedaFlores2019,Ye2017,Ye2019,David2021}
and is easily applied for correlated wave functions using VMC.\cite{Otis2020,Cuzzocrea2020}
In practice, these algorithms often use the folded-spectrum objective function\cite{Wang1994}
\begin{equation}
\Omega = \frac{\mel*{\tilde{\Psi}}{(\mathcal{H} - \omega)^2}{\tilde{\Psi}}}{\braket*{\tilde{\Psi}}{\tilde{\Psi}}}
\label{eq:omegaTarget}
\end{equation}
to target an excited state with an energy near $\omega$, before self-consistently updating
$\omega$ until a variance stationary point is reached with $\omega =E$.\cite{Ye2017,Ye2019,Cuzzocrea2020}
However, the structure of the variance landscape for approximate wave functions, and the properties
of its stationary points, are relatively unexplored.
Since the variance is equivalent to the exact square-gradient $\abs{\grad E}^2$, this landscape
can be investigated using the mapping between an approximate wave function and the exact
square-gradient landscape derived in Sec.~\ref{subsec:SquaredGradientVariance}.
Consider the RHF approximation in \ce{H2} (STO-3G) with the
doubly-occupied orbital parametrised as
\begin{equation}
\psi(\bm{r}) = \cos \phi \, \sig_g(\bm{r}) + \sin \phi \,\sig_u(\bm{r}).
\label{eq:orbParam}
\end{equation}
Variance optimisation for HF wave functions has been developed by Ye \etal\ as the iterative $\sigma$-SCF
method, which uses a variance-based analogue of the Fock matrix.\cite{Ye2017,Ye2019}
In Figure~\ref{fig:fci_rhf_variance}a, the RHF submanifold is shown as a subspace of the
exact singlet square-gradient landscape at $R=\SI{3}{\bohr}$, with
optimal $\sigma$-SCF solutions corresponding to the constrained stationary points.
Since the RHF approximation cannot reach any exact eigenstates in this system, the square-gradient is non-zero
for all RHF wave functions and there is no guarantee that the $\sigma$-SCF solutions
coincide with stationary points of the constrained energy.
This feature is illustrated in Figs.~\ref{fig:fci_rhf_variance}b and \ref{fig:fci_rhf_variance}c, where the energy and
square-gradient are compared for the occupied orbital defined in Eq.~\eqref{eq:orbParam}.
There are three constrained minima of the square-gradient for these RHF wave functions.
The first, at $\phi=0$ corresponds to the $\sig_g^2$ RHF ground state, while the other two represent
ionic configurations that are similar, but not identical, to the local maxima of the RHF energy.\cite{Ye2017}
However, the $\sig_u^2$ configuration, which forms a local minimum of the RHF energy, becomes a constrained
local maximum of the energy square-gradient.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{figure10}
\caption{Energy of restricted $\sigma$-SCF solutions corresponding
to constrained stationary points of the variance in \ce{H2} (STO-3G),
including local minima (blue) and maxima (green).
Exact singlet energies are shown for comparison (black dashed).}
\label{fig:h2_sigmaSCF}
\end{figure}
Energies corresponding to the complete set of RHF variance stationary points
for \ce{H2} (STO-3G) are shown in Fig.~\ref{fig:h2_sigmaSCF}, with variance minima and maxima denoted
by blue and green curves respectively.
These solutions closely mirror the low-energy $\sigma$-SCF states identified using the 3-21G basis set in Ref.~\onlinecite{Ye2017}.
Despite becoming a local maximum of the variance at large bond lengths, the $\sigma$-SCF approach
identifies the $\sig_u^2$ solution at all geometries.\cite{Ye2017}
Therefore, iterative methods such as $\sigma$-SCF must be capable of converging onto higher-index stationary points
of the variance.
However, not all higher-index stationary points correspond to physically meaningful solutions.
For example, an additional degenerate pair of $\sigma$-SCF maxima can be found at all bond lengths in \ce{H2},
with an energy that lies between the $\sig_g^2$ and $\sig_u^2$ solutions.
These unphysical maxima exist where the
RHF manifold passes over the square-gradient barriers created by non-stationary points on the
exact energy landscape (see Fig.~\ref{fig:fci_rhf_variance}a).
The high non-convexity of the exact square-gradient landscape means that these unphysical
higher-index stationary points of the variance are likely to be very common for approximate wave functions.
As an alternative to variance minimisation, excited state-specific wave functions can
be identified by \edit{directly searching for points where the approximate local gradient
becomes zero $\widetilde{\grad}E = \bm{0}$.
For example, Shea and Neuscamman introduced an approach that modifies the objective function Eq.~\eqref{eq:omegaTarget}
using Lagrange multipliers to
ensure that the approximate local gradient vanishes at a solution.\cite{Shea2018}
Alternatively, Hait and Head-Gordon proposed the square-gradient minimisation (SGM) algorithm that directly minimises
the square-magnitude of the local electronic gradient $\abs*{\widetilde{\grad}E}^2$.\cite{Hait2020}
While the SGM approach is closely related to minimising the exact variance,} all stationary points
of the approximate energy now become minima of local square-gradient
with $\abs*{\widetilde{\grad}E}^2=0$, as shown for the RHF wave functions of \ce{H2} (STO-3G) in Fig.~\ref{fig:fci_rhf_variance}d.
Unphysical square-gradient stationary points, such as those resembling the RHF variance
maxima in Fig.~\ref{fig:fci_rhf_variance}c, can then be easily identified with $\abs*{\widetilde{\grad}E}^2 \neq 0$.
However, since the approximate energy can have more stationary points than the exact energy, there will
be more non-stationary points of $\abs*{\widetilde{\grad}E}^2$ corresponding to inflection points between
stationary states of the approximate energy (\textit{cf}.\ Fig.~\ref{fig:fci_rhf_variance}a and \ref{fig:fci_rhf_variance}d).
These additional non-stationary points will make the approximate $\abs*{\widetilde{\grad}E}^2$ landscape even
more non-convex than the constrained variance landscape, which can make numerical
optimisation increasingly difficult.
\section{Implications for Optimisation Algorithms }
\label{sec:Discussion}
\edit{We have seen that the exact energy landscape for real wave functions has only two minima,
corresponding to negative and positive sign-permutations of the exact ground state.
In addition, there is only one sign-permuted pair of index-$k$ saddle points corresponding to the $k$-th excited state.
On the contrary, approximate methods may have higher-energy local minima and multiple index-$k$ saddle points,
although the maximum Hessian index is dictated by the number of approximate wave function parameters.
These higher-energy local minima result from the wave function constraints introduced by lower-dimensional approximations.
}
The structure of the exact energy landscape highlights the importance of developing excited state-specific
algorithms that can converge onto arbitrary saddle points of the energy.
Any type of stationary point can be identified using
iterative techniques that modify the SCF procedure,
including the maximum overlap method\cite{Gilbert2008,Besley2009}
and state-targeted energy projection.\cite{CarterFenk2020}
In addition, modified quasi-Newton optimisation
of the SCF energy\cite{Levi2020,Levi2020a} should perform well,
while state-specific CASSCF\cite{Tran2019,Tran2020}
may also benefit from similar second-order optimisation.
Alternatively, modified eigenvector-following may allow excited states with a particular Hessian
index to be targeted.\cite{Doye2002,Wales2003,Burton2021}
However, methods that search for local minima of the
energy, including SCF metadynamics\cite{Thom2008} and
direct minimisation,\cite{Voorhis2002} will perform less well
for excited states and are better suited to locating the global minimum or symmetry-broken solutions.
\edit{In addition, the structure of the energy square-gradient landscape elucidates
the challenges faced by variance optimisation approaches.
Each exact eigenstate forms a minimum of the variance and minima adjacent in energy are separated
by an index-1 saddle point with height proportional to the square of the corresponding energy difference.}
\edit{Low barriers on the exact variance landscape offer a new perspective on the
convergence drift observed in excited-state VMC calculations.\cite{Cuzzocrea2020,Otis2020}
In Ref.~\onlinecite{Cuzzocrea2020}, variance optimisation was found to
drift away from the intended target state defined by the initial guess,
passing through eigenstates sequentially in energy until converging onto the state with the lowest variance.
By definition, a deterministic minimisation algorithm cannot climb over a barrier to escape a variance minimum.
However, the statistical uncertainty of stochastic VMC calculations means that they \textit{can} climb a variance
barrier if the height is sufficiently low.
Essentially, the variance barrier becomes ``hidden'' by statistical noise.
The index-1 variance saddle points connecting states adjacent in energy may then
explain why the optimisation drifts through eigenstates sequentially in energy\cite{Cuzzocrea2020} and
why this issue is more prevalent in systems with small energy gaps.\cite{Otis2020}
Alternatively, the VMC wave function may simply not extend far
enough into the exact variance basin of attraction of the target state to create an approximate local minimum.
However, the use of highly-sophisticated wave functions in Ref.~\onlinecite{Cuzzocrea2020} would suggest that
this latter explanation is unlikely.
}
\edit{An additional concern is the presence of unphysical local variance minima or higher-index stationary
points that occur when the exact variance is constrained to an approximate wave function manifold.
For minimisation algorithms such as VMC or generalised variational principles, the presence of many
higher-index saddle points may increase the difficulty of convergence.
There is also a risk of getting stuck in a spurious local minimum on the constrained manifold which does
not correspond to a physical minimum on the exact variance landscape.
In contrast, iterative self-consistent algorithms such as $\sigma$-SCF\cite{Ye2017,Ye2019} are capable of converging onto
higher-index stationary points and it may be difficult to establish the physicality of these solutions.
Therefore, iterative variance optimisation requires careful analysis to ensure the physicality of solutions,
for example by using sufficiently accurate initial guesses.
}
\edit{Finally, although not considered in this work, the generalised variational principle developed by Neuscamman and co-workers includes Lagrange multipliers
to simultaneously optimise multiple objective functions that target a particular excited eigenstate.\cite{Shea2018,Shea2020,Hanscam2021}
The combined Lagrangian may include functionals that target a particular energy (such as Eq.~\eqref{eq:omegaTarget}),
the square-magnitude of the local gradient, orthogonality to a nearby state, or a desirable
dipole moment.\cite{Hanscam2021}
This approach is particularly suited to systems where there is a good initial guess for the excited state or its properties,
and a good choice of objective functions can significantly accelerate numerical convergence.\cite{Shea2020}
These constraints may also help to prevent the drift of stochastic variance optimisation algorithms by
increasing the barrier heights between minima with the correct target properties.
}
\section{Concluding Remarks}
\label{sec:ConcludingRemarks}
This contribution has introduced \edit{a geometric perspective on the energy landscape}
of exact and approximate state-specific electronic structure theory.
In this framework, exact ground and excited states become stationary points of an energy landscape
constrained to the surface of a unit hypersphere while approximate wave functions form constrained subspaces.
Furthermore, the Hamiltonian variance $\mel*{\Psi}{(\mathcal{H} -E)^2}{\Psi}$ is directly proportional
to the square-magnitude of the exact energy gradient.
Deriving this geometric framework allows exact and approximate excited state-specific methods to
be investigated on an equal footing and leads to the following key results:
\begin{enumerate}[itemsep=0em]
\item{Exact excited states form saddle points of the energy with the number of
downhill directions equal to the excitation level;}
\item{The local energy and second derivatives at an exact stationary point
can be used to deduce the entire energy spectrum of a system;}
\item{\edit{The exact energy landscape has only two minima, corresponding to sign-permutations
of the ground state;}}
\item{Approximate excited solutions are generally saddle points of the energy
and their Hessian index increases with the excitation energy;}
\item{Physical minima of the variance are separated from states adjacent in energy by index-1 saddle points.
The barrier height is proportional to the square of the energy difference between the two states;}
\end{enumerate}
\edit{While only the simple \ce{H2} example has been considered, these results are sufficient to establish a set of guiding principles
for developing robust optimisation algorithms for state-specific excitations.
Future work will investigate how fermionic anti-symmetry affects the relationship between exact and approximate electronic energy landscapes
for systems with multiple same-spin electrons.}
Beyond state-specific excitations, the exact energy landscape may also provide a new perspective for understanding
the broader properties of wave function approximations.
For example, this work has shown that the orbital-optimised ESMF
wave function can describe the exact ground state of dissociated \ce{H2} (STO-3G) for all bond lengths
using only the reference determinant and single excitations.
This observation suggests that the orbital-optimised ESMF ground state may provide an
alternative black-box wave function for capturing static correlation in single-bond dissociation.
Alternatively, drawing analogies between a Taylor series approximation on the exact energy landscape
and second-order perturbation theory may provide an orbital-free perspective on the divergence
of perturbative methods for strongly correlated systems.
Finally, investigating how more advanced methods such as multi-configurational SCF,\cite{RoosBook} variational CC,\cite{Cooper2010,Marie2021a} or Jastrow-modified antisymmetric geminal power\cite{Neuscamman2011,Neuscamman2013}
approximate exact stationary points on the energy landscape may inspire
entirely new ground- and excited-state wave function approaches.
\section*{Acknowledgements}
H.G.A.B. was financially supported by New College, Oxford through the Astor Junior Research Fellowship.
The author owes many thanks to Antoine Marie, Alex Thom, David Wales, David Tew, and Pierre-Fran\c{c}ois Loos for discussions
and support throughout the development of this work.
The author is also thanks the Reviewers for insightful comments that have improved this work.
\section*{References}
|
\section{Introduction}
Breaking of the discrete time-translation symmetry (DTTS) is well known in classical physics. The simplest example is a parametric oscillator, which vibrates at half the frequency of the field that modulates the oscillator eigenfrequency \cite{Landau2004a}.
Another example is the period doubling route to chaos in nonlinear dynamical systems, including periodically modulated systems \cite{Feigenbaum1978,*Feigenbaum1979,Huberman1979,Linsay1981}. DTTS breaking in driven classical systems due to many-body effects, including phase-transition like features of the onset of such symmetry breaking, has been also known \cite{Goldstein2018,Kim2006,Heo2010}.
Recently much attention has attracted DDTS breaking in driven many-body quantum systems, sometimes called the time-crystal effect, see \cite{Khemani2016a,Else2016,Zhang2017,Choi2017,Pal2018,Rovny2018,Berdanier2018,Sacha2018,Dykman2018,O'Sullivan2018} and papers cited therein. In principle, the occurrence of subharmonics in a quantum multiple-state system is not surprising. Indeed, suppose one projects all Floquet eigenvalues (quasienergies) $\varepsilon_n$ onto the first Brillouin zone $-\hbar\omega_F/2\leq \varepsilon_n<\hbar\omega_F/2$, where $n$ enumerates the states and $\omega_F$ is the drive frequency. For a large number of states, the band will be filled almost densely \cite{Hone1997,Kohn2001}. One can then find a pair of states $\Ket{n_1}$ and $\Ket{n_2}$ with the quasienergy difference $\varepsilon_{n_1}-\varepsilon_{n_2}\approx \hbar\omega_F/N$ with an integer $N>1$. If the system is in a superposition of these states, the expectation values of the dynamical variables oscillate at the frequency $\omega_F/N$. However, the amplitude of such oscillations is determined by the overlap integral of $\Ket{n_1}$ and $\Ket{n_2}$ and is often exponentially small. In addition, preparing the corresponding superposition is not necessarily trivial.
With the rapid progress in making highly coherent and well-controlled systems of qubits, a natural question to ask is: How simple is it to achieve time-translation symmetry breaking in a quantum system? One of the goals of this paper is to show explicitly that a multiple-period superposition of states with an arbitrary multiplicity can emerge already in the simplest coherent quantum system, a periodically driven qubit. Achieving this goal requires finding an appropriate driving protocol that would be easy to implement in the experiment. The argument of the previous paragraph regarding almost dense quasienergy spectrum does not apply to a qubit, as it has only two states. On the other hand, by construction the overlap integral of the corresponding states is of order unity, which should facilitate an observation of the symmetry breaking.
Another problem addressed in the paper is a qubit chain with the modulated coupling between the qubits. We consider modulation close to twice the qubit eigenfrequencies. Looking for time-symmetry breaking in such a system is motivated by the symmetry breaking that occurs in a parametrically modulated classical oscillator. Unexpectedly, we find that the dynamics of a modulated qubit chain maps onto that of the Kitaev chain. This is a consequence of the symmetry of the modulated spin system. A solution that would explicitly demonstrate the underlying symmetry should be sought first for two coupled qubits. For a long chain, on the other hand, we seek to establish the relation between the familiar onset of Majorana fermions in the topologically nontrivial regime and the onset of period doubling.
Our analysis refers to the regime where the transition frequency of the qubit(s) $\omega_0$ is large compared to the Rabi frequency, for a single qubit, and compared to the qubit coupling in the frequency units, for the qubit chain. This suggests separating the time scales and analyzing the dynamics in slow time compared to $2\pi/\omega_0$, i.e., in the rotating frame. However, the quasienergy is defined in the laboratory frame. Therefore to describe the time-symmetry breaking, the results of the analysis have to be projected back to this frame.
\section{A periodically driven qubit}
We start with a single periodically modulated qubit. The modulation includes a resonant drive with a modulated amplitude and a low-frequency modulation of the interlevel spacing of the qubit. The Hamiltonian in the laboratory frame is
\begin{align}
\label{eq:H_lab_frame}
H_{\rm lab} = \frac{1}{2}[\omega_0+ \nu(t)] \sigma^z +F(t)\sigma^x \cos\omega_F t ,
\end{align}
with $|\omega_F-\omega_0|\ll \omega_F$. We have set $\hbar = 1$; $\sigma^{x,y,z}$ are the Pauli matrices. The wave functions and the quasienergy spectrum of the spin with a time-dependent Hamiltonian have been found in several limiting cases, in particular, for a sinusoidal modulation [$F(t)=$const, $\nu(t)=0$] \cite{Ma2007},
see also
\cite{Schmidt2018a,Schmidt2018}
and references therein.
We assume that $\nu(t) $ and $F(t)$ have the form of periodic pulses,
\begin{align}
\label{eq:modulation_form}
&\nu(t) = \nu_1\sum_n\overline\delta (t-nT), \nonumber\\
& F(t) = F_0+ F_1\sum_n\overline\delta(t-nT).
\end{align}
Here, the period $T$ is a multiple of the drive period, $T=2\pi n/\omega_F$ with integer $ n \gg 1$; in this case the Hamiltonian $H_{\rm lab}$ as a whole is periodic with period $T$. We use the notation $\overline\delta(t)$ for a function, which is smooth on the time scale $\sim 1/\omega_F$, but looks like a $\delta$-function on the scale $\sim \min(T, |\omega_F-\omega_0|^{-1})$ and has unit area, $\int_{-\infty}^\infty dt \overline\delta (t)=1$. The pulses can be applied independently to the level spacing and to the driving force amplitude, but when the both parameters are modulated, we assume that they are modulated with the same period. The pulse strengths $\nu_1, F_1$ are dimensionless (to incorporate $\hbar$, we have to replace $\nu_1\to \hbar\nu_1, F_1\to \hbar F_1$).
If the modulation is comparatively weak, so that $|\nu(t)|,\,|F(t)|\ll \omega_F$, one can go to the rotating frame using the transformation $U(t)=\exp(-i\omega_Ft\sigma_z/2)$. In the rotating wave approximation (RWA) the Hamiltonian becomes
\begin{align}
\label{eq:H_RWA}
&H_{\rm RWA} = \frac{1}{2}\Delta \sigma^z +\frac{1}{2}F_0\sigma^x + \frac{1}{2}{\bf g}{\bm \sigma}\sum_n \delta(t-nT), \nonumber\\
&\Delta = \omega_0 -\omega_F, \qquad g^x = F_1, \quad g^z=\nu_1, \quad g^y=0 .
\end{align}
Here we have taken into account that, in slow time, the function $\overline\delta(t)$ becomes a $\delta$-function. The components of the vector ${\bf g}$ are determined by the modulation strength. We emphasize that the Hamiltonian $H_{\rm RWA}$ depends on time and, as $H_{\rm lab}$, is periodic with period $T$.
The time-independent part of $H_{\rm RWA}$ describes a spin in an effective magnetic field with $z$ and $x$ components being $\Delta$ and $F_0$. It is convenient to rotate the qubit in such a way that the new $z$-axis is pointing along this field, $\sigma^z +i\sigma^x = \exp(i\phi)(\tilde\sigma^z +i\tilde\sigma^x)$. The rotation angle $\phi$ is given by the familiar equation $\tan\phi=F_0/\Delta$. Then the RWA Hamiltonian becomes
\begin{align}
\label{eq:rotated_H_RWA}
H_{\rm RWA} = \frac{1}{2}\Omega \tilde\sigma^z + \frac{1}{2}\tilde{\bf g}\tilde{\bm \sigma}\sum_n \delta(t-nT),
\end{align}
where $\Omega=(\Delta^2 + F_0^2)^{1/2}$ is the Rabi frequency in the absence of the pulses and
$\tilde g^z+i\tilde g^x = \exp(-i\phi)(g^z + i g^x)$, i.e., $\tilde{\bf g}$ is the rotated vector ${\bf g}$.
Prior to analyzing the dynamics in the rotating frame described by the RWA Hamiltonian $H_{\rm RWA}$, Eq.~(\ref{eq:rotated_H_RWA}), we note that the Floquet eigenstates of this time-periodic Hamiltonian are also Floquet eigenstates of the Hamiltonian $H_{\rm lab}$ in the laboratory frame. Indeed, if $\psi(t)$ is a Floquet eigenstate of $H_{\rm RWA}$ with a rotating-frame quasienergy $\varepsilon$, this means that
\begin{align}
\label{eq:Floquet_condition}
i\dot \psi = H_{\rm RWA} \psi, \qquad \psi(t+T) = \exp(-i\varepsilon T)\psi(t).
\end{align}
Taking into account the form of the transformation to the rotating frame $U(t)$, we see that the corresponding wave function in the laboratory frame $\psi_{\rm lab}(t) = U(t) \psi(t)$ also satisfies the Floquet-eigenstate condition,
\begin{align*}
\psi_{\rm lab}(t+T)& = U(t+T)\psi(t+T) =e^{-i\omega_FT\sigma_z/2}\,e^{-i\varepsilon T}\psi_{\rm lab}(t)\\
&\equiv \exp (-i\varepsilon T-i\pi n)\psi_{\rm lab}(t),
\end{align*}
where we used that $\omega_FT = 2\pi n$. The quasienergy $\varepsilon$ in the rotating frame corresponds to the quasienergy $\varepsilon + n\pi T^{-1}$ in the laboratory frame projected on the Brillouin zone of quasienergies $(-\pi/T,\pi/T)$. For an even $n$ such a projection gives $\varepsilon$, whereas for an odd $n$ it gives $\varepsilon-(\pi/T)\,{\rm sgn}\,\varepsilon$, provided $-\pi/T\leq \varepsilon< \pi/T$. In what follows we use the term ``rotating-frame quasienergy'' for the quasienergy defined by Eq.~(\ref{eq:Floquet_condition}).
\subsection{Finding the Floquet eigenstates in the rotating frame}
To find the Floquet wave functions we start by choosing an arbitrary instant $t_0$ within a period of $H_{\rm RWA}$. We solve the time-dependent Schr\"odinger equation $i\dot \psi = H_{\rm RWA} \psi$ with the initial condition $\psi(t_0)=\psi_0$ to find $\psi(t+t_0)$. We then find the explicit form of $\psi_0$ so that the solution meets the condition $\psi(T+t_0) = \exp(-i\varepsilon T)\psi_0$. This gives both $\psi_0$ and the value of the rotating-frame quasienergy $\varepsilon$. Clearly, $\psi_0$ depends on the chosen $t_0$, since the Floquet state is time-dependent, but $\varepsilon$ does not.
Since all periods are on equal footing, we can choose $t_0$ such that $0<t_0 <T$ [equivalently, we can choose $kT<t_0<(k+1)T$ with an arbitrary integer $k$]. A Floquet eigenstate at time $t_0$ can be written as a superposition
\[\psi(t_0)\equiv \psi_0 =\alpha\vert\tilde 1\rangle +\beta\vert \tilde 0\rangle,\]
where $\vert\tilde 1\rangle$ and $\vert \tilde 0\rangle$ are the eigenstates of $\tilde\sigma^z$ with eigenvalues $1$ and $-1$, respectively; these eigenstates form a complete set, which justifies the form of $\psi(t_0)$.
The evolution of the wave function in the interval $(t_0,T-\epsilon)$ is controlled by the time-independent part of $H_{\rm RWA}$; we formally consider the limit $\epsilon\to +0$; on physical grounds, $\epsilon\ll T, \Omega^{-1}$. At the time $T-\epsilon$ the wave function becomes
\begin{align}
\label{eq:free_evolution}
\psi(T-\epsilon) =& \alpha \exp[-i\Omega(T-t_0)/2]\vert\tilde 1\rangle\nonumber\\
& +\beta\exp[i\Omega(T-t_0)/2]\vert \tilde 0\rangle \quad (\epsilon\to +0).
\end{align}
During the pulse, the evolution of the wave function is controlled by the term $\propto \tilde{\bf g}$, which is much larger than the time-independent part of $H_{\rm RWA}$. The eigenfunctions of $\tilde{\bf g}\tilde{\bm\sigma}$ can be chosen in the form
\begin{align}
\label{eq:g_eigenfunctions}
\psi_{\pm} = \zeta_{\pm}\vert\tilde 1\rangle + \eta_{\pm}\vert \tilde 0\rangle, \quad \tilde{\bf g}\tilde{\bm\sigma}\psi_{\pm} = \pm g\psi_{\pm},
\end{align}
where $g=|{\bf g}|\equiv |\tilde{\bf g}|$ and the real coefficients $\zeta_{\pm}, \eta_{\pm}$ are given by the equation
\begin{align}
\label{eq:theta_angle}
&\zeta_+ + i\zeta_- =\exp(-i\theta), \quad i\eta_+ +\eta_-=\exp(i\theta); \nonumber\\
& \tan \theta = \tilde g^x/(g+\tilde g^z).
\end{align}
The functions $\psi_{\pm}$ also form a complete set. Therefore one can write $\psi(T-\epsilon) = C_+\psi_+ + C_-\psi_-$ and express the coefficients $C_{\pm}$ as linear combinations of the coefficients $\alpha,\beta$ of the initial wave function $\psi_0$ with the weighting factors that depend on the parameters $\theta$ and $\Omega(T-t_0)$, cf. Eqs.~(\ref{eq:free_evolution}) and (\ref{eq:g_eigenfunctions}). From Eqs.~(\ref{eq:rotated_H_RWA}) and (\ref{eq:g_eigenfunctions}) we see that, immediately after the pulse, the wave function is $\psi(T+\epsilon) = C_+\psi_+\exp(-ig/2) + C_-\psi_-\exp(ig/2)$. This wave function can be again written in the basis of the functions $\vert\tilde 1\rangle$ and $\vert \tilde 0\rangle$, $\psi(T+\epsilon) = \alpha_T\vert\tilde 1\rangle +\beta_T\vert \tilde 0\rangle$. The coefficients $\alpha_T, \beta_T$ are linear combinations of $C_{\pm}$ and ultimately are linear combinations of $\alpha, \beta$.
The evolution of $\psi(t)$ in the interval $(T+\epsilon, T+t_0)$ is again controlled by the Hamiltonian $\Omega\tilde\sigma^z/2$, with $\psi(T+t_0) = \alpha_T\exp(-i\Omega t_0/2)\vert\tilde 1\rangle + \beta_T\exp(i\Omega t_0/2)\vert \tilde 0\rangle$. This expression relates $\psi(T+t_0)$ to $\psi(t_0)$. The condition (\ref{eq:Floquet_condition}) with $t=t_0$ now reads $\alpha_T\exp(i\Omega t_0/2)=\alpha\exp(-i\varepsilon T), \; \beta_T\exp(-i\Omega t_0/2) = \beta\exp(-i\varepsilon T)$. Given that $\alpha_T,\beta_T$ are linear combinations of $\alpha,\beta$, this gives a homogeneous system of linear equations for $\alpha,\beta$, which is the eigenvalue problem for the quasienergy $\varepsilon$. In fact, the corresponding characteristic equation is an equation for $\exp(-i\varepsilon T)$. It has two solutions with equal $\cos(\varepsilon T)$ and opposite in sign $\sin(\varepsilon T)$.
The values of the rotating-frame quasienergy $\varepsilon_{1,2}$ obtained this way can be chosen in the form
\begin{align}
\label{eq:quasienergy}
&\varepsilon_{1,2} = \pm T^{-1}\nonumber\\
&\times\arccos\left[\cos\frac{g}{2}\, \cos\frac{\Omega T}{2} -\cos 2\theta\,\sin\frac{g}{2}\, \sin\frac{\Omega T}{2}\right].\end{align}
They correspond to two Floquet eigenstates of the qubit in the rotating frame, $\psi_{j}(t+T)=\exp(-i\varepsilon_jT)\psi_{j}(t)$ ($j=1,2$), and thus to two Floquet eigenstates in the laboratory frame.
\begin{figure}[h]
\includegraphics[scale=0.45]{single_qubit.pdf}
\caption{The quasienergy $\varepsilon_1$ of a qubit in the rotating frame. The qubit is driven at its eigenfrequency, $\omega_F = \omega_0$. The driving amplitude is pulsed with period $T \gg 2\pi/\omega_F$, the dimensionless area of a pulse is $F_1$. The blue, magenta, green, and red curves (the curves from bottom to top, for $F_1=2\pi$) refer to the Rabi frequency $\Omega = 2T^{-1}, T^{-1},0.5T^{-1}$, and $0.05T^{-1}$.}
\label{fig:single_qubit}
\end{figure}
In Fig.~\ref{fig:single_qubit} we show the quasienergy in the case where only the amplitude of the resonant field $F(t)$ is pulsed, whereas the frequency of the qubit is not modulated, $\nu_1=0$. As seen from the figure and also from Eq.~(\ref{eq:quasienergy}), the quasienergy is a periodic function of the intensity (area) of the pulses $F_1 = g$ with period $4\pi$. It is also a periodic function of the Rabi frequency $\Omega$ with period $4\pi/T$. The magnitude of the oscillations of $\varepsilon_1$ with $F_1$ is maximal for $\Omega T = n\pi$ with an integer $n$.
A linear combination $\Psi(t) = A\psi_1(t) + B\psi_2(t)$ is not a Floquet eigenstate, $\Psi(t+T)\neq \exp(-i\varepsilon T)\Psi(t)$ with a real $\varepsilon$ for $AB\neq 0$. However, if $\varepsilon_1-\varepsilon_2 = 2\varepsilon_1 = 2\pi M/NT$ with integer $M,N$ and $N>|M|\geq 1$, we have $\Psi(t+NT) = \exp(-iM\pi)\Psi(t)$. Physical observables in such a state are oscillating with period $NT$. The amplitude of the oscillations is $\propto |AB|$ and therefore can be of order one. Another single-particle system where the difference of the quasienergies can be made equal to $2\pi/NT$ (with $N=2$ and $3$) while keeping the overlap of the corresponding wave functions $\sim 1$ is a resonantly modulated nonlinear oscillator \cite{Marthaler2007,Zhang2017b,Zhang2017c}.
\subsection{Observing multiple-period states of a qubit}
A particularly simple way to obtain multiple-period states is to pulse the level spacing of a qubit, $\omega_0\to \omega_0+\nu(t)$, with $\nu(t)$ of the form given by Eq.~(\ref{eq:modulation_form}). The periodic pulsing leads to an increment of the phase difference between the states $|1\rangle$ and $|0\rangle$ by $\nu_1$ each period $T$. It is obvious without calculation that, for $\omega_0=2\pi n/T$ with an integer $n$, the states $|1\rangle$ and $|0\rangle$ are quasienergy states in the laboratory frame with the quasienergies $\pm [\omega_0 /2+ (\nu_1/2T)]$ projected onto the Brillouin zone $(-\pi/T, \pi/T)$. Because of the relation between the quasienergies in the laboratory and rotating frames, this expression coincides with what follows from Eq.~(\ref{eq:quasienergy}). To see this one should set in Eq.~(\ref{eq:quasienergy}) $\Omega = \theta =\phi = 0, g=\nu_1$, which then gives $\varepsilon_{1,2}=\pm T^{-1}\arccos[\cos( \nu_1/2)]$.
The pulsing-induced phase shift can be revealed in a standard way using Ramsey fringes. The measurement is done in the following way. If the qubit with a pulsed level spacing is in the state $|0\rangle$, one applies a $\pi/2$ resonant (at frequency $\omega_0$) pulse at some time $t=t_i$. This pulse drives the qubit into a superposition of the quasienergy states with equal weights. If at a time $t_f$ there is applied another resonant $\pi/2$ pulse at frequency $\omega_0$, the population of the excited state, which is given by the projection of the wave function after the pulse $\Psi(t_f)$ on the wave function of the excited state $|1\rangle$, becomes
\[|\langle 1|\Psi(t_f)\rangle|^2 = \cos^2\Bigl[ \nu_1 \Bigl\lfloor (t_f - t_i)/T)\Bigr\rfloor /2\Bigr],\]
where, $\lfloor x\rfloor$ is the integer part of $x$. For $\nu_1 = 2\pi M/N$ this population is a periodic function of $t_f-t_i$ with period $NT$.
Another example is where the qubit is driven only by a close to resonance field $F(t)\cos\omega_Ft$ with a pulsed envelope, $F(t)=F_1\sum_n\overline\delta(t-nT)$. In particular, for exact resonance, $\omega_F = \omega_0$, we have in Eq.~(\ref{eq:quasienergy}) $\phi=\Omega=0, \theta=\pi/4, g=F_1$, and then $\varepsilon_{1,2}=\pm T^{-1}\arccos[\cos(F_1/2)]$. The quasienergy wave functions in the rotating frame are $\psi_{\pm} = (|0\rangle \pm |1\rangle)/\sqrt{2}$. If in the laboratory frame the qubit is initially ($t=t_i$) in the ground state $|0\rangle_{\rm lab}$, the occupation of the excited state $|1\rangle_{\rm lab}\equiv U(t)|1\rangle$ at time $t_f$ is
\[|\langle 1|\Psi(t_f)\rangle|^2 = \sin^2\Bigl[ F_1 \Bigl\lfloor (t_f - t_i)/T)\Bigr\rfloor /2\Bigr].\]
If the pulse area is a simple fraction of $2\pi$, $F_1=2\pi M/N$, the state population varies in time periodically with period $NT$.
In the both examples, establishing the broken DTTS is reduced to the standard operation of detecting the occupation of the excited state of a qubit. We emphasize again that the pulses must be smooth on the scale $\omega_0^{-1}$ but short on the scale $T$.
\section{A qubit chain with a resonantly modulated coupling}
An interesting behavior, including the possibility of DTTS breaking, is provided by a system of coupled qubits with a resonantly modulated coupling. We consider here a sinusoidal modulation at frequency $\omega_F$ close to twice the qubit transition frequency $\omega_0$, i.e., $|\omega_F-2\omega_0|\ll \omega_0$. The Hamiltonian of a chain of coupled qubits with nearest neighbor coupling and with one of the coupling parameters, $J_{xx}$, being modulated has the form
\begin{align}
\label{eq:chain_lab}
&H_{\rm lab}^{\rm (chain)}=\frac{1}{2}\omega_0 \sum_n \sigma_n^z -J_{xx}(t)\sum_n\sigma_n^x\sigma_{n+1}^x\nonumber\\
& - J_{yy}\sum_n\sigma_n^y\sigma_{n+1}^y,
\qquad J_{xx}=J_{xx}^{(0)} + 2F\cos\omega_Ft.
\end{align}
Controllable $xx$ coupling has been implemented in several types of flux qubits, cf. \cite{Yan2018} and references therein, although we are not aware of the experiments where the coupling was modulated at frequency $\sim 2\omega_0$. However, we do not immediately see physical constraints that would prohibit such a modulation, although its implementation could be accompanied by a decrease of the coherence time. The results below apply also to the case where the both coupling parameters, $J_{xx}$ and $J_{yy}$, are modulated.
In many implementations, the qubit coupling is weak compared to the qubit transition energy, $|J_{xx}|, |J_{yy}|\ll \omega_0$, which we assume to be the case. Then, as shown below, resonant driving can lead to strong effects even where it is comparatively weak, $|F|\ll \omega_0$. For small $|J_{xx}|, |J_{yy}|, |F|$ the dynamics can be conveniently analyzed by switching to the rotating frame at frequency $\omega_F/2$. The transformation is
\[U^{\rm (chain)}(t)=\exp[-i(\omega_F t/4)\sum_n(\sigma_n^z+I_n)],\]
where $I_n$ is the identity operator for an $n$th spin, i.e., the $2\times 2$ unit matrix operating in the $n$th spin Hilbert space; as will be seen below, particularly when the problem is formulated in terms of the fermion operators, introducing the operators $I_n$ into $U^{\rm (chain)}(t)$ simplifies the analysis.
In the rotating wave approximation the Hamiltonian becomes
\begin{align}
\label{eq:chain_RWA}
&H_{\rm RWA}^{\rm (chain)} = -\frac{1}{2}\mu \sum_n \sigma_n^z -\frac{1}{4}J\sum_n(\sigma_n^+\sigma_{n+1}^- + \sigma_{n+1}^+\sigma_{n}^-)\nonumber\\
& -
\frac{1}{4}F\sum_n(\sigma_n^+\sigma_{n+1}^+ + \sigma_{n+1}^-\sigma_{n}^-), \quad \mu=\frac{1}{2}\omega_F-\omega_0.
\end{align}
Here $J= J_{xx}^{(0)} + J_{yy}$ and $\sigma_n^\pm = \sigma_n^x \pm i\sigma_n^y$. In the rotating frame, the frequency detuning $\mu$ plays the role of the scaled magnetic field along the $z$-axis. The detuning is small compared to $\omega_0$, but can be of the same order of magnitude as the coupling parameter $J$ and the modulation amplitude $F$. In the expression for $H_{\rm RWA}^{\rm (chain)}$ we left out the identity operator in the spin-chain Hilbert space $-(\omega_F/4)\sum_nI_n$, as it does not affect the dynamics.
The Hamiltonian $H_{\rm RWA}^{\rm (chain)}$ is independent of time. Therefore, in contrast to the previously considered case of the pulse-modulated single qubit, the eigenvalues and eigenfunctions in the rotating frame in the RWA are given by the solution of the stationary problem
\begin{align*}
H_{\rm RWA}^{\rm (chain)}\psi = \varepsilon \psi.
\end{align*}
The rotating-frame eigenfunction $\psi$ with an RWA eigenvalue $\varepsilon$ evolves in time as
\begin{align}
\label{eq:psi_chain_vs_time}
\psi(t+T)=e^{-i\varepsilon T}\psi(t).
\end{align}
This equation holds for an arbitrary $T$, but in what follows we will be interested in $T$ being the modulation period, $T=2\pi/\omega_F$.
To relate $\varepsilon$ to the quasienergy in the laboratory frame, we introduce the parity operator $P$,
\begin{align}
\label{eq:parity}
P=\exp\Bigl[-i\frac{\pi}{2}\sum_n(\sigma_n^z+I_n)\Bigr], \quad [P,H_{\rm RWA}^{\rm(chain)}]=0.
\end{align}
Clearly, $P=P^\dagger$ and $P^2 = \prod_nI_n$. The eigenvalues of $P$ are $\pm 1$. The parity operator $P$ commutes not only with $H_{\rm RWA}^{\rm(chain)}$, but also with the Hamiltonian in the laboratory frame, $[P, H_{\rm lab}^{\rm (chain)}]=0$. The parity conservation is not a consequence of the RWA.
With the account taken of the explicit form of the unitary transformation $U^{\rm(chain)}(t)$ and Eq.~(\ref{eq:psi_chain_vs_time}), we see that the wave function $\psi_{\rm lab}(T)=U^{\rm (chain)}(t)\psi(t)$ in the laboratory frame transforms over the period as
\begin{align*}
\psi_{\rm lab}(t+T) =&\exp\Bigl[-i(\omega_F T/4)\sum_n(\sigma_n^z+I_n)\Bigr]e^{-i\varepsilon T}\psi_{\rm lab}(t)\\
&= \exp(-i\varepsilon T)P\psi_{\rm lab}(t).
\end{align*}
Therefore for even-parity states, i.e., where the eigenvalue of $P$ is $1$, the rotating-frame quasienergy $\varepsilon$ is also the quasienergy in the laboratory frame, whereas for odd-parity states, where the eigenvalue of $P$ is $-1$, the quasienergy in the laboratory frame is $\varepsilon - (\omega_F/2)\,{\rm sgn}\,\varepsilon$, i.e., it is shifted from $\varepsilon$ by half the Brillouin zone.
Importantly, the eigenstates of $H_{\rm RWA}^{\rm (chain)}$ with different parity can have the same eigenvalue $\varepsilon$, there is no level repulsion between such states. From this general argument, it is clear that, if we prepare the system in a combination of states with the same $\varepsilon$ in the rotating frame, but with different parity, so that in the laboratory frame the wave function is $U^{\rm(chain)}(t)[\alpha_{\rm even}\psi_{\rm even}(t) + \alpha_{\rm odd}\psi_{\rm odd}(t)]$, the expectation values of dynamical variables in this state will have period $2T$. In other words, this state will have broken time-translation symmetry.
\subsection{A two-qubit chain}
To illustrate the occurrence of the breaking of the time-translation symmetry in a modulated chain we start with a particularly simple case of two qubits, i.e., we assume that $n=1,2$ in Eq.~(\ref{eq:chain_RWA}). The eigenfunctions and eigenvalues of the operator $H_{\rm RWA}^{\rm (chain)}$ in this case are
\begin{align}
\label{eq:two_qubit}
&\psi_{1,2} = \cos\phi_{1,2}\Ket{00} + \sin\phi_{1,2}\Ket{11}, \quad \nonumber\\
&\psi_{3,4}= (\Ket{01}\pm \Ket{10})/\sqrt{2},\nonumber\\
&\varepsilon_{1,2}=\pm (\mu^2+F^2)^{1/2},\quad \varepsilon_{3,4} = \mp J,
\end{align}
where $\tan \phi_{1,2}=(\mu-\varepsilon_{1,2})/F$. The eigenvalue of the parity operator on the functions $\psi_{1,2}$ is 1, whereas on the functions $\psi_{3,4}$ it is $-1$.
\begin{figure}[h]
\includegraphics[scale=0.3]{two_qubit_scale2.pdf}
\caption{The rotating-frame energies of a two-qubit system with the coupling periodically modulated at frequency $\omega_F$ close to twice the qubit transition frequency $\omega_0$, Eq. (\ref{eq:two_qubit}). The dashed and solid curves refer to the states 1, 2 and 3, 4, respectively. The frequency detuning is $\frac{1}{2}\omega_F - \omega_0 \equiv \mu =0.6J$. The green circles indicate where the quasienergies in the laboratory frame differ by $\omega_F/2$. There is no anticrossing between the corresponding quasienergy levels.}
\label{fig:quasienergies}
\end{figure}
The quasienergies in the rotating frame, Eq.~(\ref{eq:two_qubit}), as functions of the driving amplitude $F$ are shown in Fig.~\ref{fig:quasienergies}. Only the values of $\varepsilon_{1,2}$, which refer to the even-parity states, vary with $F$. If $|\mu| <|J|$ (for concreteness, we assume $J>0$), by varying $F$ one can make the rotating-frame quasienergies of the states with different parity coincide, so that $\varepsilon_1 = \varepsilon_4$ and $ \varepsilon_2 = \varepsilon_3$. In this case, as explained above, linear combinations $\alpha_{\rm even}\psi_1 + \alpha_{\rm odd} \psi_4$ and $\alpha_{\rm even}'\psi_2 + \alpha_{\rm odd}' \psi_3$ are period-two states in the laboratory frame.
The analysis immediately extends to the case where the qubit frequencies are slightly different. The parity is still conserved in this case. One can easily check that the rotating-frame quasienergies of the different-parity states become equal simultaneously for two pairs of states, as in the case of equal qubit frequencies.
To observe the period doubling one can prepare the system in a superposition of states contained in $\psi_{1,2}$ and $\psi_{3,4}$, for example, in a superposition of states $|01\rangle$ and $|00\rangle$. Then if the driving is tuned so that $(\mu^2+F^2)^{1/2} = |J|$, the expectation values of physical observables will oscillate with period $2T$, the period doubling effect.
\subsection{Topologically nontrivial Floquet regime}
In the case of a longer modulated qubit chain, the analysis of the dynamics can be conveniently done using the Jordan-Wigner transformation from spins to spinless fermions. If the creation and annihilation operators of a fermion on site $n$ are $a_n^\dagger$ and $a_n$, respectively, the Hamiltonian (\ref{eq:chain_RWA}) becomes the Hamiltonian of the Kitaev chain
\begin{align}
\label{eq:Kitaev}
H_K &= -\mu\sum_n\Bigl(a_n^\dagger a_n -\frac{1}{2}\Bigr)- J\sum_n(a_n^\dagger a_{n+1} +a_{n+1}^\dagger a_n)\nonumber\\
& -F\sum_n(a_n^\dagger a_{n+1}^\dagger +a_{n+1}a_n).
\end{align}
In the fermion representation in the rotating frame, the role of the chemical potential $\mu$ is played by the frequency detuning $(\omega_F/2)-\omega_0$, cf. Eq.~(\ref{eq:chain_RWA}).
In the fermion representation, the parity operator (\ref{eq:parity}) takes a simple form
\begin{align}
\label{eq:parity_fermion}
P=\exp\Bigl(-i\pi\sum_n a_n^\dagger a_n\Bigr), \quad [P, H_K]=0.
\end{align}
The eigenvalues of $P$ are $-1$ and $1$ for odd and even number of fermions, respectively.
The properties of the Kitaev chain are well-known \cite{Kitaev2001,Alicea2012}. The considered Floquet system is topologically nontrivial for $|\mu|\equiv |\frac{1}{2}\omega_F - \omega_0|< 2|J|$. Interestingly, this condition on the detuning of the drive frequency or the strength of the qubit coupling is less restrictive than the condition $|\mu|<|J|$ that must be met, as seen from Eq.~(\ref{eq:two_qubit}), to obtain period-two states in a system of two qubits. Both conditions can be met for a given qubit system by tuning the drive frequency closer to $2\omega_0$.
For completeness, the familiar behavior of excitations in the Kitaev chain is illustrated in Fig.~\ref{fig:Kitaev} for a comparatively short modulated qubit array. The data are obtained using the method \cite{Lieb1961} and show the evolution of the first and second excited states in the rotating frame with the changing modulation frequency $\omega_F$. The lowest excited state for small $|\mu/J|$ corresponds to the excitation localized at the edges of the chain and is described by the Majorana physics \cite{Kitaev2001}. As expected, its energy becomes extremely small for small $|\mu/J|$. For example, for the chain of 16 qubits in Fig.~\ref{fig:Kitaev}, $\varepsilon_1/J <4\times 10^{-5}$ for $|\mu/J|<1$ and $F/J=0.3$, and $\varepsilon_1/J <2\times 10^{-7}$ for $|\mu/J|<1$ and $F/J=1.2$; for small $F/J$ the boundary effects make $\varepsilon_1/J$ oscillate with $|\mu/J|$ with a very small amplitude. From the point of view of obtaining small $\varepsilon_1$ for $|\mu/J|\sim 1$, the optimal range of the driving amplitude is where $F/J$ is close to 1; this range depends on the chain length.
\begin{figure}[h]
\includegraphics[scale=0.265]{kitaev_en1.pdf} \hfill
\includegraphics[scale=0.26]{kitaev_en2.pdf}
\caption{The rotating-frame energies of the first [panel (a)] and second [panel (b)] lowest excited states of the qubit chain with a resonantly modulated coupling. The energies are given by the eigenvalues of the Hamiltonian (\ref{eq:Kitaev}) for a 16-sites chain of fermions. The chemical potential of the fermions $\mu$ is given by the frequency detuning $(\omega_F-2\omega_0)/2$, see Eq.~(\ref{eq:chain_RWA}). The blue, magenta, and green curves (the lowest to the highest curves) correspond to the scaled modulation amplitude $F/J = 0.3, 1.2$, and $3$.
}
\label{fig:Kitaev}
\end{figure}
The considered implementation of the Kitaev chain is based on resonant modulation of a qubit system and is thus qualitatively different from the proposed implementations that use flux qubits or other Josephson-junction based qubit systems with no time-periodic modulation, see \cite{Narozhny2017,Backens2017} and references therein. It is also different from the superconducting cold-atom quantum wire where, for a periodically step-wise modulated chemical potential, there emerge two pairs of localized Majorana fermions with the quasienergy difference equal to $\pi/T$, with $T$ being the modulation period \cite{Jiang2011}. Bound states with 0 and $\pi$ effective quasienergies were found also in a linear optical system that mimics a periodically modulated system \cite{Kitagawa2012}. Not only our analysis refers to a system of a different type, but as we have shown, resonant modulation of a qubit chain essentially automatically ``builds in'' a pair of Majorana fermions with the quasienergy shifted from the ground state by $\omega_F/2$.
Implementing the Floquet-Kitaev chain (\ref{eq:Kitaev}) with qubits is advantageous in terms of simulating Majorana fermions and obtaining a topologically protected period-2 state in a controlled way. The ground state of the spin chain in the absence of the driving corresponds to all spins aligned along the $z$-axis, with the expectation value of $\sigma_n^z$ being $-1$ for all $n$. It corresponds to the vacuum state of the fermions, $\langle a_n^\dagger a_n\rangle =0$ for all $n$. The lowest excited state of the chain without driving has an opposite parity. In a finite-length chain the driving can be adiabatically turned on without destroying the nomenclature of the states. In particular, the driving does not change the parity. A superposition of the ground and the first excited states is then not an eigenstate of the parity operator $P$. If the driving frequency is brought close to resonance, $\omega_F\approx 2\omega_0$, so that the energy of the excited state in the RWA is extremely small, the superposition of the states has broken time-translation symmetry if measured on the time scale small compared to $\varepsilon_1^{-1}$. This time is exponentially long for small $|(\omega_F/2)-\omega_0|$. The period doubling is thus topologically protected.
\section{Conclusions}
The results of this paper explicitly show that, in the quantum regime, subharmonics, including high-order ones, can be displayed already by the simplest driven quantum system, a qubit. The proposed protocol is to drive the qubit by periodic pulses of a resonant field and/or by dc pulses that modulate the spacing between the qubit energy levels. The period of the pulses should be much longer than the reciprocal qubit transition frequency $\omega_0^{-1}$. Such driving relies on the driving conventionally used to perform qubit gate operations. The symmetry-broken states can be detected using conventional measurement protocols. Therefore an experiment on the symmetry breaking can be done with any qubits that have a high transition frequency.
The ``price'' for the simplicity of the system is that, formally, the effect requires fine tuning of the parameters of the drive. However, given that qubits have a finite coherence time, one can establish periodicity of the dynamical variables of a qubit only with a limited precision. This limitation leads to a finite width of the parameter range within which the measured period is seen as a multiple of the drive period.
The other system considered in the paper is a qubit chain with periodically modulated qubit coupling. The results refer to qubits with the transition frequency $\omega_0$ much higher than the coupling energy divided by $\hbar$, and to a resonant modulation with frequency $\omega_F$ close to $2\omega_0$. Such chain preserves the parity of the total number of spin excitations. We found that, even for two spins, the quasienergies of states with opposite parity can differ by $\omega_F/2$, so that the supersposition of these states is periodic with period $4\pi/\omega_F$.
A resonantly modulated qubit chain has a nontrivial time-translation symmetry and maps onto the Kitaev chain with the parameters controlled by the modulation. In a long chain, the state with a broken symmetry of time translation by $2\pi/\omega_F$ is topologically protected. Speaking more broadly, such a qubit chain allows one to address, in a controlled experiment, several important problems of the Majorana physics. One of them is the effect of disorder \cite{Motrunich2001}. The site disorder can be emulated just by making the transition frequencies of the qubits slightly different. By making the coupling constants $J_{xx}, J_{yy}$ and the modulation amplitude $F$ site-dependent, one can emulate hopping and pairing disorder. If the system allows incorporating the $zz$-coupling, for example, if the Hamiltonian has the term $\sum_n\sigma_n^z\sigma_{n+1}^z$, one can also explore the effects of the fermion-fermion coupling.
Quantum simulations with modulated qubit chains are invaluable in terms of studying the aforementioned effects. Fluxonium qubits \cite{Manucharyan2009} may be advantageous, since modulation at frequency $\sim 2\omega_0$ will not lead to resonant transitions in such qubits. This will allow one to avoid leaving the computational subspace and the associated heating of the qubit system, which is of potential concern for transmon qubits, for example. A chain of $\sim 20$ qubits may already demonstrate the involved physics. On the other hand, a 20-qubit chain is presumably beyond what can be studied in the near future with classical computers, as it requires diagonalizing a $2^{20}\times 2^{20}$ Hamiltonian matrix, cf.~\cite{Borgonovi2016}.
I am grateful for the discussion to Vadim Smelyanskiy, Yu Chen, Vladimir Manucharyan, Pedram Roushan, and Maxim Vavilov. This work was supported in part by the NSF, grants DMR-1708331 and DMR-1806473, and by the Google faculty research award.
|
\section{Introduction}
The layered transition metal dichalcogenides (TMDCs) have a long history of research in physics and material science \cite{Greenaway1965, Murray1972, Lee1970, Naito1982}. In recent years, this time-honored system has received increasing attention, not only because of their exotic properties such as charge density waves \cite{Lee1970, Naito1982, JJYang2012, DWShen2008, TValla2004, PChen2017} and superconductivity \cite{TValla2004, DKang2015}, but also because of their extremely large and and non-saturating magnetotransport \cite{IPletikosic2014, MNAli2014} and topological properties \cite{Soluyanov2015, ZWang2016, XCPan2015, KZhang2016} and so on. In particular, TiTe$_2$ with structures related to the so-called 1$T$ polytype is a very interesting representative of the TMDCs. Up to date, TiTe$_2$ has been studied extensively owing to its interesting structural and electronic properties \cite{PChen2017, PBAllen1994, Koike1983, deBoer1984, Claessen1992, SHarm1994, JWAllen1995, Perfetti2001, Rossnagel2001, Perfetti2002, Nicolay2006, Strocov2006, FClerc2007, Krasovskii2007, Rossnagel2009, JKhan2012, VRajaji2018, UDutta2018}. TiTe$_2$ has been considered as a model material of Fermi liquid in the field of, e.g., high-resolution angle-resolved photoemission spectroscopy (ARPES) \cite{Claessen1992, SHarm1994, JWAllen1995, Perfetti2001, Rossnagel2001, Perfetti2002, Nicolay2006}. Most recent ARPES studies have shown that the single-layer TiTe$_2$ sample undergoes charge density wave (CDW) phase transition at 92$\pm$3K \cite{PChen2017}, which is not the case for bulk TiTe$_2$ \cite{deBoer1984, Koike1983, PBAllen1994}. More interestingly, nonhydrostatic pressure studies have found that TiTe$_2$ can even exhibit superconductivity \cite{UDutta2018}. Hence, the investigation of basic electronic structure and properties of 1$T$-TiTe$_2$ material is of great importance and significance in understanding quantum physics phenomena and in exploring potential applications in advanced electronic and optoelectronic devices.
As we know, most of the macroscopic properties of an electronic material are dictated by its microscopic band structure and electronic dynamics, specifically within an electronic energy regime of a few meV near the FS. It is therefore of great importance to reveals the topological change of the FS since it is intimately related to many of its low-energy properties such as electronic transport, specific heat, magnetic susceptibility, etc. Furthermore, the shape and size of the FS can be applied for determination of the carrier density, carrier type, carrier distribution, etc. From an experimental point of view, ARPES is a very effective tool to study the electronic structure of materials. This becomes the prime motivation for us to study the electronic structure of 1$T$-TiTe$_2$ material by using modern technique such as ARPES in the present study.
It should be noted that 1$T$-TiTe$_2$ has a sandwich-like material structure, where the electrons in different layers are attracted by the van der Waals forces. Due to the weak inter-layer bonding, high-quality TiTe$_2$ single crystals hold a promise of obtaining atomically flat and defect-free surface by cleaving under ultrahigh-vacuum. When 1$T$-TiTe$_2$ was considered as a textbook Fermi liquid material, its 3D electronic structures were often ignored. But in fact, even very anisotropic physical systems remain residual three-dimensionality. As has been pointed out that although the crystal structure of 1$T$-TiTe$_2$ is mainly two-dimensional (2D), its electronic structure can have very strong 3D-like natures \cite{Rossnagel2001, Strocov2006, Rossnagel2009}. The effects of dispersion along the $k_z$ are important and should be taken into account in analyzing and interpreting of the ARPES data. In this study, we intend to measure and examine the 3D electronic structure of 1$T$-TiTe$_2$ at relatively low temperatures. Through ARPES measurement via tunable synchrotron radiation with sufficient energy and momentum resolution, we would like to reveal the 3D electronic structure of 1$T$-TiTe$_2$ in different energy bands and to observe the corresponding FSs and complex band structure. In Section II, the samples and experimental measurement using ARPES are described briefly. The results obtained from this study are presented and discussed in Section III and the main conclusions are summarized in Section IV.
\section{samples and experimental measurement}
In this study, high-quality 1$T$-TiTe$_2$ single crystals were grown from the elements by chemical vapor transport using iodine as transport agent. The 1$T$-TiTe$_2$ samples used in the study were cleaved $in$ $situ$. The details of the sample preparation were documented elsewhere \cite{Rossnagel2009}.
For experimental measurement, high resolution ARPES measurements were carried out at beamline 5-4 of the Stanford Synchrotron Radiation Lightsource (SSRL), through using a Scienta R4000 electron energy analyzer. The measurements were carried out at a temperature of 10 K and in ultrahigh vacuum with base pressure better than 3$\times$10$^{-11}$ mabr. In the experiments, the wide angle $k_x$ - $k_z$ Fermi surface (FS) mapping were measured in the photon energy range from 16 eV to 30 eV in steps of 1 eV, with varying energy resolution from 8 meV to 21 meV. Phonton energy of 21 eV and 29 eV with corresponding energy resolution of 8 meV and 17 meV were used for probing of the $k_x$ - $k_y$ Fermi surface. Moreover, a light beam with a photon energy of 20 eV with an energy resolution of $\sim$ 10 meV was chosen to study the Ti 3$d$ band structure along high symmetry $\Gamma$(A)-M(L) direction.
\section{results and discussions}
\begin{figure}[bp]
\begin{center}
\includegraphics[width=0.95\columnwidth,angle=0]{Figure1.pdf}
\end{center}
\vspace*{-0.5cm} \caption{(color online) The map of three-dimensional Fermi surface of 1$T$-TiTe$_2$. The measurements are performed in the $\Gamma$$AML$ plane ($k_x$ - $k_z$). Different values of $k_z$ are accessed by varying the photon energy between 16 and 30 eV. The grey lines indicate the final state arc for different photon energies.}
\end{figure}
In the present study, we focus our attention on the 3D electronic structure of 1$T$-TiTe$_2$ at a relatively low temperature of 10 K. The 3D electronic band structure of 1$T$-TiTe$_2$ obtained from photon energy dependent normal emission is shown in Fig. 1. The measurements were performed in a section of the high-symmetry $\Gamma$$AML$ plane. For experimental convenience, the measurements were carried out with low photon energies, which holds high momentum and energy resolution. Different values of $k_z$ were accessed by varying the photon energy between 16 and 30 eV and estimated on the basis of an inner potential $V_0$ of 14.3 eV. As can be seen from Fig. 1, all the M(L) around Ti 3$d$ band and $\Gamma$(A) centered Te 5$p$ bands show very strong $k_z$ dispersion. The shape and intensity of the most important and widely studied Ti 3$d$ vary greatly with changing photon energy. The local-density approximations (LDA) calculation predict a 3D FS consisting of three hole pockets \cite{Strocov2006, Rossnagel2009, PChen2017}- two $\Gamma$-A centered corrugated cylinders and a $\Gamma$ centered lens-shaped FS \cite{Rossnagel2009}. Corrugated cylinder features can be found from the $k_x$-$k_z$ FS map. While, we cannot distinguish the small lens-shaped Te 5$p_z$ Fermi pocket, we can only see complex FS structures around $\Gamma$ point. That is probably because of the significant $k_z$ broadening \cite{VNStrocov2003, HWadati2006, Fujimori2016}. The strong 3D dispersion and the limited $k_z$ resolution result in photoemission features containing of signals from the nearby photons.
\begin{figure*}[tbp]
\begin{center}
\includegraphics[width=1.95\columnwidth,angle=0]{Figure2.pdf}
\end{center}
\vspace*{-0.5cm} \caption{(color online) Two-dimensional Fermi surface mapping of 1$T$-TiTe$_2$ in the $k_x$-$k_y$ plane(symmetrized according to the crystal symmetry). The maps of photoemission intensity are taken with 21 eV [in (a)] and 29 eV [in (b)] photons, respectively. In (c) and (d), the FSs near $\Gamma$(A) point are treated with minimum gradient method corresponding to (a) and (b), respectively.}
\end{figure*}
In the present study, the constant photon energy FS mapping were undertaken to clarify the topological structure of the FS. Fig. 2(a) and (b) shows the FS mapping of the 1$T$-TiTe$_2$ with photon energy of 21 and 29 eV, respectively. The corresponding $k_z$ positions are indicated by grey lines in Fig. 1. These two energies are chosen for one energy to probe the FS near $\Gamma$ point and another energy to probe FS away from $\Gamma$ point and near M point. Superposed on the Fig. 2(a) and (b) are the Brilliouin zone with high-symmetry points indicated. All maps exhibit a trigonal symmetry duo to the space group $P$\={3}$m$1 of the 1$T$-type TMDCs. One can easily identify two types of the FS sheets: i) hole pockets centered at the ${\Gamma}$ point and derived from the Te 5$p$ bands and ii) the Ti 3$d$ band related to electron pockets centered at both the M(L) and M'(L') points. Noted that the strong 3D character of FS can be confirmed. The intensity, shape, and size of hole and electron pockets are changed significantly with photon energies: i) for 21 eV photon energy, the spectral weight of M(L)/M'(L') centered FS are high/weak, while for 29 eV it is just the opposite. ii) for 21 eV, the FS map show three high intensity drop-shaped features around M(L) points and three weaker elliptic features near M'(L') points. While for 29 eV, the FS map show elliptic features around all the M(L) and M'(L') points. iii) for 21eV, the size of M(L) centered electron pockets are much larger than that of M'(L') centered electron pockets. While for 29 eV photon energy, the size of M(L) centered electron pockets are significantly smaller than that of M'(L') centered electron pocket.
It can be found from Fig. 2 that, in consistent with previous calculations \cite{Strocov2006, Rossnagel2009, PChen2017}, the structure of $\Gamma$ centered hole pockets is very complex. Here we employ the minimum gradient method to examine the multiple hole pockets around $\Gamma$(A) point. This approach is supposed to be able to sharpen the weak structure \cite{YuHe2017}. The results are shown in Fig. 2 (c) and (d), corresponding to Fig. 2(a) and (b) respectively. The dashed red lines indicate the possible position of the FS. For the cases of both 21 eV and 29 eV photon energies, at least three distinct $\Gamma$(A) centered hole pockets can be resolved. The inner hole pocket is a circle and the outer two hole pockets are hexagon. On closer inspection, some very weak and shapeless features can be found between adjacent hole pockets (see Fig. 2(c)). The following Te 5$p$ band structures also confirm this. These findings are probably due to finite $k_z$ dispersion and limited $k_z$ resolution. However the most recent studies report another possibility, which suggests the ARPES contour maps should reveal band continuums \cite{PChen2017}.
Because the FS topologies around the $\Gamma$(A) point are complex and have strong 3D characteristics, it is very important to study the band structure at and away from the $\Gamma$ point. Fig. 3(a) and (b) show the ARPES images of photoelectron intensity as a function of energy and momentum for TiTe$_2$. Fig. 3(c) and (d) are 2D second-derivative images from the original data. We note that a typical semi-metal structure can be found in TiTe$_2$, namely a single Fermi level goes across the electron and valence-electron (or hole) bands so that the electrons and holes coexist in the material. The light Te 5$p$ bands approach the $\Gamma$(A) point and form the hole pockets, whereas the shallow Ti 3$d$ band forms an electron pocket near the M(L) point. Fig. 3(a) and (c), corresponding to 17 eV photon energy, away from $\Gamma$ point, clearly show two separated light Te 5$p$ bands approach the $\Gamma$(A) point and form the hole pockets and a shallow Ti 3$d$ band forms an electron pocket near the L point. For 21 eV photon excitation, close to the $\Gamma$ point, the band structures are more complex. In Fig. 3(b), two broad Te 5\emph{p} related humps are observed. It is obvious from the illustrated momentum distribution curve (MDC) at Fermi energy that it cannot be fitted with two peaks Lorentz lineshape. In Fig. 3(d), at high as 5 Te 5\emph{p} bands and a Ti 3$d$ band can be identified from the 2D second-derivative image. These features are different from the previous band structure calculations, which suggest three Te 5$p$ bands cross the Fermi level along the $\Gamma$-M direction \cite{Strocov2006, Rossnagel2009}. Our results are similar to those predicted by recent articles which pointed out the band should be continuums around the $\Gamma$ point \cite{PChen2017}.
\begin{figure}[tbp]
\begin{center}
\includegraphics[width=0.95\columnwidth,angle=0]{Figure3.pdf}
\end{center}
\vspace*{-0.5cm} \caption{(color online) ARPES maps around $\Gamma$(A) at the zone center of 1$T$-TiTe$_2$. (a) and (b) ARPES maps, taken along the $\Gamma$(A)-M(L) direction with 17 and 21 eV photon energy. The energy resolution was set to 8 and 10 meV, respectively. The inset red lines represent the MDCs at Fermi energy of Ti 5$p$ bands. The corresponding $k_z$ position are indicated by grey lines in Fig. 1. (c) and (d) measured band structures reduced with the 2D second-derivative to enhance the weak bands while maintaining band dispersion.}
\end{figure}
Now we examine the features of the Ti 3$d$ band. Fig. 4(a) shows the high-resolution ARPES spectra of Ti 3$d$ band along the $\Gamma$(A)-M(L) direction with 20 eV photon energy. This photon energy was chosen to obtain very narrow quasiparticle bandwidth, where its final state arc intersects Fermi pocket in a nearly vertical way (see Fig. 1). As been discussed previously \cite{Claessen1992, SHarm1994, JWAllen1995, Perfetti2001, Rossnagel2001, Perfetti2002, Nicolay2006}, the momentum distribution curves (MDCs) of shallowed Ti 3$d$ band can no longer be fitted with Lorentzian lineshapes. An empirical approach is used to obtain its qualitatively dispersion. Fig. 4(b) shows the energy distribution curves(EDCs) band, which was extracted by taking second-derivative of original ARPES data (Fig. 4(a)) with respect to energy. One can see an obvious kink in dispersion $\sim$ 18 meV below Fermi energy, indicated by a red arrow. The original data (Fig. 4(a)) is divided by the Fermi distribution function to visually inspect the kink. A famous double peak (peak-dip-hump) feature can be observed in the Fig. 4(d), which were widely observed in other materials \cite{Hengsberger1999, LaShell2000, Lanzara2001, Kaminski2001}. This is the first experimental observation of a kink structure in 1$T$-TiTe$_2$. However, the kink structure is widespread in other materials, such as Bi (1000) surface \cite{Hengsberger1999, LaShell2000}, High-T$_c$ superconductor \cite{Lanzara2001, Kaminski2001, Bogdanov2000, PDJohnson2001}. The kink structure was also observed in other TMDs with CDW phase \cite{TValla2000, TValla2004}, but not in materials without CDW phase \cite{DWShenPhdThesis}. These features can be understood within a many-body framework via considering the electron interactions with impurities, phonons and other electrons. After comparing the experimental results obtained from the present photoemission measurements with those from Raman spectroscopy \cite{Hangyo1983, JKhan2012, VRajaji2018} and specific heat \cite{Rossnagel2009}, we believe that the kink in the energy dispersion in TiTe$_2$ is induced by electron-phonon coupling. This mechanism is proposed on the basis of following experimental findings. i) A mode ($A$$_1g$) regarding to electron-phonon coupling near 18 meV has been found in the previous Raman spectroscopy studies \cite{Hangyo1983, JKhan2012, VRajaji2018}; ii) The most recent result obtained from specific heat measurement indicates that the phonon effect corresponds to a Debye energy of about 19 meV \cite{Rossnagel2009}; And iii) our ARPES results shown in Fig. 4 indicate that the kink is observed in the low-energy regime around the Fermi-level, which is around the optic-phonon energy measured from Fermi energy in TiTe$_2$.
\begin{figure}[tbp]
\begin{center}
\includegraphics[width=0.95\columnwidth,angle=0]{Figure4.pdf}
\end{center}
\vspace*{-0.5cm} \caption{(color online) ARPES maps for Ti 3$d$ band of 1$T$-TiTe$_2$.(a) band structure plot along $\Gamma$(A)-M(L) direction with 20 eV photon energy. The angle and energy resolution was set to 0.2$^{\circ}$ and $\sim$7 meV, respectively. (b) measured structure reduced with the second-derivative method to enhance the weak bands. The red arrow marks the kink position.(c) measured image divided by the Fermi distribution function to highlight the kink feature.(d) EDCs corresponding to (c). The red dashed line indicated the position of dip.}
\end{figure}
TiTe$_2$ has often been considered as a testing material of Fermi liquid electron gas. However, its 3D features of the electronic band structures are often ignored when spectra are fitted with a Fermi liquid lineshape. The experimental results shown and discussed in this work indicate that although the crystal structure of 1$T$-TiTe$_2$ is mainly 2D-like, its electronic band structure can display very strong 3D-like natures, namely the electronic energy spectrum of 1$T$-TiTe$_2$ depends not only on the in-plane electron wavevector $(k_x,k_y)$ but also on $k_z$ the electron wavevector along the normal direction. A strong dependence of the electronic energy spectrum upon $k_z$ in 1$T$-TiTe$_2$ is basically a consequence that it is with a sandwich-like material structure in which the electrons in different layers are attracted by the van der Waals forces. Due to inter-layer bonding and corresponding electronic interactions, the electronic band structure along the $k_z$ or normal direction should be a functional form of the electronic wavevector or momentum along this direction.
\section{Conclusions}
In this work, we have investigated the low-energy electronic structure of the layered TMDCs 1$T$-TiTe$_2$ single crystal at a relatively low temperature using state-of-the-art technique such as the ARPES. The major aim of this study is at measuring and examining the 3D features of the electronic band structure of 1$T$-TiTe$_2$ material. The main conclusions deduced from the present study are as follows. i) The electronic band structure regarding to both the Ti 3$d$ and the Te 5$p$ states show a strong 3D topology. This result demonstrates that although the crystal structure of 1$T$-TiTe$_2$ is mainly 2D-like, its electronic band structure is very much 3D-like; ii) A typical semi-metal structure can be found in TiTe$_2$ in which a single Fermi level goes across the electron and hole bands so that the electrons and holes can coexist in the material; iii) The multiple $\Gamma$(A) centered hole pockets formed by Te 5$p$ bands and three M(M') around shallow electron pockets formed by Ti 3$d$ band have been experimentally identified. This finding allows us to give a more detailed description of the electronic structure of 1$T$-TiTe$_2$ single crystal; And iv) a kink can be observed in low photon energy regime at approximately 18 meV in Ti 3$d$ band. This feature is induced mainly by electron-phonon coupling in the material. We hope these important and significant experimental findings can help us to gain an in-depth understanding of the basic electronic structure of 1$T$-TiTe$_2$ which has been considered as a testing material for Fermi liquid electron gas.
\section{Acknowledgements}
We thank Gey-Hong Gweon for his unreserved support and Kai Rossnagel for high quality samples and constructive criticism. This work was supported by the National Natural Science Foundation of China (NSFC, Grants No.11574402 and No.51502351) and by the Innovation-driven Plan in Central South University (2016CXS032). WX was supported by the NSFC (Grant No. 11574319) and the Center of Science and Technology of Hefei Academy of Science (Grant No. 2016FXZY002).
|
\section{Introduction}\label{S:Intro}
The design of computer experiments, where observations of a real physical phenomenon are replaced by simulations of a complex mathematical model (e.g., based on PDEs), has emerged as a full discipline, central to uncertainty quantification. The final objective of the simulations is often goal-oriented, that is, precisely defined. It may correspond for example to the optimization of the response of a system with respect to its input factors, or to the estimation of the probability that the response will exceed a given threshold when input factors have a given probability distribution. Achieving this objective generally requires sequential learning of the behavior of the response in a particular domain of interest for input factors: the region where the response is close to its optimum, or is close to the given threshold; see, e.g., the references in \cite{Ginsbourger2017}. When simulations are computationally expensive, sequential inference based on the direct use of the mathematical model is unfeasible due to the large number of simulations required and simplified prediction models, approximating the simulated response, have to be used. A most popular approach relies on Gaussian process modelling, where the response (unknown prior to simulation) is considered as the realization of a Gaussian Random Field (RF), with parameterized mean and covariance, and Bayesian inference gives access to the posterior distribution of the RF (after simulation). Typically, in a goal-oriented approach based on stepwise-uncertainty reduction \cite{BectBG2016, BectGLPV2012}, the prediction model is used to select the input factors to be used for the next simulation, the selection being optimal in terms of predicted uncertainty on the target. The construction of a first, possibly crude, prediction model is necessary to initialize the procedure. This amounts at approximating the behaviour of an unknown function $f$ (the model response) on a compact domain ${\mathscr X}\subset\mathds{R}^d$ (the feasible set for $d$ input factors) from a few evaluations inside the domain. That is the basic design objective we shall keep in mind throughout the paper, although we may use diverted paths where approximation/prediction will be shadowed by other objectives, integration in particular.
In general, little is known about the function {\em a priori}, and it seems intuitively reasonable to spread out points of evaluation across the available space; see \cite{BiedermannD2001}. Such space-filling designs can be obtained by optimizing a geometrical measure of dispersion or a discrepancy criterion measuring distance to uniformity. When using a Gaussian RF model, minimizing the Integrated Mean-Squared Prediction Error (IMSPE) is also a popular approach, although not very much used due to its apparent complexity, see, e.g., \cite{GP-SIAM_2014, GorodetskyK2016}. The paper promotes the use of designs optimized for integration with respect to the uniform measure for their good space-filling properties. It gives a survey of recent results on energy functionals that measure distance to uniformity and places recent approaches proposed for space-filling design, such as \cite{MakJ2017}, in a general framework and perspective encompassing design for integration, construction of the (continuous) Best Linear Unbiased Estimator (BLUE) in a location model with correlated errors, and minimization of energy (kernel discrepancy) for signed measures.
We start by a quick introduction to Bayesian function approximation and integration (Section~\ref{S:RFM}), where the function is considered as the realization of a Gaussian RF with covariance structure defined by some kernel $K$.
Exploiting recent results on the minimization of energy functionals \cite{DamelinHRZ2010, SejdinovicSGF2013, SriperumbudurGFSL2010}, we show in Section~\ref{S:KDEP} that integrally strictly positive definite kernels define strictly convex energy functionals, with an equivalence between the notions of potential and directional derivative that reveals the strong relation between discrepancy minimization and more traditional design of optimal experiments. We show that Bayesian integration is equivalent to the construction of the BLUE in a model with modified correlation structure, so that the two associated design problems coincide. We also show that the posterior variance in Bayesian integration corresponds to the minimum of a squared kernel discrepancy for signed measures with total mass one and to the minimum of an energy functional for a reduced kernel.
Since the posterior variance criterion in Bayesian integration takes a very simple form, its minimization constitutes an attractive alternative to the minimization of the IMSPE criterion for space-filling design. This is considered in Section~\ref{S:Empirical}. We consider in particular kernel herding algorithms from machine learning, which are special instances of vertex-direction methods used in optimal design and can be used for the construction of point sequences with suitable space-filling properties (any-time designs).
Several auxiliary results are given in appendix. Appendix~\ref{S:CV-properties} provides convergence properties of algorithms presented in Section~\ref{S:Empirical}. Extension to design for the simultaneous estimation of several integrals is considered in Appendix~\ref{S:BQ-several}. A Karhunen-Lo\`eve expansion of the RF model is considered in Appendix~\ref{S:KL}, that yields a Bayesian linear model for which minimization of the posterior variance in Bayesian integration corresponds to a c-optimal design problem.
\section{Random-field models for function approximation and integration}\label{S:RFM}
\subsection{Space-filling design and kernel choice for function approximation}
\label{S:GP regression}
Let $K(\cdot,\cdot)$ denote a symmetric positive definite kernel on ${\mathscr X}\times{\mathscr X}$, with associated Reproducing Kernel Hilbert Space (RKHS) ${\mathcal H}_K$. Denote $K_\mathbf{x}(\cdot)=K(\mathbf{x},\cdot)$ and $\langle \cdot,\cdot \rangle_K$ the scalar product in ${\mathcal H}_K$, so that the reproducing property gives $\langle f,K_\mathbf{x} \rangle_K=f(\mathbf{x})$ for any $f\in{\mathcal H}_K$.
Consider first the common framework where the function $f$ to be approximated is supposed to belong to ${\mathcal H}_K$. Let $\eta_n(\mathbf{x})=\sum_{i=1}^n w_i f(\mathbf{x}_i)=\mathbf{w}_n^T\mathbf{y}_n$ be a linear predictor of $f(\mathbf{x})$ based on evaluations of $f$ at the $n$-point design $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}$, with $\mathbf{x}_i\in{\mathscr X}$ for all $i$. Throughout the paper we denote $\mathbf{w}_n=(w_1,\ldots,w_n)^T$, $\mathbf{y}_n=[f(\mathbf{x}_1),\ldots,f(\mathbf{x}_n)]^T$, $\mathbf{k}_n(\cdot)=[K_{\mathbf{x}_1}(\cdot),\ldots,K_{\mathbf{x}_n}(\cdot)]^T$ and $\{\mathbf{K}_n\}_{i,j}=K(\mathbf{x}_i,\mathbf{x}_j)$, $i,j=1,\ldots,n$.
Cauchy-Schwarz inequality gives the classical result
\begin{eqnarray*}
|f(\mathbf{x})-\eta_n(\mathbf{x})| = \left| f(\mathbf{x})- \sum_{i=1}^n w_i f(\mathbf{x}_i)\right| &=& \left| \langle f, K_\mathbf{x}-\sum_{i=1}^n w_i K_{\mathbf{x}_i} \rangle_K \right| \\
&\leq& \|f\|_{{\mathcal H}_K} \, \left\|K_\mathbf{x}-\sum_{i=1}^n w_i K_{\mathbf{x}_i}\right\|_{{\mathcal H}_K} \,,
\end{eqnarray*}
where $\|f\|_{{\mathcal H}_K}$ depends on $f$ but not on $\mathbf{X}_n$, and $\rho_n(\mathbf{x},\mathbf{w})=\left\|K_\mathbf{x}-\sum_{i=1}^n w_i K_{\mathbf{x}_i}\right\|_{{\mathcal H}_K}$ depends on $\mathbf{X}_n$ (and $\mathbf{w}_n$) but not on $f$. Suppose that $\mathbf{K}_n$ has full rank. For a given $\mathbf{X}_n$, the Best Linear Predictor (BLP) minimizes
$\rho_n(\mathbf{x},\mathbf{w})$ and corresponds to $\eta_n^*(\mathbf{x}) = (\mathbf{w}_n^*)^T\mathbf{y}_n$,
with $\mathbf{w}_n^*=\mathbf{w}_n^*(\mathbf{x})=\mathbf{K}_n^{-1}\mathbf{k}_n(\mathbf{x})$, which gives ${\rho_n^*}^2(\mathbf{x})= \rho_n^2(\mathbf{x},\mathbf{w}_n^*) = K(\mathbf{x},\mathbf{x})-\mathbf{k}_n^T(\mathbf{x})\mathbf{K}_n^{-1} \mathbf{k}_n(\mathbf{x})$.
A less restrictive assumption on $f$ is to suppose that it corresponds to a realization of a RF $Z_x$, with zero mean ($\mathbb{E}\{Z_x\}=0$) and covariance $\mathbb{E}\{Z_x Z_{x'}\}=\sigma^2\, K(\mathbf{x},\mathbf{x}')$ for all $\mathbf{x}$, $\mathbf{x}'$ in~${\mathscr X}$, $\sigma>0$. Then, straightforward calculation shows that $\eta_n^*(\mathbf{x})$ is still the BLP (the posterior mean if $Z_x$ is Gaussian), and $\sigma^2\, {\rho_n^*}^2(\mathbf{x})$ is the Mean-Squared Prediction Error (MSPE) at $\mathbf{x}$. This construction corresponds to simple kriging; see, e.g., \cite{AuffrayBM2012, VazquezB2011}.
IMSPE-optimal designs minimize the integrated squared error $\IMSPE(\mathbf{X}_n)= \sigma^2\, \int_{\mathscr X} {\rho_n^*}^2(\mathbf{x}) \mbox{\rm d}\mu(\mathbf{x})$, with $\mu$ generally taken as the uniform probability measure on~${\mathscr X}$,
see, e.g., \cite{GP-SIAM_2014, GorodetskyK2016, SacksWMW89}.
IMSPE-optimal designs $\mathbf{X}_n^*$ depend on the chosen $K$. It is well known that the asymptotic rate of decrease of $\IMSPE(\mathbf{X}_n^*)$ with $n$ depends on the regularity properties of $K$ (the same is true for the integration problem); see for instance \cite{RitterWW95}. It is rather usual to take $K$ stationary (translation invariant), i.e., satisfying $K(\mathbf{x},\mathbf{x}')=\Psi(\mathbf{x}-\mathbf{x}')$ for all $\mathbf{x}$ and $\mathbf{x}'$, with $\Psi$ in some parametric class selected according to prior knowledge on the smoothness properties of $f$. A typical example is the Mat\'ern class of covariances, see \cite[Chap.~2]{Stein99}. On the other hand, for reasons explained in Section~\ref{S:Intro}, computer experiments usually involve small values of $n$, and the asymptotic behavior of the approximation error is hardly observed. Its behavior on a short horizon is much more important and strongly depends on the correlation lengths in $K$, which are difficult to choose {\em a priori}. Robustness with respect to the choice of $K$ favours space-filling designs, where the $\mathbf{x}_i$ are suitably spread over~${\mathscr X}$. Noticeably, it is shown in \cite{Schaback95} that for translation invariant and isotropic kernels (i.e., such that $K(\mathbf{x},\mathbf{x}')=\Psi(\|\mathbf{x}-\mathbf{x}'\|)$, with $\|\cdot\|$ the Euclidean distance in $\mathds{R}^d$), one has $\rho_n^2(\mathbf{x}) \leq S_K[h_r(\mathbf{x})]$ for some increasing function $S_K(\cdot)$. Here $h_r(\mathbf{x})=\max_{\|\mathbf{x}-\mathbf{x}'\|\leq r} \min_{1\leq i \leq n} \|\mathbf{x}'-\mathbf{x}_i\|$ measures the density of design points $\mathbf{x}_i$ around $\mathbf{x}$, with $r$ a fixed positive constant. It satisfies $\max_{\mathbf{x}\in{\mathscr X}} h_r(\mathbf{x}) \geq \max_{\mathbf{x}\in{\mathscr X}} h_0(\mathbf{x})=\CR(\mathbf{X}_n)$, with
\begin{eqnarray}\label{CR}
\CR(\mathbf{X}_n) = \max_{\mathbf{x}\in{\mathscr X}} \min_{1\leq i \leq n} \|\mathbf{x}-\mathbf{x}_i\| \,,
\end{eqnarray}
the \emph{covering radius} of $\mathbf{X}_n$: $\CR(\mathbf{X}_n)$ defines the smallest $r$ such that the $n$ closed balls of radius $r$ centred at the $\mathbf{x}_i$ cover~${\mathscr X}$. $\CR(\mathbf{X}_n)$ is also called the dispersion of $\mathbf{X}_n$ \cite[Chap.~6]{Niederreiter92} and corresponds to the minimax-distance criterion \cite{JohnsonMY90} used in space-filling design.
Loosely speaking, the property $\rho_n^2(\mathbf{x}) \leq S_K[h(\mathbf{x})]$ quantifies the intuition that designs with a small value of $\CR$ provide precise predictions over~${\mathscr X}$ since for any $\mathbf{x}$ in~${\mathscr X}$ there always exists a design point $\mathbf{x}_i$ at proximity where $f(\mathbf{x}_i)$ has been evaluated. Another standard geometrical criterion of spreadness is the \emph{packing radius}
\begin{eqnarray}\label{PR}
\PR(\mathbf{X}_n) = \frac12\, \min_{i\neq j} \|\mathbf{x}_i-\mathbf{x}_j\| \,.
\end{eqnarray}
It corresponds to the largest $r$ such that the $n$ open balls of radius $r$ centred at the $\mathbf{x}_i$ do not intersect; $2 \PR(\cdot)$ corresponds to the maximin-distance criterion \cite{JohnsonMY90} often used in computer experiments.
In this paper, we shall adopt the following point of view. We do not intend to construct designs adapted to a particular $K$ chosen from {\em a priori} knowledge on $f$. Neither shall we estimate the parameters in $K$ (such as correlation lengths) when $K$ is taken from a parametric class. We shall rather consider the kernel $K$ as a tool for constructing a space-filling design, the quality of which will be measured in particular through the value of $\CR$. The motivation is twofold: ($i$) the construction will be much easier than the direct minimization of $\CR$, ($ii$) it will facilitate the construction of \emph{sequences of points} suitably spread over~${\mathscr X}$ (\emph{any-time} space-filling designs).
\subsection{Bayesian quadrature}
\label{S:BQ}
Denote by ${\mathscr M}={\mathscr M}[{\mathscr X}]$ the set of finite signed Borel measures on a nonempty set~${\mathscr X}$, and by ${\mathscr M}(q)$, $q\in\mathds{R}$, the set of signed measures with total mass $q$: ${\mathscr M}(q)=\{\mu\in{\mathscr M}: \mu({\mathscr X})=q\}$. The set of Borel probability measures on~${\mathscr X}$ is denoted by ${\mathscr M}^+(1)$, ${\mathscr M}^+$ is the set of finite positive measures on~${\mathscr X}$. Typical applications correspond to~${\mathscr X}$ being a compact subset of $\mathds{R}^d$ for some $d \in \mathds{N}$.
Suppose we wish to integrate a real function defined on~${\mathscr X}$ with respect to $\mu\in{\mathscr M}^+(1)$.
Assume that $\mathsf{E}_\mu\{|f(X)|\}<+\infty$ and denote
$$
I_\mu(f)=\mathsf{E}_\mu\{f(\mathbf{X})\}=\int_{\mathscr X} f(\mathbf{x})\, \mbox{\rm d}\mu(\mathbf{x}) \,.
$$
We set a prior on $f$, and assume that $f$ is the realization of a Gaussian RF, with covariance $\sigma^2\,K(\cdot,\cdot)$, $\sigma^2>0$, and unknown mean $\beta_0$; that is, we consider the location model with correlated errors
\begin{equation}\label{model1}
f(\mathbf{x})=\beta_0+Z_x \,,
\end{equation}
where $\mathbb{E}\{Z_x\}=0$ and $\mathbb{E}\{Z_x Z_{x'}\}=\sigma^2\, K(\mathbf{x},\mathbf{x}')$ for all $\mathbf{x},\mathbf{x}'\in{\mathscr X}$, $\sigma>0$. Regression models more general than \eqref{model1} are considered in Appendix~\ref{S:BQ-several}. Here $K$ is a symmetric Positive Definite (PD) kernel; that is, $K(\mathbf{x},\mathbf{x}')=K(\mathbf{x}',\mathbf{x})$, and for all $n\in\mathds{N}$ and all pairwise different $\mathbf{x}_1,\ldots,\mathbf{x}_n\in{\mathscr X}$, the matrix $\mathbf{K}_n$ is non-negative definite; if $\mathbf{K}_n$ is positive definite, then $K$ is called Strictly Positive Definite (SPD).
Note that $K^2(\mathbf{x},\mathbf{x}') \leq K(\mathbf{x},\mathbf{x})K(\mathbf{x}',\mathbf{x}')<+\infty$ for all $\mathbf{x},\mathbf{x}'\in{\mathscr X}$ since $K$ corresponds to a covariance.
We will call a general kernel $K$ bounded when $K(\mathbf{x},\mathbf{x})<\infty$ for all $\mathbf{x}\in{\mathscr X}$, and uniformly bounded when there is a constant $C$ such that $K(\mathbf{x},\mathbf{x}) \leq C$ for all $\mathbf{x}\in{\mathscr X}$. Any PD kernel is bounded.
Similarly to Section~\ref{S:GP regression}, we denote by ${\mathcal H}_K$ the associated RKHS and by $\langle \cdot,\cdot \rangle_K$ the scalar product in ${\mathcal H}_K$. The assumption that $K$ is bounded will be relaxed in Section~\ref{S:EP} where we shall also consider singular kernels, but throughout the paper we assume that $K$ is symmetric, $K(\mathbf{x},\mathbf{x}')=K(\mathbf{x}',\mathbf{x})$ for all $\mathbf{x},\mathbf{x}'\in{\mathscr X}$. Also, we always assume, as in \cite[Sect.~2.1]{Fuglede60}, that either $K$ is non-negative on ${\mathscr X} \times {\mathscr X}$, or~${\mathscr X}$ is compact.
We set a vague prior on $\beta_0$ and assume that $\beta_0 \sim {\mathscr N}(\hat\beta_0^0,\sigma^2\,A)$ with $A\rightarrow+\infty$. This amounts to setting $1/A=0$ in all Bayesian calculations; the choice of $\hat\beta_0^0$ is then irrelevant. Suppose that $f$ has been evaluated at the $n$-point design $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}\in{\mathscr X}^n$. We assume that $\mathbf{K}_n$ has full rank. For any $\mathbf{x}\in{\mathscr X}$, the posterior distribution of $f(\mathbf{x})$ (conditional on $\sigma^2$ and $K$) is normal, with mean
\begin{equation*}\label{predict1}
\hat \eta_n(\mathbf{x}) = \hat\beta_0^n + \mathbf{k}_n^T(\mathbf{x})\mathbf{K}_n^{-1}(\mathbf{y}_n-\hat\beta_0^n \boldsymbol{1}_n)
\end{equation*}
and variance (mean-squared error)
\begin{equation}\label{MSE1}
\sigma^2\rho_n^2(\mathbf{x})= \sigma^2\,\left[K(\mathbf{x},\mathbf{x})-\mathbf{k}_n^T(\mathbf{x})\mathbf{K}_n^{-1}\mathbf{k}_n(\mathbf{x})+\frac{(1-\mathbf{k}_n^T(\mathbf{x})\mathbf{K}_n^{-1}\boldsymbol{1}_n)^2}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n} \right] \,,
\end{equation}
where
\begin{equation}\label{beta01}
\hat\beta_0^n = \frac{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\mathbf{y}_n}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n}
\end{equation}
and $\boldsymbol{1}_n$ is the $n$-dimensional vector $(1,\ldots,1)^T$,
see for instance \cite[Chap.~4]{SantnerWN2003}. The posterior mean of $I_\mu(f)$ is thus
\begin{equation}\label{In1}
\widehat I_n = \int_{\mathscr X} \hat \eta_n(\mathbf{x})\, \, \mbox{\rm d}\mu(\mathbf{x}) = \mathsf{E}_\mu\{\hat \eta_n(\mathbf{X})\} = \hat\beta_0^n + \mathbf{p}_n(\mu)^T\mathbf{K}_n^{-1}(\mathbf{y}_n-\hat\beta_0^n \boldsymbol{1}_n)\,,
\end{equation}
with
\begin{eqnarray}\label{hb}
\mathbf{p}_n(\mu)=(P_\mu(\mathbf{x}_1),\ldots,P_\mu(\mathbf{x}_n))^T \,,
\end{eqnarray}
where, for any $\nu\in{\mathscr M}$ and $\mathbf{x}\in{\mathscr X}$, we denote
\begin{eqnarray} \label{kh}
P_\nu(\mathbf{x})= \int_{\mathscr X} K(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\nu(\mathbf{x}') \,.
\end{eqnarray}
$P_\nu(\cdot)$ is called the kernel imbedding of $\nu$ into ${\mathcal H}_K$, see \cite[Def.~9]{SejdinovicSGF2013};
$P_\nu(\mathbf{x})$ is well defined and finite for any $\nu\in{\mathscr M}$ and $\mathbf{x}\in{\mathscr X}$ when $K$ is uniformly bounded. On the other hand, there always exists $\nu\in{\mathscr M}$ such that $P_\nu(\mathbf{x})$ is infinite for all $\mathbf{x}\in{\mathscr X}$ when $K$ is not uniformly bounded on~${\mathscr X}$.
The function $P_\nu(\cdot)$ is called potential in potential theory, see Section~\ref{S:EP}.
Similarly to \eqref{MSE1}, we obtain that the posterior variance of $I_\mu(f)$ becomes
\begin{eqnarray}
\sigma^2 s_n^2 &=& \sigma^2\, \left[ {\mathscr E}_K(\mu) - \mathbf{p}_n^T(\mu) \mathbf{K}_n^{-1} \mathbf{p}_n(\mu) + \frac{(1-\mathbf{p}_n^T(\mu)\mathbf{K}_n^{-1}\boldsymbol{1}_n)^2}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n} \right] \,, \label{sn1}
\end{eqnarray}
where, for any $\nu\in{\mathscr M}$, we denote
\begin{eqnarray} \label{overline K}
{\mathscr E}_K(\nu) = \int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\nu(\mathbf{x})\mbox{\rm d}\nu(\mathbf{x}') \,.
\end{eqnarray}
This is one of the key notions in potential theory, called the energy of $\nu$; see Section~\ref{S:EP}.
For $\mu$ in ${\mathscr M}^+(1)$, we have ${\mathscr E}_K(\mu) = \mathsf{E}_\mu\{K(\mathbf{X},\mathbf{X}')\}$
where $\mathbf{X}$ and $\mathbf{X}'$ are independently identically distributed (i.i.d.) with $\mu$. The quantity $-{\mathscr E}_K(\mu)$ corresponds to the quadratic entropy introduced by C.R. Rao \cite{Rao1982a}; see also Remark~\ref{R:Bregman}.
Define
\begin{eqnarray}\label{M_K^a}
{\mathscr M}_K^\alpha = \left\{\nu\in{\mathscr M}: \int_{\mathscr X} K^\alpha(\mathbf{x},\mathbf{x})\, \mbox{\rm d}|\nu|(\mathbf{x}) < +\infty \right\}\,, \ \alpha>0 \,.
\end{eqnarray}
When $\mu\in{\mathscr M}_K^{1/2}$, the reproducing property and Cauchy-Schwarz inequality imply that
\begin{eqnarray}
{\mathscr E}_K(\mu) &=& \int_{{\mathscr X}^2} \langle K(\cdot,\mathbf{x}),K(\cdot,\mathbf{x}') \rangle_K\, \mbox{\rm d}\mu(\mathbf{x})\mbox{\rm d}\mu(\mathbf{x}') \nonumber \\
&& \leq \left[\int_{\mathscr X} K^{1/2}(\mathbf{x},\mathbf{x})\,\mbox{\rm d}|\mu|(\mathbf{x})\right]^2 < +\infty\,. \label{CS:M1/2}
\end{eqnarray}
When $\beta_0$ is assumed to be known (equal to zero for instance), we simply substitute $\beta_0$ for $\hat\beta_0^n$ in \eqref{In1} and the posterior variance is
\begin{eqnarray}\label{sn10}
\sigma^2 s_{n,0}^2 = \sigma^2\, \left[ {\mathscr E}_K(\mu) - \mathbf{p}_n^T(\mu) \mathbf{K}_n^{-1} \mathbf{p}_n(\mu) \right] \,.
\end{eqnarray}
Bayesian quadrature relies on the estimation of $I_\mu(f)$ by $\widehat I_n$. An optimal design for estimating $I_\mu(f)$ should minimize $s_n^2$ given by \eqref{sn1}. One may refer to \cite{Diaconis88} for a historical perspective and to \cite{HennigOG2015} for a recent exposition on Bayesian numerical computation. The framework presented above corresponds to that considered in \cite{O'Hagan91}, restricted to the case (recommended in that paper) where the known trend function is simply the constant 1 (which corresponds to the presence of an unknown mean $\beta_0$ in the model \eqref{model1}). In Section~\ref{S:Empirical}, we shall see that $s_{n,0}^2$ is equal to the minimum value of a (squared) kernel discrepancy between the measure $\mu$ and a signed measure supported on $\mathbf{X}_n$, and that $s_n^2$ corresponds to the minimum of a squared discrepancy for signed measures that are constrained to have total mass one, and also corresponds to the minimum of an energy functional for a modified kernel $K_\mu$. Note that $\sigma^2 s_n^2 \leq \IMSPE(\mathbf{X}_n)=\sigma^2 \int_{\mathscr X} \rho_n^2(\mathbf{x})\, \mbox{\rm d}\mu(\mathbf{x})$ (which requires $\mu\in{\mathscr M}_K^1\subset{\mathscr M}_K^{1/2}$ to be well defined).
One of the key ideas of the paper is that space-filling design may be based on the minimization of $s_n^2$ rather than the minimization of $\IMSPE(\mathbf{X}_n)$.
\section{Kernel discrepancy, energy and potentials} \label{S:KDEP}
\subsection{Maximum mean discrepancy}\label{S:MMD}
Suppose now that $K$ is bounded and $f$ belongs to the RKHS ${\mathcal H}_K$. Let $\mu$ and $\nu$ be two probability measures in ${\mathscr M}^+(1)\cap{\mathscr M}_K^{1/2}$. Since $f\in{\mathcal H}_K$, using the reproducing property, we obtain $I_\mu(f)=\int_{\mathscr X} \langle f, K_\mathbf{x} \rangle_K \, \mbox{\rm d}\mu(\mathbf{x})$, $I_\nu(f)=\int_{\mathscr X} \langle f, K_\mathbf{x} \rangle_K \, \mbox{\rm d}\nu(\mathbf{x})$ and
$$
\left|I_\mu(f) - I_{\nu}(f) \right| = \left| \int_{\mathscr X} \langle f, K_\mathbf{x} \rangle_K \, \mbox{\rm d}(\mu-\nu)(\mathbf{x})\right| = \left| \langle f, P_\mu - P_\nu \rangle_K \right| \,,
$$
with $P_\mu(\cdot)$ and $P_\nu(\cdot)$ the kernel imbeddings \eqref{kh}. Define
\begin{eqnarray}\label{pseudoM-kernels}
\gamma_K(\mu,\nu) = \| P_\mu - P_\nu\|_{{\mathcal H}_K} \,.
\end{eqnarray}
Cauchy-Schwarz inequality yields the Koksma-Hlawka type inequality \cite[Chap.~2]{Niederreiter92}
$\left|I_\mu(f) - I_{\nu}(f) \right| \leq \| f\|_{{\mathcal H}_K} \gamma_K(\mu,\nu)$, and
\begin{eqnarray}\label{pseudoM-PD}
\gamma_K(\mu,\nu) = \sup_{\|f\|_{{\mathcal H}_K}=1} |I_\mu(f)-I_\nu(f)| \,,
\end{eqnarray}
see, e.g., \cite[Th.~1]{SriperumbudurGFSL2010}. Also, the expansion of $\| P_\mu - P_\nu\|^2_{{\mathcal H}_K}$ gives
\begin{eqnarray}
\gamma_K(\mu,\nu) &=& \left(\|P_\mu\|_{{\mathcal H}_K}^2 + \|P_\nu\|_{{\mathcal H}_K}^2 - 2 \langle P_\mu, P_\nu \rangle_K \right)^{1/2} \nonumber \\
&=& \left( \int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}(\nu-\mu)(\mathbf{x})\, \mbox{\rm d}(\nu-\mu)(\mathbf{x}') \right)^{1/2} \,. \label{kernel discrepancy}
\end{eqnarray}
Therefore, $\gamma_K(\cdot,\cdot)$ is at the same time a pseudometric between kernel imbeddings \eqref{pseudoM-kernels} and an integral pseudometric on probability distributions \eqref{pseudoM-PD}. It defines a kernel discrepancy between distributions \eqref{kernel discrepancy}, $\gamma_K(\cdot,\cdot)$ is also called the \emph{Maximum Mean Discrepancy} (MMD) between $\mu$ and $\nu$ in ${\mathscr M}^+(1)\cap{\mathscr M}_K^{1/2}$, see \cite[Def.~10]{SejdinovicSGF2013}.
To define a metric on the whole ${\mathscr M}^+(1)$, we need $P_\mu$ to be well defined and so that $P_\mu=P_\nu$ for $\mu$ and $\nu$ in ${\mathscr M}^+(1)$ implies $\mu=\nu$.
This corresponds to the notion of \emph{characteristic kernel}, see \cite[Def.~6]{SriperumbudurGFSL2010},
which is closely connected to the following definitions.
\begin{defi}\label{D:ISPD}
A kernel $K$ is \emph{Integrally Strictly Positive Definite} (ISPD) on ${\mathscr M}$ when ${\mathscr E}_K(\nu)>0$ for any nonzero measure $\nu\in{\mathscr M}$.
\end{defi}
\begin{defi}\label{D:CISPD}
A kernel $K$ is \emph{Conditionally Integrally Strictly Positive Definite} (CISPD) on ${\mathscr M}$ when it is ISPD on ${\mathscr M}(0)$; that is, when ${\mathscr E}_K(\nu)>0$ for all nonzero signed measures $\nu\in{\mathscr M}$ such that $\nu({\mathscr X})=0$.
\end{defi}
An ISPD kernel is CISPD.
A bounded ISPD kernel is SPD and defines an RKHS. In \cite[Lemma~8]{SriperumbudurGFSL2010}, the authors show that a uniformly bounded kernel is characteristic if and only if it is CISPD. The proof is a direct consequence of the expression \eqref{kernel discrepancy} for the MMD $\gamma_K(\mu,\nu)$. They also give (Corollary~4) a spectral interpretation of $\gamma_K(\mu,\nu)$ and show that a translation-invariant kernel such that $K(\mathbf{x},\mathbf{x}')=\Psi(\mathbf{x}-\mathbf{x}')$, with $\Psi$ a uniformly bounded continuous real-valued positive-definite function, satisfies, for any $\mu$ and $\nu$ in ${\mathscr M}^+(1)$,
$$
\gamma_K(\mu,\nu) = \left[\int_{\mathds{R}^d} \left|\phi_\mu(\boldsymbol{\omega})-\phi_\nu(\boldsymbol{\omega})\right|^2\, \mbox{\rm d}\Lambda(\boldsymbol{\omega}) \right]^{1/2} \,.
$$
Here, $\phi_\mu$ and $\phi_\nu$ denote the characteristic functions of $\mu$ and $\nu$ respectively and $\Lambda$ is the spectral Borel measure on $\mathds{R}^d$, defined by
\begin{eqnarray}\label{Lambda}
\Psi(\mathbf{x})= \int_{\mathds{R}^d} e^{-i\mathbf{x}^T\boldsymbol{\omega}}\, \mbox{\rm d}\Lambda(\boldsymbol{\omega})\,.
\end{eqnarray}
Using this spectral representation, they prove (Th.~9) that $K$ is characteristic if and only if the support of the associated $\Lambda$ coincides with $\mathds{R}^d$. For example, the sinc squared kernel $K(x,x')=\sin^2[\theta(x-x')]/(x-x')^2$, $\theta>0$, is SPD but is not characteristic (and therefore not CISPD) since the support of $\Lambda$ equals $[-2\theta,2\theta]$.
When $\gamma_K(\mu,\delta_\mathbf{x})$ is well defined for all $\mathbf{x}\in{\mathscr X}$, with $\delta_\mathbf{x}$ the Dirac delta measure at $\mathbf{x}$ (and thus in particular when $K$ is characteristic), we may consider the empirical measure $\xi_{n,e}=\xi_{n,e}(\mathbf{X}_n)=(1/n) \sum_{i=1}^n \delta_{\mathbf{x}_i}$ associated with a given design $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}$, and $\gamma_K(\mu,\xi_{n,e})$ of \eqref{pseudoM-PD} gives the worst-case integration error for $\xi_{n,e}$ when $f$ has norm one in ${\mathcal H}_K$; see Section~\ref{S:empirical}.
Typical examples of uniformly bounded ISPD, and therefore characteristic, kernels are the squared exponential kernel $K_t(\mathbf{x},\mathbf{x}')=\exp(-t\,\|\mathbf{x}-\mathbf{x}'\|^2)$, $t>0$, and the isotropic Mat\'ern kernels, in particular
\begin{eqnarray}\label{K32}
K_{3/2,\theta}(\mathbf{x},\mathbf{x}')=(1+\sqrt{3}\theta\,\|\mathbf{x}-\mathbf{x}'\|)\, \exp(-\sqrt{3}\theta\,\|\mathbf{x}-\mathbf{x}'\|) \ \mbox{ (Mat\'ern 3/2)}\,,
\end{eqnarray}
and $K_{5/2,\theta}(\mathbf{x},\mathbf{x}')=[1+\sqrt{5}\theta\, \|\mathbf{x}-\mathbf{x}'\|+ 5\theta^2\, \|\mathbf{x}-\mathbf{x}'\|^2/3]\,\exp(-\sqrt{5}\theta\, \|\mathbf{x}-\mathbf{x}'\|)$ (Mat\'ern 5/2), see, e.g., \cite{Stein99}.
Two other important examples are given hereafter.
\begin{example}[Generalized multiquadric kernel]\label{Ex:integral-Gaussian-kernels}
The sum of ISPD kernels is ISPD. Since the squared exponential kernel $K_t(\mathbf{x},\mathbf{x}')$ is ISPD for any $t>0$, the integrated kernel obtained by setting a probability distribution on $t$ is ISPD too. One may thus consider
$K(\mathbf{x},\mathbf{x}')= \int K_t(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\pi(t)$ for $\pi$ bounded and non decreasing on $[0,+\infty)$, which generates the class of continuous isotropic autocovariance functions in arbitrary dimension, see \cite{Schoenberg38} and \cite[p.~44]{Stein99}. In particular, for any $\epsilon>0$ and $s>0$, we obtain
$$
K(\mathbf{x},\mathbf{x}')= \int_0^{+\infty} K_t(\mathbf{x},\mathbf{x}')\, t^{s/2-1}\, \exp(-\epsilon\,t) \, \mbox{\rm d} t = \frac{\Gamma(s/2)}{(\|\mathbf{x}-\mathbf{x}'\|^2+\epsilon)^{s/2}} \,,
$$
showing that the generalized multiquadric kernel
\begin{eqnarray}\label{Kes}
K_{s,\epsilon}(\mathbf{x},\mathbf{x}')=(\|\mathbf{x}-\mathbf{x}'\|^2+\epsilon)^{-s/2}\,, \ \epsilon > 0\,, \ s>0 \,,
\end{eqnarray}
is ISPD, see also \cite[Sect.~3.2]{SriperumbudurGFSL2010}.
\fin
\end{example}
\begin{example}[distance-induced kernels]\label{Ex:Riesz<0}
Consider the kernels defined by
\begin{eqnarray}\label{Riesz<0}
K^{(s)}(\mathbf{x},\mathbf{x}') = - \|\mathbf{x}-\mathbf{x}'\|^{s} \,, \ s>0 \,,
\end{eqnarray}
which are CISPD for $s\in(0,2)$ \cite{SzekelyR2013}, and the related distance-induced kernels
$$
K'^{(s)}(\mathbf{x},\mathbf{x}')=\|\mathbf{x}\|^{s}+\|\mathbf{x}'\|^{s}-\|\mathbf{x}-\mathbf{x}'\|^{s} \,, \ s>0 \,.
$$
Note that ${\mathscr E}_{K'^{(s)}}(\mu)={\mathscr E}_{K^{(s)}}(\mu)$ when $\mu({\mathscr X})=0$; in \cite{SzekelyR2013} ${\mathscr E}_{K'^{(s)}}$ is called energy distance for $s=1$ and generalized energy distance for general $s\in(0,2]$.
For $s>0$, the set ${\mathscr M}_{K'^{(s)}}$ contains all signed measures $\mu$ such that $\int_{\mathscr X} \|\mathbf{x}-\mathbf{x}_0\|^{s}\, \mbox{\rm d}|\mu|(\mathbf{x})<+\infty$ for some $\mathbf{x}_0\in{\mathscr X}$. This result is a direct consequence of the triangular inequality when $s\in(0,1]$; for $s>1$ it follows from considerations involving semimetrics generated by kernels, see \cite[Remark~21]{SejdinovicSGF2013}.
$K'^{(s)}$ is CISPD for $s\in(0,2)$ ($K'^{(s)}/2$ corresponds to the covariance function of the fractional Brownian motion), but is not SPD (one has in particular, $K'^{(s)}(\boldsymbol{0},\boldsymbol{0})=0$); $K'^{(2)}$ is not CISPD since ${\mathscr E}_{K'^{(2)}}(\mu)=[\int_{\mathscr X} \mathbf{x}^T\,\mbox{\rm d}\mu(\mathbf{x})][\int_{\mathscr X} \mathbf{x}\,\mbox{\rm d}\mu(\mathbf{x})]$, $\mu\in{\mathscr M}$.
$K(x,x')=1-K^{(1)}(x,x')=1-|x-x'|$ is ISPD for ${\mathscr X}=[0,1]$.
\fin
\end{example}
\subsection{Energy and potentials}\label{S:EP}
In this section we extend the considerations of previous section to signed measures and kernels which may have singularity on the diagonal. Definitions~\ref{D:ISPD} and \ref{D:CISPD} extend to singular kernels, with Riesz kernels as typical examples.
\begin{example}[Riesz kernels]\label{Ex:Riesz}
These fundamental kernels of potential theory are defined by
\begin{eqnarray}\label{Riesz}
K_{(s)}(\mathbf{x},\mathbf{x}')=\|\mathbf{x}-\mathbf{x}'\|^{-s} \,, \ s>0 \,, \mbox{ and } K_{(0)}(\mathbf{x},\mathbf{x}')=-\log \|\mathbf{x}-\mathbf{x}'\| \,,
\end{eqnarray}
with $\mathbf{x},\mathbf{x}'\in{\mathscr X}\subset\mathds{R}^d$ and $\|\cdot\|$ the Euclidean norm. When $s\geq d$, ${\mathscr E}_{K_{(s)}}(\mu)$ is infinite for any nonzero signed measure, but for $s\in(0,d)$ $K_{(s)}$ is ISPD.
Since the logarithmic kernel $K_{(0)}(\mathbf{x},\mathbf{x}')$ has singularity at zero and tends to $-\infty$ when $\|\mathbf{x}-\mathbf{x}'\|$ tends to $+\infty$, it will only be considered for~${\mathscr X}$ compact; $K_{(0)}$ is CISPD, see \cite[p.~80]{Landkof1972}.
\fin
\end{example}
\vsp
Consider again ${\mathscr E}_K(\mu)$ given by \eqref{overline K}, with $\mu\in{\mathscr M}$. In potential theory, this quantity is called the energy of the signed measure $\mu$ for the kernel $K$. Denote
$$
{\mathscr M}_K = \left\{\nu\in{\mathscr M}: |{\mathscr E}_K(\nu)| < +\infty \right\} \,.
$$
In the following, we shall only consider kernels that are at least CISPD.
When $K$ is ISPD, ${\mathscr E}_K(\mu)$ is positive for any nonzero $\mu\in{\mathscr M}$, but when $K$ is only CISPD, ${\mathscr E}_K(\mu)$ can be negative; this is the reason for the presence of absolute value in the definition of ${\mathscr M}_K$. Note that ${\mathscr M}_K$ is the set of measures such that ${\mathscr E}_K(\mu^+)$, ${\mathscr E}_K(\mu^-)$ and ${\mathscr E}_K(\mu^+,\mu^-)=\int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\,\mbox{\rm d}\mu^+(\mathbf{x})\mbox{\rm d}\mu^-(\mathbf{x}')$ are all finite, with $\mu^+$ and $\mu^-$ denoting the positive and negative parts of the Hahn-Jordan decomposition $\mu=\mu^+-\mu^-$ of $\mu$, see \cite[Sect.~2.1]{Fuglede60}. Also note that when $K$ is bounded and defines an RKHS, ${\mathscr M}_K^\alpha \subset {\mathscr M}_K$ for any $\alpha\geq 1/2$, see \eqref{M_K^a} and \eqref{CS:M1/2}; when $K$ is uniformly bounded, ${\mathscr M}_K={\mathscr M}$.
For any $\mu\in{\mathscr M}_K$, $P_\mu(\mathbf{x})$ given by \eqref{kh} is called the potential at $\mathbf{x}$ associated with ${\mathscr E}_K(\mu)$. It is well-defined, with values in $\mathds{R}\cup\{-\infty,+\infty\}$, when $P_{\mu^+}(\mathbf{x})$ and $P_{\mu^-}(\mathbf{x})$ are not both infinite. Also, $P_\mu(\mathbf{x})$ is finite for $\mu$-almost any $\mathbf{x}$, even if $K$ is singular, when $\mu\in{\mathscr M}^+(1)\cap{\mathscr M}_K^{1/2}$.
When $K$ is ISPD, we can still define MMD through \eqref{kernel discrepancy},
\begin{eqnarray}\label{kernel discrepancyB}
\gamma_K(\mu,\nu) = {\mathscr E}_K^{1/2}(\nu-\mu) \,,
\end{eqnarray}
since ${\mathscr E}_K(\nu-\mu)$ is nonnegative whenever defined. The set ${\mathscr M}_K$ forms a pre-Hilbert space, with scalar product the mutual energy ${\mathscr E}_K(\mu,\nu)=\int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\mu(\mathbf{x})\mbox{\rm d}\nu(\mathbf{x}')$ and norm ${\mathscr E}_K^{1/2}(\mu)$.
Denote by ${\mathscr P}_K$ the linear space of potential fields $P_\mu(\cdot)$, $\mu\in{\mathscr M}_K$; when $K$ defines an RKHS ${\mathcal H}_K$, $\|P_\mu\|_{{\mathcal H}_K}={\mathscr E}_K^{1/2}(\mu)$ so that ${\mathscr P}_K\subset{\mathcal H}_K$, and ${\mathscr P}_K$ is dense in ${\mathcal H}_K$. For ${\mathscr P}_K$ to contain all functions $K_\mathbf{x}(\cdot)=K(\cdot,\mathbf{x})$, $\mathbf{x}\in{\mathscr X}$, we need $\delta_\mathbf{x}\in{\mathscr M}_K$ for all $\mathbf{x}$, which requires $K(\mathbf{x},\mathbf{x})<\infty$ for all $\mathbf{x}\in{\mathscr X}$.
For $\mu,\nu\in{\mathscr M}_K$, ${\mathscr E}_K(\mu,\nu)$ defines a scalar product $\langle P_\mu,P_\nu \rangle_{{\mathscr P}_K}$ on ${\mathscr P}_K$, with $\gamma_K(\mu,\nu)=\|P_\mu-P_\nu\|_{{\mathscr P}_K}$. Similarly to Section~\ref{S:MMD}, we obtain
\begin{eqnarray}
\gamma_K(\mu,\nu) &=& \sup_{\xi\in{\mathscr M}_K,\, {\mathscr E}_K(\xi)=1} \left| \int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\xi(\mathbf{x})\mbox{\rm d}(\mu-\nu)(\mathbf{x}') \right| \label{MMD-bis} \\
&=& \sup_{\|h\|_{{\mathscr P}_K}\leq 1} \left|I_\mu(h) - I_{\nu}(h) \right| \,; \nonumber
\end{eqnarray}
that is, a result that extends \eqref{pseudoM-PD} to general ISPD kernels.
If $K$ is only CISPD, we can also define $\gamma_K(\mu,\nu)$ in the same way when considering measures $\mu,\nu\in{\mathscr M}(1)$; we then define ${\mathscr P}_K$ as the linear space of potentials fields $P_\mu(\cdot)$, $\mu\in{\mathscr M}_K\cap{\mathscr M}(0)$, and in \eqref{MMD-bis} we restrict $\xi$ to be in ${\mathscr M}(0)$.
When $K$ is singular, there always exists $\nu$ in ${\mathscr M}_K$ such that $P_\nu(\mathbf{x}_0)=+ \infty$ for some $\mathbf{x}_0$. Consider for example Riesz kernel $K_{(s)}(\mathbf{x},\mathbf{x}')$ with $s\in(0,d)$; ${\mathscr M}_K$ contains in particular all signed measures with compact support $\mathbb{S}(\mu)$ whose potential $P_\mu(\mathbf{x})$ is bounded on $\mathbb{S}(\mu)$, see \cite[p.~81]{Landkof1972}. Take $\nu$ as the measure with density $c/\|\mathbf{x}-\mathbf{x}_0\|^{s-d}$ on~${\mathscr X}$, with $\mathbf{x}_0\in{\mathscr X}$; we have ${\mathscr E}_{K_{(s)}}(\nu)<\infty$ for~${\mathscr X}$ compact, but $P_\nu(\mathbf{x}_0)=+\infty$. As a consequence, as noted in \cite{DamelinHRZ2010}, singular kernels have little interest for integration. Indeed, take $\mu,\nu\in{\mathscr M}_K$ and $h=P_\nu\in{\mathscr P}_K$, then $|I_\mu(h)| \leq \|h\|_{{\mathscr P}_K}\,{\mathscr E}_K^{1/2}(\mu)={\mathscr E}_K^{1/2}(\nu){\mathscr E}_K^{1/2}(\mu)<\infty$ whereas $|I_{\xi_n}(h)|$ may be infinite for some discrete approximation $\xi_n$ of $\mu$ as $h$ can be infinite at some points. Singular kernels may nevertheless be used for the construction of space-filling designs, see for instance the example in Section~\ref{Ex:4.1}, and this is our motivation for considering them in the following.
The key difficulty with singular kernels is the fact that delta measures do not belong to ${\mathscr M}_K$.
An expedient solution to circumvent the problem is replace a singular kernel with a bounded surrogate. For instance, in space-filling design we may replace Riesz kernel $K_{(s)}$, $s>0$, by a generalized inverse multiquadric kernel $K_{s,\epsilon}$ given by \eqref{Kes}, and consider the limiting behaviour of the designs obtained when $\epsilon \rightarrow 0$, see Section~\ref{S:empirical}.
\subsection{Minimum energy and equilibrium measures}\label{S:minE-measures}
In this section, we show that there exist strong connections between results in potential theory and optimal design theory, where one minimizes a convex functional of $\mu\in{\mathscr M}^+(1)$, with the particularity that here the functional is quadratic.
\subsubsection{ISPD kernels and convexity of ${\mathscr E}_K(\cdot)$}
\begin{lem}\label{P:convexity1} \mbox{}
$K$ is ISPD if and only if ${\mathscr M}_K$ is convex and ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}_K$.
\end{lem}
\noindent{\em Proof.} \mbox{}
For any $K$, any $\mu$ and $\nu$ in ${\mathscr M}_K$ and any $\alpha\in[0,1]$, direct calculation gives
\begin{eqnarray} \label{convexity1}
(1-\alpha)\,{\mathscr E}_K(\mu)+\alpha\,{\mathscr E}_K(\nu) - {\mathscr E}_K[(1-\alpha)\mu+\alpha\nu] = \alpha(1-\alpha)\, {\mathscr E}_K(\nu-\mu) \,.
\end{eqnarray}
Assume that $K$ is ISPD. For any $\mu$ and $\nu$ in ${\mathscr M}_K$, the mutual energy ${\mathscr E}_K(\mu,\nu)$ satisfies $|{\mathscr E}_K(\mu,\nu)|\leq \sqrt{{\mathscr E}_K(\mu){\mathscr E}_K(\nu)} < +\infty$. Therefore, ${\mathscr E}_K(\mu-\nu)={\mathscr E}_K(\mu)+{\mathscr E}_K(\nu)-2\,{\mathscr E}_K(\mu,\nu)$ is finite and \eqref{convexity1} implies that ${\mathscr E}_K[(1-\alpha)\mu+\alpha\nu]$ is finite, showing that ${\mathscr M}_K$ is convex. Since $K$ is ISPD, ${\mathscr E}_K(\nu-\mu)>0$ for $\mu,\nu\in{\mathscr M}$, $\nu\neq \mu$, and \eqref{convexity1} implies that ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}_K$.
Conversely, assume that ${\mathscr M}_K$ is convex and ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}_K$. Any $\xi\in{\mathscr M}_K$ can be written as $\xi=\nu-\mu$ with, for instance, $\nu=2\xi$ and $\mu=\xi$, both in ${\mathscr M}_K$. If ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}_K$, \eqref{convexity1} with $\alpha\in(0,1)$ implies that ${\mathscr E}_K(\xi)>0$ when $\nu\neq \mu$, that is, when $\xi\neq 0$. Therefore, $K$ is ISPD.
\carre
\vsp
Lemma~\ref{P:convexity1} also applies to singular kernels. The lemma below concerns CISPD kernels, which are assumed to be uniformly bounded.
\begin{lem}\label{P:convexity2} \mbox{}
Assume that $K$ is uniformly bounded. Then, $K$ is CISPD if and only if ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}(1)$.
\end{lem}
\noindent{\em Proof.} \mbox{}
Since $K$ is uniformly bounded, ${\mathscr M}_K={\mathscr M}$. Assume that $K$ is CISPD. Then, ${\mathscr E}_K(\nu-\mu)>0$ for any $\mu\neq\nu\in{\mathscr M}(1)$, and \eqref{convexity1} implies that ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}(1)$.
Assume now that ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}(1)$. Take any non-zero signed measure $\xi$ in ${\mathscr M}(0)$ and consider the Hahn-Jordan decomposition $\xi=\xi^+-\xi^-$, with $\xi^+({\mathscr X})=\xi^-({\mathscr X})=c>0$. Denote $\nu=\xi^+/c$, $\mu=\xi^-/c$, with $\nu$ and $\mu$ in ${\mathscr M}^+(1)$ ($\nu$ and $\mu$ are in ${\mathscr M}_K$ since $K$ is uniformly bounded). Then, for any $\alpha\in(0,1)$, \eqref{convexity1} and the strict convexity of ${\mathscr E}_K(\cdot)$ on ${\mathscr M}(1)$ gives
${\mathscr E}_K(\xi) = c^2\, {\mathscr E}_K(\nu-\mu) >0$.
\carre
\vsp
Note that one may replace ${\mathscr M}(1)$ by ${\mathscr M}^+(1)$, or by any ${\mathscr M}(\gamma)$ with $\gamma\neq 0$, in Lemma~\ref{P:convexity2}.
\subsubsection{Minimum-energy probability measures}
In the remaining part of Section~\ref{S:minE-measures}, we assume that $K$ is such that ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}^+(1)\cap{\mathscr M}_K$ and ${\mathscr M}(1)\cap{\mathscr M}_K$, which is true under the conditions of Lemma~\ref{P:convexity1} or Lemma~\ref{P:convexity2}.
For $\mu,\nu\in{\mathscr M}_K$, denote by $F_K(\mu;\nu)$ the directional derivative of ${\mathscr E}_K(\cdot)$ at $\mu$ in the direction $\nu$,
$$
F_K(\mu;\nu) = \lim_{\alpha\rightarrow 0^+} \frac{{\mathscr E}_K[(1-\alpha)\mu+\alpha\nu]-{\mathscr E}_K(\mu)}{\alpha} \,.
$$
Straightforward calculation gives
\begin{eqnarray} \label{Dir-der}
F_K(\mu;\nu) = 2 \left[ \int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\nu(\mathbf{x})\mbox{\rm d}\mu(\mathbf{x}') - {\mathscr E}_K(\mu) \right] \,.
\end{eqnarray}
In particular, for any $\mathbf{x}\in{\mathscr X}$, the potential $P_\mu(\mathbf{x})$ associated with $\mu$ at $\mathbf{x}$ satisfies
$$
P_\mu(\mathbf{x}) = \frac12 \, F_K(\mu;\delta_\mathbf{x}) + {\mathscr E}_K(\mu) \,.
$$
\begin{remark}[Bregman divergence and Jensen difference]\label{R:Bregman}
The strict convexity of ${\mathscr E}_K(\cdot)$ implies that ${\mathscr E}_K(\nu) \geq {\mathscr E}_K(\mu)+F_K(\mu,\nu)$ for any $\mu,\nu\in{\mathscr M}_K$, with equality if and only if $\nu=\mu$. This can be used to define a Bregman divergence between measures in ${\mathscr M}_K$ (and thus between probability measures in ${\mathscr M}^+(1)\cap{\mathscr M}_K$), as
$$
B_K(\mu,\nu) = {\mathscr E}_K(\nu)-[{\mathscr E}_K(\mu)+F_K(\mu,\nu)] \,;
$$
see \cite{RaoN85}.
Direct calculation gives $B_K(\mu,\nu)={\mathscr E}_K(\nu-\mu)$ (with therefore $B_K(\mu,\nu)=B_K(\nu,\mu)$), providing another interpretation for the MMD $\gamma_K(\mu,\nu)$, see \eqref{kernel discrepancyB}.
The squared MMD is also proportional to dissimilarity coefficient, or Jensen difference, $\Delta_J(\mu,\nu)=(1/2)[{\mathscr E}_K(\mu)+{\mathscr E}_K(\nu)]-{\mathscr E}_K[\mu+\nu)/2]$ of \cite{Rao1982a}; indeed, direct calculation gives $\gamma_K^2(\mu,\nu)={\mathscr E}_K(\nu-\mu)=4\,\Delta_J(\mu,\nu)$.
\fin
\end{remark}
\vsp
Since ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}^+(1)$, there exists a unique minimum-energy probability measure. The measure $\mu_K^+\in{\mathscr M}^+(1)$ is the minimum-energy measure if and only if $F_K(\mu_K^+;\nu)\geq 0$ for all $\nu\in{\mathscr M}^+(1)$, or equivalently, since $\nu$ is a probability measure, if and only if $F_K(\mu_K^+;\delta_\mathbf{x})\geq 0$ for all $\mathbf{x}\in{\mathscr X}$. We thus obtain the following property, called equivalence theorem in the optimal-design literature.
\begin{theo}\label{P:ET-P}
When ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}^+(1)\cap{\mathscr M}_K$, $\mu_K^+\in{\mathscr M}^+(1)$ is the minimum-energy probability measure on~${\mathscr X}$ if and only if
$$
\forall\, \mathbf{x}\in{\mathscr X}\,, \ P_{\mu_K^+}(\mathbf{x}) \geq {\mathscr E}_K(\mu_K^+) \,.
$$
\end{theo}
Note that, by construction, $\int_{\mathscr X} P_{\mu_K^+}(\mathbf{x}) \,\mbox{\rm d}\mu_K^+(\mathbf{x}) = {\mathscr E}_K(\mu_K^+)$, implying $P_{\mu_K^+}(\mathbf{x})={\mathscr E}_K(\mu_K^+)$ on the support of $\mu_K^+$.
The quantity $C_K^+=[\inf_{\mu\in{\mathscr M}^+(1)} {\mathscr E}_K(\mu)]^{-1}$, with $K$ an ISPD kernel, is called the \emph{capacity} of~${\mathscr X}$ in potential theory; note that $C_K^+ \geq 0$.
The minimizing measure $\mu_K^+\in{\mathscr M}^+(1)$ is called the \emph{equilibrium measure} of~${\mathscr X}$ ($\mu_K^+$ is sometimes renormalized into $C_K^+\,\mu_K^+$, see \cite[p.~138]{Landkof1972}). Theorem~\ref{P:ET-P} thus gives a necessary and sufficient condition for a probability measure $\mu$ to be the equilibrium measure of~${\mathscr X}$.
\begin{example}[Continuation of Example~\ref{Ex:Riesz<0}]\label{Ex:Riesz<0,d>=2}
Properties of minimum-energy probability measures $\mu^+=\mu_{K^{(s)}}^+$ for $K^{(s)}$ given by \eqref{Riesz<0} with~${\mathscr X}$ a compact subset of $\mathds{R}^d$, $d\geq 2$, are investigated in \cite{Bjorck56} and \cite{PWZ2016-SP}. The mass of $\mu^+$ is concentrated on the boundary of~${\mathscr X}$, and its support only comprises extreme points of the convex hull of~${\mathscr X}$ when $s>1$; for $0<s<2$, $\mu^+$ is unique; it is supported on no more than $d+1$ points when $s>2$.
Take ${\mathscr X}={\mathscr B}_d(0,1)$. For symmetry reasons, $\mu^+$ for $0<s<2$ is uniform on the unit sphere ${\mathscr S}_d(\boldsymbol{0},1)$ and
$$
{\mathscr E}_{K^{(s)}}(\mu^+) = -\int_{{\mathscr X}^2} \|\mathbf{x}-\mathbf{x}'\|^{s} \, \mbox{\rm d}\mu^+(\mathbf{x})\mbox{\rm d}\mu^+(\mathbf{x}') = -\int_{{\mathscr X}} \|\mathbf{x}_0-\mathbf{x}'\|^{s} \, \mbox{\rm d}\mu^+(\mathbf{x}')\,,
$$
where $\mathbf{x}_0=(1,0,\ldots,0)^T$. Denote by $\psi_d(\cdot)$ the density of the first component $t=x'_1$ of $\mathbf{x}'=(x'_1,\ldots,x'_d)^T$. We obtain
$\psi_d(t) = (d-1)\, V_{d-1}\,(1-t^2)^{(d-3)/2}/(d\, V_d)$ and
$$
{\mathscr E}_{K^{(s)}}(\mu^+) = -\int_{-1}^1 [(1-t)^2+1-t^2]^{s/2}\, \psi_d(t)\mbox{\rm d} t = -\frac {{2}^{d-q-2}\Gamma (d/2) \Gamma[(d+s-1)/2]}{\sqrt {\pi }\Gamma(d+s/2-1)} \,.
$$
In particular, ${\mathscr E}_{K^{(1)}}(\mu^+)=-4/\pi$ when $d=2$ and is a decreasing function of $d$. When $s=2$, the uniform distribution on the unit sphere is also optimal, and the minimum energy equals $-2$ for all $d\geq 1$, but $\mu^+$ is not unique and the measure allocating equal weight $1/(d+1)$ at each of the $d+1$ vertices of a $d$ regular simplex with vertices on the unit sphere is optimal too.
\fin
\end{example}
\begin{example}[Continuation of Example~\ref{Ex:Riesz}]\label{Ex:Riesz-B_2}
Consider Riesz kernels $K_{(s)}$, see \eqref{Riesz}, for ${\mathscr X}={\mathscr B}_d(\boldsymbol{0},1)$, the closed unit ball in $\mathds{R}^d$.
When $s\geq d$, ${\mathscr E}_{K_{(s)}}(\nu)$ is infinite for any non-zero $\nu\in{\mathscr M}$, but for $0<s<d$ there exists a minimum-energy probability measure $\mu^+=\mu_{K_{(s)}}^+$. When $d>2$ and $s\in(0,d-2]$, $\mu^+$ is uniform on the unit sphere ${\mathscr S}_d(\boldsymbol{0},1)$ (the boundary of~${\mathscr X}$); the potential at all interior points satisfies $P_{\mu^+}(\mathbf{x}) \geq {\mathscr E}_{K_{(s)}}(\mu^+)$ with strict inequality when $s\in(0,d-2)$. When $s\in(d-2,d)$, $\mu^+$ has a density $\varphi_s(\cdot)$ in ${\mathscr B}_d(\boldsymbol{0},1)$,
$$
\varphi_s(\mathbf{x})= \frac{\pi^{-d/2}\,\Gamma(1+s/2)}{\Gamma[1-(d-s)/2]}\; \frac{1}{(1-\|\mathbf{x}\|^2)^{(d-s)/2}} \,,
$$
and the potential $P_{\mu^+}(\cdot)$ is constant in ${\mathscr B}_d(\boldsymbol{0},1)$, see, e.g., \cite[p.~163]{Landkof1972}.
When $d\leq 2$ and $s=0$, $\mu^+$ has a density in ${\mathscr B}_2(\boldsymbol{0},1)$ and $P_{\mu^+}(\cdot)={\mathscr E}_{K_{(0)}}(\mu^+)$ in ${\mathscr B}_2(\boldsymbol{0},1)$. In particular, for $d=1$, $\mu^+$ has the arsine density $1/(\pi\sqrt{1-x^2})$ in $[-1,1]$ with potential $P_{\mu^+}(x) = \log(2)-\log(||x|+\sqrt{x^2-1}|)$, $x\in\mathds{R}$ (and $P_{\mu^+}(x)=\log(2)$ for $x\in[-1,1]$).
The energy ${\mathscr E}_{K_{(s)}}$ is infinite for empirical measures associated with $n$-point designs $\mathbf{X}_n$. One may nevertheless consider the ``physical'' energy
\begin{eqnarray}\label{physical energy}
\widetilde{\mathscr E}_{K_{(s)}}(\mathbf{X}_n)=[2/n(n-1)] \sum_{1\leq i<j\leq n} \|\mathbf{x}_i-\mathbf{x}_j\|^{-s}
\end{eqnarray}
($\widetilde{\mathscr E}_{K_{(s)}}(\mathbf{X}_n)=-[2/n(n-1)] \sum_{1\leq i<j\leq n} \log\|\mathbf{x}_i-\mathbf{x}_j\|$ when $s=0$), which is finite provided that all $\mathbf{x}_i$ are distinct, see \cite{DamelinHRZ2010}. An $n$-point set $\mathbf{X}_n^*$ minimizing $\widetilde{\mathscr E}_{K_{(s)}}(\mathbf{X}_n)$ is called a set of Fekete points, and the limit $\lim_{n\rightarrow\infty} \widetilde{\mathscr E}_{K_{(s)}}^{-1}(\mathbf{X}_n^*)$ exists and is called the transfinite diameter of~${\mathscr X}$. A major result in potential theory, see, e.g., \cite{HardinS2004}, is that the transfinite diameter coincides with the capacity $C_{K_{(s)}}^+$ of~${\mathscr X}$. If $C_{K_{(s)}}^+>0$, then $\mu_{K_{(s)}}^+$ is the weak limit of a sequence of empirical probability measures associated with Fekete points in $\mathbf{X}_n^*$. In the example considered, $\widetilde{\mathscr E}_{K_{(s)}}(\mathbf{X}_n^*)$ tends to infinity when $s\geq d$, but any sequence of Fekete points is asymptotically uniformly distributed in~${\mathscr X}$; $\widetilde{\mathscr E}_{K_{(s)}}(\mathbf{X}_n^*)$ grows like $n^{s/d-1}$ for $s>d$ (and like $\log n$ for $s=d$).
\fin
\end{example}
\begin{remark}[Stein variational gradient descent and energy minimization]
Variational inference using smooth transform based on kernelized Stein discrepancy provides a gradient descent method for the approximation of a target distribution; see \cite{LiuW2016} and the references therein. The fact that the construction does not require knowledge of the normalizing constant of the target distribution makes the method particularly attractive for approximating a posterior distribution in Bayesian inference. Direct calculation shows that when the kernel is translation invariant and the target distribution is uniform, then the method corresponds to steepest descent for the minimization of ${\mathscr E}_K(\xi_{n,e})$; that is, at iteration $k$ each design point $\mathbf{x}_i^{(k)}$ is updated into
$$
\mathbf{x}_i^{(k+1)}=\mathbf{x}_i^{(k)}+ \gamma\, \sum_{i<j} \frac{\partial K(\mathbf{x},\mathbf{x}_j^{(k)})}{\partial \mathbf{x}}\bigg|_{\mathbf{x}=\mathbf{x}_i^{(k)}}
$$
for some $\gamma>0$.
The construction of space-filling design through energy minimization has already been considered in the literature. For instance, it is suggested in \cite{AudzeE77} to construct designs in a compact subset~${\mathscr X}$ of $\mathds{R}^d$ by minimizing $\widetilde{\mathscr E}_{K_{(2)}}(\mathbf{X}_n)$ given by \eqref{physical energy} (note that for $d\geq 3$ design points constructed in this way are not asymptotically uniformly distributed in~${\mathscr X}$). This approach tends to push points to the border of~${\mathscr X}$, similarly to the maximization of the packing radius $\PR(\mathbf{X}_n)$ defined by \eqref{PR}. This is generally not desirable, especially when $d$ is large.
\fin
\end{remark}
\subsubsection{Minimum-energy signed measures}
The situation is slightly different from that in previous section when we consider measures in ${\mathscr M}(1)$.
In that case, $\mu_K^*$ is the minimum-energy measure in ${\mathscr M}(1)$ if and only if $F_K(\mu_K^*;\nu)= 0$ for all $\nu\in{\mathscr M}(1)$, this condition being equivalent
to $F_K(\mu_K^*;\delta_\mathbf{x}) = 0$ for all $\mathbf{x}\in{\mathscr X}$. We thus obtain the following property.
\begin{theo}\label{P:ET-M(1)}
When ${\mathscr E}_K(\cdot)$ is strictly convex on ${\mathscr M}(1)\cap{\mathscr M}_K$, $\mu_K^*\in{\mathscr M}(1)$ is the minimum-energy signed measure with total mass one on~${\mathscr X}$ if and only if
\begin{eqnarray}\label{ET-M(1)}
\forall\, \mathbf{x}\in{\mathscr X}\,, \ P_{\mu_K^*}(\mathbf{x}) = {\mathscr E}_K(\mu_K^*) \,.
\end{eqnarray}
\end{theo}
If we define now a \emph{signed equilibrium measure} on~${\mathscr X}$ as a measure $\mu\in{\mathscr M}(1)$ such that $P_{\mu}(\mathbf{x})$ is constant on~${\mathscr X}$, from the definition of $P_\mu(\cdot)$, when such a measure exists it necessarily satisfies the condition of Theorem~\ref{P:ET-M(1)} and therefore coincides with $\mu_K^*$. Similarly to the case where one considers probability measures in ${\mathscr M}^+(1)$, we can define the (generalized) capacity of~${\mathscr X}$ for measures in ${\mathscr M}(1)$ as $C_K^*=[\inf_{\mu\in{\mathscr M}^(1)} {\mathscr E}_K(\mu)]^{-1}$, with $C_K^*=1/{\mathscr E}_K(\mu_K^*)$ when $\mu_K^*$ exists, see \cite[p.~824]{DamelinHRZ2010} (note that $C_K^*$ may be negative).
However, $\mu_K^*$ may not exist. Notice in particular that ${\mathscr M}(1)$ is not vaguely compact, contrarily to ${\mathscr M}^+(1)$ (and for Riesz kernels \eqref{Riesz} with $s<d-1$, ${\mathscr M}_{K_{(s)}}$ is not complete contrarily to ${\mathscr M}_{K_{(s)}}\cap{\mathscr M}^+$ \cite[Th.~1.19]{Landkof1972}).
\begin{example}[Continuation of Examples~\ref{Ex:Riesz<0} and \ref{Ex:Riesz<0,d>=2}]\label{Ex:Riesz<0,d=1} Take $K(x,x')=K^{(s)}(x,x')=-|x-x'|^{s}$ on ${\mathscr X}=[0,1]$, $s\in(0,2)$, see \eqref{Riesz<0}. $K$ is CISPD, and there exists a unique minimum-energy probability measure $\mu^+=\mu_{K^{(s)}}^+$ in ${\mathscr M}^+(1)$. On the other hand, below we show that minimum-energy signed measures in ${\mathscr M}(1)$ do not belong to ${\mathscr M}^+(1)$ when $s\in(1,2)$ and that there is no minimum-energy signed measure in ${\mathscr M}(1)$ when $s \geq 2$.
When $s\in(0,1)$, $\mu^+$ has a density $\varphi^{(s)}(\cdot)$ with respect to the Lebesgue measure on $[0,1]$,
$$
\varphi^{(s)}(x)= \frac{\Gamma[1-s/2]}{2^s\, \sqrt{\pi}\,\Gamma[(1-s)/2]}\; \frac{1}{[x(1-x)]^{(1+s)/2}} \,,
$$
and $P_{\mu^+}(x)={\mathscr E}(\mu^+)=-\sqrt{\pi}\,\Gamma(1-s/2)/\{2^s\,\Gamma[(1-s)/2]\,\cos(\pi s/2)\}$ for all $x\in{\mathscr X}$ (and ${\mathscr E}(\mu^+) \rightarrow -1/2$ as $s\rightarrow 1^-$). The fact that $P_{\mu^+}(x)={\mathscr E}(\mu^+)$ for all $x\in{\mathscr X}$ indicates that $\mu^+$ is the minimum-energy signed measure with total mass one when $s\in(0,1]$.
When $s\in[1,2)$, $\mu^+=(\delta_0 + \delta_1)/2$; the associated potential is $P_{\mu^+}(x)=-(|x|^{s}+|1-x|^{s})/2 \geq {\mathscr E}(\mu^+)=-1/2$, $x\in{\mathscr X}$ (note that $P_{\mu^+}(x)=-1/2$ for all $x\in{\mathscr X}$ when $s=1$).
Consider now the signed measure $\mu_w=[(1+w)/2](\delta_0+\delta_1)-w \delta_{1/2}$, $w>0$, so that $\mu_w({\mathscr X})=1$ (i.e., $\mu_w\in{\mathscr M}(1)$). Direct calculation gives
${\mathscr E}_{K^{(s)}}(\mu_w)=-(1+w)(1+w-2^{2-s}w)$, which is minimum for $w=w_*(s)=(1-2^{1-s})/(2^{2-s}-1)$ when $s<2$, with ${\mathscr E}_{K^{(s)}}(\mu_{w_*(s)})=2(1-2^{2-s})/(4-2^{s})^2$. For $s\in(1,2)$ we get ${\mathscr E}_{K^{(s)}}(\mu_{w_*(s)})< {\mathscr E}(\mu^+)=-1/2$, and there exist signed measures in ${\mathscr M}(1)$ such that ${\mathscr E}_{K^{(s)}}(\mu_w)<{\mathscr E}(\mu^+)$. Therefore, minimum-energy signed measures with total mass one are not probability measures. For $s\geq 2$, $\lim_{w\rightarrow+\infty}{\mathscr E}_{K^{(s)}}(\mu_w)=-\infty$, and there is no minimum-energy signed measure; in particular, ${\mathscr E}_{K^{(s)}}(\mu_w)=-(w+1)/2$ for $s=2$.
\fin
\end{example}
\begin{example}[Continuation of Examples~\ref{Ex:Riesz} and \ref{Ex:Riesz-B_2}]\label{Ex:Riesz3}
Consider Riesz kernels $K_{(s)}$, see \eqref{Riesz}, for ${\mathscr X}={\mathscr B}_d(\boldsymbol{0},1)$, $d>2$ and $s\in(0,d-2)$; the minimum-energy probability measure $\mu^+$ is then uniform on the unit sphere ${\mathscr S}_d(\boldsymbol{0},1)$ and the potential at all interior points satisfies $P_{\mu^+}(\mathbf{x}) > {\mathscr E}_{K_{(s)}}(\mu^+)$. Consider the signed measure $\mu_w=(1+w) \mu^+ -w \mu^{(r)}$, with $\mu^{(r)}$ uniform on the sphere ${\mathscr S}_d(\boldsymbol{0},r)$ with radius $r\in(0,1)$. Calculations similar to those in the proof of \cite[Th.~1.32]{Landkof1972} show that
${\mathscr E}_{K_{(s)}}(\mu_w)<{\mathscr E}_{K_{(s)}}(\mu^+)$ for $w$ small enough, indicating that $\mu^+$ is not the minimum-energy signed measure with total mass one.
\fin
\end{example}
\subsubsection{When minimum-energy signed measures are probability measures}\label{S:subharmonic}
\begin{theo}\label{Th:MAIN}
Assume that $K$ is ISPD and translation invariant, with $K(\mathbf{x},\mathbf{x}')=\Psi(\mathbf{x}-\mathbf{x}')$ and $\Psi$ continuous,
twice differentiable except at the origin, with $\Psi(\boldsymbol{0})<\infty$ and Laplacian $\Delta_\Psi(\mathbf{x})=\sum_{i=1}^d \partial^2 \Psi(\mathbf{x})/\partial x_i^2 \geq 0$, $\mathbf{x}\neq \boldsymbol{0}$.
Then there exists a unique minimum-energy signed measure $\mu_K^*$ in ${\mathscr M}(1)$, and $\mu_K^*$ is a probability measure.
\end{theo}
\noindent{\em Proof.} \mbox{}
The conditions of Theorem~\ref{P:ET-P} are satisfied, and there exists a unique minimum-energy probability measure $\mu^+$ such that
$P_{\mu^+}(\mathbf{x}) \geq {\mathscr E}_K(\mu^+)$ for all $\mathbf{x}\in{\mathscr X}$. It also satisfies $P_{\mu^+}(\mathbf{x}) = {\mathscr E}_K(\mu^+)$ on the support of $\mu^+$.
On the other hand, the conditions on $K$ imply that for any $\mu$ in ${\mathscr M}^+(1)$, $P_\mu(\cdot)$ is subharmonic outside the support of $\mu$, see, e.g., \cite[Sect.~I.2]{Landkof1972}. The first maximum principle of potential theory thus holds \cite[Th.~1.10]{Landkof1972}: $P_\mu(\mathbf{x})\leq c$ on the support of $\mu$ implies $P_\mu(\mathbf{x})\leq c$ everywhere. Applying this to $\mu^+$, we obtain that $P_{\mu^+}(\mathbf{x}) \leq {\mathscr E}_K(\mu^+)$ everywhere; therefore, $P_{\mu^+}(\mathbf{x}) = {\mathscr E}_K(\mu^+)$ for all $\mathbf{x}\in{\mathscr X}$. Theorem~\ref{P:ET-M(1)} implies that $\mu^+$ is the minimum-energy signed measure with total mass one.
\carre
\vsp
The central argument for the proof of the property above is that $P_\mu(\cdot)$ is subharmonic outside the support of $\mu$ for any probability measure $\mu$ with finite energy. Weaker conditions than those in the theorem may be sufficient in particular situations, such as $\Psi(\mathbf{x}-\mathbf{x}')=\psi(\|\mathbf{x}-\mathbf{x}'\|)$ with $\psi(\cdot)$ convex on $(0,\infty)$, which generalizes a result of H\'ajek (1956), see also Section~\ref{S:continuous-BLUE}.
Another generalization is to consider CISPD kernels. For example, for the kernels $K^{(s)}$ of \eqref{Riesz<0}, we have $\Delta(-\|\mathbf{x}\|^{s})=s(2-s-d)/\|\mathbf{x}\|^{2-s}$, $\mathbf{x}\neq\boldsymbol{0}$. Potentials are superharmonic for $d\geq 2$. When $d=1$, they are superharmonic for $s\in[1,2)$; they are subharmonic and satisfy the maximum principle for $s\in(0,1)$, see Example~\ref{Ex:Riesz<0,d=1}.
Extension of Theorem~\ref{Th:MAIN} to singular kernels requires advanced results from potential theory; see especially \cite{FugledeZ2018, Landkof1972}. In particular, for the Riesz kernels $K_{(s)}$ of \eqref{Riesz}, we have $\Delta(\|\mathbf{x}\|^{-s})=s(s+2-d)/\|\mathbf{x}\|^{s+2}$, $\mathbf{x}\neq\boldsymbol{0}$.
When $d>2$ and $s\in(0,d-2]$, $P_\mu$ can be proved to be superharmonic in $\mathds{R}^d$, and when
$s\in[d-2,d)$, $P_\mu$ can be proved to be subharmonic outside the support of $\mu$, $\mu^+$ being then the minimum-energy signed measure.
This is also true for the logarithmic kernel for $d\leq 2$, with $\Delta(-\log\|\mathbf{x}\|)= (2-d)/\|\mathbf{x}\|^2$, $\mathbf{x}\neq\boldsymbol{0}$.
Examples~\ref{Ex:Riesz-B_2} and \ref{Ex:Riesz3} give an illustration.
\subsection{Best Linear Unbiased Estimator (BLUE) of $\beta_0$}\label{S:BLUE}
\subsubsection{Continuous BLUE}\label{S:continuous-BLUE}
Consider again the situation of Section~\ref{S:BQ} where $\sigma^2\,K$ corresponds to the covariance of a random field $Z_x$.
Suppose that we may observe $f(\cdot)$ over~${\mathscr X}$ in order to estimate $\beta_0$ in the regression (location) model with correlated errors \eqref{model1}. Any linear estimator of $\beta_0$ takes the general form
$$
\hat\beta_0 = \hat\beta_0(\xi) = \int_{\mathscr X} f(\mathbf{x}) \,\mbox{\rm d}\xi(\mathbf{x}) = I_\xi(f)
$$
for some $\xi\in{\mathscr M}$, and $\hat\beta_0(\xi)$ is unbiased when $\xi\in{\mathscr M}(1)$. Its variance is
$$
V_\xi=\mathbb{E}\{(\hat\beta_0(\xi)-\beta_0)^2\} = \sigma^2\, {\mathscr E}_K(\xi) \,;
$$
see \cite[Sect.~4.2]{Nather85}. The existence of a minimum-energy signed measure $\mu_K^*$ is then equivalent to the existence of the continuous BLUE $\hat\beta_0^*$ for $\beta_0$, with $\hat\beta_0^*=\hat\beta_0(\mu_K^*)$; the variance of $\hat\beta_0^*$ is proportional to the minimum energy ${\mathscr E}_K(\mu_K^*)$, and Theorem~\ref{P:ET-M(1)} corresponds to Grenander's theorem \cite{Grenander50}.
Also, from that theorem, the existence of $\mu_K^*$ is equivalent to the existence of an equilibrium measure that yields a constant potential on~${\mathscr X}$. It can be related to a property of the generalized capacity $C_K^*$, as shown in the following theorem.
\begin{theo}\label{P:capacity-1}
When $K$ is ISPD, the constant function $1_{\mathscr X}$ equal to 1 on~${\mathscr X}$ belongs to the space ${\mathscr P}_K$ of potential fields if and only if there exists a minimum-energy signed measure $\mu_K^*\in{\mathscr M}(1)$, with ${\mathscr E}_K(\mu_K^*) \neq 0$. Moreover, the generalized capacity $C_K^*$ is finite and nonzero, and satisfies $\|1_{\mathscr X}\|^2_{{\mathscr P}_K}=C_K^*$.
\end{theo}
\noindent {\em Proof.}
Suppose that $1_{\mathscr X}\in{\mathscr P}_K$. There exists $\mu\in{\mathscr M}_K$ such that $P_\mu=1_{\mathscr X}$; that is, $P_\mu(\mathbf{x})=1$ for all $\mathbf{x}\in{\mathscr X}$. The definition of $P_\mu$ yields ${\mathscr E}_K(\mu)=\mu({\mathscr X})$, which is finite and strictly positive since $K$ is ISPD and $\mu\neq 0$. Denote $\mu'=\mu/\mu({\mathscr X}) \in{\mathscr M}(1)$. We obtain $P_{\mu'}(\mathbf{x})=1/\mu({\mathscr X})={\mathscr E}_K(\mu')>0$ for all $\mathbf{x}\in{\mathscr X}$. Theorem~\ref{P:ET-M(1)} implies that $\mu'$ is the minimum-energy measure $\mu_K^*$. Also, $C_K^* = 1/{\mathscr E}_K(\mu')=\mu({\mathscr X}) \neq 0$, with $\|1_{\mathscr X}\|^2_{{\mathscr P}_K}={\mathscr E}_K(\mu)$, see Section~\ref{S:EP}.
Suppose now that there exists a minimum-energy signed measure $\mu_K^*\in{\mathscr M}(1)$ with ${\mathscr E}_K(\mu_K^*) \neq 0$. Theorem~\ref{P:ET-M(1)} implies that
$P_{\mu_K^*}(\mathbf{x}) = {\mathscr E}_K(\mu_K^*)$ for all $\mathbf{x}\in{\mathscr X}$. For $\mu=\mu_K^*/{\mathscr E}_K(\mu_K^*)$, we get $P_\mu(\mathbf{x})=1$ for all $\mathbf{x}\in{\mathscr X}$, and $\|1_{\mathscr X}\|^2_{{\mathscr P}_K}={\mathscr E}_K(\mu)=1/{\mathscr E}_K(\mu_K^*)$.
\carre
\vsp
Under the conditions of Theorem~\ref{Th:MAIN}, the BLUE exists, $\hat\beta_0^*=\hat\beta_0(\mu_K^+)$, with $\mu_K^+$ the minimum-energy probability measure, and its variance equals $\sigma^2{\mathscr E}_K(\mu_K^+)$. This is also true when $K(\mathbf{x},\mathbf{x}')=\psi(\|\mathbf{x}-\mathbf{x}'\|)$ with $\psi(\cdot)$ convex on $(0,\infty)$. For $d=1$, this property was known to H\'ajek (1956), see \cite[p.~56]{Nather85}. The existence of a minimum-energy signed measure is not guaranteed in other circumstances, in particular when $K(\mathbf{x},\mathbf{x}')=\Psi(\mathbf{x}-\mathbf{x}')$ and $\Psi$ is differentiable at 0; see Example~\ref{Ex:Antoniadis} below.
\subsubsection{Discrete BLUE}\label{S:discreteBLUE}
Consider the framework of Section~\ref{S:BQ}, with the same notation, and suppose that the $n$ design points $\mathbf{x}_i$ in $\mathbf{X}_n$ are fixed.
Any linear estimator of $\beta_0$ in \eqref{model1} has then the form $\tilde \beta_0^n=\mathbf{w}_n^T \mathbf{y}_n$, with $\mathbf{w}_n=(w_1,\ldots,w_n)^T\in\mathds{R}^n$. The unbiasedness constraint imposes $\mathbf{w}_n^T\boldsymbol{1}_n=1$. The variance of $\tilde \beta_0^n$ equals $\sigma^2\mathbf{w}_n^T \mathbf{K}_n \mathbf{w}_n$, and the BLUE corresponds to the estimator $\hat\beta_0^n$ given by \eqref{beta01} (we assume that $\mathbf{K}_n$ is nonsingular). The minimum-energy signed measure in ${\mathscr M}(1)$ (here discrete) $ \mu_K^*$ is defined by the weights $\mathbf{w}_n^*=\mathbf{K}_n^{-1}\boldsymbol{1}_n/(\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n)$ set on the points in $\mathbf{X}_n$; its energy is ${\mathscr E}_K( \mu_K^*)={\mathbf{w}_n^*}^T \mathbf{K}_n \mathbf{w}_n^* = 1/(\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n)$
and the variance of the BLUE equals $\sigma^2 {\mathscr E}_K( \mu_K^*)$. Note that some components of $\mathbf{w}_n^*$ may be negative and that the potential associated with the measure $ \mu_K^*/{\mathscr E}_K( \mu_K^*)$ on ${\mathscr X}=\mathbf{X}_n$ gives the constant function $1_{\mathscr X}=\boldsymbol{1}_n$, see Theorem~\ref{P:capacity-1}. The optimal design problem for the discrete BLUE thus corresponds to the determination of the $n$-point set maximizing $\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n$.
\begin{example}\label{Ex:Antoniadis} Consider $K(x,x')=\exp(-\theta |x-x'|)$, $\theta>0$, for $x,x'\in{\mathscr X}=[0,1]$. $K$ is ISPD and satisfies
$$
1_{\mathscr X} = \frac{K(\cdot,0)+K(\cdot,1)}{2} + \frac{\theta}{2}\, \int_0^1 K(\cdot,x) \mbox{\rm d} x \,,
$$
so that $1_{\mathscr X}\in{\mathscr P}_K$, see \cite{Antoniadis84}.
The minimum-energy measure in ${\mathscr M}(1)$ is $\mu_K^*=(\delta_0+\delta_1+\theta\mu_L)/(\theta+2)$, with $\mu_L$ the Lebesgue measure on~${\mathscr X}$, and $\mu_K^*\in{\mathscr M}^+(1)$.
The BLUE of $\beta_0$ in \eqref{model1} is $\hat \beta_0^*=\int_{\mathscr X} f(x)\, \mbox{\rm d}\mu_K^*(x)$, its variance equals $\sigma^2{\mathscr E}_K(\mu_K^*)=2\sigma^2/(\theta+2)$, see \cite[p.~56]{Nather85}.
Note that $K'=K-2/(\theta+2)$ is still positive definite, but $1_{\mathscr X} \not\in {\mathcal H}_{K'}$ since $c^2 K'-1$ is not positive definite for any $c \neq 0$, see, e.g., \cite[p.~30]{BerlinetT-A2004}, \cite[p.~20]{Paulsen2009}.
Consider now the squared exponential kernel $K(x,x')=\exp(-\theta|x-x'|^2)$, $\theta>0$. The constant $1_{\mathscr X}$ does not belong to ${\mathcal H}_K$
\cite{SteinwartHS2006} and the BLUE of $\beta_0$ in \eqref{model1} is not defined for that kernel. On the other hand, the discrete BLUE \eqref{beta01} is well defined for any set of $n$ distinct points $x_i$,
$\hat \beta_0^n={\mathbf{w}_n^*}^T\mathbf{y}_n=\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\mathbf{y}_n/(\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n)$. Suppose that the $n$ points $x_i$ are equally spaced in ${\mathscr X}=[0,1]$. The process $Z_x$ in \eqref{model1} has mean square derivatives of all orders, and, roughly speaking, for large $n$ the construction of the BLUE mimics the estimation of derivatives of $f$ and the weights $w_i^*$ strongly oscillate between large positive and negative values. Figure~\ref{F:wstar}-Left shows the optimal weights $(w_i^*/|w_i^*|)(\log_{10}(\max\{|w_i^*|,1\})$, truncated to absolute values larger than 1 and in log scale, when $x_i=(i-1)/(n-1)$, $i=1,\ldots,n=101$. In Figure~\ref{F:wstar}-Right, the kernel is $K(x,x')=(1+\sqrt{5}|x-x'|+5|x-x'|^2/3)\exp(-\sqrt{5}|x-x'|)$ (Mat\'ern 5/2), so that $Z_x$ is twice mean-square differentiable; the construction of the BLUE mimics the estimation of the first and second order derivatives of $f$ at $0$ and $1$, suggesting that $1_{\mathscr X}\not\in{\mathcal H}_K$ in that case too; see \cite{DettePZ2016} for more details.
\fin
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=.49\linewidth]{wstar_cov1_theta1_p2.png} \includegraphics[width=.49\linewidth]{wstar_cov6_theta1.png}
\end{center}
\caption{\small BLUE weights $(w_i^*/|w_i^*|)(\log_{10}(\max\{|w_i^*|,1\})$ for $x_i=(i-1)/(n-1)$, $i=1,\ldots,n=101$. Left: $K(x,x')=\exp(-|x-x'|^2)$, Right: $K(x,x')=(1+\sqrt{5}|x-x'|+5|x-x'|^2/3)\exp(-\sqrt{5}|x-x'|)$ (Mat\'ern 5/2).}
\label{F:wstar}
\end{figure}
\end{example}
\vsp
Although a minimum-energy signed measure may not exist, in the next section we shall see how, for any measure $\mu\in{\mathscr M}(1)$ and any CISPD kernel $K$, we can modify $K$ in such a way that the minimum-energy signed measure for the modified kernel exists (and coincides with $\mu$).
\subsection{Equilibrium measure and kernel reduction}\label{S:D-EM-RK}
Minimum-energy signed measures, when they exist, satisfy the following property.
\begin{lem}\label{P:ET-2}
If $K$ is CISPD and if a minimum-energy signed measure $\mu_K^*$ exists in ${\mathscr M}(1)$, we have
${\mathscr E}_K(\xi) = {\mathscr E}_K[\xi-\xi({\mathscr X})\mu_K^*] + [\xi({\mathscr X})]^2 \, {\mathscr E}_K(\mu_K^*) \,, \ \forall\, \xi\in{\mathscr M}_K$.
\end{lem}
\noindent {\em Proof.}
For any $\xi\in{\mathscr M}_K$, direct calculation gives
\begin{eqnarray*}
{\mathscr E}_K[\xi-\xi({\mathscr X})\mu_K^*] &=& {\mathscr E}_K(\xi)+ [\xi({\mathscr X})]^2\, {\mathscr E}_K(\mu_K^*) - 2 \xi({\mathscr X})\, \int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\,\mbox{\rm d}\mu_K^*(\mathbf{x})\mbox{\rm d}\xi(\mathbf{x}') \\
&=& {\mathscr E}_K(\xi) - [\xi({\mathscr X})]^2 \, {\mathscr E}_K(\mu_K^*) \,,
\end{eqnarray*}
where the second equality follows from \eqref{ET-M(1)}.
\carre
\vsp
Under the conditions of Lemma~\ref{P:ET-2}, any $\xi\in{\mathscr M}(1)$ satisfies
$$
{\mathscr E}_K(\xi)={\mathscr E}_K(\xi-\mu_K^*)+ {\mathscr E}_K(\mu_K^*) \,,
$$
where the first term on the right-hand side equals the squared MMD $\gamma_K^2(\xi,\mu_K^*)$, see \eqref{kernel discrepancyB}, and the second term does not depend on $\xi$. Minimizing the energy ${\mathscr E}_K(\xi)$ is thus equivalent to minimizing the MMD $\gamma_K(\xi,\mu_K^*)$. However, ($i$) $\mu_K^*$ may not exist, ($ii$) in many situations we wish to select a measure $\xi$ having small MMD $\gamma_K(\xi,\mu)$ for \emph{a given} measure $\mu\in{\mathscr M}_K$. This is the case in particular when one aims at evaluating the integral of a function with respect to some $\mu\in{\mathscr M}^+(1)$ (Section~\ref{S:BQ}), or when we want construct a space-filling design in~${\mathscr X}$, $\mu$ being then uniform.
\subsubsection{Kernel reduction}
Take any $\mu\in{\mathscr M}_K$ such that $\mu({\mathscr X})\neq 0$. Without any loss of generality, we assume $\mu\in{\mathscr M}(1)$. Following \cite{DamelinHRZ2010}, we show how to modify the kernel $K$ in such a way that minimizing the energy ${\mathscr E}_{K_\mu}(\xi)$, $\xi\in{\mathscr M}(1)$, for the new (reduced) kernel $K_\mu$ is equivalent to minimizing $\gamma_{K_\mu}(\xi,\mu)$.
Define
\begin{eqnarray}\label{K_mu}
K_\mu(\mathbf{x},\mathbf{x}')=K(\mathbf{x},\mathbf{x}') - P_\mu(\mathbf{x}) - P_\mu(\mathbf{x}') + {\mathscr E}_K(\mu)\,,
\end{eqnarray}
see \cite{Schaback199}.
One can readily check that the energy for this new reduced kernel $K_\mu$ satisfies ${\mathscr E}_{K_\mu}(\beta\mu)=0$ for any real $\beta$ and that the potential for $\mu$ associated with $K_\mu$ satisfies $\widetilde P_\mu(\mathbf{x})=\int_{\mathscr X} K_\mu(\mathbf{x},\mathbf{x}')\, \mbox{\rm d}\mu(\mathbf{x}') = 0$ for all $\mathbf{x}$.
Next theorem indicates that, for any given $\mu$ in ${\mathscr M}(1)\cap{\mathscr M}_K$, when considering signed measures $\xi$ with total mass one, minimizing the energy ${\mathscr E}_{K_\mu}(\xi)$ is equivalent to minimizing the MMD $\gamma_K(\xi,\mu)$, provided that $K$ is CISPD.
\begin{theo}\label{P:Energy-RK}
If $K$ is CISPD, then for any $\mu\in{\mathscr M}(1)\cap{\mathscr M}_K$, we have\\
($i$) the reduced kernel $K_\mu$ defined by \eqref{K_mu} is CISPD; \\
($ii$) $\mu$ is the minimum-energy measure in ${\mathscr M}(1)$ for $K_\mu$, and
\begin{eqnarray*}\label{Energy-RK}
\forall\, \xi\in{\mathscr M}_K\,, \ {\mathscr E}_{K_\mu}(\xi)={\mathscr E}_K[\xi-\xi({\mathscr X})\mu] = {\mathscr E}_{K_\mu}[\xi-\xi({\mathscr X})\mu] \,.
\end{eqnarray*}
\end{theo}
\noindent{\em Proof.} For any nonzero $\xi\in{\mathscr M}_K$, direct calculation using \eqref{K_mu} gives
\begin{eqnarray}
{\mathscr E}_{K_\mu}(\xi) &=& {\mathscr E}_K(\xi) - 2\xi({\mathscr X})\, \int_{{\mathscr X}^2} K(\mathbf{x},\mathbf{x}')\,\mbox{\rm d}\mu(\mathbf{x})\mbox{\rm d}\xi(\mathbf{x}')+ [\xi({\mathscr X})]^2 \, {\mathscr E}_K(\mu) \nonumber \\
&=& {\mathscr E}_K[\xi-\xi({\mathscr X})\mu] \,. \label{Energy-RKB}
\end{eqnarray}
($i$) When $\xi({\mathscr X})=0$ we get ${\mathscr E}_{K_\mu}(\xi)={\mathscr E}_K(\xi)$ which is strictly positive when $\xi\neq 0$, showing that $K_\mu$ is CISPD.
($ii$) Since $[\xi-\xi({\mathscr X})\mu]({\mathscr X})=0$ and $K$ is CISPD, ${\mathscr E}_{K_\mu}(\xi)\!>\!0$ for $\xi\neq \xi({\mathscr X})\mu$, showing that $\mu$ is the (unique) minimum-energy signed measure in ${\mathscr M}(1)$ for $K_\mu$.
Since ${\mathscr E}_{K_\mu}(\mu)=0$, Lemma~\ref{P:ET-2} with $K_\mu$ substituted for $K$ implies that ${\mathscr E}_{K_\mu}(\xi)={\mathscr E}_{K_\mu}[\xi-\xi({\mathscr X})\mu]$ for any $\xi\in{\mathscr M}_K$, which, together with \eqref{Energy-RKB}, concludes the proof.
\carre
\subsubsection{Kernel reduction, BLUE and Bayesian integration}\label{S:BLUE-Kreduction}
Consider again the situation of Section~\ref{S:BLUE}, and define $\mathcal{P}_1$ as the orthogonal projection of $L^2({\mathscr X},\mu)$ onto the linear space spanned by the constant $1$; see \cite{GP-CSDA2016}. The model \eqref{model1} can then we written as
\begin{eqnarray}\label{model1p}
f(\mathbf{x})= \beta_0 + {\mathcal{P}_1}Z_x + (\Id_{L^2}-{\mathcal{P}_1})Z_x = \beta_0'+ \widetilde Z_x\,,
\end{eqnarray}
where $\beta_0'=\beta_0 + {\mathcal{P}_1}Z_x$ and $\widetilde Z_x=(\Id_{L^2}-{\mathcal{P}_1})Z_x$, with $\widetilde Z_x$ having zero mean and covariance $\mathbb{E}\{\widetilde Z_x \widetilde Z_{x'}\} = \sigma^2 K_\mu(\mathbf{x},\mathbf{x}')$.
We have seen in Section~\ref{S:BLUE} that the variance of the continuous BLUE of $\beta_0$ equals $\sigma^2 {\mathscr E}_K( \mu_K^*)$ provided that the minimum-energy signed measure $ \mu_K^*$ exists. (Note that the prior on $\beta_0'$ remains non-informative when the prior on $\beta_0$ is non-informative.) On the other hand, we obtain now that the continuous BLUE of $\beta_0'$ always exists: it coincides with $I_\mu(f)$ and its variance is $\sigma^2 {\mathscr E}_{K_\mu}(\mu)=0$. Therefore, as mentioned in introduction, Bayesian integration for the model \eqref{model1} with correlated errors is equivalent to parameter estimation in a location model with different correlation structure.
\subsection{Tensor product kernels}\label{S:TensorP}
From $d$ kernels $K_i$ respectively defined on ${\mathscr X}_i\times{\mathscr X}_i$, $i=1,\ldots,d$, we can construct a tensor product kernel as
\begin{eqnarray}\label{TP}
K^\otimes(\mathbf{x},\mathbf{x}')=\prod_{i=1}^d K_i(x_i,x'_i)\,,
\end{eqnarray}
where $\mathbf{x}=(x_1,\ldots,x_d)^T$ and $\mathbf{x}'=(x'_1,\ldots,x'_d)^T$ belong to the product space ${\mathscr X}={\mathscr X}_1 \times \cdots\times {\mathscr X}_d$. The construction is particularly useful when considering product measures on~${\mathscr X}$, since, in some sense, it allows us to decompose an integration or space-filling design problem in a high dimensional space into its one-dimensional counterparts.
Suppose that each $K_i$ is uniformly bounded and CISPD on ${\mathscr M}^{(i)}={\mathscr M}[{\mathscr X}_i]$; that is, $K_i$ is ISPD on ${\mathscr M}^{(i)}(0)$, see Definitions~\ref{D:ISPD} and \ref{D:CISPD}. One can show that this is equivalent to $K^\otimes$ being ISPD on $\otimes_{i=1}^d {\mathscr M}^{(i)}(0)$, see \cite[Th.~2]{SzaboS2017}. In the same paper, the authors prove (Th.~4) that if each $K_i$ is moreover continuous and translation invariant, then $K^\otimes$ is ISPD on ${\mathscr M}(0)$; that is, $K^\otimes$ is CISPD on ${\mathscr M}$. Their proof relies on the equivalence between the CISPD and characteristic properties for uniformly bounded kernels, and on the characterization of characteristic continuous, uniformly bounded and translation invariant kernels through a property of the support of the measure $\Lambda$ defined in \eqref{Lambda}; see Section~\ref{S:MMD}.
An important property of tensor product kernels $K^\otimes$ is that kernel reductions $K^\otimes_\mu$, see \eqref{K_mu}, are easily obtained explicitly. Indeed, when $\mu=\otimes_{i=1}^d \mu^{(i)}$ is a product measure on~${\mathscr X}$, then, for all $\mathbf{x}\in{\mathscr X}$,
\begin{eqnarray}
{\mathscr E}_{K^\otimes}(\mu) &=& \prod_{i=1}^d {\mathscr E}_{K_i}(\mu^{(i)}) \,, \label{E-T} \\
P_\mu(\mathbf{x}) &=& \prod_{i=1}^d \int_{{\mathscr X}_i} K_i(x_i,x'_i)\, \mbox{\rm d}\mu^{(i)}(x'_i) = \prod_{i=1}^d P_{\mu^{(i)}}(x_i) \,, \label{Pot-T}
\end{eqnarray}
which facilitates the calculation of ${\mathscr E}_{K^\otimes_\mu}(\xi)$, in particular when $\xi$ is a discrete measure as considered in Section~\ref{S:Empirical}. Table~\ref{Tb:EandP} gives the expressions of ${\mathscr E}_K(\mu)$ and $P_{\mu}(x)$ obtained for a few kernels, with $\mu$ uniform on ${\mathscr X}=[0,1]$; the expressions for the squared exponential and Mat\'ern kernels can be found in \cite{GinsbourgerRSDL2014}. Note that in each case ${\mathscr E}_{K_\mu}(\xi)>0$ for any $\xi\in{\mathscr M}(1)$, $\xi\neq \mu$.
\begin{table}[ht!]
\begin{center}
{\scriptsize
\caption{\small Energy ${\mathscr E}_K(\mu)$ and potential $P_{\mu}(x)$ for different kernels $K$ with $\mu$ uniform on ${\mathscr X}=[0,1]$; $P_{\mu}(x)=S_\mu(x)+S_\mu(1-x)+T_\mu(x)$; $S_\mu(\cdot)$ is continuously differentiable in $(0,1]$, $T_\mu=0$ when $K$ is translation invariant.}
\label{Tb:EandP}
\begin{tabular}{lll}
\toprule
$K(x,x')$ & ${\mathscr E}_K(\mu)$ & $S_{\mu}(x)$ [and $T_\mu(x)$] \\
\midrule
$\exp(-\theta |x-x'|)$ & $2(\theta+\mathrm{e}^{-\theta}-1)/\theta^2$ & $x(1-\mathrm{e}^{-\theta|x|})/(\theta|x|)$ \\
$K_{3/2,\theta/\sqrt{3}}(x,x')$ in \eqref{K32} & $ 2[\theta(2+\mathrm{e}^{-\theta})+3(\mathrm{e}^{-\theta}-1)]/\theta^2$ & $x[2-(2+\theta|x|)\mathrm{e}^{-\theta|x|}]/(\theta|x|)$ \\
$[(x-x')^2+\epsilon]^{-1}$ ($\epsilon\geq 0$) & $(2/\sqrt{\epsilon})\,\arctan(1/\sqrt{\epsilon})-\log(1+1/\epsilon)$ & $(1/\sqrt{\epsilon})\arctan(x/\sqrt{\epsilon})$ \\
$(|x-x'|+\epsilon)^{-1}$ ($\epsilon> 0$)& $2\, [(1+\epsilon)\,\log(1+1/\epsilon)-1]$ & $\mathrm{sign}(x) \log(1+|x|/\epsilon)$\\
$(|x-x'|+\epsilon)^{-1/2}$ ($\epsilon> 0$)& $4\epsilon^{3/2}\, [2(1+1/\epsilon)^{3/2}-2-3/\epsilon]/3$ & $2\sqrt{\epsilon}\,\mathrm{sign}(x)(\sqrt{1+|x|/\epsilon}-1)$ \\
$1-\theta\, |x-x'|$ ($0<\theta\leq 1$) & $1-\theta/3$ & $1/2-\theta x|x|/2$ \\
$|x-x'|^{-s}$ ($0<s<1$) & $2/(s^2-3s+2)$ & $x/[(1-s)|x|^s]$\\
$-\log |x-x'|$ & 3/2 & $1/2-x\log|x|$\\
$|x|+|x'|-|x-x'|$ & $2/3$ & $1/4-x|x|/2$ \ [$T_\mu(x)=|x|$]\\
$\sqrt{|x|}+\sqrt{|x'|}-\sqrt{|x-x'|}$ & $4/5$ & $1/3-2x\sqrt{|x|}/3$ \ [$T_\mu(x)=\sqrt{|x|}$]\\
\bottomrule
\end{tabular}
}
\end{center}
\end{table}
\begin{remark}[super- and subharmonicity for tensor product kernels]\label{R:subharmonicTPK} In complement of Theorem~\ref{Th:MAIN}, we may notice that when each $K_i$ in \eqref{TP} is stationary and satisfies $K_i(x,x')=\Psi_i(x-x')$, then $K^\otimes(\mathbf{x},\mathbf{x}')=\Psi(\mathbf{x}-\mathbf{x}')$ with $\Psi(\mathbf{z})=\prod_{i=1}^d \Psi_i(z_i)$. Assume that all $\Psi_i$ are twice continuously differentiable. Then, the Laplacian of $\Psi$ is
$$
\Delta_\Psi(\mathbf{x})=\sum_{i=1}^d \left[\frac{\partial^2 \Psi_i(x_i)}{\partial x_i^2} \prod_{j\neq i} \Psi_j(x_j)\right] \,. \quad\quad \fin
$$
\end{remark}
\section{Experimental design}\label{S:Empirical}
Consider an $n$-point design $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}$, with $\mathbf{x}_i\in{\mathscr X}$ for all $i$. In this section, we shall restrict our attention to finite signed measures $\xi_n=\sum_{i=1}^n w_i\, \delta_{\mathbf{x}_i}$ supported on $\mathbf{X}_n$, and denote $\mathbf{w}_n=(w_1,\ldots,w_n)^T$. As in Section~\ref{S:BQ}, we consider a measure $\mu\in{\mathscr M}^+(1)$, with special attention to space-filling design for which $\mu$ is uniform on a compact subset~${\mathscr X}$ of $\mathds{R}^d$. We assume that $K$ is SPD and $\mu$ has finite energy ${\mathscr E}_K(\mu)$, see \eqref{overline K}.
Direct calculation gives
\begin{eqnarray}
\gamma_K^2(\xi_n,\mu) = {\mathscr E}_K(\xi_n-\mu) &=& \mathbf{w}_n^T \mathbf{K}_n \mathbf{w}_n - 2 \mathbf{w}_n^T \mathbf{p}_n(\mu)+ {\mathscr E}_K(\mu) \,, \nonumber \\
&=& \sum_{i,j} w_i w_j\, K(\mathbf{x}_i,\mathbf{x}_j) - 2\, \sum_{i=1}^n w_i\,P_\mu(\mathbf{x}_i) + {\mathscr E}_K(\mu) \,, \label{JK_BQ2}
\end{eqnarray}
where $\{\mathbf{K}_n\}_{i,j}=K(\mathbf{x}_i,\mathbf{x}_j)$, $i,j=1,\ldots,n$,
and $\mathbf{p}_n(\mu)$ is given by \eqref{hb}. Note that ${\mathscr E}_K(\mu)$ and the $P_\mu(\mathbf{x}_i)$ have simple expressions when $K$ is a tensor product kernel and $\mu=\otimes_{i=1}^d \mu^{(i)}$ is a product measure on ${\mathscr X}={\mathscr X}_1\times\cdots\times{\mathscr X}_d$, see (\ref{E-T}, \ref{Pot-T}). Monte-Carlo approximation, based on a large i.i.d.\ sample from $\mu$, or a low-discrepancy sequence, can always be used instead.
\subsection{One-shot designs}\label{S:n-fixed}
\subsubsection{Support of empirical measures}\label{S:empirical}
Denote by $\xi_{n,e}=\xi_{n,e}(\mathbf{X}_n)$ the empirical measure associated with a given design $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}\subset\mathds{R}^{nd}$, $\xi_{n,e}=(1/n) \sum_{i=1}^n \delta_{\mathbf{x}_i}$. As indicated hereafter, the literature on space-filling design provides several examples of construction of $n$-point designs through the minimization of the squared MMD ${\mathscr E}_K(\xi_{n,e}-\mu)$ with respect to $\mathbf{X}_n$.
For ${\mathscr X}=[0,1]^d$, tensorised kernels based on variants of Brownian motion covariance yield $L_2$ discrepancies (symmetric, centred, wrap-around and so on); see, e.g., \cite{Hickernell1998}, \cite[Chap.~3]{FangLS2006}. For instance, for ${\mathscr X}=[0,1]$ and $K(x,x')=1-|x-x'|$ (for which the expressions of ${\mathscr E}_K(\mu)$ and $P_{\mu}(x)$ are given in Table~\ref{Tb:EandP}), ${\mathscr E}_{K_\mu}(\xi_{n,e})$ is twice the squared $L_2$ star discrepancy for $d=1$.
The ISPD kernel $K^\otimes_{s,\epsilon}(\mathbf{x},\mathbf{x}') = \prod_{i=1}^d K_{s,\epsilon}(x_i,x'_i)$, with $K_{s,\epsilon}$ given by \eqref{Kes} with $s>0$ and $\epsilon>0$, is called \emph{projection kernel} in \cite{MakJ2017b}. For very small $\epsilon$, the minimization of ${\mathscr E}_{K^\otimes_{1,\epsilon}}(\xi_{n,e})$ corresponds to the construction of a maximum-projection design, as defined in \cite{JosephGB2015}. Note that minimizing ${\mathscr E}_{K^\otimes_{s,\epsilon}}(\xi_{n,e})$ is not equivalent to minimizing ${\mathscr E}_{K^\otimes_{s,\epsilon}}(\xi_{n,e}-\mu)$: in particular, when $\mu$ is uniform on~${\mathscr X}$, which is assumed to be compact and convex, the former tends to push design points to the boundary of~${\mathscr X}$ whereas the latter keeps all points in the interior of~${\mathscr X}$; see \cite{MakJ2017b}.
In \cite{MakJ2017}, space-filling designs in a compact set ${\mathscr X}\subset \mathds{R}^d$ are constructed by minimizing ${\mathscr E}_{K^{(1)}}(\xi_{n,e}-\mu)$ for $\mu$ uniform on~${\mathscr X}$, see \eqref{Riesz<0}. They call \emph{support points} the optimal support $\mathbf{X}_n^*$, which they determine via a majorization-minimization algorithm using the property that the problem can be formulated as a difference-of-convex optimization problem. Values of ${\mathscr E}_{K^{(1)}}(\mu)$ and $P_{\mu}(\mathbf{x})$ are not available even for ${\mathscr X}=[0,1]^d$ and Monte-Carlo approximation is used.
\subsubsection{Space-filling design through Bayesian quadrature}\label{S:BQ2}
Since $s_n^2$ given by \eqref{sn1} does not depend on the function $f$ considered, a design $\mathbf{X}_n$ for Bayesian integration can in principle be chosen beforehand, by direct minimization of $s_n^2$. This corresponds to the approach followed in \cite{O'Hagan91} where several quadrature rules are tabulated (for several values of $n$). Next theorem shows the connection between the minimum of ${\mathscr E}_K(\xi_n-\mu)$ with respect to weights $\mathbf{w}_n$ and the posterior variances $s_n^2$ and $s_{n,0}^2$, see \eqref{sn1} and \eqref{sn10}. We assume that all points in $\mathbf{X}_n$ are pairwise different and $\mu$ is not fully supported on $\mathbf{X}_n$.
\begin{theo}\label{P:BQ-KD} Let $K$ be an SPD kernel and let $\mu\in{\mathscr M}^+(1)\cap{\mathscr M}_K$.
($i$) The optimal unconstrained weights $\mathbf{w}_n^*$ that minimize ${\mathscr E}_K(\xi_n-\mu)$ are $\mathbf{w}_n^*= \mathbf{K}_n^{-1}\mathbf{p}_n(\mu)$ and the corresponding measure $\xi_n^*$, with weights $\mathbf{w}_n^*$, satisfies
\begin{eqnarray}\label{JKopt}
{\mathscr E}_K(\xi_n^*-\mu) = s_{n,0}^2\,,
\end{eqnarray}
with $s_{n,0}^2$ given by \eqref{sn10}.
($ii$) The optimal weights $\hat\mathbf{w}_n$ that minimize ${\mathscr E}_K(\xi_n-\mu)$ under the constraint $\mathbf{w}_n^T\boldsymbol{1}_n=\sum_{i=1}^n w_i=1$ are
\begin{eqnarray}\label{wopt-sum=1}
\hat \mathbf{w}_n = \left(\mathbf{K}_n^{-1} - \frac{\mathbf{K}_n^{-1} \boldsymbol{1}_n\boldsymbol{1}_n^T \mathbf{K}_n^{-1}}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n} \right) \mathbf{p}_n(\mu) + \frac{\mathbf{K}_n^{-1} \boldsymbol{1}_n}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n} \,,
\end{eqnarray}
and the corresponding measure $\hat\xi_n$, with weights $\hat\mathbf{w}_n$, satisfies
\begin{eqnarray}\label{JKopt-sum=1}
{\mathscr E}_K(\hat\xi_n-\mu) = s_{n}^2\,,
\end{eqnarray}
with $s_{n}^2$ given by \eqref{sn1}; the estimator \eqref{In1} of the integral $I_\mu(f)$ is $\widehat I_n = \hat \mathbf{w}_n^T \mathbf{y}_n$.
($iii$) For any bounded signed measure $\xi_n=\sum_{i=1}^n w_i\, \delta_{\mathbf{x}_i}$ we can write
\begin{eqnarray}\label{JK-sum=any}
{\mathscr E}_K(\xi_n-\mu)=(\mathbf{w}_n-\mathbf{w}_n^*)^T\mathbf{K}_n(\mathbf{w}_n-\mathbf{w}_n^*)+{\mathscr E}_K(\xi_n^*-\mu) \,,
\end{eqnarray}
and when the weights $w_i$ sum to one, we have
\begin{eqnarray}\label{JK-sum=1}
{\mathscr E}_K(\xi_n-\mu)=(\mathbf{w}_n-\hat\mathbf{w}_n)^T\mathbf{K}_n(\mathbf{w}_n-\hat\mathbf{w}_n)+{\mathscr E}_K(\hat\xi_n-\mu).
\end{eqnarray}
\end{theo}
\noindent{\em Proof.}
The expression for $\mathbf{w}_n^*$, \eqref{JKopt} and \eqref{JK-sum=any} directly follow from the fact that ${\mathscr E}_K(\xi_n-\mu)$ is quadratic in $\mathbf{w}_n$, see \eqref{JK_BQ2}.
Since $K$ is SPD, straightforward calculation using Lagrangian theory indicates that the minimization of ${\mathscr E}_K(\xi_n-\mu)$ under the constraint $\mathbf{w}_n^T\boldsymbol{1}_n=1$ gives \eqref{wopt-sum=1}
and \eqref{JKopt-sum=1}. Suppose that $\mathbf{w}_n^T\boldsymbol{1}_n=1$, then ${\mathscr E}_K(\xi_n-\mu)=(\mathbf{w}_n-\hat\mathbf{w}_n+\hat\mathbf{w}_n-\mathbf{w}_n^*)^T\mathbf{K}_n(\mathbf{w}_n-\hat\mathbf{w}_n+\hat\mathbf{w}_n-\mathbf{w}_n^*)+{\mathscr E}_K(\xi_n^*-\mu)$ gives \eqref{JK-sum=1} since $\mathbf{K}_n(\hat\mathbf{w}_n-\mathbf{w}_n^*)$ is proportional to $\boldsymbol{1}_n$ and $(\mathbf{w}_n-\hat\mathbf{w}_n)^T\boldsymbol{1}_n=0$.
\carre
\vsp
In the discrete case considered here, the minimum-energy signed measure $\hat\xi_n$ with total mass one always exists, but note that it is not necessarily a probability measure; that is, some weights $\hat w_i$ may be negative. Theorem~\ref{P:BQ-KD} can be extended to the case where $K$ is only conditionally SPD, but the computation of optimal weights $\hat\mathbf{w}_n$ is more involved when $\mathbf{K}_n$ is singular; see Remark~\ref{R:optimal weights}.
Denote by $\widetilde\mathbf{K}_n$ the $n\times n$ matrix with elements $\{\widetilde\mathbf{K}_n\}_{i,j}=K_\mu(\mathbf{x}_i,\mathbf{x}_j)$, where $K_\mu$ is the reduced kernel \eqref{K_mu}; the corresponding vector of potential values at the $\mathbf{x}_i$ is then $\widetilde\mathbf{p}_n=(\widetilde P_\mu(\mathbf{x}_1),\ldots,\widetilde P_\mu(\mathbf{x}_n))^T=\boldsymbol{0}$.
For measures $\xi_n$ in ${\mathscr M}(1)$, in complement of ($ii$) of Theorem~\ref{P:BQ-KD}, we also have the following property.
\begin{theo}\label{P:BQ-RK}
For $K$ an SPD kernel, $\mu\in{\mathscr M}^+(1)\cap{\mathscr M}_K$ and $\xi_n\in{\mathscr M}(1)$, we have
\begin{eqnarray}\label{JK-RK}
{\mathscr E}_K(\xi_n-\mu) = {\mathscr E}_{K_\mu}(\xi_n) = \mathbf{w}_n^T \widetilde\mathbf{K}_n \mathbf{w}_n \,.
\end{eqnarray}
The posterior mean \eqref{In1} and variance \eqref{sn1} of $I_\mu(f)$ are respectively given by
\begin{eqnarray}
\widehat I_n &=& \frac{\boldsymbol{1}_n^T\widetilde\mathbf{K}_n^{-1}\mathbf{y}_n}{\boldsymbol{1}_n^T\widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n}\,, \label{In2} \\
\sigma^2 s_n^2 &=& \sigma^2 (\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n)^{-1} \,. \label{sn2}
\end{eqnarray}
\end{theo}
\noindent{\em Proof.}
Equation \eqref{JK-RK} follows from Proposition~\ref{P:Energy-RK}. Since we assumed that $\mu$ is not fully supported on $\mathbf{X}_n$ and $K$ is SPD, \eqref{JK-RK} gives $\inf_{\|\mathbf{w}_n\|=1} \mathbf{w}_n^T \widetilde\mathbf{K}_n \mathbf{w}_n >0$ , which implies that $\widetilde\mathbf{K}_n$ has full rank. Direct calculation using \eqref{K_mu} gives $\widetilde\mathbf{K}_n=\mathbf{K}_n - \mathbf{p}_n(\mu)\boldsymbol{1}_n^T - \boldsymbol{1}_n\mathbf{p}_n^T(\mu) + {\mathscr E}_K(\mu)\, \boldsymbol{1}_n\boldsymbol{1}_n^T$. The expression for $\widetilde\mathbf{K}_n^{-1}$ then yields $\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n=1/s_n^2$, with $s_n^2$ given by \eqref{sn1}, proving \eqref{sn2}. The expansion of $(\boldsymbol{1}_n^T\widetilde\mathbf{K}_n^{-1}\mathbf{y}_n)/(\boldsymbol{1}_n^T\widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n)$ gives \eqref{In1}, which proves \eqref{In2}.
\carre
\begin{remark}[Optimal weights for CISPD kernels]\label{R:optimal weights}
Lagrangian theory indicates that the solution $\hat\mathbf{w}_n$ is obtained by solving the linear equation $\mathbf{R}_n(\hat\mathbf{w}_n^T \ \lambda)^T=(\boldsymbol{0}^T \ 1)^T$, where
$$
\mathbf{R}_n = \left(
\begin{array}{cc}
\widetilde\mathbf{K}_n & \boldsymbol{1}_n \\
\boldsymbol{1}_n^T & 0 \\
\end{array}
\right) \,.
$$
When $K$ is conditionally SPD, $K_\mu$ is conditionally SPD too, and the matrix $\mathbf{R}_n$ has full rank $n+1$.
Indeed, $\mathbf{R}_n (\mathbf{z}_n^T \ z)^T=0$ implies $\boldsymbol{1}_n^T\mathbf{z}_n=0$ and $\widetilde\mathbf{K}_n\mathbf{z}_n+z\boldsymbol{1}_n=0$. Multiplying the second equation by $\mathbf{z}_n^T$, we get $\mathbf{z}_n^T\widetilde\mathbf{K}_n\mathbf{z}_n=0$. Since $K_\mu$ is conditionally SPD, this is incompatible with $\boldsymbol{1}_n^T\mathbf{z}_n=0$ unless $\mathbf{z}_n=\boldsymbol{0}$ and $z=0$. We obtain
$$
\hat\mathbf{w}_n = \frac{(\widetilde\mathbf{K}_n+\boldsymbol{1}_n\boldsymbol{1}_n^T)^{-1}\boldsymbol{1}_n}{\boldsymbol{1}_n^T(\widetilde\mathbf{K}_n+\boldsymbol{1}_n\boldsymbol{1}_n^T)^{-1}\boldsymbol{1}_n}\,,
$$
and $s_n^2=\hat\mathbf{w}_n^T\widetilde\mathbf{K}_n\hat\mathbf{w}_n= (\boldsymbol{1}_n^T(\widetilde\mathbf{K}_n+\boldsymbol{1}_n\boldsymbol{1}_n^T)^{-1}\boldsymbol{1}_n)^{-1}-1$. When $K$ is SPD and $\widetilde\mathbf{K}_n$ has full rank (Theorem~\ref{P:BQ-RK}), we recover $\hat\mathbf{w}_n = \widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n/(\boldsymbol{1}_n^T\widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n)$ and $\widehat I_n=\hat\mathbf{w}_n^T \mathbf{y}_n$ given by \eqref{In2}.
\fin
\end{remark}
\begin{remark}[BLUE and kernel reduction]\label{R:reducedK}
Equations \eqref{In2} and \eqref{sn2} indicate that $\widehat I_n$ is the BLUE of $\beta_0'$ and $\sigma^2 s_n^2$ is its variance in the model
\eqref{model1p}, $f(\mathbf{x})= \beta_0'+ \widetilde Z_x$, see Sections~\ref{S:discreteBLUE} and \ref{S:BLUE-Kreduction}. A possible interpretation is as follows.
Predictions are not modified when using the reduced kernel $K_\mu$ instead of $K$,
that is, when considering model $f(\mathbf{x})= \beta_0'+ \widetilde Z_x$ instead of \eqref{model1},
see \cite[Sect.~5.4]{GP-CSDA2016}. It implies that the expressions \eqref{In1} and \eqref{sn1} of $\widehat I_n$ and $s_n^2$ are unchanged when replacing $K$ by $K_\mu$. Since, by construction, $\widetilde \mathbf{p}_n(\mu)=\boldsymbol{0}$ and ${\mathscr E}_{K_\mu}(\mu)=0$ ($(\Id_{L^2}-{\mathcal{P}_1})Z_x$ has no contribution to the integral of $f$), we directly obtain \eqref{In2} and \eqref{sn2}.
\fin
\end{remark}
\begin{remark}[IMSPE for tensor product kernels]
The use of a tensor product kernel \eqref{TP} and a product measure $\mu=\otimes_{i=1}^d \mu^{(i)}$ on ${\mathscr X}={\mathscr X}_1\times\cdots\times{\mathscr X}_d$ facilitates the calculations of $\widetilde\mathbf{K}_n$ and ${\mathscr E}_K(\xi_n-\mu)$, see \eqref{JK_BQ2}, since ${\mathscr E}_K(\mu)$ and $P_\mu(\mathbf{x}_i)$ have the simple expressions (\ref{E-T}, \ref{Pot-T}). The calculation of the IMSPE is facilitated too, but to a lesser extend. Indeed, we have
\begin{eqnarray*}
\int_{\mathscr X} \rho_n^2(\mathbf{x})\, \mbox{\rm d}\mu(\mathbf{x}) = {\mathscr E}_K(\mu) + \frac{1}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n}-2\,\frac{\mathbf{p}_n^T(\mu)\mathbf{K}_n^{-1}\boldsymbol{1}_n}{\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n} - \tr\left[\mathbf{K}_n^{-1}\mathbf{Q}_n^\bot\mathbf{H}_n(\mu)\right],
\end{eqnarray*}
see \eqref{MSE1}, where $\mathbf{Q}_n^\bot = \mathbf{I}_n-\boldsymbol{1}_n\boldsymbol{1}_n^T\mathbf{K}_n^{-1}/(\boldsymbol{1}_n^T\mathbf{K}_n^{-1}\boldsymbol{1}_n)$, with $\mathbf{I}_n$ the $n$-dimensional identity matrix, is a projector onto the linear space orthogonal to $\boldsymbol{1}_n$, and where $\mathbf{H}_n(\mu)$ is the symmetric non-negative definite $n\times n$ matrix with elements
\begin{eqnarray*}
\{\mathbf{H}_n(\mu)\}_{j,k}=\int_{\mathscr X} K(\mathbf{x},\mathbf{x}_j)K(\mathbf{x},\mathbf{x}_k)\,\mbox{\rm d}\mu(\mathbf{x}) = \prod_{i=1}^d \int_{{\mathscr X}_i} K_i(x_i,{x_j}_i)K_i(x_i,{x_k}_i)\,\mbox{\rm d}\mu^{(i)}(x_i) \,. \fin
\end{eqnarray*}
\end{remark}
\vsp
Theorems~\ref{P:BQ-KD} and \ref{P:BQ-RK} indicate that, if $K$ is SPD, $(\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n)^{-1}$ is the minimum value of ${\mathscr E}_K(\xi_n-\mu)$ for measures $\xi_n\in{\mathscr M}(1)$. Hence, we can construct space-filling designs on a compact and convex subset~${\mathscr X}$ of $\mathds{R}^d$ by maximizing $\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n$ with respect to $\mathbf{X}_n\in\mathds{R}^{nd}$, taking $\mu$ uniform on~${\mathscr X}$. This can be performed using any unconstrained nonlinear programming algorithm, as Example~\ref{Ex:4.1} will illustrate. Note that, from \eqref{JK-RK} and Cauchy-Schwarz inequality, $(\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n)^{-1}\leq {\mathscr E}_K(\xi_{n,e}-\mu) = (\boldsymbol{1}_n^T\widetilde\mathbf{K}_n\boldsymbol{1}_n)/n^2$, the minimization of which was considered in Section~\ref{S:empirical}.
\subsection{Any-time designs}\label{S:any-time}
There exist situations where the number $n$ of design points ultimately used (for integration, or function approximation) differs from that initially planned, say $N$. It is the case in particular when function evaluations are computationally more expensive than expected, and numerical experimentation is stopped after $n<N$ simulations, or when simulations fail at some design points and testing at more than $N$ points is required to obtain $N$ valid evaluations in total. In such circumstances, it is convenient to have sequences of nested designs at one's disposal. The objective is then to construct any-time designs; that is, ordered sequences $\mathbf{x}_1,\mathbf{x}_2,\ldots$ of designs points such that any design $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}$ made of the first $n$ points of the sequence has good space-filling properties. A typical example is given by Low Discrepancy Sequences (LDS) in $[0,1]^d$, see \cite{Niederreiter92}.
\vsp
When $K$ is SPD, we may exploit expression \eqref{sn2} of the conditional variance of $I_\mu(f)$ in a greedy sequential construction: at step $n$ we choose $\mathbf{x}_{n+1}$ that minimizes $s_{n+1}^2$. This sequential construction, called Sequential Bayesian Quadrature in \cite{BriolOGO2015}, is straightforward to implement compared with global minimization of $s_n^2$, see \eqref{sn1}. Direct calculation, using formulae for the inversion of the block matrix
$$
\widetilde\mathbf{K}_{n+1}=\left(
\begin{array}{cc}
\widetilde\mathbf{K}_n & \widetilde\mathbf{k}_n(\mathbf{x}_{n+1}) \\
\widetilde\mathbf{k}_n^T(\mathbf{x}_{n+1}) & K_\mu(\mathbf{x}_{n+1},\mathbf{x}_{n+1})\\
\end{array}
\right)\,,
$$
where $\{\widetilde\mathbf{K}_n\}_{i,j}=K_\mu(\mathbf{x}_i,\mathbf{x}_j)$ and $\{\widetilde\mathbf{k}_n(\mathbf{x})\}_i=K_\mu(\mathbf{x},\mathbf{x}_i)$, $i,j=1,\ldots,n$, $\mathbf{x}\in{\mathscr X}$, gives
\begin{eqnarray}\label{sn+1^2}
s_{n+1}^2 = \left[\boldsymbol{1}_n^T\widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n+ \frac{(1-\widetilde\mathbf{k}_n(\mathbf{x}_{n+1})^T\widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n)^2} {K_\mu(\mathbf{x}_{n+1},\mathbf{x}_{n+1})-\widetilde\mathbf{k}_n^T(\mathbf{x}_{n+1})\widetilde\mathbf{K}_n^{-1}\widetilde\mathbf{k}_n(\mathbf{x}_{n+1})} \right]^{-1} \,.
\end{eqnarray}
The sequential construction is thus
$$
\mathbf{x}_{n+1}\in\Arg\max_{\mathbf{x}\in{\mathscr X}} \frac{(1-\widetilde\mathbf{k}_n(\mathbf{x})^T\widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n)^2} {K_\mu(\mathbf{x},\mathbf{x})-\widetilde\mathbf{k}_n^T(\mathbf{x})\widetilde\mathbf{K}_n^{-1}\widetilde\mathbf{k}_n(\mathbf{x})} \,.
$$
The conditional gradient algorithm of \cite{FrankW56} yields a simpler construction, particularly well adapted to the situation and also applicable when $K$ is unbounded. It relies on the sequential selection of points that minimize the current directional derivative of ${\mathscr E}_K(\xi-\mu)={\mathscr E}_{K_\mu}(\xi)$, with $\xi$ supported on design points previously selected. The algorithm is initialized at a measure $\xi^{(n_0)}$ supported on $\mathbf{X}_{n_0}\in{\mathscr X}^{n_0}$ (with for instance $n_0=1$ and $\xi^{(1)}=\delta_{\mathbf{x}_1}$ for some $\mathbf{x}_1\in{\mathscr X}$). Let $\xi^{(n)}$ denote the measure associated with the current design $\mathbf{X}_n$ of iteration $n$, with weights $w_i^{(n)}$, i.e., $\xi^{(n)}=\sum_{i=1}^n w_i^{(n)} \delta_{\mathbf{x}_i}$. Next design point is chosen in $\Arg\min_{\mathbf{x}\in{\mathscr X}} F_{K_\mu}(\xi^{(n)},\delta_\mathbf{x})$ (any minimizer can be selected in case there are several). Straightforward calculation using \eqref{Dir-der} gives $\mathbf{x}_{n+1} \in\Arg\min_{\mathbf{s}\in{\mathscr X}} [ P_{\xi^{(n)}}(\mathbf{x}) - P_\mu(\mathbf{x})]$, that is,
\begin{eqnarray}\label{x-n+1}
\mathbf{x}_{n+1} \in\Arg\min_{\mathbf{x}\in{\mathscr X}} \left[ \sum_{i=1}^n w_i^{(n)} K(\mathbf{x},\mathbf{x}_i) - P_\mu(\mathbf{x}) \right] \,.
\end{eqnarray}
Note that this construction is well defined even if $K$ is singular: in that case, it ensures that all design points are different ($\mathbf{x}_i\neq \mathbf{x}_j$ for all $i,j$); the same is true for all one-dimensional canonical projections when $K$ is the tensorised product of singular kernels.
After choosing $\mathbf{x}_{n+1}$, the measure $\xi^{(n)}$ is updated into
\begin{eqnarray}\label{xi-n+1}
\xi^{(n+1)} = (1-\alpha_n) \xi^{(n)} + \alpha_n \delta_{\mathbf{x}_{n+1}}
\end{eqnarray}
for some $\alpha_n\in[0,1]$, so that $\xi^{(n+1)}\in{\mathscr M}^+(1)$ when $\xi^{(n)}\in{\mathscr M}^+(1)$.
When $\xi^{(n_0)}$ is the empirical (uniform) measure on $\mathbf{X}_{n_0}$, the choice $\alpha_n=1/(n+1)$ implies that $\xi^{(n)}$ remains uniform on its support $\mathbf{X}_n$ for all $n$, see \cite{Wynn70} for an early contribution in the design context. The method is called \emph{kernel herding} in the machine-learning literature, see \cite{BachLJO2012, ChenWS2012, HuszarD2012}. It is shown in \cite{ChenWS2012} that ${\mathscr E}_K(\xi^{(n)}-\mu)={\mathcal O}(1/n^2)$ when ${\mathcal H}_K$ is finite dimensional, but we only have the weaker result ${\mathscr E}_K(\xi^{(n)}-\mu)={\mathcal O}(1/n)$ when ${\mathcal H}_K$ is infinite dimensional, see \cite{BachLJO2012}.
\begin{remark}
Denote $\xi^{(n+)}(\mathbf{x})=[n/(n+1)]\xi^{(n)}+[1/(n+1)]\delta_\mathbf{x}$. The direct minimization of ${\mathscr E}_K(\xi^{(n+)}(\mathbf{x})-\mu)$ with respect to $\mathbf{x}$ yields
$$
\mathbf{x}^{(n+1)} \in \Arg\min_{\mathbf{x}\in{\mathscr X}} \left[ \frac{1}{n+1}\, \sum_{i=1}^k K(\mathbf{x},\mathbf{x}_i) - P_\mu(\mathbf{x}) + \frac{1}{2(k+1)}\, K(\mathbf{x},\mathbf{x}) \right] \,,
$$
that is, a selection very close to \eqref{x-n+1} when $K(\mathbf{x},\mathbf{x})$ is constant (Mat\'ern kernel for instance). Note that this construction requires $K(\mathbf{x},\mathbf{x})<\infty$ for all $\mathbf{x}\in{\mathscr X}$, contrary to \eqref{x-n+1}.
\fin
\end{remark}
In practice $n$ is always smaller than some given $n_{\max}$, and to facilitate the construction we can restrict the choice of the $\mathbf{x}_i$ to a finite subset ${\mathscr X}_\Omega=\{\mathbf{s}_1,\ldots,\mathbf{s}_\Omega\}$ of~${\mathscr X}$, with $\Omega\gg n_{\max}$ (when ${\mathscr X}=[0,1]^d$, ${\mathscr X}_\Omega$ can be given by the first $\Omega$ points of a LDS).
For any $n\leq n_{\max}$, we can write $\mathbf{X}_n=\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}=\{\mathbf{s}_{i_1},\ldots,\mathbf{s}_{i_n}\}$,
the construction being initialized at some $n_0$-point design $\mathbf{X}_{n_0}\subset{\mathscr X}_\Omega$.
A measure $\xi$ supported on $\mathbf{X}_n$ can thus be written as
$\xi=\sum_{i=1}^\Omega \omega_i \delta_{\mathbf{s}_i}$, with $\omega_{i}=0$ when $\mathbf{s}_i\not\in\mathbf{X}_n$. Therefore, for all $n$, $\xi^{(n)}$ is fully characterized by a $\Omega$-dimensional vector $\boldsymbol{\omega}^{(n)}=(\omega_1^{(n)},\ldots,\omega_\Omega^{(n)})^T$, with $\boldsymbol{\omega}^{(n)}$ in the probability simplex $\mathds{P}_\Omega$ when $\xi^{(n)}\in{\mathscr M}^+(1)$. The updating equations (\ref{x-n+1}, \ref{xi-n+1}) then imply that $\boldsymbol{\omega}^{(n+1)}$ is obtained by moving $\boldsymbol{\omega}^{(n)}$ in the direction of a vertex of $\mathds{P}_\Omega$, hence the name \emph{vertex-direction} given to methods based on \eqref{xi-n+1} in the literature on optimal design, see, e.g., \cite[Chap.~9]{PP2013} and the references therein.
A summary of results on the rate of decrease of ${\mathscr E}_K(\xi^{(n)}-\mu)$ in this situation is given in Appendix~A. The cost of the determination of $\mathbf{x}_{n+1}$ in \eqref{x-n+1} is ${\mathcal O}(\Omega)$ (we need to compute $K(\mathbf{x},\mathbf{x}_n)$ for all $\mathbf{x}\in{\mathscr X}_\Omega$), and the cost for $n$ iterations scales as ${\mathcal O}(n\Omega)$ (including the initial cost for the computation of $P_\mu(\mathbf{x})$ for all $\mathbf{x}\in{\mathscr X}_\Omega$). An $n$-point design constructed in this way can be used as initialization for the (unconstrained) minimization of ${\mathscr E}_K(\xi_{n,e}-\mu)=\boldsymbol{1}_n^T\widetilde\mathbf{K}_n\boldsymbol{1}_n/n^2$ (Section~\ref{S:empirical}), or the maximization of $\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n$ (Section~\ref{S:BQ2}), with respect to $\mathbf{X}_n$. The resulting design $\mathbf{X}_n^*$ can in turn be used as candidate set for the greedy construction of \cite{Gonzalez85} (also called coffee-house design in \cite{Muller2000}), yielding a sequence of nested designs $\mathbf{X}_1,\ldots,\mathbf{X}_n=\mathbf{X}_n^*$.
\subsection{Illustrative examples}\label{Ex:4.1}
We take ${\mathscr X}=[0,1]^2$, $n_{\max}=100$, $n_0=1$ with $\mathbf{X}_1=\{(0.5,0.5)\}$; $\mu$ is uniform on~${\mathscr X}$ and ${\mathscr X}_\Omega$ is given by the first $2^{12}$ points of Sobol' LDS. The kernel $K$ is the tensor product of uni-dimensional Mat\'ern 3/2 covariance functions $K_{3/2,\theta}$, see \eqref{K32}.
Figure~\ref{F:designs-theta10-d2-n100}-Left shows the design $\mathbf{X}_{100}$ obtained after 99 iterations of (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$ ($\mathbf{X}_{100}$ is the support of $\xi^{(100)}$), for $\theta=10$ in $K_{3/2,\theta}$. This design has visually better space-filling properties than the first 100 points $\mathbf{S}_{100}$ of Sobol' sequence presented on the right part of the figure. This is confirmed by the numerical values of the covering and packing radii, respectively given by \eqref{CR} and \eqref{PR}: $\CR(\mathbf{X}_{100}) \simeq 0.0925 < \CR(\mathbf{S}_{100}) \simeq 0.1377$, and $\PR(\mathbf{X}_{100}) \simeq 0.0262 > \PR(\mathbf{S}_{100}) \simeq 0.0204$. Local maximization of $\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n$ (Section~\ref{S:BQ2}) with respect to $\mathbf{X}_n$, initialized at $\mathbf{X}_{100}$, yields a design $\mathbf{X}_{100}^*$ with better space-filling properties: $\CR(\mathbf{X}_{100}^*) \simeq 0.0869$ and $\PR(\mathbf{X}_{100}^*) \simeq 0.0364$. When minimizing ${\mathscr E}_K(\xi_{n,e}-\mu)$ with respect to $\mathbf{X}_n$ (Section~\ref{S:empirical}) we obtain $\CR(\mathbf{X}_{100}^*) \simeq 0.0955$ and $\PR(\mathbf{X}_{100}^*) \simeq 0.0359$. The performance is significantly worse, both in terms of $\CR$ and $\PR$, when the optimization is initialized at $\mathbf{S}_{100}$.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=.49\linewidth]{XKH_n100_cov5_th10_meth1.png} \includegraphics[width=.49\linewidth]{XSobol_n100.png}
\end{center}
\caption{\small Left: $\mathbf{X}_{100}=\mathrm{Supp}(\xi^{(100)})$ generated by (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$ and $\theta=10$ in $K_{3/2,\theta}$. Right: first 100 points $\mathbf{S}_{100}$ of Sobol' LDS.}
\label{F:designs-theta10-d2-n100}
\end{figure}
Figure~\ref{F:filling-theta10-d2-n100}-Left presents the efficiencies, in terms of covering radius \eqref{CR} (solid line) and packing radius \eqref{PR} (dashed line) of $\mathbf{S}_n$ corresponding to the first $n$ points of Sobol' sequence relatively to $\mathbf{X}_n=\mathrm{Supp}(\xi^{(n)})$, when $\xi^{(n)}$ is generated by (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$ and $\theta=10$ in $K_{3/2,\theta}$. Values smaller than one (shown by an horizontal line) indicate that $\mathbf{X}_n$ has better space-filling properties than $\mathbf{S}_n$.
A greedy coffee-house construction \cite{Gonzalez85, Muller2000} ensures $\CR(\mathbf{X}_n) \leq 2 \min_{\mathbf{X}'_n} \CR(\mathbf{X}'_n)$ and $\PR(\mathbf{X}_n) \geq (1/2)\, \max_{\mathbf{X}'_n} \PR(\mathbf{X}'_n)$, but the method is computationally expensive and tends to choose points on the border of~${\mathscr X}$. On the other hand, the construction is straightforward (and efficient) when we use the design $\mathbf{X}_{n_{\max}}^*$ obtained by maximizing $\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1}\boldsymbol{1}_n$ as candidate set. The first point $\mathbf{x}^{(1)} $ is chosen as the point in $\mathbf{X}_{n_{\max}}^*$ closest to the center of~${\mathscr X}$ (here $(1/2,1/2)$); then $\mathbf{x}^{(n+1)}$ is the point in $\mathbf{X}_{n_{\max}}^*$ furthest away from $\mathbf{X}_n=\{\mathbf{x}^{(1)},\ldots,\mathbf{x}^{(n)}\}$, for $n=1,\ldots,n_{\max}-1$. The efficiencies of $\mathbf{S}_n$, in terms of covering radius (solid line) and packing radius (dashed line), relatively to this new sequence, are plotted in Figure~\ref{F:filling-theta10-d2-n100}-Right.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=.49\linewidth]{Eff_n100_cov5_th10_meth1.png} \includegraphics[width=.49\linewidth]{Eff_n100_cov5_th10_meth1_forward.png}
\end{center}
\caption{\small Efficiencies $\CR(\mathbf{X}_{n})/\CR(\mathbf{S}_{n})$ (solid line) and $\PR(\mathbf{S}_{n})/\PR(\mathbf{X}_{n})$ (dashed line), with $\mathbf{S}_n$ corresponding to the first $n$ points of Sobol' sequence. Left: $\mathbf{X}_n=\mathrm{Supp}(\xi^{(n)})$ is generated by (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$ and $\theta=10$ in $K_{3/2,\theta}$. Right: $\mathbf{X}_n$ is generated by coffee-house design with candidate set $\mathbf{X}_{100}^*$ obtained by local maximization of $\boldsymbol{1}_n^T \widetilde\mathbf{K}_n^{-1} \boldsymbol{1}_n$.}
\label{F:filling-theta10-d2-n100}
\end{figure}
Not surprisingly, the performance of $\xi^{(n)}$ in terms of MMD are better than those of the empirical measure $\xi_{n,e}(\mathbf{S}_n)$ associated with $\mathbf{S}_n$, see Figure~\ref{F:performance-theta10-d2-n100}-Left.
Figure~\ref{F:performance-theta10-d2-n100}-Right illustrates the fact that a larger correlation length in $K$ yields a faster decrease of ${\mathscr E}_K[\xi_{n,e}(\mathbf{S}_n)-\mu]$ (compare with the figure on the left). On the other hand, this faster decrease does not mean that design points are better distributed, compare Figure~\ref{F:designs-theta1-d2-n100}-Left with Figures~\ref{F:designs-theta10-d2-n100}-Left.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=.49\linewidth]{JKH_n100_cov5_th10_meth1.png} \includegraphics[width=.49\linewidth]{JKH_n100_cov5_th1_meth1.png}
\end{center}
\caption{\small ${\mathscr E}_K(\xi^{(n)}-\mu)$ (solid line) and ${\mathscr E}_K[\xi_{n,e}(\mathbf{S}_n)-\mu]$ (dashed line) as functions of $n$ (log scale). $\mathbf{X}_n=\mathrm{Supp}(\xi^{(n)})$ is generated by (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$ and the kernel $K_{3/2,\theta}$; $\mathbf{S}_n$ corresponds to the first $n$ points of Sobol' sequence. Left: $\theta=10$ in $K_{3/2,\theta}$; Right: $\theta=1$.}
\label{F:performance-theta10-d2-n100}
\end{figure}
Finally, Figure~\ref{F:designs-theta1-d2-n100}-Right illustrates an application of Algorithm (\ref{x-n+1}, \ref{xi-n+1}) in higher dimension with a singular kernel: ${\mathscr X}=[0,1]^{10}$ and $K$ is the tensor product of the one-dimensional logarithmic kernel $K_{(0)}$ in \eqref{Riesz}. Similarly to Figure~\ref{F:filling-theta10-d2-n100}, the figure presents the efficiencies $\CR(\mathbf{X}_{n})/\CR(\mathbf{S}_{n})$ and $\PR(\mathbf{S}_{n})/\PR(\mathbf{X}_{n})$ for $n=1,\ldots,100$. Due to the large value of $d$, for a given design $\mathbf{Z}_n$, $\CR(\mathbf{Z}_n)$ defined by \eqref{CR} is under-approximated by $\CR'(\mathbf{Z}_n)=\max_{\mathbf{x}\in{\mathscr X}'_{\Omega'}}\min_{\mathbf{z}\in\mathbf{Z}_n} \|\mathbf{x}-\mathbf{z}\|$, with ${\mathscr X}'_{\Omega'}$ formed by the first $2^{19}$ points of Sobol' sequence complemented with a $3^d$ full factorial design.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=.49\linewidth]{XKH_n100_cov5_th1_meth1.png} \includegraphics[width=.49\linewidth]{PhimMgrid_PhiMm_n100_Klog_meth1_Sobol_d10.png}
\end{center}
\caption{\small Left: $\mathbf{X}_{100}=\mathrm{Supp}(\xi^{(100)})$; ${\mathscr X}=[0,1]^2$, $\mathbf{X}_n$ is generated by (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$ and the kernel $K_{3/2,\theta}$ with $\theta=1$. Right: Efficiencies $\CR(\mathbf{X}_{n})/\CR(\mathbf{S}_{n})$ (solid line) and $\PR(\mathbf{S}_{n})/\PR(\mathbf{X}_{n})$ (dashed line); ${\mathscr X}=[0,1]^{10}$, $\mathbf{X}_n=\mathrm{Supp}(\xi^{(n)})$ generated by (\ref{x-n+1}, \ref{xi-n+1}) with $\alpha_n=1/(n+1)$, and $K(\mathbf{x},\mathbf{x}')=\prod_{i=1}^{10} \log(1/|x_i-x'_i|)$, $\mathbf{S}_n$ corresponds to the first $n$ points of Sobol' sequence.}
\label{F:designs-theta1-d2-n100}
\end{figure}
|
\section{Introduction}
A {\it ruled surface} in Euclidean space $\mathbb{R}^3$ is a surface formed
by a $1$-parameter family of straight lines, called {\it rulings};
at least partly,
it admits a parametrization of the form $F(s, t)=\mbox{\boldmath $r$}(s)+t\mbox{\boldmath $e$}(s)$ with $|\mbox{\boldmath $e$}(s)|=1$,
$s \in I$, $t \in \mathbb{R}$, where $I$ is an open interval.
A {\it developable surface} is a ruled surface which is locally planar
(i.e. the Gaussian curvature is constant zero).
The parametrization map $F: I \times \mathbb{R} \to \mathbb{R}^3$ may be singular at some point $(s_0, t_0)$,
that is, the differential $dF(s_0, t_0)$ may have rank one,
and then the surface (= the image of $F$) has a particularly singular shape around that point.
In this paper, we study local diffeomorphic types of of the singular surface and its bifurcations (see Fig.1).
All maps and manifolds are assumed to be class $C^\infty$ throughout.
\begin{figure}[h]
\includegraphics[width=7cm]{h2.png}\\
\caption{\small
Deforming Mond's $H_2$-singularity via a family of ruled surfaces:
the surface has two crosscaps and one triple point.
}\label{fig1}
\end{figure}
The main feature of this paper is to combine
classical line geometry using {\it dual quaternions} \cite{Guggenheimmer, Hlavaty, PW, Selig}
and $\mathscr{A}$-classification theory of singularities of (frontal) maps $\mathbb{R}^2 \to \mathbb{R}^3$
\cite{Mond85, Ishikawa, IST, IS}.
Here $\mathscr{A}$ denotes a natural equivalence relation in singularity theory of $C^\infty$ maps;
two map-germs $f, g: (\mathbb{R}^2, 0) \to (\mathbb{R}^3,0)$
are {\it $\mathscr{A}$-equivalent} if
there exist diffeomorphism-germs $\sigma: (\mathbb{R}^2,0) \to (\mathbb{R}^2,0)$
and $\varphi: (\mathbb{R}^3,0) \to (\mathbb{R}^3,0)$ such that $g=\varphi\circ f \circ \sigma^{-1}$.
We simply say the {\it $\mathscr{A}$-type} of a map-germ to mean its $\mathscr{A}$-equivalence class.
As a weaker notion,
{\it topological $\mathscr{A}$-equivalence} is defined
by taking $\sigma$ and $\varphi$ to be homeomorphism-germs.
We also use the $\mathscr{A}$-equivalence with
the target changes being rotations $\varphi \in SO(3)$, which
is called {\it rigid equivalence} throughout the present paper.
Our aim is to classify germs of parametrization maps $F$ of ruled surfaces in $\mathbb{R}^3$
up to $\mathscr{A}$-equivalence and rigid equivalence.
\subsection{Ruled surfaces}
Geometric Algebra is a neat tool for studying motions in classical geometry;
in the case of Euclidean $3$-space, it is the algebra of dual quaternions (e.g. Selig \cite{Selig}).
As an application, any ruled surface in $\mathbb{R}^3$ is described as a curve of {\it unit dual vectors}
$$\check{\bv}: I \to \check{\mathbb{U}} \subset
\mathbb{D}^3, \quad
\check{\bv}(s)=\mbox{\boldmath $v$}_0(s)+\varepsilon \mbox{\boldmath $v$}_1(s) $
Here, $\mathbb{D} =\mathbb{R} \oplus \varepsilon \mathbb{R}$ with $\varepsilon^2=0$ is
the $\mathbb{R}$-algebra of {\it dual numbers},
and $\mathbb{D}^3=\mathbb{R}^3 \oplus \varepsilon \mathbb{R}^3$ is the space of {\it dual vectors},
and especially,
the space of unit dual vectors is defined by
$$\check{\mathbb{U}}:=\{\check{\mbox{\boldmath $v$}}=\mbox{\boldmath $v$}_0+\varepsilon \mbox{\boldmath $v$}_1 \in \mathbb{D}^3, \;
|\mbox{\boldmath $v$}_0|=1, \; \mbox{\boldmath $v$}_0\cdot \mbox{\boldmath $v$}_1=0\}$$
which is a $4$-dimensional submanifold in the $6$-dimensional space $\mathbb{D}^3$.
Obviously, $\check{\mathbb{U}}$ is diffeomorphic to the total space of the (co)tangent bundle $TS^2$,
thus it is naturally identified with the space of oriented lines in $\mathbb{R}^3$,
so $\check{\mbox{\boldmath $v$}}$ is regarded as a $1$-parameter family of oriented lines;
the ruled surface is parameterized by $F(s,t)=\mbox{\boldmath $v$}_0(s)\times \mbox{\boldmath $v$}_1(s)+t\mbox{\boldmath $v$}_0(s)$.
See \S 2.1 for the detail.
In our context, as the space of ruled surfaces in $\mathbb{R}^3$,
we consider the space $C^\infty(I, \check{\mathbb{U}})$ of curves of unit dual vectors
endowed with the Whintey $C^\infty$-topology.
Assume that our ruled surface is {\it non-cylindrical},
i.e., $\mbox{\boldmath $v$}_0'(s)\not=0$ for any $s\in I$,
then the curve $\check{\bv}$ admits the Frenet formula in $\mathbb{D}^3$ with
{\it complete} differential invariants,
the {\it dual curvature}
and the {\it dual torsion}
$$\check{\kappa}(s)=\kappa_0(s)+\varepsilon \kappa_1(s),
\quad \check{\tau}(s)=\tau_0(s)+\varepsilon \tau_1(s) \;\;\; \in \mathbb{D}.$$
Here we may take $s$ to be the arclength of the spherical curve $\mbox{\boldmath $v$}_0(s)$,
that is equivalent to $\kappa_0(s)\equiv 1$,
thus three real functions $\kappa_1, \tau_0, \tau_1$ are essential.
In particular, $\kappa_1(s_0)=0$ if and only if
$F$ is singular at $(s_0, t_0)$ for some (unique) $t_0$ (Lemma \ref{striction_curve}).
We determine which $\mathscr{A}$-types of singular germs $(\mathbb{R}^2, 0) \to (\mathbb{R}^3,0)$ appear in
generic families of ruled surfaces.
Assume that $F$ is singular at $(s_0, t_0)=(0,0)$ and $F(0,0)=0$.
From the {\it dual Bouquet formula} of $\check{\bv}$ at $s=0$ in $\mathbb{D}^3$,
we derive a {\it canonical} Taylor expansion of parameterization map $F$ (\S 3.2),
where $o(n)$ denotes Landau's notation of function-germs of order greater than $n$:
$$
\left\{
\begin{array}{ccl}
x&=&
t-\frac{1}{2}ts^2+\frac{\tau_1(0)}{2} s^3 + o(3)\\
y&=&
ts-\frac{\tau_1(0)}{2}s^2-\frac{2\tau_0(0)\kappa_1'(0)+\tau_1'(0)}{6}s^3 + o(3)\\
z&=&
\frac{\kappa_1'(0)}{2}s^2+\frac{\tau_0(0)}{2}t s^2
+\frac{\kappa_1''(0)-2\tau_0(0)\tau_1(0)}{6}s^3+ o(3)
\end{array}
\right.
$$
Then we apply to the jet of $F$ the criteria
for detecting $\mathscr{A}$-types of map-germs in Mond \cite{Mond82, Mond85}.
\begin{thm}\label{thm1}
The $\mathscr{A}$-classification of singularities of $F$
arising in generic at most $3$-parameter families of
non-cylindrical ruled surfaces is given as in Table \ref{table1};
In particular, for each $\mathscr{A}$-type in that table,
the canonical expansion with the described condition is regarded as
a normal form of the jet of ruled surface-germ under rigid equivalence.
\end{thm}
\begin{table
{
$$
\begin{array}{l | l | l | l }
&\mbox{\footnotesize normal form} & \ell & \mbox{\footnotesize cond. at $s=s_0$}\\
\hline \hline
S_0 & (x, y^2, xy) &2& \kappa_1=0, \;\; \kappa_1'\not=0 \\
\hline
S_1^\pm& (x, y^2, y^3\pm x^2y)
&3& \kappa_1=\kappa_1'=0 , \;\; \tau_1\not=0, \;\;
\kappa_1''(\kappa_1''-2\tau_0\tau_1)\gtrless 0\\
\hline
S_2& (x,y^2,y^3+ x^3y)
&4&\kappa_1=\kappa_1'=\kappa_1''=0,\;
\kappa_1^{(3)}\tau_0\tau_1\neq 0\\
B_2^\pm& (x,y^2,x^2y\pm y^5)
&& \kappa_1=\kappa_1'=0,\;
\kappa_1''=2\tau_0\tau_1\neq 0,\; b_2\gtrless 0\\
H_2& (x,xy+y^5,y^3)
&& \kappa_1=\kappa_1'=\tau_1=0,\; \kappa_1''\neq 0,\; h_2\neq 0\\
\hline
S_3^\pm &(x,y^2,y^3\pm x^4y)
&5& \kappa_1=\kappa_1'=\kappa_1''=\kappa_1^{(3)}=0,\; \kappa_1^{(4)}\tau_0\tau_1 \gtrless 0\\
C_3^\pm &(x,y^2,xy^3\pm x^3y)
&& \kappa_1=\kappa_1'=\kappa_1''=\tau_0=0, \; \tau_1\neq 0, \; \kappa_1^{(3)}(\kappa_1^{(3)}-2\tau_0'\tau_1)\gtrless 0 \\
B_3^\pm &(x,y^2,x^2y\pm y^7)
&& \kappa_1=\kappa_1'=0,\;
\kappa_1''=2\tau_0\tau_1\neq 0,\; b_2=0,\ b_3\gtrless 0
\\
H_3 &(x,xy+y^7,y^3)
&& \kappa_1=\kappa_1'=\tau_1=0,\;
\kappa_1''\neq 0,\; h_2=0,\; h_3\neq 0
\\
P_3 &(x,xy+y^3,xy^2+p_4 y^4)
&&
\kappa_1=\kappa_1'=\kappa_1'' =\tau_1= 0,\;
\tau_0\tau_1'\neq 0, \; p_4 \not=0, 1, \frac{1}{2}, \frac{3}{2}.
\end{array}
$$
}
\caption{$\mathscr{A}$-types of singularities of ruled surfaces.
In conditions, $\kappa_1', \kappa_1'', \cdots$ are derivatives at $s=s_0$ for short,
e.g. $\kappa_1'$ means $\frac{d}{ds}\kappa_1(s_0)$,
and $b_2, b_3, h_2, h_3, p_4$ are
some polynomials of those derivatives (see \S 3.2).
The letters $\lessgtr, \gtrless, \pm$ are in the same order.
In the second column, $\ell$ means $\mathscr{A}$-codimension of the map-germ. }
\label{table1}
\end{table}
Precisely saying, via a variant of Thom's transversality theorem (\S 3.3),
we show that
there exists a dense subset $\mathcal{O}$ in
the mapping space $\mathcal{R}_W$ consisting of families of non-cylindrical $\check{\bv}: I\times W \to \check{\mathbb{U}}$
with parameter space $W$ of dimension $\le 3$
so that for any family belonging to $\mathcal{O}$ and for any $\lambda \in W$,
the germ of the corresponding paramatrization map
$F(-, \lambda): I \times \mathbb{R} \to \mathbb{R}^3$ at any point $(s_0, t_0)$ is $\mathscr{A}$-equivalent to either of an immersion-germ or one of singular germs in Table \ref{table1}.
Obviously, normal forms under rigid equivalence have functional moduli:
those are nothing but $\kappa_1(s)$, $\tau_0(s)$ and $\tau_1(s)$ having the prescribed condition
on derivatives at $s=s_0$.
\begin{rem}\label{rem_realization}\upshape
({\bf Realization})
Izumiya-Takeuchi \cite{IT} firstly proved
in a rigorous way that a generic singularity of ruled surfaces is only of type crosscap $S_0$,
and Martins and Nu\~no-Ballesteros \cite{MN} showed that
any $\mathscr{A}$-simple map-germ $(\mathbb{R}^2, 0) \to (\mathbb{R}^3, 0)$ is $\mathscr{A}$-equivalent to a germ of ruled surface.
By our theorem,
$\mathscr{A}$-types which are not realized by ruled surfaces must have $\mathscr{A}$-codimension $\ge 6$.
This is sharp: for example, the $3$-jet $(x, y^3, x^2y)$, over which there are $\mathscr{A}$-orbits of codimension $6$,
is never $\mathscr{A}^3$-equivalent to $3$-jets of any non-cylindrical or cylindrical ruled surfaces (Remark \ref{5jet}).
The realizability of versal families of $\mathscr{A}$-types via
families of ruled surfaces can also be verified: for each germ in Table \ref{table1},
an $\mathscr{A}_e$-versal deformation is obtained
via deforming three invariants $\kappa_1, \tau_0, \tau_1$ appropriately (Remark \ref{versal}).
\end{rem}
\begin{rem}\label{rem_horo}\upshape
({\bf Conformal GA})
Our approach would be applicable
to other Clifford algebras and corresponding geometries.
For instance, Izumiya-Saji-Takahashi \cite{IST}
classified local singularities of horospherical flat surfaces in Lorentzian space
(conformal spherical geometry);
a horospherical surface is described by a curve in the Lie algebra $\mathfrak{so}(3,1)$.
Conformal Geometric Algebra may fit with this setting
and our approach should work.
\end{rem}
\begin{rem}\upshape
({\bf Framed curves})
Take the space of dual vectors $\mathbb{D}^3$ instead of $\check{\mathbb{U}}$.
A curve $I \to \mathbb{D}^3$ corresponds to a {\it framed curve},
which describes a $1$-parameter family of Euclidean motions of $\mathbb{R}^3$;
various geometric aspects of framed curves have
recently been studied by e.g. Honda-Takahashi \cite{HT}.
Since the dual Frenet formula is available for regular framed curves,
we may rebuild the theory by using dual quaternions.
That would be useful for singularity analysis in several topics of applied mathematics
such as 3D-interpolation via ruled/developable surfaces, 1-parameter motions of axes in robotics,
and so on (cf. \cite{PW, Selig}).
\end{rem}
\subsection{Developable surfaces}
For a non-cylindrical ruled surface,
it is developable (the Gaussian curvature zero) if and only if
$\kappa_1= 0$ identically (see \S 2).
Thus two real functions $\tau_0, \tau_1$ are complete invariants of developables.
Izumiya-Takeuchi \cite{IT} classified generic singularities of developable surfaces rigorously,
and Kurokawa \cite{Kurokawa} treated with a similar task for $1$-parameter families of developables.
We generalize their results systematically using the complete invariants.
\begin{thm}\label{thm2}
The $\mathscr{A}$-classification of singularities of $F$
arising in generic at most $2$-parameter families of non-cylindrical developable surfaces
is given as in Table \ref{table2}; In particular,
for each $\mathscr{A}$-type in that table,
the canonical expansion with the described condition is regarded as
a normal form of the jet of developable-germ under rigid equivalence.
\end{thm}
\begin{table}
{
$$
\begin{array}{l | l | l | l }
&\mbox{\footnotesize normal form} &\ell
& \mbox{\footnotesize cond. at $s=s_0$}\\
\hline\hline
cE& (x, y^2, y^3)
&1& \tau_0\not=0, \;\; \tau_1\not=0 \\
\hline
cS_0& (x, y^2, xy^3)
&2& \tau_1\not=0, \;\; \tau_0=0, \;\; \tau_0'\not=0 \\
cS_1^+& (x, y^2, y^3(x^2+y^2))
&3& \tau_1\not=0, \;\; \tau_0=\tau_0'=0, \;\; \tau_0''\not=0 \\
cC_3^+& (x, y^2, y^3(x^3+xy^2))
&4& \tau_1\not=0, \;\; \tau_0=\tau_0'=\tau_0''=0, \;\; \tau_0'''\not=0 \\
\hline
Sw& (x, xy + 2 y^3, xy^2 + 3 y^4)
&2&\tau_0\not=0, \;\; \tau_1=0, \;\; \tau_1'\not=0 \\
cA_4& (x, xy + \frac{5}{2} y^4, xy^2 + 4 y^5)
&3&\tau_0\not=0, \;\; \tau_1=\tau_1'=0, \;\; \tau_1''\not=0\\
cA_5& (x, xy + 3 y^5, xy^2 + 5y^6)^\dagger
&4&\tau_0\not=0, \;\; \tau_1=\tau_1'=\tau_1''=0, \;\; \tau_1'''\not=0\\
\hline
T_1& (x, xy + y^3, 0) + o(3)
&3&\tau_0=\tau_1=0, \;\; \tau_1'\not=0\\
T_2& (x, xy, 0) + o(3)
&4& \tau_0=\tau_1=\tau_1'=0
\end{array}
$$
}
\caption{$\mathscr{A}$-types of singularities of developable surfaces.
An exception is the type $cA_5$;
the condition implies that the germ is {\it topologically} $\mathscr{A}$-equivalent to the normal form $\dagger$
(in this case, the striction curve $\sigma$ is topologically determinative in the sense of Ishikawa \cite{Ishikawa}).}
\label{table2}
\end{table}
\begin{rem}\label{rem1}
\upshape
({\bf Realization})
In our classification process \S 4.1, we see that
non-cylindrical developables do not admit $\mathscr{A}$-types
$$cS_1^-: (x, y^2, y^3(x^2-y^2))\;\; \mbox{and}\;\; cC_3^-: (x, y^2, y^3(x^3-xy^2))$$
(for the former, it was shown in \cite{Kurokawa}),
while $cS_1^+$ and $ cC_3^+$ appear.
Furthermore, $\tau_1\not=0$ and $\tau_0=\tau_0'=\tau_0''=0$
if and only if the $5$-jet of $F$ is equivalent to $(x, y^2, 0)$, and
thus, for instance, we see that frontal singularities of
cuspidal $S$ and $B$-types
$$cS_*: (x, y^2, y^3(y^2+h(x,y^2))), \quad
cB_*: (x, y^2, y^3(x^2+h(x,y^2)))$$
($h(x,y^2)=o(2)$) never appear in our developable surfaces.
Similarly, since $\tau_1=0$ if and only if the $2$-jet is reduced to $(x, xy, 0)$,
wavefronts of cuspidal beaks/lips type $A_3^\pm$ and purse/pyramid types $D_k$ never appear.
Indeed, their $2$-jets are equivalent to $(x, 0, 0)$ and $(x^2\pm y^2, xy,0)$ respectively
(it is obvious to see no appearance of $D_k$, for the corank of our maps $F$ is at most one).
\end{rem}
A non-cylindrical developable surface, which is not a cone,
is re-parametrized as the tangent developable of the {\it striction curve} $\sigma(s)$ (Lemma \ref{striction_curve2}).
Here $\sigma(s)$ may be singular;
recall that for a possibly singular space curve,
its tangent developable is defined by
the closure of the union of tangent lines at smooth points;
indeed, it is a frontal surface,
see \S 2.4 (cf. Ishikawa \cite{Ishikawa99}).
A space curve-germ is called to be of type $(m, m+\ell, m+\ell+r)$ if
it is $\mathscr{A}$-equivalent to the germ
$$x=s^m+o(m), \;\; y=s^{m+\ell}+o(m+\ell), \;\; z= s^{m+\ell+r}+o(m+\ell+r)$$
(the curve is called to be of {\it finite type} if $m, n, \ell <\infty$).
A type of curve-germ is called {\it smoothly determinative} (resp. {\it topologically determinative})
if it determines the $\mathscr{A}$-type (resp. topological $\mathscr{A}$-type) of the tangent developable.
Ishikawa \cite{Ishikawa, Ishikawa99} gave the following complete characterization
(Mond \cite{Mond89} for the case of $m=1$, i.e. smooth curves):
\begin{itemize}
\item[(i)] smoothly determinative types are only $(1, 2, 2+r)$, $(2, 3, 4)$, $(1, 3, 4)$, $(3, 4, 5)$ and $(1, 3, 5)$;
\item[(ii)] $(m, m+\ell, m+\ell+r)$ is topologically determinative
if and only if $\ell, r$ are not both even, or $m=1$ and $\ell, r$ are both even.
\end{itemize}
Using this result,
we obtain a complete topological $\mathscr{A}$-classification of singularities of
non-cylindrical developable surfaces:
\begin{thm}\label{thm3} {\bf (Topological classification)}
For a non-cylindrical developable surface,
the germ of its striction curve $\sigma(s)$ at $s=s_0$ has the type
$$(m, m+1, m+1+r)$$
where $m-1$ and $r-1$ are orders of $\tau_1$ and $\tau_0$ at $s=s_0$, respectively:
$$\tau_1=\tau_1'=\cdots =\tau_1^{(m-2)}=\tau_0=\tau_0'=\cdots = \tau_0^{(r-2)}=0,
\quad \tau_1^{(m-1)}\tau_0^{(r-1)}\not=0.$$
In particular,
topological $\mathscr{A}$-type of
the germ of $F$ at singular points are completely determined by
orders of the dual torsion $\check{\tau}=\tau_0+\varepsilon \tau_1$.
\end{thm}
\begin{rem}\label{mond_tangent}\upshape
Theorem \ref{thm3} is regarded as the {\it dual version} of
a result of Mond \cite{Mond89} and Ishikawa \cite{Ishikawa}:
$\mathscr{A}$-type of the tangent developable of
a {\it non-singular} space curve $\sigma$ with non-zero curvature is determined
by the vanishing order of its torsion function.
This is the case that $\sigma$ is of type $(1,2,2+r)$,
and then the torsion of $\sigma$ has the same order of $\tau_0$ (Lemma \ref{striction_curve2}).
Note that in our theorem above, $\sigma(s)$ can be singular and
non-zero curvature condition is replaced by the non-cylindrical condition.
\end{rem}
\begin{rem}\upshape
Table \ref{table2} is separated into three parts.
One is the case of $\tau_1(s_0)\not=0$;
they are the tangent developables of non-singular curves of type $(1,2,2+r)$,
which are {\it frontal singularities} as mentioned in Remark \ref{mond_tangent}.
The second is the case of $\tau_0(s_0)\not=0$;
they are the tangent developables of singular curves of type
$(2,3,4)$, $(3,4,5)$ and $(4,5,6)$,
which are {\it wavefronts}
-- the former two types are smoothly determinative,
while the third one is topologically determinative, by Ishikawa's characterization.
In the remaining part, types $T_0$ and $T_1$ are tangent developable of curves of type
$(2,3,4+r)\; (r \ge 1)$.
Tangent developables of curves of types
$(1,3,3+r), (2,4,4+r)$ etc are cylindrical at $s=s_0$.
\end{rem}
\begin{rem}\upshape
Not only striction curves but also several other kind of characteristic curves on a ruled surface
can be discussed. For instance, flecnodal curves are important
in projective differential geometry of surfaces \cite{Kabata, SKSO}.
\end{rem}
The rest of this paper is organized as follows.
In \S 2, we briefly review two main ingredients for non-experts in each subject --
the first is the algebra of dual quaternions,
which is the most basic Geometric Algebra, and the second is about
useful criteria for detecting $\mathscr{A}$-types in singularity theory of maps.
In \S 3, we apply the $\mathscr{A}$-criteria to the canonical expansion of $F$
at singular points and prove Theorem \ref{thm1}.
In \S 4, we proceed to the case of developable surfaces
and prove Theorems \ref{thm2} and \ref{thm3}.
This paper is based on the first author's \cite{Tanaka}.
The second author was partly supported by JSPS KAKENHI Grant Number 15K13452.
\section{Preliminaries}
{\it Geometric Algebra} is a new look at Clifford algebras, which is nowadays recognized as a very neat tool for describing motions in Klein geometries in the context of a vast of applications to physics, mechanics and computer vision.
In \S2.1 and \S2.2, we give a very quick summary on the geometric algebra for $3$-dimensional Euclidean motions and its application to the geometry of ruled surfaces. A good compact reference is the nineth chapter of Selig \cite{Selig}
(also see \cite{Guggenheimmer, IT, ICRT, PW}).
In 2.3 and 2.4, we briefly describe some basic notions in Singularity Theory,
which will be used in Sections 3 and 4.
We deal with two classes of $C^\infty$ maps from a surface into $\mathbb{R}^3$;
ordinary smooth maps of corank at most one, i.e. $\dim \ker df\le 1$ (Mond \cite{Mond85})
and frontal maps (Ishikawa \cite{Ishikawa}, Izumiya-Saji \cite{IS}).
\subsection{Dual quaternions}
Let $\mathbb{H}$ denote the field of quaternions:
$q=a+bi+cj+dk$.
The conjugate of $q$ is $\bar{q}=a-bi-cj-dk$ and
the norm is given by $|q|=\sqrt{q\bar{q}}$.
Decompose $\mathbb{H}$ into the real and the imaginary parts,
$\mathbb{H}=\mathbb{R}\oplus \Imag\, \mathbb{H}$,
where one identifies $bi+cj+dk \in \Imag\, \mathbb{H}$ with $\mbox{\boldmath $v$}=(b, c, d)^T \in \mathbb{R}^3$
equipped with the standard inner and exterior products.
We write $q=a+\mbox{\boldmath $v$}$, then the multiplication of $\mathbb{H}$ is written as
$(a+\mbox{\boldmath $v$})(b+\mbox{\boldmath $u$})=
(ab-\mbox{\boldmath $v$}\cdot \mbox{\boldmath $u$}) + (a\mbox{\boldmath $u$}+b\mbox{\boldmath $v$} +\mbox{\boldmath $v$} \times \mbox{\boldmath $u$})$.
The quaternionic unitary group
$$\mathbb{H}_1=Sp(1)=\{q \in \mathbb{H}, |q|=1\}$$
is naturally isomorphic to $SU(2)$, that doubly covers $SO(3)$;
indeed, $\pm q\in \mathbb{H}_1$ defines the rotation $\mbox{\boldmath $x$} \mapsto q\mbox{\boldmath $x$} \bar{q}$.
The Lie algebra of $\mathbb{H}_1$ is just $\Imag\, \check{\mathbb{H}}=\mathbb{R}^3$.
Put $\mathbb{D}=\mathbb{R}[\varepsilon]/\langle \varepsilon^2 \rangle$,
and call it the algebra of {\it dual numbers}.
A dual number $a+\varepsilon b$ is invertible if $a\not=0$,
and it has a square root if $a>0$.
The $\mathbb{R}$-algebra of {\it dual quaternions} is defined by
$$\check{\mathbb{H}}:= \mathbb{D}^4 =\mathbb{H} \otimes_\mathbb{R} \mathbb{D} = \{\;
\check{q}=q_0+\varepsilon q_1 \; | \; q_0, q_1 \in \mathbb{H} \;\}.$$
That is identified with the even Clifford algebra $C\ell^+(0,3,1)$ \cite[\S 9.3]{Selig}.
The conjugate of $\check{q}$ is defined by $\check{q}^*:=\bar{q}_0+\varepsilon \bar{q}_1$,
and then $\check{q}\check{q}^*=|q_0|^2+\varepsilon {\rm Re}[q_1\bar{q}_0]$.
The Lie group of {\it unit dual quaternions} is defined by
$$\check{\mathbb{H}}_1:=\{\; \check{q} \in \check{\mathbb{H}} \; | \; \check{q}\check{q}^*=1\; \}.$$
This group is isomorphic to the semi-direct product
$\mathbb{H}_1 \ltimes \Imag\, \mathbb{H}=Sp(1) \ltimes \mathbb{R}^3$ via
the correspondance $\check{q} \leftrightarrow (q_0, q_1\bar{q}_0)$.
Then, $\check{\mathbb{H}}_1$ doubly covers $SE(3)=SO(3) \ltimes \mathbb{R}^3$,
the group of Euclidean motions of $\mathbb{R}^3$; the action $\check\Theta$ of $\check{\mathbb{H}}_1$ on $\mbox{\boldmath $x$} \in \mathbb{R}^3$ is given by
$$1+\varepsilon \check\Theta(\check{q})\mbox{\boldmath $x$}:=\check{q}(1+ \varepsilon \mbox{\boldmath $x$}) \check{q}^*
=1+\varepsilon (q_0\mbox{\boldmath $x$} \bar{q}_0+2q_1\bar{q}_0).$$
That is, $q_0$ and $2q_1\bar{q}_0$ express a rotation and a parallel transition, respectively.
The Lie algebra of $\check{\mathbb{H}}_1$ is canonically identified with
the space of {\it dual vectors}
$$\mathbb{D}^3 = \Imag\, \mathbb{H} \otimes_\mathbb{R} \mathbb{D},
\quad
\check{\mbox{\boldmath $v$}}=\mbox{\boldmath $v$}_0+\varepsilon \mbox{\boldmath $v$}_1 \quad (\mbox{\boldmath $v$}_0, \mbox{\boldmath $v$}_1 \in \Imag\, \mathbb{H}=\mathbb{R}^3),$$
which is a $\mathbb{D}$-submodule of $\check{\mathbb{H}}=\mathbb{D}^4$.
The standard inner and exterior products of $\mathbb{R}^3$ are extended to
$\mathbb{D}$-bilinear operations on $\mathbb{D}^3$;
$$\textstyle
\check{\mbox{\boldmath $u$}}\cdot \check{\mbox{\boldmath $v$}}
:= - \frac{1}{2}(\check{\mbox{\boldmath $u$}}\check{\mbox{\boldmath $v$}}+\check{\mbox{\boldmath $v$}}\check{\mbox{\boldmath $u$}}) \in \mathbb{D},
\qquad
\textstyle
\check{\mbox{\boldmath $u$}}\times \check{\mbox{\boldmath $v$}}:=
\frac{1}{2}(\check{\mbox{\boldmath $u$}}\check{\mbox{\boldmath $v$}}-\check{\mbox{\boldmath $v$}}\check{\mbox{\boldmath $u$}}) \in \mathbb{D}^3.
$$
A {\it unit dual vector} means a dual vector $\check{\mbox{\boldmath $v$}} \in \mathbb{D}^3$ with
$\check{\mbox{\boldmath $v$}}\cdot \check{\mbox{\boldmath $v$}}=1$, i.e. $ |\mbox{\boldmath $v$}_0|=1$, $\mbox{\boldmath $v$}_0\cdot \mbox{\boldmath $v$}_1=0$
(indeed, it corresponds to a $2$-blade in the Clifford algebra \cite[\S 10.1]{Selig}).
Denote the set of unit dual vectors by $\check{\mathbb{U}}$,
which is identified with
the space of oriented lines in $\mathbb{R}^3$ in the following way:
$$\mbox{oriented lines}: \mbox{\boldmath $v$}_0\times \mbox{\boldmath $v$}_1+t \mbox{\boldmath $v$}_0 \;\;
\stackrel{1:1}{\longleftrightarrow} \;\;
\mbox{unit dual vectors}: \mbox{\boldmath $v$}_0+\varepsilon \mbox{\boldmath $v$}_1.$$
This expression is very useful \cite[\S9.3]{Selig}: For instance,
\begin{enumerate}
\item[(i)]
a point $\mbox{\boldmath $a$} \in \mathbb{R}^3$ lies on the line corresponding to
a unit dual vector $\mbox{\boldmath $v$}_0+\varepsilon \mbox{\boldmath $v$}_1$
if and only if $\mbox{\boldmath $a$} \times \mbox{\boldmath $v$}_0=\mbox{\boldmath $v$}_1$;
\item[(ii)]
two lines intersect perpendicularly if and only if
corresponding unit dual vectors $\check{\mbox{\boldmath $u$}}$ and $\check{\mbox{\boldmath $v$}}$ satisfy that
$\check{\mbox{\boldmath $u$}}\cdot \check{\mbox{\boldmath $v$}}=0$.
\end{enumerate}
\subsection{Ruled and developable surfaces}
Using the identification just mentioned above,
a ruled surface is exactly described as a curve of unit dual vectors:
$$\check{\bv}: I \to \check{\mathbb{U}} \subset \mathbb{D}^3, \quad
\check{\bv}(s)=\mbox{\boldmath $v$}_0(s)+\varepsilon \mbox{\boldmath $v$}_1(s)$$
($I$ an open interval) with
$|\mbox{\boldmath $v$}_0(s)|=1$ and $\mbox{\boldmath $v$}_0(s)\cdot\mbox{\boldmath $v$}_1(s)=0$ ($s \in I$).
Interpreting it as an object in $\mathbb{R}^3$, we have a parametrization
$$
F(s,t)=\mbox{\boldmath $r$}(s)+t\mbox{\boldmath $e$}(s) \quad (\mbox{\boldmath $r$}=\mbox{\boldmath $v$}_0\times\mbox{\boldmath $v$}_1,\quad\mbox{\boldmath $e$}=\mbox{\boldmath $v$}_0).
$$
Note that $|\mbox{\boldmath $e$}(s)|=1$ and $\mbox{\boldmath $r$}\cdot \mbox{\boldmath $e$}=0$.
Let $R_{s}$ denote the ruling defined by $\check{\bv}(s)$ and put
$$R=R(\check{\bv}):=\bigcup_{s\in I} R_s \; \subset \mathbb{R}^3.$$
Formally, $\check{\bv}(s)$ looks like a $\mathbb{D}$-version of the velocity vector of a space curve.
That leads us to define the {\it curvature} $\check{\kappa}(s)$ of $\check{\bv}$ by
$$\check{\kappa}(s)=\kappa_0(s)+\varepsilon\kappa_1(s)
:=\sqrt{\check{\bv}'(s)\cdot\check{\bv}'(s)}
=|\mbox{\boldmath $v$}_0'|+
\varepsilon \frac{\;\; \mbox{\boldmath $v$}_0'\cdot\mbox{\boldmath $v$}_1'}{|\mbox{\boldmath $v$}_0'|}
\; \in \mathbb{D},
$$
provided $\check{\bv}$ is non-cylindrical, i.e., $\mbox{\boldmath $v$}_0'(s)\not=0\; (s \in I)$.
Here $\prime$ means $\frac{d}{ds}$.
From now on, we assume that
$$|\mbox{\boldmath $v$}_0'(s)|=1$$
by taking $s$ to be the arc-length of $\mbox{\boldmath $v$}_0$.
Then, $\check{\kappa}=1+\varepsilon \mbox{\boldmath $v$}_0'\cdot \mbox{\boldmath $v$}_1'$ and
thus $\check{\kappa}^{-1}=1-\varepsilon \mbox{\boldmath $v$}_0'\cdot \mbox{\boldmath $v$}_1'$.
Put
$$
\check{\bn}(s) =\mbox{\boldmath $n$}_0(s)+\varepsilon\mbox{\boldmath $n$}_1(s):=\check{\kappa}^{-1}\check{\bv}'(s),
$$
and
$$
\check{\bt}(s)={\mbox{\boldmath $t$}}_0(s)+\varepsilon{\mbox{\boldmath $t$}}_1(s):=\check{\bv}(s)\times\check{\bn}(s).
$$
Then for every $s \in I$, three dual vectors
$\check{\bv}(s)$, $\check{\bn}(s)$ and $\check{\bt}(s)$ form a basis of
the $\mathbb{D}$-module ${\rm Im}\, \check{\mathbb{H}}=\mathbb{D}^3$ satisfying
$$
\begin{array}{c}
\check{\bv}\times\check{\bn}=\check{\bt},\quad
\check{\bt}\times\check{\bv}=\check{\bn},\quad
\check{\bn}\times\check{\bt}=\check{\bv},
\\
\check{\bv}\cdot\check{\bn}=\check{\bn}\cdot\check{\bt}
=\check{\bt}\cdot\check{\bv}
=0,\quad
\check{\bv}\cdot\check{\bv}=\check{\bn}\cdot\check{\bn}
=\check{\bt}\cdot\check{\bt}=1.
\end{array}
$$
From these relations and the property (ii) of unit dual vectors mentioned before,
we see that three lines corresponding to unit dual vectors $\check{\bv}, \check{\bn}, \check{\bt}$
meet at one point and are mutually perpendicular;
in particular, $\mbox{\boldmath $v$}_0, \mbox{\boldmath $n$}_0, \mbox{\boldmath $t$}_0$ forms an orthonormal basis of $\mathbb{R}^3$.
We define the {\it torsion} $\check{\tau}(s)$ of $\check{\bv}$ by
$$\check{\tau}(s)=\tau_0(s)+\varepsilon\tau_1(s):=
\check{\bn}'(s)\cdot\check{\bt}(s) \;\; \in \mathbb{D}.$$
The following theorem is classical:
\begin{thm}\label{frenet}
{\rm (cf. Guggenheimmer \cite[\S 8.2]{Guggenheimmer}, Selig \cite[\S 9.4]{Selig})}
Assume that $s$ is the arc-length of $\mbox{\boldmath $v$}_0$, i.e. $\kappa_0(s)=|\mbox{\boldmath $v$}_0'(s)|=1$.
\begin{enumerate}
\item {\rm (Frenet formula)}
It holds that
\begin{equation}\label{frenet-eq}
\frac{d}{ds}
\left[\small\begin{array}{c}
\check{\bv}(s)\\\check{\bn}(s)\\\check{\bt}(s)
\end{array}\right]
=
\left[\small\begin{array}{ccc}
0&\check{\kappa}(s)&0
\\ -\check{\kappa}(s)&0&\check{\tau}(s)
\\0&-\check{\tau}(s)&0
\end{array}\right]
\left[\small\begin{array}{c}
\check{\bv}(s)\\\check{\bn}(s)\\\check{\bt}(s)
\end{array}\right].
\nonumber
\end{equation}
\item
The dual curvature $\check{\kappa}(s)$ and the dual torsion $\check{\tau}(s)$
are complete invariants of the ruled surface $R$ up to Euclidean motions.
That is,
for two curves $\check{\bv}_1$ and $\check{\bv}_2$,
they have the same invariants $\check{\kappa}$ and $\check{\tau}$
if and only if
ruled surfaces $R(\check{\bv}_1)$ and $R(\check{\bv}_2)$ in $\mathbb{R}^3$ are transformed to
each other by some Euclidean motion.
\item
$R(\check{\bv})$ is a developable surface (including a cone) if and only if $\kappa_1=0$ identically.
In particular, $\tau_0, \tau_1$ are complete invariants of the developable surface.
\end{enumerate}
\end{thm}
The {\it striction curve} of a ruled surface $R$ is the curve having minimal length
which meets all the rulings of $R$.
Let $F(s,t)=\mbox{\boldmath $r$}(s)+t\mbox{\boldmath $e$}(s)$ be a canonical parametrization
($\mbox{\boldmath $r$}\cdot \mbox{\boldmath $e$}=0$, $|\mbox{\boldmath $e$}|=|\mbox{\boldmath $e$}'|=1$),
then the striction curve $\sigma(s)$ is characterized by the equation
$\sigma'\cdot \mbox{\boldmath $e$}'=0$ (cf. \cite[p.218]{Selig}, \cite[Lemma 2.1]{IT}, \cite[\S5.3]{PW}).
We then have the following:
\begin{lem} \label{striction_curve0}
For a non-cylindrical ruled surface,
it holds that
\begin{enumerate}
\item
$\sigma(s)=\mbox{\boldmath $r$}(s)-(\mbox{\boldmath $r$}'(s)\cdot \mbox{\boldmath $e$}'(s))\mbox{\boldmath $e$}(s)$,
\item
$\sigma \times \mbox{\boldmath $v$}_0=\mbox{\boldmath $v$}_1$,
$\sigma \times \mbox{\boldmath $n$}_0=\mbox{\boldmath $n$}_1$ and
$\sigma \times \mbox{\boldmath $t$}_0=\mbox{\boldmath $t$}_1$,
\item
$\sigma'(s)=\tau_1(s) \mbox{\boldmath $v$}_0(s)+\kappa_1(s)\mbox{\boldmath $t$}_0(s)$,
\item
$\kappa_1 =\det(\mbox{\boldmath $e$}, \mbox{\boldmath $e$}',\mbox{\boldmath $r$}')$,
$\tau_0=\det(\mbox{\boldmath $e$},\mbox{\boldmath $e$}',\mbox{\boldmath $e$}'')$,
$\tau_1 = \sigma' \cdot \mbox{\boldmath $e$}$.
\end{enumerate}
\end{lem}
From (2) and the property (i) of unit dual vectors in \S2.1,
it follows that $\sigma(s)$ lies on each of three lines corresponding to unit dual vectors $\check{\bv}(s), \check{\bn}(s), \check{\bt}(s)$,
that is, $\sigma(s)$ is the locus of the center of moving orthogonal frames.
For completeness we prove the lemma, although it is elementary.
\
\noindent {\sl Proof} :\;
It is easy to see (1) by differentiating $\sigma(s)=\mbox{\boldmath $r$}(s)+t(s)\mbox{\boldmath $e$}(s)$.
We show (2).
First, by $\check{\bn}\cdot \check{\bv}=0$,
we see that
$\mbox{\boldmath $n$}_1\cdot \mbox{\boldmath $v$}_0=-\mbox{\boldmath $v$}_1\cdot \mbox{\boldmath $n$}_0$, and similarly $ \mbox{\boldmath $n$}_1\cdot \mbox{\boldmath $t$}_0=-\mbox{\boldmath $t$}_1\cdot \mbox{\boldmath $n$}_0$.
By the Frenet formula,
$\mbox{\boldmath $v$}_0'=\mbox{\boldmath $n$}_0$, $\mbox{\boldmath $t$}_0'=-\tau_0\mbox{\boldmath $n$}_0$, $\mbox{\boldmath $v$}_1'=\kappa_1\mbox{\boldmath $n$}_0+\mbox{\boldmath $n$}_1$ and
$\mbox{\boldmath $t$}_1'=-\tau_0\mbox{\boldmath $n$}_1-\tau_1\mbox{\boldmath $n$}_0$.
Since $\mbox{\boldmath $r$}=\mbox{\boldmath $v$}_0\times \mbox{\boldmath $v$}_1$ and $\mbox{\boldmath $e$}=\mbox{\boldmath $v$}_0$,
it follows from (1) that
$$\sigma=
-(\mbox{\boldmath $t$}_1\cdot \mbox{\boldmath $n$}_0)\mbox{\boldmath $v$}_0-(\mbox{\boldmath $v$}_1\cdot \mbox{\boldmath $t$}_0)\mbox{\boldmath $n$}_0-(\mbox{\boldmath $n$}_1\cdot \mbox{\boldmath $v$}_0)\mbox{\boldmath $t$}_0.$$
Thus
$\sigma \times \mbox{\boldmath $v$}_0=
-(\mbox{\boldmath $v$}_1\cdot \mbox{\boldmath $t$}_0)\mbox{\boldmath $n$}_0\times \mbox{\boldmath $v$}_0-(\mbox{\boldmath $n$}_1\cdot \mbox{\boldmath $v$}_0)\mbox{\boldmath $t$}_0\times \mbox{\boldmath $v$}_0
=(\mbox{\boldmath $v$}_1\cdot \mbox{\boldmath $t$}_0)\mbox{\boldmath $t$}_0+(\mbox{\boldmath $v$}_1\cdot \mbox{\boldmath $n$}_0)\mbox{\boldmath $n$}_0=\mbox{\boldmath $v$}_1$, for $\mbox{\boldmath $v$}_1\cdot \mbox{\boldmath $v$}_0=0$.
That yields (2).
Differentiating the first one of (2),
$$0=(\sigma\times \mbox{\boldmath $v$}_0)'-\mbox{\boldmath $v$}_1'=
(\sigma'\times \mbox{\boldmath $v$}_0+\sigma \times \mbox{\boldmath $n$}_0)-(\kappa_1\mbox{\boldmath $n$}_0+\mbox{\boldmath $n$}_1)
=\sigma'\times \mbox{\boldmath $v$}_0-\kappa_1\mbox{\boldmath $n$}_0$$
and similarly $\sigma'\times \mbox{\boldmath $t$}_0+\tau_1\mbox{\boldmath $n$}_0=0$.
Substitute $\sigma'=a\mbox{\boldmath $v$}_0+b\mbox{\boldmath $n$}_0+c\mbox{\boldmath $t$}_0$ for those equalities,
we obtain $a=\tau_1$, $b=0$, $c=\kappa_1$, that is (3).
Finally, (4) is easy, e.g., $\kappa_1 = \mbox{\boldmath $v$}_0'\cdot \mbox{\boldmath $v$}_1'=\mbox{\boldmath $e$}'\cdot (\mbox{\boldmath $r$}'\times \mbox{\boldmath $e$})
=\det(\mbox{\boldmath $e$}, \mbox{\boldmath $e$}',\mbox{\boldmath $r$}')$.
\hfill $\Box$
\begin{lem}\label{striction_curve}
{\rm (Izumiya et al \cite[Lemma 2.2]{IT}, \cite[\S 1]{ICRT})}
For a non-cylindrical ruled surface,
$F$ is singular at $(s_0, t_0)$ if and only if $\kappa_1(s_0)=0$ and
$t_0=-\mbox{\boldmath $r$}'(s_0)\cdot \mbox{\boldmath $e$}'(s_0)$.
The singular value $F(s_0, t_0)$ is the point $\sigma(s_0)$
where the curve $\sigma(s)$ is tangent to the ruling $R_{s_0}$ or
$\sigma'(s_0)=0$.
\end{lem}
\noindent {\sl Proof} :\;
$\frac{\partial F}{\partial s}(s_0) \times \frac{\partial F}{\partial t}(s_0)
=(\mbox{\boldmath $r$}'(s_0)+t_0\mbox{\boldmath $e$}'(s_0))\times \mbox{\boldmath $e$}(s_0)=0$
$\Leftrightarrow$ $\mbox{\boldmath $r$}'(s_0)=\alpha \mbox{\boldmath $e$}(s_0)-t_0\mbox{\boldmath $e$}'(s_0)$
for some $\alpha\not=0$
$\Leftrightarrow$
$\det(\mbox{\boldmath $e$}(s_0),\mbox{\boldmath $e$}'(s_0),\mbox{\boldmath $r$}'(s_0))=0$ and $t_0=-\mbox{\boldmath $r$}'(s_0)\cdot \mbox{\boldmath $e$}'(s_0)$.
The second claim follows from (3) in Lemma \ref{striction_curve0}.
\hfill $\Box$
\
In case of $\kappa_1=0$ identically, Lemmas \ref{striction_curve0} and \ref{striction_curve} imply that
singular points of $F$ form a non-singular curve
$s \mapsto (s, -\mbox{\boldmath $r$}'(s)\cdot \mbox{\boldmath $e$}'(s)) \in I \times \mathbb{R}$
and the image of this curve is just the striction curve $\sigma(s)$.
Note that
$\sigma(s)$ is a non-singular space curve, if $\tau_1\not=0$;
especially,
$F$ is written by $\sigma(s)+\tilde{t}\sigma'(s)$ with
$\tilde{t}=(t+\mbox{\boldmath $r$}'(s)\cdot \mbox{\boldmath $e$}'(s))/\tau_1$.
\begin{lem} {\rm (Izumiya et al \cite[\S 1]{ICRT})} \label{striction_curve2}
A non-cylindrical developable surface, which is not a cone,
is re-parametrized as the tangent developable of the striction curve $\sigma(s)$.
The curve $\sigma$ is non-singular whenever $\tau_1\not=0$, and then
the curvature $\kappa_\sigma$ and the torsion $\tau_\sigma$ of $\sigma$ are given respectively by
$$\kappa_\sigma=\frac{|\sigma'\times \sigma''|}{|\sigma'|^3}=\frac{1}{\tau_1}, \qquad
\tau_\sigma=
\frac{\det(\sigma', \sigma'', \sigma''')}{|\sigma'\times \sigma''|^2}=\frac{\tau_0}{\tau_1}.$$
\end{lem}
\subsection{$\mathscr{A}$-classification of map-germs}
A {\it singular} point of $f: M \to N$ between manifolds means a point $p \in M$ where
$df_p$ is neither injective nor surjective
(then $f(p) \in N$ is called a {\it singular value} of $f$);
we denote by $S(f)\subset M$ the set of singular points of $f$.
Two maps $\tilde{f}: U \to N$ and $\tilde{g}: V \to N$
on neighborhoods $U$ and $V$ of $p\in M$
define the same {\it map-germ at $p$}
if there is a neighborbood $W\subset U\cap V$ of $p$
so that $\tilde{f}|_W\equiv \tilde{g}|_W$;
a {\it map-germ at $p$} is an equivalence class of maps under this relation,
denoted by $f: (M, p) \to (N, f(p))$.
Two map-germs at $p$ have the same {\it $k$-jet}
if they have the same Taylor polynomials at $p$ of order $k$ in some local coordinates;
a $k$-jet is such an equivalence class of map-germs, denoted by $j^kf(p)$.
Two germs $f: (M, p) \to (N, q)$ and $g: (M', p') \to (N', q')$ are {\it $\mathscr{A}$-equivalent} if
they commute each other
via diffeomorphism-germs $\sigma$ and $\tau$:
$$
\xymatrix{
(M, p) \ar[r]^f \ar[d]_{\sigma}^{\simeq} & (N, q) \ar[d]^{\tau}_{\simeq}\\
(M', p')\ar[r]_g & (N', q')
}
$$
For simplicity,
we consider map-germs $(\mathbb{R}^m,0) \to (\mathbb{R}^n,0)$ and the $\mathscr{A}$-equivalence by
the action of diffeomorphisms $\sigma$ and $\tau$ preserving the origins.
At the $k$-jet level, {\it $\mathscr{A}^k$-equivalence} is defined.
A germ $f:(\mathbb{R}^m,0) \to (\mathbb{R}^n,0)$ is said to be {\it $k$-$\mathscr{A}$-determined} if
any germs $g:(\mathbb{R}^m,0) \to (\mathbb{R}^n,0)$ with $j^kg(0)=j^kf(0)$ is $\mathscr{A}$-equivalent to $f$;
such germs are collectively referred to as {\it finitely $\mathscr{A}$-determined} germs.
For instance, the germ $(x, y^2, xy)$ is $2$-determined.
Let $J^k(m,n)$ be the {\it jet space} consisting of all $k$ jets of $(\mathbb{R}^m,0) \to (\mathbb{R}^n,0)$,
which is identified with
the affine space of Taylor coefficients of order $r$ ($1\le r\le k$)
in a fixed system of local coordinates.
The codimension of the $\mathscr{A}$-orbit of a germ $f$ in
the space of all map-germs $(\mathbb{R}^m,0) \to (\mathbb{R}^n,0)$
is called the {\it $\mathscr{A}$-codimension} of $f$;
the $\mathscr{A}$-codimension of $f$ is finite if and only if $f$ is finitely $\mathscr{A}$-determined (see e.g. \cite{GG}).
Thanks to finite determinacy,
the process of $\mathscr{A}$-classification is reduced to a finite dimensional problem:
we stratify $J^k(m,n)$ invariantly under the $\mathscr{A}^k$-equivalence step by step
from low order $k$ and low codimension.
For instance,
using several determinacy criteria,
$\mathscr{A}$-classification of map-germs $(\mathbb{R}^2, 0) \to (\mathbb{R}^3, 0)$ up to certain codimension
has been established in Mond \cite{Mond82, Mond85}.
In \S 3, we will follow Mond's classification process.
Furthermore, in Mond \cite{Mond82, Mond89},
a special class of map-germs $(\mathbb{R}^2, 0) \to (\mathbb{R}^3, 0)$ is considered.
A map germ $f: (\mathbb{R}^2,0) \to (\mathbb{R}^3, 0)$ is of class {\it CE} (i.e. cuspidal edge),
if $\rank df(0)=1$ and the singular point set $S(f)$ is non-singular.
A germ $f$ in CE is {\it $k$-$\mathscr{A}$-determined in CE}
if any germ $g$ in CE with the same $k$-jet
as $j^kf(0)$ is $\mathscr{A}$-equivalent to $f$.
In \S 4, we will use the following criteria of determinacy in CE \cite[Lem.1.1, Prop.1.2]{Mond89}.
\begin{prop}[Mond \cite{Mond89}] \label{CE} It holds that \\
$i)$ If $f \in CE$ and $j^2f(0)=(x,y^2,0)$, then $f$ is $\mathscr{A}$-equivalent to the germ $g(x,y)=(x,y^2,y^3p(x,y^2))$ for some smooth function $p(u,v)$; \\
$ ii)$ $f(x,y)=(x,y^2, y^3)$ is $3$-determined in CE; \\
$ iii)$ $f(x,y)=(x, y^2, yp(x, y^2))$ and $g(x,y)=(x, y^2, yq(x, y^2))$ are $\mathscr{A}$-equivalent if and only if $\tilde{f}(x,y)=(x, y^2, y^3p(x, y^2))$ and $\tilde{g}(x,y)=(x, y^2, y^3q(x, y^2))$ are $\mathscr{A}$-equivalent. In particular, $f$ is $(k-2)$-determined if and only if $\tilde{f}$ is $k$-determined in CE. \end{prop}
\subsection{Singularities of frontal surfaces}
There is a special class of surfaces, called {\it frontal surfaces}.
Let $ST^*\mathbb{R}^3$ be the spherical cotangent bundle with respect to the standard metric of $\mathbb{R}^3$
equipped with the standard contact structure. Let $U$ be an open set of $\mathbb{R}^2$.
A map $\iota: U \to ST^*\mathbb{R}^3$ is called {\it isotropic} if it satisfies that
the image $d\iota(T_pU)$ is contained in the contact plane $K_{\iota(p)}$ for any $p \in U$.
A {\it frontal map}
is the composed map $f=\pi\circ \iota: U \to \mathbb{R}^3$ of an isotropic map $\iota$ and the projection
$\pi: ST^*\mathbb{R}^3\to \mathbb{R}^3$.
The image (possibly singular) surface is called to be {\it frontal}.
An isotropic immersion $\iota$ is usually called a {\it Lagrange immersion},
and $\pi\circ \iota$ and its image are called a {\it Lagrange map} and a {\it wavefront}, respectively.
Let $f: U \to \mathbb{R}^3$ be a frontal map with $\nu: U \to S^2$ so that
$\iota=(f, \nu): U \to ST^*\mathbb{R}^3=\mathbb{R}^3\times S^2$ is an isotropic map.
We identifies $T\mathbb{R}^3\simeq T^*\mathbb{R}^3$ using the standard metric,
then the unit vector $\nu$ is always orthogonal to the subspace $df(T_pU)$ at any $p \in U$.
Let $x, y$ be coordinates of $U$ and
put $\lambda(x,y)=\det \left[\frac{\partial f}{\partial x}, \frac{\partial f}{\partial y}, \nu \right](x,y)$;
then the singular point set $S(f)$ is defined by $\lambda(x,y)=0$.
If $d\lambda(p)\not=0$, then $p$ is called a {\it non-degenerate} singular point.
In particular, if $p$ is non-degenerate and $\mbox{rank}\, df_p=1$,
the germ $f$ at $p$ is of class CE.
For a developable surface with $\mbox{\boldmath $e$}\times \mbox{\boldmath $e$}'\not=0$,
set $f: U \to \mathbb{R}^3$ to be $f(s,t):=\mbox{\boldmath $r$}(s)+t\mbox{\boldmath $e$}(s)$.
Then $f$ is a frontal map;
in fact, it suffices to put $\nu=\mbox{\boldmath $e$}\times \mbox{\boldmath $e$}'/|\mbox{\boldmath $e$}\times \mbox{\boldmath $e$}'|$
(then $\frac{\partial f}{\partial t}\cdot \nu=\mbox{\boldmath $e$}\cdot \nu=0$ and
$\frac{\partial f}{\partial s} \cdot \nu=(\mbox{\boldmath $r$}'+t\mbox{\boldmath $e$}')\cdot \nu=\det(\mbox{\boldmath $r$}', \mbox{\boldmath $e$}, \mbox{\boldmath $e$}')=0$).
Note that any singularities of $f$ are non-degenerate and have corank one
(see the comment before Lemma \ref{striction_curve2}).
There are two cases:
If $\iota=(f, \nu)$ is singular,
then it is easy to see that the $2$-jet of $f$ is $\mathscr{A}^2$-equivalent to $(x,y^2,0)$, and hence
Mond's criteria for map-germs of class CE (Proposition \ref{CE}) can be applied.
If $\iota$ is non-singular, i.e. $\iota$ is a Legendre immersion,
then the $2$-jet is equivalent to $(x, xy,0)$, and thus Proposition \ref{CE} is useless.
In this case, we employ the Legendre singularity theory.
There are known useful criteria of \cite{IS}
(precisely saying, the topological type $cA_5$ is not dealt in \cite{IS} but
the same argument as in Appendix of \cite{IS} works as well):
\begin{prop} {\rm (Izumiya-Saji \cite[Theorem 8.1]{IS}) }
Let $f: U \to \mathbb{R}^3$ be a Legendre map, and $p$
a non-degenerate singular point with $\mbox{rank}\, df_p=1$.
Let $\eta$ be an arbitrary vector field around $p$
so that $\eta(q)$ spans $\ker df_q$ at any $q \in S(f)$.
Then $f$ is $\mathscr{A}$-equivalent to $cE$, $Sw$, $cA_4$ or $cA_5$
if the following condition holds:
$$
\begin{array}{l | l}
cE & \eta\lambda(p)\not=0\\
Sw & \eta\lambda(p)=0, \; \eta\eta\lambda(p)\not=0 \\
cA_4 & \eta\lambda(p)=\eta\eta\lambda(p)=0, \; \eta\eta\eta\lambda(p)\not=0 \\
cA_5 & \eta\lambda(p)=\eta\eta\lambda(p)=\eta\eta\eta\lambda(p)=0, \;
\eta\eta\eta\eta\lambda(p)\not=0.
\end{array}
$$
\end{prop}
Through the theory of frontal maps and generating functions,
Ishikawa \cite{Ishikawa, Ishikawa99}
showed that the tangent developable of a curve of type
$(m, m+\ell, m+\ell+r)$ has a parametrization $F: (\mathbb{R}^2, 0) \to (\mathbb{R}^3,0)$ defined by
\begin{eqnarray*}
x&=&t\\
y&=&s^{m+\ell}+s^{m+\ell+1}\varphi(s)+t(s^\ell+s^{\ell+1}\phi(s)),\\
z&=&(\ell+r)(m+\ell+r)\int_0^s u^r\frac{\partial y(u,t)}{\partial u} du\\
&=& (\ell+r)(m+\ell)s^{m+\ell+r}+\cdots + t(\ell(m+\ell+r)s^{\ell+r}+\cdots)
\end{eqnarray*}
with some $C^\infty$ functions $\varphi(s)$ and $\phi(s)$.
These two function must be related to invariants $\tau_0$ and $\tau_1$.
It is also shown \cite[Thm 2.1]{Ishikawa}
that the topological type of the tangent developable of a space curve is determined
by type $(m, m+\ell, m+\ell+r)$ of the curve, unless both $\ell, r$ are even,
as mentioned in Introduction.
\section{Singularities of ruled surfaces}
In this section, we prove Theorem \ref{thm1} (2);
first we give a certain stratification of the jet space of triples of functions
$(\kappa_1, \tau_0, \tau_1)$, and
then discuss a variant of Thom's transversality theorem.
\subsection{Dual Bouquet formula}
Consider a curve $\check{\bv}: I \to \mathbb{D}^3$, $\check{\bv}(s)=\mbox{\boldmath $v$}_0(s)+\varepsilon \mbox{\boldmath $v$}_1(s)$,
with $\check{\bv}\cdot \check{\bv}=1$ and $|\mbox{\boldmath $v$}_0'(s)|=1$ as in \S 2.2.
We are concerned with the germ of $\check{\bv}$ at the origin ($s_0=0$).
Throughout this section,
let $\check{\kappa}, \check{\tau}, \check{\kappa}', \check{\tau}', \cdots$ denote
their values at $s=0$ for short, e.g. $\check{\kappa}'=\check{\kappa}'(0)$,
unless specifically mentioned.
By iterated uses of the Frenet formula (Theorem \ref{frenet} (1)),
we obtain
the ``Bouquet formula" of the curve in $\mathbb{D}^3$ at $s=0$;
$$
\check{\bv}(s)=\sum_{n=0}^r
\frac{\check{\bv}^{(n)}(0)}{n!}s^n + o(r) \;\; \in \mathbb{D}^3
$$
with
{\small
\begin{eqnarray*}
\check{\bv}'(0)&=&\check{\kappa}\,\check{\bn}(0),\\
\check{\bv}''(0)&=&-\check{\kappa}^2\,\check{\bv}(0)+\check{\kappa}'\,\check{\bn}(0)+\check{\kappa}\check{\tau}\,\check{\bt}(0),\\
\check{\bv}^{(3)}(0)&=&-3\check{\kappa}\ckappa'\,\check{\bv}(0)
+(\check{\kappa}''-\check{\kappa}^3-\check{\kappa}\check{\tau}^2)\,\check{\bn}(0)
+(2\check{\kappa}'\check{\tau}+\check{\kappa}\check{\tau}')\,\check{\bt}(0), \\
\check{\bv}^{(4)}(0)&=&(\check{\kappa}^4+\check{\kappa}^2\check{\tau}^2-4\check{\kappa}\ckappa'')\check{\bv}(0)+(\check{\kappa}^{(3)}-6\check{\kappa}^2\check{\kappa}'-3\check{\kappa}'\check{\tau}^2-3\check{\kappa}\check{\tau}\ctau')\check{\bn}(0)
\\&&+(3\check{\kappa}''\check{\tau}+3\check{\kappa}'\check{\tau}'-\check{\kappa}^3\check{\tau}+\check{\kappa}\check{\tau}''-\check{\kappa}\check{\tau}^3)\check{\bt}(0), \\
\check{\bv}^{(5)}(0)
&=&(10\check{\kappa}^3\check{\kappa}'+5\check{\kappa}\ckappa'\check{\tau}^2+5\check{\kappa}^2\check{\tau}\ctau'-5\check{\kappa}\ckappa^{(3)})\check{\bv}(0)
+(\check{\kappa}^{(4)}-6\check{\kappa}^2\check{\kappa}''-6\check{\kappa}''\check{\tau}^2\\
&&-12\check{\kappa}'\check{\tau}\ctau'-3\check{\kappa}(\check{\tau}')^2-4\check{\kappa}\check{\tau}\ctau''+\check{\kappa}^3\check{\tau}^2+\check{\kappa}\check{\tau}^4)\check{\bn}(0)
+(4\check{\kappa}^{(3)}\check{\tau}+6\check{\kappa}''\check{\tau}'\\
&&+3\check{\kappa}'\check{\tau}''-9\check{\kappa}^2\check{\kappa}'\check{\tau}-\check{\kappa}^3\check{\tau}'+\check{\kappa}'\check{\tau}''+\check{\kappa}\check{\tau}^{(3)}-4\check{\kappa}'\check{\tau}^3-6\check{\kappa}\check{\tau}^2\check{\tau}')\check{\bt}(0),
\end{eqnarray*}
}
and so on.
A similar (but more naive) expansion written by Pl\"ucker coordinates, instead of dual quaternions,
can be found in a classical literature, Hlavat\'y \cite{Hlavaty}.
Since dual vectors $\{\check{\bv}(0), \check{\bn}(0), \check{\bt}(0)\}$ form a $\mathbb{D}$-basis of $\Imag\, \check{\mathbb{H}}=\mathbb{D}^3$,
we write
$$\textstyle
\check{\bv}(s)=[\, \check{\bv}(0),\check{\bn}(0),\check{\bt}(0)\,]\; \check{\mbox{\boldmath $w$}}(s)$$
where
$$\textstyle
\check{\mbox{\boldmath $w$}}(s)
=[1-\frac{1}{2}\check{\kappa}^2 s +\cdots, \check{\kappa} s + \frac{1}{2}\check{\kappa}'s + \cdots, \frac{1}{2}\check{\kappa}\check{\tau} s^2+\cdots]^T \in \mathbb{D}^3.$$
Recall that
three oriented lines in $\mathbb{R}^3$ determined by unit dual vectors
$\check{\bv}(0), \check{\bn}(0), \check{\bt}(0)$ meet at one point,
which is nothing but the striction point $\sigma(0)$,
as mentioned just after Lemma \ref{striction_curve0}.
By an Euclidean motion,
the three lines can be transformed to standard coordinate axises of $\mathbb{R}^3$, i.e.,
$\mbox{\boldmath $v$}_0(0), \mbox{\boldmath $n$}_0(0), \mbox{\boldmath $t$}_0(0)$ form
the standard basis $i, j, k$ of $\Imag\, \mathbb{H} = \mathbb{R}^3$, respectively,
and $\mbox{\boldmath $v$}_1(0)=\mbox{\boldmath $n$}_1(0)=\mbox{\boldmath $t$}_1(0)=0 \in \mathbb{R}^3$.
Namely, we may assume that
the $3\times 3$ matrix (with entries in $\mathbb{D}$) $[\,\check{\bv}(0),\check{\bn}(0),\check{\bt}(0)\,]$ is the identity,
so $\check{\bv}(s)=\check{\mbox{\boldmath $w$}}(s)$.
Then,
$$\check{\bv}(s)=\mbox{\boldmath $v$}_0(s)+\varepsilon \mbox{\boldmath $v$}_1(s)=
\left[\begin{array}{c}1\\s\\0
\end{array}\right]
+\varepsilon
\left[\begin{array}{c}
0\\ \kappa_1 s\\0
\end{array}\right]+o(1).
$$
At a point $(0, t_0) \in I \times \mathbb{R}$, the Taylor expansion of
$F(s,t)=\mbox{\boldmath $v$}_0(s)\times \mbox{\boldmath $v$}_1(s)+t\mbox{\boldmath $v$}_0(s)$
is immediately obtained;
in particular, $F(0,t_0)=[t_0,0,0]^T$ and
$$
dF(0, t_0)=
\left[\begin{array}{cc}
0&1\\noindent_0&0\\ \kappa_1&0
\end{array}\right].
$$
This gives an alternative proof of Lemma \ref{striction_curve}:
$F$ is singular at $(0, t_0)$ if and only if $\kappa_1(0)=t_0=0$
($t_0=0$ means that the point is just the striction point $\sigma(0)$ lying on the ruling).
Assume that $F$ is singular at the origin. Then
we obtain a {\it canonical Taylor expansion} of $F$:
\begin{eqnarray}\label{eq1}
&&F(s,t)=
\textstyle
\left(t-\frac{1}{2}ts^2+\frac{\tau_1}{2} s^3,\
ts-\frac{\tau_1}{2}s^2-\frac{2\tau_0\kappa_1'+\tau_1'}{6}s^3,\
\frac{\kappa_1'}{2}s^2+\frac{\tau_0}{2}t s^2+\frac{\kappa_1''-2\tau_0\tau_1}{6}s^3\right)+o(3).
\end{eqnarray}
\begin{rem}\label{truncation}\upshape
{\bf (Truncated polynomial maps)}
Let $F(s,t)$ be as in (\ref{eq1}),
and set $\bar{F}(s, t)=(\bar{\mbox{\boldmath $v$}}_0(s) \times \bar{\mbox{\boldmath $v$}}_1(s)) + t \bar{\mbox{\boldmath $v$}}_0(s)$
to be a polynomial map of order $k$ with $j^k\bar{F}(0)=j^kF(0)$.
Denote by $\bar{s}$ the arc-length of the curve $\bar{\mbox{\boldmath $v$}}_0(s)$, then $\bar{s}:=s+o(k)$,
and thus $k$-jets at $0$ of the dual curvature and the dual torsion do not change
from those of $F$.
That gives examples of polynomial ruled surfaces
with prescribed $k$-jets of $\check{\kappa}$ and $\check{\tau}$ at a point.
\end{rem}
\subsection{Recognition of singularity types}
Our task is to find appropriate diffeomorphism-germs of the source and the target
for reducing jets of $F(s,t)$ to normal forms in $\mathscr{A}$-classification
step by step; for such computations, we have used the software {\it Mathematca}.
Let $(X, Y, Z)$ be the coordinates of the target $\mathbb{R}^3$.
Below, $\kappa_1, \kappa_1', \cdots$ denote
their values at $s=0$ unless specifically mentioned.
From now on,
assume that $\kappa_1(=\kappa_1(0))=0$.
Put $y=s$ and $x=t-\frac{1}{2}ts^2+\frac{\tau_1}{2} s^3+\cdots$,
the first component of $F$ in the form (\ref{eq1}) above.
With this new coordinates $(x, y)$ of the source $\mathbb{R}^2$, we set
\begin{eqnarray}\label{2}
f(x,y)&:=&F(y, t(x, y))=(x, f_2(x,y), f_3(x,y)) \\
&=&\textstyle
\left( x,\ x y-\frac{1 }{2}\tau_{1} y^2-\frac{1 }{6}\tau_{1}' y^3,
\ \frac{1}{2}\kappa_1' y^2+\frac{1}{2}\tau_{0} x y^2
+ \frac{1}{6 }(\kappa_{1}''-2 \tau_{0} \tau_{1})y^3\right)
+o(3).\nonumber
\end{eqnarray}
Note that $f(x,y)$ is still of the form $\tilde{\mbox{\boldmath $r$}}(y)+x\tilde{\mbox{\boldmath $e$}}(y)$.
Now, we apply to this germ $f(x,y)$ the recognition trees in Mond's classification \cite[Figs.1, 2]{Mond85}.
Below, $S_k^\pm$, $B_k^\pm$, $C_k^\pm$, $H_k$ and $F_4$ denote Mond's notations of
$\mathscr{A}$-simple germs \cite{Mond85}.
\
\noindent
$\bullet\;\; ${\bf $2$-jet}:
Crosscap $S_0$ is $2$-determined, thus it follows from (\ref{2}) that
$$f\sim_\mathscr{A} S_0: (x,xy, y^2) \;\;\; \Longleftrightarrow \;\;\;
\kappa_1=0, \quad
\kappa_1' \neq 0.$$
Let $\kappa_1'=0$. Then
the $2$-jet is equivalent to either of $(x, xy,0)$ or $(x, y^2,0)$,
according to whether $\tau_1= 0$ or not.
We compute the second and third component of $f$ as
\begin{eqnarray*}
f_2&=& \textstyle
x y -\frac{1}{2} \tau_1 y^2 - \frac{1}{6}\tau_1' y^3
+ \frac{1}{24} ((8 - 4\tau_0^2) x y^3 +( - 5\tau_1 +
3 \tau_0^2 \tau_1 - 3 \tau_0 \kappa_1'' - \tau_1'' )y^4 )
\\
&&\textstyle
+ \frac{1}{120} (-15 \tau_0 \tau_0' x y^4 +
(12 \tau_0 \tau_0'\tau_1 - 9 \tau_1' +
6 \tau_0^2 \tau_1' - 6 \tau_0' \kappa_1'' -
4 \tau_0 \kappa_1^{(3)} - \tau_1^{(3)}) y^5)
\\
&&\textstyle
+o(5), \\
f_3&=&
\textstyle
\frac{1}{6} (3 \tau_0 x y^2 +( \kappa_1''- 2 \tau_0 \tau_1)y^3) +
\frac{1}{24} (4 \tau_0' x y^3+( - 3\tau_1 \tau_0' -
3 \tau_0 \tau_1'+ \kappa_1^{(3)})y^4)\\
&&\textstyle
+ \frac{1}{120} ((25 \tau_0 - 5 \tau_0^3+5 \tau_0'')xy^4
+(- 16 \tau_0 \tau_1 + 4 \tau_0^3 \tau_1
- 6 \tau_0' tau_1'-
6 \tau_0^2 \kappa_1'' \\
&&\textstyle
- 4 y^5 \tau_1 \tau_0''
- 4 y^5 \tau_0 \tau_1'' +\kappa_1^{(4)})y^5)+o(5).
\end{eqnarray*}
\
\noindent
$\bullet\;\; ${\bf $3$-jet}:
Let $\kappa_1=\kappa_1'=0$ and $\tau_1\not= 0$.
First, let us remove the term $xy$ from $f_2$;
take $\bar{x}=x$ and $\bar{y}=y-\frac{1}{\tau_1}x$,
then we see that
\begin{equation}\textstyle
j^3f(0) \sim
(x, y^2+\frac{\tau_1'}{\tau_1^3} x^2y+\frac{\tau_1'}{3\tau_1} y^3,
\kappa_1''x^2y+\frac{1}{3}\tau_1^2(\kappa_1''-2\tau_0\tau_1)y^3).
\end{equation}
The first two components can be transformed to $(x,y^2)$
by a coordinate change of $(x,y)$ with identical linear part and by a target coordinate change of $(X,Y)$,
since the plane-to-plane germ $(x,y^2)$ is $2$-determined (stable germ).
Hence $j^3f(0)$ is equivalent to one of the following:
\begin{equation}\label{3-jet1}
\left\{
\begin{array}{lll}
(x, y^2, y^3\pm x^2 y)
& \kappa_1''( \kappa_1''-2 \tau_0\tau _1)\gtrless 0, \ \tau_1\neq 0
&\cdots S_1^\pm
\\(x, y^2, y^3)
& \kappa_1'' = 0,\ \tau_0\tau_1\neq 0
&\cdots S
\\(x, y^2, x^2 y)
& \kappa_1''\ =2 \tau_0\tau_1\neq 0,
&\cdots B
\\(x, y^2, 0)
& \kappa_1''=\tau_0=0, \ \tau_1\neq 0
&\cdots C
\end{array}
\right.
\end{equation}
Note that $S_1^\pm$ is $3$-determined, thus this case is clarified.
Let $\tau_1= 0$. Then from (\ref{2}), we have
$$\textstyle
j^3f(0)\sim
\left( x,\ x y-\frac{1 }{6}\tau_{1}' y^3,\frac{1}{2}\tau_{0} x y^2 + \frac{1}{6 }\kappa_{1}'' y^3\right).
$$
In the same way as above, $j^3f(0)$ is reduced to one of the following:
\begin{equation}\label{3-jet2}
\left\{
\begin{array}{lll}
(x, xy, y^3)&
\kappa_1''\neq 0,\ \tau_1= 0
&\cdots H
\\(x, xy+y^3, x y^2)&
\kappa_1'' =\tau_1= 0,\ \tau_0\tau_1'\neq 0
&\cdots P
\\(x, xy, xy^2)&
\kappa_1''=\tau_1= \tau_1'=0, \ \tau_0\neq 0
\\(x, xy+y^3, 0)&
\kappa_1''=\tau_0=\tau_1= 0,\ \tau_1'\neq 0
\\(x, xy, 0)&
\kappa_1''=\tau_0=\tau_1= \tau_1' =0
\end{array}
\right.
\end{equation}
Each of last three types has codimension $\ge 6$, so we omit them here.
Below,
for types $S, B, \cdots, P$ in (\ref{3-jet1}) and (\ref{3-jet2}),
we detect $\mathscr{A}$-types with codimension $\le 5$ by checking higher jets and the determinacy.
\
\noindent
$\bullet \;\; ${\bf $S$-type}:
Let $\kappa_1=\kappa_1' = 0$ and $\tau_0\tau_1\neq 0$.
Then a computation shows that
$$
\begin{array}{rcl}
\textstyle
\kappa_1''=0
&\textstyle
\quad\Longrightarrow\quad
&
j^4 f(0)
\sim
\left(x, y^2, y^3-\frac{\kappa_1^{(3)}}{2\tau_0\tau_1^4}x^3 y \right)
\\
\textstyle
\kappa_1''=\kappa_1^{(3)}=0
&\textstyle
\quad\Longrightarrow\quad
&
j^5 f(0)
\sim
\left(x, y^2, y^3-\frac{\kappa_1^{(4)}}{8\tau_0\tau_1^5}x^4 y
\right)
\\
\textstyle
\kappa_1''=\kappa_1^{(3)}=\kappa_1^{(4)}=0
&\textstyle
\quad\Longrightarrow\quad
&
j^6 f(0)
\sim
\left(x, y^2, y^3-\frac{\kappa_1^{(5)}}{40\tau_0\tau_1^6}x^5 y
\right).
\end{array}
$$
Note that $S_k$ is $(k+2)$-determined (its codimension is $k+2$), thus
$f$ is of type $S_k^\pm$ $(k=2,3,4)$
if and only if
$\kappa_1=\kappa_1'=\cdots=\kappa_1^{(k)}=0$
and
$\kappa_1^{(k+1)}\tau_0\tau_1 \lessgtr 0$ (seemingly, it is so for any $k$).
\
\noindent
$\bullet \;\; ${\bf $B$-type}:
Let $\kappa_1=\kappa_1'=\kappa_1''-2 \tau_0\tau_1=0$ and $ \kappa_1''\neq 0$.
Then it would be $\mathscr{A}$-equivalent to $B_k$-type \cite[4.1:17, Table 3]{Mond85}. For instance,
$$j^5f(0)\sim (x, y^2, x^2 y+b_2 y^5)$$
with
$$
\begin{array}{rl}
b_2=&48\tau_0^2 \tau_1^2(\tau_0^2-2)-20 (\tau_0^2 (\tau_1')^2+\tau_1^2 (\tau_0')^2)-56 \tau_0 \tau_1 \tau_0' \tau_1'-24\tau_0 \tau_1(\tau_0\tau_1''+\tau_1 \tau_0'')\\
&+20 \kappa_1^{(3)}(\tau_0 \tau_1'+\tau_1 \tau_0')-5 (\kappa_1^{(3)})^2+6 \kappa_1^{(4)} \tau_0 \tau_1.
\end{array}
$$
Since $B_2$ is $5$-determined,
$$f\sim_\mathscr{A} B_2^\pm:(x,y^2,x^2y\pm y^5)
\;\; \Longleftrightarrow \;\; b_2\gtrless 0.$$
Let $b_3$ be the coefficient of $y^7$ in the last component of $j^7f(0)$,
which is written as a polynomial in derivatives of invariants at $s=0$, then
$B_3^\pm: (x,y^2,x^2y\pm y^7)$ is detected by the condition that $b_2=0$ and $b_3\not=0$.
Here $B_3$ is of codimension $5$.
\
\noindent
$\bullet \;\; ${\bf $C$-type}:
Let $\kappa_1=\kappa_1'=\kappa_1''=\tau_0=0,\ \tau_1\neq 0$.
Through
$$\textstyle \psi(X,Y,Z)=\left(\frac{1}{\tau_1}X,\ Y,\
\frac{1}{\tau_1}\left(Z-a Y^2
-b X^2 Y\right)\right)$$
with $a=\frac{1}{4}(\kappa_1^{(3)}-3\tau_1\tau_0')$,
$b=\frac{3}{2\tau_1^2}(\kappa_1^{(3)}-\tau_1\tau_0')$,
we see that
$$\textstyle
j^4 f(0)\sim
\left(x,\ y^2,\
\kappa_1^{(3)}x^3 y
+(\kappa_1^{(3)}-2 \tau_1\tau_0')x y^3\right).$$
Since $C_3$ is $4$-determined (of codimension $5$),
$$f \sim_\mathscr{A} C_3^\pm: (x, y^2, x y^3\pm x^3 y)
\;\; \Longleftrightarrow \;\;
\kappa_1^{(3)}(\kappa_1^{(3)}-2\tau_1\tau_0')\gtrless 0.$$
\
\noindent
$\bullet \;\; ${\bf $H$-type}:
Let $\kappa_1=\kappa_1'=\tau_1= 0$ and $\kappa_1''\neq 0$.
Then it would be $\mathscr{A}$-equivalent to $H_k$-type \cite[4.2.1:2]{Mond85}.
A lengthy computation shows that
$$
j^5 f(0)\sim
\left(x,\
x y + h_2 y^5,\
y^3
\right)
$$
with
$$
\begin{array}{rl}
h_2=
&\hspace{-5pt}
-15 \tau_0^2 (\tau_1')^3
-24 \tau_0' (\tau_1')^2 \kappa_1''
-36 \tau_1' (\kappa_1'')^2
-15 \tau_0^2 \tau_1' (\kappa_1'')^2
-24 \tau_0' (\kappa_1'')^3
-21 \tau_0 \tau_1' \kappa_1'' \tau_1''
\\&+20 \tau_0 (\tau_1')^2 \kappa_1^{(3)}
- \tau_0 (\kappa_1'')^2 \kappa_1^{(3)}
+5 \kappa_1'' \tau_1'' \kappa_1^{(3)}
-5 \tau_1' (\kappa_1^{(3)})^2
-4 (\kappa_1'')^2 \tau_1^{(3)}
+4 \tau_1' \kappa_1'' \kappa_1^{(4)}.
\end{array}
$$
Since $H_2$ is $5$-determined,
$$f\sim_\mathscr{A} H_2^\pm:(x,xy\pm y^5, y^3)
\;\; \Longleftrightarrow \;\; h_2\gtrless 0.$$
Let $h_3$ be the coefficient of $y^8$ in the middle component of $j^8f(0)$,
then $H_3:(x,xy+ y^8, y^3)$ is detected by $h_2=0$ and $h_3\not=0$
($H_3$ is of codimension $5$).
\
\noindent
$\bullet \;\; ${\bf $P$-type}:
Let $\kappa_1=\kappa_1'=\kappa_1'' =\tau_1= 0$ and $\tau_0\tau_1'\neq 0$.
Then we see that there is a polynomial $p_4$ in derivatives of $\kappa_1, \tau_0, \tau_1$ so that
$$f\sim_\mathscr{A}
P_3: (x,\ x y + y^3,\ x y^2 +p_4 y^4)$$
for $p_4\not=0, \frac{1}{2}, 1, \frac{3}{2}$ \cite[\S 4.2]{Mond85}.
\begin{rem}\label{C_kF_4}\upshape
{\bf (Characterization of $C_k$ and $F_4$)}
Among $\mathscr{A}$-simple germs in Mond \cite{Mond85},
$S_k^\pm$, $B_k^\pm$ and $H_k$ have been discussed above,
so there remain $C_k\; (k\ge 4)$ and $F_4$.
Consider $C$-type above and think of the next to $C_3$-type. Namely,
suppose that $\kappa_1^{(3)}(\kappa_1^{(3)}-2\tau_1\tau_0')=0$.
$\bullet \;\; $
If $\kappa_1^{(3)}=0$ and $\tau_1\tau_0'\neq 0$, then $j^4 f(0)\sim (x,y^2,xy^3)$ and
$$
\begin{array}{rcl}
\textstyle
\kappa_1^{(3)}=0
&\textstyle
\quad\Longrightarrow\quad
&
j^5 f(0)
\sim
\left(x, y^2, x y^3-\frac{\kappa_1^{(4)}}{8\tau_0'\tau_1^4}x^4 y
\right)
\\
\textstyle
\kappa_1^{(3)}=\kappa_1^{(4)}=0
&\textstyle
\quad\Longrightarrow\quad
&
j^6 f(0)
\sim
\left(x, y^2, xy^3-\frac{\kappa_1^{(5)}}{40\tau_0'\tau_1^5}x^5 y
\right)
\end{array}
$$
Since $C_k^\pm: (x, y^2, xy^3\pm x^ky)$ is $(k+1)$-determined,
we see that
$f$ is of type $C_k^\pm$ $(k=4,5)$
if and only if
$\tau_0=\kappa_1=\kappa_1'=\cdots=\kappa_1^{(k-1)}=0$
and
$\kappa_1^{(k)}\tau_0'\tau_1\lessgtr 0$ (seemingly, it is so for any $k$).
$\bullet \;\; $
If $\kappa_1^{(3)}=2\tau_1\tau_0'\neq 0$,
we have $j^4f(0)\sim (x, y^2, x^3 y)$ and
$$f \sim_\mathscr{A} F_4: (x, y^2, x^3y+y^5)
\Longleftrightarrow
3\kappa_1^{(4)}-8 \tau_0' \tau_1'-12 \tau_1 \tau_0''\not=0.$$
All cases of sufficient jets of $\mathscr{A}$-simple germs have been covered.
\end{rem}
\begin{rem}\label{5jet}\upshape{\bf (Non-realizable jets)}
Let us continue the argument in Remark \ref{C_kF_4}.
If $\kappa_1^{(3)}=\tau_0'=0$, then $f$ should be of codimension $\ge 7$ and
a computation shows that
$$
j^5 f(0)\sim
\left(x, y^2, \kappa_1^{(4)} x^4 y
+(\kappa_1^{(4)}-4\tau_1\tau_0'' )y^5
+2\sqrt{5}(\kappa_1^{(4)}-2\tau_1\tau_0'') x^2 y^3\right).
$$
In particular,
if two of
three coefficients $\kappa_1^{(4)}$, $\kappa_1^{(4)}-4\tau_1\tau_0''$,
$\kappa_1^{(4)}-2\tau_1\tau_0''$ are zero,
then all are zero.
Thus, for instance, the following $5$-jets are not equivalent to
jets of any non-cylindrical ruled surface:
$$
(x, y^2, x^4 y),
\quad(x, y^2, x^2y^3),
\quad(x, y^2, y^5).
$$
The $5$-jet $(x, y^2, y^5)$ is obviously realizable by a cylinder,
while
the $5$-jets $(x, y^2, x^4 y)$ and $(x, y^2, x^2y^3)$ are not equivalent to jets of
{\it any} ruled surfaces, even if we drop the condition $\mbox{\boldmath $e$}'(0)\not=0$.
In fact, put $F=\mbox{\boldmath $r$}(s)+t \mbox{\boldmath $e$}(s)$ with $\mbox{\boldmath $r$}(s)\cdot \mbox{\boldmath $e$}(s)=0$ and
$\mbox{\boldmath $e$}(s)=(1, 0, 0) + o(s)$.
If $F$ is singular at $(s,t)=(0,0)$ and $\mbox{\boldmath $r$}(0)=0$, then $\mbox{\boldmath $r$}(s)=o(s)$.
It is easy to see that
$F\sim_\mathscr{A} f=(x, y^2h(x,y), y^3g(x,y))$
with some functions $h, g$ of the form $p(y)+xq(y)$, and
thus the $5$-jet of $F$ is never equivalent to those two jets mentioned above.
By the same reason,
the $\mathscr{A}^3$-orbit of the $3$-jet $(x, y^3, x^2y)$ is not realized
by jets of any ruled surfaces
(the $2$-jet $(x,0,0)$ never appears in non-cylindrical ruled surfaces as seen before, and
the $3$-jet is not realizable by ruled surfaces with $\mbox{\boldmath $e$}'(0)=0$, that is seen in the same way as above).
\end{rem}
\subsection{Transversality} \label{mapping_space}
To precisely state {\it genericity} of ruled surfaces,
we need an appropriate mapping space (moduli space) equipped with a certain topology.
By the definition,
a {\it residual} subset of a mapping space is a union of countably many open dense subsets.
When maps having a prescribed condition form a residual subset,
we often say that such a map is {\it generic}, abusing words.
Let $I$ be an open interval containing $0 \in \mathbb{R}$ and
let $u$ denote the coordinate of $I$.
As the mapping space of non-cylindrical ruled surfaces, we take
$$\mathcal{R}:=\{\, \check{\bv}=\mbox{\boldmath $v$}_0+\varepsilon\mbox{\boldmath $v$}_1 \in C^\infty(I, \check{\mathbb{U}}) \, | \,
\mbox{\boldmath $v$}_0'(u)\not=0\, (u \in I) \, \}$$
equipped with Whitney $C^\infty$ topology
(As a remark, Izumiya and Takeuchi \cite{IT} and Martins and Nu\~no-Ballesteros \cite{MN}
took the space $C^\infty(I, \mathbb{R}^3\times S^2)$ instead of $C^\infty(I, \check{\mathbb{U}})$,
but the difference does not affect the matter of genericity arguments --
given a pair $(\mbox{\boldmath $r$}, \mbox{\boldmath $e$})$ of base and director curves, we simply assign a curve $\check{\bv}: I \to \check{\mathbb{U}}$ with $\mbox{\boldmath $v$}_0=\mbox{\boldmath $e$}$ and $\mbox{\boldmath $v$}_1=\mbox{\boldmath $r$}\times \mbox{\boldmath $e$}$).
Also we put
$$\mathcal{M}:=C^\infty(I, \mathbb{R}_{>0}\times \mathbb{R}^3)$$
of quadruples $(\kappa_0, \kappa_1, \tau_0, \tau_1)$ of real-valued functions with
$\kappa_0(u)>0$ equipped with Whitney $C^\infty$ topology.
Any curve $\check{\bv}(u)$ in $\mathcal{R}$ defines $\mathbb{D}$-valued functions,
$\check{\kappa}(u)$ and $\check{\tau}(u)$ (parameterized by a general parameter $u \in I$),
that produces a continuous map $\Phi: \mathcal{R} \to \mathcal{M}$.
Obviously, $\Phi$ is surjective. In fact, given
a quadruple of functions $(\kappa_0(u), \kappa_1(u), \tau_0(u), \tau_1(u)) \in \mathcal{M}$,
put a new parameter $s:=s(u)=\int_0^u \kappa_0(u) du$
and define $\kappa_1(s):=\kappa_1(u(s))$, etc.
Then, three functions $\kappa_1(s), \tau_0(s), \tau_1(s)$
determines, up to Euclidean motions, the curve $\check{\bv}(s)=\mbox{\boldmath $v$}_0(s)+\varepsilon \mbox{\boldmath $v$}_1(s)$
by solving the ordinary differential equation
determined by the Frenet formula.
The ambiguity is
fixed by the initial values $\check{\bv}(0), \check{\bn}(0), \check{\bt}(0)$,
which corresponds to the initial orthogonal axes in $\mathbb{R}^3$
at $u=0$.
Put $\check{\bv}(u):=\check{\bv}(s(u)) \in \mathcal{R}$; the set of such cruves
is exactly the preimage via $\Phi$ of the given quadruple of functions.
That infers that for a dense subset $O \subset \mathcal{M}$,
the preimage $\Phi^{-1}(O)$ is also dense in $\mathcal{R}$.
The above construction is extended for a parametric version.
Let $W$ be an open subset of $\mathbb{R}^p\; (0\le p \le 3)$,
and consider the subspace
$\mathcal{R}_W$ of $C^\infty (I\times W, \check{\mathbb{U}})$
which consists of maps $\check{\bv}(u, \lambda)=\mbox{\boldmath $v$}_0(u, \lambda)+\varepsilon \mbox{\boldmath $v$}_1(u, \lambda)$
with parameter $\lambda \in W$
satisfying $\partial \mbox{\boldmath $v$}_0/\partial u\not=0$ at any $(u, \lambda)$.
Put $\mathcal{M}_W$ to be
the mapping space of $I \times W \to \mathbb{R}_{>0}\times \mathbb{R}^3$,
and then
a surjective continuous map $\Phi:\mathcal{R}_W \to \mathcal{M}_W$
is defined in entirely the same way as above.
For a dense subset $O \subset \mathcal{M}_W$,
the preimage $\Phi^{-1}(O)$ is also in $\mathcal{R}_W$.
As seen in the previous section, we have obtained
a semi-algebraic stratification of the jet space $J^r:=\mathbb{R}^3 \times J^r(1,3)$
up to codimension $4$ ($r$ sufficiently large).
In fact,
any strata are defined by the conditions in Table \ref{table1}
of (in)equalities in Taylor coefficients $\{\kappa_1^{(k)}, \tau_0^{(k)}, \tau_1^{(k)}\}_{0 \le k\le r}$,
which form a system of coordinates of the affine space $J^r$.
Notice that
these Taylor coefficients are with respect to the arclength parameter $s$.
For each quadruple $(\kappa_0, \kappa_1, \tau_0, \tau_1) \in \mathcal{M}_W$,
we put
$$s=s(u, \lambda):=\int_0^u \kappa_0(u, \lambda) du, \qquad
\varphi(u, \lambda)=(\kappa_1(u,\lambda), \tau_0(u,\lambda), \tau_1(u,\lambda)).$$
By the assumption that $\partial s/\partial u=\kappa_0>0$,
let $\bar{\varphi}(s, \lambda):=\varphi(u(s, \lambda), \lambda)$.
Then we define
$$\Psi: I \times W \times \mathcal{M}_W \to J^r,
\qquad \Psi(u, \lambda, (\kappa_0, \varphi)) :=
j^r_s\bar{\varphi}(s(u, \lambda), \lambda),$$
where $j^r_s\bar{\varphi}$ means the $r$-jet respect to the parameter $s$.
By a version of Thom's transversality theorem (Lemma 4.6 in \cite{GG}),
there is a dense subset $O$ of $\mathcal{M}_W$
so that for any $\varphi \in O$,
the jet extension $\Psi_{\kappa_0, \varphi}: I \times W \to J^r$
is transverse to every stratum of our stratification of $J^r$.
Hence, $\Phi^{-1}(O)$ is dense in $\mathcal{R}_W$,
and for any element of $\Phi^{-1}(O)$,
only $\mathscr{A}$-singularity types listed in Table \ref{table1} appears.
This completes the proof of (2) in Theorem \ref{thm1}.
\hfill $\Box$
\begin{rem}\label{versal}\upshape
{\bf ($\mathscr{A}_e$-versal deformations)}
For each type in Table \ref{table1},
an $\mathscr{A}_e$-versal deformation of the germ is realized
by a generic family of non-cylindrical ruled surfaces.
This is directly checked by computations.
For instance, as in Table \ref{table1},
the $S_1^\pm$-singularity of ruled surface at $s=0$ is characterized by
$\kappa_1(0)=\kappa_1'(0)=0$,
$\kappa_1''(0)\not=0, 2\tau_0(0)\tau_1(0)$ and $\tau_1(0)\not=0$.
Suppose that $\varphi=(\kappa_1(s), \tau_0(s), \tau_1(s)): I \to \mathbb{R}^3$ satisfies this condition.
Define a $1$-parameter family $I \times \mathbb{R} \to \mathbb{R}^3$
by $\varphi(s,\lambda) :=\varphi(s)+(\lambda, 0, 0)$, then
obviously, its $1$-jet extension
$j^1_s\varphi$ is transverse at $(0,0)$ to
the stratum defined by $\kappa_1=\kappa_1'=0$ in $J^1=\mathbb{R}^3\times J^1(1,3)$.
This family yields a $1$-parameter family
$F(s, t, \lambda)=(t, ts - \frac{\tau_1}{2}s^2, \lambda s)+o(2)$
of ruled surfaces.
By using a coordinate change of $x=t+\cdots$
(= first component of $F(s, t, \lambda)$) and $y=s$ and some target changes,
we see that
the germ of $F(s, t, \lambda)$ is equivalent to $(x, y^2, y^3\pm x^2y+\lambda y)$,
which is an $\mathscr{A}_e$-miniversal deformation of $S_1^\pm$-singularity.
\end{rem}
\section{Singularities of developable surfaces}
\subsection{Recognition of singularity types}
For non-cylindrical developable surfaces, $\kappa_1(s)\equiv 0$ identically. Hence
the Taylor expansion of $f$ is (\ref{2}) with $\kappa_1^{(k)}= 0$ for all $k$:
$$\textstyle
f(x,y):=F(y,t(x,y))=
(x, x y - \frac{1}{2}\tau_1 y^2 - \frac{1}{6}\tau_1' y^3,
\frac{1}{2}\tau_0 x y^2 + \frac{1}{3} \tau_0\tau_1 y^3)+o(3).$$
Using the $\mathscr{A}$-criteria mentioned in \S 2,
we classify singularities arising in generic families of developable surfaces.
Notice that there are two different aspects;
singularities of frontal surfaces correspond to the case of $\tau_1\not=0$,
while singularities of wavefronts correspond to the case of $\tau_1=0$.
Below we prove Theorem \ref{thm2}.
\
\noindent
$\bullet\;\; $
{\bf Case of $\tau_1\not=0$}:
By $s = y + \tau_1^{-1}x$ and some linear change of the target,
we have $f=(x, y^2+o(2), f_3(x,y))$ with $f_3=\tau_0y^3+o(3)$.
Note that $(x, y^2)$ is $2$-determined and that
each term $x^ky^{2l}$ in $f_3$ can be removed by a coordinate change of the target
$(X, Y, Z) \mapsto (X, Y, Z-X^kY^l)$.
Use Proposition \ref{CE} in \S 2 (\cite{Mond89})
for determinacy in CE.
\begin{itemize}
\item[(i)]
If $\tau_0\not=0$, then $f\sim_\mathscr{A}(x,y^2,y^3)$, since it is $3$-determined in CE.
\item[(ii)]
Let $\tau_0=0$.
Computing the $4$-jet, we see
$$f_3=\tau_0'(6x^2y^2+8\tau_1xy^3+3\tau_1^2y^4)+o(4).$$
If $\tau_0'\not=0$, then
$f\sim_\mathscr{A} (x, y^2, x y^3)$, for
the germ is $4$-determined in CE.
Hence $f$ is of type cuspidal crosscap.
\item[(iii)]
Let $\tau_0=\tau_0'=0$.
Computing the $5$-jet, we see
$$f_3=
\tau_0''(10x^3y^2+20\tau_1x^2y^3+15\tau_1^2xy^4+4\tau_1^3y^5)+o(5).$$
If $\tau_0''\not=0$, by target changes using $X=x$ and $Y=y^2$,
terms $x^3y^2$ and $xy^4$ can be removed from $Z=f_3$,
thus we see that
$f\sim_\mathscr{A} (x, y^2, y^3(x^2+y^2))$,
for this germ is $5$-determined in CE.
That is cuspidal $S_1^+$-type. Note that cuspidal $S_1^-$ never appears.
\item[(iv)]
Let $\tau_0=\tau_0'=\tau_0''=0$.
Computing the $6$-jet, we see
$$f_3=\tau_0'''(15x^4y^2+40\tau_1x^3y^3
+45\tau_1^2x^2y^4+24\tau_1^3xy^5+5\tau_1^4y^6)+o(6).$$
If $\tau_0'''\not=0$, then
$f\sim_\mathscr{A} (x, y^2, y^3(x^3+xy^2))$,
for the germ is $6$-determined in CE.
That is cuspidal $C_3^+$-type, while cuspidal $C_3^-$ does not appear.
Note that $\tau_0=\tau_0'=\tau_0''=0$ if and only if the $5$-jet of $f$ is equivalent to $(x, y^2,0)$,
thus cuspidal $S$ and $B$-types never appear, as mentioned in Remark \ref{rem1}.
\end{itemize}
\
\noindent
$\bullet\;\; $
{\bf Case of $\tau_1=0$}:
Then
$\textstyle
f=(x, x y - \frac{1}{6}\tau_1'y^3, \frac{1}{2}\tau_0 x y^2) + o(3)$.
Note that $j^2f(0) \sim (x, xy,0)$, thus
types $A_3^\pm$ and $D_k$ never appear (Remark \ref{rem1}).
If $\tau_0=0$,
$j^3f(0)$ is equivalent to either $(x, xy+y^3,0)$ or $(x, xy,0)$,
that is of type $T_1$ or $T_2$ (codimension $3, 4$) in Table \ref{table2}.
Now assume that $\tau_0\not=0$.
Write
$$\textstyle
f=(x, f_2(x,y), f_3(x,y))=(x, x y - \frac{1}{6}\tau_1'y^3, x y^2)+o(3).$$
The singular point set $S(F)$ is defined by $(f_2)_y=(f_3)_y=0$,
and through a computation, it is simplified as $\lambda=0$ with
$$\textstyle
\lambda=x-\frac{1}{2}\tau_1'y^2 - \frac{1}{6}\tau_1''y^3
-\frac{1}{24}(\tau_1'''-3\tau_1')y^4+o(4).$$
We may take $\eta={\partial}/{\partial y}$ as a vector field which generates $\ker dF$ along $S(F)$.
Then, $\eta \lambda(0)=0$,
$\eta\eta \lambda(0) = -\tau_1'$,
$\eta\eta\eta \lambda(0) =- \tau_1''$ and
$\eta\eta\eta\eta \lambda(0) =-(\tau_1'''-3\tau_1')$.
Hence, by Izumiya-Saji's criteria in \S 2.5,
we have the conditions for detecting $Sw$, $cA_4$ and $cA_5$.
\subsection{Topological classification}
We prove Theorem \ref{thm3}.
Let $\sigma(s)$ be the striction curve of a non-cylindrical developable surface.
Assume that $\sigma(0)=0 \in \mathbb{R}^3$, and consider the germ $\sigma:(\mathbb{R},0)\to (\mathbb{R}^3,0)$.
Since $\{\mbox{\boldmath $v$}_0(s), \mbox{\boldmath $n$}_0(s), \mbox{\boldmath $t$}_0(s)\}$ form a basis of $\mathbb{R}^3$
for each $s$, we denote the $k$-th derivative by
$$
\sigma^{(k)}(s)
=
A_k(s)\mbox{\boldmath $v$}_0(s)
+B_k(s)\mbox{\boldmath $n$}_0(s)
+C_k(s)\mbox{\boldmath $t$}_0(s)
\quad(k\geq 1)
$$
where $A_k(s), B_k(s), C_k(s)$ are some functions.
Then, with respect to the basis $\{\mbox{\boldmath $v$}_0(0), \mbox{\boldmath $n$}_0(0), \mbox{\boldmath $t$}_0(0)\}$,
the expansion of $\sigma$ at $s=0$
is given by
$$\textstyle
\sigma(s)=(A_1(0)s+\frac{1}{2}A_2(0)s^2+\cdots,
B_1(0)s+\frac{1}{2}B_2(0)s^2+\cdots,
C_1(0)s+\frac{1}{2}C_2(0)s^2+\cdots).$$
Now assume that $\sigma$ is of type $(m, n_1, n_2)$, i.e.,
$$
\begin{cases}
A_1(0)=\cdots=A_{m-1}(0)=0,\ A_m(0)\neq 0,
&\\
B_1(0)=\cdots=B_{n_1-1}(0)=0,\ B_{n_1}(0)\neq 0,
&\\
C_1(0)=\cdots=C_{n_2-1}(0)=0,\ C_{n_2}(0)\neq 0.
&\end{cases}
$$
Since $\sigma'(s)=\tau_1(s)\mbox{\boldmath $v$}_0(s)$ for a developable surface (Lemma \ref{striction_curve0} (iii)),
we see that $A_1(s)=\tau_1(s)$ and $B_1(s)\equiv C_1(s)\equiv 0$.
By the Frenet formula (Theorem \ref{frenet} (1)),
\begin{align*}
\sigma^{(k+1)}
&=(\sigma^{(k)})'
=\{A_{k}\mbox{\boldmath $v$}_0
+B_{k}\mbox{\boldmath $n$}_0
+C_{k}\mbox{\boldmath $t$}_0\}'\\
&=
(A_{k}'-B_k)\mbox{\boldmath $v$}_0
+(B_{k}'+A_k- C_k \tau_0)\mbox{\boldmath $n$}_0
+(C_{k}' +B_k \tau_0)\mbox{\boldmath $t$}_0 \\
&=A_{k+1}\mbox{\boldmath $v$}_0+B_{k+1} \mbox{\boldmath $n$}_0 + C_{k+1} \mbox{\boldmath $t$}_0.
\end{align*}
Thus for $k=1$, we have
$A_2(s)=\tau_1'(s)$, $B_2(s)=\tau_1(s)$, $C_2(s)\equiv 0$, and
for $k=2$,
$A_3(s)=\tau_1''(s)-\tau_1(s)$,
$B_3(s)=2\tau_1'(s)$ and $C_3(s)=\tau_0(s)\tau_1(s)$.
For $k \ge 3$,
there are some smooth functions
$a_{k,*}(s)$, $b_{k,*}(s)$, $c_{k,*,*}(s)$ and
positive numbers
$\beta_k, \gamma_{k,0}, \cdots, \gamma_{k, k-3}>0$ such that
\begin{align*}
A_{k}(s)
&=
a_{k,0}(s)\tau_1(s)
+\cdots
+a_{k,k-2}(s)\tau_1^{(k-2)}(s)
+
\tau_1^{(k-1)}(s),
\\
B_{k}(s)
&=
b_{k,0}(s)\tau_1(s)
+\cdots
+b_{k,k-3}(s)\tau_1^{(k-3)}(s)
+\beta_k\tau_1^{(k-2)}(s),
\\
C_{k}(s)
&=
\{c_{k,0,0}(s)\tau_0(s)+\cdots
+\gamma_{k,0}\tau_0^{(k-4)}(s)\}\tau_1(s)
\\
&\quad
+\{c_{k,1,0}(s)\tau_0(s)+\cdots
+\gamma_{k,1}\tau_0^{(k-5)}(s)\}
\tau_1'(s)
+\cdots
\\
&\quad
+\{c_{k,k-4,0}(s)\tau_0(s)+\gamma_{k,k-4}\tau_0'(s)\}
\tau_1^{(k-4)}(s)
+\gamma_{k,k-3}\tau_0(s)
\tau_1^{(k-3)}(s).
\end{align*}
Hence, by the assumption on $A_k(0)$, we have
$$\tau_1(0)=\cdots=\tau_1^{(m-2)}(0)=0,\quad
\tau_1^{(m-1)}(0)\neq 0,$$
and thus
$$B_1(0)=\cdots=B_m(0)=0, \;\; B_{m+1}(0)\neq 0,
\quad C_1(0)=\cdots=C_{m+2}(0)=0.$$
In particular,
$$n_1=m+1, \;\; n_2=m+1+r\quad (r\geq 1).$$
By the above formula of $C_k(s)$ with $k=m+1+r$,
we see
$$
\tau_0(0)=\cdots=\tau_0^{(r-2)}(0)=0,
\quad
\tau_0^{(r-1)}(0)\neq 0.
$$
Conversely,
if the order of $\tau_0$ and $\tau_1$
are $r$ and $m-1$, respectively,
then the type of $\sigma$ is $(m, m+1, m+1+r)$.
This completes the proof. \hfill $\Box$
|
\section{Introduction\label{sec:Introduction}}
\noindent Inverse spectral problems on compact Riemannian manifolds ask to what extent geometric and topological data are encoded in the spectra of natural operators. There is an extremely rich literature of both positive and negative results in the case of the Laplace-Beltrami operator on compact manifolds, with Dirichlet or Neumann boundary conditions (or mixed conditions) imposed when the boundary is nonempty. The literature for other natural operators lags behind. The goal of this article is to show that most of the negative results for the Laplace-Beltrami operator in the literature, i.e., the constructions of manifolds whose Laplace-Beltrami operators are isospectral, are equally valid for other natural operators. We were motivated primarily by the surge of interest in Steklov eigenvalue problems and the related ``sloshing problem'' on compact Riemannian manifolds with boundary, so we will focus primarily on these problems. However, we will also comment on other eigenvalue problems.
\subsection{Steklov eigenvalue problems.}\label{subsec.stek} Let $(M,g)$ be a compact smooth Riemannian manifold with boundary, and let $\Delta$ be the associated Laplace-Beltrami operator.
For~$\alpha\in\mathbb{R}$ which is not in the spectrum of the Dirichlet Laplacian and for $\rho\in C^\infty(\partial M)$, the \emph{Steklov spectrum} of $M$ at frequency $\alpha$ with boundary density $\rho$, denoted by $\operatorname{Stek}_{\alpha}(M,g,\rho)$ or simply by $\operatorname{Stek}_\alpha(M,g)$ if $\rho\equiv 1$, is the collection of real numbers $\sigma$ for which there exists a nontrivial solution $u\in C^\infty(M)$ to the eigenvalue problem
\begin{equation}
\begin{cases}\Delta u=\alpha\,u\text{ on }M\smallsetminus\boundary{M}\\
\quad\normalDerivative u=\sigma\,\rho u\text{ on }\partial M,\label{eq:LB_eigenvalue_problem}
\end{cases}
\end{equation}
where $\normalDerivative u$ is the normal derivative of $u$ on the boundary.
(The problem is well-defined, since it was required that $\alpha$ not be a Dirichlet eigenvalue of $\Delta$.) In two dimensions, $\operatorname{Stek}_0(M,g,\rho)$ corresponds to the collection of squares of eigenfrequencies of a drum
all of whose mass is distributed along the boundary according to the density $\rho$ (see \cite{LambertiProvenzano}). When $\rho\equiv 1$, the Steklov spectrum $\operatorname{Stek}_\alpha(M,g)$ is precisely the eigenvalue spectrum of the \emph{Dirichlet-to-Neumann operator} $\mathcal{D}_\alpha^{M,g}\colon C^\infty(\partial M){}\to C^\infty(\partial M){}$. This operator associates to a function $v\in C^\infty(\partial M)$ the normal derivative of the unique extension $V:M\to\mathbb{R}$ of $v$ to $M$ that satisfies $\Delta V=\alpha V$. In particular, when $\alpha=0$, the extension $V$ is harmonic, so is just the solution of the Dirichlet problem with initial data $v$.
We remark that if the boundary density function $\rho$ is merely $L^{\infty}$, then \eqref{eq:LB_eigenvalue_problem} is still a well-defined eigenvalue problem, although the eigenfunctions are merely $H^1$ rather than smooth, and the boundary condition in \eqref{eq:LB_eigenvalue_problem} is interpreted in the sense of the Sobolev trace.
The Steklov spectrum was first introduced by A. Steklov in 1902 and has since found many remarkable applications; see the historical article \cite{Notices}. For example, by examining the dependence of $\operatorname{Stek}_\alpha(M,g)$ on the parameter $\alpha$, Friedlander~\cite{Friedlander1991} derived an inequality between the Neumann and
Dirichlet eigenvalues of bounded $C^{1}$-domains in~$\mathbb{R}^{n}$;
this inequality was extended to Lipschitz domains by Arendt and Mazzeo~\cite{ArendtMazzeo2012}. The study of the Steklov spectrum has recently gained impetus; see, for example, \cite{Brock2001, CianciGir, ColGitGir, ColboisElGirouard2011, FraserSchoen2016, GirouardPolterovich2012, GirouardParnovskiPolterovichSher2014, Jammes2014, PolterovichSher2015, Karp2017, YY}, and the excellent survey \cite{GP}. E.g., $\operatorname{Stek}_0(M,g)$ is known to determine the dimension and volume of~$\boundary{M}$, the geometry of $\boundary{M}$ if $\dim(M)=2$~\cite{GirouardParnovskiPolterovichSher2014}, whether a domain in $\mathbb{R}^2$ is a disk ~\cite{GirouardParnovskiPolterovichSher2014}, and whether a domain in $\mathbb{R}^3$ with connected boundary is a ball~\cite{PolterovichSher2015}.
The so-called \emph{sloshing problem}, describing oscillations of a fluid in an open container, is the special case of the Steklov problem~(\ref{eq:LB_eigenvalue_problem}) in which $\rho$ takes on only the values 0 and 1: $\rho\equiv 0$ on the walls of the container and $\rho\equiv 1$ on the free surface of the fluid.
In dimension two, the Steklov spectrum $\operatorname{Stek}_0(M,g,\rho)$ is invariant under conformal changes of metric away from the boundary; i.e., if $g'=e^fg$ with $f\equiv 0$ on $\partial M$, then $\operatorname{Stek}_0(M,g,\rho)=\operatorname{Stek}_0(M,g',\rho)$. In fact, we even have $\mathcal{D}_0^{(M,g')}=\mathcal{D}_0^{(M,g')}$. (This is immediate from the fact that the Laplacian of $g'$ is related to that of $g$ by $\Delta'=e^{-f}\Delta$ in dimension two. In higher dimensions, this equality fails.) We will say that
$(M,g,\rho)$ and $(M',g',\rho')$ are \emph{trivially Steklov isospectral} for $\alpha=0$ if there exists a diffeomorphism $F$ from $M$ to $M'$ intertwining $\rho$ and $\rho'$ such that either (i) $F:(M,g)\to (M',g')$ is an isometry or (ii) $\dim(M)=2$ and $F^*g'=e^f g$ with $f\vert_{\partial M}=0$. We caution that such conformal changes of metric will in general affect $\operatorname{Stek}_\alpha(M,g,\rho)$ for $\alpha\neq 0$, even in dimension two.
In this article we adapt to the Steklov setting the two primary techniques for constructing Laplace isospectral manifolds: Sunada's technique \cite{Sunada1985} and the torus action method (see, e.g., \cite{Gordon1994,Gordon2001,Schueth2001a,Schueth2001}). Both techniques yield pairs of Riemannian manifolds $M_1$ and $M_2$ with boundary that are simultaneously Dirichet and Neumann isospectral and that also satisfy $\operatorname{Stek}_\alpha(M_1,g_1)=\operatorname{Stek}_\alpha(M_2,g_2)$ for all $\alpha$ not in the Dirichlet spectrum. Moreover, $\operatorname{Stek}_\alpha(M_1,g_1,\rho_1)=\operatorname{Stek}_\alpha(M_2,g_2,\rho_2)$ for a large family of pairs of densities $(\rho_1,\rho_2)$. The Laplace-Beltrami operators on the boundaries are also isospectral. (In some, but not all cases, the boundaries are isometric.)
We illustrate these techniques with nontrivial examples:
\begin{itemize}
\item Pairs of (nonplanar) flat Steklov isospectral surfaces embedded in $\mathbb{R}^3$ constructed via the Sunada method;
\item Continuous families of mutually Steklov isospectral nonflat metrics on a ball in $\mathbb{R}^n$ constructed by the torus action method.
\end{itemize}
Specializing to the sloshing problem, we obtain, for example,
\begin{itemize}
\item Pairs of planar domains that are isospectral for the sloshing problem.
\end{itemize}
Referencing our results, the article \cite{ADGHRS} gives examples of Steklov isospectral orbifolds using the Sunada and torus action techniques. Example 6.1 in the same article uses direct computation to give examples of orbifold quotients $\Gamma_1\backslash B$ and $\Gamma_2\backslash B$ of Euclidean balls with $\operatorname{Stek}_0(\Gamma_1\backslash B)=\operatorname{Stek}_0(\Gamma_2\backslash B)$. Lemma 6.1 of ~\cite{ColboisElGirouard2011} establishes that cylinders over Laplace-Beltrami isospectral closed manifolds have the same Steklov spectrum, again with $\alpha=0$. To our knowledge, these examples exhaust the nontrivial examples of Steklov isospectral manifolds in the literature.
There are various notions of Dirichlet-to-Neumann operator acting on the space of $p$-forms on the boundary of a manifold. The definitions in \cite{RS} and \cite{Karp} (the latter being a modification of a definition in \cite{BelSh}) give operators with discrete spectrum. The Sunada method goes through for these Steklov spectra on $p$-forms. However, the torus action method does not. (This is not unexpected: the torus action method for the Laplace-Beltrami operator produces manifolds that are isospectral on functions, but it does not establish isospectrality for the Hodge Laplacian on $p$-forms.)
\subsection{Robin eigenvalue problems}\label{subsec.rob}
The Robin boundary value problem is dual to the Steklov eigenvalue problem in the following sense: Set $\rho\equiv 1$. Fixing a given $\sigma\in\mathbb{R}$ and interpreting (\ref{eq:LB_eigenvalue_problem}) as an eigenvalue problem for an unknown $\alpha$ converts (\ref{eq:LB_eigenvalue_problem})
into an eigenvalue problem with Robin boundary conditions. Since the Steklov isospectral manifolds that we construct satisfy $\operatorname{Stek}_\alpha(M_1,g_1)=\operatorname{Stek}_\alpha(M_2,g_2)$ for \emph{every} allowable choice of the parameter $\alpha$, they will also be isospectral for the Robin boundary value problem for every choice of the Robin parameter $\sigma$. See \cite{ArendtMazzeo2012} for historical comments on this relationship between the Steklov and Robin problems.
The Sunada and torus action methods work equally well for the mixed Robin--Neumann--Dirichlet eigenvalue problem. This problem asks for which $\alpha\in\mathbb{R}$
there exists $u\in C^\infty({M}{})$, with normal derivative
$\normalDerivative u\in C^\infty({\boundary{M}}{})$,
such that
\begin{equation}
\Delta u=\alpha\,u\text{ on }M\smallsetminus\boundary{M},\quad u=0\text{ on }D,\quad\normalDerivative u=0\text{ on }N,\quad\text{and}\quad\normalDerivative u=\sigma\,u\text{ on }S.\label{eq:LB_eigenvalue_problem_with_RDN_boundary_conditions}
\end{equation} where $\partial M=S\sqcup N\sqcup D$ (set-theoretic disjoint union) and where $\sigma$ is again a fixed Robin parameter. In case $D=\varnothing$, then the mixed Robin-Neumann problem is dual in the sense above to the Steklov problem with boundary density $\rho\equiv 1$ on $S$ and $\rho\equiv 0$ on $N$.
\subsection{Other eigenvalue problems.} Both the Sunada method and the torus action method are very robust. We remark without proof that both methods easily extend, for example, to poly-Laplacians
$\Delta^{m}$ with Dirichlet boundary conditions $u=\normalDerivative u=\normalDerivative{^{2}u}=\ldots=\normalDerivative{^{m-1}u}=0$
on $\boundary{M}$, as in the clamped plate problem where
$m=2$.
\bigskip
The paper is organized as follows: In Sections~\ref{sec:The_Sunada_method}
and~\ref{sec:The_torus_action_method}, we adapt the Sunada method and the torus action method, respectively, to the Steklov settings. Examples constructed via the two methods are given in Section~\ref{sec:Examples}. Finally, in Section~\ref{sec:density}, we construct Steklov isospectral boundary density functions: more precisely, we adapt both the Sunada method and the torus action method using an idea introduced by R. Brooks in order to construct pairs of boundary density functions $\rho_1$ and $\rho_2$ on a compact Riemannian manifold $M$ with boundary such that $\operatorname{Stek}_\alpha(M,\rho_1)=\operatorname{Stek}_\alpha(M,\rho_2)$ for all $\alpha$ not in the Dirichlet spectrum of $M$.
\section*{Acknowledgements} We thank Dorothee Schueth for suggesting Proposition~\ref{prop:Boundary_volume_form_preservation} and its proof, and we thank Leonid Friedlander and Rafe Mazzeo for informative conversations.
\section{The Sunada method\label{sec:The_Sunada_method}}
\noindent We adapt the Sunada method~\cite{Sunada1985} to the context
of the Steklov spectra.
\begin{defn}
\label{def:Gassmann_equivalence} Let $G$ be a finite group.
Two subgroups $H$ and $\mate{H}$ of $G$ are
called \emph{almost conjugate} or \emph{Gassmann equivalent}, if every $g\inG$ has equally many conjugates in $H$ and~$\mate{H}$.\end{defn}
\begin{remark}\label{rem.rep} Gassmann used such almost conjugate subgroups of a finite group to exhibit examples of pairs of nonisomorphic algebraic number fields with the same arithmetic (i.e., the same Dedekind zeta function). The formula for the character of an induced representation shows easily that $H$ and $H'$ are almost conjugate if and only if the representations of $G$ induced from the trivial one-dimensional representations of $H$ and $H'$ are equivalent: i.e., $\inducedRep{H}{G}{\boldsymbol{1}_{H}}\cong \inducedRep{\mate{H}}{G}{\boldsymbol{1}_{\mate{H}}}$, where $\boldsymbol{1}_{H}$ and $\boldsymbol{1}_{\mate{H}}$ denote the trivial one-dimensional representations of $H$ and $\mate{H}$, respectively.
\end{remark}
\begin{thm}[Sunada's Theorem adapted to the Steklov setting]
\label{thm:Sunada}Let $H$ and $\mate{H}$ be almost
conjugate subgroups of a finite group $G$. Assume that $G$ acts by isometries
on a compact Riemannian manifold $M$ with
boundary and that the restriction of the action to the subgroups $H$ and $H'$ is free. Let $\rho$ be an $L^\infty$, nonnegative, $G$-invariant function on $\partial M$. Continue to denote by $g$ and $\rho$ the Riemannian metric and the function induced on each of the orbit spaces $H\backslashM$ and $\mate{H}\backslashM$ by $g$ and $\rho$. Then
$$\operatorname{Stek}_\alpha(H\backslashM,g,\rho) =\operatorname{Stek}_\alpha(\mate{H}\backslashM,g,\rho)$$
for all $\alpha$ not in the Dirichlet spectrum of $H\backslashM$ and $\mate{H}\backslashM$. (Sunada's original theorem guarantees that the two quotient manifolds are both Dirichlet and Neumann isospectral, so the allowable choices of $\alpha$ are the same in both cases.)
\end{thm}
\begin{proof} Fix $\alpha$ and $\rho$ as in the theorem. We will abuse language and refer to solutions $u$ of Equation~(\ref{eq:LB_eigenvalue_problem}) in the Introduction as \emph{$\sigma$-eigenfunctions} for the $(\alpha,\rho)$-Steklov problem on $M$.
In what follows, if $\Gamma$ is any group acting linearly on a vector space $V$, we denote by $V^{\Gamma}$ the subspace of $\Gamma$-fixed vectors in $V$.
There are numerous simple and elegant proofs of Sunada's original theorem, some of which compare the dimension of each eigenspace in the two manifolds. These proofs go through without change in our setting. The $\sigma$-eigenfunctions for the $(\alpha,\rho)$-Steklov problem on each of the quotient manifolds $H\backslash M$ and $H'\backslash M$ pull back to $G$-invariant $\sigma$-eigenfunctions for the $(\alpha,\rho)$-Steklov problem on $M$. Thus letting $E_\sigma\subseteq C^{\infty}(M)$ be the $\sigma$-eigenspace for the $(\alpha,\rho)$-Steklov problem on $M$, we need only show that the subspaces $E_\sigma^H$ and $E_\sigma^{H'}$ of $H$-inveriant and $H'$-invariant functions, respectively, have the same dimension. Hence the proof of Theorem~\ref{thm:Sunada} reduces to the following lemma.
\begin{lem}\label{lem.dim} Let $H$ and $H'$ be almost conjugate subgroups of a finite group $G$ and let $V$ be any vector space on which $G$ acts. Then $\dim(V^H)=\dim(V^{H'})$.
\end{lem}
T. Sunada \cite{Sunada1985} gave an elementary proof of this lemma by a trace formula; see also \cite{BuserBook}, p. 295. H. Pesce~\cite{Pesce1994} gave a representation theoretic proof by applying Remark~\ref{rem.rep} along with Frobenius reciprocity to obtain
\[
\dim(V^{H})=[\boldsymbol{1}_{H}:\restrictedRep{H}{G}V]=[\inducedRep{H}{G}{\boldsymbol{1}_{H}}:V],
\]
where $[U:W]$ denotes the multiplicity of the representation $U$
in $W$. Since $\inducedRep{H}{G}{\boldsymbol{1}_{H}}$
and $\inducedRep{\mate{H}}{G}{\boldsymbol{1}_{\mate{H}}}$
are equivalent, it follows that $\dim(V^{H})=\dim(V^{\mate{H}})$.
\end{proof}
\begin{rems}\label{rem.Sunada} We note a couple of features of the Sunada construction.
\begin{enumerate}
\item\label{trans} Lemma~\ref{lem.dim} says that the vector spaces $V^H$ and $V^{H'}$ are isomorphic. In fact, the equivalence $\tau$ between the induced representations $\inducedRep{H}{G}{\boldsymbol{1}_{H}}$ and $\inducedRep{\mate{H}}{G}{\boldsymbol{1}_{\mate{H}}}$ actually yields an explicit and natural isomorphism $\tau^{\sharp}:V^{H'}\to V^H$, which Peter Buser and Pierre B\'erard \cite{Buser1986,Berard1992} called \emph{transplantation}. See also \cite{Zelditch}, \cite{BrooksGornetPerry}, \cite{GMW}.
\item If $H$ and $H'$ are conjugate subgroups of $G$, then the resulting quotient manifolds $H\backslash M$ and $H'\backslash M$ are isometric. Even when $H$ and $H'$ are not conjugate, the quotient manifolds may be accidentally isometric. Thus one must always verify nontriviality when using Sunada's technique (in fact, when using any of the known techniques for constructing isospectral manifolds).
\end{enumerate}
\end{rems}
More important for our purposes is:
\begin{remark}\label{rem.OrbSunada}
One may drop the hypothesis that $H$ and $H'$ act freely. The resulting quotients $H\backslashM$ and $\mate{H}\backslashM$ will then be Steklov isospectral good Riemannian orbifolds. (A \emph{good orbifold} is the orbit space $\mathcal{O}=\Gamma\backslash M$ of a manifold by a smooth discrete group action satisfying the condition that the isotropy group at any point is finite. A function on $\mathcal{O}$ is said to be \emph{smooth} if its pullback to $M$ is smooth. If $g$ is a Riemannian metric on $M$ and $\Gamma$ acts by isometries, then $g$ gives $\mathcal{O}$ the structure of a Riemannian orbifold. The associated Laplacian $\Delta_\mathcal{O}: C^\infty(\mathcal{O})\to C^\infty(\mathcal{O})$ is defined by $\pi^*\circ \Delta_\mathcal{O}=\Delta_M\circ\pi^*$ where $\pi:M\to\mathcal{O}$ is the projection.) We will apply the orbifold version in Example~\ref{exa.domains} when we construct planar domains that are isospectral for the sloshing problem.
\end{remark}
\begin{oep}\label{rem.sun} (i) There are various notions in the literature of a Dirichlet-to-Neumann operator acting on the space of smooth differential $p$-forms on $\partial M$ where $M$ is a smooth compact Riemannian manifold with smooth boundary. The notions of Dirichlet-to-Neumann operator on forms defined by S. Raulot and A. Savo \cite{RS} and by Karpukhin \cite{Karp} have discrete spectra. Using either of these definitions of Steklov spectrum on $p$-forms, the hypotheses of Theorem~\ref{thm:Sunada} (with $\rho\equiv 1$) guarantee that the manifolds $H\backslash M$ and $H'\backslash M$ have the same Steklov spectra on $p$-forms, for all $p$.
(ii) As noted in the introduction, taking $\rho\equiv1$ in Theorem~\ref{thm:Sunada} immediately yields isospectrality of the Robin problems on $H\backslash M$ and $H'\backslash M$ for every choice of Robin parameter. Alternatively, one can prove the Robin isospectrality directly using the same method as in the proof of Theorem~\ref{thm:Sunada}.
Moreover, one can easily modify Theorem~\ref{thm:Sunada} to address mixed Robin-Neumann-Dirichlet problems. One assumes that $\partial M=\partial_R M\sqcup \partial_N M\sqcup \partial _D M$, where each of the three subsets is $G$-invariant and where the decomposition is sufficiently nice so that the mixed Robin-Neumann-Dirichlet problem, in which Robin, Neumann, and Dirichlet conditions are imposed on $\partial_R M,\, \partial_N M$ and $\partial _D M$, respectively, is well-defined with discrete spectrum. Then the mixed problems on $H\backslash M$ and $H'\backslash M$ are isospectral, where the respective boundary conditions are imposed on $H\backslash(\partial_R M),\, H\backslash(\partial_N M)$, and $H\backslash(\partial _D M)$ and similarly for $H'$.
\end{oep}
\section{The torus action method\label{sec:The_torus_action_method}}
The torus action method was developed to construct Riemannian manifolds that have the same Laplace spectrum but that are not even locally isometric. There are several versions, e.g., ~\cite{Gordon1994,Gordon2001,Schueth2001a, Schueth2001}. We first state the version in \cite{Schueth2001} and then adapt it to the Steklov setting.
In the following, a \emph{torus} always means a nontrivial, compact, connected, abelian Lie group. Let~$T$ be a torus acting effectively by isometries on a compact, connected Riemannian manifold~$M$. The union of those orbits on which $T$ acts freely is an open, dense submanifold of $M$ (see \cite{Bredon}) that we will denote by $\principleOrbits{M}$ ; it carries the structure of a principal $T$-bundle.
\begin{thm}\label{thm.schueth}\cite{Schueth2001} Let $T$ be a torus which acts effectively on two compact,
connected Riemannian manifolds $(M,g)$ and $(M',g')$ by isometries.
For each subtorus $W\subset T$ of codimension one, suppose that there exists
a $T$-equivariant diffeomorphism $F_W: M\to M'$ such that
\begin{enumerate}
\item $F_W:M\to M'$ is volume-preserving; i.e., $F_W^* dvol_{M'}= dvol_M$ where $dvol_M$ and $dvol_{M'}$ are the Riemannian volume densities of $M$ and $M'$;
\item $F$ induces an isometry $\overline{F}_W: (W\backslash\widehat{M}, g_W)\to(W\backslash\widehat{M'},g'_W)$ where $g_W$ and $g'_W$ are the metrics induced by $g$ and $g'$ on the quotients.
\end{enumerate}
Then $(M,g)$ and $(M',g')$ are
isospectral. Moreover, if the manifolds have boundary, then they are both Dirichlet and Neumann isospectral.
\end{thm}
We now adapt this method to the Steklov setting.
\begin{thm}
\label{thm:Torus_method}
Let $T$ be a torus which acts isometrically and effectively on two compact,
connected Riemannian manifolds $(M,g)$ and $(M',g')$ with boundary. Let $\rho\in L^\infty(\partial M)$ and $\rho'\in L^\infty(\partial M')$ be $T$-invariant.
For each subtorus $W\subset T$ of codimension one, suppose that there exists
a $T$-equivariant diffeomorphism $F_W: M\to M'$ such that
\begin{enumerate}
\item\label{F_WVolPres}
$F_W:M\to M'$ is volume-preserving;
\item\label{F_WBoundaryVolPres}
$F\vert_{\partial M}:\partial M\to\partial M'$ is volume-preserving, i.e., $F_W^* dvol_{\partial M'}= dvol_{\partial M}$;
\item\label{F_WPresDensity} $F_W^*\rho'=\rho$;
\item\label{F_WIsomOnQuots} $F_W$ induces an isometry $\overline{F}_W: (W\backslash\widehat{M}, g_W)\to(W\backslash\widehat{M'},g'_W)$, where $g_W$ and $g'_W$ are the metrics induced by $g$ and $g'$ on the quotients.
\end{enumerate}
Then for each $\alpha$ not in the Dirichlet spectrum of $(M,g)$, we have
\begin{equation}\label{eq.same}\operatorname{Stek}_\alpha(M,g,\rho)=\operatorname{Stek}_\alpha(M',g',\rho').\end{equation}
(Theorem~\ref{thm.schueth} guarantees that the two quotient manifolds are Dirichlet isospectral, so the allowable choices of $\alpha$ are the same in both cases.)
\end{thm}
Before proving Theorem~\ref{thm:Torus_method}, we recall the variational characterization of the eigenvalues in $\operatorname{Stek}_\alpha(M,g,\rho)$. First recall that the boundary restriction map that takes $u\in H^1(M)\cap C^0(M)$ to $u|_{\partial M}$ extends to the compact trace operator $\operatorname{Tr}: H^1(M)\to L^2(\partial M)$. We write $u|_{\partial M}=\operatorname{Tr}(u)$. Define
\begin{equation}\label{eq.ray} R_{M,\alpha,\rho}(u)=\frac{\int_M\,\Vert\nabla u\Vert^2\,dvol_M-\alpha\int_M\, u^2\,dvol_M}{\int_{\partial M}\,u|_{\partial M}^2\,\rho\,dvol_{\partial M}}\end{equation}
Denoting the eigenvalues in $\operatorname{Stek}_\alpha(M,g,\rho)$ as
$$0=\sigma_0<\sigma_1\leq \sigma_2\leq\dots,$$
we have
\begin{equation}\label{minmax}\sigma_k=\inf_{E_k(M,\rho)}\sup_{0\neq u\in E_k}\,R_{M,\alpha,\rho}(u)\end{equation}
where the infimum is over all $k$-dimensional subspaces $E_k(M,\rho)$ of $H^1(M)$ consisting of functions whose restrictions to the boundary are $\rho$-orthogonal to the constant functions, i.e $\int_{\partial M}\,u|_{\partial M}\,\rho\,dvol_{\partial M}=0$.
\begin{proof}
[Proof of Theorem \ref{thm:Torus_method}] We adapt the proof of~\cite[Theorem 1.4]{Schueth2001}. For $W<T$ any subtorus, let $H^1(M)^W \subset H^1(M)$, $L^2(M)^W \subset L^2(M)$, $H^1(M')^W \subset H^1(M')$, and $L^2(M')^W \subset L^2(M')$
denote the subspaces of $W$-invariant functions. By Fourier decomposition with respect to the isometric action of $T$, we have
\begin{equation}\label{eq.decomp}H^1(M)=H^1(M)^T\,\oplus\,\bigoplus_W\,(H^1(M)^W\ominus H^1(M)^T)\end{equation}
and
\begin{equation}\label{eq.decomp2}L^2(\partial M)=L^2(\partial M)^T\,\oplus\,\bigoplus_W\,(L^2(\partial M)^W\ominus L^2(\partial M)^T)\end{equation}
where the sum is over all subtori $W$ of $T$ of codimension one. Multiplication by the $T$-invariant density $\rho$ preserves each of the subspaces $L^2(M)^T$ and $L^2(M)^W$. Moreover the trace operator $\mathrm{Tr}: H^1(M)\to L^2(\partial M)$ respects these decompositions. Analogous statements hold with $M$ replaced by $M'$.
As shown in~\cite{Schueth2001}, conditions \eqref{F_WVolPres} and \eqref{F_WIsomOnQuots} of Theorem~\ref{thm:Torus_method} imply that if $W$ is a subtorus of $T$ of codimension at most one and $u\in H^1(M')^W$, then
\begin{equation}\label{eq.h1}
\norm{F_{\subtorus}^{*}u}{\HOneSpace{M}{}}=\norm u{\HOneSpace{\mate{M}}{}}\qquad\text{and also}\qquad\norm{F_{\subtorus}^{*}u}{\LTwoSpace{M}}=\norm u{\LTwoSpace{\mate{M}}}.
\end{equation}
The first of the equations in~\eqref{eq.h1}, the $T$-equivariance of the maps $F_W$, and Equation~(\ref{eq.decomp}) yield an isomorphism
$$\tau: H^1(M')\to H^1(M)$$
given by
$$\tau=F_T^*\,\oplus\,\bigoplus_W\,F_W^*.$$
Hypothesis \eqref{F_WBoundaryVolPres} of the theorem and Equation~(\ref{eq.decomp2}) similarly yield an isomorphism
$$\tau_\partial:=F_T^*\,\oplus\,\bigoplus_W\,F_W^*: L^2(\partial M')\to L^2(\partial M)$$
and the diagram
$$\xymatrix{
{H^1(M')}\ar^{\tau}[r]\ar_{\operatorname{Tr}}[d]&{H^1(M')}\ar^{\operatorname{Tr}}[d]\\
{L^2(\partial M')}\ar_{\tau_{\partial}}[r]&{L^2(\partial M)}
}$$
commutes.
Hypotheses \eqref{F_WBoundaryVolPres} and \eqref{F_WPresDensity} of the theorem guarantee for each $k=1,2,\dots$ that $\tau$ maps $E_k(M',\rho')$ to $E_k(M,\rho)$ and that the denominators in the Rayleigh quotients $R_{M,\alpha,\rho}(\tau(u))$ and $R_{M',\alpha,\rho'}(u)$ coincide for each $u\in E_k(M',\rho')$.
The pair of equalities~\eqref{eq.h1} imply that the
numerators in $R_{M,\alpha,\rho}(\tau(u))$ and $R_{M',\alpha,\rho'}(u)$ also agree, and the theorem follows from Equation~(\ref{minmax}).\end{proof}
Although condition \eqref{F_WBoundaryVolPres} in Theorem~\ref{thm:Torus_method} does not appear in Theorem~\ref{thm.schueth} or in any of the other versions of the torus action method, it is actually satisfied in all of the examples that have been constructed thus far by these methods, as will be explained
in Section~\ref{sec:Examples}. Moreover, the version of the torus action method in \cite[Theorem 1.2]{Gordon2001} includes a hypothesis that the principal $T$-orbits be dense in $\boundary{M}$ and $\boundary{\mate{M}}$ in order to produce Neumann isospectral manifolds; this condition is stronger than condition \eqref{F_WBoundaryVolPres} in the following sense.
\begin{prop}
\label{prop:Boundary_volume_form_preservation} Let $M$ and $\mate{M}$ be compact, connected, orientable Riemannian manifolds with a faithful isometric action by a torus $T$ satisfying conditions~\eqref{F_WVolPres} and~\eqref{F_WIsomOnQuots} of Theorem~\ref{thm:Torus_method}. If~$\principleOrbits M\cap\boundary{M}$ is dense in $\boundary{M}$, then condition \eqref{F_WBoundaryVolPres} of Theorem~\ref{thm:Torus_method} is satisfied as well.\end{prop}
\begin{proof} Since the manifolds are orientable, condition~\eqref{F_WVolPres} says that $F_W$ pulls back the Riemannian volume form $dvol_{M'}$ of $M'$ to that of $M$.
By continuity, it suffices to show that condition \eqref{F_WBoundaryVolPres} holds at each $p\in\principleOrbits M\cap\boundary{M}$. Let $p\in \widehat{M}\cap \partial M$ and let $p'=F_W(p)$. The $T$-equivariance of $F_W$ guarantees that $p'\in\widehat{M}'$. Let $\nu$ and $\nu'$ denote the outward unit normals to $\partial M$ and $\partial M'$ at $p$ and $p'$, respectively, and let $i:\partial M\to M$ and $i':\partial M'\to M'$ be the inclusion maps. The facts that $F_W$ is an isometry and that the action of $W$ on $M$ and $M'$ preserves the boundaries imply that $\nu'-{F_W}_*(\nu)$ is tangent to $\partial M'$ and hence
\begin{equation}\label{eq.tan} (i')^*({F_W}_*(\nu)\,\lrcorner\, dvol_{M'})=(i')^*(\nu'\,\lrcorner\, dvol_{M'}) = dvol_{\partial M'}
\end{equation}
By condition (1) of Theorem~\ref{thm:Torus_method}, we have
\begin{equation}\label{eq.pull}{F_W}_*(\nu)\,\lrcorner\, dvol_{M'}={F_W}_*(\nu)\,\lrcorner\, (F_W^{-1})^* dvol_{M}=(F_W^{-1})^*(\nu\,\lrcorner\,dvol_{M})\end{equation}
Since $F_W\circ i=i'\circ F_W$, Equations~\ref{eq.tan} and \ref{eq.pull} yield
$$F_W^*(dvol_{\partial M'})=F_W^*\circ(i')^*\circ (F_W^{-1})^*(\nu\,\lrcorner\,dvol_{M})=i^*(\nu\,\lrcorner\,dvol_{M})=dvol_{\partial M}.$$
\end{proof}
\section{Examples\label{sec:Examples}}
\subsection{Examples using the Sunada technique}
There is a wealth of examples of Dirichlet or Neumann isospectral manifolds that have been constructed by the Sunada method and its various generalizations; see \cite{G:SunadaTwoDecades} and references therein. The original Sunada technique has yielded, for example, isospectral flat surfaces embedded in $\mathbb{R}^3$ \cite{Buser1988} and large finite families of mutually isospectral Riemann surfaces \cite{BGG}, which can be easily modified to produce families of mutually isospectral hyperbolic surfaces with boundary. All examples of isospectral manifolds with boundary constructed by the original Sunada technique are also Steklov isospectral.
There are various generalizations of Sunada's theorem, surveyed in \cite{G:SunadaTwoDecades}, not all of which go through directly for the Steklov spectrum. For example, the pair of Neumann isospectral flat surfaces with boundary constructed in \cite{BerardWebb1995} (one orientable, the other nonorientable) using the orbifold version of Sunada's Theorem are not Steklov isospectral, since one of the manifolds has four boundary components while its isospectral companion has only three boundary components. Yet the number of boundary components of a surface is determined by the Steklov spectrum (see \cite{GirouardParnovskiPolterovichSher2014}). See \ref{exa.domains} for some further comments.
In this subsection we illustrate the Sunada method with just a sampling of the many examples.
\subsubsection{Steklov isospectral flat surfaces embedded in $\mathbb{R}^3$}\label{exa.buser}
In \cite{Buser1988}, Peter Buser introduced the use of Schreier graphs to construct isospectral manifolds via Sunada's technique and illustrated the method by constructing a pair of Dirichlet and Neumann isospectral flat surfaces with boundary in $\mathbb{R}^3$.
For the reader's convenience, we briefly review Buser's construction before addressing the Steklov setting. Recall that if $G$ is a finite group and $S=\{s_{1},s_{2},\ldots,s_{n}\}$ is a set of nonidentity elements generating $G$, the \emph{Cayley graph} $\Gamma(G,S)$ is the $n$-regular edge-colored directed graph whose vertices are the elements of~$G$, and whose $i$-colored edges encode right multiplication by the generators $s_{i}$. More precisely, there is an $i$-colored edge from $g$ to~$g'$ if and only if $g s_i=g'$. The group $G$ acts transitively and faithfully on $\Gamma(G,S)$ by left multiplication. If $H$ is a subgroup of $G$, then the \emph{Schreier graph} $\Gamma(H\backslash G, S)$ is the quotient of $\Gamma(G,S)$ by the action of $H$. Equivalently, the vertices of the Schreier graph correspond to the elements of the space of right-cosets $H\backslash G$ and the edges indicate the right action of the elements of $S$ on $H\backslash G$. The graph theoretic version of Sunada's Theorem says that if $H_1$ and $H_2$ are almost conjugate subgroups of $G$, then for any fixed choice of generating set $S$, the adjacency operators (or Laplacians or other natural operators) associated with the Schreier graphs $\Gamma(H_1\backslash G,S)$ and $\Gamma(H_2\backslash G,S)$ are isospectral.
To construct a manifold from a Schreier graph, Buser chooses a basic tile $T$, whose piecewise-smooth boundary contains $2n$ disjoint line segments called sides, labelled $s_1$, $s_1^{-1}$, $s_2$, $s_2^{-1},\ldots,s_n,s_n^{-1}$.
Sides $s_i$ and $s_i^{-1}$ are required to have the same length. The sides need not exhaust the entire boundary of the tile. To construct a manifold $M(H\backslash G,S)$, consider a collection of $[G:H]$ identical tiles, labelled by the elements of $H\backslash G$, whose sides are glued together in pairs according to the pattern encoded by the Schreier graph. More precisely, side $s_{i}$ of tile $Hg$ is glued to side $s_{i}^{-1}$ of tile $H g s_i$. Similarly, one uses the Cayley graph $\Gamma(G,S)$ to construct a manifold $M(G,S)$. Observe that $G$ acts on $M(G,S)$ on the left, and that $M(H\backslash G,S)=H\backslash M(G,S)$. Let $\partial _0(T)$ denote the complement in $\partial T$ of the union of the sides $s_{1},s_{1}^{-1},s_{2},s_{2}^{-1},\ldots,s_{n},s_{n}^{-1}$. Buser arbitrarily chooses boundary conditions on $\partial_0 T$. The boundary conditions chosen on $\partial_0 T$ then determine the boundary conditions on the manifold $M(G,S)$ and on $M(H\backslash G,S)$ for any subgroup $H<G$.
Now suppose that $H_1$ and $H_2$ are almost conjugate subgroups of $G$. Then, as observed by Buser, Sunada's Theorem immediately yields isospectrality of $M_1:=M(H_1\backslash G,S)$ and $M_2:=M(H_2\backslash G,S)$ with respect to the given boundary conditions.
Moving to our setting, we instead choose arbitrarily an $L^\infty$ density function $\rho$ on $\partial_0 T$, thus giving rise to a density function, still denoted $\rho$, on the boundaries of $M:=M(G,S)$ and $M_i$, $i=1,2$. The density on $\partial M(G,S)$ is $G$-invariant and Theorem~\ref{thm:Sunada} yields
$$\operatorname{Stek}_\alpha(M_1,\rho)=\operatorname{Stek}_\alpha(M_2,\rho).$$
It is easy to construct an abundance of examples this way. For a concrete example, we consider the pair of flat surfaces in $\mathbb{R}^3$ given by Buser in \cite{Buser1988}. In this example, $G=\operatorname{GL}(3,\mathbb{Z}_2)$, $H_1$ is the subset of matrices with first row $(1,0,0)$, and $H_2=H_1^{t}$ is the subset consisting of transposes of elements of $H_1$. The two subgroups $H_1$ and $H_2$ are almost conjugate in $G$ (each element of $H_1$ is similar to its transpose in $H_2$) and have index 7 in $G$. Buser's surfaces are obtained by using a particular generating set $S=\{a,b\}$ of order 2 and the basic tile shown in Figure~\ref{fig:Basic_tile}. (Ignore for now the dashed line in Figure~\ref{fig:Basic_tile}; it will be used in the next example.) Buser actually used a cross-shaped tile; we have smoothed out the corners of the tile so that the resulting isospectral surfaces $M_1$ and $M_2$ are smooth.
We have not included a picture of the two surfaces here. However, Figure~\ref{fig:RS_isospectral_planar_domains} shows the quotient of each of the two surfaces by a reflection. To visualize the original surfaces, simply double the two domains in the figure across the part of the boundary indicated by double lines. Alternatively, see \cite{Buser1988}, where the surfaces constructed with a cross-shaped tile are drawn.
The surfaces are easily seen to be nonisometric; in fact they have different diameter. Since we are in dimension two, we also verify that they are not trivially Steklov isospectral when $\alpha=0$ by showing that $M_2$ is not isometric to the surface $M_1$ endowed with a metric $e^f g_E$, where $g_E$ is the Euclidean metric and where the conformal factor $f$ vanishes on the boundary. Recall that the scalar curvature of $e^f g_E$ is $4e^{-f}\Delta f$, where $\Delta$ denotes the Euclidean Laplacian. Noting that $M_2$ is flat, we conclude that $f$ must be a harmonic function. Since $f$ vanishes on the boundary, $f$ must be identically zero. Thus no such conformal equivalence exists and the surfaces are nontrivially Steklov isospectral.
\noindent
\begin{figure}
\noindent \begin{centering}
\hfill{}\subfloat[Buser's tile\label{fig:Basic_tile}]{\noindent \protect\begin{centering}
\begin{minipage}[b]{0.2\columnwidth}%
\noindent \protect\begin{center}
\raisebox{0.2mm}
{\begin{tikzpicture}[scale=0.4,thick,rotate=45]
\draw (1.77,4.77) node {$a$};
\draw (5.25,1.6) node {$a^{-1}$};
\draw (1.9,-0.92) node {$b$};
\draw (-1,2) node {$b^{-1}$};
\draw[dashed,thin] ($(0,0) + (45:1)$) -- ($(4,4) + (225:1)$);
\draw ( 1, 0) arc (0:90:1);
\draw ( 3, 0) arc (180:90:1);
\draw ( 1, 4) arc (0:-90:1);
\draw ( 3, 4) arc (180:270:1);
\draw[double] (0.97,0) -- (3.03,0);
\draw[double] (0,0.97) -- (0,3.03);
\draw[double] (0.97,4) -- (3.03,4);
\draw[double] (4,0.97) -- (4,3.03);
\end{tikzpicture}}\protect
\par\end{center}%
\end{minipage}\protect
\par\end{centering}
}\hfill{}\subfloat[Steklov-Neumann and Robin-Neumann isospectral planar domains\label{fig:RS_isospectral_planar_domains}]{\noindent \protect\begin{centering}
\begin{minipage}[b]{0.7\columnwidth}%
\noindent \protect\begin{center}
\begin{tikzpicture}[scale=0.4,thick,rotate=45]
\draw[dashed,thin] ($(3,3) + (45:1)$) -- ($(7,7) + (225:1)$);
\draw[dashed,thin] ($(7,3) + (45:1)$) -- ($(11,7) + (225:1)$);
\draw[dashed,thin] (8,3) -- (10,3);
\draw[dashed,thin] (7,4) -- (7,6);
\draw[dashed,thin] (3,4) -- (3,6);
\draw[dashed,thin] (0,7) -- (2,7);
\draw (7,0) arc (90:45:1);
\draw (11,4) arc (90:225:1);
\draw (10,7) arc (180:270:1);
\draw (6,7) arc (180:360:1);
\draw (3,8) arc (90:360:1);
\draw (3,10) arc (270:225:1);
\draw ($(-1,7)+(45:1)$) arc (45:-45:1);
\draw ($(3,3)+(135:1)$) arc (135:0:1);
\draw (7,2) arc (-90:180:1);
\draw[double] ($(7,-1)+(45:0.97)$) -- ($(11,3)+(225:0.97)$);
\draw[double] (11,3.97) -- (11,6.03);
\draw[double] (10.03,7) -- (7.97,7);
\draw[double] (6.03,7) -- (3.97,7);
\draw[double] (3,7.97) -- (3,10.03);
\draw[double] ($(3,11)+(225:0.97)$) -- ($(-1,7)+(45:0.97)$);
\draw[double] ($(-1,7)+(-45:0.97)$) -- ($(3,3)+(135:0.97)$);
\draw[double] (3.97,3) -- (6.03,3);
\draw[double] (7,-0.03) -- (7,2.03);
\end{tikzpicture}\hspace{10mm}\begin{tikzpicture}[scale=0.4,thick,rotate=45]
\draw[dashed,thin] ($(-1,7) + (45:1)$) -- ($(3,11) + (225:1)$);
\draw[dashed,thin] ($(3,3) + (45:1)$) -- ($(7,7) + (225:1)$);
\draw[dashed,thin] (4,3) -- (6,3);
\draw[dashed,thin] (7,4) -- (7,6);
\draw[dashed,thin] (3,4) -- (3,6);
\draw[dashed,thin] (0,7) -- (2,7);
\draw ($(3,3)+(135:1)$) arc (135:-45:1);
\draw (7,0) arc (90:135:1);
\draw (7,2) arc (270:0:1);
\draw (10,3) arc (180:135:1);
\draw (6,7) arc (180:315:1);
\draw (3,8) arc (90:360:1);
\draw (3,10) arc (270:180:1);
\draw (0,11) arc (360:270:1);
\draw (-1,8) arc (90:-45:1);
\draw[double] ($(-1,7)+(-45:0.97)$) -- ($(3,3)+(135:0.97)$);
\draw[double] ($(3,3)+(-45:0.97)$) -- ($(7,-1)+(135:0.97)$);
\draw[double] (7,-0.03) -- (7,2.03);
\draw[double] (7.97,3) -- (10.03,3);
\draw[double] ($(7,7)+(315:0.97)$) -- ($(11,3)+(135:0.97)$);
\draw[double] (6.03,7) -- (3.97,7);
\draw[double] (3,7.97) -- (3,10.03);
\draw[double] (2.03,11) -- (-0.03,11);
\draw[double] (-1,10.03) -- (-1,7.97);
\end{tikzpicture}\protect
\par\end{center}%
\end{minipage}\protect
\par\end{centering}
}\hfill{}
\par\end{centering}
\caption{Neumann conditions are imposed on all straight boundary parts (double-lined). The domains arise from~\cite[Figure 7]{GordonWebbWolpert1992} by
using tiles as in Figure~\ref{fig:Basic_tile}.}\label{Figure1}
\end{figure}
\subsubsection{Planar domains with isospectral sloshing problems} \label{exa.domains}
The first examples of isospectral planar domains \cite{GordonWebbWolpert1992} arose from the observation that the two isospectral flat surfaces $M_i$, $i=1,2$, described in the previous example each admit an isometric involution $\beta_i$, covering the symmetry $\beta_0$ of the basic tile in Figure~\ref{fig:Basic_tile} given by reflection across the dashed line. The quotients of the surfaces by the involutions, shown in Figure~\ref{fig:RS_isospectral_planar_domains}, are both Dirichlet and Neumann isospectral. As we will explain below, the version of Sunada's technique used to prove isospectrality does not yield Steklov isospectrality of these domains except in the special case that the density $\rho$ is identically zero on the part of the boundary indicated by double lines (the straight segments of the boundary) in Figure~\ref{fig:RS_isospectral_planar_domains}. However, if we choose $\rho$ to be zero on this part of the boundary and $\rho\equiv 1$ on the curved edges, then we do obtain isospectrality for the mixed Neumann-Steklov problem (the sloshing problem). One can also make a more general choice of $\rho$ on the curved parts of the boundary as long as consistency is maintained among the various tiles.
The proof in \cite{GordonWebbWolpert1992} of Neumann isospectrality of the planar domains goes as follows: The involutive isometries $\beta_i$, $i=1,2$, lift to an involutive isometry $\beta$ of the covering manifold $M=M(G,S)$. The isometry $\beta$ normalizes the group $G$ and each of the subgroups $H_i$, $i=1,2$. The groups $\tilde{H_1}:=H_1\rtimes\langle\beta\rangle$ and $\tilde{H_2}:= H_2\rtimes\langle\beta\rangle$ are almost conjugate subgroups of $\tilde{G}:=G\rtimes\langle\beta\rangle$. The group $\tilde{G}$ does not act freely on $M$. However, we may apply the orbifold version of Sunada's Theorem as in Remark~\ref{rem.OrbSunada} to conclude that the quotients $\tilde{H_1}\backslash M$ and $\tilde{H_2}\backslash M$ are isospectral orbifolds. The underlying spaces of these orbifolds are the domains in Figure~\ref{fig:RS_isospectral_planar_domains}. The singular sets of these orbifolds consist of the doubled line segments in Figure~\ref{fig:RS_isospectral_planar_domains}, which are reflector edges where the isotropy group has order $2$. (Note that these line segments lift to interior segments of $M$, not to boundary edges.) By the definition of smooth functions and of the Laplacian on these orbifolds (see Remark~\ref{rem.OrbSunada}), the isospectrality of the two orbifolds is equivalent to isospectrality of the underlying planar domains with Neumann boundary conditions placed on the doubled line segments of the boundary and whatever boundary conditions on the curved edges were chosen on the curved edges of the basic tile $T$ used to construct $M$.
If we choose the boundary density function $\rho\equiv 1$ on the boundary of the basic tile, the same argument yields the Steklov isospectrality of the two orbifolds, which in turn corresponds to isospectrality for the sloshing problem on the two underlying planar domains.
\begin{remark}\label{rem.trans2} We have summarized the original proof of the isospectrality of the planar domains in order to make clear the reason we can only get sloshing isospectrality rather than more general Steklov isospectrality of the planar domains. However, transplantation as in Remark~\ref{trans} yields a very simple proof by picture of the sloshing isospectrality.
\end{remark}
\subsubsection{Mixed Robin-Neumann-Dirichlet and Steklov-Neumann-Dirichlet isospectral domains}\label{exa.RND}
Levitin, Parnovski and Polterovich \cite{LPP} constructed examples of pairs of domains that are isospectral with mixed boundary conditions, including a pair consisting of a triangle and a square, whose isospectrality cannot be explained directly by Sunada's technique but can be shown by an explicit transplantation of eigenfunctions. Later Band and Parzanchevsky \cite{BP} gave a representation theoretic explanation, which was further developed and applied systematically in Herbrich \cite{Herbrich2011}.
One can similarly use transplantation directly to obtain domains that are isospectral for the mixed Robin-Neumann-Dirichlet and mixed Steklov-Neumann-Dirichlet problems. We give two examples here, both obtained by modifying the construction of the isospectral triangle and square in \cite{LPP}. The triangle and square in \cite{LPP} are each constructed by gluing together two copies of an isosceles right triangle (the basic tile); they are glued along the hypotenuse to obtain the square and along one of the legs to obtain the triangle in the isospectral pair. Figure~\ref{fig:RS_isospectral_square_and_triangle} shows two modifications of their construction, both obtained by cutting out a half disk from the basic tile.
For the mixed Robin-Neumann-Dirichlet problem, we impose Robin boundary conditions --- with the same Robin parameter on both domains in each pair --- on the curved part of the boundary indicated by a solid line in the figures, Neumann conditions on the part of the boundary indicated by doubled lines, and Dirichlet conditions on the part indicated by dashed lines. With these boundary conditions we claim that $M$ is isospectral to $M'$ and $P$ is isospectral to $P'$.
Let $u$ be an eigenfunction for the mixed problem on $M$, say with eigenvalue $\lambda$, and denote by $u_1$ and $u_2$ the restrictions of $u$ to the two tiles making up $M$ as in Figure~\ref{fig:RS_isospectral_square_and_triangle}.
We transplant $u$ to an eigenfunction $u'=T(u)$ on $M'$ whose restrictions $u_1'$ and $u_2'$ to the two tiles of $M'$ as in Figure~\ref{fig:RS_isospectral_square_and_triangle} are given by
\begin{equation}
\left(\begin{array}{c}
\mate u_{1}\\
\mate u_{2}
\end{array}\right)=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}
1 & -1\\
1 & 1
\end{array}\right)\left(\begin{array}{c}
u_{1}\\
u_{2}
\end{array}\right).\label{eq:Transplantation_in_matrix_form}
\end{equation}
In writing $u_{1}\pm u_{2}$, we implicitly identify the tiles underlying
$u_{1}$ and $u_{2}$, which involves a reflection in the dotted diagonal
of $M$. To see that $u'$ is smooth on the dotted interior segment, we observe that $u_{1}$ extends smoothly by reflection across this segment (since the segment corresponds to an edge in $M$ where $u_1$ satisfies Neumann conditions) and, similarly, $u_{2}$ smoothly extends by negative reflection across this segment (which corresponds to an edge of $M$ where $u_2$ satisfies Dirichlet conditions). It is then straightforward to verify that $u'$ is an eigenfunction with eigenvalue $\lambda$ for the mixed Robin-Neumann-Dirichlet problem. The transplantation map $T$ is invertible and isospectrality follows. The same transplantation map yields the mixed Robin-Neumann-Dirichlet isospectrality of $P$ and $P'$.
To prove the Steklov-Neumann-Dirichlet isospectrality of $M$ and $M'$ and of $P$ and $P'$, one uses the same expression for the transplantation map $T$, but now acting on Steklov-Neumann-Dirichlet eigenfunctions. Alternatively, the isospectrality is immediate from the duality between the Steklov-Neumann-Dirichlet and the Robin-Neumann-Dirichlet problem.
\noindent
\begin{figure}
\noindent \begin{centering}
\begin{tikzpicture}[scale=0.4,thick]
\draw (3.1,2.1) node {$u_{1}$};
\draw (1.9,0.8) node {$u_{2}$};
\draw (9.1,2.3) node {$\mate{u}_{1}$};
\draw (9.1,5.7) node {$\mate{u}_{2}$};
\draw (-1.0,0.8) node {$M$};
\draw (7.5,0.8) node {$\mate{M}$};
\draw (9.3,4) arc (0:360:1);
\draw[dashed] (10, 0) -- (10, 8);
\draw[double] (10, 8) -- ( 6, 4);
\draw[dashed] ( 6, 4) -- (10, 0);
\draw[dotted,thin] ( 6,4) -- (7.3,4);
\draw[dotted,thin] (9.3,4) -- (10,4);
\draw (1.3,4) arc (180:360:1);
\draw ( 0,2.7) arc (90:-90:1);
\draw[dashed] ( 0, 0) -- (4, 0);
\draw[dashed] ( 4, 0) -- (4, 4);
\draw[double] ( 4, 4) -- (3.27, 4);
\draw[double] (1.33, 4) -- (0, 4);
\draw[dashed] ( 0, 4) -- (0,2.67);
\draw[dashed] ( 0,0.73) -- (0, 0);
\draw[dotted,thin] ( 0, 4) -- ( 4, 0);
\end{tikzpicture}\hspace{2cm}\begin{tikzpicture}[scale=0.4,thick]
\draw (2.4,3.0) node {$u_{1}$};
\draw (1.0,1.8) node {$u_{2}$};
\draw (8.4,3.0) node {$\mate{u}_{1}$};
\draw (8.4,5.0) node {$\mate{u}_{2}$};
\draw (-1 ,0.8) node {$P$};
\draw (7.5,0.8) node {$\mate{P}$};
\draw (10,3.3) arc (90:270:1);
\draw (10,6.7) arc (90:270:1);
\draw[dashed] (10, 0) -- (10,1.3);
\draw[dashed] (10,3.3) -- (10,4.7);
\draw[dashed] (10,6.7) -- (10, 8);
\draw[double] (10, 8) -- ( 6, 4);
\draw[dashed] ( 6, 4) -- (10, 0);
\draw[dotted,thin] (6,4) -- (10,4);
\draw (2.7,0) arc (0:180:1);
\draw (4,3.3) arc (90:270:1);
\draw[dashed] ( 0, 0) -- (0.7,0);
\draw[dashed] (2.7, 0) -- (4, 0);
\draw[dashed] ( 4, 0) -- (4,1.3);
\draw[dashed] ( 4,3.3) -- (4, 4);
\draw[double] ( 4, 4) -- (0, 4);
\draw[dashed] ( 0, 4) -- (0, 0);
\draw[dotted,thin] ( 0, 4) -- ( 4, 0);
\end{tikzpicture}
\par\end{centering}
\protect\caption{Robin-Steklov isospectral planar pairs. They are based on the main
example in~\cite{LPP}. Isospectrality
follows from the transplantation~(\ref{eq:Transplantation_in_matrix_form}).\label{fig:RS_isospectral_square_and_triangle}}
\end{figure}
\subsection{Examples using the torus action method}
The torus action method, e.g., Theorem~\ref{thm.schueth}, has led to numerous pairs and families of Dirichlet and Neumann isospectral manifolds as well as isospectral closed manifolds. All known examples satisfy the additional condition \eqref{F_WBoundaryVolPres} of Theorem~\ref{thm:Torus_method} and therefore have isospectral Dirichlet-to-Neumann operators at all frequencies.
In fact, Proposition~\ref{prop:Boundary_volume_form_preservation} applies to all of them, yielding condition~(2)
in Theorem~\ref{thm:Torus_method}. Letting $B^ n$ and $T^{n}$ denote the $n$-dimensional ball and torus, respectively,
the examples include:
\begin{enumerate}
\item Continuous families of nonisometric metrics on $B^{n}$ for $n\geq8$~\cite{Gordon2001,Schueth2001},
and pairs of such metrics on $B^6$ and $B^7$ \cite{Schueth2001}.
These metrics can be chosen as Euclidean outside of a smaller concentric ball~\cite{Schueth2001}. \label{enu:Continuous_families_of_isospectral_balls}
\item Continuous families of metrics on $B^n\timesT^{k}$ for ${n\geq5}$ and $k\geq 2$
that are the restrictions of locally nonisometric homogeneous metrics on $\mathbb{R}^{n}\timesT^{k}$~\cite{GordonWilson1997}.
\item For $n\geq6$, if one removes a concentric ball from $B^{n}$ to obtain an annulus $M$ and takes $\rho\equiv 1$ on one of the boundary spheres and $\rho\equiv 0$ on the other, then the metrics in (1) and (2)
restrict to metrics with isospectral sloshing problems on $M$.
\end{enumerate}
\section{Isospectral density functions}\label{sec:density}
In \cite{Brooks}, R. Brooks modified Sunada's theorem in order to construct isospectral potentials for the Schr\"odinger operator. Shortly thereafter, a similar method was used to construct isospectral conformally equivalent Riemannian metrics \cite{BPY}. The technique became standard and produced many new examples. Later D. Schueth \cite{Schueth2001a} analogously modified the torus action method in order to produce isospectral potentials and isospectral conformally equivalent Riemannian metrics. In this section, we observe that similar modifications of Theorem~\ref{thm:Sunada} and Theorem~\ref{thm:Torus_method} allow us to produce isospectral boundary density functions for the Steklov spectrum. Here we carry out the modification of Theorem~\ref{thm:Sunada}; the modification of Theorem~\ref{thm:Torus_method} is similar.
\begin{thm}\label{thm:SunadaDensity} Let $M$, $G$, $H$, $H'$, $g$ and $\rho$ satisfy the hypotheses of Theorem~\ref{thm:Sunada}. Assume in addition that there exists an isometry $\tau$ of $(M,g)$, not in $G$, such that $\tau H\tau^{-1}=H'$. Then for all $\alpha$ not in the Dirichlet spectrum of $(H\backslash M,g)$, we have
$$\operatorname{Stek}_\alpha(H\backslash M,g,\rho)=\operatorname{Stek}_\alpha(H\backslash M,g,\tau^*\rho)$$
where we continue to denote by $\rho$ and $\tau^*\rho$ the boundary density functions on $H\backslash M$ induced by those on $M$.
\end{thm}
\begin{proof} By Theorem~\ref{thm:Sunada}, $\operatorname{Stek}_\alpha(H\backslash M,g,\rho)=\operatorname{Stek}_\alpha(H'\backslash M,g,\rho)$ for all $\alpha$ not in the Dirichlet spectrum of $H\backslash M$. By the additional hypothesis of Theorem~\ref{thm:SunadaDensity}, $\tau$ induces an isometry $\tau: (H\backslash M,g)\to (H'\backslash M,g)$, so we have $\operatorname{Stek}_\alpha(H'\backslash M,\rho)=\operatorname{Stek}_\alpha(H\backslash M, \tau^*\rho)$.
\end{proof}
\begin{exa}[Flat surfaces and planar domains]
In Example 5.6 in \cite{GordonWebbWolpert1992}, the tile in Figure~\ref{fig:Basic_tile} is replaced by a tile $T$ that has not only a reflection symmetry $\beta_0$ as in \ref{exa.domains} but also a rotational symmetry $\tau_0$ that commutes with $\beta_0$. The tile is pictured in Figure 15
of \cite{GordonWebbWolpert1992}. Construct $M=M(G,S)$ and $M_i=M(H_i\backslash M,S)$, $i=1,2$ exactly as in \ref{exa.buser} but using the more symmetric tile. The isometry $\tau_0$ of the basic tile lifts to an isometry $\tau$ of $M$. The isometry $\tau$ normalizes the group $G$ and $\tau A\tau^{-1}=(A^t)^{-1}$ for all $A\in G$. In particular, $\tau H_1\tau^{-1}=H_2$. Define $\partial_0 T$ as in \ref{exa.buser} and let $\rho_0:\partial_0 T\to\mathbb{R}$ be a boundary density function that is \emph{not} invariant under the restriction to $\partial_0 T$ of the rotational symmetry $\tau_0$. Denote by $\rho$ the resulting boundary density on $M$. Then the hypotheses of Theorem~\ref{thm:SunadaDensity} are satisfied with $H_1$ and $H_2$ playing the roles of $H$ and $H'$. Thus we have $\operatorname{Stek}_\alpha(M_1,\rho)=\operatorname{Stek}_\alpha(M_1, \tau^*\rho)$ for all $\alpha$ not in the Dirichlet spectrum of $M_1$.
Next we construct planar domains. In the construction in the previous paragraph, impose the additional requirement that $\rho_0$ be invariant under the reflection symmetry $\beta_0$. As in \ref{exa.domains}, the symmetry $\beta_0$ of the new basic tile lifts to isometric involutions of $M$, $M_1$, and $M_2$. Let $\mathcal{O}_i$ be the orbifold quotient of $M_i$ by the involution $\beta_i$. As before, the underlying space of $\mathcal{O}_i$ is a planar domain $D_i$ whose boundary consists of the projection to $\mathcal{O}_i$ of the boundary of $M_i$ (this part is the boundary of the orbifold) together with the a collection of straight line segments corresponding to the singular set of the orbifold. The boundary density $\rho$ on $M_i$ projects to a density function, still denoted $\rho$, on the first part of the bounary of $D_i$; we extend $\rho$ to the full boundary by setting it to be zero on the orbifold singular set. Because $\beta_0$ and $\tau_0$ commute, the isometry $\tau:M_1\to M_2$ satisfies $\tau\circ\beta_1=\beta_2\circ\tau$, and thus $\tau$ induces an isometry between the planar domains $D_1$ to $D_2$. We then have $\operatorname{Stek}_\alpha(D_1,\rho)=\operatorname{Stek}_\alpha(D_1,\tau^*\rho)$ for all $\alpha$ not in the Dirichlet spectrum of $D_1$.
\end{exa}
The modification of the torus action method is similar.
\bibliographystyle{amsalpha}
|
\section{Introduction}\label{se:intro}
One of the definitions of the word \emph{morph} that can be found in English
dictionaries is ``to gradually change into a different image''. The
Graph Drawing community defines the morph of graph drawings similarly. Namely, given two drawings $\Gamma_0$ and
$\Gamma_1$ of a graph $G$, a \emph{morph} between $\Gamma_0$ and
$\Gamma_1$ is a continuously changing family of drawings of $G$ indexed by time
$t \in [ 0,1 ]$, such that the drawing at time $t=0$ is
$\Gamma_0$ and the drawing at time $t=1$ is $\Gamma_1$.
Further, the way the Graph Drawing community adopted the word morph is
consistent with its Ancient Greek root $\mu \omega \rho \phi \acute\eta$, which
means ``shape'' in a broad sense. Namely, if both $\Gamma_0$ and
$\Gamma_1$ have a certain geometric property, it is desirable that all
the drawings of the morph also have the same property. In particular, we talk about a \emph{planar},
a \emph{straight-line}, an \emph{orthogonal}, or a \emph{convex morph} if all the intermediate drawings of
the morph are \emph{planar} (edges do not cross), \emph{straight-line} (edges are straight-line segments), \emph{orthogonal} (edges are polygonal lines composed of horizontal and
vertical segments), or \emph{convex} (the drawings are planar and straight-line, and the faces are delimited by convex polygons), respectively.
The state of the art on planar morphs covers more than 100 years, starting from the 1914/1917 works of Tietze \cite{tietze-usaeqass-14} and Smith \cite{s-ocrsui-17}. The seminal papers of Cairns \cite{c-dprc-44} and Thomassen
\cite{t-dpg-83} proved the existence of a planar straight-line morph between any two topologically-equivalent planar straight-line drawings of a graph. In the last 10 years, the attention of the research community focused on
algorithms for constructing planar morphs with few \emph{morphing steps} (see, e.g., \cite{aadddhlrssw-cpwlv-11,DBLP:journals/siamcomp/AlamdariABCLBFH17,aacdfl-mpgdpns-13-c,Angelini-optimal-14,DBLP:conf/compgeom/AngeliniLFLPR15,afpr-mpgde-13,Barrera-unidirectional,biedl2013morphing,BLS-orth,r-mvdg-14,DBLP:conf/compgeom/GoethemV18}).
Each morphing step, sometimes simply called {\em step}, is a \emph{linear} morph, in which the vertices move along straight-line (possibly distinct) trajectories at
uniform speed. A \emph{unidirectional} morph is a linear morph in which the vertex trajectories are all parallel.
It is known~\cite{DBLP:journals/siamcomp/AlamdariABCLBFH17,Angelini-optimal-14} that a planar straight-line morph with a linear number of unidirectional morphing steps exists between any two topologically-equivalent planar straight-line drawings of the same graph, and that this bound is the best possible.
\emph{Upward planarity} is usually regarded as the natural extension of planarity to directed graphs; see, e.g.,~\cite{DBLP:journals/algorithmica/BertolazziBLM94,DBLP:journals/siamcomp/BertolazziBMT98,DETT,DBLP:journals/tcs/BattistaT88,Garg:2002:CCU:586839.586865}. A drawing of a directed graph is
\emph{upward planar} if it is planar and the edges are represented by curves
monotonically increasing in the vertical direction. Despite the importance
of upward planarity, up to now, no algorithm has been devised
to morph upward planar drawings of directed graphs. This paper deals
with the following question: Given two topologically-equivalent upward planar drawings $\Gamma_0$ and $\Gamma_1$ of an upward planar directed graph $G$, does an \emph{upward planar straight-line morph} between $\Gamma_0$ and $\Gamma_1$ always exist? In this paper we give a positive answer to this question.
Problems related to upward planar graphs are usually more difficult than the corresponding problems for undirected graphs. For example, planarity can be tested in linear time \cite{Hopcroft:1974:EPT:321850.321852} while testing upward planarity is NP-complete \cite{Garg:2002:CCU:586839.586865}; all planar graphs admit planar straight-line grid drawings with polynomial area~\cite{Schnyder:1990:EPG:320176.320191} while there are upward planar graphs that require exponential area in any upward planar straight-line grid drawing \cite{DiBattista:1992:ARS:149153.149159}. Quite surprisingly, we show that, from the morphing point of view, the difference between planarity and upward planarity is less sharp; indeed, in some cases, upward planar straight-line drawings can be morphed even more efficiently than planar straight-line drawings.
More in detail, our results are as follows. Let $\Gamma_0$ and $\Gamma_1$ be topologically-equivalent upward planar drawings of an $n$-vertex upward plane graph $G$. We show algorithms to construct upward planar straight-line morphs between $\Gamma_0$ and $\Gamma_1$ with the following number of unidirectional morphing steps:
\begin{enumerate}[\bf i.]
\item $O(1)$ steps if $G$ is a reduced plane $st$-graph;
\item $O(n)$ steps if $G$ is a plane $st$-graph;
\item $O(n)$ steps if $G$ is a reduced upward plane graph;
\item $O(n\cdot f(n))$ steps if $G$ is a general upward plane graph, assuming that an $O(f(n))$-step algorithm exists to construct an upward planar morph between any two upward planar drawings of any $n$-vertex plane $st$-graph. This, together with Result \textbf{ii.}, yields an $O(n^2)$-step upward planar morph for general upward plane graphs.
\end{enumerate}
Further, we show that there exist two topologically-equivalent upward planar drawings of an $n$-vertex upward plane path such that any upward planar morph between them consists of $\Omega(n)$ morphing steps.
In order to prove Result \textbf{i.} we devise a technique that allows us to construct a morph in which each morphing step modifies either only the $x$-coordinates or only the \mbox{$y$-coordinates} of the vertices.
Result \textbf{ii.} builds on the techniques in \cite{DBLP:journals/siamcomp/AlamdariABCLBFH17} and leverages on the arrangement of low-degree vertices in upward planar drawings in order to morph maximal plane $st$-graphs. We then exploit such morphs for general plane $st$-graphs.
In order to prove Results \textbf{iii.} and \textbf{iv.} we use an inductive
technique for reducing the geometric differences
between $\Gamma_0$ and $\Gamma_1$.
The paper is organized as follows. In \autoref{se:preliminaries} we introduce preliminary definitions and notation. In \autoref{se:slow-morphs-fast-morphs} we prove a lower bound on the number of morphing steps that might be required by an upward planar morph and we present a technique for constructing upward planar morphs with few morphing steps. In
\autoref{se:st-graphs} we study upward planar morphs of plane $st$-graphs. In \autoref{se:general-graphs} we study upward planar morphs of general upward plane graphs. Finally, in \autoref{se:conclusions} we present conclusions and open problems.
\section{Preliminaries}\label{se:preliminaries}
We assume familiarity with graph drawing~\cite{DETT} and related concepts.
\paragraph{Graph drawings.} In a drawing of a graph vertices are represented by distinct points of the plane and edges are represented by Jordan arcs connecting the points representing their end-vertices. In a \emph{straight-line drawing} the edges are represented by straight-line segments. In this paper we only consider straight-line drawings. Thus, where it leads to no confusion, we will omit the term ``straight-line''. Let $\Gamma$ be a drawing of a graph $G$ and let $H$ be a subgraph of $G$. We denote by $\Gamma[H]$ the restriction of $\Gamma$ to the vertices and edges of $H$.
\paragraph{Planar drawings, graphs, and embeddings.} A drawing of a graph is \emph{planar} if no two edges intersect. A graph is \emph{planar} if it admits a planar drawing. A planar drawing partitions the plane into topologically connected regions, called \emph{faces}. The unique unbounded face is the \emph{outer face}, whereas the remaining faces are the \emph{inner faces}. Two planar drawings of a connected graph are \emph{topologically equivalent} if they have the same circular order of the edges around each vertex and the same cycle bounding the outer face. A \emph{planar embedding} is an equivalence class of planar drawings. A \emph{plane graph} is a planar graph equipped with a planar embedding. In a planar straight-line drawing an internal face (the outer face) is \emph{strictly convex} if its angles are all smaller (greater) than $\pi$. A planar straight-line drawing is \emph{strictly convex} if each face is strictly convex.
A \emph{$y$-assignment} $y_G: V(G) \rightarrow \mathbb{R}$ is an assignment of
reals to the vertices of a graph $G$.
A drawing $\Gamma$ of $G$ \emph{satisfies} $y_G$ if the $y$-coordinate in $\Gamma$ of each vertex $v \in V(G)$ is $y_G(v)$.
An \emph{$x$-assignment} $x_G$ for the vertices of $G$ is defined analogously.
\paragraph{Connectivity.} A {\em $k$-cut} in a connected graph $G$ is a set of $k$ vertices whose removal disconnects $G$. A graph is {\em $k$-connected} if it does not contain any $(k-1)$-cut; $2$-connected and $3$-connected graphs are also called {\em biconnected} and {\em triconnected} graphs, respectively. The maximal biconnected subgraphs of a graph are called {\em blocks}. A biconnected plane graph $G$ is \emph{internally $3$-connected} if, for every $2$-cut $\{u,v\}$, $u$ and $v$ are incident to the outer face of $G$ and each connected component of the graph $G - \{u,v\}$ contains a vertex incident to the outer face of $G$.
\paragraph{Directed graphs.}
In a directed graph $G$ we denote by $uv$ an edge directed from a vertex $u$ to a vertex $v$; then $v$ is a \emph{successor} of $u$, and $u$ is a \emph{predecessor} of $v$. A {\em source} is a vertex with no incoming edge; a {\em sink} is a vertex with no outgoing edge. A {\em directed path} consists of the edges $u_iu_{i+1}$, for $i=1,\dots,n-1$. A {\em directed cycle} consists of the edges $u_iu_{i+1}$, for $i=1,\dots,n$, where $u_{n+1}=u_1$. A graph without directed cycles is {\em acyclic}. A {\em transitive edge} in a directed graph $G$ is an edge $uv$ such that $G$ contains a directed path from $u$ to $v$ different from the edge $uv$. A \emph{reduced} graph is a directed graph that does not contain any transitive edges. Whenever we do not know or are not interested in the orientation of an edge connecting two vertices $u$ and $v$, we denote it by $(u,v)$. The \emph{underlying graph} of a directed graph $G$ is the undirected graph obtained from $G$ by omitting the directions from its edges. When talking about the connectivity of a directed graph we always refer to the connectivity of its underlying graph. A \emph{topological ordering} of an $n$-vertex acyclic graph $G=(V,E)$ is a numbering $\pi: V \rightarrow \{1,2,\dots,n\}$ of the vertices of $G$ such that $\pi(u) < \pi(v)$, for each edge $uv \in E$.
\paragraph{Upward planar drawings, embeddings, and morphs.}
A drawing of a directed graph is \emph{upward planar} if it is planar and each edge $uv$ is drawn as a curve monotonically increasing in the $y$-direction from $u$ to $v$. A directed graph is \emph{upward planar} if it admits an upward planar drawing.
Consider an upward planar drawing $\Gamma$ of a directed graph $G$ and consider a vertex $v$. The list $\mathcal S(v)=[w_1,\dots,w_k]$ contains the successors of $v$ in ``left-to-right order''. That is, consider a half-line $\ell$ starting at $v$ and directed leftwards; rotate $\ell$ around $v$ in clockwise direction and append a vertex $w_i$ to $\mathcal S(v)$ when $\ell$ overlaps with the tangent to the edge $(v,w_i)$. The list $\mathcal P(v)=[z_1,\dots,z_l]$ of the predecessors of $v$ is defined similarly. Then two upward planar drawings of a connected directed graph are \emph{topologically equivalent} if they have the same lists $\mathcal S(v)$ and $\mathcal P(v)$ for each vertex $v$. An \emph{upward planar embedding} is an equivalence class of upward planar drawings. An {\em upward plane graph} is an upward planar graph equipped with an upward planar embedding. If a vertex $v$ in an upward planar graph $G$ is not a source or a sink, then a planar embedding of $G$ determines $\mathcal S(v)$ and $\mathcal P(v)$. However, if $v$ is a source or a sink, then different upward planar drawings might have different lists $\mathcal S(v)$ or $\mathcal P(v)$, respectively. In fact, two upward planar drawings of an upward planar graph $G$ might not have the same upward planar embedding although the underlying graph of $G$ has the same planar embedding in the two drawings; see, for example, \blue{Fig.~\ref{fig:planar-upward-embedding}}.
\begin{figure}[tb]
\centering
\subfloat[]
{\includegraphics[page=1,width=.28\columnwidth]{Embedding1.pdf}\label{fig:planar-upward-embedding-0}}
\hfil
\subfloat[]
{\includegraphics[page=1,width=.28\columnwidth]{Embedding2.pdf}\label{fig:planar-upward-embedding-a}}
\caption{Two upward planar drawings of the same directed graph $G$ (whose underlying graph is a simple cycle) with the same planar embedding but with different upward planar embeddings. The angles labeled \texttt{large} are gray. Observe that $\mathcal S(v_8)=[v_1,v_7]$ in (a), while $\mathcal S(v_8)=[v_7,v_1]$ in (b).}\label{fig:planar-upward-embedding}
\end{figure}
For biconnected upward planar graphs a different, and yet equivalent, notion of upward planar embedding exists; this is described in the following. Consider an upward planar drawing $\Gamma$ of a biconnected upward planar graph $G$.
Let $u, v$, and $w$ be three distinct vertices that appear consecutively and in this clockwise order along the boundary of a face $f$ of $G$; note that, since $G$ is biconnected $f$ is delimited by a simple cycle. We denote by $\angle (u,v,w)$ the angle formed by the tangents to the edges $(u,v)$ and $(v,w)$ at $v$ in the interior of $f$. We say that $v$ is a \emph{sink-switch} (a \emph{source-switch}) of $f$ if the orientations of the edges $(u,v)$ and $(v,w)$ in $G$ are $uv$ and $wv$ ($vu$ and $vw$, respectively).
Further, we say that $v$ is a \emph{switch} of $f$ if it is either a sink-switch or a source-switch of $f$, and $v$ is a \emph{switch} of $G$ if it is a switch of some face of $G$. Two switches $u$ and $v$ of a face $f$ are \emph{clockwise} (\emph{counter-clockwise}) \emph{consecutive} if traversing $f$ clockwise (counter-clockwise) no switch is encountered in between $u$ and $v$.
The drawing $\Gamma$ determines a \emph{large-angle assignment}, that is, a labeling, for each face $f$ and each three clockwise consecutive switches $u$, $v$, and $w$ for $f$ of the corresponding angle $\angle (u,v,w)$ as \texttt{large}, if it is larger than $\pi$ in $\Gamma$, or \texttt{small}, it is smaller than $\pi$ in~$\Gamma$~\cite{DBLP:journals/algorithmica/BertolazziBLM94}.
Two upward planar drawings of an upward planar graph $G$ are then say to be {\em topologically equivalent} if they have the same planar embedding and the same large-angle assignment. From this notion of topological equivalence, the ones of upward planar embedding and upward plane graph can be introduced as before; again, the formerly introduced notions coincide with the just introduced ones for upward planar graphs with biconnected underlying graphs (in fact, this correspondence between the two notions could be stated for all upward planar graphs, however the definition of clockwise consecutive switches we introduced is ambiguous for upward planar graphs whose underlying graph is not biconnected). A combinatorial characterization of the \texttt{large}-\texttt{small} assignments that correspond to upward planar embeddings is given in~\cite{DBLP:journals/algorithmica/BertolazziBLM94}.
Whenever we talk about an upward planar drawing of an upward plane graph $G$, we always assume, even when not explicitly stated, that the drawing respects the upward planar embedding associated to $G$. Further, whenever we talk about a subgraph $H$ of an upward plane graph $G$, we always assume, even when not explicitly stated, that $H$ is associated with the upward planar embedding obtained from the one associated to $G$ by removing vertices and edges not in $H$.
Let $\Gamma_0$ and $\Gamma_1$ be two topologically-equivalent upward planar drawings of an upward plane graph $G$. An \emph{upward planar morph} is a continuous transformation from $\Gamma_0$ to $\Gamma_1$ indexed by time $t \in [0,1]$ in which the drawing $\Gamma_t$ at each time $t \in [0,1]$ is upward planar. We remark that each drawing $\Gamma_t$ has to respect the upward planar embedding associated to $G$; in particular, the drawings $\Gamma_0$ and $\Gamma_1$ determine the same upward planar embedding (if this were not the case, then a morph that preserves upward planarity at all times would not exist).
\paragraph{Plane $st$-graphs.}
A \emph{plane $st$-graph} is an upward plane graph with a single source $s$ and a single sink $t$, and with an upward planar embedding in which $s$ and $t$ are incident to the outer face. A plane $st$-graph always admits an upward planar straight-line drawing~\cite{DBLP:journals/tcs/BattistaT88}. A cycle in an upward plane graph is an \emph{$st$-cycle} if it consists of two directed paths. A face $f$ of an upward plane graph is an \emph{$st$-face} if it is delimited by an $st$-cycle; the directed paths delimiting an $st$-face $f$ are called \emph{left} and \emph{right boundary}, where the edge of the left boundary incident to the source-switch $s_f$ of $f$ immediately precedes the edge of the right boundary incident to $s_f$ in the clockwise order of the edges incident to $s_f$. The following is well-known.
\begin{lemma}\label{le:st-faces-iff-st-graph}
An upward plane graph is a plane $st$-graph iff all its faces are $st$-faces.
\end{lemma}
An internal vertex $v$ of a maximal plane $st$-graph $G$ is \emph{simple} if the neighbors of $v$ induce a cycle in the underlying graph of $G$.
\begin{lemma}[Alamdari et al.~\cite{DBLP:journals/siamcomp/AlamdariABCLBFH17}]\label{le:internal-vertex}
Any maximal plane $st$-graph contains a simple vertex of degree at most $5$.
\end{lemma}
\section{Slow Morphs and Fast Morphs}\label{se:slow-morphs-fast-morphs}
We start this section by proving the following lower bound.
\begin{figure}[tb]
\centering
\subfloat[$P$]
{\includegraphics[page=4,width=.28\columnwidth]{lower_bound.pdf}\label{fig:lower-bound-0}}
\hfil
\subfloat[$\Gamma_0$]
{\includegraphics[page=2,height=.28\columnwidth]{lower_bound.pdf}\label{fig:lower-bound-a}}
\hfil
\subfloat[$\Gamma_1$]
{\includegraphics[page=1,height=.28\columnwidth]{lower_bound.pdf}\label{fig:lower-bound-b}}
\caption{Illustration for \autoref{th:lower-bound}. (a) $P$; (b) $\Gamma_0$; and (c) $\Gamma_1$. For the sake of readability $\Gamma_0$ and $\Gamma_1$ have curved edges. However, the $x$-coordinates of the vertices can be slightly perturbed in order to make $\Gamma_0$ and $\Gamma_1$ straight-line.}\label{fig:lower-bound}
\end{figure}
\newcommand{There are two upward planar drawings of an $n$-vertex upward plane path such that any upward planar morph between them consists of $\Omega(n)$ steps.}{There are two upward planar drawings of an $n$-vertex upward plane path such that any upward planar morph between them consists of $\Omega(n)$ steps.}
\begin{theorem}\label{th:lower-bound}
There are two upward planar drawings of an $n$-vertex upward plane path such that any upward planar morph between them consists of $\Omega(n)$ steps.
\end{theorem}
\begin{proof}
Assume, for the sake of simplicity, that $n$ is even, and let $n=2k$. Consider the $n$-vertex upward plane path $P$ defined as follows (refer to \blue{Fig.~\ref{fig:lower-bound-0}}).
The path $P$ contains vertices $u_i$ and $v_i$, for $i=1,\dots,k$, and directed edges $u_iv_i$, for $i=1,\dots,k$, and $u_{i+1}v_i$, for $i=1,\dots,k-1$. Clearly, $P$ has a unique planar embedding $\mathcal E$; we fix the upward planar embedding of $P$ so that $\mathcal S(u_i)=[v_i,v_{i-1}]$, for $i=2,\dots,k$, and so that $\mathcal P(v_i)=[u_i,u_{i+1}]$, for $i=1,\dots,k-1$.
Let $\Gamma_0$ and $\Gamma_1$ be two upward planar straight-line drawings of $P$ in which the bottom-to-top order of the vertices is $u_1, \dots,u_k,v_k,\dots,v_1$ (see \blue{Fig.~\ref{fig:lower-bound-a}}) and $u_k, \dots,u_1,v_1,\dots,v_k$ (see \blue{Fig.~\ref{fig:lower-bound-b}}), respectively. Note that, by the upward planarity of $\Gamma_0$, the edge $u_iv_i$ has the edge $u_{i+1}v_{i+1}$ to its right in $\Gamma_0$, for $i=1,\dots,k-1$, and the edge $u_{i+1}v_i$ has the edge $u_{i+2}v_{i+1}$ to its left in $\Gamma_0$, for $i=1,\dots,k-2$. Let $\langle \Gamma_0 = \Lambda_1,\Lambda_2,\dots,\Lambda_{h+1}=\Gamma_1 \rangle$ be any upward planar morph from $\Gamma_0$ to $\Gamma_1$ that consists of $h$ morphing steps. We have the following.
\begin{claim}\label{claim:lower}
For each $j=1,2,\dots,\min\{h+1,k-1\}$, we have that:
\begin{enumerate} [(a)]
\item the vertices $u_{j},u_{j+1},\dots,u_{k-1},u_k,v_k,v_{k-1},\dots,v_{j+1},v_j$ appear in this bottom-to-top order in $\Lambda_{j}$;
\item for $i=j,\dots,k-1$, the edge $u_iv_i$ has the edge $u_{i+1}v_{i+1}$ to its right; and
\item for $i=j,\dots,k-2$, the edge $u_{i+1}v_i$ has the edge $u_{i+2}v_{i+1}$ to its left.
\end{enumerate}
\end{claim}
\begin{claimproof}
We prove the statement by induction on $j$. The statement is trivial for $j=1$, by the definition of $\Gamma_0 = \Lambda_1$.
Consider now any $j>1$. By induction, $\Lambda_{j-1}$ satisfies Properties~(a)--(c).
Suppose, for a contradiction, that there exists an index $i\in\{j,j+1,\dots,k-1\}$ such that $u_{i+1}$ lies below $u_i$ in $\Lambda_{j}$. The upward planarity of $\Lambda_{j-1}$ and $\Lambda_{j}$ implies that $v_i$ and $v_{i+1}$ both lie above $u_i$, both in $\Lambda_{j-1}$ and $\Lambda_{j}$. Further, $u_{i+1}$ lies below $v_i$ and $v_{i+1}$, both in $\Lambda_{j-1}$ and $\Lambda_{j}$; this comes from Property~(a) of $\Lambda_{j-1}$ and from the assumption that $u_{i+1}$ lies below $u_i$ in $\Lambda_{j}$. Then $u_{i+1}$ lies below the horizontal line through the lowest of $v_i$ and $v_{i+1}$ throughout the linear morph $\langle \Lambda_{j-1},\Lambda_j\rangle$. By Properties~(b) and~(c) of $\Lambda_{j-1}$, the vertex $u_{i+1}$ lies in $\Lambda_{j-1}$ inside the bounded region of the plane delimited by the edge $u_iv_i$, by the edge $u_iv_{i+1}$, and by the horizontal line through the lowest of $v_i$ and $v_{i+1}$. However, by the assumption that $u_{i+1}$ lies below $u_i$ in $\Lambda_{j}$, we have that $u_{i+1}$ lies outside the same region in $\Lambda_{j}$. Since $u_{i+1}$ does not cross the horizontal line through the lowest of $v_i$ and $v_{i+1}$ throughout the linear morph $\langle \Lambda_{j-1},\Lambda_j\rangle$, it follows that $u_{i+1}$ crosses $u_iv_i$ or $u_iv_{i+1}$ during $\langle \Lambda_{j-1},\Lambda_j\rangle$, a contradiction.
An analogous proof shows that $v_{i+1}$ lies below $v_i$ in $\Lambda_{j}$, for $i=j,j+1,\dots,k-1$. Property~(a) for $\Lambda_{j}$ follows. Properties~(b) and~(c) follow by Property~(a) and by the upward planarity of $\Lambda_{j}$. This concludes the proof of the claim.
\end{claimproof}
By \autoref{claim:lower} and since $u_k,u_{k-1}$ appear in this bottom-to-top order in $\Gamma_1=\Lambda_{h+1}$, we have that $h+1>k-1$, hence $h \in \Omega(n)$.
\end{proof}
\remove{We set as \texttt{large} the angles $\angle (u_{i},v_i,u_{i+1})$ and $\angle(v_i,u_{i+1},v_{i+1})$, for $i=1,\dots,k-1$.}
We now establish a tool that will allow us to design efficient algorithms for morphing upward planar drawings. Consider two planar straight-line drawings $\Gamma'$ and $\Gamma''$ of a plane graph $G$ with the same $y$-assignment. Since the drawings are straight-line and have the same $y$-assignment, a horizontal line $\ell$ intersects a vertex or an edge of $G$ in $\Gamma'$ if and only if it intersects the same vertex or edge in $\Gamma''$. We say that $\Gamma'$ and $\Gamma''$ are {\em left-to-right equivalent} if, for any horizontal line $\ell$, for any vertex or edge $\alpha$ of $G$, and for any vertex or edge $\beta$ of $G$ such that $\ell$ intersects both $\alpha$ and $\beta$ (in $\Gamma'$ and in $\Gamma''$), we have that the intersection of $\alpha$ with $\ell$ is to the left of the intersection of $\beta$ with $\ell$ in $\Gamma'$ if and only if the intersection of $\alpha$ with $\ell$ is to the left of the intersection of $\beta$ with $\ell$ in $\Gamma''$. The definition of {\em bottom-to-top equivalent} drawings is analogous. We have the following.
\begin{lemma}\label{le:st-y-assignment-equivalent}
Any two upward planar drawings $\Gamma'$ and $\Gamma''$ of a plane $st$-graph $G$ with the same $y$-assignment are left-to-right equivalent.
\end{lemma}
\begin{proof}
Since $G$ is a plane $st$-graph, the drawings $\Gamma'$ and $\Gamma''$ have the same faces. By \autoref{le:st-faces-iff-st-graph} such faces are $st$-faces. Also, every horizontal line $\ell$ crosses an $st$-face $f$ at most twice, and the left-to-right order of these crossings along $\ell$ is the same in $\Gamma'$ and $\Gamma''$ because the left and right boundaries of $f$ are the same in $\Gamma'$ and $\Gamma''$, given that $\Gamma'$ and $\Gamma''$ are topologically equivalent.
\end{proof}
\autoref{le:sidedness} is due to \cite{DBLP:journals/siamcomp/AlamdariABCLBFH17}. We extend it in \autoref{le:shift}.
\begin{lemma}[Alamdari et al.~\cite{DBLP:journals/siamcomp/AlamdariABCLBFH17}, Corollary 7.2]\label{le:sidedness}
Consider a unidirectional morph acting on points $p$, $q$, and $r$. If
$p$ is on one side of the oriented line through $\overline{qr}$ at the beginning and at the end of the morph, then $p$ is on the same side of the oriented line through $\overline{qr}$ throughout the morph.
\end{lemma}
\begin{lemma}\label{le:shift}
Let $\Gamma'$ and $\Gamma''$ be two left-to-right or bottom-to-top equivalent planar drawings of a plane graph. Then the linear morph $\mathcal M$ from $\Gamma'$ to~$\Gamma''$ is unidirectional and planar.
\end{lemma}
\begin{proof}
Since $\Gamma'$ and $\Gamma''$ have the same $y$-assignment ($x$-assignment), given that they are left-to-right (bottom-to-top) equivalent, it follows that all the vertices move along horizontal (vertical) trajectories. Thus, $\mathcal M$ is unidirectional.
Also, since $\Gamma'$ and $\Gamma''$ are left-to-right (bottom-to-top) equivalent, each horizontal (vertical) line crosses the \mbox{same sequence} of vertices and edges in both $\Gamma'$ and $\Gamma''$. Thus, by \autoref{le:sidedness}, $\mathcal M$ is planar.
\end{proof}
\autoref{le:shift} allows us to devise a simple morphing technique between any two upward planar drawings $\Gamma_0$ and $\Gamma_1$ of the same upward plane graph $G$, when a pair of upward planar drawings of $G$ with special properties can be computed. We say that the pair $(\Gamma_0,\Gamma_1)$ is an \emph{hvh-pair} if there exist upward planar drawings
$\Gamma'_0$ and $\Gamma'_1$ of $G$ such that:
\begin{inparaenum}[(i)]
\item \label{pr:one} $\Gamma_0$ and $\Gamma'_0$ are left-to-right equivalent,
\item \label{pr:two} $\Gamma'_0$ and $\Gamma'_1$ are bottom-to-top equivalent, and
\item \label{pr:three} $\Gamma'_1$ and $\Gamma_1$ are left-to-right equivalent.
\end{inparaenum}
Our morphing tool is expressed by the following lemma.
\newcommand{\fastmorphstatement}{Let $(\Gamma_0,\Gamma_1)$ be an hvh-pair of upward planar drawings of an upward plane graph $G$.
There is a $3$-step \mbox{upward planar morph from $\Gamma_0$ to $\Gamma_1$}.}
\begin{lemma}[Fast morph]\label{le:fast-morph}
\fastmorphstatement
\end{lemma}
\begin{proof}
By hypothesis there exist drawings $\Gamma'_0$ and $\Gamma'_1$ of $G$ satisfying Conditions~(\ref{pr:one}),~(\ref{pr:two}), and~(\ref{pr:three}) of the definition of an hvh-pair.
By \autoref{le:shift}, $\mathcal M_1 = \langle \Gamma_0, \Gamma'_0 \rangle$,
$\mathcal M_2 = \langle \Gamma'_0, \Gamma'_1 \rangle$, and $\mathcal M_3 = \langle \Gamma'_1, \Gamma_1 \rangle$ are planar linear morphs.
Therefore, $\mathcal M = \langle \Gamma_0, \Gamma'_0, \Gamma'_1, \Gamma_1 \rangle$ is a $3$-step planar morph from $\Gamma_0$ to $\Gamma_1$. In order to prove that $\mathcal M$ is an upward planar morph, we need to show that each linear morph $\mathcal M_i$ is an upward planar morph. To this aim, we only need to prove that no edge changes its orientation during $\mathcal M_i$, for $i=1,2,3$. This is trivially true for
$\mathcal M_1$ (for $\mathcal M_3$) since $\Gamma_0$ and $\Gamma'_0$ ($\Gamma_1$ and $\Gamma'_1$) induce the same $y$-assignment.
We now prove that no directed edge $uv$ changes its orientation during $\mathcal M_2$.
By Condition~(\ref{pr:two}) of the definition of an hvh-pair, the $x$-coordinate of $u$ is the same in $\Gamma'_0$ and in $\Gamma'_1$, hence it is the same throughout $\mathcal M_2$. Denote by $x'$ such $x$-coordinate. The $y$-coordinate of $u$ might be different in $\Gamma'_0$ and in $\Gamma'_1$; denote by $y'_0$ and $y'_1$ such coordinates, respectively. Consider a point $r$ that moves (at uniform speed along a straight-line trajectory) during $\mathcal M_2$ from $(x'+1,y'_0)$ in $\Gamma_0$ to $(x'+1,y'_1)$ in $\Gamma_1$. Note that $r$ moves along a vertical trajectory, hence the movement of $r$ and $\mathcal M_2$ define a unidirectional morph. Also observe that the straight-line segment $\overline{ur}$ is horizontal throughout $\mathcal M_2$; further, $v$ is above the horizontal line through $\overline{ur}$ both in $\Gamma'_0$ and in $\Gamma'_1$, by the upward planarity of $\Gamma_0$ and $\Gamma_1$ and by Conditions~(\ref{pr:one}) and~(\ref{pr:three}) of the definition of an hvh-pair. By Lemma~\ref{le:sidedness} with $p=v$, $q=u$, and $r=r$ we have that the $y$-coordinate of $v$ is greater than the $y$-coordinate of $u$ throughout $\mathcal M_2$. Hence, $\mathcal M_2$ is an upward planar morph.
\end{proof}
\newcommand{ Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex upward plane graph $G$ whose underlying graph is connected. There exist upward planar drawings $\Gamma'_0$ and $\Gamma'_1$ of an $O(n)$-vertex upward plane graph $G'$ that is a supergraph of $G$, whose underlying graph is biconnected, and such that $\Gamma'_0[G]=\Gamma_0$ and $\Gamma'_1[G]=\Gamma_1$. Further, if $G$ is reduced or an $st$-graph, then so is $ G'$.}{ Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex upward plane graph $G$ whose underlying graph is connected. There exist upward planar drawings $\Gamma'_0$ and $\Gamma'_1$ of an $O(n)$-vertex upward plane graph $G'$ that is a supergraph of $G$, whose underlying graph is biconnected, and such that $\Gamma'_0[G]=\Gamma_0$ and $\Gamma'_1[G]=\Gamma_1$. Further, if $G$ is reduced or an $st$-graph, then so is $ G'$.}
The next lemma will allow us to restrict our attention to biconnected graphs.
\begin{lemma}\label{le:biconnected-reduced-augmentation}
Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex upward plane graph $G$ whose underlying graph is connected. There exist upward planar drawings $\Gamma'_0$ and $\Gamma'_1$ of an $O(n)$-vertex upward plane graph $G'$ that is a supergraph of $G$, whose underlying graph is biconnected, and such that $\Gamma'_0[G]=\Gamma_0$ and $\Gamma'_1[G]=\Gamma_1$. Further, if $G$ is reduced or an $st$-graph, then so is $ G'$.
\end{lemma}
\begin{proof}
Initialize $G'=G$, $\Gamma'_0=\Gamma_0$, and $\Gamma'_1=\Gamma_1$. Consider a cutvertex $v$ of~$G'$.
Let $u$ and $w$ be two neighbors of $v$ belonging to different blocks of $G'$ that are consecutive in the circular order of the neighbors of $v$. By relabeling $u$ and $w$, we may assume that one of the following holds true:
\begin{itemize}
\item if $u$ and $w$ are both successors of $v$, then the edge $vw$ immediately follows the edge $vu$ in the clockwise order of the edges incident to $v$ (see \blue{Fig.}\autoref{fig:bico1});
\item if $u$ and $w$ are both predecessors of $v$, then the edge $wv$ immediately precedes the edge $uv$ in the clockwise order of the edges incident to $v$ (see \blue{Fig.}\autoref{fig:bico2}); or
\item $u$ is a successor of $v$, $w$ is a predecessor of $v$, and the edge $wv$ immediately follows the edge $vu$ in the clockwise order of the edges incident to $v$ (see \blue{Fig.}\autoref{fig:bico3}).
\end{itemize}
\begin{figure}[tb]
\centering
\subfloat[]{\includegraphics[scale=1, page=1]{biconnection}\label{fig:bico1}}
\subfloat[]{\includegraphics[scale=1, page=2]{biconnection}\label{fig:bico2}}
\subfloat[]{\includegraphics[scale=1, page=3]{biconnection}\label{fig:bico3}}
\caption{Illustration for \autoref{le:biconnected-reduced-augmentation}. The vertices $u$ and $w$ belong to two blocks $\beta_u$ and $\beta_w$, respectively, both containing the cut-vertex $v$. \label{fig:bico}}
\end{figure}
Denote by $f$ the face that is to the right of the edge $(u,v)$ when traversing such an edge according to its orientation. Note that the edge $(v,w)$ is also incident to $f$. We add to $G'$ a new vertex $v'$ inside $f$; further, we add to $G'$ two directed edges connecting $v'$ with $u$ and $w$ inside $f$. These edges are directed as the edges connecting $v$ with $u$ and $w$, respectively; that is, we add to $G'$ either the directed edge $uv'$, if $uv \in E(G')$, or the directed edge $v'u$, if $vu \in E(G')$, and either the directed edge $wv'$, if $wv \in E(G')$, or the directed edge $v'w$, if $vw \in E(G')$.
The described augmentation does not introduce any transitive edges. Further, no edge that was already in $G'$ before the augmentation becomes transitive after the augmentation; this is because the edge $(u,w)$ does not belong to $G'$, as $u$ and $w$ belong to distinct blocks of $G'$; hence $G'$ remains reduced if it was so.
In the case illustrated in \blue{Fig.}\autoref{fig:bico1} (in \blue{Fig.}\autoref{fig:bico2}), each of the blocks $\beta_u$ and $\beta_w$ of $G'$ containing $u$ and $w$ before the augmentation contains a distinct sink of $G'$ (resp.\ a distinct source of $G'$), hence $G'$ is not an $st$-graph before the augmentation. In the case illustrated in \blue{Fig.}\autoref{fig:bico3}, it might be that $G'$ is an $st$-graph before the augmentation. Note that there only two faces of the augmented graph $G'$ that do not belong to $G'$ before the augmentation. One of them is delimited by the directed paths $wvu$ and $wv'u$, hence it is an $st$-face; the other one is obtained from $f$ by replacing the directed path $wvu$ with the directed path $wv'u$, hence it is an $st$-face as long as $f$ is. It follows that, if $G'$ is an $st$-graph before the augmentation, then it remains an $st$-graph after the augmentation.
We now describe how to insert $v'$ and its incident edges into $\Gamma'_0$ and $\Gamma'_1$. By standard continuity arguments, like the ones used in the proof of F\'ary's theorem~\cite{Fary}, we have that, for $i=0,1$, there exists a sufficiently small value $\epsilon_i > 0$ such that the disk $d_i$ with radius $\epsilon_i$ centered at $v$ in $\Gamma'_i$ contains no vertex other than $v$ and is not traversed by any edge other than those incident to $v$. We place $v'$ at distance $\epsilon < \epsilon_0,\epsilon_1$ from $v$ inside $f$, as illustrated in \blue{Figs.}\autoref{fig:bico1}--\autoref{fig:bico3}; in particular, $v'$ is placed in the circular sector of $d_i$ delimited by $(u,v)$ and $(w,v)$. By selecting a sufficiently small value for $\epsilon$, the edges $(u,v')$ and $(w,v')$ can be drawn as straight-line segments that do not intersect any edge of $G'$. Further, if $\epsilon$ is sufficiently small, then the $y$-coordinate of $u$ (the $y$-coordinate of $w$) is smaller than the one of $v'$ if and only if it is smaller than the one of $v$, hence the straight-line segments representing the edges $(u,v')$ and $(w,v')$ monotonically increase in the $y$-direction from their sources to their sinks. The upward planarity of the drawings $\Gamma'_0$ and $\Gamma'_1$ of the augmented graph $G'$ follows. Note that after the augmentation we have $\Gamma'_0[G]= \Gamma_0$ and $\Gamma'_1[G]= \Gamma_0$. This is because the same equalities were satisfied before the augmentation and since the drawings of $G'$ before the augmentation were not altered during the augmentation.
Since the graph $G'$ after the augmentation contains one block less than before the augmentation, the repetition of this argument results in a biconnected graph $G'$. This concludes the proof of the lemma.
\end{proof}
\section{Plane $st$-Graphs}\label{se:st-graphs}
In this section, we show algorithms for constructing upward planar morphs between upward planar drawings of plane $st$-graphs.
\subsection{Reduced Plane $st$-Graphs}\label{sse:reduced}
We first consider plane $st$-graphs without transitive edges. We have the following.
\newcommand{Any two upward planar drawings $\Gamma_0$ and $\Gamma_1$ of a reduced plane $st$-graph $G$ form an hvh-pair.}{Any two upward planar drawings $\Gamma_0$ and $\Gamma_1$ of a reduced plane $st$-graph $G$ form an hvh-pair.}
\begin{lemma}\label{le:good-reduced}
Any two upward planar drawings $\Gamma_0$ and $\Gamma_1$ of a reduced plane $st$-graph $G$ form an hvh-pair.
\end{lemma}
\begin{proof}
By \autoref{le:biconnected-reduced-augmentation} we can assume that the reduced plane $st$-graph $G$ is biconnected. Let $\Gamma_0$ and $\Gamma_1$ be any two upward planar drawings of $G$. We show that $\Gamma_0$ and $\Gamma_1$ form an hvh-pair by exhibiting two upward planar drawings $\Gamma'_0$ and $\Gamma'_1$ of $G$ that satisfy Conditions~(\ref{pr:one}),~(\ref{pr:two}), and~(\ref{pr:three}) of the definition of an hvh-pair.
We construct drawings $\Gamma'_0$ and $\Gamma'_1$ as follows (refer to \blue{Figs.}~\ref{fig:ear-decomposition} and \ref{fig:ear-inv}).
%
Consider the {\em weak dual} multi-graph $D$ of $G$, which is defined as follows. The multi-graph $D$ has a vertex $v_f$ for each internal face $f$ of $G$ and a directed edge $v_f v_g$ if the faces $f$ and $g$ of $G$ share an edge $e$ in $G$ and $f$ lies to the left of $g$ when traversing $e$ according to its orientation. The concept of weak dual multi-graph has been used, e.g., in~\cite{RT-rpl-86,TT86-avrpg-86,TT-uavr-86}. Observe that $D$ is acyclic~\cite{TT86-avrpg-86}. We now present a structural decomposition of $G$ guided by $D$ which has been used, e.g., in~\cite{fgw-nupo-12,Mel}.
%
Let $\mathcal{T}=\{v_1, \dots, v_k\}$ be a topological ordering of the vertices of $D$ and let $P_0$ be the left boundary of the outer face of $G$. The ordering $\mathcal{T}$ defines a sequence $P_1$, $P_2$, \dots, $P_k$ of directed paths such that, for each $i=1,\dots,k$, the path $P_j$ is the right boundary of the face of $G$ corresponding to the vertex $v_j$ of~$D$. For $j=1,\dots,k$, the graph $G_j=\bigcup_{i=0}^j P_i$ is a plane $st$-graph which is obtained by attaching the directed path $P_{j}$ to two non-adjacent vertices on the right boundary of the outer face of $G_{j-1}$; further, $G_k=G$. Note that, since $G$ is a reduced plane $st$-graph, no path $P_j$ consists of a single edge.
The drawings $\Gamma'_0$ and $\Gamma'_1$ are simultaneously and iteratively constructed by adding, for $j=1,\dots,k$, the path $P_j$ to the already constructed drawings $\Gamma'_0$ and $\Gamma'_1$ of $G_{j-1}$. Note that, after $P_j$ has been drawn in $\Gamma'_0$ and $\Gamma'_1$, the right boundary of the outer face of $G_{j}$ is a directed path, hence it is represented by a $y$-monotone curve in both $\Gamma'_0$ and $\Gamma'_1$.
\begin{figure}[tb]
\centering
\subfloat[]{\includegraphics[scale=1, page=1]{ear-decomposition}\label{fig:ear-deco-p0}}
\hfil
\subfloat[]{\includegraphics[scale=1, page=2]{ear-decomposition}\label{fig:ear-deco-p1}}
\caption{Illustration for the proof of \autoref{le:good-reduced}. \protect\subref{fig:ear-deco-p0} The drawing $\Gamma'_i$ of $G_0=P_0$. \protect\subref{fig:ear-deco-p1} The drawing $\Gamma'_i$ of $G_1$.
\label{fig:ear-decomposition}
}
\end{figure}
\begin{figure}[tb]
\centering
\subfloat[]{\includegraphics[scale=1, page=1]{ear-decomposition-invariant}\label{fig:ear-inv-1}}
\subfloat[]{\includegraphics[scale=1, page=2]{ear-decomposition-invariant}\label{fig:ear-inv-2}}
\\
\subfloat[]{\includegraphics[scale=1, page=3]{ear-decomposition-invariant}\label{fig:ear-inv-3}}
\subfloat[]{\includegraphics[scale=1, page=4]{ear-decomposition-invariant}\label{fig:ear-inv-4}}
\caption{Illustration for the proof of \autoref{le:good-reduced}.
Computation of the points $a_i$ and $b_i$ in $\Gamma'_i$ for the cases $\ell=2$ \protect\subref{fig:ear-inv-1} and $\ell>2$ \protect\subref{fig:ear-inv-3}, respectively. Drawing of the path $P_j$ for the cases $\ell=2$ \protect\subref{fig:ear-inv-2} and $\ell>2$ \protect\subref{fig:ear-inv-4}.
\label{fig:ear-inv}
}
\end{figure}
For $i=0,1$, we denote by $y_i(v)$ the $y$-coordinate of a vertex $v$ in $\Gamma_i$.
We obtain drawings $\Gamma'_0$ and $\Gamma'_1$ of $G_0 = P_0$ by placing its vertices along the line $x=0$ at the $y$-coordinates they have in $\Gamma_0$ and $\Gamma_1$, respectively, and by drawing its edges as straight-line segments (see \blue{Fig.}\autoref{fig:ear-deco-p0}). It is easy to see that $\Gamma_0[G_0]$, $\Gamma'_0$, $\Gamma'_1$, and $\Gamma_1[G_0]$ fulfill Conditions~(\ref{pr:one})-(\ref{pr:three}) of the definition of an hvh-pair.
Suppose now that, for some $j\in \{1,\dots,k\}$, the drawings $\Gamma'_0$ and $\Gamma'_1$ are upward planar straight-line drawings of $G_{j-1}$ such that $\Gamma_0[G_{j-1}]$, $\Gamma'_0$, $\Gamma'_1$, and $\Gamma_1[G_{j-1}]$ fulfill Conditions~(\ref{pr:one})-(\ref{pr:three}) of the definition of an hvh-pair.
We show how to add the path $P_j= u_0 u_1 \dots u_{\ell-1} u_\ell$ to both $\Gamma'_0$ and $\Gamma'_1$ so that the resulting drawings together with $\Gamma_0[G_{j}]$ and $\Gamma_1[G_{j}]$ fulfill Conditions~(\ref{pr:one})-(\ref{pr:three}) of the definition of an hvh-pair. Note that $u_0$ and $u_\ell$ belong to the right boundary of the outer face of $G_{j-1}$, hence they are already present in $\Gamma'_0$ and $\Gamma'_1$. Since all the edges of $P_j$ are going to be drawn as straight-line segments, it suffices to show how to draw the internal vertices of $P_j$ in $\Gamma'_0$ and $\Gamma'_1$.
%
For $i=0,1$, we assign to the internal vertices of $P_j$ in $\Gamma'_i$ the same $y$-coordinates they have in $\Gamma_i$. Also, we assign to all such vertices, in both drawings, the same $x$-coordinate $x^*_j$, which has a ``sufficiently large'' value determined as follows.
If $j=1$, then we set $x^*_j=1$ (see \blue{Fig.}~\ref{fig:ear-deco-p1}).
If $j>1$, then we proceed as follows. Refer to~\autoref{fig:ear-inv}.
For $i=0,1$, let $\rho_i(u_0)$ be a ray emanating from $u_0$ with positive slope, directed rightwards, not intersecting $\Gamma'_i$, except at $u_0$, and such that the intersection point $a_i$ of $\rho_i(u_0)$ with the horizontal line $y= y_i(u_1)$ lies to the right of every vertex in $\Gamma'_i$. Analogously, for $i=0,1$, let $\rho_i(u_\ell)$ be a ray emanating from $u_\ell$ with negative slope, directed rightwards, not intersecting $\Gamma'_i$,
except at $u_\ell$, and such that the intersection point $b_i$ of $\rho_i(u_\ell)$ with the horizontal line $y= y_i(u_{\ell-1})$ lies to the right of every vertex in $\Gamma'_i$. Refer to \blue{Figs.}~\ref{fig:ear-inv-1} and~\ref{fig:ear-inv-3} for the cases in which $\ell=2$ and $\ell>2$, respectively.
Observe that $u_1$ and $u_{\ell-1}$ coincide if $P_j$ is a path of length $2$, i.e., if $\ell = 2$.
We set $x^*_j$ as the maximum of the $x$-coordinates of $a_0$, $b_0$, $a_1$, and $b_1$.
The obtained drawings $\Gamma'_0$ and $\Gamma'_1$ of $G_j$ are planar, as the slopes of the straight-line segments representing the edge $u_0 u_1$ in $\Gamma'_0$ and $\Gamma'_1$ are smaller than or equal to those of $\rho_0(u_0)$ and $\rho_1(u_0)$, respectively, as the slopes of the straight-line segments representing the edge $u_{\ell-1} u_\ell$ in $\Gamma'_0$ and $\Gamma'_1$ are larger than or equal to those of $\rho_0(u_\ell)$ and $\rho_1(u_\ell)$, respectively, and as the vertical line $x=x^*_j$ lies to the right of all the vertices of $G_{j-1}$ both in $\Gamma'_0$ and in $\Gamma'_1$. The drawings $\Gamma'_0$ and $\Gamma'_1$ are upward, given that the vertices of $G_j$ have the same $y$-coordinates they have in $\Gamma_0$ and $\Gamma_1$, respectively, and given that $\Gamma_0$ and $\Gamma_1$ are upward drawings.
In order to conclude the proof, we show that the obtained drawings, together with $\Gamma_0[G_j]$ and $\Gamma_1[G_j]$, fulfill Conditions~(\ref{pr:one})--(\ref{pr:three}) of the definition of an hvh-pair.
%
Since ($\Gamma_0$,$\Gamma'_0$) and ($\Gamma_1$, $\Gamma'_1$) are pairs of upward planar drawings of $G_j$ with the same $y$-assignment, by \autoref{le:st-y-assignment-equivalent}, Conditions (\ref{pr:one}) and (\ref{pr:three}) hold true. In order to prove Condition~(\ref{pr:two}), first recall that by construction $\Gamma'_0$ and $\Gamma'_1$ have the same $x$-assignment. Also, since all the intermediate vertices of any path $P_j$ are drawn on the vertical line $x=x^*_j$ in both $\Gamma'_0$ and $\Gamma'_1$, and the circular ordering of the edges around each vertex is the same in both $\Gamma'_0$ and $\Gamma'_1$, we have that the sequence of vertices and edges crossed by each vertical line in $\Gamma'_0$ and $\Gamma'_1$ is the same, thus implying Condition~(\ref{pr:two}).
\end{proof}
Combining \autoref{le:fast-morph} with \autoref{le:good-reduced} we obtain the following result.
\begin{theorem}\label{th:st-reduced-graphs}
Let $\Gamma_0$ and $\Gamma_1$ be any two upward planar drawings of a reduced plane $st$-graph. There is a $3$-step upward planar morph from $\Gamma_0$ to $\Gamma_1$.
\end{theorem}
\subsection{General Plane $st$-Graphs}\label{sse:general-st-graph}
We now turn our attention to general plane $st$-graphs. We restate here, in terms of plane $st$-graphs, a result by Hong and Nagamochi~\cite{DBLP:journals/jda/HongN10} that was originally formulated in terms of hierarchical plane (undirected) graphs.
\begin{theorem}[Hong and Nagamochi~\cite{DBLP:journals/jda/HongN10}, Theorem~8]\label{th:hongANDnaga-biconnected}
Consider an internally $3$-connected plane $st$-graph~$G$ and let $y_G$ be a $y$-assignment for the vertices of $G$ such that each vertex $v$ is assigned a value $y_G(v)$ that is greater than those assigned to its predecessors. There exists a strictly-convex upward planar drawing of $G$ satisfying $y_G$.
\end{theorem}
We use \autoref{th:hongANDnaga-biconnected} to prove the following lemma, which allows us to restrict our attention to maximal plane $st$-graphs.
\newcommand{\theoremstgraphmaximal}{Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex plane $st$-graph $G$.
Suppose that an algorithm $\mathcal A$ exists that constructs an $f(r)$-step upward planar morph between any two upward planar drawings of an $r$-vertex maximal plane $st$-graph.
Then there exists an $O(f(n))$-step upward planar morph from $\Gamma_0$ to $\Gamma_1$.}
\begin{lemma}\label{th:st-graphs-maximal}
\theoremstgraphmaximal
\end{lemma}
\begin{proof}
By \autoref{le:biconnected-reduced-augmentation} we can assume that $G$ is biconnected.
We augment $G$ to a maximal plane $st$-graph $G^*$ as follows (refer to \blue{Fig.~\ref{fig:triangulating-a}}). For each internal face $f$ of $G$ we add to $G$: (i) a vertex $v_f$ into $f$, (ii) a directed edge from the source-switch $s_f$ of $f$ to $v_f$, and (iii) directed edges from $v_f$ to every other vertex incident to $f$. We also add a vertex $v^*$ into the outer face of $G$, and add directed edges from $v^*$ to all the vertices incident to the outer face of $G$. The resulting graph $G^*$ is a maximal plane $st$-graph (in particular it is internally-$3$-connected) and contains $O(n)$ vertices.
\begin{figure}[tb]
\centering
\subfloat[\label{fig:triangulating-a}]
{\includegraphics[page=1,height=.3\textwidth]{figures/Triangulating}}
\hfil
\subfloat[\label{fig:triangulating-b}]
{\includegraphics[page=2,height=.3\textwidth]{figures/Triangulating}}
\caption{Illustration for the proof of \autoref{th:st-graphs-maximal}. (a) A biconnected plane $st$-graph $G$ (shown with black vertices and thick edges) and its augmentation to a maximal plane $st$-graph $G^*$ (by the addition of the white vertices and white-arrowed edges). (b) The construction of an upward planar drawing $\Gamma^*_i$ of $G^*$ in which each vertex of $G^*$ that is also in $G$ has the same $y$-coordinate as in $\Gamma_i$.}\label{fig:triangulating}
\end{figure}
Denote by $y^0_G$ the $y$-assignment for the vertices of $G$ that is induced by $\Gamma_0$. We define a $y$-assignment $y^0_{G^*}$ for the vertices of $G^*$ by setting:
\begin{itemize}
\item $y^0_{G^*}(v)=y^0_{G}(v)$ for each vertex $v\in V(G)$;
\item for each vertex $v_f$ of $G^*$ inserted into an internal face $f$ of $G$, a value for $y^0_{G^*}(v_f)$ that is larger than $y^0_{G^*}(s_f)$ and smaller than $y^0_{G^*}(v)$, for every other vertex $v$ incident to $f$; and
\item for the vertex $v^*$ of $G^*$ inserted into the outer face of $G$, a value for $y^0_{G^*}(v^*)$ that is smaller than $y^0_{G^*}(v)$, for every vertex $v\neq v^*$ of $G^*$.
\end{itemize}
We similarly define a $y$-assignment $y^1_{G^*}$ for the vertices of $G^*$ using the $y$-coordinates of $\Gamma_1$.
Note that, for $i=0,1$, each vertex $v$ of $G^*$ has been assigned a value $y^i_{G^*}$ that is greater than those assigned to its predecessors. We can hence use \autoref{th:hongANDnaga-biconnected} to construct upward planar drawings $\Gamma^*_0$ and $\Gamma^*_1$ of $G^*$ satisfying $y^0_{G^*}$ and $y^1_{G^*}$, respectively (refer to \blue{Fig.~\ref{fig:triangulating-b}}).
By \autoref{le:st-y-assignment-equivalent} we have that the drawings $\Gamma^*_0[G]$ and $\Gamma_0$ are left-to-right equivalent. Therefore, by \autoref{le:shift}, the linear morph $\mathcal M_0$ from $\Gamma_0$ to $\Gamma^*_0[G]$ is unidirectional and planar. Such a morph is also upward since both $\Gamma_0$ and $\Gamma^*_0[G]$ are upward planar and left-to-right equivalent. Analogously, the linear morph $\mathcal M_1$ from $\Gamma^*_1[G]$ to $\Gamma_1$ is upward planar.
We now apply algorithm $\mathcal A$ to construct an $O(n)$-step upward planar morph from $\Gamma^*_0$ to $\Gamma^*_1$ and restrict such a morph to the vertices and edges of $G$ to obtain an $O(n)$-step upward planar morph $\mathcal M_{01}$ from $\Gamma^*_0[G]$ to $\Gamma^*_1[G]$.
An upward planar morph $\mathcal M$ from $\Gamma_0$ to $\Gamma_1$ is finally obtained as the concatenation of $\mathcal M_0$, $\mathcal M_{01}$, and $\mathcal M_1$. The number of steps of $\mathcal M$ is equal to the number of steps of $\mathcal M_{01}$ plus two, hence it is in $O(n)$. This concludes the proof.
\end{proof}
In the following we will present an algorithm that constructs an upward planar morph between two upward planar drawings of a maximal plane $st$-graph. Before doing so, we need to introduce one more tool. The \emph{kernel} of a polygon $P$ is the set of points $p$ inside or on $P$ such that, for any point $q$ on $P$, the open segment $\overline{pq}$ lies inside $P$. We have the following.
\begin{lemma}\label{le:convexify-around-vertex}
Let $\Gamma$ be an upward planar drawing of an internally $3$-connected plane $st$-graph $G$, let $f$ be an internal $st$-face of $G$, and let $P$ be the polygon representing $f$ in $\Gamma$.
There exists an upward planar drawing $\Gamma'$ of $G$ such that the polygon representing the boundary of $f$ is strictly convex and $\mathcal M = \langle \Gamma, \Gamma' \rangle$ is a unidirectional upward planar morph.
Further, if $v$ is a vertex incident to $f$ that is in the kernel of $P$ in~$\Gamma$, then $v$ is in the kernel of the polygon representing the boundary of $f$ throughout $\mathcal M$.
\end{lemma}
\begin{proof}
Denote by $y_G$ the $y$-assignment for the vertices of $G$ induced by $\Gamma$. By \autoref{th:hongANDnaga-biconnected}, there exists a strictly-convex upward planar drawing $\Gamma'$ of $G$ satisfying $y_G$.
Thus, by~\autoref{le:st-y-assignment-equivalent} and since $G$ is a plane $st$-graph, we have that $\Gamma$ and $\Gamma'$ are left-to-right-equivalent drawings. By~\autoref{le:shift}, the linear morph $\mathcal M$ from $\Gamma$ to $\Gamma'$ is unidirectional and planar. Since $\Gamma$ and $\Gamma'$ are upward and left-to-right equivalent, it follows that $\mathcal M$ is an upward planar morph.
Consider now a vertex $v$ incident to $f$ that is in the kernel of $P$ in~$\Gamma$. Since the polygon representing the boundary of $f$ in $\Gamma'$ is strictly convex, $v$ is also in the kernel of such a polygon. Augment $G$ to a graph $G_*$ by introducing (suitably oriented) edges connecting $v$ to the vertices incident to $f$ that are not already adjacent to~$v$. Draw these edges in $\Gamma$ and $\Gamma'$ as straight-line segments, obtaining two drawings $\Gamma_*$ and $\Gamma_*'$ of $G_*$. Since $v$ is in the kernel of the polygon representing the boundary of $f$ both in $\Gamma$ and in $\Gamma'$, and since $\Gamma$ and $\Gamma'$ are upward planar and left-to-right equivalent, we have that $\Gamma_*$ and $\Gamma_*'$ are left-to-right equivalent upward planar drawings of $G_*$. By the same arguments used for $\mathcal M$, we have that the linear morph $\mathcal M_* = \langle \Gamma_*, \Gamma_*'\rangle$ is planar. Hence, $v$ is in the kernel of the polygon representing the boundary of $f$ throughout $\mathcal M$.
\end{proof}
Given two upward planar straight-line drawings $\Gamma_0$ and $\Gamma_1$ of a maximal plane $st$-graph $G$, our strategy for constructing an upward planar morph from $\Gamma_0$ to $\Gamma_1$ consists of the following steps:
\begin{enumerate}[(1)]
\item we find a simple vertex $v$ of $G$ of degree at most $5$;
\item we remove $v$ and its incident edges from $G$, $\Gamma_0$, and $\Gamma_1$, obtaining upward planar drawings $\Gamma^-_0$ and $\Gamma^-_1$ of an upward plane graph $G^-$;
\item we triangulate $G^-$, $\Gamma^-_0$, and $\Gamma^-_1$ by inserting edges incident to a former neighbor $u$ of $v$, obtaining upward planar drawings $\Gamma'_0$ and $\Gamma'_1$ of a maximal plane $st$-graph $G'$;
\item we apply induction in order to construct an upward planar morph $\mathcal M'$ from $\Gamma'_0$ to $\Gamma'_1$; and
\item we remove from $\mathcal M'$ the edges incident to $u$ that are not in $G$ and insert $v$ and its incident edges in $\mathcal M'$, thus obtaining an upward planar morph from $\Gamma_0$ to $\Gamma_1$.
\end{enumerate}
In order for this strategy to work, we need $u$ to satisfy certain properties, which are expressed in the upcoming definition of {\em distinguished neighbor}; further, we need to perform one initial (and one final) upward planar morph so to convexify the polygon representing what will be called a {\em characteristic cycle}.
Let $v$ be a simple vertex with degree at most $5$ in a maximal plane $st$-graph $G$. Let $G(v)$ be the subgraph of $G$ induced by $v$ and its neighbors.
A predecessor $u$ of $v$ in $G$ is a \emph{distinguished predecessor} if it satisfies the following properties: (a) for each predecessor $w$ of $v$, there is a directed path in $G(v)$ from $w$ to $v$ through~$u$; (b) $u$ is the only predecessor of $v$ if its degree is $3$; and (c) $v$ has at most two predecessors if its degree is $4$ or $5$.
A successor $u$ of $v$ in $G$ is a \emph{distinguished successor} if it satisfies the following properties: (a) for each successor $w$ of $v$, there is a directed path in $G(v)$ from $v$ to $w$ through $u$; (b) $u$ is the only successor of $v$ if its degree is $3$; and (c) $v$ has at most two successors if its degree is $4$ or $5$.
A neighbor of $v$ is a \emph{distinguished neighbor} if it is either a distinguished predecessor or successor of $v$. Examples of distinguished neighbors are in \autoref{fi:distinguished-neighbors}. We are going to exploit the following.
\begin{figure}[t]
\centering
\subfloat[]
{\includegraphics[page=1,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-a}}
\hfil
\subfloat[]
{\includegraphics[page=3,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-b}}
\hfil
\subfloat[]
{\includegraphics[page=4,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-c}}
\hfil
\subfloat[]
{\includegraphics[page=5,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-d}}
\hfil
\subfloat[]
{\includegraphics[page=2,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-e}}
\hfil
\subfloat[]
{\includegraphics[page=6,height=.18\textwidth]{figures/characteristic}\label{fi:characteristic-f}}
\hfil
\subfloat[]
{\includegraphics[page=7,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-g}}
\hfil
\subfloat[]
{\includegraphics[page=8,height=.2\textwidth]{figures/characteristic}\label{fi:characteristic-h}}
\caption{Distinguished predecessors (enclosed by red squares), distinguished successors (enclosed by red circles), and characteristic cycles (filled yellow). Note that, in (e), (g), and (h), the vertex $s_2$ is not a distinguished successor of $v$; indeed, although for every successor $w$ of $v$ there is a directed path in $G(v)$ from $v$ to $w$ through $s_2$, we have that $v$ has more than two successors.}
\label{fi:distinguished-neighbors}
\end{figure}
\newcommand{The vertex $v$ has at most one distinguished predecessor, at most one distinguished successor, and at least one distinguished neighbor.}{The vertex $v$ has at most one distinguished predecessor, at most one distinguished successor, and at least one distinguished neighbor.}
\begin{lemma}\label{le:distinguished}
The vertex $v$ has at most one distinguished predecessor, at most one distinguished successor, and at least one distinguished neighbor.
\end{lemma}
\begin{proof}
Suppose, for a contradiction, that $v$ has (at least) two distinguished predecessors $u_1$ and $u_2$. Since $u_1$ is a distinguished predecessor of $v$ and $u_2$ is a predecessor of $v$, it follows that $G$ contains a directed path $u_2 \dots u_1 v$; further, since $u_2$ is a distinguished predecessor of $v$ and $u_1$ is a predecessor of $v$, it follows that $G$ contains a directed path $u_1 \dots u_2 v$. The union of these directed paths contains a directed cycle, a contradiction to the fact that $G$ is an $st$-graph. It follows that $v$ has at most one distinguished predecessor. An analogous argument proves that $v$ has at most one distinguished successor.
Let $P$ and $S$ be the sets of predecessors and successors of $v$ in $G$, respectively.
If the degree of $v$ is $3$, then either $|P|=1$ or $|S|=1$. In the former case the only predecessor of $v$ is a distinguished predecessor of $v$, while in the latter case the only successor of $v$ is a distinguished successor of $v$.
Assume next that the degree of $v$ is $4$ or $5$. We prove that, if $|P|\leq 2$, then $v$ has at least one distinguished predecessor. If $|P|=1$, then the only predecessor of $v$ is a distinguished predecessor of $v$. Further, if $|P|=2$, then let $s$ and $p$ be the two predecessors of $v$ in $G$. Since $G$ is maximal, it contains either the directed edge $sp$ or the directed edge $ps$. In the former case $p$ is a distinguished predecessor of $v$, while in the latter case $s$ is a distinguished predecessor of $v$. An analogous proof shows that, if $|S|\leq 2$, then $v$ has at least one distinguished successor. This completes the proof, given that $|P|\leq 2$ or $|S|\leq 2$, since the degree of $v$ is at most $5$.
\end{proof}
We define the \emph{characteristic cycle} $C(v)$ as follows. Let $c_G(v)$ be the subgraph of $G$ induced by the neighbors of $v$. Since $v$ is simple, the underlying graph of $c_G(v)$ is a cycle.
If $c_G(v)$ is an $st$-cycle, as in \blue{Figs.~\ref{fi:characteristic-a}}, \blue{~\ref{fi:characteristic-b}}, \blue{~\ref{fi:characteristic-c}}, and \blue{~\ref{fi:characteristic-d}}, then $C(v):=c_G(v)$; in particular, this is always the case if $v$ has degree~$3$.
Otherwise, $c_G(v)$ has two sources $s_1$ and $s_2$ and two sinks $t_1$ and $t_2$. Throughout the rest of this section, we always assume that, if $c_G(v)$ has two sources $s_1$ and $s_2$ and two sinks $t_1$ and $t_2$, then $G$ contains the edges $s_1v$ and $vs_2$; indeed, the cases in which $G$ contains the edges $s_2v$ and $vs_1$, or $t_1v$ and $vt_2$, or $t_2v$ and $vt_1$ are analogous. This assumption implies that $v$ has at least three successors, namely $s_2$, $t_1$, and $t_2$, and hence no distinguished successor. Suppose also, w.l.o.g., that $s_1, t_1, s_2$, and $t_2$ appear in this clockwise order along $c_G(v)$.
If $v$ has degree $4$, as in \blue{Fig.~\ref{fi:characteristic-e}}, then $C(v)$ is composed of the edges $s_1v$, $vs_2$, $s_2t_2$, and $s_1t_2$.
Otherwise, $v$ has degree $5$, as in \blue{Figs.~\ref{fi:characteristic-f}}, \blue{~\ref{fi:characteristic-g}}, and\blue{~\ref{fi:characteristic-h}}. Let $v_1$ be the distinguished predecessor of $v$. The directed path $P_1 = v_1vs_2$ splits $c_G(v)$ into two paths $P_2$ and $P_3$ with length $2$ and $3$, respectively. Then $C(v)$ is composed of $P_1$ and $P_3$. We have the following structural lemma.
\newcommand{\characteristiccycleproperties}{
The characteristic cycle $C(v)$ is an $st$-cycle which contains all the distinguished neighbors of $v$. Further, all the vertices of $c_G(v)$ not belonging to $C(v)$ are adjacent to all the distinguished neighbors of $v$.
}
\begin{lemma}\label{le:characteristic-cycle-properties}
\characteristiccycleproperties
\end{lemma}
\begin{proof}
If $c_G(v)$ is an $st$-cycle, then by construction $C(v)$ coincides with $c_G(v)$, hence $C(v)$ is an $st$-cycle which contains all the neighbors (and in particular all the distinguished neighbors) of $v$, and there are no vertices of $c_G(v)$ not belonging to $C(v)$. In the following we hence assume that $c_G(v)$ is not an $st$-cycle.
If $v$ has degree $4$, then by construction $C(v)$ consists of two directed paths, namely $s_1vs_2t_2$ and $s_1t_2$, hence it is an $st$-cycle which contains the only distinguished neighbor (namely $s_1$) of $v$. The only vertex of $c_G(v)$ not belonging to $C(v)$, namely $t_1$, is adjacent to $s_1$.
If $v$ has degree $5$, then observe that $v$ is neither a source nor a sink of $C(v)$, as $C(v)$ contains the directed edges $v_1v$ and $vs_2$; further, $s_2$ is neither a source nor a sink of $C(v)$, as $C(v)$ contains the directed edge $vs_2$ and the directed edge of $P_3$ outgoing $s_2$. Since the underlying graph of $C(v)$ is a cycle with $5$ vertices, it follows that $C(v)$ has one source and one sink, hence it is an $st$-cycle. Further, by construction $C(v)$ contains $v_1$, hence it contains all the distinguished neighbors of $v$. Finally, the only vertex of $c_G(v)$ not belonging to $C(v)$ is the internal vertex of $P_2$, which is adjacent to $v_1$.
\end{proof}
Characteristic cycles are used in order to prove the following.
\newcommand{\distinguishedkernel}{
Let $\Gamma$ be any upward planar drawing of $G$. There is a unidirectional upward planar morph $\langle \Gamma, \Gamma' \rangle$, where in $\Gamma'$ the distinguished neighbors of $v$ are in the kernel of the polygon representing~$c_G(v)$.
}
\begin{lemma}\label{le:distinguished-kernel}
\distinguishedkernel
\end{lemma}
\begin{proof}
By Lemma~\ref{le:characteristic-cycle-properties} the distinguished neighbors of $v$ belong to $C(v)$. If the polygon $P$ representing $C(v)$ in $\Gamma$ is convex, then each distinguished neighbor of $v$ {\em sees} the other vertices of $C(v)$, meaning that the open straight-line segment connecting any distinguished neighbor of $v$ with any other vertex of $C(v)$ lies inside $P$, and hence inside the polygon representing $c_G(v)$ in $\Gamma$. Again by Lemma~\ref{le:characteristic-cycle-properties} the vertices of $c_G(v)$ that are not in $C(v)$ are adjacent to the distinguished neighbors of $v$, which hence see every vertex of $c_G(v)$. It follows that the distinguished neighbors of $v$ are in the kernel of the polygon representing $c_G(v)$ in $\Gamma$, and we can just define $\Gamma':=\Gamma$; note that no morph is actually needed in order to obtain the desired drawing $\Gamma'$ from $\Gamma$.
If $P$ is not convex, then we show how a unidirectional upward planar morph can be employed in order to transform $\Gamma$ into an upward planar drawing $\Gamma'$ in which the polygon representing $C(v)$ is convex, thus bringing us back to the previous case. Let $G^\circ$ be the subgraph of $G$ obtained by removing all the vertices and edges in the interior of $C(v)$ and let $\Gamma^\circ$ be $\Gamma[G^\circ]$. Observe that only $v$ and its incident edges might be removed from $G$ in order to obtain $G^\circ$.
We prove that $G^\circ$ is $3$-connected. Suppose, for a contradiction, that $G^\circ$ contains a $2$-cut $\{a,b\}$. If $v$ was removed from $G$ in order to obtain $G^\circ$, then $\{a,b,v\}$ is a $3$-cut of $G$. Since $G$ is maximal, any $3$-cut induces a separating triangle, i.e., a $3$-cycle with vertices both on the inside and on the outside. However, since $v$ is simple, it is not part of any separating triangle, a contradiction. Assume next that $v$ was not removed from $G$ in order to obtain $G^\circ$. If $v \in \{a,b\}$, then $\{a,b\}$ is also a $2$-cut of $G$, contradicting the fact that $G$ is maximal (and hence $3$-connected). Finally, if $v \notin \{a,b\}$, then $\{a,b,v\}$ is a $3$-cut of $G$, and a contradiction can be derived as in the case in which $v$ was removed from $G$.
Since $G^\circ$ is $3$-connected and $C(v)$ is an $st$-cycle (by Lemma~\ref{le:characteristic-cycle-properties}), we can apply Lemma~\ref{le:convexify-around-vertex} to construct an upward planar drawing $\Gamma^\diamondsuit$ of $G^\circ$ such that $C(v)$ is strictly convex in $\Gamma^\diamondsuit$ and $\langle \Gamma, \Gamma^\diamondsuit \rangle$ is a unidirectional upward planar morph.
We obtain our desired upward planar drawing $\Gamma'$ of $G$ from $\Gamma^\diamondsuit$ as follows.
If $G^\circ$ contains $v$, then we simply augment $\Gamma^\diamondsuit$ by drawing the edges that are in $G$ but not in $G^\circ$ as straight-line segments, thus obtaining $\Gamma'$. The convexity of $C(v)$ in $\Gamma^\diamondsuit$ implies that no crossings are introduced because of this augmentation. Further, as in the proof of Lemma~\ref{le:convexify-around-vertex}, we have that $\Gamma$ and $\Gamma'$ have the same $y$-assignment, hence by Lemma~\ref{le:st-y-assignment-equivalent} they are left-to-right equivalent, and thus by Lemma~\ref{le:shift} the linear morph $\langle \Gamma, \Gamma' \rangle$ is unidirectional and upward planar.
If $G^\circ$ does not contain $v$, then we need to determine a placement for $v$ in $\Gamma^\diamondsuit$ in order to obtain $\Gamma'$. We insert $v$ in the interior of the convex polygon representing $C(v)$ in $\Gamma^\diamondsuit$, so that its $y$-coordinate is the same as in $\Gamma$. We draw the edges incident to $v$ as straight-line segments. This ensures that $\Gamma$ and $\Gamma'$ have the same $y$-assignment and hence, as in the previous case, that the linear morph between them is unidirectional and upward planar.
\end{proof}
The following concludes our discussion on maximal plane $st$-graph.
\newcommand{Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex maximal plane $st$-graph $G$. There is an $O(n)$-step upward planar \mbox{morph from $\Gamma_0$ to $\Gamma_1$.}}{Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex maximal plane $st$-graph $G$. There is an $O(n)$-step upward planar \mbox{morph from $\Gamma_0$ to $\Gamma_1$.}}
\begin{theorem}\label{th:st-triangulated-graphs}
Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex maximal plane $st$-graph $G$. There is an $O(n)$-step upward planar \mbox{morph from $\Gamma_0$ to $\Gamma_1$.}
\end{theorem}
\begin{proof}
The proof is by induction on $n$. In the base case we have $n=3$, hence $\Gamma_0$ and $\Gamma_1$ are two triangles. We show that $\Gamma_0$ and $\Gamma_1$ form an hvh-pair. Denote by $u$ and $w$ the source and the sink of $G$, respectively. Observe that the third vertex of $G$, call it $v$, is on the same side of the edge $uw$ in $\Gamma_0$ and in $\Gamma_1$, as $\Gamma_0$ and in $\Gamma_1$ have the same upward planar embedding; assume that $v$ lies to the right of $uw$, the other case is symmetric. For $i=0,1,$ let $\Gamma'_i$ be a drawing of $G$ such that the $x$-coordinate of $u$ and $w$ is $0$, the $x$-coordinate of $v$ is $1$, and $y$-coordinate of each vertex is the same as in $\Gamma_i$. It is easy to see that $\Gamma'_0$ and $\Gamma'_1$ are upward planar drawings of $G$ and that these drawings, together with $\Gamma_0$ and $\Gamma_1$, satisfy Conditions~(\ref{pr:one})--(\ref{pr:three}) of the definition of an hvh-pair. Thus, by \autoref{le:fast-morph}, there exists a $3$-step upward planar morph from $\Gamma_0$ to $\Gamma_1$.
Suppose next that $n > 3$. By Lemma~\ref{le:internal-vertex}, $G$ contains a simple vertex $v$ of degree at most $5$. By Lemma~\ref{le:distinguished}, $v$ has at least one distinguished neighbor, which we denote by $u$. Assume for the remainder of the proof that $u$ is a predecessor of $v$, the case in which it is a successor of $v$ being symmetric. By Lemma~\ref{le:distinguished-kernel}, there exists a unidirectional upward planar morph from $\Gamma_0$ to an upward planar drawing, which we denote again by $\Gamma_0$, in which $u$ lies in the kernel of $c_G(v)$. Analogously, by means of a unidirectional upward planar morph, we can ensure that $u$ lies in the kernel of $c_G(v)$ in $\Gamma_1$.
In order to obtain the desired morph from $\Gamma_0$ to $\Gamma_1$ we are going to apply induction. For this sake we define an $(n-1)$-vertex maximal plane $st$-graph $G'$, and two upward planar drawings $\Gamma_0'$ and $\Gamma_1'$ of it. The graph $G'$ is obtained from $G$ by removing $v$ and by inserting a directed edge $uq$ for each successor $q$ of $v$ that is not adjacent to $u$ in $G$. These edges are all added inside $c_G(v)$. Note that, by the definition of distinguished predecessor, either $u$ is the only predecessor of $v$, or $v$ has one predecessor $p$ different from $u$, where $G$ contains the directed edge $pu$. The drawings $\Gamma_0'$ and $\Gamma_1'$ are obtained from $\Gamma_0$ to $\Gamma_1$, respectively, by removing $v$ and its incident edges and by drawing the edges of $G'$ \mbox{not in $G$ as straight-line segments. }
For $i=0,1$, since $u$ lies in the kernel of the polygon representing $c_G(v)$ in $\Gamma_i$, we have that $\Gamma'_i$ is planar. We prove that $\Gamma_i'$ is upward. Every successor $q$ of $v$ has a $y$-coordinate larger than the one of $v$ in $\Gamma_i$; since $u$ has a $y$-coordinate smaller than the one of $v$ in $\Gamma_i$, it follows that the edge from $u$ to $q$ is monotonically increasing in the $y$-direction in $\Gamma'_i$. Since all the edges of $G'$ that are also in $G$ are drawn as in $\Gamma_i$ and since $\Gamma_i$ is an upward drawing, it follows that $\Gamma'_i$ is an upward planar drawing of $G'$. Observe that $G'$ is an $st$-graph; indeed, it suffices to note that the edges that are removed from $G$ do not result in any new source or sink in $G'$: (i) no successor $q$ of $v$ becomes a source in $G'$, as a directed edge $uq$ is inserted in $G'$ if it is not in $G$; (ii) no predecessor $p$ of $v$ different from $u$, if any, becomes a sink in $G'$, as the directed edge $pu$ belongs to $G$; and (iii) $u$ does not become a sink in $G'$ as $v$ has at least one successor in $G$. Finally, note that $G'$ is maximal, since $G$ is maximal and the edges added to $G'$ triangulate the interior of $c_G(v)$. It follows that $G'$ is a maximal plane $st$-graph.
By induction, there is an upward planar morph $\mathcal M'=\langle \Gamma'_0 = \Lambda_0, \Lambda_1, \dots, \Lambda_k=\Gamma'_1\rangle$ from $\Gamma'_0$ to $\Gamma'_1$. In the following we transform $\mathcal M'$ into an upward planar morph $\mathcal M$ between two upward planar drawings $\Delta_0$ and $\Delta_1$ of $G$. This will be done by inserting $v$ at a suitable point in the drawing of $G'$ at any time instant of the morph $\mathcal M'$ and by drawing the edges incident to $v$ as straight-line segments. We will later show that $\mathcal M$ is actually composed of $k$ linear morphs.\footnote{This insertion problem has been studied and solved in~\cite{DBLP:journals/siamcomp/AlamdariABCLBFH17} for planar morphs of undirected graphs. Here we cannot immediately reuse the results in~\cite{DBLP:journals/siamcomp/AlamdariABCLBFH17}, as we need to preserve the upwardness of the drawing throughout the morph. However, the property that every drawing of $G'$ in $\mathcal M'$ is upward significantly simplifies the problem of inserting $v$ in $\mathcal M'$ so to obtain an upward planar morph of $G$.}
Let $\varepsilon>0$ be a sufficiently small value such that the following properties are satisfied throughout $\mathcal M'$:
\begin{enumerate}[(a)]
\item for each successor $q$ of $v$ in $G$, it holds true that $y(q) > y(u) + \varepsilon$;
\item if $v$ has a predecessor $p\neq u$ in $G$, then $y(p) < y(u) - \varepsilon$; and
\item for any segment $s$ of $c_G(v)$ not incident to $u$, the line through $s$ does not intersect the disk $\delta$ with radius $\varepsilon$ centered at $u$.
\end{enumerate}
Since $\mathcal M'$ is an upward planar morph and since $G'$ contains edges from $u$ to every successor $q$ of $v$ and from every predecessor $p$ of $v$ to $u$, it follows that such a value $\varepsilon$ exists; in particular, standard continuity arguments, like the ones used in the proof of F\'ary's Theorem~\cite{Fary}, ensure that Property~(c) is satisfied for a sufficiently small value $\varepsilon>0$.
We distinguish the cases in which $v$ has degree $3$ or greater than $3$ in $G$.
If $v$ has degree $3$ in $G$, then let $a$, $b$, and $c$ be the neighbors of $v$ in $G$, where $a=u$. We choose three values $\alpha, \beta$, and $\gamma$, as discussed below, and then place $v$ at the point $\alpha \cdot a + \beta \cdot b + \gamma \cdot c$ at any time instant of $\mathcal M'$ ($a$, $b$, and $c$ here represent the points at which the corresponding vertices are placed at any time instant of $\mathcal M'$). We choose $\alpha, \beta$, and $\gamma$ as any positive values such that $\alpha + \beta + \gamma = 1$ and such that the point $\alpha \cdot a + \beta \cdot b + \gamma \cdot c$ lies in $\delta$ throughout $\mathcal M'$. Note that $v$ lies inside the triangle $c_G(v)$ for any positive values of $\alpha, \beta$, and $\gamma$ such that $\alpha + \beta + \gamma = 1$ (indeed, the position of $v$ is a convex combination of the ones of $a$, $b$, and $c$); further, choosing $\alpha$ sufficiently close to $1$ ensures that $v$ is at distance at most $\varepsilon$ from $u$, and hence lies inside $\delta$, throughout $\mathcal M'$.
Suppose now that $v$ has degree $4$ or $5$ in $G$. Since $u$ is a distinguished predecessor of $v$, we have that $v$ has at most two predecessors in $G$, one of which is $u$. If $v$ has no predecessor other than $u$, then $v$ has at least three successors in $G$; let $w$ be a successor of $v$ not adjacent to $u$ in $G$. If $v$ has a predecessor $p$ different from $u$, then $v$ has at least two successors in $G$; let $w$ be the one adjacent to $p$ in $G$. Note that, in both cases, the directed edge $uw$ belongs to $G'$ but not to $G$ and connects $u$ with a successor $w$ of $v$. We compute a value $\lambda$, as discussed below, and then place $v$ at the point $\lambda \cdot u + (1-\lambda) \cdot w$ at any time instant of $\mathcal M'$ ($u$ and $w$ here represent the points at which the corresponding vertices are placed at any time instant of $\mathcal M'$). We choose $\lambda$ as any positive value smaller than $1$ such that the point $\lambda \cdot u + (1-\lambda) \cdot w$ lies in $\delta$ throughout $\mathcal M'$. Note that $v$ is on the straight-line segment representing the edge $uw$ for any positive value of $\lambda$ smaller than $1$ (indeed, the position of $v$ is a convex combination of the ones of $u$ and $w$); further, choosing $\lambda$ sufficiently close to $1$ ensures that $v$ is at distance at most $\varepsilon$ from $u$, and hence lies inside $\delta$, throughout $\mathcal M'$.
In both cases, the choice of $\varepsilon$ ensures that at any time instant of $\mathcal M$ the drawing of $G$ is upward planar. In particular, Properties~(a) and~(b), together with the fact that every drawing of $G'$ in $\mathcal M'$ is upward, directly ensure that every drawing of $G$ in $\mathcal M$ is upward. Further, Property~(c), together with the fact that every drawing of $G'$ in $\mathcal M'$ is planar, ensures that every drawing of $G$ in $\mathcal M$ is planar. In particular, every point of $\delta$ sees every point of any straight-line segment $s$ of $c_G(v)$ not incident to $u$ in the interior of the polygon representing $c_G(v)$; hence the directed edges from $v$ to its successors cause no crossings throughout $\mathcal M$. Further, if $v$ has a predecessor $p$ different from $u$, the fact that $v$ lies on the straight-line segment connecting $u$ with the neighbor of $p$ in $G$ ensures that $p$ sees $v$ in the interior of the polygon representing $c_G(v)$ throughout $\mathcal M$.
We now prove that $\mathcal M$ consists of $k$ morphing steps. Assume that the degree of $v$ is $3$, the discussion for the case in which the degree of $v$ is $4$ or $5$ being analogous and simpler. Denote by $\Delta_i$ the drawing of $G$ obtained from $\Lambda_i$ by placing $v$ at the point $\alpha \cdot a_i + \beta \cdot b_i + \gamma \cdot c_i$, as discussed above, where by $a_i$, $b_i$, and $c_i$ denote the positions of the vertices $a$, $b$, and $c$ in $\Lambda_i$, respectively. Hence, at any time $t\in [0;1]$ of the linear morph $\langle \Delta_i,\Delta_{i+1}\rangle$, the position of $v$ is
\begin{eqnarray*}
&& (1-t)\cdot(\alpha \cdot a_i + \beta \cdot b_i + \gamma \cdot c_i) + t\cdot(\alpha \cdot a_{i+1} + \beta \cdot b_{i+1} + \gamma \cdot c_{i+1})\\
&=& \alpha((1-t)\cdot a_i + t \cdot a_{i+1})+ \beta((1-t)\cdot b_i + t \cdot b_{i+1})+ \gamma((1-t)\cdot c_i + t \cdot c_{i+1}).
\end{eqnarray*}
Hence, the position of $v$ at any time instant of the linear morph $\langle \Delta_i,\Delta_{i+1}\rangle$ is given by the convex combination with coefficients $\alpha$, $\beta$, and $\gamma$ of the positions of $a$, $b$, and $c$. It follows that the upward planar morph $\mathcal M$ defined above coincides with the $k$-step morph $\langle \Delta_0, \Delta_1, \dots, \Delta_k\rangle$.
Finally, denote by $f(n)$ the number of morphing steps of the described algorithm. We have $f(3)=3$ and $f(n)=4+f(n-1)$, if $n>3$. Indeed, in the inductive case the upward planar morph from $\Gamma_0$ to $\Gamma_1$ consists of:
\begin{itemize}
\item a first morphing step from the given drawing $\Gamma_0$ to the drawing $\Gamma_0$ in which $u$ lies in the kernel of $c_G(v)$;
\item a second morphing step $\langle \Gamma_0,\Delta_0\rangle$, where in $\Gamma_0$ the vertex $u$ lies in the kernel of $c_G(v)$ (note that only $v$ moves during this morphing step);
\item the morph $\mathcal M=\langle \Delta_0, \Delta_1, \dots, \Delta_k\rangle$, whose number $k$ of steps is the same as in $\mathcal M'$, which is the inductively constructed morph of the $(n-1)$-vertex maximal plane $st$-graph $G'$; hence $k=f(n-1)$;
\item a second to last morphing step $\langle \Delta_k,\Gamma_1\rangle$, where in $\Gamma_1$ the vertex $u$ lies in the kernel of $c_G(v)$ (note that only $v$ moves during this morphing step); and
\item a final morphing step from the drawing $\Gamma_1$ in which $u$ lies in the kernel of $c_G(v)$ to the given drawing $\Gamma_1$.
\end{itemize}
The recurrence equation for $f(n)$ solves to $f(n)=4n-9$. This concludes the proof of the theorem.
\end{proof}
\newcommand{Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex plane $st$-graph. There exists an $O(n)$-step upward planar morph from $\Gamma_0$ to $\Gamma_1$.}{Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex plane $st$-graph. There exists an $O(n)$-step upward planar morph from $\Gamma_0$ to $\Gamma_1$.}
We finally get the following.
\begin{corollary}\label{co:st-main}
Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex plane $st$-graph. There exists an $O(n)$-step upward planar morph from $\Gamma_0$ to $\Gamma_1$.
\end{corollary}
\begin{proof}
The statement follows by \autoref{th:st-graphs-maximal} and \autoref{th:st-triangulated-graphs}.
\end{proof}
\section{Upward Plane Graphs}\label{se:general-graphs}
In this section we deal with general upward plane graphs. In order to morph two upward planar drawings of an upward plane graph $G$ we are going to augment the upward planar drawings of $G$ to two upward planar drawings of a (possibly reduced) plane $st$-graph $G'$ and then to use the results of Section~\ref{se:st-graphs} for morphing the obtained upward planar drawings of $G'$. The augmentation process itself uses upward planar morphs. In the following we formally describe this strategy.
Let $G$ be an upward plane graph whose underlying graph is biconnected, let $f$ be a face of $G$ which is not an $st$-face, and let $u$, $v$, and $w$ be three clockwise consecutive switches of $f$. Further, let $v^-$ and $v^+$ be the vertices preceding and succeeding $v$ in clockwise order along the boundary of $f$, respectively, and let $u^-$ and $u^+$ be the vertices preceding and succeeding $u$ in clockwise order along the boundary of $f$, respectively. We say that $[u, v, w]$ is a \emph{pocket} for $f$ if $\angle (v^-,v,v^+)= \texttt{small}$ and $\angle (u^-,u,u^+) = \texttt{large}$. The following is well-known.
\begin{lemma}[Bertolazzi et al.~\cite{DBLP:journals/algorithmica/BertolazziBLM94}]\label{lem:small-small-large}
Let $G$ be an upward plane graph whose underlying graph is biconnected and let $f$ be a face of $G$ that is not an $st$-face. Then, there exists a pocket $[u, v, w]$ for $f$.
\end{lemma}
Next, we give a lemma that shows how to ``simplify'' a face of an upward plane graph that is not an $st$-graph, by removing one of its pockets.
\newcommand{\morphingbetweenaugmentations}{
Let $G$ be an $n$-vertex (reduced) upward plane graph whose underlying graph is biconnected, let $f$ be a face of $G$ that is not an $st$-face, let $[u, v, w]$ be a pocket for $f$, and let $\Gamma$ be an upward planar drawing of $G$.
Suppose that an algorithm $\mathcal A$ ($\mathcal A_R$) exists that constructs an $f(r)$-step ($f_R(r)$-step) upward planar morph between any two upward planar drawings of an $r$-vertex (reduced) plane $st$-graph.
Then, there exists an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph from $\Gamma$ to an upward planar drawing $\Gamma^*$ of $G$ in which $u$ sees $w$ inside $f$ and in which $u$ lies below $w$, if the directed path between $u$ and $v$ along the boundary of $f$ is directed from $v$ to $u$, or $u$ lies above $w$, otherwise.
}
\begin{lemma}\label{lem:morphing-between-augmentations}
\red{\morphingbetweenaugmentations}
\end{lemma}
\begin{proof}
Suppose that the directed path $p_{vu}$ between $u$ and $v$ along the boundary of $f$ is directed from $v$ to $u$ (refer to \blue{Fig.}~\ref{fig:augmentation-a}); the case in which it is directed from $u$ to $v$ can be treated symmetrically.
\begin{figure}[tb]
\centering
\subfloat[\label{fig:augmentation-a}]
{\includegraphics[page=1,width=.33\textwidth]{figures/gadget-cready}}
\hfil
\subfloat[]
{\includegraphics[page=1,width=.33\textwidth]{figures/gadget4}\label{fig:augmentation-b}}
\hfil
\subfloat[]
{\includegraphics[page=1,width=.33\textwidth]{figures/gadget5}\label{fig:augmentation-c}}
\caption{Illustrations for the proof of \autoref{lem:morphing-between-augmentations}. (a) The upward planar drawing $\Gamma$ of $G$; in particular, the illustration shows the face $f$ whose boundary contains the pocket $[u, v, w]$. (b) The upward planar drawing $\Gamma'$ of $G$ (plus the directed paths $p'$ and $p''$). (c) The upward planar drawing $\Gamma^*$ of $G$ (plus the directed edge $uw$). }\label{fig:augmentations}
\end{figure}
The proof is structured as follows.
First, we show that there exists an upward planar drawing $\Gamma'$ of~$G$ such that (see \blue{Fig.}~\ref{fig:augmentation-b}):
\begin{enumerate}[(i)]
\item \label{prop:i} the upward planar drawings of two directed paths $p'$ and $p''$ from $u$ to $w$ can be inserted in $\Gamma'$ in the interior of $f$; and
\item \label{prop:ii} there exists an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph $\mathcal M'$ from $\Gamma$ to $\Gamma'$.
\end{enumerate}
Second, we show that there exists an upward planar drawing $\Gamma^*$ of $G$ such that (see \blue{Fig.}~\ref{fig:augmentation-c}):
\begin{enumerate}[(i)]\setcounter{enumi}{2}
\item\label{prop:iii} $u$ sees $w$ inside $f$ and $u$ lies below $w$, and
\item\label{prop:iv}there exists an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph $\mathcal M^*$ from $\Gamma'$ to $\Gamma^*$.
\end{enumerate}
The lemma follows from the existence of the drawings $\Gamma'$ and $\Gamma^*$ above, since composing $\mathcal M'$ with $\mathcal M^*$ yields the desired upward planar morph with $O(f(n))$ steps (with $O(f_R(n))$ steps).
The drawing $\Gamma'$ is constructed in four phases.
In {\bf phase 1} we augment $\Gamma$ to an upward planar drawing $\Gamma_1$ of an upward plane graph $G_1$. Refer to \blue{Fig.}~\ref{fig:preliminary-augmentation-a}. Let $p_{vu}=v u_1 \dots u_k u$ and $p_{vw}=v w_1 \dots w_h w$ be the directed paths from $v$ to $u$ and from $v$ to $w$, respectively, that belong to the boundary of $f$. In order to construct $\Gamma_1$ and $G_1$ we insert, for some sufficiently small $\epsilon>0$, the following directed paths inside $f$ into $\Gamma$ and $G$:
\begin{figure}[tb]
\centering
\subfloat[\label{fig:preliminary-augmentation-a}]
{\includegraphics[page=1,width=.33\textwidth]{figures/gadget}}
\hfil
\subfloat[]
{\includegraphics[page=1,width=.33\textwidth]{figures/gadget2}\label{fig:preliminary-augmentation-b}}
\hfil
\subfloat[]
{\includegraphics[page=1,width=.33\textwidth]{figures/gadget3}\label{fig:preliminary-augmentation-c}}
\caption{Construction of the drawing $\Gamma'$; only what happens to the face $f$ is shown. (a) The drawing $\Gamma_1$ of $G_1$. The face $g$ is gray. (b) The drawing $\Gamma_2$ of $G_2$. (c) The drawing $\Gamma_4$ of $G_4$; the paths $p'$ and $p''$ are thick and red.}\label{fig:preliminary-augmentations}
\end{figure}
\begin{enumerate}[(a)]
\item a directed path $q_1=v v'w'_1\dots w'_h w$, where $w'_i$ is on the same horizontal line as $w_i$, at horizontal distance $\epsilon/2$ from it, while $v'$ is above $v$, at distance $\epsilon/2$ from it, along the bisector of the angle $\angle (v^-,v,v^+)$;
\item a directed path $q_2=v' v''w''_1\dots w''_h w$, where $w''_i$ is on the same horizontal line as $w_i$, at horizontal distance $\epsilon$ from it, while $v''$ is above $v$ and $v'$, at distance $\epsilon$ from $v$, along the bisector of the angle $\angle (v^-,v,v^+)$;
\item a directed path $q_3=v''u'_1\dots u'_k u'$, where $u'_i$ is on the same horizontal line as $u_i$, at horizontal distance $\epsilon$ from it, while $u'$ is above $u$, at distance $\epsilon$ from it, along the bisector of the angle $\angle (u^-,u,u^+)$; and
\item a directed edge $uu'$.
\end{enumerate}
It is easy to see that the resulting drawing $\Gamma_1$ has no crossings and that the directed paths $q_1$, $q_2$, $q_3$, and $uu'$ are upwardly drawn inside $f$, provided that $\epsilon$ is small enough. Note that there is an $st$-face $g$ of $G_1$ that is delimited by the directed path composed of $vv'v''$ and of $q_3$ and by the directed path composed of $p_{vu}$ and of $uu'$.
In {\bf phase 2} we augment $\Gamma_1$ to an upward planar drawing $\Gamma_2$ of a plane $st$-graph $G_2$. Refer to \blue{Fig.}~\ref{fig:preliminary-augmentation-b}. In order to construct $\Gamma_2$, we insert edges drawn as straight-line segments into every face of $\Gamma_1$, except for $g$, until no further edge can be inserted while maintaining planarity; each inserted edge is oriented from the endpoint with the lowest $y$-coordinate to the endpoint with the highest $y$-coordinate in $\Gamma_1$ (if the end-points of an edge have the same $y$-coordinate, then we insert two new adjacent vertices, slightly above and below the middle point of that edge, and then keep on inserting edges). This concludes the construction of the drawing $\Gamma_2$ of $G_2$. Since $\Gamma_1$ is upward and planar, it follows that $\Gamma_2$ is upward and planar as well. Further, $G_2$ is an $st$-graph by Lemma~\ref{le:st-faces-iff-st-graph}, since $g$ is an $st$-face, as argued above, and since all the faces of $G_2$ different from $g$ are also delimited by $st$-cycles, as otherwise more edges could have been introduced while maintaining the planarity of $\Gamma_2$; note that every internal face of $\Gamma_2$ different from $g$ is delimited by an upwardly drawn $3$-cycle, while more than $3$ vertices might be incident to the outer face of $\Gamma_2$.
In {\bf phase 3} we replace each directed edge $uv$ of $G_2$ that does not belong to $G$ (and has been inserted in phase 1 or 2) with a directed path $(u,w_{uv},v)$ and insert $w_{uv}$ at an arbitrary internal point of the edge $uv$ in $\Gamma_2$. Clearly, the resulting graph $G_3$ is a plane $st$-graph and it is reduced if $G$ is. Further, the resulting drawing $\Gamma_3$ is an upward planar drawing of $G_3$.
In {\bf phase 4} we augment $G_3$ to a plane $st$-graph $G_4$ by adding two directed edges $uv'$ and $uv''$ inside $g$. Observe that $G_4$ is a plane $st$-graph, by Lemma~\ref{le:st-faces-iff-st-graph}, since the directed edges $uv'$ and $uv''$ split $g$ into three $st$-faces. Further, $G_4$ is reduced if $G$ is. Let $p'$ be the directed path composed of $uv'$ and of the (subdivided) directed path $q_1$ and let $p''$ be the directed path composed of $uv'$ and of the (subdivided) directed path $q_2$. Observe that the $st$-cycle $\mathcal D$ of $G_4$ composed of $p'$ and $p''$ does not enclose any vertex of $G$, although it encloses vertices of $G_4$ not in $G$. We construct an upward planar straight-line drawing $\Gamma_4$ of $G_4$ by means of, e.g., the algorithm by Di Battista and Tamassia~\cite{DBLP:journals/tcs/BattistaT88}.
Now let $\Gamma' = \Gamma_4[G]$. Property~(i) then follows from the fact that upward planar drawings of the directed paths $p'$ and $p''$ from $u$ to $w$ can be inserted in $\Gamma'$ as they are drawn in $\Gamma_4$. Further, since $G_3$ has $O(n)$ vertices, by applying algorithm $\mathcal A$ ($\mathcal A_R$) we can construct an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph $\mathcal M_{3,4}$ from $\Gamma_3$ to $\Gamma_4[G_3]$. Since $\Gamma=\Gamma_3[G]$ and $\Gamma' = \Gamma_4[G]$, the restriction of $\mathcal M_{3,4}$ to the vertices and edges of $G$ provides an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph $\mathcal M'$ from $\Gamma$ to $\Gamma'$, which proves Property~(ii).
The drawing $\Gamma^*$ is constructed as follows.
First, we remove from $\Gamma_4$ and $G_4$ all the vertices and edges enclosed by $\mathcal D$. Let $\Gamma_5$ and $G_5$ be the resulting drawing and the resulting graph, respectively. We have that $G_5$ is a plane $st$-graph by Lemma~\ref{le:st-faces-iff-st-graph}; indeed, the face $d$ of $G_5$ delimited by $\mathcal D$ is an $st$-face, as $\mathcal D$ is composed of the directed paths $p'$ and $p''$, and every other face of $G_5$ is an $st$-face since it is also a face of $G_4$, which is a plane $st$-graph. Moreover, $\Gamma_5$ is an upward planar drawing of $G_5$, given that $\Gamma_4$ is an upward planar drawing of $G_4$.
Second, we augment $G_5$ to a plane $st$-graph $G_6$ by inserting the directed edge $uw$ inside $d$. We construct an upward planar straight-line drawing $\Gamma_6$ of $G_6$ by means of, e.g., the algorithm by Di Battista and Tamassia~\cite{DBLP:journals/tcs/BattistaT88}.
Now let $\Gamma^* = \Gamma_6[G]$. Property~(iii) then follows from the fact that the directed edge $uw$ lies inside $f$ and is upwardly drawn in $\Gamma_6$. Further, since $G_6$ has $O(n)$ vertices, by applying algorithm $\mathcal A$ ($\mathcal A_R$) we can construct an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph $\mathcal M_{5,6}$ from $\Gamma_5$ to $\Gamma_6[G_5]$. Since $\Gamma'=\Gamma_5[G]$ and $\Gamma^* = \Gamma_6[G]$, the restriction of $\mathcal M_{5,6}$ to the vertices and edges of $G$ provides an $O(f(n))$-step (an $O(f_R(n))$-step) upward planar morph $\mathcal M^*$ from $\Gamma'$ to $\Gamma^*$, which proves Property~(iv). This concludes the proof of the lemma.
%
\end{proof}
We are now ready to prove the following.
\newcommand{\metaalgorithmplane}{Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of an $n$-vertex (reduced) upward plane graph $G$.
Suppose that an algorithm $\mathcal A$ ($\mathcal A_R$) exists that constructs an $f(r)$-step (an $f_R(r)$-step) upward planar morph between any two upward planar drawings of an $r$-vertex (reduced) plane $st$-graph.
There exists an $O(n\cdot f(n))$-step (an $O(n \cdot f_R(n))$-step) upward planar morph from $\Gamma_0$ to~$\Gamma_1$.}
\begin{theorem}\label{th:meta-algorithm-plane}
\metaalgorithmplane
\end{theorem}
\begin{proof}
By \autoref{le:biconnected-reduced-augmentation}, we can assume that $G$ is biconnected.
Denote by $\ell(G)$ the number of switches labeled \texttt{large} in the upward planar embedding of $G$. In order to prove the statement, we show that there exists a \mbox{$((2\ell(G)-3) \cdot f(n))$-step} (a $((2\ell(G)-3) \cdot f_R(n))$-step) upward planar morph from $\Gamma_0$ to $\Gamma_1$, if $G$ is a (reduced) upward plane graph. Since $\ell(G) \in O(n)$, the statement follows. The proof is by induction on $\ell(G)$.
In the base case $\ell(G)=2$ and thus $G$ is a (reduced) plane $st$-graph (the two switches labeled \texttt{large} are those incident to the outer face of $G$). Hence, by applying algorithm $\mathcal A$ ($\mathcal A_R$) to $\Gamma_0$ and $\Gamma_1$, we obtain an $f(n)$-step (an $f_R(n)$-step) upward planar morph from $\Gamma_0$ to $\Gamma_1$.
In the inductive case $\ell(G)>2$. Then there exists a face $f$ of $G$ that is not an $st$-face. Thus, by \autoref{lem:small-small-large}, there exists a pocket $[u,v,w]$ for $f$. By \autoref{lem:morphing-between-augmentations}, we can construct upward planar drawings $\Gamma'_0$ and $\Gamma'_1$ of $G$ in which $u$ sees $w$ inside $f$ and in which $u$ lies below $w$ (assuming that a directed path exists in $f$ from $v$ to $u$, the other case being symmetric), and such that there exists an $f(n)$-step (an $f_R(n)$-step) upward planar morph $\mathcal M_{start}$ from $\Gamma_0$ to $\Gamma'_0$ and an $f(n)$-step (an $f_R(n)$-step) upward planar morph $\mathcal M_{finish}$ from $\Gamma'_1$ to $\Gamma_1$.
Let $G^*$ be the plane graph obtained from $G$ by splitting $f$ with a directed edge $uw$. The graph $G^*$ is an upward plane graph whose upward planar embedding is constructed by assigning to each switch in $G^*$ the same label \texttt{small} or \texttt{large} it has in $G$. Then $\ell(G^*) = \ell(G)-1$, since $u$ is not a switch in $G^*$. Further, $G^*$ is reduced if $G$ is reduced, since there exists no directed path in $G$ passing first through $u$ and then through $w$, given that $u$ is a sink of $G$.
Let $\Gamma^*_0$ and $\Gamma^*_1$ be the upward planar straight-line drawings of $G^*$ obtained by drawing the directed edge $uw$ as a straight-line segment connecting $u$ and $w$ in $\Gamma'_0$ and in $\Gamma'_1$, respectively. By the inductive hypothesis and since $V(G^*)=V(G)$, we can construct a $((2\ell(G^*)-3)\cdot f(n))$-step (a $((2\ell(G^*)-3)\cdot f_R(n))$-step) upward planar morph from $\Gamma^*_0$ to $\Gamma^*_1$. Observe that, since $G \subset G^*$, restricting each drawing in $\mathcal M^*$ to the vertices and edges of $G$ yields a $\big( (2\ell(G)-5)\cdot f(n) \big)$-step upward planar morph $\mathcal M^-$ of $G$ from $\Gamma'_0$ to $\Gamma'_1$. Therefore, by concatenating morphs $\mathcal M_{start}$, $\mathcal M^-$, and $\mathcal M_{finish}$, we obtain a $\big( (2\ell(G)-3) \cdot f(n) \big)$-step (a $\big( (2\ell(G)-3) \cdot f_R(n) \big)$-step) upward planar morph of $G$ from $\Gamma_0$ to $\Gamma_1$. This concludes the proof.
\end{proof}
\autoref{th:st-reduced-graphs}, \autoref{co:st-main}, and \autoref{th:meta-algorithm-plane} imply the following main result.
\begin{theorem}\label{th:upward-morph}
Let $\Gamma_0$ and $\Gamma_1$ be two upward planar drawings of the same $n$-vertex (reduced) upward plane graph. There exists an $O(n^2)$-step (an $O(n)$-step) upward planar morph from $\Gamma_0$ to $\Gamma_1$.
\end{theorem}
\section{Conclusions and Open Problems}\label{se:conclusions}
In this paper, we addressed for the first time the problem of morphing upward planar straight-line drawings. We proved that an upward planar morph between any two upward planar straight-line drawings of the same upward plane graph always exists; such a morph consists of a quadratic number of linear morphing steps. The quadratic bound can be improved to linear for reduced upward plane graphs and for plane $st$-graphs, and to constant for reduced plane $st$-graphs. All our algorithms can be implemented in polynomial time in the real RAM model.
Our algorithms assume the (undirected) connectivity of the upward planar graph whose drawings have to be morphed. However, we believe that the techniques presented in~\cite{DBLP:journals/siamcomp/AlamdariABCLBFH17} in order to deal with disconnected graphs can be applied also to our setting with only minor modifications.
Several problems are left open by our research. In our opinion the most interesting question is whether an $O(1)$-step upward planar morph between any two upward planar drawings of the same maximal plane $st$-graph exists. In case of a positive answer, by \autoref{th:st-graphs-maximal} and \autoref{th:meta-algorithm-plane}, an optimal $O(n)$-step upward planar morph would exist between any two upward planar drawings of the same $n$-vertex upward plane graph. In case of a negative answer, it would be interesting to find broad classes of upward plane graphs that admit upward planar morphs with a sub-linear number of steps. In particular, we ask whether series-parallel digraphs~\cite{DBLP:journals/ijcga/BertolazziCBTT94,DBLP:journals/siamcomp/CohenBTT95} admit upward planar morphs with $O(1)$ steps.
\bibliographystyle{abbrv}
|
\chapter*{Introducci\'on}\label{Introduccion}
\addcontentsline{toc}{chapter}{Introducci\'on}
\pagenumbering{arabic
Los datos de conteo pueden aparecer bajo diferentes circunstancias. En un marco univariante, la distribuci\'on Poisson es la distribuci\'on que con mayor frecuencia ha sido empleada para modelar tales datos (ver por ejemplo, Haight (1967, pp. 100--107) \cite{Hai67}, Johnson y Kotz (1969, pp. 88--90) \cite{JoKo69}, Sahai y Khurshid (1993) \cite{SaKh93}).
En la pr\'actica, los datos de conteo bivariantes surgen en varias disciplinas diferentes y la distribuci\'on Poisson bivariante (DPB), siendo una generalizaci\'on de la distribuci\'on Poisson, juega un rol importante al momento de modelarlos, siempre que dichos datos presenten una correlaci\'on no negativa.
Esta distribuci\'on ha sido usada para modelar datos que aparecen en un amplio rango de campos incluyendo medicina, para mediciones en el pretratamiento y en el postratamiento en los mismos pacientes, o el n\'umero de accidentes por trabajador en una factor\'ia durante dos intervalos dados de tiempo (ver por ejemplo, Hamdan (1972) \cite{Ham72}); en marketing, para el n\'umero de compras de diferentes productos; en epidemiolog\'ia, para el n\'umero de incidentes de distintos tipos de muertes en una serie de distritos; en deportes, para el n\'umero de goles marcados por cada uno de los dos equipos oponentes en un partido de balompi\'e (ver por ejemplo, Maher (1982) \cite{Mah82}, Karlis y Ntzoufras (2000) \cite{KaNt00}, (2003a) \cite{KaNt03a}, (2003b) \cite{KaNt03b}, (2005) \cite{KaNt05}, Rue y Salvesen (2000) \cite{RuSa00}); en biolog\'ia, para el n\'umero de semillas y plantas que crecen en una parcela (ver por ejemplo, Lakshminarayana, Pandit y Rao (1999) \cite{LaPaRa99}); en econometr\'ia, para el n\'umero de cambios de trabajo voluntarios e involuntarios (ver por ejemplo, Jung y Winkelmann (1993) \cite{JuWi93}); en datos de turismo (ver por ejemplo, Berkhout y Plug (2004) \cite{BePl04}); en seguros de coches (ver por ejemplo, Berm\'udez (2009) \cite{Ber09}); en sanidad (ver por ejemplo, Karlis y Ntzoufras (2005) \cite{KaNt05}, Karlis y Tsiamyrtzis (2008) \cite{KaTs08}); en la industria textil, para el n\'umero de dos tipos de defectos en muestras de fibras de textil (ver por ejemplo, Ho y Singer (2001) \cite{HoSi01}), entre muchos otros.
En el caso multivariante, la distribuci\'on Poisson ha sido usada para modelar redes sociales multi-relacionales (ver por ejemplo, Dai, Chua y Lim (2012) \cite{DaChLi12}).
Contrastar la bondad de ajuste de las observaciones dadas con un modelo probabil\'istico es un aspecto crucial del an\'alisis de datos. Para el caso univariante, se han construido muchos tests de bondad de ajuste con la finalidad de comprobar si los datos provienen de una distribuci\'on Poisson (para una revisi\'on detallada, ver G\"urtler y Henze, (2000) \cite{GuHe00}). En comparaci\'on, la literatura sobre tests de bondad de ajuste para la DPB es m\'as bien escasa. Hasta donde conocemos, podemos mencionar el test propuesto por Crockett (1979) \cite{Cro79}, el test desarrollado por Loukas y Kemp (1986) \cite{LoKe86}, que se basa en una extensi\'on del \'indice de dispersi\'on univariante, y el test sugerido por Rayner y Best (1995) \cite{RaBe95}, que consiste en una modificaci\'on del test dado por Loukas y Kemp (1986) \cite{LoKe86}. La principal desventaja de estos tests de bondad de ajuste es que no son consistentes contra cada alternativa fija.
El objetivo de esta texto es proponer y estudiar tests de bondad de ajuste para la DPB, que sean consistentes. Dado que la funci\'on generatriz de probabilidad (fgp) caracteriza la distribuci\'on de un vector aleatorio y se puede estimar consistentemente por la funci\'on generatriz de probabilidad emp\'irica (fgpe), los tests que proponemos son funciones de la fgpe. El primer test estad\'istico compara la fgpe de los datos con un estimador de la fgp de la DPB. Luego, mostramos que la fgp de la DPB es la \'unica fgp que satisface cierto sistema de ecuaciones diferenciales parciales, lo cual nos lleva a proponer dos tests estad\'isticos basados en el an\'alogo emp\'irico de dicho sistema, uno de ellos de tipo Cram\'er-von Mises y el otro se basa en los coeficientes de los polinomios de la versi\'on emp\'irica. Los tests que proponemos pueden ser vistos como extensiones al caso bivariante de algunos tests de bondad de ajuste dise\~nados para el caso univariante.
Con el fin de decidir cu\'ando rechazar la hip\'otesis nula, debemos conocer la distribuci\'on nula del test estad\'istico o, al menos, una aproximaci\'on de la misma. Puesto que, para los tests propuestos, no es posible obtener las distribuciones nulas para tama\~nos de muestra finito, las aproximamos por las asint\'oticas, las cuales resultaron depender de cantidades desconocidas, por lo tanto no son \'utiles como estimaciones de la distribuci\'on nula. As\'i, para aproximar la distribuci\'on nula, proponemos un estimador bootstrap param\'etrico.
En cuanto a la potencia, obtuvimos que los tests que proponemos son consistentes contra alternativas fijas. Adem\'as, analizamos el comportamiento asint\'otico de los tests estad\'isticos bajo alternativas contiguas y encontramos que son capaces de detectar alternativas que convergen a la nula a raz\'on de $n^{-1/2}$.
Todas las propiedades estudiadas de los tests introducidos en este texto son asint\'oticas, es decir, describen el comportamiento de los tests para muestras de tama\~no grande. Con la finalidad de evaluar el comportamiento de los tests propuestos para muestras de tama\~no finito, realizamos un estudio de simulaci\'on. En todos los casos considerados, el m\'etodo bootstrap proporciona una buena aproximaci\'on a la distribuci\'on nula. En cuanto a la potencia, a diferencia de los tests estad\'isticos dise\~nados por los investigadores Crockett (1979) \cite{Cro79}, Loukas y Kemp (1986) \cite{LoKe86}, y Rayner y Best (1995) \cite{RaBe95}, los tests que proponemos fueron capaces de detectar todas las alternativas seleccionadas.
Con el prop\'osito de mantener la notaci\'on tan simple como sea posible, desarrollamos el an\'alisis te\'orico para el caso bivariante, pero los m\'etodos se pueden extender de manera natural al caso multivariante.
El presente texto se organiza de la siguiente manera. En el Cap\'itulo \ref{Resultados previos y definiciones} presentamos algunos resultados preliminares que nos servir\'an en los cap\'itulos siguientes, tambi\'en damos la definici\'on de la DPB con algunas de sus propiedades y adem\'as, mostramos contrastes de bondad de ajuste para la distribuci\'on Poisson tanto para datos univariantes como para datos bivariantes.
El Cap\'itulo \ref{Estadisticos-tipoCramer-von-Mises} contiene los dos primeros tests estad\'isticos que proponemos, que son de tipo Cram\'er-von Mises. Aqu\'i, tambi\'en mostramos la distribuci\'on asint\'otica nula de los tests estad\'isticos y proporcionamos estimadores bootstraps consistentes. En la parte final, estudiamos la potencia de los tests propuestos frente a alternativas fijas y locales.
En el Cap\'itulo \ref{Estadistico-Wn} presentamos el tercer estad\'istico que proponemos y describimos sus caracter\'isticas. Estudiamos su distribuci\'on asint\'otica nula y aproximamos su distribuci\'on nula por medio de un estimador bootstrap consistente. Tambi\'en, analizamos su potencia frente a alternativas fijas y contiguas.
El Cap\'itulo \ref{Resultados-numericos} est\'a dedicado a mostrar los resultados de un estudio de simulaci\'on y la aplicaci\'on de los tests propuestos a dos conjuntos de datos reales. Dicho estudio de simulaci\'on fue realizado con el objetivo de evaluar el comportamiento de los tests que proponemos y comparar la potencia tanto entre ellos como con otros tests que encontramos en la literatura estad\'istica.
En el Cap\'itulo \ref{Expresiones-matematicas} entregamos las expresiones matem\'aticas de los tests que hemos desarrollado y damos algunos detalles t\'ecnicos que son muy \'utiles al momento de implementar algoritmos o subrutinas en alg\'un lenguaje de programaci\'on.
El Cap\'itulo \ref{Extensiones} muestra c\'omo los tests propuestos se pueden extender al caso multivariante general.
\chapter{Resultados previos y definiciones} \label{Resultados previos y definiciones}
\section{Notaci\'on}\label{Notacion}
\begin{itemize}
\item Todos los vectores a utilizar en este texto ser\'an vectores filas y $v^\top$ es el traspuesto del vector fila $v$.
\item Para cualquier vector $v,\, v_k$ denota su $k-$\'esima coordenada.
\item $\mathbb{N}_0=\{0,1,2,3,\ldots\}$.
\item $\Theta=\{\theta= \,(\theta_1,\theta_2,\theta_3)\in \mathbb{R}^3\ :\ \theta_1>\theta_3, \ \theta_2>\theta_3,\ \theta_3> 0\}$.
\item $\langle\, \cdot \, ,\cdot\,\rangle_{_V}$ denota el producto escalar en el espacio vectorial $V$.
\item $\|\cdot\|_{_V}$ denota la norma en el espacio vectorial $V$ y $\|\cdot\|$ denota la norma Euclidea de $\mathbb{R}^d,\, d\in \mathbb{N}$.
\item $I\{A\}$ denota la funci\'on indicadora del conjunto $A$.
\item $P_{\vartheta}$ denota la ley de probabilidades de una DPB con par\'ametro $\vartheta$.
\item $E_{\vartheta}$ denota la esperanza con respecto a la funci\'on de probabilidad $P_{\vartheta}$.
\item $\mathop{\longrightarrow} \limits^{L}\ $ denota convergencia en ley (o en distribuci\'on).
\item $\mathop{\longrightarrow} \limits^{P}\ $ denota convergencia en probabilidad.
\item $\mathop{\longrightarrow} \limits^{c.s.}\ $ denota convergencia casi segura (c.s.).
\item Si $\{C_n\}$ es una sucesi\'on de variables aleatorias (v.a.) y $\, \epsilon\in \mathbb{R}$, entonces
\begin{itemize}
\item \vskip -.1 cm $C_n=O_{_P}(n^{-\epsilon})\,$ significa que $\,n^{\epsilon}\,C_n\,$ est\'a acotada en probabilidad, cuando $\,n\rightarrow\infty$.
\item $C_n=o_{_P}(n^{-\epsilon})\,$ significa que $\,n^{\epsilon}\,C_n\ \mathop{\longrightarrow} \limits^{\!P} \ 0$, cuando $\,n\rightarrow\infty$.
\item $C_n=o(n^{-\epsilon})\,$ significa que $\,n^{\epsilon}\,C_n\ \mathop{\longrightarrow} \limits^{\!c.s.} \ 0$, cuando $\,n\rightarrow\infty$.
\end{itemize}
\end{itemize}
\section{Resultados preliminares}
Para demostrar la Proposici\'on \ref{Conv-FuncGenProbBiv}, dada a continuaci\'on, primero presentaremos un lema, que es una extensi\'on del Lema 8.2.6 en Athreya y Lahiri (2006, p. 242) \cite{AtLa06}. Previamente, consideremos la siguiente notaci\'on: para un conjunto arbitrario $S$, $\partial S$ e $int S$ denotan los conjuntos de puntos frontera y puntos interiores de $S$, respectivamente.
\begin{lema} \label{polya-d}
Sean $\{f_n\}_{n\geq 1}$ y $f$ una colecci\'on de funciones reales y no decrecientes definidas sobre $Q=[b_1,c_1]\times [b_2,c_2]\times\cdots\times [b_d,c_d] \subseteq \mathbb{R}^d$, con $-\infty<b_j\leq c_j<\infty$, $j=1,2,\ldots,d$. Sea $D=D_1\cup D_2 \cup D_3$, donde $D_1$ es el conjunto de v\'ertices de $Q$, $D_2$ es un conjunto denso en $\partial Q$ y $D_3$ es un conjunto denso en $int Q$. Si $f$ es continua en $Q$ y
\[\left|f_{n}(x)-f(x)\right|\to 0, \quad \forall x \in D,\]
entonces \[\displaystyle \sup_{x\in Q}|f_n(x)-f(x)| \to 0.\]
\end{lema}
\nt {\bf Demostraci\'on} \hspace{2pt} Sea $\varepsilon>0$ arbitrario pero fijo. Como $f$ es una funci\'on continua y $Q$ es un conjunto compacto, entonces $f$ es uniformemente continua en $Q$ y por tanto
\begin{equation}\label{f(x)-f(y)<e}
\exists\delta=\delta(\varepsilon)>0\ \, \text{tal que}\ \, \forall x,y\in Q:\, \|x-y\|<\delta\, \Longrightarrow \, |f(x)-f(y)|<\varepsilon.
\end{equation}
Sean
\[H_i=[u_{1i},v_{1i}]\times [u_{2i},v_{2i}]\times\cdots\times [u_{di},v_{di}],\ \ i=1,2,\ldots,M,\]
con $\|u_i-v_i\|< \delta$, $u_i=(u_{1i},u_{2i},\ldots,u_{di})\in D$ y $v_i=(v_{1i},v_{2i},\ldots,v_{di})\in D$, $1\leq i \leq M$, de modo que
\[Q\subseteq \bigcup_{i=1}^M H_i.\]
Consideremos
\begin{equation}\label{Delta-n}
\Delta_n=\max_{x\in V} \left |f_n(x)-f(x)\right |,
\end{equation}
donde $V=\{u_i,v_i: i=1,2,\ldots,M\}$.
Por la convergencia puntual de $f_n(\cdot)$ a $f(\cdot)$ en $D$, obtenemos
\begin{equation}\label{Delta-n-a-cero}
\Delta_n \to 0.
\end{equation}
Por otro lado, para cada $x\in Q$, existe $i\in\{1,2,\ldots,M\}$, tal que $x\in H_i$, con lo cual, $u_i\leq x\leq v_i$.
La monoticidad de las funciones $f_n$ y $f$, junto con (\ref{f(x)-f(y)<e}) y (\ref{Delta-n}), implican
\[f_n(x)-f(x)\leq f_n(v_i)-f(u_i)=f_n(v_i)-f(v_i)+f(v_i)-f(u_i)\leq \Delta_n+\varepsilon.\]
An\'alogamente
\[f_n(x)-f(x)\geq -\Delta_n-\varepsilon.\]
De (\ref{Delta-n-a-cero}) y puesto que $\varepsilon >0$ es arbitrario se logra
\[\sup_{x\in Q}|f_n(x)-f(x)| \to 0.\ \square\]
Feuerverger (1989) \cite{Fue89} demostr\'o que la funci\'on generatriz de momentos emp\'irica converg\'ia c.s. a la funci\'on generatriz de momentos (fgm), en conjuntos compactos, para v.a.
Aunque existe una relaci\'on uno a uno entre la fgm y la fgp, hay problemas en el origen, por esta raz\'on, probaremos a continuaci\'on, que la fgpe converge c.s. a la fgp, para el caso multivariante en general, en conjuntos de la forma $Q=[b_1,c_1]\times [b_2,c_2]\times\cdots\times [b_d,c_d]$, donde $0\leq b_j\leq c_j<\infty,\ 1\leq j\leq d$, $d\in\mathbb{N}$.
Tambi\'en probaremos, que las derivadas de la fgpe convergen c.s. a las derivadas de la fgp, en conjuntos $Q$ como los definidos en el p\'arrafo anterior. Sea $d\in\mathbb{N}$. Para cualquier funci\'on $h:S\subseteq\mathbb{R}^d \to \mathbb{R}$, denotaremos
\[D^{k_1k_2\cdots k_d}h(u)=\frac{\partial^k }{\partial u_1^{k_1} \partial u_2^{k_2}\cdots \partial u_d^{k_d}}\ h(u),\]
$\forall k_1,k_2,\ldots,k_d\in \mathbb{N}_0$ tal que $k=k_1+k_2+\cdots+k_d$.
Sean $\boldsymbol{X}_1=(X_{11},\ldots,X_{d1}), \boldsymbol{X}_2=(X_{12},\ldots,X_{d2}),\ldots, \boldsymbol{X}_n=(X_{1n},\ldots,X_{dn})$ vectores aleatorios independientes e id\'enticamente distribuidos (iid) definidos sobre el espacio de probabilidad $(\Omega,\mathcal{A},P)$ y que toman valores en $\mathbb{N}_0^d$. En lo que sigue, sea
\[g_n(u)=\frac{1}{n}\sum_{i=1}^n u_1^{X_{1i}}u_2^{X_{2i}}\cdots u_d^{X_{di}},\quad u\in W,\]
la fgpe de $\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$, para alg\'un conjunto apropiado $W \subseteq \mathbb{R}^d$. De manera m\'as formal,
\[g_n(u,\omega)=\frac{1}{n}\sum_{i=1}^n u_1^{X_{1i}(\omega)}u_2^{X_{2i}(\omega)}\cdots u_d^{X_{di}(\omega)},\quad u\in W,\quad \omega\in \Omega,\]
pero la dependencia de $\omega$ es usualmente suprimida.
\begin{proposicion}\label{Conv-FuncGenProbBiv}
Sean $\boldsymbol{X}_1,\ldots,\boldsymbol{X}_n$ vectores aleatorios iid de $\boldsymbol{X}=(X_1,\ldots,X_d)\in \mathbb{N}_0^d$. Sea $g(u)=E\!\left(u_1^{X_1}u_2^{X_2}\cdots u_d^{X_d}\right)$ la fgp de $\boldsymbol{X}$, definida sobre $W \subseteq \mathbb{R}^d$. Adem\'as, sean $0\leq b_j\leq c_j<\infty,\ 1\leq j\leq d$, tal que $Q=[b_1,c_1]\times [b_2,c_2]\times\cdots\times [b_d,c_d] \subseteq W$, entonces
\begin{equation}\label{gn-converge-g}
\sup_{u\in Q} |g_n(u)-g(u)|\ \mathop{\longrightarrow}\limits^{c.s.} \ 0.
\end{equation}
Si $D^{k_1k_2\cdots k_d}g(u)$ existe en $Q$, entonces
\begin{equation}\label{Dgn-converge-Dg}
\sup_{u\in Q} \left|D^{k_1k_2\cdots k_d} g_n(u)-D^{k_1k_2\cdots k_d} g(u)\right|\ \mathop{\longrightarrow}\limits^{c.s.} \ 0.
\end{equation}
\end{proposicion}
\nt {\bf Demostraci\'on} \hspace{2pt} Sea $D$ un conjunto denso numerable en $Q$, de acuerdo al Lema \ref{polya-d}. Por la ley fuerte de los grandes n\'umeros, existe un conjunto $A_D \in \mathcal{A}$ tal que $P(A_D)=1$ y para todo $\omega \in A_D$: \[g_n(u,\omega)\ \mathop{\longrightarrow}\limits^{c.s.} \ g(u), \forall u \in D.\]
Puesto que $X_i\geq 0\,$ y $\,u_i\geq 0$, $1\leq i \leq d$, tenemos que $g_n(u,\omega)$ y $g(u)$ son funciones no decrecientes.
Adem\'as, $g$ es una funci\'on continua definida sobre $Q$.
Ahora, (\ref{gn-converge-g}) se sigue del Lema \ref{polya-d}.
La demostraci\'on de (\ref{Dgn-converge-Dg}) sigue pasos similares a los dados para demostrar (\ref{gn-converge-g}). $\square$\\
Para probar algunos de nuestros resultados aplicaremos los lemas que establecemos a continuaci\'on, los que presentamos aqu\'i para facilitar la lectura de nuestras demostraciones.
\begin{lema}[\textnormal{M\'etodo Delta}]\label{Metodo-Delta}
(Lehmann y Romano(2005, p. 436) \cite{LehRom05})\
Supongamos que $X$ y $X_1, X_2, \ldots\,$ son vectores aleatorios en $\mathbb{R}^k$. Supongamos que $\ \displaystyle\tau_n(X_n-\mu)\ \mathop{\longrightarrow}^L \ X$ donde $\mu$ es un vector constante y $\tau_n$ es una sucesi\'on de constantes tales que $\ \tau_n\to \infty$.
\begin{enumerate}
\item [$(a)$] Supongamos que $g$ es una funci\'on desde $\mathbb{R}^k$ a $\mathbb{R}$ que es diferenciable en $\mu$ con gradiente (vector de primeras derivadas parciales) de dimensi\'on $1\times k$ en $\mu$ igual a $\dot{g}(\mu)\neq \mathbf{0}$. Entonces, \[\tau_n\left[\,g(X_n)-g(\mu)\right] \ \mathop{\longrightarrow}^L \ \dot{g}(\mu)X^{\top}.\] En particular, si $X$ es normal multivariante en $\mathbb{R}^k$ con vector de media $\mathbf{0}$ y matriz de covarianzas $\Sigma$, entonces \[\tau_n\left[\,g(X_n)-g(\mu)\right] \ \mathop{\longrightarrow}^L\ N\left(\mathbf{0},\dot{g}(\mu)\,\Sigma\ \dot{g}(\mu)^{\top}\right).\]
\item [$(b)$]\vskip .2 cm M\'as generalmente, supongamos que $g=(g_1,g_2,\ldots,g_q)^{\top}$ es una funci\'on desde $\mathbb{R}^k$ a $\mathbb{R}^q$, donde $g_i$ es una funci\'on desde $\mathbb{R}^k$ a $\mathbb{R}$ que es diferenciable en $\mu$. Sea $D$ una matriz no nula de orden $q\times k$, donde el elemento $(i,j)$ es $\displaystyle\frac{\partial}{\partial y_j}g_i(y_1,y_2,\ldots,y_k)$ evaluada en $\mu$. Entonces \begin{equation}\notag \tau_n\left[\,g(X_n)-g(\mu)\right]=\tau_n\left[\,g_1(X_n)-g_1(\mu), \ldots,g_q(X_n)-g_q(\mu)\right]^{\top} \ \mathop{\longrightarrow}^L \ DX^{\top}. \end{equation} En particular, si $X$ es normal multivariante en $\mathbb{R}^k$ con vector de media $\mathbf{0}$ y matriz de covarianzas $\Sigma$, entonces \begin{equation}\notag\tau_n\left[\,g(X_n)-g(\mu)\right] \ \mathop{\longrightarrow}^L\ N\left(\mathbf{0},D\Sigma D^{\top}\right). \end{equation}
\end{enumerate}
\end{lema}
Para los lemas siguientes, con $H$ denotaremos un espacio de Hilbert de dimensi\'on infinita, separable y real.
\begin{lema}\label{TCL-Hilbert} (Teorema central del l\'imite en espacios de Hilbert, van der Vaart y Wellner (1996, pp. 50--51)\cite{vanWel96}) Si $X_1, X_2, X_3, \ldots$ son elementos aleatorios iid medibles \mbox{Borel} en un espacio de Hilbert $H$ con media cero (es decir, $E\left(\langle X_1, h\rangle_{_H}\right)=0$ para cada $h$), y $E\left(\|X_1\|^2_{_H}\right)< \infty$, entonces la sucesi\'on $\displaystyle \frac{1}{\sqrt{n}}\sum_{i=1}^n X_i$ converge en distribuci\'on a la variable Gaussiana $G$. La distribuci\'on de $G$ est\'a determinada por la distribuci\'on de sus marginales $\langle G, h\rangle_{_H}$, que se distribuyen seg\'un una ley $N\left(0,E\left(\langle X, h\rangle^2_{_H}\right)\right)$ para cada $h\in H$.
\end{lema}
\begin{lema} \label{kundu} (Teorema 1.1 en Kundu et al. (2000) \cite{KuMaMu00}) Sea $\{e_k: \, k \geq 0\}$ una base ortonormal de $H$. Para cada $n\geq 1$, sea $Y_{n1}, Y_{n2}, \ldots, Y_{nn}$ un arreglo triangular de elementos aleatorios independientes $H$-valuados con medias cero y momentos segundos finitos, sea $Y_n=\displaystyle \sum_{j=1}^n Y_{nj}$. Sea $C_n$ el operador de covarianza de $Y_n$. Supongamos que se cumplen las siguientes condiciones:
\begin{itemize} \itemsep=0pt
\item[(i)] $\displaystyle \lim_{n\to \infty} \langle C_ne_k,e_l\rangle_{_{\!H}}=a_{kl}$, existe para todo $ k,l \geq 0$.
\item[(ii)] \vskip .1 cm $\displaystyle \lim_{n\to \infty}\sum_{k=0}^{\infty}\langle C_ne_k,e_k\rangle_{_{\!H}}=\sum_{k=0}^{\infty} a_{kk}<\infty $.
\item[(iii)] \vskip .1 cm $\displaystyle \lim_{n\to \infty} L_n(\varepsilon, e_k)=0$ para cada $\varepsilon>0$ y cada $k \geq 0$, donde, para $b \in H$, \[L_n(\varepsilon, b)=\sum_{j=1}^n E\left(\langle Y_{nj}, b\rangle_{_{\!H}}^2 \,I{\left\{|\langle Y_{nj}, b\rangle_{_{\!H}}|>\varepsilon\right\}}\right).\]
\end{itemize}
Entonces \[Y_n \stackrel{L}{\longrightarrow} N(0,C),\] en $H$, donde el operador de covarianza $C$ es caracterizado por
$\langle C f,e_l\rangle_{_{\!H}}=\displaystyle \sum_{k=0}^{\infty}\langle f,e_k \rangle_{_{\!H}} a_{kl}$, para cada $l \geq 0$.
\end{lema}
\section{Definici\'on de la distribuci\'on Poisson bivariante}
Se han dado varias definiciones para la DPB (ver por ejemplo, Kocherlakota y Kocherlakota (1992, pp. 87--90) \cite{KoKo92}, para una revisi\'on detallada). En este texto, consideraremos la siguiente, que es la que ha recibido m\'as atenci\'on en la literatura estad\'istica (ver por ejemplo, Holgate (1964) \cite{Hol64}; Johnson, Kotz y Balakrishnan (1997) \cite{JoKoBa97}).
\begin{definicion}\label{Dist_Pois_Biv}
(Johnson, Kotz y Balakrishnan (1997, pp. 124--125) \cite{JoKoBa97}) \hspace{2pt} Sean
\[X_1=Y_1+Y_{3}\, \quad \textnormal{ y } \quad X_2=Y_2+Y_{3}\, ,\
donde $\,Y_1,\,Y_2\,$ e $\,Y_{3}\,$ son v.a. Poisson, mutuamente independientes con medias dadas por $\theta'_1=\theta_1-\theta_{3}>0$, $\theta'_2=\theta_2-\theta_{3}>0\ $ y $\ \theta_{3}\geq0$, respectivamente.
A la distribuci\'on conjunta del vector $(X_1, X_2)$ se le denomina {\bf distribuci\'on Poisson bivariante} (DPB) con par\'ametro $\theta=(\theta_1, \theta_2, \theta_{3})$, lo cual denotaremos mediante $(X_1, X_2)\sim PB(\theta_1, \theta_2, \theta_{3})$ o simplemente $(X_1, X_2)\sim PB(\theta)$.
La funci\'on de probabilidad conjunta de $X_1$ y $X_2$ est\'a dada por
\[P_{\theta}(X_1=x_1,X_2=x_2)=\exp(\theta_{3}-\theta_{1}-\theta_2) \sum_{i=0}^{\min\{x_{1},x_2\}} \frac{(\theta_{1}-\theta_{3})^{x_{1}-i}\,(\theta_2-\theta_{3})^{x_2-i}\, \theta_{3}^i}{(x_1-i)!\ (x_2-i)!\ i!},\]
donde $x_1, x_2\in \mathbb{N}_0$.
\end{definicion}
\begin{observacion}
Si $\theta_3=0$, entonces $X_1=Y_1$ y $X_2=Y_2$, y por tanto $X_1$ y $X_2$ son independientes, pues $Y_1$ e $Y_2$ son v.a. Poisson mutuamente independientes. En los Cap\'itulos \ref{Estadisticos-tipoCramer-von-Mises} y \ref{Estadistico-Wn} supondremos que $\theta_3>0$. El caso $\theta_3=0$ ser\'a descrito en el Cap\'itulo \ref{Extensiones}.
\end{observacion}
Tal como sucede en el caso univariante, una de las formas como se obtuvo la funci\'on de probabilidad conjunta de la DPB fue como el l\'imite de la distribuci\'on binomial bivariante.
En primer lugar daremos la definici\'on de la distribuci\'on binomial bivariante y luego presentaremos el resultado que aproxima la distribuci\'on binomial bivariante a la DPB, cuya demostraci\'on puede verse en Hamdan y Al-Bayyati (1969) \cite{HaAl69}.
\begin{definicion}\label{Dist-Binom-Biv}(Johnson, Kotz y Balakrishnan (1997, p. 125) \cite{JoKoBa97}) \hspace{2pt}
Supongamos que cada individuo de una poblaci\'on es clasificado ya sea como $A$ o $A^{c}$ y simult\'aneamente como $B$ o $B^{c}$, con probabilidades dadas por
\begin{center}
\begin{tabular}[t]{c|cc|c}
&$B$&$B^c$&\\
\hline
$A$&$p_{11}$&$p_{10}$&$p_1$\\
$A^c$&$p_{01}$&$p_{00}$&$q_1$\\
\hline &$p_2$&$q_2$&$1$
\end{tabular}
\end{center}
Consideremos una muestra aleatoria de tama\~no $n$, seleccionada con reemplazo de la poblaci\'on anterior. Sean las v.a.:
$X_1=$ n\'umero de individuos que son clasificados como $A$,
$X_2=$ n\'umero de individuos que son clasificados como $B$.
Estas v.a. tienen distribuci\'on binomial bivariante conjunta, con funci\'on de pro\-babilidad
\[P(X_1=x_1, X_2=x_2) = \sum_{k=0}^{\min\{x_1,
x_2\!\}}\frac{n! \ p_{11}^k \, p_{10}^{x_1-k}\, p_{01}^{x_2-k}\,
p_{00}^{n+k-x_1-x_2}}{k!\ (x_1-k)!\ (x_2-k)!\ (n+k-x_1-x_2)!}\,,\]
lo que se representar\'a por $(X_1,X_2)\sim BB(n; p_{10}, p_{01}, p_{11})$.
\end{definicion}
\begin{teorema}\label{Aprox-Binom-Biv}
(Hamdan y Al-Bayyati (1969) \cite{HaAl69}) \newline
Sea $(X_1,X_2)\sim BB(n; p_{10}, p_{01}, p_{11})$. Suponer que $\,p_{10}, p_{01}, p_{11}\rightarrow 0$, cuando $\,n \rightarrow \infty,\,$ de modo que $\,n p_{10}=\theta_1-\theta_3, \, n p_{01}=\theta_2-\theta_3\,$ y $\,n p_{11}=\theta_3$. Entonces
\begin{align}
\lim_{n \rightarrow \infty}P(X_1=x_1 &, X_2=x_2)\notag\\
&\!=
\exp(\theta_3-\theta_1-\theta_2)\sum_{k=0}^{\min\{x_1,
x_2\}}\frac{(\theta_1-\theta_3)^{x_1-k} (\theta_2-\theta_3)^{x_2-k}\,
\theta_3^k}{(x_1-k)!\, (x_2-k)!\, k!}, \ x_1, x_2\in \mathbb{N}_0.\notag
\end{align}
\end{teorema}
\section[Algunas caracter\'isticas y propiedades de la DPB]{Algunas caracter\'isticas y propiedades de la DPB}
A continuaci\'on presentamos ciertos resultados de la DPB que nos ser\'an de utilidad en el desarrollo de este texto. Quiz\'as el de mayor importancia es el dado a continuaci\'on, pues es la base de nuestro trabajo.
Para ello consideremos el vector aleatorio $(X_1,X_2)\sim PB(\theta)$, como el establecido en la Definici\'on \ref{Dist_Pois_Biv}.
\subsection{Funci\'on generatriz de probabilidad}
Para deducir la fgp de la DPB, recordemos que la fgp de una v.a. $X$ que se distribuye Poisson (univariante) con par\'ametro $\lambda > 0$, se define y calcula mediante
\[E_{\lambda}\!\left(t^X\right) = \sum_{x=0}^\infty t^x \frac{\lambda^x\exp(-\lambda)}{x!}
= \exp(-\lambda)\sum_{x=0}^\infty \frac{(\lambda t)^x}{x!}=
\exp\{\lambda(t-1)\},\ \forall\,
t\in \mathbb{R}.\]
Ahora, como las v.a. $Y_1, Y_2$ e $Y_3$ se distribuyen seg\'un una ley de Poisson, entonces sus fgp se pueden expresar por
\[E_{\theta'_1}\!\left(t^{Y_1}\right)=\exp\{\theta'_1(t-1)\},\ \, E_{\theta'_2}\!\left(t^{Y_2}\right)=\exp\{\theta'_2(t-1)\},\ \, E_{\theta_3}\!\left(t^{Y_3}\right)=\exp\{\theta_3(t-1)\},\, \forall\,
t\!\in\! \mathbb{R}.\]
Adem\'as, como las v.a. $Y_1, Y_2$ e $Y_3$ son mutuamente independientes, entonces, la fgp conjunta, $g(u;\theta)$, de la DPB se obtiene mediante
\begin{align}
g(u;\theta)&=E_{\theta}\!\left(u_1^{X_1}\,u_2^{X_2} \right),\notag\\[.3 cm]
&=E_{\theta}\!\left(u_1^{Y_1+Y_3}\,u_2^{Y_2+Y_3}\right)= E_{\theta'_1}\!\left(u_1^{Y_1}\right)E_{\theta'_2}\!\left(u_2^{Y_2}\right) E_{\theta_3}\!\left\{(u_1u_2)^{Y_3}\right\}\notag\\[.3 cm]
&=\exp\{\theta'_1(u_1-1)+\theta'_2(u_2-1)+ \theta_3(u_1u_2-1)\}, \notag\\[.3 cm]
&=\exp\{(\theta_1-\theta_3)(u_1-1)+(\theta_2-\theta_3)(u_2-1)+ \theta_3(u_1u_2-1)\}, \notag\\[.3 cm]
&=\exp\{\theta_1(u_1-1)+\theta_2(u_2-1)+\theta_3(u_1-1)(u_2-1) \},\label{fgp3-DPB}
\end{align}
$\forall\,u=(u_1,u_2)\in \mathbb{R}^{2}$, $\forall\,\theta\in\Theta$.
\subsection{Funci\'on generatriz de momentos}
Recordemos que la fgm de una v.a. $X$ que se distribuye seg\'un una ley de Poisson (univariante) con par\'ametro $\lambda > 0$, se define y calcula mediante
\[M_X(t) = E_{\lambda}\!\left\{\exp(tX)\right\}=\exp(-\lambda)\sum_{x=0}^\infty
\frac{\left(\lambda e^t\right)^x}{x!}= \exp\!\left\{\lambda\left(e^t-1\right)\right\},\ \forall\, t\in \mathbb{R}.\]
Puesto que las v.a. $Y_1, Y_2$ e $Y_3$ se distribuyen seg\'un una ley de Poisson, entonces sus fgm se pueden escribir como
\[M_{Y_1}(t)=\exp\!\left\{\theta'_1\!\left(e^t\!-\!1\right)\right\}\!, \, M_{Y_2}(t)=\exp\!\left\{\theta'_2\!\left(e^t\!-\!1\right)\right\}\!, \, M_{Y_3}(t)=\exp\!\left\{\theta_3\!\left(e^t\!-\!1\right)\right\}\!, \forall\,t\!\in\! \mathbb{R}.\]
Adem\'as, como las v.a. $Y_1, Y_2$ e $Y_3$ son mutuamente independientes, entonces, la fgm conjunta, se obtiene como
\begin{align}
M_{(X_1, X_2)}(u_1, u_2) &= E_{\theta}\!\left\{\exp(u_1X_1+u_2X_2)\right\}\notag\\[.3 cm]
&= E_{\theta}\!\left[\exp\!\left\{u_1Y_1+u_2Y_2+(u_1+u_2)Y_3\right\}\right]\notag\\[.3 cm]
&= E_{\theta'_1}\!\!\left(e^{u_1 Y_1}\right)E_{\theta'_2}\!\!\left(e^{u_2 Y_2}\right)E_{\theta_3}\!\!\left\{e^{(u_1+u_2)Y_3}\right\}\notag\\[.3 cm]
&=\exp\!\left\{\theta'_1(e^{u_1}-1)+\theta'_2(e^{u_2}-1)+\theta_3
(e^{u_1+\,u_2}-1)\right\},\notag\\[.3 cm]
&=\exp\!\left\{(\theta_1-\theta_3)\!\left(e^{u_1}-1\right)+(\theta_2- \theta_3)\!\left(e^{u_2}-1\right)+\theta_3\!\left(e^{u_1+u_2}-1\right)\right\}, \notag\\[.3 cm]
&=\exp\!\left\{\theta_1\!\left(e^{u_1}-1\right)+\theta_2\!\left(e^{u_2}-1\right)+ \theta_3\!\left(e^{u_1}-1\right)\!\left(e^{u_2}-1\right)\right\},\notag
\end{align}
$\forall\,u=(u_1,u_2)\in \mathbb{R}^{2}$, $\forall\,\theta\in\Theta$.
\subsection{Momentos}\label{Momentos-DPB}
Recordemos que si una v.a. $X$ se distribuye seg\'un una ley de Poisson (univariante) con par\'ametro $\lambda > 0$, entonces su $k-$\'esimo momento en torno al cero es
\[\mu_k^{'}=E_{\lambda}\!\left(X^k\right)= \sum_{i=0}^k \lambda^i\,S(k,i),\]
donde $S(a,b)$ es llamado {\bf n\'umero Stirling de segundo tipo} y satisface las relaciones
\begin{align}
S(n,0)&=0,\ S(0,0)=S(n,1)=S(n,n)=1, \ \text{para}\ n\in \mathbb{N},\notag\\[.2 cm]
S(n,j)=S&(n-1,j-1)+jS(n-1,j), \ \text{para}\
j=1,2,\ldots,n-1.\notag
\end{align}
En particular, para $k=1,2,3,$ obtenemos
\[E_{\lambda}\!\left(X\right)=\lambda,\quad E_{\lambda}\!\left(X^2\right)= \lambda+\lambda^2,\quad E_{\lambda}\!\left(X^3\right)=\lambda+3\lambda^2+ \lambda^3.\]
El correspondiente momento central de $X$ es
\[\mu_k=E_{\lambda}\!\left\{(X-\lambda)^k\right\} = \sum_{i=0}^k
\binom{k}{i}(-\lambda)^{k-i}\, \mu_i^{'}.\]
Para el modelo Poisson bivariante, el $r=(r_1,r_2)-$\'esimo momento en torno al origen es
\begin{align}
\mu_{r_1,r_2}^{'}& =\mu_r^{'}(X_1,X_2)= E_{\theta}\!\left(X_1^{r_1}X_2^{r_2}\right)= E_{\theta}\!\left\{(Y_1+Y_3)^{r_1}(Y_2+Y_3)^{r_2}\right\}\notag\\[.2 cm]
&=E_{\theta}\!\left\{\sum_{i_1=0}^{r_1} \binom{r_1}{i_1} Y_1^{i_1}\
Y_3^{r_1-i_1}\,\sum_{i_2=0}^{r_2} \binom{r_2}{i_2} Y_2^{i_2}\ Y_3^{r_2-i_2}\right\}\notag\\[.2 cm]
&=\sum_{i_1=0}^{r_1}\sum_{i_2=0}^{r_2}\binom{r_1}{i_1}\binom{r_2}{i_2}
E_{\theta'_1}\!\left(Y_1^{i_1}\right)\, E_{\theta'_2}\!\left(Y_2^{i_2}\right)\, E_{\theta_3}\!\left(Y_3^{r_1+r_2-i_1-i_2}\right)\notag\\[.2 cm]
&=\sum_{i_1=0}^{r_1}\sum_{i_2=0}^{r_2}\binom{r_1}{i_1}\binom{r_2}{i_2}
\sum_{j_1=0}^{i_1}(\theta_1-\theta_3)^{j_1}S(i_1,j_1)
\sum_{j_2=0}^{i_2}(\theta_2-\theta_3)^{j_2} S(i_2,j_2)\notag\\
&\hspace{76 mm} \times\sum_{j_3=0}^{r_1+r_2-i_1-i_2}\theta_3^{j_3}
S(r_1+r_2-i_1-i_2,j_3).\notag
\end{align}
En particular,
\begin{align}
E_{\theta}(X_k) &= \theta_{k},\ \ \text{para} \ k=1,2,\notag\\[.15 cm]
E_{\theta}(X^2_k) &= \theta_{k}+\theta_{k}^2,\ \ \text{para} \ k=1,2,\notag\\[.15 cm]
E_{\theta}(X_1\,X_2) &= \theta_{1}\,\theta_{2}+\theta_3,\notag\\[.15 cm]
E_{\theta}(X_1^2\,X_2) &= \theta_{1}\,\theta_{2}+\theta_{1}^2\,\theta_{2}+2\,\theta_{1}\,\theta_3+ \theta_3,\notag\\[.15 cm]
E_{\theta}(X_1\,X_2^2) &= \theta_{1}\,\theta_{2}+\theta_{1}\,\theta_{2}^2+2\,\theta_{2}\,\theta_3+ \theta_3,\notag\\[.15 cm]
E_{\theta}(X_1^2\,X_2^2) &= \theta_{1}\,\theta_{2}+\theta_{1}\,\theta_{2}^2+\theta_{1}^2\,\theta_{2}+ \theta_{1}^2\,\theta_{2}^2+4\,\theta_{1}\,\theta_{2}\,\theta_3+2\,\theta_{1} \,\theta_3+2\,\theta_{2}\,\theta_3+\theta_3+2\,\theta_3^2.\notag
\end{align}
\begin{observacion} De los resultados dados en las ecuaciones anteriores se sigue que la DPB tiene momentos de todos los \'ordenes.
\end{observacion}
Escribiremos el correspondiente $r=(r_1,r_2)-$\'esimo momento central para el modelo Poisson bivariante mediante
\begin{align}
\mu_{r_1,r_2} &= \mu_r(X_1,X_2)=E_{\theta}\! \left[\{X_1-E_{\theta}(X_1)\}^{r_1}\{X_2-E_{\theta}(X_2)\}^{r_2}\right] \notag\\[.15 cm] &=E_{\theta}\!\left[\{X_1-\theta_1\}^{r_1}\{X_2-\theta_2\}^{r_2}\right]\notag\\[.15 cm]
&= \sum_{i=0}^{r_1}\sum_{j=0}^{r_2}\binom{r_1}{i}\binom{r_2}{j} (-\theta_1)^{r_1-i}(-\theta_2)^{r_2-j}E_{\theta}\!\left(X_1^iX_2^j\right) \notag\\[.15 cm]
&= \sum_{i=0}^{r_1}\sum_{j=0}^{r_2}\binom{r_1}{i}\binom{r_2}{j} (-\theta_1)^{r_1-i}(-\theta_2)^{r_2-j} \mu_{i,j}^{'}.\notag
\end{align}
Por otra parte, como las v.a. $X_1$ y $X_2$ se distribuyen seg\'un una ley de Poisson (univariante), entonces sus varianzas son
\[\text{var}(X_k) = \theta_k,\ k=1, 2.\]
Adem\'as, como $Y_1$, $Y_2$ e $Y_3$ son variables independientes, la covarianza entre $X_1$ y $X_{2}$ es
\begin{align}
\text{cov}\left(X_1,X_2\right) &= \text{cov}\left(Y_1+Y_3,
Y_2+Y_3\right)\notag\\[.2 cm]
&= \text{cov}\left(Y_1, Y_2\right)+\text{cov}\left(Y_1, Y_3\right)+\text{cov}\left(Y_3, Y_2\right)+\text{cov}\left(Y_3, Y_3\right)\notag\\[.2 cm]
&=\text{var}\left(Y_3\right)=\theta_3\,.\notag
\end{align}
Por lo tanto, el coeficiente de correlaci\'on entre $X_1$ y
$X_2$ es
\[\rho=\text{corr}\left(X_1,X_2\right)=
\frac{\text{cov}\left(X_1,X_2\right)}{\sqrt{\text{var}(X_1)\ \text{var}(X_2)}}=\frac{\theta_3}{\sqrt{\theta_1\,
\theta_2}}\,.\]
Este coeficiente de correlaci\'on no puede exceder de
$\theta_3\left(\min\!\left\{\theta_1,\theta_2\right\}\right)^{-1}$, porque
\[\min\left\{\theta_1,\theta_2\right\}\leq \theta_k,\ k=1,2, \text{ luego } \theta_1\,\theta_2\geq \left(\min\left\{\theta_1,\theta_2\right\}\right)^2.\]
O bien, como lo se\~nal\'o Holgate (1964) \cite{Hol64},
\[0<\rho\leq \min\!\left(\sqrt{\frac{\theta_1}{\theta_2}},
\sqrt{\frac{\theta_2}{\theta_1}}\,\right),\]
pues, de la Definici\'on \ref{Dist_Pois_Biv}
\[\sqrt{\frac{\theta_1}{\theta_2}}=\sqrt{\frac{\theta_1}{\theta_2} \frac{\theta_1}{\theta_1}}=\frac{\theta_1}
{\sqrt{\theta_1 \theta_2}}=\frac{\theta'_1+\theta_3}
{\sqrt{\theta_1 \theta_2}}\geq \rho,\]
de igual forma se consigue que $\rho\leq \sqrt{\frac{\theta_2}{\theta_1}}$.
\subsection{F\'ormulas de recursi\'on}
Para calcular num\'ericamente los valores de
$P(r, s;\theta)=P_{\theta}(X_1=r, X_2=s)$, son \'utiles las
relaciones de recurrencia presentadas por Johnson, Kotz y Balakrishnan (1997, p. 125) \cite{JoKoBa97} que se enuncian como sigue
\begin{proposicion} (Johnson et al. (1997, p. 125) \cite{JoKoBa97}) Si $(X_1, X_2)\sim PB(\theta_{1}, \theta_{2}, \theta_3)$, entonces
\begin{equation}\label{form-rec-DPB}
\begin{array}{l}
rP(r, s;\theta)= (\theta_1-\theta_{3}) P(r-1, s;\theta)+\theta_{3}\, P(r-1, s-1;\theta),\\[.3 cm]
sP(r, s;\theta) = (\theta_2-\theta_{3}) \,P(r, s-1;\theta)+\theta_{3}\, P(r-1, s-1;\theta),\\[.3 cm]
\text{Si} \ r<0\ \ \ \text{o} \ \ s<0,\ \text{entonces considerar} \ P(r, s;\theta)=0.
\end{array}
\end{equation}
\end{proposicion}
Adem\'as de las relaciones dadas en la Proposici\'on anterior, tambi\'en nos ser\'an de gran utilidad las relaciones de recurrencia dadas en Johnson, Kotz y Balakrishnan (1997, p. 127) \cite{JoKoBa97}, que se enuncian como sigue.
\begin{proposicion} Sea $(X_1, X_2)\sim PB(\theta_{1}, \theta_{2}, \theta_3)$, entonces
\begin{align}
\frac{\partial}{\partial\theta_1}P(r, s;\theta) &= P(r-1, s;\theta)- P(r, s;\theta),\notag \\[.1 cm]
\frac{\partial} {\partial\theta_2}P(r, s;\theta) &=P(r, s-1;\theta)-P(r, s;\theta)\label{form-rec-derDPB} \\[.1 cm]
\frac{\partial}{\partial\theta_3}P(r, s;\theta) &= P(r-1, s-1;\theta)-P(r-1, s;\theta)-P(r, s-1;\theta)+P(r, s;\theta).\notag
\end{align}
\end{proposicion}
\subsection{Estimaci\'on puntual}\label{Estimacion-puntual}
Los estimadores m\'as usados com\'unmente son los obtenidos por los siguientes m\'etodos:
\begin{itemize}
\item M\'etodo de m\'axima verosimilitud (Kocherlakota y Kocherlakota (1992, pp. 43-45) \cite{KoKo92}).
{\bf Propiedades} Seg\'un Kocherlakota y Kocherlakota (1992, p. 45) \cite{KoKo92}, el estimador de m\'axima verosimilitud, $\hat{\theta}_{MV}$, es consistente, asint\'oticamente normal y asint\'oticamente eficiente para $\theta$.
\begin{itemize}
\item El estimador m\'aximo veros\'imil de $\theta$ debe satisfacer
\[\hat{\theta}_k=\bar{X}_{k}=\frac{1}{n}\sum_{i=1}^n X_{ki}\, ,\ \ k=1,2\quad \text{y} \quad \bar{R}=1,\]
donde
\[\bar{R}=\frac{1}{n}\mathop{\sum}\limits_{i=1}^n\frac{P(X_{1i}-1,
X_{2i}-1;\theta)}{P(X_{1i}, X_{2i};\theta)}\,.\]
\begin{observacion}
El estimador m\'aximo veros\'imil del par\'ametro $\theta_3$ no tiene una forma expl\'icita y por lo tanto debe ser calculado por m\'etodos num\'ericos utilizando la ecuaci\'on $\,\bar{R}=1$. En nuestro caso, para calcular el estimador $\hat{\theta}_3$ empleamos el m\'etodo iterativo de Newton-Raphson est\'andar.
\end{observacion}
La matriz de varianzas y covarianzas asint\'otica de los estimadores m\'aximo veros\'{\i}miles, $\Sigma_{MV}$, est\'a dada por
\[\Sigma_{MV}=\left(\begin{array}{ccc}
\theta_1 & \theta_3 & \theta_3 \\[.2 cm]
\theta_3 & \theta_{2} & \theta_3 \\[.2 cm]
\theta_3 & \theta_3 & \displaystyle\frac{\theta^2_3\left(\theta_1+\theta_2- 2\theta_3\right)\left(Q-1\right)-\theta_3^2+(\theta_1-2\theta_3) (\theta_2-2\theta_3)}{(\theta_{1}\theta_{2}-\theta_3^2)(Q-1)- \theta_{1}-\theta_2+2\theta_3}
\end{array}\right),\]
donde
\begin{equation}\label{Q}
Q=\sum_{r=1}^\infty\,\sum_{s=1}^\infty\frac{P^2(r-1,s-1;\theta)} {P(r,s;\theta)} \,.
\end{equation}
\item Por las propiedades de los estimadores m\'aximo veros\'imiles, se tiene \[\sqrt{n}\left(\hat{\theta}_{MV}-\theta\right)\ \mathop{\longrightarrow}^L \ N_3(\mathbf{0},\Sigma_{MV}).\]
\end{itemize}
\item \vskip .1 cm M\'etodo de los momentos (Kocherlakota y Kocherlakota (1992, pp. 34-35) \cite{KoKo92}).
\begin{itemize}
\item El estimador de $\theta$ por el m\'etodo de los momentos, $\hat{\theta}_{MM}$, es consistente y est\'a dado por
\[\hat{\theta}_k=\bar{X}_k, \, k=1,2 \ \ \ \text{y}\ \ \ \hat{\theta}_3=s_{11}=\frac{1}{n}\sum_{i=1}^n X_{1i}X_{2i}-\bar{X}_1\bar{X}_2.\]
Por otro lado, la matriz de varianzas y covarianzas asint\'otica de los estimadores por el m\'etodo de los momentos, $\Sigma_{MM}$, est\'a dada por
\[\Sigma_{MM} =\left(\begin{array}{ccc}
\theta_1 & \theta_3 & \theta_3 \\[.2 cm]
\theta_3 & \theta_{2} & \theta_3 \\[.2 cm]
\theta_3 & \theta_3 & \theta_{1}\theta_{2}+\theta_3+\theta_3^2
\end{array}\right).\]
\item Por ser $\hat{\theta}_{MM}$ funci\'on de momentos muestrales, aplicando el M\'etodo Delta (Teorema \ref{Metodo-Delta}) se tiene que
\[\sqrt{n}\left(\hat{\theta}_{MM}-\theta\right)\ \mathop{\longrightarrow}^L \ N_3(\mathbf{0},\Sigma_{MM}).\]
\end{itemize}
\item M\'etodo del doble cero (Kocherlakota y Kocherlakota (1992, pp. 42-43) \cite{KoKo92}).
\begin{itemize}
\item El estimador de $\theta$ por el m\'etodo del doble cero, $\hat{\theta}_{DC}$, es consistente y est\'a dado por
\[\hat{\theta}_k=\bar{X}_k, \ k=1,2 \ \ \ \text{y}\ \ \ \hat{\theta}_3=\bar{X}_1+\bar{X}_2+\log \hat{\phi},\]
donde $\phi=\exp(\theta_3-\theta_1-\theta_2)$ y $\hat{\phi}$ es la proporci\'on observada de los datos $(0,0)$ en la muestra.
La matriz de varianzas y covarianzas asint\'otica de los estimadores por el m\'etodo del doble cero, $\Sigma_{DC}$, est\'a dada por
\[\Sigma_{DC}=
\left(\begin{array}{ccc}
\theta_1 & \theta_3 & -\theta_1\,\phi \\
\theta_3 & \theta_2 & -\theta_2\,\phi \\
-\theta_1\,\phi & -\theta_2\,\phi & \phi\left(1-\phi\right) \end{array}
\right).\]
\item Por ser $\hat{\theta}_{DC}$ funci\'on de momentos muestrales, aplicando el M\'etodo Delta (Teorema \ref{Metodo-Delta}) se tiene que
\[\sqrt{n}\left(\hat{\theta}_{DC}-\theta\right)\ \mathop{\longrightarrow}^L \ N_3(\mathbf{0},\Sigma_{DC}).\]
\end{itemize}
\item M\'etodo de los puntos pares (Kocherlakota y Kocherlakota (1992, pp. 40-41) \cite{KoKo92}).
\begin{itemize}
\item El estimador de $\theta$ por el m\'etodo de los puntos pares, $\hat{\theta}_{PP}$, es consistente y est\'a dado por
\[\hat{\theta}_1=\bar{X}_1, \ \ \hat{\theta}_2=\bar{X}_2 \ \ \ \text{y} \ \ \ \hat{\theta}_3=\frac{1}{2}\,(\bar{X}_1+\bar{X}_2)+\frac{1}{4}\log \! \left(\frac{2\hat{A}}{n}-1\right), \ \text{si}\ \hat{A}>\frac{n}{2}\,,\]
donde $2A=n\left[1+\exp\{-2(\theta_1+\theta_2-2\theta_3)\}\right]$ y $\hat{A}$ es la suma de las frecuencias de ocurrencia de los pares $(X_{1i}, X_{2i})$, $i=1,2,\ldots,n$ para los cuales el valor de ambas variables tiene la misma paridad, es decir, en ambas variables el valor es par o en ambas variables el valor es impar.
La matriz de varianzas y covarianzas asint\'otica de los estimadores por el m\'etodo de los puntos pares, $\Sigma_{PP}$, est\'a dada por
\[\Sigma_{PP}=
\left(\begin{array}{ccc}
\theta_1 & \theta_3 & \theta_3 \\[.2 cm]
\theta_3 & \theta_2 & \theta_3 \\[.2 cm]
\theta_3 & \theta_3 & \frac{1}{4}(6\theta_3-\theta_1-\theta_2)+\frac{1}{16}\!\left[\exp\{4(\theta_1+ \theta_2-2\theta_3)\}-1\right] \\
\end{array}
\right).\]
\item Por ser $\hat{\theta}_{PP}$ funci\'on de momentos muestrales, aplicando el M\'etodo Delta (Teorema \ref{Metodo-Delta}) se tiene que
\[\sqrt{n}\left(\hat{\theta}_{PP}-\theta\right)\ \mathop{\longrightarrow}^L \ N_3(\mathbf{0},\Sigma_{PP}).\]
\end{itemize}
\item M\'etodo de los puntos pares condicionados (Papageorgiou y Loukas (1988) \cite{PaLo88}).
\begin{itemize}
\item El estimador de $\theta$ por el m\'etodo de los puntos pares condicionados, $\hat{\theta}_{PC}$, es consistente y est\'a dado por
\[\hat{\theta}_1=\bar{X}_1, \ \ \hat{\theta}_2=\bar{X}_2 \ \ \ \text{y} \ \ \ \hat{\theta}_3=\bar{X}_2+\frac{1}{2}\log \! \left(\frac{2\hat{A}}{\hat{A}+\hat{B}}-1\right), \ \text{si}\ \hat{A}>\hat{B}\,,\]
donde $2A=B\left[1+\exp\{2(\theta_3-\theta_2)\}\right]$, $A$ y $B$ son las sumas de las frecuencias observadas en los puntos $(0,2s)$ y $(0,2s+1)$, $s\in \mathbb{N}_0$, respectivamente.
La matriz de varianzas y covarianzas asint\'otica de los estimadores por el m\'etodo de los puntos pares condicionados, $\Sigma_{PC}$, est\'a dada por
\[\Sigma_{PC}=\frac{1}{4}
\left( \begin{array}{ccc}
C & D & E \\
D & F & F-G+H \\
E & F-G+H & F+2(H-G)+J \\
\end{array}
\right),\]
donde
\begin{equation}\label{Notacion-PPC}
\begin{array}{ll}
\alpha=\exp(-\theta_1), & \beta=\exp\{2(\theta_3-\theta_2)\},\\[.3 cm]
C=\alpha(\beta+1)\{2+\alpha(\beta+1)\}, & D=\alpha^2\left(\beta^2-1\right),\\[.3 cm] E=\displaystyle\frac{D(\theta_1\!+\!\alpha^2\beta)\!+\!2\theta_3\alpha^2 (\beta\!+\!1)\!-\!\alpha^2\beta H}{\alpha^2 \beta}, & F=\alpha(1-\beta)\{2\!+\!\alpha(\beta-1)\}, \\[.3 cm] G=\beta^{-1}\theta_1(\beta-1)^2, & H=(\beta+1)\{\theta_3(\beta^{-1}+1)-2\theta_2\}, \\[.3 cm] J=\displaystyle\frac{\theta_1(\beta\!-\!1)^2\!+\!2\theta_3(\beta^2\!-\!1) \!+\!\theta_2(\beta\!+\!1)^2}{\alpha^2 \beta^2} .
\end{array}
\end{equation}
\item Por ser $\hat{\theta}_{PC}$ funci\'on de momentos muestrales, aplicando el M\'etodo Delta (Teorema \ref{Metodo-Delta}) se tiene que
\[\sqrt{n}\left(\hat{\theta}_{PC}-\theta\right)\ \mathop{\longrightarrow}^L \ N_3(\mathbf{0},\Sigma_{PC}).\]
\end{itemize}
\end{itemize}
{\bf Eficiencia relativa asint\'otica}
La eficiencia relativa asint\'otica es la raz\'on entre la varianza generalizada (determinante de la matriz de varianzas y covarianzas) de los estimadores m\'aximo veros\'imiles y la del estimador bajo consideraci\'on. Las expresiones de las varianzas generalizadas de los estimadores discutidos anteriormente son:
\begin{align}
|\Sigma_{MV}|&=\displaystyle\frac{(\theta_1-\theta_3)^2(\theta_{2}-\theta_3)^2} {(\theta_1\theta_2-\theta_3^2)(Q-1)-\theta_1-\theta_2+2\theta_3}, \notag\\[.25 cm]
|\Sigma_{MM}|&=\theta_1^2\theta_2^2+\theta_1\theta_2\theta_3- (\theta_{1}+ \theta_2)\theta_3^2+\theta_3^3-\theta_3^4, \notag\\[.25 cm]
|\Sigma_{DC}|&=\left(\theta_1\theta_2-\theta_3^2\right) \{\exp(\theta_1+\theta_2-\theta_3)-1\}-\theta_{1}\theta_{2}(\theta_{1}+ \theta_{2}-2\,\theta_3), \notag\\[.25 cm]
|\Sigma_{PP}|&=\frac{1}{16}\Bigl[\left(\theta_1\theta_2- \!\theta_3^2\right)\left[\exp\{4(\theta_1+\theta_2-2\theta_3)\}- 1\right] \Bigr.\notag\\ &\Bigl.\hspace{15mm}+4\left\{2\theta_3 (\theta_{1}-\theta_{3})(\theta_{2}-\theta_3)- \theta_{1}(\theta_{2}-\theta_3)^2-\theta_{2}(\theta_{1}- \!\theta_3)^2\right\}\Bigr], \notag\\[.25 cm]
|\Sigma_{PC}|&=\frac{1}{64}\left[F\{CJ\!+\!E(2D\!-\!E)\}-(G\!-\!H) \{2D(E\!-\!D)\!+\!C(G\!-\!H)\}-D^2(F\!+\!J)\right],\notag
\end{align}
donde $Q$ est\'a definido en (\ref{Q}), y $C, D, E, F, G, H$ y $J$ est\'an definidas en (\ref{Notacion-PPC}).\vskip .3 cm
Un an\'alisis m\'as detallado sobre la eficiencia relativa asint\'otica del estimador de $\theta$ se puede encontrar en los trabajos desarrollados por los investigadores que han tratado el tema, ver por ejemplo, Holgate (1964) \cite{Hol64}, Loukas et al. (1986) \cite{LoKePa86}, Papageorgiou y Loukas (1988) \cite{PaLo88}, Paul y Ho (1989) \cite{PaHo89}, y Kocherlakota y Kocherlakota (1992) \cite{KoKo92}. Cabe se\~nalar que la eficiencia relativa asint\'otica para el estimador de $\theta$ por el m\'etodo de los puntos pares condicionados est\'a solamente tratado en Papageorgiou y Loukas (1988) \cite{PaLo88}.
\section{Bondad de Ajuste}
Un aspecto crucial en cualquier an\'alisis de datos es contrastar la bondad de ajuste de las observaciones con el modelo probabil\'istico supuesto, es decir, se contrasta si los datos provienen de la poblaci\'on que se supone.
Dadas las observaciones $X_1, X_2, \ldots, X_n$ iid, con distribuci\'on $F$, el objetivo es contrastar la hip\'otesis nula $H_0: ``F=F_0"$. La hip\'otesis alternativa ser\'a $H_1: ``F\neq F_0"$, $F_0$ puede ser una distribuci\'on totalmente especificada, o bien, especificada salvo por un n\'umero finito de par\'ametros.
\subsection{Tests de bondad de ajuste en dimensi\'on uno}\label{Test-bondad-ajuste-dim-1}
En este apartado nuestro inter\'es es contrastar
\begin{equation}\label{H0-DPUnivariante}
\begin{array}{l}
H_0 : X\sim P(\vartheta), \ \text{para alg\'un}\ \vartheta>0,\\[.2 cm]
H_1 : X \ \text{no se distribuye}\ P(\vartheta), \ \forall\,\vartheta>0.
\end{array}
\end{equation}
Sean $X_1, X_2,\ldots, X_n$ observaciones independientes de una v.a. $X$ que toma valores enteros no negativos. Sean
\[F(k\,;\vartheta)=\exp(-\vartheta)\sum_{j=0}^k\frac{\vartheta^j}{j!}, \ k\in \mathbb{N}_0,\]
la funci\'on de distribuci\'on de la distribuci\'on Poisson $P(\vartheta)$ y $F$ la funci\'on de distribuci\'on desconocida de $X$. Adem\'as, $\hat{\vartheta}_n=\hat{\vartheta}_n(X_1, X_2,\ldots, X_n)=\bar{X}_n$ denota el estimador m\'aximo veros\'imil de $\vartheta$.
Como es bien sabido (ver por ejemplo, G\"urtler y Henze (2000) \cite{GuHe00}), el estad\'istico $\chi^2$, que es una herramienta cl\'asica para contrastar la bondad de ajuste, presenta dos inconvenientes al tratar de comprobar la hip\'otesis que los datos se distribuyen Poisson: $(a)$ selecci\'on de celdas y $(b)$ estimaciones del par\'ametro $\vartheta$ para datos agrupados.\vskip .2 cm
G\"urtler y Henze (2000) \cite{GuHe00} presentan una variedad de procedimientos para contrastar (\ref{H0-DPUnivariante}). La distribuci\'on nula de los estad\'isticos considerados dependen del verdadero y desconocido valor del par\'ametro. Ni siquiera su distribuci\'on asint\'otica es libre, es decir, tambi\'en depende del par\'ametro. Por ello, estos autores emplean un bootstrap param\'etrico para estimar los valores cr\'iticos o los $p-$valores.
En las siguientes subsecciones estudiaremos algunos de los contrastes presentados por G\"urtler y Henze (2000) \cite{GuHe00}.
\subsubsection[\'Indice de dispersi\'on de Fisher. Estad\'istico $\hat{U}^{\,2}_{n2}$]{\'Indice de dispersi\'on de Fisher. Estad\'istico $\boldsymbol{\hat{U}^{\,2}_{n2}}$}
\nt {\bf \'Indice de dispersi\'on de Fisher}
Puesto que la media y la varianza de la distribuci\'on Poisson son iguales, entonces el cociente entre sus estimadores deber\'ia estar cercano a 1. Espec\'ificamente,
\begin{equation}\label{ecuac2.27-1}
\frac{\hat{\sigma}^2}{\hat{\vartheta}}=\frac{1}{n} \frac{\sum_{j=1}^n(X_j-\bar{X}_n)^2}{\bar{X}_n}\approx 1.
\end{equation}
Un procedimiento para contrastar el test estad\'istico dado en (\ref{H0-DPUnivariante}) es rechazar $H_0$ para valores peque\~nos o grandes del \'indice de dispersi\'on de Fisher, dado por
\[D_n=\sum_{j=1}^n\frac{(X_j-\bar{X}_n)^2}{\bar{X}_n}.\]
\nt {\bf Estad\'istico $\boldsymbol{\hat{U}^2_{n2}}$}
De (\ref{ecuac2.27-1}) y del \'indice de dispersi\'on de Fisher, resulta
$$D_n\approx n\quad \Longleftrightarrow\quad (D_n-n)^2\approx 0.$$
Por lo tanto, para contrastar las hip\'otesis en (\ref{H0-DPUnivariante}), se rechaza $H_0$ para valores ``grandes" \, del estad\'istico
\[\hat{U}^{\,2}_{n2}=\left\{\frac{1}{\sqrt{2n}}(D_n-n)\right\}^2.\]
Como
\[\hat{U}_{n2}^{\,2}=\frac{1}{2n}{(D_n-n)^2}\ \mathop{\longrightarrow}\limits_{H_0}^{L}\ \chi^2_{1},\]
(ve\'ase por ejemplo Rayner, Thas y Best (2009, p. 156) \cite{RaThBe09}).
As\'i, se rechaza $\,H_0: X\sim P(\vartheta)\ $ si $\ P_{H_0}\!\left(\hat{U}_{n2}^{\,2}>c\right)=\alpha,\ $ donde $\ c=\chi^2_{1,1-\alpha}$. Por lo tanto, se rechaza $\,H_0: X\sim P(\vartheta)\ $ si
$$\frac{1}{2n}\left(\sum_{j=1}^n\frac{(X_j- \bar{X}_n)^2}{\bar{X}_n}-n \right)^{\!\!2}\,>\,\chi^2_{1,1-\alpha}\,.$$
\begin{observacion}
El test basado en el estad\'istico $\hat{U}_{n2}^{\,2}$ no es un test consistente pues est\'a construido en base a momentos, espec\'ificamente se centra en la propiedad que la media y la varianza para la distribuci\'on Poisson son iguales.
\end{observacion}
\subsubsection{Contrastes basados en la funci\'on generatriz de probabilidad emp\'irica}\label{Contrastes-basados-en-fgpe}
\nt {\bf Estad\'istico} $\boldsymbol{R_{n,a}}$
\vskip .1 cm
Puesto que la distribuci\'on de $X$ est\'a caracterizada por su fgp, obtenemos que si $X$ se distribuye seg\'un una ley de Poisson, entonces su fgp est\'a dada por
$$g(t;\vartheta)=E_\vartheta\!\left(t^X\right)=\exp\{\vartheta(t-1)\},$$
la cual puede ser estimada por su fgpe
\[g_n(t)=\frac{1}{n}\sum_{j=1}^n t^{X_j} \ \ \text{de}\ \ X_1, X_2,\ldots, X_n.\]
\vskip .1 cm
Por tanto, parece natural basar un contraste de $H_0$ sobre
\begin{equation}\label{ecuac2.29}
G_n(t)=\sqrt{n}\left(g_n(t)-g(t\,;\hat{\vartheta}_n)\right),\ 0\leq t\leq 1.
\end{equation}
\vskip .2 cm
N\'otese que, bajo $H_0$, $G_n(t)$ es la diferencia de dos estimadores consistentes de $g(t;\vartheta)$. Por lo que un test razonable para contrastes deber\'ia rechazar $H_0$ para valores ``grandes" \,de $R_{n,a}$ definido por
\[R_{n,a}=\int_0^1 G^2_n(t) t^a dt,\]
donde $\,a\geq 0\,$ es una constante. Elegir un valor grande de $\,a\,$ significa colocar m\'as peso cerca del punto extremo $\,t=1$.
\begin{observacion}
Para $a=0$, se obtiene el estad\'istico sugerido por Rueda et al. (1991) \cite{RuPeOR91} y el caso general, cuando $a\geq 0$, fue propuesto por Baringhaus et al. (2000) \cite{BaGuHe00}.
\end{observacion}
\vskip .4 cm
\nt {\bf Estad\'istico} $\boldsymbol{T_{n,a}}$
\vskip .1 cm
Es otro estad\'istico para contrastar $H_0$, el cual est\'a basado en el hecho que $\,g(t;\vartheta)\,$ es la \'unica fgp que satisface la ecuaci\'on diferencial
\[\vartheta\, g(t)-g'(t)=0,\]
donde $g(t)$ denota la fgp de la variable aleatoria $X$.
\vskip .2 cm
Si $H_0$ es cierta, entonces
\[\bar{X}_n\, g_n(t)-g'_n(t)\approx 0,\ \forall t.\]
Por lo que un test razonable para contrastar la hip\'otesis nula deber\'ia rechazar $H_0$ para valores ``grandes" \, de $\,T_{n,a}\,$ definido por
\[T_{n,a}=n\int_0^1\Bigl\{\bar{X}_n\, g_n(t)-g'_n(t)\Bigr\}^{\!2} t^a\, dt,\]
donde $\,a\geq 0\,$ es una constante. Elegir un valor grande de $\,a\,$ significa colocar m\'as peso cerca del punto extremo $\,t=1$.
\vskip .3 cm
La distribuci\'on de los estad\'isticos $\,R_{n,a}\,$ y $\,T_{n,a}\,$ es desconocida para tama\~nos de muestra finito. Un modo cl\'asico de aproximar la distribuci\'on nula es mediante la distribuci\'on asint\'otica nula.
\vskip .3 cm
La convergencia d\'ebil de los estad\'isticos $\,R_{n,a}\,$ y $\,T_{n,a}\,$ est\'a dada en el resultado siguiente, cuya prueba sigue pasos similares a los dados en la demostraci\'on del Teorema 2.1 en Baringhaus y Henze (1992) \cite{BaHe92}.
\begin{teorema}\label{TeoConvDebilRnaTna}
(G\"urtler y Henze (2000) \cite{GuHe00})\ Sea $\{X_{n,1},X_{n,2},\ldots,X_{n,n}\},\ n\geq 1$, un arreglo triangular de v.a. iid en cada fila, tales que $X_{n,1}\sim P(\vartheta_n)$, donde $0<\vartheta=\mathop{\lim}\limits_{n\rightarrow \infty} \vartheta_n$ existe. Sean $\,G_n(t)$ definido como en (\ref{ecuac2.29}) y
$$\tilde{G}_n(t)=\sqrt{n}\left(\bar{X}_n\,g_n(t)-g'_n(t)\right),\, 0\leq t\leq 1,$$
donde
$$g_n(t)=\frac{1}{n}\sum_{j=1}^n t^{X_{n,j}},\quad g(t\,;\hat{\vartheta}_n)=\exp\!\left\{\hat{\vartheta}_n(t-1)\right\}$$
$$\hat{\vartheta}_n=\hat{\vartheta}_n(X_{n,1},X_{n,2},\ldots,X_{n,n})= \bar{X}_n=\frac{1}{n}\sum_{j=1}^n X_{n,j}.$$
Entonces se cumple lo siguiente:
\begin{itemize}
\item [$(i)$] $\displaystyle R_{n,a}\mathop{\longrightarrow}\limits^{L} \int_0^1 Z^2(t)\,t^a dt,\ a\geq0$, donde $Z(\cdot)$ es un proceso Gaussiano centrado con n\'ucleo de covarianza
$$K(u,v)=\exp\{\vartheta(uv-1)\}-\{1+\vartheta(u-1)(v-1)\}\exp\{\vartheta(u+v-2)\}, \ 0\leq u,v\leq 1.$$
\item [$(ii)$] $\displaystyle T_{n,a}\mathop{\longrightarrow}\limits^{L} \int_0^1 \tilde{Z}^2(t)\,t^a dt,\ a\geq0$, donde $\tilde{Z}(\cdot)$ es un proceso Gaussiano centrado con n\'ucleo de covarianza
$$\tilde{K}(u,v)=\vartheta\left\{1+\vartheta(1-u)(1-v)\right\} \,\exp\{\vartheta(uv-1)\} -\vartheta \exp\{\vartheta(u+v-2)\},$$
donde $\,0\leq u,v\leq 1$.
\end{itemize}
\end{teorema}
Adem\'as, G\"urtler y Henze (2000) \cite{GuHe00} demuestran que los test basados en los estad\'isticos $R_{n,a}\,$ y $\, T_{n,a}$ son consistentes frente a alternativas fijas.\\
Un modo alternativo de aproximar la distribuci\'on nula de los estad\'isticos antes citados es usar un estimador bootstrap param\'etrico, que se describe como sigue:\vskip .25 cm
Denotemos por $W_n$ a uno cualquiera de los estad\'isticos anteriores ($R_{n,a}\,$ o $\, T_{n,a}$). Sea $\,H_{n,\vartheta}(t)=P_{\vartheta}(W_n\leq t)\,$ la funci\'on de distribuci\'on de la distribuci\'on nula de $W_n$ cuando $\vartheta$ es el verdadero valor del par\'ametro. El m\'etodo bootstrap estima el cuantil de $\,H_{n,\vartheta}(t)$ mediante el $(1-\alpha)$-cuantil de $H_{n,\,\hat{\vartheta}_n}$, que se aproxima mediante simulaci\'on, empleando los siguientes pasos:
\begin{itemize}
\item [$\bullet$] Generar $B$ seudo-muestras aleatorias de tama\~no $n$ con distribuci\'on $P(\hat{\vartheta}_n)$, es decir, generar $X^*_{j1},X^*_{j2}, \ldots, X^*_{jn}, j\!=\!1,2,\ldots, B$ iid de acuerdo a $P(\hat{\vartheta}_n)$.
\item [$\bullet$] \vskip .2 cm Calcular $\,W^*_{j,n}=W_n(X^*_{j1}, X^*_{j2},\ldots, X^*_{jn})$, para $j=1,2 ,\ldots, B$.
\item [$\bullet$] \vskip .2 cm Sea $\,H^*_{n,B}(t)=\displaystyle\frac{1}{B}\sum_{j=1}^B I{\{W^*_{j,n}\leq t\}}\,$ para la funci\'on de distribuci\'on emp\'irica de $W^*_{1,n}$, $W^*_{2,n},\ldots, W^*_{B,n}\,$ y $\,W^*_{1:B}\leq W^*_{2:B}\leq\ldots\leq W^*_{B:B}\,$ para sus estad\'isticos de orden.
\vskip .2 cm
Finalmente, el valor cr\'itico, $c^*_{n,B}$, est\'a dado por
\[c^*_{n,B}=
\left\{\begin{array}{ll}
W^*_{B(1-\alpha):B}\,, & \text{si}\ B(1-\alpha)\ \text{es un entero,} \\ [.3 cm]
W^*_{[B(1-\alpha)]+1:B}\,, & \text{en otro caso,}
\end{array}
\right.\]
donde $[y]$ es la parte entera de $y$.
\end{itemize}
\nt {\bf Estad\'istico} $\boldsymbol{V_n}$
Motivados por el hecho que $\,\frac{\partial^2}{\partial t^2}\log g(t;\vartheta)\equiv 0$, Nakamura y P\'erez-Abreu (1993) \cite{NaPe93} propusieron la suma de los cuadrados de los coeficientes del polinomio $\,g^2_n(t)\,\frac{\partial^2}{\partial t^2}\log g_n(t)$ como un estad\'istico para contrastar (\ref{H0-DPUnivariante}).
Como
\begin{equation}\label{Der2-log(gn)}
\frac{\partial^2}{\partial t^2}\log g_n(t)=\frac{g_n(t) \displaystyle\frac{\partial^2}{\partial t^2} g_n(t)-\left\{\frac{\partial}{\partial t} g_n(t)\right\}^2}{g^2_n(t)},
\end{equation}
llamando $N_n(t)$ al numerador y puesto que $g_n(t)=\displaystyle\sum_{i=1}^n t^{X_i}$, obtenemos
\begin{align}
N_n(t)&=g_n(t) \frac{\partial^2}{\partial t^2} g_n(t)-\left\{\frac{\partial}{\partial t} g_n(t)\right\}^2\notag\\
&=\frac{1}{n^2}\sum_{i,j=1}^n X_i(X_i-X_j-1)t^{X_i+X_j-2}.\notag
\end{align}
Debido a que $g^2_n(t)>0, \forall t$, entonces de (\ref{Der2-log(gn)})
\[\frac{\partial^2}{\partial t^2}\log g_n(t)\equiv 0\ \Longleftrightarrow \ N_n(t)=0, \ \forall t.\]
Haciendo $X_{(n)}=\max\{X_1, X_2,\ldots, X_n\}$ y notando que $N_n(t)$ es un polinomio aleatorio en $t$ de grado $d_n=2X_{(n)}-2$, obtenemos \[N_n(t)=\sum_{k=0}^{d_n} a_k t^k,\] donde
\[a_k=\frac{1}{n^2}\sum_{i,j=1}^n X_i(X_i-X_j-1)I{\{X_i+X_j-2=k\}},\ 0\leq k \leq d_n.\]
Por lo tanto
\[N_n(t)=0, \ \forall t\ \Longleftrightarrow \ a_k=0,\ 0\leq k \leq d_n.\]
As\'i, un estad\'istico para contrastar $H_0$ es considerar la suma de los cuadrados de los coeficientes polinomiales, esto es,
\begin{align}
V_n&=\sum_{k=0}^{d_n} a_k^2\notag\\
&=\frac{1}{n^4}\sum_{i,j,k,l=1}^n X_i(X_i-X_j-1)X_k(X_k-X_l-1)I{\{X_i+X_j=X_k+X_l\}},\notag
\end{align}
que expresa a $V_n$ como un $V$-estad\'istico con un n\'ucleo de orden 4.
La distribuci\'on nula y la distribuci\'on asint\'otica nula de $V_n$ son ambas desconocidas. No obstante, Nakamura y P\'erez-Abreu (1993) \cite{NaPe93} observaron num\'ericamente que la distribuci\'on nula de
\[V_n^*=\frac{nV_n}{(\bar{X}_n)^{1\text{.}45}}\]
es aproximadamente independiente de $\vartheta$.
\subsection{Tests de bondad de ajuste en dimensi\'on dos}\label{Test-bondad-ajuste-dim-2}
El objetivo de este apartado es contrastar
\begin{equation}\label{H0-DPBivariante}
\begin{array}{l}
H_0 : (X_1,X_2)\sim PB(\theta_1, \theta_2,\theta_3),\ \text{para alg\'un}\ \theta_1, \theta_2,\theta_3>0,\\[.2 cm]
H_1 : (X_1,X_2)\ \,\text{no se distribuye}\ PB(\theta_1, \theta_2,\theta_3), \ \forall\, \theta_1, \theta_2,\theta_3>0.
\end{array}
\end{equation}
Hasta donde conocemos, existen tres tests estad\'isticos para contrastar (\ref{H0-DPBivariante}). A continuaci\'on los describiremos brevemente y expondremos lo esencial de cada uno de ellos.
\subsubsection[Test $T$ de Crockett]{Test $\boldsymbol{T}$ de Crockett}
El estad\'istico propuesto por Crockett (1979) \cite{Cro79}, $T$, est\'a basado en una forma cuadr\'atica en $Z_{X_1}=S^2_{X_1}-\bar{X}_1$ y $Z_{X_2}=S^2_{X_2}-\bar{X}_2$, donde $\bar{X}_1, \bar{X}_2, S^2_{X_1}$ y $S^2_{X_2}$ son las medias y varianzas muestrales, respectivamente. El objetivo es encontrar la matriz de varianzas y covarianzas de $Z_{X_1}$ y $Z_{X_2}$.
Se tiene que var$(Z_{X_1})=$ var$(S^2_{X_1})+$ var$(\bar{{X_1}})-2$ cov$(S^2_{X_1},\bar{X}_1)$. De Stuart y Ord (1987, Volumen I, p. 338) \cite{StOr87}, obtenemos que
\[\text{var}(S_{X_1}^2)=\frac{\mu_4-\mu_2^2}{n},\]
donde $\mu_k$ es el $k$-\'esimo momento central definido en la Subsecci\'on \ref{Momentos-DPB}.
Como $\mu_1=0$, de Stuart y Ord (1987, Volumen I, ejemplo 10.2, p. 323) \cite{StOr87}
\[\text{cov}(S_{X_1}^2,\bar{X}_1)=\frac{\mu_3}{n}.\]
As\'i,
\[\text{var}(Z_{X_1})=\frac{\theta_1+3\theta_1^2-\theta_1^2}{n}+\frac{\theta_1}{n}- 2\frac{\theta_1}{n}=2\frac{\theta_1^2}{n}\]
y similarmente,
\[\text{var}(Z_{X_2})=2\frac{\theta_2^2}{n}.\]
Puesto que,
\[\text{cov}(Z_{X_1}, Z_{X_2}) = \text{cov}(S^2_{X_1},S^2_{X_2})+\text{cov}(\bar{X}_1,\bar{X}_2)- \text{cov}(S^2_{X_1},\bar{X}_2)-\text{cov}(\bar{X}_1,S^2_{X_2}),\]
de la primera ecuaci\'on de (2.1.19) en Kocherlakota y Kocherlakota (1992, p. 40) \cite{KoKo92}
\[\text{cov}(S^2_{X_1},S^2_{X_2})=\text{cov}(m_{2,0},m_{0,2})=\frac{\mu_{2,2}- \mu_{2,0}\mu_{0,2}}{n},\]
donde $\mu_{r,s}$ es el $(r,s)$-\'esimo momento central, descrito en la Subsecci\'on \ref{Momentos-DPB}, adem\'as
\[m_{r,s}=\frac{1}{n}\sum_{i=1}^n (X_{1i}-\bar{X}_1)^r (X_{2i}-\bar{X}_2)^s.\]
De Kocherlakota y Kocherlakota (1992, p. 43) \cite{KoKo92} \[\text{cov}(\bar{X}_1,\bar{X}_2)=\frac{1}{n}\,\text{cov}(X_1,X_2)= \frac{\theta_3}{n}\] y de (2.1.17) en Kocherlakota y Kocherlakota (1992, p. 40) \cite{KoKo92}
\[\text{cov}(S^2_{X_1},\bar{X}_2)=\text{cov}(m_{2,0},m_{0,1})=\frac{\mu_{2,1}-\mu_{2,0}\mu_{0,1}+2\mu_{1,1} \mu_{1,0}\mu_{0,0}-\mu_{2,1}\mu_{0,0}-2\mu_{1,0}\mu_{1,1}}{n}.\]
Haciendo los c\'alculos respectivos obtenemos
\[\text{cov}(Z_{X_1}, Z_{X_2})=2\frac{\theta_3^2}{n}.\]
As\'i, si $V$ denota la matriz de varianzas y covarianzas de $Z=(Z_{X_1}, Z_{X_2})$, entonces los resultados de esta secci\'on muestran que
\[V=\frac{2}{n}\left(
\begin{array}{cc}
\theta_1^2 & \theta_3^2 \\[.1 cm]
\theta_3^2 & \theta_2^2
\end{array}
\right).
\]
Usando los estimadores $\bar{X}_1, \bar{X}_2$ y $m_{1,1}=S_{X_1X_2}$ (covarianza muestral) para $\theta_1, \theta_2$ y $\theta_3$, respectivamente, Crockett (1979) \cite{Cro79} demuestra que, bajo $H_0$,
\[T=Z\hat{V}^{-1}Z^{\top}\ \mathop{\longrightarrow}\limits^{L}\ Y\sim \chi^2_{2},\]
con lo cual se consigue el estad\'istico propuesto por dicho autor, dado por
\[ T= \frac{n}{2}\frac{\bar{X}_2^{2} \left(S^2_{X_1}-\bar{X}_1\right)^2- 2S^2_{X_1X_2}\left(S^2_{X_1}-\bar{X}_1\right) \left(S^2_{X_2}-\bar{X}_2\right)+ \bar{X}_1^{2} \left(S^2_{X_2}-\bar{X}_2\right)^2} {\bar{X}_1^{2} \bar{X}_2^{2}-S^4_{X_1X_2}}.\]
As\'i, se rechaza $H_0$ si
\[T \geq \chi^2_{2,\alpha},\]
donde $\chi^2_{2, \alpha}$, para $0<\alpha<1$, denota el percentil $\alpha$ superior de la distribuci\'on $\chi^2$ con 2 grados de libertad.
\subsubsection[Test $I_B$ de Loukas y Kemp]{Test $\boldsymbol{I_B}$ de Loukas y Kemp}
Loukas y Kemp (1986) \cite{LoKe86} desarrollaron un test para la DPB basado en lo que llamaron \'indice de dispersi\'on bivariante y denotaron por $I_B$, pues es una extensi\'on del \'indice de dispersi\'on univariante.
Con el objetivo de obtener $I_B$, consideremos ahora una muestra aleatoria simple $(X_{11}, X_{21})$, $(X_{12}, X_{22}), \ldots, (X_{1n}, X_{2n})$, tal que $(X_{1k}, X_{2k})\sim PB(\theta_1,\theta_2,\theta_3)$, para $\, k=1, 2, \ldots, n$, donde $\theta_1>\theta_3,\,\theta_2>\theta_3\,$ y $\,\theta_3>0$.
Si $\theta_1,\theta_2$ y $\theta_3$ son conocidos, entonces el \'indice de dispersi\'on bivariante toma la forma
$$I_B=\frac{1}{1-\rho^2}\sum_{i=1}^n\left(W_{1i}^2-2\rho W_{1i}W_{2i}+W_{2i}^2\right),$$
donde $\,W_{ji}=\displaystyle\frac{X_{ji}-\theta_j}{\sqrt{\theta_j}}, j=1,2, \ i=1,2,\ldots,n\ $ y $\ \rho=\displaystyle\frac{\theta_3}{\sqrt{\theta_1\theta_2}}$.\vskip .3 cm
Bajo las condiciones del Teorema 1 en Loukas y Kemp (1986) \cite{LoKe86}, dichos autores muestran que $I_B$ se distribuye aproximadamente como una variable $\chi^2$ con $2n$ grados de libertad.\vskip .25 cm
En la situaci\'on pr\'actica usual, cuando $\, \theta_1,\, \theta_2\, $ y $\, \theta_3\,$ son desconocidos, entonces Loukas y Kemp (1986) \cite{LoKe86} definen el \'indice de dispersi\'on bivariante como
\begin{align}
I_B&=\frac{\mathop{\sum}\limits_{i=1}^n\left\{\displaystyle\frac{(X_{1i}- \bar{X}_1)^2}{\bar{X}_1}-\frac{2S_{12}(X_{1i}- \bar{X}_1)(X_{2i}-\bar{X}_2)}{\bar{X}_1 \bar{X}_2}+\frac{(X_{2i}-\bar{X}_2)^2}{\bar{X}_2} \right\}}{1-\displaystyle\frac{S_{12}^2}{\bar{X}_1 \bar{X}_2}}\notag\\[.2 cm]
&=\frac{n(\bar{X}_2S_1^2-2S_{12}^2+\bar{X}_1S_2^2)} {\bar{X}_1\bar{X}_2-S_{12}^2},\notag
\end{align}
donde $\bar{X}_1$ y $\bar{X}_2$ son las medias muestrales, $S^2_1$ y $S^2_2$ son la varianzas muestrales y $S_{12}$ es la covariaza muestral.
Como en este proceso se debe estimar el par\'ametro $\theta$, entonces se pierden tres grados de libertad y por tanto, este nuevo $I_B$ se distribuye aproximadamente como una $\chi^2$ con $2n-3$ grados de libertad, como lo mencionan Loukas y Kemp (1986) \cite{LoKe86}.\vskip .2 cm
As\'i, se rechaza $H_0$ si
\[I_B \geq \chi^2_{2n-3,\alpha},\]
donde $\chi^2_{2n-3, \alpha}$, para $0<\alpha<1$, denota el percentil $\alpha$ superior de la distribuci\'on $\chi^2$ con $2n-3$ grados de libertad.
\begin{observacion} Si $\theta_1,\theta_2$ y $\theta_3$ son desconocidos, entonces Loukas y Kemp (1986) \cite{LoKe86} aproximan $\theta_1$ y $\theta_2$ por sus estimadores m\'aximo veros\'imiles, esto es, $\hat{\theta}_1=\bar{X}_1$ y $\,\hat{\theta}_2=\bar{X}_2$, y el estimador de $\theta_3$ que, como debe calcularse por m\'etodos num\'ericos, lo aproximan por la covarianza muestral, es decir, $\hat{\theta}_3=S_{12}$.
\end{observacion}
\subsubsection[Test $NI_B$ de Rayner y Best]{Test $\boldsymbol{NI_B}$ de Rayner y Best}
Rayner y Best (1995) \cite{RaBe95} expresaron el estad\'istico de Loukas y Kemp (1986) \cite{LoKe86}, $I_B$, como sigue
\begin{equation}\label{ecuac3.63}
I_B=\frac{n}{1-\hat{\rho}^{\,2}}\!\left(\frac{S_{X_1}^2}{\bar{X}_1}-2\, \frac{S_{X_1X_2}^2}{\bar{X}_1\bar{X}_2}+\frac{S_{X_2}^2} {\bar{X}_2}\right),
\end{equation}
donde $\displaystyle\hat{\rho}=\frac{S_{X_1X_2}}{\sqrt{\bar{X}_1\bar{X}_2}}$ es un estimador de $\rho$, $\bar{X}_1$ y $\bar{X}_2$ son las medias muestrales, $S^2_{X_1}$ y $S^2_{X_2}$ son la varianzas muestrales, y $S_{X_1X_2}$ es la covariaza muestral.\vskip .2 cm
De (\ref{ecuac3.63}) se observa que
\[\text{si}\ \ \hat{\rho}^2=\frac{S_{X_1X_2}^2}{\bar{X}_1\bar{X}_2} >\frac{1}{2} \left(\frac{S_{X_1}^2}{\bar{X}_1}+\frac{S_{X_2}^2} {\bar{X}_2}\right),\ \text{entonces} \ \, I_B<0.\]
Por lo tanto, cuando $\rho$ crece hay una probabilidad creciente que $I_B$ sea negativo y su distribuci\'on no sea bien aproximada por una $\chi^2$. Para remediar esta situaci\'on, proponen utilizar un nuevo estad\'istico, lo llaman $NI_B$ y lo definen mediante la expresi\'on
$$NI_B=\frac{n}{1-r^2}\left(\frac{S_{X_1}^2}{\bar{X}_1}-2\,r^2 \sqrt{\frac{S_{X_1}^2S_{X_2}^2}{\bar{X}_1\bar{X}_2}}+\frac{S_{X_2}^2} {\bar{X}_2}\right)\!,$$
donde, el estimador de $\rho$ es el coeficiente de correlaci\'on muestral, dado por
\[r=\frac{\mathop{\sum}\limits_{i=1}^n(X_{1i}-\bar{X}_{1})(X_{2i}-\bar{X}_{2})} {\sqrt{\mathop{\sum}\limits_{i=1}^n(X_{1i}-\bar{X}_{1})^2\ \mathop{\sum}\limits_{i=1}^n(X_{2i}-\bar{X}_{2})^2}}.\]
Al igual que $I_B$, bajo $H_0$, $NI_B\cong \chi_{2n-3}^2$. Por lo tanto, se rechaza $H_0$ si
\[NI_B \geq \chi^2_{2n-3,\alpha},\]
donde $\chi^2_{2n-3, \alpha}$, para $0<\alpha<1$, denota el percentil $\alpha$ superior de la distribuci\'on $\chi^2$ con $2n-3$ grados de libertad.
\begin{observacion} Los tests estad\'isticos $T, I_B$ y $NI_B$ no son consistentes, pues est\'an construidos en base a los momentos, espec\'ificamente se basan en que los dos primeros momentos poblacionales son iguales.
\end{observacion}
\begin{observacion} Para efectos de programar los estad\'isticos $T, I_B$ y $NI_B$, Rayner y Best (1995) \cite{RaBe95} hacen notar que Crockett (1979) \cite{Cro79} usa $(n-1)$ como divisor para calcular $S^2_{X_1}$ y $S^2_{X_2}$, y divisor $n$ en $S_{X_1X_2}$, en cambio Loukas y Kemp (1986) \cite{LoKe86} usan el divisor $n$ para calcular $S^2_{X_1}, S^2_{X_2}$ y $S_{X_1X_2}$.
\end{observacion}
\chapter{Estad\'isticos tipo Cram\'er-von Mises}\label{Estadisticos-tipoCramer-von-Mises}
\section{Tests estad\'isticos} \label{Tests-estadisticos}
Sean $\boldsymbol{X}_1=(X_ {11}, X_ {21}), \boldsymbol{X}_2=(X_ {12}, X_ {22}), \ldots, \boldsymbol{X}_n=(X_ {1n}, X_ {2n})$ vectores aleatorios iid de $\boldsymbol{X}=(X_1,X_2)\in \mathbb{N}_0^2$. Bas\'andonos en la muestra aleatoria $\boldsymbol{X}_1,\boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$, nuestro objetivo es contrastar la hip\'otesis
\[H_0 : (X_1,X_2)\sim PB(\theta_1,\theta_2,\theta_3), \ \text{para alg\'un}\ (\theta_1,\theta_2,\theta_3)\in\Theta,\]
contra la alternativa
\[H_1 : (X_1,X_2)\nsim PB(\theta_1,\theta_2,\theta_3), \ \forall(\theta_1,\theta_2,\theta_3)\in\Theta.\]
Con este prop\'osito, aprovecharemos algunas de las propiedades de la fgp que nos permitir\'an proponer los siguientes dos tests estad\'isticos.
\begin{enumerate}
\item De acuerdo a la Proposici\'on \ref{Conv-FuncGenProbBiv}, un estimador consistente de la fgp es la fgpe. Si $H_0$ es verdadera y $\hat{{\theta}}_n$ es un estimador consistente de ${\theta}$, entonces $g(u;\hat{{\theta}}_n)$ estima consistentemente la fgp de los datos.
Como la distribuci\'on de $\boldsymbol{X}$ es determinada de forma \'unica por su fgp, $g(u)$, $u\in [0,1]^2$, un test razonable para contrastar $H_0$ deber\'ia rechazar la hip\'otesis nula para valores ``grandes'' de $R_{n,w}(\hat{{\theta}}_n)$ definido por
\begin{equation}\label{Estad-Rnw} R_{n,w}(\hat{{\theta}}_n)=\int_0^1\int_0^1 G^2_n(u;\hat{{\theta}}_n)w(u)du,
\end{equation}
donde \[G_n(u;\hat{\theta}_n)= \sqrt{n}\left\{g_n(u)-g(u; \hat{{\theta}}_n)\right\},\] $\hat{\theta}_n=\hat{{\theta}}_n(\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n)=(\hat{\theta}_{1n}, \hat{\theta}_{2n},\hat{\theta}_{3n})$ es un estimador consistente de $\theta$ y $w(u)$ es una funci\'on medible de peso tal que $w(u)\geq 0, \ \forall u\in [0,1]^2$, y
\begin{equation} \label{int-funcion-peso}
\int_0^1\int_0^1 w(u)du < \infty.
\end{equation}
Este \'ultimo supuesto sobre $w$ asegura que la integral doble en (\ref{Estad-Rnw}) es finita para cada $n$ fijo.
$R_{n,w}(\hat{{\theta}}_n)$ es una extensi\'on bivariante de los estad\'isticos propuestos por Rueda et al. (1991) \cite{RuPeOR91} y Baringhaus et al. (2000) \cite{BaGuHe00} para contrastar la bondad de ajuste a la distribuci\'on Poisson univariante, tal como se vi\'o en la Subsecci\'on \ref{Contrastes-basados-en-fgpe}.
\item Puesto que la fgp de la distribuci\'on Poisson univariante con par\'ametro $\lambda$ es la \'unica fgp que satisface $g'(t)=\lambda g(t)$, Baringhaus y Henze (1992) \cite{BaHe92} propusieron un test estad\'istico que se basa en el an\'alogo emp\'irico de esta ecuaci\'on, presentado en la Subsecci\'on \ref{Contrastes-basados-en-fgpe}.
Con el fin de extender este test al caso bivariante, primero tenemos que encontrar una ecuaci\'on o un sistema de ecuaciones cuya \'unica soluci\'on sea la fgp de la DPB. La siguiente proposici\'on establece dicho sistema. \begin{proposicion}\label{Soluc-sistema-de-dos-EDP} Sea $G_2=\{g:[0,1]^2\to \mathbb{R}$, tal que $g$ es una fgp y $\frac{\partial}{\partial u_i}g (u_1,u_2)$, $i=1,2$, existen $\forall (u_1,u_2) \in [0,1]^2\}$. Sea $g(u_1,u_2;\theta)$ como en (\ref{fgp3-DPB}). Entonces $g(u_1,u_2;\theta)$ es la \'unica fgp en $G_2$ que satisface el siguiente sistema \begin{equation}\label{EDPs-fgp} D_i(u;\theta)=0, \quad i=1,2, \quad \forall u \in [0,1]^2,\end{equation} donde
\[\begin{array}{ll} D_1(u;\theta)&=\displaystyle\frac{\partial}{\partial u_1}g(u_1,u_2)-\{\theta_1+\theta_3(u_2-1)\}g(u_1,u_2),\\[.35 cm] D_2(u;\theta) &=\displaystyle\frac{\partial }{\partial u_2}g(u_1,u_2)-\{\theta_2+\theta_3(u_1-1)\}g(u_1,u_2).\end{array}\] \end{proposicion}
{\bf Demostraci\'on} \hspace{2pt} Sea $(X_1,X_2)$ un vector aleatorio y sea $g(u_1,u_2)=E\left(u_1^{X_1} u_2^{X_2}\right)$ su fgp. Entonces, de la primera expresi\'on en (\ref{EDPs-fgp}) \begin{equation}\label{EDP1-Modif-fgp} \frac{\partial}{\partial u_1}\log g(u_1,u_2)=\theta_1+\theta_3(u_2-1).\end{equation}
Integrando (\ref{EDP1-Modif-fgp}) sobre $u_1$, obtenemos \[\begin{array}{rcl}
g(u_1,u_2) & = & \exp\{\phi_1(u_2)+\theta_1 u_1+\theta_3(u_2-1) u_1\},\\[.25 cm]
& = & \exp\{\phi(u_2)+\theta_1(u_1-1)+\theta_3(u_1-1)(u_2-1)\}, \end{array}\]
donde $\phi(u_2)=\phi_1(u_2)+\theta_1+\theta_3(u_2-1)$.\vskip .2 cm
Procediendo similarmente, de la segunda ecuaci\'on en (\ref{EDPs-fgp}) obtenemos
\[g(u_1,u_2)=\exp\{\varphi(u_1)+\theta_2(u_2-1)+\theta_3(u_1-1)(u_2-1)\}.\] As\'i, necesariamente $\varphi(u_1)=\theta_1(u_1-1)$ y $\phi(u_2)=\theta_2(u_2-1)$, en otras palabras, la fgp de la DPB es la \'unica soluci\'on de (\ref{EDPs-fgp}). $\square$\vskip .3 cm
Por la Proposici\'on \ref{Conv-FuncGenProbBiv}, $g(u)$ y sus derivadas pueden ser estimadas consistentemente por la fgpe y las derivadas de la fgpe, respectivamente. As\'i, si $H_0$ fuera cierta, entonces las funciones
\begin{equation}\label{EDPs-empiricas}
\begin{array}{l}
D_{1n}(u;\hat{\theta}_n )=\displaystyle\frac{\partial }{\partial u_1}g_n(u_1,u_2)-\left\{\hat{\theta}_{1n}+\hat{\theta}_{3n}(u_2-1)\right\} g_n(u_1,u_2), \\[.35 cm]
D_{2n}(u;\hat{\theta}_{n})=\displaystyle\frac{\partial }{\partial u_2}g_n(u_1,u_2)-\left\{\hat{\theta}_{2n}+\hat{\theta}_{3n}(u_1-1)\right\} g_n(u_1,u_2),
\end{array}\end{equation}
deber\'ian estar cerca de 0, donde $\hat{\theta}_n$ es un estimador consistente de $\theta$. As\'i, para contrastar $H_0$ consideramos el siguiente test estad\'istico
\begin{equation}\label{Estad-Snw}
S_{n,w}(\hat{\theta}_n)=n\int_0^1\int_0^1 \left\{D^2_{1n}(u;\hat{\theta}_n )+D^2_{2n}(u;\hat{\theta}_{n} )\right\}w(u)du,
\end{equation}
donde la funci\'on de peso $w(u)$ es como la definida en el \'item anterior.
Notar que la fgp de $\boldsymbol{X}$ est\'a en $G_2$ s\'i y s\'olo si $E(X_i)$ existe, $i=1,2$.
\end{enumerate}
En los dos casos anteriores, un test razonable para contrastar $H_0$ deber\'ia rechazar la hip\'otesis nula para valores grandes de cada test estad\'istico. Ahora, para determinar qu\'e son los valores grandes en cada caso, debemos calcular la distribuci\'on nula de cada test estad\'istico o al menos una aproximaci\'on de cada una de ellas.
Puesto que las distribuciones nulas son desconocidas, trataremos de aproximarlas. Un modo cl\'asico de estimar la distribuci\'on nula es mediante la distribuci\'on asint\'otica nula. En la siguiente secci\'on estudiaremos esta situaci\'on.
\section{Aproximaci\'on de la distribuci\'on nula} \label{Aproximacion-distribucion-nula}
\subsection{Distribuci\'on asint\'otica nula}\label{Distribucion-asintotica-nula}
Como un primer intento de aproximar la distribuci\'on nula de $R_{n,w}(\hat{\theta}_n)$ y de $S_{n,w}(\hat{\theta}_n)$ obtendremos sus distribuciones asint\'oticas nulas. Con este prop\'osito, supondremos que el estimador $\hat{\theta}_n$ satisface la siguiente condici\'on de regularidad.
\begin{supuesto}\label{hat(theta)-theta}
Bajo $H_0$, si $\theta= (\theta_{1},\theta_{2},\theta_{3})\!\in\!\Theta$ denota el verdadero valor del par\'ametro, entonces
\[\sqrt{n}\left(\hat{\theta}_n-\theta\right) =\frac{1}{\sqrt{n}}\sum_{i=1}^n \boldsymbol{\ell}\left(\boldsymbol{X}_{i}; \theta\right)+\boldsymbol{o}_{_P}(1),\]
donde $\boldsymbol{\ell}:\mathbb{N}^2_0\times \Theta \longrightarrow \mathbb{R}^3$ es tal que \[E_{\theta}\bigl\{\boldsymbol{\ell} (\!\boldsymbol{X}_{\!{1}}; \theta)\bigr\}=\mathbf{0}\in \mathbb{R}^3\] y \[J(\theta)=E_{\theta}\!\left\{ \boldsymbol{\ell}\!\left(\boldsymbol{X}_{\!{1}}; \theta\right)^\top\boldsymbol{\ell} \!\left(\boldsymbol{X}_{\!{1}}; \theta\right)\right\}< \infty.\]
Aqu\'i, $\boldsymbol{o}_{_P}(1)$ es un vector que consta de 3 elementos $o_{_P}(1)$.
\end{supuesto}
\begin{observacion}\label{metodos-aprox}
El Supuesto \ref{hat(theta)-theta} no es restrictivo pues lo cumplen los estimadores m\'as usados com\'unmente, como son los citados en la Subsecci\'on \ref{Estimacion-puntual}.
\end{observacion}
Para obtener la distribuci\'on asint\'otica nula de $R_{n,w}(\hat{\theta}_n)$ y de $S_{n,w}(\hat{\theta}_n)$, la plataforma de trabajo que usaremos ser\'a el espacio de Hilbert separable $\mathcal{H}=L^2\left([0,1]^2,w\right)$ definido por
\[\mathcal{H}=\left\{\varphi:[0,1]^2\rightarrow \mathbb{R}, \text{ medible, tal que } \| \varphi\|_{_\mathcal{H}}^2=\int_0^1\int_0^1 \varphi(u)^2 w(u)du<\infty\right\},\]
con producto escalar
$ \langle \phi, \psi\rangle_{_\mathcal{H}}=\int_0^1\int_0^1 \phi(u)\psi(u) w(u)du$.
En dicho espacio de Hilbert $\mathcal{H}$, tenemos que
\begin{itemize}
\item
$R_{n,w}(\hat{\theta}_n)= \|G_n(\hat{{\theta}}_n)\|_{_\mathcal{H}}^{2},\ \, \text{con}\ \, G_n(u;\hat{{\theta}}_n)=\displaystyle\frac{1}{\sqrt{n}}\sum_{i=1}^n G(\boldsymbol{X}_i;\hat{\theta}_n;u),\ \, u\in [0,1]^2$,
donde
\[G(\boldsymbol{X}_i;\hat{\theta}_n;u)= u_1^{X_{1i}}\,u_2^{X_{2i}}-g(u;\hat{\theta}_n),\, i=1,2,\ldots,n.\]
\item
$S_{n,w}(\hat{\theta}_n)= \|Z_{1n}\|_{_\mathcal{H}}^{2}+ \|Z_{2n}\|_{_\mathcal{H}}^{2},\ \ \text{con}\ \ Z_{kn}(u)=\displaystyle\frac{1}{\sqrt{n}}\sum_{i=1}^n V_k(\boldsymbol{X}_i;\hat{\theta}_n;u),\ \,k=1,2$,$\ u\in [0,1]^2$, donde
\[V_1(\boldsymbol{X}_i;\hat{\theta}_n;u)= X_{1i}\,I{\{X_{1i}\geq 1\}}\,u_1^{X_{1i}-1}\,u_2^{X_{2i}}-\left\{\hat{\theta}_{1n}+ \hat{\theta}_{3n}(u_2-1)\right\}u_1^{X_{1i}}\,u_2^{X_{2i}},\]
\[V_2(\boldsymbol{X}_i;\hat{\theta}_n;u)= X_{2i}\,I{\{X_{2i}\geq 1\}}\,u_1^{X_{1i}}\,u_2^{X_{2i}-1}-\left\{\hat{\theta}_{2n}+ \hat{\theta}_{3n}(u_1-1)\right\}u_1^{X_{1i}}\,u_2^{X_{2i}},\]
$i=1,2,\ldots,n$.
\end{itemize}
El siguiente resultado proporciona la distribuci\'on asint\'otica nula de $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$.
\begin{teorema}\label{ConvDebil-Rnw-Snw}
Sean $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid de $\boldsymbol{X}=(X_{1},X_{2})\sim PB(\theta)$. Supongamos que se cumple el Supuesto \ref{hat(theta)-theta}. Entonces
\begin{itemize}
\item [(a)] $R_{n,w}(\hat{\theta}_n)=\|R_n\|_{_\mathcal{H}}^2+r_n,$\\ donde $P_{{\theta}}(|r_n|>\varepsilon)\to 0$, $\forall \varepsilon>0$, $R_n(u)=\displaystyle\frac{1}{\sqrt{n}}\sum_{i=1}^n R^0(\boldsymbol{X}_i;{\theta};{u})$, con
$$R^0(\boldsymbol{X}_i;{\theta}; {u})=u_1^{X_{1i}} u_2^{X_{2i}}-g({u};{\theta}) \left\{1+\boldsymbol{\ell}\left(\boldsymbol{X}_{i}; {\theta}\right)(u_1\!-1,u_2\!-1,(u_1\!-1)(u_2\!-1))^\top\right\},$$
\item [(b)] \begin{equation} \label{Snw-Sn} S_{n,w}(\hat{\theta}_n)= \|S_{1n}\|_{_\mathcal{H}}^2+\|S_{2n}\|_{_\mathcal{H}}^2+s_n,\end{equation}
donde $P_{{\theta}}(|s_n|>\varepsilon)\to 0$, $\forall \varepsilon>0$, $S_{kn}(u)=\displaystyle\frac{1}{\sqrt{n}}\sum_{i=1}^n S^0_k(\boldsymbol{X}_i;\theta; u),\ k=1,2,\ $ con
\begin{align}
S^0_1(\boldsymbol{X}_i;\theta; u)& = X_{1i}\,I{\{X_{1i}\geq 1\}}\,u_1^{X_{1i}-1}\,u_2^{X_{2i}}-\left\{\theta_{1}+\theta_{3}(u_2-1) \right\} u_1^{X_{1i}}\,u_2^{X_{2i}}\notag\\[.1 cm] &\hspace{8mm}-g(u;\theta) \boldsymbol{\ell}\left(\boldsymbol{X}_i; \theta\right)(1,0,u_2-1)^{\top},\notag\\[.25 cm]
S^0_2(\boldsymbol{X}_i;\theta; u)& = X_{2i}\,I{\{X_{2i}\geq 1\}}\,u_1^{X_{1i}}\,u_2^{X_{2i}-1}-\left\{\theta_{2}+\theta_{3}(u_1-1) \right\} u_1^{X_{1i}}\,u_2^{X_{2i}}\notag\\[.1 cm] &\hspace{8mm}-g(u;\theta) \boldsymbol{\ell}\left(\boldsymbol{X}_i; \theta\right)(0,1,u_1-1)^{\top}.\notag
\end{align}
\end{itemize}
Adem\'as,
\[R_{n,w}(\hat{\theta}_n)\mathop{\longrightarrow}\limits^{L} \sum_{j\geq 1}\lambda_j^R\chi^2_{1j}.\]
\[S_{n,w}(\hat{\theta}_n)\mathop{\longrightarrow}\limits^{L} \sum_{j\geq 1}\lambda_j^S\chi^2_{1j}.\]
donde $\chi^2_{11},\chi^2_{12},\ldots$ son v.a. independientes $\chi^2$ con 1 grado de libertad y los conjuntos $\{\lambda_j^R\}_{j\geq 1}$ y $\{\lambda_j^S\}_{j\geq 1}$ son los autovalores no nulos de los operadores $C_R(\theta)$ y $C_S(\theta)$, respectivamente, definidos sobre el espacio de funciones
$\left\{\tau:\mathbb{N}_0^2\to \mathbb{R}, \text{tal que} \ E_{\theta}\!\left\{\tau^2(\boldsymbol{X})\right\}<\infty,\forall \theta\in\Theta\right\}$, como sigue
\begin{equation} \label{Operador-C}
C_R(\theta) \tau({x})= E_{\theta}\{h_R({x},\boldsymbol{Y};\theta) \tau(\boldsymbol{Y})\} \quad \text{y}\quad C_S(\theta) \tau({x})= E_{\theta}\{h_S({x},\boldsymbol{Y};\theta) \tau(\boldsymbol{Y})\}
\end{equation}
donde
\[h_R({x},{y};\theta)=\int_0^1\int_0^1 R^0({x}; \theta; u) R^0({y}; \theta; u)w(u)du,\]
\begin{equation} \label{h-S} h_S({x},{y};\theta)=\int_0^1\int_0^1 \sum_{k=1}^2 S_k^0({x}; \theta; u) S_k^0({y}; \theta; u)w(u)du.
\end{equation}
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt} \hspace{2pt} Solamente presentaremos la demostraci\'on de la parte (b), pues la demostraci\'on de la parte (a) sigue pasos similares.
Por definici\'on,
$S_{n,w}(\hat{\theta}_n)= \|Z_{1n}\|_{_\mathcal{H}}^{2} +\|Z_{2n}\|_{_\mathcal{H}}^{2}$, con
\[Z_{kn}(u)=\frac{1}{\sqrt{n}}\sum_{i=1}^n V_k(\boldsymbol{X}_i;\hat{\theta}_n;u), \ k=1,2.\]
Por desarrollo en serie de Taylor,
\begin{equation}\label{Z-kn-aprox}
Z_{kn}(u)=\frac{1}{\sqrt{n}}\sum_{i=1}^n V_k(\boldsymbol{X}_i;\theta; u)+\frac{1}{n}\sum_{i=1}^nQ^{(1)}_k(\boldsymbol{X}_i;\theta; u)\,\sqrt{n}(\hat{\theta}_n-\theta)^{\top}+q_{kn},
\end{equation}
$k=1,2$, donde
\[q_{kn}=\frac{1}{2}\sqrt{n}\bigl(\hat{\theta}_n-\theta\bigr) \frac{1}{\sqrt{n}}\frac{1}{n}\sum_{i=1}^n Q^{(2)}_k\bigl(\boldsymbol{X}_i;\widetilde{\theta};u\bigr) \,\sqrt{n}\bigl(\hat{\theta}_n-\theta\bigr)^{\top},\]
$\widetilde{\theta}=\alpha \hat{\theta}_n+(1-\alpha)\theta$, para alg\'un $0<\alpha<1\,$ y $\,Q^{(2)}_k\bigl(x;\vartheta; u\bigr)$ es la matriz de orden $3\times 3$ que contiene las derivadas de segundo orden de $V_k\bigl(x;\vartheta; u\bigr)$ con respecto a $\vartheta$, para $k=1,2$, \[Q^{(1)}_k(x;\vartheta; u)=
\left( Q^{(1)}_{k1}(x;\vartheta; u), Q^{(1)}_{k2}(x;\vartheta; u), Q^{(1)}_{k3}(x;\vartheta; u)\right),\]
donde $\ Q^{(1)}_{kj}(x;\vartheta;u)=\frac{\partial}{\partial \vartheta_j} V_k(x;\vartheta; u)$, para $k=1,2$, $j=1,2,3$.\vskip .15 cm
Como las primeras derivadas son
\begin{equation}\label{Deriv-Vk}
Q^{(1)}_1(x;\vartheta;u)=-u_1^{x_1} u_2^{x_2}(1,0,u_2-1),\ \ \ Q^{(1)}_2(x;\vartheta;u)=-u_1^{x_1} u_2^{x_2}(0,1,u_1-1),
\end{equation}
entonces, $\left|Q^{(1)}_{kj}\!\left(\boldsymbol{X}_{\!1};\theta; u\right)\right|\!\leq\! 1$, para $k=1,2$, $j=1,2,3$, $\forall u \!\in\! [0,1]^2$.
As\'i, considerando (\ref{int-funcion-peso}), resulta
\begin{equation}\label{E-Qkj-acotada}
E_{\theta}\left\{\left\|Q^{(1)}_{kj}\left(\boldsymbol{X}_{1};\theta; u\right)\right\|_{_\mathcal{H}}^2\right\}<\infty,\ k=1,2,\ j=1,2,3.
\end{equation}
Como, para $k=1,2$, $j=1,2,3$, tenemos que
\[E_{\theta}\!\left[\left\{\frac{1}{n}\sum_{i=1}^n Q^{(1)}_{kj}(\boldsymbol{X}_i;\theta; u) \right\}^2\right]= \frac{1}{n} E_{\theta}\!\left[\left\{Q^{(1)}_{kj}(\boldsymbol{X}_1;\theta; u) \right\}^2\right]+\frac{n-1}{n}\!\left[E_{\theta}\! \left\{Q^{(1)}_{kj} (\boldsymbol{X}_1;\theta; u) \right\}\right]^2,\]
entonces
\begin{align}
E_{\theta}\!\left[\left\{\frac{1}{n}\sum_{i=1}^n Q^{(1)}_{kj} (\boldsymbol{X}_i;\theta; u)-E_{\theta}\! \left\{Q^{(1)}_{kj} (\boldsymbol{X}_1;\theta; u)\right\} \right\}^2\right]&\notag\\[.2 cm] &\hspace{-38mm}=\frac{1}{n}E_{\theta}\!\left[\left\{Q^{(1)}_{kj} (\boldsymbol{X}_1;\theta; u)\right\}^2\right]-\frac{1}{n}\!\left[E_{\theta}\! \left\{Q^{(1)}_{kj} (\boldsymbol{X}_1;\theta; u) \right\}\right]^2\notag\\[.2 cm]
&\hspace{-38mm}\leq\frac{1}{n}E_{\theta}\!\left[\left\{Q^{(1)}_{kj} (\boldsymbol{X}_1;\theta; u)\right\}^2\right].\notag
\end{align}
Usando la desigualdad de Markov y (\ref{E-Qkj-acotada}), se logra
\begin{align}
&P_{\theta}\!\left[\left\| \frac{1}{n}\sum_{i=1}^n Q^{(1)}_{kj}(\boldsymbol{X}_i;\theta; u)-E_{\theta}\!\left\{Q^{(1)}_{kj}(\boldsymbol{X}_1;\theta; u)\right\} \right\|_{_\mathcal{H}}>\varepsilon\right]\notag\\[.2 cm]
&\hspace{85mm}\leq \frac{1}{n\, \varepsilon^2}\,E_{\theta}\!\left[\left\|Q^{(1)}_{kj} (\boldsymbol{X}_1;\theta; u)\right\|_{_\mathcal{H}}^2\right]\to 0,\notag
\end{align}
para $k=1,2$, $j=1,2,3$.
As\'i, obtenemos que
\begin{equation}\label{E-Qk}
\frac{1}{n}\sum_{i=1}^n Q^{(1)}_{k}(\boldsymbol{X}_i;\theta; u)\mathop{\longrightarrow} \limits^{P}E_{\theta}\left\{Q^{(1)}_{k}(\boldsymbol{X}_1;\theta; u)\right\}=
\left\{
\begin{array}{ll}
-g(u;\theta)(1,0,u_2-1), & \hbox{si } k=1,\\[.2 cm]
-g(u;\theta)(0,1,u_1-1), & \hbox{si } k=2,
\end{array}
\right.
\end{equation}
en $\mathcal{H}$.
De (\ref{Deriv-Vk}) se sigue que $Q^{(2)}_k\!\bigl(x;\vartheta; u\bigr)= \boldsymbol{0}\,$ (la matriz nula de orden $3\times 3$), $k=1,2$ y as\'i $q_{kn}=0$.
Ahora, por el Supuesto \ref{hat(theta)-theta} y como $q_{kn}=0$, entonces, (\ref{Z-kn-aprox}) se puede escribir como
\begin{align}
Z_{kn}(u)&=\frac{1}{\sqrt{n}}\sum_{i=1}^n \left[V_k(\boldsymbol{X}_i;\theta; u)+E_{\theta}\left\{Q^{(1)}_{k}(\boldsymbol{X}_1;\theta; u)\right\} \boldsymbol{\ell}\left(\boldsymbol{X}_i; \theta\right)^{\top}\right]\notag\\[.2 cm]
&\hspace{22mm}+ E_{\theta}\left\{Q^{(1)}_{k}(\boldsymbol{X}_1;\theta; u)\right\}\, \boldsymbol{o}_{_P}(1)^{\top}.\notag
\end{align}
As\'i, de esta \'ultima ecuaci\'on, y de (\ref{E-Qkj-acotada}) y (\ref{E-Qk})
\[
Z_{kn}(u)=S_{kn}(u)+s_{kn},
\]
donde $\|s_{kn}\|_{\mathcal{H}}=o_{_P}(1)$, $k=1,2$.
Por otra parte, observar que
\[
\|S_{1n}\|_{_\mathcal{H}}^{2}+\|S_{2n}\|_{_\mathcal{H}}^{2}=\frac{1}{n} \sum_{i=1}^n\sum_{j=1}^n h_S(\boldsymbol{X}_i,\boldsymbol{X}_j;\theta),\]
donde, $h_S(x,y;\theta)$ es como el definido en (\ref{h-S}) y satisface $\,h_S(x,y;\theta)= h_S(y,x;\theta)$,
$E_{\theta}\left\{h_S^2(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)\right\}< \infty\,$ y $\,E_{\theta} \left\{|h_S(\boldsymbol{X}_1,\boldsymbol{X}_1;\theta)|\right\} < \infty$.
Adem\'as, por la Proposici\'on \ref{Soluc-sistema-de-dos-EDP} y el Supuesto \ref{hat(theta)-theta}
\[E_{\theta}\!\left\{h_S(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)\right\}= \int_0^1\int_0^1 \sum_{k=1}^2 E_{\theta}\!\left\{S_k^0(\boldsymbol{X}_1; \theta; u)\right\} E_{\theta}\!\left\{S_k^0(\boldsymbol{X}_2; \theta; u)\right\}w(u)\,du=0.\]
Por \'ultimo, como $h_S$ es degenerado, $E_{\theta}\!\left\{h_S(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)/ \boldsymbol{X}_1\right\}=0$, entonces, por el Teorema 6.4.1.B en Serfling (1980) \cite{Ser80}
\begin{equation}\label{norma-Skn-converge}
\|S_{1n}\|_{_\mathcal{H}}^{2}+\|S_{2n}\|_{_\mathcal{H}}^{2}
\mathop{\longrightarrow} \limits^{\!L}\ \sum_{j\geq 1}\lambda_j^S\,\chi^2_{1j}\,,
\end{equation}
donde $\chi^2_{11},\chi^2_{12},\ldots$ y el conjunto $\{\lambda_j^S\}_{j\geq 1}$ son como los definidos en el Teorema.\vskip .15 cm
De (\ref{norma-Skn-converge}) se desprende que $\|S_{1n}\|_{_\mathcal{H}}^{2}+\|S_{2n}\|_{_\mathcal{H}}^{2}=O_P(1)$.
Ahora, como \[\|Z_{kn}\|_{_\mathcal{H}}^{\,2}=\|S_{kn}\|_{_\mathcal{H}}^{\,2}+ \|s_{kn}\|_{_\mathcal{H}}^{\,2}+2\langle S_{kn},s_{kn}\rangle_{_\mathcal{H}}, \ k=1,2,\]
\nt y puesto que $\ \langle S_{kn},s_{kn}\rangle_{_\mathcal{H}}\leq\|S_{kn}\|_{_\mathcal{H}}\, \|s_{kn}\|_{_\mathcal{H}}=o_{_P}(1)$,
entonces
\[S_{n,w}(\hat{\theta}_n)\mathop{\longrightarrow} \limits^{\!L}\ \sum_{j\geq 1}\lambda_j^S\,\chi^2_{1j}\,.\]
Lo cual concluye la demostraci\'on del resultado. $\square$\\
La distribuci\'on asint\'otica nula de $R_{n,w}(\hat{\theta}_n)$ y de $S_{n,w}(\hat{\theta}_n)$ depende del desconocido verdadero valor del par\'ametro $\theta$, por tanto, en la pr\'actica, no proporcionan una soluci\'on \'util al problema de estimar la distribuci\'on nula de los respectivos tests estad\'isticos. Esto podr\'ia solucionarse al reemplazar $\theta$ por $\hat{\theta}_n$.
Pero una dificultad mayor es determinar los conjuntos $\{\lambda_j^R\}_{j\geq 1}$ y $\{\lambda_j^S\}_{j\geq 1}$, puesto que, en la mayor\'ia de los casos, calcular los autovalores de un operador no es una tarea simple y en nuestro caso, debemos obtener tambi\'en las expresiones de $h_R(x,y;\theta)$ y $h_S(x,y;\theta)$, que no son f\'aciles de encontrar, pues dependen de la funci\'on $\boldsymbol{\ell}$, que por lo general no tiene una expresi\'on sencilla.
As\'i, en la siguiente subsecci\'on consideramos otra forma de aproximar la distribuci\'on nula de los tests estad\'isticos, el m\'etodo bootstrap param\'etrico.
\subsection{Aproximaci\'on bootstrap de la distribuci\'on nula}\label{Aproximacion-bootstrap}
Un modo alternativo de estimar la distribuci\'on nula es mediante el m\'etodo bootstrap param\'etrico.
Sean $\boldsymbol{X}_{1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid que toman valores en $\mathbb{N}_0^2$. Supongamos que $\hat{\theta}_n= \hat{\theta}_n(\boldsymbol{X}_{\!1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}) \in \Theta$.
Sean $\boldsymbol{X}^*_{1}, \boldsymbol{X}^*_{2},\ldots, \boldsymbol{X}^*_{n}$ vectores aleatorios iid de una poblaci\'on que se distribuye seg\'un la ley $PB(\hat{\theta}_{1n},\hat{\theta}_{2n}, \hat{\theta}_{3n})$, dado $\boldsymbol{X}_{1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ y sea $R^*_{n,w}(\hat{\theta}_n^*)$ la versi\'on bootstrap de $R_{n,w}(\hat{\theta}_n)$ obtenida al reemplazar $\boldsymbol{X}_{\!1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ y
$\hat{\theta}_n= \hat{\theta}_n(\boldsymbol{X}_{\!1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n})$ por $\boldsymbol{X}^*_{\!1}, \boldsymbol{X}^*_{2},\ldots, \boldsymbol{X}^*_{n}$ y $\hat{\theta}_n^*= \hat{\theta}_n(\boldsymbol{X}^*_{1}, \boldsymbol{X}^*_{2},\ldots, \boldsymbol{X}^*_{n})$, respectivamente, en la expresi\'on de $R_{n,w}(\hat{\theta}_n)$.
Adem\'as, sea $P_*$ la ley de probabilidad condicional bootstrap, dado $\boldsymbol{X}_{1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$.
En forma an\'aloga se describe $S^*_{n,w}(\hat{\theta}_n^*)$, la versi\'on bootstrap de $S_{n,w}(\hat{\theta}_n)$.
Para probar que el m\'etodo bootstrap se puede usar para aproximar consistentemente la distribuci\'on nula de $R_{n,w}(\hat{\theta}_n)$, o de $S_{n,w}(\hat{\theta}_n)$, supondremos las siguientes hip\'otesis que son un poco m\'as fuertes que el Supuesto \ref{hat(theta)-theta}.
\begin{supuesto}\label{E(l*l)-l-cont} Se cumple el Supuesto \ref{hat(theta)-theta} y las funciones $\boldsymbol{\ell}$ y $J$ satisfacen
\begin{enumerate}
\item $\sup_{\vartheta\in\,\Theta_0} E_{\vartheta}\Bigl[\|\boldsymbol{\ell} (\boldsymbol{X}; \vartheta)\|^2 I{\left\{\|\boldsymbol{\ell}(\boldsymbol{X}; \vartheta)\|>\gamma \right\}}\Bigr]\rightarrow 0$, cuando $\gamma \to \infty$, donde $\Theta_0 \subseteq \Theta$ es una vecindad abierta de $\theta$.
\item $\boldsymbol{\ell}(\boldsymbol{X};\vartheta)$ y $J(\vartheta)$ son continuas como funciones de $\vartheta$ en $\vartheta=\theta$ y $J(\vartheta)$ es finita $\forall \vartheta \in \Theta_0$.
\end{enumerate}
\end{supuesto}
Tal como se estableci\'o en la Observaci\'on \ref{metodos-aprox}, el Supuesto \ref{E(l*l)-l-cont} no es restrictivo pues lo cumplen estimadores com\'unmente usados.
Para mostrar que el m\'etodo bootstrap aproxima consistentemente a la distribuci\'on nula de $R_{n,w}(\hat{\theta}_n)$, o de $S_{n,w}(\hat{\theta}_n)$, nos ser\'a \'util el siguiente resultado, que lo volveremos a usar m\'as adelante.
\begin{lema}\label{P(hat(t))*l(hat(t))-P(t)*l(t)}
Supongamos que se verifica el Supuesto \ref{E(l*l)-l-cont} y que $\hat{\theta}_n
\mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un $\theta \in \Theta$. Entonces
\begin{itemize}
\item [(a)] $\displaystyle \sum_{r,s \geq 0} r\left|\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|=o(1),\ i=1,2,3,$
\item [(b)] $\displaystyle \sum_{r,s \geq 0} s\left|\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|=o(1),\ i=1,2,3,$
\end{itemize}
donde $P(r,s;\vartheta)=P_{\vartheta}(X_1=r, X_2=s),\, \forall \vartheta\in \Theta_0$.
\end{lema}
\nt {\bf Demostraci\'on} \hspace{2pt} Solamente presentaremos la demostraci\'on de la parte (b), pues la demostraci\'on de la parte (a) sigue pasos similares.
Como
\[\sum_{r,s \geq 0}=\sum_{r,s \geq 0;\, \|\boldsymbol{\ell}(r,s; \vartheta)\|>\gamma}+\sum_{r,s \geq 0;\, \|\boldsymbol{\ell}(r,s; \vartheta)\|\leq \gamma},\]
consideraremos por separado cada uno de estos sumandos.
\nt \underline{Caso $\|\boldsymbol{\ell}(r,s; \vartheta)\|>\gamma$} \hspace{2pt} Tengamos presente que $E_{\theta}\left(X^2_2\right)=\theta_2+\theta_2^2$.
Sea $\varepsilon>0$ arbitrario pero fijo. Por el Supuesto \ref{E(l*l)-l-cont} (1), $\exists \gamma=\gamma(\varepsilon)>0$ tal que
\begin{equation}\label{Sup-acotado}
\sup_{\|\boldsymbol{\ell}(r,s;\vartheta)\|>\gamma} \sum_{r,s} \|\boldsymbol{\ell}(r,s;\vartheta)\|^2 P(r,s;\vartheta)<\frac{\varepsilon^2}{4\left(\theta_2+\theta_2^2\right)}, \ \ \forall \vartheta \in \Theta_0
\end{equation}
y por tanto
\begin{equation}\label{Sup-suma-acotado}
\sup_{\|\boldsymbol{\ell}(r,s;\vartheta)\|>\gamma} \sum_{r,s} s\left|\boldsymbol{\ell}_i(r,s;\vartheta) P(r,s;\vartheta)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|<\varepsilon,
\end{equation}
$i=1,2,3$, pues, por la desigualdad de Cauchy-Schwarz
\begin{align}
&\sum_{r,s} s\left|\boldsymbol{\ell}_i(r,s;\vartheta) P(r,s;\vartheta)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|\notag\\
&\ \ \ \ \leq \sum_{r,s} s \left|\boldsymbol{\ell}_i(r,s;\vartheta)\right| P(r,s;\vartheta) + \sum_{r,s} s \left|\boldsymbol{\ell}_i(r,s;\theta)\right| P(r,s;\theta)\notag\\
&\ \ \ \ \leq \!\left(\theta_2+\theta_2^2\right)^{\!\frac{1}{2}}\! \left\{\sum_{r,s} \|\boldsymbol{\ell}(r,s;\vartheta)\|^2 P(r,s;\vartheta)\right\}^{\!\!\frac{1}{2}}\! + \! \left(\theta_2+\theta_2^2\right)^{\!\frac{1}{2}} \!\left\{\sum_{r,s} \|\boldsymbol{\ell}(r,s;\theta)\|^2 P(r,s;\theta)\right\}^{\!\!\frac{1}{2}}\!.\notag
\end{align}
De lo anterior y considerando (\ref{Sup-acotado}), obtenemos (\ref{Sup-suma-acotado}).\vskip .1 cm
\nt \underline{Caso $\|\boldsymbol{\ell}(r,s; \vartheta)\|\leq\gamma$} \hspace{2pt} Para esta situaci\'on, primero mostraremos la validez de (b) cuando $r,s\geq M$ y luego para $r,s< M$, donde $M$ es una constante positiva que determinaremos m\'as adelante.
Como $\sum_{r,s \geq 0} s P(r,s;\vartheta)=E_{\vartheta}(X_2)=\vartheta_2$ y las $P(r,s;\vartheta)$ son continuas como funciones de $\vartheta$, entonces dado $\varepsilon_1> 0, \,\exists M=M(\varepsilon_1)>0$, de modo que
\[\sum_{r,s < M} s P(r,s;\vartheta)> \vartheta_2-\frac{\varepsilon_1}{2\gamma}, \ \forall \vartheta \in \Theta_1\subseteq\Theta_0,\]
siempre que $\|\vartheta-\theta\|<\delta_1$, para cierto $\delta_1 >0$.
Por tanto, si $\|\boldsymbol{\ell}(r,s;\vartheta)\| \leq \gamma$, entonces
\begin{equation}\label{Suma-acotada}
\sum_{r,s\geq M} s\left|\boldsymbol{\ell}_i(r,s;\vartheta) P(r,s;\vartheta)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|<\varepsilon_1,\ \forall \vartheta\in \Theta_0 \cap\left\{\|\vartheta-\theta\|<\delta_1\right\},\ \,
\end{equation}
$ i=1,2,3$, pues
\[\vartheta_2=\sum_{r,s \geq 0} s P(r,s;\vartheta)=\sum_{r,s \geq M} s P(r,s;\vartheta)+\sum_{r,s<M} s P(r,s;\vartheta)>\sum_{r,s \geq M} s P(r,s;\vartheta)+\vartheta_2-\frac{\varepsilon_1}{2\gamma},\]
de donde
\[\sum_{r,s \geq M} s P(r,s;\vartheta)<\frac{\varepsilon_1}{2\gamma}.\]
Considerando esta \'ultima desigualdad y el hecho que $|\boldsymbol{\ell}_i(r,s;\vartheta)|\leq\|\boldsymbol{\ell}(r,s; \vartheta)\|\leq\gamma$, para $i=1,2,3$, entonces
\begin{align}
\sum_{r,s\geq M} s &\left|\boldsymbol{\ell}_i(r,s;\vartheta) P(r,s;\vartheta) - \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|\notag \\
&\hspace{43mm}\leq\gamma\left\{ \sum_{r,s\geq M} s P(r,s;\vartheta) + \sum_{r,s\geq M} s P(r,s;\theta) \right\}< \varepsilon_1,\notag
\end{align}
lo cual prueba (\ref{Suma-acotada}).
Para $r,s< M$. Como $\boldsymbol{\ell}_i(r,s;\vartheta)P(r,s;\vartheta)$ son funciones continuas en $\vartheta$, para $ i=1,2,3$, se tendr\'a que $\,\forall \varepsilon_2> 0, \,\exists \delta_2=\delta_2(\varepsilon_2)>0$, tal que $\|\vartheta-\theta\|<\delta_2$ implica que
\[\sup_{r,s < M}\left| \boldsymbol{\ell}_i(r,s;\vartheta)P(r,s;\vartheta) -\boldsymbol{\ell}_i(r,s;\theta)P(r,s;\theta)\right|<\frac{\varepsilon_2}{M^3}\]
y por tanto,
\begin{align}
\sum_{r,s< M} s &\left|\boldsymbol{\ell}_i(r,s;\vartheta) P(r,s;\vartheta)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|\notag\\
&\hspace{15mm}<M\sum_{r,s< M} \left|\boldsymbol{\ell}_i(r,s;\vartheta) P(r,s;\vartheta)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|< \varepsilon_2,\notag
\end{align}
siempre que $\,\vartheta \in \left\{\|\vartheta-\theta\|<\delta_2\right\}\cap \Theta_0$, para cierto $\delta_2>0$. Lo cual concluye la demostraci\'on, pues $\varepsilon_1$ y $\varepsilon_2$ son arbitrarios. $\square$\vskip .3 cm
Tambi\'en nos ser\'a de utilidad el siguiente resultado.
\begin{proposicion}\label{Nucleos-Rnw-Snw}
Sean $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid de $\boldsymbol{X}=(X_{1},X_{2}) \in \mathbb{N}_0^2$. Supongamos que se verifica el Supuesto \ref{E(l*l)-l-cont} y que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un $\theta \in \Theta$. Entonces
\begin{itemize}
\item [(a)] $\displaystyle\sup_{u,v \in [0,1]^2} \left| K^R_n(u,v)-K^R(u,v)\right|\stackrel{c.s.}{\longrightarrow}0$, donde
\begin{align}
K^R_n(u,v)&=E_*\left\{R^0(\boldsymbol{X}_1^*;\hat{\theta}_n; u) R^0(\boldsymbol{X}_1^*;\hat{\theta}_n; v)\right\}, \notag\\[.2 cm]
K^R(u,v)&=E_{\theta}\left\{R^0(\boldsymbol{X}_1;\theta; u) R^0(\boldsymbol{X}_1;\theta; v)\right\},\notag
\end{align} con $R^0(\boldsymbol{X}_1;\theta; u)$ el definido en el Teorema \ref{ConvDebil-Rnw-Snw}(a).
\item [(b)] \vskip .1 cm $\displaystyle\sup_{u,v \in [0,1]^2} \left| K^S_n(u,v)-K^S(u,v)\right|\stackrel{c.s.}{\longrightarrow}0$, donde
\begin{align}
K^S_n(u,v)&=\left(E_*\left\{S_i^0(\boldsymbol{X}_1^*;\hat{\theta}_n; u) S_j^0(\boldsymbol{X}_1^*;\hat{\theta}_n; v)\right\}\right), 1\leq i,j \leq 2. \notag\\[.2 cm]
K^S(u,v)&=\Bigl(E_{\theta}\left\{S_i^0(\boldsymbol{X}_1;\theta; u) S_j^0(\boldsymbol{X}_1;\theta; v)\right\}\Bigr), 1\leq i,j \leq 2.\notag
\end{align} con $S_i^0(\boldsymbol{X}_1;\theta; u)$ el definido en el Teorema \ref{ConvDebil-Rnw-Snw}(b), $\,i=1,2$.
\end{itemize}
\end{proposicion}
\nt {\bf Demostraci\'on} \hspace{2pt}
Solamente presentaremos la demostraci\'on de la parte (b), pues la demostraci\'on de la parte (a) sigue pasos similares.
El Supuesto \ref{E(l*l)-l-cont} y $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$ implican que
\begin{equation}\label{E*S-EtS}
\sup_{u,v \in [0,1]^2} \left| E_*\left\{S_i^0(\boldsymbol{X}_1^*;\hat{\theta}_n; u) S_j^0(\boldsymbol{X}_1^*;\hat{\theta}_n; v)\right\}- E_{\theta}\left\{S_i^0(\boldsymbol{X}_1;\theta; u) S_j^0(\boldsymbol{X}_1;\theta; v)\right\}\right| \stackrel{c.s.}{\longrightarrow}0,
\end{equation}
$1\leq i,j \leq 2$.
Para probar este resultado, primero hacemos la resta de las esperanzas en (\ref{E*S-EtS}) y encontramos, por ejemplo, que
\begin{align}
&E_*\!\left\{S_1^0(\boldsymbol{X}_1^*;\hat{\theta}_n; u) S_2^0(\boldsymbol{X}_1^*;\hat{\theta}_n; v)\right\}- E_{\theta}\!\left\{S_1^0(\boldsymbol{X}_1;\theta; u) S_2^0(\boldsymbol{X}_1;\theta; v)\right\}\notag\\[.21 cm]
&=v_1u_2\left[\left\{\hat{\theta}_{3n}+\phi_1(r;\hat{\theta}_n) \phi_2(r;\hat{\theta}_n)\right\} g(r;\hat{\theta}_n)-\left\{\theta_3+ \phi_1(r;\theta) \phi_2(r;\theta)\right\}g(r;\theta)\right]\notag\\[.21 cm]
&\quad -v_1\left\{\phi_1(r;\hat{\theta}_n) \phi_2(v;\hat{\theta}_n) g(r;\hat{\theta}_n)- \phi_1(r;\theta)\phi_2(v;\theta) g(r;\theta)\right\}\notag\\[.21 cm]
&\quad -u_2\left\{\phi_1(u;\hat{\theta}_n) \phi_2(r;\hat{\theta}_n) g(r;\hat{\theta}_n)-\phi_1(u;\theta)\phi_2(r;\theta) g(r;\theta)\right\}\notag\\[.21 cm]
&\quad +\phi_1(u;\hat{\theta}_n) \phi_2(v;\hat{\theta}_n) g(r;\hat{\theta}_n)- \phi_1(u;\theta)\phi_2(v;\theta) g(r;\theta)\notag\\[.21 cm]
&\quad+A(u)^{\top}\!\left\{g(u;\hat{\theta}_n)g(v;\hat{\theta}_n)J(\hat{\theta}_n) -g(u;\theta)g(v;\theta)J(\theta)\right\}B(v)\notag\\[.21 cm]
&\quad+\!\left[g(v;\hat{\theta}_n)\phi_1(u;\hat{\theta}_n) E_*\!\!\left\{u_1^{X^*_{1}}u_2^{X^*_{2}} \! \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\!\right\}\!-\!g(v;\theta) \phi_1(u;\theta)E_{\theta}\!\left\{u_1^{X_{1}}u_2^{X_{2}} \boldsymbol{\ell}(\boldsymbol{X}_{1};\theta)\right\}\right]B(v)\notag\\[.21 cm]
&\quad+\!\left[g(u;\hat{\theta}_n)\phi_2(v;\hat{\theta}_n) E_*\!\!\left\{v_1^{X^*_{1}}v_2^{X^*_{2}}\! \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\!\right\}\!-\!g(u;\theta) \phi_2(v;\theta)E_{\theta}\{v_1^{X_{1}}v_2^{X_{2}} \boldsymbol{\ell}(\boldsymbol{X}_{1};\theta)\}\right]A(u),\notag\\[.21 cm]
&\quad-\left[g(v;\hat{\theta}_n)E_*\!\!\left\{\psi_1(u;\boldsymbol{X}^*_1; \hat{\theta}_n)\right\}-g(v;\theta)E_{\theta}\!\left\{\psi_1(u;\boldsymbol{X}_1; \theta)\right\}\right]B(v)\notag\\[.21 cm]
&\quad-\left[g(u;\hat{\theta}_n)E_*\!\!\left\{\psi_2(v;\boldsymbol{X}^*_1; \hat{\theta}_n)\!\right\}-g(u;\theta)E_{\theta}\! \left\{\psi_2(v;\boldsymbol{X}_{\!1}; \theta)\right\}\right]A(u)\notag
\end{align}
donde
\begin{align}
&r_1=u_1v_1,\ r_2=u_2v_2,\ r=(r_1,r_2),\ \ A(u)=(1,0,u_2-1)^{\top},\ B(v)=(0,1,v_1-1)^{\top},\notag\\[.22 cm]
&\phi_1(w;\vartheta)=\vartheta_1+\vartheta_3(w_2-1), \quad \phi_2(w;\vartheta)=\vartheta_2+\vartheta_3(w_1-1),\quad w=(w_1,w_2),\notag\\[.22 cm]
&\psi_1(u;\boldsymbol{X}_1;\vartheta)=X_1\,I{\{X_1\geq 1\}}\,u_1^{X_1-1}\,u_2^{X_2}\,\boldsymbol{\ell}(\boldsymbol{X}_1;\vartheta), \notag\\[.22 cm]
&\psi_2(u;\boldsymbol{X}_1;\vartheta)=X_2\,I{\{X_2\geq 1\}}\,u_1^{X_1}\,u_2^{X_2-1}\,\boldsymbol{\ell}(\boldsymbol{X}_1;\vartheta), \notag
\end{align}
Las otras diferencias de (\ref{E*S-EtS}) tienen expresiones an\'alogas a la diferencia de esperanzas anterior. Las situaciones que aparecen en tales esperanzas son de alguno de los siguientes casos:
\begin{itemize}
\item Caso 1: sea $f(u,v;\vartheta)$ una funci\'on polinomial en las variables $u, v\in [0,1]^2$ y $\vartheta\in\Theta_1\subseteq\Theta_0$, donde $\Theta_1$ es un conjunto compacto que contiene a $\theta$. Luego, $f(u,v;\theta)g(u;\theta)$ es continua como funci\'on de $u, v$ y $\theta$, por lo tanto, es uniformemente continua en $[0,1]^2\times\Theta_1$ y como $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, entonces \[\sup_{u,v\in [0,1]^2} \left|f(u,v;\hat{\theta}_n)g(u;\hat{\theta}_n)- f(u,v;\theta)g(u;\theta)\right| =o(1).\]
\item \vskip .2 cm Caso 2: como $g(u;\theta)g(v;\theta)J(\theta)$ es continua como funci\'on de $\theta$, entonces es uniformemente continua en $\Theta_1$, donde $\Theta_1$ es un conjunto compacto que contiene a $\theta$. Adem\'as, como $u,v\in [0,1]^2\,$ y $\,\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, entonces
\[\sup_{u,v\in [0,1]^2} \left|A(u)^{\top} \left[g(u;\hat{\theta}_n)g(v;\hat{\theta}_n)J(\hat{\theta}_n) -g(u;\theta)g(v;\theta)J(\theta)\right]B(v)\right|=o(1).\]
\item Caso 3: sea $f(u,v;\vartheta)$ una funci\'on polinomial en las variables $u, v\in [0,1]^2$ y $\vartheta\in\Theta_1\subseteq\Theta_0$, donde $\Theta_1$ es un conjunto compacto que contiene a $\theta$. Por lo tanto $f$ es continua como funci\'on de $\vartheta$ y tambi\'en lo es $g(u;\vartheta)$, luego $f(u,v;\hat{\theta}_n)=f(u,v;\theta)+o(1)$ y $g(u;\hat{\theta}_n)=g(u;\theta)+o(1)$, con lo cual
\begin{align}
f(u,v;\hat{\theta}_n)&g(u;\hat{\theta}_n)E_*\! \left\{v_1^{X^*_{1}}v_2^{X^*_{2}} \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\right\}- f(u,v;\theta)g(u;\theta)E_{\theta}\!\left\{v_1^{X_{1}}v_2^{X_{2}} \boldsymbol{\ell}(\boldsymbol{X}_1;\theta)\right\}\notag\\[.2 cm]
&\hspace{4mm}=f(u,v;\theta)g(u;\theta)\left[E_*\!\left\{v_1^{X^*_{1}}v_2^{X^*_{2}} \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\right\}- E_{\theta}\!\left\{v_1^{X_{1}}v_2^{X_{2}} \boldsymbol{\ell}(\boldsymbol{X}_1;\theta)\right\}\right]\notag\\[.2 cm]
&\hspace{12mm}+o(1)\{f(u,v;\theta) +g(u;\theta)+1\}E_*\!\left\{v_1^{X^*_{1}}v_2^{X^*_{2}} \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\right\}.\notag
\end{align}
Usando el Lema \ref{P(hat(t))*l(hat(t))-P(t)*l(t)} y el hecho que $\hat{\theta}_n\in \Theta_0$, resulta
\begin{align}
E_*\!\left\{v_1^{X^*_{1}}v_2^{X^*_{2}} \boldsymbol{\ell}_i(\boldsymbol{X}^*_1;\hat{\theta}_n)\right\}&- E_{\theta}\{v_1^{X_{1}}v_2^{X_{2}} \boldsymbol{\ell}_i(\boldsymbol{X}_1;\theta)\}\notag\\[.1 cm]
&=\sum_{r,s \geq 0} v_1^r v_2^s\left\{\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)-\boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right\}\notag\\[.1 cm]
&\leq \sum_{r,s \geq 0}\left|\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)-\boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|=o(1),\notag
\end{align}
para $i=1,2,3$. Por lo tanto,
\[\left|E_*\!\left\{v_1^{X^*_{1}}v_2^{X^*_{2}} \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\right\}- E_{\theta}\!\left\{v_1^{X_{1}}v_2^{X_{2}} \boldsymbol{\ell}(\boldsymbol{X}_1;\theta)\right\}\right|=\mathbf{o}(1).\]
Ahora, puesto que $J(\hat{\theta}_n)< \infty$, entonces
\[\left\|E_*\!\left\{v_1^{X^*_{1}}v_2^{X^*_{2}} \boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\right\}\right\|\leq \!\left\{ E_*\!\left(\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\|^2\right) \right\}^{1/2}< \infty.\]
Adem\'as, como $|g(u;\theta)|\leq 1$ y $|f(u,v;\theta)|< \infty$, $\forall u, v\in [0,1]^2$, entonces
\[\sup_{u,v\in [0,1]^2}\left|\left[\phi(u,v;\hat{\theta}_n) E_*\!\left\{\varphi(v;\boldsymbol{X}^*_{1};\hat{\theta}_n)\right\} -\phi(u,v;\theta)E_{\theta}\{\varphi(v;\boldsymbol{X}_{1};\theta)\} \right]A(u)\right|=o(1),\]
donde $\,\phi(u,v;\vartheta)=f(u,v;\vartheta) g(u;\vartheta)\ $ y $\ \varphi(v;\boldsymbol{X}_{1};\vartheta)=v_1^{X_{1}}v_2^{X_{2}} \boldsymbol{\ell}(\boldsymbol{X}_{1};\vartheta)$.
\item \vskip .15 cm Caso 4: como $g(u;\hat{\theta}_n)=g(u;\theta)+o(1)$, pues es continua como funci\'on de $\vartheta\in \Theta_0$, entonces
\begin{align}
&g(u;\hat{\theta}_n)\,E_*\!\left\{X^*_{2}\,I{\{X^*_{2}\geq 1\}}\,v_1^{X^*_{1}}\,v_2^{X^*_{2}-1}\boldsymbol{\ell}(\boldsymbol{X}^*_1; \hat{\theta}_n)\right\}\notag\\[.1 cm]
&\hspace{40 mm}-g(u;\theta)E_{\theta}\left\{X_{2}\,I{\{X_{2}\geq 1\}}\,v_1^{X_{1}}\,v_2^{X_{2}-1}\boldsymbol{\ell}(\boldsymbol{X}_1; \theta)\right\}\notag\\[.2 cm]
&\hspace{35 mm} =g(u;\theta)\biggl[E_*\!\left\{X^*_{2}\,I{\{X^*_{2}\geq 1\}}\,v_1^{X^*_{1}}\,v_2^{X^*_{2}-1}\boldsymbol{\ell}(\boldsymbol{X}^*_1; \hat{\theta}_n)\right\}\biggr.\notag\\
&\hspace{55 mm}\biggl.-E_{\theta}\left\{X_{2}\,I{\{X_{2}\geq 1\}}\,v_1^{X_{1}}\,v_2^{X_{2}-1}\boldsymbol{\ell}(\boldsymbol{X}_1; \theta)\right\}\biggr]\notag\\[.2 cm]
&\hspace{55 mm}+o(1)\,E_*\!\left\{X^*_{2}\,I{\{X^*_{2}\geq 1\}}\,v_1^{X^*_{1}}\,v_2^{X^*_{2}-1}\boldsymbol{\ell}(\boldsymbol{X}^*_1; \hat{\theta}_n)\right\}.\notag
\end{align}
Usando el Lema \ref{P(hat(t))*l(hat(t))-P(t)*l(t)} y el hecho que $\hat{\theta}_n\in \Theta_0$, obtenemos
\begin{align}
E_*\!\left\{X^*_{2}I{\{X^*_{2}\geq 1\}}v_1^{X^*_{1}}v_2^{X^*_{2}-1}\boldsymbol{\ell}_i(\boldsymbol{X}^*_1; \hat{\theta}_n)\right\}&-E_{\theta}\!\left\{X_{2}I{\{X_{2}\geq 1\}}v_1^{X_{1}}v_2^{X_{2}-1}\boldsymbol{\ell}_i(\boldsymbol{X}_1; \theta)\right\}\notag\\[.2 cm]
&\hspace{-38mm}=\sum_{r\geq 0,s \geq 1}s \,v_1^r v_2^{s-1}\left\{\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)-\boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right\}\notag\\[.2 cm]
&\hspace{-38mm}\leq\sum_{r,s \geq 0} s \left|\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)-\boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|=o(1),\notag
\end{align}
para $i=1,2,3$. Por lo tanto,
\[\left|E_*\!\left\{\psi_2(v;\boldsymbol{X}_1^*;\hat{\theta}_n)\right\} -E_{\theta}\!\left\{\psi_2(v;\boldsymbol{X}_1;\theta)\right\}\right| =\mathbf{o}(1),\]
donde $\, \psi_2(v;\boldsymbol{X}_1;\vartheta)=X_2\,I{\{X_2\geq 1\}}\,v_1^{X_1}\,v_2^{X_2-1}\boldsymbol{\ell}(\boldsymbol{X}_1;\vartheta)$.
Por otra parte,
\[E_*\!\left\{\psi_2(v;\boldsymbol{X}_1^*;\hat{\theta}_n)\right\}\leq \left\{E_*\!\left(X^{*2}_{2}\right)\right\}^{1/2} \left\{E_*\!\left(\|\boldsymbol{\ell}(\boldsymbol{X}^*_1; \hat{\theta}_n)\|^2\right)\right\}^{1/2}< \infty,\]
con lo cual,
\[\sup_{u,v \in [0,1]^2}\left|\left[g(u;\hat{\theta}_n)E_*\! \left\{\psi_2(v;\boldsymbol{X}^*_1;\hat{\theta}_n)\!\right\}-g(u;\theta) E_{\theta}\!\left\{\psi_2(v;\boldsymbol{X}_1;\theta)\right\}\right]A(u) \right|=o(1).\]
\end{itemize}
La presentaci\'on de estos cuatro casos generales muestran que se verifica (\ref{E*S-EtS}) y con ello se consigue el resultado. $\square$
\begin{teorema}\label{Cons-boot-Rnw-Snw}
Sean $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid de $\boldsymbol{X}=(X_{1},X_{2}) \in \mathbb{N}_0^2$. Supongamos que se verifica el Supuesto \ref{E(l*l)-l-cont} y que $\hat{\theta}_n\mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un $\theta \in \Theta$. Entonces
\begin{itemize}
\item [(a)]$\displaystyle\sup_{x\in\mathbb{R}} \left|P_*\left\{R^*_{n,w}(\hat{\theta}^{*}_n) \leq x\right\}-P_{\theta}\left\{R_{n,w} (\hat{\theta}_n )\leq x\right\}\right|\ \mathop{\longrightarrow}\limits^{c.s.}\ 0.$
\item [(b)]$\displaystyle\sup_{x\in\mathbb{R}} \left|P_*\left\{S^*_{n,w}(\hat{\theta}^{*}_n) \leq x\right\}-P_{\theta}\left\{S_{n,w} (\hat{\theta}_n )\leq x\right\}\right|\ \mathop{\longrightarrow}\limits^{c.s.}\ 0.$
\end{itemize}
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt}Solamente presentaremos la demostraci\'on de la parte (b), pues la demostraci\'on de la parte (a) sigue pasos similares.
Por definici\'on, $S_{n,w}^*(\hat{\theta}_n^*)= \|Z_{1n}^*\|_{_\mathcal{H}}^{2} +\|Z_{2n}^*\|_{_\mathcal{H}}^{2}$, con
\[Z_{kn}^*(u)=\frac{1}{\sqrt{n}}\sum_{i=1}^n V_k(\boldsymbol{X}_i^*;\hat{\theta}_n^*; u),\, k=1,2.\]
Siguiendo pasos similares a los dados en la demostraci\'on del Teorema \ref{ConvDebil-Rnw-Snw}(b) se puede ver que
\[Z_{kn}^*(u)=S_{kn}^*(u)+s^*_{kn},\ k=1,2,\]
con $\|s_{kn}^*\|_{\mathcal{H}}=o_{_{P_*}}(1)$ c.s., $k=1,2$, donde $S_{kn}^*(u)$ es definido como $S_{kn}(u)$ que aparece en el Teorema \ref{ConvDebil-Rnw-Snw}, con
$\boldsymbol{X}_i$ y $\theta$ reemplazados por $\boldsymbol{X}^*_i$ y $\hat{\theta}_n$, respectivamente.
Por conveniencia anal\'itica, ahora consideraremos el siguiente espacio de Hilbert separable:
\begin{align}
\mathcal{H}_1&=\Biggl\{\varphi:[0,1]^2\rightarrow \mathbb{R}^2, \text{ donde } \varphi(u)=(\varphi_1(u), \varphi_2(u)) \text{ es una funci\'on medible tal que} \Biggr.\notag\\[.2 cm]
&\hspace{45 mm} \Biggl.\|\varphi\|_{_{\mathcal{H}_1}}^{2}\!=\!\! \int_{[0,1]^2}\! \left\{\varphi_1^2(u)+ \varphi_2^2(u)\right\}w(u)du<\infty \Biggr\}.\notag
\end{align}
con producto escalar \[\langle(\phi_1, \phi_2), (\psi_1,\psi_2) \rangle_{_{\mathcal{H}_1}}=\int_{[0,1]^2} \!\! \left\{\phi_1(u)\psi_1(u)+\phi_2(u)\psi_2(u)\right\} w({u})d{u}<\infty.\]
Claramente, $\|\varphi\|_{_{\mathcal{H}_1}}^{2}=
\|\varphi_1\|_{_{\mathcal{H}}}^{2}+\|\varphi_2\|_{_{\mathcal{H}}}^{2}$.
Sea \[Y_n^*(u)=\sum_{i=1}^n Y_{ni}^*(u)\]
donde
\[Y_{ni}^*(u)= \frac{1}{\sqrt{n}}\,S(\boldsymbol{X}_i^*;\hat{\theta}_n; u),\ 1\leq i \leq n,\]
\begin{equation}\label{vector-S}
S(\boldsymbol{X}_i^*;\hat{\theta}_n; u)=\left(S_1^0(\boldsymbol{X}_i^*;\hat{\theta}_n; u), S_2^0(\boldsymbol{X}_i^*;\hat{\theta}_n; u)\right),\ 1\leq i \leq n.
\end{equation}
Observar primero que $Y_{ni}^*(u)$ tiene medias cero y momentos segundos finitos, para $1\leq i \leq n$, siempre que $\hat{\theta}_n\in \Theta_0$, lo cual ocurre c.s. porque $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$.\vskip .15 cm
Consideremos el n\'ucleo de covarianza $\ K^S_n(u,v)= E_*\!\left\{Y_n^*(u)^\top Y_n^*(v)\right\}$, luego
\[K^S_n(u,v)=\left(E_*\left\{S_i^0(\boldsymbol{X}_1^*;\hat{\theta}_n; u) S_j^0(\boldsymbol{X}_1^*;\hat{\theta}_n; v)\right\}\right), 1\leq i,j \leq 2.\]
Adem\'as, sea $K^S(u,v)=E_{\theta}\left\{S(\boldsymbol{X}_1;{\theta}; u)^{\top} \, S(\boldsymbol{X}_1;{\theta}; v)\right\}$ donde $S$ es como el definido en (\ref{vector-S}), es decir,
\[K^S(u,v)=\Bigl(E_{\theta}\left\{S_i^0(\boldsymbol{X}_1;\theta; u) S_j^0(\boldsymbol{X}_1;\theta; v)\right\}\Bigr), 1\leq i,j \leq 2.\]
Sea $\mathcal{Z}$ un proceso Gaussiano centrado cuyo operador de covarianza $C$ es caracterizado por
\begin{align}
\langle C f, h\rangle_{_{\mathcal{H}_1}}=cov\left(\langle \mathcal{Z}, f\rangle_{_{\!\mathcal{H}_1}}, \langle \mathcal{Z}, h\rangle_{_{\!\mathcal{H}_1}}\right)&=E_{\theta}\left\{\langle \mathcal{Z}, f\rangle_{_{\!\mathcal{H}_1}} \langle \mathcal{Z}, h\rangle_{_{\!\mathcal{H}_1}}\right\}\notag\\[.2 cm]
&=\int_{[0,1]^4} f(u)K^S(u,v)h(v)^{\top} w(u)w(v)dudv.\label{Oper-cov-C-Yn}
\end{align}
Por el Lema \ref{TCL-Hilbert}, $(S_{1n},S_{2n})\stackrel{L}{\longrightarrow} \mathcal{Z}$ en $\mathcal{H}_1$, cuando los datos son iid provenientes del vector aleatorio $\boldsymbol{X}\sim PB(\theta)$, donde $S_{kn}(u)$ es como el definido en el Teorema \ref{ConvDebil-Rnw-Snw}(b), $k=1,2$.
Sea $\{e_k: \, k \geq 0\}$ una base ortonormal de $\mathcal{H}_1$. A continuaci\'on probaremos que se cumplen las condiciones (i)-(iii) del Lema \ref{kundu}.
\begin{itemize}
\item [(i)] Sea $C_n$ el operador de covarianza de $Y_n^*$, esto es, sean $f,h \in \mathcal{H}_1$, luego \begin{align} \langle C_n f, h\rangle_{_{\mathcal{H}_1}}&=cov(\langle Y_n^*, f\rangle_{_{\!\mathcal{H}_1}}, \langle Y_n^*, h\rangle_{_{\!\mathcal{H}_1}})=E_{*}\{\langle Y_n^*, f\rangle_{_{\!\mathcal{H}_1}} \langle Y_n^*, h\rangle_{_{\!\mathcal{H}_1}}\}\notag\\[.2 cm] &=\int_{[0,1]^4} f(u)E_*\left\{Y_n^{*}(u)^\top Y_n^*(v)\right\}h(v)^\top w(u)w(v)dudv\notag\\[.2 cm] &=\int_{[0,1]^4} f(u)K^S_n(u,v)h(v)^{\top} w(u)w(v)dudv.\label{def-Cn-Yn} \end{align}
De (\ref{Oper-cov-C-Yn}), (\ref{def-Cn-Yn}) y de la Proposici\'on \ref{Nucleos-Rnw-Snw}(b),
\begin{align} \lim_{n\to \infty}\,\langle C_n e_k, e_l\rangle_{_{\mathcal{H}_1}} &=\lim_{n\to \infty}\int_{[0,1]^4} e_k(u)K^S_n(u,v)e_l(v)^{\top}w(u)w(v)du dv\notag\\ &=\int_{[0,1]^4} e_k(u)K^S(u,v)e_l(v)^{\top}w(u)w(v)dudv=\langle C e_k, e_l\rangle_{_{\mathcal{H}_1}} = a_{kl}.\notag \end{align}
\item[(ii)] De (\ref{Oper-cov-C-Yn}), (\ref{def-Cn-Yn}), de la Proposici\'on \ref{Nucleos-Rnw-Snw}(b) y del \'item (i), se obtiene
\begin{align} \lim_{n\to \infty}\sum_{k=0}^{\infty} \langle C_{n} e_k,e_k\rangle_{_{\mathcal{H}_1}} &= \lim_{n\to \infty}\sum_{k=0}^{\infty} \int_{[0,1]^4} e_k(u)K^S_n(u,v)e_k(v)^{\top}w(u)w(v)du dv\notag\\ &=\sum_{k=0}^{\infty} \int_{[0,1]^4} e_k(u)K^S(u,v)e_k(v)^{\top}w(u)w(v)du dv\notag\\ &=\sum_{k=0}^{\infty} \langle Ce_k,e_k\rangle_{_{\mathcal{H}_1}}= \sum_{k=0}^{\infty} a_{kk}<\infty,\notag \end{align}
pues, de la primera ecuaci\'on en (\ref{Oper-cov-C-Yn}) y la igualdad de Parseval
\[\sum_{k=0}^{\infty} a_{kk}=\sum_{k=0}^{\infty} \langle C e_k,e_k\rangle_{_{\mathcal{H}_1}}=\sum_{k=0}^{\infty} E_{\theta}\!\left\{\langle \mathcal{Z},e_k\rangle^2_{_{\mathcal{H}_1}}\right\}= E_{\theta}\!\left\{\|\mathcal{Z}\|_{_{\mathcal{H}_1}}^{\,2}\right\}< \infty.\]
\item[(iii)] Como $|g(u;\hat{\theta}_n)|\leq 1$, entonces \[|S_k^0(\boldsymbol{X}_i^*;\hat{\theta}_n;u)| \leq X^*_{ki}+\hat{\theta}_{kn}+\hat{\theta}_{3n}+\sqrt{2} \|\boldsymbol{\ell}(\boldsymbol{X}_i^*;\hat{\theta}_n)\|,\ \forall u \in [0,1]^2,\ k=1,2.\] Sea $\, 0<M=\int_{[0,1]^2}w(u)du<\infty$, luego
\[\langle Y_{ni}^*,e_k \rangle^2_{_{\mathcal{H}_1}}\leq\frac{4M}{n} \left(\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|+ \frac{1}{2}A_{i}\right)^2,\]
donde $A_{i}=(X^*_{1i}+\hat{\theta}_{1n}+\hat{\theta}_{3n})+ (X^*_{2i}+\hat{\theta}_{2n}+\hat{\theta}_{3n})$, siempre que $\hat{\theta}_n \in \Theta_0$, lo cual ocurre c.s. puesto que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$.
Ahora, como
\[\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|+ \frac{1}{2}A_i\leq
\left\{
\begin{array}{ll}
2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|, & \hbox{si}\ \, A_i\leq 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|\\[.25 cm]
A_i, & \hbox{si}\ \, 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|< A_i
\end{array}
\right.,
\]
entonces
\[\varepsilon< \left|\langle Y_{ni}^*,e_k \rangle_{_{\mathcal{H}_1}}\right| \leq \frac{2\sqrt{M}}{\sqrt{n}}\left(\|\boldsymbol{\ell}(\boldsymbol{X}^*_i; \hat{\theta}_n)\|+ \frac{1}{2}A_i\right),\]
implica que
\begin{align}
I\left\{\left|\langle Y_{ni}^*,e_k \rangle_{_{\mathcal{H}_1}}\right|>\varepsilon\right\} &\leq I\left\{\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|+ \frac{1}{2}A_i>\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}}\right\}\notag\\[.2 cm]
&= I\left\{\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|> \gamma, \ A_i\leq 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\| \right\}\notag\\[.2 cm] &\qquad +I\left\{A_i>\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}},\ A_i> 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\| \right\},\notag
\end{align}
donde $\gamma=\gamma(n,\varepsilon)=\frac{\sqrt{n}\,\varepsilon}{4\sqrt{M}}$ es tal que $\gamma\to \infty\,$ cuando $\,n\to \infty$, pues $\,0<M< \infty$. \vskip .3 cm
Por lo tanto,
\begin{align}
&\langle Y_{ni}^*,e_k\rangle^2_{_{\mathcal{H}_1}} I\left\{\left|\langle Y_{ni}^*,e_k \rangle_{_{\mathcal{H}_1}}\right|>\varepsilon\right\}\notag\\[.2 cm]
&\hspace{30 mm}\leq \frac{16M}{n} \ \|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|^2 \,I\left\{\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\|> \gamma,\ A_i\leq 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\| \right\}\notag\\[.2 cm]
&\hspace{30 mm}\qquad+\frac{4M}{n}A_i^2\,I\left\{A_i>\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}}, \ A_i> 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n)\| \right\}.\notag
\end{align}
En consecuencia,
\[L_n(\varepsilon, e_k)=\sum_{i=1}^n E_*\!\left(\langle Y_{ni}^*,e_k\rangle^2_{_{\mathcal{H}_1}} I{\left\{\left|\langle Y_{ni}^*,e_k \rangle_{_{\mathcal{H}_1}}\right|>\varepsilon\right\}}\right)\leq L_{1n}(\varepsilon, e_k)+L_{2n}(\varepsilon, e_k),\]
donde,
\begin{align} L_{1n}(\varepsilon, e_k)&=\frac{16M}{n}\sum_{i=1}^n E_*\!\left[\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\|^2 \,I{\left\{\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\|> \gamma,\ A_1\leq 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\| \right\}}\right]\notag\\[.2 cm] &=16M E_*\!\left[\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\|^2 \,I{\left\{\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\|> \gamma,\ A_1\leq 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\| \right\}}\right]\notag\\[.2 cm] &\leq 16M \sup_{\hat{\theta}_n\in \Theta_0}E_{\hat{\theta}_n}\!\!\left[\|\boldsymbol{\ell} (\boldsymbol{X}^*_1;\hat{\theta}_n)\|^2\, I{\left\{\|\boldsymbol{\ell} (\boldsymbol{X}^*_1;\hat{\theta}_n)\| >\gamma\right\}}\right]\notag \end{align} y como $\gamma\to \infty$ cuando $n\to \infty$, entonces, por el Supuesto \ref{E(l*l)-l-cont}(1) se concluye que
\begin{equation}\label{L1n}\lim_{n\to \infty}L_{1n}(\varepsilon,e_k)=0.\end{equation}
Considerando la desigualdad de H\"{o}lder y el hecho que la funci\'on indicadora es tal que $0\leq I\{C\}\leq 1$, podemos escribir \begin{align} L_{2n}(\varepsilon, e_k)&=\frac{4M}{n}\sum_{i=1}^n E_*\!\left(A_1^2\,I{\left\{A_1>\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}},\ A_1> 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\| \right\}}\right)\notag\\[.2 cm]
&=4M E_*\!\left(A_1^2\,I{\left\{A_1>\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}},\ A_1> 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\| \right\}}\right)\notag\\[.2 cm]
&\leq 4M \left\{E_*\!\left(A_1^3\right)\right\}^{2/3}\left[ E_*\!\left(I{\left\{A_1>\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}},\ A_1> 2\|\boldsymbol{\ell}(\boldsymbol{X}^*_1;\hat{\theta}_n)\| \right\}}\right)\right]^{1/3}\notag\\[.2 cm]
&= 4M \left\{E_*\!\left(A_1^3\right)\right\}^{2/3}\left\{ P_*\!\left(A_1>t \right)\right\}^{1/3},\notag \end{align} donde $t=t(n,\varepsilon)=\frac{\sqrt{n}\,\varepsilon}{2\sqrt{M}}$ es tal que $t\to \infty$ cuando $n\to \infty$, pues $\,0<M< \infty$.\vskip .2 cm
Puesto que $|A_1|=A_1$ y $E_*\!\left(A_1^3\right)< \infty$, el Corolario 1.14 (ii) en Serfling (1980, p. 47) implica que $P_*\!\left(A_1>t \right)\to 0$ cuando $t\to \infty$. Adem\'as, como $\, 0<M< \infty$, entonces
\begin{equation}\label{L2n}
\lim_{n\to \infty}L_{2n}(\varepsilon,e_k)=0.\end{equation}
De (\ref{L1n}) y (\ref{L2n}), concluimos que
\[\lim_{n\to \infty}L_n(\varepsilon,e_k)=0,\]
de donde se cumple la condici\'on (iii) de Kundu et al. (2000).
\end{itemize}
As\'i, del Lema \ref{kundu}, tenemos que
$Y_n^*\mathop{\longrightarrow} \limits^{L} \mathcal{Z}$, c.s., en $\mathcal{H}_1$. Como hab\'iamos mencionado antes, $\mathcal{Z}$ es tambi\'en el l\'imite d\'ebil de $(S_{1n},S_{2n})$ cuando los datos son iid del vector aleatorio $\boldsymbol{X}\sim PB(\theta)$. Finalmente, el resultado sigue del teorema de la aplicaci\'on continua. $\square$
\begin{observacion}
Es importante observar que el resultado del Teorema \ref{Cons-boot-Rnw-Snw} se cumple si $H_0$ es verdadera o no.
Si de hecho, $H_0$ es verdadera, el Teorema \ref{Cons-boot-Rnw-Snw} implica que la distribuci\'on condicional de $R^*_{n,w}(\hat{\theta}^{\,*}_n)$ est\'a cercana c.s. a la distribuci\'on nula de $R_{n,w}(\hat{\theta}_n)$.
Si $H_0$ es falsa, entonces el Teorema \ref{Cons-boot-Rnw-Snw} nos dice que la distribuci\'on condicional de $R^*_{n,w}(\hat{\theta}^{\,*}_n)$ y la distribuci\'on de $R_{n,w}(\hat{\theta}_n)$, cuando la muestra es tomada de una poblaci\'on con distribuci\'on $PB(\theta)$, est\'an c.s. cercanas, donde $\theta$ es el l\'imite c.s. de $\hat{\theta}_n$.
\end{observacion}
Sea
$$r^*_{n,w,\alpha}=\inf\!\left\{x:P_*(R^*_{n,w}(\hat{\theta}^*_n)\geq x)\leq \alpha\right\}$$
el percentil superior $\alpha$ de la distribuci\'on bootstrap de $R_{n,w}(\hat{\theta}_n)$.
Del Teorema \ref{Cons-boot-Rnw-Snw}, la funci\'on test
$$\Psi^*_R=\left\{
\begin{array}{ll}
1, & \text{si}\ R_{n,w}(\hat{\theta}_n)\geq r^*_{n,w,\alpha}, \\[.15 cm]
0, & \text{en caso contrario},
\end{array}
\right.$$
o equivalentemente, el test que rechaza $H_0$ cuando
$$p_{_R}^*=P_*\!\left(R^*_{n,w}(\hat{\theta}^{\,*}_n) \geq R_{obs}\right)\leq \alpha,$$
es asint\'oticamente correcto en el sentido que, cuando $H_0$ es cierta
$$\lim_{n\to \infty} P_{\theta}\!\left(\Psi^*_R=1\right)=\alpha,$$
donde $R_{obs}$ es el valor observado del test estad\'istico $R_{n,w} (\hat{\theta}_n)$.\\
Similarmente, sea
$$s^*_{n,w,\alpha}=\inf\!\left\{x:P_*(S^*_{n,w}(\hat{\theta}^*_n)\geq x)\leq \alpha\right\}$$
el percentil superior $\alpha$ de la distribuci\'on bootstrap de $S_{n,w}(\hat{\theta}_n)$.
Del Teorema \ref{Cons-boot-Rnw-Snw}, la funci\'on test
$$\Psi^*_S=\left\{
\begin{array}{ll}
1, & \text{si}\ S_{n,w}(\hat{\theta}_n)\geq s^*_{n,w,\alpha}, \\[.15 cm]
0, & \text{en caso contrario},
\end{array}
\right.$$
o equivalentemente, el test que rechaza $H_0$ cuando
$$p_{_S}^*=P_*\!\left(S^*_{n,w}(\hat{\theta}^{\,*}_n) \geq S_{obs}\right)\leq \alpha,$$
es asint\'oticamente correcto en el sentido que, cuando $H_0$ es cierta
$$\lim_{n\to \infty} P_{\theta}\!\left(\Psi^*_S=1\right)=\alpha,$$
donde $S_{obs}$ es el valor observado del test estad\'istico $S_{n,w} (\hat{\theta}_n)$.\\
Como es usual, en la pr\'actica, $r^*_{n,w,\alpha}\,$ o $\,p_{_R}^*$ deben ser aproximados por simulaci\'on. Dados $\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$ vectores aleatorios iid de $\boldsymbol{X}=(X_1,X_2)\in\mathbb{N}_0^2$, siguiremos los siguientes pasos:
\begin{itemize}
\item [$(1)$] Calcular el estimador $\hat{\theta}_n$ y calcular $R_{obs}$, el valor de $R_{n,w}(\hat{\theta}_n)$ para la muestra original.
\item [$(2)$] Generar una muestra bootstrap, digamos, $\boldsymbol{X}_1^*, \boldsymbol{X}_2^*,\ldots,\boldsymbol{X}_n^*$ iid desde la distribuci\'on $PB(\hat{\theta}_n)$.
\item [$(3)$] Sobre la base del bootstrap, calcular el estimador $\hat{\theta}^{\,*}_n$ y el valor del test estad\'istico, digamos $R^*$.
\item [$(4)$] Calcular $R_{n,w}(\hat{\theta}^*_n)$ para cada muestra bootstrap y denotar por $R^*_b,\ b=1,2,\ldots,B$, el respectivo valor resultante.
\item [$(5)$] Aproximar el $p-$valor bootstrap, $p_{_R}^*$, por medio de la expresi\'on $$\hat{p}_{_R}=\frac{card\{b:R^*_b\geq R_{obs}\}}{B},$$ o aproximar el punto cr\'itico, $r^*_{n,w,\alpha}$, por $R^*_{a:B}$, donde $a=[(1-\alpha)B]+1$, $[x]$ es la parte entera de $x$, y $R^*_{1:B}, R^*_{2:B},\ldots,R^*_{B:B}$ son los valores $R^*_b,\ b=1,2,\ldots,B$, en orden creciente.
\end{itemize}
\begin{observacion}\label{Boot-Snw} Para aproximar $s^*_{n,w,\alpha}$ se sigue el mismo procedimiento anterior, con los cambios obvios.
\end{observacion}
\section{Alternativas} \label{Alternativas_Rnw-Srw}
En esta secci\'on estudiaremos el comportamiento de los tests propuestos $\Psi^*_R$ y $\Psi^*_S$ bajo alternativas fijas y locales.
\subsection{Alternativas fijas} \label{Alternativas-fijas}
Como uno de nuestros objetivos es la consistencia de los tests de bondad de ajuste, lo pr\'oximo que haremos es estudiar este t\'opico para los tests que proponemos. Con este prop\'osito, primero obtendremos el l\'imite c.s. de $\frac{1}{n} R_{n,w}$ y $\frac{1}{n} S_{n,w}$.
\begin{teorema}\label{Cons-Rnw-Snw}
Sean $\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$ vectores aleatorios iid de $\boldsymbol{X}=(X_1,X_2)\in\mathbb{N}_0^2$ con fgp $g(u)$. Si $\hat{{\theta}}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un ${\theta}\in \mathbb{R}^3$, entonces
\begin{itemize}
\item[(a)] $\displaystyle\frac{1}{n}R_{n,w}(\hat{{\theta}}_n)\ \mathop{\longrightarrow}\limits^{c.s.} \ \int_0^1\int_0^1\Bigl\{g(u)- g(u;{\theta})\Bigr\}^2 w(u) du =\eta(g;{\theta}).$
\item[(b)] Si $g(u) \in G_2$,
$\displaystyle\frac{1}{n} S_{n,w}(\hat{\theta}_n) \ \mathop{\longrightarrow}\limits^{c.s.}\int_0^1\int_0^1 \left\{D^2_1(u;\theta)+D^2_2(u;\theta)\right\}w(u)du=\xi(g;\theta).$
\end{itemize}
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt} Solamente presentaremos la demostraci\'on de la parte (b), pues la demostraci\'on de la parte (a) sigue pasos similares.
Por definici\'on
\[\frac{1}{n}S_{n,w}(\hat{\theta}_n)=\|D_{1n}\|^2_{_\mathcal{H}}+ \|D_{2n}\|^2_{_\mathcal{H}},\]
donde $D_{kn}=D_{kn}(u;\hat{\theta}_n ), k=1,2$, est\'an definidas en (\ref{EDPs-empiricas}).
De las ecuaciones (\ref{EDPs-fgp}) y (\ref{EDPs-empiricas}), obtenemos
\begin{align}
\left|D_{1n}(u;\hat{\theta}_n )-D_1(u;\theta)\right|& \leq r_1+|\theta_1-\hat{\theta}_{1n}| + |\theta_3-\hat{\theta}_{3n}|+ |\hat{\theta}_{1n}+\hat{\theta}_{3n}|r_0, \notag\\
\left|D_{2n}(u;\hat{\theta}_n )-D_2(u;\theta)\right|&
\leq r_2+|\theta_2-\hat{\theta}_{2n}| + |\theta_3-\hat{\theta}_{3n}|+ |\hat{\theta}_{2n}+\hat{\theta}_{3n}|r_0,\notag
\end{align}
$\forall u \in [0,1]^2$, donde
\[r_0=\sup_{u\in [0,1]^2}\left|g_n(u)- g(u)\right|\]
y
\[r_i=\sup_{u\in [0,1]^2}\left|\frac{\partial } {\partial u_i}g_n(u)-\frac{\partial }{\partial u_i}g(u)\right|,\ i=1,2.\]
De la Proposici\'on \ref{Conv-FuncGenProbBiv}, $r_i=o(1)$, $i=0,1,2$. El hecho que $|\theta_i-\hat{\theta}_{in}| =o(1)$, $i=1,2,3$, junto con (\ref{int-funcion-peso}), implican que
\[\frac{1}{n}S_{n,w}(\hat{\theta}_n)=\|D_{1}\|^2_{_\mathcal{H}}+ \|D_{2}\|^2_{_\mathcal{H}}+o(1),\]
lo que demuestra el resultado porque
$\|D_{1}\|^2_{_\mathcal{H}}+ \|D_{2}\|^2_{_\mathcal{H}}=\xi(g;\theta)$. $\square$\\
Como una consecuencia de los Teoremas \ref{ConvDebil-Rnw-Snw}, \ref{Cons-boot-Rnw-Snw} y \ref{Cons-Rnw-Snw}, el siguiente resultado da la consistencia de los tests $\Psi^*_R$ y $\Psi^*_S$.
\begin{corolario}\label{Altern-fijas-Rnw-Snw}
Sean $\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$ vectores aleatorios iid de $\boldsymbol{X}\in\mathbb{N}_0^2$ con fgp $g(u)$. Supongamos que se cumplen las hip\'otesis de los Teoremas \ref{ConvDebil-Rnw-Snw} y \ref{Cons-boot-Rnw-Snw}.
\begin{itemize}
\item[(a)] Si $\eta(g;\theta)>0$, entonces
$P(\Psi^*_R=1) \to 1.$
\item[(b)] Si $g(u) \in G_2$ y $\xi(g;\theta)>0$, entonces $P(\Psi^*_S=1) \to 1.$
\end{itemize}\end{corolario}
\begin{observacion}
Notar que $\eta(g;{\theta})\geq 0$ ($\xi(g;\theta) \geq 0$). Si $w>0$ en casi todo $[0,1]^2$, entonces $\eta(g;\theta)= 0$ ($\xi(g;\theta)= 0$) s\'i y s\'olo si $H_0$ es cierta.
Por lo tanto, la consistencia de los tests propuestos est\'a garantizada simplemente tomando una funci\'on de peso que sea positiva en casi todo $[0,1]^2$.
\end{observacion}
\subsection{Alternativas contiguas} \label{Alternativas-contiguas-Rnw-Swn}
El resultado de consistencia dado en el Corolario \ref{Altern-fijas-Rnw-Snw} no distingue entre medidas de probabilidad alternativas de $P$. Una mejor discriminaci\'on se obtiene considerando alternativas para las cuales la potencia tiende a valores menores o iguales que 1. Esto se logra reemplazando una alternativa fija por una sucesi\'on de alternativas que converge a la nula a una cierta velocidad, que se suelen denominar alternativas contiguas.\vskip .2 cm
Con este objetivo consideremos ahora un arreglo triangular $\boldsymbol{X}_{n,1}, \boldsymbol{X}_{n,2},\ldots,\boldsymbol{X}_{n,n}$ de vectores aleatorios bivariantes independientes por filas que toman valores en $\mathbb{N}_0^2$ y funci\'on de probabilidad conjunta $P_n(x_1,x_2)$ dada por
\begin{equation}\label{P(n)=P(theta)+b}
P_n(x_1,x_2)=P_{\theta}(x_1,x_2)\left\{1+\frac{1}{\sqrt{n}} \, b_n(x_1,x_2)\right\},
\end{equation}
donde $P_{\theta}(x_1,x_2)$ es la funci\'on de probabilidad del vector aleatorio bivariante que tiene una distribuci\'on $PB(\theta)$, para alg\'un $\theta \in \Theta$, y $b_n(x_1,x_2)$ satisface las siguientes condiciones.
\begin{supuesto}\label{b(x,y)}
\begin{enumerate}
\item [$(1)$] $E_{\theta}\{b_n(X_1,X_2)\}=0$, $\forall n$.
\item [$(2)$] $b_n(x_1,x_2)\to b(x_1,x_2)$, $\forall (x_1,x_2)\in \mathbb{N}_0^2$.
\item [$(3)$] $\displaystyle \sup_{n} E_{\theta}\!\left\{b_n(X_1,X_2)^4\right\} <\infty$.
\end{enumerate}
\end{supuesto}
El Supuesto \ref{b(x,y)}(1) asegura que $\sum_{x_1,\,x_2 \in \mathbb{N}_0}P_n(x_1,x_2)=1$; los Supuestos \ref{b(x,y)}(2) y (3) contienen condiciones de tipo t\'ecnico que usaremos en las demostraciones.\vskip .2 cm
El Teorema \ref{ConvDebil-Rnw-Snw} establece que, cuando $H_0$ es cierta, $R_{n,w}(\hat{\theta}_n)$ converge en distribuci\'on a una combinaci\'on lineal de variables $\chi^2$ independientes con 1 grado de libertad, donde los pesos son los autovalores del operador $C_R(\theta)$ dado en (\ref{Operador-C}). Sea $\{\phi_j^R\}$ el conjunto de autofunciones ortonormales correspondiente a los autovalores $\{\lambda_j^R\}$ de $C_R(\theta)$.\vskip .2 cm
Observar que lo expuesto anteriormente para $R_{n,w}(\hat{\theta}_n)$, tambi\'en es v\'alido para $S_{n,w}(\hat{\theta}_n)$, considerando el operador $C_S(\theta)$ con sus conjuntos $\{\phi_j^S\}$ y $\{\lambda_j^S\}$ de autofunciones ortonormales y de autovalores, respectivamente.\vskip .2 cm
El siguiente Teorema da la ley l\'imite de estos estad\'isticos bajo las alternativas $P_n(x_1,x_2)$ en (\ref{P(n)=P(theta)+b}).
\begin{teorema}\label{Altern-contiguas-S_n,w}
Sea $\boldsymbol{X}_{n,1}, \boldsymbol{X}_{n,2},\ldots,\boldsymbol{X}_{n,n}$ un arreglo triangular de vectores aleatorios bivariantes que son independientes por filas y que toman valores en $\mathbb{N}_0^2$, con funci\'on de probabilidad dada por $P_n(x_1,x_2)$ definida en (\ref{P(n)=P(theta)+b}). Supongamos que se cumplen los Supuestos \ref{hat(theta)-theta} y \ref{b(x,y)}. Entonces
\begin{itemize}
\item[(a)] $R_{n,w}(\hat{\theta}_n)\, \mathop{\longrightarrow} \limits^{L} \, \displaystyle\sum_{k=1}^{\infty} \lambda_k^R\!\left(Z_k+c_k^R\right)^2,$
donde $\,c_k^R=\mathop{\sum}\limits_{x_1,x_2}b(x_1,x_2)\,\phi_k^R(x_1,x_2)$ y $Z_1,Z_2,\ldots\,$ son variables normales est\'andar independientes.
\item[(b)] \vskip .3 cm $S_{n,w}(\hat{\theta}_n)\ \mathop{\longrightarrow} \limits^{L} \ \displaystyle\sum_{k=1}^{\infty} \lambda_k^S\!\left(Z_k+c_k^S\right)^2,$
donde $\,c_k^S=\mathop{\sum}\limits_{x_1,x_2}b(x_1,x_2)\,\phi_k^S(x_1,x_2)\,$ y $Z_1,Z_2,\ldots\,$ son variables normales est\'andar independientes.
\end{itemize}
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt}
Sea $A_n \subseteq \mathbb{N}_0^2$ tal que $P_{\theta}(A_n)\to 0$, entonces
\[ P_n(A_n)=P_{\theta}(A_n)+\frac{1}{\sqrt{n}}\sum_{(x_1,x_2)\in A_n}P_{\theta}(x_1,x_2)b_n(x_1,x_2).\]
De la desigualdad de H\"older y del Supuesto \ref{b(x,y)}(3) se sigue que
\begin{align}
&\sum_{(x_1,x_2)\in A_n}P_{\theta}(x_1,x_2)b_n(x_1,x_2)\notag\\
&\hspace{35mm}\leq \left\{\sum_{(x_1,x_2)\in A_n}P_{\theta}(x_1,x_2) \right\}^{3/4}
\left\{\sum_{(x_1,x_2)\in A_n}P_{\theta}(x_1,x_2)b_n(x_1,x_2)^4 \right\}^{1/4} \notag\\[.25 cm]
&\hspace{35mm}\leq \sup_{n} E_{\theta}\left\{b_n(X_1,X_2)^4\right\} <\infty,\notag
\end{align}
y por tanto $P_n(A_n)\to 0$, esto es, $P_n$ es contigua a $P_{\theta}$.
A continuaci\'on veremos que, bajo $P_n$, $\sqrt{n}(\hat{\theta}_n-\theta)$ converge en ley a una distribuci\'on normal.
Con este objetivo, sean $\boldsymbol{X}_1, \boldsymbol{X}_2, \ldots, \boldsymbol{X}_n$ vectores aleatorios iid con funci\'on de probabilidad com\'un $P_{\theta}$ y sea
\[l_n=\log \frac{P_n(\boldsymbol{X}_1)P_n(\boldsymbol{X}_2)\ldots P_n(\boldsymbol{X}_n)}
{P_{\theta}(\boldsymbol{X}_1)P_{\theta}(\boldsymbol{X}_2)\ldots P_{\theta}(\boldsymbol{X}_n)}.\]
Sea $Z_{n,i}=b_n(\boldsymbol{X}_i)$, $1\leq i \leq n$. Entonces, por desarrollo en serie de Taylor,
\begin{equation} \label{ln}
l_n=\sum_{i=1}^n \log\left(1+\frac{1}{\sqrt{n}}Z_{n,i} \right)=\frac{1}{\sqrt{n}}\sum_{i=1}^n Z_{n,i}-\frac{1}{2n}\sum_{i=1}^n Z_{n,i}^2+\varrho_n,
\end{equation}
donde $\varrho_n=\varrho_n(Z_{n,1},Z_{n,2}, \ldots, Z_{n,n})$,
\[\varrho_n=\sum_{i=1}^n \frac{2}{3!\,n\sqrt{n}\,a^3_{in}} Z_{n,i}^3,\]
con $a_{in}=a_{in}(Z_{n,i})$
\[a_{in}=1+\frac{\alpha_i}{\sqrt{n}}\,Z_{n,i},\]
para alg\'un $0<\alpha_i<1$, $1\leq i \leq n$.
A continuaci\'on estudiamos el l\'imite de cada uno de los sumandos en el lado derecho de (\ref{ln}).
Sea $\varepsilon$ una constante positiva. Adem\'as, sea $M_n=\sqrt{n}/2$ y consideremos $\tilde{\varrho}_n=\varrho_n(Z_{n,1}I\{|Z_{n,1}|\leq M_n\}, \ldots, Z_{n,n}I\{|Z_{n,n}|\leq M_n\})$.
Notemos que $\tilde{a}_{in}=a_{in}(Z_{n,i}I\{|Z_{n,i}|\leq M_n\})\geq \frac{1}{2}, 1\leq i \leq n$, y por lo tanto
\[E_{\theta}(|\tilde{\varrho}_n|)\leq \frac{2^3}{3}\frac{1}{\sqrt{n}} \,\displaystyle \sup_m E_{\theta}\! \left\{|b_m(\boldsymbol{X}_1)|^3 \right\},\]
lo cual implica que
\begin{equation} \label{resto1}
P_{\theta}(|\tilde{\varrho}_n|>\varepsilon )\to 0.
\end{equation}
Como
\begin{align}
P_{\theta}\left(|\tilde{\varrho}_{n} -\varrho_{n}|>\varepsilon\right) &\leq P_{\theta}\left(\bigcup_{i=1}^n\left\{|Z_{n,i}|> M_n \right\}\right)\notag\\
&\leq \sum_{i=1}^n P_{\theta}\left(|Z_{n,i}|> M_n \right)\notag\\[.15 cm]
&\leq \frac{n}{M_n^4}\,\sup_m E_{\theta}\!\left\{b_m(\boldsymbol{X}_1)^4 \right\}\to 0.\label{resto2}
\end{align}
De (\ref{resto1}) y (\ref{resto2}), concluimos que
\begin{equation} \label{resto}
P_{\theta}\left(|\varrho_{n}|>\varepsilon\right)\to 0.
\end{equation}
Por el Supuesto \ref{b(x,y)},
$E_{\theta}(Z_{n,i})=0,$
y
\[\sigma_n^2=Var_{\theta}(Z_{n,i}) \to \sigma^2=E_{\theta}\left\{b(X_1,X_2)^2\right\}<\infty.\]
Adem\'as,
\begin{align}
E_{\theta}\left[Z_{n,i}^2\,I\{|Z_{n,i}|>na\} \right] &\leq E_{\theta}^{1/2}\left(Z_{n,i}^4 \right)P^{1/2}_{\theta}\!
\left(|Z_{n,i}|>na\right)\notag\\[.2 cm]
&\leq\frac{1}{(na)^4}\left[\sup_n E_{\theta}\left\{b_n(X_1,X_2)^4 \right\} \right]^2\to 0, \quad \forall a>0.\notag
\end{align}
Por el teorema central del l\'imite para arreglos triangulares (Teorema 1.9.3 en Serfling, 1980, pp. 31-32, \cite{Ser80}) se sigue que
\begin{equation}\label{Zi-conv-N}
\frac{1}{\sqrt{n}}\sum_{i=1}^n Z_{n,i}\ \mathop{\longrightarrow} \limits^{L}\ N(0,\sigma^2).
\end{equation}
Se tiene que
\[E_{\theta}\left(\frac{1}{n}\sum_{i=1}^n Z_{n,i}^2\right)=\sigma_n^2\to \sigma^2,\]
\[Var_{\theta}\left(\frac{1}{n}\sum_{i=1}^n Z_{n,i}^2\right)=\frac{Var_{\theta}(Z_{n,i}^2)}{n}\leq \frac{1}{n}\sup_n E_{\theta}\left\{b_n(X_1,X_2)^4\right\}\to 0,\]
de donde
\begin{equation}\label{Zi^2-conv-sigma^2}
\frac{1}{2n}\sum_{i=1}^n Z_{n,i}^2 \mathop{\longrightarrow} \limits^{P}\ \frac{1}{2}\sigma^2.
\end{equation}
Ahora, del Supuesto \ref{hat(theta)-theta}, (\ref{ln}) y (\ref{resto})--(\ref{Zi^2-conv-sigma^2}), la sucesi\'on
\[\left( \sqrt{n}(\hat{\theta}_n-\theta), l_n \right)\]
converge en ley a una distribuci\'on normal multivariante de dimensi\'on 4, cuando los datos provienen de $P_{\theta}$. As\'i, por el tercer Lema de Le Cam (Corolario 12.3.2 en Lehmann y Romano (2005) \cite{LehRom05}), concluimos que, cuando los datos provienen de $P_n$, $\sqrt{n}(\hat{\theta}_n-\theta)$ converge en distribuci\'on a una ley normal, lo cual implica que es acotada en probabilidad.\vskip .2 cm
Siguiendo pasos similares a los dados en la demostraci\'on del Teorema \ref{ConvDebil-Rnw-Snw} (b), podemos probar que (\ref{Snw-Sn}) tambi\'en se cumple cuando los datos provienen de $P_n$, con $P_n(|s_n|>\varepsilon)\to 0$, $\forall \varepsilon>0$.
As\'i, cuando los datos tienen la funci\'on de probabilidad $P_n$, aplicando el Teorema 2.3 en Gregory (1977) \cite{Gre77}, obtenemos que
\[\|S_{1n}\|_{_\mathcal{H}}^2+\|S_{2n}\|_{_\mathcal{H}}^2 \mathop{\longrightarrow} \limits^{L} \ \sum_{k=1}^{\infty} \lambda_k^S\left(Z_k+c_k^S\right)^2,\]
y con ello se consigue el resultado en (b). La demostraci\'on del resultado en (a) sigue los mismos pasos. $\square$\vskip .2 cm
Del Teorema \ref{Altern-contiguas-S_n,w}, concluimos que el test $\Psi^*_R$ ($\Psi^*_S$) es capaz de detectar alternativas como las establecidas en (\ref{P(n)=P(theta)+b}), que convergen a la DPB a una raz\'on de $n^{-1/2}$.
\chapter{Estad\'istico $\boldsymbol{W_n(\hat{\theta}_n)}$}\label{Estadistico-Wn}
\section{Definici\'on del test estad\'istico} \label{Def-Test-estadistico}
En este cap\'itulo proponemos otro test para contrastar la hip\'otesis nula, cuyo test estad\'istico lo deduciremos de la Proposici\'on \ref{Soluc-sistema-de-dos-EDP}. Para ello consideremos que se verifican las hip\'otesis en dicha proposici\'on.
Sea $(X_1,X_2)\in \mathbb{N}^2_0$ un vector aleatorio y sea $g(u_1,u_2)=E\left(u_1^{X_1} u_2^{X_2}\right)$ su fgp, luego, por definici\'on
\[g(u)=\sum_{r,s\geq 0}u_1^r u_2^s P(r,s),\]
donde $P(r,s)=P(X_1=r,X_2=s)$.\vskip .2 cm
De la ecuaci\'on anterior y de las definiciones de $D_1(u;\theta)$ y $D_2(u;\theta)$ dadas en la Proposici\'on \ref{Soluc-sistema-de-dos-EDP}, podemos escribir
\begin{align}
D_1(u;\theta)&=\sum_{r\geq 0}\sum_{s\geq 0}\left\{(r+1)P(r+1,s)-(\theta_1-\theta_3)P(r,s)-\theta_3P(r,s-1)\right\}u_1^r u_2^s,\notag \\[.2 cm]
D_2(u;\theta) &=\sum_{r\geq 0}\sum_{s\geq 0}\left\{(s+1)P(r,s+1)-(\theta_2-\theta_3)P(r,s)-\theta_3P(r-1,s)\right\}u_1^r u_2^s.\notag
\end{align}
Consideremos ahora las versiones emp\'iricas de las ecuaciones anteriores, esto es, consideremos $D_{1n}(u;\hat{\theta}_n)$ y $D_{2n}(u;\hat{\theta}_n)$ definidas en (\ref{EDPs-empiricas}). Si $H_0$ fuera cierta entonces $D_{1n}(u;\hat{\theta}_n)$ y $D_{2n}(u;\hat{\theta}_n)$ deber\'ian ser pr\'oximas a 0, $\forall u \in [0,1]^2$. Esta proximidad puede interpretarse de varias formas. En el cap\'itulo anterior ya vimos que esto era equivalente a
$\int \{D_{1n}(u;\hat{\theta}_n)^2+D_{2n}(u;\hat{\theta}_n)^2\}w(u)du\approx 0$.\vskip .2 cm
Veamos otra interpretaci\'on siguiendo un razonamiento similar al hecho en Nakamura y P\'erez-Abreu (1993) \cite{NaPe93} para el caso univariante, que presentamos al final de la Secci\'on \ref{Test-bondad-ajuste-dim-1}. Para ello observemos que
\begin{align}
D_{1n}(u;\hat{\theta}_n)&=\sum_{r\geq 0}\sum_{s\geq 0} d_1(r,s;\hat{\theta}_n) u_1^ru_2^s,\notag\\[.2 cm]
D_{2n}(u;\hat{\theta}_n)&=\sum_{r\geq 0}\sum_{s\geq 0} d_2(r,s;\hat{\theta}_n) u_1^ru_2^s,\notag
\end{align}
donde
\begin{align}
d_1(r,s;\hat{\theta}_n)&=(r+1)p_n(r+1,s)-(\hat{\theta}_{1n}-\hat{\theta}_{3n}) p_n(r,s)-\hat{\theta}_{3n} p_n(r,s-1),\notag\\[.2 cm]
d_2(r,s;\hat{\theta}_n)&=(s+1)p_n(r,s+1)-(\hat{\theta}_{2n}-\hat{\theta}_{3n}) p_n(r,s)-\hat{\theta}_{3n} p_n(r-1,s),\notag
\end{align}
y
\[p_n(r,s)=\frac{1}{n}\sum_{k=1}^n I{\{X_{1k}=r,X_{2k}=s\}}\]
es la frecuencia relativa emp\'irica del par $(r,s)$. Por tanto, $D_{in}(u;\hat{\theta}_n)= 0$,
$\forall u \in [0,1]^2$, $i=1,2$, s\'i y s\'olo si los coeficientes de $u_1^ru_2^s$ en las expansiones anteriores son nulos $\forall r,s \geq 0$. Esto nos lleva a considerar el siguiente estad\'istico para contrastar $H_0$:
\begin{equation}\label{Estad-Wn-bivariante}
W_n(\hat{\theta}_n)=\sum_{r\geq 0}\sum_{s\geq 0}\{d_1(r,s;\hat{\theta}_n)^2+d_2(r,s;\hat{\theta}_n)^2\}=
\sum_{r,s = 0}^M\{d_1(r,s;\hat{\theta}_n)^2+d_2(r,s;\hat{\theta}_n)^2\},
\end{equation}
donde $M=\max \{X_{1(n)}, X_{2(n)}\}$, $X_{k(n)}=\max_{1\leq i\leq n}X_{ki}$, $k=1,2$.
Teniendo en cuenta que
\[d_k(r,s;\hat{\theta}_n)=\frac{1}{n} \sum_{i=1}^n \phi_{krs}(\boldsymbol{X}_i; \hat{\theta}_n),\quad k=1,2,\]
con
\begin{equation}\label{psi_1rs-2rs}
\begin{array}{l}
\phi_{1rs}(x;{\theta})=(r+1)I{\{x_1=r+1,x_2=s\}}-(\theta_1-\theta_3) I{\{x_1=r,x_2=s\}}\\[.3 cm]
\hspace{85mm}-\theta_3I{\{x_1=r,x_2=s-1\}},\\[.5 cm]
\phi_{2rs}(x;{\theta})=(s+1)I{\{x_1=r,x_2=s+1\}}-(\theta_2-\theta_3) I{\{x_1=r,x_2=s\}}\\[.3 cm]
\hspace{85mm}-\theta_3I{\{x_1=r-1,x_2=s\}},
\end{array}
\end{equation}
donde $x=(x_1,x_2)$, entonces el estad\'istico $W_n(\hat{\theta}_n)$ puede ser expresado como sigue:
\[W_n(\hat{\theta}_n)=\frac{1}{n^2}\sum_{i,j=1}^n h(\boldsymbol{X}_i,\boldsymbol{X}_j; \hat{\theta}_n),\]
donde
\begin{equation}\label{definicion-h}
\begin{array}{l}
h(x,y; {\theta})=h_1(x,y; {\theta})+h_2(x,y; {\theta}),\\[.35 cm]
h_k(x,y; {\theta})=\displaystyle\sum_{r \geq 0}\sum_{s \geq 0} \phi_{krs}(x; {\theta}) \phi_{krs}(y; {\theta}),\quad k=1,2,
\end{array}
\end{equation}
con $x=(x_1,x_2)$ e $y=(y_1,y_2)$.\\
Un test razonable para contrastar $H_0$ deber\'ia rechazar la hip\'otesis nula para valores grandes de $W_n(\hat{\theta}_n)$. Ahora, para determinar qu\'e son los valores grandes, debemos calcular la distribuci\'on nula de $W_n(\hat{\theta}_n)$ o al menos una aproximaci\'on de ella.
Como la distribuci\'on nula de $W_n(\hat{\theta}_n)$ es desconocida, trataremos de estimarla empleando el modo cl\'asico, esto es, aproximaremos la distribuci\'on nula mediante la distribuci\'on asint\'otica nula. En la siguiente secci\'on estudiaremos esta situaci\'on.
\section[Aproximaci\'on de la distribuci\'on nula de $W_n(\hat{\theta}_n)$]{Aproximaci\'on de la distribuci\'on nula de $\boldsymbol{W_n(\hat{\theta}_n)}$}\label{Distribucion-nula-Wn}
\subsection{Distribuci\'on asint\'otica nula}\label{Distribucion-asintotica-nula-Wn}
El siguiente resultado proporciona la distribuci\'on asint\'otica nula de $W_n(\hat{\theta}_n)$.
\begin{teorema}\label{TeoConvDebil-Wn}
Sean $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ v.a. iid de $\boldsymbol{X}=(X_{1},X_{2})\sim PB(\theta_{1},\theta_{2},\theta_{3})$. Supongamos que se cumple el Supuesto \ref{hat(theta)-theta}. Entonces
\[nW_n(\hat{\theta}_n)=\frac{1}{n}\sum_{i,j=1}^n h_d(\boldsymbol{X}_i,\boldsymbol{X}_j; \theta)+\rho_n,\]
donde $P_{{\theta}}(|\rho_n|>\varepsilon)\to 0$, $\forall \varepsilon>0$,
\begin{equation}\label{nucleo_de_Wn_aproximado}
h_d(x,y;\theta)=h(x,y;\theta)+\boldsymbol{\ell}(x;\theta) S \boldsymbol{\ell}(y;\theta)^{\top},
\end{equation}
con $h(x,y;\theta)$ definido en (\ref{definicion-h}), $x=(x_1,x_2)$, $y=(y_1,y_2)$, $S=\sum_{r,s\geq 0}A_{rs}$ y $A_{rs}$ es la matriz sim\'etrica dada por
\begin{equation}\label{Matriz-Ars}
A_{rs}=\left(
\begin{array}{ccc}
a^2 & 0 & a(b-a) \\
0 & a^2 & a(c-a) \\
a(b-a) & a(c-a) & (b-a)^2+(c-a)^2
\end{array}
\right),
\end{equation}
con $\,a=P(r,s;\theta),\ b=P(r,s-1;\theta),\ c=P(r-1,s;\theta)$. Adem\'as
\[nW_n(\hat{\theta}_n)\ \mathop{\longrightarrow} \limits^{\!L}\ \sum_{j\geq 1}\lambda_j\chi^2_{1j},\]
donde $\chi^2_{11},\chi^2_{12},\ldots$ son v.a. independientes $\chi^2$ con 1 grado de libertad y el conjunto $\{\lambda_j\}$ son los autovalores no nulos del operador $C(\theta)$ definido sobre el espacio de funciones $\{\tau:\mathbb{N}_0^2\to \mathbb{R}, \text{tal que} \ E_{\theta}\!\left\{\tau^2(\boldsymbol{X})\right\}<\infty,\forall \theta\in\Theta\}$, como sigue
\begin{equation}\label{Operador_C_para_Wn}
C(\theta) \tau(x)= E_{\theta}\{h_d(x,\boldsymbol{Y};\theta) \tau(\boldsymbol{Y})\}.
\end{equation}
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt} Por definici\'on
\[W_n(\hat{\theta}_n)=\frac{1}{n^2}\sum_{i,j=1}^n h(\boldsymbol{X}_i,\boldsymbol{X}_j; \hat{\theta}_n).\]
Por desarrollo en serie de Taylor
\begin{align}
&W_n(\hat{\theta}_n)\notag\\[.2 cm]
&=\!\frac{1}{n^2}\!\sum_{i,j=1}^n\! \left\{\!h(\boldsymbol{X}_i,\boldsymbol{X}_j; \theta)\!+Q^{(1)}(\boldsymbol{X}_i,\boldsymbol{X}_j;\theta) (\hat{\theta}_n\!-\theta)^{\top}\!\!+\frac{1}{2}(\hat{\theta}_n\!-\theta) Q^{(2)}(\boldsymbol{X}_i,\boldsymbol{X}_j;\theta) (\hat{\theta}_n\!-\theta)^{\top}\!\right\}\notag\\[.2 cm]
&\hspace{4mm}\left.+\frac{1}{3!\,n^2}\sum_{i,j=1}^n\, \sum_{k_1,k_2,k_3=1}^3 \! \frac{\partial^3} {\partial \vartheta_{k_1}\partial \vartheta_{k_2}\partial \vartheta_{k_3}}h(\boldsymbol{X}_i,\boldsymbol{X}_j; \vartheta)\right|_{\vartheta=\tilde{\theta}} \!\! (\hat{\theta}_{k_1n}\!-\theta_{k_1})(\hat{\theta}_{k_2n}\!-\theta_{k_2}) (\hat{\theta}_{k_3n}\!-\theta_{k_3}),\label{Wn-aprox}
\end{align}
donde $Q^{(k)}(x,y;\theta)$ representan las $k-$\'esimas derivadas de $h(x,y;\vartheta)$ respecto de $\vartheta$ evaluadas en $\theta$, $k=1,2$, y $\tilde{\theta}=\alpha \hat{\theta}_n+(1-\alpha)\theta$, para alg\'un $\,0<\alpha<1$.
Las derivadas de $h(x,y;\theta)$ est\'an dadas por
\begin{align}
\frac{\partial}{\partial \vartheta_{1}}h(x,y;\theta)&=-\sum_{r,s\geq 0}\Bigl[ I{\{x_1=r,x_2=s\}}\phi_{1rs}(y;\theta)+I{\{y_1=r,y_2=s\}} \phi_{1rs}(x;\theta)\Bigr],\notag\\[.13 cm]
\frac{\partial}{\partial \vartheta_{2}}h(x,y;\theta)&=-\sum_{r,s\geq 0}\Bigl[ I{\{x_1=r,x_2=s\}}\phi_{2rs}(y;\theta)+I{\{y_1=r,y_2=s\}}\phi_{2rs}(x;\theta) \Bigr],\notag\\[.13 cm]
\frac{\partial}{\partial \vartheta_{3}}h(x,y;\theta)&=\sum_{r,s\geq 0}\Bigl(I{\{x_1=r,x_2=s\}}-I{\{x_1=r,x_2=s-1\}}\Bigr) \phi_{1rs}(y;\theta)\notag\\ &\hspace{12mm}+\sum_{r,s\geq 0}\Bigl(I{\{y_1=r,y_2=s\}}-I{\{y_1=r,y_2=s-1\}}\Bigr) \phi_{1rs}(x;\theta)\notag\\
&\hspace{12mm}+\sum_{r,s\geq 0}\Bigl(I{\{x_1=r,x_2=s\}}-I{\{x_1=r-1,x_2=s\}}\Bigr) \phi_{2rs}(y;\theta)\notag\\ &\hspace{12mm}+\sum_{r,s\geq 0}\Bigl(I{\{y_1=r,y_2=s\}}-I{\{y_1=r-1,y_2=s\}}\Bigr) \phi_{2rs}(x;\theta),\notag\\[.1 cm]
\frac{\partial^2}{\partial \vartheta_{1}^2}h(x,y;\theta)&=2\sum_{r,s\geq 0} I{\{x_1=r,x_2=s\}}I{\{y_1=r,y_2=s\}},\notag\\
\frac{\partial^2}{\partial \vartheta_{1}\partial\vartheta_{2}}h(x,y;\theta)&=0,\notag\\[.13 cm]
\frac{\partial^2}{\partial \vartheta_{1}\partial\vartheta_{3}}h(x,y;\theta)&=-\!\sum_{r,s\geq 0} \Bigl[I{\{x_1=r,x_2=s\}}\bigl(I{\{y_1=r,y_2=s\}}-I{\{y_1=r,y_2=s-1\}}\bigr) \Bigr.\notag\\
&\hspace{10mm}\Bigl.+ I\{y_1=r,y_2=s\}\bigl(I{\{x_1=r,x_2=s\}}-I{\{x_1=r,x_2=s-1\}}\bigr) \Bigr],\notag\\[.1 cm]
\frac{\partial^2}{\partial \vartheta_{2}^2}h(x,y;\theta)&=2\sum_{r,s\geq 0} I{\{x_1=r,x_2=s\}}I{\{y_1=r,y_2=s\}},\notag\\
\frac{\partial^2}{\partial \vartheta_{2}\partial\vartheta_{3}}h(x,y;\theta)&=-\!\sum_{r,s\geq 0} \Bigl[I{\{x_1=r,x_2=s\}}(I{\{y_1=r,y_2=s\}}-I{\{y_1=r-1,y_2=s\}}) \Bigr.\notag\\
&\hspace{10mm}\Bigl. + I{\{y_1=r,y_2=s\}} \bigl(I{\{x_1=r,x_2=s\}}-I{\{x_1=r-1,x_2=s\}}\bigr)\Bigr],\notag\\[.1 cm]
\frac{\partial^2}{\partial \vartheta_{3}^2}h(x,y;\theta)&=2\sum_{r,s\geq 0}\Bigl[\bigl(I{\{x_1=r,x_2=s\}}-I{\{x_1=r,x_2=s-1\}}\bigr)\Bigr. \notag\\ &\hspace{17mm}\Bigl.\times\bigl(I{\{y_1=r,y_2=s\}}-I{\{y_1=r,y_2=s-1\}}\bigr) \Bigr] \notag\\
&\hspace{12mm} +2\sum_{r,s\geq 0}\Bigl[\bigl(I{\{x_1=r,x_2=s\}}- I{\{x_1=r-1,x_2=s\}}\bigr)\Bigr. \notag\\ &\hspace{29mm}\Bigl. \times\bigl(I{\{y_1=r,y_2=s\}}-I{\{y_1=r-1,y_2=s\}}\bigr)\Bigr].\notag
\end{align}
Adem\'as,
\[\frac{\partial^3} {\partial \vartheta_{i}\partial \vartheta_{j}\partial \vartheta_{k}}h(x,y;\theta)=0,\ 1\leq i,j,k\leq 3.\]
De lo anterior se obtiene que
\[\begin{array}{ll}
E_{\theta}\left\{Q^{(1)}(\boldsymbol{X}_1,\boldsymbol{X}_1; {\theta})\right\}&= 2(\theta_1-\theta_3,\theta_2-\theta_3,4\theta_3-\theta_1-\theta_2)<\infty,\\[.2 cm]
E_{\theta}\left\{Q^{(1)}(\boldsymbol{X}_1,\boldsymbol{X}_2; {\theta})\right\}&=\mathbf{0}, \ (\text{vector nulo de } \mathbb{R}^3),\\[.2 cm]
E_{\theta}\left\{Q^{(2)}(\boldsymbol{X}_1,\boldsymbol{X}_1; {\theta})\right\}&= -2 \left(
\begin{array}{rrr}
-1 & 0 & 1 \\
0 & -1 & 1 \\
1 & 1 & -4
\end{array}
\right)<\infty,
\\[.2 cm]
E_{\theta}\left\{Q^{(2)}(\boldsymbol{X}_1,\boldsymbol{X}_2; {\theta})\right\}&= 2 \sum_{r,s\geq 0} A_{rs}<\infty,
\end{array}\]
donde $A_{rs}$ es la matriz definida en (\ref{Matriz-Ars}).
Puesto que
\[\frac{1}{n^2}\sum_{i,j=1}^n Q^{(k)}(\boldsymbol{X}_i,\boldsymbol{X}_j;\theta)=
\frac{1}{n^2}\sum_{i\neq j} Q^{(k)}(\boldsymbol{X}_i,\boldsymbol{X}_j; {\theta})+
\frac{1}{n^2}\sum_{i=1}^n Q^{(k)}(\boldsymbol{X}_i,\boldsymbol{X}_i; {\theta}),\ k=1,2,\]
y por la ley fuerte de los grandes n\'umeros para U-estad\'isticos (v\'ease por ejemplo Teorema 5.4 en Serfling (1980) \cite{Ser80}),
\[\frac{1}{n^2}\sum_{i\neq j} Q^{(k)}(\boldsymbol{X}_i,\boldsymbol{X}_j; {\theta})\mathop{\longrightarrow}\limits^{c.s.}
E_{\theta}\left\{Q^{(k)}(\boldsymbol{X}_1,\boldsymbol{X}_2; {\theta})\right\}<\infty,\ k=1,2,\]
y por la ley fuerte de los grandes n\'umeros,
\[\frac{1}{n}\sum_{i=1}^n Q^{(k)}(\boldsymbol{X}_i,\boldsymbol{X}_i; {\theta})\mathop{\longrightarrow}\limits^{c.s.}E_{\theta}\left\{Q^{(k)} (\boldsymbol{X}_1,\boldsymbol{X}_1; {\theta})\right\}<\infty,\ k=1,2.\]
As\'i, (\ref{Wn-aprox}) se puede escribir como
\[W_n(\hat{\theta}_n)=\frac{1}{n^2}\sum_{i,j=1}^n h(\boldsymbol{X}_i,\boldsymbol{X}_j; \theta)+\sqrt{n}(\hat{\theta}_n-\theta) \frac{1}{n}S \sqrt{n}(\hat{\theta}_n-\theta)^{\top}+o_{_P}(1),\]
donde $S$ es la matriz sim\'etrica dada por $S=\frac{1}{2}\,E_{\theta}\!\left\{Q^{(2)}(\boldsymbol{X}_1,\boldsymbol{X}_2; {\theta})\right\}$.\vskip .2 cm
Considerando ahora el Supuesto \ref{hat(theta)-theta}, se obtiene
\[W_n(\hat{\theta}_n)=\frac{1}{n^2}\sum_{i,j=1}^n h_d(\boldsymbol{X}_i,\boldsymbol{X}_j; \theta)+\varepsilon_n,\]
donde $h_d(x,y;\theta)$ es el definido en (\ref{nucleo_de_Wn_aproximado}) y $\varepsilon_n=o_{_P}(n^{-1})$.\vskip .2 cm
Notar que $h_d(x,y;\theta)=h_d(y,x;\theta)$, pues $\left\{\boldsymbol{\ell}(x;\theta) S\boldsymbol{\ell}(y;\theta)^{\top}\right\}^{\top}=\boldsymbol{\ell}(y;\theta) S\boldsymbol{\ell}(x;\theta)^{\top}\in \mathbb{R}$, y bajo $H_0$, resulta
\[E_{\theta}\!\left\{|h_d(\boldsymbol{X}_1,\boldsymbol{X}_1;\theta)|\right\} < \infty \quad \mbox{y} \quad E_{\theta}\!\left\{h_d(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)^2\right\}< \infty. \]
Adem\'as, por el Supuesto \ref{hat(theta)-theta} y el hecho que $E_{\theta}\!\left\{h(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)\right\}=0$, dan como resultado $E_{\theta}\{h_d(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)\}=0 $.
Por \'ultimo, como $h_d$ es degenerado, $E_{\theta}\!\left\{h_d(\boldsymbol{X}_1,\boldsymbol{X}_2;\theta)/ \boldsymbol{X}_1\right\}=0$, entonces, por el Teorema 6.4.1.B en Serfling (1980) \cite{Ser80},
\[\frac{1}{n}\sum_{i,j=1}^n h_d(\boldsymbol{X}_i,\boldsymbol{X}_j; \theta)
\mathop{\longrightarrow} \limits^{\!L}\ \sum_{j\geq 1}\lambda_j\,\chi^2_{1j},\]
con lo cual se consigue el resultado. $\square$\\
Los mismos comentarios hechos al final de la Subsecci\'on \ref{Distribucion-asintotica-nula} para $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ se pueden hacer para $W_n(\hat{\theta}_n)$, es decir, la distribuci\'on asint\'otica nula no proporciona una estimaci\'on \'util a la distribuci\'on nula de $W_n(\hat{\theta}_n)$. As\'i, en la siguiente secci\'on consideramos otra forma de aproximar la distribuci\'on nula del test estad\'istico, el m\'etodo bootstrap.
\subsection[Estimador bootstrap de la distribuci\'on nula de $W_n(\hat{\theta}_n)$]{Estimador bootstrap de la distribuci\'on nula de $\boldsymbol{W_n(\hat{\theta}_n)}$}\label{Estimador bootstrap}
Como mencionamos en la Secci\'on \ref{Aproximacion-bootstrap}, un modo alternativo de aproximar la distribuci\'on nula es mediante el m\'etodo bootstrap. Para ello, sean $\boldsymbol{X}_{1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid que toman valores en $\mathbb{N}_0^2$ tales que $\hat{\theta}_n= \hat{\theta}_n(\boldsymbol{X}_{\!1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}) \in \Theta$.
Sean $\boldsymbol{X}^*_{1}, \boldsymbol{X}^*_{2},\ldots, \boldsymbol{X}^*_{n}$ v.a. iid de una poblaci\'on que se distribuye seg\'un la ley $PB(\hat{\theta}_{1n},\hat{\theta}_{2n}, \hat{\theta}_{3n})$, dado $\boldsymbol{X}_{1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ y sea $W^*_{n}(\hat{\theta}^*_n)$ la versi\'on bootstrap de $W_{n}(\hat{\theta}_n)$ obtenida al reemplazar $\boldsymbol{X}_{\!1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ y
$\hat{\theta}_n\!=\! \hat{\theta}_n(\boldsymbol{X}_{\!1}, \boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n})$ por $\boldsymbol{X}^*_{\!1}, \boldsymbol{X}^*_{2},\ldots, \boldsymbol{X}^*_{n}$ y $\hat{\theta}^*_n= \hat{\theta}_n(\boldsymbol{X}^*_{1}, \boldsymbol{X}^*_{2},\ldots, \boldsymbol{X}^*_{n})$, respectivamente, en la expresi\'on de $W_{n}(\hat{\theta}_n)$.
Como en el Cap\'itulo \ref{Estadisticos-tipoCramer-von-Mises}, sea $P_*$ la ley de probabilidad condicional bootstrap, dado $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$.
Para mostrar que el m\'etodo bootstrap aproxima consistentemente a la distribuci\'on nula de $W_n(\hat{\theta}_n)$ nos ser\'an \'utiles las siguientes expresiones.
Observar que del segundo sumando de (\ref{nucleo_de_Wn_aproximado}) podemos escribir que
\begin{align}
\boldsymbol{\ell}(x;\theta)A_{rs}\,\boldsymbol{\ell}(y;\theta)^{\top}&= \boldsymbol{\ell}(x;\theta)
\left(\begin{array}{ccc}
a & 0 & 0 \\
0 & a & 0 \\
b-a & c-a & 0
\end{array}\right)
\left(\begin{array}{ccc}
a & 0 & b-a \\
0 & a & c-a \\
0 & 0 & 0
\end{array}\right)\boldsymbol{\ell}(y;\theta)^{\top}\notag\\
&=\left(\boldsymbol{\ell}(x;\theta) (a,0,b-a)^{\top},\boldsymbol{\ell}(x;\theta) (0,a,c-a)^{\top}\right)\left(
\begin{array}{l}
(a,0,b-a)\boldsymbol{\ell}(y;\theta)^{\top}\\[.1 cm]
(0,a,c-a)\boldsymbol{\ell}(y;\theta)^{\top}
\end{array}
\right)\notag\\[.2 cm]
&=\boldsymbol{\ell}(x;\theta)\, (a,0,b-a)^{\top} \boldsymbol{\ell}(y;\theta) \,(a,0,b-a)^{\top}\notag\\[.2 cm]
&\hspace{8mm}+\boldsymbol{\ell}(x;\theta)\,(0,a,c-a)^{\top} \boldsymbol{\ell}(y;\theta)\,(0,a,c-a)^{\top}.\notag
\end{align}
Sean
\begin{equation}\label{psi_3rs-4rs}
\phi_{3rs}(x;\theta)=\boldsymbol{\ell}(x;\theta) (a,0,b-a)^{\top}\quad\text{y}\quad \phi_{4rs}(x;\theta)=\boldsymbol{\ell}(x;\theta)(0,a,c-a)^{\top},
\end{equation}
donde $x=(x_1,x_2)$.
Con esta notaci\'on,
\begin{equation}\label{l-Ars-l}
\boldsymbol{\ell}(x;\theta)A_{rs}\,\boldsymbol{\ell}(y;\theta)^{\top}= \phi_{3rs}(x;\theta)\phi_{3rs}(y;\theta)+ \phi_{4rs}(x;\theta)\phi_{4rs}(y;\theta).
\end{equation}
Tambi\'en nos ser\'a de utilidad el siguiente resultado.
\begin{proposicion}\label{Nucleos-Wn}
Sean $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid de $\boldsymbol{X}=(X_{1},X_{2}) \in \mathbb{N}_0^2$. Supongamos que se verifica el Supuesto \ref{E(l*l)-l-cont} y que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un $\theta \in \Theta$. Entonces
\[\sup_{p,q,u,v\in \mathbb{N}_0} \left| K_n(p,q,u,v)-K(p,q,u,v)\right| \stackrel{c.s.}{\longrightarrow}0,\]
donde
\begin{align}
K_n(p,q,u,v)&=\left(E_*\{\phi_{ipq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{juv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\}\right), 1\leq i,j \leq 4.\notag\\[.2 cm]
K(p,q,u,v)&=\Bigl(E_{\theta}\!\left\{\phi_{ipq}(\boldsymbol{X}_1;\theta) \phi_{juv}(\boldsymbol{X}_1;\theta)\right\}\Bigr), 1\leq i,j \leq 4.\notag
\end{align}
con $\,\phi_{1rs}(\boldsymbol{X}_1;\vartheta)$, $\phi_{2rs}(\boldsymbol{X}_1;\vartheta)$ definidos en (\ref{psi_1rs-2rs}), y $\,\phi_{3rs}(\boldsymbol{X}_1;{\vartheta})$ y $\phi_{4rs}(\boldsymbol{X}_1;{\vartheta})$ definidos en (\ref{psi_3rs-4rs}).
\end{proposicion}
\nt {\bf Demostraci\'on} \hspace{2pt}
El Supuesto \ref{E(l*l)-l-cont} y el hecho de que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$ implican que
\begin{equation}\label{E*-Et}
\sup_{p,q,u,v \in \mathbb{N}_0} \left| E_*\{\phi_{ipq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{juv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\}- E_{\theta}\{\phi_{ipq}(\boldsymbol{X}_1;\theta) \phi_{juv}(\boldsymbol{X}_1;\theta)\}\right| \stackrel{c.s.}{\longrightarrow}0, \end{equation}
donde $1\leq i,j \leq 4$.
Para probar este resultado consideremos los siguientes casos:
\begin{itemize}
\item [(a)] $i,j$ tales que $1\leq i,j \leq 2$,
\item [(b)] $i,j$ tales que $3\leq i,j \leq 4$ y
\item [(c)] $i,j$ tales que $i=1,2,$ $j=3,4$.
\end{itemize}
Notar primero que por el Teorema del Valor Medio
\[P(r,s;\hat{\theta}_n)-P(r,s;\theta)=\nabla_{\vartheta}P(r,s;\tilde{\theta}) (\hat{\theta}_n-\theta)^{\top},\ \tilde{\theta}=\alpha \hat{\theta}_n+(1-\alpha)\theta, \ \text{para alg\'un} \ 0<\alpha <1,\]
donde $\nabla_{\vartheta}P(r,s;\tilde{\theta})$ denota el vector gradiente de $P(r,s;\vartheta)$ evaluado en $\tilde{\theta}$. Por las relaciones de recurrencia dadas en (\ref{form-rec-derDPB}), obtenemos que
\[\nabla_{\vartheta}P(r,s;\tilde{\theta})=(a-c,b-c,d-a-b+c),\]
donde $a=P(r-1,s;\tilde{\theta}), b=P(r,s-1;\tilde{\theta}), c=P(r,s;\tilde{\theta})$ y $d=P(r-1,s-1;\tilde{\theta})$.
Como $0\leq P(r,s;\tilde{\theta})\leq 1$ y puesto que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, resulta
\begin{equation}\label{P-o(1)}
\sup_{r,s \in \mathbb{N}_0} \left|P(r,s;\hat{\theta}_n)-P(r,s;\theta)\right|=o(1).
\end{equation}
\begin{itemize}
\item En el caso (a), una situaci\'on es la siguiente
\begin{align}
E_*&\{\phi_{1pq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{1uv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\}\notag\\[.18 cm]
&\hspace{5mm} =\left\{
\begin{array}{ll}
\begin{array}{l}
(p+1)^2P(p+1,q;\hat{\theta}_n)+\hat{\theta}_{3n}^2P(p,q-1;\hat{\theta}_n)\\[.27 cm]
\hspace{33mm}+(\hat{\theta}_{1n}-\hat{\theta}_{3n})^2 P(p,q;\hat{\theta}_n)
\end{array}
& \hbox{si } u=p, v=q,\\[.7 cm]
-(p+1)(\hat{\theta}_{1n}-\hat{\theta}_{3n}) P(p+1,q;\hat{\theta}_n), & \hbox{si } u=p+1, v=q, \\[.3 cm]
-(p+1)\hat{\theta}_{3n} P(p+1,q;\hat{\theta}_n), & \hbox{si } u=p+1, v=q+1, \\[.3 cm]
-p(\hat{\theta}_{1n}-\hat{\theta}_{3n}) P(p,q;\hat{\theta}_n), & \hbox{si } u=p-1, v=q, \\[.3 cm]
-p\hat{\theta}_{3n} P(p,q-1;\hat{\theta}_n), & \hbox{si } u=p-1, v=q-1,\\[.3 cm]
\hat{\theta}_{3n}(\hat{\theta}_{1n}-\hat{\theta}_{3n}) P(p,q;\hat{\theta}_n), & \hbox{si } u=p, v=q+1, \\[.3 cm]
\hat{\theta}_{3n}(\hat{\theta}_{1n}-\hat{\theta}_{3n}) P(p,q-1;\hat{\theta}_n), & \hbox{si } u=p, v=q-1,
\end{array}
\right.\notag
\end{align}
\nt y an\'alogo para $E_{\theta}\left\{\phi_{1pq}(\boldsymbol{X}_1;\theta) \phi_{1uv}(\boldsymbol{X}_1;\theta)\right\}$, con las modificaciones obvias.\vskip .25 cm
Al hacer la resta de las esperanzas anteriores nos encontramos con las siguientes tres situaciones diferentes
\begin{itemize}
\item [(a.1)] $t(\hat{\theta}_n) P(r,s;\hat{\theta}_n)-t(\theta)P(r,s;\theta)$,
\item [(a.2)] $r\!\left\{t(\hat{\theta}_n) P(r,s;\hat{\theta}_n)-t(\theta)P(r,s;\theta)\!\right\}$,
\item [(a.3)] $r^2\!\left\{P(r,s;\hat{\theta}_n)-P(r,s;\theta)\!\right\}$,
\end{itemize}
donde $t(\vartheta)=\vartheta^k_3$, $t(\vartheta)=(\vartheta_1-\vartheta_3)^k$, para $k=1,2$, o bien $t(\vartheta)=\vartheta_3(\vartheta_1-\vartheta_3)$.\vskip .3 cm
Para la situaci\'on (a.1), puesto que $t(\vartheta)$ es continua como funci\'on de $\vartheta\in \Theta_0$ y como $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, entonces $t(\hat{\theta}_n)=t(\theta)+o(1)$, por lo tanto
\begin{equation}\label{t*P}
t(\hat{\theta}_n) P(r,s;\hat{\theta}_n)-t(\theta)P(r,s;\theta)= t(\theta)\{P(r,s;\hat{\theta}_n)-P(r,s;\theta)\}+o(1)P(r,s;\hat{\theta}_n),
\end{equation}
como $t(\theta)< \infty$, de (\ref{P-o(1)}) resulta
\[\sup_{r,s\in \mathbb{N}_0}\left|t(\hat{\theta}_n) P(r,s;\hat{\theta}_n)-t(\theta)P(r,s;\theta)\right|=o(1).\]
Para la situaci\'on (a.3), por desarrollo en serie de Taylor
\[r^2\!\left\{P(r,s;\hat{\theta}_n)-P(r,s;\theta)\right\}= \sum_{i=1}^3r^2\frac{\partial}{\partial \theta_i}P(r,s;\tilde{\theta})
(\hat{\theta}_{in}-{\theta}_i),\]
donde $\tilde{\theta}=\alpha \hat{\theta}_n+(1-\alpha)\theta$, para alg\'un $0<\alpha <1$.\vskip .2 cm
Veamos el primer sumando y an\'alogo el resto:
Como $\frac{\partial}{\partial \theta_1}P(r,s;\tilde{\theta})= P(r-1,s;\tilde{\theta})-P(r,s;\tilde{\theta})$ y $E_{\vartheta}(X^2_1)< \infty, \forall \vartheta\in \Theta_0$, entonces
\begin{align}
r^2\frac{\partial}{\partial \theta_1}P(r,s;\tilde{\theta})
(\hat{\theta}_{1n}-{\theta}_1)&=(\hat{\theta}_{1n}-{\theta}_1)r^2 \{P(r-1,s;\tilde{\theta})-P(r,s;\tilde{\theta})\}\notag\\
&\leq |\hat{\theta}_{1n}-{\theta}_1|\sum_{r,s\geq 0}r^2\{P(r-1,s;\tilde{\theta})+P(r,s;\tilde{\theta})\}\notag\\[.1 cm]
&\leq|\hat{\theta}_{1n}-{\theta}_1|\left[E_{\tilde{\theta}}\{(X_1+1)^2\}+ E_{\tilde{\theta}}(X_1^2)\right]=o(1).\notag
\end{align}
De donde
\begin{align}
&r^2\!\left\{P(r,s;\hat{\theta}_n)-P(r,s;\theta)\right\}\notag\\[.1 cm]
&\hspace{5mm}\leq \!\left\{|\hat{\theta}_{1n}\!-{\theta}_1|+2|\hat{\theta}_{3n}\! -{\theta}_3|\right\} \left[E_{\tilde{\theta}}\left\{(X_1+1)^2\right\}+ E_{\tilde{\theta}}\!\left(X_1^2\right)\right] +2|\hat{\theta}_{2n}\!-{\theta}_2|E_{\tilde{\theta}}\!\left(X_1^2\right).\notag
\end{align}
Por lo tanto
\[\sup_{r,s\in \mathbb{N}_0}\left|r^2\!\left\{P(r,s;\hat{\theta}_n)-P(r,s;\theta)\right\}\right|=o(1).\]
Para la situaci\'on (a.2) y considerando (\ref{t*P}), obtenemos
\begin{align}
&r\left\{t(\hat{\theta}_n) P(r,s;\hat{\theta}_n)-t(\theta)P(r,s;\theta)\right\} \notag \\
&\hspace{48mm}=r\, t(\theta) \left\{P(r,s;\hat{\theta}_n)-P(r,s;\theta)\right\}+rP(r,s;\hat{\theta}_n)\, o(1).\notag
\end{align}
Puesto que $t(\theta)< \infty$, $E_{\vartheta}(X_1)< \infty, \forall \vartheta\in \Theta_0$ y siguiendo pasos similares a los dados en la situaci\'on (a.3) anterior se consigue que
\[\sup_{r,s\in \mathbb{N}_0}\left|r\left\{t(\hat{\theta}_n) P(r,s;\hat{\theta}_n)-t(\theta)P(r,s;\theta)\!\right\}\right|=o(1).\]
\item Para el caso (b), tenemos, por ejemplo
\begin{align}
E_*\!\left\{\phi_{3pq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{3uv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\!\right\}&= E_*\!\left\{\boldsymbol{\ell} (\boldsymbol{X}_1^*;\hat{\theta}_n) B(p,q;\hat{\theta}_n)^{\top} \boldsymbol{\ell}(\boldsymbol{X}_1^*;\hat{\theta}_n) B(u,v;\hat{\theta}_n)^{\top}\!\right\} \notag\\[.1 cm] &= B(p,q;\hat{\theta}_n)J(\hat{\theta}_n) B(u,v;\hat{\theta}_n)^{\top},\notag\\[.2 cm]
E_{\theta}\!\left\{\phi_{3pq}(\boldsymbol{X}_1;\theta) \phi_{3uv}(\boldsymbol{X}_1;\theta)\right\}&=B(p,q;\theta)J(\theta) B(u,v;\theta)^{\top}, \notag
\end{align}
donde
\begin{equation}\label{B(r,s)}
B(r,s;\vartheta)=(P(r,s;\vartheta),0, P(r,s-1;\vartheta)-P(r,s;\vartheta)).
\end{equation}
Como, por el Supuesto \ref{E(l*l)-l-cont}, $J(\vartheta)$ es continua como funci\'on de $\vartheta\in \Theta_0$ y $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, entonces $J(\hat{\theta}_n)=J(\theta)+o(1)$. Adem\'as, por el Teorema del Valor Medio
\begin{equation}\label{Bhat(t)-Bt}
B(r,s;\hat{\theta}_n)^{\top}-B(r,s;\theta)^{\top}= grad(P) (\hat{\theta}_n-\theta)^{\top}=\mathbf{o}(1),
\end{equation}
pues
\[grad(P)=\left(\begin{array}{c}
\nabla_{\theta}P(r,s;\tilde{\theta}) \\
\mathbf{0} \\
\nabla_{\theta}P(r,s-1;\tilde{\theta})- \nabla_{\theta}P(r,s;\tilde{\theta})
\end{array}\right)
< \infty,\]
donde $\tilde{\theta}=\alpha \hat{\theta}_n+(1-\alpha)\theta$, para alg\'un $0<\alpha <1$.
Por tanto
\begin{align}
&\left|E_*\!\left\{\phi_{3pq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{3uv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\!\right\}- E_{\theta}\!\left\{\phi_{3pq}(\boldsymbol{X}_1;\theta) \phi_{3uv}(\boldsymbol{X}_1;\theta)\right\}\right|\notag\\[.2 cm]
&\hspace{42mm}\leq \left|\left\{B(p,q;\hat{\theta}_n)-B(p,q;\theta)\right\} J(\theta)B(u,v;\hat{\theta}_n)^{\top}\right| \notag\\[.2 cm]
&\hspace{47mm}+ \left|B(p,q;\theta)J(\theta)\!\left\{B(u,v;\hat{\theta}_n)^{\top}- B(u,v;\theta)^{\top}\right\}\right|+o(1).\notag
\end{align}
Puesto que $J(\vartheta)<\infty$ y $B(r,s;\vartheta)<\infty$, $\forall \vartheta\in\Theta_0, \forall r,s\in \mathbb{N}_0$, entonces se logra
\[\sup_{p,q,u,v\in \mathbb{N}_0}\left|E_*\!\left\{\phi_{3pq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{3uv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\!\right\}- E_{\theta}\!\left\{\phi_{3pq}(\boldsymbol{X}_1;\theta) \phi_{3uv}(\boldsymbol{X}_1;\theta)\right\}\right| \stackrel{c.s.}{\longrightarrow}0.\]
Aplicando el mismo procedimiento se demuestra (\ref{E*-Et}) para $3\leq i,j\leq 4$.
\item \vskip .4 cm En el caso (c), una situaci\'on es
\[E_*\!\left\{\phi_{1pq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{3uv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\right\}- E_{\theta}\{\phi_{1pq}(\boldsymbol{X}_1;\theta) \phi_{3uv}(\boldsymbol{X}_1;\theta)\}=A_1+A_2+A_3,\]
donde
\begin{align}
A_1&=(p+1)P(p+1,q;\hat{\theta}_n) \boldsymbol{\ell}(p+1,q;\hat{\theta}_n)B(u,v;\hat{\theta}_n)^{\top}\notag\\[.1 cm] &\hspace{50 mm}-(p+1)P(p+1,q;\theta) \boldsymbol{\ell}(p+1,q;\theta)B(u,v;\theta)^{\top},\notag\\[.2 cm
A_2&=(\theta_1-\theta_3)P(p,q;\theta) \boldsymbol{\ell}(p,q;\theta)B(u,v;\theta)^{\top}\notag\\[.1 cm]
&\hspace{50 mm} -(\hat{\theta}_{1n}-\hat{\theta}_{3n})P(p,q;\hat{\theta}_n) \boldsymbol{\ell}(p,q;\hat{\theta}_n)B(u,v;\hat{\theta}_n)^{\top},\notag \\[.2 cm
A_3&=\theta_3 P(p,q-1;\theta) \boldsymbol{\ell}(p,q-1;\theta)B(u,v;\theta)^{\top}\notag\\[.1 cm]
&\hspace{50 mm} -\hat{\theta}_{3n} P(p,q-1;\hat{\theta}_n) \boldsymbol{\ell}(p,q-1;\hat{\theta}_n)B(u,v;\hat{\theta}_n)^{\top} \nota
\end{align}
y $B$ es definido como en (\ref{B(r,s)}).\vskip .2 cm
Veamos primero que $\displaystyle\sup_{r,s,u,v\in \mathbb{N}_0}\left|A_3\right|=o(1)$. De manera an\'aloga se demuestra que $\displaystyle\sup_{r,s,u,v\in \mathbb{N}_0}\left|A_2\right|=o(1)$.
\begin{align}
\hat{\theta}_{3n} P(r,s;\hat{\theta}_n) \boldsymbol{\ell}&(r,s;\hat{\theta}_n)B(u,v;\hat{\theta}_n)^{\top}
-\theta_3 P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)B(u,v;\theta)^{\top}\notag\\[.15 cm]
&=(\hat{\theta}_{3n}-\theta_3)P(r,s;\hat{\theta}_n) \boldsymbol{\ell}(r,s;\hat{\theta}_n) B(u,v;\hat{\theta}_n)^{\top}\notag\\[.15 cm]
&\ \ \ +\theta_3P(r,s;\theta)\boldsymbol{\ell}(r,s;\theta) \left\{B(u,v;\hat{\theta}_n)^{\top}- B(u,v;\theta)^{\top}\right\}\notag\\[.15 cm]
&\ \ \ +\theta_3\left\{P(r,s;\hat{\theta}_n) \boldsymbol{\ell}(r,s;\hat{\theta}_n)-P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)\right\} B(u,v;\hat{\theta}_n)^{\top}.\label{Tercer-suma}
\end{align}
Como $J(\vartheta)$ es finita $\forall \vartheta \in \Theta_0$, esto implica que $|P(r,s;\vartheta)\boldsymbol{\ell}_i(r,s;\vartheta)|\leq L$, para cierta constante positiva $L>0$, $\forall r,s\in \mathbb{N}_0$,
$\forall \vartheta \in \Theta_0$, $i=1,2,3$.\vskip .2 cm
Puesto que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, de (\ref{Bhat(t)-Bt}) resulta
\begin{equation}\label{Bi_hat(theta)-Bi(theta)}
\sup_{u,v\in \mathbb{N}_0}\left|B_i(u,v;\hat{\theta}_n)-B_i(u,v;\theta)\right|=o(1),\ i=1,2,3.
\end{equation}
Por lo tanto
\[\sup_{r,s,u,v\in \mathbb{N}_0}\left|(\hat{\theta}_{3n}-\theta_3)P(r,s;\hat{\theta}_n) \boldsymbol{\ell}(r,s;\hat{\theta}_n) B(u,v;\hat{\theta}_n)^{\top}\right|=o(1),\]
\[\sup_{r,s,u,v\in \mathbb{N}_0}\left|\theta_3P(r,s;\theta)\boldsymbol{\ell}(r,s;\theta) \left\{B(u,v;\hat{\theta}_n)^{\top}- B(u,v;\theta)^{\top}\right\}\right|=o(1).\]
Para el tercer sumando de (\ref{Tercer-suma}), el Lema \ref{P(hat(t))*l(hat(t))-P(t)*l(t)} junto con que $\hat{\theta}_n\in\Theta_0$ y el hecho que $|B_i(u,v;\hat{\theta}_n)|\leq 1$, $i=1,2,3$, implican
\[\sup_{r,s,u,v\in \mathbb{N}_0}\left|\theta_3\left\{P(r,s;\hat{\theta}_n) \boldsymbol{\ell}(r,s;\hat{\theta}_n)-P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)\right\} B(u,v;\hat{\theta}_n)^{\top}
\right|=o(1).\]
Veamos ahora que $\ \sup_{r,s,u,v\in \mathbb{N}_0}\left|A_1\right|=o(1)$. Para ello, notar que
\begin{align}
rP(r,s;\hat{\theta}_n) \boldsymbol{\ell}(r,s;\hat{\theta}_n) & B(u,v;\hat{\theta}_n)^{\top}-rP(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)B(u,v;\theta)^{\top}\notag\\
&=r\left\{P(r,s;\hat{\theta}_n)\boldsymbol{\ell}(r,s;\hat{\theta}_n)- P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta) \right\}B(u,v;\hat{\theta}_n)^{\top}\notag\\
&\hspace{18mm}+r P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)\left\{B(u,v;\hat{\theta}_n)^{\top}\!- B(u,v;\theta)^{\top} \right\}.\notag
\end{align}
Usando el Lema \ref{P(hat(t))*l(hat(t))-P(t)*l(t)}(a) junto con que $\hat{\theta}_n\in\Theta_0$, obtenemos
\begin{align}
&\sup_{r,s\in\mathbb{N}_0}r \left|\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)-\boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|\notag\\
&\hspace{17mm}\leq\sum_{r,s \geq 0} r\left|\boldsymbol{\ell}_i(r,s;\hat{\theta}_n) P(r,s;\hat{\theta}_n)- \boldsymbol{\ell}_i(r,s;\theta) P(r,s;\theta)\right|=o(1),\ i=1,2,3.\notag
\end{align}
De esta \'ultima relaci\'on y puesto que $|B_i(u,v;\hat{\theta}_n)|\leq 1$, $i=1,2,3$, nos resulta
\[\sup_{r,s,u,v\in \mathbb{N}_0}\left|r\left\{P(r,s;\hat{\theta}_n) \boldsymbol{\ell}(r,s;\hat{\theta}_n)-P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)\right\} B(u,v;\hat{\theta}_n)^{\top}
\right|=o(1).\]
Por la desigualdad de Cauchy-Schwarz y puesto que $J(\theta)< \infty$, entonces
\begin{align}
\sup_{r,s\in\mathbb{N}_0}r|\boldsymbol{\ell}_i(r,s;\theta)| P(r,s;\theta) &\leq \sum_{r,s\geq 0}r|\boldsymbol{\ell}_i(r,s;\theta)| P(r,s;\theta)\notag\\[.1 cm]
&\leq \left\{E_{\theta}\!\left(X_1^2\right)\right\}^{\!1/2} \left\{ E_{\theta}\!\left(\|\boldsymbol{\ell}(\boldsymbol{X}_1;\theta)\|^2\right)\right\}^{\!\!1/2}< \infty,\ i=1,2,3.\notag
\end{align}
Esta \'ultima desigualdad junto con (\ref{Bi_hat(theta)-Bi(theta)}) implican \[\sup_{r,s,u,v\in \mathbb{N}_0}\left|r P(r,s;\theta) \boldsymbol{\ell}(r,s;\theta)\left\{B(u,v;\hat{\theta}_n)^{\top}- B(u,v;\theta)^{\top} \right\}
\right|=o(1).\]
As\'i, concluimos que $\ \sup_{r,s,u,v\in \mathbb{N}_0}\left|A_1\right|=o(1)$.
\end{itemize}
Los casos (a), (b) y (c) muestran que se verifica (\ref{E*-Et}) para $1\leq i,j\leq 4$, con lo cual se demuestra la proposici\'on. $\square$
Ahora probaremos que el m\'etodo bootstrap estima consistentemente la distribuci\'on nula de $W_n(\hat{\theta}_n)$.
\begin{teorema}\label{Cons-boot-Wn}
Sean $\boldsymbol{X}_{1},\boldsymbol{X}_{2},\ldots, \boldsymbol{X}_{n}$ vectores aleatorios iid de $\boldsymbol{X}=(X_{1},X_{2}) \in \mathbb{N}_0^2$. Supongamos que se verifica el Supuesto \ref{E(l*l)-l-cont} y que $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un $\theta \in \Theta$. Entonces
\[\sup_{x\in\mathbb{R}}\!\left|P_*\left(nW^*_{n}(\hat{\theta}^*_n) \leq x\right)-P_{\theta}\!\left(nW_{n}(\hat{\theta}_n)\leq x\right)\right|\ \mathop{\longrightarrow}\limits^{c.s.}\ 0.\]
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt} Por definici\'on
\[W^*_n(\hat{\theta}^*_n)=\frac{1}{n^2}\sum_{i,j=1}^n h^*(\boldsymbol{X}^*_i,\boldsymbol{X}^*_j;\hat{\theta}^*_n).\]
Siguiendo pasos similares a los dados en la demostraci\'on del Teorema \ref{TeoConvDebil-Wn} se puede ver que
\[W^*_n(\hat{\theta}^*_n)=W^*_{1n}(\hat{\theta}^*_n),\]
donde
\[W^*_{1n}(\hat{\theta}^*_n)=\frac{1}{n^2}\sum_{i,j=1}^n h^*_d(\boldsymbol{X}^*_i,\boldsymbol{X}^*_j;\hat{\theta}_n)+\xi^*,\]
con $\xi^*=o_{_{P_*}}(n^{-1})$ c.s. y $h^*_d$ es definido como $h_d$ en (\ref{nucleo_de_Wn_aproximado}) con $\boldsymbol{X}_i$, $\boldsymbol{X}_j$ y $\theta$ reemplazados por $\boldsymbol{X}^*_i$,$\boldsymbol{X}^*_j$ y $\hat{\theta}_n$, respectivamente.
De (\ref{definicion-h}) y (\ref{l-Ars-l}), $h^*_d$ se puede escribir como
\begin{align}
h^*_d(\boldsymbol{X}^*_i,\boldsymbol{X}^*_j;\hat{\theta}_n)&= h^*_1(\boldsymbol{X}^*_i,\boldsymbol{X}^*_j;\hat{\theta}_n)+ h^*_2(\boldsymbol{X}^*_i,\boldsymbol{X}^*_j;\hat{\theta}_n)+ \sum_{r,s\geq 0}\boldsymbol{\ell}(\boldsymbol{X}^*_i;\hat{\theta}_n) A^*_{rs}\boldsymbol{\ell}(\boldsymbol{X}^*_j;\hat{\theta}_n)^{\top}\notag\\[.1 cm]
&=\sum_{1\leq k\leq 4} \sum_{r,s\geq 0} \phi_{krs}(\boldsymbol{X}^*_i;\hat{\theta}_n) \phi_{krs}(\boldsymbol{X}^*_j;\hat{\theta}_n),\notag
\end{align}
con lo cual
\[W^*_{1n}(\hat{\theta}^*_n)=\sum_{i,j=1}^n \frac{1}{n^2} \sum_{1\leq k\leq 4} \sum_{r,s\geq 0} \phi_{krs}(\boldsymbol{X}^*_i;\hat{\theta}_n) \phi_{krs}(\boldsymbol{X}^*_j;\hat{\theta}_n).\]
La demostraci\'on se completar\'a verificando las condiciones (i)-(iii) del Lema \ref{kundu} para $W^*_{1n}(\hat{\theta}^*_n)$. Para ello, consideremos el siguiente espacio de Hilbert separable
\[\mathcal{H}_2=\left\{z=z(r,s)=(z_{1rs},z_{2rs},z_{3rs},z_{4rs})_{r,s\geq 0}\ \ \text{tal que}\ \ \|z\|_{_{\mathcal{H}_2}}^2=\sum_{1\leq k\leq 4}\,\sum_{r,s\geq 0} \!z_{krs}^2<\infty\right\}.\]
con producto escalar $\,\langle x,y\rangle_{_{\!\mathcal{H}_2}}=\displaystyle\sum_{1\leq k\leq 4} \sum_{r,s\geq 0}x_{krs}y_{krs}<\infty$.
En este espacio de Hilbert se tiene que
\[nW^*_{1n}(\hat{\theta}^*_n)=\sum_{1\leq k\leq 4}\sum_{r,s\geq 0}\left\{\frac{1}{\sqrt{n}}\sum_{1\leq i\leq n} \phi_{krs}(\boldsymbol{X}^*_i;\hat{\theta}_n)\right\}^2 \equiv\|\Phi_n^*\|_{_{\mathcal{H}_2}}^2,\]
donde $\Phi_n^*$ est\'a dado por
\[\Phi_n^*(r,s)=\sum_{i=1}^n \Phi_{ni}^*(r,s),\] \[\Phi_{ni}^*(r,s)=\frac{1}{\sqrt{n}}\Phi(\boldsymbol{X}^*_i;\hat{\theta}_n;r,s), 1\leq i\leq n,\]
\begin{equation}\label{vector-Phi}
\Phi(\boldsymbol{X}^*_i;\hat{\theta}_n;r,s)= \left(\phi_{1rs}(\boldsymbol{X}^*_i;\hat{\theta}_n), \phi_{2rs}(\boldsymbol{X}^*_i;\hat{\theta}_n), \phi_{3rs}(\boldsymbol{X}^*_i;\hat{\theta}_n), \phi_{4rs}(\boldsymbol{X}^*_i;\hat{\theta}_n)\right), 1\leq i\leq n,
\end{equation}
\nt con $\phi_{1rs}(x;{\theta})$ y $\phi_{2rs}(x;{\theta})$ definidos en (\ref{psi_1rs-2rs}), $\phi_{3rs}(x;{\theta})$ y $\phi_{4rs}(x;{\theta})$ definidos en (\ref{psi_3rs-4rs}), donde $x$ y $\theta$ son reemplazados por $\boldsymbol{X}^*_i$ y $\hat{\theta}_n$, respectivamente.\vskip .3 cm
Por el Supuesto \ref{E(l*l)-l-cont} y por (\ref{form-rec-DPB}), $\Phi_{ni}^*(r,s)$ tiene medias nulas y momentos segundos finitos, $1\leq i \leq n$, pues $\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$.
Sean $(p,q),(u,v)\in \mathbb{N}_0^2$ y consideremos el n\'ucleo de covarianza $K_n(p,q,u,v)=E_*\left\{\Phi_n^*(p,q)^\top \Phi_n^*(u,v)\right\}$, luego
\[K_n(p,q,u,v)=\left(E_*\{\phi_{ipq}(\boldsymbol{X}_1^*;\hat{\theta}_n) \phi_{juv}(\boldsymbol{X}_1^*;\hat{\theta}_n)\}\right), 1\leq i,j \leq 4.\]
Adem\'as, sea $K(p,q,u,v)=E_{\theta}\left\{\Phi(\boldsymbol{X}_1;\theta;p,q)^{\top} \Phi(\boldsymbol{X}_1;\theta;u,v)\right\}$ donde $\Phi$ es como el definido en (\ref{vector-Phi}), es decir,
\[K(p,q,u,v)=\Bigl(E_{\theta}\!\left\{\phi_{ipq}(\boldsymbol{X}_1;\theta) \phi_{juv}(\boldsymbol{X}_1;\theta)\right\}\Bigr), 1\leq i,j \leq 4.\]
Sea $\mathcal{Z}$ la sucesi\'on Gaussiana centrada cuyo operador de covarianza $C$ est\'a caracterizado por
\begin{align}
\langle C x,y\rangle_{_{\!\mathcal{H}_2}}=cov\left(\langle \mathcal{Z}, x\rangle_{_{\!\mathcal{H}_2}}, \langle \mathcal{Z}, y\rangle_{_{\!\mathcal{H}_2}}\right)&=E_{\theta}\left\{\langle \mathcal{Z}, x\rangle_{_{\!\mathcal{H}_2}} \langle \mathcal{Z}, y\rangle_{_{\!\mathcal{H}_2}}\right\}\notag\\[.2 cm]
&=\sum_{p,q,u,v\geq 0} x(p,q)K(p,q,u,v)y(u,v)^{\top}.\label{Oper-cov-C}
\end{align}
Por el Lema \ref{TCL-Hilbert}, $(\Phi_{1n},\Phi_{2n},\Phi_{3n},\Phi_{4n})\stackrel{L}{\longrightarrow} \mathcal{Z}$ en $\mathcal{H}_2$, cuando los datos son iid provenientes de $\boldsymbol{X}\sim PB(\theta)$, donde \[\Phi_{kn}(r,s)=\frac{1}{\sqrt{n}}\sum_{1\leq i\leq n}\phi_{krs}(\boldsymbol{X}_i;\theta), 1\leq k\leq 4.\]
Para nuestro caso una base ortonormal conveniente en $\mathcal{H}_2$ es la dada por el conjunto $\left\{e^k_{r_0s_0}: r_0,s_0\geq 0, k=1,2,3,4\right\}$, donde
\[\begin{array}{c}
e^1_{r_0s_0}(r,s)=\bigl(I{\{r=r_0,s=s_0\}},0,0,0\bigr)_{r,s\geq 0},\\[.35 cm]
e^2_{r_0s_0}(r,s)=\bigl(0,I{\{r=r_0,s=s_0\}},0,0\bigr)_{r,s\geq 0},\\[.35 cm]
e^3_{r_0s_0}(r,s)=\bigl(0,0,I{\{r=r_0,s=s_0\}},0\bigr)_{r,s\geq 0},\\[.35 cm]
e^4_{r_0s_0}(r,s)=\bigl(0,0,0,I{\{r=r_0,s=s_0\}}\bigr)_{r,s\geq 0}.
\end{array}\]
Ahora mostraremos que se verifican las condiciones (i)-(iii) del Lema \ref{kundu}.
\begin{itemize}
\item [(i)] Sea $C_n$ el operador de covarianza de $\Phi_n^*$, esto es, $\forall x,y \in \mathcal{H}_2$ \begin{align} \langle C_n x,y\rangle_{_{\!\mathcal{H}_2}}&=cov(\langle \Phi_n^*, x\rangle_{_{\!\mathcal{H}_2}}, \langle \Phi_n^*, y\rangle_{_{\!\mathcal{H}_2}})=E_{*}\!\left\{\langle \Phi_n^*, x\rangle_{_{\!\mathcal{H}_2}} \langle \Phi_n^*, y\rangle_{_{\!\mathcal{H}_2}}\right\}\notag\\[.2 cm] &=\sum_{p,q,u,v\geq 0} x(p,q)E_*\!\left\{\Phi_n^*(p,q)^\top \Phi_n^*(u,v)\right\}y(u,v)^{\top}\notag\\[.2 cm] &=\sum_{p,q,u,v\geq 0} x(p,q)K_n(p,q,u,v)y(u,v)^{\top}.\label{def-Cn} \end{align}
De (\ref{Oper-cov-C}), (\ref{def-Cn}) y de la Proposici\'on \ref{Nucleos-Wn}
\begin{align} \lim_{n\to \infty}\langle C_n e^k_{p_0q_0}, e^l_{u_0v_0}\rangle_{_{\!\mathcal{H}_2}} &=\lim_{n\to \infty} \sum_{p,q,u,v\geq 0} e^k_{p_0q_0}(p,q)K_n(p,q,u,v) e^l_{u_0v_0}(u,v)^{\top}\notag\\[.2 cm] &=\sum_{p,q,u,v\geq 0} e^k_{p_0q_0}(p,q)K(p,q,u,v)e^l_{u_0v_0}(u,v)^{\top}\notag\\[.2 cm]
&=\langle C e^k_{p_0q_0}, e^l_{u_0v_0}\rangle_{_{\!\mathcal{H}_2}}= a^{kl}_{p_0q_0u_0v_0}.\notag
\end{align}
\item [(ii)] De (\ref{Oper-cov-C}), (\ref{def-Cn}), de la Proposici\'on \ref{Nucleos-Wn} y del \'item (i), se obtiene
\begin{align} \lim_{n\to \infty} \sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} &\langle C_{n} e^k_{p_0q_0},e^k_{p_0q_0}\rangle_{_{\!\mathcal{H}_2}}\notag\\[.2 cm]
&= \lim_{n\to \infty}\sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} \sum_{p,q,p,q\geq 0} e^k_{p_0q_0}(p,q)K_n(p,q,p,q) e^l_{p_0q_0}(p,q)^{\top}\notag\\[.2 cm]
&= \sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} \sum_{p,q,p,q\geq 0} e^k_{p_0q_0}(p,q)K(p,q,p,q) e^l_{p_0q_0}(p,q)^{\top}\notag\\[.2 cm]
&=\sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} \langle C e^k_{p_0q_0},e^k_{p_0q_0}\rangle_{_{\!\mathcal{H}_2}}\notag\\[.2 cm]
&= \sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} a^{kk}_{p_0q_0p_0q_0}<\infty,\notag \end{align}
pues, de la primera ecuaci\'on en (\ref{Oper-cov-C}) y la igualdad de Parseval
\begin{align}
\sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} a^{kk}_{p_0q_0p_0q_0}&=\sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} \langle C e^k_{p_0q_0},e^k_{p_0q_0}\rangle_{_{\!\mathcal{H}_2}}\notag\\[.2 cm]
&=\sum_{1\leq k\leq 4}\sum_{p_0,q_0\geq 0} E_{\theta}\!\left\{\langle \mathcal{Z},e^k_{p_0q_0}\rangle_{_{\!\mathcal{H}_2}}^2\right\}= E_{\theta}\!\left\{\|\mathcal{Z}\|_{_{\mathcal{H}_2}}^{\,2}\right\}< \infty.\notag
\end{align}
\item [(iii)] Sean $r_0,s_0\in \mathbb{N}_0$, fijos,
\begin{equation}\label{PI-Psi-e1}
\langle \Phi^*_{ni},e^k_{r_0s_0}\rangle_{_{\mathcal{H}_2}}^2=\frac{1}{n} \phi^2_{kr_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n), k=1,2,3,4.
\end{equation}
Para $k=1$, tenemos
\[|\langle \Phi^*_{ni},e^1_{r_0s_0}\rangle_{_{\mathcal{H}_2}}|=\frac{1}{\sqrt{n}} |\phi_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n)|\leq r_0+1+|\hat{\theta}_{1n}- \hat{\theta}_{3n}|+\hat{\theta}_{3n}.\]
Por tanto, dado $\varepsilon >0$, existir\'a $n_0=n_0(r_0,\varepsilon)\in \mathbb{N}_0$ tal que
\[\frac{1}{\sqrt{n}} |\phi_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n)|\leq \frac{1}{\sqrt{n}}(r_0+1+|\hat{\theta}_{1n}- \hat{\theta}_{3n}|+\hat{\theta}_{3n})\leq \varepsilon, \forall n> n_0.\]
De donde,
\begin{equation}\label{I-varepsilon}
I{\left\{\frac{1}{\sqrt{n}} |\phi_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n)|> \varepsilon\right\}}=
\left\{
\begin{array}{ll}
1, & \hbox{si } n\leq n_0,\\
0, & \hbox{si } n>n_0.
\end{array}
\right.
\end{equation}
Se tiene que
\[L_n(\varepsilon,e^1_{r_0s_0})=\frac{1}{n}\sum_{i=1}^n E_*\!\left[\phi^2_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n) I{\left\{\frac{1}{\sqrt{n}} |\phi_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n)|> \varepsilon\right\}}\right].\]
Adem\'as, de (\ref{psi_1rs-2rs})
\begin{align}
\phi^2_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n)&=(r_0+1)^2 \, I{\{X_{1i}^*=r_0+1,X_{2i}^*=s_0\}}\notag\\[.2 cm]
&\hspace{-4mm}+(\hat{\theta}_{1n}-\hat{\theta}_{3n})^2\, I{\{X_{1i}^*=r_0,X_{2i}^*=s_0\}}+\hat{\theta}_{3n}^2 I{\{X_{1i}^*=r_0,X_{2i}^*=s_0-1\}}.\notag
\end{align}
Con esta \'ultima igualdad y (\ref{I-varepsilon})
\begin{align}
&\sum_{i=1}^nE_*\!\left[\phi^2_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n) I{\left\{\frac{1}{\sqrt{n}} |\phi_{1r_0s_0}(\boldsymbol{X}_i^*;\hat{\theta}_n)|> \varepsilon\right\}}\right]\notag\\
&=\sum_{i=1}^n\sum_{r,s\geq 0}\left[(r_0+1)^2 \, I{\{r=r_0+1,s=s_0\}}+(\hat{\theta}_{1n}-\hat{\theta}_{3n})^2 \, I{\{r=r_0,s=s_0\}}\right.\notag\\
&\hspace{22mm}\left.+\hat{\theta}_{3n}^2 \, I{\{r=r_0,s=s_0-1\}}\right]I{\left\{\frac{1}{\sqrt{n}} |\phi_{1r_0s_0}(r,s;\hat{\theta}_n)|> \varepsilon\right\}}\,P(r,s;\hat{\theta}_n)\notag\\
&\leq\sum_{i=1}^{n_0}\sum_{r,s\geq 0}\left[(r_0+1)^2 \, I{\{r=r_0+1,s=s_0\}}+(\hat{\theta}_{1n}-\hat{\theta}_{3n})^2 \, I{\{r=r_0,s=s_0\}}\right.\notag\\
&\hspace{22mm}\left.+\hat{\theta}_{3n}^2 \, I{\{r=r_0,s=s_0-1\}}\right]\,P(r,s;\hat{\theta}_n)\notag\\
& =\sum_{i=1}^{n_0}\!\left\{(r_0+1)^2P(r_0+1,s_0;\hat{\theta}_n) +(\hat{\theta}_{1n}\!-\hat{\theta}_{3n})^2P(r_0,s_0;\hat{\theta}_n)+ \hat{\theta}_{3n}^2P(r_0,s_0\!-1;\hat{\theta}_n)\!
\right\}\notag\\
& \leq n_0\left\{(r_0+1)^2 +(\hat{\theta}_{1n}-\hat{\theta}_{3n})^2+ \hat{\theta}_{3n}^2
\right\}.\notag
\end{align}
Por tanto
\[L_n(\varepsilon,e^1_{r_0s_0})\leq \frac{n_0}{n}\!\left\{(r_0+1)^2 +(\hat{\theta}_{1n}-\hat{\theta}_{3n})^2+ \hat{\theta}_{3n}^2
\right\}.\]
Puesto que $r_0$ y $\varepsilon$ son fijos y $n_0=n_0(r_0,\varepsilon)$, entonces esta \'ultima desigualdad converge a cero. As\'i
\[\lim_{n\to \infty}L_n(\varepsilon,e^1_{r_0s_0})=0.\]
An\'alogamente se logra que $\lim_{n\to \infty}L_n(\varepsilon,e^2_{r_0s_0})=0$.\vskip .2 cm
De (\ref{PI-Psi-e1}), para $k=3$ y de (\ref{psi_3rs-4rs})
\[\langle \Phi^*_{ni},e^3_{r_0s_0}\rangle_{_{\mathcal{H}_2}}=\frac{1}{\sqrt{n}} \boldsymbol{\ell}(\boldsymbol{X}_i^*;\hat{\theta}_n) d(r_0,s_0;\hat{\theta}_n)^{\top},\]
donde $d(r,s;\hat{\theta}_n)=(P(r,s;\hat{\theta}_n),0, P(r,s-1;\hat{\theta}_n)-P(r,s;\hat{\theta}_n))$ y por tanto
\begin{align}
&L_n(\varepsilon,e^3_{r_0s_0})\notag\\
&\hspace{4mm}=\frac{1}{n}\sum_{i=1}^n E_*\!\left[\left\{ \boldsymbol{\ell}(\boldsymbol{X}_i^*;\hat{\theta}_n) d(r_0,s_0;\hat{\theta}_n)^{\top}\right\}^2 I{\left\{\frac{1}{\sqrt{n}} \left| \boldsymbol{\ell}(\boldsymbol{X}_i^*;\hat{\theta}_n) d(r_0,s_0;\hat{\theta}_n)^{\top}\right|> \varepsilon\right\}}\right].\notag
\end{align}
Ahora, como
\begin{equation}\label{(l*d)^2}
\left|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)d(r,s;\hat{\theta}_n)^{\top}\right| \leq \|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\| \|d(r,s;\hat{\theta}_n)\| \leq \sqrt{2}\|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\|,
\end{equation}
entonces
\[\varepsilon < \frac{1}{\sqrt{n}} | \boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)d(r_0,s_0;\hat{\theta}_n)^{\top}| \leq \frac{\sqrt{2}}{\sqrt{n}}\|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\|,\]
por tanto
\begin{align}
I{\left\{\frac{1}{\sqrt{n}} \left| \boldsymbol{\ell}(\boldsymbol{X}_i^*;\hat{\theta}_n) d(r_0,s_0;\hat{\theta}_n)^{\top}\right|> \varepsilon\right\}} &\leq I{\left\{\frac{\sqrt{2}}{\sqrt{n}}\,\|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\|> \varepsilon\right\}}\notag\\[.2 cm]
&=I{\left\{\|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\| >\gamma\right\}},\notag
\end{align}
donde $\gamma=\gamma(n)=\varepsilon \frac{\sqrt{n}}{\sqrt{2}}$ es tal que $\gamma\to \infty$ cuando $n\to \infty$.\vskip .2 cm
Adem\'as, de (\ref{(l*d)^2}), se tiene que $\,\left( \boldsymbol{\ell}(\boldsymbol{X}_i^*;\hat{\theta}_n) d(r_0,s_0;\hat{\theta}_n)^{\top}\right)^2\!\!\leq 2 \|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\|^2$ y como $\boldsymbol{X}_1^*,\boldsymbol{X}_2^*,\ldots,\boldsymbol{X}_n^*$ son vectores aleatorios iid de $\boldsymbol{X}^*\sim PB(\hat{\theta}_n)$, entonces
\begin{align}
L_n(\varepsilon,e^3_{r_0s_0})&\leq\frac{2}{n}\sum_{i=1}^n E_*\!\left(\|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\|^2 I{ \left\{\|\boldsymbol{\ell} (\boldsymbol{X}_i^*;\hat{\theta}_n)\| >\gamma\right\}}\right)\notag\\[.1 cm]
&\leq 2 \sup_{\hat{\theta}_n\in \Theta_0}E_{\hat{\theta}_n}\!\!\left(\|\boldsymbol{\ell} (\boldsymbol{X}^*;\hat{\theta}_n)\|^2 I{ \left\{\|\boldsymbol{\ell} (\boldsymbol{X}^*;\hat{\theta}_n)\| >\gamma\right\}}\right),\notag
\end{align}
y como $\gamma\to \infty$ cuando $n\to \infty$, entonces, por el Supuesto \ref{E(l*l)-l-cont} (1) se concluye que
\[\lim_{n\to \infty}L_n(\varepsilon,e^3_{r_0s_0})=0.\]
De la misma forma se prueba que $\lim_{n\to \infty}L_n(\varepsilon,e^4_{r_0s_0})=0$.
\end{itemize}
As\'i, por el Lema \ref{kundu}, tenemos que
$\Phi_n^*\mathop{\longrightarrow} \limits^{L} \mathcal{Z}$, c.s., en $\mathcal{H}_2$. Como ya hab\'iamos mencionado, $\mathcal{Z}$ es tambi\'en el l\'imite d\'ebil de $(\Phi_{1n},\Phi_{2n},\Phi_{3n},\Phi_{4n})$ cuando los datos son iid provenientes de $\boldsymbol{X}\sim PB(\theta)$. Finalmente, el resultado se obtiene por el Teorema de la aplicaci\'on continua. $\square$\\
Similarmente a lo expuesto en la parte final de la Secci\'on \ref{Aproximacion-distribucion-nula}, sea
$$w^*_{n,\alpha}=\inf\!\left\{x:P_*(W^*_n(\hat{\theta}^*_n)\geq x)\leq \alpha\right\}$$
el percentil superior $\alpha$ de la distribuci\'on bootstrap de $W_n(\hat{\theta}_n)$.
Del Teorema \ref{Cons-boot-Wn}, la funci\'on test
$$\Psi^*_W=\left\{
\begin{array}{ll}
1, & \text{si}\ W_n(\hat{\theta}_n)\geq w^*_{n,\alpha}, \\[.15 cm]
0, & \text{en caso contrario},
\end{array}
\right.$$
o equivalentemente, el test que rechaza $H_0$ cuando
$$p_{_W}^*=P_*\!\left(W^*_n(\hat{\theta}^{\,*}_n) \geq W_{obs}\right)\leq \alpha,$$
es asint\'oticamente correcto en el sentido que, cuando $H_0$ es cierta
$$\lim_{n\to \infty} P_{\theta}\!\left(\Psi^*_W=1\right)=\alpha,$$
donde $W_{obs}$ es el valor observado del test estad\'istico $W_n (\hat{\theta}_n)$.
\begin{observacion}\label{Boot-Wn} Para aproximar $w^*_{n,\alpha}$ se sigue el mismo procedimiento presentado al final de la Secci\'on \ref{Aproximacion-distribucion-nula}, con los cambios obvios.
\end{observacion}
\newpage
\section{Alternativas} \label{Alternativas-Wn}
En esta secci\'on estudiaremos el comportamiento del test $\Psi^*_W$ bajo alternativas fijas y locales.
\subsection{Alternativas fijas}\label{Alternativas-fijas-Wn}
Como una de nuestras metas es obtener tests de bondad de ajuste que sean consistentes, lo pr\'oximo que haremos es estudiar este t\'opico para el nuevo test que proponemos. Con este fin, primero obtendremos el l\'imite c.s. de $W_n(\hat{\theta}_n)$.
\begin{teorema}\label{Convergencia-de-W_n(theta-hat)}
Sean $\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$ vectores aleatorios iid de $\boldsymbol{X}=(X_1,X_2)\in\mathbb{N}_0^2$,
con $E(X_k^2)<\infty$, $k=1,2$. Sea $p(r,s)=P(X_1=r,X_2=s)$. Si
$\hat{\theta}_n \mathop{\longrightarrow}\limits^{c.s.} \theta$, para alg\'un $\theta\in \mathbb{R}^3$, entonces
\[W_n(\hat{\theta}_n)\ \mathop{\longrightarrow}\limits^{c.s.}\sum_{r,s \geq 0}\left\{a_1(r,s;\theta)^2+a_2(r,s;\theta)^2\right\}=\nu(P;\theta),\]
donde
\begin{align}
a_1(r,s;\theta)&=(r+1)p(r+1,s)-(\theta_1-\theta_3)p(r,s)-\theta_3p(r,s-1), \notag\\[.2 cm]
a_2(r,s;\theta)&=(s+1)p(r,s+1)-(\theta_2-\theta_3)p(r,s)-\theta_3p(r-1,s).\notag
\end{align}
\end{teorema}
\begin{observacion}\label{eta(P;theta) geq 0}
Notar que $\nu(P;\theta)\geq 0$. Adem\'as, de la Proposici\'on \ref{Soluc-sistema-de-dos-EDP} obtenemos que, $\nu(P;\theta)=0$ s\'i y s\'olo si $H_0$ es cierta.
\end{observacion}
\nt {\bf Demostraci\'on del Teorema \ref{Convergencia-de-W_n(theta-hat)}} \hspace{2pt}
Se tiene que
\begin{equation}\label{daux1}
d_1(r,s;\hat{\theta}_n)=d_1(r,s;{\theta})-(\hat{\theta}_{1n}-\theta_1)p_n(r,s)+ (\hat{\theta}_{3n}-\theta_3)\{p_n(r,s)-p_n(r,s-1)\}.
\end{equation}
Tambi\'en se tiene que
\[\sum_{r,s\geq 0}d_1(r,s;{\theta})^2=\frac{1}{n^2}\sum_{i\neq j} h_1(\boldsymbol{X}_i,\boldsymbol{X}_j; {\theta})+
\frac{1}{n^2}\sum_{i=1}^n h_1(\boldsymbol{X}_i,\boldsymbol{X}_i; {\theta}).\]
Por la ley fuerte de los grandes n\'umeros para U-estad\'isticos (v\'ease por ejemplo Teorema 5.4 en Serfling (1980) \cite{Ser80}),
\[\frac{1}{n^2}\sum_{i\neq j} h_1(\boldsymbol{X}_i,\boldsymbol{X}_j; {\theta})\mathop{\longrightarrow}\limits^{c.s.}
E\{h_1(\boldsymbol{X}_1,\boldsymbol{X}_2; {\theta})\}=\sum_{r,s \geq 0}a_1(r,s;\theta)^2.\]
Como $E(X_1^2)<\infty$, por la ley fuerte de los grandes n\'umeros,
\[\frac{1}{n}\sum_{i=1}^n h_1(\boldsymbol{X}_i,\boldsymbol{X}_i; {\theta})\mathop{\longrightarrow}\limits^{c.s.} E\{h_1(\boldsymbol{X}_1,\boldsymbol{X}_1; {\theta})\}=E\left\{
\sum_{r,s \geq 0}\phi_{1rs}(\boldsymbol{X}_1; {\theta})^2\right\}<\infty.\]
Por tanto,
\begin{equation}\label{daux2}
\sum_{r,s\geq 0}d_1(r,s;{\theta})^2\mathop{\longrightarrow}\limits^{c.s.}
\sum_{r,s \geq 0}a_1(r,s;\theta)^2.
\end{equation}
Como $p_n(r,s)^2\leq p_n(r,s)$, $\forall r,s \geq 0$, y $\sum_{r,s\geq 0}p_n(r,s)=1$, se tiene que
\begin{equation}\label{daux3}
(\hat{\theta}_{1n}-\theta_1)^2 \sum_{r,s\geq 0} p_n(r,s)^2 \leq (\hat{\theta}_{1n}-\theta_1)^2 =o(1),
\end{equation}
y an\'alogamente,
\begin{equation}\label{daux4}
(\hat{\theta}_{3n}-\theta_3)^2\sum_{r,s\geq 0}\{p_n(r,s)-p_n(r,s-1)\}^2=o(1).
\end{equation}
De (\ref{daux1})--(\ref{daux4}),
\begin{equation}\label{daux5}
\sum_{r,s\geq 0}d_1(r,s;\hat{\theta}_n)^2\mathop{\longrightarrow}\limits^{c.s.}
\sum_{r,s \geq 0}a_1(r,s;\theta)^2.
\end{equation}
Procediendo similarmente
\begin{equation}\label{daux6}
\sum_{r,s\geq 0}d_2(r,s;\hat{\theta}_n)^2\mathop{\longrightarrow}\limits^{c.s.}
\sum_{r,s \geq 0}a_2(r,s;\theta)^2.
\end{equation}
Finalmente, el resultado se tiene de (\ref{daux5}) y (\ref{daux6}). $\square$\\
Como una consecuencia de los Teoremas \ref{TeoConvDebil-Wn}, \ref{Cons-boot-Wn} y \ref{Convergencia-de-W_n(theta-hat)}, y la Observaci\'on \ref{eta(P;theta) geq 0}, el siguiente resultado da la consistencia del test $\Psi^*_W$.
\begin{corolario}\label{Altern-fijas-Wn}
Sean $\boldsymbol{X}_1, \boldsymbol{X}_2,\ldots,\boldsymbol{X}_n$ vectores aleatorios iid de $\boldsymbol{X}\in\mathbb{N}_0^2$. Supongamos que se cumplen las hip\'otesis de los Teoremas \ref{TeoConvDebil-Wn} y \ref{Cons-boot-Wn}.
Si $H_0$ no es cierta, entonces $P(\Psi^*_W=1) \to 1.$
\end{corolario}
\subsection{Alternativas contiguas} \label{Alternativas-contiguas-Wn}
Como se estableci\'o en el Teorema \ref{TeoConvDebil-Wn}, cuando $H_0$ es cierta, $nW_n(\hat{\theta}_n)$ converge d\'ebilmente a una combinaci\'on lineal de v.a. $\chi^2$ independientes con 1 grado de libertad, donde los pesos son los autovalores del operator $C(\theta)$ dado en (\ref{Operador_C_para_Wn}).
Sea $\{\phi_j\}$ el conjunto ortonormal de autofunciones correspondientes a los autovalores $\{\lambda_j\}$ de $C(\theta)$.
El siguiente Teorema da la ley l\'imite del test estad\'istico bajo las alternativas $P_n(x_1,x_2)$ en (\ref{P(n)=P(theta)+b}).
\begin{teorema}\label{Altern-contiguas-Wn}
Sea $\boldsymbol{X}_{n,1}, \boldsymbol{X}_{n,2},\ldots,\boldsymbol{X}_{n,n}$ un arreglo triangular de vectores aleatorios bivariantes que son independientes por filas y que toman valores en $\mathbb{N}_0^2$, con funci\'on de probabilidad dada por $P_n(x_1,x_2)$ definida en (\ref{P(n)=P(theta)+b}). Supongamos que se cumplen los Supuestos \ref{hat(theta)-theta} y \ref{b(x,y)}. Entonces
\[nW_n(\hat{\theta}_n)\ \mathop{\longrightarrow} \limits^{L} \ \sum_{j=1}^{\infty} \lambda_j\left(Z_j+c_j\right)^2,\]
donde $c_j=\mathop{\sum}\limits_{x_1,x_2}b(x_1,x_2)\phi_j(x_1,x_2)$ y $Z_1,Z_2,\ldots$ son variables normales est\'andar independientes.
\end{teorema}
\nt {\bf Demostraci\'on} \hspace{2pt} Sigue los mismos pasos que la demostraci\'on del Teorema \ref{Altern-contiguas-S_n,w}, por lo tanto la omitiremos. $\square$\vskip .2 cm
Del Teorema \ref{Altern-contiguas-Wn}, concluimos que el test $\Psi^*_W$ es capaz de detectar alternativas como las establecidas en (\ref{P(n)=P(theta)+b}), que convergen a la DPB a una raz\'on de $n^{-1/2}$.
\newpage
\[\begin{array}{c}\\ \\ \end{array}\]
\newpage
\chapter{Resultados num\'ericos} \label{Resultados-numericos}
Las propiedades estudiadas en las Secciones \ref{Aproximacion-distribucion-nula}, \ref{Alternativas_Rnw-Srw}, \ref{Distribucion-nula-Wn} y \ref{Alternativas-Wn} son asint\'oticas, esto es, tales propiedades describen el comportamiento de los tests propuestos para muestras de tama\~no grande. Para estudiar la bondad de la aproximaci\'on bootstrap para muestras de tama\~no finito, as\'i como tambi\'en para comparar la potencia de los tests propuestos tanto entre ellos como con otros tests, hemos llevado a cabo un experimento de simulaci\'on. En esta secci\'on describimos dicho experimento y damos un resumen de los resultados que hemos obtenido.
Todos los c\'alculos computacionales realizados en este texto se llevaron a cabo mediante el uso de programas escritos en el lenguaje R \cite{R}.
Como competidores de los tests estad\'isticos propuestos, hemos considerado los tests dados por Crockett (1979) \cite{Cro79} (denotado por $T$), Loukas y Kemp (1986) \cite{LoKe86} (denotado por $I_B$), y Rayner y Best (1995) \cite{RaBe95} (denotado por $NI_B$). La expresi\'on de cada test estad\'istico, as\'i como la regiones cr\'iticas correspondientes ya las presentamos en el Cap\'itulo \ref{Resultados previos y definiciones}.
Para calcular $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ es necesario dar una forma expl\'icita a la funci\'on de peso $w$. En el caso univariante, Baringhaus et al. (2000) \cite{BaGuHe00} consideraron la funci\'on de peso $w(u)=u^a$, con $a\geq 0$. Varias extensiones son posibles. Aqu\'i hemos tenido en cuenta la siguiente
\begin{equation}\label{funcion-peso-explicita}
w(u;a_1,a_2)=u_1^{a_1}u_2^{a_2}.
\end{equation}
Observar que las \'unicas restricciones que hemos impuesto a la funci\'on de peso son que $w$ sea positiva casi en todas partes en $[0,1]^2$ y la establecida en (\ref{int-funcion-peso}). La funci\'on $w(u;a_1,a_2)$ dada en (\ref{funcion-peso-explicita}) cumple dichas condiciones siempre que $a_i>-1$, $i=1,2$.
En el Cap\'itulo \ref{Expresiones-matematicas} damos las expresiones matem\'aticas de los estad\'isticos $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ para la funci\'on de peso dada en (\ref{funcion-peso-explicita}). Tambi\'en damos la expresi\'on para el estad\'istico $W_n(\hat{\theta}_n)$.
Adem\'as, mostramos los resultados que obtuvimos al aplicar los tests propuestos a dos conjuntos de datos reales.
\section{Datos simulados}\label{Simulacion-datos}
Para estudiar la bondad de la aproximaci\'on bootstrap en muestras de tama\~no finito tanto para los tests propuestos, como para las aproximaciones a la distribuci\'on nula de los tests estad\'isticos $T$, $I_B$ y $NI_B$, que est\'an basados en sus distribuciones asint\'oticas nulas, generamos muestras de tama\~no $n=30(20)70$ de la distribuci\'on $PB(\theta_1,\theta_2,\theta_3)$, con $\theta_1=\theta_2=$1.0 y $\theta_3$ tal que el coeficiente de correlaci\'on, $\rho=\theta_3/\sqrt{\theta_1\,\theta_2}$ sea igual a 0.25, 0.50 y 0.75, con el fin de examinar la bondad de las aproximaciones para datos con una correlaci\'on baja, media y alta, respectivamente.
Para estimar el par\'ametro $\theta$ empleamos el m\'etodo de m\'axima verosimilitud como se describe en Kocherlakota y Kocherlakota (1992, pp. 103--105) \cite{KoKo92}. Luego, aproximamos los $p-$valores bootstrap, $p^*$, de los tests propuestos, para ello, en el caso de $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ usamos la funci\'on de peso dada en (\ref{funcion-peso-explicita}) para $(a_1,a_2)\in \{(0,0), (1,0)\}$ y generamos $B=500$ muestras bootstrap. Tambi\'en, calculamos los $p-$valores (asint\'oticos) asociados a los tests estad\'isticos $T$, $I_B$ y $NI_B$.
El procedimiento anterior lo repetimos $1000$ veces y calculamos la fracci\'on de los $p-$valores estimados que resultaron ser menores o iguales que 0.05 y 0.10, que son las estimaciones de las probabilidades del error tipo I para $\alpha=$ 0.05 y 0.10 (en las tablas, denotamos esto como f05 y f10), respectivamente.
Si las aproximaciones consideradas fuesen exactas, entonces los $p-$valores calculados deber\'ian ser una muestra aleatoria de una distribuci\'on uniforme en el intervalo $(0,1)$. Por lo tanto, para medir la bondad de las aproximaciones consideradas, calculamos el $p-$valor del test estad\'istico de uniformidad de Kolmogorov-Smirnov ($KS$) para cada conjunto de 1000 $p-$valores obtenido para cada uno de los tests estad\'isticos.
Repetimos el experimento anterior para $\theta_1=$ 1.5, $\theta_2=$ 1.0 y $\theta_3$ de manera que el coeficiente de correlaci\'on,
$\rho=\theta_3/\sqrt{\theta_1\,\theta_2}$, fuera aproximadamente igual a 0.25, 0.50 y 0.75. En este caso, ya que $\theta_1\neq \theta_2$,
adem\'as de $(a_1,a_2)\in \{(0,0), (1,0)\}$ consideramos
$(a_1,a_2)=(0,1)$ para $R_{n,a}$ y $S_{n,a}$, con el fin de examinar el efecto de dar un peso diferente a cada uno de los componentes cuando estos tienen distintas esperanzas.
Las Tablas \ref{Resultados-igual-E-1-simulacion-error-I}-- \ref{Resultados-diferente-E-3-simulacion-error-I} resumen los resultados obtenidos. En dichas tablas, denotamos por $R_{n,a}$ y $S_{n,a}$ a los tests estad\'isticos $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$, respectivamente, cuando la funci\'on de peso $w$ toma la forma dada por (\ref{funcion-peso-explicita}), para alg\'un $a=(a_1,a_2)$.
Observando los valores dados en las tablas, concluimos que la aproximaci\'on asint\'otica de los $p-$valores de los tests estad\'isticos $T$, $I_B$ y $NI_B$ no da resultados satisfactorios para los casos tratados. Por el contrario, el bootstrap proporciona una aproximaci\'on precisa a la distribuci\'on nula de $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ en todos los casos tratados. La aproximaci\'on bootstrap a la distribuci\'on nula de $W_n(\hat{\theta}_n)$ es adecuada para $n\geq 50$.
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la probabilidad del error tipo I. Caso $E(X_1)=E(X_2)$, $\rho=$ 0.25} \label{Resultados-igual-E-1-simulacion-error-I}
\begin{center}
{\normalsize
\begin{tabular}{|l|cc|c|cc|c|cc|c|}
\cline{2-10}\multicolumn{1}{c|}{}
&\multicolumn{3}{c|}{$n=30$}& \multicolumn{3}{c|}{$n=50$} & \multicolumn{3}{c|}{$n=70$}\\ \hline
\begin{tabular}{l}
$\theta=$ (1.0,1.0,0.25) \\
$\rho=$ 0.25
\end{tabular}
& f05 & f10 & $KS$ & f05 & f10 & $KS$ & f05 & f10 & $KS$\\
\hline
$R_{n,(0,0)}$ & 0.035 & 0.091 & 0.818621 & 0.047 & 0.100 & 0.459543 & 0.041 & 0.083 & 0.329116\\
$S_{n,(0,0)}$ & 0.046 & 0.090 & 0.902243 & 0.049 & 0.097 & 0.818621 & 0.041 & 0.089 & 0.769894\\
\hline
$R_{n,(1,0)}$ & 0.035 & 0.093 & 0.257432 & 0.050 & 0.100 & 0.769894 & 0.042 & 0.086 & 0.413150\\
$S_{n,(1,0)}$ & 0.039 & 0.094 & 0.665399 & 0.048 & 0.098 & 0.995881 & 0.042 & 0.093 & 0.559560\\
\hline
$W_{n}$ & 0.022 & 0.056 & 1.00e-05 & 0.033 & 0.078 & 0.111356 & 0.038 & 0.090 & 0.612128\\
\hline
$T$ & 0.011 & 0.031 & $<$ 2.2e-16 & 0.046 & 0.092 & 0.060937 & 0.013 & 0.038 & $<$ 2.2e-16 \\
$I_B$ & 0.027 & 0.061 & $<$ 2.2e-16 & 0.098 & 0.144 & 0.001642 & 0.022 & 0.054 & $<$ 2.2e-16 \\
$NI_B$ & 0.010 & 0.034 & $<$ 2.2e-16 & 0.068 & 0.111 & 0.003452 & 0.013 & 0.033 & $<$ 2.2e-16 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la probabilidad del error tipo I. Caso $E(X_1)=E(X_2)$, $\rho=$ 0.50} \label{Resultados-igual-E-2-simulacion-error-I}
\begin{center}
{\normalsize
\begin{tabular}{|l|cc|c|cc|c|cc|c|}
\cline{2-10}\multicolumn{1}{c|}{}
&\multicolumn{3}{c|}{$n=30$}& \multicolumn{3}{c|}{$n=50$} & \multicolumn{3}{c|}{$n=70$}\\ \hline
\begin{tabular}{l}
$\theta=$ (1.0,1.0,0.50) \\
$\rho=$ 0.50
\end{tabular}
& f05 & f10 & $KS$ & f05 & f10 & $KS$ & f05 & f10 & $KS$\\
\hline
$R_{n,(0,0)}$ & 0.047 & 0.091 & 0.508494 & 0.044 & 0.101 & 0.129364 & 0.048 & 0.100 & 0.413150\\
$S_{n,(0,0)}$ & 0.049 & 0.094 & 0.863178 & 0.045 & 0.098 & 0.329116 & 0.046 & 0.100 & 0.459543\\
\hline
$R_{n,(1,0)}$ & 0.049 & 0.097 & 0.665399 & 0.039 & 0.096 & 0.863178 & 0.054 & 0.097 & 0.129364\\
$S_{n,(1,0)}$ & 0.045 & 0.092 & 0.459543 & 0.043 & 0.093 & 0.291736 & 0.053 & 0.096 & 0.459543\\
\hline
$W_{n}$ & 0.022 & 0.061 & 0.013476 & 0.032 & 0.077 & 0.111356 & 0.037 & 0.081 & 0.111356\\
\hline
$T$ & 0.064 & 0.095 & 0.018402 & 0.024 & 0.039 & $<$ 2.2e-16 & 0.021 & 0.053 & 0.000179 \\
$I_B$ & 0.152 & 0.186 & $<$ 2.2e-16 & 0.073 & 0.119 & $<$ 2.2e-16 & 0.051 & 0.081 & $<$ 2.2e-16 \\
$NI_B$ & 0.080 & 0.132 & 0.054004 & 0.018 & 0.049 & $<$ 2.2e-16 & 0.007 & 0.035 & $<$ 2.2e-16 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la probabilidad del error tipo I. Caso $E(X_1)=E(X_2)$, $\rho=$ 0.75} \label{Resultados-igual-E-3-simulacion-error-I}
\begin{center}
{\normalsize
\begin{tabular}{|l|cc|c|cc|c|cc|c|}
\cline{2-10}\multicolumn{1}{c|}{}
&\multicolumn{3}{c|}{$n=30$}& \multicolumn{3}{c|}{$n=50$} & \multicolumn{3}{c|}{$n=70$}\\ \hline
\begin{tabular}{l}
$\theta=$ (1.0,1.0,0.75) \\
$\rho=$ 0.75
\end{tabular}
& f05 & f10 & $KS$ & f05 & f10 & $KS$ & f05 & f10 & $KS$\\
\hline
$R_{n,(0,0)}$ & 0.045 & 0.097 & 0.769894 & 0.060 & 0.107 & 0.995881 & 0.056 & 0.107 & 0.508494\\
$S_{n,(0,0)}$ & 0.053 & 0.092 & 0.369615 & 0.061 & 0.106 & 0.769894 & 0.050 & 0.099 & 0.902243\\
\hline
$R_{n,(1,0)}$ & 0.052 & 0.096 & 0.995881 & 0.059 & 0.106 & 0.960002 & 0.050 & 0.109 & 0.769894\\
$S_{n,(1,0)}$ & 0.050 & 0.096 & 0.818621 & 0.059 & 0.104 & 0.934732 & 0.048 & 0.106 & 0.329116\\
\hline
$W_{n}$ & 0.029 & 0.076 & 0.024117 & 0.036 & 0.085 & 0.111356 & 0.038 & 0.088 & 0.129364\\
\hline
$T$ & 0.025 & 0.049 & $<$ 2.2e-16 & 0.034 & 0.065 & 1.00e-07 & 0.024 & 0.058 & 5.30e-06 \\
$I_B$ & 0.116 & 0.140 & $<$ 2.2e-16 & 0.141 & 0.162 & $<$ 2.2e-16 & 0.129 & 0.153 & $<$ 2.2e-16 \\
$NI_B$ & 0.045 & 0.074 & 6.10e-06 & 0.033 & 0.081 & $<$ 2.2e-16 & 0.029 & 0.063 & $<$ 2.2e-16 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la probabilidad del error tipo I. Caso $E(X_1)\neq E(X_2)$, $\rho\approx$ 0.25} \label{Resultados-diferente-E-1-simulacion-error-I}
\begin{center}
{\normalsize
\begin{tabular}{|l|cc|c|cc|c|cc|c|}
\cline{2-10}\multicolumn{1}{c|}{}
&\multicolumn{3}{c|}{$n=30$}& \multicolumn{3}{c|}{$n=50$} & \multicolumn{3}{c|}{$n=70$}\\ \hline
\begin{tabular}{l}
$\theta=$ (1.5,1.0,0.31) \\
$\rho=$ 0.25311
\end{tabular}
& f05 & f10 & $KS$ & f05 & f10 & $KS$ & f05 & f10 & $KS$\\
\hline
$R_{n,(0,0)}$ & 0.048 & 0.101 & 0.257432 & 0.053 & 0.110 & 0.508494 & 0.047 & 0.100 & 0.989545\\
$S_{n,(0,0)}$ & 0.049 & 0.104 & 0.769894 & 0.052 & 0.104 & 0.172476 & 0.045 & 0.092 & 0.508494\\
\hline
$R_{n,(1,0)}$ & 0.046 & 0.097 & 0.413150 & 0.053& 0.104 & 0.863178 & 0.046 & 0.104 & 0.559560\\
$S_{n,(1,0)}$ & 0.042 & 0.098 & 0.291736 & 0.049 & 0.104 & 0.459543 & 0.051 & 0.104 & 0.508494\\
\hline
$R_{n,(0,1)}$ & 0.055 & 0.106 & 0.718379 & 0.061 & 0.112 & 0.902243 & 0.048 & 0.090 & 0.863178\\
$S_{n,(0,1)}$ & 0.049 & 0.103 & 0.459543 & 0.051 & 0.102 & 0.329116 & 0.048 & 0.089 & 0.612128\\
\hline
$W_{n}$ & 0.022 & 0.066 & 0.041633 & 0.036 & 0.076 & 0.111356 & 0.037 & 0.082 & 0.111356\\
\hline
$T$ & 0.018 & 0.046 & 1.00e-07 & 0.021 & 0.060 & 0.000318 & 0.053 & 0.093 & 0.099411 \\
$I_B$ & 0.031 & 0.060 & $<$ 2.2e-16 & 0.013 & 0.028 & $<$ 2.2e-16 & 0.108 & 0.161 & 0.000589 \\
$NI_B$ & 0.016 & 0.041 & $<$ 2.2e-16 & 0.010 & 0.018 & $<$ 2.2e-16 & 0.076 & 0.136 & 0.048751 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la probabilidad del error tipo I. Caso $E(X_1)\neq E(X_2)$, $\rho\approx$ 0.50} \label{Resultados-diferente-E-2-simulacion-error-I}
\begin{center}
{\normalsize
\begin{tabular}{|l|cc|c|cc|c|cc|c|}
\cline{2-10}\multicolumn{1}{c|}{}
&\multicolumn{3}{c|}{$n=30$}& \multicolumn{3}{c|}{$n=50$} & \multicolumn{3}{c|}{$n=70$}\\ \hline
\begin{tabular}{l}
$\theta=$ (1.5,1.0,0.62) \\
$\rho=$ 0.50623
\end{tabular}
& f05 & f10 & $KS$ & f05 & f10 & $KS$ & f05 & f10 & $KS$\\
\hline
$R_{n,(0,0)}$ & 0.050 & 0.094 & 0.329116 & 0.048 & 0.093 & 0.902243 & 0.054 & 0.110 & 0.665399\\
$S_{n,(0,0)}$ & 0.046 & 0.100 & 0.769894 & 0.048 & 0.091 & 0.769894 & 0.045 & 0.094 & 0.226206\\
\hline
$R_{n,(1,0)}$ & 0.049 & 0.093 & 0.718379 & 0.046 & 0.092 & 0.769894 & 0.050 & 0.104 & 0.718379\\
$S_{n,(1,0)}$ & 0.043 & 0.095 & 0.612128 & 0.048 & 0.087 & 0.818621 & 0.050 & 0.097 & 0.508494\\
\hline
$R_{n,(0,1)}$ & 0.050 & 0.089 & 0.612128 & 0.043 & 0.100 & 0.718379 & 0.054 & 0.100 & 0.665399\\
$S_{n,(0,1)}$ & 0.052 & 0.091 & 0.459543 & 0.043 & 0.099 & 0.459543 & 0.052 & 0.089 & 0.508494\\
\hline
$W_{n}$ & 0.026 & 0.055 & 0.003013 & 0.037 & 0.071 & 0.111356 & 0.039 & 0.079 & 0.111356\\
\hline
$T$ & 0.056 & 0.088 & 0.000526 & 0.050 & 0.104 & 0.011917 & 0.049 & 0.096 & 0.001109 \\
$I_B$ & 0.147 & 0.201 & $<$ 2.2e-16 & 0.169 & 0.223 & $<$ 2.2e-16 & 0.147 & 0.196 & $<$ 2.2e-16 \\
$NI_B$ & 0.094 & 0.152 & 0.000622 & 0.082 & 0.145 & 0.006666 & 0.076 & 0.120 & 0.078967 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la probabilidad del error tipo I. Caso $E(X_1)\neq E(X_2)$, $\rho\approx$ 0.75} \label{Resultados-diferente-E-3-simulacion-error-I}
\begin{center}
{\normalsize
\begin{tabular}{|l|cc|c|cc|c|cc|c|}
\cline{2-10}\multicolumn{1}{c|}{}
&\multicolumn{3}{c|}{$n=30$}& \multicolumn{3}{c|}{$n=50$} & \multicolumn{3}{c|}{$n=70$}\\ \hline
\begin{tabular}{l}
$\theta=$ (1.5,1.0,0.92) \\
$\rho=$ 0.75118
\end{tabular}
& f05 & f10 & $KS$ & f05 & f10 & $KS$ & f05 & f10 & $KS$\\
\hline
$R_{n,(0,0)}$ & 0.059 & 0.102 & 0.369615 & 0.053 & 0.094 & 0.508494 & 0.046 & 0.091 & 0.960002\\
$S_{n,(0,0)}$ & 0.055 & 0.105 & 0.612128 & 0.049 & 0.092 & 0.769894 & 0.048 & 0.088 & 0.508494\\
\hline
$R_{n,(1,0)}$ & 0.051 & 0.096 & 0.413150 & 0.051 & 0.099 & 0.459543 & 0.050 & 0.091 & 0.863178\\
$S_{n,(1,0)}$ & 0.053 & 0.101 & 0.459543 & 0.051 & 0.092 & 0.769894 & 0.045 & 0.086 & 0.665399\\
\hline
$R_{n,(0,1)}$ & 0.058 & 0.104 & 0.718379 & 0.049 & 0.098 & 0.459543 & 0.046 & 0.089 & 0.863178\\
$S_{n,(0,1)}$ & 0.058 & 0.108 & 0.863178 & 0.050 & 0.091 & 0.559560 & 0.046 & 0.094 & 0.329116\\
\hline
$W_{n}$ & 0.037 & 0.081 & 0.000714 & 0.042 & 0.079 & 0.111356 & 0.037 & 0.083 & 0.149677\\
\hline
$T$ & 0.029 & 0.059 & 1.70e-06 & 0.057 & 0.094 & 0.008821 & 0.078 & 0.109 & 0.065401\\
$I_B$ & 0.091 & 0.116 & $<$ 2.2e-16 & 0.209 & 0.239 & $<$ 2.2e-16 & 0.196 & 0.220 & $<$ 2.2e-16\\
$NI_B$ & 0.021 & 0.051 & $<$ 2.2e-16 & 0.089 & 0.152 & 0.001554 & 0.094 & 0.149 & 0.003483\\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
Para estudiar la potencia repetimos el experimento anterior para muestras de tama\~no $n=50$ para varias alternativas: la distribuci\'on binomial bivariante $BB(m; p_1, p_2, p_3)$, donde $p_1+p_2-p_3\leq 1,\, p_1,p_2\geq p_3>0$; la distribuci\'on binomial negativa bivariante $BNB(\nu; \gamma_0, \gamma_1, \gamma_2)$, donde $\nu\!\in\! \mathbb{N}, \gamma_0,\gamma_1\!>\!\gamma_2\!>\!0$; mixturas de la DPB de la forma $pPB(\theta)+(1\!-\!p)PB(\lambda), 0<p<1$, denotada por $PPB(p;\theta,\lambda)$; la distribuci\'on Neyman tipo A bivariante $NTAB(\lambda;\lambda_1,\lambda_2,\lambda_3)$, donde $0\!<\!\lambda_1+\lambda_2+\lambda_3\!\leq 1$; y la distribuci\'on serie logar\'itmica bivariante $SLB(\lambda_1,\lambda_2,\lambda_3)$, donde $0\!<\!\lambda_1+\lambda_2+\lambda_3\!<\!1$ (ver, por ejemplo, Kocherlakota y Kocherlakota \cite{KoKo92} para una descripci\'on de estas distribuciones).
Para generar las muestras aleatorias de las distribuciones bivariantes que usamos como alternativas, implementamos algoritmos computacionales siguiendo los procedimientos de simulaci\'on dados en Kocherlakota y Kocherlakota \cite{KoKo92}.
La distribuci\'on binomial bivariante y la distribuci\'on Neyman tipo A bivariante han sido utilizadas como alternativas en otros art\'iculos relacionados (v\'ease, por ejemplo, Loukas y Kemp (1986) \cite{LoKe86}, Rayner y Best (1995) \cite{RaBe95}). Las otras distribuciones alternativas empleadas son distribuciones an\'alogas a las utilizadas por G\"urtler y Henze (2000) \cite{GuHe00} en el caso univariante.
Los par\'ametros para estas alternativas fueron elegidos de manera que el \'indice de dispersi\'on de cada componente del vector aleatorio estuvieran cercanos de 1, concretamente, de tal manera que $|var(X_i)/E(X_i)-1| <1$, $i=1,2$. En este sentido, las alternativas consideradas est\'an ``cerca'' de la DPB.
Por otra parte, consideramos $a_1=a_2=0$, porque para esta elecci\'on de $a_1$ y $a_2$ los test estad\'isticos $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ que resultan ocupan menos tiempo computacional que para otras combinaciones de $a$.
Las Tablas \ref{Resultados-simulacion-potencia-1} y \ref{Resultados-simulacion-potencia-2} muestran las alternativas consideradas y la potencia estimada para el nivel de significaci\'on nominal $\alpha=$ 0.05. Estas tablas tambi\'en muestran el \'indice de dispersi\'on de cada componente del vector aleatorio, as\'i como el coeficiente de correlaci\'on. Al examinar los resultados en estas tablas llegamos a la conclusi\'on de que los tests que hemos propuesto son capaces de detectar todas las alternativas tratadas, mientras que, en general, los otros tests no pueden detectar la mayor\'ia de las alternativas, especialmente los tests basados en $I_B$ y $NI_B$.
De acuerdo a los resultados de las Tablas \ref{Resultados-simulacion-potencia-1} y \ref{Resultados-simulacion-potencia-2} podemos decir que las potencias de $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$ est\'an muy cerca, siendo $R_{n,w}(\hat{\theta}_n)$ un poco m\'as potente que $S_{n,w}(\hat{\theta}_n)$ en la mayor\'ia de los casos tratados. El test basado en $W_n(\hat{\theta}_n)$ es algo menos potente que los otros dos tests propuestos, en las alternativas ensayadas.
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la potencia. Alternativas: $BB(m; p_1, p_2, p_3)$, $BNB(\nu; \gamma_0, \gamma_1, \gamma_2)$ y $PPB(p;\theta,\lambda)$ }\label{Resultados-simulacion-potencia-1}
\begin{center}
{\normalsize
\begin{tabular}{|l|ccc|cccccc|}
\hline
Alternativa & $\frac{var(X_1)}{E(X_1)}$ & $\frac{var(X_2)}{E(X_2)}$& $\rho$ & $R_{n,(0,0)}$ & $S_{n,(0,0)}$ & $W_{n}$ & $T$ & $I_B$ & $NI_B$\\
\hline\hline
$BB(1;0$\text{.}$41,0$\text{.}$02,0$\text{.}$01)$ & 0.590 & 0.980 & 0.026 & 0.863 & 0.881 & 0.829 & 0.111 & 0.000 & 0.000\\
$BB(1;0$\text{.}$41,0$\text{.}$03,0$\text{.}$02)$ & 0.590 & 0.970 & 0.092 & 0.845 & 0.866 & 0.779 & 0.115 & 0.000 & 0.000\\
$BB(2;0$\text{.}$61,0$\text{.}$01,0$\text{.}$01)$ & 0.390 & 0.990 & 0.080 & 0.988 & 0.953 & 0.948 & 0.931 & 0.004 & 0.004\\
$BB(1;0$\text{.}$61,0$\text{.}$03,0$\text{.}$02)$ & 0.390 & 0.970 & 0.020 & 1.000 & 1.000 & 0.999 & 0.945 & 0.000 & 0.000\\
$BB(2;0$\text{.}$71,0$\text{.}$01,0$\text{.}$01)$ & 0.290 & 0.990 & 0.064 & 1.000 & 0.996 & 1.000 & 1.000 & 0.000 & 0.000\\
\hline\hline
$BNB(1;0$\text{.}$92,0$\text{.}$97,0$\text{.}$01)$ & 1.920 & 1.970 & 0.486 & 0.928 & 0.890 & 0.524 & 0.843 & 0.622 & 0.974\\
$BNB(1;0$\text{.}$97,0$\text{.}$97,0$\text{.}$01)$ & 1.970 & 1.970 & 0.493 & 0.933 & 0.884 & 0.526 & 0.860 & 0.633 & 0.975\\
$BNB(1;0$\text{.}$97,0$\text{.}$97,0$\text{.}$02)$ & 1.970 & 1.970 & 0.493 & 0.936 & 0.901 & 0.518 & 0.846 & 0.616 & 0.975\\
$BNB(1;0$\text{.}$98,0$\text{.}$98,0$\text{.}$01)$ & 1.980 & 1.980 & 0.495 & 0.944 & 0.906 & 0.530 & 0.855 & 0.607 & 0.980\\
$BNB(1;0$\text{.}$99,0$\text{.}$99,0$\text{.}$01)$ & 1.990 & 1.990 & 0.498 & 0.932 & 0.893 & 0.510 & 0.850 & 0.585 & 0.973\\
\hline\hline
$PPB(0$\text{.}$35;(0$\text{.}$2,0$\text{.}$2,0$\text{.}$1);(0$\text{.}$9,1$\text{.}$0,0$\text{.}$6))$ & 1.170 & 1.202 & 0.762 & 0.850 & 0.845 & 0.622 & 0.528 & 0.000 & 0.000\\
$PPB(0$\text{.}$37;(0$\text{.}$2,0$\text{.}$2,0$\text{.}$1);(0$\text{.}$9,0$\text{.}$9,0$\text{.}$8))$ & 1.178 & 1.178 & 0.956 & 0.748 & 0.754 & 0.642 & 0.458 & 0.001 & 0.001\\
$PPB(0$\text{.}$37;(0$\text{.}$2,0$\text{.}$2,0$\text{.}$1);(0$\text{.}$9,1$\text{.}$0,0$\text{.}$2))$ & 1.178 & 1.212 & 0.461 & 0.920 & 0.901 & 0.582 & 0.653 & 0.000 & 0.000\\
$PPB(0$\text{.}$40;(0$\text{.}$2,0$\text{.}$2,0$\text{.}$1);(0$\text{.}$9,1$\text{.}$0,0$\text{.}$3))$ & 1.190 & 1.226 & 0.566 & 0.964 & 0.941 & 0.556 & 0.759 & 0.000 & 0.000\\
$PPB(0$\text{.}$40;(0$\text{.}$2,0$\text{.}$3,0$\text{.}$1);(1$\text{.}$0,0$\text{.}$9,0$\text{.}$1))$ & 1.226 & 1.131 & 0.370 & 0.957 & 0.918 & 0.516 & 0.765 & 0.000 & 0.000 \\
\hline
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\caption{Resultados de simulaci\'on para la potencia. Alternativas: $NTAB(\lambda;\lambda_1,\lambda_2,\lambda_3)$ y $SLB(\lambda_1,\lambda_2,\lambda_3)$ }\label{Resultados-simulacion-potencia-2}
\begin{center}
{\normalsize
\begin{tabular}{|l|ccc|cccccc|}
\hline
Alternativa & $\frac{var(X_1)}{E(X_1)}$ & $\frac{var(X_2)}{E(X_2)}$& $\rho$ & $R_{n,(0,0)}$ & $S_{n,(0,0)}$ & $W_{n}$ & $T$ & $I_B$ & $NI_B$\\
\hline\hline
$NTAB(0$\text{.}$41;0$\text{.}$01,0$\text{.}$01,0$\text{.}$98)$ & 1.990 & 1.990 & 0.995 & 0.934 & 0.922 & 0.853 & 0.681 & 0.001 & 0.880\\
$NTAB(0$\text{.}$50;0$\text{.}$01,0$\text{.}$01,0$\text{.}$98)$ & 1.990 & 1.990 & 0.995 & 0.931 & 0.916 & 0.831 & 0.685 & 0.001 & 0.875\\
$NTAB(0$\text{.}$70;0$\text{.}$01,0$\text{.}$01,0$\text{.}$98)$ & 1.990 & 1.990 & 0.995 & 0.937 & 0.928 & 0.777 & 0.727 & 0.001 & 0.893\\
$NTAB(0$\text{.}$71;0$\text{.}$01,0$\text{.}$01,0$\text{.}$98)$ & 1.990 & 1.990 & 0.995 & 0.945 & 0.943 & 0.796 & 0.736 & 0.000 & 0.903\\
$NTAB(0$\text{.}$75;0$\text{.}$01,0$\text{.}$01,0$\text{.}$98)$ & 1.990 & 1.990 & 0.995 & 0.954 & 0.937 & 0.776 & 0.732 & 0.000 & 0.911\\
\hline\hline
$SLB(0$\text{.}$25,0$\text{.}$15,0$\text{.}$10)$ & 0.690 & 0.779 & 0.104 & 1.000 & 1.000 & 0.995 & 0.354 & 0.034 & 0.031\\
$SLB(5d/7,d/7,d/7)^*$ & 1.000 & 1.000 & 0.289 & 0.945 & 0.999 & 0.973 & 0.258 & 0.177 & 0.171\\
$SLB(3d/4,d/8,d/8)^*$ & 1.000 & 1.000 & 0.267 & 0.950 & 1.000 & 0.981 & 0.274 & 0.183 & 0.164\\
$SLB(7d/9,d/9,d/9)^*$ & 1.000 & 1.000 & 0.250 & 0.948 & 0.999 & 0.979 & 0.270 & 0.169 & 0.166\\
$SLB(0$\text{.}$51,0$\text{.}$01,0$\text{.}$02)$ & 0.668 & 0.981 & 0.098 & 1.000 & 1.000 & 0.999 & 0.429 & 0.046 & 0.041\\
\hline
\multicolumn{10}{l}{\normalsize $^*\, d=1-\exp(-1)\approx$ 0.63212.}
\end{tabular}}
\end{center}
\end{table}
\end{landscape}
Tambi\'en comparamos los tests estad\'isticos $R_{n,w}(\hat{\theta}_n)$, $S_{n,w}(\hat{\theta}_n)$ y $W_n(\hat{\theta}_n)$ desde un punto de vista computacional. La Tabla \ref{Tiempo-CPU} muestra el tiempo de CPU promedio consumido por cada uno de los tests estad\'isticos. En este aspecto, $W_n(\hat{\theta}_n)$ es mucho m\'as eficiente que $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$. Los c\'alculos fueron realizados en un servidor que tiene las siguientes caracter\'isticas t\'ecnicas: Intel(R) Xeon(R) CPU X3430 @ 2.40GHz.
\begin{table}
\caption{Tiempo de CPU promedio (en segundos).}\label{Tiempo-CPU}
\begin{center
\begin{tabular}{|l|r|r|r|}
\cline{2-4}\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{\ \ \ $n=30$\ \ \ } &\multicolumn{1}{c|}{\ \ \ $n=50$\ \ \ } & \multicolumn{1}{c|}{\ \ \ $n=70$\ \ \ } \\
\hline
$R_{n,(0,0)}$ & 39678.35\ \ & 42275.29\ \ & 44567.67\ \ \\
$S_{n,(0,0)}$ & 3019.81\ \ & 7296.59\ \ & 14202.57\ \ \\
$W_n$ & 1452.07\ \ & 1807.28\ \ & 2142.86\ \ \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Conjuntos de datos reales}\label{Conjuntos-datos-reales}
Ahora, aplicaremos los tests propuestos a dos conjuntos de datos reales. El primer conjunto de datos comprende el n\'umero de plantas de las especies \emph{Lacistema aggregatum} y \emph{Protium guianense} en cada uno de 100 cuadrantes contiguos. El primer autor que analiz\'o estos datos y los present\'o en detalle, fue Holgate (1966) \cite{Hol66}, despu\'es fueron analizados por Gillings (1974) \cite{Gil74}, Crockett (1979) \cite{Cro79}, Loukas y Kemp (1986) \cite{LoKe86}, Kocherlakota y Kocherlakota (1992) \cite{KoKo92} y tambi\'en por Rayner y Best (1995) \cite{RaBe95}. Crockett (1979) \cite{Cro79}, Loukas y Kemp (1986) \cite{LoKe86} y Rayner y Best (1995) \cite{RaBe95} examinaron dichos datos para averiguar si correspond\'ian a un modelo Poisson bivariante, concluyeron que la DPB no modelaba correctamente los datos mencionados.
El segundo conjunto de datos se refiere a la demanda de atenci\'on en sanidad en Australia, que fueron analizados por Karlis y Tsiamyrtzis (2008) \cite{KaTs08}. Los datos se refieren a la encuesta de Salud en Australia en el periodo 1977--1978. El tama\~no de la muestra es bastante grande ($n=5190$). Estos autores utilizaron dos variables: el n\'umero de consultas con un m\'edico o un especialista ($X_1$) y el n\'umero total de medicamentos prescritos y no prescritos utilizados en los \'ultimos dos d\'ias previos a la encuesta ($X_2$). Los datos se analizaron bajo el supuesto de que $(X_1,X_2)$ ten\'ia una DPB.
La Tabla \ref{Resultados-datos-reales} muestra los $p-$valores para contrastar la bondad de ajuste a una DPB para los dos conjuntos de datos reales empleando los tres tests estad\'isticos que hemos propuesto. Para ello, en el caso de $R_{n,w}(\hat{\theta}_n)$ y $S_{n,w}(\hat{\theta}_n)$, usamos
$(a_1,a_2)\in \{(0,0), (1,0), (0,1)\}$. Para los datos de las plantas, todos los tests rechazaron la hip\'otesis nula, esto es, que los datos no son bien modelados por una DPB. Este resultado coincide con los obtenidos por los investigadores que analizaron previamente este conjunto de datos. A la misma conclusi\'on se llega al emplear los datos de salud.
\begin{table}[h]
\caption{
Resultados para los conjuntos de datos reales.}\label{Resultados-datos-reales}
\begin{center}
\begin{tabular}{|l|c|c|}
\cline{2-3} \multicolumn{1}{c|}{}
& \multicolumn{1}{c|}{Plantas} & \multicolumn{1}{c|}{Salud}\\ \hline
$R_{n,(0,0)}$ & 0.002 & 0.000\\
$R_{n,(1,0)}$ & 0.004 & 0.000\\
$R_{n,(0,1)}$ & 0.004 & 0.000\\ \hline\hline
$S_{n,(0,0)}$ & 0.002 & 0.002\\
$S_{n,(1,0)}$ & 0.006 & 0.000\\
$S_{n,(0,1)}$ & 0.008 & 0.000\\ \hline\hline
$W_n$ & 0.042 & 0.000\\ \hline\hline
\multicolumn{1}{|c|}{$\hat{\theta}_n$} & (0.64000,\,0.94000,\,0.19852) &
(0.30173,\,1.21830,\,0.12518) \\ \hline
\end{tabular}
\end{center}
\end{table}
\chapter{Expresiones matem\'aticas de los tests y algunos aspectos computacionales}\label{Expresiones-matematicas}
Las expresiones matem\'aticas en las siguientes secciones son necesarias para implementar computacionalmente los tests estad\'isticos que hemos propuesto.
\section[C\'alculo del test estad\'istico $R_{n,w}(\hat{\theta}_n)$]{C\'alculo del test estad\'istico $\boldsymbol{R_{n,w}(\hat{\theta}_n)}$ } \label{Calculo-estadistico-Rnw}
Para $w$ definido como en (\ref{funcion-peso-explicita}) y para $a_1, a_2\in \mathbb{N}_0$, de (\ref{Estad-Rnw}), obtenemos
\begin{align}
R_{n,w}(\hat{\theta}_n)&=\frac{1}{n}\int_0^1\int_0^1 \left[\sum_{i=1}^n\left[ u_1^{X_{1i}}u_2^{X_{2i}}- e^{\hat{\theta}_{1n}(u_1-1)+\hat{\theta}_{2n} (u_2-1) +\hat{\theta}_{3n}(u_1-1)(u_2-1)}\right]\right]^{2} u_1^{a_1}\, u_2^{a_2} \, du\notag\\[.2 cm]
&=\frac{1}{n}\sum_{i=1}^n \sum_{j=1}^n\int_0^1\int_0^1 u_1^{X_{1i}+X_{1j}+a_1}u_2^{X_{2i}+ X_{2j}+a_2}\,du\notag\\[.2 cm]
&\ \ \ - 2\,e^{\hat{\theta}_{3n}-\hat{\theta}_{1n}- \hat{\theta}_{2n}} \!\sum_{i=1}^n\!\int_0^1 \!u_1^{X_{1i}+a_1} \,e^{\left\{\left(\hat{\theta}_{1n}- \hat{\theta}_{3n}\right)u_1\!\right\}}\!\int_0^1\! u_2^{X_{2i}+a_2} \,e^{\left\{\left(\hat{\theta}_{2n}-\hat{\theta}_{3n}+ \hat{\theta}_{3n}u_1\right)u_2\!\right\}}\,du\notag\\[.2 cm]
&\ \ \ +n\,e^{\left\{2\left(\hat{\theta}_{3n}-\hat{\theta}_{1n}- \hat{\theta}_{2n}\right)\right\}} \int_0^1u_1^{a_1}\,e^{\left\{2\left(\hat{\theta}_{1n}- \hat{\theta}_{3n}\right)u_1\right\}}\int_0^1u_2^{a_2}\, e^{\left\{2\left(\hat{\theta}_{2n}- \hat{\theta}_{3n}+\hat{\theta}_{3n}u_1\right) u_2\right\}}\, du.\label{Rnw-1}
\end{align}
Para las integrales de (\ref{Rnw-1}) que son de la forma $\displaystyle\int_0^1 t^k\exp(rt)\,dt,\,$ con $\,k\in \mathbb{N}_0, \ r>0$, integrando por partes, resulta
\begin{equation}\label{Rnw-2}
\int_0^1 t^k\,e^{rt}\,dt=\sum_{i=0}^{k} \frac{(-1)^i\, k!\exp(r)}{(k-i)!\,r^{i+1}}+\frac{(-1)^{k+1}\,k!}{r^{k+1}}\,,\ \forall\, k\!\in\mathbb{N}_0,\ r>0.
\end{equation}
Desarrollando (\ref{Rnw-1}) y empleando (\ref{Rnw-2}) con $t=u_2$, y para $k$ y $r$ apropiados, obtenemos
\begin{align}
&R_{n,w}(\hat{\theta}_n)=\frac{1}{n}\sum_{i=1}^n \sum_{j=1}^n \frac{1}{(X_{1i} +X_{1j}+a_1+1)(X_{2i}+X_{2j}+a_2+1)}\notag\\[.4 cm]
&\hspace{4mm}-2\exp(\hat{\theta}_{3n}-\hat{\theta}_{1n}- \hat{\theta}_{2n})\sum_{i=1}^n(X_{2i}+a_2)!\left\{\sum_{k=0}^{X_{2i}+ a_2}\! \frac{(-1)^k\, \exp(\hat{\theta}_{2n}- \hat{\theta}_{3n})}{(X_{2i}+a_2-k)!} \right.\notag\\[.35 cm]
&\hspace{9mm}\left.\times\int_0^1\! \frac{u_1^{X_{1i}+a_1}\exp\{\hat{\theta}_{1n}u_1\}\ du_{1}} {(\hat{\theta}_{2n}-\hat{\theta}_{3n}+\hat{\theta}_{3n}u_1 )^{k+1}} +(-1)^{X_{2i}+a_2+1}\! \int_0^1\! \frac{u_1^{X_{1i}+a_1}\exp\{(\hat{\theta}_{1n} -\hat{\theta}_{3n})u_1\}} {(\hat{\theta}_{2n}-\hat{\theta}_{3n}+\hat{\theta}_{3n}u_1 )^{X_{2i}+a_2+1}} du_1\!\right\}\notag\\[.4 cm]
&\hspace{4mm}+n\,a_2!\exp\{2( \hat{\theta}_{3n}\!-\hat{\theta}_{1n} \!-\hat{\theta}_{2n})\}\left\{\sum_{k=0}^{a_2} \frac{(-1)^k\exp\{2(\hat{\theta}_{2n}-\hat{\theta}_{3n})\}}{(a_2-k)!\ 2^{k+1}}\right.\notag\\[.35 cm]
&\hspace{9mm}\left.\times\int_0^1 \frac{u_{\!1}^{a_1}\exp(2\hat{\theta}_{1n}u_{1})\ }{(\hat{\theta}_{2n}-\hat{\theta}_{3n}+ \hat{\theta}_{3n}u_1)^{k+1}}du_{1}+\frac{(-1)^{a_2+1}} {2^{a_2+1}}\int_0^1\frac{u_{1}^{a_1}\exp\{2(\hat{\theta}_{1n}\!-\hat{\theta}_{3n}) u_1\}}{(\hat{\theta}_{2n}-\hat{\theta}_{3n}+ \hat{\theta}_{3n}u_1)^{a_2+1}}du_1\right\}.\label{Rnw-3}
\end{align}
Para las integrales de (\ref{Rnw-3}) que son de la forma $\displaystyle\int_0^1 \frac{t^k\exp(rt)}{(b+ct)^m}\,dt$, con $r,b>0$, $c\geq 0,\ k\in\mathbb{N}_0, \ m\in\mathbb{N}$, integrando por partes, se logra:
\begin{itemize}
\item [$\bullet$] Para $\,c=0$, salvo constantes, se obtiene una integral como la dada por (\ref{Rnw-2})
\item [$\bullet$] Para $\,c>0$, haciendo el cambio de variables $\ y=b+ct,\,$ se obtiene \begin{align} \int_0^1\frac{t^k\exp(rt)}{(b+ct)^m}\,dt& =\frac{\exp\left(-\frac{rb}{c}\right)} {c^{k+1}} \int_b^{b+c}\frac{(y-b)^k\exp\left(\frac{r}{c}y\right)}{y^m}\,dy\notag\\[.2 cm] &= \frac{\exp\left(-\frac{rb}{c}\right)}{c^{k+1}}\sum_{i=0}^k \binom{k}{i} (-b)^{k-i} \int_b^{b+c}\frac{y^i\exp\left(\frac{r}{c}y\right)}{y^m}\,dy\notag\\[.2 cm] &=\frac{\exp\!\left(-\frac{rb}{c}\right)}{c^{k+1}}\sum_{i=0}^k \!\binom{k}{i} (-b)^{k-i} \!\left\{\!\!
\begin{array}{ll} \displaystyle\int_b^{b+c}\frac{\exp\left(\frac{r}{c}y\right)}{y^{m-i}}\,dy, &\text{si}\ \,i<m, \\[.5 cm]
\displaystyle\int_b^{b+c}\!\!y^{i-m}\exp\!\left(\frac{r}{c}y\right)dy, & \text{si}\ i\geq m.
\end{array}
\right.\label{Rnw-4}
\end{align}
Para las integrales de (\ref{Rnw-4}), integrando nuevamente por partes, resulta
\begin{align}
&\int_b^{b+c}y^{i-m}\exp\!\left(\frac{r}{c}y\right)\,dy\notag\\[.2 cm]
&\quad\quad =(i-m)! \exp\!\left(\frac{rb}{c}\right)\sum_{j=0}^{i-m}\frac{(-1)^j} {(i-m-j)!}\!\left(\frac{c}{r}\right)^{j+1}\!\left\{\exp(r)(b+c)^{i-m-j}\!-b^{i-m-j} \right\}.\label{Rnw-5}
\end{align}
\begin{align}
&\int_b^{b+c}\frac{\exp(\frac{r}{c}y)}{y^{m-i}}\,dy= \frac{1}{(m-i-1)!}\left[\left(\frac{r}{c}\right)^{m-i-1} \int_{\frac{rb}{c}}^{\frac{r(b+c)}{c}} \frac{\exp(t)}{t}dt+\exp\!\left(\frac{rb}{c}\right)\right.\notag\\[.2 cm]
&\hspace{2mm}\left.\times\!
\left\{\!\!
\begin{array}{ll}
0, & \!\!\text{si}\ m\!-\!i=1,\\[.2 cm]
\displaystyle\sum_{j=1}^{m-i-1}\!\frac{(m-i-j-1)!\ r^{j-1}}{c^{j-1}\, \{b(b+c)\}^{m-i-j}} \left\{(b+c)^{m-i-j}-b^{m-i-j}\exp(r)\right\}, & \!\!\text{si}\ m\!-\!i\geq 2.
\end{array}
\right.\!\!
\right]\!,\label{Rnw-6}
\end{align}
\end{itemize}
As\'i, $R_{n,w}(\hat{\theta}_n)$ se puede obtener de las ecuaciones (\ref{Rnw-2}), (\ref{Rnw-4}), (\ref{Rnw-5}) y (\ref{Rnw-6}).\vskip .1 cm
Claramente, (\ref{Rnw-6}) no tiene una forma cerrada, pues contiene a la integral $\int \frac{\exp(t)}{t}dt$, la cual es muy sensible para ciertas combinaciones de los valores $r, b$ y $c$, de modo que el c\'alculo computacional de dicha integral sufre problemas de redondeo o diverge y por lo tanto no podemos obtener su valor para ciertas combinaciones de $r, b$ y $c$, lo cual nos lleva a quedar sin poder calcular $R_{n,w}(\hat{\theta}_n)$.
Por razones computacionales, la siguiente representaci\'on
\begin{equation}\label{Rnw-7}
R_{n,w}(\hat{\theta}_n) = n \sum_{i,j,k,l=0}^{\infty} \frac{\left(p_n(i,j) -p(i,j;\hat{\theta}_n)\right)\left(p_n(k,l)- p(k,l;\hat{\theta}_n)\right)}{(i+k+a_1+1)(j+l+a_2+1)}
\end{equation}
result\'o ser m\'as apropiada, donde $p(i, j;\theta)=P_{\theta}(X_1=i, X_2=j)$ y $p_n(i,j)$ es la frecuencia relativa del par $(i, j)$, dada por
\[p_n(i,j)=\frac{1}{n}\sum_{k=1}^n I{\{X_{1k}=i,X_{2k}=j\}}.\]
Adem\'as, (\ref{Rnw-7}) es menos restrictiva, pues permite que $a_1>-1$ y $a_2>-1$.
Un truncamiento de las cuatro series infinitas en $M+15$ arroj\'o valores suficientemente precisos del test estad\'istico y un buen rendimiento de la subrutina correspondiente, donde $\displaystyle M=\max \{ X_{1(n)},X_{2(n)}\}$, $X_{k(n)}=\max_{1\leq i\leq n}X_{ki} $, $k=1,2$.
\section[C\'alculo del test estad\'istico $S_{n,w}(\hat{\theta}_n)$]{C\'alculo del test estad\'istico $\boldsymbol{S_{n,w}(\hat{\theta}_n)}$ } \label{Calculo-estadistico-Snw}
Para $w$ definido como en (\ref{funcion-peso-explicita}), y para $a_1>-1$ y $a_2>-1$, de (\ref{Estad-Snw}), obtenemos
\begin{align}
&S_{n,w}(\hat{\theta}_n)\notag\\
&=\frac{1}{n}\int_0^1\!\int_0^1 \!\left[\sum_{i=1}^n\left\{X_{1i}I{\{X_{1i}\geq 1\}}u_1^{X_{1i}-1}u_2^{X_{2i}}\!-\{\hat{\theta}_{1n}\!+ \hat{\theta}_{3n}(u_2-1)\}u_1^{X_{1i}}u_2^{X_{2i}}\right\}\right]^2 \!\! u_1^{a_1}u_2^{a_2}du\notag\\[.1 cm]
&\ \ +\frac{1}{n}\int_0^1\!\int_0^1\!\left[\sum_{i=1}^n\left\{X_{2i}I{\{X_{2i}\geq 1\}}u_1^{X_{1i}}u_2^{X_{2i}-1}\!-\!\{\hat{\theta}_{2n}\!+ \hat{\theta}_{3n}(u_1-1)\}u_1^{X_{1i}}u_2^{X_{2i}}\right\}\right]^2\! \! u_1^{a_1}u_2^{a_2}du\notag\\[.1 cm]
&=\frac{1}{n}\sum_{i=1}^n\sum_{j=1}^n \left(S_{1ij}+S_{2ij}\right),\notag
\end{align}
donde
\begin{align}
S_{1ij}&=\frac{X_{1i}\,I{\{X_{1i}\geq 1\}}\,X_{1j}\,I{\{X_{1j}\geq 1\}}}{(X_{1ij}-1) (X_{2ij}+1)}+ \frac{(\hat{\theta}_{1n}- \hat{\theta}_{3n})^2} {(X_{1ij}+1)(X_{2ij}+1)}\notag\\[.1 cm]
&\hspace{10mm}-\frac{\hat{\theta}_{3n}\Bigl(X_{1i}\, I{\{X_{1i}\geq 1\}}+X_{1j}\,I{\{X_{1j}\geq 1\}}\Bigr)} {X_{1ij}\,(X_{2ij}+2)}+\frac{2\,\hat{\theta}_{3n}(\hat{\theta}_{1n}- \hat{\theta}_{3n})} {(X_{1ij}+1)(X_{2ij}+2)}\notag\\[.1 cm]
&\hspace{10mm}- \frac{(\hat{\theta}_{1n}- \hat{\theta}_{3n})\Bigl(X_{1i}\,I{\{X_{1i}\geq 1\}}+ X_{1j}\,I{\{X_{1j}\geq 1\}}\Bigr)} {X_{1ij}\,(X_{2ij}+1)}+\frac{\hat{\theta}_{3n}^{\ 2}}{(X_{1ij}+1)(X_{2ij}+3)}\,,\notag
\end{align}
\begin{align}
S_{2ij}&=\frac{X_{2i}\,I{\{X_{2i}\geq 1\}}\,X_{2j}\,I{\{X_{2j}\geq 1\}}}{(X_{1ij}+1) (X_{2ij}-1)}+ \frac{(\hat{\theta}_{2n}- \hat{\theta}_{3n})^2} {(X_{1ij}+1)(X_{2ij}+1)} \notag\\[.2 cm]
&\hspace{10mm}-\frac{\hat{\theta}_{3n}\Bigl(X_{2i}\, I{\{X_{2i}\geq 1\}}+X_{2j}\,I{\{X_{2j}\geq 1\}}\Bigr)} {(X_{1ij}+2)\,X_{2ij}}+\frac{2\,\hat{\theta}_{3n}(\hat{\theta}_{2n}- \hat{\theta}_{3n})} {(X_{1ij}+2)(X_{2ij}+1)}\notag\\[.2 cm]
&\hspace{10mm}- \frac{(\hat{\theta}_{2n}- \hat{\theta}_{3n})\Bigl(X_{2i}\,I{\{X_{2i}\geq 1\}}+ X_{2j}\,I{\{X_{2j}\geq 1\}}\Bigr)} {(X_{1ij}+1)\,X_{2ij}}+\frac{\hat{\theta}_{3n}^{\ 2}}{(X_{1ij}+3)(X_{2ij}+1)}\,,\notag
\end{align}
con $X_{1ij}=X_{1i}+X_{1j}+a_1\ $ y $\ X_{2ij}=X_{2i}+X_{2j}+a_2$, $1\leq i,j \leq n$.
\section[C\'alculo del test estad\'istico $W_n(\hat{\theta}_n)$]{C\'alculo del test estad\'istico $\boldsymbol{W_n(\hat{\theta}_n)}$} \label{Calculo-estadistico-Wn}
De (\ref{Estad-Wn-bivariante}) y (\ref{psi_1rs-2rs}), obtenemos
\begin{align}
W_n(\hat{\theta}_n)&=\sum_{r,s = 0}^M \left[\sum_{i=1}^n\left\{(r+1)p_n(r+1,s) - (\hat{\theta}_{1n}-\hat{\theta}_{3n})p_n(r,s) -\hat{\theta}_{3n}p_n(r,s-1)\right\}\right]^2\notag\\[.2 cm]
&\hspace{7mm} +\sum_{r,s = 0}^M \left[\sum_{i=1}^n\left\{(s+1)p_n(r,s+1) - (\hat{\theta}_{2n}-\hat{\theta}_{3n})p_n(r,s) -\hat{\theta}_{3n}p_n(r-1,s)\right\}\right]^2\!,\notag
\end{align}
donde $M=\max \{X_{1(n)}, X_{2(n)}\}$, $X_{k(n)}=\max_{1\leq i\leq n}X_{ki}$, $k=1,2$.
\newpage
\[\begin{array}{c}\\ \\ \end{array}\]
\newpage
\chapter{Extensiones} \label{Extensiones}
\section{El caso de dos variables Poisson independientes}
El caso de dos variables Poisson independientes ocurre cuando $\theta_3=0$, que fue excluido puesto que, seg\'un la definici\'on dada en la Secci\'on \ref{Notacion}, $\theta_3>0$. La raz\'on para no considerar este importante caso es que para que la aproximaci\'on bootstrap funcione, necesitamos que $\theta$ sea un punto interior del espacio param\'etrico $\Theta$.
Si estamos interesados en un test de bondad de ajuste para un modelo Poisson independiente, entonces podemos, o bien, considerar el test estad\'istico $R_{n,w}(\hat{\theta}_{0n})$, $S_{n,w}(\hat{\theta}_{0n})$ o $W_n(\hat{\theta}_{0n})$, con $\hat{\theta}_{0n}=(\hat{\theta}_{1n}, \hat{\theta}_{2n}, 0)$, es decir, el mismo test estad\'istico como antes con $\hat{\theta}_{3n}$ fijo e igual a su valor poblacional bajo la hip\'otesis nula, $\theta_3=0$, o bien, podemos usar el Teorema de Raikov y aplicar un test de bondad de ajuste para la distribuci\'on Poisson univariante a la suma de los componentes.
Los resultados de los Cap\'itulos \ref{Estadisticos-tipoCramer-von-Mises} y \ref{Estadistico-Wn} seguir\'an siendo v\'alidos para estos tests estad\'isticos.
\section[El caso general $d$-variante]{El caso general $\boldsymbol{d}$-variante}
Hasta ahora hemos asumido que los datos son bivariantes. En esta secci\'on mostramos que los tests que hemos propuesto en este texto se pueden extender de manera natural al caso $d$-variante general, $d\in \mathbb{N}$. Sea
\[X_1=Y_1+Y_{d+1}, \,\, X_2=Y_2+Y_{d+1},\,\, \ldots, \,\, X_d=Y_d+Y_{d+1},\]
donde $Y_1, Y_2, \ldots, Y_d, Y_{d+1}$ son v.a. Poisson (univariantes) mutuamente independientes con medias $\theta'_1=\theta_1-\theta_{d+1}>0, \ldots, \theta'_d=\theta_d-\theta_{d+1}>0$ y $\theta_{d+1} \geq 0$, respectivamente.\vskip .2 cm
La distribuci\'on conjunta del vector $(X_1, X_2, \ldots, X_d)$ es llamada distribuci\'on Poisson $d$-variante con par\'ametro $\theta=(\theta_1,\theta_2,\ldots, \theta_d,\theta_{d+1})$ (ver, por ejemplo, Johnson, Kotz y Balakrishnan (1997, p. 139) \cite{JoKoBa97}).
La fgp conjunta de $(X_1, X_2, \ldots, X_d)$ est\'a dada por
\begin{equation}\label{fgp-multivariante}
g(u;\theta)=
\exp\left\{\sum_{i=1}^d \theta_i\left(u_i-1\right)+ \theta_{d+1}
\left(\prod_{i=1}^du_i-\sum_{i=1}^du_i+d-1\right)\right\},
\end{equation}
$\forall u \in \mathbb{R}^d$.
Claramente, el test basado en el estad\'istico $R_{n,w}(\hat{\theta}_{n})$ se puede aplicar para contrastar
\[H_{0d}:(X_1, X_2, \ldots, X_d) \text{ tiene una distribuci\'on Poisson } d\text{-variante},\]
para cualquier $d$ fijo.
El siguiente resultado muestra la extensi\'on $d$-variante de la Proposici\'on
\ref{Soluc-sistema-de-dos-EDP}, lo cual nos permitir\'a proponer una extensi\'on de los tests basados en $S_{n,w}(\hat{\theta}_{n})$ y en $W_n(\hat{\theta}_{n})$.
\begin{proposicion}\label{Soluc-sistema-d-variante}
Sea $G_d=\{g:[0,1]^d\to \mathbb{R}$, tal que $g$ es una fgp y $\frac{\partial}{\partial u_i}g (u_1,\ldots, u_d)$, existen $\forall u \in [0,1]^d$, $1\leq i\leq d\}$. Sea $g(u;\theta)$ como definida en (\ref{fgp-multivariante}). Entonces $g(u;\theta)$ es la \'unica fgp en $G_d$ que satisface el siguiente sistema
\begin{equation}\label{EDPs-fgp-d-variante}
D_{id}(u;\theta)= 0,\quad 1\leq i\leq d,\ \ \forall u \in [0,1]^d,
\end{equation}
donde
\[
D_{id}(u;\theta)=\frac{\partial }{\partial u_i}g(u)-
\left\{\theta_i+\theta_{d+1}\left(\prod _{j\neq i }u_j-1\right)\right\}g(u),\quad 1\leq i\leq d.
\]
\end{proposicion}
\nt {\bf Demostraci\'on} \hspace{2pt} Consideremos el vector aleatorio $(X_1,X_2,\ldots,X_d)\in \mathbb{N}^d_0$, cuya fgp est\'a dada por $g(u_1,u_2,\ldots,u_d)=E\left(u_1^{X_1} u_2^{X_2}\cdots u_d^{X_d}\right)$. Entonces, para $i=1$ en (\ref{EDPs-fgp-d-variante}) \begin{equation}\label{EDP1-fgp-d-variante} \frac{\partial}{\partial u_1}\log g(u_1,u_2,\ldots,u_d)=\theta_1+\theta_{d+1}\left(\prod _{j\neq 1 }u_j-1\right).\end{equation}
Integrando (\ref{EDP1-fgp-d-variante}) sobre $u_1$, obtenemos
\begin{align}
&g(u_1,u_2,\ldots,u_d)= \exp\left\{\phi_1(u_2,u_3,\ldots,u_d)+\theta_1 u_1+\theta_{d+1}\left(\prod _{j=1}^d u_j-u_1\right)\right\},\notag\\[.2 cm]
&\hspace{19mm}=\exp\!\left\{\varphi_1(u_2,u_3,\ldots,u_d)+\theta_1(u_1-1) +\theta_{d+1}\!\left(\prod_{j=1}^d u_j-\!\sum_{j=1}^d u_j+d-1\right)\right\},\notag
\end{align}
donde $\,\varphi_1(u_2,u_3,\ldots,u_d)=\phi_1(u_2,u_3,\ldots,u_d)+\theta_1+ \theta_{d+1}\displaystyle\left(\sum_{j=2}^d u_j-d+1\right)$.\vskip .5 cm
Procediendo similarmente, de (\ref{EDPs-fgp-d-variante}), para $i=2,3,\ldots,d$, obtenemos
\begin{align}
&g(u_1,u_2,\ldots,u_d)= \exp\left\{\phi_2(u_1,u_3,\ldots,u_d)+\theta_2 u_2+\theta_{d+1}\left(\prod _{j=1}^d u_j-u_2\right)\right\},\notag\\[.2 cm]
&\hspace{19mm}=\exp\!\left\{\varphi_2(u_1,u_3,\ldots,u_d)+\theta_2(u_2-1) +\theta_{d+1}\!\left(\prod_{j=1}^d u_j-\sum_{j=1}^d u_j+d-1\right)\right\},\notag
\end{align}
donde $\,\varphi_2(u_1,u_3,\ldots,u_d)=\phi_2(u_1,u_3,\ldots,u_d)+\theta_2+ \theta_{d+1}\displaystyle\left(\sum_{j=1,j\neq 2}^d u_j-d+1\right)$.\vskip .6 cm
As\'i, sucesivamente hasta que
\begin{align}
&g(u_1,u_2,\ldots,u_d)= \exp\left\{\phi_d(u_1,u_2,\ldots,u_{d-1})+\theta_d u_d+\theta_{d+1}\left(\prod _{j=1}^d u_j-u_d\right)\right\},\notag\\[.2 cm]
&\hspace{18mm}=\exp\!\left\{\varphi_d(u_1,u_2,\ldots,u_{d-1})+\theta_d(u_d-1) +\theta_{d+1}\!\left(\prod_{j=1}^d u_j-\!\sum_{j=1}^d u_j+d-1\right)\right\},\notag
\end{align}
donde $\,\varphi_d(u_1,u_2,\ldots,u_{d-1})=\phi_d(u_1,u_2,\ldots,u_{d-1})+\theta_d+ \theta_{d+1}\displaystyle\left(\sum_{j=1}^{d-1} u_j-d+1\right)$.
Por lo tanto, necesariamente debe ocurrir que
\begin{align}
\varphi_1(u_2,u_3,\ldots,u_d)&=\sum_{j=2}^d \theta_j (u_j-1),\notag\\[.1 cm]
\varphi_2(u_1,u_3,\ldots,u_d)&=\sum_{\substack{j=1\\ j\neq 2}}^d \theta_j (u_j-1),\notag\\[.1 cm]
&\vdots\notag\\
\varphi_d(u_1,u_2,\ldots,u_{d-1})&=\sum_{j=1}^{d-1} \theta_j (u_j-1),\notag
\end{align}
en otras palabras, la fgp de la DP $d$-variante es la \'unica soluci\'on de (\ref{EDPs-fgp-d-variante}). $\square$\\
De la Proposici\'on \ref{Conv-FuncGenProbBiv}, $g(u)$ y sus derivadas pueden ser estimadas consistentemente por medio de la fgpe y las derivadas de la fgpe, respectivamente.
As\'i, si $H_{0d}$ fuera cierta, entonces las funciones
\[
D_{in}(u;\hat{\theta}_n )=\frac{\partial }{\partial u_i}g_n(u_1,u_2, \ldots, u_d)-\left\{\hat{\theta}_{i,n}+\hat{\theta}_{d+1,n}\left(\prod_{j\neq i }u_j-1\right)\right\} g_n(u_1,u_2, \ldots, u_d),
\]
$1\leq i \leq d$, deber\'ian estar pr\'oximas a 0, donde $\hat{\theta}_n$ es un estimador consistente de $\theta$.
Por lo tanto, para probar $H_{0d}$ podr\'iamos considerar el siguiente test estad\'istico
\[S_{d,n,w}(\hat{\theta}_n)=n\int_{[0,1]^d} \left\{D^2_{1n}(u;\hat{\theta}_n )+\cdots+D^2_{dn}(u;\hat{\theta}_{n} )\right\}w(u)\ du,\]
donde $w(u)$ es una funci\'on de peso medible, no negativa, con integral finita en $[0,1]^d$.
Haciendo las modificaciones correspondientes, se pueden conseguir resultados similares a los establecidos en el Cap\'itulo \ref{Estadisticos-tipoCramer-von-Mises}.
\vskip .2 cm
Para el test estad\'istico basado en $W_n(\hat{\theta}_{n})$, consideremos que se verifican las hip\'otesis de la Proposici\'on \ref{Soluc-sistema-d-variante}.
Sea $(X_1,\ldots,X_d)\in \mathbb{N}^d_0\,$ un vector aleatorio y sea $\,g(u_1,\ldots,u_d)=E\left(u_1^{X_1} u_2^{X_2}\cdots u_d^{X_d}\right)$ su fgp, luego, por definici\'on
\[g(u)=\sum_{r_1,r_2,\ldots,r_d\geq 0}u_1^{r_1} u_2^{r_2}\cdots u_d^{r_d} P(r_1,r_2,\ldots,r_d),\]
donde $P(r_1,r_2,\ldots,r_d)=P(X_1=r_1,X_2=r_2,\ldots,X_d=r_d)$.\vskip .2 cm
De lo anterior y de (\ref{EDPs-fgp-d-variante}), podemos escribir
\begin{align}
D_{1d}(u;\theta)&=\sum_{r_1,r_2,\ldots,r_d\geq 0}\Bigl\{(r_1+1) P(r_1+1,r_2,r_3,\ldots,r_d)-(\theta_1-\theta_{d+1}) P(r_1,r_2,\ldots,r_d)\Bigr.\notag \\
&\hspace{46mm}\Bigl.-\theta_{d+1} P(r_1,r_2-1,r_3-1,\ldots,r_d-1) \Bigr\}u_1^{r_1} u_2^{r_2}\cdots u_d^{r_d},\notag \\[.2 cm]
D_{2d}(u;\theta)&=\sum_{r_1,r_2,\ldots,r_d\geq 0}\Bigl\{(r_2+1) P(r_1,r_2+1,r_3,\ldots,r_d)-(\theta_2-\theta_{d+1}) P(r_1,r_2,\ldots,r_d)\Bigr.\notag \\
&\hspace{46mm}\Bigl.-\theta_{d+1} P(r_1-1,r_2,r_3-1,\ldots,r_d-1) \Bigr\}u_1^{r_1} u_2^{r_2}\cdots u_d^{r_d},\notag\\
&\vdots\notag\\
D_{dd}(u;\theta)&=\sum_{r_1,r_2,\ldots,r_d\geq 0}\Bigl\{(r_d+1) P(r_1,r_2,\ldots,r_{d-1},r_d+1)-(\theta_d-\theta_{d+1}) P(r_1,r_2,\ldots,r_d)\Bigr.\notag \\
&\hspace{43mm}\Bigl.-\theta_{d+1} P(r_1-1,r_2-1,\ldots,r_{d-1}-1,r_d) \Bigr\}u_1^{r_1} u_2^{r_2}\cdots u_d^{r_d}.\notag
\end{align}
Consideremos ahora las versiones emp\'iricas $D_{1n}(u;\hat{\theta}_n), D_{2n}(u;\hat{\theta}_n), \ldots, D_{dn}(u;\hat{\theta}_n)$, de las ecuaciones anteriores.
Si $H_{0d}$ fuera cierta entonces $D_{1n}(u;\hat{\theta}_n), D_{2n}(u;\hat{\theta}_n), \ldots, D_{dn}(u;\hat{\theta}_n)$ deber\'ian ser pr\'oximas a 0, $\forall u \in [0,1]^d$. Esta proximidad a cero la podemos interpretar como ya lo hicimos al inicio del Cap\'itulo \ref{Estadistico-Wn}, para el caso bivariante. Para ello observemos que
\[D_{in}(u;\hat{\theta}_n)=\sum_{r_1,r_2,\ldots,r_d\geq 0} b_i(r_1,r_2,\ldots,r_d;\hat{\theta}_n) u_1^{r_1} u_2^{r_2}\cdots u_d^{r_d},\ \ 1\leq i\leq d\]
donde
\begin{align}
b_1(r_1,r_2,\ldots,r_d;\hat{\theta}_n)&=(r_1+1) p_n(r_1+1,r_2,r_3,\ldots,r_d)-(\hat{\theta}_{1n}-\hat{\theta}_{d+1,n}) p_n(r_1,r_2,\ldots,r_d)\notag \\[.1 cm]
&\hspace{45mm}-\hat{\theta}_{d+1,n} p_n(r_1,r_2-1,r_3-1,\ldots,r_d-1),\notag\\[.2 cm]
b_2(r_1,r_2,\ldots,r_d;\hat{\theta}_n)&=(r_2+1) p_n(r_1,r_2+1,r_3,\ldots,r_d)-(\hat{\theta}_{2n}-\hat{\theta}_{d+1,n}) p_n(r_1,r_2,\ldots,r_d)\notag \\[.1 cm]
&\hspace{45mm}-\hat{\theta}_{d+1,n} p_n(r_1-1,r_2,r_3-1,\ldots,r_d-1),\notag\\[.1 cm]
&\vdots\notag\\
b_d(r_1,r_2,\ldots,r_d;\hat{\theta}_n)&=(r_d+1) p_n(r_1,r_2,\ldots,r_{d-1},r_d+1)-\!(\hat{\theta}_{dn}\!-\hat{\theta}_{d+1,n}) p_n(r_1,r_2,\ldots,r_d)\notag \\[.1 cm]
&\hspace{44mm}-\hat{\theta}_{d+1,n} p_n(r_1-1,r_2-1,\ldots,r_{d-1}-1,r_d),\notag
\end{align}
y
\[p_n(r_1,r_2,\ldots,r_d)=\frac{1}{n}\sum_{k=1}^n I{\{X_{1k}=r_1,X_{2k}=r_2,\ldots,X_{dk}=r_d\}},\]
es la frecuencia relativa emp\'irica de la $d$-tupla $(r_1,r_2,\ldots,r_d)$.
Por lo tanto, $D_{in}(u;\hat{\theta}_n)= 0$,
$\forall u \in [0,1]^d$, $1\leq i\leq d$, s\'i y s\'olo si los coeficientes de $u_1^{r_1} u_2^{r_2}\cdots u_d^{r_d}$ en las expansiones anteriores son nulos $\forall r_1,r_2,\ldots,r_d \geq 0$. Esto nos lleva a considerar el siguiente estad\'istico para contrastar $H_{0d}$:
\begin{align}
W_{d,n}(\hat{\theta}_n)&=\sum_{r_1,r_2,\ldots,r_d\geq 0}\left\{\sum_{i=1}^d b_i(r_1,r_2,\ldots,r_d;\hat{\theta}_n)^2\right\}\notag\\[.2 cm]
&=\sum_{r_1,r_2,\ldots,r_d= 0}^M\left\{\sum_{i=1}^d b_i(r_1,r_2,\ldots,r_d;\hat{\theta}_n)^2\right\},\notag
\end{align}
donde $M=\max \{X_{1(n)}, X_{2(n)},\ldots,X_{d(n)}\}$, $X_{k(n)}=\max_{1\leq j\leq n}X_{kj}$, $1\leq k\leq d$.
Haciendo los cambios respectivos, se pueden conseguir resultados similares a los establecidos en el Cap\'itulo \ref{Estadistico-Wn}.
|
\section{Introduction}
\label{sec:intro}
Whispered speech refers to a form of spoken communication in which the vocal folds do not vibrate and, therefore, there is no periodic glottal excitation. This can be intentional (e.g.,~speaking in whispers), or as a result of disease or trauma (e.g.,~patients suffering from aphonia after a total laryngectomy). The lack of pitch reduces the expressiveness and naturalness of the voice. Moreover, it can be a serious impediment for speech intelligibility in tonal languages~\cite{Chen2018} or in the presence of other interfering sources (i.e.,~cocktail party problem~\cite{Popham2018}). The conversion from whispered to voiced speech, either by reconstructing partially existent pitch contours or by generating completely new ones, is an area of research that not only has relevant practical and real-world applications, but also fosters the development of advanced speech conversion systems.
In general, existing methods
for whispered-to-voiced speech conversion either follow a data-driven or an analysis-by-synthesis approach. In the data-driven approach, machine learning is used to estimate the pitch from the available speech parameters (e.g.,~mel frequency cepstral coefficients; MFCCs). Then, a vocoder is used to synthesize speech from those by, for instance, predicting fundamental frequencies and voiced/unvoiced decisions from frame-based spectral information of the whispered signal, using Gaussian mixture models (GMMs)~\cite{Toda2008,Nakamura2011,Nakamura2012} or deep neural networks~\cite{Gonzalez2017a}. The analysis-by-synthesis approach follows a similar methodology to code-excited linear prediction~\cite{Morris2002,Ahmadi2008,Sharifzadeh2010}. To estimate pitch parameters, a common strategy is to derive those from other parameters available in the whispered signal, such as estimated speech formants~\cite{Li2014}.
A key application of whisper-to-voiced speech conversion is to provide individuals with aphonia with a more naturally sounding voice. People who have their larynx removed as a treatment for cancer inevitably lose their voice. To speak again, laryngectomees can resort to a number of methods, such as the voice valve, which produces an unnatural, whispered sound, or the electrolarynx, a vibrating device placed against the neck that generates the lost glottal excitation but, nonetheless, produces a robotic voice due to its constant vibration. In recent years, the use of whisper-to-speech reconstruction methods~\cite{Toda2008,Nakamura2011,Nakamura2012,Fuchs2012}, or silent speech interfaces~\cite{Denby2010,Gonzalez2017a,Gonzalez2017b} in which an acoustic signal is synthesised from non-audible speech-related biosignals such as the movements of the speech organs, have started to be investigated to provide laryngectomees with a better and more naturally sounding voice.
In this paper, and in contrast to previous approaches, we present a speaker-dependent end-to-end model for voiced speech generation based on generative adversarial networks (GANs)~\cite{GANsPaper}. With an end-to-end model directly performing the conversion between waveforms, we avoid the explicit extraction of spectral information, the error-prone prediction of intermediate parameters like pitch, and the use of a vocoder to synthesize speech from such intermediate parameters. With a GAN learning the mapping from whispered to voiced speech, we avoid the specification of an error loss over raw audio and make our model truly generative, thus being able to produce new realistic pitch contours. Our results show this novel pitch generation as an implicit process of waveform restoration. To evaluate our proposal, we compared the pitch contour distributions predicted by our proposal with those obtained by a regression model, observing that our proposal is able to attain more natural pitch contours than those predicted by a regression model, including a more realistic variance factor that relates to more expressiveness.
The remainder of the paper is structured as follows. In section~\ref{sec:method} we describe the method used to restore the voiced speech with GANs. The experimental setup is described in section~\ref{sec:setup}, which includes descriptions of the dataset, an RNN baseline and the hyper-parameters for our GAN model. Finally, sections~\ref{sec:results} and~\ref{sec:conclusions} contain the discussions of results and conclusions respectively.
\section{Generative Adversarial Networks for Voiced Speech Restoration}
\label{sec:method}
The proposed model is an improvement over our previous work on speech enhancement using GANs (SEGAN)~\cite{pascual2017segan,pascual2018koreansegan} in order to handle speech reconstruction tasks. SEGAN was designed as a speaker- and noise-agnostic model to generate clean/enhanced versions of aligned noisy speech signals. From now on, we change signal names for the new task, so we rather work with natural, voiced (i.e. restored) and whispered speech signals. To adapt the architecture to the task of voiced speech restoration, we decide to remove the audio alignment requirement, as the data we use has slight misalignments between input and output speech (see Section \ref{ssec:task} for more details). In addition, we introduce a number of improvements that consistently stabilize and facilitate its training after direct regularization over the waveform is removed. These modifications also refine the generated quality at the generator output when regression is removed.
\subsection{SEGAN}
We now outline the most basic aspects of SEGAN, specifically highlighting the ones that have been subject to change. For the sake of brevity we refer the reader to the original paper and code~\cite{pascual2017segan} for more detailed explanations on the old architecture and setup. The SEGAN generator network ($G$) embeds an input noisy waveform chunk into the latent space via a convolutional encoder. Then, the reconstruction is made in the decoder by `deconvolving' back the latent signals into the time domain. $G$ features skip connections with constant factors (acting as identity functions) to promote that low-level features could escape a potentially unnecessary compression from the encoder. Such skip connections also improve training stability, as they allow gradients to flow better across the deep structure of G, which has a total of 22~layers. In the denoising setup, an L$_1$ regularization term helped centering output predictions around 0, discouraging $G$ to explore bizarre amplitude magnitudes that could make the discriminator network ($D$) converge to easy discriminative solutions for the fake adversarial case.
\subsection{Adapted SEGAN}
The SEGAN architecture has been adapted to cope with misalignments in the input/output signals as mentioned before, as well as to achieve a more stable architecture and to produce better quality outputs. In the current setup, similarly to the original SEGAN mechanism, we inject whisper data to $G$, which compresses it and then recovers a version of the utterance with prosodic information.
To cope with misalignments, we get rid of the L$_1$ regularization term, as this was forcing a one-to-one correspondence between audio samples, assuming input and output had the same phase. In its place we use a softer regularization which works in the spectral domain, similar to the one used in the parallel Wavenet~\cite{oord2017parallel}. We use a non-averaged version of this loss though, as we work with large frames during training (16,384 samples per sequence), and averaging the spectral frames over this large span could be ineffective. Moreover, we calculate the loss as an absolute distance in decibels between the generated speech and the natural one.
\begin{figure}[!t]
\centering
\includegraphics[width=0.85\linewidth]{whisper2clean_ibserspeech.png}
\caption{Generator network architecture. Skip connection with learnable $\ve{a}_l$ are depicted with purple boxes. These are summed to each intermediate activation of the decoder. Encoder and decoder are like in the original SEGAN~\cite{pascual2017segan}, but with half the amount of layers and doubled pooling per layer.}
\label{fig:silentsegan}
\end{figure}
The spectral regularization is added to the adversarial loss coming from $D$ with a weighting factor $\lambda$. In SEGAN, $D$ is a learnable comparative loss function between natural or voiced signals and whispered ones. This means we have a \texttt{(natural, whispered)} paired input as a real batch sample and \texttt{(voiced, whispered)} as a fake batch sample. In contrast, $G$ has to make \texttt{(voiced,whispered)} true, thus being the adversarial objective. In the current setup, we add an additional fake signal in $D$ that will enforce the preservation of intelligibility when we forward data through $G$: the \texttt{(natural,random\_natural\_shuffle)} pair. This pair tries to send messages to $G$ about a bad behavior whenever the content between both chunks, the one coming from $G$ and the reference one, changes. Note that we are using the least-squares GAN form (LSGAN) in the adversarial component, so our loss functions, for $D$ and $G$ respectively, become
\begin{equation*}
\begin{aligned}
\underset{D}\min~V(D) & = \frac{1}{3}\E_{\ve{x},\tilde{\ve{w}}\sim p_{\text{data}}(\ve{x}, \tilde{\ve{w}})}[(D(\ve{x},\tilde{\ve{w}}) - 1)^{2}] +\\
& + \frac{1}{3}\E_{\ve{z}\sim p_{\ve{z}}(\ve{z}),\tilde{\ve{w}}\sim p_{\text{data}}(\tilde{\ve{w}})}[D(G(\ve{z},\tilde{\ve{w}}),\tilde{\ve{w}})^{2}] \\
& + \frac{1}{3}\E_{\ve{x},\ve{x}^{r}\sim p_{\text{data}}(\ve{x})}[D(\ve{x},\ve{x}^r)^{2}] \\
\underset{G}\min~V(G) & = \E_{\ve{z}\sim p_{\ve{z}}(\ve{z}),\tilde{\ve{w}}\sim p_{\text{data}}(\tilde{\ve{w}})}[(D(G(\ve{z},\tilde{\ve{w}}),\tilde{\ve{w}}) - 1)^{2}],
\end{aligned}
\end{equation*}
where $\tilde{\ve{w}}\in\mathbb{R}^T$ is the whispered utterance, $\ve{x}\in\mathbb{R}^T$ is the natural speech, $\ve{x}^r\in\mathbb{R}^T$ is a randomly chosen natural chunk within the batch, $G(\ve{z}, \tilde{\ve{w}})\in\mathbb{R}^T$ is the enhanced speech, and [$D(\ve{x}, \tilde{\ve{w}}), D(G(\ve{z}, \tilde{\ve{w}}), \tilde{\ve{w}}), D(\ve{x}, \ve{x}^r)$] are the discriminator decisions for each input pair. All of these signals are vectors of length $T$ samples except for $D$ outputs, which are scalars. $T$ is a hyper-parameter fixed during training but it is variable during test inference.
After removing the regularization factor L$_1$, the generator output can explore large amplitudes whilst adapting to mimic the speech distribution. As a matter of fact, this collapsed the training whenever the $\tanh$ activation was placed in the output layer of $G$ to bound its output to $[-1, 1]$, because the amplitude grew quickly with aggressive gradient updates and $\tanh$ would not allow $G$ to properly update anymore due to saturation. The way to correct this was bounding the gradient of $D$ by applying spectral normalization as proposed in~\cite{miyato2018spectral}. The discriminator does not have any batch normalization technique in this implementation, and its architecture is the same as in our previous work.
\begin{figure*}[htpb!]
\centering
\fbox{\includegraphics[width=0.25\linewidth, height=130pt]{clean_spec}}
\fbox{\includegraphics[width=0.25\linewidth, height=130pt]{whis_spec}}
\fbox{\includegraphics[width=0.25\linewidth, height=130pt]{dewhis_spec}}
\caption{\label{fig:wavs_specs}From left to right: natural speech, whispered speech (input to $G$) and output from $G$ as voiced signal.}
\end{figure*}
The new design of $G$ is shown in Figure~\ref{fig:silentsegan}. It remains as a fully convolutional encoder-decoder structure with skip connections, but with two changes. First, we reduce the number of layers by augmenting the pooling factor from 2 to 4 at every encoder-decoder layer. This is in line with preliminary experiments on the denoising task, where increasing pooling has been effective to improve objective scores for that task. Second, we introduce learnable skip connections, and these are now summed instead of concatenated to decoder feature maps. We thus have now learnable vectors $\ve{a}_{l}$ which multiply every channel of its corresponding shuttle layer $l$ by a scalar factor $\alpha_{l,k}$. These factors are all initialized to one. Hence, at the $j$-th decoder layer input we have the addition of the $l$-th encoder layer response following
\begin{equation*}
\ve{h}_{j} = \ve{h}_{j-1} + \ve{a}_l\odot\ve{h}_{l} ,
\end{equation*}
where $\odot$ is an element-wise product along channels.
\section{Experimental Setup}
\label{sec:setup}
To evaluate the performance of our technique, a clinical application involving the generation of audible speech from captured movement of the speech articulators is tested. More details about the experimental setup in terms of dataset, baseline and hyper-parameters for our proposed approach are given below.
\subsection{Task and Dataset}
\label{ssec:task}
In our previous work~\cite{Gonzalez2017a,Gonzalez2017b}, a silent speech system aimed at helping laryngectomy patients to recover their voices was described. The system comprised an articulator motion capture device~\cite{Fagan2008}, which monitored the movement of the lips and tongue by tracking the magnetic field generated by small magnets attached to them, and a synthesis module, which generated speech from articulatory data. To generate speech acoustics, recurrent neural networks (RNNs) trained on parallel articulatory and speech data were used. The speech produced by this system had a reasonable quality when evaluated on normal speakers, but it was not completely natural owing to limitations when estimating the pitch (i.e.,~the capturing device did not have access to any information about the glottal excitation).
In this work, we are interested on determining whether the proposed adapted SEGAN could improve those signals by generating more natural and realistic prosodic contours. To evaluate this, we have articulatory and speech data available, recorded simultaneously for 6 healthy British subjects (2 females and 4 males). Each speaker has recorded a random subset of the CMU Arctic corpus~\cite{Arctic} (25 minutes for each speaker, approximately). Then, whispered speech was generated from the articulatory data by using the RNN-based articulatory-to-speech synthesiser described in~\cite{Gonzalez2017b}. In this work, these whispered signals are taken as the input to SEGAN, which acts as a post-filter enhancing the naturalness of the signals. For each whispered signal we have a natural version, which is the original speech signal recorded by the subject. To simplify our first modeling approach we used one male and one female speakers, namely M4 and F1, and built two speaker-dependent SEGAN models. These speakers are selected for the better level of intelligibility of their whisper data within their genders. We want to note, however, that both female speakers are less intelligible in their whisper form than male speakers. These two speakers' data is split into two sets: (1) training, with approximately 90\% of the utterances and (2) test, with the remaining approximate 10\%. In order to have augmented data we follow the same chunking method as in our previous work~\cite{pascual2017segan} but window strides are one order of magnitude smaller. Hence we have a canvas of 16,384 samples ($\approx$ 1 second at 16kHz) every 50\,ms, in contrast with the previous 0.5\,s.
\begin{figure*}[!t]
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{m4_pitch_histogram}
\endminipage\hfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{pitch_histogram_f1}
\endminipage
\caption{\label{fig:meanpitch_hists} Histograms of pitch values in Hertz per utterance for male speaker (left) and female speaker (right). The three systems appearing are natural signals; RNN baseline voiced predictions with vocoder features; and voiced speech using SEGAN.}
\end{figure*}
\begin{figure*}[!t]
\includegraphics[width=\linewidth]{pitch_contours}
\caption{\label{fig:pitch_contours} Section of pitch contour of a test utterance of the male speaker calculated with Ahocoder from 4 different sources: Natural data (blue); RNN baseline (orange); Voiced with seed 100 (green) and voiced with seed 200 (red). Here it is shown how changing the seed indeed creates different plausible contours.}
\end{figure*}
\subsection{SEGAN Setup}
We use the same kernel widths of 31 as we had in~\cite{pascual2017segan}, both when encoding and decoding and for both $G$ and $D$ networks. The feature maps are incremental in the encoder and decremental in the decoder, having \{64, 128, 256, 512, 1024, 512, 256, 128, 64, 1\} in the generator and \{64, 128, 256, 512, 1024\} in the discriminator convolutional structures. The discriminator has a linear layer at the end with a single output neuron, as in the original SEGAN setup. The latent space is constructed with the concatenation of the thought vector $\ve{c}\in\mathbb{R}^{\frac{T}{1024}\times 1024}$ with the noise vector $\ve{z}\in\mathbb{R}^{\frac{T}{1024}\times 1024}$, where $\ve{z}\sim \mathcal{N}(0,I)$.
Both networks are trained with Adam~\cite{kingma2014adam} optimizer, with the two-timescale update rule (TTUR)~\cite{heusel2017gans}, such that $D$ will have a four times faster learning rate to virtually emulate many iterations in $D$ prior to updating G. This way, we have $D$ learning rate 0.0004 and $G$ learning rate 0.0001, with $\beta_1 = 0$ and $\beta_2 = 0.9$, which are the same schedules based on recent successful approaches to faster and stable convergent adversarial training~\cite{zhang2018self}. All signals processed by the GAN, either in the input/output of $G$ or the input of D, are pre-emphasized with a 0.95 factor, as it proved to help coping with some high-frequency artifacts in the de-noising setup. When we generate voiced data out of $G$ we de-emphasize it with the same factor to get the final result.
\subsection{Baseline}
To assess the performance of SEGAN in this task we have as reference the RNN-based articulatory-to-speech system from our previous work~\cite{Gonzalez2017b} and the natural data for each modeled speaker. The recurrent model is used to predict both the spectral (i.e.,~MFCCs) and pitch parameters (i.e.,~fundamental frequency, aperiodicities and unvoiced-voiced decision) from the articulatory data, so the source is articulatory data and not whispered speech in that case. The STRAIGHT vocoder~\cite{Kawahara1999} is then employed to synthesise the waveform from the predicted parameters.
\section{Results}
\label{sec:results}
We analyze the statistics of the generated pitch contours for the RNN, SEGAN and natural data. Figure~\ref{fig:meanpitch_hists} depicts the histograms of all contours extracted from predicted/natural waveforms. Ahocoder~\cite{ahocoder2014} was used to extract $\log$F0 curves, which are then converted to Hertz scale. Then, all voiced frames were selected and concatenated per each of the three systems. We come up with a long stream for each system and for the two genders. It can seen that, for both genders, voiced histograms (corresponding to SEGAN) have a broader variance than RNN ones, closer to the natural signal shape. This is understandable if we consider that the RNN was trained with a regression criterion that optimizes its output towards the mean of the pitch distribution. This ends up producing a monotonic prosody effect, normally manifested as a robotic sounding that can be heard in the audio samples referenced below. This indicates that the adversarial procedure can generate more natural pitch values.
Figure~\ref{fig:pitch_contours} shows pitch contours generated by SEGAN with different random seeds. We have to note that each random seed generates a different latent vector $\ve{z}$, so the stochasticity creates novel curves that look plausible. It also can be noted that SEGAN made some errors in determining the correct voicing decision for some speech segments. We may enforce a better behavior in a future version of the system with an auxiliary unvoiced/voiced classifier in the output of G.
Finally, figure~\ref{fig:wavs_specs} shows examples of waveforms and spectrograms for natural, whispered and voiced signals. We can appreciate how, for a small chunk of waveform, the generator network is able to refine low frequencies and gets rid of high frequency noises to approximate the natural data. Preliminary listening tests suggest that this model can achieve a good natural voiced version of the speech, but some artifacts intrinsic to the convolutional architecture (specially in high-frequencies) have to be palliated. This observation is in line with what was also prompted in the WaveGAN~\cite{donahue2018synthesizing} work, and this is also one of the potential reasons of the effectiveness of using pre-emphasis. We refer the reader to the audio samples to have a feeling of the current quality of our system~\footnote{http://veu.talp.cat/whispersegan/}.
\section{Conclusions}
\label{sec:conclusions}
We presented a speaker-dependent end-to-end generative adversarial network to act as a post-filter of whispered speech to deal with a pathological application. We adapted our previous speech enhancement GAN architecture to overcome misalignment issues and still obtained a stable GAN architecture to reconstruct voiced speech. The model is able to generate novel pitch contours by only seeing the whispered version of the speech at its input. The method generates richer curves than the baseline, which sounds monotonic in terms of prosody. Future lines of work include an even more end-to-end approach by going sensor-to-speech. Also, further study is required to alleviate intrinsic high frequency artifacts provoked by the type of decimation-interpolation architecture we base our design on.
\section{Acknowledgements}
This research was supported by the project TEC2015-69266-P (MINECO/FEDER, UE).
\clearpage
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The Relativistic Heavy Ion Collider at BNL and the Large Hadron Collider at CERN create a hot and dense matter similar to the early universe microseconds after the Big Bang. Such facilities enable us to study strongly interacting matter at the extreme temperature in the laboratory~\cite{Shuryak:2017,Chen:2018}. Analysis of multi-particle correlations is a powerful tool in exploring the underlying mechanism of particle production in hot QCD matter~\cite{M.Aaboud:2017,C.Aidala:2017,L.Adamczyk:2015,C.Adare:2015,L.Adamczyk:2014,S.Chatrchyan:2011,V.Khachatryan:2010,B.Alver:2010_1,B.Alver:2010,B.I.Abelev:2009,R.E.Ansorge:1988}. For example, inclusive two-particle $\Delta\eta$-$\Delta\phi$ correlations have been found to include two components: direct two-particle correlations and an effective "long-range" correlation due to event-by-event fluctuations of the overall particle multiplicity~\cite{M.Aaboud:2017,C.Aidala:2017,L.Adamczyk:2015,C.Adare:2015,L.Adamczyk:2014,S.Chatrchyan:2011,V.Khachatryan:2010,B.Alver:2010_1,B.Alver:2010,B.I.Abelev:2009,R.E.Ansorge:1988}. For small systems such as $pp$ collisions, one physics mechanism underlying all correlations is the global conservation of energy and momentum as well as the net strangeness, baryon number, and electric charge \cite{J.Adam:2017}. Data in p-Pb \cite{S.Chatrchyan:2013_1,G.Aad:2013,B.Abelev:2013} and $pp$ collisions~\cite{B.Alver:2007,V.Khachatryan:2010,J.Adam:2017} show that two particles separated by many units of pseudorapidity prefer to have similar azimuthal angles, thus their correlation function is peaked at $\Delta\phi$ = 0. The mechanism of the peak is dominated by effects associated with the fragmentation of hard-scattered partons from the same minijet~\cite{J.Adam:2017}, resonance decays, and the femtoscopic correlation~\cite{E.P.Rogochaya:2017,Alice:2018-3}. Exactly the same phenomenon was observed in heavy-ion collisions \cite{S.Chatrchyan:2011}, where the anisotropic flows are believed to originate from the hydrodynamical evolution~\cite{CMS-2018} or the anisotropic parton escape in transport models~\cite{L.H:2016,Z.W.Lin:2016}.
Recently, the ALICE Collaboration measured two-particle correlations in $pp$ collisions for low $p_T$ particles (below 2.5 GeV/c) at $\sqrt{s}$ = 7 TeV; it found a pronounced near-side depression in baryon-baryon correlations, which did not show up in meson-meson or baryon-antibaryon correlation functions~\cite{J.Adam:2017}. The ALICE Collaboration also compared with several Monte Carlo (MC) model calculations, where the models are unable to reproduce even qualitatively the depletion in the data. This may suggest the need to modify the particle production mechanism or the fragmentation functions in the Monte Carlo models~\cite{J.Adam:2017}. Here we perform a study of two-particle correlations with a different model: a multi-phase transport (AMPT) model. We study the dynamical evolution of the correlations in different stages including partonic interactions, hadronization, and hadronic interactions; we then investigate the underlying physics responsible for the depletion of the baryon-baryon correlations in the near side.
\section{Model and methodology}
Both the string melting version and default version of the AMPT model \cite{Z.W.Lin:2005} are applied in this work. The AMPT model was developed to simulate heavy-ion collisions in a wide colliding energy range from AGS to LHC. It consists of four main components: the initial conditions, partonic rescatterings, the conversion from partonic matter into hadronic matter, and hadronic interactions. The initial conditions, which include the spatial and momentum distributions of minijet partons and soft excited strings, are obtained from the HIJING model \cite{M.Gyulassy:1994}. In the string melting version \cite{Lin:2001zk}, excited strings are melt into quarks and antiquarks. Scatterings among the partons are modeled by Zhang's parton cascade model (ZPC) \cite{B.Zhang:1998}, which includes two-body scatterings with the cross section obtained from the perturbative QCD calculation with a screening mass. In the default AMPT model (denoted as AMPT-Default), most of the energy produced in the overlap volume of a heavy ion collision is in hadronic strings and thus not included in the parton cascade. In the string melting version of AMPT (denoted as AMPT-Melting), on the other hand, all excited hadronic strings in the overlap volume are converted into partons. In the AMPT-Default model, partons are recombined with their parent string when they stop interacting, and the resulting strings are converted to hadrons via the Lund string fragmentation \cite{B.Andersson:1983}. In the AMPT-Melting model, a spatial quark coalescence model is used to combine partons into hadrons. Dynamics of the subsequent hadronic matter is then described by the extended version of a relativistic transport (ART) model \cite{B.A.Li:1995}. So far the AMPT model has been often used in studies of heavy ion collisions. For example, it has been successful in describing multiple observables in relativistic heavy ion collisions at RHIC and LHC \cite{Z.W.Lin:2005, Z.W.Lin:2014,G.L.Ma:2016}, including pion HBT correlations \cite{Z.W.Lin:2002}, two-particle azimuthal angular correlations or longitudinal decorrelation~\cite{G.L.Ma:2006,J.Xu:2011,L.G.Pang:2015}. It has also been used to study the particle production mechanism~\cite{Y.J.Ye:2016,Liu:2017}.
Many new phenomena have been observed at LHC energies, such as the large anisotropic flows developed at p-Pb collisions~\cite{S.Chatrchyan:2013,G.Aad:2013B,B.B.Abelev:2013} and even in high multiplicity $pp$ collisions~\cite{V.Khachatryan:2010,G.Aad:2016,V.Khachatryan:2016}, which indicate that a quark-gluon plasma have been developed in the small collision systems at LHC energies~\cite{L.H:2016,Song:2017,Adare:2017wlc,J.L.Nagle:2018}. In earlier studies, the AMPT version v1.26t5/v2.26t5 \cite{ampt} has been used to study $pp$ and p-Pb collisions at LHC energies. This version \cite{Z.W.Lin:2014,G.L.Ma:2016} describes well the long range azimuthal correlations present in small collision systems~\cite{G.L.Ma:2014,A.Bzdak:2014}.
We use this version in our study, where the parton cross section is set to 1.5 mb~\cite{He:2017}.
In order to compare with experimental measurements of correlations between trigger and associated particles, we follow exactly the same analysis method as used by the ALICE Collaboration~\cite{J.Adam:2017}. The two-particle correlation function as a function of relative azimuthal angle $\Delta\phi$ and relative pseudorapidity $\Delta\eta$ between the particle pair of interest is defined as:
\begin{equation}
C(\Delta\eta,\Delta\phi) = \frac{S(\Delta\eta,\Delta\phi)/N_{pairs}^{signal}}{B(\Delta\eta,\Delta\phi)/N_{pairs}^{mixed}},
\end{equation}
where $S$($\Delta\eta$,$\Delta\phi$) is the distribution of correlated pairs and $B$($\Delta\eta$,$\Delta\phi$) is the reference distribution reflecting the single-particle acceptance. $S$ is constructed from particle pairs coming from the same event:
\begin{equation}
S(\Delta\eta,\Delta\phi) = \frac{d^2N_{pairs}^{signal}}{d{\Delta\eta}d\Delta\phi},
\end{equation}
where $N_{pairs}^{signal}$ is the number of particle pairs. $B$ is constructed using an event-mixing technique:
\begin{equation}
B(\Delta\eta,\Delta\phi) = \frac{d^2N_{pairs}^{mixed}}{d{\Delta\eta}d\Delta\phi},
\end{equation}
where $N_{pairs}^{mixed}$ is the number of particle pairs mixed from different events. In the AMPT model, particles from each event are combined with particles in the same event to build $S$($\Delta\eta$,$\Delta\phi$), while they are combined with particles from other events to build $B$($\Delta\eta$,$\Delta\phi$). Each event is mixed with 10 other events in this study to improve the statistical power of the reference estimation, and the impact parameter direction of the AMPT events is rotated randomly in the transverse plane for the $B$($\Delta\eta$,$\Delta\phi$) calculations. Another check by mixing event with similar event plane direction is carried out. The difference between different background reconstructions is negligible. In order to further investigate the correlation or anti-correlation of the particle pairs quantitatively, a one-dimensional (1-D) $\Delta\phi$ correlation function can be constructed from the 2-D correlation function by integrating over $\Delta\eta$ as
\begin{equation}
C(\Delta\phi) = A\times\frac{\int S(\Delta\eta,\Delta\phi)d\Delta\eta}{\int B(\Delta\eta,\Delta\phi)d\Delta\eta},
\end{equation}
where the normalization constant $A$ is given by $N_{pairs}^{mixed}/N_{pairs}^{signal}$.
\section{Results \& discussions}
\subsection{Two particle correlations in AMPT model}
We first use the AMPT model to calculate minimum bias $pp$ events at $\sqrt{s}$ = 7 TeV. The correlation functions for different particle pairs from AMPT-Melting are shown in Fig.~\ref{fig1:2D-CorF-melting}, and results from AMPT-Default are shown in Fig.~\ref{fig2:2D-CorF-default}. In these figures, we have applied the same kinematic selection criteria as used in the experiment measurement~\cite{J.Adam:2017} in order to directly compare with the data. They include a pseudorapidity range $|\eta|<0.8$ for all particles and a particle-type-dependent $p_T$ selection due to the detector capability: $p_T >$ 0.5 GeV/c for p($\bar{p}$), $p_T >$ 0.3 GeV/c for $K^{\pm}$, $p_T >$ 0.2 GeV/c for $\pi^{\pm}$, and $p_T >$ 0.6 GeV/c for $\Lambda$($\bar{\Lambda}$). The particle-anti-particle correlation functions (panels a1, b1, c1, d1, e1, f1, g1, h1) in Fig.~\ref{fig1:2D-CorF-melting} and Fig.~\ref{fig2:2D-CorF-default} show a clear near-side peak structure, where baryon-antibaryon correlations are qualitatively similar to the meson's. The only difference is the magnitude and width of the near-side peak, where the magnitude is higher for mesons and lower for baryon-antibaryon. For the same particle pairs of mesons and baryons (c.f. Fig.~\ref{fig1:2D-CorF-melting} and Fig.~\ref{fig2:2D-CorF-default}), the near side peak may represent the contributions from mini-jet interactions with the medium.
\begin{figure}[!htb]
\centering
\includegraphics[scale=0.20]{figures_fix/fig1_a1_pi1pi2_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_a2_pi1pi1_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_a3_pi2pi2_100.pdf}\\
\includegraphics[scale=0.20]{figures_fix/fig1_b1_k1k2_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_b2_k1k1_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_b3_k2k2_100.pdf}\\
\includegraphics[scale=0.20]{figures_fix/fig1_c1_p1p2_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_c2_p1p1_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_c3_p2p2_100.pdf}\\
\includegraphics[scale=0.20]{figures_fix/fig1_d1_l1l2_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_d2_l1l1_100.pdf}
\includegraphics[scale=0.20]{figures_fix/fig1_d3_l2l2_100.pdf}\\
\caption{The $\Delta\phi$-$\Delta\eta$ correlation functions for different particle pairs from the string melting AMPT model for $pp$ collisions at $\sqrt{s}$ = 7 TeV. }
\label{fig1:2D-CorF-melting}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[scale=0.20]{figures_fix/fig2_a1_pi1pi2_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_a2_pi1pi1_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_a3_pi2pi2_100_default.pdf}\\
\includegraphics[scale=0.20]{figures_fix/fig2_b1_k1k2_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_b2_k1k1_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_b3_k2k2_100_default.pdf}\\
\includegraphics[scale=0.20]{figures_fix/fig2_c1_p1p2_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_c2_p1p1_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_c3_p2p2_100_default.pdf}\\
\includegraphics[scale=0.20]{figures_fix/fig2_d1_l1l2_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_d2_l1l1_100_default.pdf}
\includegraphics[scale=0.20]{figures_fix/fig2_d3_l2l2_100_default.pdf}
\caption{Similar to Fig.~\ref{fig1:2D-CorF-melting} but from the default AMPT model.}
\label{fig2:2D-CorF-default}
\end{figure}
Similar to the measured correlation functions in $pp$ collisions at LHC energies, a distinct near-side peak at ($\Delta\eta$,$\Delta\phi$) $\sim$ (0,0) is observed for meson-meson pairs~\cite{V.Khachatryan:2010,V.Khachatryan:2016,G.Aad:2016}. As discussed in earlier studies~\cite{V.Khachatryan:2010,V.Khachatryan:2016,G.Aad:2016,J.Adam:2017}, the peak is a combination of several effects, such as the fragmentation of hard-scattered partons, higher mass resonance decays, femtoscopic correlations and Coulomb interaction among charge particles. The AMPT model includes most of the physics process except the femtoscopic effect, thus it reproduces the peak structure of the correlation functions. A more interesting feature is the pronounced depression of near-side distribution in the $\bar{p}$-$\bar{p}$ and $\bar{\Lambda}$-$\bar{\Lambda}$ correlation functions as shown in panel (c3) and (d3) of Fig.~\ref{fig1:2D-CorF-melting}. This depression structure only shows up in the AMPT-Melting model but not in the AMPT-Default version [c.f. Fig.~\ref{fig2:2D-CorF-default}]. It could be due to additional parton cascade and different hadronization process between the two versions of the AMPT model, as we shall discuss in more details next.
\begin{figure}[!htb]
\includegraphics[scale=0.32]{figures_proj_fix/fig3_a_pionplus_pionminus}
\includegraphics[scale=0.32]{figures_proj_fix/fig3_b_kaonplus_kaonminus}\\
\includegraphics[scale=0.32]{figures_proj_fix/fig3_c_proton_protonbar}
\includegraphics[scale=0.32]{figures_proj_fix/fig3_d_lambda_lambdabar}
\caption{One-dimensional $\Delta\phi$ correlation functions for $\pi^+$-$\pi^-$, $K^+$-$K^-$, $p$-$\bar{p}$, and $\Lambda$-$\bar{\Lambda}$ in $pp$ collisions. Open symbols are AMPT results with t(hadron) = 0 fm/c (hadronic stage off) or t(hadron)=20 fm/c (hadronic stage on). Solid points are experimental data~\cite{J.Adam:2017}.}
\label{fig3:1D-CorF-unlike-sign}
\end{figure}
To further compare with the experimental data \cite{J.Adam:2017}, we project the 2-D correlation functions in Fig.~\ref{fig1:2D-CorF-melting} and Fig.~\ref{fig2:2D-CorF-default} over the $|\Delta\eta| < $1.3 window to the $\Delta\phi$ axis. Figure~\ref{fig3:1D-CorF-unlike-sign} shows the particle-anti-particle pair correlation functions along the $\Delta\phi$ axis in the AMPT model.
We have chosen the hadronic rescattering time $t_H$ of 0 or 20 fm/c, with $t_H$ = 0 fm/c representing the results with the hadronic cascade turned off, to investigate the hadronic rescattering contributions to the correlation functions. In general, both AMPT-Default and AMPT-Melting can qualitatively describe the experimental data, including the near side peak structure and the away side flat distributions. It is expected that mini-jets gives a dominant contribution to the structure of the two-particle angular correlations in $pp$ collisions. Quantitatively, both versions of the AMPT model with sufficient hadronic interaction time ($t_H$ = 20 fm/c) better describe the $K^+$-$K^-$ correlation function than that with no hadronic interaction time, indicating that high mass resonance decays and hadronic scatterings contribute significantly to the correlation function. In addition, the string melting AMPT model provides a better description of the correlation function data of baryon-antibaryon pairs than the default AMPT model; this suggests that, in addition to hadronic interactions, partonic interactions and hadronization are also important for baryon correlation functions in $pp$ collisions at LHC energies.
\begin{figure}[!htb]
\includegraphics[scale=0.32]{figures_proj_fix/fig4_a1_pionplus_pionplus}
\includegraphics[scale=0.32]{figures_proj_fix/fig4_a2_pionminus_pionminus}\\
\includegraphics[scale=0.32]{figures_proj_fix/fig4_b1_kaonplus_kaonplus}
\includegraphics[scale=0.32]{figures_proj_fix/fig4_b2_kaonminus_kaonminus}\\
\includegraphics[scale=0.32]{figures_proj_fix/fig4_c1_proton_proton}
\includegraphics[scale=0.32]{figures_proj_fix/fig4_c2_protonbar_protonbar}\\
\includegraphics[scale=0.32]{figures_proj_fix/fig4_d1_lambda_lambda}
\includegraphics[scale=0.32]{figures_proj_fix/fig4_d2_lambdabar_lambdabar}
\caption{Similar to Fig.~\ref{fig3:1D-CorF-unlike-sign} but for correlation functions of pair with the same charge.}
\label{fig4:1D-CorF-like-sign}
\end{figure}
We show the same-charge correlation functions from the AMPT model in Fig.~\ref{fig4:1D-CorF-like-sign}. For $\pi$-$\pi$ correlations, our model calculations describe the shape of the data reasonably well but have a lower magnitude for the near side peak. This is likely due to the abscence of quantum statistics effects in the AMPT model. Similar findings were reached based on the comparison of PYTHIA calculations to the experimental data by the ALICE Collaboration~\cite{J.Adam:2017}.
On the proton-proton correlation function, results from AMPT-Melting and AMPT-Default both show no depression in the near side.
However, the anti-proton-anti-proton correlation functions from the AMPT-Melting version
are closer to the data than that from the AMPT-Default calculations.
Since the proton-proton correlation function may suffer from resonance weak decay contributions, we have investigated this effect by forcing all the final $\Lambda$ and $\Sigma^0$ to decay in the model calculations. The results including these weak decays are shown as blue open stars in Fig.~\ref{fig4:1D-CorF-like-sign}, and we see that the effect is small.
Regarding the $\Lambda$-$\Lambda$ and $\bar{\Lambda}$-$\bar{\Lambda}$ correlation functions, the AMPT-Melting version well describes the experimental data, including a clear depression structure on the near side and a strong enhancement on the away side. The results between $t_H$ = 0 fm/c and $t_H$ = 20 fm/c for $\Lambda$s are almost identical. It could be due to the fact that the current AMPT model does not contain resonance decay contribution to $\Lambda$ (other than from $\Sigma^0$ decays), or the role of hadronic scattering is tiny. We also find that results from the default AMPT model are similar to results from PYTHIA calculations~\cite{J.Adam:2017} and can not describe the experiment data.
Similar studies using the Monte Carlo event generators with different input parameters of PYTHIA~\cite{T.Sjostrand:2006} were done by the ALICE Collaboration~\cite{J.Adam:2017}. The results of MC models reproduce reasonably well the meson pair correlations but fail to reproduce the baryon correlations. For those MC model studies\cite{J.Adam:2017}, first of all, significant differences are also seen for baryon-antibaryon pairs, where the magnitude of the near-side peak is much higher in all MC models than the ALICE data. Here we have seen that the AMPT-Melting version reproduces the experimental results reasonably well [c.f. Fig.~\ref{fig3:1D-CorF-unlike-sign}]. Furthermore, no depression is observed for protons and $\Lambda$s for any of the MC models using by ALICE~\cite{J.Adam:2017}. Instead, a near-side peak is present for particle-particle pairs in the PYTHIA results~\cite{J.Adam:2017}. In our calculations, an interesting observation is that a depression in the near side is present in Fig.\ref{fig4:1D-CorF-like-sign} on $\Lambda$-$\Lambda$ and $\bar{\Lambda}$-$\bar{\Lambda}$ correlations. However, no depression is observed for proton-proton or $\bar{p}$-$\bar{p}$ correlations in from our AMPT calculations; instead a near-side peak is present in proton-proton correlations.
\subsection{Two particle correlations in AMPT model with new quark coalescence}
\begin{figure}[!htb]
\includegraphics[scale=0.24]{figures_proj_fix/fig5_a1_pionplus_pionplus_newCoal}
\includegraphics[scale=0.24]{figures_proj_fix/fig5_a2_pionminus_pionminus_newCoal}\\
\includegraphics[scale=0.24]{figures_proj_fix/fig5_b1_kaonplus_kaonplus_newCoal}
\includegraphics[scale=0.24]{figures_proj_fix/fig5_b2_kaonminus_kaonminus_newCoal}\\
\includegraphics[scale=0.24]{figures_proj_fix/fig5_c1_proton_proton_newCoal}
\includegraphics[scale=0.24]{figures_proj_fix/fig5_c2_protonbar_protonbar_newCoal}\\
\includegraphics[scale=0.24]{figures_proj_fix/fig5_d1_lambda_lambda_newCoal}
\includegraphics[scale=0.24]{figures_proj_fix/fig5_d2_lambdabar_lambdabar_newCoal}
\caption{Comparison of two-particle correlations between the old and new quark coalescence in the string melting AMPT model.}
\label{fig4:1D-CorF-like-sign-newCoal}
\end{figure}
The version of the AMPT-Melting model used until here has some limitations which may affect our results. For example, the current (i.e. ``old'') coalescence component in the AMPT version v2.26t5 forces the numbers of mesons, baryons, and antibaryons in an event to be separately conserved through the quark coalescence process, where only the net-baryon number needs to be conserved. The recent development on a new quark coalescence component~\cite{He:2017} in the AMPT model removes this forced separate conservations in the old quark coalescence model, and it has been shown to provide a better description of baryon productions at LHC energies~\cite{He:2017}. In Fig.~\ref{fig4:1D-CorF-like-sign-newCoal} we compare the correlations from the old and new quark coalescence. The effect between the difference quark coalescence model is tiny on meson-meson correlations. For baryon-baryon correlations, however, the new quark coalescence lead to different results, where the baryon-baryon correlation is now almost the same as the corresponding antibaryon-antibaryon correlation (as expected at this high energy).
\subsection{The transverse momentum dependence of two particle correlations in AMPT model}
\begin{figure}[!htb]
\includegraphics[scale=0.32]{figures_proj_fix/fig6_a_proton_proton_lowPt}
\includegraphics[scale=0.32]{figures_proj_fix/fig6_b_protonbar_protonbar_lowPt}
\includegraphics[scale=0.32]{figures_proj_fix/fig6_c_lambda_lambda_lowPt}
\includegraphics[scale=0.32]{figures_proj_fix/fig6_d_lambdabar_lambdabar_lowPt}
\caption{
One-dimensional $\Delta\phi$ correlation functions of p-p, $\bar{p}$-$\bar{p}$, $\Lambda$-$\Lambda$ and $\bar{\Lambda}$-$\bar{\Lambda}$ for $p_T <$ 1.0 GeV/c from the AMPT model. Open symbols represent AMPT calculations with different configurations as illustrated in the figure. Solid points are experiment data~\cite{J.Adam:2017}. }
\label{fig5:CF-lowpt}
\end{figure}
The two-particle angular correlations may depend on the transverse momentum. We investigate this effect by studying the baryon-baryon correlation functions of baryons with $p_T<$ 1.0 GeV/c in the model calculations, shown in Fig.~\ref{fig5:CF-lowpt}. In comparison to the model results for the full $p_T$ window [c.f. Fig.~\ref{fig4:1D-CorF-like-sign}], clear differences are found. An anti-correlation structure is observed, and proton-proton correlations from the new quark coalescence model are closer to the experimental data.
\subsection{The electric charge dependence of two particle correlations in AMPT model}
The two-particle angular correlations may also depend on the electric charge of the pairs. Figure~\ref{fig6:CF-p-la} presents results from the AMPT-Melting model on p-$\Lambda$ and $\bar{p}$-$\bar{\Lambda}$ correlations in $pp$ collisions at LHC energies. It is seen that the shape of the correlation function is similar to the results present in Fig.~\ref{fig4:1D-CorF-like-sign-newCoal} and Fig.~\ref{fig5:CF-lowpt}. Our results are consistent with the experimental findings, where the depression is a characteristic attribute connected solely to the baryonic nature of a particle~\cite{J.Adam:2017}.
\begin{figure}[!htb]
\includegraphics[scale=0.32]{figures_proj_fix/fig7_a_proton_lambda}
\includegraphics[scale=0.32]{figures_proj_fix/fig7_b_protonbar_lambdabar}
\includegraphics[scale=0.32]{figures_proj_fix/fig7_c_proton_lambda_lowPt}
\includegraphics[scale=0.32]{figures_proj_fix/fig7_d_protonbar_lambdabar_lowPt}
\caption{
One-dimensional $\Delta\phi$ correlation functions of p-$\Lambda$, $\bar{p}$-$\bar{\Lambda}$ in comparison with the experimental data for $p_T<$ 2.5 GeV/c (upper panels) and $p_T<$ 1.0 GeV/c (lower panels). Open symbols represent AMPT calculations with different configurations as illustrated in the figure. Solid points are experiment data~\cite{J.Adam:2017}. }
\label{fig6:CF-p-la}
\end{figure}
\subsection{The parton cross section dependence of two particle correlations in AMPT model}
Next, we address the parton scattering effect on the two-particle correlations by using different parton scattering cross sections in the string melting AMPT model,
in order to investigate the separate contributions from parton cascade and from quark coalescence hadronization.
We compare the results for 0 mb, 1.5 mb, 3 mb and 6 mb in Fig.~\ref{fig7:CF-cross-section}, where the 0 mb results represent the hadronization contribution only.
The parton cross section is seen to have a small effect for $\pi$-$\pi$ correlations but a large effect for both proton-proton and anti-proton-anti-proton correlations; this suggests that parton interactions play an important role in baryon pair correlations at LHC energies.
\begin{figure}[!htb]
\includegraphics[scale=0.28]{figures_proj_fix/fig8_a_pionplus_pionplus_newCoal}
\includegraphics[scale=0.28]{figures_proj_fix/fig8_b_proton_proton_newCoal}
\includegraphics[scale=0.28]{figures_proj_fix/fig8_c_protonbar_protonbar_newCoal}
\caption{Two-particle correlations from AMPT-Melting with the new coalescence at different parton cross sections.}
\label{fig7:CF-cross-section}
\end{figure}
\subsection{The system size dependence of two particle correlations in AMPT model}
Finally, we extend the study to p-Pb collisions to see whether such a depletion structure of correlations will be present in small systems from $pp$ to p-Pb collisions. Figure~\ref{fig8:CF-p-Pb} shows the same-charge particle pair correlations for mesons and for baryons in the string melting AMPT with the new quark coalescence. The correlations are shown for a low multiplicity interval and a high multiplicity interval separately, where the parton stage lifetime may be different. We see the usual correlations for the meson pairs but a clear depression on the near side for the baryon pairs. These results indicate that such a depression structure of low $p_T$ baryon pair correlations are present in both $pp$ and p-Pb collisions at LHC energies.
\begin{figure}[!htb]
\includegraphics[scale=0.32]{figures_proj_fix/fig9_a_ntmax150_mul_L50_newCoal}
\includegraphics[scale=0.32]{figures_proj_fix/fig9_b_ntmax150_mul_M100_newCoal}
\caption{Two-particle correlations for the multiplicity interval $<$50 (left panel) and $>$100 in p-Pb collisions at $\sqrt{s}= 5.02$ TeV from AMPT-Melting with the new coalescence.}
\label{fig8:CF-p-Pb}
\end{figure}
\section{conclusion}
We have carried out a detailed study of two-particle angular correlations in $pp$ and p-Pb collisions at LHC energies in the framework of a multi-phase transport model, with the focus on understanding the origin of anti-correlation between baryon pairs observed in the experiment. We find that mini-jet and hadronic scatterings are both important components in order to describe the experimental data, especially for the meson pairs. In addition, only the string melting AMPT model can qualitatively describe the near side depression in the angular correlations of baryon pairs, which suggests that quark coalescence and parton scatterings are essential to describe the particle productions in $pp$ collisions at LHC energies. The new quark coalescence model for string melting AMPT improves the description on experimental data. By comparing the correlation results with difference parton scattering cross sections, it is also clear that parton scatterings are important for baryon pair correlations at LHC energies. Extension to p-Pb collisions with the AMPT model predicts similar baryon-baryon correlations as observed in $pp$ collisions.
\section{acknowledgements}
We gratefully acknowledge discussions with Dr. G. L. Ma. This work was supported in part by the Major State Basic Research Development Program in China under Contract No. 2014CB845400 and No. 2015CB856904, the National Natural Science Foundation of China under contract Nos. 11775288, 11421505, 11628508 and 11520101004.
|
\section{Introduction}\label{B-sec1}
This paper is a continuation of the first author's paper:~\cite{fujino-slc-trivial}.
We strongly recommend the reader to look
at \cite[1.~Introduction]{fujino-slc-trivial} before starting to read this paper.
In \cite{fujino-slc-trivial}, we
introduced the notion of basic slc-trivial fibrations, which
is a kind of canonical bundle formula for reducible varieties, and
investigated some fundamental properties.
For the precise definition of basic slc-trivial fibrations,
see \cite[Definition 4.1]{fujino-slc-trivial} or
Definition \ref{C-def3.1} below.
The following statement is one of the main results of \cite{fujino-slc-trivial}.
\begin{thm}[{\cite[Theorem 1.2]{fujino-slc-trivial}}]\label{B-thm1.1}
Let $f\colon (X, B)\to Y$ be a basic slc-trivial fibration and
let $\mathbf B$ and $\mathbf M$ be the induced
discriminant and moduli $\mathbb Q$-b-divisors of $Y$ respectively.
Then we have the following properties:
\begin{itemize}
\item[(i)] $\mathbf K+\mathbf B$ is $\mathbb Q$-b-Cartier, and
\item[(ii)] $\mathbf M$ is b-potentially nef, that is,
there exists a proper birational morphism $\sigma \colon Y'\to Y$
from a normal variety $Y'$ such that
$\mathbf M_{Y'}$ is a potentially nef $\mathbb Q$-divisor on $Y'$ and
that $\mathbf M=\overline{\mathbf M_{Y'}}$.
\end{itemize}
\end{thm}
For the definition and some basic properties of b-divisors,
see \cite[2.3.2 b-divisors]{corti} and \cite[Section 2]{fujino-slc-trivial}.
\medskip
On moduli $\mathbb Q$-b-divisors, we have the following conjecture,
which is still widely open.
\begin{conj}[{b-semi-ampleness conjecture,
see \cite[Conjecture 1.4]{fujino-slc-trivial}}]
\label{B-conj1.2}
Let $f \colon (X, B)\to Y$ be a basic
slc-trivial fibration.
Then the moduli $\mathbb Q$-b-divisor $\mathbf M$ is
b-semi-ample.
\end{conj}
The main purpose of this paper is to prove the following theorem.
\begin{thm}[Main Theorem]\label{B-thm1.3}
Let $f \colon (X, B)\to Y$ be a basic slc-trivial fibration
such that $Y$ is complete.
Let $\mathbf M$ be the moduli $\mathbb Q$-b-divisor
associated to $f \colon (X, B)\to Y$.
Assume that
there exists a proper birational
morphism $\sigma \colon Y'\to Y$ from a normal
variety $Y'$ such that
$\mathbf M=\overline {\mathbf M_{Y'}}$ with $\mathbf M_{Y'}\equiv 0$. Then
$\mathbf M_{Y'}\sim _{\mathbb Q}0$ holds.
\end{thm}
Theorem \ref{B-thm1.3} solves Conjecture \ref{B-conj1.2}
when the moduli $\mathbb Q$-b-divisor $\mathbf M$ is
b-numerically trivial.
It is obviously a generalization of \cite[Theorem 3.5]{ambro-moduli} and
\cite[Theorem 1.3]{floris}.
More precisely, Florin Ambro and Enrica Floris
proved Theorem \ref{B-thm1.3} for klt-trivial fibrations and lc-trivial fibrations,
respectively.
\medskip
As a direct consequence of Theorem \ref{B-thm1.3},
we have the following result:~Corollary \ref{B-cor1.4}.
It says that the b-semi-ampleness
conjecture
(see Conjecture \ref{B-conj1.2}) holds true when the base space is a curve.
Note that Corollary \ref{B-cor1.4} was already proved for klt-trivial fibrations
by Florin Ambro (see \cite[Theorem 0.1]{ambro-shokurov}).
\begin{cor}\label{B-cor1.4}
Let $f \colon (X, B)\to Y$ be a basic slc-trivial fibration
with $\dim Y=1$.
Then the moduli $\mathbb Q$-divisor $M_Y$ of $f \colon
(X, B)\to Y$ is
semi-ample.
\end{cor}
For the proof of Theorem \ref{B-thm1.3}, we closely follow Floris's arguments in
\cite{floris}. We adapt her proof of Theorem \ref{B-thm1.3} for lc-trivial fibrations
to our setting. As is well known, the
main ingredient of \cite[Theorem 0.1]{ambro-shokurov},
\cite[Theorem 3.5]{ambro-moduli}, and \cite[Theorem 1.3]{floris}
is Deligne's result on local subsystems of polarizable variations
of $\mathbb Q$-Hodge structure (see \cite[Corollaire (4.2.8)]{deligne}).
\medskip
In \cite{fujino-fujisawa}, the first and the second
authors discussed variations of mixed
Hodge structure toward applications for
higher-dimensional algebraic varieties (see also \cite{ffs}).
One of the most important
applications of \cite{fujino-fujisawa} is the proof of the
projectivity of the coarse moduli spaces of stable
varieties in \cite{fujino-ann}. Then the first author
introduced the notion of basic slc-trivial fibrations
in \cite{fujino-slc-trivial} in order to make results
in \cite{fujino-fujisawa} useful
for various geometric applications. The
first and the third authors established that
every quasi-log canonical pairs have only
Du Bois singularities in \cite{fujino-haidong} by using
\cite{fujino-slc-trivial}.
We strongly recommend the reader to look
at \cite[1.~Introduction]{fujino-slc-trivial} for
more details. In this paper, we prove
\cite[Conjecture 1.4]{fujino-slc-trivial} under
some special assumption. We freely use the
formulation introduced in \cite{fujino-slc-trivial}
and the
arguments in this paper heavily depend on \cite{fujino-fujisawa}.
\medskip
We briefly explain the organization of this paper.
In Section \ref{A-sec2},
we fix the notation and recall some definitions for the reader's convenience.
In Section \ref{C-sec3}, we quickly recall the notion of basic slc-trivial fibrations
and some definitions following \cite{fujino-slc-trivial}.
In Section \ref{d-sec4}, we see that
the cyclic group action constructed in
\cite[Section 6]{fujino-slc-trivial} preserves some parts of weight filtrations of
the variation of mixed Hodge structure.
Section \ref{C-sec5} is devoted to the proof of Theorem \ref{B-thm1.3}.
By using the result obtained in Section \ref{d-sec4},
we reduce Theorem \ref{B-thm1.3} to
Deligne's result on local subsystems of polarizable variations
of $\mathbb Q$-Hodge structure.
\begin{ack}
The first author was partially
supported by JSPS KAKENHI Grant Numbers
JP16H03925, JP16H06337.
The second author was partially supported by JSPS KAKENHI
Grant Number JP16K05107.
The authors would like to thank Takeshi Abe for
useful discussions and comments.
\end{ack}
\begin{conventions}
We work over $\mathbb C$, the complex number field, throughout
this paper. We freely use the basic
notation of the minimal model program as in
\cite{fujino-fundamental} and \cite{fujino-foundations}.
A {\em{scheme}} means a separated scheme of
finite type over $\mathbb C$.
A {\em{variety}} means a reduced scheme, that is,
a reduced separated scheme of finite type over $\mathbb C$.
In this paper, a variety may be reducible.
However, we sometimes assume that a variety is irreducible without
mentioning it explicitly if there is no danger of confusion.
The set of integers (resp.~rational numbers)
is denoted by $\mathbb Z$ (resp.~$\mathbb Q$).
The set of positive rational numbers (resp.~integers)
is denoted by $\mathbb Q_{>0}$ (resp.~$\mathbb Z_{>0}$).
\end{conventions}
In this paper, we do not use $\mathbb R$-divisors.
We only use $\mathbb Q$-divisors.
\section{Preliminaries}\label{A-sec2}
In this section, we quickly recall some basic definitions and notation
for the reader's convenience.
For the details, see \cite[Section 2]{fujino-slc-trivial}.
\medskip
Let us start with the definition of {\em{simple normal crossing pairs}}.
\begin{defn}[Simple normal crossing pairs]\label{A-def2.1}
We say that the pair $(X, B)$ is {\em{simple normal crossing}}
at a point $a\in X$ if $X$ has a Zariski open neighborhood
$U$ of $a$ that can be embedded in a smooth
variety $M$, where $M$ has a regular system of parameters
$(x_1, \ldots, x_p, y_1, \ldots, y_r)$ at $a=0$ in
which $U$ is defined by a monomial equation
$$
x_1\cdots x_p=0
$$
and
$$
B=\sum _{i=1}^r b_i (y_i=0)|_U, \quad
b_i\in \mathbb Q.
$$
We say that $(X, B)$ is a {\em{simple normal crossing pair}}
if it is simple normal crossing at every point of $X$.
If $(X, 0)$ is a simple normal crossing pair, then
$X$ is called a {\em{simple normal crossing variety}}.
If $(X, B)$ is a simple normal crossing pair and
$B$ is reduced, then $B$ is called a {\em{simple
normal crossing divisor}} on $X$.
Let $(X, B)$ be a simple normal crossing pair
such that all the coefficients of $B$ are
less than or equal to one.
Let $\nu \colon X^\nu\to X$ be the normalization of $X$.
We put $K_{X^\nu}+\Theta=\nu^*(K_X+B)$, that is,
$\Theta$ is the sum of
the inverse images of $B$ and the singular locus of $X$.
By assumption, all the coefficients of $\Theta$ are less than or equal to one.
Therefore, it is easy to see that $(X^\nu, \Theta)$ is sub log canonical.
In this situation, we simply say that $W$ is
a {\em{stratum}} of $(X, B)$ if $W$ is an irreducible component of $X$ or
$W$ is the $\nu$-image of some log canonical center of $(X^\nu, \Theta)$.
We note that a stratum of a simple normal crossing variety $X$
means a stratum of a simple normal crossing pair
$(X, 0)$.
\end{defn}
We write the precise definition of {\em{semi-log canonical pairs}},
{\em{slc centers}}, and {\em{slc strata}} for the reader's convenience.
For the details of semi-log canonical pairs, we recommend the reader to
see \cite{fujino-fund-slc}.
\begin{defn}[Semi-log canonical pairs]\label{x-def2.2}
Let $X$ be an equidimensional scheme which
satisfies Serre's $S_2$ condition and
is normal crossing in codimension one. Let $\Delta$
be an effective $\mathbb Q$-divisor on $X$
such that no irreducible component of $\Supp \Delta$
is contained in the singular locus of $X$ and that
$K_X+\Delta$ is $\mathbb Q$-Cartier.
We say that $(X, \Delta)$ is a {\em{semi-log canonical}} pair if
$(X^\nu, \Delta_{X^\nu})$ is log canonical
in the usual sense, where $\nu:X^\nu\to X$
is the normalization of $X$ and
$K_{X^\nu}+\Delta_{X^\nu}=\nu^*(K_X+\Delta)$,
that is, $\Delta_{X^\nu}$ is the sum of the inverse
images of $\Delta$
and the conductor of $X$. An {\em{slc center}} of $(X, \Delta)$
is the $\nu$-image of an lc center of $(X^\nu, \Delta_{X^\nu})$.
An {\em{slc stratum}} of $(X, \Delta)$
means either an slc center of $(X, \Delta)$ or an
irreducible component of $X$.
\end{defn}
We recall various definitions and operations of
($\mathbb Q$-)divisors.
We note that we are mainly interested in {\em{reducible}} varieties
in this paper.
\begin{say}[Divisors]\label{x-say2.3}
Let $X$ be a scheme with structure sheaf $\mathcal O_X$ and let
$\mathcal K_X$ be the sheaf of total quotient rings of $\mathcal O_X$.
Let $\mathcal K^*_X$ denote the (multiplicative)
sheaf of invertible elements in $\mathcal K_X$,
and $\mathcal O^*_X$ the sheaf of invertible
elements in $\mathcal O_X$.
We note that $\mathcal O_X\subset \mathcal K_X$
and $\mathcal O^*_X\subset
\mathcal K^*_X$ hold.
A {\em{Cartier divisor}} $D$ on $X$ is a global section of
$\mathcal K^*_X/\mathcal O^*_X$, that is,
$D$ is an element of $\Gamma(X, \mathcal K^*_X/\mathcal O^*_X)$.
A {\em{$\mathbb Q$-Cartier divisor}} is an element of
$\Gamma (X, \mathcal K^*_X/\mathcal O^*_X)\otimes
_{\mathbb Z}\mathbb Q$. Let $D_1$ and $D_2$
be two $\mathbb Q$-Cartier divisors
on $X$. Then $D_1$ is {\em{linearly}}
(resp.~{\em{$\mathbb Q$-linearly}})
{\em{equivalent}} to $D_2$, denoted by $D_1\sim D_2$ (resp.~$D_1
\sim _{\mathbb Q}D_2$), if
$$
D_1=D_2+\sum _{i=1}^k r_i (f_i)
$$
such that $f_i\in \Gamma (X, \mathcal K^*_X)$ and $r_i\in \mathbb Z$
(resp.~$r_i\in \mathbb Q$) for every $i$.
We note that $(f_i)$ is a {\em{principal Cartier divisor}}
associated to $f_i$, that is,
the image of $f_i$ by
$$
\Gamma (X, \mathcal K^*_X)\to
\Gamma(X, \mathcal K^*_X/\mathcal O^*_X).
$$
Let $f \colon X\to Y$ be a morphism between schemes.
If there exists a $\mathbb Q$-Cartier
divisor $B$ on $Y$ such that
$D_1\sim _{\mathbb Q} D_2+f^*B$, then
$D_1$ is said to be {\em{relatively $\mathbb Q$-linearly
equivalent to $D_2$}}.
It is denoted by $D_1\sim _{\mathbb Q, f}D_2$ or
$D_1\sim _{\mathbb Q, Y} D_2$.
\medskip
From now on, let $X$ be an equidimensional scheme. We note
that $X$ is not necessarily regular in codimension one.
A ({\em{Weil}}) {\em{divisor}} $D$ on $X$ is a finite formal
sum
$$
D=\sum _i d_iD_i,
$$
where $D_i$ is an irreducible reduced closed subscheme of $X$
of pure codimension one and $d_i$ is an integer
for every $i$ such that $D_i\ne D_j$ for every $i\ne j$.
If $d_i \in \mathbb Q$ for every $i$,
then $D$ is called a {\em{$\mathbb Q$-divisor}}.
Let $D=\sum _i d_i D_i$ be a $\mathbb Q$-divisor as above.
We put
\begin{equation*}
D^{\leq 1}=\sum _{d_i\leq 1}d_i D_i, \quad
D^{<1} =\sum _{d_i<1}d_iD_i, \quad
D^{= 1}=\sum _{d_i= 1} D_i, \quad \text{and} \quad
\lceil D\rceil =\sum _i \lceil d_i \rceil D_i,
\end{equation*}
where $\lceil d_i\rceil$ is the integer defined by $d_i\leq
\lceil d_i\rceil <d_i+1$. Let $D$ be a $\mathbb Q$-divisor.
We also put
$$
\lfloor D\rfloor=-\lceil -D\rceil.
$$
We call $D$ a {\em{subboundary}}
$\mathbb Q$-divisor if $D=D^{\leq 1}$ holds.
When $D$ is effective and $D=D^{\leq 1}$ holds,
we call $D$ a {\em{boundary}} $\mathbb Q$-divisor.
We further assume that
$f \colon X\to Y$ is a surjective morphism onto an irreducible
variety $Y$.
Then we put
$$
D^v=\sum _{f(D_i)\subsetneq Y}d_i D_i \quad
\text{and} \quad D^h=D-D^v,
$$
and call $D^v$ the {\em{vertical part}}
and $D^h$ the {\em{horizontal part}} of $D$
with respect to $f \colon X\to Y$, respectively.
\medskip
Finally, let $D$ be a $\mathbb Q$-Cartier divisor on a
complete normal irreducible variety $X$.
If $D\cdot C=0$ for any complete curve $C$ on $X$, then
$D$ is said to be {\em{numerically trivial}}. When $D$ is
numerically trivial, we simply write $D\equiv 0$.
\end{say}
Let us recall the definition of {\em{potentially nef divisors}}
introduced by the first author in \cite{fujino-slc-trivial}.
\begin{defn}[{Potentially nef divisors, see
\cite[Definition 2.5]{fujino-slc-trivial}}]\label{x-def2.4}
Let $X$ be a normal
irreducible variety and let $D$ be a divisor on $X$.
If there exist a completion $X^\dag$ of $X$,
that is, $X^\dag$ is a complete normal
variety and contains $X$ as a dense Zariski open set, and
a nef divisor $D^\dag$ on $X^\dag$ such that
$D=D^\dag|_X$, then $D$ is called
a {\em{potentially nef}} divisor on $X$.
A finite $\mathbb Q_{>0}$-linear
combination of potentially nef divisors is called
a {\em{potentially nef}} $\mathbb Q$-divisor.
\end{defn}
Although it is dispensable,
the following definition is very useful when we state our results (see
Theorems \ref{B-thm1.1} and \ref{B-thm1.3}).
We note that the {\em{$\mathbb Q$-Cartier
closure}} of a $\mathbb Q$-Cartier $\mathbb Q$-divisor
$D$ on a normal variety $X$ is the $\mathbb Q$-b-divisor
$\overline D$ with trace
$$
\overline D _Y=f^*D,
$$
where $f \colon Y\to X$ is a proper birational morphism
from a normal variety $Y$.
\begin{defn}[{see \cite[Definition 2.12]{fujino-slc-trivial}}]\label{x-def2.5}
Let $X$ be a normal irreducible variety.
A $\mathbb Q$-b-divisor $\mathbf D$ of $X$
is {\em{b-potentially nef}}
(resp.~{\em{b-semi-ample}}) if there
exists a proper birational morphism $X'\to X$ from a normal
variety $X'$ such that $\mathbf D=\overline {\mathbf D_{X'}}$, that
is, $\mathbf D$ is the $\mathbb Q$-Cartier closure of $\mathbf D_{X'}$, and that
$\mathbf D_{X'}$ is potentially nef
(resp.~semi-ample).
A $\mathbb Q$-b-divisor $\mathbf D$ of $X$ is {\em{$\mathbb Q$-b-Cartier}}
if there is a proper birational morphism $X'\to X$ from a normal
variety $X'$ such that $\mathbf D=\overline{\mathbf D_{X'}}$.
Let $X$ be a complete normal irreducible variety.
A $\mathbb Q$-b-divisor $\mathbf D$ of $X$ is
{\em{b-numerically trivial}} (resp.~{\em{$\mathbb Q$-b-linearly trivial}})
if there exists a proper birational morphism
$X'\to X$ from a complete normal variety $X'$ such that
$\mathbf D=\overline{\mathbf D_{X'}}$ with $\mathbf D_{X'}\equiv 0$
(resp.~$\mathbf D_{X'}\sim _{\mathbb Q}0$).
\end{defn}
For the details of (b-)potentially nef divisors,
we recommend the reader to see \cite[Section 2]
{fujino-slc-trivial}.
\section{Quick review of basic slc-trivial fibrations}\label{C-sec3}
In this section, we quickly recall some definitions
of {\em{basic slc-trivial fibrations}} in \cite[Section 4]{fujino-slc-trivial}.
We recommend the reader to see \cite[1.15]{fujino-slc-trivial}
for some historical comments.
\medskip
We introduce the notion of basic slc-trivial fibrations.
\begin{defn}[{Basic slc-trivial fibrations,
see \cite[Definition 4.1]{fujino-slc-trivial}}]\label{C-def3.1}
A {\em{pre-basic slc-trivial fibration}} $f \colon (X, B)\to Y$ consists of
a projective surjective morphism
$f \colon X\to Y$ and a simple normal crossing pair $(X, B)$ satisfying
the following properties:
\begin{itemize}
\item[(1)] $Y$ is a normal irreducible variety,
\item[(2)] every stratum of $X$ is dominant onto $Y$ and
$f_*\mathcal O_X\simeq \mathcal O_Y$,
\item[(3)] $B$ is a $\mathbb Q$-divisor such that $B=B^{\leq 1}$ holds
over
the generic point of $Y$, and
\item[(4)] there exists
a $\mathbb Q$-Cartier $\mathbb Q$-divisor $D$ on $Y$ such that
$$
K_X+B\sim _{\mathbb Q}f^*D.
$$
\end{itemize}
If a pre-basic slc-trivial fibration $f \colon (X, B)\to Y$ also satisfies
\begin{itemize}
\item[(5)] $\rank f_*\mathcal O_X(\lceil -B^{<1}\rceil)=1$,
\end{itemize}
then it is called a {\em{basic slc-trivial fibration}}.
\end{defn}
Roughly speaking, if $X$ is irreducible
and $(X, B)$ is sub kawamata log terminal
(resp.~sub log canonical) over the generic point of $Y$,
then it is a klt-trivial fibration (resp.~an lc-trivial fibration).
\medskip
In order to define discriminant $\mathbb Q$-b-divisors and
moduli $\mathbb Q$-b-divisors for basic slc-trivial fibrations,
we need the notion of induced (pre-)basic slc-trivial fibrations.
\begin{say}[{Induced (pre-)basic slc-tirival
fibrations, see \cite[4.3]{fujino-slc-trivial}}]\label{C-say3.2}
Let $f \colon (X, B)\to Y$ be a \linebreak
(pre-)basic slc-trivial fibration
and let $\sigma \colon Y'\to Y$ be a generically finite
surjective morphism from a normal irreducible variety $Y'$.
Then we have an {\em{induced {\em{(}}pre-{\em{)}}basic slc-trivial fibration}}
$f' \colon (X', B_{X'})\to Y'$, where
$B_{X'}$ is defined by $\mu^*(K_X+B)=K_{X'}+B_{X'}$, with
the following commutative diagram:
$$
\xymatrix{
(X', B_{X'}) \ar[r]^{\mu} \ar[d]_{f'} & (X, B)\ar[d]^{f} \\
Y' \ar[r]_{\sigma} & Y,
}
$$
where $X'$ coincides with
$X\times _{Y}Y'$ over a nonempty Zariski open set of $Y'$.
More precisely, $X'$ is a simple normal crossing variety with a morphism
$X'\to X\times _Y Y'$ that is an isomorphism over
a nonempty Zariski open set of $Y'$ such that
$X'$ is projective over $Y'$ and that every stratum of $X'$ is dominant onto
$Y'$.
\end{say}
Now we are ready to define {\em{discriminant
$\mathbb Q$-b-divisors}} and
{\em{moduli $\mathbb Q$-b-divisors}} for basic slc-trivial fibrations.
\begin{say}[{Discriminant and
moduli $\mathbb Q$-b-divisors,
see \cite[4.5]{fujino-slc-trivial}}]\label{C-say3.3}
Let $f \colon (X, B)\to Y$ be a \linebreak
(pre-)basic
slc-trivial fibration as in Definition \ref{C-def3.1}.
Let $P$ be a prime divisor on $Y$.
By shrinking $Y$ around the generic point of $P$,
we assume that $P$ is Cartier. We set
$$
b_P=\max \left\{t \in \mathbb Q\, \left|\,
\begin{array}{l} {\text{$(X^\nu, \Theta+t\nu^*f^*P)$ is sub log canonical}}\\
{\text{over the generic point of $P$}}
\end{array}\right. \right\},
$$
where $\nu \colon X^\nu\to X$ is the normalization and
$K_{X^\nu}+\Theta=\nu^*(K_X+B)$, that is,
$\Theta$ is the sum of the inverse images of $B$ and the singular
locus of $X$, and
set $$
B_Y=\sum _P (1-b_P)P,
$$
where $P$ runs over prime divisors on $Y$.
Then it is easy to see that
$B_Y$ is a well-defined $\mathbb Q$-divisor on
$Y$ and is called the {\em{discriminant
$\mathbb Q$-divisor}} of $f \colon (X, B)\to Y$. We set
$$
M_Y=D-K_Y-B_Y
$$
and call $M_Y$ the {\em{moduli $\mathbb Q$-divisor}} of $f \colon
(X, B)\to Y$.
By definition, we have
$$
K_X+B\sim _{\mathbb Q}f^*(K_Y+B_Y+M_Y).
$$
Let $\sigma\colon Y'\to Y$ be a proper birational morphism
from a normal variety $Y'$ and let $f' \colon (X', B_{X'})\to Y'$ be
an induced (pre-)basic slc-trivial fibration
by $\sigma \colon Y'\to Y$.
We can define $B_{Y'}$, $K_{Y'}$ and $M_{Y'}$ such that
$\sigma^*D=K_{Y'}+B_{Y'}+M_{Y'}$,
$\sigma_*B_{Y'}=B_Y$, $\sigma _*K_{Y'}=K_Y$
and $\sigma_*M_{Y'}=M_Y$. We note that
$B_{Y'}$ is independent of the choice of $(X', B_{X'})$,
that is, $B_{Y'}$ is well defined. Hence
there exist a unique $\mathbb Q$-b-divisor $\mathbf B$
such that
$\mathbf B_{Y'}=B_{Y'}$ for every $\sigma \colon Y'\to Y$ and a unique
$\mathbb Q$-b-divisor $\mathbf M$ such that $\mathbf M_{Y'}=M_{Y'}$ for
every $\sigma \colon Y'\to Y$.
Note that $\mathbf B$ is called the {\em{discriminant $\mathbb Q$-b-divisor}} and
that $\mathbf M$ is called
the {\em{moduli $\mathbb Q$-b-divisor}} associated to $f \colon (X, B)\to Y$.
We sometimes simply say that $\mathbf M$ is
the {\em{moduli part}} of $f \colon (X, B)\to Y$.
\end{say}
For the full details of this section, we recommend the reader to see
\cite[Section 4]{fujino-slc-trivial}.
\section{On variation of mixed Hodge structure}\label{d-sec4}
This section heavily depends on \cite[Sections 4 and 7]{fujino-fujisawa}.
We strongly recommend the reader to take a quick look at \cite[Section 4]{fujino-fujisawa}
before reading this section.
\medskip
Let us quickly recall \cite[Theorem 7.1]{fujino-fujisawa}, which is
one of the main ingredients of
\cite{fujino-slc-trivial} (see \cite[Section 3]{fujino-slc-trivial}).
\begin{thm}[{\cite[Theorem 7.1]{fujino-fujisawa}}]\label{d-thm4.1}
Let $(V, T)$ be a simple normal crossing pair such that $T$ is
reduced and let
$h \colon V\to Y$ be a projective surjective morphism onto a smooth
variety $Y$.
Assume that every stratum of $(V, T)$ is dominant onto $Y$.
Let $\Sigma$ be a simple normal crossing divisor on $Y$ such that
every stratum of $(V, T)$ is smooth over $Y^*=Y\setminus \Sigma$.
We put $V^*=h^{-1}(Y^*)$, $T^*=T|_{V^*}$, and $d=\dim V-\dim Y$.
Let $\iota \colon V^*\setminus T^*\hookrightarrow V^*$ be the natural
open immersion.
Then the local system $R^k(h|_{V^*})_*\iota_!\mathbb Q_{V^*\setminus T^*}$
underlies a graded polarizable admissible variation of
$\mathbb Q$-mixed Hodge structure on $Y^*$ for every $k$.
We put $\mathcal V^k_{Y^*}=R^k(h|_{V^*})_*\iota_!\mathbb Q_{V^*\setminus T^*}
\otimes \mathcal O_{Y^*}$ for
every $k$.
Let
$$
\cdots \subset F^{p+1}(\mathcal V^k_{Y^*})\subset F^p(\mathcal V^k_{Y^*})
\subset F^{p-1}(\mathcal V^k_{Y^*})\subset \cdots
$$
be the Hodge filtration.
We assume that all the local monodromies on the local system
$R^k(h|_{V^*})_*\iota_!\mathbb Q_{V^*\setminus T^*}$ around $\Sigma$
are unipotent for every $k$.
Then $R^kh_*\mathcal O_V(-T)$ is isomorphic to the canonical extension
of
$$
\Gr ^0_F(\mathcal V^k_{Y^*})=F^0(\mathcal V^k_{Y^*})/F^1(\mathcal V^k_{Y^*}),
$$
which is denoted by $\Gr^0_F(\mathcal V^k_Y)$, for every $k$.
By taking the dual, we have $$R^{d-k}h_*\omega_{V/Y}(T)\simeq
\left(\Gr^0_F(\mathcal V^k_Y)\right)^*$$ for every $k$.
\end{thm}
For the details of Theorem \ref{d-thm4.1}, we recommend the reader to
see \cite[Sections 4 and 7]{fujino-fujisawa} (see also \cite{ffs}).
We note that the reader can find basic definitions of variations of
mixed Hodge structure in \cite[Section 3]{fujino-fujisawa}.
\medskip
Let us introduce the notion of {\em{birational maps}}
of simple normal crossing
pairs.
\begin{defn}[Birational maps of simple normal crossing
pairs]\label{d-def4.2}
Let $(V_1, T_1)$ and $(V_2, T_2)$ be
simple normal crossing pairs such that
$T_1$ and $T_2$ are reduced.
Let $\alpha \colon V_1\dashrightarrow V_2$ be a proper birational map.
Assume that
there exist Zariski open sets $U_1$
and $U_2$ of $V_1$ and $V_2$ respectively such that
$U_1$ contains
the generic point of
any stratum of $(V_1, T_1)$, $U_2$ contains
the generic point of any stratum of $(V_2, T_2)$,
and $\alpha$ induces an isomorphism between
$(U_1, T_1|_{U_1})$ and $(U_2, T_2|_{U_2})$.
Then we call $\alpha$ a {\em{birational map between
$(V_1, T_1)$ and $(V_2, T_2)$}}.
\end{defn}
As an easy application of \cite[Lemma 6.2]{fujino-fujisawa} and
\cite[Theorem 1.4]{bierstone}, we can prove the
following useful lemma.
\begin{lem}\label{d-lem4.3}
Let $(V_1, T_1)$ and $(V_2, T_2)$ be simple normal crossing
pairs
such that $T_1$ and $T_2$ are reduced.
Let $\alpha \colon V_1\dashrightarrow V_2$ be a birational map
between $(V_1, T_1)$ and $(V_2, T_2)$.
Then there exists a commutative diagram
\begin{equation}\label{d-eq4.1}
\xymatrix{
& (V', T') \ar[dl]_-{p_1}\ar[dr]^-{p_2}& \\
(V_1, T_1) \ar@{-->}[rr]_-\alpha&& (V_2, T_2),
}
\end{equation}
where $(V', T')$ is a simple normal crossing pair
such that $T'$ is reduced, and $p_i$ is a proper
birational morphism between $(V', T')$ and
$(V_i, T_i)$ for $i=1, 2$.
In this situation, $p_i$ induces a natural one-to-one correspondence
between the set of strata of $(V', T')$ and that of $(V_i, T_i)$ for
$i=1, 2$.
Let $S$ be any stratum of $(V', T')$.
Then we have $$Rp_{i*}\mathcal O_S\simeq
\mathcal O_{p_i(S)}$$ for
$i=1, 2$.
Moreover, we have $$Rp_{i*}\mathcal O_{V'}(-T')\simeq
\mathcal O_{V_i}(-T_i)$$ for
$i=1, 2$.
\end{lem}
\begin{proof}
By \cite[Theorem 1.4]{bierstone}, we can take
a desired commutative diagram \eqref{d-eq4.1}, where
$p_i$ is a proper birational morphism
between $(V', T')$ and $(V_i, T_i)$ such that
$p_i$ is an isomorphism over $U_i$ for $i=1, 2$.
By \cite[Lemma 6.2]{fujino-fujisawa},
we have $Rp_{i*}\mathcal O_{V'}(-T')\simeq
\mathcal O_{V_i}(-T_i)$ for $i=1, 2$.
Let $S$ be a stratum of $(V', T')$.
Then $p_i(S)$ is a stratum of $(V_i, T_i)$ since
$p_i$ is a birational
morphism between $(V', T')$ and $(V_i, T_i)$ for $i=1, 2$.
Therefore, $p_i(S)$ is a smooth
irreducible variety and $p_i \colon S\to p_i(S)$
is obviously birational for $i=1, 2$.
This implies that $Rp_{i*}\mathcal O_S\simeq
\mathcal O_{p_i(S)}$ for $i=1, 2$.
Since $p_i \colon V'\to V_i$ is a proper
birational morphism between $(V', T')$ and
$(V_i, T_i)$, it is easy to see that there
exists a natural one-to-one correspondence between
the set of strata of $(V', T')$ and that of
$(V_i, T_i)$ for $i=1, 2$.
\end{proof}
\begin{rem}\label{x-rem4.4}
In Lemma \ref{d-lem4.3}, we assume that
$\alpha\colon (V_1, T_1)\dashrightarrow (V_2, T_2)$ is
projective over a fixed scheme $Y$, that is,
there exists the following commutative diagram
$$
\xymatrix{
(V_1, T_1) \ar[dr]_-{h_1}\ar@{-->}[rr]^-\alpha&& (V_2, T_2)
\ar[dl]^-{h_2}\\
&Y&
}
$$
such that $h_1$ and $h_2$ are projective.
Then we see that we can make $V'$ projective over $Y$ by the
proof of Lemma \ref{d-lem4.3}.
\end{rem}
We define a somewhat artificial
condition for birational maps of simple normal crossing pairs.
We will use it in Lemma \ref{x-lem4.6} below.
For the basic definitions of semi-simplicial
varieties, see, for example, \cite[Section 5.1]{peters-steenbrink}.
\begin{defn}\label{x-def4.5}
Let $(V, T)$ be a simple normal crossing
pair such that
$T$ is reduced.
Let $\alpha \colon V\dashrightarrow V$ be a birational map
between $(V, T)$ and $(V, T)$
in the sense of Definition \ref{d-def4.2}.
We say that $\alpha$ satisfies condition $(\bigstar)$
if there exists a commutative diagram
\begin{equation}\label{d-eq4.2}
\xymatrix{
& (V', T') \ar[dl]_-{p_1}\ar[dr]^-{p_2}& \\
(V, T) \ar@{-->}[rr]_-\alpha&& (V, T)
}
\end{equation}
with the following properties:
\begin{itemize}
\item[(1)] $(V', T')$ is a simple normal crossing pair such that $T'$ is reduced.
\item[(2)] $p_i$ is a proper
birational morphism between $(V', T')$ and $(V, T)$ in the sense of
Definition \ref{d-def4.2} for $i=1, 2$.
\item[(3)] There are semi-simplicial resolutions
$\varepsilon _T \colon T_\bullet \to T$
and $\varepsilon _V \colon V_\bullet \to V$, that is,
$T_\bullet$ and $V_\bullet$ are semi-simplicial varieties, $\varepsilon_T$ and
$\varepsilon _V$ are argumentations and of cohomological descent,
such that $V_p$ and $T_q$ are disjoint unions of some strata of $(V, T)$ for all
$p$ and $q$ and
that they fit in the following commutative diagram
\begin{equation}\label{d-eq4.3}
\xymatrix{T_\bullet\ar[d]_-{\varepsilon_T}\ar[r]^-\phi& V_\bullet
\ar[d]^-{\varepsilon _V}\\
T \ar[r]_j& V,
}
\end{equation}
where $\phi$ is a morphism of semi-simplicial varieties
and $j$ is the natural closed embedding.
Moreover, $\varepsilon_T \colon S\to \varepsilon _T(S)$
(resp.~$\varepsilon_V \colon S\to \varepsilon _V(S)$) is
a natural isomorphism for any irreducible component $S$ of $T_\bullet$
(resp.~$V_\bullet$). We note that
$S$ is a stratum of $(V, T)$.
\item[(4)] There are semi-simplicial varieties $\varepsilon_{T'} \colon
T'_\bullet
\to T'$ and $\varepsilon _{V'} \colon V'_\bullet \to V'$ such that
$\varepsilon _{T'}$ and $\varepsilon _{V'}$ are argumentations,
$V'_p$ and $T'_q$ are disjoint unions of
some strata of $(V', T')$ for all $p$ and $q$
and that they fit in the following commutative diagram
\begin{equation}\label{d-eq4.4}
\xymatrix{T'_\bullet\ar[d]_-{\varepsilon_{T'}}\ar[r]^-{\phi'}& V'_\bullet
\ar[d]^-{\varepsilon _{V'}}\\
T' \ar[r]_{j'}& V',
}
\end{equation}
where $\phi'$ is a morphism of semi-simplicial varieties and
$j'$ is the natural closed embedding.
As in (3), $\varepsilon_{T'} \colon S'\to \varepsilon _{T'}(S')$
(resp.~$\varepsilon_{V'} \colon S'\to \varepsilon _{V'}(S')$) is
a natural isomorphism for any irreducible component $S'$ of $T'_\bullet$
(resp.~$V'_\bullet$). We note that $S'$ is a stratum of $(V', T')$.
\item[(5)] The following commutative diagram
\begin{equation}\label{d-eq4.5}
\xymatrix{&& T' \ar[dll]_-{p_1|_{T'}}\ar[dd]^(.60){j'}\ar[drr]^-{p_2|_{T'}}&& \\
T\ar[dd]_-j \ar@{-->}[rrrr]|-\hole ^(.40){\alpha|_T}&&&& T\ar[dd]^-j \\
&& V'\ar[dll]_-{p_1}\ar[drr]^-{p_2} && \\
V \ar@{-->}[rrrr]_-\alpha&&&& V
}
\end{equation}
can be lifted to a commutative diagram
\begin{equation}\label{d-eq4.6}
\xymatrix{&& T'_\bullet \ar[dll]_-{p_1|_{T'_\bullet}}\ar[dd]^(.60){\phi'}
\ar[drr]^-{p_2|_{T'_\bullet}}&& \\
T_\bullet\ar[dd]_-\phi \ar@{-->}[rrrr]|-\hole ^(.40){\alpha|_{T_\bullet}}&&&& T_\bullet\ar[dd]^-\phi \\
&& V'_\bullet\ar[dll]_-{p_1|_{V'_\bullet}}\ar[drr]^-{p_2|_{V'_\bullet}} && \\
V _\bullet\ar@{-->}[rrrr]_-{\alpha_\bullet}&&&& V_\bullet
}
\end{equation}
over \eqref{d-eq4.5}
by \eqref{d-eq4.3} and \eqref{d-eq4.4}
such that $p_1|_{V'_p}, p_2|_{V'_p}, \alpha_p, p_1|_{T'_q},
p_2|_{T'_q}$, and $\alpha|_{T'_q}$ are birational maps
of smooth varieties for all $p$ and $q$.
\item[(6)] If $\alpha\colon (V, T)\dashrightarrow (V, T)$ is
projective over a fixed scheme $Y$, that is,
there exists the following commutative diagram
$$
\xymatrix{
(V, T) \ar[dr]_-h\ar@{-->}[rr]^-\alpha&& (V, T)
\ar[dl]^-h\\
&Y&
}
$$ such that $h$ is projective,
then
$V'$ is also projective over $Y$.
\end{itemize}
\end{defn}
The main purpose of this section is to establish the
following result, which will play a crucial role in the proof of
Theorem \ref{B-thm1.3} in Section \ref{C-sec5}.
\begin{lem}\label{x-lem4.6}
We use the same notation and assumption as in Theorem \ref{d-thm4.1}.
We assume that $Y$ is a curve.
We further assume that $(V, T+\Supp h^*\Sigma)$ is a simple
normal crossing pair and that all the local monodromies
on the local system $R^jh_*\mathbb Q_{S^*}$ around
$\Sigma$ are unipotent for any stratum $S$ of $(V, T)$ and
all $j$, where $S^*=S|_{V^*}$.
Let $\alpha \colon V\dashrightarrow V$ be a birational map between $(V, T)$
and $(V, T)$ over $Y$.
We assume that
$\alpha$ satisfies condition $(\bigstar)$ in Definition \ref{x-def4.5}.
Then $\alpha$ induces isomorphisms
$$
\alpha^* \colon W_m\Gr^0_F(\mathcal V^k_Y)\overset{\sim}{\longrightarrow}
W_m\Gr^0_F(\mathcal V^k_Y)
$$
for all $m$ and $k$, where $W$ denotes the canonical extension of
the weight filtration.
Let $G$ be a finite group which acts on $(V, T)$ birationally over $Y$
such that every element $\alpha\in G$ satisfies condition $(\bigstar)$
in Definition \ref{x-def4.5}.
Then $G$ acts on $W_m\Gr^0_F(\mathcal V^k_Y)$ for all $m$ and $k$.
\end{lem}
In the proof of Lemma \ref{x-lem4.6}, we will use some arguments and
constructions in \cite[Section 4]{fujino-fujisawa}.
\begin{proof}[Proof of Lemma \ref{x-lem4.6}]
\setcounter{step}{0}
By assumption, $\alpha$ satisfies condition $(\bigstar)$ in Definition \ref{x-def4.5}.
Therefore, we can take a commutative diagram
\begin{equation*}\label{eq-zu7}
\xymatrix{
& (V', T') \ar[dl]_-{p_1}\ar[dr]^-{p_2}& \\
(V, T) \ar@{-->}[rr]_-\alpha&& (V, T)
}
\end{equation*}
as in \eqref{d-eq4.2}. We note that
$V'$ is projective
over $Y$.
From now on, we will use the same notation as in
Definition \ref{x-def4.5}.
We put $u=h\circ j\circ\varepsilon_T \colon T_\bullet \to Y$ and
$v=h\circ \varepsilon _V \colon V_\bullet \to Y$.
We set $E_\bullet =v^{-1}(\Sigma)_{\mathrm{red}}$ and
$F_\bullet=u^{-1}(\Sigma)_{\mathrm{red}}$.
Since $(V, T+\Supp h^*\Sigma)$ is a simple normal crossing
pair by assumption,
$E_\bullet$ and $F_\bullet$ are simple
normal crossing divisors on $V_\bullet$ and $T_\bullet$, respectively.
As in the proof of \cite[Lemma 4.12]{fujino-fujisawa},
we can construct a complex $C(\phi^*)$
on $Y$ equipped with filtrations $W$ and $F$ such that
$H^k(C(\phi^*))\simeq
\mathcal V^k_Y$, where
$\mathcal V^k_Y$ is the canonical extension of
$\mathcal V^k_{Y^*}=R^k(h|_{V^*})_*
\iota_!\mathbb Q_{V^*
\setminus T^*}\otimes \mathcal O_{Y^*}$, for every $k$.
We note that the filtration $W$ is denoted by $L$ in \cite[Lemma 4.12]
{fujino-fujisawa}.
\begin{step}\label{d-step1}
The spectral sequence
$$
E^{p,q}_1(C(\phi^*), F)=H^{p+q}(\Gr^p_FC(\phi^*))\Rightarrow
H^{p+q}(C(\phi^*))
$$
degenerates at $E_1$
(see the proof of \cite[Lemma 4.12]{fujino-fujisawa}
and \cite[13.3]{fujino-slc-trivial}).
Therefore, we have the following short exact sequences
$$
\xymatrix{
0 \ar[r]& \ar[r]^-{s^{p+q}}H^{p+q}(F^1C(\phi^*)) & \ar[r]^-{t^{p+q}}H^{p+q}(C(\phi^*))
& H^{p+q}(\Gr^0_FC(\phi^*)) \ar[r]& 0
}
$$
for all $p$ and $q$.
We note that $F^0C(\phi^*)=C(\phi^*)$ by construction.
Let us consider the following commutative diagram.
$$
\xymatrix{
0 \ar[r]& \ar[r]^-{s^{p+q}}H^{p+q}(F^1C(\phi^*)) & \ar[r]^-{t^{p+q}}H^{p+q}(C(\phi^*))
& H^{p+q}(\Gr^0_FC(\phi^*)) \ar[r]& 0\\
& & H^{p+q} (W_{-p}C(\phi^*))\ar[r]\ar[u]_-{a^{p+q}_{-p}}&
H^{p+q}(W_{-p} \Gr^0_FC(\phi^*))\ar[u]_-{b^{p+q}_{-p}}&
}
$$
By definition,
we have
$$
F^1H^{p+q}(C(\phi^*))=\mathrm{Im}\, s^{p+q}
$$
and
$$
W_q H^{p+q}(C(\phi^*))=\mathrm{Im}\,a^{p+q}_{-p}
$$
for all $p$ and $q$.
We put
\begin{equation}\label{d-eq4.7}
W_qH^{p+q}(\Gr^0_FC(\phi^*)):=\mathrm{Im}\, b^{p+q}_{-p}
\end{equation}
for all $p$ and $q$.
Then the map $t^{p+q}$
induces
\begin{equation}\label{d-eq4.8}
\xymatrix{
\Gr^0_FH^{p+q}(C(\phi^*)) \ar[r]^-{\sim} & H^{p+q} (\Gr^0_FC(\phi^*)) \\
W_q \Gr^0_FH^{p+q} (C(\phi^*)) \ar[r]_-{i^{p+q}_q}\ar@{^{(}->}[u]&
W_q H^{p+q} (\Gr^0_FC(\phi^*)) \ar@{^{(}->}[u]
}
\end{equation}
for all $p$ and $q$.
We will prove that $i^{p+q}_q$ are isomorphisms
for all $p$ and $q$ in Step \ref{d-step2}.
\end{step}
\begin{step}\label{d-step2}
Let us analyse the spectral sequence
\begin{equation}\label{d-eq4.9}
E^{p, q}_1(C(\phi^*), W)\Rightarrow H^{p+q}(C(\phi^*))
\end{equation}
in detail.
Let $\Omega_{V_{p+1}/Y}(\log E_{p+1})$ and
$\Omega_{T_p/Y}(\log F_p)$ be relative logarithmic de Rham complexes
of $v_{p+1} \colon V_{p+1}\to Y$ and $u_p \colon T_p\to Y$, respectively.
Then we have
$$
\left(E^{p, q}_1(C(\phi^*), W), F\right)=
\left(R^q(v_{p+1})_*\Omega_{V_{p+1}/Y}(\log E_{p+1}), F\right)
\oplus \left(R^q(u_p)_*\Omega_{T_p/Y}(\log F_p), F\right)
$$
by construction. We note that the
differentials of the spectral sequence \eqref{d-eq4.9}
are strictly compatible with the filtration induced by $F$
(see \cite[(1.1.5)]{deligne}, \cite[Remark 3.2]{fujino-fujisawa},
and \cite[A.~3.1]{peters-steenbrink})
and that the spectral sequence \eqref{d-eq4.9}
degenerates at $E_2$.
We do not repeat the proof of the above facts here.
For the proof, see the first part of the proof of
\cite[Lemma 4.12]{fujino-fujisawa}
and \cite[13.3]{fujino-slc-trivial}.
The following argument corresponds to the strictness of the filtration $F$
on the $E_0$-term of the spectral sequence
$E^{p,q}_r(C(\phi^*), W)$ (see \cite[13.3]{fujino-slc-trivial}).
By \cite[(2.11) Theorem]{steenbrink},
$R^b(u_p)_*\Omega^a_{T_p/Y}(\log F_p)$ is
locally free for any $a$, $b$, and $p$.
Therefore, the spectral sequence
$$
R^b(u_p)_*\Omega^a_{T_p/Y}(\log F_p)
\Rightarrow
R^{a+b}(u_p)_*\Omega_{T_p/Y}(\log F_p)
$$
degenerates at $E_1$. In particular,
$$
\Gr^0_FR^q(u_p)_*\Omega_{T_p/Y}(\log F_p)\simeq
R^q(u_p)_*\mathcal O_{T_p}
$$
holds for any $p$, $q$.
By the same way, we see
that
$$\Gr^0_FR^q(v_{p+1})_*\Omega_{V_{p+1}/Y}(\log E_{p+1})\simeq
R^q(v_{p+1})_*\mathcal O_{V_{p+1}}
$$
holds for any $p$, $q$.
Thus we have
\begin{equation}\label{d-eq4.10}
\begin{split}
&\Gr^0_FE^{p,q}_1(C(\phi^*), W) \\&=\Gr^0_FR^q(v_{p+1})_*
\Omega_{V_{p+1}/Y}(\log E_{p+1})
\oplus \Gr^0_F R^q(u_p)_*\Omega_{T_p/Y}(\log F_p)\\
&\simeq R^q(v_{p+1})_*\mathcal O_{V_{p+1}}\oplus
R^q(u_p)_*\mathcal O_{T_p}.
\end{split}
\end{equation}
By taking $\Gr^0_F$ of the spectral sequence
\eqref{d-eq4.9},
we obtain the following spectral sequence
\begin{equation*}
E^{p,q}_1(\Gr^0_FC(\phi^*), W)\Rightarrow
H^{p+q}(\Gr^0_FC(\phi^*)).
\end{equation*}
Note that
$$
\Gr^0_FE^{p,q}_1(C(\phi^*), W)\simeq
E^{p,q}_1(\Gr^0_FC(\phi^*), W)
$$
holds as we saw in \eqref{d-eq4.10}. Moreover,
$$
\Gr^0_FE^{p, q}_r(C(\phi^*), W) \simeq E^{p, q}_r(\Gr^0_FC(\phi^*), W)
$$
holds
for every $r\geq 0$
by the lemma on two filtrations
(see \cite[Propositions (7.2.5) and (7.2.8)]{deligne2} and
\cite[Theorem 3.12]{peters-steenbrink}).
Hence, we obtain
\begin{equation}\label{d-eq4.11}
\begin{split}
\Gr^0_F\Gr^W_q\!H^{p+q}(C(\phi^*))&\simeq \Gr^0_FE^{p,q}_2(C(\phi^*), W)
\\ &\simeq E^{p,q}_2(\Gr^0_FC(\phi^*), W)
\simeq
\Gr^W_q\!H^{p+q}(\Gr^0_FC(\phi^*))
\end{split}
\end{equation}
for all $p$ and $q$.
We note that the filtration $W$ on $H^{p+q}(\Gr^0_FC(\phi^*))$
is the one defined in \eqref{d-eq4.7}. We also note that
$\Gr^0_F\Gr^W_q\!H^{p+q}(C(\phi^*))$ is canonically
isomorphic to $\Gr^W_q\!\Gr^0_FH^{p+q}(C(\phi^*))$.
Thus, we can check that
\begin{equation}\label{d-eq4.12}
\xymatrix{
i^{p+q}_q \colon W_q\Gr^0_FH^{p+q}(C(\phi^*)) \ar[r]&
W_qH^{p+q}(\Gr^0_FC(\phi^*))
}
\end{equation}
in \eqref{d-eq4.8} are isomorphisms
for all $p$ and $q$ inductively by using \eqref{d-eq4.8} and \eqref{d-eq4.11} .
\end{step}
\begin{step}\label{d-step3}
In this proof, we did not define the filtration $W$ on
$C(\phi^*)$ explicitly. For the details of
the filtration $W$ on $C(\phi^*)$,
which is denoted by
$L$ in \cite[Section 4]{fujino-fujisawa}, see
(4.2.1) and (4.8.2) in \cite[Section 4]{fujino-fujisawa}.
By construction, we have
\begin{equation*}
\begin{split}
W_{-p} \Gr^0_FC(\phi^*)^n &=
W_{-p-1}(Rv_*\mathcal O_{V_\bullet})^{n+1}\oplus
W_{-p}(Ru_*\mathcal O_{T_\bullet})^n\\
&= \bigoplus_{s\geq p+1} (R(v_s)_*\mathcal O_{V_s})^{n+1-s}
\oplus \bigoplus _{t\geq p} (R(u_t)_*\mathcal O_{T_t})^{n-t}.
\end{split}
\end{equation*}
Therefore, by Lemma \ref{d-lem4.3} and the commutative
diagram \eqref{d-eq4.6} in Definition \ref{x-def4.5},
$\alpha$ induces isomorphisms
$$
\alpha^* \colon W_{-p}\Gr^0_FC(\phi^*)\overset{\sim}{\longrightarrow}
W_{-p}\Gr^0_FC(\phi^*)
$$
for all $p$.
Thus $\alpha$ induces isomorphisms
$$
\alpha^* \colon W_qH^{p+q}(\Gr^0_FC(\phi^*))
\overset{\sim}{\longrightarrow} W_qH^{p+q}(\Gr^0_FC(\phi^*))
$$
for all $p$ and $q$ by the following commutative diagram
\begin{equation*}
\xymatrix{
H^{p+q}(W_{-p}\Gr^0_FC(\phi^*)) \ar[d]^-\wr_-{\alpha^*}\ar[r]
& H^{p+q}(\Gr^0_FC(\phi^*))\ar[d]^-\wr_-{\alpha^*}
\\
H^{p+q}(W_{-p}\Gr^0_FC(\phi^*))\ar[r]&
H^{p+q}(\Gr^0_FC(\phi^*))
}
\end{equation*}
and the definition of the filtration $W$ in \eqref{d-eq4.7}.
Hence, we obtain isomorphisms
$$
\alpha^* \colon W_mH^k(\Gr^0_FC(\phi^*))\overset{\sim}
{\longrightarrow}W_mH^k(\Gr^0_FC(\phi^*))
$$
for all $m$ and $k$ by putting
$p=k-m$ and $q=m$.
By \eqref{d-eq4.12} and the fact that
$\mathcal V^k_Y\simeq H^k(C(\phi^*))$, we obtain the
desired isomorphisms
$$
\alpha^* \colon W_m\Gr^0_F(\mathcal V^k_Y)\overset{\sim}
{\longrightarrow}W_m\Gr^0_F(\mathcal V^k_Y)
$$
for all $m$ and $k$.
\end{step}
When the group $G$ acts on $(V, T)$ birationally over $Y$
such that every element $\alpha\in G$ satisfies
condition $(\bigstar)$ in Definition
\ref{x-def4.5}, it is easy to see that
$G$ also acts on $W_m\Gr^0_F(\mathcal V^{k}_Y)$ for all $m$ and
$k$ by the
above result.
\end{proof}
We make an important remark on dual variations of
mixed Hodge structure. We will use it
in Step \ref{C-step4} in the proof of Theorem \ref{B-thm1.3}.
\begin{rem}
[{see \cite[Remarks 3.15 and 7.4]{fujino-fujisawa}}]\label{x-rem4.7}
We use the same notation and assumption as in Lemma \ref{x-lem4.6}.
Let us consider the dual local system of $R^k(h|_{V^*})_*\iota_!\mathbb Q_{
V^*\setminus T^*}$ and
the dual variation of mixed Hodge structure on it.
Then the locally free sheaf $(\mathcal V^k_{Y^*})^*$ carries
the Hodge filtration $F$ and the weight filtration $W$ defined
as in \cite[Remark 3.15]{fujino-fujisawa}.
By the construction of the Hodge filtration $F$,
$$
\Gr^0_F\!\left((\mathcal V^k_Y)^*\right)\simeq
\left(\Gr^0_F(\mathcal V^k_Y)\right)^*
$$
holds, where $(\mathcal V^k_Y)^*$ is
the canonical extension of $(\mathcal V^k_{Y^*})^*$.
More generally,
$$
\Gr^{-p}_F\!\left((\mathcal V^k_Y)^*\right)\simeq
\left(\Gr^p_F(\mathcal V^k_Y)\right)^*
$$
holds for every $p$.
We note that
$\Gr^0_F\!\left((\mathcal V^k_Y)^*\right)=F^0\!\left((\mathcal V^k_Y)^*\right)$,
the canonical extension of the lowest piece of the Hodge filtration.
By taking the dual of Lemma \ref{x-lem4.6},
$G$ acts on $W_m\Gr^0_F\!\left((\mathcal V^k_Y)^*\right)$ for
every $m$,
where $W$ denotes the canonical extension of
the weight filtration of $(\mathcal V^k_{Y^*})^*$.
We note that we have
$$
\Gr^W_m\Gr^p_F\!\left((\mathcal V^k_Y)^*\right)
\simeq \left( \Gr^W_{-m}\Gr^{-p}_F(\mathcal V^k_Y)\right)^*
$$
for all $p$ and $m$ by construction.
\end{rem}
We close this section with the following lemma, which is
more or less well known to the experts (see \cite{zucker},
\cite{peters}, \cite{kollar}, and \cite{fujino-fujisawa2}).
We will use it in the proof of Theorem \ref{B-thm1.3} in Section \ref{C-sec5}.
\begin{lem}\label{d-lem4.7}
Let $C$ be a smooth projective curve
and let $C_0$ be a non-empty Zariski open set of $C$.
Let $V_0$ be a polarizable variation of $\mathbb Q$-Hodge
structure over $C_0$ with unipotent monodromies around $\Sigma=C\setminus
C_0$. Let $F^b$ be the canonical extension of the lowest piece of
the Hodge filtration.
Let $\mathcal L$ be a line bundle on $C$ which is a direct summand of $F^b$.
Assume that $\deg_C\mathcal L=0$.
Then $\mathcal L|_{C_0}$ is a flat subbundle of $F^b|_{C^0}$.
\end{lem}
\begin{proof}
Let $h_0$ be the smooth hermitian metric on $\mathcal L|_{C_0}$ induced
by the Hodge metric of $F^b|_{C_0}$.
Then $\frac{\sqrt{-1}}{2\pi} \Theta_{h_0}(\mathcal L|_{C_0})$ is a semipositive
smooth $(1, 1)$-form on $C_0$.
We note that $\Theta_{h_0}(\mathcal L|_{C_0})$ is the curvature tensor of the Chern
connection of $(\mathcal L|_{C_0}, h_0)$.
Then $$\deg_C\mathcal L=\frac{\sqrt{-1}}{2\pi}\int _{C_0}
\Theta_{h_0}(\mathcal L|_{C_0})$$ holds (see, for example, \cite[Theorem 5.1]{kollar}).
Note that the right hand side is an improper integral.
By assumption, $\deg _C\mathcal L=0$.
This implies that $\Theta_{h_0}(\mathcal L|_{C_0})=0$.
Therefore, $\mathcal L|_{C_0}$ is a flat subbundle of $F^b|_{C_0}$.
\end{proof}
\begin{rem}\label{d-rem4.8}
In Lemma \ref{d-lem4.7},
the smooth hermitian metric $h_0$ on $\mathcal L|_{C_0}$ can be
extended naturally to a singular hermitian metric $h$
on $\mathcal L$ in the sense of Demailly such that
$\sqrt{-1}\Theta_h(\mathcal L)$ is positive in the sense of
currents and that the Lelong
numbers of $h$ are zero everywhere.
For the details, see \cite[Theorem 1.1]{fujino-fujisawa2}.
\end{rem}
\section{Proof of Theorem \ref{B-thm1.3}}\label{C-sec5}
In this section, we prove Theorem \ref{B-thm1.3} and
Corollary \ref{B-cor1.4}.
\medskip
Let us prepare an easy lemma.
By this lemma, we can reduce the problem to the case where the base
space is a curve.
\begin{lem}\label{C-lem5.1}
Let $Y$ be a smooth projective irreducible variety with $\dim Y\geq 2$ and
let $N$ be a numerically trivial Cartier divisor on $Y$.
Let $H$ be a smooth ample Cartier divisor on $Y$ such
that $H$ contains no irreducible components of $\Supp N$.
Then $N\sim 0$ if and only if $N|_H\sim 0$.
\end{lem}
\begin{proof}
We consider the following long exact sequence
\begin{equation*}
\begin{split}
0&\to H^0(Y, \mathcal O_X(N-H))\to
H^0(Y, \mathcal O_Y(N))\to H^0(H, \mathcal O_H(N|_H))
\\&\to H^1(Y, \mathcal O_Y(N-H))\to \cdots.
\end{split}
\end{equation*}
It is obvious that $H^0(Y, \mathcal O_Y(N-H))=0$.
By the Kodaira vanishing theorem,
we have $H^1(Y, \mathcal O_Y(N-H))=0$.
Therefore,
$H^0(Y, \mathcal O_Y(N))\simeq
H^0(H, \mathcal O_H(N|_H))$ holds.
In particular, $N\sim 0$ if and only if $N|_H\sim 0$.
\end{proof}
Let us start the proof of Theorem \ref{B-thm1.3}.
We adapt Floris's proof of
Theorem \ref{B-thm1.3} for lc-trivial
fibrations (see \cite{floris}) to our setting, that is,
basic slc-trivial fibrations.
\begin{proof}[Proof of Theorem \ref{B-thm1.3}]
This proof heavily depends on \cite[Section 6]{fujino-slc-trivial}.
Let $\sigma\colon Y'\to Y$ be a projective
birational morphism from a smooth
projective variety $Y'$. By considering
the induced basic slc-trivial fibration by $\sigma\colon Y'\to Y$,
we may assume that $Y$ is a smooth projective variety.
\setcounter{step}{0}
\begin{step}\label{C-step1}
In this step, we construct a cyclic cover of the generic
fiber of $f:X\to Y$ following \cite[6.1 and 6.2]{fujino-slc-trivial}.
Let $f \colon (X, B)\to Y$ be a basic slc-trivial fibration.
Let $F$ be a general fiber of $f \colon X\to Y$.
We put
$$
b=\min\{m \in \mathbb Z_{>0}\, |\, m(K_F+B_F)=m(K_X+B)|_F\sim 0\}.
$$
Then we can write
\begin{equation}\label{C-eq5.1}
K_X+B+\frac{1}{b}(\varphi)=f^*(K_Y+B_Y+M_Y)
\end{equation}
with $\varphi\in \Gamma(X, \mathcal K^*_X)$,
where $B_Y$ is the discriminant $\mathbb Q$-divisor
and $M_Y$ is the moduli $\mathbb Q$-divisor
of $f \colon (X, B)\to Y$.
By taking some suitable blow-ups
(see \cite[Theorem 1.4 and Section 8]{bierstone} and
\cite[Lemma 2.11]{fujino-ann}),
we may assume that $\Supp (B-f^*(B_Y+M_Y))$ is a simple
normal crossing divisor on $X$,
$(B^h)^{=1}$ is Cartier,
and every stratum of $(X, (B^h)^{=1})$ is dominant
onto $Y$.
We take the $b$-fold cyclic cover $\pi \colon \widetilde X\to X$ associated
to \eqref{C-eq5.1}, that is,
$$
\widetilde X=\Spec _X \bigoplus _{i=0}^{b-1}
\mathcal O_X(\lfloor i\Delta\rfloor),
$$
where $\Delta=K_{X/Y}+B-f^*(B_Y+M_Y)$.
We note that $\pi:\widetilde X\to X$ is a finite Galois
cover by construction (see \cite[Proposition 6.3 (i)]{fujino-slc-trivial}).
We put $K_{\widetilde X}+B_{\widetilde X}=\pi^*(K_X+B)$.
By construction, it is easy to see that
$(B^h_{\widetilde X})^{=1}=\pi^*((B^h)^{=1})$ and
that $(\widetilde X, (B^h_{\widetilde X})^{=1})$ is
semi-log canonical.
Moreover, every slc stratum of $(\widetilde X, (B^h_{\widetilde X})^{=1})$
is dominant onto $Y$.
We take a projective
birational morphism $d \colon V\to \widetilde X$
from a simple normal crossing variety $V$ such that
$d$ is an isomorphism
over the generic point of every slc stratum of $(\widetilde X,
(B^h_{\widetilde X})^{=1})$ by \cite[Theorem 1.4]{bierstone}. We
put $K_V+B_V=d^*(K_{\widetilde X}+B_{\widetilde X})$.
Then we
get the following commutative diagram
\begin{equation}\label{C-eq5.2}
\xymatrix{
(X, B)\ar[d]_-f & \widetilde X \ar[dl]_-{\widetilde f}
\ar[l]_-\pi& (V, B_V)\ar[dll]^-h\ar[l]_-d\\
Y & &
}
\end{equation}
with $g=\pi\circ d$.
By taking a suitable birational modification of $Y$ and
considering induced (pre-)basic slc-trivial fibrations as in
\cite[6.2]{fujino-slc-trivial},
we may further assume that the following properties hold for
$$
K_X+B+\frac{1}{b}(\varphi)=f^*(K_Y+B_Y+M_Y)
$$
and
$$
h \colon (V, B_V)
\overset{g}{\longrightarrow} (X, B)\overset{f}{\longrightarrow} Y.
$$
\begin{itemize}
\item[(a)] $Y$ is a smooth projective
irreducible variety, and $X$ and $V$ are projective
simple
normal crossing varieties.
\item[(b)] There exist simple normal crossing divisors
$\Sigma_X$, $\Sigma_V$, and $\Sigma_Y$
on $X$, $V$, and $Y$, respectively.
\item[(c)] $f$ and $h$ are projective surjective
morphisms.
\item[(d)] The supports of
$B$, $B_V$, and $B_Y$, $M_Y$ are
contained in
$\Sigma_X$, $\Sigma_V$, and $\Sigma_Y$, respectively.
\item[(e)] Every stratum of $(X, \Sigma^h_X)$
and $(V, \Sigma^h_V)$ is smooth
over $Y\setminus \Sigma_Y$.
\item[(f)] $f^{-1}(\Sigma_Y)\subset \Sigma_X$, $f(\Sigma^v_X)\subset
\Sigma_Y$, and $h^{-1}(\Sigma_Y)\subset
\Sigma_V$, $h(\Sigma^v_V)\subset
\Sigma_Y$.
\item[(g)] $(B^h)^{=1}$ and $(B^h_V)^{=1}$ are Cartier.
\end{itemize}
We note that conditions (a)--(g) above are nothing but the conditions
stated just before \cite[Proposition 6.3]{fujino-slc-trivial}.
As we saw in the proof of \cite[Theorem 1.2]{fujino-slc-trivial}
(see \cite[Section 9]{fujino-slc-trivial}),
$\mathbf M=\overline{\mathbf M_Y}$ holds and
$\mathbf M_Y$ is a nef $\mathbb Q$-divisor on $Y$.
By assumption, $\mathbf M_Y\equiv 0$.
If $\nu\colon Y''\to Y$ is a finite surjective morphism from a
smooth
projective irreducible variety $Y''$,
then it is easy to see that $\mathbf M_Y\sim _{\mathbb Q}0$ if and
only if $\nu^*\mathbf M_Y\sim _{\mathbb Q}0$.
Therefore, by taking a unipotent reduction
(see \cite[Lemma 7.3]{fujino-slc-trivial}),
we may further assume that
\begin{itemize}
\item[(A)] for any irreducible component $P$ of
$\Supp \Sigma_Y$, there exists a prime divisor $Q$ on $V$ such that
$\mult _Q(-B_V+h^*B_Y)=0$,
$h(Q)=P$, and
$\mult _Qh^*P=1$,
\item[(B)] all the local monodromies on the local system
$$R^{\dim V-\dim Y} (h|_{V^*})_*\iota_!\mathbb Q_{V^*\setminus
(B_{V^*}^h)^{=1}}$$ around
$\Sigma_Y$ are unipotent, where
$Y^*=Y\setminus \Sigma_Y$, $V^*=h^{-1}(Y^*)$, $B_{V^*}=(B_V)|_{V^*}$,
and $\iota\colon V^*\setminus (B_{V^*}^h)^{=1}\hookrightarrow V^*$ is the natural
open immersion, and
\item[(C)] all the local monodromies on the local system
$R^kh_*\mathbb Q_{S^*}$ around $\Sigma_Y$ are unipotent for any stratum $S$
of $(V, (B^h_V)^{=1})$ and every $k$, where $S^*=S|_{V^*}$.
\end{itemize}
Note that the above assumptions (A) and (B) are nothing but
the assumptions in (iv) and (v) in \cite[Proposition 6.3]{fujino-slc-trivial}.
We also note that
we do not treat the assumption (C) in the original statement of
\cite[Lemma 7.3]{fujino-slc-trivial}.
Therefore, we have to make $N_j$ in the proof of
\cite[Lemma 7.3]{fujino-slc-trivial} sufficiently divisible
in order to make the monodromy on the local system
$R^kh_*\mathbb Q_{S^*}$ around
$P_j$, an irreducible component of $\Sigma_Y$, unipotent for any
stratum $S$ of $(V, (B^h_V)^{=1})$ and
every $k$ when we take a finite cover $\nu \colon Y''\to Y$ for a unipotent
reduction (see \cite[Lemma 7.3]{fujino-slc-trivial}).
\end{step}
\begin{step}\label{C-step2}
We assume that $\dim Y\geq 2$.
Then we take a general ample Cartier divisor
$H$ on $Y$ and put $Z=f^*H$ and
$W=h^*H$.
In this situation,
$$
K_X+Z+B+\frac{1}{b}(\varphi)=f^*(K_Y+H+B_Y+M_Y).
$$
By adjunction,
$$
K_Z+B|_Z+\frac{1}{b}(\varphi|_Z)=f^*(K_H+B_Y|_H+M_Y|_H)
$$
holds.
It is not difficult to see that
$f|_Z \colon (Z, B|_Z)\to H$ is a basic slc-trivial fibration
and $$h|_W \colon (W, B_V|_W)\overset {g|_W}
{\longrightarrow} (Z, B|_Z)\overset{f|_Z}{\longrightarrow}
H$$ satisfies conditions (a)--(g),
(A), (B), and (C) in Step \ref{C-step1}.
We note that $B_Y|_H=B_H$ and $M_Y|_H=M_H$ hold,
where $B_H$ (resp.~$M_H$) is
the discriminant (resp.~moduli) $\mathbb Q$-divisor
of $f|_Z \colon (Z, B|_Z)\to H$.
By Lemma \ref{C-lem5.1},
$M_Y\sim _{\mathbb Q}0$ if and only if $M_Y|_H\sim
_{\mathbb Q}0$.
Therefore, we can replace $f \colon (X, B)\to Y$ with $f|_Z \colon
(Z, B|_Z)\to
H$.
By repeating this reduction finitely many times,
we may assume that $Y$ is a smooth projective
curve.
\end{step}
\begin{step}\label{C-step3}
In Step \ref{C-step1},
we have already seen that $\pi:\widetilde X\to X$ is Galois.
Let $G=\mathbb Z/b\mathbb Z$ be the Galois group of
the $b$-fold cyclic cover
$\pi \colon \widetilde X\to X$.
The action of $G$ on $\widetilde X$ preserves the slc strata
of $(\widetilde X, (B^h_{\widetilde X})^{=1})$ by construction.
Therefore,
any element $\alpha$ of $G$ induces a birational map between
$(V, T)$ and $(V, T)$ over $X$, where
$T=(B^h_V)^{=1}$.
From now on, we will check that $\alpha$ satisfies
condition $(\bigstar)$ in Definition \ref{x-def4.5}.
As usual, we can take a commutative diagram
\begin{equation*}
\xymatrix{
& (V', T') \ar[dd]^(.35){g'}\ar[dl]_-{p_1}\ar[dr]^-{p_2}& \\
(V, T) \ar[dr]_-g\ar[ddr]_-h\ar@{-->}[rr]^(.40){\alpha}&& (V, T)\ar[dl]^-g
\ar[ddl]^-h\\
& X\ar[d]^-f& \\
& Y&
}
\end{equation*}
by using \cite[Theorem 1.4]{bierstone}, where $(V', T')$ is a simple
normal crossing pair such that
$T'$ is reduced, and $p_i$ is a projective
birational morphism between $(V', T')$ and $(V, T)$ for
$i=1, 2$.
We put $C=(B^h)^{=1}$.
The irreducible decomposition of $X$ and $C$ are given by
$$
X=\bigcup _{i\in I} X_i, \quad\text{and}
\quad C=\bigcup _{\lambda\in \Lambda} C_\lambda
$$
respectively as in \cite[4.14]{fujino-fujisawa}.
We put $V=\bigcup _{i\in I}V_i$ and $V_i=\bigcup _j V_{i_j}$,
where $V_{i_j}$ runs over irreducible components
of $V$ such that
$g(V_{i_j})=X_i$.
We put $T=\bigcup _{\lambda\in \Lambda} T_\lambda$
and $T_\lambda=\bigcup _l T_{\lambda_l}$,
where $T_{\lambda_l}$ runs over irreducible components of $T$
such that $g(T_{\lambda_l})=C_\lambda$.
Note that $T_{\lambda}$ and $V_i$ are disjoint unions
of some strata of $(V, T)$.
By applying the same construction as above to
$(V', T')$ and $g':=g\circ p_1=g\circ p_2 \colon V'\to X$,
we get
$V'=\bigcup _{i\in I} V'_i$ and $T'=\bigcup _{\lambda\in \Lambda}
T'_\lambda$.
We apply the same construction as in \cite[4.14]{fujino-fujisawa}
to $V=\bigcup _{i\in I}V_i$ and $T=\bigcup _{\lambda\in \Lambda}
T_\lambda$
(resp.~$V'=\bigcup _{i \in I} V'_i$ and $T'=\bigcup _{\lambda\in \Lambda}
T'_\lambda$) instead of
$X=\bigcup _{i\in I}X_i$ and $D=\bigcup _{\lambda\in \Lambda}
D_\lambda$ in \cite[4.14]{fujino-fujisawa}.
Then we can construct semi-simplicial
resolutions $\varepsilon _T \colon T_\bullet \to T$ and
$\varepsilon _V \colon V_\bullet \to V$
(resp.~$\varepsilon _{T'} \colon T'_\bullet \to T'$ and $\varepsilon _{V'}
\colon V'_\bullet \to V'$).
By construction, these semi-simplicial
resolutions satisfy the conditions stated in Definition \ref{x-def4.5}.
Therefore, $\alpha$ satisfies condition $(\bigstar)$.
This is what we wanted.
\end{step}
\begin{step}\label{C-step4}
We note that $M_Y$ is a Cartier divisor on $Y$ and
that $\mathcal O_Y(M_Y)$ is a direct summand
of
$$\left(\Gr^0_F(\mathcal V^d_Y)\right)^*\simeq
\Gr^0_F\!\left((\mathcal V^d_Y)^*\right),
$$ where
$d=\dim X-\dim Y$ (see \cite[Proposition 6.3]{fujino-slc-trivial}).
More precisely, by construction,
$\mathcal O_Y(M_Y)$ is an eigensheaf of rank one corresponding
to the eigenvalue
$\zeta^{-1}$ of
$$h_*\omega_{V/Y}\left((B^h_V)^{=1}\right)\simeq
\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)$$
by the group action of $G=\mathbb Z/b\mathbb Z$,
where $\zeta$ is a fixed primitive $b$-th root of unity
(see the proof of \cite[Proposition 6.3]{fujino-slc-trivial}).
We take an integer $l$ such that
\begin{equation*}
\mathcal O_Y(M_Y)\subset W_l\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)
\ \
\text{and} \quad
\mathcal O_Y(M_Y)\not \subset W_{l-1}\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)
\end{equation*}
hold.
Thus we can easily see that
$\mathcal O_Y(M_Y)$ is an eigensheaf of rank one corresponding
to the eigenvalue $\zeta^{-1}$ of $W_l\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)$
and that $\mathcal O_Y(M_Y)\cap W_{l-1}\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)=\{0\}$
in $W_l\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)$. We note that
$G$ acts on $W_m\Gr^0_F\!\left((\mathcal V^d_Y)^*\right)$ for every $m$ by
Lemma \ref{x-lem4.6} and Remark \ref{x-rem4.7}.
Since $\deg M_Y=0$ by assumption,
$\mathcal O_Y(M_Y)|_{Y^*}$ defines a local subsystem of
$\Gr^W_l\!\left((\mathcal V^d_{Y^*})^*\right)$
by Lemma \ref{d-lem4.7}.
We note that
$$\Gr^W_l\Gr^0_F\!\left((\mathcal V^d_{Y^*})^*\right)\simeq
\Gr^0_F\Gr^W_l\!\left((\mathcal V^d_{Y^*})^*\right)
=F^0\Gr^W_l\!\left((\mathcal V^d_{Y^*})^*\right)\subset
\Gr^W_l\!\left((\mathcal V^d_{Y^*})^*\right)$$
holds since we have
$F^1\Gr^W_l\!\left((\mathcal V^d_{Y^*})^*\right)=0$ by
the construction of the dual Hodge filtration (see \cite[Remark 3.15]
{fujino-fujisawa} and Remark \ref{x-rem4.7}).
Therefore, there exists a positive integer $a$
such that $\mathcal O_Y(aM_Y)|_{Y^*}\simeq
\mathcal O_{Y^*}$ by \cite[Corollaire (4.2.8) (iii) b)]{deligne}.
This is because $\Gr^W_l\!\left((\mathcal V^d_{Y^*})^*\right)$ is a polarizable
variation of $\mathbb Q$-Hodge structure.
Thus we get $\mathcal O_Y(aM_Y)\simeq
\mathcal O_Y$ by taking the canonical extension.
This is what we wanted.
\end{step}
Hence, we obtain $\mathbf M_{Y'}\sim _{\mathbb Q}0$.
\end{proof}
We close this section with the proof of
Corollary \ref{B-cor1.4}.
\begin{proof}[Proof of Corollary \ref{B-cor1.4}]
By \cite[Lemma 4.12]{fujino-slc-trivial},
we may assume that $Y$ is a smooth projective
curve. We always have $\deg M_Y\geq 0$ since $M_Y$
is nef by \cite[Theorem 1.2]{fujino-slc-trivial}.
If $\deg M_Y>0$,
then it is obvious that
$M_Y$ is ample.
If $\deg M_Y=0$, then
$M_Y$ is numerically trivial.
In this case, by Theorem \ref{B-thm1.3},
$M_Y\sim _{\mathbb Q}0$ holds.
Therefore, we see that $M_Y$ is always
semi-ample.
\end{proof}
|
\section*{Acknowledgments}
\hspace*{1em}
\begin{acknowledgments}
We thank V. Braun, J.-W. Chen, M. G\"ockeler, Y.-S. Liu, Y.-B. Yang, F. Yuan, Y. Zhao and R.-L. Zhu for helpful discussions.
This work was partially supported by the U.S. Department of Energy Office of Science, Office of Nuclear Physics under Award Number DE-FG02-93ER-40762 and DE-SC0011090 (XJ), and also within the framework of the TMD Topical Collaboration (XJ and AS). XJ acknowledges support from the Alexander von Humboldt Foundation. JHZ and AS are supported by the SFB/TRR-55 grant ``Hadron Physics from Lattice QCD". This work is also supported in part by National Natural
Science Foundation of China under Grant
No.11575110, 11655002, 11735010, by Natural Science Foundation of Shanghai under Grant No.~15DZ2272100.
\end{acknowledgments}
|
\section{Introduction}
A fundamental problem in crystallography is to understand why atoms or molecules typically arrange themselves into crystalline order at low temperature. The challenge to prove rigorously that global minimizers of certain configurational energies arrange in a periodic lattice is referred to as the \emph{crystallization problem} \cite{Blanc}. Starting with seminal works in one dimension \cite{Gardner,Hamrick,Radi,Ventevogel}, the last decades have witnessed a remarkable interest in this mathematical problem for systems consisting of \emph{identical} particles.
At zero or very low temperature, atomic interactions are expected to be governed solely by the geometry of the atomic arrangement. Identifying configurations with the respective positions of identical atoms $\lbrace x_1, \ldots, x_n \rbrace$, one is concerned with the minimization of a configurational energy $\mathcal{E}( \lbrace x_1,\ldots,x_n \rbrace )$ which comprises classical interaction potentials. Crystallization then consists in proving that any ground state of $\mathcal{E}$ is a subset of a periodic lattice. \EEE
Despite of its paramount theoretical and
applicative relevance, rigorous results on crystallization are scarce and even under simplified assumptions the problem presents many difficulties. In this paper, we contribute to these fundamental mathematical questions by investigating crystallization of particle systems consisting of \emph{two different atomic types}, also called \emph{dimers}.
It is well known that the geometry of molecular compounds often differs from that of their
components. An example in this direction is sodium chloride (NaCl) which has a FCC structure (Face Centered Cubic), whereas
sodium crystals are BCC (Body Centered Cubic) and chlorine crystals are orthorhombic. The goal of this contribution is to illustrate such a phenomenon in a simplified setting by deriving a two-dimensional crystallization result in the hexagonal lattice for dimers whose components instead would likely assemble themselves in a triangular lattice.
Similar to the problem for systems of identical particles, we follow the classical molecular-mechanical frame of configurational energy minimization. We identify configurations of $n$ particles with their respective \emph{positions} $\{x_1,\ldots,x_n\} \subset \mathbb{R}^{2} \EEE$ and additionally with their \emph{types} $\{q_1,\ldots,q_n\} \in \{-1,1\}^n\EEE$. Our goal is to determine global minimizers of a corresponding interaction energy $ \mathcal{E}(\{(x_1,q_1),\ldots,(x_n,q_n)\})$.
More specifically, the energy $\mathcal{E} = \mathcal{E}_{\mathrm{a}} + \mathcal{E}_{\mathrm{r}}$ is assumed to consist of two short-ranged two-body interaction potentials $\mathcal{E}_{\mathrm{a}}$ and $\mathcal{E}_{\mathrm{r}}$, where $\mathcal{E}_{\mathrm{a}}$ represents the interactions between atoms of different type and $\mathcal{E}_{\mathrm{r}}$ encodes the energy contributions of atoms having the same type. The potential $\mathcal{E}_{\mathrm{a}}$ is attractive-repulsive and favors atoms sitting at some specific reference distance, whereas $\mathcal{E}_{\mathrm{r}}$ is a pure repulsive term. The latter is supposed to be strong enough at short distances such that the hexagonal lattice, with the
two atomic types alternating, is energetically preferred with respect to closer packed structures. (Note that weaker short-ranged repulsion may indeed favor crystallization in the square lattice, see \cite{FriedrichKreutzSquare}.)
An example of a dimer satisfying qualitatively the assumptions on the attractive and repulsive potentials is graphitic \EEE boron nitride \cite{DresselhausM, Wang}, which, similar to graphene \cite{Geim}, is a one-atom thick layer of atoms arranged in a hexagonal lattice. Moreover, our choice of the interaction potentials is motivated by the modeling of ions in ionic compounds. Indeed, the two interaction energies may be interpreted as (simplified) Coulomb-interactions between ions of equal and opposite charge. In this spirit, we will frequently refer to the particles as \emph{ions} and to the atomic types $\{q_1,\ldots,q_n\} \in \{-1,1\}^n$ as positive or negative \emph{charges}.
The paper contains the following three main results:
\begin{itemize}
\item[(1)] Theorem \ref{TheoremEnergyGroundstates}: Under suitable assumptions on the attractive and the repulsive potentials, we characterize the ground-state energy of $n$-particle configurations of ions in the plane. In particular, we quantify exactly the surface energy related to atoms at the boundary of the ensemble. Moreover, we show that in our modeling frame ground states are repulsion free, i.e., $\mathcal{E}_{\rm r}(\{(x_1,q_1),\ldots,(x_n,q_n)\}) = 0$.
\smallskip
\item[(2)] Theorem \ref{TheoremGroundstatesleq31}: We characterize the geometry of ground states. We show that each global minimizer of the configurational energy is essentially a connected subset of the regular hexagonal lattice with the two atomic types alternating. This characterization holds except for possibly one defect occurring at the boundary which can be identified explicitly either as a single atom or as a simple octagonal cycle in the structure.
\smallskip
\item[(3)] Theorem \ref{TheoremCharge}: We consider the \emph{net charge} defined by $\mathcal{Q}(\{(x_1,q_1),\ldots,(x_n,q_n)\}) = \sum_{i=1}^n q_i$ or, in other words, the (signed) difference of the number of the two atomic types. We provide a fine asymptotic characterization of the net charge as the number of particles grows. More precisely, we show that its fluctuation can be at most of order ${\rm O}(n^{1/4})$, i.e., $|\mathcal{Q}(\{(x_1,q_1),\ldots,(x_n,q_n)\})| \leq c n^{1/4}$ for some constant $c>0$ independent of $n$. By giving an explicit construction we further prove that the scaling is sharp.
\end{itemize}
To the best of our knowledge, the present paper provides a first rigorous mathematical crystallization result for two-dimensional molecular compounds. Let us mention that in case of different types of particles, simulations are abundant, but rigorous results seem to be limited to \cite{Betermin,B43,B195}. We also refer to \cite{Betermin2} for a recent study about optimal charge distributions on Bravais lattices. \EEE Our analysis is largely inspired by many contributions on identical particle systems. We include here only a minimal crystallization literature overview and refer the reader to the recent review \cite{Blanc} for a more general perspective.
Mathematical results on crystallization in one dimension, proving or disproving
periodicity of ground-state configurations for different choices of
the energy, are by now quite classical, see, e.g., \cite{Gardner,Hamrick,Radi,Ventevogel}. In two dimensions for a finite number of identical particles, ground states have been proved to be
subsets of the triangular lattice by {\sc Heitman, Radin, and Wagner} \cite{HR,Radin,Wagner83} for sticky and soft, purely two-body interaction energies. Recently, a new perspective on this problem was given by {\sc De Luca and Friesecke} \cite{Lucia} using a discrete Gauss-Bonnet method from discrete differential geometry. Including angular potentials favoring $2\pi/3$ bond-angles, which is typically the case for graphene, crystallization in the hexagonal lattice has been proved by {\sc Mainini and Stefanelli} \cite{Mainini}, see also \cite{emergence} for a related classification of ground states. In a similar fashion, if $\pi/2$ bond-angles are favored, the
square lattice can be recovered \cite{Mainini-Piovano}. Under less restrictive assumptions on the potentials, various results have been obtained in the thermodynamic limit \cite{ELi, Smereka15}, namely as the number of
particles $n$ tends to infinity, most notably the seminal paper by {\sc Theil} \cite{Theil}. The crystallization problem in three dimension appears to be very difficult and only few rigorous results \cite{Flateley1,Flateley2, cronut, Suto06} are currently available.
In two dimensions, a fine characterization of ground-state geometries is possible: as $n$ increases, one observes the emergence of a polygonal cluster, the {\it Wulff shape} \cite{Yuen}. For the triangular \cite{Schmidt-disk},
the hexagonal \cite{Davoli15}, and the square lattice \cite{Mainini-Piovano}, ground states differ from
the Wulff shape by at most ${\rm O}(n^{3/4})$ particles and at most by ${\rm O}(n^{1/4})$ in Hausdorff distance. This sharp bound is called the $n^{3/4}$ law \cite{Davoli16}.
Our general proof strategy follows the induction method on bond-graph layers developed in \cite{HR, Mainini-Piovano, Mainini, Radin}. In particular, our study is related to \cite{Mainini}, where crystallization for identical particles in the hexagonal lattice has been investigated under three-body angular potentials. Actually, the ground-state energy in the present context coincides with the one obtained there.
Concerning the characterization of ground-state geometries, however, richer geometric structures with respect to the ones of \cite{Mainini} are possible since configurations are in general less rigid without angular potentials: simple examples show that ground states may contain
distortions and defects at the boundary such as octagons. From a technical point of view, novel geometric arguments in the
induction step are necessary, complementing the available approaches, see Remark \ref{rem: method} for more details. Let us mention that, in any case, certain quantitative requirements on the potentials are indispensable since other assumptions might favor an assemblence of the atoms in other periodic structures. (In particular, we refer to the forthcoming \EEE paper \cite{FriedrichKreutzSquare} for the square-lattice case.)
For the characterization of the net charge in terms of the sharp scaling ${\rm O}(n^{1/4})$, we use the fact that in ground states the two atomic types are alternately arranged in the hexagonal lattice. Morever, the argument fundamentally relies on the $n^{3/4}$-law \cite{Davoli15} which states that ground states differ from
the Wulff shape by at most ${\rm O}(n^{3/4})$ particles, or equivalently and more relevant in our context, by ${\rm O}(n^{1/4})$ in Hausdorff distance. Let us emphasize that in our model the net charge is not a priori given but a crucial part of the minimization problem. The investigation of ground-state geometries under preassigned net charge is a different challenging problem which we defer to future studies.
The article is organized as follows. In Section 2 we introduce the precise mathematical setting and state the main results. In Section 3 we discuss elementary geometric properties of ground states. In Section 4 we construct explicitly some configurations in order to provide sharp upper bounds for the ground-state energy and the (signed) net charge. In Section 5 we introduce the concept of boundary energy and prove corresponding estimates which are instrumental for the induction method used in the proof of Theorem \ref{TheoremEnergyGroundstates} and \ref{TheoremGroundstatesleq31}. In Section 6 we give a lower bound for the ground-state energy matching the one of the configurations constructed in Section 4. In Section 7 we characterize the geometry of ground states and finally Section 8 is devoted to the proof of a $n^{1/4}$-law for the net charge of ground states.
\section{Setting and main results}
In this section we introduce our model, give some basic definitions, and present our main results.
\subsection{Configurations and interaction energy}
We consider two-dimensional particle systems consisting of two different atomic types and model their interaction by classical potentials in the frame of Molecular Mechanics \cite{Molecular, Friesecke-Theil15,Lewars}.
Let $n \in \mathbb{N}$. We indicate the \textit{configuration} of $n$ particles by
$$C_n = \{ (x_1,q_1),\ldots,(x_{n},q_{n}) \} \subset (\mathbb{R}^2 \times \lbrace -1,1 \rbrace)^n,$$
identified with the respective \emph{atom positions} $X_n = (x_1,\ldots,x_n) \in \mathbb{R}^{2n}$ together with their \emph{types} $Q_n = ( q_1,\ldots,q_n ) \in \lbrace -1,1\rbrace^n$. Referring to a model for ionic compounds, we will frequently call $Q_n$ the \emph{charges} of the atoms, $q = 1$ representing \emph{cations} and $q=-1$ representing \emph{anions}. Indeed, our choice of the empirical potentials (see below) is inspired by ions in ionic compounds which are primarily held together by the electrostatic forces between the net negative charge of the anions and net positive charge of the cations \cite{Pauling}.
We \EEE define the phenomenological energy $\mathcal{E}: (\mathbb{R}^2 \times \lbrace -1,1 \rbrace)^{n}$ $ \to \overline{\mathbb{R}}$ of a given configuration $\lbrace (x_1,q_1),\ldots,(x_n,q_n) \rbrace \in (\mathbb{R}^2 \times \lbrace -1,1 \rbrace)^{n}$ by
\begin{align}\label{Energy}
\mathcal{E}(C_n) = \frac{1}{2}\sum_{\underset{q_i = q_j}{i \neq j}} V_{\mathrm{r}}(|x_i-x_j|) + \frac{1}{2}\sum_{\underset{q_i \neq q_j}{i \neq j}} V_{\mathrm{a}}(|x_i-x_j|),
\end{align}
where $V_{\mathrm{r}}, V_{\mathrm{a}} : [0,+\infty)\to \overline{\mathbb{R}}$ are a \emph{repulsive potential} and an \emph{attractive-repulsive potential}, respectively. The factor $\frac{1}{2}$ accounts for the fact that every contribution is counted twice in the sum. The two potentials are pictured schematically in Fig.~\ref{FigurePotentials}. Let $r_0 \in [1,(2\sin(\frac{\pi}{7}))^{-1})$ and note that $r_0 \le \sqrt{2}$. The attractive-repulsive potential $V_{\mathrm{a}}$ satisfies
\begin{align*}
{\rm [i]}& \ \ V_{\mathrm{a}}(r) =+\infty \text{ for all $r < 1$},\\
{\rm [ii]}& \ \ V_{\mathrm{a}}(r)=-1 \text{ if and only if $r=1$ and $V_{\mathrm{a}}(r) >-1$ otherwise},\\
{\rm [iii]}& \ \ V_{\mathrm{a}}(r) \leq 0 \text{ for all } r\geq 1 \text{ with equality for all $r > r_0$.}
\end{align*}
The distance $r =1 $ represents the (unique) equilibrium distance of two atoms with opposite charge. The constraint $V_{\mathrm{a}}(r) =+\infty $ is usually called the \emph{hard-interaction} or the \emph{hard sphere} assumption, see e.g.\ \cite{Blanc, Mainini}. \EEE The choice of $V_{\rm a}$ reflects a balance between a long-ranged Coulomb attraction and a short-ranged repulsive force when a pair of ions comes too close to each other. Assumption [iii] restricts the interaction range and ensures that the \emph{bond graph} (see Section \ref{sec: prelimi}) is planar. \EEE
The repulsive potential $V_{\mathrm{r}}$ satisfies
\begin{align*}
{\rm [iv]} & \ \ \text{$V_{\mathrm{r}}(r) = +\infty$ for all $r<1$ and $0 \leq V_{\mathrm{r}}(r) <+\infty$ for all $r\geq 1$},\\
{\rm [v]} & \ \ \text{$V_{\mathrm{r}}$ is non-increasing and convex for $r \geq 1$,}\\
{\rm [vi]} & \ \ \text{$V_{\mathrm{r}}(\sqrt{2}r_0) > 3$,} \\
{\rm [vii]} & \ \ \text{$V_{\mathrm{r}}(r) = 0 $ iff $r \geq \sqrt{3}$.}
\end{align*}
Assumption [iv] is the hard-interaction assumption for the repulsive potential. The natural assumption [v] is satisfied for example for repulsive Coulomb interactions. We remark that some quantitative requirement of the form [vi] and [vii] are indispensable to obtain a crystallization result in the hexagonal lattice. Indeed, for vanishing $V_{\rm r}$, ground states could be patches of the triangular lattice. Other quantitative assumptions on the repulsive potential will favor an assemblence of the atoms in the square-lattice as we prove in \cite{FriedrichKreutzSquare}. Also note that two-body pair interactions for identical particle systems often \EEE favor crystallization in the triangular lattice \cite{Radin, Theil}, when the interaction is of attractive-repulsive-type (instead of pure repulsive-type), e.g.\ for the Lennard Jones or the Morse potential. \EEE This reflects the aforementioned \EEE fact that the geometry of molecular compounds often differs from that of their
components.
A main assumption is [vii], i.e., the repulsion vanishes at $\sqrt{3}$ (the distance between second neighbors in the hexagonal lattice). Note that, if $V_{\rm r}(\sqrt{3})$ is instead assumed to be positive and sufficiently large, then crystallization in the hexagonal lattice is not expected as, e.g., a one-dimensional chain of atoms with alternating charges \EEE is energetically favorable. For small, positive $V_{\rm r}(\sqrt{3})$ we still expect crystallization in the hexagonal lattice, but the analysis is much more demanding and beyond the scope of the present contribution.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw[->](0,-2)--++(0,6) node[anchor= east] {$V_{\mathrm{a}}(r)$};
\draw[->](-1,0)--++(5,0) node[anchor =north] {$r$};
\draw[decorate, decoration={snake,amplitude=.4mm,segment length=1.5mm}] (0,3) node[anchor =east]{$+\infty$}--++(1,0);
\draw[dashed,thin](1,-1) --++(0,1) node[anchor =north east] {$ 1 $}--++(0,3);
\draw[fill=black](0,-1) node[anchor = east] {$-1$}++(1,0) circle(.025);
\draw[thick](1.6,0)--++(2.4,0);
\draw[thick](1,-1)--(1.1,-0.6);
\draw[thick](1.1,-0.6) parabola[bend at end] (1.6,0) node[anchor=north]{$r_0$};
\begin{scope}[shift={(8,0)}]
\draw[->](0,-2)--++(0,6) node[anchor= east] {$V_{\mathrm{r}}(r)$};
\draw[->](-1,0)--++(5,0) node[anchor =north] {$r$};
\draw[decorate, decoration={snake,amplitude=.4mm,segment length=1.5mm}] (0,3) node[anchor =east]{$+\infty$}--++(1,0);
\draw[dashed,thin](1,0) node[anchor =north] {$ 1 $}--++(0,3);
\draw[thick](1,.75) parabola[bend at end] ({sqrt(3)},0) node[anchor=north]{$\sqrt{3}$};
\draw[thick]({sqrt(3)},0)--(4,0);
\draw[dashed,thin]({sqrt(3)},-.075)++(150:-.125)--++(150:1.125);
\end{scope}
\end{tikzpicture}
\caption{The potentials $V_{\mathrm{a}}$ and $V_{\mathrm{r}}$.}
\label{FigurePotentials}
\end{figure}
Finally, we require the following \emph{slope conditions} \EEE
\begin{itemize}
\item[{\rm[viii]} ] $ \displaystyle
V'_{\mathrm{r},-}(\sqrt{3}) < {-\frac{3}{\pi}}, \ \ \ \ \ \ \ \frac{1}{r-1}(V_{\mathrm{a}}(r)- V_{\mathrm{a}}(1)) > -2V'_{\mathrm{r},+}(1) \ \ \ \text{for all } r \in (1,r_0],
$
\end{itemize}
where the functions $V_{\mathrm{r},+}',V_{\mathrm{r},-}'$ denote the right and left derivative, respectively. (They exist due to convexity of $V_{\rm r}$.) These conditions are reminiscent of the \emph{soft-interaction} assumption by {\sc Radin} \cite{Radin} and the \emph{slope condition} for an angular potential by {\sc Mainini and Stefanelli} \cite{Mainini}. In particular, we assume that the repulsion grows linearly out of $\sqrt{3}$ and that, roughly speaking, the slope of $V_{\mathrm{a}}$ is steep enough compared to the slope of $V_{\mathrm{r}}$. We highlight that without assumptions of this kind finite crystallization is presently not known, not even for identical particle systems. Assumption [viii] \EEE is only needed in Lemma \ref{LemmaBoundaryEnergy} where the energy contribution of atoms at the boundary of the configuration is \EEE estimated. See also Remark \ref{rem: angle/bond} for an in-depth explanation of the assumption and \EEE for a comparison of our model to \EEE \cite{Mainini}. Note that [viii] also restricts the choice of possible $r_0$.
Assumptions [ii] and [vii] ensure that the (infinite) hexagonal lattice (see \eqref{eq: hex-lattice} below) is \emph{stress free}. An assumption of this kind is necessary as otherwise for finite particle systems surface relaxation at the boundary of the configuration could occur, eventually ruling out \EEE finite crystallization in any lattice. \EEE
We remark that the assumptions are chosen here for the sake of maximizing simplicity rather than generality. Some conditions, e.g., about the hard-interaction assumption or the exact value in [vi], could be weakened at the expense of more elaborated arguments. \EEE In particular, two different repulsive potentials for cations and anions could be considered as long as [iv]-[viii] hold for both potentials.
In the following we assume that conditions $[\mathrm{i}]$-$[\mathrm{viii}]$ are always satisfied.
\subsection{Basic notions}\label{sec: prelimi}
In this section we collect some basic notions. Consider a configuration $C_n \in (\mathbb{R}^{2} \times \{\ -1,1\})^{n}$ with finite energy consisting of the positions $X_n = (x_1,\ldots,x_n) \EEE \in \mathbb{R}^{2n}$ and the charges $Q_n = ( q_1,\ldots,q_n ) \EEE \in \lbrace -1,1 \rbrace^n$.
\textbf{Neighborhood, bonds, angles:} For $i \in \{1,\ldots,n\}$ we define the \textit{neighborhood of} $x_i\in \mathbb{R}^2$ by
\begin{align}\label{eq: neighborhood}
\mathcal{N}(x_i) = \left(X_n \setminus \{x_i\}\right) \cap \lbrace x \in \mathbb{R}^2: \ |x-x_i| \le r_0 \rbrace,
\end{align}
where $r_0$ is defined in [iii]. \EEE If $x_j \in \mathcal{N}(x_i)$, we say that $x_i$ and $x_j$ are \emph{bonded}. We will say that $x_i$ is $k$-bonded if $\# \mathcal{N}(x_i) = k$. Given $x_j,x_k \in \mathcal{N}(x_i)$, we define the \textit{bond-angle} between $x_j,x_i,x_k$ as the angle between the two vectors $x_k-x_i$ and $x_j-x_i$ (choose anti-clockwise orientation, for
definiteness). In general, we say that it is an angle at $x_i$.
\EEE
\textbf{The bond graph:} The set of atomic positions $X_n \subset \mathbb{R}^{2n}$ together with the set of bonds $\{\{x_i,x_j\}: x_j \in \mathcal{N}(x_i)\}$ forms a graph which we call the \textit{bond graph}. Since for finite energy configurations we get \EEE $\mathrm{dist}(x_i,X_n \setminus \{x_i\})\geq 1$ and $x_j \in \mathcal{N}(x_i)$ only if $|x_i-x_j| \leq r_0<\sqrt{2}$, we have that their bond graph is planar. Indeed, given a quadrangle with all sides and one diagonal in $[1,r_0]$, the second diagonal is at least $\sqrt{2} >r_0$. Note that assumption [iii] states that the attractive interactions are restricted to nearest neighbors in the bond graph only. If no ambiguity arises, the number of bonds in the bond graph will be denoted by $b$, i.e.,
\begin{align}\label{eq:b}
b= \# \{\{x_i,x_j\}: x_j \in \mathcal{N}(x_i)\}.
\end{align}
We say a configuration is \emph{connected} if each two atoms are joinable through a simple path in the bond graph. \EEE Any simple cycle of the bond graph is a \textit{polygon}.
\textbf{Acyclic bonds:} We say that a bond is \textit{acyclic} if it is not contained in any simple cycle of the bond graph. Among acyclic bonds we distinguish between \textit{flags} and \textit{bridges}. We call a bridge an acyclic bond which is contained in some simple path connecting two vertices which are included in two distinct cycles. All other acyclic bonds are called flags, see Fig.~\ref{FigureFlagsBridges}.
\begin{figure}[htp]
\centering
\begin{tikzpicture}
\draw[thick](0,0)--++(1,0)--++(60:1)--++(10:1);
\draw[thick](1,0)--++(-50:1)--++(30:1);
\draw[thick](1,0)++(60:1)--++(160:1);
\draw[fill=black](0,0)circle(.05)++(1,0)circle(.05)++(60:1)circle(.05)++(10:1)circle(.05);
\draw[fill=black](1,0)circle(.05)++(-50:1)circle(.05)++(30:1)circle(.05);
\draw[fill=black](1,0)++(60:1)++(160:1)circle(.05);
\draw[fill=black](5,-.5)circle(.05)++(1,0)circle(.05)++(72:1)circle(.05)++(144:1)circle(.05)++(216:1)circle(.05)++(288:1);
\draw(5,-.5)--++(1,0)--++(72:1)--++(144:1)--++(216:1)--++(288:1);
\draw[thick](5,-.5)++(1,0)++(72:1)--++(20:1)--++(80:1);
\draw[fill=black](5,-.5)++(1,0)++(72:1)++(20:1)circle(.05)++(80:1)circle(.05);
\draw[fill=black](5,-.5)++(1,0)++(72:1)++(144:1)++(216:1)++(144:1)circle(.05);
\draw[thick](5,-.5)++(1,0)++(72:1)++(144:1)++(216:1)--++(144:1);
\draw(9,-1)--++(30:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1)--++(330:1);
\draw(9,-1)++(30:1)++(90:1)++(210:-1)--++(330:1)--++(30:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1);
\draw[fill=black](9,-1)circle(.05)++(30:1)circle(.05)++(90:1)circle(.05)++(150:1)circle(.05)++(210:1)circle(.05)++(270:1)circle(.05)++(330:1);
\draw[fill=black](9,-1)++(30:1)++(90:1)++(210:-1)circle(.05)++(330:1)circle(.05)++(30:1)circle(.05)++(90:1)circle(.05)++(150:1)circle(.05)++(210:1)circle(.05)++(270:1);
\draw[dashed](9,-1)++(30:1)++(90:1)--++(210:-1);
\end{tikzpicture}
\caption{Examples of flags (bold) and a bridge (dashed)}
\label{FigureFlagsBridges}
\end{figure}
\textbf{Defects:} By elementary polygons we denote polygons which do not contain any non-acyclic bonds in its interior region. \EEE We call an elementary \EEE polygon in the bond graph which is not a hexagon a \emph{defect}. A configuration is said to be \emph{defect-free} if all its elementary \EEE polygons are hexagons. We also introduce the \textit{excess of edges} $\eta$ by
\begin{align}\label{Excess}
\eta= \sum_{j\geq 6} (j-6) f_j,
\end{align}
where $f_j$ denotes the number of elementary \EEE polygons with $j$ vertices in the bond graph. The excess of edges is a tool to quantify the number of {defects} in the bond graph. \EEE { Note that the summation in (\ref{Excess}) runs over $j\geq 6$. This is due to the fact that we use this definition only for configurations whose bond graph contains only $k$-gons with $k \geq 6$, cf.~Lemma \ref{LemmaPolygon}. We remark that in \cite{Lucia} the \textit{excess of edges} (with respect to the triangular lattice) is referred to as \emph{defect measure}.}
In the following we refer to $C_n$ instead of $X_n$ when speaking about its bond graph, acyclic bonds, or connectedness properties, when no confusion may occur.
\textbf{Charges:} We say that a configuration satisfying
\begin{align}\label{SamechargeNeighbourhood}
\mathcal{N}(x_i) \cap \left\{x_j \in X_n: q_j=q_i\right\} = \emptyset \text{ for all } i \in \{1,\ldots,n\}
\end{align}
has \textit{alternating charge distribution}. Moreover, we call a configuration \emph{repulsion-free} if $|x_i - x_j| \ge \sqrt{3}$ for all $x_i \neq x_j$ with $q_i = q_j$. The \emph{net charge} of a configuration is defined as the (signed) difference of the number of the two atomic types, i.e.,
\begin{align} \label{DefinitionNetCharge}
\mathcal{Q}(C_n) := \sum_{i=1}^n q_i.
\end{align}
\EEE
\textbf{Ground state:} A configuration $C_n$ is called \textit{ground state} for the interaction energy (\ref{Energy}) if for all $C'_n \subset ( \mathbb{R}^{2} \times \{ -1, 1\})^{n}$ there holds
\begin{align*}
\mathcal{E}(C_n) \leq \mathcal{E}(C_n').
\end{align*}
In other words, $C_n$ minimizes (\ref{Energy}) among all possible configurations consisting of $n$ atoms.
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=0.8]
\foreach \i in {0,...,3}{
\foreach \k in {0,...,4}{
\foreach \j in {0,2,4}{
\draw[thin](\k*1.5,{\k*sqrt(3)*.5})++(-\i*1.5,{\i*sqrt(3)*.5})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*1.5,{\k*sqrt(3)*.5})++(-\i*1.5,{\i*sqrt(3)*.5})++(\j*60:1) circle(.05);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*1.5,{\k*sqrt(3)*.5})++(-\i*1.5,{\i*sqrt(3)*.5})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*1.5,{\k*sqrt(3)*.5})++(-\i*1.5,{\i*sqrt(3)*.5})++(\j*60:1) circle(.05);
}
}
}
\end{tikzpicture}
\caption{A configuration with atomic positions forming a subset of the hexagonal lattice. The configuration has alternating charge distribution where white indicates $q=1$ and black indicates $q=-1$. \EEE }
\label{FigureHexagonal-new}
\end{figure}
\textbf{The hexagonal lattice:} We define the \textit{hexagonal lattice} by
\begin{align}\label{eq: hex-lattice}
\mathcal{L} := \EEE \left\{p v_1 + q v_2 + r v_0 : p,q \in \mathbb{Z}, r \in \{0,1\}\right\},
\end{align}
with $v_1 = (0,\sqrt{3})$, $v_2=\frac{1}{2}(3,\sqrt{3})$ and $v_0=(1,0)$. We identify it with its bond graph defined above.
Note that the hexagonal lattice is planar, connected, and all edges have unit length. Since it is bipartite, we can associate to all positions $X_n\subset \mathcal{L}$ of atoms a competitor for the minimization of (\ref{Energy}) by two-coloring (and hereby choosing the charge of) the atoms. The corresponding configuration $C_n$ \EEE is pictured in Fig.~\ref{FigureHexagonal-new}. \EEE Then by assumption [ii], [iii], and [vii] \EEE we have
\begin{align*}
\mathcal{E}(C_n) = -b,
\end{align*}
since all atoms of same charge have at least distance $\sqrt{3}$ and all atoms of $X_n$ are bonded only to atoms of \EEE opposite charge. This means that for subsets of the hexagonal lattice the energy is computed (up to sign) by counting the number of bonds.
\subsection{Main results}
In this section we state our main results. We will derive a rigorous planar crystallization result in the spirit of \cite{HR, Mainini-Piovano, Mainini, Radin}. Afterwards, we will investigate the \emph{net charge} of ground-state configurations. Our first result characterizes the energy of ground states and shows that all ground states are connected. To this end, we introduce the function
\begin{align}\label{eq: beta definition}
\beta(n) := \frac{3}{2}n - \sqrt{\frac{3}{2}n}
\end{align}
for $n \in \mathbb{N}$. By $\lfloor t \rfloor$ we denote the integer part of $t \in \mathbb{R}$. \EEE
\begin{theorem}[Ground-state energy]\label{TheoremEnergyGroundstates} Let $n \geq 1$ and let $C_n$ \EEE be a ground state. Then $C_n$ is connected and satisfies \EEE
\begin{align}\label{Energygroundstates}
\mathcal{E}(C_n) = \EEE -b = -\lfloor \beta(n)\rfloor.
\end{align}
\end{theorem}
\begin{remark}\label{rem: repulsionsfree}
{\normalfont
In view of assumptions ${\rm [ii]}$ and [vii], we have that $\mathcal{E}(C_n) \geq -b$ with equality if and only if the configuration is repulsion-free and all bonds have unit length. In particular, Theorem \ref{TheoremEnergyGroundstates} implies that ground states satisfy both properties.
}
\end{remark}
\EEE
The next result states that ground states are essentially subsets of the hexagonal lattice with alternating charge distribution. Moreover, they satisfy some topological properties. Without further notice, all following statements regarding the geometry of ground states hold up to isometry. \EEE
\begin{theorem}[Characterization of ground states]\label{TheoremGroundstatesleq31} Let $n \geq 30$ and let $C_n$ be a ground state. \EEE Then except for possibly one flag, \EEE $C_n$ is a connected subset of the hexagonal lattice with alternating charge distribution. Moreover, $C_n$ is defect-free and does not contain any bridges. \EEE
\end{theorem}
\begin{remark}\label{rem: main}
{\normalfont (i) More precisely, we show that the bond graph of ground states consists only of hexagonal cycles except for at most two flags. Since ground states are repulsion-free and all bonds have unit length, at least one of the flags is contained in the hexagonal lattice. If two flags exist and are connected, we note that one of them can be rotated in a continuous way without changing the energy.
(ii) The ground states for $n \le 29$ can also be characterized, but due to the smallness of the structures, more degeneracies can occur. In particular, for $n=8,9,12,15, 18, 21, 29$ \EEE there might be one octagon at the boundary. (We refer to Remark \ref{rem: octogons} for more details.) For $n < 10$ the structure can be much more flexible, see Fig.~\ref{FigureFlexible}.
}
\end{remark}
\EEE
Our third main result addresses the net charge. Recall (\ref{DefinitionNetCharge}).
\begin{theorem}[Net charge]\label{TheoremCharge}
The following properties for the net charge of ground states hold:
\begin{itemize}
\item[(i)] (Charge control) There is a universal constant $c>0$ such that for all $n \in \mathbb{N}$ and all ground states $C_n$ the charge satisfies $|\mathcal{Q}(C_n)| \le cn^{1/4}$.
\item[(ii)] (Sharpness of the $n^{1/4}$-scaling) There exists an increasing \EEE sequence of integers $(n_j)_j$ and ground states $(C_{n_j})_j$ such that
$$\liminf_{j \to +\infty} n_j^{-1/4}|\mathcal{Q}(C_{n_j})|>0. $$
\end{itemize}
\end{theorem}
The proof of our main results will be given in Sections \EEE 3-8. In particular, in Section 4 we construct explicitly configurations and give the upper bound for the ground-state energy (Theorem \ref{TheoremEnergyGroundstates}) as well as the proof of Theorem \ref{TheoremCharge}(ii). In Section 6 we conclude the proof of Theorem \ref{TheoremEnergyGroundstates} by providing the lower bound. Section 7 is then devoted to the characterization of ground-state geometries for $n \ge 30$ \EEE (Theorem \ref{TheoremGroundstatesleq31}). Finally, in Section 8 we prove Theorem \ref{TheoremCharge}(i).
\section{Elementary geometric properties}
In this section we provide some elementary geometric facts for ground states independently of $n \in \mathbb{N}$. \EEE
In the proofs, we will use the following convention: we say that we \textit{relocate} $(x,q) \in C_n$, and write $C_n-\{(x,q)\}$, \EEE by considering the configuration $(C_n\cup \{(x+\tau,q)\})\setminus \{(x,q)\} $ \EEE, where $\tau \in \mathbb{R}^2$ is chosen such that
\begin{align*}
\mathrm{dist}(X_n,x+\tau)\geq \sqrt{3}.
\end{align*}
We relocate a set of atoms $A \subset C_n$ by relocating successively every $(x,q) \in A$ and write $C_n -A$. Note that $ C_n-A$ still consists of $n$ particles. \EEE
The first lemma addresses the neighbors. Recall \eqref{eq: neighborhood} and \eqref{SamechargeNeighbourhood}.\EEE
\begin{lemma}[Neighbors and charge distribution\EEE]\label{LemmaNeighborhood}
Let $C_n$ be a ground state. Then $C_n$ has alternating charge distribution and \EEE
\begin{align}\label{NeighbourhoodCard}
\#\mathcal{N}(x_i) \leq 3 \text{ for all } i \in \{1,\ldots,n\}.
\end{align}
\end{lemma}
\begin{proof} Since $C_n$ \EEE is a ground state, it holds that $\mathcal{E}(C_n)<+\infty$. Therefore, by assumption [i],[iv] \EEE
\begin{align}\label{eq: distance}
\mathrm{dist}(x_i, X_n \EEE \setminus\{x_i\}) \geq 1 \text{ for all }i\in \{1,\ldots, n \EEE \}.
\end{align}
For brevity, we define $\mathcal{N}_{\rm rep}(x_i) = \mathcal{N}(x_i) \cap \left\{x_j \in X_n: q_j=q_i\right\}$ for all $i \in \lbrace 1 ,\ldots, n \rbrace$. We give the proof of the statement in two \EEE steps. \EEE
\medskip
\textbf{Claim 1:} We have \EEE
\begin{align*}
\# \mathcal{N}_{\rm rep}(x_i) \leq 1 \text{ for all } i \in \{1,\ldots,n\},
\end{align*}
and if $\#\mathcal{N}(x_i) \leq 4$, then
\begin{align*}
\# \mathcal{N}_{\rm rep}(x_i) =0.
\end{align*}
\noindent \emph{Proof of Claim 1}: First, \eqref{eq: distance} \EEE and $r_0 < (2\sin(\frac{\pi}{7}))^{-1}$ entail by an elementary geometric argument that \EEE
\begin{align}\label{eq: six neighbors}
\#\mathcal{N}(x_i) \leq 6 \text{ for all } i \in \{1,\ldots,n\}.
\end{align}
In fact, if $\#\mathcal{N}(x_i) \ge 7$, two neighbors of $x_i$ would necessarily have distance smaller than $1$. \EEE
Now assume by contradiction that $\# \mathcal{N}_{\rm rep}(x_i) \geq 2$. Note that \EEE every bond between points of different charge contributes at least $-1$ to the energy by $[\mathrm{ii}]$. This along with $V_{\rm r} \ge 0$ and the fact that the energy per neighbor of same charge is larger than $3$ (see $[\mathrm{v}]$ and $[\mathrm{vi}]$) allows us to relocate $(x_i,q_i)$: we \EEE obtain
\begin{align*}
\mathcal{E}( C_n - \{(x_{i},q_i)\} ) < \mathcal{E}(C_n) + \#\left(\mathcal{N}(x_i) \setminus \mathcal{N}_{\rm rep}(x_i)\right) - 3\#\mathcal{N}_{\rm rep}(x_i) \le \mathcal{E}(C_n) + 4 - 6 < \EEE \mathcal{E}(C_n).
\end{align*}
This contradicts the fact that $C_n$ is a ground state. The argument in the case $\#\mathcal{N}(x_i)\leq 4$ and $\#\mathcal{N}_{\rm rep}(x_i) \geq 1$ is similar: \EEE due to the fact that every bond between points of different charge contributes at least $-1$ to the energy and the contribution of neighbors of the same charge is larger than $3$, \EEE we have
\begin{align*}
\mathcal{E}( C_n - \{(x_{i},q_i)\} \EEE ) < \mathcal{E}(C_n) + \#\left(\mathcal{N}(x_i) \setminus \mathcal{N}_{\rm rep}(x_i)\right) - 3\#\mathcal{N}_{\rm rep}(x_i) \le \mathcal{E}(C_n) + 3 - 3 \EEE = \EEE \mathcal{E}(C_n).
\end{align*}
This contradiction shows that \EEE Claim 1 holds true.
\textbf{Claim 2:} \EEE $\#\mathcal{N}(x_i) \leq 3$ for all $i \in \{1,\ldots,n\}$.
\noindent \emph{Proof of Claim 2}: Assume by contradiction that there exists $i \in \{1,\ldots,n\}$ such that $\#\mathcal{N}(x_i) \geq 4$. By Claim 1 we may suppose that there exist \EEE $\{x_0,\ldots,x_3\}\subset \mathcal{N}(x_{i})$ with $q_j =-q_i$ $j=0,\ldots,3$. We let $\theta_j \in [0,2\pi)$ be the angle between $x_j,x_i,x_{j+1}$. (Here and in the following the indices have to be understood modulo $4$.) \EEE We can choose $j_0 \in \{0,\ldots,3\}$ such that
\begin{align*}
\theta_{j_0}+\theta_{j_0-1} \leq \frac{1}{4}\sum_{j=0}^3 (\theta_{j}+\theta_{j-1} ) = \pi.
\end{align*}
Note that $|x_{j_0}-x_{j_0+1}| \leq 2r_0 \sin\left(\theta_{j_0}/2\right), |x_{j_0}-x_{j_0-1}| \leq 2r_0\sin\left(\theta_{j_0-1}/2\right)$. By concavity of $\sin(x)$ for $x \in [0,\pi]$ and the fact that $\sin(x)$ is increasing for $x\in [0,\frac{\pi}{2}]$, we have
\begin{align*}
|x_{j_0}-x_{j_0+1}| + |x_{j_0}-x_{j_0-1}| &\leq r_0\left( 2\sin\left(\frac{\theta_{j_0}}{2}\right) + 2\sin\left(\frac{\theta_{j_0-1}}{2}\right)\right)\notag \\&\leq 4r_0 \sin\left(\frac{\theta_{j_0}+\theta_{j_0-1}}{4}\right)\leq 4r_0 \sin\left(\frac{\pi}{4}\right)=2\sqrt{2}r_0.
\end{align*}
Since $V_{\mathrm{r}}$ is convex and non-increasing (see $[\mathrm{v}]$), we find
\begin{align*}
V_{\mathrm{r}}\left(|x_{j_0}-x_{j_0+1}|\right) + V_{\mathrm{r}}\left(|x_{j_0}-x_{j_0-1}|\right) \geq 2V_{\mathrm{r}}\left(\frac{1}{2}\left(|x_{j_0}-x_{j_0+1}|+ |x_{j_0}-x_{j_0-1}|\right)\right) \geq 2V_{\mathrm{r}}(\sqrt{2}r_0).
\end{align*}
Using $[\mathrm{ii}]$, $[\mathrm{vi}]$, \EEE $V_{\rm r} \ge 0$, and \eqref{eq: six neighbors} \EEE we finally observe
\begin{align*}
\mathcal{E}(C_n-\{ (x_{j_0},q_{j_0})\})&\leq \mathcal{E}(C_n)-2V_{\mathrm{r}}(\sqrt{2}r_0) + \# \mathcal{N}(x_{j_0}) \EEE \leq \mathcal{E}(C_n)-2V_{\mathrm{r}}(\sqrt{2}r_0) + 6 < \mathcal{E}(C_n).
\end{align*}
This contradicts the fact that $C_n$ is a ground state and concludes the proof of Claim 2. \EEE
Now, \eqref{NeighbourhoodCard} holds, see \EEE Claim 2. \EEE The property of alternating charge (see \eqref{SamechargeNeighbourhood}) \EEE follows from \eqref{NeighbourhoodCard} and the second statement of Claim 1. This concludes \EEE the proof of Lemma \ref{LemmaNeighborhood}.
\end{proof}
We now investigate simple cycles in the bond graph. \EEE
\begin{lemma}[Polygons\EEE]\label{LemmaPolygon} Let $C_n$ be a ground state. Then all polygons have at least $6$ edges and the number of edges is always even. \EEE
\end{lemma}
\begin{proof}
As $C_n$ has alternating charge distribution, two \EEE successive vertices of a path of the bond graph of a ground state have to be of different charge. This prohibits cycles of odd length. In particular, \EEE this rules out triangles and pentagons. We are thus left to prove that there is no simple square. \EEE
Now assume by contradiction that the bond graph of a ground state contains a simple square. As two successive vertices of the square have different charge, \EEE the vertices connected by the diagonal have the same charge. Denote the lengths of the two diagonals by $d_1$ and \EEE $d_2$, respectively. An elementary computation shows
\begin{align} \label{ineq: di}
d_1^2+d_2^2 \leq 4r_0^2.
\end{align}
In fact, denoting the vertices of the square by $x_1,\ldots,x_4$ (counterclockwise) one can use the elementary expansion
$$|x_1 - x_3|^2 + |x_2 - x_4|^2 = |x_1 - x_2|^2 + |x_2 - x_3|^2 + |x_3 - x_4|^2 + |x_4 - x_1|^2 - |x_1 - x_2 + x_3 - x_4|^2 $$
along with the fact that each bond length is smaller or equal to $r_0$. \EEE Now (\ref{ineq: di}) together with Young's inequality gives
\begin{align*}
(d_1 +d_2)^2 \leq 2(d_1^2 +d_2^2)\leq 8r_0^2.
\end{align*}
Since $V_{\mathrm{r}}$ is convex and non-increasing (see ${\rm [v]}$), \EEE we obtain
\begin{align*}
V_{\mathrm{r}}(d_1) +V_{\mathrm{r}}(d_2)\geq 2 V_{\mathrm{r}}\left(\frac{1}{2}( d_1+d_2)\right) \geq 2V_{\mathrm{r}}(\sqrt{2}r_0).
\end{align*}
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=.8]
\draw[thin](0,0)node[anchor =east]{$x_1$}--++(1,0)--++(70:1)--++(-1,0)--++(70:-1);
\draw[thin](0,0)--++(220:1);
\draw[fill= white](0,0)++(220:1) circle(.05);
\draw[thin](1,0)--++(-20:1);
\draw[fill= black](1,0)++(-20:1) circle(.05);
\draw[thin](70:1)--++(100:1);
\draw[fill= black](70:1)++(100:1) circle(.05);
\draw[thin](1,0)++(70:1)--++(50:1);
\draw[fill= white](1,0)++(70:1)++(50:1) circle(.05);
\draw[fill=black](0,0)circle(.05)++(1,0)++(70:1)circle(.05);
\draw[fill=white](0,0)++(1,0)circle(.05)++(70:1)++(180:1)circle(.05) node[anchor =east]{$x_2$};
\end{tikzpicture}
\caption{ Relocating $(x_1,q_1)$ and $(x_2,q_2)$ \EEE strictly decreases the energy.}
\label{FigureBondgraphsquare}
\end{figure}
We relocate two neighboring particles in the square, denoted by $(x_1,q_1)$ and $(x_2,q_2)$. \EEE By \eqref{NeighbourhoodCard} and the fact that $x_1$ and $x_2$ share a bond, we observe that hereby \EEE we remove at most $5$ bonds between atoms \EEE of different charge, see Fig.~\ref{FigureBondgraphsquare}. Using $[\mathrm{ii}]$, $[\mathrm{vi}]$, \EEE and $V_{\rm r} \ge 0$, we obtain
\begin{align*}
\mathcal{E}(C_n- \{(x_1,q_1),(x_2,q_2)\}) \leq \mathcal{E}(C_n) -2V_{\mathrm{r}}(\sqrt{2}r_0) + 5 < \mathcal{E}(C_n) -1 < \mathcal{E}(C_n).
\end{align*}
This contradicts the fact that $C_n$ is a ground state and concludes the proof.
\end{proof}
\begin{lemma}\label{lemma:bondangles} Let $C_n$ be a ground state. Then all the bond angles $\theta$ satisfy
\begin{align}\label{ineq:bondangles}
\frac{\pi}{3}\leq \theta \leq \frac{5\pi}{3}.
\end{align}
\end{lemma}
\begin{proof} Assume by contradiction that there exists a bond angle $\theta$ between $x_1,x_0,x_2$ satisfying $\theta < \frac{\pi}{3}$. As $|x_1-x_0|, |x_2-x_0|\leq r_0$, an easy trigonometric argument shows that then also $|x_1-x_2| \leq r_0$. Therefore, the points $\{x_0,x_1,x_2\}$ form a triangle in the bond graph. This contradicts Lemma \ref{LemmaPolygon} and the first inequality in \eqref{ineq:bondangles} follows. Now, if there exists a bond angle $\theta > 5\pi/3$, then, since the bond angles at every point sum to $2\pi$, there exists a bond angle $\tilde{\theta} < \pi/3$, which is impossible. Therefore, also the second inequality in \eqref{ineq:bondangles} is proved.
\end{proof}
\EEE
\section{Upper bound on the ground-state energy}
This section is devoted to the explicit construction of configurations $D_n$ \EEE with alternating charge distribution which are \EEE subsets of the hexagonal lattice. These configurations provide a reference energy value for every $n$, namely $\mathcal{E}(D_n) \EEE=-\lfloor\beta(n)\rfloor$, cf. \eqref{eq: beta definition}. This already gives the upper bound in \eqref{Energygroundstates}. Moreover, we will construct explicitly configurations $C_n$ \EEE with net charge (see \eqref{DefinitionNetCharge}) \EEE of the order $n^{1/4}$ which establishes \EEE Theorem \ref{TheoremCharge}(ii). We defer the lower bound on the ground-state energy, the upper bound on the net charge, and the characterization of the ground states to the subsequent sections. \EEE
By the special geometry of the hexagonal lattice, it is quite natural to give an interpretation of the two terms appearing in $\beta$. The leading order term of the energy is given by $-\frac{3}{2}n$. It is related to the fact that each atom which is not contained in the exterior face of the bond graph is bonded to exactly \EEE three atoms of opposite \EEE charge and every bond connects two atoms. The repulsive term in the energy is zero for such configurations since the distance of each \EEE two atoms \EEE of the same charge is bigger or equal than $\sqrt{3}$. The additional lower order correction term in the energy is \EEE due to the fact that a certain proportion of the \EEE atoms touching the exterior face of the bond graph is \EEE only bonded to two atoms of opposite charge. Their cardinality scales like $\sqrt{n}$.
\subsection{Special subsets of the hexagonal lattice}\label{sec: daisy}
We exhibit special configurations that are subsets of the hexagonal lattice with energy \EEE $-\lfloor \beta(n)\rfloor$. This provides an upper bound for the ground-state energy. We give \EEE a recursive construction for these geometries following the ideas in \cite[Section 4,5]{Mainini}. \EEE
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=.5]
\foreach \j in {0,2,4}{
\draw[thin](\j*60:1)--(\j*60+60:1);
\draw[fill=black](\j*60:1) circle(.07);
}
\foreach \j in {1,3,5}{
\draw[thin](\j*60:1)--(\j*60+60:1);
\draw[fill=white](\j*60:1) circle(.07);
}
\begin{scope}[shift={(4.5,0)}]
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin](\k*60+30:{sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*60+30:{sqrt(3)})++(\j*60:1) circle(.07);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*60+30:{sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*60+30:{sqrt(3)})++(\j*60:1) circle(.07);
}
}
\end{scope}
\begin{scope}[shift={(12,0)}]
\foreach \i in {1,2}{
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin](\k*60+30:{\i*sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*60+30:{\i*sqrt(3)})++(\j*60:1) circle(.07);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*60+30:{\i*sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*60+30:{\i*sqrt(3)})++(\j*60:1) circle(.07);
}
}
}
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin](\k*60:3)++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*60:3)++(\j*60:1) circle(.07);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*60:3)++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*60:3)++(\j*60:1) circle(.07);
}
}
\end{scope}
\end{tikzpicture}
\caption{ $D_6$, $D_{24}$ and $D_{54}$.\EEE}
\label{FigureDaisies}
\end{figure}
First, we \EEE provide the construction for $n=6k^2$, $k\in \mathbb{N}$. For $k=1$ we define $D_{6}$ \EEE to be a regular hexagon $\{x_1,\ldots,x_6\}$, where the points $x_1,\ldots,x_6$ are arranged in a counter-clockwise sense as the vertices of the hexagon and with charges $q_i=(-1)^{i+1}, i=1,\ldots,6$. Once we have constructed $D_{6(k-1)^2}$\EEE, we construct $D_{6k^2}$ \EEE by attaching hexagons on all boundary sides of $D_{6(k-1)^2}$ \EEE such that every atom has two or three atoms \EEE of opposite charge in its neighborhood. (This is possible since the hexagonal lattice is bipartite.) \EEE These configurations are pictured in Fig.~\ref{FigureDaisies}. Due to their symmetry, these configurations are called \emph{daisies} and we will indicate their atomic positions also by $X_{6k^2}^{\rm daisy}$. Note that daisies have net charge zero. \EEE
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=0.5]
\foreach \i in {0,1,2}{
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin,dashed](\k*60+30:{3*sqrt(3)})++(\j*60:1)++(\k*60+150:{\i*sqrt(3)})--++(\j*60+120:1);
}
\foreach \j in {1,3,5}{
\draw[thin,dashed](\k*60+30:{3*sqrt(3)})++(\k*60+150:{\i*sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
}
}
}
\draw[thin](120:3)++(60:2)--++(120:1)--++(180:1)--++(60:-1)--++(180:1)--++(60:-1);
\draw[thin](120:3)++(60:2)++(120:1)++(180:1)++(60:-1)--++(-60:1);
\draw[thin](120:3)++(60:2)++(120:1)++(180:1)++(60:-2)++(180:1)--++(-60:1);
\draw[fill=white](120:3)++(60:2)++(120:1) circle(.07);
\draw[fill=black](120:3)++(60:2)++(120:1)++(180:1) circle(.07);
\draw[fill=white](120:3)++(60:2)++(120:1)++(180:1)++(60:-1) circle(.07);
\draw[fill=black](120:3)++(60:2)++(120:1)++(180:2)++(60:-1) circle(.07);
\draw[fill=white](120:3)++(60:2)++(120:1)++(180:2)++(60:-2) circle(.07);
\foreach \i in {1,2}{
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin](\k*60+30:{\i*sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*60+30:{\i*sqrt(3)})++(\j*60:1) circle(.07);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*60+30:{\i*sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*60+30:{\i*sqrt(3)})++(\j*60:1) circle(.07);
}
}
}
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin](\k*60:3)++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*60:3)++(\j*60:1) circle(.07);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*60:3)++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*60:3)++(\j*60:1) circle(.07);
}
}
\draw[->](-3.25,6.1)node[anchor =east]{$6k^2+1 $}--++(2,-.75);
\draw[->](-4.25,5.5)node[anchor =east]{$6k^2+2 $}--++(2,-.25);
\draw[->](-4.25,5.75)++(60:-1)node[anchor =east]{$6k^2+3 $}--++(2,-.5);
\draw[->](-5.25,5.2)++(60:-1)node[anchor =east]{$6k^2+4 $}--++(2,0);
\draw[->](-5.25,5.2)++(60:-2)node[anchor =east]{$6k^2+5 $}--++(2,0);
\end{tikzpicture}
\caption{Construction of the $(6k^2+m)$-configuration for $k=3$ and $m=5$.}
\label{FigureCkn}
\end{figure}
Now for $n\neq 6k^2$ we can assume that $n=6k^2+m$ for some $1\leq m< 12k+6$. We start \EEE from $D_{6k^2}$ \EEE and add a new atom \EEE to the bond graph in such a way that it gets bonded to the leftmost \EEE among the uppermost atoms \EEE of $D_{6k^2}$ \EEE and that it has distance larger \EEE or equal than $\sqrt{3}$ to all the other atoms. We choose the charge to be the opposite charge of the leftmost \EEE among the uppermost atoms. Then we add atoms in a counter-clockwise fashion such that the latest atom added is bonded to the atom \EEE added in the previous step and possibly to some other atom of $D_{6k^2}$.\EEE Moreover, its distance to all the other atoms is at least $\sqrt{3}$. \EEE We choose the charge to be the opposite of the charge of the atom \EEE added in the previous step. One can realize that this defines uniquely \EEE a procedure in order to add $m$ atoms as shown in Fig.~\ref{FigureCkn}.
The first main result of this section is the following upper bound for the ground-state energy. \EEE
\begin{proposition}[Upper bound for ground-state energy\EEE]\label{PropositionDaisy} For each $n \in \mathbb{N}$, let $D_n$ be the configuration introduced above. There holds
$
\mathcal{E}(D_n) = -\lfloor \beta(n)\rfloor.
$
In particular, for all ground states $C_n$ there holds
$
\mathcal{E}(C_n) \leq -\lfloor \beta(n)\rfloor.
$\EEE
\end{proposition}
\begin{proof} The proof follows as in \cite[Proposition 4.1, Proposition 5.1]{Mainini}, additionally \EEE noting that in the provided construction, all atoms \EEE in the bond graph are bonded \EEE to points of opposite charge only.
\end{proof}
\subsection{Daisies with an additional trapezoid\EEE}\label{sec: daisy + something}
Note that the above configurations have net charge in $\lbrace -1 ,0 , 1\rbrace$. Starting with a daisy and attaching a trapezoid \EEE in a suitable way, we can also construct configurations with energy $- \lfloor \beta(n) \rfloor$ having a net charge of order $n^{1/4}$. The following construction is inspired by the one of \cite{Davoli15} applied in connection to the derivation of the so-called $n^{3/4}$-law.
We choose $k \in \mathbb{N}$ with $k \ge 252$, $r \in 6 \mathbb{N}$ with $r \le \sqrt{k/7}$, and let
\begin{align}\label{eq: the-n-def}
n := 6k^2 + 2kr + \frac{1}{6}r^2 + 1 = 6\hat{k}^2+1,
\end{align}
where $\hat{k} = k + r/6$. We construct a configuration $C_n$ as follows. We start from $C_{6k^2}$ and add a new atom to the bond graph in such a way that it gets bonded to the leftmost among the uppermost atoms of $C_{6k^2}$ and that it has distance greater or equal than $\sqrt{3}$ to all the other atoms. We choose the charge to be the opposite charge of the leftmost among the uppermost atoms. (Say, charge $1$ for definiteness.) \EEE Then we add atoms in a counter-clockwise way similar to the previous construction until `the line is completed'.
Overall, we add $2k-1$ atoms by this procedure where $k$ atoms have charge $1$ and $k-1$ atoms have charge $-1$. We then repeat this construction by adding another row of atoms on the top where we need to add $2(k-1)-1$ atoms. We repeat this until we have added $r$ rows of atoms corresponding to
$$ \sum_{j= k - r+1}^k (2j-1) = 2kr - r^2$$
added atoms. Note that in each row the number of added atoms of charge $1$ exceeds the number of added atoms of charge $-1$ by exactly one. The construction is sketched in Fig.~\ref{FigureDaisyParallelogram}. \EEE
Finally, in a last row we add $m :=7r^2/6 +1$ atoms. Note that by the assumptions $k \ge 5$ and $r \le \sqrt{k/7}$ we have
\begin{align*}
\frac{7}{6}r^2 + 1 \le k/6+ 1 \le \frac{k - r}{2}, \EEE
\end{align*}
i.e., the chain of atoms added in row $r$ is `long enough' so that $m$ atoms can be added in the last row. Note that the resulting configuration consists of $n$ atoms, see \eqref{eq: the-n-def}.
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=0.2]
\foreach \k in {0,...,5}{
\begin{scope}[rotate=60*\k]
\draw[thin](8,0)++(60:2)++(-60:-1)++(0:1)++(60:2)++(120:2)--++(0:-1);
\foreach \j in {0,...,5} {
\draw[thin](8,0)++(60:-2+\j)++(-60:3-\j)--++(0:1)--++(60:1)--++(120:1);
\draw[fill=white](8,0)++(60:-2+\j)++(-60:3-\j)++(0:1)circle(.07)++(60:1)++(120:1)circle(.07);
\draw[fill=black](8,0)++(60:-2+\j)++(-60:3-\j)++(0:1)++(60:1)circle(.07)++(120:1);
}
\end{scope}
}
\begin{scope}[rotate=60*2]
\draw[thin, dash pattern=on 1pt off 1pt on 1pt off 1pt](9,0)++(60:-2+5)++(-60:4-5)++(0:1)++(60:1)++(120:1)--++(0:-1);
\foreach \j in {0,...,5} {
\draw[thin,dash pattern=on 1pt off 1pt on 1pt off 1pt](9,0)++(60:-2+\j)++(-60:4-\j)--++(0:1)--++(60:1)--++(120:1);
}
\draw[thin, dash pattern=on 1pt off 1pt on 1pt off 1pt](10,0)++(60:3)++(-60:5-5)++(0:1)++(60:1)++(120:1)--++(0:-1);
\foreach \j in {0,...,4} {
\draw[thin,dash pattern=on 1pt off 1pt on 1pt off 1pt](10,0)++(60:-1+\j)++(-60:4-\j)--++(0:1)--++(60:1)--++(120:1);
}
\end{scope}
\foreach \k in {0,...,5}{
\begin{scope}[rotate=60*\k]
\draw[thin](7,0)++(60:2)++(-60:-2)++(0:1)++(60:1)++(120:1)--++(0:-1);
\foreach \j in {0,...,4} {
\draw[thin](7,0)++(60:-2+\j)++(-60:2-\j)--++(0:1)--++(60:1)--++(120:1);
\draw[fill=white](7,0)++(60:-2+\j)++(-60:2-\j)++(0:1)circle(.07)++(60:1)++(120:1)circle(.07);
\draw[fill=black](7,0)++(60:-2+\j)++(-60:2-\j)++(0:1)++(60:1)circle(.05)++(120:1);
}
\end{scope}
}
\foreach \k in {0,...,5}{
\foreach \j in {-1,1}{
\draw[thin](60*\k:\j*7)--++(120+60*\k:\j*1);
\draw[thin](60*\k:\j*7)--++(240+60*\k:\j*1);
}
}
\foreach \k in {0,...,2}{
\draw[fill=black](120*\k:7)circle(.07);
\draw[fill=white](120*\k+60:7)circle(.07);
}
\foreach \j in {0,...,2}{
\draw[thin](\j*60:1)--++(120+\j*60:1);
\draw[thin](\j*60+180:1)--++(300+\j*60:1);
}
\foreach \j in {0,...,2}{
\draw[fill=black](\j*120:1)circle(.05);
\draw[fill=white](\j*120+60:1)circle(.07);
}
\foreach \m in {0,1,2,3,4,5}{
\foreach \k in {1,...,3}{
\foreach \i in {1,...,6}{
\begin{scope}[shift={(\m*60+90:{-sqrt(3)})}]
\begin{scope}[shift={(\i*60+30:{\k*sqrt(3)})}]
\foreach \j in {0,...,2}{
\draw[thin](\j*60:1)--++(120+\j*60:1);
\draw[thin](\j*60+180:1)--++(300+\j*60:1);
}
\foreach \j in {0,...,2}{
\draw[fill=black](\j*120:1)circle(.07);
\draw[fill=white](\j*120+60:1)circle(.07);
}
\end{scope}
\end{scope}
}
}
}
\end{tikzpicture}
\caption{Construction of a daisy \EEE with an additional trapezoid. \EEE Two rows of atoms have been added where the first and the last atom in the added rows have charge $+1$. Thus, the net charge of the configuration is $+2$.\EEE}
\label{FigureDaisyParallelogram}
\end{figure}
Concerning the energy, we observe that in the row where $2j-1$ atoms are added \EEE we add exactly $3j-2$ bonds to the bond graph for $k-r+1 \le j \le k$. \EEE In the ultimate row we add $1 + 3(m-1)/2 = 1+ 7r^2/4$ bonds. Consequently, using Proposition \ref{PropositionDaisy} the energy of $C_n$ is given by
\begin{align*}
\mathcal{E}(C_n) & = \mathcal{E}(C_{6k^2}) - \sum_{j=k-r+1}^k (3j-2) - 1 - 7r^2/4 = -9k^2 + 3k - 3kr + 3r^2/2 + r/2 - 1 - 7r^2/4 \\& = -\frac{3}{2}6(k + r/6)^2 + \sqrt{\frac{3}{2}6(k+r/6)^2} -1 \EEE = -\lfloor \beta( 6\hat{k}^2) \rfloor - 1 \EEE = -\lfloor \beta(n) \rfloor.
\end{align*}
We now determine the \emph{net charge} of the configuration. Recall that the configuration $C_{6k^2}$, from which we started our construction, has net charge zero. As explained above, in each added row the number of added atoms of charge $1$ exceeds the number of added atoms of charge $-1$ by exactly one, i.e., $\mathcal{Q}(C_n) = r+1$.
We are now in the position to give the proof of Theorem \ref{TheoremCharge}(ii). To this end, consider the sequence of integers $n_j = 6(j+r_j/6)^2+1$, where $r_j = 6 \lfloor \frac{1}{6} \sqrt{j/7} \rfloor$, and the ground states $C_{n_j}$ constructed above. Note that by $r_j \le j$ we get $n_j \le 9j^2$. Thus, we calculate
$$\mathcal{Q}(C_{n_j}) = r_j+1 \ge \sqrt{j/7}-5 \ge \frac{1}{\sqrt{21}}n_j^{1/4} - 5.$$
This yields $\displaystyle\liminf_{j \to +\infty} n_j^{-1/4}|\mathcal{Q}(C_{n_j})|>0$.
\EEE
\section{Boundary Energy}
In this section we introduce the concept of boundary energy and prove a corresponding estimate which will be fundamental for the characterization of ground states and their energy in Section \ref{sec: characterization-energy}, \ref{sec: characterization}. \EEE
\textbf{Boundary atoms, boundary energy:} Within the bond graph, we say that an atom is a \textit{boundary atom\EEE} if it is not contained in the interior region of any simple cycle. Otherwise, we call it \emph{bulk atom}. \EEE We denote the union of the boundary atoms \EEE by $\partial X_n$. A \textit{boundary bond} is a bond containing a boundary atom. All other bonds are called \textit{bulk bonds}. Similarly, a bond-angle will be called \emph{boundary angle} or \emph{bulk angle}, depending on whether the associated atom is a boundary atom or not. \EEE
We denote by $d$ the number of boundary atoms of $X_n$. Given $C_n$, we define its \textit{bulk}, denoted by $C_n^{\mathrm{bulk}}$, as the sub-configuration obtained by dropping all boundary atoms (and the corresponding charges). Similarly, the particle positions are indicated by $X_n^{\mathrm{bulk}}$. \EEE With the above definition, we have that the bulk is an \EEE $(n-d)$-atom configuration. We define $\mathcal{E}^{\mathrm{bulk}}(C_n)$ as the energy of $C_n^{\mathrm{bulk}}$. We then have two contributions to the energy of $C_n$, namely $\mathcal{E}^{\mathrm{bnd}}$ and $\mathcal{E}^{\mathrm{bulk}}$, where
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) := \mathcal{E}(C_n) - \mathcal{E}^{\mathrm{bulk}}(C_n).
\end{align*}
\textbf{Maximal polygon:} We introduce an additional notion in the case that $C_n$ is connected and does not contain acyclic bonds. In this case, the bond graph is delimited by a simple cycle which we call the \emph{maximal polygon}. We denote the atoms of the maximal polygon by $\{x_1,\ldots,x_d\}$ and the interior angle at the point $x_i \in \partial X_n$ by $\theta_i$. \EEE Moreover, we indicate by
\begin{align*}
&I_2=\{ x_i \in \partial X_n: \EEE \#\mathcal{N}(x_i)=2\},\\&I_3=\{ x_i \in \partial X_n: \EEE \#\mathcal{N}(x_i)=3\},
\end{align*}
the set of $2$-bonded and $3$-bonded boundary atoms, respectively. If $C_n$ is a ground state, we note that $\# I_2 + \# I_3 = d$ by Lemma \ref{LemmaNeighborhood}. \EEE
We now provide an estimate for the boundary energy. Its proof is inspired by \cite[Lemma 6.2]{Mainini}. The precise estimates, however, deviate significantly from the study in \cite{Mainini} due to the presence of the repulsive potential $V_{\rm r}$ instead of an angular potential. We defer a discussion in that direction after the proof, see Remark \ref{rem: angle/bond}. \EEE
\begin{lemma}[Boundary energy\EEE]
\label{LemmaBoundaryEnergy}
Let $n \geq 6$ and let $C_n$ be a connected ground state with no acyclic bonds. Then
\begin{align}\label{BoundaryEnergyEstimate}
\mathcal{E}^{\mathrm{bnd}}(C_n) \geq -\frac{3}{2}d +3
\end{align}
with equality only if all the three \EEE following conditions are satisfied:
\begin{align}
& \text{All boundary bonds are of unit length},\label{eq: 1}\\
&\#I_3 = \frac{1}{2}d -3,\label{eq: 2}\\
&\theta_i = \frac{4\pi}{3} \text{ if } x_i \in I_3 \text{ and } \theta_i = \frac{2\pi}{3} \text{ if } x_i \in I_2.\label{eq: 3}
\end{align}
\end{lemma}
\begin{remark}\label{rem: bdy}
{\normalfont (i) Note by Lemma \ref{LemmaNeighborhood} that $C_n$ has alternating charge distribution and thus $d$ is even.
(ii) Observe by \eqref{eq: 2} that equality in \eqref{BoundaryEnergyEstimate} implies that $\frac{3}{2}d-3$ bonds contribute to the boundary energy. Thus, equality in \eqref{BoundaryEnergyEstimate} together with [ii] and [vii] imply that for all boundary atoms $x_i$ one has $\min\lbrace |x_i - x_j|: \ j \in \lbrace 1, \ldots, n \rbrace, j\neq i, \ q_j = q_i \rbrace \ge \sqrt{3}$.\EEE}
\end{remark}
\begin{proof} Suppose that $\{x_1,\ldots,x_d\}$ are ordered such that \EEE $x_i\in \mathcal{N}(x_{i+1})$, $i=1,\ldots,d$. Here and in the following, we use \EEE the identification $x_{d+1} = x_1$ and $x_{0} = x_d$. For a $3$-bonded atom $x_i$, denote by $x_i^b \in X_n \setminus \{x_{i-1},x_{i+1}\}$ the atom that is connected to $x_i$ with the third bond. The boundary energy can be estimated by
\begin{align}\label{eq: boundary energy}
\begin{split}
\mathcal{E}^{\mathrm{bnd}}(C_n) &\geq \sum_{i=1}^d \Big(\frac{1}{2}\big( V_{\mathrm{a}}(|x_i-x_{i+1}|)+V_{\mathrm{a}}(|x_i-x_{i-1}|)\big)+V_{\mathrm{r}}(|x_{i+1}-x_{i-1}|) \Big) \\&\quad+ \sum_{x_i \in I_3}\Big( V_{\mathrm{a}}(|x_i-x_i^b|) + V_{\mathrm{r}}(|x_{i-1}-x_i^b|) + V_{\mathrm{r}}(|x_{i+1}-x_i^b|) \Big).
\end{split}
\end{align}
Here, we split the interactions into two sums. The first sum represents \EEE the energy between successive boundary atoms. The second sum is a lower bound for the energy of $3$-bonded \EEE atoms that may be bonded to other boundary atoms or to bulk atoms: note that we might double count (negative) attractive interaction if also $x_i^b$ is a boundary atom. However, we never double count (positive) repulsive interaction. To see this, suppose that a repulsive interaction would be double counted in the sum of the right hand side, i.e., there exist $i,j \in \{1,\ldots,d\}$, $i \neq j$, such that, e.g., $\{x_{j-1},x_j^b\} = \{x_{i-1},x_i^b\}$. More precisely, as $i\neq j$, this means $x_{j-1} = x^b_i$ and $x_{i-1} =x_j^b$. This implies that $x_i$ as well as $x_j$ are both bonded to $x_{j-1}$ and $x_{i-1}$. Then, however, the atoms $\{x_j,x_{j-1}, x_{i-1}, x_i\}$ form a square in the bond graph which contradicts Lemma \ref{LemmaPolygon}. \EEE
For a $3$-bonded atom $x_i$, denote by $\theta_i^1,\theta_i^2\in [0,2\pi]$ the two angles forming $\theta_i$ enclosed by the three bonds at $x_i$. Finally, we define $\delta : = \frac{\#I_2+2\#I_3}{d}$ and note that $\delta \in [1,2]$ since $\# I_2 + \# I_3 =d$. \EEE
We will prove that
\begin{align}\label{boundarylowerbound-new}
\mathcal{E}^{\mathrm{bnd}}(C_n) & \ge -\delta d + \sum_{x_i \in I_2} V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i}{2}\Big)\Big) + \sum_{x_i \in I_3} \sum_{j=1,2} V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i^j}{2}\Big)\Big) \EEE \nonumber \\& \geq - \delta d +\delta d V_{\mathrm{r}}\Big(2\sin\Big(\frac{\pi(d-2)}{2\delta d}\Big)\Big),
\end{align}
where the first inequality is strict if not all lengths of boundary bonds \EEE are equal to $1$. We defer the proof of this estimate and its strict version \EEE and first show that it implies the statement of the lemma. \EEE
First, introducing \EEE
\begin{align*}
\alpha(\delta) := \frac{\pi(d-2)}{2\delta d},
\end{align*}
estimate \eqref{boundarylowerbound-new} can be written as \EEE
\begin{align}\label{Ebndestimate}
\begin{split}
\mathcal{E}^{\mathrm{bnd}}(C_n) &\geq -\delta d + \delta dV_{\mathrm{r}}\Big(2\sin\Big(\frac{\pi(d-2)}{2\delta d}\Big)\Big) = \delta d\Big(
V_{\mathrm{r}}\Big(2\sin\Big(\alpha(\delta)\Big)\Big)-1\Big).
\end{split}
\end{align}
We obtain (\ref{BoundaryEnergyEstimate}) by minimizing the right hand side of (\ref{Ebndestimate}) with respect to $\delta$. To see this, \EEE set $\delta_0 = \frac{3}{2}-\frac{3}{d}$. For $\delta \leq \delta_0$ we have
\begin{align*}
\alpha(\delta) \geq \alpha(\delta_0) = \frac{\pi}{3}.
\end{align*}
By $[\mathrm{vii}]$ \EEE we get $V_{\mathrm{r}}( 2\sin(\alpha(\delta)))=0$ for all $1\leq \delta\leq \delta_0$\EEE. Therefore, we find
\begin{align}\label{deltageqdelta1}
\delta d\Big( \EEE V_{\mathrm{r}}\Big(2\sin\Big(\alpha(\delta)\Big)\Big)-1\Big)= -\delta d\geq -\delta_0 d=-\Big(\frac{3}{2}d-3\Big)
\end{align}
and we obtain estimate (\ref{BoundaryEnergyEstimate}) \EEE for $\delta\leq \delta_0$. Now for $\delta >\delta_0$, we have $\alpha(\delta) < \alpha(\delta_0)$. By $(\mathrm{v})$ \EEE we get
$$V_{\mathrm{r}}(2\sin(\alpha(\delta))) \geq V_{\mathrm{r}}(2\sin(\alpha(\delta_0))) + 2V_{\mathrm{r},-}^\prime(2\sin(\alpha(\delta_0))(\sin(\alpha(\delta))-\sin(\alpha(\delta_0)).$$
Then by \EEE the fact that $\sin(\theta)$ is concave for $\theta \in [0,\pi]$, $V_{\mathrm{r},-}^\prime(\sqrt{3}) < - 3/\pi<0$ by $[\mathrm{viii}]$, $\alpha(\delta_0) = \frac{\pi}{3}$, \EEE and $\alpha(\delta)-\alpha(\delta_0) < 0$ \EEE we derive
\begin{align}\label{deltageqdelta0}
\begin{split}
V_{\mathrm{r}}(2\sin(\alpha(\delta))) &\geq V_{\mathrm{r}}(\sqrt{3}) + 2V_{\mathrm{r},-}^\prime (\sqrt{3})\cos(\alpha(\delta_0))\big(\alpha(\delta)-\alpha(\delta_0)\big) \\&=V_{\mathrm{r}}(\sqrt{3}) + V_{\mathrm{r},-}^\prime(\sqrt{3})\Big(\frac{\pi(d-2)}{2\delta d}-\frac{\pi}{3}\Big) > -\frac{3}{\pi}\Big(\frac{\pi(d-2)}{2\delta d}-\frac{\pi}{3}\Big) \\&=\frac{1}{\delta d}\Big(\delta d-\frac{3}{2}d+3\Big).
\end{split}
\end{align}
Here, we also used that $\cos(\alpha(\delta_0)) = \frac{1}{2}$ and $V_{\mathrm{r}}(\sqrt{3}) = 0$. \EEE From the previous calculation and (\ref{Ebndestimate}), estimate (\ref{BoundaryEnergyEstimate}) \EEE follows also for $\delta >\delta_0$. \EEE
We now show that we have strict inequality in (\ref{BoundaryEnergyEstimate}) if one of the conditions \eqref{eq: 1}-\eqref{eq: 3} is violated. First, we have a strict inequality in \eqref{boundarylowerbound-new} if a boundary bond is not of unit length and therefore also in \eqref{Ebndestimate}. (Recall that we defer the proof of \eqref{boundarylowerbound-new} and its strict version to the end of the proof.) \EEE If \eqref{eq: 2} is violated, we find $\delta \neq \delta_0$ after a short computation. Then we obtain strict inequalities from \eqref{deltageqdelta1} and \eqref{deltageqdelta0}, respectively. Finally, let use suppose that \eqref{eq: 3} is violated. We can assume that $\delta=\delta_0$ and \eqref{eq: 1}-\eqref{eq: 2} hold as otherwise the inequality in (\ref{BoundaryEnergyEstimate}) is strict. If \EEE equality holds in (\ref{BoundaryEnergyEstimate}), then equality also holds in \eqref{boundarylowerbound-new}. As $V_{\rm r}(2\sin(\alpha(\delta_0)))= 0$, \EEE this implies
\begin{align*}
\sum_{ x_i \in I_2} V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i}{2}\Big)\Big) + \sum_{x_i \in I_3} \Big( V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i^1}{2}\Big)\Big) + V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i^2}{2}\Big)\Big) \Big)=0.
\end{align*}
In view of ${\rm [iv]}$ and ${\rm [vii]}$\EEE, this gives
\begin{align}\label{eq: anglerange}
\theta_i \in [\tfrac{2\pi}{3},\tfrac{4\pi}{3}] \ \ \ \text{for all $x_i \in I_2$}, \ \ \ \ \ \theta^1_i, \theta^2_i \in [\tfrac{2\pi}{3},\tfrac{4\pi}{3}] \ \ \ \text{for all $x_i \in I_3$}.
\end{align}
Under the assumption that \eqref{eq: 3} is violated, using \eqref{eq: anglerange} and $\theta_i = \theta_i^1 + \theta_i^2$ we find some $x_i \in I_3$ with $\theta_i > \frac{4\pi}{3}$ or some $x_i \in I_2$ with $\theta_i > \frac{2\pi}{3}$. Then \eqref{eq: anglerange} implies \EEE
\begin{align*}
\pi(d-2) = \EEE \frac{4\pi}{3}\Big(\frac{1}{2}d-3\Big)+ \frac{2\pi}{3}\Big(\frac{1}{2}d +3\Big) =\frac{4\pi}{3}\#I_3 + \frac{2\pi}{3}\#I_2 < \sum_{x_i \in I_2} \theta_i + \sum_{x_i \in I_3} \theta_i = \pi(d-2),
\end{align*}
where the last step follows from the fact that the maximal polygon has $d$ vertices. \EEE This is a contradiction and shows strict inequality in \eqref{BoundaryEnergyEstimate} if \eqref{eq: 3} is violated.
To complete the proof, it remains to show \eqref{boundarylowerbound-new} and its strict version. \EEE In the case of a \EEE $2$-bonded $x_i$, define $r^1_i = |x_i-x_{i-1}|$, $r^2_i = |x_i-x_{i+1}|$. In the case of a \EEE $3$-bonded $x_i$, define additionally \EEE $r^3_i = |x_i-x_{i}^b|$.
By the cosine rule we obtain
\begin{align}\label{xidistance}
\begin{split}
&|x_{i+1}-x_{i-1}|= \ell(\theta_i,r_i^1,r_i^2), \ \ \ |x_i^b-x_{i-1}| = \ell(\theta_i^1,r_i^1,r_i^3), \ \ \ |x_i^b-x_{i+1}| = \ell(\theta_i^2,r_i^2,r_i^3),
\end{split}
\end{align}
where we have used the shorthand
\begin{align}\label{eq: ell}
\ell(\theta,r_1,r_2) = \sqrt{r_1^2 + r_2^2 - 2r_1r_2 \cos(\theta)}.
\end{align}
\EEE We want to prove that for every boundary atom $x_i$ its contribution to the energy can be controlled by the energy contribution in a modified configuration which has \EEE the same angles but unit bond lengths instead of $r_i^1, r_i^2$. Recall that \EEE by $[\mathrm{viii}]$ we have for all $r \in (1,r_0]$ \EEE
\begin{align}\label{eq: strict inequality}
\frac{1}{2(r-1)}(V_{\mathrm{a}}(1)-V_{\mathrm{a}}(r)) < V'_{\mathrm{r},+}(1). \EEE
\end{align}
Let $\theta \in [\pi/3,5\pi/3]$, $r_1,r_2 \in [1,r_0]$. \EEE Then $\ell(\theta,r_1,r_2)\geq 1$, $V_{\mathrm{r},+}'(r)\leq 0$ for $r \geq 1$ and $V_{\mathrm{r},+}'$ is monotone increasing \EEE due to the convexity assumption in $(\mathrm{v})$ \EEE on $V_{\mathrm{r}}$. Moreover, we have $\partial_{r_1} \ell(\theta,r_1,r_2)\leq 1$ by an elementary computation. (This can also be seen by a geometric argumentation: by changing the length of $r_1$, the length change of $\ell$ is always smaller or equal to \EEE the length change of $r_1$, equal only if $\theta\in\lbrace 0 , \pi \rbrace$.) We therefore obtain for $r, s \in [1,r_0]$ \EEE
\begin{align*}
\frac{1}{2(r-1)}(V_{\mathrm{a}}(1)-V_{\mathrm{a}}(r)) \le V_{\mathrm{r},+}'(1) \EEE \leq V_{\mathrm{r},+}'(\ell(\theta,s,r_2)) \leq V_{\mathrm{r},+}'(\ell(\theta,s,r_2)) \partial_{ r_1 }\ell(\theta,s,r_2).
\end{align*}
Integrating this from $1$ to $r$ in the variable $s$ and using the fundamental theorem of calculus, we get \EEE
\begin{align*}
\frac{1}{2}\left( V_{\mathrm{a}}(1)-V_{\mathrm{a}}(r) \right)\leq \int_{1}^r V_{\mathrm{r},+}'(\ell(\theta,s,r_2)) \partial_{r_1} \EEE \ell(\theta,s,r_2) \mathrm{d}s = V_{\mathrm{r}}(\ell(\theta,r,r_2))-V_{\mathrm{r}}(\ell(\theta,1,r_2)).
\end{align*}
Applying this estimate in the second \EEE as well as in the third \EEE component of $\ell(\theta,r_1,r_2)$ with $1$ \EEE and $r_2$ respectively, we derive \EEE
\begin{align}\label{Estimateunitlength}
V_{\mathrm{a}}(1) + V_{\mathrm{r}}(\ell(\theta,1,1)) \leq \frac{1}{2}\left( V_{\mathrm{a}}(r_1)+V_{\mathrm{a}}(r_2)\right) +V_{\mathrm{r}}(\ell(\theta,r_1,r_2)).
\end{align}
Note that, due to ${\rm [i]}$, \eqref{eq: neighborhood}, and Lemma \ref{lemma:bondangles}, for $2$-bonded $x_i $, we have $r_i^1,r_i^2 \in [1,r_0] $, $\theta_i \in [\pi/3,5\pi/3]$, and for $3$-bonded $x_i$ we have $r_i^1,r_i^2, r_i^3 \in [1,r_0]$, and $\theta_i^1,\theta_i^2 \in [\pi/3,5\pi/3]$. Now for all $2$-bonded $x_i$, using (\ref{Estimateunitlength}) with $r_i^1$, $r_i^2$, and $\theta_i$, \EEE we have
\begin{align}\label{Twobondedestimate}
\frac{1}{2}\left( V_{\mathrm{a}}(r_i^1) + V_{\mathrm{a}}(r_i^2)\right) + V_{\mathrm{r}}(\ell(\theta_i,r_i^1,r_i^2)) \geq V_{\mathrm{a}}(1) + V_{\mathrm{r}}(\ell( \theta_i, \EEE 1,1)).
\end{align}
On the other hand, for all \EEE $3$-bonded $x_i$, using (\ref{Estimateunitlength}) twice with $r_i^1$, $r_i^3 $, and $\theta_i^1$ and $r_i^2$, $r_i^3 $, and $\theta_i^2$\EEE, we have
\begin{align}\label{Threebondedestimate}
\begin{split}
&\frac{1}{2}\left( V_{\mathrm{a}}(r_i^1) +V_{\mathrm{a}}(r_i^3)\right) + V_{\mathrm{r}}(\ell(\theta_i^1,r_i^1,r_i^3)) \geq V_{\mathrm{a}}(1) + V_{\mathrm{r}}(\ell(\theta_i^1,1,1)), \\&
\frac{1}{2} \left(V_{\mathrm{a}}(r_i^2) + V_{\mathrm{a}}(r_i^3)\right) + V_{\mathrm{r}}(\ell(\theta_i^2,r_i^2,r_i^3)) \geq V_{\mathrm{a}}(1) + V_{\mathrm{r}}(\ell(\theta_i^2,1,1)).
\end{split}
\end{align}
Using (\ref{xidistance}), (\ref{Twobondedestimate})-(\ref{Threebondedestimate}), ${\rm [ii]}$, and $V_{\rm r} \ge 0$ we obtain by \eqref{eq: boundary energy}
\begin{align}\label{boundaryestimate1}
\mathcal{E}^{\mathrm{bnd}}(C_n) &\geq \sum_{i=1}^d \Big(\frac{1}{2}\big( V_{\mathrm{a}}(|x_i-x_{i+1}|)+V_{\mathrm{a}}(|x_i-x_{i-1}|)\big)+V_{\mathrm{r}}(|x_{i+1}-x_{i-1}|) \Big)\notag \\&\quad+ \sum_{x_i \in I_3}\Big( V_{\mathrm{a}}(|x_i-x_i^b|) + V_{\mathrm{r}}(|x_{i-1}-x_i^b|) + V_{\mathrm{r}}(|x_{i+1}-x_i^b|) \Big)\notag \\& = \EEE \sum_{x_i \in I_2} \Big(\frac{1}{2}\big(V_{\mathrm{a}}(r_i^1) + V_{\mathrm{a}}(r_i^2)\big) + V_{\mathrm{r}}(\ell(\theta_i,r_i^1,r_i^2)) \Big) + \sum_{x_i \in I_3} V_{\rm r}(|x_{i+1}-x_{i-1}|) \EEE \notag \\&\quad+\sum_{x_i \in I_3} \Big(\frac{1}{2}\big(V_{\mathrm{a}}(r_i^1)+V_{\mathrm{a}}(r_i^2) + 2 V_{\mathrm{a}}(r_i^3)\big) + V_{\mathrm{r}}(\ell(\theta_i^1,r_i^1,r_i^3))+ V_{\mathrm{r}}(\ell(\theta_i^2,r_i^2,r_i^3)) \Big) \notag\\&\geq -(\#I_2+2\#I_3) + \sum_{i \in I_2} V_{\mathrm{r}}(\ell(\theta_i,1,1)) + \sum_{i \in I_3} \sum_{j=1,2} V_{\mathrm{r}}(\ell(\theta_i^j,1,1)).
\end{align}
For later purposes, we remark that this inequality is strict if one bond has not unit length. This follows from the strict inequality in \eqref{eq: strict inequality}. \EEE
Recall $\delta= \frac{\#I_2+2\#I_3}{d}$ and note that $\ell(\theta,1,1) = 2\sin(\theta/2)$ by \eqref{eq: ell}. Using \EEE $\theta_i = \theta_i^1 + \theta_i^2$ for $x_i \in I_3$ and \eqref{boundaryestimate1} \EEE we obtain
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) &\geq -\delta d + \sum_{x_i \in I_2} V_{\mathrm{r}}(\ell(\theta_i,1,1)) + \sum_{x_i \in I_3} \sum_{j=1,2}V_{\mathrm{r}}(\ell(\theta_i^j,1,1)) \EEE \notag \\& = -\delta d + \sum_{x_i \in I_2} V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i}{2}\Big)\Big) + \sum_{x_i \in I_3} \sum_{j=1,2} V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_i^j}{2}\Big)\Big). \EEE
\end{align*}
This yields the first inequality in \eqref{boundarylowerbound-new}. We note that this inequality is strict if one bond has not unit length since then \eqref{boundaryestimate1} is strict.
The second inequality in \eqref{boundarylowerbound-new} follows by a convexity argument: \EEE
\EEE since $V_{\mathrm{r}}$ is convex and non-increasing by ${\rm [v]}$ \EEE and $\sin(\theta/2)$ is concave for $\theta \in [0,2\pi]$, we have for all $\lambda \in [0,1]$ \EEE
\begin{align*}
\lambda V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_1}{2}\Big)\Big)+ (1-\lambda)V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta_2}{2}\Big)\Big) &\geq V_{\mathrm{r}}\Big(\lambda 2\sin\Big(\frac{\theta_1}{2}\Big) + (1-\lambda) 2\sin\Big(\frac{\theta_2}{2}\Big)\Big) \\& \geq V_{\mathrm{r}}\Big( 2\sin\Big(\frac{\lambda\theta_1+(1-\lambda)\theta_2}{2}\Big)\Big).
\end{align*}
Hence, $\theta \mapsto V_{\mathrm{r}}\Big(2\sin\Big(\frac{\theta}{2}\Big)\Big)$ is a convex function for $\theta \in [0,2\pi]$. This together with \EEE the fact that $\# I_2 + 2 \# I_3 = \delta d$ and
$$
\pi(d-2)=\sum_{i=1}^d \theta_i = \sum_{x_i \in I_2} \theta_i + \sum_{x_i \in I_3}(\theta_i^1+\theta_i^2)
$$
yields the second inequality in \eqref{boundarylowerbound-new}. This \EEE concludes the proof. \EEE
\end{proof}
\begin{remark}\label{rem: angle/bond}
(i) We briefly explain assumption [viii] from a technical point of view. The condition prevents two phenomena concerning surface relaxation: the first is the occurrence of more atoms on the boundary of the configuration than one would expect for hexagonal configurations. This is achieved \EEE by the first condition of [viii], cf.\ \eqref{deltageqdelta0}. The second phenomenon is the presence of elastically deformed boundary bonds. This is prevented by the second condition, cf.\ \eqref{Estimateunitlength}-\eqref{Threebondedestimate}.
(ii) \EEE At this stage, let us highlight the difference of our analysis to \cite{Mainini}. In \cite{Mainini}, an empirical angular potential is considered which penalizes deviations of the bond-angles from $\frac{2\pi}{3}$, modeling covalent bonding for carbon nanostructures. In our model for ionic compounds, the energy contribution can also be expressed in terms of the bond-angle. More precisely, by \eqref{eq: ell} we have energy contributions of the form
$$V_{\rm r} \big( \ell(\theta,r_1,r_2) \big) = V_{\rm r}\Big(\sqrt{r_1^2 + r_2^2 - 2r_1r_2 \cos(\theta)}\Big).$$
In view of assumption [vii], only lengths $\ell(\theta,r_1,r_2) < \sqrt{3}$ are penalized. In particular, as $r_1,r_2 \ge 1$, bond-angles $\theta \ge \frac{2\pi}{3}$ never penalize the energy and, if the bond lengths $r_1$ and $r_2$ exceed one, also bond-angles less than $\frac{2\pi}{3}$ might not penalize the energy.
In principle, this implies more geometric flexibility of ground-state configurations with respect to \cite{Mainini}. This calls for refined arguments for controlling the boundary energy and, in particular, for characterizing the ground-state geometries in Section \ref{sec: characterization}. Let us highlight that, in spite of the weaker penalization of bond-angles deviating from $\frac{2\pi}{3}$, it is still possible to prove that ground states assemble themselves in the hexagonal lattice.
\end{remark}
We recall the definition of $b$ in \eqref{eq:b}. Recall also \EEE the excess of edges $\eta = \sum_{j\geq 6} (j-6) f_j $ introduced in \eqref{Excess}, where $f_j$ denotes the number of elementary polygons \EEE with $j$ vertices in the bond graph. \EEE By Lemma \ref{LemmaPolygon} for any ground state we have that $\eta \in 2\mathbb{N}$. Moreover, \EEE $\eta=0$ if only if the bond graph consists of hexagons only. We now use this notion to estimate the cardinality of the bulk. \EEE It will turn out useful in Section \ref{sec: characterization} to exclude the existence \EEE of other elementary polygons \EEE than hexagons. \EEE
\begin{lemma}[Cardinality of the bulk\EEE]\label{LemmaBoundaryestimate} Suppose that $C_n$ is a connected ground state \EEE and that it does not contain any acyclic bonds. Then
\begin{align*}
n-d = 4b \EEE + 6+\eta - 5n.
\end{align*}
\end{lemma}
\begin{proof} By $f_j$ we denote the number of elementary polygons \EEE in the bond graph with $j$ vertices and set $f = \sum_{j \ge 3} f_j$. Note that $f$ is the number of faces of the bond graph (omitting the exterior face). \EEE We have
\begin{align*}
\sum_{j \ge 3} jf_j \EEE = 2 b\EEE -d,
\end{align*}
since by the summation on the left all bonds contained in the maximal polygon are counted only once whereas all other bonds \EEE are counted twice. From Lemma \ref{LemmaPolygon} and the definition of $\eta$ \EEE we obtain
\begin{align*}
6 f = \EEE 2 b \EEE -d-\eta.
\end{align*}
Using this together with Euler's formula $n- b + f \EEE =1$ (omitting the exterior face) we get
\begin{align*}
n-d = 4b \EEE +6+\eta - 5n.
\end{align*}
\end{proof}
\section{Characterization of the ground-state energy}\label{sec: characterization-energy}
This section is devoted to the proof of Theorem \ref{TheoremEnergyGroundstates}. We only need to provide a lower bound on the ground-state energy since the upper bound has already been obtained by an explicit construction, see Proposition \ref{PropositionDaisy}.
\EEE
We state two algebraic lemmas that will be used in the sequel.
\begin{lemma}\label{LemmaSquareroot}
Let $j,n,m \in \mathbb{N}$ and let \EEE $x \in \mathbb{R} $ satisfy
\begin{align*}
\frac{m}{4} - \frac{5}{4}n \EEE \ge x \geq -\frac{3}{2}n + j +\sqrt{\frac{3}{2}(-4x-5n+m)}.
\end{align*}
Then $x \geq -\frac{3}{2}n +j -3 + \sqrt{\frac{3}{2}(-4j+m+n+6)} $.
\end{lemma}
\begin{proof} The proof is elementary: we note \EEE that the function
\begin{align*}
x \mapsto x +\frac{3}{2}n - j -\sqrt{\frac{3}{2}(-4x-5n+m)}
\end{align*}
is strictly increasing \EEE and vanishes for $x =-\frac{3}{2}n +j -3 + \sqrt{\frac{3}{2}(-4j+m+n+6)}$.
\end{proof}
We use the following properties of the function $\beta$ which has been defined in \eqref{eq: beta definition}. \EEE
\begin{lemma}\label{LemmaPropertiesbeta} The function $\beta : \mathbb{N} \to \mathbb{R}$ satisfies
\begin{itemize}
\item[1)] $\lfloor \beta(n-1)\rfloor +1 \leq \lfloor \beta(n)\rfloor$.
\item[2)] $\lfloor \beta(m)\rfloor + \lfloor \beta(n-m)\rfloor +1 \leq \lfloor \beta(n)\rfloor$ for all $n\geq 12$, $n \geq m \geq 6$ and equality holds if and only if $n=12$ and $m=6$.
\item[3)] $\lfloor \beta(n)\rfloor \geq \lfloor \beta(n-k)\rfloor +2 + k$ for all $n \ge 13$ and $ n \ge k \ge 6$. \EEE
\item[4)] $\lfloor \beta(n)\rfloor \geq \lfloor \beta(n-5)\rfloor + 7$ for all $n \geq 13$ except for $n \in \lbrace 15,18,21,29\rbrace$. \EEE
\end{itemize}
\end{lemma}
\begin{proof}
The proof of 1) and 2) is elementary \EEE and can be found in \cite[Lemma 6.4, 6.5]{Mainini}. It relies on monotonicity and convexity properties of $\beta$. As a preparation for 3) and 4), we observe that for $n \geq 41$, $k = 5$ or for $n \ge 17$, $k=6$ or for $n \ge 13$, $n \ge k \ge 7$ one has
\begin{align}\label{eq: help}
\frac{3}{2}n -\sqrt{\frac{3}{2}n}\geq \frac{3}{2}(n-k) -\sqrt{\frac{3}{2}(n-k)} + 2+ k.
\end{align}
Indeed, after some manipulations, we see that this is equivalent to $\sqrt{\frac{3}{2}n} +\sqrt{\frac{3}{2}(n- k ) \EEE } \geq 3k/(k-4)\EEE$. The latter holds true for $n \geq 41$, $k = 5$ or for $n \ge 17$, $k=6$ or for $n \ge 13$, $n \ge k \ge 7$.
One can check directly that $\lfloor \beta(n)\rfloor = \lfloor \beta(n-6)\rfloor +8$ for $n =13,\ldots,16$, see Table \ref{table2}. \EEE This together with \eqref{eq: help} yields 3). Property 4) follows from \eqref{eq: help} and an explicit computation for the cases $13 \le n \le 40$, see Table \ref{table}. \EEE
\end{proof}
\begin{table}[h]\centering
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$n$ & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 &15 \\ \hline
$\lfloor\beta(n)\rfloor$ & 1 & 2 & 3&4&6&7&8&9&11&12&13&15&16&17\\
\hline
\hline
$n$ & 16 &17 &18 &19 &20 &21 & 22 & 23 & 24 &25 & 26 &27 &28 &29 \\
\hline
$\lfloor\beta(n)\rfloor$ &19 &20&21&23&24&25&27&28&30&31&32&34&35&36 \\
\hline
\end{tabular}
\vspace{8mm}
\caption{The function \EEE $\lfloor\beta(n)\rfloor$ for $2 \leq n \leq 29$. The table together with Theorem \ref{TheoremEnergyGroundstates} \EEE can be used to see that the configurations in Fig.~\ref{FigureFlexible}, Fig.~\ref{FigureBridge}, and Fig.~\ref{FigureOctagons} are ground states.\EEE} \label{table2}
\end{table}
\begin{table}[h]\centering
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$n$ & 13 & 14 &15 & 16 &17 &18 &19 &20 &21 & 22 & 23 & 24 &25 & 26 \\ \hline
$\gamma(n)$ & 7&7&6&7&7&6&7&7&6&7&7&7&7&7\\
\hline
\end{tabular}
\vspace{0.8cm}
\hspace{0.02cm}
\vspace{0.3cm}
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$n$ &27 &28 &29 &30 &31 & 32 & 33 & 34 &35 & 36 &37 &38 &39 &40 \\
\hline
$\gamma(n)$ &7&7&6&7&7&7&7&7&7&7&7&7&7&7 \\
\hline
\end{tabular}
\vspace{-3mm}
\caption{The function $\gamma(n) := \lfloor \beta(n) \rfloor - \lfloor \beta(n-5)\rfloor$ for $13 \le n \le 40$.\EEE} \label{table}
\end{table}
Before we proceed with the proof of Theorem \ref{TheoremEnergyGroundstates}, we will consider the cases \EEE $1 \le n \le 6$ which will serve as the induction base. \EEE
\begin{lemma}[Cases $1 \le n \le 6$]\label{lemma: small n}
For $1 \le n \le 6$ every ground state $C_n$ is connected and satisfies $\mathcal{E}(C_n) = -b = \lfloor \beta(n) \rfloor$.
\end{lemma}
\begin{proof}
Let $n \leq 5$. By Lemma \ref{LemmaPolygon} we have that the bond graph does not contain any polygon \EEE with less than or equal to $5$ edges. \EEE Hence, the bond graph is cycle free and $b \leq n-1$. This provides the lower bound $
\mathcal{E}(C_n) \geq -(n-1)$ by Remark \ref{rem: repulsionsfree}. \EEE Clearly, one can construct configurations with $n$ atoms and $n-1$ bonds having alternating charge distribution. Note also that $b = n-1 = \lfloor \beta(n) \rfloor\EEE$ is only possible if the configuration is connected.
In the case $n=6$, note that the number of polygons $f$ satisfies $f \le 1$ since by Lemma \ref{LemmaPolygon} we have that every polygon has at least $6$ edges. \EEE By Euler's formula we get $6-b+f \geq 1$ where the inequality is due to the fact that we may have more than one connected component. This implies $b \leq 6$ and thus $\mathcal{E}(C_6) \geq -b$ by Remark \ref{rem: repulsionsfree}. Exactly a regular hexagon with alternating charge distribution and unit bond length yields a configuration with energy equal to $-6 = -\lfloor \beta(6) \rfloor$. This concludes the proof. \EEE
\end{proof}
We now proceed with the proof of Theorem \ref{TheoremEnergyGroundstates}. We follow the strategy devised in \cite[Theorem 6.1]{Mainini} with the adaptions needed due to the presence of different atomic types and repulsive potentials \EEE instead of angle potentials. In contrast to \cite{Mainini}, however, we split the proof of the characterization of the ground-state energy and the characterization of the ground states. Indeed, the latter is more involved in our setting and the investigation is deferred to Section \ref{sec: characterization}.
\begin{proof}[Proof of Theorem \ref{TheoremEnergyGroundstates}] We start by noting that every ground state $C_n$ has alternating charge distribution by Lemma \ref{LemmaNeighborhood}. \EEE By Proposition \ref{PropositionDaisy} we have that the ground-state energy satisfies
\begin{align}\label{Groundstateineq1}
\mathcal{E}(C_n) \leq -\lfloor \beta(n)\rfloor.
\end{align}
We proceed by induction. Suppose that the statement has been proven for all $m < n$ (for $1 \le m \le 6$ see Lemma \ref{lemma: small n}). We first show connectedness of the ground state (Claim 1) and then the energy equality (Claim 2). \EEE
\textbf{Claim 1:} $C_n$ is connected.
\noindent \emph{Proof of Claim 1}: \EEE Assume by contradiction that $C_n$ was \EEE not connected, i.e., \EEE $C_n$ consists of two or more connected components. Let $C'_m$ and $C'_{n-m}$ be two sub-configurations consisting of $m$ and $n-m$ atoms, respectively, which do not have any bonds between them. \EEE If $m\geq 6, n\geq 12$, we can apply the induction hypothesis, Lemma \ref{LemmaPropertiesbeta} $2)$, and $V_{\rm r} \ge 0$ \EEE to get
\begin{align*}
\mathcal{E}(C_n) \ge \mathcal{E}(C_m') + \mathcal{E}(C_{n-m}') \EEE \geq -\lfloor\beta(n-m)\rfloor -\lfloor\beta(m)\rfloor >-\lfloor \beta(n)\rfloor.
\end{align*}
If \EEE $m< 6$, we can apply Lemma \ref{lemma: small n} \EEE and Lemma \ref{LemmaPropertiesbeta} $1)$ iteratively $m$ times \EEE to get
\begin{align*}
\mathcal{E}(C_n) \geq -\lfloor\beta(n-m)\rfloor - \lfloor\beta(m)\rfloor = -\lfloor\beta(n-m)\rfloor - (m-1)\EEE > -\lfloor \beta(n)\rfloor.
\end{align*}
The case $n<12$ is already included in this argument since \EEE at least one connected component consists of less than $6$ \EEE atoms. In \EEE view of (\ref{Groundstateineq1}), we obtain a contradiction to the fact that $C_n$ is a ground state.
\textbf{Claim 2:} Energy equality $\mathcal{E}(C_n) = -b = -\lfloor \beta(n) \rfloor$.
\noindent \emph{Proof of Claim 2}: \EEE We divide the proof into three steps. We first treat the case that $C_n$ contains acyclic bonds (Step 1). Afterwards, we consider only configurations $C_n$ without acyclic bonds and show $\mathcal{E}(C_n) = -b$ (Step 2) and $\mathcal{E}(C_n) = -\lfloor \beta(n) \rfloor$ (Step 3).
\emph{Step 1: $C_n$ contains acyclic bonds:} If there exist flags, we can find an atom $x_i$ such that removing $x_i$ removes exactly one flag. We can count the energy contribution of this flag by at least $-1$ and we estimate the energy of the rest of the configuration by induction. By Lemma \ref{LemmaPropertiesbeta} 1) we get
\begin{align*}
\mathcal{E}(C_n) \ge -1 + \mathcal{E}(C_n \setminus \lbrace (x_i,q_i)\rbrace) \EEE \geq -1 -\lfloor \beta(n-1)\rfloor \geq -\lfloor \beta(n)\rfloor.
\end{align*}
Equality also shows that $C_n \setminus \lbrace (x_i,q_i)\rbrace$ has $\lfloor \beta(n-1) \rfloor$ bonds by induction and $C_n$ has $\lfloor \beta(n-1) \rfloor+1 = \lfloor \beta(n) \rfloor$ bonds. \EEE
We now suppose that a bridge exists. \EEE Consider the two sub-configurations $C'_m$ and $C'_{n-m}$ which are connected by the bridge. By the definition of bridges we have that both $C'_m$ and $C'_{n-m}$ contain at least one simple cycle and therefore, by Lemma \ref{LemmaPolygon}, $m,n-m \geq 6$. The energy contribution of the bridge is greater or equal to $-1$. Using the induction assumption and Lemma \ref{LemmaPropertiesbeta} 2) we get
\begin{align*}
\mathcal{E}(C_n) \ge \mathcal{E}(C'_m) + \mathcal{E}(C'_{n-m}) - 1 \EEE \ge -\lfloor \beta(m)\rfloor - \lfloor \beta(n-m)\rfloor -1 \geq -\lfloor \beta(n)\rfloor.
\end{align*}
As before, equality also implies that $C_n$ has $\lfloor \beta(m)\rfloor + \lfloor \beta(n-m)\rfloor +1 = \lfloor \beta(n) \rfloor$ bonds.
\emph{Step 2: $\mathcal{E}(C_n) = -b$ for connected $C_n$ with no acyclic bonds:} Suppose by contradiction that \EEE the statement was false. Then, by Remark \ref{rem: repulsionsfree}, \EEE there exist $x_1,x_2 \in X_n$ such that $q_1=q_2$ and $|x_1-x_2|<\sqrt{3}$ or $q_1 =-q_2$ and $1<|x_1-x_2|\leq r_0$. If \EEE $x_1 \in \partial X_n$ or $x_2 \in \partial X_n$, by \EEE using Lemma \ref{LemmaBoundaryEnergy} and Remark \ref{rem: bdy}(ii) \EEE we have the strict inequality \EEE
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) > -\frac{3}{2}d + 3
\end{align*}
and by induction assumption we have
\begin{align*}
\mathcal{E}^{\mathrm{bulk}}(C_n) \geq -\lfloor \beta(n-d)\rfloor.
\end{align*}
On the other hand, if we have $x_1, x_2 \notin \partial X_n$, we calculate by Lemma \ref{LemmaBoundaryEnergy} and induction \EEE
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) \geq -\frac{3}{2}d + 3, \ \ \ \ \ \ \ \mathcal{E}^{\mathrm{bulk}}(C_n) > -\lfloor \beta(n-d)\rfloor.
\end{align*}
Here, the estimate for the bulk part is strict. Indeed, since in this case $C_n^{\rm bulk}$, which consists of $n-d$ particles, \EEE is not repulsion-free or has bonds longer than 1, \EEE by the induction assumption it cannot be a ground state, see Remark \ref{rem: repulsionsfree}. \EEE
In both cases it holds that
\begin{align*}
\mathcal{E}(C_n) > -\left\lfloor \frac{3}{2}n -\sqrt{\frac{3}{2}(n-d)}\right\rfloor +3.
\end{align*}
Since the right hand side is an integer, we obtain
\begin{align}\label{Integerinequality1}
-(\lfloor -\mathcal{E}(C_n)\rfloor +1) \geq -\frac{3}{2}n +\sqrt{\frac{3}{2}(n-d)} + 3.
\end{align}
Recall that we assumed $\mathcal{E}(C_n) > -b$ by contradiction, \EEE which implies $-(\lfloor-\mathcal{E}(C_n)\rfloor+1) \geq -b$. \EEE Now by Lemma \ref{LemmaBoundaryestimate} we obtain
\begin{align*}
n-d \geq 4(\lfloor-\mathcal{E}(C_n)\rfloor+1) + 6 -5n,
\end{align*}
where we used that $\eta \ge 0$. \EEE Using the above inequality and (\ref{Integerinequality1}) we obtain
\begin{align*}
-(\lfloor -\mathcal{E}(C_n)\rfloor +1) \geq -\frac{3}{2}n +\sqrt{\frac{3}{2}( 4 \EEE (\lfloor-\mathcal{E}(C_n)\rfloor+1) + 6 -5n)} + 3.
\end{align*}
Now we can use Lemma \ref{LemmaSquareroot} with $j=3$, $m=6$, and $x = -(\lfloor -\mathcal{E}(C_n)\rfloor +1)$ \EEE to obtain
\begin{align*}
-(\lfloor -\mathcal{E}(C_n)\rfloor +1) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}n}.
\end{align*}
The last inequality implies $\mathcal{E}(C_n) > -\lfloor \beta(n) \rfloor$ \EEE contradicting \eqref{Groundstateineq1}.
\emph{Step 3: $\mathcal{E}(C_n) = -\lfloor \beta(n) \rfloor$ for connected $C_n$ with no acyclic bonds:} \EEE
Due to (\ref{Groundstateineq1}), it suffices to prove $\mathcal{E}(C_n) \geq -\lfloor \beta(n)\rfloor$. Again we proceed by induction. By Lemma \ref{LemmaBoundaryestimate} and the induction assumption \EEE we obtain
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) \geq -\frac{3}{2}d+3,\quad \mathcal{E}^{\mathrm{bulk}}(C_n) \geq -\lfloor \beta(n-d)\rfloor.
\end{align*}
This gives
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n +\sqrt{\frac{3}{2}(n-d)} + 3.
\end{align*}
By Lemma \ref{LemmaBoundaryestimate} and Step 2 \EEE we obtain $n-d \geq - 4 \mathcal{E}(C_n) \EEE + 6 -5n$.
This yields
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}(-4\mathcal{E}(C_n)-5n +6)} + 3.\EEE
\end{align*}
Applying Lemma \ref{LemmaSquareroot} with $j=3$, $m=6$, and $x = \mathcal{E}(C_n)$ \EEE we obtain $\mathcal{E}(C_n) \geq -\beta(n)$. \EEE Finally, since $\mathcal{E}(C_n)$ is an integer due to Step 2, we conclude $\mathcal{E}(C_n) \ge -\lfloor \beta(n) \rfloor$. \EEE
\end{proof}
For later purposes, we observe that the calculations of Step 2 and Step 3 in the previous proof can be refined. Recall the definition of the excess of edges $\eta$ in \eqref{Excess}. \EEE
\begin{lemma}[Refined energy inequality for $\eta$]\label{lemma: eta energy}
Let $n \ge 6$ and let $C_n$ be a ground state with no acyclic bonds. \EEE Then
$$ \mathcal{E}(C_n) \ge -\frac{3}{2}n + \sqrt{\frac{3}{2}(n+\eta)}.$$
If $\mathcal{E}^{\rm bnd}(C_n) > -\frac{3}{2}d + 3$ or $\mathcal{E}^{\rm bulk}(C_n) > \lfloor \beta(n-d)\rfloor$, then
$$ \mathcal{E}(C_n) \ge -\frac{3}{2}n + \sqrt{\frac{3}{2}(n+\eta-4)}+1.$$
\end{lemma}
\begin{proof}
By Lemma \ref{LemmaBoundaryEnergy} and Theorem \ref{TheoremEnergyGroundstates} applied on $C_n^{\mathrm{bulk}}$ we have
\begin{align}\label{eq: strict?}
\mathcal{E}^{\mathrm{bnd}}(C_n) \ge -\frac{3}{2}d +3, \ \ \ \ \ \mathcal{E}^{\mathrm{bulk}}(C_n) \geq -\lfloor \beta(n-d)\rfloor.
\end{align}
By Lemma \ref{LemmaBoundaryestimate} and Theorem \ref{TheoremEnergyGroundstates} we get $n-d = -4\mathcal{E}(C_n) + 6 + \eta -5n$. This together with the summation of the two terms in \eqref{eq: strict?} yields \EEE
\begin{align}\label{eq: strict?2}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}(-4\mathcal{E}(C_n) -5n+6+\eta)} +3.
\end{align}
By applying Lemma \ref{LemmaSquareroot} with $j=3$ and $m=6+\eta$ we obtain
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}(n+\eta)}.
\end{align*}
Finally, if $\mathcal{E}^{\rm bnd}(C_n) > -\frac{3}{2}d + 3$ or $\mathcal{E}^{\rm bulk}(C_n) > \lfloor \beta(n-d)\rfloor$, i.e., one inequality in \eqref{eq: strict?} is strict, we can replace $3$ by $4$ in \eqref{eq: strict?2} since $\mathcal{E}(C_n)$ is an integer. Then applying again Lemma \ref{LemmaSquareroot} with $j=4$, $m=6+\eta$ we obtain
$
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}(n+\eta-4)} +1.
$
\end{proof}
\section{Characterization of ground states}\label{sec: characterization}
In this section we characterize the ground states of (\ref{Energy}). We do not provide a complete characterization for $n < 10$ since the system is highly flexible in those cases. Some of the ground states for $n<10$ are pictured in Fig.~\ref{FigureFlexible}. We will start by providing some geometric facts about ground states. Afterwards, \EEE we formulate and prove the first main result of the section which shows \EEE that ground states consist of hexagonal cycles except for possibly (at most) two flags or one octagon at the boundary, see Proposition \ref{TheoremGroundstatesleq-new}. \EEE Finally, for $n \ge 30$, we will be able to prove \EEE that no octagons occur which will conclude \EEE the proof of Theorem \ref{TheoremGroundstatesleq31}. The proof will also show that, among the ground states for $10 \le n \le 29$, an octagon can only occur in the cases $n=12, 15,18,21,29$, \EEE see Remark \ref{rem: octogons} and Fig.~\ref{FigureOctagons}. \EEE
\begin{remark}\label{rem: method}
{\normalfont
We briefly remark that in the following our strategy deviates considerably from the one in \cite{Mainini} due to the different modeling assumptions concerning repulsive and angular potentials, see Remark \ref{rem: angle/bond} for details. On the one hand, for $n \le 29$ indeed more flexible structures may occur \EEE which are not subsets of the hexagonal lattice. On the other hand, although we eventually will prove that for $n \ge 30$ ground states essentially assemble themselves in the hexagonal lattice, we need a different approach compared to \cite{HR, Mainini-Piovano, Mainini, Radin}: differently to the proof by induction performed there, we cannot use the property that ground states are subsets of the hexagonal lattice in the induction hypothesis (consider, e.g., the step from $29$ to $30$). \EEE Therefore, finer geometric considerations are necessary which are developed in two steps: first, we prove by induction that the breaking of the hexagonal symmetry due to presence of non-hexagonal elementary \EEE polygons can only occur on the boundary. Then we show that for $n \ge 30$ the existence of such defects leads to an energy exceeding \eqref{Energygroundstates}.
}
\end{remark}
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=.8]
\draw[thin](0,0)--(3,0);
\foreach \j in {0,1}{
\draw[fill=black](2*\j,0) circle(.05);
\draw[fill=white](2*\j+1,0) circle(.05);
}
\begin{scope}[shift={(4,0)}]
\draw[thin](0,0)--++(70:1)--++(10:1)--++(-10:1)--++(-70:1)--++(70:-1)--++(190:1)--++(170:1)--++(110:1);
\draw[fill=white](0,0)++(70:1)circle(.05)++(10:1)++(-10:1)circle(.05)++(-70:1)++(70:-1)circle(.05)++(190:1)++(170:1)circle(.05)++(110:1);
\draw[fill=black](0,0)circle(.05)++(70:1)++(10:1)circle(.05)++(-10:1)++(-70:1)circle(.05)++(70:-1)++(190:1)circle(.05)++(170:1)++(110:1);
\end{scope}
\begin{scope}[shift={(10.5,-.75)}]
\draw[thin](0,0)--++(-60:1);
\draw[thin](120:1)++(180:1)--++(180:1);
\draw[thin](120:1)++(60:1)--++(60:1);
\draw[fill=white](0,0)++(-60:1)circle(.05);
\draw[fill=white](120:1)++(180:1)++(180:1)circle(.05);
\draw[fill=white](120:1)++(60:1)++(60:1)circle(.05);
\draw[thin](60:1)--++(120:1)--++(180:1)--++(240:1)--++(300:1)--++(0:1)--++(60:1);
\draw[fill=black](60:1)++(120:1)circle(.05)++(180:1)++(240:1)circle(.05)++(300:1)++(0:1)circle(.05);
\draw[fill=white](60:1)circle(.05)++(120:1)++(180:1)circle(.05)++(240:1)++(300:1)circle(.05);
\end{scope}
\begin{scope}[shift={(12,0)},rotate=-90]
\draw[thin](0,0)--++(30:1)--++(90:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1)--++(270:1)--++(330:1);
\draw[thin](0,0)++(150:1)++(90:1)++(90:1)--++(150:1);
\draw[fill=black](0,0)++(150:1)++(90:1)++(90:1)++(150:1)circle(.05);
\draw[fill=black](0,0)circle(.05)++(30:1)++(90:1)circle(.05)++(90:1)++(150:1)circle(.05)++(210:1)++(270:1)circle(.05)++(270:1)++(330:1);
\draw[fill=white](0,0)++(30:1)circle(.05)++(90:1)++(90:1)circle(.05)++(150:1)++(210:1)circle(.05)++(270:1)++(270:1)circle(.05)++(330:1);
\end{scope}
\end{tikzpicture}
\caption{Some ground states for $n<10$.}
\label{FigureFlexible}
\end{figure}
\subsection{Geometric properties of ground states}
In this section we collect some geometric properties of ground states. \EEE
We start with an elementary property.
\begin{lemma}[Bridges\EEE]\label{lemma: bridges}
Ground states for $n \ge 13$ do not contain bridges.
\end{lemma}
\begin{proof}
Suppose that a bridge exists. Consider the two sub-configurations $C'_m$ and $C'_{n-m}$ which are connected by the bridge. As the energy contribution of the bridge is greater or equal to $-1$, we get by Theorem \ref{TheoremEnergyGroundstates}, Lemma \ref{LemmaPropertiesbeta} 2), the fact that $n \ge 13$, and $V_{\rm r} \ge 0$ \EEE
\begin{align*}
\mathcal{E}(C_n) \ge -\lfloor \beta(m)\rfloor - \lfloor \beta(n-m)\rfloor -1 > -\lfloor \beta(n)\rfloor.
\end{align*}
This contradicts the fact that $C_n$ is a ground state.
\end{proof}
\begin{figure}[htp]
\begin{tikzpicture}
\draw[thin](0,0)--++(0:1)--++(60:1)--++(120:1)--++(180:1)--++(240:1)--++(300:1);
\draw[thin](2,0)++(60:1)++(300:1)--++(0:1)--++(60:1)--++(120:1)--++(180:1)--++(240:1)--++(300:1);
\draw[thin](1,0)++(60:1)--++(0:1);
\draw[fill=black](0,0)circle(.05)++(0:1)++(60:1)circle(.05)++(120:1)++(180:1)circle(.05)++(240:1)++(300:1);
\draw[fill=white](0,0)++(0:1)circle(.05)++(60:1)++(120:1)circle(.05)++(180:1)++(240:1)circle(.05)++(300:1);
\draw[fill=black](2,0)++(60:1)++(300:1)circle(.05)++(0:1)++(60:1)circle(.05)++(120:1)++(180:1)circle(.05)++(240:1)++(300:1);
\draw[fill=white](2,0)++(60:1)++(300:1)++(0:1)circle(.05)++(60:1)++(120:1)circle(.05)++(180:1)++(240:1)circle(.05)++(300:1);
\end{tikzpicture}
\caption{A ground state configuration for $n=12$ containing a bridge.}
\label{FigureBridge}
\end{figure}
\begin{remark}\label{rem: bridge}
{\normalfont
Lemma \ref{lemma: bridges} is sharp in the sense that for $n=12$ there exists a ground state that contains a bridge connecting two hexagons (and then $\eta=0$)\EEE, cf.\ Table~\ref{table2} and Figure~\ref{FigureBridge}. Also note that for $n \le 11$ ground states cannot contain bridges as each polygon in the bond graph has at least $6$ vertices by Lemma \ref{LemmaPolygon}.
}
\end{remark}
\EEE
The next lemma states that the number of flags is at most two. \EEE Let us mention that this property also applies to the ground states of \cite{Mainini} although this has not been observed explicitly there. \EEE
\begin{lemma}[Flags\EEE] \label{LemmaOctagonflag} Let $n \geq 10$ and let $C_n$ be a ground state. Then the bond graph of $C_n$ contains at most $2$ flags.
\end{lemma}
\begin{proof}
Assume by contradiction \EEE that there exist $j \ge 3$ \EEE flags. Using the fact that a flag contributes at least $-1$ to the energy and applying \EEE Theorem \ref{TheoremEnergyGroundstates} on the sub-configuration obtained after removing the flags, we have $\mathcal{E}(C_n) \geq -j -\lfloor \beta(n-j)\rfloor$. By \eqref{eq: beta definition} we then get
\begin{align*}
\mathcal{E}(C_n) & \ge -\Big(\frac{3}{2}n -\frac{1}{2}j -\sqrt{\frac{3}{2}(n-j)}\Big) = -\Big(\frac{3}{2}n - \sqrt{\frac{3}{2}n} +\sqrt{\frac{3}{2}n}-\frac{1}{2}j -\sqrt{\frac{3}{2}(n-j)}\Big) \\& = -\Big( \frac{3}{2}n - \sqrt{\frac{3}{2}n} -\frac{1}{2}j +\frac{\frac{3}{2}j}{\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-j)}}\Big).
\end{align*}
The function $ f(j) := \EEE -\frac{1}{2}j +\frac{\frac{3}{2}j}{\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-j)}}$ is non-increasing in $j$, non-positive, and we have that $f(3) \leq -1$ \EEE for $n\geq 16$. With the above estimate this implies \begin{align*}
\mathcal{E}(C_n) > -\lfloor \beta(n)\rfloor
\end{align*}
which leads to a contradiction to Theorem \ref{TheoremEnergyGroundstates} in the cases $n\geq 16$. The cases $10 \leq n \leq 15 $ can be checked directly
by comparing the above formula $\mathcal{E}(C_n) \geq -j -\lfloor \beta(n-j)\rfloor, j\geq 3$, with $\lfloor\beta(n)\rfloor, n=10,\ldots,15$, cf.\ Table~\ref{table2}.
\end{proof}
Lemma \ref{LemmaOctagonflag} is sharp in the sense that for $n=9$ there exists a ground state that contains three flags in its bond graph, cf. Table~\ref{table2} and the third configuration in Fig.~\ref{FigureFlexible}. \EEE
\textbf{Equilibrated atoms:} We say an atom $x \in X_n$ is \emph{equilibrated} if all bond-angles at $x$ lie in $\lbrace \frac{2\pi}{3}, \frac{4\pi}{3} \rbrace$. By $\mathcal{A}$ we denote the atoms which are \emph{not} equilibrated. By $\mathcal{A}_{\rm bulk} \subset \mathcal{A} $ we denote the bulk atoms which are not equilibrated. Note that if $\mathcal{A}=\emptyset$ and $C_n$ \EEE is connected, then $X_n$ is a subset of the hexagonal lattice \EEE $\mathcal{L}$. The following properties will be useful in the sequel.
\begin{lemma}[Regular hexagons and bond-angles]\label{LemmaHexagon}
Let $C_n$ be a ground state.
Then all hexagons are regular with unit bond length and have \EEE alternating charge. All bond-angles $\theta$ satisfy $\frac{2\pi}{3} \le \theta \le \frac{4\pi}{3}$. If $x \in \mathcal{A}$, then $x$ is \EEE $2$-bonded and the bond angles lie in $(\frac{2\pi}{3},\frac{4\pi}{3})$.
\end{lemma}
\begin{proof}
By Theorem \ref{TheoremEnergyGroundstates} and Remark \ref{rem: repulsionsfree} we have that all bonds in the bond graph are of unit length and that the configuration is repulsion-free. An additional necessary condition for equality is that all bond-angles $\theta$ satisfy
\begin{align}\label{eq: larger than 2pi3}
\frac{2\pi}{3} \leq \theta \leq \frac{4\pi}{3}.
\end{align}
In fact, suppose that $x_1,x_0,x_2$ form the angle $\theta$. Since $x_1,x_2$ are neighbors of $x_0$, we have $q_1=q_2$ by Lemma \ref{LemmaNeighborhood}. The above mentioned necessary conditions imply $|x_1-x_0| = |x_2-x_0|=1$ and $V_{\mathrm{r}}(|x_1-x_2|)=0$. The latter only holds if $|x_1-x_2| \ge \sqrt{3}$ by $(\mathrm{vii})$. Simple trigonometry then yields $\frac{2\pi}{3} \leq \theta \leq \frac{4\pi}{3}$
From this discussion \EEE we derive that the edges of each hexagon necessarily need to have length $1$ and the interior angles are larger or equal to $2\pi/3$. As the sum of the interior angles in a planar hexagon equals \EEE $4\pi$, we get that each interior angle is $2\pi/3$, i.e., each hexagon is indeed a regular hexagon with unit bond length. The charge is alternating by Lemma \ref{LemmaNeighborhood}. \EEE
Consider an atom $x_i$ with a bond-angle $\theta_1$ which satisfies $\frac{4\pi}{3}>\theta_1 > \frac{2\pi}{3}$. Suppose by contradiction that $x_i$ had more than two bonds (i.e., three bonds, see Lemma \ref{LemmaNeighborhood}). Summing up all the three bond-angles $\theta_1,\theta_2,\theta_3$ at $x_i$ we get that $\min\lbrace \theta_2,\theta_3 \rbrace < 2\pi/3$. This, however, contradicts \eqref{eq: larger than 2pi3}.
\end{proof}
\begin{figure}[htp]
\begin{tikzpicture}[scale =2]
\draw[thin](0,0)--++(-60:1)--++(0:1)--++(60:1);
\draw[thin](0,0)--++(60:1)--++(0:1);
\draw[dashed,thin](0,0)++(60:1)++(0:1)--++(-60:1);
\draw[fill=black](0,0)circle(.05)++(-60:1)++(0:1)circle(.05)++(60:1);
\draw[fill=white](0,0)++(-60:1)circle(.05)++(0:1)++(60:1)circle(.05);
\draw[fill=white](0,0)++(60:1)circle(.05)++(0:1);
\draw[fill=black](0,0)++(60:1)++(0:1)circle(.05);
\draw(0,0) node[anchor=north east]{$x_{i+2} $} ++(-60:1)node[anchor=north]{$x_{i+3} $}++(0:1)node[anchor=north]{$x_{i+4} $}++(60:1)node[anchor=north west]{$x_{i+5} $};
\draw[thin](0,0)++(60:1) node[anchor=south]{$x_{i+1} $}++(0:1) node[anchor=south]{$x_{i} $};
\draw[thin](-60:.4) arc(-60:60:.4);
\draw[thin](0,0) node[anchor=west]{$\theta_{i+2} $};
\draw[thin](60:1)++(240:.4) arc(240:360:.4);
\draw[thin](60:1) node[anchor=north west]{$\theta_{i+1} $};
\draw[thin](-60:1)++(0:.4) arc(0:120:.4);
\draw[thin](-60:1) node[anchor=south west]{$\theta_{i+3} $};
\draw[thin](-60:1)++(0:.6) arc(180:60:.4);
\draw[thin](-60:1)++(1,0) node[anchor=south]{$\theta_{i+4} $};
\end{tikzpicture}
\caption{Part of an \EEE octagon with $4$ consecutive angles $\theta_{i+1},\ldots, \theta_{i+4}$ \EEE equal to $\frac{2\pi}{3}$.}
\label{Figureangles}
\end{figure}
We observe \EEE that octagons contain non-equilibrated atoms. More precisely, we have the following statement.
\begin{lemma}[Octagon]\label{lemma: octa}
Let $C_n$ be a ground state containing an octagon $\{x_0,\ldots,x_7\}$ in the bond graph. Let $\theta_i$, $i=0,\ldots,7$, be the interior angles of the octagon. Then we have
\begin{align*}
(i) & \ \ (\theta_i,\theta_{i+1},\theta_{i+2},\theta_{i+3}) \neq \left(\frac{2\pi}{3},\frac{2\pi}{3},\frac{2\pi}{3},\frac{2\pi}{3}\right) \text{ for all } i=0,\ldots,7,\\
(ii) & \ \ \theta_i < \frac{4\pi}{3} \text{ for all } i=0,\ldots,7.
\end{align*}
Here, the indices have to be understood $\mathrm{mod}\, 8$.
\end{lemma}
\begin{proof}
By Remark \ref{rem: repulsionsfree} and Lemma \ref{LemmaHexagon} we have that
\begin{align}\label{E=Eimplication}
\frac{2\pi}{3} \leq \theta_i \leq \frac{4\pi}{3} \text{ and } |x_{i+1}-x_i|=1 \text{ for all } i=0,\ldots,7.
\end{align}
Suppose by contradiction that (i) was wrong. If there are more than three consecutive angles of size $\frac{2\pi}{3}$, as indicated in Fig.~\ref{Figureangles}, then the bond graph would contain an additional hexagon and at least one triangle or square. This is a contradiction to the fact that the bond graph contains an octagon.
We now show (ii). \EEE Assume by contradiction that, without restriction, $\theta_0 =\frac{4\pi}{3}$. Then by (\ref{E=Eimplication}) and the fact that the interior angles of the octagon sum to $6\pi$, there holds for all $i=1,\ldots,7$
\begin{align*}
\theta_i + \frac{4\pi}{3} + 6\frac{2\pi}{3} \leq\sum_{j=0}^7 \theta_j= 6\pi.
\end{align*}
This implies $\theta_i \leq \frac{2\pi}{3}$ for all $i=1,\ldots,7$. Again using (\ref{E=Eimplication}) this yields $\theta_i = \frac{2\pi}{3}$ for all $i=1,\ldots, 7$ which contradicts (i).
\end{proof}
\EEE
The following lemma investigates the properties of a configuration in which a non-equilibrated bulk atom is present. Roughly speaking, it states that the existence of such an atom induces the existence of more non-equilibrated atoms and a certain excess of edges $\eta$. Note that at this point our analysis deviates significantly from \cite{Mainini}: in a model with angle potentials favoring $\frac{2\pi}{3}$ angles, \EEE it is obvious that non-equilibrated atoms cannot exist in ground states. \EEE
\begin{lemma}[Non-equilibrated atoms\EEE]\label{LemmaOctagon} Let $C_n$ be a ground state with no acyclic bonds in the bond graph. Assume that $\# \mathcal{A}_{\rm bulk} \ge 1$. Then one of the following holds:
\begin{itemize}
\item[i)] $\# \mathcal{A} \ge 2$ and \EEE $\eta \geq 6$,
\item[ii)] $\# (\mathcal{A} \setminus \mathcal{A}_{\rm bulk}) \ge 1$ and $\eta = 4$. \EEE
\end{itemize}
\end{lemma}
\begin{proof}
We first prove that each polygon containing non-equilibrated atoms has at least eight vertices and contains at least two non-equilibrated atoms. Then we show that $\eta \ge 4$ and that in the case $\eta < 6$, we have $\# (\mathcal{A} \setminus \mathcal{A}_{\rm bulk}) \ge 1$. The statement clearly follows from these claims. \EEE
\textbf{Claim 1:} Each polygon containing a non-equilibrated atom has at least eight vertices.
\noindent \emph{Proof of Claim 1:} Due to Lemma \ref{LemmaHexagon}, both angles at non-equilibrated atoms lie in $(\frac{2\pi}{3},\frac{4\pi}{3})$ and therefore \EEE each polygon containing a non-equilibrated atom is not a hexagon. Thus, it has at least eight vertices by Lemma \ref{LemmaPolygon}. \EEE
\EEE
\textbf{Claim 2:} Each polygon contains either no or at least two non-equilibrated atoms. \EEE
\noindent \emph{Proof of Claim 2:} Consider a polygon with $k$ edges which contains a non-equilibrated atom \EEE with interior angle $\theta_1 \neq \frac{2\pi}{3}, \frac{4\pi}{3}$. By Lemma \ref{LemmaPolygon} we have that $k \in 2\mathbb{N}$. Assume by contradiction that all the other angles $\theta_2, \ldots, \theta_k$ \EEE are either $\frac{2\pi}{3}$ or $\frac{4\pi}{3}$. We have
\begin{align*}
\sum_{j=1}^k \theta_j =\pi(k-2).
\end{align*}
Since we assumed that $\theta_j$, $j \ge 2$, \EEE are integer multiples of $\frac{2\pi}{3}$, we have
\begin{align*}
\theta_1 + k' \frac{2\pi}{3}= \pi(k-2),
\end{align*}
where $k' \in \mathbb{N}$ is given by $k'=\#\{j : \theta_j =\frac{2\pi}{3}\}+2\#\{j : \theta_j =\frac{4\pi}{3}\}$. This implies
\begin{align*}
k =\frac{\theta_1}{\pi}+ 2 + \frac{2}{3}k'.
\end{align*}
Since both $k,k' \in \mathbb{N}$ and $\frac{2\pi}{3} < \theta_1 < \frac{4\pi}{3}$ by Lemma \ref{LemmaHexagon}, \EEE there exists only a solution to the equation if $\theta_1 =\pi$ and $k' \in 3\mathbb{N}$. This implies that $k$ is odd: a contradiction.
\textbf{Claim 3:} \EEE We have $\eta \ge 4$. If $\eta<6$, then $\# (\mathcal{A} \setminus \mathcal{A}_{\rm bulk}) \ge 1$. \EEE
\noindent \emph{Proof of Claim 3:} \EEE Since $\# \mathcal{A}_{\rm bulk} \ge 1$, there exists a non-equilibrated bulk atom. As $C_n$ does not have acyclic bonds, we observe that this atom is a vertex of at least two polygons. \EEE Claim 1 then yields that there have to be at least two polygons with at least eight vertices, i.e., $\eta \geq 4$, $\eta \in 2\mathbb{N}$. \EEE
It remains to show that, if $\eta =4$, then $\# (\mathcal{A} \setminus \mathcal{A}_{\rm bulk}) \ge 1$. Assume by contradiction that $\eta=4$ and $\mathcal{A} \setminus \mathcal{A}_{\rm bulk} = \emptyset$. As $\eta=4$, the two non-hexagons have to be octagons. By Claim 1, \EEE the assumption $\mathcal{A} \setminus \mathcal{A}_{\rm bulk} = \emptyset$, and the fact that $C_n$ has no acyclic bonds, \EEE we find that all non-equilibrated atoms are contained in both octagons.
Denote the interior angles in the first octagon different from $ \{\frac{2\pi}{3},\frac{4\pi}{3}\}$ by $\alpha_1,\ldots,\alpha_k$, where $1 \le k \le 8$. Similarly, the angles in the second octagon different from $ \{\frac{2\pi}{3},\frac{4\pi}{3}\}$ \EEE are denoted by $\beta_1,\ldots,\beta_k$, where without restriction $\alpha_i$ and $\beta_i$ lie at the same atom. Note that $\beta_i = 2\pi - \alpha_i$ for $i=1,\ldots,k$ as non-equilibrated atoms are $2$-bonded. Due to Lemma \ref{lemma: octa}(ii), all other interior angles of the octagons are $\frac{2\pi}{3}$. Thus, by the interior angle sum of the octagons we obtain
\begin{align*}
\sum_{j=1}^k \alpha_j + (8-k) \EEE \frac{2\pi}{3} = 6\pi, \ \ \ \ \sum_{j=1}^k \beta_j + (8-k) \EEE \frac{2\pi}{3} = 6\pi.
\end{align*}
Using $\sum_{j=1}^k \beta_j = 2\pi k- \sum_{j=1}^k \alpha_j $ and summing the two equations, we obtain the unique solution $k=2$. In particular, this implies $\# \mathcal{A} = 2$. Denote the two atoms in $\mathcal{A}$ by $x_1$ and $x_2$. From Lemma \ref{lemma: octa}(i) we get that the atoms $x_1$ and $x_2$ lie `on opposite sides' of the octagons, \EEE i.e., the shortest path in the bond graph connecting $x_1$ and $x_2$ has length $4$. Then we also see that $\alpha_1=\alpha_2=\beta_1=\beta_2=\pi$ by a simple geometric argument. (An octagon with this geometry is depicted in the rightmost configuration in Fig.~\ref{FigureFlexible}.) \EEE Finally, this yields that the two octagons are identical up to an isometry. This, however, contradicts the fact that both non-equilibrated atoms $x_1$ and $x_2$ are contained in both octagons. Thus, $\eta=4$ and $\mathcal{A} \setminus \mathcal{A}_{\rm bulk} = \emptyset$ is not possible. This concludes the proof. \EEE
\end{proof}
Based on Lemma \ref{LemmaOctagon}, we now show that non-equilibrated bulk atoms cannot exist in ground states with no acyclic bonds. \EEE
\begin{lemma}[Non-equilibrated bulk atoms\EEE] \label{LemmaAbulk} Let $n \geq 1$ and let $C_n$ be a ground state with no acyclic bonds. \EEE Then $\mathcal{A}_{\mathrm{bulk}} = \emptyset$.
\end{lemma}
\begin{proof} We prove the statement by induction. We first note that the statement is true for $1 \le n \le 9$. In fact, in this case the bond graph contains at most one polygon by Lemma \ref{LemmaPolygon} and Lemma \ref{LemmaHexagon}. This \EEE implies $\mathcal{A}_{\mathrm{bulk}} \subset X_n^{\mathrm{bulk}}=\emptyset$. Let $n \ge 10$. \EEE We assume that the result has been proven for $m<n$ and proceed to show the statement for $n$. Assume by contradiction that $\mathcal{A}_{\mathrm{bulk}} \neq \emptyset$. By Lemma \ref{LemmaOctagon} there are two cases to consider:
\begin{itemize}
\item[i)] $\#(\mathcal{A}\setminus \mathcal{A}_{\mathrm{bulk}}) \geq 1$ and $\eta=4$.
\item[ii)] $\# \mathcal{A} \geq 2$ and $\eta \geq 6$.
\end{itemize}
\noindent \emph{Proof for Case $\mathrm{i)}$:} By Lemma \ref{LemmaBoundaryEnergy} we obtain the strict inequality
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) >-\frac{3}{2}d +3.
\end{align*}
By Lemma \ref{lemma: eta energy} (with strict inequality) this gives
\begin{align}\label{EnergyestimateOctagon}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}(n+\eta-4)}+1.
\end{align}
Note that $\eta=4$ and therefore we have
\begin{align}\label{Strictenergyboundetabig}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}n}+1 > -\lfloor \beta(n)\rfloor
\end{align}
which contradicts Theorem \ref{TheoremEnergyGroundstates} and the fact that $C_n$ is a ground state.
\noindent \emph{Proof for Case $\mathrm{ii})$}: We are now in the case that $\#\mathcal{A} \geq 2$ and $\eta \geq 6$. By the previous case we can assume that $\mathcal{A} \setminus \mathcal{A}_{\mathrm{bulk}} =\emptyset$. This implies $\#\mathcal{A}_{\mathrm{bulk}} \geq 2$. \EEE After removing the boundary, we can suppose that
\begin{align}\label{eq: bulk-equ}
\mathcal{E}(C_n^{\rm bulk}) = - \lfloor \beta(n-d) \rfloor.
\end{align}
Indeed, if $\mathcal{E}(C_n^{\rm bulk}) > - \lfloor \beta(n-d) \rfloor$, we derive by Lemma \ref{lemma: eta energy} (with strict inequality) that \eqref{EnergyestimateOctagon} holds. Since $\eta \ge 4$, we get a contradiction exactly as in Case i), see \eqref{Strictenergyboundetabig}. Hence, $C_n$ as well as $C_n^{\mathrm{bulk}}$ are ground states. We now distinguish the following cases:
\begin{itemize}
\item[a)] $C_n^{\mathrm{bulk}}$ contains at least two flags,
\item[b)] $C_n^{\mathrm{bulk}}$ contains a bridge,
\item[c)] $C_n^{\mathrm{bulk}}$ contains less than two flags and no bridge. \EEE
\end{itemize}
\EEE
\noindent \emph{Proof for Case $\mathrm{a})$}: We use Lemma \ref{LemmaBoundaryEnergy} to obtain
\begin{align}\label{ineq: bnd2flags}
\mathcal{E}^{\mathrm{bnd}}(C_n) \geq -\frac{3}{2}d + 3.
\end{align}
Using the fact that a flag contributes at least $-1$ to the energy and applying Theorem \ref{TheoremEnergyGroundstates} on the sub-configuration obtained after removing exactly two flags from $C_n^{\mathrm{bulk}}$, \EEE we obtain
\begin{align}\label{ineq: bulk2flags}
\mathcal{E}^{\rm bulk}(C_n) = \mathcal{E}(C_n^{\mathrm{bulk}}) \EEE \geq \EEE -2 - \lfloor\beta(n-d-2)\rfloor \geq -\frac{3}{2}(n-d) +\sqrt{\frac{3}{2}(n-d-2)} + 1.
\end{align}
Combining (\ref{ineq: bnd2flags})-(\ref{ineq: bulk2flags}) we obtain
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n +\sqrt{\frac{3}{2}(n-d-2)} + 4.
\end{align*}
By Lemma \ref{LemmaBoundaryestimate} and $b = - \mathcal{E}(C_n)$ we obtain $
\mathcal{E}(C_n) \geq -\frac{3}{2}n + 4 + \sqrt{\frac{3}{2}(-4\mathcal{E}(C_n) + 4+\eta-5n)}.
$
Lemma \ref{LemmaSquareroot} for $j=4$, $m= 4+ \eta$, and $x = \mathcal{E}(C_n)$ yields
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}(n+10+\eta-16)} + 1 \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}n}+1 > -\lfloor \beta(n)\rfloor,
\end{align*}
where we used $\eta \geq 6$. This contradicts Theorem \ref{TheoremEnergyGroundstates} and the fact that $C_n$ is a ground state. \EEE
\noindent \emph{Proof for Case $\mathrm{b})$}: In view of Remark \ref{rem: bridge}, $C_n^{\mathrm{bulk}}$ can only contain a bridge if $n-d = 12$ and $C_n^{\mathrm{bulk}}$ consists of two regular hexagons connected with a bridge. This contradicts $\#\mathcal{A}_{\mathrm{bulk}} \geq 2$. \EEE
\noindent \emph{Proof for Case $\mathrm{c})$}: Denote by $l \in \lbrace 0,1 \rbrace$ the number of flags of $C_n^{\mathrm{bulk}}$ and let $C_{n-d-l}^*$ be the configuration which arises by removing $l$ atoms from $C_n^{\mathrm{bulk}}$ such that $C_{n-d-l}^*$ has no acyclic bonds. Observe that $\mathcal{E}(C_{n-d-l}^*)-l \le \mathcal{E}(C_n^{\mathrm{bulk}})$. Then also $C_{n-d-l}^*$ is a ground state since otherwise $\mathcal{E}(C_n^{\rm bulk}) > - \lfloor \beta(n-d) \rfloor$ by Lemma \ref{LemmaPropertiesbeta} 1) which contradicts \eqref{eq: bulk-equ}. \EEE As $\# \mathcal{A}_{\rm bulk} \ge 2$ and $l \le 1$, $C_{n-d-l}^*$ contains a non-equilibrated atom. Thus, due to the fact that all hexagons in the bond graph are regular (see Lemma \ref{LemmaHexagon}), we have that $\eta^{*} = \eta(C_{n-d-l}^*) \geq 2$, where $\eta^{*}$ denotes the excess of edges \EEE of $C_{n-d-l}^*$. By the induction assumption we have that $C_{n-d-1}^*$ has no non-equilibrated bulk atom and thus \EEE has a non-equilibrated boundary atom. Thus, strict inequality holds in \eqref{BoundaryEnergyEstimate} for $C_{n-d-l}^*$, see Lemma \ref{LemmaBoundaryEnergy}. Therefore, by Lemma \ref{lemma: eta energy} (with strict inequality) applied for $C_{n-d-l}^*$ and $\eta^{*} \geq 2$ we obtain \EEE
\begin{align*}
\mathcal{E}^{\mathrm{bulk}}(C_n) = \mathcal{E}(C_n^{\mathrm{bulk}}) \ge \mathcal{E}(C_{n-d-l}^*)-l \EEE \geq -\frac{3}{2}(n-d-l) + \sqrt{\frac{3}{2}(n-d-l-2)}+1-l.
\end{align*}
Using Lemma \ref{LemmaBoundaryEnergy} for $C_n$ and summing $\mathcal{E}^{\mathrm{bnd}}(C_n)$ and $\mathcal{E}^{\mathrm{bulk}}(C_n)$, we derive
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + 4 + \sqrt{\frac{3}{2}(n-d-2-l)}+ \frac{l}{2}.
\end{align*}
By Lemma \ref{LemmaBoundaryestimate} and $\eta \ge 6$ \EEE we obtain
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + 4 + \frac{l}{2} + \sqrt{\frac{3}{2}(-4\mathcal{E}(C_n) + 10 \EEE -5n-l)}.
\end{align*}
Lemma \ref{LemmaSquareroot} for $j=4+\frac{l}{2}$, $m= 10 \EEE -l$, and $x = \mathcal{E}(C_n)$ yields
\begin{align*}
\mathcal{E}(C_n) &\geq -\frac{3}{2}n +1 +\frac{l}{2} + \sqrt{\frac{3}{2}(n-3l)} = -\frac{3}{2}n +\sqrt{\frac{3}{2}n} + 1 +l \Big(\frac{1}{2} - \frac{9}{2(\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-3l)})}\Big). \EEE
\end{align*}
This estimate can be used to calculate $\mathcal{E}(C_n) \geq -\frac{3}{2}n + \sqrt{\frac{3}{2}n}+1 >-\lfloor\beta(n)\rfloor$ for all $n \ge 16$ when $l=1$ or for all $n \geq 10$ when $l=0$. In the cases $10 \le n \le 15$, $l=1$, one can use $\mathcal{E}(C_n) \geq -\frac{3}{2}n +1 +\frac{1}{2} + \sqrt{\frac{3}{2}(n-3)}$ and compare this estimate directly with $\lfloor \beta(n)\rfloor$ to obtain $\mathcal{E}(C_n) > -\lfloor \beta(n)\rfloor$, \EEE cf.\ Table \ref{table2}. In every case, this yields a contradiction to the fact that $C_n$ is a ground state. \EEE
\end{proof}
\subsection{Characterization of ground states: proof of Theorem \ref{TheoremGroundstatesleq31}}
In this section we prove Theorem \ref{TheoremGroundstatesleq31}. \EEE
\textbf{Boundary $k$-gon:} We say that a $k$-gon in the bond graph is a \textit{boundary} $k$-gon, whenever it shares at least one edge with the unbounded face.
The following proposition is the main ingredient for the proof of Theorem \ref{TheoremGroundstatesleq31}.
\begin{proposition}\label{TheoremGroundstatesleq-new} Let $n \geq10$ and let $C_n$ be a ground state. Then the bond graph consists only of hexagonal cycles except for at most two flags and at most one boundary octagon. The bond graph cannot contain both flags and an octagon at the same time. \EEE
\end{proposition}
Once this proposition is proven, for the proof of Theorem \ref{TheoremGroundstatesleq31} it remains to show that in the case $n \ge 30$ no octagons may occur. In fact, Theorem \ref{TheoremGroundstatesleq31} then follows from Theorem \ref{TheoremEnergyGroundstates}, Remark \ref{rem: main}, Lemma \ref{LemmaNeighborhood}, and Lemma \ref{lemma: bridges}. For \EEE $n = 9$, a ground state may contain an octagon and a flag at the same time, see the rightmost configuration \EEE in Fig.~\ref{FigureFlexible}. In this sense, the assumption $n \geq10$ in \EEE Proposition \ref{TheoremGroundstatesleq-new} is sharp.
\begin{proof}[Proof of Proposition \ref{TheoremGroundstatesleq-new}]
Let $C_n$ be a ground state. Recall by Theorem \ref{TheoremEnergyGroundstates} that $\mathcal{E}(C_n) = -b=-\lfloor \beta(n)\rfloor$ and that $C_n$ \EEE is connected. We divide the proof into several steps. First, we prove that ground states contain only hexagonal cycles if $\mathcal{A}=\emptyset$ (Claim 1). Then, in the case $\mathcal{A} \neq \emptyset$, we prove that at most one boundary octagon may exist (Claim 2). Finally, we show that the existence of a non-hexagonal cycle excludes the existence of flags (Claim 3). The statement follows from Claim 1 - Claim 3 and Lemma \ref{LemmaOctagonflag}. \EEE The claims are proven by contradiction.
\noindent\textbf{Claim 1:} If $\mathcal{A}= \emptyset$, $C_n$ is defect-free.
\noindent \emph{Proof of Claim 1}: Suppose that there exists a $k$-gon, $k \ge 8$. Since $C_n$ \EEE is connected and $\mathcal{A}=\emptyset$, we have that $C_n$ \EEE is a connected subset of the hexagonal lattice. We observe that then our energy coincides with the one considered in \cite{Mainini}. We can repeat the argument in the proof of \cite[Proposition 6.7]{Mainini}, i.e., we can move boundary atoms inside this $k$-gon and observe that one can strictly lower the energy. This contradicts the fact that $C_n$ is a ground state.
\noindent\textbf{Claim 2:} For every ground state without flags there exists at most one $k$-gon, $k\ge 8$. If it exists, it has to be a boundary octagon.
\noindent \emph{Proof of Claim 2}: In view of Claim 1, we can suppose $\mathcal{A} \neq \emptyset$ and $\eta \geq 2$. As $\eta \ge 2$, $C_n$ contains no bridge (see Lemma \ref{lemma: bridges}, Remark \ref{rem: bridge}) and thus no acyclic bonds. \EEE By Lemma \ref{LemmaAbulk} we have that $\mathcal{A}_{\mathrm{bulk}}=\emptyset$ and therefore $\mathcal{A}\setminus \mathcal{A}_{\mathrm{bulk}} \neq \emptyset$. Applying Lemma \ref{LemmaBoundaryEnergy} and Lemma \ref{lemma: eta energy} (with strict inequality) we obtain that
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n +1 + \sqrt{\frac{3}{2}(n+\eta-4)}.
\end{align*}
In the case $\eta \geq 4$ we obtain a contradiction to the fact that $C_n$ is a ground state. Therefore, we can assume that \EEE $\eta=2$, i.e., there exists exactly one octagon. We \EEE apply Lemma \ref{LemmaHexagon} to find that hexagons are regular which implies $\mathcal{A}$ has to be contained in the octagon of the bond graph. By Lemma \ref{LemmaAbulk} \EEE we observe $\mathcal{A} \subset \partial X_n$. This implies that the octagon is a boundary octagon.
\noindent\textbf{Claim 3:} A ground state cannot contain both a $k$-gon, $k \ge 8$, and a flag.
\noindent \emph{Proof of Claim 3}: Assume by contradiction that there exists a $k$-gon and $l$ flags in the bond graph. (As before in Claim 2, by Remark \ref{rem: bridge} there are no bridges.) \EEE Using the fact that a flag contributes at least $-1$ to the energy and removing the flags we obtain a sub-configuration $C_{n-l}$ satisfying $ \mathcal{E}(C_{n}) \ge \mathcal{E}(C_{n-l}) - l$. Then also $C_{n-l}$ is a ground state since otherwise $\mathcal{E}(C_{n})> - \lfloor \beta(n) \rfloor$ by Lemma \ref{LemmaPropertiesbeta} 1) which contradicts Theorem \ref{TheoremEnergyGroundstates}. As $C_{n-l}$ has no acyclic bonds, we can use Lemma \ref{LemmaAbulk} to find \EEE $\mathcal{A}_{\mathrm{bulk}}=\emptyset$ and therefore $\mathcal{A}\setminus \mathcal{A}_{\mathrm{bulk}} \neq \emptyset$, where $\mathcal{A}_{\mathrm{bulk}}, \mathcal{A}$ correspond to configuration $C_{n-l}$. Applying Lemma \ref{LemmaBoundaryEnergy} and Lemma \ref{lemma: eta energy} on $C_{n-l}$ \EEE (with strict inequality for $\eta \ge 2$), we obtain
$$ \mathcal{E}(C_n) \geq \mathcal{E}(C_{n-l}) - l \ge -l -\frac{3}{2}(n-l) +1 + \sqrt{\frac{3}{2}(n-l-2)}. $$
Setting $j = l +2$, we find
\begin{align*}
\mathcal{E}(C_n)\ge -\Big( \frac{3}{2}n - \sqrt{\frac{3}{2}n} -\frac{1}{2}j +\frac{\frac{3}{2}j}{\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-j)}}\Big).
\end{align*}
Recall that $j \ge 3$. \EEE At this point, we can follow verbatim the proof of Lemma \ref{LemmaOctagonflag} to get the contradiction $\mathcal{E}(C_n) > -\lfloor \beta(n)\rfloor$ for each $n \ge 10$. This contradicts Theorem \ref{TheoremEnergyGroundstates} and the fact that $C_n$ is a ground state. \EEE
\end{proof}
As a final preparation for the proof of Theorem \ref{TheoremGroundstatesleq31}, we need the following elementary geometric lemma. \EEE
\begin{lemma}\label{LemmaOctagon3bdd} Let $C_n$ be a ground state with $\eta = 2$ that does not contain any acyclic bonds. Let $\{x_0,\ldots,x_7\}$ be the \EEE octagon in the bond graph. Set $X^3:= \{x_i \in \{0,\ldots,7\}: x_i \text{ is } 3\text{-bonded}\}$. \EEE We have the following:
\begin{itemize}
\item[i)] If $ \# X^3 \EEE \leq 3$, then $X^3$ \EEE is connected,
\item[ii)] If $\#X^3 \in \{4,5\}$, then there exist $j_1,j_2,j_3 \in \{0,\ldots,7\}$ such that $X_n \setminus \{x_{j_1},x_{j_2},x_{j_3}\}$ is not connected, $x_{j_1},x_{j_2},x_{j_3}$ are $2$-bonded, \EEE and
\begin{align}\label{eq: remove3}
\mathcal{E}\big(C_n \setminus \{ (x_{j_1},q_{j_1}), (x_{j_2}, q_{j_2}),(x_{j_3},q_{j_3}) \} \big) \le \EEE \mathcal{E}(C_n)+5.
\end{align}\EEE
\item[iii)] If $\# X^3 \geq 6$, then $C_n^{\mathrm{bulk}}$ is not connected.
\end{itemize}
\end{lemma}
\begin{proof} Let $\{x_0,\ldots,x_7\}$ be the \EEE octagon in the bond graph. All following \EEE statements are seen $\mathrm{mod}\, 8$ with respect to the numeration of $i$. Assume that $x_{i+1},x_{i-1} \in \mathcal{N}(x_i)$ for all $i=0,\ldots,7$ and $x_i \notin \mathcal{N}(x_j) $ for all $j \notin \{i-1,i+1\}$. Let $\theta_i$, $i=0,\ldots,7$, be \EEE the interior angles of the octagon. We start with three preliminary observations. \EEE
\textbf{Claim 1:} If $x_i$ is $2$-bonded, \EEE then $x_i \in \partial X_n$.
\noindent \emph{Proof of Claim 1:} Consider a $2$-bonded $x_i$ and assume \EEE by contradiction that $x_i \in X_n^{\mathrm{bulk}}$. By Lemma \ref{lemma: octa}(ii) \EEE we have that $\theta_i < \frac{4\pi}{3}$ and therefore the other angle $\alpha$ at $x_i$ satisfies $\alpha >\frac{2\pi}{3}$. Since $x_{i}$ is a bulk \EEE atom, Lemma \ref{LemmaHexagon} and $\alpha > \frac{2\pi}{3}$ imply that $x_{i}$ is contained in the octagon and in another polygon which is not a hexagon \EEE since for hexagons all interior angles are equal to $\frac{2\pi}{3}$. This contradicts $\eta=2$.
\textbf{Claim 2:} If $x_i$ is $3$-bonded, then $x_{i-1}$ or $x_{i+1}$ is also $3$-bonded. \EEE
\noindent \emph{Proof of Claim 2:} Assume by contradiction that there exists a $3$-bonded $x_i$ such that $x_{i-1}$ and $x_{i+1}$ are $2$-bonded. Denote by $z$ the third neighbor of $x_i$. \EEE Since the bond graph of $C_n$ does not contain acyclic bonds, $x_i$ and $z$ are contained in a polygon. This contains necessarily either $x_{i-1}$ or $x_{i+1}$, say $x_{i-1}$. \EEE Thus, since $x_{i-1}$ is $2$-bonded, this polygon contains also $x_{i-2}$. Then \EEE $x_{i-1}$ has to be a bulk \EEE atom. This, however, contradicts Claim 1. \EEE
\textbf{Claim 3:} If $x_{i-1},x_i,x_{i+1}$ are $3$-bonded, then $x_{i-2}$ and $x_{i+2}$ are $2$-bonded.
\noindent \emph{Proof of Claim 3:} If there were four consecutive $3$-bonded atoms, there would be four consecutive interior angles of size $\frac{2\pi}{3}$, see Lemma \ref{LemmaHexagon}. This, however, contradicts Lemma \ref{lemma: octa}(i). \EEE
We now proceed with the proof of the statement. \EEE Set $k=\# X^3\EEE$. First, by Lemma \ref{LemmaHexagon}, Lemma \ref{lemma: octa}(ii), \EEE and the fact that
$
\sum_i \theta_i = 6\pi
$
we have that $k \leq 6$.
If $k=0$, there is nothing to prove. $k=1$ is not possible due to Claim 2. If $k=2$, again due to Claim 2, the $3$-bonded atoms \EEE have to be bonded. \EEE If $k=3$, the $3$-bonded atoms \EEE necessarily need to be of the form $x_{i-1},x_i,x_{i+1}$ for some $i=0,\ldots,7$. Otherwise, we have a contradiction to Claim 2. This proves $\mathrm{i)}$.
Now suppose that $k \in \{4,5\}$. By Lemma \ref{LemmaHexagon} we have that, if $x_i$ is $3$-bonded, then $\theta_i = \frac{2\pi}{3}$. By Lemma \ref{lemma: octa}(i) the $3$-bonded atoms cannot be of the form $\{x_i,\ldots,x_{i+k-1}\}$. Hence, there exist $ 0 \le i_1< i_1+1 < j_1 <i_2 < i_2 +2 <j_2$ such that \EEE $x_{i_1},x_{i_1+1}, x_{i_2}, \EEE x_{i_2+1}$ are $3$-bonded and $x_{j_1},x_{j_2},x_{j_2+1}$ are $2$-bonded. By Claim 1 we get that, if $x_j$ is $2$-bonded, then $x_j \in \partial X_n$. Therefore, $x_{i_1}$ and $x_{i_2}$ are connected only through paths going through $x_{j_1}$ or $x_{j_2}$ as otherwise one of the two atoms $x_{j_1}$ and $x_{j_2}$ would not be contained in the boundary. Thus, the set $X_n \setminus\{x_{j_1},x_{j_2},x_{j_2+1}\}$ is not connected. Since $x_{j_1},x_{j_2},x_{j_2+1}$ are $2$-bonded and $x_{j_2}$ is bonded to $x_{j_2+1}$, we remove exactly $5$ bonds. As each bond contributes at least $-1$ to the energy, this yields \EEE \eqref{eq: remove3}. \EEE
This proves $\mathrm{ii)}$.
Finally, suppose $k=6$. By Claim 3 we have that there exists $i \in \{0,\ldots,7\}$ such that $x_i,x_{i+1},x_{i+2},x_{i+4},x_{i+5},x_{i+6}$ are $3$-bonded. Since no acyclic bonds are present in the bond graph, we have that $x_{i+1},x_{i+5} \in X_n^{\mathrm{bulk}}$. Moreover, arguing as in case $\mathrm{ii)}$, $x_{i+1}$ is not connected with $x_{i+5}$ \EEE in $X_n^{\mathrm{bulk}}$. This \EEE proves $\mathrm{iii)}$.
\end{proof}
\begin{figure}[htp]
\centering
\begin{tikzpicture}[scale=.8]
\begin{scope}[shift={(0,-2)}]
\draw[thin](0:0)--++(60:1)--++(0:2)--++(300:1)--++(240:1)--++(180:2)--++(120:1);
\draw[thin](0:0)--++(180:1)--++(120:1)--++(60:1)--++(0:1)--++(-60:1);
\draw[fill=black](0:0)circle(.05)++(60:1)++(0:1)circle(.05)++(0:1)++(300:1)circle(.05)++(240:1)++(180:1)circle(.05)++(180:1)++(120:1);
\draw[fill=white](0:0)++(60:1)circle(.05)++(0:1)++(0:1)circle(.05)++(300:1)++(240:1)circle(.05)++(180:1)++(180:1)circle(.05)++(120:1);
\draw[fill=black](0:0)++(180:1)++(120:1)circle(.05)++(60:1)++(0:1)circle(.05)++(-60:1);
\draw[fill=white](0:0)++(180:1)circle(.05)++(120:1)++(60:1)circle(.05)++(0:1)++(-60:1);
\end{scope}
\begin{scope}[shift={(0,1)}]
\draw[thin](0:0)--++(60:1)--++(0:2)--++(300:1)--++(240:1)--++(180:2)--++(120:1);
\draw[thin](0:0)--++(180:1)--++(120:1)--++(60:1)--++(0:1)--++(-60:1);
\draw[thin](0:0)++(60:1)++(0:2)--++(60:1)--++(0:1)--++(-60:1)--++(-120:1)--++(-180:1);
\draw[fill=black](0:0)++(60:1)++(0:2)++(60:1)circle(.05)++(0:1)++(-60:1)circle(.05)++(-120:1)++(-180:1);
\draw[fill=white](0:0)++(60:1)++(0:2)++(60:1)++(0:1)circle(.05)++(-60:1)++(-120:1)circle(.05)++(-180:1);
\draw[fill=black](0:0)circle(.05)++(60:1)++(0:1)circle(.05)++(0:1)++(300:1)circle(.05)++(240:1)++(180:1)circle(.05)++(180:1)++(120:1);
\draw[fill=white](0:0)++(60:1)circle(.05)++(0:1)++(0:1)circle(.05)++(300:1)++(240:1)circle(.05)++(180:1)++(180:1)circle(.05)++(120:1);
\draw[fill=black](0:0)++(180:1)++(120:1)circle(.05)++(60:1)++(0:1)circle(.05)++(-60:1);
\draw[fill=white](0:0)++(180:1)circle(.05)++(120:1)++(60:1)circle(.05)++(0:1)++(-60:1);
\end{scope}
\begin{scope}[shift={(7,0)}]
\draw[thin](0:0)--++(60:1)--++(0:2)--++(300:1)--++(240:1)--++(180:2)--++(120:1);
\draw[thin](0:0)--++(180:1)--++(120:1)--++(60:1)--++(0:1)--++(-60:1);
\draw[thin](0:0)++(60:1)++(0:2)--++(60:1)--++(0:1)--++(-60:1)--++(-120:1)--++(-180:1);
\draw[thin](0:0)++(-60:1)--++(-120:1)--++(-180:1)--++(-240:1)--++(-300:1);
\draw[thin](0:0)++(-60:1)++(0:3)++(60:1)--++(-60:1)--++(-120:1)--++(-180:1)--++(-240:1);
\draw[fill=black](0:0)++(-60:1)++(-120:1)circle(.05)++(-180:1)++(-240:1)circle(.05)++(-300:1);
\draw[fill=white](0:0)++(-60:1)++(-120:1)++(-180:1)circle(.05)++(-240:1)++(-300:1);
\draw[fill=black](0:0)++(-60:1)++(0:3)++(60:1)++(-60:1)circle(.05)++(-120:1)++(-180:1)circle(.05)++(-240:1);
\draw[fill=white](0:0)++(-60:1)++(0:3)++(60:1)++(-60:1)++(-120:1)circle(.05)++(-180:1)++(-240:1);
\draw[fill=black](0:0)++(60:1)++(0:2)++(60:1)circle(.05)++(0:1)++(-60:1)circle(.05)++(-120:1)++(-180:1);
\draw[fill=white](0:0)++(60:1)++(0:2)++(60:1)++(0:1)circle(.05)++(-60:1)++(-120:1)circle(.05)++(-180:1);
\draw[fill=black](0:0)circle(.05)++(60:1)++(0:1)circle(.05)++(0:1)++(300:1)circle(.05)++(240:1)++(180:1)circle(.05)++(180:1)++(120:1);
\draw[fill=white](0:0)++(60:1)circle(.05)++(0:1)++(0:1)circle(.05)++(300:1)++(240:1)circle(.05)++(180:1)++(180:1)circle(.05)++(120:1);
\draw[fill=black](0:0)++(180:1)++(120:1)circle(.05)++(60:1)++(0:1)circle(.05)++(-60:1);
\draw[fill=white](0:0)++(180:1)circle(.05)++(120:1)++(60:1)circle(.05)++(0:1)++(-60:1);
\end{scope}
\end{tikzpicture}
\caption{ Illustration of the three cases in Lemma \ref{LemmaOctagon3bdd}.\EEE}
\label{Fig: Separating Octagon}
\end{figure}
We are now in a position to prove \EEE Theorem \ref{TheoremGroundstatesleq31}.
\begin{proof}[Proof of Theorem \ref{TheoremGroundstatesleq31}]
As observed below Proposition \ref{TheoremGroundstatesleq-new}, it suffices to check that for $n \ge 30$ the bond graph of a ground state $C_n$ does not contain an octagon. Assume by contradiction that $n\geq 30$ \EEE and that the bond graph contains an \EEE octagon. By Proposition \ref{TheoremGroundstatesleq-new} we know that there do not exist any flags in the bond graph and the bond graph contains only one (boundary) \EEE octagon and hexagons otherwise. By Lemma \ref{lemma: bridges} the bond graph does not contain any acyclic bonds. Denote by $\{x_0,\ldots,x_7\}$ the octagon in the bond graph. We need to consider the three cases
$${\rm (a)} \ \ \# X^3 \le 3, \ \ \ \ \ {\rm (b)} \ \ \# X^3 \in \{4,5\}, \ \ \ \ \ {\rm (c)} \ \ \# X^3 \ge 6, $$
where $X^3:= \{x_i \in \{0,\ldots,7\}: x_i \text{ is } 3\text{-bonded}\}$. (See Fig.~\ref{Fig: Separating Octagon} for an illustration.) \EEE
\noindent \emph{Proof of Case $(\mathrm{a})$:}
Since $\# X^3 \EEE \leq 3$, by Lemma \ref{LemmaOctagon3bdd} we have that there are $k$ $2$-bonded atoms with $k\geq 5$ and they form a connected set. Hence, removing these $2$-bonded atoms, we remove exactly $k+1$ bonds. Estimating the energy of every bond by $-1$ we get by Theorem \ref{TheoremEnergyGroundstates}
\begin{align*}
\mathcal{E}(C_n) \geq -(k+1) - \lfloor\beta(n-k)\rfloor.
\end{align*}
This implies $\mathcal{E}(C_n) > \lfloor \beta(n) \rfloor$. Indeed, this follows from Lemma \ref{LemmaPropertiesbeta} 3)-4) and the fact that $k\ge 5$, $n \ge 30$. \EEE This gives a contradiction in Case (a). \EEE
\noindent \emph{Proof of Case $(\mathrm{b})$:} By the assumption and Lemma \ref{LemmaOctagon3bdd} we have that there exist $j_1,j_2,j_3 \in \{0,\ldots,7\}$ such that
\begin{align*}
\mathcal{E}\big(C_n \setminus \{ (x_{j_1},q_{j_1}), (x_{j_2}, q_{j_2}),(x_{j_3},q_{j_3}) \} \big) \le \EEE \mathcal{E}(C_n)+5 \EEE
\end{align*}
and $C_n \setminus \{ (x_{j_1},q_{j_1}), (x_{j_2}, q_{j_2}),(x_{j_3},q_{j_3}) \} $ \EEE is not connected. Denote by $n_1,n_2 \in \mathbb{N}, n_1+n_2 =n-3$ the cardinality of the two connected components of $C_n \setminus \{ (x_{j_1},q_{j_1}), (x_{j_2}, q_{j_2}),(x_{j_3},q_{j_3}) \} $ \EEE which do not have any bonds between them. Since the bond graph of $C_n$ does not contain any acyclic bonds, as explained at the beginning of the proof, \EEE we have $n_1,n_2 \geq 6$. By Lemma \ref{LemmaPropertiesbeta} 2) and Theorem \ref{TheoremEnergyGroundstates} \EEE we obtain
\begin{align*}
\mathcal{E}(C_n) \ge \EEE \mathcal{E}\big(C_n \setminus \{ (x_{i},q_{i}): \ i = i_1,i_2,i_3\} \big) - 5 \ge \EEE -\lfloor\beta(n_1)\rfloor - \lfloor \beta(n_2)\rfloor - 5 > -\lfloor\beta(n-3)\rfloor - 4.
\end{align*}
By Theorem \ref{TheoremEnergyGroundstates} we have that $\mathcal{E}(C_n)$ is an integer. This implies
\begin{align}\label{eq: n1-NNN}
\mathcal{E}(C_n) &\geq -\lfloor\beta(n-3)\rfloor - 3 \geq -\frac{3}{2}n +\sqrt{\frac{3}{2}(n-3)} + \frac{3}{2}\notag\\&= -\frac{3}{2}n + \sqrt{\frac{3}{2}n} +\frac{3}{2}- \frac{9}{2\Big(\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-3)}\Big)}.
\end{align}
It is elementary to check that for $n\geq 16$ we have
$$
\frac{3}{2}- \frac{9}{2\Big(\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-3)}\Big)} \geq 1.
$$
This is a contradiction to the fact that $C_n$ is a ground state.
\noindent \emph{Proof of Case $(\mathrm{c})$:} By the assumption and Lemma \ref{LemmaOctagon3bdd} we have that $C_n^{\mathrm{bulk}}$ is not connected. As the bond graph contains an octagon, there exists a non-equilibrated atom. Lemma \ref{LemmaAbulk} \EEE implies $\mathcal{A}_{\rm bulk} = \emptyset$ and thus a boundary atom is not equilibrated. Then by Lemma \ref{LemmaBoundaryEnergy} we get
\begin{align*}
\mathcal{E}^{\mathrm{bnd}}(C_n) > -\frac{3}{2}d +3.
\end{align*}
As $C_n^{\mathrm{bulk}}$ is not connected, $C_n^{\mathrm{bulk}}$ cannot be a ground state. \EEE Applying Theorem \ref{TheoremEnergyGroundstates} to $C_n^{\mathrm{bulk}}$ we thus obtain
\begin{align*}
\mathcal{E}^{\mathrm{bulk}}(C_n)>-\lfloor \beta( n-d) \EEE \rfloor.
\end{align*}
By Theorem \ref{TheoremEnergyGroundstates}, we have \EEE $\mathcal{E}(C_n)=-b$ and therefore $\mathcal{E}^{\mathrm{bulk}}(C_n)$ and $\mathcal{E}^{\mathrm{bnd}}(C_n)$ are integers. Hence, we obtain
\begin{align*}
\mathcal{E}(C_n)\ge -\lfloor \beta( n-d) \EEE \rfloor -\frac{3}{2}d +5.
\end{align*}
Using Lemma \ref{LemmaBoundaryestimate} and $\eta =2$ we get
\begin{align*}
\mathcal{E}(C_n) \geq -\frac{3}{2}n + 5 -\sqrt{\frac{3}{2}(-4\mathcal{E}(C_n)-5n+8)}.
\end{align*}
This together with Lemma \ref{LemmaSquareroot} applied for $j=5$, $m= 8$, and $x = \mathcal{E}(C_n)$ leads to
\begin{align}\label{eq: n1}
\mathcal{E}(C_n) \geq -\frac{3}{2}n +2 + \sqrt{\frac{3}{2}(n-6)}= -\frac{3}{2}n +\sqrt{\frac{3}{2}n}+2 - \frac{9}{\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-6)}}.
\end{align}
For $n \geq 17$ we have
\begin{align*}
2 - \frac{9}{\sqrt{\frac{3}{2}n}+\sqrt{\frac{3}{2}(n-6)}} \geq 1
\end{align*}
which leads to a contradiction to the fact that $C_n$ is a ground state.
\end{proof}
\begin{remark}\label{rem: octogons}
{\normalfont
Inspection of the previous proof shows that among $13 \le n \le 29$ only for $n=15,18,21,29$ boundary octagons may occur. Indeed, in Case $(\mathrm{a})$ this follows from Lemma \ref{LemmaPropertiesbeta} 4), see particularly Table \ref{table}. In Case $(\mathrm{b})$ and Case $(\mathrm{c})$ we obtain a contradiction for each $13 \le n \le 29$: in Case $(\mathrm{b})$, we necessarily have $n \ge 16$ (see upper left configuration in Fig.~\ref{Fig: Separating Octagon}) and thus a contradiction in \eqref{eq: n1-NNN}. In Case $(\mathrm{c})$, we necessarily have $n \ge 22$ (see rightmost configuration in Fig.~\ref{Fig: Separating Octagon}) and thus a contradiction in \eqref{eq: n1}.
For $n=10,11$, the presence of an octagon is excluded by Proposition \ref{TheoremGroundstatesleq-new} and the fact that each hexagon can share at most $2$ atoms with an octagon (see Lemma \ref{LemmaHexagon} and Lemma \ref{lemma: octa}). Consequently, for $n \le 29$, ground states may contain a boundary octagon only for $n=8,9,12,15,18,21,29$. This is indeed possible as shown in Fig.~\ref{FigureFlexible} and Fig.~\ref{FigureOctagons}, cf.\ Table~\ref{table2}. \EEE
}
\end{remark}
\begin{figure}[htp]
\begin{tikzpicture}[scale=0.6]
\begin{scope}[shift={(-4,0)}]
\draw[thin](2,0)++(60:1)--++(1,0)--++(1,0)--++(300:1)--++(240:1)--++(180:1)--++(180:1);
\draw[fill=white](2,0)++(60:1)++(1,0)circle(.05)++(1,0)++(300:1)circle(.05)++(240:1)++(180:1)circle(.05)++(180:1);
\draw[fill=black](2,0)++(60:1)++(1,0)++(1,0)circle(.05)++(300:1)++(240:1)circle(.05)++(180:1)++(180:1);
\foreach \k in {0,...,5}{
\foreach \j in {0,2,4}{
\draw[thin](\k*60+30:{sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=black](\k*60+30:{sqrt(3)})++(\j*60:1) circle(.05);
}
\foreach \j in {1,3,5}{
\draw[thin](\k*60+30:{sqrt(3)})++(\j*60:1)--++(\j*60+120:1);
\draw[fill=white](\k*60+30:{sqrt(3)})++(\j*60:1) circle(.05);
}
}
\end{scope}
\begin{scope}[shift={(7,-4)},rotate=-90]
\draw[thin](0,0)--++(30:1)--++(90:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1)--++(270:1)--++(330:1);
\draw[thin](0,0)--++(270:1)--++(330:1)--++(30:1)--++(90:1)--++(150:1);
\draw[thin](0,0)++(270:1)--++(210:1)--++(150:1)--++(90:1)--++(30:1);
\draw[fill=black](0,0)++(270:1)++(330:1)circle(.05)++(30:1)++(90:1)circle(.05)++(150:1);
\draw[fill=white](0,0)++(270:1)circle(.05)++(330:1)++(30:1)circle(.05)++(90:1)++(150:1);
\draw[fill=black](0,0)++(270:1)++(210:1)circle(.05)++(150:1)++(90:1)circle(.05)++(30:1);
\draw[fill=white](0,0)++(270:1)++(210:1)++(150:1)circle(.05)++(90:1)++(30:1);
\draw[fill=black](0,0)circle(.05)++(30:1)++(90:1)circle(.05)++(90:1)++(150:1)circle(.05)++(210:1)++(270:1)circle(.05)++(270:1)++(330:1);
\draw[fill=white](0,0)++(30:1)circle(.05)++(90:1)++(90:1)circle(.05)++(150:1)++(210:1)circle(.05)++(270:1)++(270:1)circle(.05)++(330:1);
\end{scope}
\begin{scope}[shift={(7,-4)}]
\draw[thin](-2,0)++(300:1)++(180:1)--++(240:1)--++(300:1)--++(360:1)--++(60:1);
\draw[fill=black](-2,0)++(300:1)++(180:1)++(240:1)circle(.05)++(300:1)++(360:1)circle(.05)++(60:1);
\draw[fill=white](-2,0)++(300:1)++(180:1)++(240:1)++(300:1)circle(.05)++(360:1)++(60:1);
\draw[thin](-1,0)++(120:1)--++(180:1)--++(240:1)--++(300:1)--++(360:1);
\draw[fill=black](-1,0)++(120:1)++(180:1)++(240:1)circle(.05)++(300:1)++(360:1);
\draw[fill=white](-1,0)++(120:1)++(180:1)circle(.05)++(240:1)++(300:1)circle(.05)++(360:1);
\end{scope}
\begin{scope}[shift={(5,0)}]
\draw[thin](-1,0)++(120:1)--++(180:1)--++(240:1)--++(300:1)--++(360:1);
\draw[fill=black](-1,0)++(120:1)++(180:1)++(240:1)circle(.05)++(300:1)++(360:1);
\draw[fill=white](-1,0)++(120:1)++(180:1)circle(.05)++(240:1)++(300:1)circle(.05)++(360:1);
\begin{scope}[rotate=-90]
\draw[thin](0,0)--++(30:1)--++(90:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1)--++(270:1)--++(330:1);
\draw[thin](0,0)--++(270:1)--++(330:1)--++(30:1)--++(90:1)--++(150:1);
\draw[thin](0,0)++(270:1)--++(210:1)--++(150:1)--++(90:1)--++(30:1);
\draw[fill=black](0,0)++(270:1)++(330:1)circle(.05)++(30:1)++(90:1)circle(.05)++(150:1);
\draw[fill=white](0,0)++(270:1)circle(.05)++(330:1)++(30:1)circle(.05)++(90:1)++(150:1);
\draw[fill=black](0,0)++(270:1)++(210:1)circle(.05)++(150:1)++(90:1)circle(.05)++(30:1);
\draw[fill=white](0,0)++(270:1)++(210:1)++(150:1)circle(.05)++(90:1)++(30:1);
\draw[fill=black](0,0)circle(.05)++(30:1)++(90:1)circle(.05)++(90:1)++(150:1)circle(.05)++(210:1)++(270:1)circle(.05)++(270:1)++(330:1);
\draw[fill=white](0,0)++(30:1)circle(.05)++(90:1)++(90:1)circle(.05)++(150:1)++(210:1)circle(.05)++(270:1)++(270:1)circle(.05)++(330:1);
\end{scope}
\end{scope}
\begin{scope}[shift={(0,-5)}, rotate=-90]
\draw[thin](0,0)--++(30:1)--++(90:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1)--++(270:1)--++(330:1);
\draw[thin](0,0)--++(270:1)--++(330:1)--++(30:1)--++(90:1)--++(150:1);
\draw[thin](0,0)++(270:1)--++(210:1)--++(150:1)--++(90:1)--++(30:1);
\draw[fill=black](0,0)++(270:1)++(330:1)circle(.05)++(30:1)++(90:1)circle(.05)++(150:1);
\draw[fill=white](0,0)++(270:1)circle(.05)++(330:1)++(30:1)circle(.05)++(90:1)++(150:1);
\draw[fill=black](0,0)++(270:1)++(210:1)circle(.05)++(150:1)++(90:1)circle(.05)++(30:1);
\draw[fill=white](0,0)++(270:1)++(210:1)++(150:1)circle(.05)++(90:1)++(30:1);
\draw[fill=black](0,0)circle(.05)++(30:1)++(90:1)circle(.05)++(90:1)++(150:1)circle(.05)++(210:1)++(270:1)circle(.05)++(270:1)++(330:1);
\draw[fill=white](0,0)++(30:1)circle(.05)++(90:1)++(90:1)circle(.05)++(150:1)++(210:1)circle(.05)++(270:1)++(270:1)circle(.05)++(330:1);
\end{scope}
\begin{scope}[shift={(-6,-5)},rotate=-90]
\draw[thin](0,0)--++(30:1)--++(90:1)--++(90:1)--++(150:1)--++(210:1)--++(270:1)--++(270:1)--++(330:1);
\draw[thin](0,0)++(270:1)--++(210:1)--++(150:1)--++(90:1)--++(30:1);
\draw[thin](0,0)--++(90:-1);
\draw[fill=white](90:-1)circle(.05);
\draw[fill=black](0,0)++(270:1)++(210:1)circle(.05)++(150:1)++(90:1)circle(.05)++(30:1);
\draw[fill=white](0,0)++(270:1)++(210:1)++(150:1)circle(.05)++(90:1)++(30:1);
\draw[fill=black](0,0)circle(.05)++(30:1)++(90:1)circle(.05)++(90:1)++(150:1)circle(.05)++(210:1)++(270:1)circle(.05)++(270:1)++(330:1);
\draw[fill=white](0,0)++(30:1)circle(.05)++(90:1)++(90:1)circle(.05)++(150:1)++(210:1)circle(.05)++(270:1)++(270:1)circle(.05)++(330:1);
\end{scope}
\end{tikzpicture}
\caption{The ground states containing an octagon for $n=12,15,18,21,29$ (up to isometry and changing of the charges).}
\label{FigureOctagons}
\end{figure}
\EEE
\section{Characterization of the net charge}
This final section is devoted to the proof of Theorem \ref{TheoremCharge}(i). We recall that part (ii) of the statement has already been addressed by an explicit construction in Section \ref{sec: daisy + something}.
We start with some preliminary definitions. First, recall the definition of the hexagonal lattice $\mathcal{L}$ in \eqref{eq: hex-lattice}. Set $u_1 = (1,0)$, $u_2 = (\frac{1}{2},\frac{\sqrt{3}}{2})$, $u_3 = (-\frac{1}{2},\frac{\sqrt{3}}{2})$, $u_4 = (-1,0)$.
\begin{definition}[Zig-zag paths]\label{def: zig-zag}
(i) A tuple $(p_1,\ldots,p_m) \subset \mathcal{L}$ is called a \emph{zig-zag path} if there exists $k \in \lbrace 1,2,3 \rbrace$ such that
\begin{align}\label{eq: zig-zag}
(a)& \ \ p_j - p_{j-1} \in \lbrace u_k, u_{k+1} \rbrace \ \ \ \text{ for all } j \in \lbrace 2,\ldots,m\rbrace, \notag\\
(b) & \ \ p_{j+1} - p_{j} \neq p_j - p_{j-1} \ \ \ \text{ for all } j \in \lbrace 2,\ldots,m-1\rbrace.
\end{align}
(ii) We say that two zig-zag paths have the same \emph{orientation} if the same $k \in \lbrace 1,2,3 \rbrace$ appears in \eqref{eq: zig-zag}.
(iii) Given a configuration $C_n$ with $X_n \subset \mathcal{L}$, we say the zig-zag path $(p_1,\ldots,p_m) \subset \mathcal{L}$, $m \ge 3$, is a \emph{bridging zig-zag path} for $C_n$ if $p_1, p_m \in X_n\EEE$ and $p_2,\ldots,p_{m-1} \notin X_n\EEE$.
\end{definition}
Let $C_n$ be a ground state for $n \ge 30$ with no acyclic bonds. \EEE By Theorem \ref{TheoremGroundstatesleq31} we get that \EEE $X_n$ is defect-free and satisfies $X_n \subset \mathcal{L}$ (up to isometry). \EEE Let $(p_1,\ldots,p_m)$ be a bridging zig-zag path for $C_n$. Since $C_n$ \EEE is connected and defect-free, we find that $\mathcal{L} \setminus (X_n \cup \bigcup_{j=2}^{m-1} p_j)$ consists of two (one of them possibly empty) connected components. Exactly one of these components is bounded which we denote by $\mathcal{B}( X_n; \EEE ( p_1, \ldots, p_m ))$.
\begin{lemma}[Bridging zig-zag paths of ground states]\label{lemma: zig-zag paths}
Let $n \ge 30$. Each ground state with no acyclic bonds \EEE has at most one bridging zig-zag path.
\end{lemma}
\begin{proof}
As a preparation, we observe the following: suppose that there exists a bridging zig-zag path for a ground state $C_n$. Then we can choose a bridging zig-zag path $(p_1,\ldots,p_m) \subset \mathcal{L}$ for $C_n$ such that $\mathcal{B}(X_n;(p_1,\ldots,p_m)) = \emptyset$. To see this, we proceed as follows. Take an arbitrary bridging zig-zag path $(\bar{p}_1,\ldots,\bar{p}_{\bar{m}})$ and consider $\bar{\mathcal{B}} := \mathcal{B}(X_n;(\bar{p}_1,\ldots,\bar{p}_{\bar{m}}))$. If $\mathcal{\bar{B}} = \emptyset$, we have concluded. If $\mathcal{\bar{B}} \neq \emptyset$, we can consider another bridging zig-zag path $(\tilde{p}_1,\ldots, \tilde{p}_{\tilde{m}})$ for $C_n$ having the same orientation as $(\bar{p}_1,\ldots,\bar{p}_{\bar{m}})$ and satisfying $\tilde{p}_2,\ldots, \tilde{p}_{\tilde{m}-1} \in \mathcal{\bar{B}}$. Define $\tilde{\mathcal{B}} := \mathcal{B}(X_n;(\tilde{p}_1,\ldots,\tilde{p}_{\tilde{m}}))$ and note that $\tilde{\mathcal{B}} \subset \bar{\mathcal{B}}$, $\#\tilde{\mathcal{B}} < \# \bar{\mathcal{B}}$. This construction can be iterated and after a finite number of iteration steps \EEE we find a bridging zig-zag path $(p_1,\ldots,p_m) \subset \mathcal{L}$ for $C_n$ such that $\mathcal{B}(X_n;(p_1,\ldots,p_m)) = \emptyset$. \EEE
Now suppose by contradiction \EEE that there was a ground state $C_n$, $n \ge 30$, with no acyclic bonds \EEE which contains two bridging zig-zag paths. Consider a bridging zig-zag path $(p_1,\ldots, p_m)$ satisfying $\mathcal{B}(X_n;(p_1,\ldots,p_m)) = \emptyset$ and define $X_n' = X_n \cup \bigcup_{j=2}^{m-1} p_j$. \EEE Clearly, we can assign charges to the atoms $p_2 ,\ldots, p_{m-1}$ \EEE to obtain a configuration $C_n'$ with \EEE alternating charge distribution. Note that $C_n'$ consists of $n+m-2$ atoms. We now estimate the energy of $C_n'$.
First, we observe that between the atoms of the path $(p_1,\ldots,p_m)$ there are $m-1$ bonds. Since $\mathcal{B}(X_n;(p_1,\ldots,p_m)) = \emptyset$, either each \EEE $p_2,p_4, \ldots, p_{2 \lfloor (m-1)/2 \rfloor }$ or each \EEE $p_3,p_5,\ldots, p_{2 \lfloor m/2 \rfloor -1 }$ is \EEE bonded to an atom of $C_n$. (The latter set is \EEE empty if $m=3$.) Thus, we obtain
\begin{align*}
\mathcal{E}(C_n') \le \mathcal{E}(C_n) - (m-1) - \min\lbrace \lfloor (m-1)/2 \rfloor , \lfloor m/2 \rfloor -1 \rbrace \le \mathcal{E}(C_n) - \frac{3}{2}m + \frac{5}{2}.
\end{align*}
In particular, as $C_n$ was supposed to be a ground state, this implies $\mathcal{E}(C_n') \le - \lfloor \beta(n) \rfloor - \frac{3}{2}m + \frac{5}{2} \le - \lfloor \beta(n+m-2) \rfloor$. Here the second inequality is elementary to check. This shows that $C_n'$ is a ground state.
As $C_n$ has two bridging zig-zag paths, there is at least one \EEE bridging zig-zag path for $C_n'$. We now repeat the above procedure. Choose $(p'_1,\ldots, p'_{m'})$ with $\mathcal{B}(X'_n;(p'_1,\ldots,p'_{m'})) = \emptyset$ and define a configuration $C_n''$ with $\# X_n'' \EEE = n+m+m'-4$, with alternating charge distribution and consisting of the atoms $X_n' \cup \bigcup_{j=2}^{m'-1} p'_j$. Arguing as before, we calculate
\begin{align*}
\mathcal{E}(C_n'') \le \mathcal{E}(C'_n) - \frac{3}{2}m' + \frac{5}{2} \le \mathcal{E}(C_n) - \frac{3}{2}(m+m') + 5.
\end{align*}
Since $C_n$ was supposed to be a ground state, this implies
\begin{align*}
\mathcal{E}(C_n'') &\le - \left\lfloor \frac{3}{2}n - \sqrt{\frac{3}{2}n} \right\rfloor - \frac{3}{2}(m+m') + 5 \\&\le - \left\lfloor\frac{3}{2} (n+ m + m'-4) - \sqrt{\frac{3}{2}(n + m + m'-4) }\right\rfloor -\frac{1}{2}.
\end{align*}
This implies $\mathcal{E}(C_n'') < -\lfloor \beta(n + m + m'-4) \rfloor$ which contradicts Theorem \ref{TheoremEnergyGroundstates}. This concludes the proof.
\end{proof}
Recall the construction of daisies in Section \ref{sec: daisy}. From \cite{Davoli15} we obtain the following result.
\begin{proposition}[Deviation from Wulff-shape]\label{prop: davoli}
Let $ n \ge 30$ and let $C_n$ be a ground state with no acyclic bonds. \EEE Then, possibly after translation, we find two daisies $X^{\rm daisy}_{6k_1^2} \subset \mathcal{L}$ and $X^{\rm daisy}_{6k_2^2} \subset \mathcal{L}$ with $X^{\rm daisy}_{6k_1^2} \subset X_n \EEE \subset X_{6k_2^2}^{\rm daisy}$ such that
$$0 < k_2 - k_1 \le cn^{1/4}, $$
where $c>0$ is a universal constant independent of $n$ and $C_n$.
\end{proposition}
\begin{proof}
From Theorem \ref{TheoremGroundstatesleq31} and the fact that $C_n$ does not have acyclic bonds \EEE we get that the ground state $C_n$ is a subset of the hexagonal lattice. Moreover, $C_n$ is repulsion-free, see Remark \ref{rem: repulsionsfree}. Thus, the energy of a ground state coincides with the one in \cite{Davoli15}, see \cite[Equation (6)]{Davoli15}. The claim then follows from \cite[Theorem 1.2]{Davoli15}.
\end{proof}
We are now in the position to prove Theorem \ref{TheoremCharge}.
\begin{proof}[Proof of Theorem \ref{TheoremCharge}]
As discussed at the beginning of the section, \EEE it remains to prove part (i) of the statement. In view of Theorem \ref{TheoremGroundstatesleq31} and Remark \ref{rem: main}(i), it suffices to treat the case that $C_n$ does not have acyclic bonds. \EEE We apply Proposition \ref{prop: davoli} to find two daisies with $X^{\rm daisy}_{6k_1^2} \subset X_n \subset X^{\rm daisy}_{6k_2^2}$. It is elementary to see that $X^{\rm daisy}_{6k_2^2} \setminus X^{\rm daisy}_{6k_1^2}$ can be written as the union of $6(k_2-k_1)$ zig-zag paths as introduced in Definition \ref{def: zig-zag}. We claim that $X_n \setminus X^{\rm daisy}_{6k_1^2}$ can be written as the union of at most $6(k_2-k_1)+ 1 \EEE$ zig-zag paths.
To see this, thanks to Lemma \ref{lemma: zig-zag paths}, observe that at most one of the $6(k_2-k_1)$ zig-zag paths may contain a bridging zig-zag path. Denote this zig-zag path by $\mathcal{P}_1=(p_1,\ldots,p_m)$ and the bridging zig-zag path by $(p_{k_1},\ldots,p_{k_2})$, where $1 \leq k_1 < k_2\leq m$. We indicate the two zig-zag paths $(p_1,\ldots,p_{k_1}) \cap X_n $ and $(p_{k_2},\ldots,p_{m}) \cap X_n$ by $ \gamma_0$ and $\gamma_1 $ respectively.
For the remaining zig-zag paths segmenting $X^{\rm daisy}_{6k_2^2} \setminus X^{\rm daisy}_{6k_1^2}$, denoted by $\mathcal{P}_k$, $k=2,\ldots,6(k_2-k_1)$, we define $\gamma_k:=\mathcal{P}_k \cap X_n$. Recalling that only $\mathcal{P}_1$ may contain a bridging zig-zag path, we get that $\gamma_k$ is a zig-zag path for all $k=0,\ldots,6(k_2-k_1)$. Moreover, $\gamma_k \subset X_n$ and thus $X_n \setminus X_{6k^2_1}^\mathrm{daisy}$ can be written as the union of the $6(k_2-k_1)+ 1 \EEE$ (possibly empty) zig-zag paths $\gamma_k$, $k=0,\ldots,6(k_2-k_1)$. \EEE
Recall that $C_n$ has alternating charge distribution and therefore the net charge of each zig-zag path is in $\lbrace -1,0,1\rbrace$. Also recall from Section \ref{sec: daisy} that daisies always have net charge zero. This implies that the net charge of the configuration $C_n$ satisfies
$$|\mathcal{Q}(C_n) | \le 6(k_2-k_1)+1\EEE.$$
The statement follows from the fact that $k_2 - k_1 \le cn^{1/4}$, see Proposition \ref{prop: davoli}.
\end{proof}
\section*{Acknowledgements}
M.\ F.\ acknowledges support from the Alexander von Humboldt Stiftung.
L.\ K.\ acknowledges support
from the Austrian Science Fund (FWF) project P~29681, and from the Vienna Science and Technology Fund (WWTF), the
City of Vienna, and the Berndorf Private Foundation through Project MA16-005. The authors would like to thank Ulisse Stefanelli for turning their attention to this problem.
\EEE
|
\section{Introduction}
\label{sec: Intro}
Methods for pulse shaping~\cite{Glaser2015} to control quantum processes have allowed important advances in different domains, ranging from the steering of photo-chemical processes~\cite{shapiro2003principles,Balint2008,Sola2018} to the optimization of gate operations in quantum computing~\cite{Khaneja2005_grape,Motzoi2009}.
In quantum information, optimal control theory (OCT) is typically used to generate target unitary operators~ \cite{Khaneja2005_grape, Caneva2011_crabe, Kelly2014, Egger2014, Machnes2018_goat}.
Within the field of superconducting qubits~\cite{Devoret2013} OCT has been successfully applied to design various qubit gates in different hardware implementations \cite{Motzoi2009, Egger2013a, Schutjens2013_wahwah, Liebermann2017, Heeres2017} as well as to identify optimal operating conditions, such as the quasi-dispersive regime~\cite{Goerz2017}.
In parallel to OCT, local control theory (LCT) has also emerged as a valuable approach to control the dynamics of quantum systems by shaping external fields.
In particular, LCT has already been successfully applied to steer photo-chemical reactions in molecular systems~\cite{Marquetand2007, Volker2009, Curchod2015}.
In LCT, an external field is designed on-the-fly under the constraint that it monotonically increases the quantum population of a selected target state when starting from a given initial state~ \cite{Kosloff1992, Curchod2011}.
While OCT is based on a computationally intensive variational approach, which requires computing the full time evolution of the system at each optimization step, LCT can generate pulses that produce the desired population transfer by computing the evolution of the system only once.
Although LCT does not necessarily provide a time-optimal pulse, thanks to its remarkable computational efficiency and conceptual simplicity, it can nonetheless become the method of choice for the design of state preparation pulses.
In this paper, we focus our investigation on the application of LCT to generate state preparation pulses for fixed-frequency superconducting qubits coupled via tunable couplers.
In Sec.~\textcolor{red}{\ref{sec: methods}} we introduce LCT and show how to apply it to a setup made-up of fixed-frequency transmon qubits coupled by a tunable coupler~\cite{McKay2016, Roth2017}.
Sec.~\ref{subsec:bare_lct} presents and discusses the pulses generated by the LCT algorithm.
Sec.~\ref{subsec:lct_optimization}-\ref{subsec:truncation} reports on a procedure aimed at further optimizing their properties such as bandwidth, pulse length, and gate reversibility.
\section{Methods}
\label{sec: methods}
\subsection{Theoretical background}
\label{subsec: theory}
We consider $n$ fixed-frequency qubits all mutually interacting through a single flux-tunable qubit, called tunable coupler (TC)~\cite{McKay2016}.
Such systems combine the long coherence time of fixed-frequency transmon qubits with the high controllability of flux-tunable coupling elements.
The system is described by the Hamiltonian~\cite{Roth2017}
\begin{align}
\hat{H}(t) =& -\frac{1}{2} \sum_{i=1}^{n} \omega_{i} \hat{\sigma}_{i}^{z} + \sum_{i=1}^{n} g_{i} \left( \hat{\sigma}_{i}^{+} \hat{\sigma}_{\text{TC}}^{-} +\hat{\sigma}_{i}^{-} \hat{\sigma}_{\text{TC}}^{+} \right) \nonumber \\
&-\frac{1}{2} \omega_{\text{TC}}(t) \hat{\sigma}_{\text{TC}}^{z} \text{ ,} \label{eq1}
\end{align}
in units of $\hbar=1$.
The qubit $i$ and TC raising and lowering operators are $\hat\sigma_i^+$, $\hat\sigma_i^-$, $\hat\sigma_\text{TC}^+$ and $\hat\sigma_\text{TC}^-$, respectively, while the number operators are $\hat\sigma^z_i$ and $\hat\sigma^z_\text{TC}$.
Qubit $i$ has frequency $\omega_i$ and couples with strength $g_i$ to the TC.
The frequency of the TC, $\omega_\text{TC}(t)$, is controlled by a current $I(t)$ brought close to the TC by a high-speed flux bias line, see Fig.~\ref{fig1}(a).
The resulting flux $\Phi(t)$ threading through the TC SQUID loop changes the frequency of the TC according to
\begin{align} \label{Eqn:TCPhi}
\omega_{\text{TC}}(t) = \omega^{0}_{\text{TC}} \sqrt{ | \cos{(\pi \Phi(t) / \Phi_{0})} | } \text{,}
\end{align}
where $\Phi_{0}$ is the magnetic flux quantum \cite{Koch2007}.
The full system wave function $\ket{\Psi(t)}$ then evolves according to the time-dependent Schr\"{o}dinger equation
\begin{equation}
\imath \partial_t \ket{\Psi(t)} = \hat{H}(t)\ket{ \Psi(t) }\text{.}
\label{eq6}
\end{equation}
The population $\langle\hat{P}_\phi\rangle$ of any $n$-qubit target state $\ket{\phi}$ is governed by
\begin{equation}
\partial_t \langle \hat{P}_{\phi} \rangle = \imath \langle \left[ \hat{H}(t), \hat{P}_{\phi} \right] \rangle\text{,} \label{eq5}
\end{equation}
where $\hat{P}_{\phi} = \ket{ \phi }\!\!\bra{ \phi }$ is the corresponding projector operator and $\langle \dots \rangle$ denotes the expectation value with respect to $\ket{ \Psi(t)}$.
In our model, the only free, tunable parameter is the frequency of the tunable coupler $\omega_{\text{TC}}(t)$.
We will, thus, employ LCT to increase the population in $\ket{\phi}$ by shaping $\omega_\text{TC}(t)$ on-the-fly.
The TC frequency can be decomposed into a time-independent and a time-dependent part $\omega_{\text{TC}}(t) = \omega_{\text{TC}} + \delta\omega_{\text{TC}}(t)$ \cite{Curchod2011, Curchod2015}.
This splits the Hamiltonian $\hat{H}(t)$ into a time-dependent $\hat{H}^{\prime}(t)=-\delta\omega_\text{TC}(t) \hat\sigma_\text{TC}^{z}/2$ and a drift term
\begin{align}
\hat{H}_{\text{d}} =& -\frac{1}{2} \sum_{i=1}^{n} \omega_{i} \hat{\sigma}_{i}^{z} + \sum_{i=1}^{n} g_{i} \left( \hat{\sigma}_{i}^{+} \hat{\sigma}_{\text{TC}}^{-} +\hat{\sigma}_{i}^{-} \hat{\sigma}_{\text{TC}}^{+} \right) \nonumber \\
& -\frac{1}{2} \omega_{\text{TC}} \hat{\sigma}_{\text{TC}}^{z}\text{.} \label{eq3}
\end{align}
The drift $\omega_{\text{TC}}$ term depends on the constant DC flux bias applied to the TC~\cite{McKay2016}.
When the target state $\ket{\phi}$ is an eigenvector $\ket{\psi_{j}}$ of the drift Hamiltonian $\hat{H}_{\text{d}}$ the projector operator $\hat{P}_{\phi}$ commutes with $\hat{H}_{\text{d}}$
and Eq.~\eqref{eq5} simplifies to
\begin{equation}
\partial_t \langle \hat{P}_{j} \rangle = -\frac{\imath}{2} \delta\omega_{\text{TC}}(t) \langle \left[ \hat{\sigma}_{\text{TC}}^{z}, \hat{P}_{j} \right] \rangle \text{.} \label{eq7}
\end{equation}
LCT induces a monotonous increase of the target state population by generating a $\delta\omega_{\text{TC}}(t)$ pulse that guarantees the positivity of the right hand side of Eq.~\eqref{eq7}.
For our setup this condition is achieved by changing the frequency of the TC according to
\begin{equation}
\delta \omega_{\text{TC}}(t) = \frac{\imath}{2} \lambda
\langle \left[ \hat{\sigma}_{\text{TC}}^{z}, \hat{P}_{j} \right] \rangle^*
\text{.} \label{eq8}
\end{equation}
The coupling parameter $\lambda$ controls the magnitude by which the control field $\delta \omega_{\text{TC}}(t)$ is changed.
Its value can be tuned as long as the resulting pulse $\delta \omega_{\text{TC}}(t)$ can be implemented in realistic experimental setups.
In the case of large systems with many possible states, the implementation of the LCT scheme can become numerically challenging.
However, when some of the (high energy) states do not contribute to the dynamics, we can restrict the action of the LCT algorithm to a subspace of the full Hilbert space using the projector operator $\hat{P}_{n'}=\sum_{k=1}^{n'} \ket{ \psi_{k} }\!\bra{ \psi_{k} }$ over the first $n'$ eigenvectors (assumed to be ordered according to their corresponding eigenvalues). Equation (\ref{eq8}) then simplifies to
\begin{align} \label{eq_proj}
\delta \omega_{\text{TC}}(t) \simeq \\
- \lambda \text{ Im} \Big( \sum_{k}^{n'} & \bra{ \psi_{j} } \hat{\sigma}_{\text{TC}}^{z} \ket{ \psi_{k} } \braket{ \psi_{k} | \Psi(t) }
\braket{ \psi_{j} | \Psi(t) }^* \notag
\Big) \text{.}
\end{align}
Since the TC frequency cannot exceed $\omega^0_\text{TC}$, see Eq.~\eqref{Eqn:TCPhi}, $\delta\omega_{\text{TC}}(t)$ is confined to the interval $[-\omega^0_{\text{TC}},0]$.
Thus, it is necessary to impose a restriction on the $\lambda$-factor in order to avoid reaching the upper bound of $\delta\omega_{\text{TC}}(t)$.
This is accomplished by capping the value of $\delta\omega_{\text{TC}}(t)$ to 0 (i.e. taking $\min[\delta\omega_{\text{TC}}(t),0]$) and by
constraining the magnitude of $\lambda$ such as $\delta\omega_{\text{TC}}(t)>-\omega_{\text{TC}}^{0}$.
The LCT algorithm can be summarized in two steps: First, the instantaneous state
is propagated for a short time interval $[ t, t + \delta t]$ under $\hat H(t)$.
Second, the resulting wavefunction $\ket{ \Psi(t+\delta t)}$ is used to update the external field using Eq.~\eqref{eq_proj}.
These two steps are repeated using the updated control field until the desired population transfer is achieved.
A smooth external driving pulse is obtained when $\delta t$ is made sufficiently small.
For practical purposes, when the target state $\ket{ \phi }$ does not overlap with the initial system wavefunction $\ket{ \Psi(0) }$ a small fraction $\eta$ of the target state is added into the initial wavefunction
\begin{align} \label{Eqn:seed}
\ket{ \Psi^{\prime}(0) } = \sqrt{\eta} \ket{ \psi_{j} } + \sqrt{1-\eta} \ket{ \Psi(0) }
\end{align}
to ensure that the LCT algorithm converges.
\subsection{System}
\label{subsec: system}
We apply LCT to a system composed of $n=2$ qubits, see Eq.~\eqref{eq1} and Fig.~\ref{fig1}(a).
The qubits, labeled Q1 and Q2, are set at realistic~\cite{McKay2016} frequency values, $\omega_{1}/(2 \pi)=5.890~\rm{GHz}$ and $\omega_{2} / (2 \pi)=5.031~\rm{GHz}$, respectively.
They are coupled with strengths $g_{1}/(2 \pi)=100~\rm{MHz}$ and $g_{2} / (2 \pi)=71~\rm{MHz}$ to a TC with a maximal frequency $\omega^{0}_{\text{TC}}/(2 \pi) = 7.445~\rm{GHz}$.
The control pulses are designed in such a way that $\delta \omega_{\text{TC}}$ is 0 at the beginning and the end of the simulation.
The model can be accurately described using the the first two states of the qubits and of the TC since the higher energy states do not affect the process of interest, namely the population transfer between the states $\ket{0}$ and $\ket{1}$ of Q1 and Q2.
The eigenvectors of the drift Hamiltonian in Eq.~\eqref{eq3}, labeled $\ket{q_{1}q_{2}q_{\text{TC}}}$, are used to identify the $2^3$ system states, see Fig.~\ref{fig1}(b).
Due to the modest size of the problem, we do not need to introduce projectors as described in Eq.~\eqref{eq_proj}.
When $\delta\omega_\text{TC}$ is swept from $0$ to $-3~\rm{GHz}$, we observe two avoided level crossings between the TC state and the qubit states, see Fig.~\ref{fig1}(c).
The associated nonadiabatic coupling terms, obtained with the Hellmann-Feynman expression~\footnote{The nonadiabatic coupling terms $d_{jk}$ between pairs of the full system \eqref{eq1} eigenstates $\ket{\psi_{j}}$ and $\ket{\psi_{k}}$ with corresponding eigenenergies $\varepsilon_{j}$ and $\varepsilon_{k}$ given by relation
$
\hat{H}(\delta \omega_{\text{TC}}) \ket{\psi_{j}(\delta \omega_{\text{TC}})} = \varepsilon_{j}(\delta \omega_{\text{TC}}) \ket{\psi_{j}(\delta \omega_{\text{TC}})}
$
for a certain $\delta \omega_{\text{TC}}$ pulse value are determined with the Hellmann-Feynman expression \cite{Singh1989}
$
d_{jk}(\delta \omega_{\text{TC}}) = \frac{\bra{\psi_{j}(\delta \omega_{\text{TC}})} \frac{\partial \hat{H}(\delta \omega_{\text{TC}})}{ \partial \delta \omega_{\text{TC}}} \ket{\psi_{k}(\delta \omega_{\text{TC}})}}{\varepsilon_{j}(\delta \omega_{\text{TC}}) - \varepsilon_{k}(\delta \omega_{\text{TC}})}
$
}
are shown in Fig.~\ref{fig1}(d).
The LCT algorithm will make use of these avoided level crossings to transfer population between the two qubits.
\begin{center}
\begin{figure}[t]
\begin{tikzpicture}
\node at (0,4.25) {\includegraphics[width=0.48\textwidth, clip, trim=100 100 475 225]{Figures/TCSketch.pdf}};
\node at (0,0) {\includegraphics[scale=0.4]{Figures/LCT_fig_1.pdf}};
\node at (-3.55,6.00) {(a)};
\node at (-1.45,2.475) {(c)};
\node at (-1.45,-0.05) {(d)};
\node at (-3.55,2.475) {(b)};
\node at (-0.08,3.80) {I(t)};
\node at (-2.75,3.80) {Qubit 1};
\node at ( 2.40,3.80) {Qubit 2};
\node at (-0.50,6.00) {TC};
\end{tikzpicture}
\caption{(a) Sketch of two fixed-frequency transmon qubits coupled using a tunable coupler.
(b) Energy level scheme of the undriven system.
(c) Evolution of the eigenstates of the full Hamiltonian in Eq.~(\ref{eq1}) as a function of $\delta \omega_{\text{TC}}$. The labels $\ket{010}$ (dotted-dashed orange), $\ket{100}$ (dashed blue) and $\ket{001}$ (dotted green) refer to the eigenstates of the Hamiltonian when the coupling is set to zero.
(d) The nonadiabatic couplings $d_{1,2}$ (dotted-dashed orange), $d_{2,3}$ (dashed blue) and $d_{1,3}$ (dotted green), as functions of $\delta \omega_{\text{TC}}$.
}
\label{fig1}
\end{figure}
\end{center}
\section{Results and Discussion}
\label{sec: results_and_discussion}
\subsection{LCT pulse}
\label{subsec:bare_lct}
\begin{figure*}[htbp!]
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[width=0.98\textwidth]{Figures/LCT_fig_2.pdf}};
\draw[-latex] (-6.88,0.95) -- +(1.15,0.0) node[pos=0.5,below] {$t_\text{on}$};
\node at (-6.55,2.2) {($\text{a}_1$)};
\node at (-6.55,0) {($\text{a}_2$)};
\node at (-6.55,-3.35) {($\text{a}_3$)};
\node at (-1.55,2.2) {($\text{b}_1$)};
\node at (-1.55,0) {($\text{b}_2$)};
\node at (-0.94,-3.35) {($\text{b}_3$)};
\node at (3.45,2.2) {($\text{c}_1$)};
\node at (3.45,0) {($\text{c}_2$)};
\node at (3.45,-3.35) {($\text{c}_3$)};
\end{tikzpicture}
\caption{
($\text{a}_1$) LCT pulse designed to transfer population from state $\ket{100}$ to state $\ket{010}$ (Inset: full 450 ns pulse).
The parameter $\lambda$ was set to
$12500$.
($\text{b}_1$) Frequency filtered pulse $\delta\omega_\text{TC}^\text{filt}(t)$ used as an initial condition to design the second local control pulse.
($\text{c}_1$) LCT pulse designed to transfer population from state $\ket{100}$ to state $\ket{010}$ when using the pulse in ($\text{b}_1$) as an initial condition.
($\text{a}_2$), ($\text{b}_2$) and ($\text{c}_2$) population transfer resulting from the pulses in ($\text{a}_1$), ($\text{b}_1$) and ($\text{c}_1$), respectively.
($\text{a}_3$), ($\text{b}_3$) and ($\text{c}_3$) Fourier transforms of the pulses in ($\text{a}_1$), ($\text{b}_1$) and ($\text{c}_1$), respectively.
The dashed lines indicate the harmonics corresponding to frequency differences between the qubits.
}
\label{FigLC1}
\end{figure*}
In this section, we design a LCT
pulse that achieves population transfer from the state $\ket{100}$ to the state $\ket{010}$, i.e. that brings the excitation from Q1 to Q2.
We assume that the TC is biased at the flux sweet spot $\Phi(t=0)=0$.
Since the initial and final states are orthonormal, we use the state preparation in Eq.\ (\ref{Eqn:seed}) with $\eta=10^{-6}$ to initialize the LCT algorithm.
Figure \ref{FigLC1}($\text{a}_1$) shows a $150~\rm{ns}$ long LCT pulse obtained for $\lambda=12500$.
This pulse makes the tunable coupler energy level oscillate between the two avoided level-crossings depicted in Fig.~\ref{fig1}(c).
As the TC $\ket{001}$ state passes through the first avoided crossing at $-1.56~\rm{GHz}$ a fraction of the qubit population in $\ket{100}$ is transferred to the TC.
Part of this population is then transferred to the second qubit (state $\ket{010}$) once the second avoided crossing at $-2.40~\rm{GHz}$ is reached.
The TC oscillates with a complicated frequency pattern dominated by the harmonics of the transition between the two qubits, $(\omega_1-\omega_2)/(2\pi)=859~\rm{MHz}$ and by other components below $1~\rm{GHz}$ as shown by the power spectrum of the pulse in Fig.~\ref{FigLC1}($\text{a}_3$).
It is important to note that, despite the many frequencies appearing on the Fourier transform of the LCT pulse, no other transition further than the targeted ones are excited during LCT process.
As expected from Eqs. \eqref{eq7} and \eqref{eq8}, after an initial delay $t_\text{on}$ of about $170~\rm{ns}$, the population of the target state ($\ket{010}$) increases monotonically with time while the populations of the other states considered in the simulation shows important high frequency oscillations.
At the end of the transfer process ($\sim 300~\rm{ns}$), the initial population has been almost entirely transferred to the target state, achieving a mismatch $1-P_{\ket{010}}$ of less than $10^{-6}$, where $P_{\ket{010}}$ is the population of the target state.
While very promising, this first `high fidelity' LCT pulse has a highly complex spectrum and its implementation requires instruments with a large bandwidth.
\subsection{Optimization of LCT pulses}
\label{subsec:lct_optimization}
Because of limits set by the control instruments, large bandwidth pulses are impractical to generate.
We therefore need a procedure to refine the LCT pulse, which allows to confine the bandwidth within a reasonable range.
To this end, we apply a high frequency filter
to the LCT pulse obtained in the previous section and use it as a `reference' to generate an improved pulse using the LCT algorithm.
This new reference corresponds to the term $\delta\omega^\text{filt}_\text{TC}(t) $ in Eq.~\eqref{Eq_LCT_in_3_parts}.
In practice, we decompose the new LCT pulse into three different components
\begin{align} \label{Eq_LCT_in_3_parts}
\omega_\text{TC}(t)=
\omega^0_\text{TC} + \delta\omega^\text{filt}_\text{TC}(t) + \delta\omega^{\text{lct},2}_\text{TC}(t) \, .
\end{align}
Only the component $\delta\omega^{\text{lct},2}_\text{TC}(t)$, initially set to $0$, will be generated on-the-fly using the LCT algorithm,
while the first two terms
are kept fixed.
As in the previous section, the pulse $\delta\omega_\text{TC}^{\text{lct},2}$ is shaped on-the-fly using the requirement that the right-hand side of Eq.~\eqref{eq7}, $\partial_t \langle \hat{P}_{j} \rangle$, remains positive.
The filtered pulse $\delta\omega^\text{filt}_\text{TC}(t)$ in Fig.~\ref{FigLC1}($\text{b}_1$) is obtained by applying a high frequency cut-off at $0.4~\rm{GHz}$ to the pulse in Fig.~\ref{FigLC1}($\text{a}_1$). The corresponding spectra before and after the application of the filter are shown in Fig.~\ref{FigLC1}($\text{a}_3$) and Fig.~\ref{FigLC1}($\text{b}_3$), respectively.
This operation removes much of the complex structure of the pulse while preserving its overall shape (Fig.~\ref{FigLC1}($\text{b}_1$)).
As expected, the pulse composed by the first two components in Eq.~\eqref{Eq_LCT_in_3_parts} fails to transfer the population to the target qubit (Fig.~\ref{FigLC1}($\text{b}_2$)).
However, using LCT we can generate a new time-dependent field, i.e. $\delta\omega^{\text{lct},2}_\text{TC}(t)$ in Eq.~\eqref{Eq_LCT_in_3_parts} with coupling parameter $\lambda_2$, which restores this property.
In particular, we are able to design new LCT pulses with a narrow bandwidth and an error $1-P_{\ket{010}}<10^{-6}$ using a wide range of $\lambda_2$ in the interval $[100, 1000]$, see Fig.~\ref{FigLC1}($\text{c}_1$-$\text{c}_3$).
In addition, the population transfer is now completed in only $\sim 30~\rm{ns}$ (see Fig.~\ref{FigLC1}($\text{c}_2$)) compared to the initial $120~\rm{ns}$ obtained with the first LCT run described in Sec.~\ref{subsec:bare_lct} (Fig.~\ref{FigLC1}($\text{a}_2$)).
This improvement results from the nature of the `reference' pulse, i.e. the sum $\omega_\text{TC}^{0} + \delta\omega^\text{filt}_\text{TC}(t) $, forcing the TC frequency in the energy range that matches the separation between the two avoided crossings shown in Fig.~\ref{fig1}(c).
Note that the truncation of the power spectrum in Fig.~\ref{FigLC1}($\text{c}_3$) above 1GHz (1.5GHz) without further optimization will reduce the fidelity to $10^{-4}$ ($10^{-5}$).
\subsection{Reverse processes}
\label{subsec:reverse}
So far, the LCT pulses were generated to accomplish a well defined transition from a given initial state to a final state.
Therefore, we cannot expect that by applying the same pulse to the final state it can revert the process and transfer the population back to the initial state.
For instance, applying the pulse in Fig.~\ref{FigLC1}($\text{a}_1$) (generated for the population transfer from $\ket{100}$ to $\ket{010}$) to the reverse process (from $\ket{010}$ back to $\ket{100}$) we only achieve an imperfect transfer that leaves $29\%$ of the population in the TC.
Interestingly, we found that the amount of population trapped in the TC is particularly sensitive to the value of the parameter $\lambda_2$.
Therefore, the LCT pulse can be further optimized to increase the efficiency of the reverse transfer by tuning $\lambda_2$.
Note that changes to $\lambda_2$ do not affect the success of the population transfer from $\ket{100}$ to $\ket{010}$, since the conditions (initial and final states) and the reference pulse $\delta\omega_\text{TC}(t)$ are kept fixed.
Exploiting this fact, we illustrate a procedure for the recursive optimization of the direct and reverse population transfers between the states $\ket{100}$ and $\ket{010}$, starting from the bandwidth optimized pulse derived in Sec.~\ref{subsec:lct_optimization}.
For an initial choice of $\lambda_2$, we derive a first LCT pulse for the direct process ($\ket{100}$ to $\ket{010}$) and then test it for the reverse transfer ($\ket{010}$ to $\ket{100}$).
If this fails to accomplish a population transfer back to the initial state $\ket{100}$ with an error $1-P_{\ket{100}}$ less than $10^{-6}$ we update the parameter $\lambda_2$ and recompute the pulse using the LCT algorithm.
This procedure is repeated until the reverse population transfer fidelity reaches a maximum.
The Nelder-Mead algorithm \cite{Nelder1965} is used to optimize $\lambda_{2}$.
In some cases, we noticed that maximizing the population transferred during the reverse process required a change of the frequency cut-off values for $\delta\omega_\text{TC}^\text{filt}(t)$.
\begin{figure}[t]
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[width=0.45\textwidth]{Figures/Flow_chart.pdf}};
\node at (-2.1cm,2.40cm) {ref. pulse: $\omega^0_{\text{TC}}$};
\node at (-1.35cm,-0.25cm) {ref. pulse: $\omega^0_{\text{TC}}+\delta\omega^{\text{filt}}_{\text{TC}}(t)$};
\end{tikzpicture}
\caption{
Flow chart showing the iterative procedure used to obtain pulses with a smaller bandwidth and capable of transferring population when the initial and target states are exchanged.
A new LCT calculation is performed each time the parameter $\lambda_2$ is updated or when a new frequency cut-off is applied.
}
\label{Scheme1}
\end{figure}
A flow chart of the algorithm used to obtain a narrow bandwidth pulse able to transfer the qubit population in both directions is shown in Fig.~\ref{Scheme1}.
For the setup in Fig.~\ref{fig1} and the parameter discussed in Section~\ref{subsec: system}, the produced LCT pulse is given in Fig.~\ref{FigLC4}(a) together with the population dynamics for the direct and reverse processes respectively shown in Fig.~\ref{FigLC4}(b) and (c).
The final LCT pulse in Fig.~\ref{FigLC4}(a) can further be used to inspire a new class of fully analytical and ultrashort pulses that can be used for state preparation (see Appendix~\ref{Appendix1}).
\begin{figure}[t]
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[width=0.475\textwidth]{Figures/LCT_fig_3.pdf}};
\node at (2.5,4.5) {$\tau$};
\node at (-2.3,1.9) {(a)};
\node at (-2.3,-0.55) {(b)};
\node at (-2.3,-3.15) {(c)};
\end{tikzpicture}
\caption{
(a) Final pulse generated by the algorithm depicted in Fig.~\ref{Scheme1} (final frequency cut-off at $0.45~\rm{GHz}$, $\lambda_{2}=437.4$).
The inset shows the tail of the pulse (after the time $\tau$), which can be substituted with the half-Gaussian function (dashed red line).
(b) Evolution of the system when all the population is initially in Q1.
(c) Evolution of the system when all the population is initially in Q2.
The populations shown in (b) and (c) are calculated using the pulse with the shortened tail.
}
\label{FigLC4}
\end{figure}
\subsection{Pulse truncation}
\label{subsec:truncation}
The pulses obtained using the algorithm shown in Fig.~\ref{Scheme1} still have a long tail in the time domain that is inherited from the original, fixed $\lambda$, LCT calculation (Section~\ref{subsec:bare_lct}).
Since the tail does not contribute to the population transfer, see e.g. Fig.~\ref{FigLC1}($\text{c}_2$), the pulses can be shortened by imposing a Gaussian decay after a critical time $\tau$ using the half Gaussian function $\alpha \exp\{-(t-\tau)^2/(2\sigma^{2})\}$ for $t \geq \tau$ shown in Fig.~\ref{FigLC4}(a).
The optimal value of $\tau$ is obtained by including it in the optimization process shown in Fig.~\ref{Scheme1}, while its
initial value is selected as the time required by the original pulse in Fig.~\ref{FigLC4}(a) to reach 99\% of population transfer for the reverse process in Fig.~\ref{FigLC4}(c).
For a chosen $\sigma$ value, this leads to an optimized pulse where population transfer fidelities $1-P_{\ket{010}}$ and $1-P_{\ket{100}}$ are both less than $10^{-6}$ for the forth and back population transfer, respectively.
Finally, the choice of the parameter $\alpha$ is imposed by the need to guarantee continuity at the transition point.
\section{Conclusion}
\label{sec:conclusion}
In this work we propose using local control theory (LCT) to manipulate qubit populations in an architecture where fixed-frequency superconducting qubits are coupled using tunable couplers.
Given the initial and target states, LCT constructs a pulse on-the-fly by computing the time evolution
only once.
The only tunable parameter is the intensity of the applied pulse (controlled by $\lambda$ in Eq.~\eqref{eq8}).
$\lambda$ influences the shape and length of the resulting LCT pulse, giving the possibility to shorten the transfer time below 50 ns while keeping a high fidelity for the process.
The LCT algorithm was extended to design pulses that can achieve a complete population transfer in both directions between the initial and the target states.
This extension of the LCT algorithm comprises an additional optimization step over the parameter $\lambda$.
LCT can also serve as a starting point for a deterministic procedure to further reduce the complexity of the pulse, see Appendix.
This opens up a new avenue of research to design efficient gates for different applications of quantum computing~\cite{moll_quantum_optimization_2018,Egger2018, Barkoutsos2018a}.
Further work will investigate the sensitivity of LCT pulses to the different parameters characterizing the model Hamiltonian (Eq.~\eqref{eq1}), as well as using LCT in systems with more elements where frequency crowding may become an issue.
Such systems could for instance include $n>2$ qubits coupled to the same tunable coupler.
\section{Acknowledgment}
The authors acknowledge stimulating discussions with Marco Roth and Nikolaj Moll. We also acknowledge generous computational time from the Croatian National Grid Infrastructure (CRO-NGI) and the Irish Centre for High-End Computing (ICHEC).
This work has received funding from the European Union's Horizon 2020 research and innovation program under the grant agreement No. 676531 (project E-CAM).
|
\section{Introduction}
\label{s:intro}
With the increased accessibility of high energy, coherent light
sources, there has been a resurgence of interest in inverse imaging
problems where the measurement can be interpreted as the modulus of
the Fourier transform of an unknown object (see, for example,
\cite{chapman2006,elser2003,elser2007,fienup1987,miao1999,miao2015,Osherovich,bendory17}).
In a typical set-up, an object with electron density $\rho(\bx)$ is
irradiated with a planar beam of coherent x-rays and a measurement of
the scattered wave is made in the far field, i.e.\ the Fraunhofer
regime. The wave's phase information is not directly measurable, and
the data collected are interpreted as samples of the intensity of the
Fourier transform, $\{|\hrho(\xi_j)|^2:\:j\in \cJ\}$. The
reconstruction problem is then largely reduced to that of
``recovering'' the phases of the complex numbers
$\{\hrho(\xi_j):\:j\in \cJ\}$. This experimental approach is referred
to as coherent diffraction imaging (CDI).
\begin{remark}
The {\em phase retrieval} problem came to prominence in x-ray crystallography,
where $\rho(\bx)$ is assumed to be a periodic function with some unit cell and
$\{\hrho(\xi_j)\}$ are the coefficients of a discrete Fourier series, defined
only on a regular lattice. In this context, without additional information,
the phase retrieval problem is obviously ill-posed. One can assign any value to the phase
of each $\hrho(\xi_j)$ to produce a periodic image. To circumvent this
problem, for sufficiently small molecules, direct methods that rely on the
non-negativity of the electron density and algebraic relations (Karle-Hauptman
determinants) proved to be very powerful \cite{millane1990}. For larger
structures, a variety of experimental approaches have been introduced to supply
additional information that permits the reconstruction of the phase information
needed to reconstruct the original crystal \cite{LaddPalmer}.
\end{remark}
In the present paper, we are interested in the setting where
$\rho(\bx)$ is an essentially arbitrary, but compactly supported,
function. The study of this problem dates back to 1952, when Sayre
noted that for amorphous {\em non-crystalline} objects, one may obtain
values of the intensity $|\hrho(\xi_j)|^2$ on a finer mesh than in the
periodic case \cite{Sayre1952}, since the spectrum is continuous. In
essence, he proposed that one {\em ``over-sample''} $|\hrho(\xi)|^2$ by
a factor of two in each direction then recover $\rho(\bx)$ as the solution to an
overdetermined, constrained nonlinear least squares problem. The
constraints come from some prior knowledge of $\rho$, such as its
support, whether it is non-negative, etc. It turns out that Sayre's
conjecture is essentially correct in more than one spatial
dimension. More precisely, a finite approximation to phase retrieval
problem introduced below in Section~\ref{s.dcpp}, with support as the
auxiliary information, has a solution, generically unique up to
``trivial associates;'' see~\cite{barakat1984,hayes1982,hayes1987,bendory17}.
In the continuum case, the trivial associates are obtained by applying
operations to $\rho$ that leave $|\hrho|$ invariant: translations
($\rho(\bx) \rightarrow \rho(\bx-\ba)$ for some $\ba$) and inversion
($\rho(\bx) \rightarrow \rho(-\bx)$).
The most common additional information used in phase retrieval is a
support constraint: that $\rho(\bx)$ is nonzero only within some
closed and bounded region $D$ in ${\mathbb R}^d$. It is with reference
to an estimate for the support of $\rho(\bx)$ that one speaks about
over-sampling its magnitude Fourier transform. If $R\supset D$ is the
smallest rectangle covering $D$, then the Fourier data
$\{|\hrho(\bk_j)|\}$ must be sampled on a grid fine enough to
represent a periodic function with fundamental cell a rectangle whose
side lengths are twice those of $R$. In the phase retrieval
literature, this is called ``double oversampling.'' In fact some
degree of oversampling is clearly needed so that the sampled data
contains adequate information about the support of $\rho(\bx)$ for
reconstruction to be possible. Double oversampling is the assumption
required for the proof of the basic uniqueness theorem (Hayes's
theorem) in coherent diffraction imaging; it also allows the autocorrelation function of $\rho,$
\begin{equation}
\rho\star\rho(\bx)=\int\rho(\bx+\by)\rho(\by)d\by,
\end{equation}
to be reconstructed from the magnitude Fourier data without aliasing artifacts.
The earliest practical method for solving the phase retrieval problem
is a variant of the alternating projection algorithm due to Saxton and
Gerchberg \cite{gerchberg1972}. The basic idea, which was first
introduced in the context of Banach spaces by von Neumann as a method
to find the intersections of convex sets, is quite general. In phase
retrieval we let $A$ denote the collection of images $\rho$ with the
given magnitude Fourier data and let $B$ denote the set of images
which satisfy the support constraint. Given a function $f(\bx)$,
projection onto $A$ corresponds to computing its Fourier transform
$\hf$, keeping the phase information from $\hf$ and replacing the
modulus with the measured data $|\hrho(\xi)|$. We denote this
operator by $P_A$. Projection onto $B$ corresponds to multiplying
$f(\bx)$ by the characteristic function of $D$. We denote this
operator by $P_B$. Alternating projection can then be written as the
following iteration:
\begin{equation}
\rho_{k+1} = P_B\circ P_A(\rho_k),
\label{apiter}
\end{equation}
with some initial guess $\rho_0$.
This algorithm has a long history when $A$ and $B$ are convex sets,
which we do not
seek to review here (see, for example, \cite{bauschke1996}). It has
also received a lot of study in the non-convex setting
\cite{andersson2013,bauschke2002,borwein}.
Unfortunately, alternating projection often
converges to fixed points unconnected to the reconstruction problem at
hand. To overcome this, Fienup proposed a new class of so-called
hybrid input-output (HIO) algorithms
\cite{bauschke2002,fienup1982,fienup1987}, which
were placed into the larger framework of \emph{difference-maps} by
Elser and collaborators, see~\cite{elser2003,elser2007}.
The fixed points
of these algorithms all specify correctly reconstructed objects.
We will describe this method in detail below in sections
\ref{sec4} and \ref{sec5}.
Note that methods from continuous optimization have also been
applied to this problem; see \cite{Osherovich}.
Despite the enormous effort that has gone into finding robust algorithms,
the state of the art is generally unsatisfactory and reconstructions
are typically not very accurate. That is to say, the
phase retrieval problem with a support constraint has all the hallmarks
of an ill-posed problem. Our
main purpose in this paper is to describe recent work aimed at understanding
what aspects of the phase retrieval problem render it ill-posed, and how this
knowledge can be used to modify the experimental protocols to obtain better
conditioned inverse problems.
\begin{remark}
The {\em phase retrieval problem}
sometimes refers to the more general setting where $\rho(\bx)$ is
unknown and measurements $M(k)$ are of the form
\[ M(k) = | \langle \rho, a_k \rangle |, \]
for some set of querying functions $a_k$. Here,
$\langle \rho, a \rangle$ denotes the inner product of the two
functions.
If the phase information
were available, then solving for $\rho$ would correspond
to a linear least squares problem.
Without the phase information, the problem is non-convex.
When the map from $\rho$ to $M(k)$ is invertible, a variety of
optimization methods have been developed based, for example, on
semidefinite relaxation or gradient descent
\cite{candes2015a,candes2015b}.
Unfortunately, the phase retrieval problem of interest in x-ray
scattering does not satisfy the necessary hypotheses for these methods to
apply, namely that the forward map is injective and the solution is unique.
The recent paper \cite{alaifari2017} contains a detailed analysis of phase retrieval
in the invertible case and an interesting discussion of stability in that context.
\end{remark}
\subsection{The Discrete Classical Phase Retrieval Problem}\label{s.dcpp}
For the sake of simplicity, we analyze a finite-dimensional analogue
of the phase retrieval problem described above, which, in the limit of
infinitely many samples, converges to the continuum problem.
We assume that $\rho$ is \emph{real valued}, and imagine that the
unknowns are the samples $f_{\bj}=\rho\left(\frac{\bj}{N}\right),$
where $\bj\in J$ are points in a finite cubical integer lattice,
$J=\{0,1,\dots,2N-1\}^d \subset \bbZ^d$.
Here $d$ is the ambient dimension ($d=2$ in our examples, but
3D phase retrieval is also possible \cite{chapman2006}).
The vector $\bF := (f_{\bj}:
\bj\in J)$ denotes an image, which can be viewed as a uniform
pixelization of a density function $\rho$ lying in $[0,2)^d$. We use
the notation $\bbR^J$ to denote the set of all possible such
images. Using $J$ as the index set is somewhat non-standard in the
engineering literature.
The measured data values are
modeled as $a_\bk := |\hat f_{\bk}|$,
where
\begin{equation}
\hat f_{\bk}=\sum_{\bj\in J}f_{\bj}\exp\left(\frac{2\pi i\bj\cdot\bk}{2N}\right)~,
\qquad \bk\in J~,
\label{dft}
\end{equation}
is the usual $d$-dimensional discrete Fourier transform (DFT)
taking the $(2N)^d$ pixel values to $(2N)^d$ frequency data.
We call $a_\bk$ \emph{magnitude DFT} data, and denote the data vector
by $\ba := (a_\bk: \bk\in J)$. We define the measurement map,
$\cM:\bbR^J\to\bbR_+^J,$ by setting
\begin{equation}
\cM(\bF):=(|\hf_{\bk}|:\:\bk\in J).
\end{equation}
\begin{remark}
One may connect the above discrete model to the continuous case as follows.
From \eqref{dft}, the indices $\bk$ are $2N$-periodic in each
dimension. Let $\tilde\bk$ be the periodic folding of $\bk$ into
the origin-centered cube $\{-N,-N+1,\dots,N-1\}^d$. Then define the
spatial frequencies $\xi_\bk := \pi \tilde{\bk}$ for $\bk\in J$,
which lie in the cube $[-\pi N,\pi N]^d$. The sum in \eqref{dft} can be
interpreted as approximate samples of $(2N)^d\hat\rho(\xi_\bk)$,
where the Fourier integral over $[0,2)^d$ has been approximated by a
$2N$-point trapezoid quadrature (in each dimension)
at the nodes $\bj/N$. Thus for
a continuous function $\rho$ the sequence $(2N)^{-d} \hat f_{\bk}$
tends to the exact Fourier transform as $N\to\infty$.
An alternative (but less physically realistic) interpretation is:
\eqref{dft} gives point samples of the {\em exact}
Fourier transform of a ``sum of point masses'' scatterer model
$\rho(\bx) := \sum_{\bj\in J} f_{\bj} \delta(\bx-\bj/N)$.
\end{remark}
The advantage of a discrete model over the continuous one is that it
admits, given a support condition to be presented shortly, an exact
solution that is generically unique up to trivial associates. In
contrast, for the continuous $\rho$ problem, given any finite
collection of samples of $|\hrho(\bk)|$, there is an infinite
dimensional space of functions with these Fourier coefficients, which
also satisfy the support constraint.
\begin{definition}
Given a magnitude DFT data vector $\ba = (a_\bk: \bk\in J)$,
the {\em magnitude torus}, denoted by $\TA_{\ba}$, is the
collection of images $\bF$ in $\bbR^J$ with this magnitude DFT data,
i.e.
\begin{equation}
\TA_{\ba}=\{\bF\in\bbR^J:\:|\hat f_\bk| = a_\bk \text{ for all }\bk\in J\}.
\end{equation}
\end{definition}
\begin{remark}
Note that $\TA_{\ba}$ is either empty (if $\ba$ does not obey the
inversion symmetry demanded by \eqref{dft} for a real image), or is a
real torus, or union of tori, of dimension equal to approximately half
of the cardinality of $J$. The reason for the approximate nature, and
the possible existence of multiple connected components (which are all
tori), is the fact that some data is forced to obey various
symmetries; for example $\hat f_{\mathbf{0}}$ is always real, whereas
most DFT data is generically complex. Vanishing DFT coefficients
also lower the dimension of the torus.
\end{remark}
The prior information about the image is encoded as a second set
$B\subset \bbR^J$. The \emph{discrete classical phase retrieval problem} is
then the problem of finding points in the intersection $\TA_{\ba}\cap B$;
see Fig.~\ref{tangentfig}. For an image $\bF$ we denote its
true support by
\begin{equation}
S_{\bF}:=\{\bj\in J:\:f_{\bj}\neq 0\}.
\label{Sf}
\end{equation}
If $S_{\bF} \subset S\subset J,$ then we say that
$S$ is an estimate for the support of $\bF$, and let
\begin{equation}
B_S:=\{\bF\in\bbR^J:\: f_{\bj}=0\text{ for }\bj\notin S\}.
\end{equation}
This is clearly a linear subspace of $\bbR^J$.
To define the operations in the finite, discrete case that generate the
set of trivial associates of an image $\bF$ we need to extend the
image to be $2N$-periodic. That is, the integer indices are defined mod
$2N$ in each dimension. With this understood, the image $f_{\bj}$ is
defined for all $\bj\in\bbN^d,$ with its restriction to $\bj\in J$
representing a single period. We call such images $J$-periodic. This
is consistent with the formula for the DFT, which defines $\hf_{\bk}$
for all $\bk\in\bbN^d,$ and the inverse formula, which defines
$f_{\bj}$ for all $\bj\in\bbN^d,$ as $J$-periodic images.
With this periodic extension, the
operations that generate the trivial associates are:
\begin{enumerate}
\item If $\bv\in J$, then the {\bf translate} of $\bF$ by $\bv$ is defined
by its components
\begin{equation}\label{eqn2}
f^{(\bv)}_{\bj} := f_{\bj-\bv}~, \qquad \bj\in J~.
\end{equation}
\item The {\bf inversion} $\check{f}$ is defined by its components
\begin{equation}\label{eqn3}
\check{f}_{\bj} := f_{-\bj}~, \qquad \bj\in J~.
\end{equation}
\end{enumerate}
\noindent
Each image in the set of all $2(2N)^d$ trivial
associates has the same magnitude DFT data as $\bF$.
We now state a well-known uniqueness theorem due to Hayes
\cite{barakat1984,hayes1982,hayes1987}, concerning the support
constraint. If $S$ is contained in a rectangular subset of $J$, with
side lengths at most half the corresponding side-lengths of $J$,
i.e.\ at most $N$, and $S_{\bF} \subset S$, then the set
$\TA_{\ba}\cap B_S$ is finite, and generically consists of trivial
associates of a single point in this set.
\begin{definition} \label{ssdef}
An image $\bF$ with support $S_{\bF}\subset S$, for a set $S$
as above, is said to have \emph{small support}.
\end{definition}
Note that small support corresponds to the density function
$\rho$ having support lying within a $d$-dimensional cube of side length 1.
In some experimental situations one has a constraint
$\bF\in B_+$, where
\begin{equation}
B_+:=\{\bF:\: f_{\bj}\geq 0,\text{ for all }\bj\in J\}.
\label{Bplus}
\end{equation}
This auxiliary condition alone does not uniquely specify a set of trivial
associates, or even a finite set. It is easy to show, however,
that the squared magnitude data
$(|\hat f_{\bk}|^2:\:\bk\in J)$ are the DFT coefficients of the
autocorrelation image
\begin{equation}
[\bF\star\bF]_{\bj}=\sum_{\bl\in J}f_{\bl}f_{\bj+\bl}.
\end{equation}
In~\cite{BEGM} we prove that if the support of $\bF\star\bF$ is
sufficiently small, then the set of non-negative images in $\TA_{\ba}$
is finite, and generically consists of trivial associates of a single
element. This follows because, for $\bF\in B_+$, if the support of
$\bF\star\bF$ is sufficiently small, as a subset of $J,$ then it
provides a non-trivial upper bound on the support of $\bF$, for which
we can conclude uniqueness, up to trivial associates. As all points
$\bF\in\TA_{\ba}$ have the same autocorrelation image, without a
non-negativity constraint, the support of the autocorrelation image
does \emph{not}, in general, provide a bound on the support of the
image itself. We refer to the problem of finding the
intersection $\TA_{\ba}\cap B_+$ as {\em phase retrieval with
non-negativity constraints}.
\begin{definition} \label{adequatedef}
We say that the set $B,$ defined by some auxiliary conditions,
is {\em adequate} if the intersection $\TA_{\ba}\cap
B$ is a finite set for any $\ba$ in the range of $\cM$.
(Thus, if the support of $\bF\star\bF$ is
sufficiently small, then non-negativity of the image is adequate
data for phase retrieval.)
\end{definition}
\subsection{Well-posedness of the Discrete Phase Retrieval Problem}
The concept of well-posedness for an inverse problem comprises two
distinct questions: the first is the uniqueness of the solution and
the second concerns the continuity properties of the local inverse
map near a solution. For the phase retrieval problem, uniqueness
should be understood as uniqueness up to trivial associates.
Neither aspect of well-posedness has been analyzed in detail for
the phase retrieval problem. From the proof of Hayes' uniqueness
theorem, it follows that there are pairs of images, $\bF_1, \bF_2$,
which are not trivial associates, with identical magnitude DFT data
and support for which $\|\bF_1-\bF_2\|\approx 1$. Yet, since uniqueness is
generic, we can find images, $\bF'_1, \bF'_2$, as near to $\bF_1,
\bF_2$ as we like for which the phase retrieval problem does have a
unique solution. Moreover we can assume that the supports of $\bF'_1,
\bF'_2$ are the same as those of $\bF_1, \bF_2$. From this
observation it is clear that, at any finite precision, this problem
does not always have a unique solution, even up to trivial associates. A
similar observation was made by Fienup and Seldin
in~\cite{FienupSeldin:90}.
A primary concern here is that of understanding
what makes it difficult to find points in $\TA_{\ba}\cap B,$ even if
it is assumed that the set $B$ is selected so that this intersection
consists of finitely many points. In coherent diffraction imaging,
the cardinality of the index
set $J$ is in the hundreds of thousands, millions, or even
billions (depending on whether a two-dimensional or three-dimensional
object is being imaged and at what resolution).
High dimensionality certainly complicates the problem at
hand, but it is not the root cause of its difficulty. Rather, it
is the geometry \emph{near to} points in $\TA_{\ba}\cap B$ that
renders this problem so difficult. This sort of local geometry is usually discussed in
terms of the relationship of the fibers of the tangent bundles to
$\TA_{\ba}$ and $B$ at points of intersection. For the remainder of
this discussion we focus on the support condition case $B=B_S$,
which is itself a linear subspace.
\begin{definition}
For a point $\bF\in\TA_{\ba},$ the {\em fiber of the tangent bundle}
to $\TA_{\ba}$ at $\bF$, denoted by $T_{\bF}\TA_{\ba}$, is the affine
subspace of $\bbR^J$ through $\bF$ that is the best linear
approximation to $\TA_{\ba}$ near to $\bF$.
\end{definition}
\begin{remark}
An embedded submanifold of a Euclidean space has a ``best''
approximating affine subspace if it is at least $\cC^1.$ A
magnitude torus is a product of round circles and is therefore a
real analytic subspace of $\bbR^J.$
\end{remark}
\begin{definition}
Let $\bF\in \TA_{\ba}\cap B_S$.
The intersection
of $\TA_{\ba}$ with $B_S$ is {\em transversal} at $\bF$ if
\begin{equation}
T_{\bF}\TA_{\ba}\cap B_S= \{\bF\},
\end{equation}
that is, the affine space
$T_{\bF}\TA_{\ba}$ intersects the linear subspace $B_S$ only at $\bF$.
\end{definition}
If the intersection $\bF$ is transversal, then the geometry of
$\TA_{\ba}\cup B_S$ near to $\bF$ is accurately modeled by a
neighborhood of $\bF$ in $T_{\bF}\TA_{\ba}\cup B_S.$ In this case,
the conditioning of the problem of finding a point $\bF\in
\TA_{\ba}\cap B_S$ is determined by the angles between
$T_{\bF}\TA_{\ba}$ and $B_S.$ If there are positive dimensional
subspaces of $T_{\bF}\TA_{\ba}$ and $B_S$ that make a very small
angle with one another, then the condition number, though finite, will be very large.
For high dimensional non-linear submanifolds of $\bbR^J$ one does not
generally expect to have an explicit description of the fibers of the
tangent bundle. It is a remarkable feature of magnitude tori that such
a description is accessible. Using this description we can show that
the intersections between $\TA_{\ba}$ and $B_S$ are {\em typically
not transversal}. If the intersection at $\bF$ is not transversal,
then $T_{\bF}\TA_{\ba}\cap B_S$ is a positive dimensional affine
subspace. In this case the local geometry of $T_{\bF}\TA_{\ba}\cup
B_S$ near to $\bF$ does not resemble that of $\TA_{\ba}\cup B_S;$
the linearized problem does not have a unique solution. Formally
speaking, the condition number of the non-linear problem is infinite.
Moreover, in this case, linear analysis fails to adequately describe
the behavior of algorithms for finding intersection points.
Above we defined $\cM:\bbR^J\to \bbR_+^J$ as the ``forward operator''
(in the language of inverse problems), i.e.\ the measurement map from
an image to its corresponding DFT magnitude data $\ba:=(|\hat
f_{\bk}|:\:\bk\in J)$. The inverse image $\cM^{-1}(\ba)$ of a point
$\ba\in\bbR_+^J$ is simply its magnitude torus $\TA_{\ba}$. Suppose
now that $\bF\in\TA_{\ba}$ has small support contained in the set
$S$. Then, by Hayes' theorem, the set $\cM^{-1}(\ba)\cap B_S$ is
finite, and non-empty. This remains true if $\bF$ is replaced by a
nearby point $\bF'\in B_S$. Therefore, the map
$\cM\restrictedto_{B_S}$ has a local inverse, defined on the manifold
of consistent data $\cM(B_S)$, near to $\cM(\bF)$. We denote this
local inverse by $\cM^{-1}_{\bF,S}.$ As noted above, the other issue
that arises in a discussion of well-posedness is the continuity of
this local inverse. In~\cite{BEGM} we prove the following theorem.
\begin{theorem}
The local inverse $\cM^{-1}_{\bF,S}$ satisfies a Lipschitz estimate if
and only if the intersection of $\TA_{\ba}$ with $B_S$ at $\bF$
is transversal.
\end{theorem}
By itself,
the failure to have a local Lipschitz inverse leads to a kind of ill-conditioning.
When the intersection is non-transversal, the local inverse is,
at best, H\"older continuous of order $\alpha <1$,
which implies an infinite condition number.
In fact the number of
accurate digits possible in the reconstructed image
cannot exceed $\alpha d$ when the data is available with a
relative precision of $d$ digits.
For phase retrieval, it is often the case that $\alpha\leq \frac 1 2$
(see Fig. \ref{tangentfig}(b) for an illustration with $\alpha=\frac 1 2$).
More critical, however, is that non-transversality
stalls the convergence
for standard reconstruction algorithms,
even in the ideal case of noise-free data,
as discussed in section \ref{sec4}.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{transversal.pdf}
\caption{ Illustration of two types of intersection in image space
between the magnitude torus (where $\ba = \cM(\bF)$) and the
constraint $B_s$. (a) Transversal case. The angle between the
tangent space and $B_S$ is positive. Also visible (left-most black
dot) is a trivial associate of the image $\bF$. Finding $\bF$ given
the data $\ba$ is (locally) well-conditioned. (b) Non-transversal
case. The angle between the fiber of the tangent bundle at $\bF$ and
$B_S$ is zero (since we sketch in $\mathbb{R}^3$ we are forced to
show $B_S$ lying within $T_{\bF}\TA_{\ba}$; in general this is only
true for a subspace of $B_S$). In the case shown, the distance from
the torus grows quadratically with distance from $\bF$ for points in
$B_S$. High-order contact of this type is much more problematic in
high dimensions, where the dimension of the fibers of the tangent bundle can be
large.}
\label{tangentfig}
\end{figure}
\subsection{Contents of the Paper}
An outline of the paper follows: in sections \ref{sec2} and
\ref{sec3}, we discuss the geometry of phase retrieval with support
constraints. We turn to the practical consequences of our analysis in
sections \ref{sec4} and \ref{sec5}, and extend the analysis to the
case of non-negativity constraints in section \ref{sec5.2}. In
section \ref{sec7}, we consider the possibility of alternate
experimental protocols that yield better conditioned inverse problems.
We draw heavily here on results from the text~\cite{BEGM}, which
contains, among other things, complete proofs of the main theorems
used (as well as more detailed numerical experiments).
\section{The Tangent Bundle to $\TA_{\ba}$} \label{sec2}
In this section we let $\bF$ denote an image with small support.
The question of transversality of the intersection at
$\bF\in\TA_{\ba}\cap B$ concerns the relationship between the fiber
of the tangent bundle to $\TA_{\ba}$ at $\bF$
and a linear approximation to the set $B$. For any $\bF\in\TA_{\ba}$
we let $T_{\bF}\TA_{\ba}$ denote the fiber of the tangent bundle to
$\TA_{\ba}$ at $\bF$. The fiber of normal bundle at $\bF,$
$N_{\bF}\TA_{\ba},$ is the affine subspace through $\bF$ orthogonal to
$T_{\bF}\TA_{\ba}.$ We have defined these fibers as affine subspaces
of the ambient space $\bbR^J,$ and let $T^0_{\bF}\TA_{\ba},
N^0_{\bF}\TA_{\ba},$ denote the linear subspaces of $\bbR^J$ so that
\begin{equation}
T_{\bF}\TA_{\ba}= \bF+T^0_{\bF}\TA_{\ba}\quad \text{ and }\quad
N_{\bF}\TA_{\ba}=\bF+ N^0_{\bF}\TA_{\ba}.
\end{equation}
Recall that for $B=B_S,$ where $S$ is an estimate for the support of $\bF$,
$B_S$ is a linear subspace and the intersection is transversal if and
only if $T_{\bF}\TA_{\ba}\cap B_S=\{\bF\}$.
\begin{remark}
If $B=B_+$ (see \eqref{Bplus}) then the
intersection lies on $\pa B_+,$ which is not a smooth submanifold of
$\bbR^J,$ but rather a stratified space. This renders the concept of
transversality more subtle to define. As the $\pa B_+$ is ``piecewise
linear'' in that it is locally a union of orthants in linear spaces of
various dimensions, it again makes sense to say that the intersection
is transversal provided that $T_{\bF}\TA_{\ba}\cap\pa
B_+=\{\bF\}.$ Since it is conceptually much simpler (and more general
in its applicability), most of our
discussion of transversality uses a support constraint as auxiliary
information. In Section~\ref{sec5.2} we briefly discuss the transversality of
the intersection with $\pa B_+.$
\end{remark}
\subsection{The Tangent Bundle in the DFT Representation}
The key to analyzing these intersections is to have an
explicit, readily computable description of the fibers of the tangent
bundle to $\TA_{\ba}.$ In this section we give two such descriptions.
The DFT \eqref{dft}, which we denote by $\cF$,
maps the torus $\TA_{\ba}$ onto a torus in $\bbC^J$
defined by
\begin{equation}
\begin{split}
\hTA_{\ba}&:= \cF \TA_{\ba} =
\{\hbf:\: |\hat f_{\bj}|=\ba_{\bj}, \; \text{ for }\bj\in J\}\\
&=\{(e^{i\theta_{\bj}}a_{\bj}:\:\bj\in J)\text{ for all }:\: \btheta\in \bbR^J\}.
\end{split}
\end{equation}
Taking $\theta_{\bj}$-derivatives
gives a very simple description of the tangent
bundle: for $\hbf\in\hTA_{\ba}$,
\begin{equation}
T^0_{\hbf}\hTA_{\ba}=\Span_{\bbR}\{i\frac{\hat f_{\bj}}{|\hat f_{\bj}|}\be^{\bj}:\bj\in J\},
\end{equation}
where the standard basis vector
$\be^{\bj}\in\bbR^J$ has a $1$ in the $\bj$th location and is
otherwise zero. It is the ``real-span'' because
$\hTA_{\ba}$ is a real submanifold of $\bbC^J.$
As the images, $\bF,$ we consider are real, this is reflected in a
symmetry of $\hbf$: for each index $\bj\in J$ there is a conjugate
index $\bj':=2(N-1)\bone-\bj,$ where $\bone=(1,\dots,1),$ for which
\begin{equation}
\hat f_{\bj'}=\overline{\hat f_{\bj}}~;
\end{equation}
note that $(\bj')'=\bj.$ In fact, the fiber of the tangent bundle is the span of a smaller
set of vectors:
\begin{equation}
T^0_{\hbf}\hTA_{\ba}=\Span_{\bbR}\left\{i\left[\frac{\hat f_{\bj}}{|\hat f_{\bj}|}\be^{\bj}-
\frac{\overline{\hat f_{\bj}}}{|\hat f_{\bj}|}\be^{\bj'}\right]:\bj\in J\right\}.
\end{equation}
The fiber of the normal bundle has a similar
description:
\begin{equation}
N^0_{\hbf}\hTA_{\ba}=\Span_{\bbR}\left\{\left[\frac{\hat f_{\bj}}{|\hat f_{\bj}|}\be^{\bj}+
\frac{\overline{\hat f_{\bj}}}{|\hat f_{\bj}|}\be^{\bj'}\right]:\bj\in J\right\}.
\end{equation}
Up to a scale factor, the DFT is a unitary map, and therefore
$T_{\bF}\TA_{\ba}=\cF^{-1}[T_{\hbf}\hTA_{\ba}],$ though this is not a
very explicit, or useful description.
\subsection{The Tangent Bundle in the Image Representation}
We now give a second description,
in the image domain, of bases for
the tangent and normal bundles to $\TA_{\ba},$ whose elements share
many properties with that of the image itself. Recall that, for
$\bv\in J,$ the translate, $\bF^{(\bv)},$ of $\bF$ by $\bv$ is defined
in~\eqref{eqn2}. Introduce the following images, which are the difference and sum of
translates of the image by $\bv$ and $-\bv,$
\begin{equation}
\btau^{\bv}:= \bF^{(\bv)}-\bF^{(-\bv)}~,
\qquad
\bnu^{\bv}:= \bF^{(\bv)}+\bF^{(-\bv)}~.
\end{equation}
\begin{theorem}\label{thm1}
Suppose that $\ba\in\bbR^J$ is the DFT magnitude data of a real
image, $\bF.$ For any point $\bF\in \TA_{\ba}$, we have that
\begin{equation}
T^0_{\bF}\TA_{\ba}=\Span_{\bbR}\{\btau^{\bv}:\, \bv\in J\}~,\qquad
N^0_{\bF}\TA_{\ba}=\Span_{\bbR}\{\bnu^{\bv}:\, \bv\in J\}.
\end{equation}
\end{theorem}
\begin{remark} The images $\{\bF^{(\bv)}\}$ are of course just trivial
associates of $\bF$, whose existence makes the solution of the phase
retrieval problem non-unique. In fact the distances between the
trivial associates are fairly large; an effective algorithm
defined by a map with strong contraction properties would not have
problems on this account. The theorem describes a far more insidious
effect of the existence of trivial associates,
as explained in the next paragraph:
{\em it often renders the
intersections of $\TA_{\ba}$ and $B_S$ non-transversal.}
As we shall see in
the next section, this adversely affects the continuity properties
of the inverse map, which is entirely algorithm-independent. In
Section~\ref{sec4} we see that it also vastly diminishes the contraction
properties of the maps used to define phase retrieval algorithms,
which inevitably leads to stagnation and even poorer reconstructions
than would be expected from the results of Section~\ref{sec3}.
\end{remark}
Given an image, $\bF$, there is a subset $J_t$ of $J$ so that
$\{\btau^{\bv}:\bv\in J_t\},$ is a basis for the vector space
$T^0_{\bF}\TA_{\ba}.$ If $S$ is a realistic estimate for the support
of $\bF$, then there is usually a non-empty subset $J_{it}\subset J_t$
such that the tangent vectors $\{\btau^{\bv}:\:\bv\in J_{it}\}$ also
have support in $S$. In this case
\begin{equation}
T^0_{\bF}\TA_{\ba}\cap B_S \;\supset\; \Span_{\bbR}\{\btau^{\bv}:\:\bv\in J_{it}\},
\end{equation}
which implies that the intersection at $\bF$ is \emph{not}
transversal.
For a subset $W\subset J$ the $p$-pixel neighborhood, $W_p,$ of $W$ is
defined to be
\begin{equation}\label{eqn2.10.002}
W_p\overset{d}{=}\{\bj\in J:\:
\exists \bk\in W\text{ with } \|\bk-\bj\|_{\infty}\leq p\}.
\end{equation}
Once again this distance should be understood in the $J$-periodic sense.
A simple combinatorial argument shows that
\begin{equation}\label{eqn2.11.002}
\dim T_{\bF}\TA_{\ba}\cap B_{S_{\bF,p}}\geq 2p(p+1),
\end{equation}
showing that a looser support constraint leads to a greater failure of
transversality.
\subsection{The Convolution Property of $T_{\bF}\TA_{\ba}$}\label{sec2.3.004}
In this section we examine a surprising property of the tangent bundle
to magnitude torus $\TA_{\ba}$ defined by a image $\bh$ that is a
convolution of two other images, that is
\begin{equation}
\bh=\bF\ast \bg.
\end{equation}
In this section it is important to recall that we regard images in
$\bbR^{J}$ as periodic, i.e. as elements of $\bbR^{\bbN^d}$ with
indices in $J\subset\bbN^d$ representing a single period.
This discussion requires some additional notation. For
$\bF\in\bbR^J,$ we let
\begin{equation}
\ba_{\bF}\overset{d}{=}(|\hf_{\bk}|:\:\bk\in J),
\end{equation}
so that $\TA_{\ba_{\bF}}$ is the magnitude torus defined by $\bF,$ and we
let
\begin{equation}
\btau^{(\bv)}_{\bF}\overset{d}{=}\bF^{(\bv)}-\bF^{(-\bv)}.
\end{equation}
Suppose that $J_t\subset J$ is chosen so that $\{\btau^{\bv}_{\bF}:\: \bv\in
J_t\}$ is a basis for $T_{\bF}\TA_{\ba_{\bF}}.$ If $\balpha\in\bbR^{J_t},$
then we let
\begin{equation}
\btau^{\balpha}_{\bF}\overset{d}{=}\sum_{\bv\in
J_t}\alpha_{\bv}\btau^{(\bv)}_{\bF}\in T_{\bF}\TA_{\ba_{\bF}}.
\end{equation}
Recall that (periodic) discrete convolution is defined by
\begin{equation}
[\bF\ast\bg]_{\bj}=\sum_{\bk\in J} f_{\bj-\bk}g_{\bk}.
\end{equation}
The support of a convolution satisfies
\begin{equation}\label{eqn2.15.002}
S_{\bF\ast\bg}\subset S_{\bF}+S_{\bg},
\end{equation}
where we recall that if $X,Y\subset J,$ then $X+Y\overset{d}{=}\{\bj+\bk:\:\bj\in
X,\, \bk\in Y\}\mod J.$ The DFT coefficients of a convolution satisfy
\begin{equation}\label{2.16.002}
\widehat{\bF\ast\bg}_{\bj}=\hf_{\bj}\hg_{\bj}.
\end{equation}
It is an elementary computation to show that
\begin{equation}\label{eqn2.15.01}
\btau^{(\bv)}_{\bF\ast\bg}=\btau^{(\bv)}_{\bF}\ast\bg=\bF\ast\btau^{(\bv)}_{\bg}.
\end{equation}
For generic images, $\bF$ and $\bg,$ a single index set $J_t$ can be
used so that $\{\btau^{\bv}_{\bF}:\: \bv\in J_t\}$ is a basis for
$T_{\bF}\TA_{\ba_{\bF}},$ and $\{\btau^{\bv}_{\bg}:\: \bv\in J_t\}$ is a
basis for $T_{\bg}\TA_{\ba_{\bg}}.$ Letting $\balpha\in \bbR^{J_t}$, it
follows from~\eqref{eqn2.15.01} that
\begin{equation}\label{eqn2.20.003}
\btau^{\balpha}_{\bF\ast \bg}=\btau^{\balpha}_{\bF}\ast\bg=\bF\ast\btau^{\balpha}_{\bg}.
\end{equation}
More succinctly we can write:
\begin{equation}
T_{\bF\ast\bg}\TA_{\ba_{\bF\ast\bg}}=T_{\bF}\TA_{\ba_{\bF}}\ast\bg=
\bF\ast T_{\bg}\TA_{\ba_{\bg}}.
\end{equation}
If we let
\begin{equation}
S_{\btau^{\balpha}_{\bF}}=\{\bj\in J:\:\btau^{\balpha}_{\bF,\bj}\neq 0\},
\end{equation}
then~\eqref{eqn2.15.002} and~\eqref{eqn2.20.003} imply that
\begin{equation}
S_{\btau^{\balpha}_{\bF\ast \bg}}\subset S_{\btau^{\balpha}_{\bF}}+S_{\bg}.
\end{equation}
Suppose that $\bg$ is an image for which there exists an $\balpha\in J_t$ so that
\begin{equation}\label{eqn2.21.002}
S_{\btau_{\bg}^{\balpha}}\subset S_{\bg}.
\end{equation}
The vector field $\bF\ast\btau_{\bg}^{\balpha}\in
T_{\bF\ast\bg}\TA_{\ba_{\bF\ast\bg}}$ then satisfies
\begin{equation}\label{eqn2.22.002}
S_{\bF\ast\btau_{\bg}^{\balpha}}\subset S_{\bg}+S_{\bF}.
\end{equation}
On its face the condition in~\eqref{eqn2.21.002} seems very unlikely to hold for
any $\balpha,$ as it is equivalent to the system of
linear equations for $\balpha\in J_t$:
\begin{equation}\label{eqn2.23.002}
\btau_{\bg,\bj}^{\balpha}=0\text{ for }\bj\in S^c_{\bg}.
\end{equation}
Since $\dim T_{\bg}\TA_{\ba_{\bg}}=|J_t|\approx|J|/2,$ and $|S_{\bg}^c|= |J|-|S_g|\geq
\frac{3|J|}{4},$ for an image with small support, the equations
in~\eqref{eqn2.23.002} appear to be overdetermined.
It turns out that if $\bg$ is inversion symmetric, that is
\begin{equation}
\bg_{\bj}=\bg_{-\bj},
\end{equation}
then the basis vectors for $T_{\bg}\TA_{\ba_{\bg}}$ satisfy the
equations
\begin{equation}
\btau^{(\bv)}_{\bg,\bj}=- \btau^{(\bv)}_{\bg,-\bj}\text{ for all
}\bv\in J_t,\,\bj\in J.
\end{equation}
For the case of an inversion symmetric image, half of the equations
in~\eqref{eqn2.23.002} imply the other half, from which the following
theorem follows easily.
\begin{theorem}\label{thm2}
Let $\bg\in\bbR^J$ be inversion symmetric, then the solution space
to the equations in~\eqref{eqn2.23.002} has dimension at least
\begin{equation}
|J_t|- \frac{|J|-|S_{\bg}|}{2}\approx \frac{|S_{\bg}|}{2}.
\end{equation}
\end{theorem}
\begin{remark} This theorem also holds for images that
are anti-symmetric, i.e. $\bg_{\bj}=-\bg_{-\bj},$ and images that are inversion
symmetric (or anti-symmetric) with respect to any point in $J.$
\end{remark}
An inversion symmetric image has real DFT coefficients, and therefore
that phase retrieval problem for such an image reduces to the much
easier sign retrieval problem. But suppose that $\bF$ is an arbitrary
image with small support, and $\bg$ is inversion symmetric
so that $\bF\ast\bg$ also has small support. The theorem along
with~\eqref{eqn2.22.002} imply that
\begin{equation}
\dim T_{\bF\ast\bg}\TA_{\ba_{\bF\ast\bg}}\cap
B_{S_{\bF}+S_{\bg}}\geq \frac{|S_{\bg}|}{2}.
\end{equation}
For non-negative images $S_{\bF\ast\bg}=S_{\bF}+S_{\bg};$ in any case, the sum
is a reasonable estimate for $S_{\bF\ast\bg}.$ The failure of
transversality, with $S=S_{\bF\ast\bg}$ is \emph{inherited} by
$\bF\ast\bg$ from $\bg,$ even though $\bF\ast\bg$ has no obvious symmetries.
\subsection{Examples of the Failure of Transversality for Convolutions}\label{sec2.4.004}
The analysis in the previous section shows that for images that are
convolutions with inversion symmetric images, the failure of
transversality occurs even if we use the exact support of the image to
define the support constraint. In order to diminish the effects of
noise, it is a very common practice to multiply measured data by a
smooth cut-off function, such as a Gaussian. Since our measurements are
in the DFT domain, equation~\eqref{2.16.002} shows that this is
equivalent to convolving the unknown image with a Gaussian. Our
analysis suggests that this makes the problem of recovering the phase
much more difficult. In this section we present the results of
numerical experiments that demonstrate this phenomenon.
For our numerical experiments, we use images defined by a sum
of radial functions,
\begin{equation}
\rho(\bx)=\sum_{i=1}^IR_i(\bx)~.
\end{equation}
In the simplest case, each function $R_i$ is a scaled characteristic function
of a disc,
\begin{equation}
R^0_i(\bx)=\alpha_i\chi_{[0,r_i]}(\|\bx-\bcc_i\|)~,
\end{equation}
where the intensities $\alpha_i>0$, radii $r_i>0$, and centers $\bcc_i$
are set randomly; see left side of Fig.~\ref{fig3}(a).
In this case the function $\rho$ is piecewise constant; in the imaging
literature one would say that $\rho$ represents a \emph{hard} object.
The discrete image $\bF$ is then generated from point samples of $\rho$
on a regular grid.
We also generate smoother images by
(discrete) convolution of this $\bF$ with the discretely sampled Gaussian,
\begin{equation}
G_k(\bx) = c_k e^{-\frac{16N^2\|\bx\|^2}{(k+1)^2}}~, \qquad k>0,
\label{smoother}
\end{equation}
where $k$ controls the smoothness,
and where $c_k$ is chosen to make $\|G_k\|_1=1$.
(The unsmoothed case we denote by $k=0$.)
In Figure~\ref{fig3} we show examples of such images with
smoothness levels $0$ (unsmoothed), $4$, and $8$.
\begin{figure}[h]
\begin{centering}
\centerline{\includegraphics[width=13cm]{phantoms_smthness3.pdf}}
\caption{Images similar to those used in numerical experiments
below of various smoothness levels: $k=0, 4, 8.$ The parameter $k$ scales
the width of the Gaussian in \eqref{smoother} used to smooth the
image. These are $256\times 256$ images ($N=128$), with the object
contained in a $128\times 128$ square.}\label{fig3}
\end{centering}
\end{figure}
Since the smoothing in \eqref{smoother} results in full support,
we instead set a threshold $\epsilon\approx\epsmach$
appropriate for finite-precision arithmetic,
and define the support to be
\begin{equation}
S_{\bF}\overset{d}{=}\{\bj\in J:\:|f_{\bj}|\geq \epsilon\}~.
\end{equation}
The $p$-pixel neighborhoods, $S_{\bF,p}$ of $S_{\bF}$ are defined as
in~\eqref{eqn2.10.002}. To see how $\dim T_{\bF}\TA_{\ba}\cap
B_{S_{\bF,p}}$ depends on the degree of smoothing, $k$, and the size
of the support padding, $p$, we generate 5 random samples,
$\{\bF_i^{k}:\:i=1,\dots,5\},$ for each smoothness level
$k=0,1,2,3,4.$
Since the intersection dimension computation requires a dense SVD
of a matrix of size $\bigO(N^2\times N^2)$, the study is
limited to small images; we choose the image size so that the
double oversampled image is $64\times 64$. For these images
$ \dim T_{\bF_i^k}\TA_{\ba}=2046\text{ if }k=0$ and $1984\text{ if }k>0.$
The dimension decreases when $k>0$ as the symmetries of $G_k$
forces certain DFT coefficients to vanish.
For each sample image, we numerically compute $U$, an orthonormal
basis for $T^0_{\bF_i^{k}}\TA_{\ba}$, and $V$, an orthonormal basis
for $B_{S_{\bF_i^k,p}}$, and then compute the SVD of $H=V^tU$. In
exact arithmetic, $\dim T_{\bF_i^{k}}\TA_{\ba}\cap B_{S_{\bF_i^k,p}}$
is the number of singular value, $\{\sigma_n(H)\},$ equal to 1. Since we
work in finite-precision arithmetic, such singular values are only
approximately 1. Figure~\ref{fig4} contains plots of
$\log_{10}(1-\sigma_n).$ Note that there is a sharp transition from
singular values within $10^{-15}$ of 1 to smaller ones, which is
remarkably consistent across the samples. As the support of the
Gaussian, $G_k,$ (at machine precision) grows with $k,$ this is
essentially as predicted by Theorem~\ref{thm2}.
We summarize these dimension measurements in Table~\ref{tab1}. The
dimensions shown correspond to the number of singular values greater
than $1-10^{-15}$. The $k=0$ row is precisely $2p(p+1)$ as predicted
in~\eqref{eqn2.11.002}. The table also has a $p=0$ column where we
have used the exact support, $S_{\bF_i^{k}},$ to define the support
constraint. As predicted by Theorem~\ref{thm2}, for $k>0$ the
dimensions of these intersections are non-zero. As $k$ increases, the
support of $G_k$ grows, and this produces a larger and larger
intersection between $T_{\bF_i^{k}}\TA_{\ba}$ and $B_{S_{\bF_i^k}}.$
As $p$ increases the dimension of these intersections grow beyond
$|S_{G_k}|/2:$ the larger $p$ is, the more ways there are to obtain
tangent vectors of the form $\btau_{G_k}^{\balpha}\ast\bF^0$ so that
$S_{\btau_{G_k}^{\balpha}}+S_{\bF^k_i}\subset S_{\bF^k_i,p}.$ In
examples below we show that even a very low dimensional failure of
transversality can lead to stagnation in standard reconstruction
algorithms.
\begin{figure}[h]
\centering
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{sing_val_tspintrsct_p1.pdf}
\caption{1-pixel neighborhoods}
\end{subfigure}\qquad
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{sing_val_tspintrsct_p2.pdf}
\caption{2-pixel neighborhoods}
\end{subfigure}
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{sing_val_tspintrsct_p3.pdf}
\caption{3-pixel neighborhoods}
\end{subfigure}\qquad
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{sing_val_tspintrsct_p4.pdf}
\caption{4-pixel neighborhoods}
\end{subfigure}
\caption{Singular values of $V^tU$ from images with varying degrees of smoothness
$k=0,1,2,3,4$ (corresponding to the colors: red, green, blue,
cyan, magenta). Plots produced by 5 random $64\times
64$-examples are shown for $p=1,2,3,4.$ The plots show
$\log_{10}(1-\sigma_n),$ where $\{\sigma_n\}$ are the singular
values of $H.$ }\label{fig4}
\end{figure}
\begin{table}[h]
\begin{center}
\begin{tabular}{| c || c | c | c |c|c|}
\hline
smthns\textbackslash supp &$p=0$& $p=1$ & $p=2$ & $p=3$& $p=4$ \\ \hline\hline
$k=0$&0 &4 &12 &24 &40 \\ \hline
$k=1$&18 &34 &54 &78 &106 \\ \hline
$k=2$&38 &64 &92 &124 &160 \\ \hline
$k=3$&61 &88 &120 &156 &196 \\ \hline
$k=4$&85 &119 &155 &195 &239 \\ \hline
\end{tabular}
\end{center}
\caption{Table showing the typical dimensions of
$T_{\bF^k_i}{\TA_{\ba}}\cap {B_{S_{\bF^k_i,p}}}$ for $p=1,2,3,4$ and
varying degrees of smoothness. }\label{tab1}
\end{table}
This experiment can be repeated using images defined as samples, $\bF^k,$ of
functions of the form
\begin{equation}\label{eqn2.35.004}
\rho(\bx)=\sum_{i=1}^IR^k_i(\bx),
\end{equation}
where
\begin{equation}\label{eqn2.36.004}
R^k_i(\bx)=\alpha_i\chi_{[0,r_i]}(\|\bx-\bcc_i\|)\|\bx-\bcc_i\|^k.
\end{equation}
These functions increase in smoothness with $k,$ but are not defined
as convolutions. For each $k\in\{0,1,2,3,4\}$ we produce 5 random
choices of $\rho.$ In this case we find that the dimensions of the
intersections are independent of $k,$ satisfying $ \dim
T_{\bF^k_i}\TA_{\ba_{\bF^k_i}}\cap B_{S_{\bF^k_i,p}}=2p(p+1);$ in this
case $ \dim T_{\bF^k_i}\TA_{\ba_{\bF^k_i}}=2046$ for all $k.$ In
Figure~\ref{fig4.1} we show plots of $(\log(1-\sigma_n):\:
n=1,\dots,200),$ for each choice $\bF^k_i,\, i=1,\dots,5,$ with
$p=1,3.$ From these plots we see that $\dim
T_{\bF^k_i}\TA_{\ba_{\bF^k_i}}\cap B_{S_{\bF^k_i,p}}$ does not depend
on $k,$ or the choice of example, and that the number of directions in
which $T_{\bF^k_i}\TA_{\ba_{\bF^k_i}}$ meets $B_{S_{\bF^k_i,p}}$ at
very small angles does increase with $k.$ These small angles also
dramatically stall the convergence of standard algorithms for finding
these intersections.
\begin{figure}[h]
\centering
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{sing_val_tspintrsct_p1_0.pdf}
\caption{1-pixel neighborhoods}
\end{subfigure}\qquad
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{sing_val_tspintrsct_p3_0.pdf}
\caption{3-pixel neighborhoods}
\end{subfigure}
\caption{The first 200 singular values of $H=V^tU$ from images
with varying degrees of smoothness $k=0,1,2,3,4$ (corresponding
to the colors: red, green, blue, cyan, magenta). Plots produced
by 5 random $64\times 64$-examples are shown for $p=1,3.$ The
plots show $(\log_{10}(1-\sigma_n):\: n=1,\dots,200),$ where
$\{\sigma_n\}$ are the singular values of $H.$ }\label{fig4.1}
\end{figure}
\section{Transversality, Well-Posedness and Microlocal Non-uniqueness}\label{sec3}
We turn now to an analysis of the effects of a non-transversal
intersection on the computational difficulty of the phase retrieval
problem. While the results in this section are algorithm independent,
they have direct implications about the loss of solution accuracy
given finite precision data and computations.
The results in this section are related to and, in part,
inspired by those in~\cite{cahill2016}
and~\cite{alaifari2017}.
\subsection{Transversality and Well-Posedness}
Recall from the introduction
that $\cM:\bbR^J\to \bbR_+^J,$ denotes the measurement map
$\cM(\bF)=(|\hat f_{\bj}|:\:\bj\in J)$.
First note that this map is Lipschitz continuous in the $2$-norm,
\begin{equation}
\|\cM(\bF)-\cM(\bg)\|_2 \;\le\; C_{\cM}\|\bF-\bg\|_2~,
\label{lip}
\end{equation}
with $C_{\cM} = \sqrt{|J|} = (2N)^{d/2}$,
which follows from the Plancherel theorem for the DFT \eqref{dft}
and the triangle inequality.
The inverse image of a point
$\ba\in\bbR_+^J$ is the magnitude torus $\TA_{\ba}$.
Suppose that
$\bF\in\TA_{\ba}$ has small support contained in the set $S$, then
the set $\cM^{-1}(\ba)\cap B_S$ is finite, and non-empty,
and the local inverse $\cM^{-1}_{\bF,S}$ is defined on a neighborhood
of $\ba=\cM(\bF)$ in the set of consistent data $\cM(B_S)$.
As is typical in the field of inverse problems, in order for the
problem to be locally well posed at $\bF$ it is necessary for this local
inverse to be a Lipschitz map. That is, there must be a neighborhood
$U\subset B_S$ of $\bzero$ and a constant $C>0$, such that for
$\delta\bF\in U$, we have the estimate
\begin{equation}\label{eqn19}
C\|\delta\bF\|_2\leq \|\cM(\bF)-\cM(\bF+\delta\bF)\|_2,
\end{equation}
so that if $\ba=\cM(\bF),$ and $\ba+\delta\ba=\cM(\bF+\delta\bF),$ then
\begin{equation}
\|\cM^{-1}_{\bF,S}(\ba)-\cM^{-1}_{\bF,S}(\ba+\delta\ba)\|_2\leq
\frac{1}{C}\|\delta\ba\|_2.
\end{equation}
Now assume that $\TA_{\ba}\cap B_S$ is non-transversal at $\bF,$ and
let $\btau\in T_{\bF}\TA_{\ba}\cap B_S$ be a unit vector, then the
definition of the tangent bundle, and \eqref{lip}, imply that there
is a constant $c$, dependent on $\bF$, so that
\begin{equation}\label{eqn21}
\|\cM(\bF)-\cM(\bF+t\btau)\|\leq ct^2~,
\qquad \mbox{ for all sufficiently small real $t$~.}
\end{equation}
In this case the best general bound one can hope for is that
\begin{equation}
\|\cM^{-1}_{\bF,S}(\ba)-\cM^{-1}_{\bF,S}(\ba+\delta\ba)\|_2 \; \leq \;
\frac{1}{C}\sqrt{\|\delta\ba\|_2}~.
\end{equation}
This implies that the local inverse is, at best, H\"older
continuous of order $\frac 12$, and therefore has an unbounded condition number. As noted
in the introduction, it also implies that if the measurements have $d$
significant digits, then, generally, it will be impossible to
reconstruct an image with more than $\frac{d}{2}$ significant digits.
In~\cite{BEGM}, we prove the following result:
\begin{theorem}\label{transversality_thm}
For an image $\bF\in B_S$, let $\TA_{\ba}$ denote the magnitude torus
defined by $\ba=\cM(\bF)$. Suppose that $a_{\bj}\neq 0$ for all
$\bj\in J$. There are positive constants, $\eta$, $C$, so that, if
$\delta\bF\in B_S$ and $\|\delta\bF\|_2<\eta$, then
\begin{equation}
C\|\delta\bF\|_2 \;\leq\; \|\cM(\bF)-\cM(\bF+\delta\bF)\|_2
\end{equation}
if and only if $T_{\bF}\TA_{\ba}\cap B_S=\{\bF\}$.
\end{theorem}
The theorem says that, if a support condition is the auxiliary
information that is available, and the DFT data is generic
(non-vanishing), then the phase retrieval problem can only be
well-conditioned near to $\bF\in \TA_{\ba}\cap B_S$ if this
intersection is transversal. This statement is intrinsic to the phase
retrieval problem, i.e.\ is algorithm independent. The results in
section~\ref{sec4} indicate that, with a realistic support condition,
these intersections are very rarely transversal (see Table~\ref{tab1}
and Figure~\ref{fig4.1} above). As our numerical experiments below
show, this failure of transversality can also dramatically harm the
convergence properties of standard algorithms.
\subsection{$\epsilon$-Non-Uniqueness} \label{nonuniquesec}
As discussed in the introduction, the solution to the phase retrieval problem
with support condition is not always unique up to trivial
associates. From the discussion in the previous section we already know that
the conditioning of the phase retrieval problem depends subtly on the unknown
image, and the precise nature of the auxiliary information. In this section we
explore various ways in which this problem can fail to have a unique solution
to a given precision $\epsilon>0.$ Suppose that there are two images
$\bF_1,\bF_2,$ and a subset $S\subset J,$ adequate for generic uniqueness, such
that
\begin{enumerate}
\item The norms $\|\bF_1\|_{2}=\|\bF_2\|_2,$ but the minimum distance between trivial
associates of $\bF_1$ and $\bF_2$ is much larger than $\epsilon\|\bF_1\|_{2}.$
\item The sets $\{\bj:\epsilon<|f_{i\bj}|\}\subset S,$ for $i=1,2.$
\item $\|\cM(\bF_1)-\cM(\bF_2)\|_{2}<\epsilon.$
\end{enumerate}
then we say that the solution to the phase retrieval problem defined by the
data $(\cM(\bF_1),S)$ is $\epsilon$-non-unique. In the remainder of this
section we describe two distinct mechanisms
that lead to $\epsilon$-non-uniqueness.
\subsubsection{Consequences of Genuine Non-Uniqueness}
The fact that a discrete image, with sufficiently small support, is
generically determined by the magnitude DFT data is a consequence of
the classical theorem that polynomials in two or more variables are
generically irreducible over the complex numbers. If $(f_{\bj}:\bj\in
J)$ is the image, then its $\bZ$-transform is
\begin{equation}
\BF(\bZ)=\sum_{\bj\in J}f_{\bj}\bZ^{-\bj},
\end{equation}
where $\bZ^{-\bj}=z_1^{-j_1}\cdots z_d^{-j_d}.$ There is a minimal integer
vector $\bm$ so that $\bZ^{\bm}\BF(\bZ)$ is a polynomial.
Suppose that $\bF$ is an image whose $\bZ$-transform, $\BF(\bZ)$ is reducible,
in the sense that there are polynomials, $\BF_1,\BF_2$ in $\bZ$ such that
\begin{equation}
\BF(\bZ)=\bZ^{\bn}\BF_1(\bZ)\BF_2(\bZ),
\end{equation}
for some integer vector $\bn.$ If $\bF_1$ and $\bF_2$ are images with
$\bZ$-transforms $\BF_1,$ $\BF_2,$ (up to a factor of $\bZ^{\bm_i}$ for some
$\bm_i$) then, up to a translation, $\bF=\bF_1\ast\bF_2,$ where $\ast$ denotes
discrete convolution. If no trivial associate of either $\bF_1$ or $\bF_2$ is
inversion symmetric, then the image $\bF'=\bF_1\ast\check{\bF}_2$ is not
a trivial associate of $\bF$ and, typically, the minimum distance between the
trivial associates of $\bF$ and $\bF'$ is large. If $\bF_1$ and $\bF_2$ are
non-negative, then so are $\bF$ and $\bF',$ and the smallest rectangles
containing each image coincide.
Suppose that $\bF$ and $\bF'$ both have small support contained in a
set $S,$ which is small enough to generically imply uniqueness, up to
trivial associates, and let $0<\epsilon$ be chosen with $\epsilon\ll
\|\bF-\bF'\|_2.$ Because uniqueness is generic we can modify these two
images to obtain generic images $\bg$ and $\bg',$ so that the norms
satisfy the estimates $\|\bg-\bF\|_2<\epsilon/2$,
$\|\bg'-\bF'\|_2<\epsilon/2$, and $S_{\bg}=S_{\bF}$,
$S_{\bg'}=S_{\bF'}$. The data $(\cM(\bg),S)$ defines both a phase
retrieval problem with a unique solution, up to trivial associates,
and an $\epsilon$-non-unique problem. That is, we can construct images
$\bg$ and $\bg',$ with support in $S,$ and nearly
identical magnitude DFT data:
$$\|\cM(\bg)-\cM(\bg')\|\leq \epsilon,$$
but satisfying
$$\|\bg-\bg'\|_2\geq \|\bF-\bF'\|_2-\epsilon>\!\!>0.$$
\begin{figure}[h]
\begin{centering}
\begin{subfigure}[t]{.8\textwidth}
\centerline{\includegraphics[width=11cm]{nonuniq_imgs.pdf}}
\caption{A pair of images with essentially identical magnitude-DFT data
that are not trivial associates, and their common support. For
clarity the left and center images show the central $256\times 256$
portion of the original $512\times 512$ image.}
\end{subfigure}\qquad
\begin{subfigure}[t]{.8\textwidth}
\centerline{\includegraphics[width=11cm]{nonuniq_ft_difs.pdf}}
\caption{$\log_{10}$-Magnitude-DFT data for the images above, and the
difference between the images themselves.
The image on the right shows the central $256\times 256$ portion of the original $512\times 512$ image.}
\end{subfigure}
\caption{ An illustration of true non-uniqueness in the phase retrieval
problem.}\label{fig2.1}
\end{centering}
\end{figure}
We conclude this section with an example of a pair of non-negative images,
$\bF,\bF'$, with exactly the same support and magnitude-DFT data such that
$\|\bF-\bF'\|_{2}\approx .48\|\bF\|_2.$ The minimum distance between trivial
associates is about $.18\|\bF\|_2,$ but the closest trivial associates have
rather different supports. These images are obtained as described above with
$\bF_1$ and $\bF_2$ non-negative images whose supports are inversion symmetric,
but the images themselves are not. The left and middle images in
Figure~\ref{fig2.1}(a) show the central $256\times 256$ portion of $\bF$ and $\bF',$
and the right image shows their common support. The left image in
Figure~\ref{fig2.1}(b) is the $\log_{10}$-magnitude-DFT of both images,
the right image is the central $256\times 256$ portion of the difference of the
two images.
What is striking about this example is how perfectly ordinary the
images and their magnitude-DFT data look. The only criterion that we
know of (in the continuum model) to exclude this phenomenon is that it
cannot occur in an image with jump discontinuities, because the
convolution of two bounded measurable functions is continuous. For the
discrete model it is difficult to make this statement precise, as
there are images $\bF=\bF_1\ast \bF_2,$ where, say, $\bF_1$ is a ``sum
of $\delta$-functions,'' which provide counterexamples. Note that,
for discrete images, a sum of $\delta$-functions is modeled by an
image with support in a set of isolated pixels. In fact such examples
can be found in~\cite{FienupSeldin:90}. A complete (asymptotic)
analysis of this problem might require an analysis, as $q$ tends to
infinity, of the density of the subset of reducible polynomials of
degree $q$ within the set of all polynomials of this degree. For
results in this direction see~\cite{KaltofenMay2003}
\subsubsection{Microlocal Non-Uniqueness}\label{sec3.2}
There is a second mechanism that leads to $\epsilon$-non-uniqueness,
which we call \emph{microlocal non-uniqueness}. We now explain the
mechanism underlying this phenomenon. For the construction, let $S$
be a set so that images supported in $S$ have small support. Let
$\bF\in\TA_{\ba}$ be an image that can be decomposed as a sum,
\begin{equation}
\bF=\bF_1+\cdots+\bF_k,
\end{equation}
where the components have the following properties:
\begin{enumerate}
\item For each $1\leq l\leq k$ we have $\{\bj:\: \epsilon<|f_{l\bj}|\}\subset S.$
\item Each pair $1\leq l\neq m\leq k$ has distinct spectral $\epsilon$-support,
\begin{equation}
\{\bj:\: \epsilon<|\hat f_{l\bj}|\}\cap \{\bj:\: \epsilon<|\hat f_{m\bj}|\}=\emptyset.
\end{equation}
\end{enumerate}
From the second condition, it follows that, for $\bV=\{\bv_l\in J:1\leq
l\leq m\},$ and $\Beta=(\beta_1,\dots,\beta_k),$ a binary string, the set of images
\begin{equation}\label{eqn26}
\bF^{\bV,\Beta}=\sum_{l=1}^k(-1)^{\beta_l}\bF_l^{(\bv_l)},
\end{equation}
all have the same DFT magnitude data to precision $k\epsilon.$ Indeed,
we are also free to replace some of the $\bF_l$ with their
inversions $\cbf_l$. For a realistic estimate $S$ of the support,
images $\bF^{\bV,\Beta}$ defined by a collection of small
translations $\bV$ also have their supports within $S$ to precision
$k\epsilon$. In this way a large collection of images, which are not
trivial associates, can be constructed that belong to the intersection
$\TA_{\ba}\cap B_S$ up to a fixed, very small, error. The data for any one of
these images is $m\epsilon$-non-unique, for some fixed $m$.
We close this section with an example of a pair images, whose
difference is $O(1)$, but with identical support, and magnitude DFT
data to precision $\epsilon=10^{-12}.$ In this example, which is
shown in Figure~\ref{fig2}, $k=4.$
\begin{example}\label{illposed}
In $d=2$, let
the four images $\bF_1$ through $\bF_4$ be defined by their components
\begin{equation}\label{eqn28}
f_{i\bj}=e^{-\sigma_i^2|\bj-\bl_i|^2}\cos(\langle\bk_i,\bj-\bl_i\rangle),\quad i=1,2,3,4,
\end{equation}
where
\begin{equation}
\bk_1=\bzero,\quad \bk_2=(70, 60),\quad \bk_3=(-60,70),\quad \bk_4=(200,200),
\end{equation}
and $N=512$, so $J=\{0,1,\dots,1023\}^2$.
Then we construct
\begin{align}
\begin{split}
\bF_a &= \bF_{1}+\bF_{2}+\bF_{3}+\bF_{4}~,\\
\bF_b &= \bF_{1}+\bF^{(\bv_2)}_{2}+\bF^{(\bv_3)}_{3}+\bF^{(\bv_4)}_{4},
\label{illposedimgs}
\end{split}
\end{align}
where the translation vectors are given by
$$\bv_2 = (-8,0), \quad \bv_3 = (0,-8), \quad \bv_4 = (8,8).$$
Figure~\ref{fig2}(a,b) shows a plot of $\bF_a$ and $\bF_b$.
The support sets are defined as
$S=\{\bj:|f_{1\bj}+f^{(\bv_2)}_{2\bj}+f^{(\bv_3)}_{3\bj}+f^{(\bv_4)}_{4\bj}|
>10^{-12}\}.$
In both cases this is a disk of diameter $475$ pixels, thus the support is
small.
Yet the magnitude DFT data of these images are equal to precision
$10^{-15}$, thus phase retrieval is
incapable of distinguishing $\bF_a$ from $\bF_b$,
even if the data is measured to, say, 12 digits of accuracy.
\end{example}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{.4\textwidth}
\centerline{\includegraphics[width= 5.5cm, height=5.5cm]{img_sum_ref.pdf}}
\caption{A sum of 4 component Gaussians.}
\end{subfigure}\qquad
\begin{subfigure}[t]{.4\textwidth}
\centerline{\includegraphics[width= 5.5cm, height=5.5cm]{img_sum3.pdf}}
\caption{A different sum of 4 component Gaussians.}
\end{subfigure}
\begin{subfigure}[t]{.4\textwidth}
\centerline{\includegraphics[height= 4.5cm]{DFT_ref.pdf}}
\caption{The $\log_{10}$-DFT magnitude data for the object in (a).}
\end{subfigure}\qquad
\begin{subfigure}[t]{.4\textwidth}
\centerline{\includegraphics[height= 4.5cm]{DFT_diff3.pdf}}
\caption{The $\log_{10}$ of the magnitude of the differences of
the DFT magnitude data for the objects in [a] and [b].}
\end{subfigure}
\caption{The top row shows two different objects $\bF_a$ and $\bF_b$ defined
in \eqref{illposedimgs}.
More precisely, we plot the central $256\times 256$ portion of the
$1024\times 1024$ array used in constructing these examples.
(c) is a plot of
of the $\log_{10}$ of the magnitude DFT data for the object in (a), while (d) is
the $\log_{10}$ of the difference of the magnitude DFT data for the objects in
(a) and (b). Note that the maximum difference is about $10^{-15}.$}
\label{fig2}
\end{figure}
\begin{remark}
The reader may note that our construction is somewhat pathological,
since the DFT of the image consists of well-separated Gaussian
``islands" of non-zero data, which leads to easier detection of this
sort of $\epsilon$-non-uniqueness. The examples in Section
\ref{nonuniquesec} are less pathological and this form of
non-uniqueness is more difficult to detect. It remains an open problem
to describe the class of images for which $\epsilon$-uniqueness can be
proven, even for very small values of $\epsilon.$
\end{remark}
\section{Algorithms for Phase Retrieval}\label{sec4}
We now see what the results of the previous sections imply about the
behavior of standard algorithms used for phase retrieval. These
algorithms are defined by iterating maps, which are, in turn, built
from ``closest point maps.'' If $W$ is a subset of $\bbR^J,$ then
$P_W(\bF)$ is defined to be the point in $W$ closest to $\bF$ with
respect to the Euclidean distance. If $W$ is a linear subspace then
$P_W$ is the orthogonal projection. If $W$ is convex then $P_W$ is
defined and continuous everywhere, whereas for a non-convex set,
these maps are defined, and continuous, on the complement of a
positive codimensional subset.
\begin{figure}
\includegraphics[width=2.5in]{diffmap1.pdf}
\includegraphics[width=2.9in]{diffmap0.pdf}
\caption{The hybrid iterative map $\bF^{n+1} = D_{AB}(\bF^n)$
defined by \eqref{DAB},
in the setting where $A$ (red curve) and $B$ (blue curve)
are 1D manifolds in $\mathbb{R}^2$.
(a) $\bF \in A \cap B$ is a transversal intersection
(here the center manifold is the single point $\bF$).
(b) Non-transversal case, with generic quadratic separation between
the manifolds.
Each plot shows the iterates $\bF^0,\bF^1,\dots$ (grey dots),
and the construction of the update vector (green solid line)
$\bF^1-\bF^0$ as the difference between a projected reflection
(green star) and a projection (green square).
Note that in each plot the green solid and dotted lines are equal as displacement vectors.
\label{f:diffmap}
}
\end{figure}
In the phase retrieval problem, let us assume that the unknown $\bF$
and its support $S$ are adequate in the sense of definition
\eqref{adequatedef}, with the magnitude torus $\TA_{\ba}$ defined by
the DFT magnitude data $\ba=\ba_{\bF}$.
As a torus, $\TA_{\ba}$ is obviously not a convex
set. The alternating projection method \eqref{apiter}, which we
write here in the form
\begin{equation}
\bF^{n+1} := P_{\TA_{\ba}}\circ P_{B}(\bF^{n})
\end{equation}
is well known to be prone to converge to points that are not in
$\TA_{\ba}\cap B.$ The stable fixed points of the alternating
projection map are points $\tilde{\bF}\in\TA_{\ba}$ such that
$\tilde{\bF}$ and $P_B(\tilde{\bF})$ jointly form a non-zero, local
minimum of the Euclidean distance between the two sets. Empirically
these exist in great profusion, and the alternating projection method
rarely (if ever) converges to a point in $\TA_{\ba}\cap B.$ See
\cite{BEGM} for a more extensive discussion.
In an attempt to avoid such false local minima and improve
convergence, a variety of modifications to the alternating projection
map have been introduced that involve reflection operators as well as
projections. Quite a few variants have appeared in the literature
\cite{bauschke2002,chapman2006,elser2003,elser2007,fienup1987,miao1999,miao2015},
and we will limit our attention to a special case of Fienup's hybrid
input-output (HIO) method \cite{fienup1982}, which is also a special
case of the ``difference map'' approach due to Elser {\em et al.}
\cite{elser2003}. Letting $A$ and $B$ now denote general sets, with
closest point projections $P_A$ and $P_B$, the map that we iterate is
\begin{equation}
D_{AB}(\bF)\;:=\;\bF+P_{A}\circ R_B(\bF)-P_B(\bF),
\label{DAB}
\end{equation}
where $R_B$ is the ``reflection'' around $B$ defined by
$R_B(\bF):=2P_B(\bF)-\bF$; see Figure~\ref{f:diffmap}. This is
Fienup's HIO method with $\beta = 1$ and a specific instance of
Elser's, difference map as well. If $A$ is a linear subspace, then
$D_{AB}$ is also the Douglas-Rachford map, which is defined to be
\begin{equation}
T_{B,A}=\frac{1}{2}\left[R_A\circ R_B+\Id\right],
\end{equation}
see~\cite{borwein}.
Since we are not testing all possible HIO or difference map variants,
we will call the specific method we use here a ``hybrid iterative map."
If $\bF^*$ is a fixed point of $D_{AB}$ then
\begin{equation}
P_{A}\circ R_B(\bF^*)=P_B(\bF^*),
\end{equation}
in other words, the point $\bF^{**}:=P_B(\bF^*)$ lies in $A\cap B$. The
iterates are defined by $\bF^{n+1}=D_{AB}(\bF^n)$, and approximate
reconstructions are given by
\begin{equation}
\br^n := P_B(\bF^n)~.
\label{r}
\end{equation}
If the iterates
converge, then, assuming that $P_B$ is continuous at the limit point, the
sequence $ \{ \br^n \}$ converges to a point in $A\cap B.$
The fixed point set of $D_{AB}$ can be much larger than the set of
intersections. Given a point $\bF\in A\cap B$, we let
\begin{equation}
L_A:=P_A^{-1}(\bF) \qquad \text{ and } \quad L_B:=P_B^{-1}(\bF).
\end{equation}
The \emph{center manifold} defined by $\bF$ (see Figure~\ref{f:diffmap})
is then the set
\begin{equation}
\cC_{AB}^{\bF}:=R_{B}^{-1}(L_A)\cap L_B.
\end{equation}
The center manifold for any $\bF\in A \cap B$ contains the part of the
fixed point set for the map $D_{AB},$ which ``points to'' $\bF.$ Given
an image $\bF$ with support in a small support set $S$ (see Definition
\ref{ssdef}), and choosing $B=B_S$ and $A=\TA_{\ba}$, then the part of
the center manifold near to $\bF$ is given by
$\cC_{\TA_{\ba}B_S}^{\bF}=(\bF+B_S^{\bot})\cap N_{\bF}\TA_{\ba}$, and with
$\dim\cC_{\TA_{\ba}B_S}^{\bF}>|J|/4$. However the map
$D_{\TA_{\ba}B_S}\restrictedto_{\cC_{\TA_{\ba}B_S}^{\bF}}=\Id$, which
is only neutrally stable.
While it is true that the fixed point sets are contained in the center
manifolds defined by points in $\TA_{\ba}\cap B_S,$ there are other
subsets that are attracting. If the pair of points $(\bF_1,\bF_2)\in\TA_{\ba}\times B_S$
defines a critical point of the map $d_{\TA_{\ba}B_S}:\TA_{\ba}\times
B_s\to\bbR_+,$
\begin{equation}
d_{\TA_{\ba}B_S}(\bF,\bg)=\|\bF-\bg\|_2,
\end{equation}
then the set
\begin{equation}
\cC_{\TA_{\ba} B_S}^{\bF_1,\bF_2}\overset{d}{=}(\bF_2+B_S^{\bot})\cap N_{\bF_1}\TA_{\ba}
\end{equation}
contains the line segment from $\bF_1$ to $\bF_2.$ Indeed this
intersection is again a subset with dimension about $|J|/4$ (if
$d=2$). From low dimensional examples it appears that these sets can
define attracting basins, even if the critical point at
$(\bF_1,\bF_2)$ is not a local minimum. The existence of these
attracting sets seems to complicate the dynamics of hybrid map iterations.
The fiber of the tangent bundle
$T_{\bF}\TA_{\ba}$ at $\bF\in \TA_{\ba}$ is the best linear
approximation to $\TA_{\ba}$ near to $\bF$; hence a linearization of
the problem of locating points in $\TA_{\ba}\cap B_S$ is to locate
points in the intersection of the affine subspaces
$T_{\bF}\TA_{\ba}\cap B_S$. With this as motivation we first analyze
the behavior of the map $D_{AB}(\bF)$ when $A$ and $B$ are
linear subspaces.
\subsection{Linear Subspaces}\label{sec4.1}
For the case of a linear subspace, $W\subset\bbR^N,$ the map $P_W$ is
the orthogonal projection onto $W$ and $R_W$ is the orthogonal
reflection with fixed point set $W$. Let $A$ and $B$ denote linear
subspaces of $\bbR^N$.
Let us first consider the linear model for the
benign \emph{transversal} intersection case.
We have $A\cap B=\{\bzero\}$, i.e.\ a single
isolated point,
and $\dim A+\dim B < N$ as befits the phase retrieval application.
(For example, for $d=2$, $\dim A = |J|/2$ and $\dim B \le |J|/4$ when the constraint is adequate.)
To analyze the iteration defined by $D_{AB}$ we split $\bbR^N$ into
the following subspaces $A$, $B$, and $C:=(A+B)^{\bot}=A^{\bot}\cap
B^{\bot}.$ The subspace $C$ is the center manifold defined by
$\{\bzero\}=A\cap B$ for this case. Let $U$, $V$, $Y$ denote matrices whose
columns are orthonormal bases for $A$, $B$ and $C$ respectively. If
$\bF=U\bx_1+V\bx_2+Y\bx_3,$ then, in this representation, the map
$D_{AB}$ takes the form
\begin{equation}
D_{AB}(\bx_1,\bx_2,\bx_3)
= \left(\begin{matrix} 2H^tH& H^t&0\\
-H& 0 & 0\\0&0&\Id\end{matrix}\right)\left(\begin{matrix}\bx_1\\\bx_2\\\bx_3\end{matrix}\right),
\end{equation}
where $H=V^tU$.
In~\cite{BEGM}, it is shown that the upper $2\times 2$ block matrix
$\left(\begin{matrix} 2H^tH& H^t\\
-H& 0 \end{matrix}\right)$
is a contraction, and therefore the
map is contracting in directions normal to the center manifold $C$, and
$\lim_{n\to\infty}D^n_{AB}(\bx_1,\bx_2,\bx_3)=(\bzero,\bzero,\bx_3)$.
This contraction is visible as the convergent spiral in
Figure~\ref{f:diffmap}(a), where $C=\{\bF\}$.
Its rate of contraction is determined by largest singular value of $H$.
The limit point then yields the desired solution
under the projection $P_B$.
The correct linear model for a \emph{non-transversal} intersection is
similar, but $A\cap B=F$ is now a subspace of positive dimension. We now
split $\bbR^N$ as $A_0+B_0+F+C,$ where $A_0=A\cap F^{\bot},$
$B_0=B\cap F^{\bot},$ and $C=(A+B)^{\bot}.$ If $U$, $V$, $X$, $Y$
denote matrices whose columns are
orthonormal bases for $A_0$, $B_0$, $F$, $C$ respectively, then, with
$\bF=U\bx_1+V\bx_2+X\bx_3+Y\bx_4,$ we have:
\begin{equation}
D_{AB}(\bx_1,\bx_2,\bx_3,\bx_4)=\left(\begin{matrix} 2H^tH&H^t&0&0\\
-H & 0& 0 & 0 \\
0 & 0& \Id & 0 \\0&0&0&\Id\end{matrix}\right)\left(\begin{matrix}\bx_1\\ \bx_2
\\\bx_3\\\bx_4\end{matrix}\right).
\end{equation}
As before, the leading $2\times 2$ block is a contraction, and therefore
\[ \lim_{n\to\infty}D^n_{AB}(\bx_1,\bx_2,\bx_3,\bx_4)=
(\bzero,\bzero,\bx_3,\bx_4). \]
Crucially, in this linear model $D_{AB}$ is the identity operator in both
the $F$- and $C$-directions.
For the non-linear phase retrieval problem, the intersections of
interest are, as shown above, generally non-transversal. In this case,
linearization at the intersection point tells one nothing about the
map's behavior, even very near to the center manifold. More
precisely, because the intersections of $\TA_{\ba}\cap B_S$ are
isolated points $\bF$, the affine subspace $C=N_{\bF}\TA_{\ba}\cap
B_S^{\bot}$ remains the linear model for the center manifold in the
non-linear case. The subspace $F=T^0_{\bF}\TA_{\ba}\cap B_S$
constitutes a positive-dimensional set normal to the center manifold
where the map $D_{AB}$ is not known to be contracting.
In the 2-dimensional, quadratic non-transversal case shown in
Figure~\ref{f:diffmap}(b), one observes geometric contraction towards
$C$. In simple, low dimensional examples of this sort convergence,
even in the non-transversal case, is often observed. In fact the
problem becomes, in some sense, easier as non-transversality causes
the dimension of the target center manifold to increase. On the other
hand, very small, but non-zero angles lead to very slow convergence.
The spiral trajectory in Figure~\ref{f:diffmap}(a) is also
note-worthy, as it indicates that the linearization of $D_{AB}$ at the
limit point has complex eigenvalues, which is a phenomenon that
persists in the phase retrieval problem. These questions are discussed
in detail in~\cite{BEGM}.
In the phase retrieval problem, where the dimension $|J|$ is large,
and the geometry is much more complicated, non-transversality seems to
preclude convergence. Indeed, it is common to find that the hybrid
map iteration \emph{stagnates} at a substantial distance from the
center manifold, whenever $\dim T^0_{\bF}\TA_{\ba}\cap B_S>0.$
\begin{definition}
An algorithm has stagnated if the distances from subsequent iterates to the
nearest exact intersection point remain almost constant, and much larger than
machine precision; moreover the distances between successive iterates are
also essentially constant, and much larger than machine precision.
\end{definition}
This behavior is almost always observed when using hybrid iterative
map-based algorithms on noise-free data coming from
images that are not tightly constrained by the support mask,
as we show next.
The
failure of transversality not only renders the problem of finding points in
$\TA_{\ba}\cap B$ ill-posed, but also prevents standard algorithms for finding
these points from converging.
In applications, a variety of possible support information is possible,
such as a bounding rectangle, bounding disc, etc.
Here we use quite an optimistic estimate for our
knowledge of the support, namely that the true support is known
up to a ``padding'' of $p$ pixels. The notion of a $p$-pixel
neighborhood is defined in~\eqref{eqn2.10.002}.
In applications this might possibly derive from
knowledge of a lower-resolution version of the
target image; note that it includes much more information than
merely a reasonably accurate bounding rectangle.
\section{The Performance of the Hybrid Iterative Maps} \label{sec5}
The theory presented in the previous sections makes rather specific predictions
as to how the hybrid iterative map will behave on various sorts
of images, and different sorts of auxiliary information.
For the support constraint, one expects to see that the
iterates $\{\bF^n\}$ of such a map stagnate, and that
the differences between the
approximate reconstructions $\{\br^n\}$ and the nearest exact intersection
point, $\bF$, should lie mostly in directions belonging to
$T_{\bF}\TA_{\ba}\cap B_S$.
In Section~\ref{sec5.1} we show that
these predictions are largely verified in
practice.
In Section~\ref{sec5.2}, we instead consider the non-negativity
constraint. While we still assume that the image has small support,
that information is not explicitly used.
When $B=B_+$, $\pa B_+$ is not a smooth space,
but is rather stratified by the number of vanishing coordinates.
The strata are orthants in Euclidean spaces of various dimensions and
the intersections with
$\TA_{\ba}$ lie on the boundary of the orthant.
In light of this, the intersection $T_{\bF}\TA_{\ba}\cap\pa B_+$ is a
reasonable measure of the transversality of the intersection
$\TA_{\ba}\cap B_+$ at $\bF$.
The more coordinates that vanish at a point, the more
directions in which $\pa B_+$ is strictly convex near to that point.
This suggests that the intersections
between $\TA_{\ba}$ and $B_+$ have a better chance to be transversal,
and therefore hybrid map-based algorithms should work better with
this auxiliary information. We will see, in Section~\ref{sec5.2}, that both
expectations are indeed true.
If $\bF\in\TA_{\ba}$ is a non-negative image, then it is obvious that
the zeroth DFT coefficient $\hat f_{\bzero}=\|\bF\|_1,$ the
$\ell_1$-norm of $\bF$. As follows from the triangle inequality, the $\ell_1$-norm is strictly
minimized on the magnitude torus $\TA_{\ba}$ exactly at such
single-signed images. Hence
for non-negative images one can use the $\ell_1$-norm to define a
different constraint, and therefore different algorithms. Let
$B^1_{r}$ denote the $\ell_1$ ball of radius $r=|\hat f_{\bzero}|$. The
analysis of the intersection $T_{\bF}{\TA_{\ba}}\cap {\pa B_+},$ where
$\bF$ is a non-negative image in $\TA_{\ba},$ has the somewhat
unexpected consequence that
\begin{equation}\label{eqn42}
T_{\bF}{\TA_{\ba}}\cap {\pa B_+}=T_{\bF}{\TA_{\ba}}\cap {\pa B^1_r}.
\end{equation}
That is, the failure of transversality of these two intersections agree
exactly, and therefore algorithms based on using $B=B_+$ can be expected
to behave similarly to those using $B=B^1_r.$
We find that this is true, on average,
though, as the maps involved are non-linear, individual runs of these
algorithms can behave quite differently. This is also briefly explored
in Section~\ref{sec5.2}.
\begin{figure}[t]
\centering
\begin{subfigure}[H]{.8\textwidth}
\centerline{\includegraphics[height= 3.75cm]{cnvng2-2.pdf}}
\caption{ The iterates lie in an attracting basin with
$\dim T_{\bF^{(2,2)}}\TA_{\ba}\cap B_{S_2}=0.$ }
\end{subfigure}
\begin{subfigure}[H]{.8\textwidth}
\centerline{\includegraphics[height= 3.75cm]{noncnvng21.pdf}}
\caption{The iterates lie in an attracting basin with
$\dim T_{\bF^{(1,2)}}\TA_{\ba}\cap B_{S_2}=1.$ }
\end{subfigure}
\caption{An illustration of how the convergence properties of the
hybrid iterative map using $D_{\TA_{\ba} B_{S_2}}$
depend on the dimension of the
$\dim T_{\bF^{(\bv)}}\TA_{\ba}\cap B_{S_2}$. See Example~\ref{exam1}.
The true errors are shown in
blue, residual \eqref{E} in red.}
\label{fig5}
\end{figure}
\subsection{The Support Constraint}\label{sec5.1}
In this section we examine the dependence of hybrid iterative maps
on smoothness ($k$) and padding of the support ($p$) for images of the
types used in Section~\ref{sec2.4.004}. When it is clear which image
is intended, we use $S_{p}$ to refer to $S_{\bF,p}.$
\begin{example}\label{exam1}
This example gives compelling evidence for the central importance of
the failure of transversality. We employ a piecewise
constant $256\times 256$ (i.e.\ $2N=256$) image $\bF$ with support
condition $S_{2}$, which is the exact support padded by $p=2$ pixels as
defined in~\eqref{eqn2.10.002}. The intersection $\TA_{\ba}\cap
B_{S_2}$ contains 25 points, which are the trivial associates
$\{\bF^{(\bv)}: \; \|\bv\|_{\infty}\leq2\}$.
The dimension of each intersection
$T_{\bF^{(\bv)}}\TA_{\ba}\cap B_{S_2}$ depends on $\bv$. At
$\bv=\bzero$ this dimension attains the maximum of 12.
Each of the center manifolds
$\{\cC^{\bF^{(\bv)}}_{\TA_{\ba}B_{S_2}}:\|\bv\|_{\infty}\leq 2\}$
defines a basin of attraction for the hybrid map $D_{\TA_{\ba}B_{S_2}}.$
Starting at a random point on $\TA_{\ba}$ the iterates
seem to eventually fall into one of these basins of attraction.
Letting $\bF^{0}$ denote the starting point,
and writing $B = B_{S_2}$,
we have $\bF^{n}=D_{\TA_{\ba}B}(\bF^{n-1})$ the $n$th iterate. These
points are eventually close to points on a center manifold, but not
very close to the point in $\TA_{\ba}\cap B$ that defines it.
The sequence of approximate reconstructions is defined by \eqref{r}.
The plots in Figure~\ref{fig5} show the \emph{true error}
$\|\br^n-\bF^{(\bv)}\|_2$ (in blue), where
$\bF^{(\bv)}$ is the trivial associate of the true image
closest to $\br^n$, and the
\emph{residual} (in red), which is defined to be
\begin{equation}
E(\bF^n) := \|P_{B}(\bF^n)-P_{\TA_{\ba}}\circ R_{B}(\bF^n)\|_2~.
\label{E}
\end{equation}
Throughout this paper, the true errors are plotted in blue, and the
residuals in red. In a real experiment, only the residual is observable.
Recalling that $\ba$ is the data vector, the Lipschitz bound \eqref{lip}
implies
$$
\|\cM(\br^n) - \ba\|_2 \le C_{\cM} \|\br^n - P_{\TA_{\ba}}(\br^n)\|_2
\leq C_{\cM} \|\br^n - P_{\TA_{\ba}}\circ R_{B}(\bF^n)\|_2 =C_{\cM} E(\bF^n),
$$
where the middle inequality follows from the definition of $P_{\TA_{\ba}}$. This
inequality shows that the data residual norm (left side), a measure
of the extent to which the approximate reconstructions satisfy the DFT-magnitude
constraints, is controlled by our plotted quantity \eqref{E}.
For the hybrid map we also have $E(\bF^n)=\|\bF^{n+1}-\bF^{n}\|$,
so \eqref{E} also provides an indicator as to whether the iterates
are converging. Our plots are semilog-plots with a logarithmic
$y$-axis; a linear decrease therefore indicates exponential
(geometric) decay.
In Figure~\ref{fig5}(a) the iterates have settled into the attracting basin
defined by the associate
$\bF^{(2,2)}$. At this point the intersection with $B_{S_2}$ is
transversal; it is quite apparent that, by the 17,000th iterate, the
approximate reconstructions have converged to this intersection point to
machine precision. In Figure~\ref{fig5}(b) the iterates have settled into the
attracting basin defined by $\bF^{(1,2)},$ for which $\dim
T_{\bF^{(1,2)}}\TA_{\ba}\cap B_{S_2}=1$.
The iterates appear to have largely
stagnated after about the 35,000th iterate, with an true error of about $10^{-2}$
and a residual of about $10^{-4}.$ It follows from~\eqref{eqn21}, and the
observation
that the residual is about the square of the true error, that the
differences
$\br^n-\bF^{(1,2)}$ are likely to lie largely along a common tangent
direction. Indeed, a more careful analysis of these differences, given
in~\cite{BEGM}, verifies this expectation.
\end{example}
\begin{figure}[t]
\centering
\begin{subfigure}[H]{.45\textwidth}
\includegraphics[height= 4.5cm]{k0_p1_50reps_no_conv.pdf}
\caption{True errors and residuals for $D_{\TA_{\ba} B_{S_p}}$ with $k=0, p=1.$}
\end{subfigure}\qquad
\begin{subfigure}[H]{.45\textwidth}
\includegraphics[height= 4.5cm]{k0_p3_50reps_no_conv.pdf}
\caption{True errors and residuals for $D_{\TA_{\ba} B_{S_p}}$ with $k=0, p=3.$}
\end{subfigure}
\caption{The convergence properties of the hybrid map algorithm
$D_{\TA_{\ba} B_{S_p}},$ with $k=0$; $p=1,3,$ 50 random restarts on
a $256\times 256$-image. See Example~\ref{e:startpt}.}
\label{Fig6}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[H]{.45\textwidth}
\includegraphics[height= 4.5cm]{k2_p1_50reps_no_conv.pdf}
\caption{True errors and residuals for $D_{\TA_{\ba} B_{S_1}}$ with
$k=2, p=1,$ without convolution.}
\end{subfigure}\qquad
\begin{subfigure}[H]{.45\textwidth}
\includegraphics[height= 4.5cm]{k2_p3_50reps_no_conv.pdf}
\caption{True errors and residuals for $D_{\TA_{\ba} B_{S_3}}$ with
$k=2, p=3,$ without convolution.}
\end{subfigure}
\begin{subfigure}[H]{.45\textwidth}
\includegraphics[height= 4.5cm]{k1_p1_50reps_conv.pdf}
\caption{True errors and residuals for $D_{\TA_{\ba} B_{S_1}}$ with
$k=2, p=1,$ with convolution.}
\end{subfigure}\qquad
\begin{subfigure}[H]{.45\textwidth}
\includegraphics[height= 4.5cm]{k1_p3_50reps_conv.pdf}
\caption{True errors and residuals for $D_{\TA_{\ba} B_{S_3}}$ with
$k=2, p=3,$ with convolution.}
\end{subfigure}
\caption{In (a,b) we examine the convergence properties of the
hybrid map $D_{\TA_{\ba} B_{S_p}},$ on a $k=2$ image
created without convolution; $p=1,3,$ and, in (c,d) on an image
created by convolving the $k=0$ image with a Gaussian, chosen so
that the power spectra matches the $k=2$ case as well as possible.
See Example~\ref{e:startpt}.
As
before, $p=1,3.$ For each plot we used 50 random restarts on a
$256\times 256$-image; the axes on all four plots are the same.}
\label{Fig6.1}
\end{figure}
\begin{example}\label{e:startpt}
In this example we explore the effects of choosing different
starting points for a variety of images with different levels of
smoothness, using an algorithm based on the maps $D_{\TA_{\ba}
B_{S_p}},$ with $p=1,3.$ Figures~\ref{Fig6}(a,b) shows the
behavior of 10,000 iterates of these algorithms for a piecewise
constant image, $k=0,$ and $p=1,3;$ Figures~\ref{Fig6.1}(a,b) are
similar, but with a smoother image, for which $k=2,$ and $p=1,3.$ In
this experiment the image is defined as a sum of functions,
as in~\eqref{eqn2.35.004}--\eqref{eqn2.36.004}, with $k=2.$ Finally
in Figures~\ref{Fig6.1}(c,d) we show the results of a similar
experiment where the image is smoothed by convolving with a
Gaussian. The width of the Gaussian is selected so that the power
spectra of the images used in $(a,b)$ and $(c,d)$ are as similar as
possible. For each image we show the true errors (blue) and residuals
(red) for 50 random initial conditions.
For a piecewise constant image ($k=0$) an algorithm based on
iterating $D_{\TA_{\ba} B_{S_1}}$ seems to converge in essentially
every trial, albeit with a wide range of rates. Some of the true error
curves exhibit scalloping behavior. The linearizations of
$D_{\TA_{\ba} B_{S_1}}$ at such limiting fixed points have the
surprising property that they are \emph{not} contractions. Instead
they are highly non-normal maps, with complex eigenvalues of modulus
less than $1$. A very simple example of this phenomenon appears in
Figure~\ref{f:diffmap}(a). Other trajectories seems to be
contracting uniformly toward a fixed point. With a looser support
constraint, an algorithm based on $D_{\TA_{\ba} B_{S_3}}$ seems to have
stagnated for all trials, except one. Each of the images $\{\bF^{(\bv)}:\:\|\bv\|_{\infty}\leq
3\}$ defines an attracting basin for this map. Only at the
``corners'' $\{\bF^{(\pm3,\pm3)}\}$ is the intersection with
$B_{S_3}$ transversal. For the single trial that shows convergence
$\bv=(-3,3);$ this is the only trial that found an attracting basin
defined by a transversal intersection.
It is apparent that a smoother image and/or looser support
constraint makes it much harder for these algorithms to converge. In
the experiment whose results are plotted in Figure~\ref{Fig6.1}(a),
the image has $k=2$ and we use a 1-pixel neighborhood of the true
support for the support constraint. The scalloping curves strongly
indicate that the iterates have fallen into an attracting basin
defined by a transversal intersection, and that these iterates are
very slowly converging to a fixed point. The slow convergence
(relative to the $k=0$ case) is a result of the much smaller
non-zero angles between $T_{\bF^{(\bv)}}\TA_{\ba}$ and $B_{S_1},$
caused by the smoothness, even for translates, $\bv,$ where the
intersection is transversal. The scalloping of the (unobservable)
true errors is reflected in a similar scalloping in the (observable)
residuals. In other experiments the iterates seem to be very slowly
convergent, or perhaps have stagnated. In these cases the residual
is roughly the square of the true error, indicating approach along a
direction lying in $T_{\bF^{(\bv)}}\TA_{\ba}\cap B_{S_1}.$
The plots in Figure~\ref{Fig6.1}(b) indicate that all trials have
stagnated, though they do seem to fall into two distinct groups. In
the first group, the true error is close to $1,$ suggesting that the
iterates have not found an attracting basin defined by a true
intersection. In the second group the true errors are below $10^{-1}$
and the residuals are often even smaller than the squares of the
true errors. Empirically, this seems to occur when the iterates find an
attracting basin, not defined by a true intersection, but close to
one that is. This sort of behavior can persist for a very large
number of iterates (millions, at least).
For the plots in Figure~\ref{Fig6.1}(c,d), we use images that are
defined by convolving with a Gaussian. As shown in
Sections~\ref{sec2.3.004}--\ref{sec2.4.004}, this leads to a large
$\dim T_{\bF^{(\bv)}}\TA_{\ba}\cap B_{S_p},$ even for $\bv$ with
$S_{\bF^{(\bv)}}\subset S_p.$ The most striking comparison is
between Figure~\ref{Fig6.1}(a) and Figure~\ref{Fig6.1}(c): in (a)
essentially all experiments terminate with an true error less than
$10^{-1}$ and many appear to be slowly converging, with true errors often
less than $10^{-3},$ whereas in (c), all but 2 cases
have stagnated with a true error very close to $10^{-1},$ and a residual
close to $10^{-4}.$ In one case the true error appears to have stagnated
at about $10^{-3}$ and in another, the true error is $10^{-2},$ and
appears to be decreasing geometrically. In (b) and (d) the
looseness of the support constraint appears to be the dominant
source of difficulty. It is notable how different (a) and (b) are, but
how similar (c) and (d) are. With a non-transversal intersection,
more precise support information does little to improve the behavior
of the algorithm.
\end{example}
From these examples we see that algorithms based on hybrid iterative
maps often stagnate at a very substantial distance from any
true intersection point. Even for a piecewise constant image, the
iterates stagnate, most of the time, once the support constraint becomes
somewhat imprecise. The quantitative relationship between the true
errors and the residuals often indicates approach along common
tangent directions.
\subsection{The Positivity and $\ell_1$ Constraints}\label{sec5.2}
We turn now to the usage of non-negativity as auxiliary information,
and begin by recalling that non-negativity alone does not suffice for
generic uniqueness up to trivial associates. However, if we also assume
that the autocorrelation image $(\bF\star\bF$) has sufficiently small
support, then this does indeed define an adequate constraint for
$\TA_{\ba}\cap B_+$ to consist of finitely many points, which are
generically trivial associates. A special case of the uniqueness
result proved in~\cite{BEGM} is
\begin{theorem}
Let $M$ be a positive integer, let $J=\{-2M,-2M+1,\dots,2M\}^d$, and
let $\TA_{\ba}$ be the
magnitude torus defined by a non-negative image $\bF\in\bbR^J$ for which
$S_{\bF\star\bF}\subset\{-M_0,-M_0+1,\dots,M_0\}^d$,
where $M_0$ is the largest integer not exceeding $4M/3$.
Then the intersection $\TA_{\ba}\cap B_+$ consists of finitely many
points, which, generically, are trivial associates of $\bF$.
\end{theorem}
\noindent
For a non-negative image, $\bF,$ the usual containment
$S_{\bF\star\bF}\subset S_{\bF}-S_{\bF}$ is an equality. This fact
allows one to deduce an upper bound on $S_{\bF}$ from the bound on
$S_{\bF\star\bF}.$ The theorem then follows from Hayes' uniqueness
theorem.
\begin{remark}It should be noted that the autocorrelation image is determined
by the measured data $\{|f_{\bj}|^2:\:\bj\in J\}$ and therefore the
support condition on $\bF\star\bF$ is, in principle, verifiable. The
theorem is stated for images of size $(4M+1)^d$; there is an analogous
result for images of any size, whose precise statement depends on the
dimensions of the image mod 4; see~\cite{BEGM}.
\end{remark}
The analysis in the case of the support constraint suggests that the
``transversality'' of the intersection at $\bF\in\TA_{\ba}\cap B_+$
will strongly influence the behavior of algorithms based on the map
$D_{\TA_{\ba}B_+}.$ When $\TA_{\ba}\cap B_+$ is finite, this
intersection actually lies in $\pa B_+,$ which is not smooth, but
is a piecewise affine space. Therefore a
reasonable measure of the failure of transversality is
$T_{\bF}\TA_{\ba}\cap \pa B_+.$ It turns out that to study these
intersections it is very helpful to consider the $\ell_1$-norm as a
function on $T_{\bF}\TA_{\ba}.$ In~\cite{BEGM}, it is shown that the
$\ell_1$-norm on this affine subspace assumes its minimum value at
$\bF.$ The intersection at $\bF$ is transversal, i.e. locally
$T_{\bF}\TA_{\ba}\cap B_+=\{\bF\},$ if and only if this is a strict
minimum. This analysis also establishes the equality in
equation~\eqref{eqn42}, and shows that the intersection is a proper
convex cone lying in an orthant of a Euclidean space. The analysis
leads to a practical method for computing these intersections in
concrete examples.
Using this approach, we have considered many examples of the type
defined in~\eqref{eqn2.35.004}--\eqref{eqn2.36.004} with various
values of $k>0,$ and have never found an example with a
non-transversal intersection. We have also carried out these
computations for a collection of $128\times 128$ images, defined by
convolution of a piecewise constant image with $G_k,$ with values of
$k$ ranging from $0$ to $6.$ The results are shown in
Table~\ref{tab2}; the $\dim T_{\bF}\TA_{\ba}\cap B_+$ increases slowly
with $k.$ As $\pa B_+$ is strictly convex in many directions near to
points in $\TA_{\ba}\cap B_+,$ one might expect algorithms based on
$D_{\TA_{\ba}B_+}$ to work better than those based on
$D_{\TA_{\ba}B_{S_p}},$ In the following example we show that, even
for images defined by convolution, this is, indeed, the case.
\begin{table}[h]
\begin{center}
\begin{tabular}{| c || c | c | c |c|c|c|c|}
\hline
$k$ & $0$ & $1$& $2$ & $3$& $4$&$5$&$6$ \\ \hline\hline
$\dim T_{\bF}\TA_{\ba}\cap\pa B_+$&0 &0 &4 &10 &18 &22 &34 \\ \hline
\end{tabular}
\end{center}
\caption{Table showing the dimensions of $T_{\bF}{\TA_{\ba}}\cap \pa B_+$ for
$k=0,1,2,3,4,5,6.$ }\label{tab2}
\end{table}
\begin{example}
{\rm In Figure~\ref{fig6} we show the results of 10,000 iterates of
$D_{\TA_{\ba}B_+}$ with 25 random starting points for each of four
$256\times 256$ images (i.e.\ $N=128$). The images, constructed
using~\eqref{eqn2.35.004}--\eqref{eqn2.36.004}, have varying degrees
of smoothness with $k=0,2,4$ in (a), (b), (c), respectively. For
comparison, in (d) we show the results with an image defined by
convolution with a Gaussian, where the width is selected so the
power spectrum is similar to the $k=2$ case. The plots in (a) show
geometric convergence, with a wide range of rates. In (b) and (c)
most examples quickly achieve errors in the $10^{-2}-10^{-3}$ range, and then the error plots display the characteristic scalloping
behavior seen in Figure~\ref{Fig6.1}(a). These trajectories are, in
fact, very slowly converging. The plots shown in (d) indicate that,
for images defined by convolution, the algorithm again stagnates,
though the ultimate true error is a little smaller with the
positivity constraint than with the support constraint. This
reflects the fact that $\pa B_+$ is considerably more convex, near
to a point in $\TA_{\ba}\cap B_+$ than a linear subspace like
$B_{S_p}.$ Once again the quadratic relationship between the true
error and the residual indicates that the trajectory ultimately lies
along a common tangent direction.
\begin{figure}[h]
\centering
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{k0_pos_25reps_no_conv.pdf}
\caption{$k=0$}
\end{subfigure}\qquad
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{k2_pos_25reps_no_conv.pdf}
\caption{$k=2$}
\end{subfigure}
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{k4_pos_25reps_no_conv.pdf}
\caption{$k=4$}
\end{subfigure}\qquad
\begin{subfigure}[H]{.4\textwidth}
\centering
\includegraphics[height= 4cm]{k1_pos_25reps_conv.pdf}
\caption{Convolution image}
\end{subfigure}
\caption{25 random restarts of 10,000 iterates of $D_{\TA_{\ba}B_+}$ on
non-negative images with varying degrees of smoothness.}\label{fig6}
\end{figure}
}
\end{example}
\section{Overcoming the difficulty of classical phase retrieval} \label{sec7}
The foregoing sections provide compelling evidence that the intrinsic
difficulty in recovering the phase lies in the local geometry of the
intersections of a magnitude torus $\TA_{\ba}$ with a subset, $B,$
defined by the choice of auxiliary information. To improve the
situation one needs to break what is essentially an infinitesimal
symmetry in order to render these intersections more transversal.
In practice, this can be achieved by collecting different
experimental data:
pthychography has become an important tool for this, consisting essentially of
rastering across the unknown
image with a mask, making a scattering measurement for each location.
This provides a much larger and richer data set to work with at the cost
of a longer, more involved experiment. Hybrid maps and other
iterative phase retrieval methods work well with such data sets and converge
quite rapidly.
Another way to obviate the classical phase retrieval problem is to record
in the near-field of the sample (the Fresnel regime).
For further discussion of pthychography, we refer
the reader to \cite{Dierolf2010,Marchesini2013,Rodenburg2007,Thibault2008}
and the references therein. For a discussion of the mathematical issues in
near-field imaging, see \cite{Hohage2017}.
We limit our attention here to the
coherent diffraction imaging setting (CDI), since it retains some
advantages (including speed/timescale of acquisition), and would
become an even more powerful technology if its associated
phase retrieval problem could be
addressed robustly.
We propose two experimental modifications which could attain that end.
\subsection{Sharp Cut-off Mask}
In biological applications one is
often seeking to image a sample of soft tissue. If one could
cut the sample along a sharp edge, the object would be non-smooth. Moreover,
knowledge of the precise shape would provide for an
accurate support constraint, which would in
turn break the infinitesimal translational symmetry. In practice, it
is better to use a mask that is not invariant under the
inversion symmetry; see~\eqref{eqn3}.
In spite of the fact that the material being imaged may be soft,
it is possible to
obtain very high resolution. The examples shown in Figure~\ref{fig8} were
selected as the best results from 20 random initial conditions, for each of the
two experimental set-ups. Figure~\ref{fig8}(a) shows the result obtained
when running an algorithm defined by the map $D_{\TA_{\ba}B_{S_1}}$ on the
smooth image without a sharp cut-off. The set $S_1$ is the 1-pixel neighborhood
of the set $\{\bj:\:|f_{\bj}|\geq 10^{-10}\}.$ As expected the iterates
quickly stagnate, whereas, in Figure~\ref{fig8}(b) we see that a sharp cut-off
allows for geometric convergence, where we use as a support constraint
the 1-pixel neighborhood of the region bounded by the sharp cut-off. With a
2-pixel support neighborhood the performance degrades markedly. For both of
these experiments the 19 other runs yielded results that were only slightly worse.
\begin{figure}[H]
\centering
\begin{subfigure}[t]{.85\textwidth}
\centering
\includegraphics[width= 10cm]{soft_ref.pdf}
\caption{ 20,000 iterates of $D_{\TA_{\ba}B_{S_1}}$ for a soft
object ($k=6$). }
\end{subfigure}
\begin{subfigure}[t]{.85\textwidth}
\centering
\includegraphics[width = 10cm]{hard_cutoff.pdf}
\caption{20,000 iterates of $D_{\TA_{\ba}B_{S_1}}$ for a soft
object ($k=6$) cut-off with a sharp mask.}
\end{subfigure}
\caption{An illustration of how the convergence properties of a
hybrid iterative map based on $D_{\TA_{\ba} B_{S_1}}$ are affected
by a sharp
cut-off mask. The images in the top row, from left to right are the
support-mask, the reference image, the reconstructed image. The bottom
plots show the true error (blue) and the residual \eqref{E} (red).}
\label{fig8}
\end{figure}
\subsection{External Holography}
A second experimental modification (and perhaps one that is easier to
carry out), consists of what we will refer to as {\em external
holography}. For this, we imagine placing a \emph{known} hard
object in the exterior of the (perhaps soft) object one would like to
image. Several related ideas appear in the literature. One is called
double blind Fourier holography, and was recently considered
in~\cite{leshem2016,raz2014}, with a reconstruction method based on a
mixture of Fourier and linear algebraic ideas. Another approach, using
more complex reference objects, is found in the recent work of
Barmherzig, Cand\`es, et al.,
see~\cite{Candesetal2019a,Candesetal2019b}. The reconstruction method
in this approach is largely algebraic.
Here, we simply make use of the hybrid map based on
$D_{\TA_{\ba} B_{S_1}},$ where the support mask is the 1-pixel neighborhood
of the smallest rectangle that encloses the object, along with the
1-pixel neighborhood of the exterior hard object.
The {\em shape} of the external object must be precisely
known; its location is less important and can be determined as part of
an iteration step.
As we see in Figure~\ref{fig9}, which uses the same soft
object as that employed in Figure~\ref{fig8}(a), inclusion of the external object
leads to geometric convergence.
Using the 2-pixel neighborhood of the external
object leads to results similar to those in Figure~\ref{fig8}(a). Once
again, we have shown the best outcome obtained from 20 independent
trials. Some of the other trials gave markedly worse results than the
one shown here. If we use the exact support of the external object,
then the results consistently resemble those in Figure~\ref{fig8}(b).
\begin{figure}[H]
\centering
\includegraphics[width = 10cm]{ext_obj.pdf}
\caption{An illustration of how the convergence properties of the
hybrid map $D_{\TA_{\ba} B_{S_1}}$ are improved by the
addition of a hard external object. See the caption of the previous
figure for explanation of the plots and curves.}
\label{fig9}
\end{figure}
It is worth noting that
external holography overcomes the microlocal
non-uniqueness described in Section~\ref{sec3.2}. For simplicity, suppose the
$\bF_0$ has a decomposition as $\bF_0=\bF_1+\bF_2,$ where $\bF_1$ and $\bF_2$
are compact objects for which the supports of $\hbf_1$ and $\hbf_2$ are
disjoint to a high degree of accuracy. Let $\bg$ denote the external object;
for translation directions $\bv_1,\bv_2,$ we observe that,
for all $\bk$,
\begin{multline}\label{eqn44}
|f_{1\bk}+f_{2\bk}+g_{\bk}|^2-|f^{(\bv_1)}_{1\bk}+f^{(\bv_2)}_{2\bk}+g_{\bk}|^2
=\\
2\Re\left[f_{1\bk}\overline{g_{\bk}}(1-e^{\frac{\pi i\bk\cdot\bv_1}{N}})
+f_{2\bk}\overline{g_{\bk}}(1-e^{\frac{\pi i\bk\cdot\bv_2}{N}})\right]
+O(|f_{1\bk}f_{2\bk}|).
\end{multline}
Assuming that DFT coefficients $g_{\bk}$ decay slowly, the first term on the right
hand side of~\eqref{eqn44} is typically many orders of magnitude
larger than the error term $O(|f_{1\bk}f_{2\bk}|).$ This shows that
a hard external object effectively breaks the microlocal translational
symmetry that leads to $\epsilon$-non-uniqueness.
\section{Conclusions}
In this paper, we have described a framework for analyzing the
classical phase retrieval problem, where only the magnitude of the
Fourier transform of an unknown object is measured, typically in
combination with some information about its support. Perhaps most
alarmingly, we have shown that the problem is classically
ill-posed---that is, with typical support information, the locally
defined inverse map is only H\"older continuous. Moreover, one can
easily construct objects that are quite distinct, but have
supports, and magnitude DFT data that are indistinguishable to any
precision $\epsilon>0$. While some such counterexamples are clearly
pathological, others are not (as shown in section
\ref{nonuniquesec}). This leads to two open mathematical questions:
how dense is the set of $\epsilon$-non-unique objects in the space of all
objects, and can one determine, for a given data set, whether phase
retrieval is even possible, at a given precision?
Assuming that, in the generic case, the problem is solvable, we have
also shown that phase retrieval typically involves finding the
intersection of two sets that do not meet transversally. It is
precisely this failure of transversality that prevents the local
inverse from being Lipschitz continuous. The formal linearization at a
non-transversal intersection has infinite condition number. Beyond
this, small angles between the set $B,$ defined by the auxiliary data,
and the fiber of the tangent bundle to $\TA_{\ba},$ at the
intersection point, prevent standard iterative methods from
converging. These effects are mitigated by having an object with a
sharp boundary and accurate support information. As the external
holography example shows, it suffices to have these properties for a
component of the object being imaged.
The mathematical foundations of this paper are presented in detail in \cite{BEGM},
and we are currently working on modifications of the experimental protocol
(other than pthychography or near-field imaging, see \cite{Hohage2017})
that will lead to better-posed inverse problems. The results of that work will
be reported at a later date.
|
\part*{Supplementary material}
\section*{DD sequence evolution computation}
The DD sequence for quadrupole cancellation reads
$F\left(\frac{T}{3}\right)\Pi^{-x}\left(\frac{\pi}{2}\right) \Pi^{-y}\left(\frac{\Omega_{0}T}{3}\right)\Pi^{y}\left(\frac{\Omega_{0}T}{3}\right)\Pi^{x}\left(\frac{\pi}{2}\right)$.
In order to compute the superposition phase accumulated through the sequence, we first write explicitly
\begin{multline}
\Pi^{-y}\left(\frac{\Omega_{0}T}{3}\right)\Pi^{y}\left(\frac{\Omega_{0}T}{3}\right)= \\ \exp\left[i\frac{1}{\hbar}\left(Q_{J}\frac{T}{3}(\textbf{J}^{2}-\frac{3}{2}(\textbf{J}^{2}_{z}+\textbf{J}^{2}_{x}))-\Omega_{0}\frac{T}{3}\textbf{J}_{y}\right)\right] \\ \exp\left[i\frac{1}{\hbar}\left(Q_{J}\frac{T}{3}(\textbf{J}^{2}-\frac{3}{2}(\textbf{J}^{2}_{z}+\textbf{J}^{2}_{x}))+\Omega_{0}\frac{T}{3}\textbf{J}_{y}\right)\right]= \\
\exp\left[i \frac{Q_{J}T}{\hbar}(\frac{2}{3}\textbf{J}^{2}-(\textbf{J}^{2}_{z}+\textbf{J}^{2}_{x}))\right]
\end{multline}
We now note that the operators $\Pi^{x}\left(\frac{\pi}{2}\right)$ and $\Pi^{-x}\left(\frac{\pi}{2}\right)$ act as a rotation of $\frac{\pi}{2}$ angle around the $x$ axis. Therefore, the operation of the first four operators is equivalent to the operation of the evolution operator
\begin{equation}
U_{1}=\exp\left[i\frac{Q_{J}T}{\hbar}(\frac{2}{3}\textbf{J}^{2}-(\textbf{J}^{2}_{y}+\textbf{J}^{2}_{x}))\right].
\end{equation} The final free evolution part is
\begin{equation}
U_{2}=\exp\left[i\frac{\delta T}{3\hbar} \textbf{J}_{z}+\frac{Q_{J}T}{\hbar}\left(\frac{1}{3}\textbf{J}^{2}-\textbf{J}^{2}_{z}\right)\right].
\end{equation}
$U_{1}$ and $U_{2}$ commute and therefore the total sequence evolution results in
\begin{multline}
U_{2}U_{1}= \\
\exp\left[i\frac{\delta T}{3\hbar} \textbf{J}_{z}+i\frac{Q_{J}T}{\hbar}\left(\textbf{J}^{2}-\left(\textbf{J}^{2}_{x}+\textbf{J}^{2}_{y}+\textbf{J}^{2}_{z}\right)\right)\right]= \\
\exp\left[i\frac{\delta T}{3\hbar} \textbf{J}_{z}\right]
\end{multline}
matching the result stated in the main text.
\section*{Fast magnetic noise}
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Supple_Low_High_Amp.pdf}
\caption{\textbf{Parity fringes with low and high RF power.} Each plot shows two parity fringes taken with the quadrupole cancellation sequence: one for Rabi frequency $\Omega_{0}\approx2\pi\times90\text{ kHz}$ (grey circles) with the parity fit to the theoretical model (lack solid line), and one for Rabi frequency $\Omega_{0}\approx2\pi\times47\text{ kHz}$ (dark blue circles) along with its corresponding fit to the theoreical model (red solid line). The plots from top to bottom show the parity fringes for the (1,2),(2,3),(1,3) ion pairs, corresponding to the ion numbers in figure 2 in the main text.}
\label{Low_High_Amp_comparison}
\end{figure}
In our experiment, lower Ramsey fringe contrast was observed in experiments cancelling the quadrupole shift with respect to experiments measuring it. We believe it is due to fast magnetic noise components that were not compensated. These noise components overlap with our dynamical decoupling sequence spectrum. When choosing a lower RF Rabi frequency, the contrast improves. The results of experimental comparison between two DD pulse Rabi frequencies are shown in Fig. \ref{Low_High_Amp_comparison}. Here we compared the parity oscillations for two different RF field amplitudes. The Rabi frequencies were approximately $90$ and $47$ kHz for high and low amplitudes, respectively. The parity fringes were fitted to the expression
\begin{equation}
a\times\exp\left[-\frac{t}{b}\right]\cos\left[2\pi\times c t\right]
\label{fringe_fitting_model}
\end{equation}
where $a,b,c$ are fit parameters corresponding to the fringe amplitude, fringe decay constant and fringe frequency. Comparison between the amplitude and decay fit parameters for all three of the ion pair correlations is shown in Fig. \ref{Low_High_Amp_fit_parameters}. From the plot it is evident that with lower RF power the contrast increases by roughly 10\% compared to the higher power. In addition, the decay time is larger in the lower power experiment. We then conclude that the lower contrast depends on the RF power, and can be minimized. Therefore, it is not a limitation to our sequence. In the experiment shown in Fig. 2c (III,IV) in the main text, the lower RF amplitude was used.
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Supple_Low_High_Amp_fit_parameters_amp_decay.pdf}
\caption{\textbf{Comparison between fringe amplitude and decay time for high and low RF power.} \textbf{(Top, Bottom)} Maximum likelihood estimation for the parameter $a,b$ in equation \ref{fringe_fitting_model} respectively, with error-bars marking 95\% confidence interval. Blue (Orange) circles mark the estimated parameter for low (high) Rabi frequency.}
\label{Low_High_Amp_fit_parameters}
\end{figure}
\section*{Radio-frequency ac stark-shifts and magnetic field shifts cancellation}
Due to the fact that the first part of our quadrupole cancellation sequence uses continuous RF drive resonant with the excited state Zeeman manifold separation, the ground state Zeeman manifold is ac stark-shifted. We define the magnetic response in angular frequency per magnetic field of the ground state as $\chi_{g}$ and of the excited state as $\chi_{e}$. The magnetic quantum numbers of the ground and excited states are denoted as $m_{g}$ and $m_{e}$ respectively, and for a given RF power we define $\Omega_{g}$ and $\Omega_{e}$ as the ground and excited state magnetic Rabi frequencies. \\
Applying the sequence mentioned in Eq. 5 in the main text results in a final phase shift between the excited and ground state that does not include laser phase drifts. In the first part of the sequence, $\frac{2}{3}$ of the total interrogation time $T$, the excited state accumulates phase due to quadrupole shift contributions only (since the magnetic field drifts are averaged out by the drive and the drive phase cancels due to the rotary echo). However, the ground state accumulated phase has two contributing terms: magnetic field drifts part $\frac{2}{3}\times\chi_{g}m_{g}BT$ where $B$ is the ambiant magnetic field, and ac stark shift term $\pm\frac{2}{3}\times\frac{\Omega_{g}^{2}}{2\delta}T$, where $\delta=\left(\chi_{e}-\chi_{g}\right)B$ and the shift's sign is the sign of $m_{g}$. Since these shifts are dependent differently on $\Omega_{g}$ and $B$, they cannot cancel each other in a robust way. In the second part of the experiment, the last $\frac{1}{3}T$, the ions are free-evolving according to Eq. 5 in the main text. While this free evolution cancels the quadrupole shift phase entirely, magnetic field phase continues to accumulate in both excited and ground states. This phase between ground and excited states takes the form of $\frac{1}{3}\left(-\chi_{g}m_{g}+\chi_{e}m_{e}\right)BT$. In total, the final superposition not taking into account laser drifts can be written as $\frac{1}{\sqrt{2}}\left(\exp{i\phi}\left|g\right\rangle+\left|e\right\rangle\right)$ where $\phi$ is given by
\begin{equation}
\phi=\chi_{g}m_{g}BT-\frac{1}{3}\chi_{e}m_{e}BT\pm\frac{2}{3}\times\frac{\Omega_{g}^{2}}{2\delta}T.
\label{unwanted_phase}
\end{equation}
These shifts can be eliminated with the insertion of echo pulses on the ground state manifold. These pulses switches the sign phase accumulation in the first and third terms in Eq. \ref{unwanted_phase}.
\paragraph*{Magnetic field cancellation.} Disregarding the last term in Eq. \ref{unwanted_phase}, we can see that plugging in $^{88}\text{Sr}^{+}$ parameters $\chi_{g}=2.802 \text{ MHz}/\text{G}$ and
$\chi_{e}=1.68 \text{ MHz}/\text{G}$ and using the transition of $m_{g}=-\frac{1}{2}$ and $m_{e}=-\frac{3}{2}$, applying an echo pulse on the ground state manifold at some $\tau$ results in the phase:
\begin{eqnarray}
&&\phi=\chi_{g}m_{g}B\tau-\frac{1}{3}\chi_{e}m_{e}B\tau \nonumber \\ &&+\left(-\chi_{g}m_{g}B\left(T-\tau\right)-\frac{1}{3}\chi_{e}m_{e}B\left(T-\tau\right)\right) \nonumber \\
&&=2.802\left(-\frac{1}{2}\right)B\tau-\frac{1}{3}1.68\left(-\frac{3}{2}\right)B\tau \nonumber \\ &&+\left(-2.802\left(-\frac{1}{2}\right)B\left(T-\tau\right)-\frac{1}{3}1.68\left(-\frac{3}{2}\right)B\left(T-\tau \nonumber \right)\right)
\end{eqnarray}
Equating this phase to zero we get $\tau=\frac{4}{5}$, which corresponds to the echo time in the last sequence of figure 1 in the main text. A second ground-state echo is then applied just before the final optical Ramsey pulse in order to retrieve the initial ground level within the ground Zeeman manifold. \\
\paragraph*{Magnetic field and ac stark shift cancellation.} In this case, we again use two ground state echo pulses. The first pulse is used to cancel the light shift phase. To that aim, it should be applied at $\tau_{1}=\frac{1}{3}T$. This timing also cancels any magnetic phase accumulation during the first $\frac{2}{3}$ of the interrogation. Therefore, we are left with the magnetic field shifts from the $\frac{2}{3}T$ to $T$. Here, the pulse time should be chosen in order to null the remained phase
\begin{eqnarray}
&&\phi_{\text{remain}}=\left(\chi_{g}m_{g}-\chi_{e}m_{e}\right)B\left(\tau_{2}-\frac{2}{3}T\right)\nonumber \\
&&+\left(-\chi_{g}m_{g}-\chi_{e}m_{e}\right)B\left(T-\tau_{2}\right).
\end{eqnarray}
Equating $\phi_{\text{remain}}=0$ gives the second pulse time
\begin{equation}
\tau_{2}=\frac{1}{6}\left(\frac{\chi_{e}m_{e}}{\chi_{g}m_{g}}+5\right)T. \nonumber
\end{equation}
Depending on the chosen states, a situation in which $\tau_{2}>T$ might occur, for example in the choice of states above. In this case, instead of inserting a second $\pi$ pulse in the ground state manifold it should be inserted in the excited state manifold, and its time should be calculated appropriately. The above sequence is a minimal control sequence, requiring only two echo pulses. However, we comment that more pulses and more sophisticated echo sequences can be used, that can be more efficient in filtering out magnetic noise. \\
We also note here that since the ground-state ac stark-shift depends on $\Omega_{g}^{2}$, it is beneficial to reduce the RF intensity as much as possible. The optimal value should depend on both quadrupole shift magnitude, magnetic field magnitude and magnetic field power spectral density. For example, taking $\Omega_{e}=2\pi\times5\text{ kHz}$ corresponds to $\Omega_{g}$ with the same order of magnitude. Assuming the same parameters of $^{88}\text{Sr}^{+}$, and assuming $1^{\text{st}}$ order Zeeman shift $\approx10\text{ Hz}$ at $50 \text{ Hz}$ frequency shift, the light shift will then be in the several Hz's level, which is easily reduced to the mHz level by echos as mentioned above.
\onecolumn
\section*{Deviations from ideal continuous DD}
We would like to estimate how well does the continuous DD approximation made in equation 4 in the main text describes the real evolution. The approximation states that the continuous DD evolution operator
$\exp\left[i\frac{T}{3}\left(\delta J_{z}+Q_{J}\left(J^{2}-3J_{z}^{2}\right)-\Omega_{0}J_{y}\right)\right]$ has a similar operation to the evolution operator
$\exp\left[i\frac{T}{3}\left(Q_{J}\left(J^{2}-\frac{3}{2}\left(J_{z}^{2}+J_{x}^{2}\right)\right)-\Omega_{0}J_{y}\right)\right]$. Instead of focusing on part of the DD evolution, we evaluate the difference between the actual and approximated entire continuous DD part of our sequence evolution. Here, the actual evolution takes the form of
\begin{multline}
U_{act}\left(T\right)=\exp\left[-i\frac{\pi}{2}\boldsymbol{J}_{x}\right]\exp\left[i\frac{T}{3}\left(\delta \boldsymbol{\boldsymbol{J}}_{z}+Q_{J}\left(\boldsymbol{J}^{2}-3J_{z}^{2}\right)-\Omega_{0}\boldsymbol{J}_{y}\right)\right]\\
\exp\left[i\frac{T}{3}\left(\delta \boldsymbol{J}_{z}+Q_{J}\left(\boldsymbol{J}^{2}-3\boldsymbol{J}_{z}^{2}\right)+\Omega_{0}\boldsymbol{J}_{y}\right)\right]
\exp\left[i\frac{\pi}{2}\boldsymbol{J}_{x}\right]=\\
\exp\left[i\frac{\Omega_{0} T}{3}\left(-\frac{\delta}{\Omega_{0}} \boldsymbol{J}_{y}+\frac{Q_{J}}{\Omega_{0}}\left(\boldsymbol{J}^{2}-3J_{y}^{2}\right)-\boldsymbol{J}_{z}\right)\right]\exp\left[i\frac{\Omega_{0} T}{3}\left(-\frac{\delta}{\Omega_{0}} \boldsymbol{J}_{y}+\frac{Q_{J}}{\Omega_{0}}\left(\boldsymbol{J}^{2}-3J_{y}^{2}\right)+\boldsymbol{J}_{z}\right)\right],
\label{actual_DD_operator}
\end{multline}
and the approximated evolution is
\begin{equation}
U_{app}\left(T\right)=\exp\left[i\frac{2T}{3}\left(Q_{J}\left(\boldsymbol{J}^{2}-\frac{3}{2}\left(\boldsymbol{J}_{x}^{2}+\boldsymbol{J}_{y}^{2}\right)\right)\right)\right]. \label{approximate DD operator}
\end{equation}
We define:
$\boldsymbol{V}=-\frac{\delta}{\Omega_{0}}\boldsymbol{J}_{y}-\frac{3Q_{J}}{2\Omega_{0}}\left(\boldsymbol{J}_{y}^{2}-\boldsymbol{J}_{x}^{2}\right)$ and $\boldsymbol{H}=\frac{Q_{J}}{\Omega_{0}}\left(\boldsymbol{J}^{2}-\frac{3}{2}\left(\boldsymbol{J}_{y}^{2}+\boldsymbol{J}_{x}^{2}\right)\right)$, and we note that $\boldsymbol{V}$ has zeros for its diagonal elements, while $\boldsymbol{H}$ is diagonal with diagonal elements different from zero. Using these definitions we rewrite the operators $U_{act}$ and $U_{app}$:
\begin{equation}
U_{act}\left(T\right)=\exp\left[i\frac{\Omega_{0} T}{3}\left(\boldsymbol{V}+\boldsymbol{H}-\boldsymbol{J}_{z}\right)\right]\exp\left[i\frac{\Omega_{0} T}{3}\left(\boldsymbol{V}+\boldsymbol{H}+\boldsymbol{J}_{z}\right)\right],
\label{actual V,H,Jz}
\end{equation}
\begin{equation}
U_{app}\left(T\right)=\exp\left[i\frac{2\Omega_{0} T}{3}\boldsymbol{H}\right]. \label{approximate V,H,Jz}
\end{equation}
For a mutual eigenstate $\left|m\right\rangle$ of $\boldsymbol{J}^{2}$ and $\boldsymbol{J}_{z}$, we define $\left\langle m\right|U_{act}\left(T\right)\left|m\right\rangle=r_{act}\exp\left(i \phi_{act}\right)$ and $\left\langle m\right|U_{app}\left(T\right)\left|m\right\rangle=\exp\left(i \phi_{app}\right)$. The latter has magnitude of unity due to the fact that $\left|m\right\rangle$ is an eigenstate of $\boldsymbol{H}$. We denote the $\boldsymbol{H}$ eigenvalue corresponding to $\left|m\right\rangle$ by $h_{m}$. With this notation, we write $\phi_{app}=\frac{\Omega_{0}T}{3}\times2 h_{m}$
We now assume for simplicity that $\frac{\delta}{\Omega_{0}}=\frac{Q_{J}}{\Omega_{0}}=p\ll1$, and note that the operators $\boldsymbol{V},\boldsymbol{H}$ and the number $h_{m}$ are proportional to $p$. We now turn to prove that the elements that contribute to the phase $\phi_{act}-\phi_{app}$ with the lowest power in $p$ are proportional to $p^{3} \Omega_{0}T$.
\subsection*{Theoretical proof for $p^{3}$ scaling}
We examine the quantity $\left\langle m\right|U_{act}\left(T\right)U_{app}^{\dagger}\left(T\right)\left|m\right\rangle=r_{act}\exp\left(i\left(\phi_{act}-\phi_{app}\right)\right)=g\left(p,\Omega_{0},\right)+if\left(p,\Omega_{0},\right)$, where $f$ and $g$ are real functions describing real and imaginary parts.
\subsubsection*{Reduction to imaginary part}
We claim that the lowest power of $p$ in a Taylor expansion of the phase $\phi_{act}-\phi_{app}$ is equal to the lowest power of $p$ in the Taylor expansion of the imaginary part $f$. Since $g\left(0,\Omega_{0}\right)=1$ and $f\left(0,\Omega_{0}\right)=0$ (see Eq. \ref{actual_DD_operator} and Eq. \ref{approximate DD operator}), we can write
$g\left(p,\Omega_{0}\right)=1+\left(\text{powers of }p\right)$, and $f\left(p,\Omega_{0}\right)=0+\left(\text{powers of }p\right)$. This leads to the conclusion that the lowest power of $p$ in the expansion of the phase of $\left\langle m\right|U_{act}U_{app}^{\dagger}\left(T\right)\left|m\right\rangle$, which is $\arctan\left(\frac{f\left(p,\Omega_{0}\right)}{g\left(p,\Omega_{0}\right)}\right)$, has the same lowest power of $p$ as $f\left(p,\Omega_{0}\right)$. Therefore, it is sufficient to prove that $f\left(p,\Omega_{0}\right)$ does not contain a linear or quadratic $p$ term.
\subsubsection*{$p^{3}$ scale}
We now expand $\left\langle m\right|U_{act}U_{app}^{\dagger}\left(T\right)\left|m\right\rangle$ in a Taylor series:
\begin{equation}
\left\langle m\right|U_{act}U_{app}^{\dagger}\left|m\right\rangle=\sum_{n,k}\frac{1}{n!k!}\left(i\frac{\Omega_{0}T}{3}\right)^{n+k} \underset{T_{nk}}{\underbrace{\left\langle m\right|\left(\boldsymbol{V}+\boldsymbol{H}-h_{m}\boldsymbol{I}+\boldsymbol{J}_{z}\right)^{n}\left(\boldsymbol{V}+\boldsymbol{H}-h_{m}\boldsymbol{I}-\boldsymbol{J}_{z}\right)^{k}\left|m\right\rangle}}, \label{Taylor series}
\end{equation}
where $\boldsymbol{I}$ is the unity operator and where we defined $T_{nk}$ as a sum coefficient. $T_{nk}$ can be written as summation over terms proportional to
\begin{equation}
\left\langle m\right|...\left[\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)^{l_{i}}\boldsymbol{J}_{z}^{j_{i}}\left(-\boldsymbol{J}_{z}\right)^{j'_{i}}\boldsymbol{V}^{v_{i}}\right]\left[\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)^{l_{i+1}}\boldsymbol{J}_{z}^{j_{i+1}}\left(-\boldsymbol{J}_{z}\right)^{j'_{i+1}}\boldsymbol{V}^{v_{i+1}}\right]...\left|m\right\rangle, \label{T_{nk} terms}
\end{equation}
where for each $i=1,2,3,...,n+k$ one and only one of $\{l_{i},j_{i},j'_{i},v_{i}\}\text{ is } 1$ and the rest are zeros, $\sum_{i}\left(j_{i}\right)\le n$, $\sum_{i}\left(j'_{i}\right)\le k$ and $\sum_{i}\left(l_{i}+j_{i}+j'_{i}+v_{i}\right)=n+k$.
We now prove two statements:
\textbf{(*) All terms in which $\sum_{i}\left(v_{i}\right)=0$ cancel out.}
\textbf{proof:}
Since $\left(\boldsymbol{J}_{z}\right)$ and $\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)$ commute, these terms can be written as:
$\left\langle m\right|\boldsymbol{J}_{z}^{\sum_{i}j_{i}}\left(-\boldsymbol{J}_{z}\right)^{\sum_{i}j'_{i}}\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)^{\sum_{i}l_{i}}\left|m\right\rangle$. In cases where $\sum_{i}l_{i}\ne 0$, the above expression equals $\left\langle m\right|\boldsymbol{J}_{z}^{\sum_{i}j_{i}}\left(-\boldsymbol{J}_{z}\right)^{\sum_{i}j'_{i}}\left|m\right\rangle\left(h_{m}-h_{m}\right)^{\sum_{i}l_{i}}=0$. In cases where $\sum_{i}l_{i}= 0$, the sum over all terms with $n+k=d$ takes the form $\sum_{n}\frac{n!}{\left(n\right)!\left(d-n\right)!}\left\langle m\right|\left(\boldsymbol{J}_{z}\right)^{n}\left(-\boldsymbol{J}_{z}\right)^{k}\left|m\right\rangle=\left\langle m\right|\left(\boldsymbol{J}_{z}-\boldsymbol{J}_{z}\right)^{d}\left|m\right\rangle=0$.
\textbf{(**) Terms in which $\sum_{i}\left(v_{i}\right)=1$ always vanish.}
\textbf{proof:}
Due to the fact that $\left(\boldsymbol{V}\right)$ has only zeros as its diagonal elements, it transforms a state $\left|m\right\rangle$ to a state orthogonal to it. On the other hand, both $\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)$ and $\boldsymbol{J}_{z}$ are diagonal in the $\left|m\right\rangle$ basis, and therefore transforms a state $\left|m\right\rangle$ to a state proportional to it. As a result, the state
$\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)^{a}\boldsymbol{J}_{z}^{b}\left(-\boldsymbol{J}_{z}\right)^{c}\boldsymbol{V}\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)^{a'}\boldsymbol{J}_{z}^{b'}\left(-\boldsymbol{J}_{z}\right)^{c'}\left|m\right\rangle$, for integer $a,b,c,a',b',c'$, must be orthogonal to $\left|m\right\rangle$.
Statements \textbf{(*)} and \textbf{(**)} prove that all terms proportional to $p$ vanish.
Only $T_{nk}$ terms in which $n+k$ is odd contribute to $f\left(p,\Omega_{0}\right)$. Therefore, we prove that terms proportional to $p^{2}$ cancel out for odd $n+k$. Terms proportional to $p^{2}$ are divided to three types:
1 - $\sum_{i}v_{i}=0$ and $\sum_{i}l_{i}=2$
2 - $\sum_{i}v_{i}=1$ and $\sum_{i}l_{i}=1$
3 - $\sum_{i}v_{i}=2$ and $\sum_{i}l_{i}=0$
Types 1 and 2 vanish due to the two statements \textbf{(*)} and \textbf{(**)} above. We now deal with type 3. As is written above, only terms with odd $n+k$ contribute to $f\left(p,\Omega_{0}\right)$. We assume a term in $T_{nk}$ in which $n+k$ is odd and $\sum_{i}v_{i}=2$ and $\sum_{i}l_{i}=0$. We must have $\sum_{i}j_{i}+\sum_{i}j'_{i}=n+k-2$, and that means that either $\sum_{i}j_{i}$ is odd and $\sum_{i}j'_{i}$ is even, or $\sum_{i}j_{i}$ is even and $\sum_{i}j'_{i}$ is odd. In either case, a similar term with switched $\sum_{i}j_{i}$ and $\sum_{i}j'_{i}$ must appear in $T_{kn}$, and due to opposite sign, these terms will cancel in summation.
\subsubsection*{Bound for the residual phase}
We would now show that the phase $\phi_{act}-\phi_{app}$ has a dominant part proportional to $p^{3}\Omega_{0}T$. We will show it in an example for $J=\frac{5}{2}$ spin and for the state $\left|m=\frac{5}{2}\right\rangle$. We now look at all the imaginary terms in Eq. \ref{Taylor series}, meaning terms for which $n+k$ is odd, and are proportional to $p^{3}$. These terms must be in one of two forms:
\begin{enumerate}
\item $\left\langle m\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{3}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|m\right\rangle$,
\item $\left\langle m\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)\left(\boldsymbol{J}_{z}\right)^{g_{3}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|m\right\rangle$.
\end{enumerate}
Note that $g_{1}+g_{2}+g_{3}+g_{4}=n+k-3$. In addition, these terms must have even number of $\left(-\boldsymbol{J}_{z}\right)$ and even number of $\left(\boldsymbol{J}_{z}\right)$, because all terms that do not satisfy this condition cancel in summation with a corresponding term in $T_{kn}$.
We now calculate each of these two terms for some choice of $g_{1},g_{2},g_{3},g_{4}$ for $\left|m=\frac{5}{2}\right\rangle$:
We begin with type 1 term. The operator $\boldsymbol{V}$ couples a state $\left|m\right\rangle$ to the states $\left|m+1\right\rangle,\left|m+2\right\rangle,\left|m-1\right\rangle,\left|m-2\right\rangle$, with the coefficients $v_{m\rightarrow m+1},v_{m\rightarrow m+2},v_{m\rightarrow m-1},v_{m\rightarrow m-2}$ respectively.
We can therefore compute:
\begin{multline}
\left\langle \frac{5}{2}\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{3}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|\frac{5}{2}\right\rangle=\\
\left(\frac{5}{2}\right)^{g_{1}+g_{4}}\left[\left(\frac{3}{2}\right)^{g_{2}}\left(\frac{1}{2}\right)^{g_{3}}V_{\frac{3}{2}\rightarrow\frac{5}{2}}V_{\frac{1}{2}\rightarrow\frac{3}{2}}V_{\frac{5}{2}\rightarrow\frac{1}{2}}+\left(\frac{1}{2}\right)^{g_{2}}\left(\frac{3}{2}\right)^{g_{3}}V_{\frac{1}{2}\rightarrow\frac{5}{2}}V_{\frac{3}{2}\rightarrow\frac{1}{2}}V_{\frac{5}{2}\rightarrow\frac{3}{2}}\right]
\end{multline}
Next, we compute type 2 term:
\begin{eqnarray}
\left\langle \frac{5}{2}\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\left(\boldsymbol{H}-h_{\frac{5}{2}}\boldsymbol{I}\right)\left(\boldsymbol{J}_{z}\right)^{g_{3}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|\frac{5}{2}\right\rangle=\\
\left(\frac{5}{2}\right)^{g_{1}+g_{4}}\left[V_{\frac{3}{2}\rightarrow\frac{5}{2}}V_{\frac{5}{2}\rightarrow\frac{3}{2}}\left(h_{\frac{3}{2}}-h_{\frac{5}{2}}\right)\left(\frac{3}{2}\right)^{g_{2}+g_{3}}+V_{\frac{1}{2}\rightarrow\frac{5}{2}}V_{\frac{5}{2}\rightarrow\frac{1}{2}}\left(h_{\frac{1}{2}}-h_{\frac{5}{2}}\right)\left(\frac{1}{2}\right)^{g_{2}+g_{3}}\right]
\end{eqnarray}
When we plug the numbers $V_{\frac{5}{2}\rightarrow\frac{3}{2}}=p (-i\frac{\sqrt{5}}{2})$, $V_{\frac{5}{2}\rightarrow\frac{1}{2}}=p (3\sqrt{\frac{5}{2}})$, $V_{\frac{3}{2}\rightarrow\frac{1}{2}}=p (i\sqrt{2})$, $h_{\frac{3}{2}}-h_{\frac{5}{2}}=-6p$ and $h_{\frac{1}{2}}-h_{\frac{5}{2}}=-9p$ we obtain:
\begin{multline}
\left\langle \frac{5}{2}\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{3}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|\frac{5}{2}\right\rangle=\\
p^{3}\left(-\frac{15}{2}\right)\left(\frac{5}{2}\right)^{g_{1}+g_{4}}\left[\left(\frac{3}{2}\right)^{g_{2}}\left(\frac{1}{2}\right)^{g_{3}}+\left(\frac{1}{2}\right)^{g_{2}}\left(\frac{3}{2}\right)^{g_{3}}\right]
\end{multline}
and
\begin{multline}
\left\langle m\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\left(\boldsymbol{H}-h_{m}\boldsymbol{I}\right)\left(\boldsymbol{J}_{z}\right)^{g_{3}}\boldsymbol{V}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|m\right\rangle=\\
p^{3}\left(\frac{5}{4}\right)\left(\frac{5}{2}\right)^{g_{1}+g_{4}}\left[\left(-6\right)\left(\frac{3}{2}\right)^{g_{2}+g_{3}}+\left(-9\right)\left(\frac{1}{2}\right)^{g_{2}+g_{3}}\right]
\end{multline}
These calculations show that both terms have the same sign. In addition, we write that replacing each of the operators $\boldsymbol{V}$ and $\boldsymbol{H}-h_{\frac{5}{2}}\boldsymbol{I}$ with the operator $\left(h_{\frac{1}{2}}-h_{\frac{5}{2}}\right)\boldsymbol{I}$ yields a more negative number for any choice of $g_{1},g_{2},g_{3},g_{4}$:
\begin{multline}
\left\langle \frac{5}{2}\right|\left(\boldsymbol{J}_{z}\right)^{g_{1}}\left(h_{\frac{1}{2}}-h_{\frac{5}{2}}\right)\boldsymbol{I}\left(\boldsymbol{J}_{z}\right)^{g_{2}}\left(h_{\frac{1}{2}}-h_{\frac{5}{2}}\right)\boldsymbol{I}\left(\boldsymbol{J}_{z}\right)^{g_{3}}\left(h_{\frac{1}{2}}-h_{\frac{5}{2}}\right)\boldsymbol{I}\left(\boldsymbol{J}_{z}\right)^{g_{4}}\left|\frac{5}{2}\right\rangle= p^{3}\left(-27\right)\left(\frac{5}{2}\right)^{g_{1}+g_{2}+g_{3}+g_{4}}
\end{multline}
This means that the sum
\begin{equation}
p^{3}\left(-27\right)\sum_{d\ge3\mbox{ odd}}\sum_{n=0}^{d}\frac{1}{n!\left(d-n\right)!}\left(i\frac{\Omega_{0}T}{3}\right)^{d}\left(\frac{5}{2}\right)^{d-3}
\end{equation}
places a conservative bound on the residual phase. This sum is easily calculated and is equal to
\begin{equation}
i p^{3} \left(\frac{2}{5}\right)^{3} \left(-27\right)\left(\sin\left(2\frac{\Omega_{0} T}{3}\times\frac{5}{2}\right)+\left(2\frac{\Omega_{0} T}{3}\times\frac{5}{2}\right)\right)
\end{equation}
which is the sum of a linear and a bounded oscillating terms, both scale as $p^{3}$.
To conclude, we proved that the residual frequency difference, $f_{r}$, between the operation of actual DD operator and the approximated one enters to leading order as a linear combination of terms scaling as $\frac{\delta^{3}}{\Omega_{0}^{3}}\Omega_{0}$, $\frac{\delta^{2}Q_{J}}{\Omega_{0}^{3}}\Omega_{0}$, $\frac{\delta Q_{J}^{2}}{\Omega_{0}^{3}}\Omega_{0}$ and $\frac{Q_{J}^{3}}{\Omega_{0}^{3}}\Omega_{0}$, with an added bounded oscillating term.
In addition, the statements \textbf{(*)} and \textbf{(**)} also prove that $r_{act}$ deviates from 1 only as of $p^{2}$ to first order.
\textbf{Note: For clarity, the proof assumed $\frac{\delta}{\Omega_{0}}=\frac{Q_{J}}{\Omega_{0}}=p$. A similar proof holds when using $p_{1}=\frac{\delta}{\Omega_{0}}$ and $p_{2}=\frac{Q_{J}}{\Omega_{0}}$, and the conclusion above is still derived.}
\newpage
\subsection*{numerical verification for $p^{3}$ scaling}
\begin{figure}[!h]
\centering
(a) \includegraphics[width=0.43\textwidth]{supple_lost_population.pdf}
(b) \includegraphics[width=0.3\textwidth]{supple_residual_freq_shift_different_m_levels.pdf}
\caption{\textbf{Population lost and residual phase in continuous DD sequence.} \textbf{(a)} Population lost for different $m$ initial states. The plots show numerically calculated $1-\left|\left\langle m\right|U_{app}\left(T\right)\left|m \right\rangle\right|^{2}$ as a function of the time $T$, for initial states $\left|m\right\rangle=\left|\frac{5}{2}\right\rangle,\left|\frac{3}{2}\right\rangle,\left|\frac{1}{2}\right\rangle$. \textbf{(b)} Residual frequency $f_{r}$ scaling with $p$, for different initial $m$ states. The plots show slope extracted from of linear fit to $\phi_{act}\left(T\right)-\phi_{app}\left(T\right)$ for $\left|m\right\rangle=\left|\frac{5}{2}\right\rangle,\left|\frac{3}{2}\right\rangle,\left|\frac{1}{2}\right\rangle$ vs different values of $p\Omega_{0}$. The fits are to third order polynomial in $p$. We note that here we show the shift's absolute value. The shift for state $\left|m\right\rangle=\left|\frac{1}{2}\right\rangle$ has opposite sign with respect to the other states.}
\label{Pop_Lost}
\end{figure}
We calculate the complex amplitudes $\left\langle m\right|U_{act}\left(T\right)\left|m\right\rangle$ and $\left\langle m\right|U_{app}\left(T\right)\left|m\right\rangle$ for interrogation times from $T=0 \text{ s}$ to $T=2 \text{ s}$. In the calculation we set $\Omega_{0}=2\pi\times50 \text{kHz}$.
We would like to verify numerically that indeed $U_{act}\approx1$, since population-leak to other states in the $J$ manifold would result in a loss of contrast in the optical Ramsey fringe. To that aim, we set the conservative values $\delta=Q_{J}=2\pi\times 100\text{ Hz}$, and calculate $1-\left|\left\langle m\right|U_{app}\left(T\right)\left|m \right\rangle\right|^{2}$. The result are shown in Fig. \ref{Pop_Lost}a.
Fig. \ref{Pop_Lost}a shows that this loss in contrast agrees with the scaling of $p^{2}=\left(\frac{\delta}{\Omega_{0}}\right)^{2}=\left(\frac{Q{J}}{\Omega_{0}}\right)^{2}$, with a pre-factor of order 10. These parameters show that this amplitude deviation is small and can be neglected. We therefore need only to account for the phase acquired by the initial state.
\begin{figure}[!b]
\centering
\includegraphics[width=0.4\textwidth]{supple_residual_phase_linear_and_oscillations.pdf}
\caption{\textbf{Linear and oscillating term in $\phi_{act}-\phi_{app}$.} The residual phase was calculated for the state $\left|\frac{5}{2}\right\rangle$, $Q_{J}=2\pi\times10 \text{ Hz}$, $\delta=2\pi\times1 \text{ Hz}$ and $\Omega_{0}=2\pi\times 50 \text{ kHz}$. A dominant linear trend is shown, with added oscillating term.}
\label{linear and oscillations}
\end{figure}
We turn to evaluate the phase acquired by the initial state when $U_{act}\left(T\right)$ or $U_{app}\left(T\right)$ are applied. Our figure of merit is $\left|\phi_{act}\left(T\right)-\phi_{app}\left(T\right)\right|$. Again, here we take $\delta=Q_{J}:=p\Omega_{0}$. Using these definitions, residual frequency shift $f_{r}$ arising from the DD sequence corresponds to $\frac{\partial}{\partial T}\left|\phi_{act}\left(T\right)-\phi_{app}\left(T\right)\right|$. This derivative is estimated through a maximum likelihood fit to a linear function. Fig. \ref{Pop_Lost}b shows the scaling of $f_{r}$ as a function of $p$.
These calculation results exhibit a cubic scaling in $p$. As an example, for $p\Omega_{0}=2\pi\times10\text{ Hz}$, the residual shift from the continuous DD part in the sequence is less than $40\mathrm{ \mu Hz}$, which means less than $\approx 9.1\times 10^{-20}$ relative frequency shift.
Next, we verify that the phase-form of a linear term added to an oscillating term is justified. Fig. \ref{linear and oscillations} shows such a behavior for the state $\left|\frac{5}{2}\right\rangle$, $Q_{J}=2\pi\times10 \text{ Hz}$, $\delta=2\pi\times1 \text{ Hz}$ and $\Omega_{0}=2\pi\times 50 \text{ kHz}$. We note, however, that the plot is only qualitative, since it shows a single frequency component, where more frequency components exist. That is due to sample aliasing of the numerical calculation.
We also calculated the expected shift for our DD sequence applied on three ions with two parameters choices, in which $Q_{J}\ne\delta$. The first was using our evaluated experiment parameters: $Q_{J}/2\pi=28,42,28\text{ Hz}$ and $\delta/2\pi=-33,10,53\text{ Hz}$, including our magnetic field gradient and a possible imperfect RF resonance calibration of $10\text{ Hz}$. First, second and third numbers correspond to ion 1 (left), ion 2 (middle) and ion 3 (right). The second was a typical experimental parameters, where $Q_{J}$ is the same as in our experiment and $\delta/2\pi=5,5,5\text{ Hz}$. The resulting residual shift in $\text{mHz}$ is given below.
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
\multicolumn{4}{|c|}{$m=\frac{5}{2}$} \\
\hline
& Ion 1 & Ion 2 & Ion 3 \\
\hline
Our experiment parameters & 0.63 & 1.5 & 0.92 \\
\hline
Typical clock parameters & 0.45 & 1.5 & 0.45\\
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
\multicolumn{4}{|c|}{$m=\frac{3}{2}$} \\
\hline
& Ion 1 & Ion 2 & Ion 3 \\
\hline
Our experiment parameters & 0.23 & 0.89 & 0.17 \\
\hline
Typical clock parameters & 0.26 & 0.89 & 0.26\\
\hline
\end{tabular}
\end{center}
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
\multicolumn{4}{|c|}{$m=\frac{1}{2}$} \\
\hline
& Ion 1 & Ion 2 & Ion 3 \\
\hline
Our experiment parameters & -0.85 & -2.4 & -1.1 \\
\hline
Typical clock parameters & -0.71 & -2.4 & -0.71\\
\hline
\end{tabular}
\end{center}
\end{document}
|
\section{\label{sec:introduction}Introduction}
Atomically thin two-dimensional (2D) materials are being investigated
for a range of applications including emerging beyond-CMOS electronic devices,
thermoelectrics, and optoelectronics.
A number of these materials have ``ring-shaped'' valence bands.
These materials include the
semiconducting III-VI monochalcogenides, GaS, GaSe, InS, and InSe
\cite{darshanamhat, zolyomi_GaX,zolyomi_InX, GaS_photodetector_AnPingHu, Hennig_GroupIII_ChemMat, SGLouie_GaSe_arxiv, WYao_GaS_GaSe_arxiv, guo2017band},
bilayer graphene when subject to a vertical electric field \cite{Fermi_ring_Neto_PRB07,Falko_BLG_Lifshitz_PRL14, MacDonald_bi_gap_PRB07},
monolayers of Bi$_2$Se$_3$\cite{darshanamhat} and Bi$_2$Te$_3$\cite{Zahid_Lake,Lundstrom_Jesse_Bi2Te3,Udo_Bi2Se3},
few-layers of Bi$_2$Se$_3$ intercalated with 3d transition metals\cite{li2016gate},
monolayer SnO\cite{seixas2016multiferroic, houssa2017hole},
2D hexagonal lattices of group-VA elements \cite{sevinccli2017quartic},
and hexagonal group-V binary compounds\cite{nie2017room}.
A ring-shaped valence band edge results in a $1/\sqrt{E}$ singularity in the 2D density of states
and a step function turn on of the density of modes at the valence band edge
\cite{Lundstrom_Jesse_Bi2Te3,darshanamhat,GaSe_Ajayan_NL13,Ajayan_InSe,GaSe_Geohagen_ACSNano}.
At low temperatures, density functional theory calculations show that the singularity in the
density of states leads to a ferromagnetic phase transition at sufficient hole
doping in GaS and GaSe\cite{SGLouie_GaSe_arxiv, WYao_GaS_GaSe_arxiv}.
More recent calculations find that such a transition is a general property of the
Mexican hat dispersion \cite{seixas2016multiferroic}.
The ring-shaped dispersion affects ionized impurity scattering through the density of states,
the momentum transfer required to scatter around the ring, and the
momentum dependence of the screening.
The question we address is what is the influence of the ``ring-shaped'' dispersion
on the temperature, density, and Fermi energy dependence of the ionized
impurity scattering rates and ionized impurity limited mobility.
Prior studies have theoretically investigated the role of ionized impurity scattering
in two--dimensional materials with a parabolic dispersion.
Ionized impurity scattering can severely limit the mobility
in the transition metal dichalcogenides such as
MoS$_{2}$ \cite{DJena_PRX} and
give rise to an unexpected temperature
dependence of the mobility \cite{ong2013mobility}.
It has been predicted that reducing the doping can enhance the linear screening
response within the Thomas-Fermi theory \cite{kolomeisky2016anomalous}.
The role of screening on charged
impurity scattering and charged impurity limited mobility
in materials with a ring-shaped dispersion
has not yet been addressed.
We address this question
using an analytical bandstructure model with parameters extracted from first principles calculation.
Screening is included within the random phase approximation.
Polarization functions and scattering rates are analyzed,
and the ionized impurity limited hole mobility of
the III-VI materials, GaS, GaSe, InS, and InSe, are compared.
\section{\label{sec:theory}Theory}
The materials and geometry of the problem consist of a monolayer 2D semiconducting material
on a insulating substrate encapsulated by an insulating capping layer
which could be the same as the substrate.
Example insulating materials are BN or SiO$_2$.
The structure is illustrated in Fig. \ref{fig:device} with SiO$_2$ for the
substrate and BN for the capping layer.
A cylindrical coordinate system is used with ${\bf r}$ a vector in the $x$--$y$ plane.
The origin is at the center of the semiconductor.
Charged impurities will be considered for two different positions, in the center of the 2D semiconductor, $z=0$, and
on the surface of the substrate, $z=-d$.
Accounting for the $5$ {\AA} thickness of a monolayer III-VI semiconductor and the $3$ {\AA} van der Waals gap,\cite{darshanamhat}
we use $d=5.5$ {\AA} for the charged impurities on the surface of the substrate.
The value of the impurity density used in all calculations is $10^{12}$ cm$^{-2}$.
All calculated scattering rates are linearly proportional to the impurity density,
and all mobilities are inversely proportional to the impurity density, so any
calculated values can be scaled for different impurity densities.
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{device_v2.eps}
\caption{\label{fig:device}
Monolayer GaS between SiO$_2$ substrate and BN capping layer.
The black atom depicts an ionized impurity inside the capping layer
5.5{\AA} from the channel.
And the blue atom depicts a ionized impurity inside the monolayer
GaS.
}
\end{figure}
The investigation of the effect of the Mexican hat dispersion on screening, scattering,
and mobility, begins with the model quartic dispersion
\begin{equation}
E(k) = \epsilon_h -\frac{\hbar^{2}k^{2}}{2m^*} + \frac{1}{4\epsilon_{h}}\left(\frac{\hbar^{2}k^{2}}{2m^*}\right)^2 .
\label{eq:MH}
\end{equation}
Quartic models have been previously used to investigate interactions in
biased bilayer graphene \cite{Fermi_ring_Neto_PRB07}, multiferroic 2D materials \cite{seixas2016multiferroic},
and electronic and thermoelectric properties of group III-VI and group VA 2D
materials \cite{darshanamhat,sevinccli2017quartic}.
We define our momentum-energy relation such that
the hole kinetic energy is positive,
the valence band edge is at $E=0$, and
negative energies correspond to
energies in the band gap.
The term ${\epsilon}_h$ in Eq. (\ref{eq:MH}) is the height of the `hat' at $k=0$ and $m^*$ is the magnitude of the
effective mass at $k=0$ (the top of the hat).
The addition of the constant term ${\epsilon}_h$ in Eq. (\ref{eq:MH}), shifts the dispersion
so that the minimum energy, corresponding to the band edge, occurs at $E=0$.
For energies $0 < E < {\epsilon}_h$ the Mexican hat dispersion has two Fermi
wavevectors
corresponding to the two branches of the dispersion.
In this energy region, the Fermi surface consists of two concentric circles
shown in Fig. \ref{fig:dispersion_dos}(a).
The radii of the two circles are
$k_1 = \frac{\sqrt{2 m^* {\epsilon}_h}}{\hbar} \sqrt{1 - \sqrt{E/{\epsilon}_h} }$
and
$k_2 = \frac{\sqrt{2 m^* {\epsilon}_h}}{\hbar} \sqrt{1 + \sqrt{E/{\epsilon}_h} }$.
At the band edge, $E=0$, the two circles merge into a single circle with a radius
of $k_0 = 2 \sqrt{m^*{\epsilon}_h} / \hbar$.
The effective mass at the band edge determined from
$\frac{1}{m^*(k_0)} = \left. \frac{\partial^2 E}{\hbar^2 \partial k^2} \right|_{k=k_0}$
is $m^*/2$.
The single--spin densities of states for each individual $k$-space ring are identical
and equal to
\begin{equation}
\begin{array}{ll}
D_{1}(E) = D_2(E) = \frac{m^*}{2 \pi \hbar^2} \sqrt{ \frac{{\epsilon}_h}{E} } \;\;\;\;\;\;\;
&
(0 \leq E \leq {\epsilon}_h),
\end{array}
\end{equation}
The total single-spin density of states is given by the sum and is equal to
\begin{equation}
D(E)=
\left\{
\begin{array}{ll}
\frac{m^*}{\pi \hbar^2} \sqrt{ \frac{{\epsilon}_h}{E} } \;\;\;\;
&
(0 \leq E \leq {\epsilon}_h)\\
\frac{m^*}{2 \pi \hbar^2 } \sqrt{ \frac{{\epsilon}_h}{E} }
&
\left( {\epsilon}_h < E \right) \; .
\end{array}
\right.
\label{eq.DE_hat}
\end{equation}
The density of states is plotted in Fig. \ref{fig:dispersion_dos}(b) using $m^* = 0.409$ m$_0$ and ${\epsilon}_h = 0.11$ eV,
which are similar to the values for monolayer GaS \cite{darshanamhat}.
The density of states diverges as $1/\sqrt{E}$ at the band edge, and it is equal to the
single--spin parabolic density of states, $\frac{m^*}{2\pi \hbar^2}$, at the top of the hat.
A parabolic dispersion $E(k) = \frac{\hbar^{2}k^{2}}{2m^*}$ will be used as a reference and for comparison.
The parabolic and Mexican hat dispersions and density of states are plotted together in Fig. \ref{fig:dispersion_dos}.
An effective mass of $m^* = 0.409 m_0$ is used for both dispersions.
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{dispersion_DOS.eps}
\caption{\label{fig:dispersion_dos}
(a) Comparison of a parabolic (blue) and Mexican hat dispersion (blue).
The height of the Mexican hat band at $k=0$ is $\epsilon_{h} = 0.11$ eV.
(b) Density of states of the parabolic band (blue) and Mexican hat
dispersion (red).
The parabolic and Mexican hat dispersion both have an effective mass of
$0.409$ m$_0$.
}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{charge_density_all.eps}
\caption{\label{fig:charge_density}
Carrier concentration of GaS, GaSe, InS and InSe as a function of Fermi
level $E_F$ for (a) $5$ K, (b) $77$ K and (c) $300$ K.
Parameters used for materials are tabulated in Table
\ref{tab:mat_params}.
}
\end{figure}
The two--dimensional Fourier transform of the bare Coulomb potential for a point charge at position
$x = y = 0$, $z = z_0$ is
\begin{equation}
v(q) = \frac{e^2 e^{-q|z-z_0|}}{2\epsilon q},
\label{eq:vqbare}
\end{equation}
where $e$ is the charge of electron, $\epsilon$ is the average static dielectric constant
and $q$ is the momentum transfer.
Since all of the III-VI materials have relative dielectric constants in the range of
$3 - 4$, we will use the dielectric constant of the semiconductor.
Within the random phase approximation,
the screened Coulomb potential is
\begin{equation}
V(q,z) = \frac{v(q)}{1 - \Pi(q) v(q)} .
\label{eq:coulomb_pot}
\end{equation}
Substituting Eq. (\ref{eq:vqbare}) into Eq. (\ref{eq:coulomb_pot}) gives
the 2D RPA screened potential,
\begin{align}
V(q) &= \frac{e^{2}}{2\epsilon (q e^{q|z-z_0|} - \frac{e^2}{2\epsilon}\Pi (q))}
\nonumber \\
&= \frac{e^{2}}{2\epsilon (q e^{q|z-z_0|} + {q_\lambda(q)})},
\label{eq:coulomb_pot2}
\end{align}
where ${q_\lambda(q)} \equiv -\frac{e^2}{2\epsilon}\Pi (q)$ is the wavevector dependent
inverse screening length.
In the static limit, the polarization function is
\cite{maldague1978many}
\begin{equation}
\Pi (q) = \frac{2}{A} \sum_{\bf{k}}
\frac{ f(E_{{\bf k} + {\bf q}}) - f(E_{{\bf k}}) }
{E_{{\bf k} + {\bf q}} - E_{{\bf k}}}
\label{eq:Pi}
\end{equation}
where $A$ is the area, $E_{\bf k}$ is the eigenenergy at wavevector ${\bf k}$, and
$f(E)$ is the Fermi-Dirac function.
The factor of 2 is for spin degeneracy, since the Mexican hat bands in the III-VI materials
are spin degenerate.
For both the Mexican hat and parabolic dispersions, $E_{\bf k}$ is only a function of the magnitude of $k$.
Therefore, we define the variable,
\begin{equation}
k_+ = \left|\bf{k}+\bf{q}\right| = \sqrt{k^2 + q^2 + 2kq\cos{\theta}},
\end{equation}
and calculate the polarization from Eq. \ref{eq:Pi},
\begin{equation}
\Pi (q) = \frac{1}{2\pi^2} \int_{0}^{\infty} dkk
\int^{2\pi}_0 d\theta \frac{f(E(k_+)) - f(E(k))}
{E(k_+) - E(k)} .
\end{equation}
In the limit ${\bf q} \rightarrow 0$,
the polarization function becomes the negative of the thermally averaged density
of states at the Fermi level,
\begin{equation}
\Pi (q=0) = \int_{0}^{\infty} dE D(E) \frac{\partial f}{\partial E} ,
\label{eq:pi_analytical}
\end{equation}
where $D(E)$ is the density of states.
Using the ${\bf q} \rightarrow 0$ limit for $\Pi({\bf q})$ in Eq. (\ref{eq:coulomb_pot2}),
gives the Thomas-Fermi form of the 2D screened Coulomb potential with an inverse
screening length of $\frac{e^2}{2{\epsilon}} D(E_F)$.
For the Mexican hat dispersion
this is problematic, since
the density of states diverges near the band edge.
Note that in defining the polarization function in Eq. (\ref{eq:Pi}), $\Pi < 0$.
To calculate the momentum relaxation time, we need the matrix elements of the
RPA Coulomb potential.
We assume separable wavefunctions of the form
$\braket{{\bf r}}{{\bf k}} = \frac{1}{\sqrt{A}} e^{i{\bf k} \cdot {\bf r}} \sqrt{\delta(z)}$
and take the matrix elements of $\tilde{V}({\bf r}) = \int \frac{d^2q}{4\pi^2} V(q) e^{i{\bf q} \cdot {\bf r}}$
to obtain
$\bra{{\bf k}} \tilde{V} \ket{{\bf k}'} \equiv V_{{\bf k},{\bf k}'} = \frac{1}{A} V(|{\bf k} - {\bf k}'|)$.
The Fermi's golden rule expression for the inverse momentum relaxation time is given by
\begin{equation}
\frac{1}{\tau (k)} =
\frac{N_I 2\pi}{\hbar}
\sum_{{\bf k}'}
\lvert V_{{\bf k}',{\bf k}} \rvert^2 \delta(E_{{\bf k}'}-E_{{\bf k}})
\left( 1 - \frac{{\bf v}({\bf k}) \cdot {\bf v}({\bf k}')}{\left| {\bf v}({\bf k}) \right| ^2} \right),
\label{eq:Fermi_Au}
\end{equation}
where $N_I$ is the number of charged impurities.
For the Mexican hat dispersion, the group velocity ${\bf v}$ is opposite to the direction
of ${\bf k}$ on the inner ring and parallel to ${\bf k}$ on the outer ring.
On a given branch of the Mexican hat dispersion, $E({\bf k})$ is only a function of the magnitude
of ${\bf k}$.
Therefore, by converting the sum over ${\bf k}'$ into an integral and explicitly keeping track of the
two branches of the dispersion, Eq. (\ref{eq:Fermi_Au}) becomes
\begin{equation}
\frac{1}{\tau(k)} =
\frac{n_I e^4}{4 {\epsilon}^2 \hbar}
\sum_{\nu=1}^2
D_\nu (E)
\int_0^{2\pi} \! \! \! \! d\theta \:
\frac{ \left( 1 - \frac{\bf{v(k'_\nu)}\cdot v(\bf{k})}{v^2(k)} \right) }
{\left(q \: e^{qz_0} + {q_\lambda(q)} \right)^2}
\label{eq:tau}
\end{equation}
where the sum is over the two Fermi rings,
$q=\left| {\bf k}'_\nu -{\bf k} \right| = \sqrt{{k'_\nu }^{2}+ k^2 - 2k'_\nu k \cos\theta}$,
$k$ and $k'_\nu$ correspond to the radii of the concentric iso-energy rings in Fig. \ref{fig:dispersion_dos},
$D_\nu (E)$ is the final single--spin density of states corresponding to
ring $\nu$,
${\bf v}({\bf k}'_\nu)$ is the final group velocity of
ring $\nu$,
and $n_I$ is the impurity density per unit area.
The value of $z_0$ is either zero for impurities placed at the center of the semiconducting monolayer
or $5.5$ {\AA} for impurities placed on the substrate.
The last term on the right of Eq. (\ref{eq:Fermi_Au}) is $1 - \frac{v'}{v}\cos(\theta_{{\bf v},{\bf v}'})$
where $\theta_{{\bf v},{\bf v}'}$ is the angle between
the group velocity of state ${\bf k}$ and the group velocity of state ${\bf k}'$.
This term is the relative change in the component of the velocity that is parallel to the initial
velocity $v$.
When the final velocity $v'$ is in the same direction and greater than the initial velocity $v$,
then scattering from $v$ to $v'$ causes the carrier to speed up and
gives a negative contribution to the momentum relaxation time \cite{1986_Neg_tau_m}.
This situation occurs for carriers that are initially near the top of the hat in Fig. \ref{fig:dispersion_dos}(a)
and then scatter to the outer ring.
However, the negative values are restricted to a range of angles centered around $180^\circ$,
and the integral over $\theta$ in Eq. (\ref{eq:tau}) is always positive.
The carrier mobility is determined from the
average group velocity driven by an external
electric field oriented in the $x$--direction.
To linear order, this is
\begin{equation}
\langle v_x \rangle = \frac{\sum_{\bf k} v_x ({\bf k} ) f_A ( {\bf k} )} {\sum_{{\bf k}} f_0 ({\bf k})},
\label{eq:vx}
\end{equation}
where $ f_A (\bf k) $ is the asymmetric component of the non-equilibrium distribution function.
Within the relaxation time approximation, the asymmetric distribution function can be written as,
\begin{equation}
f_A({\bf k}) = - \tau (k) \frac{e {\cal E}_x}{\hbar}\frac{\partial f_0({\bf k})}{\partial k} \cos \theta,
\label{eq:fA}
\end{equation}
where $ f_0 ({\bf k}) $ is the equilibrium Fermi function, ${\cal E}_x $ is the electric field along
the transport direction and $ \theta $ is the direction of ${\bf k}$ with respect to the $k_x$ axis.
The mobility is directly evaluated from its definition,
$\mu = \langle v_x \rangle / {\cal E}_x$.
Substituting (\ref{eq:fA}) into (\ref{eq:vx}), the final expression for carrier mobility is
\begin{equation}
\mu = - \frac{e}{2 \pi \hbar^2 p}
\int_{0}^{\infty} \! \! \! dk
\: k \: \tau(k) \, \frac{\partial f_0}{\partial \epsilon} \left(\frac{\partial \epsilon}{\partial k}\right)^2,
\label{eq:mu}
\end{equation}
where the spin--degenerate 2D hole density $p = \frac{2}{A}\sum_{\bf k} f_0({\bf k})$.
\section{\label{sec:results}Results}
The polarization function $\Pi(q)$ gives
wavevector dependent screening.
In a two dimensional material with parabolic dispersion,
the density of states is constant which results
in a constant polarization function for $q<2k_F$ at low temperature.
In a Mexican hat dispersion,
the singular density of states gives
a strong wavevector dependence
to the polarization function at low temperature.
It also increases the overall magnitude of the polarization function.
The wavevector dependent inverse screening length ${q_\lambda(q)}$
is added to the momentum transfer $q$
in the denominator of Eq. (\ref{eq:coulomb_pot2}),
and the sum determines the magnitude and wavevector dependence
of the screened Coulomb interaction.
Therefore, we begin by analyzing $q_\lambda$ as a function of $q$
for the Mexican hat dispersion.
To provide a point of reference,
we first show in Fig.~\ref{fig:pi}(a) the well--known
wavevector dependent inverse screening length $q_\lambda$
resulting from a parabolic dispersion
with the Fermi level fixed at $40$ meV above the band edge.
At low temperature and for wavevectors smaller than $2 k_F$,
the magnitude of $q_\lambda$ is simply
$\frac{e^2}{2 \epsilon} \: \frac{m^*}{\pi \hbar^2}$,
i.e. $\frac{e^2}{2 \epsilon}$ times the density of states at the Fermi level.
This is equal to $3.78$ $m_r / {\epsilon}_r = 0.499 \; {\rm \AA}^{-1}$ where $m_r = 0.409$ is the relative effective mass
and ${\epsilon}_r = 3.1$ is the relative dielectric constant.
Since the density of states is constant, the resulting inverse screening length is
constant up until the momentum transfer is greater than $2 k_F$.
At higher temperatures, the polarization function can be written as a convolution of the
zero--temperature polarization and a thermal broadening function \cite{maldague1978many}.
The result is that the sharp $q$-dependent features become smeared out at finite temperatures.
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{pi_fig_v4_PD.eps}
\caption{
\label{fig:pi}
(a)--(c) $q_\lambda(q) \equiv -\frac{e^2}{2{\epsilon}} \Pi(q)$
at three different temperatures: 5 K, 77 K and 300 K.
(a) ${q_\lambda(q)}$ for a parabolic dispersion with $E_F = 40$ meV.
(b) ${q_\lambda(q)}$ for a Mexican hat dispersion with $E_F = +5$ meV.
(c) ${q_\lambda(q)}$ for a Mexican hat dispersion with $E_F = -3$ meV.
The inset shows the two iso-energy rings in momentum space of the Mexican hat dispersion.
The momentum transfer $q$ between two rings is shown.
(d) ${q_\lambda}(q)$ for a Mexican hat dispersion for a fixed charged density of $10^{13}$ cm$^{-2}$.
The corresponding Fermi energies at each temperature are shown on the graph.
For both the parabolic and Mexican hat dispersions, the band structure
parameters are the same as those used in Fig. \ref{fig:dispersion_dos},
and the dielectric constant is ${\epsilon} = 3.1 {\epsilon}_0$.
}
\end{figure}
Unlike the parabolic dispersion
where scattering occurs within a single Fermi ring,
Coulomb scattering in a Mexican hat dispersion occurs within
and between two concentric rings for energies up to $\epsilon_h$, which
defines the height of the Mexican hat dispersion.
Furthermore, the density of states is singular
at the band edge.
To understand the implications of these features,
the inverse screening length is plotted,
as a function of the momentum transfer, $q$,
for different values of the Fermi energy
in Fig.~\ref{fig:pi}(b-c) and
for a fixed carrier density in Fig.~\ref{fig:pi}(d).
Fig. \ref{fig:pi}(b) shows
the inverse screening length of the Mexican hat dispersion
at 3 different temperatures with
the Fermi level fixed at $5$ meV above the band edge.
The low--temperature ($T=5$ K) curve has a strong $q$ dependence
that arises from the bandstructure.
There are two Fermi wavevectors for $0 < E_F < \epsilon_{0}$
denoted as $k_{F_1}$ and $k_{F_2}$ and
illustrated in the inset of Fig. \ref{fig:pi}(c).
The two Fermi wavevectors result in three features
for $\Pi (q) $ at $T=5$ K.
These features correspond to momentum transfers of
$ q = 2k_{F_1} $,
$ q = 2k_{F_2} $, and
$ q = k_{F_2} - k_{F_1}$.
Just as with the parabolic dispersion,
there is a sharp change in the derivative of $\Pi(q)$
when $q$ is twice the Fermi wavevector,
except now there are two Fermi wavevectors.
The third and largest peak occurs when $q = k_{F_2} - k_{F_1}$,
which is the minimum momentum required to
transfer between the two Fermi rings.
This can be viewed as a type of Fermi surface nesting.
Increasing the temperature smooths out these sharp features, and
at $T=300$ K, $\Pi(q)$ smoothly decreases with increasing $q$.
When the Fermi level is $3$ meV below the band edge as in Fig. \ref{fig:pi}(c),
the screening at $T=5$ K is essentially zero since there are no carriers,
and the qualitative features of the polarization functions at
$77$ K and and $300$ K are
the same as those in Fig. \ref{fig:pi}(b)
with a small reduction in the overall magnitude resulting from
the reduced carrier density.
Fig. \ref{fig:pi}(d) shows the inverse screening lengths at a fixed carrier
density of $ 10^{13}$ cm$^{-2}$ for different temperatures.
Now, the Fermi level moves with temperature as shown in the legend.
At $5$ K, the Fermi level is $1.6$ meV above the band edge, and
the small $q$ peak becomes very large as the Fermi level approaches
the $1/\sqrt{E}$ singularity in the density of states.
At $77$ K and $300$ K,
the Fermi levels are in the band gap, and
the polarization functions are similar to those in Fig. \ref{fig:pi}(c).
Now, we consider the magnitude and angle dependence
of the matrix elements $\bra{{\bf k}} \tilde{V} \ket{{\bf k}'}$
of the screened Coulomb potential,
given by Eq. (\ref{eq:coulomb_pot2})
with $q = |{\bf k} - {\bf k}'|$ and $z = z_0$.
Fig. \ref{fig:crt_exp} shows polar plots of the screened Coulomb potential
with $n_s = 10^{13}$ cm$^{-2}$ at two different temperatures and energies.
The polar angle $\theta$ is the angle between ${\bf k}$ and ${\bf k}'$.
The relevant ${q_\lambda}$ plots are shown in Fig. \ref{fig:pi}(d).
For a fixed energy,
scattering can occur within the inner ring (${\bf k}$ and ${\bf k}'$ both lie on the inner ring),
within the outer ring (${\bf k}$ and ${\bf k}'$ both lie on the outer ring),
or between the inner ring and the outer ring (${\bf k}$ and ${\bf k}'$ lie on different rings).
These 3 different matrix elements are denoted `Inner,' `Outer,' and `Inter,' respectively,
in Fig. \ref{fig:crt_exp}.
We first consider the low-temperature $T=5$ K matrix elements
at an energy of 2.5 meV above the band edge
shown in Fig. \ref{fig:crt_exp}(a).
At the carrier density of $n_s = 10^{13}$ cm$^{-2}$, $E_F = 1.6$ meV,
The wavevector dependent screening ${q_\lambda}$
corresponds to the upper curve in Fig. \ref{fig:pi}(d), and a
more detailed view is shown in Fig. \ref{fig:crt_exp}(c).
At $E = 2.5$ meV,
the radius of the inner ring $k_1 = 0.142 \: {\rm \AA}^{-1}$,
the radius of the outer ring $k_2 = 0.165 \: {\rm \AA}^{-1}$,
and $k_2 - k_1 = 0.023 \: {\rm \AA}^{-1}$.
At $\theta = 0^\circ$, $q=0$ for the inner and outer ring matrix elements
and $q=k_2 - k_1 = 0.023 \: {\rm \AA}^{-1}$ for the inter ring matrix element.
At $q=0$, ${q_\lambda} = 9.1 \: {\rm \AA}^{-1}$, and at $q=0.023 \:{\rm \AA}^{-1}$, ${q_\lambda} = 9.9 \:{\rm \AA}^{-1}$.
Thus, at $\theta = 0^\circ$, all three scattering mechanisms are strongly suppressed by
the screening.
The $\theta = 0^\circ$ inter ring scattering is a backscattering process,
since the two rings have opposite velocities.
Thus, the small $q$ inter ring backscattering is strongly suppressed by the screening.
The values of $q$, ${q_\lambda}$ and $q + {q_\lambda}$ are plotted in Fig. \ref{fig:crt_exp}(c).
The value of ${q_\lambda}$ in the range of $0 \leq q \leq 2 k_{F_2}$ is much larger than $q$.
This means that for $q \leq 2k_{F_2}$,
the $q$ dependence of $V(q)$ is determined solely by the $q$ dependence of the polarization,
and the bare momentum transfer $q$ is negligible in comparison.
Since ${q_\lambda}$ falls rapidly as $q$ increases, the RPA screened Coulomb potential
in a Mexican hat bandstructure favors large angle scattering.
This is opposite to the trend resulting from the bare $1/q$ Coulomb interaction.
The large outer-ring matrix elements for $\theta$ between $150^\circ$ and $210^\circ$
arise because the momentum transfer around the outer ring becomes larger than $2k_{F_2}$.
The kink at $120^\circ$ corresponds to the peak in ${q_\lambda}$ at $2k_{F_1}$.
At low temperature, the polarization strongly suppresses the magnitude of the
matrix elements at the Fermi level.
Only for those energies several $k_BT$ above the Fermi level can the momentum transfer become
large enough that the polarization becomes negligible, and $V(q)$ returns to a $1/q$ dependence.
This large momentum transfer corresponds to backscattering across the outer ring.
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{crt_exp_v6.eps}
\caption{
\label{fig:crt_exp}
Polar plots of the matrix elements of
the RPA screened Coulomb potential as a function
of scattering angle at (a) $T=5$ K and $E=2.5$ meV
and (b) $T=300$ K and $E=25$ meV.
The polar angle $\theta$ is the angle between ${\bf k}$ and ${\bf k}'$.
The legend refers to the 3 curves in each polar plot.
``Inner'' denotes matrix elements with ${\bf k}$ and ${\bf k}'$ both on the inner ring,
``Outer'' denotes matrix elements with ${\bf k}$ and ${\bf k}'$ both on the outer ring,
and
``Inter'' denotes matrix elements with ${\bf k}$ on the inner ring and ${\bf k}'$ on the outer ring.
(c) $q$, ${q_\lambda}$, and $q+{q_\lambda}$ as a function of $q$ corresponding to (a).
(d) $q$, ${q_\lambda}$, and $q+{q_\lambda}$ as a function of $q$ corresponding to (b).
The carrier density is fixed at $10^{13}$ cm$^{-2}$ for all figures.
}
\end{figure}
Fig. \ref{fig:crt_exp}(b) shows the $T=300$ K matrix elements at an
energy of 25 meV above the band edge.
As the temperature increases to 300 K, both the magnitude and the angular dependence
of the matrix elements change considerably compared to those at $T=5$ K.
This is a result of the large change in the polarization function as shown in
Fig. \ref{fig:pi}(d).
An enlarged view of the $T=300$ K ${q_\lambda}$ curve is shown in Fig. \ref{fig:crt_exp}(d).
The Fermi level now lies below the band edge at $E_F=-26$ meV.
Compared to the $T=5$ K polarization,
the magnitude of the polarization decreases by an order of magnitude at the bandedge,
the sharp features disappear, and ${q_\lambda}$ monotonically decreases as $q$ increases.
However, the overall decrease of $q + {q_\lambda}$ over the range of relevant $q$ values is relatively small.
At $E = 25$ meV, $k_1 = 0.11 \:{\rm \AA}^{-1}$ and $k_2 = 0.187 \:{\rm \AA}^{-1}$.
At $q=0$, ${q+q_\lambda} = 0.861 \:{\rm \AA}^{-1}$, and at $q = 2k_2 = 0.374 \:{\rm \AA}^{-1}$, ${q+q_\lambda} = 0.462 \:{\rm \AA}^{-1}$.
Thus, the maximum increase in the matrix element going from $\theta=0$ to $\theta = 180^\circ$
is a factor of $1.8$, which is shown for the matrix elements of the outer ring
in Fig. \ref{fig:crt_exp}(b).
Over the entire range of relevent momentum transfer $q$, the $T=300$ K
polarization is much less than the $T=5$ K polarization, so that the matrix elements are
uniformly larger at $T=300$ K compared to those at $T=5$ K.
Since the scattering rate is proportional to $|V(q)|^2$, the scattering rates will
be significantly larger at room temperature compared to those at low temperature.
The integrand that determines the momentum scattering rates at a given energy $E$,
given by Eq. (\ref{eq:tau}),
contains not only $|V(q)|^2$, but
also the final density of states and the relative change in the velocity which can be
positive or negative.
The $[1 - \frac{v'}{v}\cos(\theta_{{\bf v}, {\bf v}'})]$ term further reduces the small
angle intra-ring matrix elements, which are already small due to the large polarization at small $q$.
The integrand of Eq. (\ref{eq:tau}) is plotted in Fig. \ref{fig:S_int} at $T=300$ K, $E_F = -26$ meV,
and $E = 100$ meV.
Fig. \ref{fig:S_int}(a) shows the angle-dependent scattering rate for the initial $k$ on the inner ring,
and Fig. \ref{fig:S_int}(b) shows the angle-dependent scattering rate for the initial $k$ on the outer ring.
Note that the energy $E=100$ meV is $10$ meV below the top of the hat in Fig. \ref{fig:dispersion_dos}(a).
At this energy, the magnitude of the group velocity of a state on the inner ring is much less that
of a state on the outer ring.
For inter ring scattering from the inner ring to the outer ring, $v' > v$,
and a forward scattering process with $\theta_{{\bf v},{\bf v}'}=0$
causes the $[1 - \frac{v'}{v}\cos(\theta_{{\bf v}, {\bf v}'})]$ term in the integrand to become negative.
The forward scattering process with $\theta_{v,v'} = 0^\circ$ corresponds to
backscattering in $k$-space with $\theta = 180^\circ$, where
$\theta$ is the angle between the initial state $k$ on the inner ring
and the final state $k'$ on the outer ring.
Thus, in Fig. \ref{fig:S_int}(a),
the negative values of $1/\tau(\theta)$,
shown by the blue curve, are centered around $\theta = 180^\circ$.
Backscattering with $\theta_{v,v'} = 180^\circ$ corresponds to forward scattering
in $k$-space with $\theta = 0^\circ$, and the corresponding positive values of
$1/\tau(\theta)$ are shown by the red curve centered around $\theta = 0^\circ$.
When scattering from the outer ring to the inner ring, $v'/v < 1$, so that
$1/\tau(\theta)$ is positive for all angles as shown in Fig. \ref{fig:S_int}(b).
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{polar_intra_inter_v3.eps}
\caption{
Polar plots of $1/\tau(\theta)$ given by the integrand of Eq. (\ref{eq:tau})
for (a) $k$ on the inner ring and (b) $k$ on the outer ring.
Inter-ring and intra-ring rates are indicated by the legend.
Inter-ring contributions can be either positive or negative.
(c) Four components of the total scattering rate as a function of energy.
The components are indicated by the legends where,
for example, ``Inner-outer'' denotes the initial state on the inner ring and the
final state on the outer ring.
\label{fig:S_int}
}
\end{figure}
Fig. \ref{fig:S_int}(c) shows the 4 components of the total scattering
rate as a function of energy at $T=300$ K.
The energy $100$ meV corresponds to the polar plots shown in (a) and (b).
As the energy approaches the top of the hat, $110$ meV,
the radius $k_1$ of the inner ring goes to zero,
so that $q = |{\bf k}_2 - {\bf k}_1|$
becomes independent of $\theta$.
The denominator in Eq. (\ref{eq:coulomb_pot2}) is then independent of $\theta$,
the $\cos(\theta_{v,v})$ term integrates to zero,
and the integral over $\theta$ gives $2\pi$.
Thus, in the limit $E$ approaches ${\epsilon}_h$ from below,
the integral in Eq. (\ref{eq:tau}) can be performed analytically
for both inter-ring scattering and intra-ring scattering within the inner ring.
At $E={\epsilon}_h$, the single-spin density of states of both the inner ring and the outer ring
are equal to $\frac{m^*}{2\pi \hbar^2}$.
For inter-ring scattering, $q = k_2$, and the inter-ring scattering rate is
\begin{align}
\frac{1}{\tau_{\rm inter}}&= \frac{n_I e^4}{4 {\epsilon}^2 \hbar} \frac{m^*}{2 \pi \hbar^2}
\frac{ 2 \pi }{\left(k_2 + {q_\lambda}(k_2) \right)^2}
\nonumber \\
& =
4 \pi^2 \alpha^2 \frac{m_0 c^2}{\hbar} \frac{m_r}{{\epsilon}_r^2} \frac{ n_I }{\left(k_2 + {q_\lambda}(k_2) \right)^2}
\nonumber \\
& = 2.06 \times 10^{13} \; {\rm s^{-1}},
\label{eq:tau_eh}
\end{align}
where $k_2 = 2 \sqrt{2 m^* {\epsilon}_h}/ \hbar = 0.217 \: {\rm \AA}^{-1}$
and ${q_\lambda}(k_2) = 0.364 \: {\rm \AA}^{-1}$.
In the second line of Eq. (\ref{eq:tau_eh}), $\alpha$ is the fine structure constant,
$m_0$ is the bare electron mass, $c$ is the speed of light, $m_r = 0.409$ is the relative mass,
and ${\epsilon}_r = 3.1$ is the relative dielectric constant.
For intra-ring scattering within the inner ring,
$q \rightarrow 0$, and the intra-ring scattering rate becomes
\begin{equation}
\frac{1}{\tau_{\rm intra}} =
4 \pi^2 \alpha^2 \frac{m_0 c^2}{\hbar} \frac{m_r}{{\epsilon}_r^2} \frac{ n_I }{ q_{\lambda}^2(0) }
= 9.38 \times 10^{12} \; {\rm s^{-1}},
\label{eq:tau_eh_intra1}
\end{equation}
where ${q_\lambda}(0) = 0.861 \; {\rm \AA}^{-1}$.
The reduction of $1/\tau_{\rm intra}$ with respect to $1/\tau_{inter}$
is solely the result of the increased value of ${q_\lambda}$ as $q \rightarrow 0$.
The largest component to the total scattering rate is from scattering
within the outer ring.
Scattering within the outer ring allows for the largest momentum
transfer $q$ and thus the smallest values of ${q_\lambda}$.
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{crt_general_v5.eps}
\caption{
Momentum scattering rates for a charge density of $ 10^{13}$cm$^{-2}$ at
(a) $5$ K, (b) $77$ K and (c) $300$ K.
The band parameters correspond to the ones used in
Fig. \ref{fig:dispersion_dos}, and the polarization functions are shown in Fig.
\ref{fig:pi}(d).
(d) Maximum values of $q$ for scattering from the outer ring ($2k_2$)
or within the inner ring ($2k_1$) as a function of energy.
\label{fig:crt_gen}
}
\end{figure}
The total scattering rates for an initial state on the inner or the outer ring
are shown in Fig. \ref{fig:crt_gen} for the same
charge density ($10^{13}\;{\rm cm}^{-2}$) and temperatures as in Fig. \ref{fig:pi}(d).
The parameters are also the same as the ones used in the calculation of the screened
Coulomb matrix elements in Figs. \ref{fig:crt_exp} and \ref{fig:S_int}.
At $T=5$ K, as a result of the extremely large polarization,
the scattering rate is suppressed for energies below 14 meV.
At $E=14$ meV, $2k_2 = 0.36 \; {\rm \AA}^{-1}$.
At energies below 14 meV, the polarization is large for all possible
momentum transfer $q$, the matrix elements of the screened Coulomb potential
are reduced, and the scattering rate is reduced.
The low-energy minimum occurs at $E=2.5$ meV, when the minimum inter-ring scattering
momentum $q = 0.023 \; {\rm \AA}^{-1}$ is where the polarization function has its maximum value.
As the energy decreases below $2.5$ meV towards the band edge, the $1/\sqrt{E}$ density of states term
in Eq. (\ref{eq:tau}) takes over, and the rate increases as $E \rightarrow 0$.
For momentum transfer $q \gtrsim 0.36 \; {\rm \AA}^{-1}$, the polarization is negligible,
and the RPA screened potential reverts to the bare unscreened potential as shown in
Fig. \ref{fig:crt_exp}(c).
As the energy increases above 14 meV, unscreened backscattering takes place within the
outer ring.
The energy dependence for higher energies is governed by the energy dependence of the density of states
and the $1/q^2 \approx 1/4k_2^2$ dependence of the matrix element squared.
The radius $k_1$ of the inner ring is maximum at $E=0$ and decreases with increasing energy.
Thus, the polarization relevant to the inner ring matrix elements increases with energy,
causing the matrix elements to decrease.
The density of states monotonically decreases and the scattering rate for states on the inner ring
monotonically decreases with energy.
The total rate is dominated by the intra-ring scattering of the outer ring.
At $T=77$ K, the polarization loses its sharp features and its magnitude is everywhere reduced
causing an overall increase of the scattering rates
and a monotonic decrease with energy.
This trend is more pronounced at $T=300$ K where there is relatively little change in the sum
$q + {q_\lambda}$ over the range of relevant energies, and the energy dependencies of the rates
are determined by the $1/\sqrt{E}$ dependence of the density of states.
The total scattering rates for GaS and InSe are shown in Fig. \ref{fig:crt_mat}
for temperatures of $5$ K, $77$ K and $300$ K.
The temperature dependence of the overall magnitudes of the scattering rates
are determined by the magnitudes of the
matrix elements squared of the screened Coulomb potential, which, in turn, are determined by the temperature
dependence of the polarization functions as shown in Figs. \ref{fig:pi}(d) and \ref{fig:crt_exp}.
When the energy is equal to the height of the hat,
the contribution from the inner-ring scattering disappears giving an abrupt decrease
in the total scattering rate at $T=77$ K and $300$ K.
At $T=5$ K, the scattering rate from the inner ring is always small compared to that of the outer ring
(except right at the band edge), so that the small discontinuity at $E={\epsilon}_h$
is primarily the result of the disappearance of the inter-ring scattering from the outer ring
to the inner ring.
For energies above the top of the hat,
the rates become almost identical differing by at most a factor of 1.2 for InSe.
The fine differences result from the details of the different Fermi levels combined with the different thermal
broadening for each different temperature.
The large decrease in the $T=5$ K, low-energy
scattering rate for InSe compared to GaS is the result of the larger polarization
in InSe due to its larger mass and larger density of states.
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{crt_material.eps}
\caption{\label{fig:crt_mat}
Energy dependence of the total momentum relaxation rates
for (a) GaS and (b) InSe for 3 different temperatures.
The charge density is fixed at $10^{13}$cm$^{-2}$.
Parameters used for GaS and InSe are tabulated in Table \ref{tab:mat_params}.
}
\end{figure}
The temperature and charge density dependence of the mobility
are plotted in Fig. \ref{fig:mu_rpa}.
Both the temperature dependence and the density dependence of the mobility
are primarily governed by the temperature and density dependence of the polarization.
The initial decrease in mobility with temperature results from the decrease in
polarization with temperature as shown in Fig. \ref{fig:pi}(d).
The decrease in screening, increases the matrix element squared which increases the scattering rate
and decreases the mobility.
At $T=300$ K, there is a significant contribution to the integrand ($\mu(E)$) of Eq. (\ref{eq:mu}) from
energies above ${\epsilon}_h$.
Once $E = {\epsilon}_h$ starts to fall inside the thermal window defined by $-\partial f_0/\partial E$
in Eq. (\ref{eq:mu}), the sudden decrease in $1/\tau(E)$ shown in Fig. \ref{fig:crt_mat},
gives rise to a corresponding increase in $\mu(E)$, so that the integral begins
to increase with temperature.
The `turn-on' or `thermal activation' of the mobility starts to be seen at lower temperatures for
lower carrier densities as shown in Fig. \ref{fig:mu_rpa}(a).
For lower carrier densities, screening is less, the matrix elements and scattering rates are larger
at lower energies,
the low-energy values of $\mu(E)$ are reduced, and the discontinuity at $E={\epsilon}_h$ is larger
so that the higher energies give a disproportionally larger contribution to the mobility.
For a fixed temperature,
as the charge density increases, the screening increases, which reduces the matrix element squared
and the scattering rates and increases the mobility as seen in Fig. \ref{fig:mu_rpa}(b).
At a charge density of $5 \times 10^{13}$ cm$^{-2}$, the mobility is between $100 - 200$ cm$^2$/V$\cdot$s
for the 3 temperatures, 5 K, 77 K, and 300 K.
\begin{table}[b]
\centering
\begin{tabular}{c c c c}
\hline
\hline
\multirow{2}{*}{Material}\Tstrut &
\multirow{2}{*}{\begin{tabular}{@{}c@{}}Effective mass \\ m* (m$_0$)
\end{tabular}} & \multirow{2}{*}{\begin{tabular}{@{}c@{}}Height of the
hat\\ ($\epsilon_h$) (meV) \end{tabular}} & \multirow{2}{*}{\begin{tabular}{@{}c@{}}Relative\\ permittivity \end{tabular}} \\
& & \\ [0.5ex]
\hline
GaS\Tstrut & 0.409 & 111.2 & 3.10 \\ [0.5ex]
GaSe & 0.600 & 58.7 & 3.55 \\ [0.5ex]
InS & 0.746 & 100.6 & 3.08 \\ [0.5ex]
InSe & 0.926 & 34.9 & 3.38 \\ [0.5ex]
\hline
\hline
\end{tabular}
\caption{Effective mass and height of the hat for III-VI materials with
Mexican hat \cite{darshanamhat}.}
\label{tab:mat_params}
\end{table}
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{material_rpa.eps}
\caption{
\label{fig:mu_rpa}
Charged impurity limited hole mobility of GaS
(a) as a function of temperature for a carrier density of
10$^{12}$ cm$^{-2}$ (blue) and 10$^{13}$ cm$^{-2}$ (red) and
(b) as a function of carrier density at $5$ K (blue), $77$ K (red) and
$300$ K (green) for a fixed charge density of 10$^{13}$ cm$^{-2}$.
The impurity density $n_I$ is fixed at $10^{12}$ cm$^{-2}$.
}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{mobility_thickness_v3.eps}
\caption{
(a) Charged impurity limited monolayer hole mobility as a function of
temperature of
GaS, GaSe, InS and InSe with the charged impurities
in the middle of the channel ($d = 0$, dashed line)
and on the substrate ($d=5.5 \; {\rm \AA}$, solid line).
(b) $qe^{qd}$, ${q_\lambda}$, $q + {q_\lambda}$, and $qe^{qd} + {q_\lambda}$ for GaS at $T=77$ K
where $d = 5.5$ \AA.
The hole density $n_s = 10^{13}$ cm$^{-2}$, and
the charged impurity density $n_I$ is fixed at $10^{12}$ cm$^{-2}$.
\label{fig:mu_vs_T}
}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth,keepaspectratio]{mobility_cd_sweep.eps}
\caption{
Charged impurity limited monolayer hole mobilities as a function of carrier density at
(a) $T=77$ K and (b) $T=300$ K for the 4 III-VI materials as indicated by the legends.
Solid lines result from charged impurities on the substrate ($z_0 = 5$ {\AA}),
and the dashed lines result from charged impurities in the middle of the channel
($z_0 = 0$).
The impurity density $n_I$ is fixed at $10^{12}$ cm$^{-2}$.
\label{fig:mobility_vs_density}
}
\end{figure}
The temperature dependence of the 4 III-VI p-type materials are shown in Fig. (\ref{fig:mu_vs_T})
for a fixed hole density of $10^{13}$ cm$^{-2}$ and two different positions of the
charged impurities, in the middle of the channel ($z_0 = 0 $\AA) and on the substrate ($z_0 = 5.5$ \AA).
The relevant material parameters are given in Table \ref{tab:mat_params}.
The general trends of the temperature dependence follow those seen in Fig. \ref{fig:mu_rpa}.
The low temperature mobilities order according to the effective masses with the lower masses
correlating with the higher mobilities.
However, the dependence is weaker than a $1/m^*$ dependence.
The minimum and maximum effective mass differ by a factor of 2.3, and
at $T=5$ K, the mobilities differ by a factor of 1.4.
The difference in mobilities increases to a maximum of 2 near the beginning
of the high-temperature crossover where the mobilities start to increase.
The cross-over begins at a lower temperatures for the materials with a smaller
value of ${\epsilon}_h$, since lower temperatures can thermally excite carriers above the top of the hat.
Moving the charged impurities from the middle of the channel to the substrate
increases the mobility, as would be expected, since the charged impurities are further
away from the carriers.
However, it also lowers the temperature of the high-temperature crossover, which is not an
obvious consequence.
The reason lies in the large enhancement of the bare, large-wavevector screening as shown in
Fig. \ref{fig:mu_vs_T}(b) for GaS at $T=77$ K.
For GaS, the bare term $qe^{qd}$ in the denominator becomes larger than ${q_\lambda}$
at $q=0.22 \; {\rm \AA}^{-1}$.
The minimum value of $2k_2$ is $0.31 \; {\rm \AA}^{-1}$ at the band edge,
and at the top of the hat, $2k_2 = 0.43 \; {\rm \AA}^{-1}$.
At that value of $q$, the denominator $qe^{qd} + {q_\lambda}$ is larger
than at $q=0$, so that backscattering
across the outer ring is strongly suppressed
giving a large enhancement to $\mu(E)$ for energies $E={\epsilon}_h$.
The hole density dependence of the charged impurity limited
mobility at $T=77$ K and $T=300$ K is shown in Fig. (\ref{fig:mobility_vs_density}).
The mobility monotonically increases with hole density $p_s$ for
a fixed charged impurity density $n_I$.
This trend would be expected due to increased screening resulting from the higher hole density.
At the highest hole densities considered of $3 \times 10^{13}$ cm$^{-2}$, with the
charged impurities on the substrate, the
$T = 300$ K mobilities lie between 500 and 800 cm$^2$/V$\cdot$s for all 4 materials.
With the impurities at the center of the channel, the mobilities decrease
one order of magnitude and lie in the range of 50 to 80 cm$^2$/V$\cdot$s.
All mobilities are calculated for a charged impurity density of $n_I = 10^{12}$ cm$^{-2}$,
and the mobilities are inversely proportional to $n_I$, so that all mobility values shown
can be easily scaled for arbitrary values of $n_I$.
\section{Conclusions}
The Mexican hat type bandstructure that occurs in the valence band of monolayer
and few layer III-VI materials and other 2D materials
gives rise to unique screening properties.
The singular density of states
at the band edge and the two Fermi wavevectors up to the height of the hat,
lead to large screening and strong wavevector dependence of the screening.
The wavevector dependence of the screened Coulomb interaction is so strong, that
the temperature and density dependence of the matrix element squared
is the dominant factor determining the overall trends with respect
to temperature and density.
The reduction of polarization with temperature causes an initial increase
in scattering and decrease in mobility with increasing temperature.
Short wavevector inter-ring backscattering and scattering
within the smaller ring is always suppressed by the
large polarization at small $q$.
When the the charged impurities lie in the middle of the 2D channel,
the wavevector dependence of the polarization
favors large wavevector backscattering across the outer ring.
When the charged impurities lie on the substrate, the bare
screening increases rapidly at larger wavevectors suppressing
the backscattering within the outer ring.
For charged impurities on the substrate, the polarization suppresses
the small wavevector scattering and the exponential wavevector dependence of the bare
Coulomb interaction suppresses the large wavevector scattering across the outer ring
leading to an overall increase in mobility.
The suppression of the large wavevector scattering also
reduces the temperature at which the mobility starts to increase when the charged
impurities are on the substrate.
The mobility monotonically increases with hole density up to the maximum
value considered of $3 \times 10^{13}$ cm$^{-2}$ where it reaches a maximum
value of 800 cm$^2$/V$\cdot$s for GaSe with the charged impurities located on the substrate.
Placing the impurities in the center of the channel reduces the maximum value by
an order of magnitude.
All mobility values are calculated for a charged impurity density of $n_I = 10^{12}$
cm$^{-2}$ and scale inversely proportionally to $n_I$.
\noindent
\begin{acknowledgements}
We acknowledge helpful discussion with Dr. Yafis Barlas.
This work was supported by FAME, one of six centers of STARnet,
a Semiconductor Research Corporation program sponsored by MARCO and DARPA.
This work used the Extreme Science and Engineering Discovery
Environment (XSEDE), which is supported by National
Science Foundation Grant No. ACI-1548562 and allocation
ID TG-DMR130081.
\end{acknowledgements}
|
\section{Introduction}
Following the goal of establishing a complete parameterization of cluster sink-strengths (CSS) for any type of clusters mobilities including both species mixed mobilities (i.e. rotations of glide directions, noted $1DR-1DR$), we have established in a companion paper \cite{Adjanor1} analytical expressions for $1D-1D$ absorption including the diffusion anisotropy analog case (diffusion coefficient ratio being different than one). This task is essential to a rate-equation cluster dynamics (RECD) modeling of microstructure evolution of defect clusters populations according to the state of knowledge on loops mobilities, which are of primary importance regarding material properties evolutions in both fission and fusion reactors, but also nucleation and growth processes at thermal equilibrium.
In section \ref{Results}, the effective CSS are estimated by OKMC for the most general case of $1DR-1DR$ CSS with finite rotation energies. Estimates from simulation are required because no general analytical formula is at hand. This is done for a very large set of conditions in terms of D-ratio (${\mathcal D}=D_{B}/D_{A}$) and of rotation energies $E_{A}$, $E_{B}$ couples. The evolution of CSS with these three parameters is then rationalized by proposing a semi-analytical formula, matching the analytical formulas for the limiting cases of $3D_{A}-3D_{B}$, $1D_{A}-3D_{B}$ and $1D_{A}-1D_{B}$ CSS and fitting the transition between them thanks to a combination of sigmoid-type functions. The fit is made on one couple of concentrations $(C_{A}, C_{B})$, and reproduces with similar levels of accuracy other couples of concentrations, thus assessing for the broad validity of this semi-analytical formula.
Given its good ability to reproduce the effective CSS over the many orders of magnitude over which they evolve, the agreement can be considered as very reasonable. The interpretation of the CSS evolution heavily relies on the D-ratio exponents: as highlighted on limiting cases when the D-ratio varies, the CSS mostly varies like $(D_{A}/D_{B})^\delta$, the exponent being characteristic of types of the mobilities. A map of effective $\delta$ values helps the interpretations and the precision of semi-analytical formula heavily relies on its main trends. On this map, domains of different mobility dimensionality clearly appear, and their characteristic exponents are connected with the limiting cases for $1D_{A}-1D_{B}$ and $1D_{A}-3D_{B}$ anisotropic analog of the CSS ($(D_{A} \ne D_{B}$). But in the limited range of D-ratios investigated for the purpose of high precision fitting, these domains do not close: even for quite large D-ratios, the fixed sink limit is not reached. At section \ref{Closure}, we investigate the closure of these domains, i.e. the convergence of effective CSS towards the analytical CSS with respect to a fixed sink when the D-ratio gets lower ($\Delta=\log_{10}(D_A/D_B)$ higher). As we shall see, it is only when reaching much smaller D-ratios than for our typical set of conditions that the slowest specie can be considered as immobile regarding CSS expressions. As discussed at section \ref{Discussion} and validated at section \ref{sectionCD} by comparing RECD to extensive OKMC simulation of complete microstructure evolutions, this implies a quite broad relevance of the established general CSS expression, even when the effective mobilities of species are considerably lowered after trapping by impurities or elastic fields. A more general consequence of this, the prevalence of aggregation (termed here as growth by mutual mobility) over Ostwald ripening is also discussed.
\section{Method and Results} \label{Results}
When dealing with 1D mobilities, the convergence of CSS estimates is known to be very slow \cite{Malerba}. In the Annex \ref{General}, we detail the OKMC simulation procedure for such adequate calculation, and we also show in the Annex \ref{Search} that the conditions for the true convergence of object kinetic Monte-Carlo (OKMC) CSS estimates are the extremely delicate and demanding. Nevertheless, thanks to validated simple models, the conditions for estimates convergence can be established using three criteria: the first criterion states that lowest number species ($Min(N_A,N_B)$) should be greater than the smallest box dimension in unit cells, the second criterion states that the number of reactions to be performed should be at least equal to the total number of species $N_{A}+N_{B}$ and the last criterion is that a few tens of CSS estimates should be used to assess their standard deviations to mean ratio.
In the calculations presented in the rest of this article, the three preceding convergence criteria are met, so the computed CSS are now referred as "effective CSS" $\kappa^2_{\text{eff}}$ instead of estimates of the CSS.
First, we present results for the $C_{A}=C_{B}$ for a few $(E_{A},E_{B})$ combinations, in order to establish trends.
\begin{figure}
\includegraphics[width=0.8\textwidth]{k2_est.jpg}
\caption{Logarithm of the effective CSS $\kappa^2_{\text {eff}}$ in $\si{cm^{-2}}$ as a function of the radius and of $\Delta=\log_{10}(D_A/D_B)$ for the conditions $(E_A=2, E_B=2)\ (\si{eV})$ and $C_A=C_B$ ranging from $\num{3.4e+16}$,$\num{6.8e+16}$, $ \num{1.7e+17}$, $\num{3.4e+17}$, $\num{5.1e+17}$, $\num{6.8e+17}$ $\si{cm^{-3}}$ (from blue to red surfaces). }
\label{k2_EA@2.0_EB@2.0}
\end{figure}
Figure \ref{k2_EA@2.0_EB@2.0} presents effective CSS computed from simulation with $(E_{A}=2\ eV,E_{B}=2\ eV)$ at various concentrations. With both rotation energies as high as $2\ eV$ at $573\ K$, the particles will have a very large mean-free path before rotation compared to box size, so they can safely be treated as purely 1D-mobile and thus comparison with the analytical $\kappa^2_{1D-1D}$ is relevant.
For each $R$ value, the linear relation between $\Delta=\log_{10}(D_{A}/D_{B})$ and $\log_{10}(\kappa^2_{\text{eff}})$ on the figure indeed suggests a power-law dependency of the CSS on the D-ratio. Detailed analysis indeed shows that it follows the expect characteristic exponent $-1/3$ excepted from the diffusion anisotropy analogy highlighted in the companion paper.
The evolution of CSS surfaces with $R$ are clearly more complex, and changing the representation to $\log_{10}(R)$ would not reveal a simple and general scaling. Again, this is not surprising, as according to Eq. \ref{k2_CaCb}, the CSS should depend on $R$ through ${\overline{R_{\text{eff}}}}$ and the inverse logarithm of $R^3$.
From the preceding part, we can conclude that the proposed analytical expression match well the computed effective CSS for the $1D-1D$ case. Other limiting cases like $1D-3D$ and $3D-3D$ (rotation energy couples set to $(E_{A}=0\ eV,E_{B}=2\ eV)$ and $(E_{A}=0\ eV,E_{B}=0\ eV)$ respectively) were checked to follow their related limiting case CSS expressions, although for the sake of conciseness they corresponding graphs are not shown here. However, this approach reaches its limits when dealing with intermediate values of rotation energies which do not correspond to a well defined limiting case: with a direct plot of a set of CSS surfaces depending on D-ratio and $R$, it can be difficult to determine which CSS analytical expression is the best match. Moreover, pursing the approach of estimating CSS for many different couples of concentrations, there might be few hope of rationalizing the results: concentrations relevant to typical irradiation or nucleation conditions span over too much order of magnitudes.
We will thus adopt a different approach in the next sections.
\subsection{Semi-analytical expression for general sink-strengths} \label{semiAnalytical}
We will now focus on establishing a semi-analytical expression of CSS for the general case of $(E_{A}, E_{B}) \in [0,\ 2]\times[0,\ 2]$ ($\si{eV}$).
and reproducing the identified analytical expressions for limiting cases. The functions allowing for the transition between the limiting cases will be fitted on simulation results which makes the approach not fully analytical but only semi-analytical. For the purpose of this fitting procedure, a first set of $1728$ conditions has been simulated. It corresponds to the following parameters ranges:
\begin{eqnarray}
&&C_{A}=C_{B}=\num{1e17} \si{cm^{-3}}, \notag\\
&&R_{A}=R_{B}=0.5\ \si{nm},\notag\\
&&(E_{A}, E_{B}) \in \{0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 2.0\}^2 (\si{eV}).\notag\\
&&D_{A}+D_{B}=\num{3.12e-5} \si{cm^{2} s^{-1}}, \notag\\
&&\Delta=-\log_{10}({\mathcal D}) \in \{0, 0.27, 0.54, 0.81, 1.09, 1.36, 1.63, 1.90, 2.18, 2.45, 2.72, 3\}. \notag\\
\end{eqnarray}
Note that these parameters have been chosen with some guidance from the parameterization in table \ref{tableMalerba} from the Annex, but that this choice does not impact the generality of the results to come, thanks their semi-analytic character.
One advantage on having the condition $C_{A}=C_{B}$ fixed to a not too low density is that it eases the search for the optimal simulation parameters (such as internal variables for OKMC's "link-cell" type neighbor finding algorithm, box size, single run duration etc.) valid for the whole set of conditions. Thus, we could obtain coefficients of variation ${\sigma(\kappa^2_{est})}/{\overline{(\kappa^2_{est})}}$ of $1\%$ on average over all the $1728$ conditions and about $5\%$ in the worse cases. The average number of estimates (runs with different initial placements of defects) is equal to ten.
\begin{figure}
\includegraphics[width=0.8\textwidth]{cube-montage_NEW.jpg}
\includegraphics[width=0.8\textwidth]{cube-montage_FIT.jpg}
\caption{(a) two rotated views of the same $\log_{10}(\kappa^2_{\text{eff}})$ (the CSS being in $cm^{-2}$) isosurfaces in the $(E_A, E_B, \Delta=\log_{10}(D_{A}/D_{B}))$ space. (b) same views of isosurfaces obtained from the semi-analytical formula Eq. \ref{semiEmpirical}.}
\label{cube1}
\end{figure}
The results are shown on the top panel of Fig. \ref{cube1}. They are displayed in the form of CSS-isosurfaces (represented using the Mayavi library \cite{Mayavi}). The overall shape is quite complex but we can nevertheless draw some trends. First we see that for values of A species' rotation energy $E_{A}$ close to zero, the CSS reaches the highest values. This is also quite independent of both $E_{B}$ and the diffusion coefficients ratio: the corresponding isosurfaces are almost flat and parallel to the $E_{A}=0$ plane. This is because when $D_{A}$ is greater $D_{B}$, if $A$ is purely or almost purely 3D-mobile, the type of mobility of $B$ has very little influence on absorption probability, the overwhelming efficiency of 3D-mobility will completely dominate.
Another case is when $D_{B}$ is equal or very close to $D_{A}$, $E_{A}$ and $E_{B}$ should play symmetric roles for the CSS. This is indeed what is observed in the Fig. \ref{k2_DA-DA}. The hyperbolic-like shape of the contour lines can be simply interpreted by the fact that because $D_A=D_B$, for $E_A$ fixed, the CSS depend very weakly on $E_{B}$, so iso-CSS lines should be parallel to $E_{B}$ (and conversely for $E_B$ fixed). Also, we note that globally, whatever the couple $(E_{A},E_{B})$ values the CSS varies quite weakly in this isotropic diffusion analog case $D_A=D_B$. Nevertheless, it will be important to reproduce these hyperbolic-like shapes, because by extrusion and non-uniform shear along the $z$-axis they generate the complex isosurfaces of Fig. \ref{cube1}.
\begin{figure}
\includegraphics[width=0.8\textwidth]{k2_DA-DA.jpg}
\caption{Contours of the decimal logarithm of effective CSS in $\si{cm^{-3}}$ for $D_A=D_B$ in the $(E_A,E_B)$ plane.}
\label{k2_DA-DA}
\end{figure}
Another salient feature of Fig. \ref{cube1} is that a series of isosurfaces parts are almost identical up to a constant translation along the $z$-axis. This is particularly striking for large values of $E_{A}$. In this logarithmic representation, the constant spacing along $\Delta$ between isosurfaces is the signature of the power-law dependencies highlighted in the limiting cases. This clearly holds for specific parts of the $(E_A,E_B,\Delta)$-parameter space: the plateau values reached by the surfaces notably when both $E_{A}$ and $E_{B}$ are large suggest a quite wide range of validity of limiting case analytical formulas. Between these plateau regions there are very clear transition zones where isosurfaces have sigmoid shapes.
All these observations will now guide us to build a semi-analytical CSS expression for any couple of rotation energies, whose transition coefficients will be adjusted on the present data set. Hereafter, we will use the well known "sigmoid" function (also known as the logistic function):
\begin{equation}
\sigma(\lambda, \varepsilon, x)=\frac{1}{1+\exp\left[-\lambda(x- \varepsilon)\right]}.
\end{equation}
As sketched on Fig. \ref{delta-scheme0}, when $D_{A}=D_{B}$, one very general formulation of the CSS may be:
\begin{align*}
&\kappa^2\left(E_{A}, E_{B}, D_{A}, D_{A}\right) = \notag\\
&f(E_{A})\left( \left( 1-g(E_{B}) \right) \kappa^2_{1D-3D} + g(E_{B}) \kappa^2_{1D-1D} \right) \notag\\
&+ \left(1- f'(E_{A})\right) \notag\\
&\times\left( \left( 1-g'(E_{B})\right) \kappa^2_{3D-3D} + g'(E_{B}) \kappa^2_{3D-1D} \right)\label{MostGeneralKappa}\\
\end{align*}
with the functions $f$, $f'$, $g$ and $g'$ having the constraints: $f(0)=f'(0)=g(0)=g'(0)=0$ and $f(1)=f'(1)=g(1)=g'(1)=1$ so that the four limiting cases are recovered:
\begin{eqnarray}
&&\kappa^2\left(0, 0, D_{A}, D_{A}\right) = \kappa^2_{3D-3D}\left(D_{A}, D_{A}\right) \notag\\
&&\kappa^2\left(0, \infty, D_{A}, D_{A}\right) = \kappa^2_{3D-1D}\left(D_{A}, D_{A}\right) \notag\\
&&\kappa^2\left(\infty, 0, D_{A}, D_{A}\right) = \kappa^2_{1D-3D}\left(D_{A}, D_{A}\right) \notag\\
&&\kappa^2\left(\infty, \infty, D_{A}, D_{A}\right) = \kappa^2_{1D-1D}\left(D_{A}, D_{A}\right) \notag
\end{eqnarray}
\begin{figure}
\includegraphics[width=0.8\textwidth]{delta-scheme0.jpg}
\caption{Schematic view of the relation between the fitting function $f$, $f'$, $g$ and $g'$ and the limiting cases for the CSS when $D_{A}=D_{B}$ in the $(E_{A},E_{B})$ plane.}
\label{delta-scheme0}
\end{figure}
If we impose that $E_{A}$ and $E_{B}$ play symmetric roles ($\kappa^2\left(E_{A}, E_{B}, D_{A}, D_{A}\right) =\kappa^2\left(E_{B}, E_{A}, D_{A}, D_{A}\right)$) and use $\kappa^2_{3D-1D}\left(D_{A}, D_{A}\right)=\kappa^2_{1D-3D}\left(D_{A}, D_{A}\right)$, then we must have $f=g$ and $f'=g'$ and we can cast the CSS to the following form:
\begin{eqnarray}
\kappa^2\left(E_{A}, E_{B}, D_{A}, D_{A}\right) &= & f(E_{A})\left( \left( 1-f(E_{B}) \right) \kappa^2_{1D-3D} + f(E_{B}) \kappa^2_{1D-1D} \right) \notag\\
&+&\left( 1-f'(E_{A}) \right)\left( \left( 1-f'(E_{B})\right) \kappa^2_{3D-3D} + f'(E_{B}) \kappa^2_{1D-3D} \right)\label{generalKappa}
\end{eqnarray}
In practice, experimenting various fitting possibilities, it turns out that we can make this form even simpler taking:
\begin{eqnarray}
f'(x)=f(x) = \sigma(\lambda_f,\varepsilon_f,x) \\
\end{eqnarray}
and a simple fitting procedure leads to $\lambda_f=8$ and $\varepsilon_f=0.2$, with overall 5\% discrepancy with the simulated data set and a maximum deviation of about 15\% for the worse point.
Following the previous findings, we will now assume that the fit can be extended to $D_{A}>D_{B}$ scaling it with the ratio of diffusion coefficients to a power that depends of the dimensionality of the mobilities:
\begin{eqnarray}
\kappa^2\left(E_{A}, E_{B}, D_{A}, D_{B}\right) &=& \kappa^2\left(E_{A}, E_{B}, D_{A}, D_{A}\right) \\
& & \times \left(\frac{D_{A}}{D_{B}}\right)^{\delta(E_{A}, E_{B})}
\end{eqnarray}
We are now left with the task of defining the exponent function $\delta$. The values of delta that should be matched according to the analog limiting cases are sketched on Fig. \ref{delta-scheme}:
\begin{itemize}
\item $(E_{A}=0,\ E_{B}=0) \ (\si{eV})$ corresponds to $3D-3D$ absorption rates which simply depend on $(D_{A}+D_{B})$, that makes $(1+(D_{A}/D_{B})^{-1})$ when factoring for $D_B$ the CSS, which is close to $1=(D_{A}/D_{B})^0$ when $D_{A} \gg D_{B}$ in other words a close to zero $\delta$ value,
\item $(E_{A}=2,\ E_{B}=2)$ corresponds to $1D-1D$ absorption rates with a characteristic exponent of $-1/3$, consistently with the related limiting case \cite{Adjanor1},
\item $(E_{A}=2,\ E_{B}=0)$ similarly corresponds to $1D-3D$ with $D_{A}>D_{B}$ absorption rates leading to a $-1/2$ exponent \cite{Adjanor1}
\begin{figure}
\includegraphics[width=0.8\textwidth]{delta-scheme.jpg}
\caption{Schematic view of the analytical expressions for the limiting cases (corresponding the evolution of CSS along the vertical edges of the cube) for the CSS in the $(E_{A},E_{B}, \Delta)$ space. The functions $f$, $f'$, $g$ and $g'$ were shown to be identical by symmetry and practical arguments.}
\label{delta-scheme}
\end{figure}
\end{itemize}
One very simple choice for an exponent function $\delta$ meeting these requirements is:
\begin{equation}
\delta(E_{A}, E_{B}) = \sigma(\varepsilon,\lambda, E_{A}) \left(-\frac{1}{2}+\frac{1}{6} \sigma(\varepsilon',\lambda', E_{B})\right). \label{delta}
\end{equation}
The first sigmoid multiplies the whole expression allowing to reproduce a smooth transition from the 3D to 1D mobility of A species, which should indeed be in the leading term since they have the largest diffusion coefficient. The second sigmoid just accounts for the behavior of the B-species which are less mobile, as reflected by the weaker influence of this term.
Performing a basic fitting procedure for this function $\delta$ leads to $\lambda=10$ and $ \varepsilon=0.5$ for both sigmoid functions.
This simple procedure leads to quite a reasonable fit of the simulation data as they are reproduced with about 16\% discrepancy on average, with a maximum error of about a factor $2.5$, while the whole set of data varies about a factor $50$. Depending on the application such a precision might be sufficient or not. For example for very precise computation of dislocation loops nucleation out of irradiation, all the errors on the estimated absorption rates for monomers up to critical size clusters will cumulate and may have an impact on the overall estimated nucleation rates. On the other hand, for microstructural evolution under irradiation, where the cascade cluster production may allow clusters to grow bypassing the classical nucleation path and where, due to colossal supersaturations, the expected critical radii may be extremely small, then reproducing the order of magnitude of CSS may be seen as a reasonable approximation when the purpose is to compare RECD calculation to experimental observations often having even larger intrinsic uncertainties.
\begin{figure}
\includegraphics[width=0.8\textwidth]{delta-eff.jpg}
\caption{Effective exponents $\delta_{\text{eff}}$ isosurfaces, allowing to identify CSS domains.}
\label{delta-eff}
\end{figure}
Representing the actual map of effective $\delta$ values,
\begin{equation}
\delta_{\text{eff}}\left(E_A,E_B,\log_{10} \left( \frac{D_A}{D_B}\right)\right) = {\frac{\partial \log_{10} \left( \kappa^2_{\text{eff}}\right)}{\partial \log_{10} \left( \frac{D_A}{D_B}\right)}}
\end{equation}
from the set of estimated CSS shows us that trying to improve the fit would require a more complex expression of the function $\delta$ Eq. \ref{delta}. This is shown on Fig. \ref{delta-eff}. The figure may also be seen as "reaction dimensionality diagram", as it permits, basing on the values of the exponent, to identify zones in the $(E_{A}, E_{B}, \Delta)$-space associated with the 1D-1D type reaction (the zone delimited by the purple surface on the figure), the 1D-3D zone (the blue zone), and the rest which can be seen as a close to 3D-3D zone, apart from a small zone where $D_{A} \simeq D_{B}$, $E_{A}\simeq 0$, $E_{B} \simeq 2eV$ (in red on the figure) where the exponent approaches $1/6$. Due to its very small extend in the $\Delta$-dimension this red zone wasn't identified so far in the analysis of limiting cases, but it can be interpreted considering the analog cases highlighted in \cite{Adjanor1}
\begin{flalign}\label{anisotropicCSS}
&\frac{\partial C_{A}}{\partial t}= - 8 R {\overline D} \left( \frac{D_z} {D_{\rho}} \right)^{1/6} C_{A} C_{B}, \text{for $D_z \gg D_{\rho}$}
\end{flalign}
which according to the authors of this development \cite{GoseleSeeger,Woo2} should be rightful for $D_z \gg D_{\rho}$.
Using the analogy approach of the companion paper, we may take $D_z=D_A+D_B$ and $D_{\rho}=D_{B}$. Actually, the red zone is for quite small $\Delta$ values so, if we admit that the relation also holds for moderate diffusion "anisotropy" ($D_z \gtrsim D_{\rho}$) then, the characteristic exponent $1/6$ is explained. The fact that this zone actually has a very limited range and does not extend to larger $\Delta$ values is just due to the chosen representation: because the CSS are studied with $(D_A+D_B)$ constant, $D_B$ vanishes along the $z$-axis and the evolution is then only driven by $E_A \simeq 0$ and its $\delta=0$ typical exponent.
\begin{table}
\begin{tabular}{|c|c|c|c|c|}
\hline
$(C_{A},C_{B}) $ & $\mathcal{C}$ & $\varsigma$ & $M$ & $Max({\kappa^2_{\text{eff}}})$ \\
$(\si{cm^{-3}})$ & Eq.\ref{correlation} & Eq.\ref{stddev-log}& Eq. \ref{maxes} & $/Min({\kappa^2_{\text{eff}}})$ \\
& & & & \\
\hline
$(\num{1e17},\num{1e17})$ & $0.98$ & $0.087$ & 2.37 & 49 \\
\hline
$(\num{5e16},\num{5e16})$ & $0.97$ & $0.080$ & 2.38 & 21 \\
\hline
$(\num{5e17},\num{5e17})$ & $0.98$ & $0.078$ & 2.36 & 16 \\
\hline
$(\num{5e15},\num{5e15})$ & $0.97$ & $0.095$ & 2.37 & 29 \\
\hline
$(\num{2e15},\num{2e15})$ & $0.97$ & $0.105$ & 2.37 & 31 \\
\hline
$(\num{2e15},\num{1e16})$ & $0.97$ & $0.097$ & 2.39 & 25 \\
\hline
$(\num{5e16},\num{2e15})$ & $0.98$ & $0.091$ & 2.35 & 19 \\
\hline
\end{tabular}
\caption{Correlation coefficients $\mathcal{C}$ for $\log_{10}(\kappa^2_{\text{eff}})$ versus $\log_{10}(\kappa^2_{\text{fit}})$ and standard deviations $\varsigma$ of the residual of the logarithmic CSS Eq. \ref{stddev-log}. The next column is the maximum discrepancy, $M=Max({\kappa^2_{\text{eff}}/\kappa^2_{\text{fit}}})$ over the validation data sets. For comparison the largest ratio of effective CSS $Max({\kappa^2_{\text{eff}}})/Min({\kappa^2_{\text{eff}}})$ is given on the last column for each data set. The total radius used is always $1\ \si{nm}$ and the box dimensions range from 300 to 4000 unit lattices depending on concentrations and $L_{min}$.\label{validationSet}}
\end{table}
We will now validate the semi-analytical formula outside of its fitting data set.
Due to high computational cost \cite{Note5} of well converged CSS for high rotation energies and low concentrations, only a limited set of concentration pairs could be used. However, we tested the formula both for equal concentrations and different ones. Results are presented on table \ref{validationSet}.
We note that the correlation coefficient for $\log_{10}(\kappa^2_{\text{eff}})$ versus $\log_{10}(\kappa^2_{\text{fit}})$
\begin{equation} \label{correlation}
\mathcal{C} = Corr(\log_{10}(\kappa^2_{\text{eff}}), \log_{10}(\kappa^2_{\text{fit}}))
\end{equation}
are very close to one. These correlations being on decimal logarithms, it shows that, at least in terms of orders of magnitude, the formula capture almost perfectly the CSS evolution over the parameter ranges.
The standard deviations on logarithms are:
\begin{equation} \label{stddev-log}
\varsigma = \left(\frac{1}{n}\sum^{n}\left(\log_{10}(\kappa^2_{\text{fit}})-\log_{10}(\kappa^2_{\text{eff}})\right)^2\right)^{1/2}
\end{equation}
and they are all of the same order ($0.09$) for all validation sets. This gives an estimate of the "average error" (coefficient of variation) of the semi-analytical formula: about 20\%, as $10^{0.09}\simeq1.2$. Globally, this is a quite good validation of the semi-analytical fit. At first sight, the quite large maximum discrepancies
\begin{equation} \label{maxes}
M=Max\left(\frac{\kappa^2_{\text{eff}}}{\kappa^2_{\text{fit}}},\frac{\kappa^2_{\text{fit}}}{\kappa^2_{\text{eff}}}\right)
\end{equation}
displayed on table \ref{validationSet} could be considered as a source of inaccuracy. To that concern first, it should be noted that these large deviations mostly correspond to intermediate values of rotation energies, as can be seen from the isosurfaces of the residuals on Fig. \ref{residue_isosurfaces}. We can see that the residuals of logarithms
\begin{equation}
r=\left[ \log_{10}(\kappa^2_{\text{eff}})-\log_{10}(\kappa^2_{\text{fit}}) \right]^2 \label{logarithmic_residue}
\end{equation}
is everywhere very close to $0$ (blank zones are below $0.00746$, whose square root elevated to power ten corresponds to values below the 20\% average discrepancy) except when $E_{A}\simeq 0.5\ eV$, $E_{B}\simeq 0.0\ eV$ and $\Delta \gtrsim 2$. In that very small orange zone, it reaches $0.14$ whose square root corresponds to the logarithm of the maximum discrepancy $M$ for the first condition in table \ref{validationSet} ($\log_{10}(2.37)\simeq\sqrt{0.14}$): the zones of quite high deviation from the fit are always quite small and actually correspond to transition zones.
\begin{figure}
\includegraphics[width=0.8\textwidth]{residue_isosurfaces.jpg}
\caption{Isosurfaces of residuals of logarithms from Eq. \ref{logarithmic_residue} for the reference conditions ($T=573\ K, C_{A}=C_{B}=\num{1e17} \si{cm^{-3}}, R_{A}=R_{B}=1 \si{nm}$)}
\label{residue_isosurfaces}
\end{figure}
The small extent of significant residue zones indicates that the approach consisting of entering limiting cases in the fitting formula is overall relevant and that it is mostly the choice of the transition function involving two basic logistic functions that could be improved. It appears that parity of the logistic function (it is an odd function when changing its origin to $(\varepsilon,1/2)$) does not allow to match perfectly the evolution of the effective CSS both before and after $ \varepsilon$. This asymmetry needed to reproduce finer details of the effective CSS over the whole range may arise from the need of $\kappa_{1DR-0}$ expressions to describe them for example near $E_{A}\simeq 0.5\ eV$, $E_{B}\simeq 0.0\ eV$ at large D-ratios. Nevertheless, to keep it as simple as possible these limiting cases where not included the semi-analytical formula.
A possibility that does not require to include more limiting cases would be to use other functions to model the transition (asymmetric sigmoid like the generalized logistic function) which do not impose this symmetry. This was tested and achieved some partial improvement but at the cost of doubling the number of parameters. Fitting the transition zones with splines would probably be an even better option, but the number of parameters would then be as large as the product of the degree of the splines by the number of fitting point.
Notwithstanding these possible refinements of the semi-analytical fitting for transition regions, casting all possible CSS into a general formula should be seen a significant improvement of the CSS description compared to crude simplifications met in the literature consisting in approximating them with 3D CSS. Indeed we see from table \ref{validationSet} that varying rotation energies couples and diffusion ratio the CSS vary with a factor 50 as the dimensions of mobilities go from $1D-1D$ to $3D-3D$ and $\Delta$ is below three. Thus, the approximation of treating them as purely 3D would result for the present example in overestimating also by a factor 50. This as to be compared with the 20\% overall accuracy allowed by the semi-analytical formula, that we recast here in its most explicit expansion:
\begin{align}
&\kappa^2_{\text{fit}}(D_A,D_B, E_{A}, E_{B}) = \left(\frac{D_A}{D_B}\right)^{\delta(E_A,E_B)} \notag\\
&\times \left\{\left(1-\sigma_{1}(E_{A})\right)\kappa^2_{3D_{A}-3D_{A}} +\sigma_{1}(E_{A})\left[ \left(1-\sigma_{1}(E_{B})\right)\kappa^2_{3D_{A}-3D_{A}} + \sigma_{1}(E_{B})\kappa^2_{1D_{A}-1D_{A}} \right]\right\}\\ \label{semiEmpirical}
\end{align}
with
\begin{align}
&\delta(E_A,E_B)=\sigma_2(E_{A}) \left(-\frac{1}{2}+\frac{1}{6} \sigma_2(E_{B})\right), \notag\\
&\sigma_{1}(x)=\sigma(\lambda=8, \varepsilon=0.2,x), \notag\\
&\sigma_{2}(x)=\sigma(\lambda=10, \varepsilon=0.5,x), \notag\\
\end{align}
here additional simplifications have been made compared to the generic form Eq. \ref{generalKappa}: $\kappa^2_{1D_{A}-3D_{A}}$ has been simply replaced by $\kappa^2_{3D_{A}-3D_{A}}$, with imperceptible loss of overall accuracy. Of course, that does not mean that these two CSS expressions have close values int general (it is clearly wrong for small $\Delta$), but rather that most of their relative evolution for increasing $\Delta$ can be cast into a leading term, the D-ratio power delta term (or scaling term).
It is also interesting to note that quite counter-intuitively, simple reductions of the dimensionality of the mobilities are often misleading whereas dimensionality equivalences are relevant:
having both species 1D-migrating and $D_{B}$ as small as $D_{B}/D_{A}=10^{-3}$ does not allow to treat $B$ as immobile, as the effective CSS is far from the analytical $1D-0$ CSS, whereas it matches well the $2D-0$ CSS with proper accounting of scaling term.
This important point will be further investigated in the next section.
Finally, note that this semi-analytical formula is directly valid only at $T_{0}=573 K$. As such, it is also bound to the first nearest neighbor jump distance used in the OKMC simulations for the fitting set. To extend it to other temperatures and other lattices than BCC, a simple modification is proposed in the Annex \ref{extension_to_other_temperatures}.
\subsection{Closure of the CSS domains} \label{Closure}
In the previous section, we noted that even for diffusion coefficients ratios as low as ${\mathcal D}=10^{-3}$, the least mobile specie may not be treated as immobile, as the most relevant analytical CSS is the properly corrected $1D-1D$ expression. It is important for modeling concerns to investigate to which extends this holds, and what is the typical diffusion ratio where the $1D-0$ CSS starts to be more relevant.
The semi-analytical formula that we have established allows for reasonably accurate estimation of CSS depending on $(E_A, E_B, \Delta)$, but due to its numerical cost, the fitting set was limited to $0 \le \Delta \le 3$.
On that part of the $(E_A, E_B, \Delta)$-space, the two main absorption-rate domains are clearly visible ($3D-1D$ and $1D-1D$) but none of them are closed at $\Delta=3$. As a consequence the validity of the formula as such is restricted to that range, because there must be some D-ratio above which $B$ should be considered as immobile and the CSS tend to $3D-0$ and $1D-0$ expressions respectively. For the former, the transition does not need to be explicitly accounted as $3D-3D$ CSS expressions encompasses the case $D_{B}=0$, whereas for the $1D-1D$ to $1D-0 $ transition the reaction rates do not even have the same reaction order. Thus the range of $\Delta$ was extended to $\Delta=6$ in a new calculation set where the sampling criteria were slightly relaxed. As displayed on Fig. \ref{montage_BIGGER} the domain for $1D-3D$ CSS now closes by $\Delta=4.5$ and that of $1D-1D$ CSS closes about $5$. From $\log_{10}(D_{A}/D_{B})=3.75$ to $7.5$ some isosurfaces are wavy because the sampling set is smaller than for the former conditions. This could not be easily improved as the present result already required about 2.2 million CPU hours on Xeon Sandy Bridge 2.6 GHz cores. Being more precise on the closures in the $(E_A, E_B, \Delta)$-space would require quite a lot more computational resources, so we should rather focus on more restricted case, the $(E_{A}=2\ eV, E_{B}=2\ eV)$ case.
\begin{figure}
\includegraphics[width=0.8\textwidth]{montage_BIG_B123.jpg}
\caption{Three views of the effective exponents $\delta_{\text{eff}}$ isosurfaces.}
\label{montage_BIGGER}
\end{figure}
To that end, a few sets of simulations (for a few $(C, R)$ couples, $C=C_{A}=C_{B}$) for $\Delta$ ranging from $0$ to $9$ was ran. The results are shown on Fig. \ref{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES} which represents the normalized quantity:
\begin{equation}\label{conformal}
Y(\Delta)=\frac{\kappa^2_{\text{eff}}-\kappa^2_{1D-0}}{\kappa^2_{2D-0}-\kappa^2_{1D-0}}.
\end{equation}
This quantity is equal to one when $\kappa^2_{\text{eff}}=\kappa^2_{2D-0}$ and when its tends to zero then $\kappa^2_{\text{eff}}$ tends to $\kappa^2_{1D-0}$. When varying concentrations and radii, it helps to check if a common trend for the transition towards ${1D-0}$ stands out, when concentration couples vary. We see that it is indeed the case for the typical volume fraction conditions investigated. The question whether this simple behavior extends to a much wider range of $(C_A, C_B, R)$ conditions could be delicate and is not addressed here. We rather focus on proposing a practical correction for the vanishing scaling factor $({\cal D})^{-1/3}$ that would clearly lead to an underestimation of CSS from some point when $\Delta$ is small. For the investigated conditions, the $1D-1D$ to $1D-0$ transition happens to follow a common trend that is well fitted by the generalized sigmoid,
\begin{equation}\label{generalizedSigmoid}
Y_{\text{fit}}(\Delta)=a-\frac{b}{(c+d \exp(e \Delta-f))^g}
\end{equation}
with $a=0.0071, b=-1.26, c=2.23, d=0.932, e=3.81, f=0.98, g=0.26$ as shown on Fig. \ref{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES}.
\begin{figure}
\includegraphics[width=0.8\textwidth]{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES.jpg}
\caption{Transition of the normalized CSS from $1D-1D$ to $1D-0$ for four different conditions. The generalized sigmoid fit follows Eq. \ref{generalizedSigmoid}. In the caption, $\Phi$ represents the defects volume fractions and the concentrations $C_{A}=C_{B}$ are given in $\si{cm^{-3}}$.
}
\label{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES}
\end{figure}
From the representation of $\kappa^2_{\text{eff}}/\kappa^2_{1D-0}$ at Fig. \ref{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES_ZOOM}, we can conclude that, depending on the $C$ and $R$ the effective CSS reaches the analytical $\kappa^2_{1D-0}$ within a typical 5-10 \% uncertainty of estimates by $5<\Delta<6$, except for the blue curve whose atypical behavior is explained by a very large volume fraction. Formally, we can identify a critical $\Delta$ value by equating the 1D-0 and 1D-1D CSS which leads to
\begin{equation}
\Delta^{*} = -3 \log_{10} \left( \frac{3}{\overline{\alpha}} R^3 C\right), \label{critical}
\end{equation}
where ${\overline{\alpha}}\simeq\frac{4}{\ln(\pi^2 C R^3/{2})}$ is the usual parameter entering in the $1D-1D$ CSS expression \cite{Adjanor1}.
The evaluation of $\Delta^{*}$ for the conditions of Fig. \ref{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES_ZOOM} leads to values mostly between $6$ and $8$, which corresponds reasonably well to the critical values from the figure. Nevertheless the approach from the Eq.\ref{critical} and the actual evolution towards the critical values do not compare further: it is clear that the transition is very gradual and that both $1D-1D$ and $1D-0$ mechanisms operate simultaneously even above the critical value. To the simplest way, the cross-over between mobile and fixed sinks related CSS may be cast with semi-analytical formula as:
\begin{eqnarray}
Max\left(\kappa^2_{\text{fit}}(C_A, C_B, R, E_A, E_B, D_A, D_B),\right.\\
\left.\kappa^2_{1DR-0}(C_A, C_B, R, E_A, D_A)\right).
\end{eqnarray}
A more elaborated way of reproducing the transition would be to use the fit Eq. \ref{generalizedSigmoid} and replace it back in Eq. \ref{conformal}:
\begin{eqnarray}
\kappa^2_{\text{fit}} = Y_{\text{fit}}(\Delta)(\kappa^2_{2D-0}-\kappa^2_{1D-0})+\kappa^2_{1D-0},
\end{eqnarray}
but this should be valid only for pure $1D-1D$ and establishing a $Y_{\text{fit}}$ function valid for any rotation energy couple would require a more complex function with much more parameters.
\begin{figure}
\includegraphics[width=0.8\textwidth]{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES_ZOOM.jpg}
\caption{Convergence of the ratio of the effective CSS over the analytical $1D-0$ CSS for different conditions. In the legend, $\Phi$ represents the volume fractions and the concentrations are given in $\si{cm^{-3}}$.}
\label{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES_ZOOM}
\end{figure}
\section{Application to cluster dynamics} \label{sectionCD}
We now briefly expose the results of the application of the semi-analytical CSS on cluster dynamics simulations. This development allows to properly account for the agglomeration of defect clusters with mixed 1D/3D mobility i.e. any couple of rotation energies. Some additional technical details on the implementation in the code CRESCENDO \cite{Jourdan} can be found in the corresponding section of paper I. Here we focus on the ability of the present extension to reproduce an OKMC microstructure simulation close to the state-of-art in terms of complexity of random-walks. A concentration of $\num{2e16}\ \si{cm^3}$ interstitial monomers is initially placed in a quasi-cubic box of about $2000$ lattice constants length. The parameterization will allow the formed clusters to be mobile up to the size $60$. Increasing cluster content from $1$ to $60$ monomers, the diffusion coefficients are decreasing according to the power law $n^{2/3}$ from $\num{1.8e-5}$ to $\num{1.16e-6}\ $ and their rotation energies increase from $0$ to $2\ \si{eV}$ (linearly up to size $12$ and then the energy reaches a plateau value of constant value of $2\ \si{eV}$). Regarding rotation energies, these parameters are intended to sketch some state-of-the-art OKMC parameterizations found in the literature \cite{chiapetto2015nanostructure}. To be comparable with RECD, OKMC simulation must consist of hundreds to thousands of runs with different random seeds: a given OKMC run generally ends up with a few clusters resulting in a sparse distribution, while in RECD significant concentrations of defects are often found in quite smeared and continuous distributions. Figure \ref{2e16_redogaia_b2000} shows the comparison between the two simulation methods. It is worth mentioned that, although the set of one thousand lengthy OKMC runs represent a considerable amount of computer resources compared to the RECD simulation (OKMC requires here several millions times more individual CPU time than RECD), it is still not enough to fully characterize the distribution as it appears discontinuous due to the lack of sampling at the greatest sizes. Nevertheless, on the existing OKMC points the agreement with RECD is completely satisfactory, apart from the fact that at the intermediate simulation times the OKMC distribution is rougher and matches RECD distribution with some delay. Aside mentioned sampling limitations, this discrepancy is also quite likely to be a small box size effect: at intermediate evolution times, the peak of the cluster distribution is quite spread and it is barely above the minimum OKMC concentration for a single run ($C_{min}=1/V_{box}$, the dashed line of Fig. \ref{2e16_redogaia_b2000}). This clearly generates box size effects that are quite challenging to overcome with reasonable computational resources.
\begin{figure}
\includegraphics[width=1\textwidth]{2e16_redogaia_b2000.png}
\caption{Defect clusters distribution (x-axis is the cluster size in number of monomers, and the y-axis in the cluster concentration in \si{cm^3}). Starting from a population of $2 \times 10^{16}\ \si{cm^{-3}}$ SIA, the time evolution of the defect clusters population with mixed mobility was obtained by averaging one thousand OKMC runs (points). The comparison with RECD calculation is given by the continuous lines. The dashed line represents the minimum concentration (the inverse of the box volume) accountable in OKMC simulations.}
\label{2e16_redogaia_b2000}
\end{figure}
\section{Discussion: the relevance of mixed mobility sinks strength to dislocation loop evolution modeling}\label{Discussion}
Dislocation loops are complex objects whose mobilities are driven by the singular character of the stress field they wear. This is also the physical origin of the rotation energies \cite{Okita} (that we have taken here as black box parameter) as the rotation phenomenon shares some similarities with cross-slip, but with specific attempt frequencies and activation energies due to the closed chain nature of loops \cite{Dudarev}. A consequence of this is that the rotation probability of a loop may not always bear a simple parameterization with a fixed rotation energy that only depends on its size: it may also depend on the elastic fields that surrounds it. By construction, in cluster dynamics we can only hope to treat these types of spatial effects in a very indirect way and this would be at the cost of quite heavy modifications of the original formalism.
Another important phenomenon is the trapping of loop by chemical heterogeneities. Carbon and nitrogen, even at a few tens of ppm concentrations can have a huge impact on mobility of loops in ferritic materials at moderate temperature. Cottrell atmospheres of carbon atoms may form in the tensile zone of dislocation loops thus lowering the effective mobility of loops, and in the presence of vacancies, vacancy-carbon complexes may form \cite{Anento} and trap dislocation loops for long times. The effective migration energies of loops are thus significantly increased by a function of the trap-loop binding energy. In other words, untrapped species are expected to be in lower number densities than trapped ones in many conditions, so, at first sight, one may believe that $1DR-1DR$ CSS may play a secondary role in the microstructure kinetics compared to $1DR-0$ ones. But we shall keep in mind that:
\begin{itemize}
\item At high temperature, the trapping is less effective. There could even exist a cross-over between trapping and it opposite effect, the elastic confinement of loops by over-sized solute atoms \cite{Hudson2005}.
\item The vicinity of fixed sinks (dislocation network and grain boundaries) are often depleted of traps. It is indeed in these regions that greater populations of small dislocation loops (in the form of TEM "black dots") can be seen in complex alloys with many solutes and impurities such as industrial RPV steels at moderate doses. This is a local effect but nevertheless crucial as it conditions the local mobility of dislocation which is a key point for mechanical properties evolution under irradiation.
\item Even if loops spend most of the time trapped, clusters from the resulting smaller population of untrapped loops may react quite intensively and depending on the detailed conditions create a specific population.
\item At least from some point, the traps should be saturated and the new free defect clusters produced in cascades should be more likely to interact with other mobile clusters, than to be trapped. In the parameters set taken as an example in this study (see table \ref{tableMalerba}) the single interstitials are considered as 3D-mobile and from size 3 they start having a mixed mobility. Note also, that in other systems, the mono-interstitials may be crowdions and exhibit pure 1D-motion \cite{Amino2016}. In all these cases the $1DR-1DR$ CSS must be taken into account.
\item As we have seen in section \ref{Closure}, even with six orders of magnitude difference for the diffusion coefficients of two interacting 1D-mobile species, the $1DR-0$ CSS approximation is a serious underestimation of the effective CSS, contrary to the adapted $1D-1D$ expression. We can be more quantitative for this last but not least point: a popular way of accounting for trapping, inspired by the developments of Krishan \cite{Krishan} and others, is to simply lower the effective diffusion coefficient by a factor $\exp\left(-E_{\text{trap}}/k_B T \right)$, where $E_{\text{trap}}$ is the binding energy between the cluster and the trap. Actually, this corresponds to a lower bound for the effective diffusion coefficient as the trapping efficiency is assumed to be ideal. At $573\ K$, a trapping energy around $0.68\ \si{eV}$ corresponds to the factor $10^{-6}$ reduction of effective mobility. So, below this threshold trapping energy, a weakly trapped cluster should definitely be considered as mobile with respect a free clusters of the same size, as the CSS error resulting from treating the interaction as $1D-0$ would be large according to figure \ref{k2_calc_S_k2_2Dand1D_vs_D1_-S-_D2_FEW_CURVES_ZOOM}. There are several defect sizes complying with this condition in the parameterization exemplified in the appendix's table \ref{tableMalerba}: these are the smallest clusters but also, the most mobile and thus the most influential on the kinetics.
\end{itemize}
With these considerations in mind, mobile clusters aggregation (by diffusion of both reaction partners) should, at least be considered as a channel for loop growth competing with Ostwald ripening, and could even overtake it in specific conditions. In a nutshell, for a population of trapped clusters, Ostwald ripening will dominate only if the cluster dissolution rate is higher than the de-trapping rate.
For the parameterization exemplified in the annex, trapping energies of loops range from $0.17$ to $0.6\ \si{eV}$ below a cluster size threshold value \cite{Note6} and then $1.1\ \si{eV}$ above the threshold. The corresponding binding energies values are always much higher that trapping energies at each cluster size (binding energies typically range $0.87\ \si{eV}$ up to the formation energy of interstitial which is extremely high). So if the raw diffusion coefficients of interstitial clusters are comparable in terms of orders of magnitude to that of the monomer, then the loops will be more likely to untrap rather than dissolve and feed the Ostwald ripening process. Of course we may not prematurely conclude that aggregation always prevails, because high diffusion also promotes higher effective rates of emission (absorption rates appear in emission rates), but we stress out that the competition between the two processes must be considered.
This shed new light on the range of validity of some of the major nucleation theories: in the LSW formalism, both reaction channels are indeed considered but the formulation of agglomeration rates restricts its application to 3D mobilities. Since no expression for absorption rates involving two $1DR$ species with arbitrary diffusion coefficients was available, the applicability of general results on growth asymptotics to these cases was unclear. Now we see that, once proper corrections are applied, absorption rates between such species lead to second order kinetics rates just as 3D ones, and classical results may be qualitatively valid.
\section{Summary and conclusions}
Because the CSS dependence on concentrations, radii, diffusion coefficient and rotation energies is unknown and non-linear, complete fitting of CSS on all of its variables may represent quite a formidable task: any decent griding would probably require millions of conditions, all of which having quite different convergence behavior as the estimated CSS will span over several orders of magnitude. To tackle this problem of the large set of conditions, the proposed approach is the following: fixing the defect concentrations and the radii, we generated a first data set griding on the diffusion coefficient ratio and on the rotation energies. Then, we combined the analytical expressions developed in paper I \cite{Adjanor1} as limiting cases, thus allowing us to extend the expressions to arbitrary concentrations and radii, as was further check by analysis of the residuals. The transition between the limiting cases can be modeled by two combinations of sigmoids: a first combination of sigmoids is used to describe the transition between the CSS when $D_A=D_B$, and the second couple of sigmoids is intended to reproduce the presence of the quite distinct mobility domains (consistently with the limiting cases) which appear clearly when mapping the effective exponents of the diffusion ratio correcting factor. This term appears to be most crucial to capture the many order of magnitudes over which the CSS vary when the diffusion coefficient ratio goes from one to small values ($\Delta$ going from $0$ to large values). We also looked for the point from which the slowest specie should actually be considered as immobile in the reaction. The CSS with respect to a fixed sink would then become more relevant. This limit happens to be surprisingly low, which stresses out the necessity to use the $1DR-1DR$ CSS fit up to very small diffusion ratios. This should also be the case for the most important defect cluster couples (those mostly produced by cascades and that actually lead the dynamics) in the most typical irradiation situation discussed and would still hold when effective diffusion coefficients are be considerably lowered by trapping. The relevance of $1DR-1DR$ reactions is thought to be quite general: depending on detailed conditions and especially on temperature, aggregation of mobile clusters may compete with Ostwald ripening.
Finally, in the application of the RECD implementation, an in-depth validation of the formula was allowed thanks to a massive set of OKMC simulations with large box sizes requiring at least millions times more CPU time than RECD. As in moderate stiffness cases, RECD numerical schemes allow for very large time steps, this opens the way to very long-term (decades of physical time) simulations of the microstructure evolution fully accounting for the mixed mobility defect clusters which is intrinsically out of reach of any event-based method such as KMC.
\section{Acknowledgements}
Lorenzo Malerba is warmly acknowledged for initially pointing out the necessity of considering the mixed mobilities in KMC simulations, which initiated this work on their RECD counterpart, as well as Christophe Domain who additionally performed a benchmarking with its own implementation of effective CSS calculation within LAKIMOCA. Jean-Paul Crocombette is thanked for his suggestion to think about a "phase-diagram"-like representation of the CSS evolution with the mobility parameter. Thomas Jourdan is also acknowledged for collaboration on the implementation into the code CRESCENDO.
\section{Funding}
This project has received funding from the Euratom research and training program 2014-2018 under grant agreement No 661913 (SOTERIA).
|
\section{Statement of the Result}
Here we consider the stochastic process version of the Linial--Meshulam random simplicial complex model. Recall that the Linial--Meshulam random simplicial complex model (introduced in \cite{LM}), denoted $Y_d(n, p)$ for $d$ a fixed dimension, $n \in \N$, and $p = p(n) \in [0, 1]$, is the probability space on $d$-dimensional simplicial complexes with complete $(d - 1)$-skeleton generated by including each possible $d$-dimensional face independently with probability $p$. Accordingly the (discrete-time) stochastic process version of $Y_d(n, p)$, which we denote here as $\Y_d(n)$, following \cite{LP2}, is a Markov process $Y_d(n, 0) \subseteq Y_d(n, 1) \subseteq \cdots \subseteq Y_d(n, \binom{n}{d + 1})$ where $Y_d(n, 0)$ is the complete $(d - 1)$-complex on $n$ vertices and $Y_d(n, k)$ is generated by adding a $d$-dimensional face to $Y_d(n, k - 1)$ chosen uniformly at random from among all $d$-dimensional faces not included in $Y_d(n, k - 1)$.
For a topological property $P$ and a single instance of $\Y_d(n)$, the \emph{hitting time} for property $P$ is defined to be the minimal $m$ so that $Y_d(n, m)$ satisfies property $P$. A statement $S$ about $\Y_d(n)$ is said to hold \emph{with high probability} if the probability $S$ holds tends to $1$ as $n \to \infty.$
Our main new result on $\Y_d(n)$ is the following theorem; following tradition, we call a $(d-1)$-dimensional face \emph{isolated} if it is not covered by any $d$-dimensional face.
\begin{theorem}\label{mainresult}
Fix $d \geq 2$. With high probability the $(d - 1)$st homology group of $Y = \Y_d(n)$ with integer coefficients vanishes exactly when the last isolated $(d - 1)$-dimensional face is covered by a $d$-dimensional face. That is the hitting time for the property that no $(d - 1)$-dimensional face of $Y$ is isolated exactly coincides with the hitting time for the property that $H_{d - 1}(Y; \Z) = 0$.
\end{theorem}
\noindent In fact we prove a slightly stronger result in Corollary \ref{isolatedfaces}, which shows that slightly before the final isolated $(d - 1)$-dimensional face is covered, the $(d - 1)$st homology group is a free abelian group with rank given by the number of isolated faces.
The $d = 2$ case of Theorem \ref{mainresult} was previously established by \L uczak and Peled \cite{LP2}. Moreover, the $d = 1$ is the classic result of Bollobas and Thomason \cite{BT} that the stochastic random graph becomes connected at the exact moment its last isolated vertex is covered by an edge.
\section{Background}
The Linial--Meshulam model is a higher-dimensional generalization of the Erd\H{o}s--R\'{e}nyi random graph, and one of the most fundamental results in random graph theory is the following theorem of Erd\H{o}s and R\'{e}nyi which establishes the connectivity threshold for $G(n, p)$.
\begin{unnumtheorem}[\cite{ER}]
For $c < 1$ and $p = \frac{c \log n}{n}$, with high probability $G \sim G(n, p)$ is not connected, and for $c > 1$ and $p = \frac{c \log n}{n}$, with high probability $G \sim G(n, p)$ is connected.
\end{unnumtheorem}
This statement can be given a homological reformulation by observing that a graph is connected if and only if its zeroth reduced homology group is trivial.
This motivates the general definition of \emph{homological connectivity over abelian group $R$} for higher dimensional simplicial complexes.
A $d$-dimensional simplicial complex $X$ is said to be homologically connected over $R$ provided that $\tilde{H_i}(X; R) = 0$ for all $i \leq d - 1$. In $Y_d(n, p)$, all complexes have complete $(d - 1)$-skeleton so homological connectivity over $R$ of $Y \sim Y_d(n, p)$ is equivalent to $\tilde{H}_{d - 1}(Y; R) = 0$. In the case that $R= \Z,$ we will simply say the complex is \emph{homologically connected}. Indeed, by the universal coefficient theorem, a complex $X$ is homologically connected if and only if it is homologically connected over all abelian groups $R.$
Generalizing the connectivity result of Erd\H{o}s and R\'{e}nyi, Linial and Meshulam prove the following.
\begin{unnumtheorem}[\cite{LM}]
For $c < 2$ and $p = \frac{c \log n}{n}$, with high probability $Y \sim Y_2(n, p)$ satisfies $H_1(Y; \Z/2\Z) \neq 0$, and for $c > 2$ and $p = \frac{c \log n}{n}$, with high probability $Y \sim Y_2(n, p)$ satisfies $H_1(Y; \Z/2\Z) = 0$.
\end{unnumtheorem}
For any $d$-complex $Y$, $H_{d - 1}(Y; \Z/2\Z) = 0$ implies that $H_{d - 1}(Y; \Q) = 0$. By the universal coefficient theorem, therefore, the above result of Linial and Meshulam implies that $H_1(Y;\Z)$ is finite for $Y \sim Y_2(n, \frac{c \log n}{n})$ and $c > 2$. However, the result \emph{does not} imply that $H_1(Y, \Z) = 0$ in this case; a priori, it may be some other finite group. Thus, unlike the case of the Erd\H{o}s--R\'{e}nyi random graph, it is not sufficient to consider only $\Z/2\Z$ coefficients to show that $H_{d - 1}(Y)$ is trivial. Generalizing the Linial--Meshulam result to higher dimensions and to other coefficients rings, Meshulam and Wallach prove the following result.
\begin{unnumtheorem}[\cite{MW}]
Fix $d \geq 1$, and let $R$ be a fixed finite abelian group. For $c < d$ and $p = \frac{c \log n}{n}$, with high probability $Y \sim Y_d(n, p)$ satisfies $H_{d - 1}(Y; R) \neq 0$, and for $c > d$ and $p = \frac{c \log n}{n}$, with high probability $Y \sim Y_d(n, p)$ satisfies $H_{d - 1}(Y; R) = 0$.
\end{unnumtheorem}
If $Y$ is a $d$-dimensional simplicial complex with $H_{d - 1}(Y; \Z/q\Z) = 0$ for all primes $q$ then $H_{d - 1}(Y; \Z) = 0$. However, the theorem of Meshulam and Wallach does not rule out the possibility that $H_{d - 1}(Y; \Z)$ has $q$-torsion for a sequence of primes $q$ which grow with $n$. Indeed until now the question of the homological connectivity threshold (with integer coefficients) has been open for all $d \geq 3$. Previously, the best result for $d \geq 3$ about the homological connectivity threshold was the following result of Hoffman, Kahle and Paquette:
\begin{unnumtheorem}[\cite{HKP2}]
For $d \geq 2$ and $p \geq \frac{80d \log n}{n}$, with high probability $Y \sim Y_d(n, p)$ satisfies $H_{d - 1}(Y) = 0$.
\end{unnumtheorem}
For $d = 2$, the main hitting-time result of \cite{LP2} establishes that $\frac{2 \log n}{n}$ is the sharp threshold for the first homology group with integer coefficients to vanish in $Y_2(n, p)$, and our result here generalizes this hitting-time result to higher dimensions.
We should also mention that over the field $\Z/2\Z$, the hitting-time result has been established. If one considers homology with $\Z/2\Z$ coefficients the hitting-time result was previously known in the $d = 2$ case due to Kahle and Pittel \cite{KP}, and more recently the $\Z/2\Z$ version of the hitting-time for homological connectivity was proved by Cooley et al. \cite{CdGKS} in all dimensions. In addition a hitting-time result that for $\Q$-coefficients is proved in \cite{HKP}.
\section{Cocycle counting}
While the $d = 2$ case of Theorem \ref{mainresult} has already been established in \cite{LP2}, we develop a new approach based on the methods of Meshulam and Wallach \cite{MW}. They develop the technique of
\emph{cocycle counting} to show that for any fixed finite abelian group $R$, $H_{d - 1}(Y_d(n, p); R) = 0$ when $p = \frac{d \log n + \omega(1)}{n}.$
That is, rather than considering homology, they consider cohomology and bound the probability that $Y \sim Y_d(n, p)$ has a nontrivial cocycle. When $R$ is fixed and finite this may be accomplished by showing that the expected cardinality $\mathbb{E}|H^{d - 1}(Y; R)|$ tends to $0.$
We will adapt this technique to work over many fields simultaneously. Following \cite{MW}, we start by defining some useful notation.
\begin{definition}
For a $(d - 1)$-cochain $\phi$ of the simplex on $n$ vertices with coefficients in any field $R$, the \emph{weight} of $\phi$, denoted $w(\phi)$, is defined to be the minimum of the support of $\phi'$ for all $\phi'$ with $\phi - \phi'$ a coboundary, and $b(\phi)$ is defined to be the number of $d$-dimensional faces $\sigma$ in the simplex on $n$ vertices so that $\partial_d^*(\phi)(\sigma) \neq 0$.
\end{definition}
With this notation in hand, Meshulam and Wallach prove the following coisoperimetric inequality \cite[Proposition 3.1]{MW}.
\begin{lemma}[Coisoperimetric inequality]\label{coisoperimetric}
For any abelian group $R$ and any $(d - 1)$-chain of the simplex on $n$ vertices,
$$b(\phi) \geq \dfrac{n w(\phi)}{d + 1}.$$
\end{lemma}
\noindent Note that although $b(\phi)$ and $w(\phi)$ depend on the underlying ring $R$, the coisoperimetric inequality is uniform over all abelian groups $R$.
We will now sketch the basic cocycle counting method.
Observe that if $\phi$ is a $(d - 1)$-cochain then the probability that it is a cocycle over $Y \sim Y_d(n, p)$ is $(1 - p)^{b(\phi)}$.
Consider each equivalence class of cochains modulo coboundaries and choose a minimal-support element from each equivalence class. If $\phi$ is a cochain with support size $k$ and weight equal to $k$, then by Lemma \ref{coisoperimetric} the probability that $\phi$ is a cocycle is at most $(1 - p)^{nk/(d + 1)}$. Thus if $R$ is a fixed finite field of size $r$, we have that for $Y \sim Y_d(n, p),$
$$
\Pr(H^{d - 1}(Y; R) \neq 0) \leq \sum_{k = 1}^{\binom{n}{d}} \binom{\binom{n}{d}}{k} (r - 1)^k (1 - p)^{nk/(d + 1)}.$$
Indeed there are $\binom{\binom{n}{d}}{k}$ choices for the support of a cochain of weight $k$, and from there at $(r - 1)$ choices for the coefficient associated to each facet in the support. It follows that if $p = \frac{c \log n}{n}$ for $c > d(d + 1)$ then $H^{d - 1}(Y; R) = 0$ for $Y \sim Y_d(n, p)$. In order to improve on this, Mesulam and Wallach find a better bound than $\binom{\binom{n}{d}}{k} (r - 1)^k$ for the number of nontrivial cochains and use the coisopermetric inequality in a more subtle way.\\
In the current situation we want to show that homology with integer coefficients vanishes.
Our approach is based on the elementary observation that for any simplicial complex $X,$ if $H_{d - 1}(X; \Z/q\Z) = 0$ for all primes $q$ then $H_{d - 1}(X; \Z) = 0$.
So, we adapt the cocycle counting method of Meshulam and Wallach to work over $\Z/q\Z$ for all primes $q$ simultaneously.
It is worth pointing out that a direct first-moment argument alone cannot work if $q$ is very large. If we sample $Y$ from $Y_d(n, \frac{c \log n}{n})$ then the probability that $Y$ has no $d$-dimensional faces is $\exp(-\Theta(n^d \log n))$. In this case, the dimension of $H_{d - 1}(Y; R)$ for any field $R$ is $\binom{n - 1}{d}$, so if $R = \Z/q\Z$, for $q$ a prime larger than $\exp(n^d)$, then the expected number of cocycles over $R$ is at least $\exp(\Theta(n^{2d}) - \Theta(n^d \log n)) \rightarrow \infty$. Thus is critical that we eliminate the $(r - 1)^k$ term from the cocycle counting method.\\
To do so, rather than consider $(d-1)$-cochains, we consider $(d - 1)$-dimensional complexes and bound the probability that any such complexes support a cocycle over any prime-order finite field. We will make use of the coisoperimetric inequality (Lemma \ref{coisoperimetric}), but now in a more geometric way. We first define the following geometric analogue to $b(\phi)$.
\begin{definition}
For a fixed field $R$ and any $(d - 1)$-dimensional subcomplex $X$ of the simplex on $n$ vertices we define $b(X, R)$ as:
$$b(X, R) := \inf \{b(\phi) : \text{$\phi$ is supported exactly on $X$ with coefficients in $R$ with $w(\phi) = |X|$} \}.$$
Note that $b(X, R)$ can be infinity but only in the situation where there are no cochains of minimum weight supported on $X$.
We also define $b(X)$ to be the infimum of $b(X, R)$ over $R = \Z/q\Z$ for all primes $q$ and $R = \Q$.
\end{definition}
Now $b(X)$ is closely related to $b(\phi)$ where $X$ is a $(d - 1)$-complex and $\phi$ is a cochain supported on $X$, but it removes everything about an underlying coefficient ring. Meshulam and Wallach also define a geometric quantity $\beta(X),$ closely related to $b(\phi)$ for $\phi$ supported on $X,$ as follows.
\begin{definition}
For a $(d - 1)$-dimensional subcomplex $X$ of the simplex on $n$ vertices we define $\beta(X)$ to be the number of $d$-dimensional faces which contain exactly one $(d - 1)$-dimensional face of $X$.
\end{definition}
We make use of this definition too in Section \ref{largecocycles}. In outlining their proof in \cite{MW}, Meshulam and Wallach point out that the coisoperimetric inequality does not hold if $b(\phi)$ is replaced with $\beta(X)$ and that this is a major obstacle to applying their technique to prove that integer homology vanishes. Nonetheless, it is a useful quantity because for any cochain $\phi$ with coefficients in $R$, minimally-supported on $X,$ one has $\beta(X) \leq b(X) \leq b(\phi)$. \\
So while $\beta(X)$ does not satisfy the coisoperimetric inequality, the geometric quantity $b(X)$ \emph{does} satisfy it. Indeed, due to the uniformity in $R$ in the coisoperimetric inequality (Lemma~\ref{coisoperimetric}),
$$b(X) \geq \frac{n |X|}{d + 1}.$$
There is one potential disadvantage to using $b(X)$ in place of $b(\phi)$. If $R$ is a coefficient ring and if $\phi$ is a cochain over $R$ which is minimally supported on $X$, then it is clear from the definition of $b(\phi)$ that the probability that $\phi$ is a cocycle of $Y_d(n, p)$ is $(1 - p)^{b(\phi)}$. However, if instead we wish to bound the probability that $X$ is the support of a cocycle over any prime-order finite field, we no longer have the simple bound $(1 - p)^{b(X)}$. Nonetheless, as we will show in Lemma \ref{probabilitybound}, this bound is true up to a lower--order correction.
To frame Lemma \ref{probabilitybound}, we begin by introducing the following notation.
\begin{definition}
For a fixed field $R$ and any $(d - 1)$-dimensional subcomplex $X$ of the simplex on $n$ vertices, we let $z(X, R)$ denote the event that there exists a cocycle $\phi$ over $R$ with $w(\phi) = |X|$ and $\supp(\phi) = X$. We let $z(X)$ denote the event that there exists $R$ in $\{\Z/q\Z : \text{$q$ is prime and at most $\sqrt{d + 1}^{|X|}$} \} \cup \{\Q\}$ so that $z(X, R)$ holds.
\end{definition}
The choice of $\sqrt{d + 1}^{|X|}$ in the definition of $Z(X)$ comes from a bound on the size of the torsion group of the cokernel of an integral matrix, given as Claim \ref{matrixbound}. This claim is essentially Proposition 3 of \cite{Soule} who credits it to Gabber. However, we do have a sharper exponent in our bound than in \cite{Soule} ($t^{\rank(M)}$ compared to $t^{\min\{n, m\}}$). This sharper exponent is not necessary to our application here, but in the interest of keeping the proof self-contained we give a proof of Claim \ref{matrixbound}, and no additional work is require to obtain the sharper exponent.
\begin{claim}\label{matrixbound}
If $M$ is a matrix with integer entries so that the norm of every column of $M$ is at most $t$, then the torsion part of the cokernel of $M$, denoted $\coker(M)_T,$ has size at most $t^{\rank_\Q(M)}$
\end{claim}
\begin{proof}
Let $M$ be a matrix which satisfies our assumptions. First, define $N$ to be a restriction of $M$ to a maximal set of $\Q$-linearly independent columns of $M$. We have that $\coker(M)_T \leq \coker(N)_T$. Indeed, this immediate as clearly $\Im_{\Z}(N) \leq \Im_{\Z}(M)$. \\
Now we want to construct a square matrix from $N$ in a canonical way. Beginning with $N$ let $i_1$ be the smallest index in $\{1, ..., m\}$ so that the standard basis vector $e_{i_1}$ is not in the $\Q$-span of $N$; add $e_{i_1}$ to $N$. Now let $i_2$ be the smallest index in $\{1,..., m\}$ so that $e_{i_2}$ is not in the $\Q$-span of $N$ and $e_{i_1}$, add $e_{i_2}$ to the matrix. Continue in this way to arrive at a (necessarily square) matrix $N'$. We check that $\coker(N)_T \leq \coker(N')_T$. \\
We check the subgroup inclusion inductively. Suppose $v$ is a torsion element of the span of the columns of $N$ together with standard basis vector $e_{i_1}, e_{i_2}, \cdots, e_{i_k}$, but $v$ is not a torsion element of the cokernel after adding the column $e_{i_{k + 1}}$. Then $v$ can be written as an integral linear combination of columns of $N$ and standard basis vectors $e_{i_1}, e_{i_2}, \cdots, e_{i_k}, e_{i_{k + 1}}$, with nonzero coefficient $\alpha$ on $e_{i_{k + 1}}$. However since $v$ is a torsion element of the cokernel before adding $e_{i_{k + 1}}$ we have that there exists an nonzero integer $s$ so that $sv$ is in the integer span of the columns $N$ and $e_{i_1}, e_{i_2}, \cdots, e_{i_k}$. This gives us two ways to write $sv$ as a linear combination of columns of $N$ and $e_{i_1}, e_{i_2}, \cdots, e_{i_k}, e_{i_{k + 1}}$, one with coefficient $s\alpha$ on $e_{i_{k + 1}}$ and one with coefficient 0 on $e_{i_{k + 1}}$. By linear independence of the columns of $N'$ we have that $s \alpha = 0$, a contradiction. Thus $\coker(N)_T \leq \coker(N')_T$ and we complete the proof.
\end{proof}
Similar bounds appear in \cite{HKP2, Kalai, LP2}. In \cite{Kalai}, it is shown that if $X$ is a $d$-dimensional simplicial complex on $n$ vertices then $|H_{d - 1}(X)_T| \leq \sqrt{d + 1}^{\binom{n - 2}{d}}$. Claim \ref{matrixbound} essentially gives a local version of this result.\\
\begin{lemma}\label{probabilitybound}
For any $c > (d-1/2),$
there is an $n_0$ sufficiently large
so that
for all $n \geq n_0,$
for all $(d - 1)$-dimensional complexes $X$ with $b(X) = (1 - \theta)nk \geq nk/(d + 1),$ where $k := |X|$,
and for $Y \sim Y_d(n, \lceil \frac{c \log n}{n} \binom{n}{d + 1} \rceil)$,
the probability that $z(X)$ holds is at most $n^{-(1 - \theta) (d - 1/2)k}.$
\end{lemma}
\begin{proof}
Fix a $(d - 1)$-dimensional complex with $b(X) = (1 - \theta) nk$ where $k := |X|$, and fix a field $R$ to be either $\Z/q\Z$ for $q \leq \sqrt{d + 1}^{k}$ or to be $\Q$. We will bound the probability of $z(X, R)$ for $R$ fixed and then take a union bound over all at-most $\sqrt{d + 1}^{k}$ necessary fields to bound $z(X)$. We will use the Linial--Meshulam stochastic process $\Y_d(n) = \{Y_d(n,i)\}_{i=0}^{\binom{n}{d+1}}$ to sample from $Y_d(n,m)$ where $m := \lceil \frac{c \log n}{n}\binom{n}{d + 1} \rceil$. For each $i$, let $z(X, R, i)$ be the event that $X$ is the support of a cocycle over $R$ of weight $k$ in $Y_d(n, i)$. At each step $i,$ we let $\X(i)$ denote the dimension of the kernel of the coboundary matrix of $Y_d(n, i)$ restricted to the columns associated to $X$. Clearly if $\X(m) = 0$, then $X$ is not the support of a cocycle over $R$.\\
For each $i$ let $p_i$ denote the probability that $\X(i + 1) < \X(i)$. Now if $X$ is the support of a cocycle of weight $k$ over $R$ in $Y_d(n, i)$ then
$$p_i \geq \frac{b(X)}{\binom{n}{d + 1}} \geq \frac{(1 - \theta)nk}{\binom{n}{d + 1}}.$$
We want to bound the following probability
$$\Pr((\X(m) > 0) \cap z(X, R, m)).$$
Note that $z(X, R, m)$ implies $\X(m) > 0$ so the probability above is actually equal to $\Pr(z(X, R))$. Clearly,
\begin{eqnarray*}
\Pr(\X(m) > 0 \cap z(X, R, m)) &\leq& \Pr\left((\X(m) > 0) \cap z(X, R, m) \mid \forall i \leq m, p_i \geq\frac{(1 - \theta)nk}{\binom{n}{d + 1}} \right) \\
&&+ \Pr\left((\X(m) > 0) \cap z(X, R, m) \mid \exists i \leq m, p_i <\frac{(1 - \theta)nk}{\binom{n}{d + 1}} \right) \\
&\leq& \Pr \left( \X(m) > 0 \mid \forall i \leq m, p_i \geq\frac{(1 - \theta)nk}{\binom{n}{d + 1}}\right) \\
&& + \Pr\left(z(X, R, m) \mid \exists i \leq m, p_i <\frac{(1 - \theta)nk}{\binom{n}{d + 1}} \right).
\end{eqnarray*}
Now the second summand is zero. Indeed while $z(X, R, i)$ holds $p_i \geq \frac{(1 - \theta)nk}{\binom{n}{d + 1}}$ and if $z(X, R, i)$ fails for some $i \leq m$, then so does $z(X, R, m)$. The goal is to bound the first summand. This will be accomplished by comparison to a binomial random variable. Let $B$ be a binomial random variable with $m$ trials and success probability $\psi = \dfrac{(1 - \theta) nk}{\binom{n}{d + 1}}$. Since $\X(0) = k$, it follows that
$$\Pr\left(\X(m) > 0 \mid \forall i \leq m, p_i \geq \frac{(1 - \theta) nk}{\binom{n}{d + 1}} \right) \leq \Pr(B < k).$$
We bound the probability that $B$ is less than $k,$ which since $k$ is less than $\mathbb{E}(B) = m\psi = \Omega(k \log n),$ allows the following version of Chernoff's inequality to apply:
\[
\Pr(B < k) \leq \exp( mH_\psi(k/m)),
\]
where $H_\psi(x) = x\log(\psi/x) + (1-x)\log( (1-\psi)/(1-x)).$
Observe that $k$ is no more than $n^d,$ and that $n^d/m = O(1/\log n).$
Hence uniformly in $k \leq n^d,$ we have that
\[
\begin{aligned}
mH_\psi(k/m)
&=
k\log(m\psi/k)
+(m-k)\log(1-\psi)
-(m-k)\log(1-k/m) \\
&\leq
k\log(m\psi/k)
-m\psi + k(1+\psi) + O(k^2/m) \\
&\leq
-m\psi
+k\log\log n
+k(1+O(1/\log n)).
\end{aligned}
\]
Thus there is a constant $C>0$ so that for all $n$ large enough and for all $k \leq n^d$
\begin{eqnarray*}
\Pr(B < k) \leq n^{-(1 - \theta)ck}(C\log n)^{k}.
\end{eqnarray*}
Now we sum over all fields $\Z/q\Z$ with $q \leq \sqrt{d + 1}^k$ and the field $\Q$ to get that the probability that $X$ is the support of a cocycle over any such field is at most:
\begin{equation} \label{eq:truebound}
\Pr(z(X)) \leq n^{-(1 - \theta)ck}(C\sqrt{d+1}\cdot \log n)^{k}.
\end{equation}
This gives the desired bound for all $n$ sufficiently large.
\end{proof}
\begin{remark}
We should point out here that there is nothing particularly meaningful about the choice of $(d - 1/2)$ in the lemma. Later, this turns out to be a convenient value to have, and so we use it here to simplify some notation.
Also, recalling the definition of $b(X),$ there is some cochain $\phi$ over some field $R$ such that $\phi$ is supported on $X$ and so that $b(\phi)=b(X).$ For this particular cochain, the probability it is a cocycle is $(1 - p)^{b(X)}= n^{-(1 - \theta)ck}$. Hence we have
\[
n^{-(1 - \theta)ck} \leq \Pr(z(X)) \leq n^{-(1 - \theta)ck}(C\sqrt{d+1}\cdot \log n)^{k},
\]
showing that the above bound is accurate up to subleading factors.
\end{remark}
\section{Overview of the proof}
Now that we have defined $z(X)$, we can give an outline of the proof of Theorem \ref{mainresult}. Essentially the idea of the proof will be to prove that with high probability $Y \sim Y_d(n, \frac{c \log n}{n})$ satisfies three particular conditions for $c > d - 1/2$ and then to show that these three conditions deterministically imply Theorem \ref{mainresult}. The majority of the work of this paper is to prove the former, which we state below as Lemma \ref{keylemma}. To simplify notation, for a $d$-dimensional simplicial complex, we use ``face", ``facet", and ``ridge" to refer to $d$-dimensional faces, $(d - 1)$-dimensional faces, and $(d - 2)$-dimensional faces respectively. We also say that a cocycle $\phi$ is inclusion-minimal over a field $R$ if there is no cocycle over $R$ supported on any proper subset of the support of $\phi$.
\begin{lemma}\label{keylemma}
Fix $d \geq 2$ and $c > d - 1/2$, then with high probability $Y \sim Y_d(n, \frac{c \log n}{n})$ satisfies the following three conditions:
\begin{enumerate}
\item $z(X)$ fails to hold for all $(d - 1)$-subcomplexes $X$ with $|X| \geq n/(3d)$
\item $Y$ has no inclusion minimal $(d-1)$-cocycles of support size $k$ over any field for $2 \leq k \leq n/(3d)$.
\item $Y$ has no isolated facets that meet at a ridge.
\end{enumerate}
\end{lemma}
These three conditions, in turn, imply the desired homology vanishing on $Y$, as a consequence of the following \emph{deterministic} lemma.
\begin{lemma}\label{deterministic}
Suppose that $Y$ is a $d$-dimensional simplicial complex with complete $(d - 1)$-skeleton so that conditions 1 and 2 from Lemma \ref{keylemma} hold, then $H_{d - 1}(Y)$ is a free abelian group of rank equal to the number of isolated facets of $Y$. Moreover if all three conditions hold then the stochastic process of adding $d$-dimensional faces uniformly at random to $Y$ will result in a complex $Y' \supset Y$ which has $H_{d - 1}(Y') = 0$ exactly at the moment the final isolated facet of $Y$ is covered.
\end{lemma}
\noindent In the proof of this lemma we encounter the term \emph{strongly-connected} which we define here and will use again in Section \ref{smallcocycles}.
\begin{definition}
For a $d$-dimensional simplicial complex $X$ we define the \emph{dual graph} $G(X)$ to be the graph whose vertex set is the set $d$-dimensional faces of $X$ with an edge between $\sigma$ and $\tau$ provided that $\sigma$ and $\tau$ intersect at a $(d - 1)$-dimensional face. We say that $X$ is \emph{strongly-connected} if its dual graph is connected. Equivalently, $X$ is strongly connected if for every $\sigma$ and $\tau$ there is a path $\sigma = \sigma_1, \sigma_2, \sigma_3, \cdots, \sigma_k = \tau$, so that for every $1 \leq i \leq k - 1$, $|\sigma_i \cap \sigma_{i + 1}| = d - 1$.
\end{definition}
\begin{proof}[Proof of Lemma~\ref{deterministic}]
First suppose that $Y$ is such that conditions 1 and 2 hold. Then we have that over $\Q$ every inclusion-minimal cocycle of $Y$ is an isolated facet. Thus $H^{d - 1}(Y; \Q)$ is generated by isolated facets of $Y$. It follows that the same holds for $H^{d - 1}(Y; \Z)$ (recall that torsion subgroups ``shift up" one dimension when we change from homology to cohomology, so we don't immediately have that $H_{d - 1}(Y; \Z)$ is free). We claim that $\beta^{d - 1}(Y; \Z/q\Z) = \beta^{d - 1}(Y; \Q)$ for every prime $q$. This will imply that there is no torsion in homology and so $H^{d - 1}(Y; \Z)$ will be isomorphic to $H_{d - 1}(Y; \Z)$ proving the first part of the claim. Suppose there is a prime $q$ so that $\beta^{d - 1}(Y; \Z/q\Z) > \beta^{d - 1}(Y; \Q)$. Then $Y$ has a nontrivial cocycle $\phi$ with coefficients in $\Z/q\Z$ that is not the image of an integral cocycle modulo $q$. Let $X$ be the support of a minimal-weight representative of $\phi$. We may assume that $X$ has no isolated facets (otherwise we could subtract from $\phi$ an appropriate multiple of a cochain supported on exactly an isolated facet of $X$ to arrive at a new cocycle over $\Z/q\Z$ which is not the image of a cocycle over $\Z$ and has smaller support). \\
Now over $\Z$ we have $\partial_d^* |_X(\phi) = q \psi$ for some integral vector $\psi$. Moreover by conditions 1 and 2, $q > \sqrt{d + 1}^{|X|}$ (and $|X| \geq n/(3d)$). But this $q$ is too large relative to $|X|$ for the cokernel of $\partial_d^*|_X$ to have $q$-torsion by Claim \ref{matrixbound}, thus $q \psi$ in the image of $\partial_d^*|_X$ over $\Z$ implies that $\psi$ is also in the image over $\Z$. Thus there exists an integral vector $\phi'$ so that $\phi - q \phi'$ is supported on $X$ and is a cocycle over $\Z$. However since $X$ has no isolated facets and $H^{d - 1}(X, \Z)$ is generated by isolated facets we have that $\phi - q \phi'$ is a coboundary over $\Z$, but then modding out by $q$ gives us that $\phi$ is a coboundary over $\Z/q\Z$ contradicting our assumption on $\phi$. This proves the first part of the claim. \\
Now suppose that $Y$ satisfies conditions 1, 2, and 3 and let $Y'$ be the complex at the moment in the stochastic process where the last isolated facet of $Y$ is covered. We wish to show that $H_{d - 1}(Y') = 0$. We will prove this by induction on the number of isolated facets of $Y$. By conditions 1 and 2, $H_{d - 1}(Y)$ is a free abelian group generated by the isolated facets of $Y$. Thus if $Y$ has no isolated facets then $H_{d - 1}(Y) = 0$ and $Y' = Y$ so we have the result. For the inductive step, we prove that if $Y$ satisfies conditions 1, 2, and 3 then for any face $\sigma$ which could be added to $Y$, we have that $Y \cup \{\sigma\}$ still satisfies 1, 2, and 3. This will prove that 1, 2, and 3 hold at every step and eventually we cover some isolated facet of the complex and then can apply induction. \\
Conditions 1 and 3 are clearly monotone, so we only have to show that condition 2 is monotone under our other assumptions. Suppose not. Let $\sigma$ be a face so that $Y$ satisfies conditions 1, 2, and 3 but $Y \cup \{\sigma\}$ does not satisfy condition 2. Let $\phi$ be an inclusion-minimal cocycle, with weight and support size at least two, over some field $R$ for $Y \cup \{\sigma\}$. Since $\phi$ is a cocycle of $Y \cup \{\sigma \}$ it is a cocycle for $Y$. But because $Y$ satisfies 1 and 2, we have that $H^{d - 1}(Y; R)$ is generated by isolated facets. Therefore the support of $\phi$ is a union of isolated facets. By inclusion-minimality (after adding $\sigma$) the support of $\phi$ is strongly connected too. However, by 3 we have that the support of $\phi$ must be a single isolated facet, so $\phi$ has support size one, contradicting our assumption. This shows that $Y \cup \{\sigma\}$ satisfies condition 2 and we finish the proof by induction.
\end{proof}
Before we get to the proof of Theorem \ref{mainresult} we state the following corollary which characterizes the structure of $H_{d - 1}(Y)$ for $Y \sim Y_d(n, \frac{c \log n}{n})$ and $c > d - 1/2$. The proof is immediate from Lemmas \ref{keylemma} and \ref{deterministic}:
\begin{corollary}\label{isolatedfaces}
If $c > d - 1/2$, then with high probability $H_{d - 1}(Y)$ for $Y \sim Y_d(n, \frac{c \log n}{n})$ is a free abeilan group of rank equal to the number of isolated facets of $Y$.
\end{corollary}
Together with a first moment argument showing that $Y \sim Y_d(n, \frac{c \log n}{n})$ has no isolated facets with high probability for $c > d,$
this corollary establishes the sharp threshold for integral homology to vanish in $Y_d(n, p)$. Now we give the proof of the main result.
\begin{proof}[Proof of Theorem \ref{mainresult}]
Consider an instance of $\Y_d(n)$, and let $m_0$ be the hitting time for the event that the final isolated facet of $\Y_d(n)$ is covered. Clearly $Y_d(n, i)$ has nontrivial $(d - 1)$st homology group for $i < m_0$. That is the hitting time for $(d-1)$st homology to vanish is not earlier than the hitting time for the final isolated facet to be covered. It therefore suffices to show that with high probability $H_{d - 1}(Y_d(n, m_0)) = 0$. \\
First generate $Y \sim Y_d \left( n, \frac{(d - 1/4) \log n}{n} \right)$. If $Y$ has isolated facets, then run the stochastic Linial--Meshulam process starting at $Y$ and continuing until the moment the last isolated facets is covered. In the case that $Y$ has isolated facets, this generates a complex $Y_d(n, m_0)$ in $\Y_d(n)$. By Lemma \ref{keylemma}, $Y$ satisfies the three stated conditions with high probability and so by Lemma \ref{deterministic}, the probability that $Y_d(n, m_0)$ has nontrivial $(d - 1)$st homology group given that $Y$ has isolated facets is $o(1)$. Thus it suffices to check that the probability that $Y$ has no isolated facets is also $o(1)$, but this follows from a straightforward second moment argument.
\end{proof}
The rest of the paper will be devoted to proving Lemma \ref{keylemma}. Condition 1 will be referred to as the ``large cocycle" condition and will be proved in Section \ref{largecocycles}. Condition 2 will be referred to as the ``small cocycle" condition and will be proved in Section \ref{smallcocycles}. Condition 3 is much easier and we prove it now
\begin{lemma}\label{noadjacent}
If $Y \sim Y_d(n, c \log n/n)$ and $c > (d + 1)/2$ then with high probability $Y$ does not contain two isolated facets that meet at a ridge.
\end{lemma}
\begin{proof}
We will use the first moment method. Two isolated facets that meet a ridge is a subcomplex with two $(d - 1)$-dimensional faces, which are both isolated, and $d + 1$ vertices. The number of such complexes is at most $\binom{n}{d + 1} \binom{d + 1}{d-1}$. The probability that both facets are isolated is at most $(1 - p)^{2(n - d) - 1}$. Thus the expected number of pairs of isolated facets that meet at a ridge is at most
\begin{eqnarray*}
\binom{n}{d + 1} \binom{d + 1}{d-1}(1 - p)^{2(n - d) - 1} &\leq& n^{d + 1} (d + 1)^2 \exp \left( -\frac{c \log n}{n} (2(n - d) - 1) \right) \\
&\leq& n^{d + 1}(d + 1)^2 n^{-2c(1 - o(1)) }
\end{eqnarray*}
This is $o(1)$ since $2c > (d + 1)$.
\end{proof}
\section{Large cocycles}\label{largecocycles}
The goal of this section is to prove the following lemma about the large cocycle condition. This will be accomplished by using Lemma \ref{probabilitybound} together with an enumeration result from \cite{MW} to bound the probability that $Y \sim Y_d(n, \frac{c \log n}{n})$ contains a $(d - 1)$-dimensional subcomplex $X$, with $|X| \geq n/(2d)$, for which $z(X)$ holds.
\begin{lemma}\label{nolargecocycles}
If $Y \sim Y_d(n, \frac{c \log n}{n})$ and $c > d - \frac{1}{2}$ then with high probability $z(X)$ fails to hold for all $(d - 1)$-dimensional complexes on $n$ vertices of size at least $n/(3d)$
\end{lemma}
Similar to the approach in \cite{MW}, but now avoiding having to deal with coefficients, we want to count the number of $(d - 1)$-complexes on $n$ vertices with $b(X) = (1 - \theta) n |X|$. To do so we recall that $\beta(X) \leq b(X)$, and we count the number of complexes with $\beta(X) \leq (1 - \theta)n|X|$. We make use of the following lemma from Meshulam--Wallach.
\begin{lemma}[Claim 4.2 from \cite{MW}]\label{claim42}
Let $0 < \epsilon \leq 1/2$ and then for $n$ large enough and $X$ so that $\beta(X) \leq (1 - \theta)|X|(n - d)$ for some $0 < \theta \leq 1$, there exists a subfamily $S \subseteq X$ of size less than $C \frac{|X|}{n} + 2\log\frac{1}{\epsilon\theta}$ such that $\Gamma(S) := \{ \tau \in X : |\tau \cap \sigma| = d - 1 \text{ for some $\sigma \in S$}\}$ has size at least $(1 - \epsilon)\theta |X|$, where $C$ is a constant depending only on $\epsilon$ and $d$.
\end{lemma}
We are now ready to use Lemma \ref{claim42} to count the number of complexes $X$ with $\beta(X) \leq (1 - \theta)|X|n$. This will upper bound the number of complexes $X$ with $b(X) = (1 - \theta)|X|n$.
\begin{lemma}[Modification to Proposition 4.1 from \cite{MW}]\label{betabound}
For $n$ large enough, $k \geq n/(3d)$, and $\theta \geq 1/(2d)$ there exists a constant $c = c(d)$ so that the number of $(d - 1)$-complexes $X$ with $|X| = k$ and $\beta(X) \leq (1 - \theta)kn$ is at most
$$\left( c n^{(d - 1)(1 - \theta(1 - \frac{1}{2d^2}))} \right)^k$$
\end{lemma}
We give the proof here, essentially as it appears in \cite{MW}, though we omit any consideration of an underlying coefficient ring. The proof follows directly from Lemma \ref{claim42}.
\begin{proof}
For $n$, $k$, and $\theta$, let $\mathcal{F}_n(k, \theta)$ denote the collection of $(d - 1)$-dimensional subcomplexes $X$ of the simplex on $n$ vertices with $k$ facets and $\beta(X) \leq (1 - \theta)kn$.
If $X \in \mathcal{F}_n(k, \theta)$ then $$\beta(X) \leq \left(1 - \frac{\theta n - d}{n - d} \right) k (n - d)$$
Suppose that $\theta \geq 1/(2d)$ and let $\theta' = \dfrac{\theta n - d}{n - d}$ and $\epsilon = 1/(2d^2)$, by Lemma \ref{claim42} when $n$ is large enough we obtain for every $X \in \mathcal{F}_n(k, \theta')$ where $k \geq n/(3d)$ a set $S$ of size at most $C\frac{k}{n}$, for some constant $C$ depending on $\epsilon$,\footnote{This is where we use that fact that $k \geq n/3d$ and $\theta \geq 1/(2d)$. Indeed the $C$ from Lemma \ref{claim42} gives us a bound of $C \frac{|X|}{n} + 2\log \frac{1}{\epsilon \theta} \leq C \frac{|X|}{n} + 2 \log (4d^3) \leq C \frac{|X|}{n} + 6d \log (4d^3) \frac{|X|}{n}$. So the $C$ in this proof should be the $C$ in Lemma \ref{claim42} plus $6d \log (4d^3)$. Of course we could set any $\delta, \theta_0 >0$ and assume that $k \geq \delta n$ and $\theta \geq \theta_0$, but the choices of $1/(3d)$ and $1/(2d)$ respectively are convenient in other parts of our paper.} so that $\Gamma(S)$ has size at least $(1 - 1/(2d^2)) \theta' k$. Thus we get a map taking $X$ in $\mathcal{F}_n(k, \theta)$ to $(S, \Gamma(S), X - \Gamma(S))$. Since the latter two coordinates of this 3-tuple give a partition of $X$, this map is injective, so the cardinality of $\mathcal{F}_n(k, \theta)$ is at most the number of such tuples. Therefore it is at most
\begin{eqnarray*}
\left(\sum_{i = 0}^{Ck/n} \binom{\binom{n}{d}}{i} \right) \left( 2^{\frac{Ck}{n} dn} \right) \left( \sum_{j = 0}^{k - \theta'k(1 - 1/(2d^2))} \binom{\binom{n}{d}}{j}\right)
\end{eqnarray*}
Now the first two factors in the product above are at most $c_1^k$ and $c_2^k$ respectively for some constants $c_1$ and $c_2$ depending on $d$. Thus for $n$ large enough and $k \leq \binom{n}{d}/2$, we have
\begin{eqnarray*}
|\mathcal{F}_n(k, \theta)| \leq (c_1c_2)^k k (1 - \theta'(1 - 1/(2d^2))) \binom{n^d}{k (1 - \theta'(1 - 1/(2d^2)))}
\end{eqnarray*}
Therefore there exists a constant $c$ so that for $n/(3d) \leq k \leq \binom{n}{d}/2$,
\begin{eqnarray*}
|\mathcal{F}_n(k, \theta)| \leq c^k n^{(d - 1) (1 - \theta (1 - \frac{1}{2d^2}))k}
\end{eqnarray*}
This finishes the proof of the lemma in the case that $k \leq \binom{n}{d}/2$. In the case that $k$ is larger than $\binom{n}{d}/2$, we may use the trivial bound of $2^{\binom{n}{d}}$ on $|\mathcal{F}_n(k, \theta)|$ and so there is nothing to prove.
\end{proof}
We are now ready to combine Lemma \ref{betabound} with Lemma \ref{probabilitybound} to show that with high probability $z(X)$ fails to hold for every $(d- 1)$-dimensional complex $X$ on $n$ vertices with at least $n/(3d)$ facets.
\begin{lemma}\label{noZ}
If $Y \sim Y_d(n, \frac{c \log n}{n} \binom{n}{d + 1})$ for $c > d - 1/2$ then with high probability $z(X)$ fails to hold for all $(d - 1)$-dimensional complexes on $n$ vertices of size at least $n/(3d)$.
\end{lemma}
\begin{proof}
For any $(d - 1)$-dimensional subcomplex $X$, we have that the probability of $z(X)$ is at most $n^{-(1 - \theta)(d - 1/2)k}$ where $b(X) = (1 - \theta)nk$ by Lemma \ref{probabilitybound}. Now if we define $f_n(k, \theta)$ to be the number of $X$ with $|X| = k$ and $b(X) = (1 - \theta)nk$, then we wish to show that
$$\sum_{k \geq n/(3d)} \sum_{\theta \in \Upsilon} f_n(k, \theta) n^{-(1 - \theta)(d - 1/2)k} = o(1),$$
where we set $\Upsilon = \{\theta : (1 - \theta)nk \in \Z, (1 - \theta)nk \leq n^{d + 1}, \theta \leq d/(d + 1)\}.$
For $\theta \in \Upsilon_1 := \{ \theta \in \Upsilon : \theta < 1/(2d)\}$, we cannot apply Lemma \ref{betabound}, but the trivial bound on $\sum_{\theta} f(k, \theta) \leq \binom{n^d}{k}$ works instead. Indeed we have,
\begin{eqnarray*}
\sum_{k \geq n/(3d)} \sum_{\theta \in \Upsilon_1} f(k, \theta) n^{-(1 - \theta)(d - 1/2)k} &\leq& \sum_{k \geq n/(3d)} n^{(d-1)k} n^{-(1 - 1/(2d))(d - 1/2)k}\\
&\leq& \sum_{k \geq n/(3d)} n^{-k/(4d)} = o(1).
\end{eqnarray*}
For $\theta \in \Upsilon_2 := \{ \theta \in \Upsilon : \theta \geq 1/(2d)\}$ we will apply Lemma \ref{betabound}, as $b(X) = (1 - \theta)nk$ implies that $\beta(X) \leq (1 - \theta)nk$ so $f_n(k, \theta) \leq |\mathcal{F}_n(k, \theta)|$. We have
\begin{eqnarray*}
\sum_{k \geq n/(3d)}
\sum_{\theta \in \Upsilon_2}
f(k, \theta) n^{-(1 - \theta)(d - 1/2)k}
&\leq&
\sum_{k = n/(3d)}^{n^d}
\sum_{\theta \in \Upsilon_2}
\left( c n^{(d - 1)(1 - \theta(1 - \frac{1}{2d^2}))} \right)^k n^{-(1 - \theta)(d - 1/2)k} \\
&\leq&
\sum_{k = n/(3d)}^{n^d}
\sum_{\theta \in \Upsilon_2}
\left(c n^{-\frac{1}{2} + (\frac{1}{2} + \frac{1}{2d} - \frac{1}{2d^2})\theta} \right)^k \\
&\leq&
\sum_{k = n/(3d)}^{n^d}
\sum_{\theta \in \Upsilon_2}
\left(c n^{-\frac{1}{2} + (\frac{1}{2} + \frac{1}{2d} - \frac{1}{2d^2})\frac{d}{d + 1}} \right)^k \\
&\leq&
\sum_{k = n/(3d)}^{n^d}
\sum_{\theta \in \Upsilon_2}
\left(c n^{-\frac{1}{2d(d + 1)}} \right)^k \\
&\leq& n^{d + 1}
\sum_{k = n/(3d)}^{n^d}
\left(c n^{-\frac{1}{2d(d + 1)}} \right)^k
\leq n^{2d + 1}n^{-\Theta(n)}
= o(1).
\end{eqnarray*}
\end{proof}
Finally, we use a simple coupling argument to finish the proof of Lemma \ref{nolargecocycles}.
\begin{proof}[Proof of Lemma \ref{nolargecocycles}]
Fix $c' \in (d - 1/2, c)$. Let $Y_1 \sim Y_d(n, \frac{c' \log n}{n} \binom{n}{d + 1})$. With high probability $Y_1 \subseteq Y$. Indeed, by a routine application of Chernoff's bound, if $Y$ is distributed as $Y_d(n, \frac{c \log n}{n})$ then the probability that $Y$ has fewer than $\frac{c' \log n}{n} \binom{n}{d + 1}$ faces is at most
$$\exp \left( - \frac{ \frac{\log n}{n} \binom{n}{d + 1} (c - c')^2}{2c} \right) = o(1).$$
And by Lemma \ref{noZ}, with high probability $z(X)$ fails to hold for all $X$ on $n$ vertices of size at least $n/(2d)$ in $Y_1$, and hence the same holds in $Y$ since the kernel of the $d$th coboundary map of $Y$ is contained in the kernel of the $d$th coboundary map of $Y_1$.
\end{proof}
\section{Small cocycles}\label{smallcocycles}
To show that the small cocycle condition holds for $Y \sim Y_d(n, \frac{c \log n}{n})$ with high probability for $c > d - 1/2$ holds we rely on the fact that the support of an inclusion minimal cocycle is strongly-connected.
Strongly connected $(d - 1)$-complexes with $k$ facets on $n$ vertices
are relatively few, in comparison to those that are not strongly connected.
Specifically, we have the following:
\begin{lemma}\label{stronglyconnected}
The number of strongly-connected $(d - 1)$-complexes with $k$ facets on $n$ vertices is at most
\[
n^{d+k-1} (2d)^{k}.
\]
\end{lemma}
\begin{proof}
A complex $X$ is strongly connected if and only if the dual graph of $G(X)$ is a connected subgraph of the dual graph $H$ of the $(d-1)$--skeleton of the full simplex on $n$ vertices. As each $G(X)$ has a spanning tree, we estimate the number of $X$ above by the number of rooted subtrees of $H$ with $k$ vertices. There are at most
\[
\binom{n}{d} \cdot 2^{k-1} \cdot (dn)^{k-1}
\]
such rooted subtrees, enumerated in breadth--first--search order. In this enumeration, the $\binom{n}{d}$ counts the number of choices for the root, the $2^{k-1}$ counts the ways to partition the remaining $k-1$ into the sizes of the neighborhoods of each vertex in the breadh--first--search, and the $(dn)^{k-1}$ overestimates the number of ways to pick the neighborhoods.
\end{proof}
\begin{lemma}
If $c > (d + 1)/2$ then with high probability $Y \sim Y_d(n, c \log n / n)$ has no inclusion-minimal $(d-1)$-cocycles of support size $k$ over any field for $2 \leq k \leq \log n$.
\end{lemma}
\begin{proof}
We use the first moment method. If $X$ is the support of an inclusion-minimal cocycle of $Y$ over some field then $X$ is a strongly-connected $(d - 1)$ complex. Moreover, if $|X| = k \leq \log n$ then at least $n - d \log n$ vertices do not belong to $X$. Now if $X$ is to be the support of a cocycle over any field, then any face $\tau$ obtained as the union of a $(d - 1)$-dimensional face of $X$ and a vertex outside of $X$ must be excluded from $Y$. Thus the probability that a fixed $X$ of size $k \leq \log n$ is the support of a cocycle over any field is at most $(1 - p)^{k(n - d \log n)}$.
Since $X$ must be strongly connected, the number of choices for $X$ with size $k$ is at most $(2d)^k n^{d - 1 + k}$ by Lemma~\ref{stronglyconnected}. Applying the union bound over $k \in \{2, 3, ..., \log{n}\}$ for the probability that there exists an inclusion-minimal cocycle of support size $k$ we have:
\begin{eqnarray*}
\sum_{k = 2}^{\log{n}} (2d)^kn^{d - 1 + k} (1 - p)^{kn(1 - o(1))} &\leq& n^{d - 1} \sum_{k = 2}^{\log{n}}(2d)^k n^k n^{-ck(1 - o(1))} \\
&\leq& n^{d - 1} (\log{n}) n^{2 - 2c + o(1)} \\
&=& (\log{n}) n^{d - 1 - 2(c - 1) + o(1)} = o(1).
\end{eqnarray*}
\end{proof}
\begin{lemma}
If $c > 3/2$, then with high probability $Y \sim Y_d(n, c \log n /n)$ has no inclusion minimal $(d - 1)$-cocycles of support size $k$ over any field for $\log{n} \leq k \leq n/(3d)$.
\end{lemma}
\begin{proof}
If $X$ is the support of such a cocycle then there are at least $n - n/3 = 2n/3$ vertices of $Y$ outside of $X$. Taking this consideration and the same argument as the proof of $k \leq \log{n}$, we bound the following to prove the lemma:
\begin{eqnarray*}
\sum_{k = \log{n}}^{n/(3d)} n^{d - 1 + k} (1 - p)^{2kn/3} &\leq& n^{d - 1} \sum_{k = \log{n}}^{n/(3d)} n^k n^{-2ck/3} \\
&\leq& n^d (n^{2c/3 - 1})^{-\log{n}}\\
&=& n^{-\Theta(\log{n})}.
\end{eqnarray*}
\end{proof}
This finishes the proof of Lemma \ref{keylemma}, and hence the proof of our main result.
\section{Conclusion}
Our result finally establishes $p = \frac{d \log n}{n}$ as the sharp threshold for homological connectivity of $Y_d(n, p)$. Moreover, Corollary \ref{isolatedfaces} tells us about the structure of the $(d - 1)$st homology group immediately before it vanishes. However, the following two questions are closely related to our main result and remain open.
\begin{itemize}
\item What is the homological connectivity threshold for the random hypergraph model? This model is similar to the Linial--Meshulam model except that one does not start with the complete $(d - 1)$-skeleton. Rather the $d$-dimensional faces are included independently and the complex is obtained by taking the downward closure of the top-dimensional faces. In \cite{CdGKS}, Cooley et al.\ show the hitting-time result for homological connectivity with $\Z/2\Z$ coefficients in the random hypergraph model. Their result establishes that the sharp threshold for homological connectivity with $\Z/2\Z$-coefficients for the random hypergraph model is at $\dfrac{d \log n}{2n}$. Can our methods be adapted to obtain the corresponding result with integer coefficients?
\item The question about torsion in homology of $Y_d(n, p)$ is raised in \cite{KLNP, LP2}. Namely, experimental evidence strongly suggests that shortly before the first nontrivial cycle\footnote{That is the first top homology class not generated by boundaries of the $(d + 1)$-simplex} appears in the \emph{top} homology group of $\Y_d(n)$, there is an exceptionally large (on the order of $\exp(\Theta(n^d))$) torsion group which appears in the $(d - 1)$st homology group. Outside of this however, it is believed that $\Y_d(n)$ has no torsion in homology; \cite{LP2} formulates this conjecture precisely. Our paper in fact grew out of an attempt to prove the stronger result that for $c$ a sufficiently large constant and $p = c/n$, one has that with high probablity the $(d - 1)$st homology group of $Y_d(n, p)$ is torsion free. However this problem remains open.
\end{itemize}
|
\section{Introduction}
Configuration space is one of the most useful constructions in algebraic topology.
In particular, configuration space of unordered points in a topological space has been shown to
be a nice model for homotopical approximation to mapping spaces.
It dates back to the works of Milgram\cite{milgram-iterated-loop-spaces}, May\cite{may-geometry-of-iterated},
and Segal\cite{segal-configuration-spaces-and},
which states that unordered points in $\RR^n$ with labels in a based space $X$ is weakly homotopy equivalent to $\Omega^n \Sigma^n X,$
if $X$ is path-connected. Then it is generalized to a setting with a manifold replaces
$\RR^n,$
and with a partial monoid replaces a space $X$ as labels
\cite{bodigheimer-stable-splittings-of}, \cite{macduff-configuration-spaces},
\cite{salvatore-configuration-spaces-with}, \cite{shimakawa-configuration-spaces-with}.
Based on these works and others, a new notion called factorization homology is
developed recently \cite{ayala-francis-factorization-homology}(see
also \S 6 of \cite{knudsen-configuration-spaces-in}).
We consider a configuration space of intervals in $\RR$, in which two
intervals are pasted when meeting endpoints have opposite properties,
that is, one is open and the other is closed,
and a half-open interval annihilates when its length approaches zero.
These transformations of intervals are called cutting-pasting and
creation-annihilation, respectively.
If we attach labels on these intervals to consider a configuration
space of intervals in $\RR$
with labels in some based space $X,$ then
we obtain a space which is
weakly homotopy equivalent to $\Omega\Sigma X,$
the space of based loops in the reduced suspension of X \cite{okuyama-space-of-intervals}.
In this paper, we consider a configuration space of intervals with
labels in a partial abelian monoid $M.$
We can use partial sum to define a sort of interaction of intervals.
So our configuration space has as its topology one which reflects this
interaction as well as transformations
come from cutting-pasting and creation-annihilation mentioned above.
The purpose of this paper is to construct such a space
and show that there exists a weak homotopy
equivalence analogous to \cite{okuyama-space-of-intervals}.
When we regard a based space $X$ as a topological partial abelian
monoid by a trivial sum, then the reduced suspension
construction can be considered as a special case of
the classifying space construction
$BM$ of a partial abelian monoid $M$\cite{segal-configuration-spaces-and}.
We can state our theorem as follows.
\begin{thm}\label{thm:main}
Let $M$ be a (discrete) partial abelian monoid whose elements are self-insummable.
Then the configuration space of intervals in $\RR$ with labels in $M$
is weakly homotopy equivalent to $\Omega B M.$
\end{thm}
Self-insummability of a partial abelian monoid which appears in the above theorem
is explained in \S 2. This condition is used in Lemma \ref{lem:unique-representation}.
This paper is organized as follows.
In \S 2.1, we give a definition of partial abelian monoid. In \S 2.2, we give a definition of a product of
partial abelian monoids, which is shown to be useful for constructing configuration spaces.
Configuration spaces of intervals $I$ and $I_M$ are defined in \S 3.1 and \S 3.2, respectively.
To prove the weak equivalence, we construct a thickening $\widetilde{I}_M$ of $I_M$ and
maps
$$
I_M \leftarrow \widetilde{I}_M \to \Omega BM,
$$
which induce isomorphisms on homotopy groups.
The thickening and other auxiliary spaces are defined in \S 3.3.
In \S 4, we construct of a map $\alpha : \widetilde{I}_M \to \Omega B M.$
Then in \S 5, we give a proof of the main theorem.
In this paper, intervals in $\RR^1$ with two different roles occur.
One is as an object of which configuration is to be considered and the other is as just
a subset of $\RR^1$ for usual usage, that is, a domain of configuration, a domain of definition of a function,
or an interval for a homotopy. For the latter, we use a standard notation such as
$ ( a , b ] = \{ x ~|~ a < x \leq b \}.$ For the former, we make some convention based on
another standard notation such as $ ] a , b ] = \{ x ~|~ a < x \leq b \}.$ See \S 3.1 for the convention.
We work in the category of compactly generated weak Hausdorff spaces and
base points are assumed to be non-degenerate.
\section{Partial abelian monoid}
\subsection{Partial abelian monoid}
The notion of partial abelian monoid plays an important role in our work.
It is an abelian monoid with partially defined sum, whose definition is given
as follows :
\begin{defn}\label{defn:pam}
A {\bf topological partial abelian monoid} is a space $M$ with base point $0$
equipped with a subspace
$M_2$ of $M\times M$
and a map $m : M_2\to M$ which satisfies
\begin{enumerate}
\item $M\vee M\subset M_2,$ and $m(a,0)=m(0,a)=a$,
\item $(a,b)\in M_2$ if and only if $(b,a)\in M_2,$ and $m(a,b)=m(b,a),$
\item $(m(a,b),c)\in M_2$ if and only if $(a,m(b,c))\in M_2,$ and $
m(m(a,b),c)=m(a,m(b,c)) .$
\end{enumerate}
We denote $m(a,b)=a+b.$ By associativity, we can speak of
summable $n$-tuples in $M^n$, and we denote the subset of summable $n$-tuples by $M_n.$
\end{defn}
A partial abelian monoid $M$ is said to be self-insummable if every non-zero element
is self-insummable, that is, $\Delta \cap M_2 = \{ ( 0, 0 ) \}.$
Partial abelian monoid is also called partial commutative monoid.
Abelian partial monoid defined in \cite{shimakawa-configuration-spaces-with} is the same concept in philosophy,
but it is more general than the partial abelian monoid given above.
\subsection{A product of partial abelian monoids}
For a topological space $Y,$ let $\multi{Y}=\coprod_{n\geq 0}\sympro^n{Y}.$
It can be considered as the free abelian monoid generated by $Y_+=Y\cup \{0\}$ with an appropriate
topology, or equivalently, as $\symproinf{Y_+}$, an infinite symmetric product introduced in
\cite{dold-thom-quasifaserungen-und-unendliche}.
In this paper, however, we often treat an element of $\multi{Y}$
as a finite multiset --- a finite ``set'' with
repeated elements~\cite{stanley-enumerative-combinatorics}. The notions such as
cardinality, submultiset or union of multisets are defined so as to respect multiplicities
of elements. We denote by $ x_1\dotplus \dots \dotplus x_n$ to mean the multiset consisting of
elements $x_1,\dots, x_n$ where each element is repeated a number of times equal to
its multiplicity. The union of multisets $\alpha$ and $\beta$ is denoted by $\alpha \dotplus \beta.$
Thus if $\alpha = a_1\dotplus\dots \dotplus a_r$ and $\beta = b_1\dotplus \dots \dotplus b_s$ then
$\alpha \dotplus \beta = a_1\dotplus\dots\dotplus a_r\dotplus b_1\dotplus \dots\dotplus b_s.$
To give an element of $\multi{Y}$ is equivalent to giving a map $\sigma : S\to Y,$
where $S$ is a finite set. In this form, a submultiset of $\sigma$ can be given as a restriction
map $\sigma|T : T \to Y$ to a subset $T$ of $S.$ If $M$ is a partial abelian monoid
and $\sigma = m_1\dotplus\dots\dotplus m_r \in \multi{M},$
we say that $\sigma$ is summable if $(m_1,\dots, m_r)$, in any order, is a summable $r$-tuple.
Otherwise, we say that $\sigma$ is insummable. On the other hand, we say that $\sigma$ is
pairwise insummable if, for any subset $T\subset S$ of cardinality two, $\sigma|T$ is insummable.
Let $M$ and $N$ be partial abelian monoids.
We denote by $p_1 : M\times N \to M$ and $p_2 : M\times N \to N$ the projections from $M\times N$
onto its first and second factors respectively.
Let $\sigma \in \multi{M\times N}$ and $\sigma : S \to M\times N$ be its representation as a map.
Consider the following property for $\sigma$ : for any subset $T$ of $S,$
if one of $p_i\circ (\sigma|T)$ is pairwise insummable then
the other is summable.
We denote by $\pretensor{M,N}$ the subspace of $\multi{M\times N}$ consisting of $\sigma$ with
this property.
Let $\sim$ be the least equivalence
relation on $\pretensor{M,N}$ which satisfies the following three conditions:
\begin{enumerate}
\item[(R1)] If $m_1$ or $n_1$ is zero then
$$ (m_1,n_1)\dotplus\dots\dotplus(m_r,n_r) \sim (m_2,n_2)\dotplus\dots\dotplus (m_r,m_r),$$
\item[(R2)] If $m_1 = m'_1 + m''_1$ then
$$ (m_1,n_1)\dotplus\dots\dotplus (m_r , n_r )
\sim (m'_1,n_1)\dotplus(m''_1, n_1)\dotplus (m_2, n_2)\dotplus\dots \dotplus (m_r , n_r),$$
\item[(R3)] If $n_1=n'_1 + n''_1$ then
$$ (m_1,n_1)\dotplus\dots\dotplus (m_r , n_r )
\sim (m_1,n'_1)\dotplus (m_1, n''_1)\dotplus (m_2, n_2) \dotplus \dots \dotplus (m_r , n_r).$$
\end{enumerate}
Let $T_k$ be the subspace of $\pretensor{M, N}$
of elements of cardinality less than or equal to $k.$ Then the relation $\sim$ on
$\pretensor{M,N}$ induces a relation on $T_k.$
Then we denote by $M\otimes N$ the union $\bigcup_{k\geq 0} (T_k / \sim)$
of quotient spaces. Let $\pi_\otimes : \pretensor{M,N} \to M\otimes N$ be the natural map.
Two elements $[\alpha], [\beta]$ in $M\otimes N$ are summable if we can choose
their representatives $\alpha, \beta$ in $\pretensor{M,N}$ so that their sum $\alpha \dotplus \beta$ taken in $\multi{M\times N}$
is contained in $\pretensor{M,N}.$
Thus, $M\otimes N$ is a partial abelian monoid in a natural way.
The product of partial abelian monoids defined above covers diverse examples.
\begin{examples*}
A based space $X$ can be regarded as a trivial partial abelian monoid by setting
$X_2 = X\vee X$ and $\mu : X\vee X \to X$ the folding map. Then
$X\otimes M$ is the configuration space of finite points in $X$ with labels in $M$
such that only summable labels occur simultaneously.
\begin{enumerate}
\item For two based spaces $X, X' , $
their product $X\otimes X'$ coincides with their smash product
$X \wedge X'.$
\item \label{ex:classifying}Viewing $S^1$ as a based space, we get
$S^1\otimes M = BM$ the classifying space defined in
\cite{segal-configuration-spaces-and}. In particular, if $M$ is a monoid this coincides with the
McCord model of the classifying space\cite{mccord-classifying-spaces-and}.
\item Let $X$ be a compact based space and $M = Gr := \sqcup Gr_n(\RR^\infty)$
be the infinite Grassmannian with a partial sum defined only for two vector spaces
which are perpendicular to each other. Then
$X\otimes Gr = F(X)$ coincides with the configuration space stated in
\cite{segal-k-homology-theory}. This configuration space realizes connective $K$-homology.
In \cite{tamaki-twisting-segal's-k-homology}, a similar construction is given, but it is enriched by an operad
to make twisting on $K$-theory, thus larger than $X\otimes Gr.$
\end{enumerate}
We denote by $\fin{Y}$ the set of finite subsets of a space $Y$ and
give it a topology by
$$\fin{Y} = \{\emptyset \} \sqcup Y \sqcup (Y^2\setminus\Delta)/\Sigma_2
\sqcup (Y^3 \setminus \Delta_3)/\Sigma_3 \sqcup \cdots ,$$
where $\Delta_n\subset Y^n$ is the fat diagonal.
We regard $\fin{Y}$ as a partial abelian monoid by disjoint union of sets.
We can realize configuration spaces in which non-summable labels can occur simultaneously,
by using $\fin{Y}.$
\begin{enumerate}
\setcounter{enumi}{3}
\item \label{ex:config} If $C_n = \fin{\RR^n}$ then
$C_n \otimes X = C_n(X)$
is the configuration space of finite points in $\RR^n$ with labels in $X$
\cite{segal-configuration-spaces-and}.
\item \label{ex:shimakawa}$\fin{Y}\otimes M = C^M(Y)$ is
the configuration space of finite points in $Y$ with labels in $M$
defined in \cite{shimakawa-configuration-spaces-with}. Note that in \cite{shimakawa-configuration-spaces-with}, the notion of
abelian partial monoid is more general concept than our notion of partial
abelian monoid.
\end{enumerate}
Among examples in an extreme case, when $M$ or $N$ is a monoid, we have:
\begin{enumerate}
\setcounter{enumi}{5}
\item $ X\otimes \NN = \symproinf{X}$, the infinite symmetric product on a based space $X$ introduced
in \cite{dold-thom-quasifaserungen-und-unendliche}.
\item For abelian groups $A, B, $ their product $A\otimes B$ defined here
is the usual tensor product of modules.
\end{enumerate}
\end{examples*}
\section{Configuration space of intervals}
\subsection{Configuration space of intervals without labels}
Let $H$ be the half-plane in $\RR^2$ given by
$ H = \{ (u, v) \in \RR^2 ~|~ u \leq v\}$
and $P = \{ \pm 1\}$ be the set of ``parities'' considered as a space with discrete
topology. To any point $( u , v ; p , q )$ in the direct product $H \times P^2$
with $u < v,$
we assign an interval
$$
J = \{ x \in \RR ~|~ u <_p x <_q v\}
$$
in $\RR,$ where the symbol
``$<_p$'' denotes the inequality ``$\leq$'' if $p = +1$ and ``$<$'' if $p = -1.$
We also denote $J = \intervalconvention{p}{a , b}{q}.$
If $u = v,$ it is not an interval anymore, but we allow it only if
$p\neq q$ and call it a degenerate interval. For notational convention, let $\bar{p} = -p$ for any $p \in P.$
Let
$$
\interval = \{ (u , v ; p, q ) \in H\times P^2~|~u \leq v , ~p\neq q \mbox{~if~}u = v\}
$$
and we identify the element of $\interval$ with the interval assigned to it.
Let $u_L$ and $u_R$ denote the projections $\interval \to \RR$ onto its first and
second factors respectively, and
$p_L$ and $p_R$ denote the projections onto its third and fourth projections $\interval \to P$ respectively.
Under the identification stated above, $u_L(J)$ or $u_R(J)$ is the coordinate of the left or right
endpoint of $J$, respectively, while $p_L(J)$ or $p_R(J)$ is the parity of the left or right endpoint,
respectively.
For any two elements $J_1, J_2 \in \interval,$
we denote $J_1 \leq J_2$ if either (1) $u_R( J_1 ) < u_L(J_2)$ or (2) $u_R(J_1) \leq u_L(J_2)$ and $p_R(J_1) \neq p_L(J_2).$
We denote $J_1 < J_2$ if $u_R(J_1) < u_L(J_2).$
Let $L_r$ be the subspace of $\interval^r$ given by
$$
L_r = \left\{ ( J_1 , \dots , J_r ) \in \interval^r~\left|~\begin{array}{l}
J_1\leq \dots \leq J_r
\end{array}\right\}\right. _,
$$
Then $L_r$ is the configuration space of $r$ bounded intervals in $\RR$
such that endpoints of the same parity do not collide, with additional limiting
points of creation and annihilation provided by degenerate intervals.
Later, these additional points are identified with the
empty configuration. Meanwhile, if endpoints of two distinct intervals, one open and the other closed,
collide then two meeting intervals are pasted and the configuration of these two intervals
is identified with the configuration of one interval.
To establish these identifications,
let $\sim$ be the least equivalence relation on $\bigsqcup_{r\geq 0} L_r$ which satisfies
the following two relations :
\begin{enumerate}
\item $( J_1, \dots, J_r ) \sim (J_1, \dots, J_{i-1}, J_{i+1}, \dots , J_r ) $ if $u_L(J_i) = u_R(J_i).$
\item $( J_1, \dots, J_r ) \sim (J_1, \dots, J_{i-1}, K , J_{i+2} ,\dots , J_r ) $ if $u_R(J_i) = u_L(J_{i+1}),$
where $K = ( u_L( J_i ) , u_R(J_{i+1}), p_L( J_i ), p_R( J_{i+1}) ).$
\end{enumerate}
Let $I$ be the quotient space of $\bigsqcup_{r\geq 0} L_r$ by this equivalence relation.
The image of $L_0$ is a single point and denoted by $\emptyset,$ which is
treated as the base point of $I.$
Let $L'_r$ be the subspace of $L_r$ given by
$$
L'_r = \left\{ ( J_1 , \dots , J_r ) \in \interval^r~\left|~\begin{array}{l}
J_1< \dots < J_r, \\
u_L(J_i) < u_R(J_{i}) \mbox{~for each~} i = 1,\dots ,r
\end{array}\right\}\right. _.
$$
As sets there is a bijective correspondence between
$\bigsqcup_{r\geq 0} L'_r$ and $I.$ We call
an element in $L'_r$ corresponding to an element $\xi$ of $I$
the {\bf reduced representative} of $\xi.$
Let $\pi_\interval$ denote the composite $\interval \twoheadrightarrow L_1 \hookrightarrow I.$
\subsection{Configuration space of intervals with partially summable labels}
Let $U = (a, b)$ be an open interval in $\RR.$ We consider two special
types of elements in $\multi{\interval \times M}.$
\begin{enumerate}
\item[(E1)] $e = ( J , n ) $ with one of the following :
\begin{enumerate}
\item $J = ] a, b [$
\item $J = ] a, w [$ or $J = ] a, w ],~ a < w < b$
\item $J = ] w, b [$ or $J = [ w, b [,~ a < w < b$
\item $J = \intervalconvention{p}{w_1, w_1}{q} ,~ a < w_1 < w_2 < b, p+q = 0$
\end{enumerate}
\item[(E2)] $e = ( J_1 , n ) \dotplus ( J_2 , n )$ with
$J_1 = \left] a , w_1 \right|_p , J_2 =\,_q| w_2 , b [ $ and $a\leq w_1 < w_2 < b$, $p+q=0$
\end{enumerate}
where $n$ is a non-zero element in $M,$ which is denoted by $n(e).$
We call such $e$ {\bf an elementary configuration in $U$.}
Let $ ( J_1, m_1 ) \dotplus \dots \dotplus ( J_r, m_r ) $ be a representative of $\xi \in \multi{\intM}$.
Suppose that $i_1, \dots , i_\lambda$ are all the subscripts $i$ such that $J_i \cap U \neq \emptyset.$
Then let $\xi|_U$ denote the multiset in $\multi{\intM}$ represented by
$ ( J_{i_1}\cap U , m_{i_1} )\dotplus \dots\dotplus (J_{i_\lambda}\cap U , m_{i_\lambda}) ).$
Let $\pi_{mul} : \multi{\intM} \to \multi{I\times M}$ be the map induced by
$\pi_\interval\times id_M : \intM \to I\times M$ and
let $\pretensor{\interval , M} = \pi_{mul}^{-1}(\pretensor{I, M}) \subset \multi{\interval\times M}.$
An element $\xi \in \pretensor{\interval, M}$ is said to be
{\bf admissible} if for any $t\in \RR$ there exists an open interval $U= (a,b)$ which contains
$t$ such that $\xi|_U$ is represented by $e_1 \dot{+} \dots \dot{+} e_r$ for some elementary
configurations $e_1, \dots , e_r$ in $U$ such that $(n(e_1), \dots , n(e_r)) \in M_r.$
In this case, such a representative
$e_1 \dot{+} \dots \dot{+} e_r$ is called an admissible sum of elementary configurations.
If, moreover, there exist $\ep>0$ and an interval $U$ can be taken as $U = (t-\ep, t+\ep)$ for all $t$,
then we say that $\xi$ is {\bf $\ep$-admissible}.
Let $V = ( a, b )$ be an open interval with $b-a > \ep.$ We say that an $\ep$-admissible
element $\xi$ is supported by $V$ if $\xi |_{(a+\ep/2, b-\ep/2)} = \xi.$
Let $W, W(\ep)$, and $W(\ep, V)$ be the subspace of $\pretensor{\interval, M}$ which consists of
admissible elements, $\ep$-admissible elements, and $\ep$-admissible elements supported by $V$,
respectively.
If $\ep > \ep'$ and $V\subset V',$ then we have a natural inclusion $W(\ep, V) \subset W(\ep', V').$
Let $I_M$ be the image in $I\otimes M$ of $W$
under the natural map $\pi_\otimes \circ \pi_{mul}.$
Let also $I_M(\ep)$ and $I_M(\ep, V)$ be the image in
$I\otimes M$ of $W(\ep)$ and $W(\ep, V)$, respectively ,
under $\pi_\otimes\circ \pi_{mul}.$
We call
$$I_M = \bigcup_{\ep>0, V} I_M(\ep, V).$$
a {\bf configuration space of intervals with partially summable labels,}
where its topology is the weak topology of the union.
\subsection{Thickening, Moore type variant, and the total space}
If $V = ( 0 ,s )$ then let $I_M( \ep , s ) = I_M ( \ep , V )$ and $W( \ep, s ) = W( \ep, V ).$
We define
\[
\widetilde{I}_M = \bigcup_{\ep > 0, s \geq 0} I_M(\ep, s ) \times \{\ep\}\times\{s\}
\]
and give it the topology as a subspace of $I_M \times \RR^2.$
If $s \leq \ep,$
$I_M(\ep, s )$ consists of one point, the element $\emptyset$ in $I_M$ which represents
the empty configuration. As a base point of $\widetilde{I}_M,$ we take
$(\emptyset, 1,0).$
\begin{prop}
The projection $p : \widetilde{I}_M\to I_M$ onto the first component is a weak homotopy equivalence.
\end{prop}
\begin{proof}
Let $f : S^n \to I_M$ be a map which represents an element $\alpha \in \pi_n(I_M).$
The image of $f$ is a compact (Hausdorff) subspace of $I_M,$
since $S^n$ is compact and all our spaces are weak Hausdorff and compactly generated
\cite{mccord-classifying-spaces-and}.
Then we can find $\ep >0$ and $s>0$ such that $I_M(\ep , s )$ contains the image of $f$ (
weak Hausdorff case of Lemma 9.3 in \cite{steenrod-convenient-category-of}).
However, we have a homotopy equivalence
$I_M(\ep ,s ) = I_M(\ep, s ) \times \{\ep\} \times \{s\}
\hookrightarrow \widetilde{I}_M.$
Composing this inclusion, we get a map
$f' : S^n \to \widetilde{I}_M$ which maps to $\alpha$ under $p_*.$
This proves that $p_* : \pi_n(\widetilde{I}_M)\to \pi(I_M)$ is surjective.
For injectivity, suppose $\alpha = [f] \in \pi_n(\widetilde{I}_M)$ maps to
$\pi_*(\alpha) = [g] \in\pi_n(I_M),$ which is zero. Let $H : S^n\times I \to I_M$ be a homotopy
between $g$ and $*.$ Then by the compactness, we
can assume that the image of $H$ is contained in $I_M(\ep , s )$ for some $\ep $ and $s.$
It is similar to the previous case, to show that there exists a map $H' : S^n\times I \to \widetilde{I}_M.$
$H'$ gives a null homotopy of some map $f' : S^n \to \widetilde{I}_M,$ which is homotopic to $f.$
\end{proof}
We proceed to define $E_M.$
For any $J = \intervalconvention{p}{u, v}{q}$ in $\interval,$ we define its ``mirror image'' $J^\mu$ as
$J^\mu = \intervalconvention{\bar{q}}{-v,-u}{\bar{p}}.$ This defines
an involution on $\interval,$ which induces an involution on $I_M,$ which is also denoted by $\mu.$
The space of fixed points $E_M = I_M^\mu$ can be considered as the space of
intervals in $[0, \infty )$ such that $0\in \RR$ works as a ``vanishing point.''
For $s > 0,$ let $E_M(\ep, s) $ be the space $E_M(\ep, s) = E_M \cap I_M(\ep, (-s, s)).$
Then we define
\[
\widetilde{E}_M = \bigcup_{\ep > 0, s \geq \ep} E_M(\ep, s ) \times \{\ep\}\times\{s\} \subset E_M \times \RR^2
\]
and give it the relative topology.
If $s = \ep,$
$E_M(\ep, \ep )$ contains only one point, the element $\emptyset$ in $E_M$ which represents
the empty configuration. As a base point of $\widetilde{E}_M,$ we take
$(\emptyset, 1,1).$
\begin{prop}
$\widetilde{E}_M$ is contractible.
\end{prop}
\begin{proof}
When $h : \RR\to \RR$ is some increasing function, we put $h(J) = \intervalconvention{p}{h(u) , h(v)}{q}$
for any interval $J = \intervalconvention{p}{u, v}{q}\in \interval.$
If $h(u) = h(v)$ then $h(J)\in \interval$ represents a (degenerate) interval only if $p\neq q.$
Let $h_t : ( -s, s ) \to ( -s, s ) ~(0 \leq t\leq 1)$ be the homotopy defined by
$$
h_t( u ) = \left\{
\begin{array}{cc}
u- ts & u\geq ts\\
0 & |u| < ts\\
u + ts & u\leq -ts
\end{array}
\right.
$$
Then we define a homotopy $H_{t} : E_M(\ep, s) \to E_M(\ep, s)$ as follows. If
$\xi = (J_1, m_1)\dotplus \dots\dotplus (J_r, m_r) \in \multi{\interval \times M}$ is a
$\mu$-invariant representative
of $\xi \in E_M(\ep, s)$ then $H_t(\xi)$ be the element of $E_M(\ep,s)$ represented by
$$
H_t(\xi) = (h_t(J_1), m_1)\dotplus \dots\dotplus (h_t(J_r) , m_r).
$$
For each $i,$ $h_t(J_i)$ is at least a degenerate interval since our representative is $\mu$-invariant.
It is easy to see that $H_t$ defines a contraction of $\widetilde{E}_M.$
\end{proof}
We define $d : \multi{\interval \times M} \to \multi{\interval \times M}$ by
$$d( ( J_\lambda, m_\lambda )_\lambda ) = ( J_\lambda, m_\lambda )_\lambda \dotplus ( J^\mu_\lambda, m_\lambda )_\lambda. $$
By restricting $d$ to $\adm{\ep, s},$ we define an embedding $I_M(\ep,s) \to E_M(\ep, s)$ then an embedding $\widetilde{I}_M \to \widetilde{E}_M.$
In the following sections, we only give definitions and proofs for the case $\ep = 1.$
Required changes for other cases are obvious.
\section{Construction of $\alpha : \widetilde{I}_M \to \Omega' BM$}
We are going to define a map $\alpha : \widetilde{I}_M \to \Omega' BM$
in the following steps:
Let $U_t = ( t - 1, t + 1 )$ and $V_t = ( t -1/2, t + 1/2 )$ and
$\adms = \adm{1,s} \times \{s\}.$
\begin{enumerate}
\item We can define a map $\omega(J) : V_t\to S^1$ for each interval $J\in (H\cap U^2_t)\times P^2.$
Then we define
$$\omega_t : \adms \to \multi{\map{V_t, S^1} \times M}.$$
Let $G$ denote the image of $\adms$ under $\omega_t.$
\item \label{stepmu}
We can define a map $s_t : G \to \multi{\map{V_t, S^1} \times M}.$
We denote its image by $G'.$
\item $G'$ is mapped into $\map{V_t, \pretensor{S^1, M}}$ under the sequence of natural maps
\begin{eqnarray*}
\multi{\map{V_t, S^1} \times M} & \to & \multi{\map{V_t, S^1\times M} } \\
& \to & \map{V_t, \multi{S^1 \times M} } .
\end{eqnarray*}
\item Composition of maps given in (1)$\sim$(3) and the map induced by
the quotient $\pretensor{S^1, M} \to S^1\otimes M = BM$
gives us a map
$\alpha_t : \adms \to \map{V_t, BM}.$
\item $\alpha_t ( [\xi]) \in \map{ V_t, \pretensor{S^1, M}}$
for two distinct $t\in [0, s]$ coincide on their intersection of the domain of definition, so we
can achieve extension to get an element in
$\Omega'_s\pretensor{S^1, M}.$ Thus we get a map
$$ \adms \to \Omega'_s BM$$
\item We can take quotient on the source to define a map
$\alpha(\ep, s) : I_M(\ep, s) \to \loopsBM.$
\item We consider all the $s > 0$ and
take a union to get a continuous map $\alpha : \widetilde{I}_M \to \Omega' BM.$
\end{enumerate}
Continuity of the maps are clear from its construction.
\begin{center}
\begin{tikzcd}
\adms \ar{r}{\omega_t}\ar[end anchor = north west]{rdddd}[left]{\alpha_t}
& G \ar{d}{s_t
\ar[hook]{r} & \multi{\map{V_t, S^1} \times M}
\\
& G' \ar[hook]{r}\ar{dd} & \multi{\map{V_t, S^1} \times M}\ar{d}\\
& & \multi{\map{V_t, S^1\times M} }\ar{d} \\
& \map{V_t, \pretensor{S^1, M} }\ar[hook]{r}\ar{d} & \map{V_t, \multi{S^1\times M} }\\
& \map{V_t, BM }.\\
\end{tikzcd}
\end{center}
\subsection{A map $\omega_t : \adms \to \multi{\map{V_t, S^1}\times M}$}
Let $\mathcal K$ be the subspace of $\interval $ defined by
$$\mathcal K = \{ J = \intervalconvention{p}{ u, v}{q} ~|~ v-u > 1 \mbox{~if~} p=q\}$$
For any $J = \intervalconvention{p}{ u, v}{q} \in\mathcal K,$ let $\Omega(J) : \RR \to S^1$ be
a map defined as follows:
\begin{enumerate}
\item if $v - u > 1$ then
$$\omega (J)(s) = \left\{
\begin{array}{ll}
p(s-u-1/2) & (\mbox{if~} u-1/2 < s \leq u+1/2)\\
0 & (\mbox{if~} u+1/2 < s \leq v-1/2)\\
q(s-v+1/2) & (\mbox{if~} v - 1/2 < s \leq v + 1/2)\\
1 & (\mbox{otherwise})
\end{array}
\right.$$
\item if $v - u \leq 1$ then (in this case $p+q = 0$)
$$\omega (J)(s) = \left\{
\begin{array}{ll}
p(s-u-1/2) & (\mbox{if~} u - 1/2 < s \leq v - 1/2)\\
p(v-u-1) & (\mbox{if~} v - 1/2 < s \leq u + 1/2)\\
q(s-v+1/2) & (\mbox{if~} u + 1/2 < s \leq v+1/2)\\
1 & (\mbox{otherwise})
\end{array}
\right.$$
\end{enumerate}
The correspondence $J\mapsto \omega(J)$ gives us map $\omega :\mathcal K \to \map{\RR^1, S^1}.$
We are now going to define a map $\omega_t : W_s \to \multi{\map{V_t, S^1}\times M}.$
\begin{enumerate}
\item If $e = ( J , m ) $ is an elementary configuration in $(t-1,t+1)$ of type (E1),
\begin{enumerate}
\item we put $\tilde{J} = J$
\item if $J = ] a, w [$ then we put $\tilde{J} = [a, w[,$ otherwise $\tilde{J} = J,$
\item if $J = ] w, b [$ then we put $\tilde{J} = ]w, b]$ otherwise $\tilde{J} = J,$
\item we put $\tilde{J} = J$
\end{enumerate}
We put $\tilde{e} = ( \tilde{J} , m ) \in \multi{\interval\times M}.$
\item If $f = (K, n)\dotplus (K',n) $ is an elementary configuration in $]t-1,t+1[$ of type (E2),
\begin{enumerate}
\item $\tilde{K} = K,$ and if $K' = ] w_2, b [$ then we replace it with $\tilde{K'} = ] w_2, b],$ otherwise $\tilde{K'} = K',$
\item $\tilde{K'} = K,$ and if $K = ] a, w_1 [$ then we replace it with $\tilde{K} = [a , w_1[,$ otherwise $\tilde{K} = K,$
\end{enumerate}
\end{enumerate}
We put $\tilde{f} = ( \tilde{K} , n )\dotplus (\tilde{K'},n) \in \multi{\interval\times M}.$
Resulting element $\tilde{e} \in \multi{\interval\times M}$ is called the replacement of $e.$
For $\xi = e_1 \dotplus \dots \dotplus e_r \dotplus f_1 \dotplus \dots \dotplus f_s,$
let $r(\xi ) = \tilde{e}_1 \dotplus \dots \dotplus \tilde{e}_r \dotplus \tilde{f}_1 \dotplus \dots \dotplus \tilde{f}_s.$
We can define
$\omega_t : \adms \to \multi{\map{V_t, S^1}\times M}$
by the following diagram:
\begin{center}
\begin{tikzcd}
\adms \ar{r}{r} \ar{rrd}[below]{\omega_t}&
\multi{\mathcal K\times M} \ar{r} &
\multi{\map{\RR^1, S^1}\times M} \ar{d}{\mbox{res}}\\
& & \multi{\map{V_t, S^1}}_,
\end{tikzcd}
\end{center}
where the second map in the upper horizontal line is $\multi{\omega\times id_M}.$
This map is essentially the one defined in \cite{okuyama-space-of-intervals} for the
case of the space of intervals with labels in a trivial partial abelian monoid.
Let $G$ denote the image of $\omega_t.$
\subsection{A map $s_t : G \to \multi{\map{V_t, S^1} \times M}$}
We can define a `sum' $s_t : G \to \multi{\map{V_t, S^1} \times M}.$
For any element $\xi\in \adms,$ $\xi_t = \xi|_{U_t}$ is of the form
$$\xi_t = e_1 \dotplus \dots \dotplus e_r \dotplus f_1 \dotplus \dots \dotplus f_s,$$
where $e_i$ is an elementary configuration of type (E1) for each
$1\leq i\leq r$ and $f_j$ is that of type (E2) for each $1\leq j \leq s.$
Let $e_i = ( J_i, m_i ) $ and $f_j = ( K_j, n_j )\dotplus ( K'_j, n_j )$ so that
each $J_i$ is one of (F0)$\sim$ (F3)
and each $K_i$ is (F1) and each $K'_i$ is (F2) shown below.
\begin{enumerate}
\item[(F0)] $J = ] t-1, t+1 [,$
\item[(F1)] $J = ] t-1, w [$ or $J = ] t-1 , w ]$, where $t-1 < w < t+1,$
\item[(F2)] $J = ] w, t+1 [$ or $J = [ w, t+1 [$, where $t-1 < e < t+1,$
\item[(F3)] $J = ] w_1 , w_2 ]$ or $ J = [ w_1 , w_2 [,$ where $t-1 < w_1 \leq w_2 < t+1.$
\end{enumerate}
Accordingly, $\omega_t(\xi_t) \in G$ is represented by
$$
(\varphi_1, m_1) \dotplus\dots\dotplus(\varphi_r , m_r)
\dotplus
(\psi_1, n_1) \dotplus (\psi', n_1 )\dotplus\dots \dotplus (\psi_s, n_s) \dotplus (\psi'_s, n_s),
$$
where $\varphi_i = \omega(J_i)|_{V_t}$ and $\psi_j = \omega(K_j)|_{V_t}, \psi'_j = \omega(K'_j)|_{V_t}.$
\begin{lem}\label{lem:unique-representation}
Suppose that every non-zero element of $M$ is self-insummable.
Then any element of $G$ is of the form
$$
(\varphi_1, m_1) \dotplus\dots\dotplus(\varphi_r , m_r)
\dotplus
(\psi_1, n_1)\dotplus (\psi', n_1 )\dotplus\dots \dotplus (\psi_s, n_s) \dotplus (\psi'_s, n_s) ,
$$
where $\varphi_i = \omega(J_i)$ is an image of an element of type (E1)
and $\psi_j = \omega(K_j), \psi'_j = \omega(K'_j)$ constitute a pair of an image
of an element of type (E2).
This representation is unique up to order.
\end{lem}
\begin{proof}
Let $m_1\dotplus \dots \dotplus m_r \dotplus n_1\dotplus n_1\dotplus \dots \dotplus n_s\dotplus n_s$ be an element of
$\multi{M}$ such that $(m_1, \dots ,m_r , n_1 ,\dots ,n_s) \in M_{r+s}.$ Since, by assumption, $(n_i , n_i ) \not\in M_2,$ it
is easy to see that this representation is unique.
\end{proof}
Let $ (\psi, n )\dotplus ( \psi', n ) \in \multi{\map{V_t, S^1}\times M}$ be the image of an
elementary configuration $(K, n)\dotplus (K', n) $ of type (E2).
Let
\[
K = \left] t-1 , v \right|_{q} \mbox{~and~} K' = \,_{\bar{q}}| u', t+1 [.
\]
Then we define
$\psi \oplus \psi' : V_t \to S^1$
by
\[
(\psi \oplus \psi') (s) =
\left\{\begin{array}{ll}
\psi(s) & (\mbox{if~} s \leq u'-1/2)\\
q (u'-v) & (\mbox{if~} u'- 1/2 \leq s < v+1/2)\\
\psi'(s) & (\mbox{if~} s > v_1+1/2).\\
\end{array}\right.
\]
Now a map $s_t : G \to \multi{ \map{ V_t, S^1} \times M }$ is defined by
\begin{eqnarray*}
s_t ( \omega (\xi_t ) ) & = & \sum_{i} (\omega(J_i), m_i)
\dotplus \sum_{j} (\omega(K_j)\oplus \omega(K'_j), n_j) .
\end{eqnarray*}
It is clear that $s_t : G \to \multi{ \map{ V_t, S^1} \times M}$
is continuous.
\subsection{Final steps}
It is readily seen that
$G'$ is mapped into $\map{V_t, \pretensor{S^1, M}}$ under the sequence of natural maps
\begin{eqnarray*}
\multi{\map{V_t, S^1} \times M} & \to & \multi{\map{V_t, S^1\times M} } \\
& \to & \map{V_t, \multi{S^1 \times M} } .
\end{eqnarray*}
Thus $\mu\circ\omega_t $ followed by this sequence gives a map
$ \adms \to \map{ V_t, \pretensor{S^1, M} }.$
Next, we take the quotient $\pretensor{S^1, M} \to S^1\otimes M = BM,$ then maps altogether
gives us a map $\alpha_t : \adms \to \map{V_t, BM}.$
Let $[\xi] \in \adms,$ then
$\alpha_t ( [\xi ]) \in \map{ V_t, BM}$
for two distinct $t\in [0, s]$ coincide on their intersection of the domain of definition and
we can take the union of them to get an element in
$\Omega'_s BM.$ Thus we get a map
$ \adms \to \loopsBM.$
However, this map factors through the natural quotient map
$\adms \to \ims{s}$ as the following diagram,
\begin{center}
\begin{tikzcd}
\adms \ar{r} \ar{d}& \Omega'_s BM\\
\ims{s}\ar{ru}[below right]{\alpha(s)}
\end{tikzcd}
\end{center}
so we have a map $\alpha(s) : \ims{s} \to \loopsBM.$
Finally, we consider all the $s > 0$ and
take a union to get a continuous map $\alpha : \widetilde{I}_M \to \Omega' BM.$
Let $P' X$ denote the space of Moore paths on $X$;
\[
P'X = \{ f : [ 0, s ] \to X ~|~ \mbox{continuous} , f(s) = *\}.
\]
Then a map $\beta : \widetilde{E}_M \to P'BM$ is defined by using $\alpha^\ep_{(-s,s)}.$
A map $p : \widetilde{E}_M \to BM$ is defined by the composite
$\widetilde{E}_M \to P'BM \to BM,$
where $P'BM \to BM$ is the evaluation at $0.$
\begin{center}
\begin{tikzcd}
\widetilde{E}_M \ar{r}{\beta} \ar{rd}[below]{p} & P' BM \ar{d}{ev_0}\\
& BM.
\end{tikzcd}
\end{center}
\section{Proof of the main theorem}
Proof of the main theorem is based on the following proposition.
\begin{prop}\label{prop:quasifibration}
Let $M$ be a (discrete) self-insummable partial abelian monoid.
The map $p : \widetilde{E}_M \to BM$ is a quasifibration with fiber $\widetilde{I}_M.$
\end{prop}
Assuming the above proposition, we can give a
\begin{proof}[Proof of Theorem 1]
In the following diagram,
\begin{center}
\begin{tikzcd}
\widetilde{I}_M \ar{r}\ar{d} & \widetilde{E}_M\ar{r}\ar{d} & BM \ar{d}\\
\Omega'BM \ar{r}& P' BM \ar{r}& BM\\
\end{tikzcd}
\end{center}
lower horizontal line is the path-loop fibration and
vertical map on the right is identity map and vertical map in the middle is
a weak homotopy equivalence, since it is
a map between weakly-contractible spaces,
hence so is the vertical map on the left.
\end{proof}
This section is devoted to giving a proof of
Proposition \ref{prop:quasifibration}.
\subsection{Dold-Thom criterion (1)}
For any element $\xi$ of $E_M,$ we can find a representative of the form
$$S_- \dotplus S_0 \dotplus S_+\in
\multi{\interval\times M},$$
which satisfies the following conditions:
\begin{enumerate}
\item $S_+|_{( 0,\infty )} = S_+$
\item $(S_-)^\mu = S_+$
\item $S_0$ is a multiset of the following form
$$ (J_1, m_1)\dotplus \dots\dotplus (J_r, m_r)$$
where $J_i = \intervalconvention{\bar{p_i}}{-u_i, u_i}{p_i}$ for some $u_i >0$ and $p_i \in P$ for
each $ i= 1, \dots , r .$
\end{enumerate}
We can also take such a representative in a reduced form.
Then from $S_0$ we construct a new multiset
\[
S'_0 = ( K_1, m_1)\dotplus \dots\dotplus (K_r, m_r)
\]
with
$K_i = [ 0 , u_i |_{p_i}.$ We call the element
$S'_0 \dotplus S_+ \in \multi{\interval \times M}$
{\bf a representative of the positive part} of $\xi.$
In the following, we use the standard filtration $F_l BM~(l\geq 0)$ of
$BM = S^1 \otimes M,$ appeared in \S 2.
\begin{lem}\label{lem:key-lemma1}
Let $l \geq 1.$
If we denote $V = F_l BM \setminus F_{l-1} BM,$ then there exists a homotopy
equivalence $p^{-1} V \simeq V\times \widetilde{I}_M.$
Accordingly, $V$ is a distinguished set with respect to $p.$
\end{lem}
\begin{proof}
We define a map $\varphi : p^{-1}V\to \widetilde{I}_M$ by
taking the positive part and attach a ``cap'' on the left.
The cap consists of
left-open intervals of length 1, which is contained in
$( 0, 2 )$. To state the construction precisely,
let $ (p^{-1} V)_s$ denote the subspace
$p^{-1} V \cap \ems{s}$ of $p^{-1}V.$
We define a map
$\varphi_s : (p^{-1}V)_s \to \ims{s+2}$ as follows.
Let $(\xi, s ) \in (p^{-1}V)_s$ and let
$S'_0 \dotplus S_+$ be a representative of the positive part of $\xi.$
Among all the intervals occurring in this representation, we collect
ones which have $u_L(J) \leq 1/2.$ There should be exactly
$j$ of such intervals in number, so we may assume that they constitute
a multiset
$$
(J_1, m_1) \dotplus \dots\dotplus (J_j, m_j).
$$
Then the ``cap'' of $S'_0 \dotplus S_+$ is the multiset
$$
C = (K_1, m_1)\dotplus \dots \dotplus (K_j, m_j)
$$
where
$K_i = \left] 1 - u_L(J_i), 2 - u_L(J_i) \right|_{p}, p = \overline{p_L(J_i)}.$
Now, $\varphi_s ( \xi )$ be the element of $\ims{s+2}$ represented by
$$
C \dotplus \translation{2}(S'_0 \dotplus S_+ ),
$$
where $\translation{2}$ denotes the translation by $2.$
Now
$\varphi : p^{-1} V \to V\times \widetilde{I}_M $ is defined by
$(\xi, s) \mapsto (\varphi_s(\xi) , s+2 ).$
Thus we have defined a map
\[
( p, \varphi ) : p^{-1} V \to V\times \widetilde{I}_M.
\]
We define a map $\psi : V\times \widetilde{I}_M \to p^{-1}V$
by inserting ``the standard lift of the $V$-component'' to the $\widetilde{I}_M$-component.
The standard lift of $v\in V$ is an element of
$\ems{2}$ which maps to $v$ under $p,$ whose
positive part consists of right closed intervals.
To state the construction precisely,
we define a map $\psi_s : V\times \ims{s} \to (p^{-1}V)_{s+2}.$
For any element $v = t_1\otimes m_1 + \dots + t_j\otimes m_j \in V, $ we set
\[ L_i = \,_{p_i}| |t_i|/2, |t_i|/2 +1 ] \in \interval \]
where $p_i = -t_i/|t_i|$ if $t_i \neq 0$ and
$p_i$ is any of $\pm 1$ if $t_i = 0.$ Then ``the standard lift'' of $v$ is the multiset
$$
L = ( L_1 , m_1 )\dotplus \dots\dotplus ( L_j , m_j ) .
$$
Then we define
$\psi_s(v, \xi)$ to be the element of $(p^{-1}V)_{s+2}$ represented by
the multiset
$ L \dotplus \translation{2}(X) \dotplus \mu(L \dotplus \translation{2}(X)).$
Now
$\psi : V\times \widetilde{I}_M \to p^{-1}V$ is defined by
$( v, (\xi, s)) \mapsto (\psi_s(v, \xi) , s+2).$
Now, we define a homotopy $H' : p^{-1}V \to p^{-1}V $ between $\psi \circ ( p, \varphi ) $ and $id_{p^{-1}V}.$
We can do this by pushing everything to $2\in \RR$ and erases the part of the configuration on the left of $2.$
For this purpose, let $h'_t : \RR\to \RR ~(0\leq t\leq 1)$ be the homotopy defined by
$$
h'_t( u ) = \left\{
\begin{array}{cc}
u- 2 t & u\geq 2 t\\
0 & |u| < 2t\\
u + 2 t & u\leq -2 t
\end{array}
\right.
$$
Then we define a homotopy $H'_{t} : (p^{-1}V)_s\to (p^{-1}V)_s$ as follows. If
$\xi = (J_1, m_1)\dotplus \dots\dotplus (J_r, m_r) \in \multi{\interval \times M}$ is a
representative of
$\xi \in \Image [ \psi \circ ( p, \varphi ) ]$
then let $H'_t(\xi)$ be the element of $(p^{-1}V)_s$ represented by
$$
(\translation{2}h'_t \translation{-2}(J_1), m_1)\dotplus \dots\dotplus (\translation{2}h'_t \translation{-2}(J_r) , m_r).
$$
For each $i,$ $h_t(J_i)$ is at least a degenerate interval by our construction.
It is easy to see that $H'_t$ defines a homotopy $H' : p^{-1}V\to p^{-1}V$ between
$\psi \circ ( p, \varphi ) $ and $id_{p^{-1}V}.$
Next, using the same homotopy $h'_t : \RR\to \RR~(0\leq t\leq 1),$ we define a homotopy
$H''_t : V\times \widetilde{I}_M\to V\times \widetilde{I}_M$ between
$ (p, \varphi)\circ \psi $ and $id_{V\times \widetilde{I}_M}$
as follows.
If
$\xi = (J_1, m_1)\dotplus \dots\dotplus (J_r, m_r) \in \multi{\interval \times M}$ is a
representative of $\eta \in \Image [(p, \varphi) \circ \psi],$ then
$H''_t(\eta)$ be the element of $(p^{-1}V)_s$ represented by
$$
(\translation{2}h'_t \translation{-2}(J_1), m_1)\dotplus \dots\dotplus (\translation{2}h'_t \translation{-2}(J_r) , m_r).
$$
For each $i,$ $h_t(J_i)$ is at least a degenerate interval by our construction.
It is easy to see that $H''_t$ defines a homotopy
$H''_t : V\times \widetilde{I}_M\to V\times \widetilde{I}_M$ between
$ (p, \varphi)\circ \psi $ and $id_{V\times \widetilde{I}_M}.$
\end{proof}
\subsection{Dold-Thom criterion (2)}
Before proceeding, we investigate a topological property of $p^{-1}z$ for $z\in BM.$
The following notation is useful.
Let $\elemone{ v, p, m }$ denote the elementary configuration of type E1 given by
$ (J, m)$ with
\[
J = \intervalconvention{\bar{p}}{ -v, v }{p}
\]
and let
$\elemtwo{ v, p, m }$ denote the elementary configuration of type E2 given by
$ (K, n) \dotplus (K', n) $ with
\[
K' = \left] -1 , -v \right|_{\bar{p}}, K = \,_p| v, 1 [
\]
put
\[ e(t,m) = \elemone{ (1-t)/2,t/ |t|, m } \]
and
\[ f(t,n) = \elemtwo{ t/2 , -t/|t|, n }. \]
Furthermore, for any $m\in M,$ let $Z(m_0)$ denote a finite sum of
configurations of the form $E( u_k, p_k, c_k)|_{ ( -1,1 )}$ with
$ 1/2\leq |u| \leq 1$ and $\sum_k c_k = m_0.$
For any $z \in BM,$ we have a unique representation
$$ z = 0\otimes m_0 + t_1 \otimes m_1 + \dots + t_s\otimes m_s $$
such that $ -1 < t_1< t_2 < \dots < t_s < 1, t_i \neq 0$ and $ m_i \neq 0 $ for $1\leq i \leq s. $
Let $m \in M,$ and let $P(m)$ denote the set of partition of $m$ into a sum of two
elements in $M,$ that is,
$ P(m) = \{ (a, b) \in M_2 ~|~ a + b = m\}. $
Let $\alpha = ( ( a_1 , b_1 ) ,\dots, (a_s, b_s) ) \in P(m_1)\times \dots\times P(m_s),$
and $F_\alpha$ denote the subspace of $\widetilde{E}_M$ which consists
of $\xi \in \widetilde{E}_M$ such that
$\xi|_{(-1, 1)}$ has a representative of the form
\begin{equation}
Z(m_0) \dotplus \sum_{i=1}^{s} e(t_i , a_i) \dotplus \sum_{i=1}^{s} f(t_i, b_i) \dotplus \sum_{j=1}^{q} f(u_j, n_j)
\end{equation}
$1 \leq |u_j| < 2$ for each $j.$
where $0< |t_i| < 1$ for each $i$ and $1 \leq |u_j| < 2$ for each $j.$
Then $F_\alpha $ is contained in
$p^{-1}z.$ Similarly, let $H_\alpha$ denote the subspace of $F_\alpha$ which consists
of $\xi$ for which $\xi|_{(-3,3)}$ has a representative of the form (1), but now with an
extra condition that $q=0$
and $f(t,b)$ are altered by $f'(t,b)$ and $Z_0(m_0)$ is of a special form as follows:
\[
f'(t,b) = ( K' , b) \dotplus (K, b )
\]
with
\[
K' = \left] -3 , -t/2 \right|_{\bar{p}} , K = \,_{p}\!\left| t/2, 3\right[ ,
\]
where $p = t/ |t|,$
and $Z_0(m_0) = (~\left] -3 , 3 \right[ , m_0) .$
Notice that $F_{\alpha}$ is not path-connected in general, but $F_\alpha$
and $F_\beta$ can not be connected by a continuous
path in $p^{-1} z$ whenever $\alpha\neq \beta.$
The same statement is true for $H_\alpha.$
In the following lemma, $F= F_\alpha$ and $H= H_\alpha.$
\begin{lem}\label{lem:deformation_retraction}
$H$ is a deformation retract of $F.$
\end{lem}
\begin{proof}
To construct a retraction $r : F \to H, $ we consider maps
$\sigma : [ 0, \infty ) \to [ 0, \infty ) $ and $\tau : [ 1/2, \infty ) \to [ 1, \infty ) $
defined by
\begin{eqnarray*}
\sigma(v) & = & \left\{\begin{array}{ll}
v & (0\leq v \leq 1/2)\\
3v-1 & (v \geq 1/2)\\
\end{array}\right.\\
\tau( v ) & = & \left\{\begin{array}{ll}
v+1/2 & ( 1/2\leq v \leq 3/4 )\\
3v-1 & ( v \geq 3/4 )
\end{array}\right.
\end{eqnarray*}
we extend $\sigma$ and $\tau$ by $\sigma( v) = -\sigma(-v)$ and $\tau(v) = - \tau( -v ) $ for negative $v.$
For any element of $\xi \in F,$ $\xi|_{ ( -1, 1 ) } $ has a representative
of the form
$$ \sum_{i=1}^{s} e(t_i , a_i) \dotplus \sum_{i=1}^{s} f(t_i, b_i)\dotplus \sum_{i=1}^{q} f(u_j, c_j),$$
In the following argument, we use the fact that since $b_j'$ is self-insummable,
it is distinct from any $n_j'.$
If $\varphi : U \to V$ is a function with gradient greater or equal to 1 defined on an open interval
$U\subset \RR$ with values in
another interval $V\subset \RR,$ and if $h = ( J_i, m_i ) \in \multi{\interval\times M}$
is a multiset in which each $J_i = \intervalconvention{p_i}{u_i, v_i}{q_i}$ corresponds to an interval contained in $U,$
then by $\varphi_* h,$ we mean a multiset $ ( J_i', m_i ),$ where
$J_i' = \intervalconvention{p_i}{\varphi(u_i), \varphi(v_i) }{q_i}.$
We put
$$
\xi' = \sum_{i=1}^s \sigma_*(e (t_i , a_i) ) \dotplus \sum_{i=1}^s \sigma_*(f(t_i, b_i))
\dotplus \sum_{i=1}^{q} \tau_*( f(u_j, c_j) ).
$$
Then we put
$$r(\xi ) = \xi' + \tau_*(\xi |_{ (-\infty, -1] \cup [1 , \infty ) }).$$
This defines a deformation retraction $ r : F \to H.$
\end{proof}
\begin{lem}\label{lem:key-lemma2} For any $l \geq 1, $
there exists an open set $\openset\subset F_{\filternumber} BM$ which contains
$F_{\filternumber-1} BM$ and
homotopies $h_t: \openset\to \openset$ and $H_t : p^{-1}\openset\to p^{-1}\openset$
such that
\begin{enumerate}
\item $h_0 = id_{\openset}$, $h_t(F_{\filternumber-1} BM) \subset F_{\filternumber-1} BM$ and
$h_1(\openset) \subset F_{\filternumber-1} BM$,
\item $H_0 = id_{p^{-1}\openset}$ and $p\circ H_t = h_t\circ p$ for all $t$, and
\item $H_1 : p^{-1}z \to p^{-1} h_1(z)$ is a weak homotopy equivalence for all $z\in \openset.$
\end{enumerate}
\end{lem}
\begin{proof}
Let $U'$ be the subspace of $(S^1 \times M)^\filternumber$ which consists of
elements $$((t_1, m_1),\dots, (t_\filternumber, m_\filternumber))$$ such that
there exists at least one $i$ with $|t_i| > 1/2.$
Then let $U$ be the subspace of $F_j BM$ which consists of elements which is represented by
elements in $U'.$
We define a homotopy $h'_t : [-1, 1] \to [-1, 1]~(0 \leq t \leq 1 )$ by
\begin{equation} \label{eqn:ht'}
h'_t( u ) = \left\{
\begin{array}{cc}
-1 & -1\leq u\leq \frac{t}{2} -1\\
2u/ (2-t) & |u| < 1-\frac{t}{2}\\
1 & 1-\frac{t}{2} \leq u\leq 1.
\end{array}
\right.
\end{equation}
Then we define a homotopy $h_t : \openset\to \openset$ by
$$
( (t_1, m_1) , \dots, (t_\filternumber, m_\filternumber ))\mapst
( ( h'_t( t_1 ), m_1 ) , \dots, (h'_t(t_\filternumber) , m_\filternumber ) ).
$$
Next, we construct a homotopy $H_t : (p^{-1}O)_s \to (p^{-1}O)_{s+5/2}.$
Consider a homotopy $\lambda_t : [ 0, \infty ) \to [ 0, \infty ) $ defined by
\[
\lambda_t(v) = \left\{\begin{array}{ll}
0 & (0\leq v \leq t/4)\\
(4v - t)/ (4-2t) & (t/4\leq v \leq 1/2)\\
(3t+1) v - 3t/2 & ( 1/2 \leq v \leq 1)\\
v + 3t/2 & (v \geq 1)
\end{array}\right.
\]
We also denote by $\lambda_t$ the self-homotopy of $\interval$
given by
\begin{equation}\label{eqn:lambda_t}
\intervalconvention{p}{a,b}{q}\mapsto \intervalconvention{p}{\lambda_t(a),\lambda_t(b)}{q}.
\end{equation}
Then let $\widetilde{\lambda}_t = \multi{ \lambda_t \times id_M }$
be the self-homotopy of $ \multi{ \interval \times M} .$
Similarly, using a homotopy $\nu_t : [ 0, \infty ) \to [ 0, \infty ) $ defined by
\[
\nu_t( v ) = \left\{\begin{array}{ll}
(t+1)v & ( 0\leq v \leq 3/4 )\\
(3t+1) v - 3t/2 & ( 3/4 \leq v \leq 1)\\
v + 3t/2 & (v \geq 1)
\end{array}\right._,
\]
we define a self-homotopy of $\interval$ by
\begin{equation}\label{eqn:nu_t}
\intervalconvention{p}{a,b}{q}\mapsto \intervalconvention{p}{\nu_t(a),\nu_t(b)}{q}.
\end{equation}
Then we can define a self-homotopy $\widetilde{\nu}_t$ of $\multi{\interval\times M}$ by
$\widetilde{\nu}_t = \multi{ \nu_t \times id_M }.$
Let $S_{-} \dotplus S_0 \dotplus S_{+} \in \multi{ \interval\times M}$ be a representative
of a given element $\xi \in (p^{-1}O)_s$ as in \S 4.4. Then let $H'_t(\xi)$ be
an element of $(p^{-1}O)_{s+5/2}$ represented by
$\widetilde{\lambda}_t( S_0 ) \dotplus \widetilde{\nu}_t( S_+ ) \dotplus \mu\widetilde{\nu}_t(S_+).$
Now $H_t : (p^{-1}O)_s \to (p^{-1}O)_{s+5/2}$ is defined by $H_t( \xi, s ) = ( H'_t(\xi), s+5/2).$
It is clear that $H_t$ covers $h_t.$\smallskip\\
To complete the proof, we show that $H_1 : p^{-1}z \to p^{-1} h_1(z)$ is in fact a
homotopy equivalence. By renumbering $t_1, \dots , t_s,$ we may assume that
there exists a number $s_0$ such that
$| t_i | > 1/2$ for $i\leq s_0$ and $|t_i| \leq 1/2$ otherwise.
Then $h_1 z = 0\otimes m_0 + h_1(t_{s_0+1}) \otimes m_{s_0+1} + \dots + h_1(t_s) \otimes m_s.$
Let
\[\alpha' = ( (a_{s_0+1}, b_{s_0+1}), \dots , ( a_s , b_s ) ) \in P(m_{s_0+1})\times\dots\times P(m_s).\]
By Lemma \ref{lem:deformation_retraction}, we have a deformation retraction
$r: F'_\alpha \to H'_\alpha.$
We construct a map $g: H_{\alpha'} \to F_{\alpha}$ such that $g\circ r$ is
a homotopy inverse of $H_1|_{F_\alpha}.$ Given an element of $H_{\alpha'},$
we want to recover the information of $F_\alpha$ by using the data given by $\alpha.$
For this, we insert ``the standard lift of $z$ of type $\alpha$''
defined by the multiset
\begin{equation}
\zeta = Z'(m_0) \dotplus \sum_{i=1}^{s} e(t_i , a_i) \dotplus \sum_{i=1}^{s} f'(t_i, b_i)
\end{equation}
at the origin, where
$Z'(m_0) = ( [ -1, 1 [ , m_0 )$
and $f'(t,b)$ denotes a configuration which is constructed by
changing the parity of only one endpoint among four occurring in $f(t,b)$ so that
the resulting configuration consists of two half-open intervals.
For $\xi \in H_{\alpha'},$ let $S'_0 \dotplus S_+$ be a representative of the
positive part of $\xi.$
Then we may assume that
$$S_0' = (K_0, m_0)\dotplus (K_1,m_1) \dotplus \dots \dotplus(K_r, m_r)$$
with $u_L(K_i) = 0, p_L(K_i) = 1.$
For $1\leq i\leq r,$ we alter $p_L(K_i)$ by the value $-p_R(K_i)$ and
denote by $S_0''$ the resulting multiset.
Now if we define
$g : H_{\alpha'}\to F_\alpha$ by
$$
g(\xi ) = \zeta \dotplus \mu ( \translation2 ( S_0'' \dotplus S_+ )).
$$
then it is easy to see that
$ g\circ r$ is a homotopy inverse to $H_1 |_{F_\alpha}.$
\end{proof}
\begin{proof}[Proof of Propostion \ref{prop:quasifibration}]
$F_0 BM = \{*\}$ is obviously a distinguished set with respect to $p.$
By induction, assume that $F_{l-1} BM$ is a distinguished set. Then
by Lemma \ref{lem:key-lemma2} and Hilfssatz 2.10 of
\cite{dold-thom-quasifaserungen-und-unendliche}, we have an open set
$O\subset F_{l}$ which contains $F_{l-1} BM,$ which is a distinguished set.
Then by Lemma \ref{lem:key-lemma1} and
Satz 2.2 of \cite{dold-thom-quasifaserungen-und-unendliche},
we have that $F_{l} BM$ is a distinguished set. Then by Satz 2.15 of
\cite{dold-thom-quasifaserungen-und-unendliche},
$p : \widetilde{E}_M \to BM$ is a quasifibration.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}\label{1}
Mathematical formalism of quantum theory introduced non-commutativity of
observables. Hei\-sen\-berg's uncertainty relation elucidates an operational meaning of the non-commutativity
as a limitation to the simultaneous measurability of a pair of observables.
In 1927, using the famous $\gamma$-ray microscope thought experiment,
Heisenberg \cite{Hei27} claimed that canonically conjugate observables $Q, P$
can be measured simultaneously only with the relation
\beql{Hei27}
\ep(Q)\ep(P)\ge\frac{\hbar}{2}
\eeq
for the ``mean errors'' $\ep(Q), \ep(P)$.
However, his formal derivation of this relation from the well-established relation
\begin{equation}
\sigma(Q)\sigma(P)\ge\frac{\hbar}{2} \label{Ken27}
\end{equation}
for the standard deviations $\sigma(Q), \sigma(P)$ due to Heisenberg \cite{Hei27} and Kennard \cite{Ken27}
needs an additional assumption such as a quantitative version of the repeatability hypothesis \cite{15A2}.
Although the repeatability hypothesis was commonly accepted at that time
(cf.~Schr\"{o}dinger \cite[Section 8]{Sch35}, Dirac \cite[p.~36]{Dir58}, and von Neumann \cite[p.~335]{vN32E}),
this hypothesis has been completely abandoned in the modern quantum mechanics \cite{DL70},
in which quantum measurements are generally described by completely positive instruments \cite{84QC}.
In such a general description of quantum measurements, Heisenberg's relation \eq{Hei27}
loses its universal validity \cite{88MS,02KB5E}.
An alternative relation universally valid for arbitrary measurements,
arbitrary pairs of observables, and arbitrary states was derived only recently by
one of the authors \cite{03HUR,03UVR,03UPQ,04URJ,04URN,05UUP},
and has recently received considerable attention.
The validity of the above new relation, as well as a stronger version of this relation
\cite{Bra13,Bra14,14EDR}, were experimentally tested with neutrons
\cite{12EDU,13VHE,16A3}
and with photons
\cite{RDMHSS12,13EVR,WHPWP13,RBBFBW14,14ETE}.
Other approaches generalizing Heisenberg's original relation
can be found, for example, in
\cite{App98c,Hal04,Wer04,BLW13,BLW14PRA,BLW14JMP
BLW14RMP,LYFO14},
apart from the information theoretical approach \cite{14NDQ,15A1}.
All the above studies of universally valid uncertainty relations have been restricted to
quantum systems with finite degrees of freedom.
In the present paper, we show universally valid uncertainty relations for quantum systems
described by general $\sigma$-finite von Neumann algebras including systems
with infinite degrees of freedom, based on quantum measurement theory
for systems with infinite degrees of freedom recently established
by the present authors \cite{16A1}.
The authors aim to clarify the mathematical essence of derivations of
universally valid uncertainty relations and to give a simpler proof than ever before.
The universally valid uncertainty relations derived by Branciard \cite{12EDU,13VHE}
for pure states and by one of the authors \cite{14EDR} for mixed states
considerably strengthened Ozawa's original relation \cite{03UVR,03UPQ}
and is considered as the strongest relation ever.
Their generalizations to general $\sigma$-finite von Neumann algebras of course match
our aim.
To achieve this purpose the theory of the standard form of von Neumann algebras plays
a crucial role.
The bound, denoted by $D_{AB}$ below, in uncertainty relations
is written in the language of the theory.
We would like to emphasize that
the theory of the standard form of von Neumann algebras can be
a more powerful tool than ever for quantitative analysis in quantum information theory.
This study may be counted as an instance supporting the opinion that ``the fields of operator
algebras and quantum information can benefit each other in multiple ways" \cite[ll.~6--7]{KGR16}.
We adopt the (von Neumann) algebraic formulation of quantum theory herein.
Observables of a physical system are described
by self-adjoint operators affiliated with a von Neumann algebra,
and physical situations and experimental settings of the system
are described by normal states on the von Neumann algebra.
Let $\mathcal{M}$ be a von Neumann algebra on a Hilbert space $\mathcal{H}$
and $(S,\mathcal{F})$ a measurable space.
We assume that von Neumann algebras are $\sigma$-finite in the present paper.
Let $\mathcal{M}_{s.a.}$ denote the set of self-adjoint elements of $\mathcal{M}$.
Let $\mathcal{M}_\ast$ denote the predual of $\mathcal{M}$,
let $\mathcal{M}_{\ast,+}$ denote the set of positive elements of $\mathcal{M}_\ast$,
and $\mathcal{S}_n(\mathcal{M})$ the set of normal states on $\mathcal{M}$.
For every $\rho\in\mathcal{M}_\ast$ and $M\in\mathcal{M}$, let
$\langle \rho,M\rangle$ denote the pairing of $\mathcal{M}_\ast$ and $\mathcal{M}$, i.e.,
$\langle \rho,M\rangle=\rho(M)$.
In quantum mechanics, the observables are described by the von Neumann algebra
$\mathcal{M}=\textrm{\boldmath $B$}(\mathcal{H})$,
the set of bounded linear operators on a Hilbert space $\mathcal{H}$,
and the set of states, described by the density operators on $\mathcal{H}$, corresponds
to the set $\mathcal{S}_n(\mathcal{M})$ of normal states on $\textrm{\boldmath $B$}(\mathcal{H})$.
In the modern quantum mechanics having abandoned
the repeatability hypothesis,
the Heisenberg type uncertainty relation
\begin{equation}
\varepsilon(A)\ep(B)\geq C_{AB}
\end{equation}
for an arbitrary pair of observables $A,B$ is known to hold only for a limited class of measurements,
for instance, for jointly unbiased joint measurements \cite{AK65,AG88,Ish91,91QU},
where $\varepsilon$ denotes the measurement error (to be defined in Section 3)
and $C_{AB}=C_{A,B,\rho}$ is defined by
\begin{equation}
C_{A,B,\rho}=\frac{1}{2}|\langle\rho,-i[A,B]\rangle|.
\end{equation}
In 2003, one of the authors \cite{03UVR,03UPQ} derived a universally valid uncertainty relation
\begin{equation}\label{Ozawa}
\varepsilon(A)\ep(B)+\varepsilon(A)\sigma(B)+\sigma(A)\ep(B)\geq C_{AB},
\end{equation}
where $\sigma(A)=\sigma(A;\rho)$ is
the standard deviation of an observable $A$ in a normal state $\rho$ defined by
\begin{equation}
\sigma(A;\rho)=(\rho(A^2)-\rho(A)^2)^{\frac{1}{2}}=\langle\rho,(A-\rho(A))^2\rangle^{\frac{1}{2}}.
\end{equation}
In 2013, Branciard \cite{Bra13,Bra14} strengthened the above relation as
\begin{equation}\label{eq:Branc}
\varepsilon(A)^2\sigma(B)^2+\sigma(A)^2\ep(B)^2+2\varepsilon(A)\ep(B)
\sqrt{\sigma(A)^2\sigma(B)^2-C_{AB}^2}
\geq C_{AB}^2.
\end{equation}
Subsequently, this relation \eq{Branc} was further improved by one of the authors
\cite{14EDR} replacing the lower bound
$C_{AB}$ by a more stringent one $D_{AB}$ as
\begin{equation}\label{Ozawa2}
\varepsilon(A)^2\sigma(B)^2+\sigma(A)^2\ep(B)^2+2\varepsilon(A)\ep(B)
\sqrt{\sigma(A)^2\sigma(B)^2-D_{AB}^2}
\geq D_{AB}^2,
\end{equation}
where $D_{AB}=D_{A,B,\rho}$ is defined by
\begin{equation}
D_{A,B,\rho}= \frac{1}{2}\mathrm{Tr}[|\sqrt{\tilde{\rho}}(-i[A,B])\sqrt{\tilde{\rho}}|],
\end{equation}
and $\tilde{\rho}$ is a (unique) density operator on $\mathcal{H}$ such that $\rho(M)=\mathrm{Tr}[M \tilde{\rho}]$
for all $M\in\textrm{\boldmath $B$}(\mathcal{H})$.
The fact that the bound $D_{AB}$
depends on the choice of the observable algebra relevant to the measuring
interaction is important.
This is easily seen by the example below.
\begin{example}[]
Let $\mathcal{N}=\mathrm{M}_2(\mathbb{C})\otimes\mathrm{M}_2(\mathbb{C})\cong\mathrm{M}_4(\mathbb{C})$
be a von Neumann algebra as the observable algebra,
$\sigma_x$, $\sigma_y$ and $\sigma_z$ the Pauli matrices
and $\omega_\psi=\langle \psi| (\cdot)\psi \rangle$ a normal state on
$\mathrm{M}_2(\mathbb{C})\otimes\mathrm{M}_2(\mathbb{C})\cong\mathrm{M}_4(\mathbb{C})$,
where $\psi$ is defined by
\begin{equation}
\psi=\sqrt{\dfrac{1}{2}}(e_{z\uparrow}\otimes e_{z\downarrow}- e_{z\downarrow}\otimes e_{z\uparrow}),
\end{equation}
and $e_{z\uparrow}$ and $e_{z\downarrow}$ are eigenvectors of $\sigma_z$
corresponding to eigenvalues $+1$ and $-1$, respectively,
i.e., $\sigma_z e_{z\uparrow}= e_{z\uparrow}$, $\sigma_z e_{z\downarrow}=- e_{z\downarrow}$.
Here, we set
$\mathcal{M}=\mathrm{M}_2(\mathbb{C})\otimes\mathbb{C}1\cong\mathrm{M}_2(\mathbb{C})$,
which is of course a von Neumann subalgebra of
$\mathcal{N}=\mathrm{M}_2(\mathbb{C})\otimes\mathrm{M}_2(\mathbb{C})\cong\mathrm{M}_4(\mathbb{C})$.
Then the restriction $\omega_\psi|_{\mathrm{M}_2(\mathbb{C})}$ of $\omega_\psi$
to $\mathrm{M}_2(\mathbb{C})$ satisfies
\begin{equation}
\omega_\psi|_{\mathrm{M}_2(\mathbb{C})}(M)=
\langle \psi|(M\otimes 1)\psi \rangle=\dfrac{1}{2}\mathrm{Tr}[M]
\end{equation}
for all $M\in \mathrm{M}_2(\mathbb{C})$.
This is an example that every normal state on a von Neumann algebra is defined by a vector state
on a larger von Neumann algebra than the original one.
Then we have
\begin{align}
D_{\sigma_x\otimes 1,\sigma_y\otimes 1,\omega_\psi}&=
C_{\sigma_x\otimes 1,\sigma_y\otimes 1,\omega_\psi}=
\dfrac{1}{2}|\langle \psi|[\sigma_x\otimes 1,\sigma_y\otimes 1]\psi \rangle|
=\dfrac{1}{2}|\langle \psi|2i(\sigma_z\otimes 1)\psi \rangle|=0, \\
D_{\sigma_x,\sigma_y,\frac{1}{2}\mathrm{Tr}}&=\dfrac{1}{2}\mathrm{Tr}
\left| \sqrt{\dfrac{1}{2}}[\sigma_x,\sigma_y]\sqrt{\dfrac{1}{2}}\right|
=\dfrac{1}{2}\mathrm{Tr}
\left| \dfrac{1}{2} 2i\sigma_z\right|=\dfrac{1}{2}\mathrm{Tr}[1]=1.
\end{align}
Therefore, $D_{\sigma_x\otimes 1,\sigma_y\otimes 1,\omega_\psi}< D_{\sigma_x,\sigma_y,\frac{1}{2}\mathrm{Tr}}$.
That is, when we consider the system as a range where the observable algebra is $\mathcal{N}$,
the bound value is $0$; however, when only observables contained in $\mathcal{M}$
are considered, the bound value is $1$.
\end{example}
Generally, it holds that the bound becomes smaller
as we consider systems with higher degrees of freedom.
This is because the class of physically admissible measurements is the wider
the more observables can be involved in the interaction Hamiltonian for the
measurement.
Therefore, the designation of the observable algebra (of the system to be measured)
substantially contributes to the determination of the bound.
We would like to emphasize that the use of $D_{AB}$ is firmly valuable.
For more detailed discussions we refer the reader to \cite{14EDR,16A3}.
\sloppy
In this paper, we extend Eq.(\ref{Ozawa2}) to the setting of $\sigma$-finite
von Neumann algebras.
Let $(\mathcal{M},\mathcal{H},\mathcal{P},J)$ be a standard form
with a $\sigma$-finite von Neumann algebra $\mathcal{M}$, i.e.,
$(\mathcal{M},\mathcal{H},\mathcal{P},J)$ is a quadruple consisting of
a Hilbert space $\mathcal{H}$,
a $\sigma$-finite von Neumann algebra $\mathcal{M}$ on $\mathcal{H}$,
a self-dual cone $\mathcal{P}$ of $\mathcal{H}$
and the modular conjugation $J$ of $\mathcal{M}$.
We refer the reader to Section \ref{2} (and \cite{BR02,Haa75,Tak02})
for the theory of the standard form of von Neumann algebras.
Let $A$ and $B$ be elements of $\mathcal{M}_{s.a.}$
and $\rho$ a normal state on $\mathcal{M}$.
We define a normal functional $\omega_{A,B,\rho}$ on $\mathcal{M}$ by
\begin{equation}
\omega_{A,B,\rho}(M)=\langle \xi_\rho|M J(-i[A,B])J\xi_\rho\rangle
\end{equation}
for all $M\in\mathcal{M}$, where $\xi_\rho$ is a unique unit vector of $\mathcal{P}$ such that
$\langle \rho,M\rangle=\langle \xi_\rho|M \xi_\rho\rangle$
for all $M\in\mathcal{M}$. Here we redefine $D_{A,B,\rho}$ by
\begin{equation}
D_{A,B,\rho}= \frac{1}{2}\Vert \omega_{A,B,\rho}\Vert.
\end{equation}
This coincides with the original one in the case of $\mathcal{M}=\textrm{\boldmath $B$}(\mathcal{H})$.
It is then obvious that
\begin{equation}
D_{A,B,\rho}=\frac{1}{2}\Vert \omega_{A,B,\rho}\Vert\geq \frac{1}{2}|\langle \omega_{A,B,\rho},1 \rangle|
=\frac{1}{2}|\langle \xi_\rho|J(-i[A,B])J\xi_\rho\rangle|=C_{A,B,\rho}.
\end{equation}
Our purpose is to derive Eq.(\ref{Ozawa2}) in terms of $D_{A,B,\rho}$ redefined above
in the setting of general $\sigma$-finite von Neumann algebras.
For the original case of $\mathcal{M}=\textrm{\boldmath $B$}(\mathcal{H})$,
Ozawa \cite{14EDR} presented two proofs of the derivation of Eq.(\ref{Ozawa2}):
One is based on the method, called the ``canonical purification", in terms of
the dual Hilbert space $\mathcal{H}^\ast$ of $\mathcal{H}$,
and the other is based on the representation of $\textrm{\boldmath $B$}(\mathcal{H})$
on the Hilbert space of Hilbert-Schmidt class operators on $\mathcal{H}$.
Our proof herein is a natural unification of those methods
via the theory of the standard form of von Neumann algebras
and can also be applied to any measurements described
by CP instruments which cannot be realized by any measuring processes.
In Section \ref{2}, we introduce quantum measurement theory
for quantum systems described by general $\sigma$-finite von Neumann algebras.
In Section \ref{3}, we define the error and the disturbance of measurements
used in this paper.
In Section \ref{4}, we show the main theorem of the paper, that is,
a universally valid uncertainty relation for measurement error and disturbance
in general quantum systems described by $\sigma$-finite von Neumann algebras.
In Section \ref{5}, we also show a universally valid uncertainty relation
for simultaneous measurements.
In Section \ref{6}, we conclude the paper with some remarks.
\section{Preliminaries on Quantum Measurement}\label{2}
Here we introduce quantum measurement theory based on completely positive (CP) instruments defined on
general ($\sigma$-finite) von Neumann algebras,
which enables us to describe processes of measurement in
quantum systems with infinite degrees of freedom, expecially, in quantum fields.
The previous investigation \cite{16A1} by the authors
much contributes to the development of the theory and it mathematics.
Our attempt herein, the establishment of universally valid uncertainty relations
in general quantum systems, is its succeeding program.
Extending the scope of application of uncertain relations to quantum fields
is essential for developing both foundations of quantum theory
and quantitative analysis in quantum field theory.
In particular, many physicists, inspired by quantum information theory,
are recently very interested in the latter.
Therefore, our study has potential demand in physics and is not just mathematical concern.
This section provides us with preliminaries on recent quantum measurement theory enough to understand
the physical setting and mathematical proof of universally valid uncertainty relations.
We refer the reader to \cite{04URN,14MFQ,16A1}
for detailed expositions of quantum measurement theory
based on CP instruments and measuring processes.
First we shall define the concept of CP instrument
describing output probabilities and dynamical changes of states caused by physically realizable measurement.
Let $\mathcal{M}$ be a von Neumann algebra on a Hilbert space $\mathcal{H}$
and $(S,\mathcal{F})$ a measurable space.
Let $\mathrm{P}(\mathcal{M}_\ast)$ denote the set of positive linear maps on $\mathcal{M}_\ast$.
\begin{definition}[Instrument \text{\cite{DL70,Dav76,84QC}}]
A map $\mathcal{I}:\mathcal{F}\rightarrow \mathrm{P}(\mathcal{M}_\ast)$ is called
an instrument $\mathcal{I}$ for $(\mathcal{M},S)$
if it satisfies the following two conditions:\\
$(1)$ $\Vert\mathcal{I}(S)\rho\Vert=\Vert\rho\Vert$ for all $\rho\in \mathcal{M}_{\ast,+}$;\\
$(2)$ For every $\rho\in\mathcal{M}_\ast$, $M\in\mathcal{M}$ and
countable mutually disjoint sequence $\{\Delta_j\}\subset\mathcal{F}$,
\begin{equation}
\langle \mathcal{I}(\cup_j \Delta_j)\rho, M \rangle
=\sum_j \langle \mathcal{I}(\Delta_j)\rho, M \rangle.
\end{equation}
An instrument $\mathcal{I}$ for $(\mathcal{M},S)$ is called a completely positive instrument,
or a CP instrument for short, if $\mathcal{I}(\Delta)$ is completely positive for every $\Delta\in\mathcal{F}$.
\end{definition}
An instrument $\mathcal{I}$ for $(\mathcal{M},S)$ represents a measuring apparatus
${\bf A}({\bf x})$ with output variable ${\bf x}$ taking values in the measurable space $(S,\mathcal{F})$,
and specifies both the probability measure $\mathrm{Pr}\{{\bf x}\in(\cdot)\Vert\rho\}$
of ${\bf x}$ and the family $\{\rho_{\{{\bf x}\in\Delta\}}\}_{\Delta\in\mathcal{F}}$
of states after the measurement in each normal state $\rho$ on $\mathcal{M}$, which are given by
\begin{align}
\mathrm{Pr}\{{\bf x}\in\Delta\Vert\rho\}&=\Vert\mathcal{I}(\Delta)\rho\Vert,\hspace{5mm}\Delta\in\mathcal{F},\\
\rho_{\{{\bf x}\in\Delta\}}&=
\left\{
\begin{array}{ll}
\dfrac{\mathcal{I}(\Delta)\rho}{\Vert\mathcal{I}(\Delta)\rho\Vert},
&\quad(\mathrm{if}\;\mathrm{Pr}\{{\bf x}\in\Delta\Vert\rho\}>0),\\
\mathcal{I}(S)\rho, &\quad(\mathrm{otherwise}),
\end{array}
\right.
\end{align}
respectively, where each state $\rho_{\{{\bf x}\in\Delta\}}$ realizes after the measurement
when $\rho$ is prepared before the measurement and output values of ${\bf x}$
not contained in $\Delta$ is ignored during data processing.
Conversely, an instrument $\mathcal{I}$ for $(\mathcal{M},S)$ is defined by
a measuring apparatus ${\bf A}({\bf x})$ with
output variable ${\bf x}$ taking values in the measurable space $(S,\mathcal{F})$
if for all $M\in\mathcal{M}_{s.a.}$
the joint probability measure $\mathrm{Pr}\{(M,{\bf x})\in\Delta\Vert\rho\}$
on $(\mathbb{R}\times S,\mathcal{B}(\mathbb{R})\times\mathcal{F})$
of the successive measurement carried out by ${\bf A}({\bf x})$ and
the measurement of $M$ in this order
is an affine function of $\rho\in\mathcal{S}_n(\mathcal{M})$.
Here, the map $\Gamma\times\Delta\in\mathcal{B}(\mathbb{R})\times\mathcal{F}
\mapsto\mathrm{Pr}\{(M,{\bf x})\in\Gamma\times\Delta\Vert\rho\}$ is defined by
\begin{equation}
\mathrm{Pr}\{M\in\Gamma,{\bf x}\in\Delta\Vert\rho\}=\mathrm{Pr}\{{\bf x}\in\Delta\Vert\rho\}
\mathrm{Pr}\{M\in\Gamma\Vert\rho_{\{{\bf x}\in\Delta\}}\}
\end{equation}
and is then uniquely extended into the probability measure on
$(\mathbb{R}\times S,\mathcal{B}(\mathbb{R})\times\mathcal{F})$, where
$\mathrm{Pr}\{M\in\Gamma,{\bf x}\in\Delta\Vert\rho\}=\mathrm{Pr}\{(M,{\bf x})\in\Gamma\times\Delta\Vert\rho\}$.
In addition, it is known that an instrument is CP if and only if
the measuring apparatus that defines the given instrument satisfies
the condition called the trivial extendability \cite{04URN,14MFQ,16A1}.
Now we consider a map $\mathcal{I}:\mathcal{M}\times\mathcal{F}\rightarrow \mathcal{M}$
satisfying the following three conditions:\\
$(i)$ For every $\Delta\in\mathcal{F}$, $M\mapsto \mathcal{I}(\Delta,M)$ is
a normal positive linear map on $\mathcal{M}$.\\
$(ii)$ $\mathcal{I}(1,S)=1$.\\
$(iii)$ For every $\rho\in\mathcal{M}_\ast$, $M\in\mathcal{M}$ and
countable mutually disjoint sequence $\{\Delta_j\}\subset\mathcal{F}$,
\begin{equation}
\langle \rho, \mathcal{I}(M,\cup_j \Delta_j) \rangle
=\sum_j\langle \rho, \mathcal{I}(M,\Delta_j)\rangle.
\end{equation}
There is a one-to-one correspondence between an instrument for $(\mathcal{M},S)$
and a map $\mathcal{I}:\mathcal{M}\times\mathcal{F}\rightarrow \mathcal{M}$
satisfying the above three conditions, which is given by the relation
\begin{equation}
\langle \mathcal{I}(\Delta)\rho, M \rangle=\langle \rho, \mathcal{I}(M,\Delta)\rangle
\end{equation}
for all $\rho\in\mathcal{M}_\ast$, $M\in\mathcal{M}$ and $\Delta\in\mathcal{F}$.
Thus we also call the map $\mathcal{I}:\mathcal{M}\times\mathcal{F}\rightarrow \mathcal{M}$
an instrument $(\mathcal{M},S)$.
We then define a probability operator-valued measure $\Pi_\mathcal{I}$
associated with an instrument $\mathcal{I}$ for $(\mathcal{M},S)$ by
$\Pi_\mathcal{I}(\Delta)=\mathcal{I}(1,\Delta)$ for all $\Delta\in\mathcal{F}$.
Let $\Pi:\mathcal{B}(\mathbb{R})\rightarrow\mathcal{M}$ be a probability operator-valued measure.
For every $n\in\mathbb{N}$, we define the symmetric operator $\Pi^{(n)}$ by
\begin{equation}
\langle\xi|\Pi^{(n)}\eta\rangle = \int x^n\;d\langle \xi|\Pi(x)\eta\rangle
\end{equation}
for any $\xi,\eta\in\mathrm{dom}(\Pi^{(n)})$, where we define the domain $\mathrm{dom}(\Pi^{(n)})$ by
\begin{equation}
\mathrm{dom}(\Pi^{(n)})=\{\xi\in\mathcal{H}\;|\;\int_\mathbb{R}x^{2n}\;d\langle \xi| \Pi(x)\xi\rangle
<\infty\}.
\end{equation}
Every CP instrument admits the following representation theorem.
\begin{proposition}[\text{\cite[Proposition 4.2]{84QC}}]\label{Naimark-Ozawa}
For any CP instrument $\mathcal{I}$ for $(\mathcal{M},S)$,
there are a Hilbert space $\mathcal{K}$,
a spectral measure $E:\mathcal{F}\rightarrow \textrm{\boldmath $B$}(\mathcal{K})$,
a nondegenerate normal representation
$\pi:\mathcal{M}\rightarrow\textrm{\boldmath $B$}(\mathcal{K})$
and an isometry $V\in\textrm{\boldmath $B$}(\mathcal{H},\mathcal{K})$ satisfying
\begin{align}
\mathcal{I}(M,\Delta) &=V^\ast \pi(M)E(\Delta)V, \label{CPrep}\\
E(\Delta)\pi(M) &=\pi(M)E(\Delta)
\end{align}
for any $\Delta\in\mathcal{F}$ and $M\in\mathcal{M}$,
and $\mathcal{K}=\overline{\mathrm{span}}(\pi(\mathcal{M})E(\mathcal{F})V\mathcal{H}))$.
\end{proposition}
The quadruple $(\mathcal{K},E,\pi,V)$ in the above proposition is unique up to unitary equivalence.
We call the quadruple $(\mathcal{K},E,\pi,V)$ a minimal dilation of $\mathcal{I}$.
Let $\mathcal{X}$ and $\mathcal{Y}$ be C$^\ast$-algebras on Hilbert spaces $\mathcal{H}$ and $\mathcal{K}$,
respectively.
The minimal tensor product of $\mathcal{X}$ and $\mathcal{Y}$, denoted by
$\mathcal{X}\otimes_{\mathrm{min}}\mathcal{Y}$, is defined by the completion of
the algebraic tensor product $\mathcal{X}\otimes_{\mathrm{alg}}\mathcal{Y}$
of $\mathcal{X}$ and $\mathcal{Y}$
(as a $^\ast$-subalgebra of $\textrm{\boldmath $B$}(\mathcal{H}\otimes\mathcal{K})$)
by the norm topology of $\textrm{\boldmath $B$}(\mathcal{H}\otimes\mathcal{K})$.
Let $\mathcal{M}$ and $\mathcal{N}$ be von Neumann algebras on Hilbert spaces $\mathcal{H}$ and $\mathcal{K}$,
respectively. The $W^\ast$-tensor product of $\mathcal{M}$ and $\mathcal{M}$, denoted by
$\mathcal{M}\;\overline{\otimes}\;\mathcal{N}$, is defined by the completion of
$\mathcal{M}\otimes_{\mathrm{alg}}\mathcal{N}$ by the ultraweak topology
of $\textrm{\boldmath $B$}(\mathcal{H}\otimes\mathcal{K})$.
We refer the reader to \cite{EL77,Tak79} for details on tensor products of operator algebras.
By Proposition \ref{Naimark-Ozawa} \cite[Proposition 3.3]{16A1},
for every CP instrument $\mathcal{I}$ for $(\mathcal{M},S)$,
there exists a unique unital CP map
$\Psi_\mathcal{I}:\mathcal{M}\otimes_{\mathrm{min}} L^\infty(S,\mathcal{I})\rightarrow\mathcal{M}$
such that $\Psi_\mathcal{I}(M\otimes[\chi_\Delta])=\mathcal{I}(M,\Delta)$
for all $M\in\mathcal{M}$ and $\Delta\in\mathcal{F}$.
Next, we shall define the concept of measuring process,
which is nothing but a quantum mechanical modeling of measuring apparatus.
Let $\mathcal{M}$ and $\mathcal{N}$ be von Neumann algebras.
For every $\sigma\in \mathcal{N}_\ast$, the map
$\mathrm{id}\otimes\sigma:\mathcal{M}\;\overline{\otimes}\;\mathcal{N}
\rightarrow\mathcal{M}$ is defined by
$\langle \rho\otimes\sigma,X \rangle=\langle \rho,(\mathrm{id}\otimes\sigma)(X) \rangle$
for all $X\in\mathcal{M}\;\overline{\otimes}\;\mathcal{N}$
and $\rho\in\mathcal{M}_\ast$.
\begin{definition}[Measuring process \text{\cite[Definition 3.4]{16A1}}]
A measuring process $\mathbb{M}$ for $(\mathcal{M},S)$
is a quadruple $\mathbb{M}=(\mathcal{K},\sigma,F,U)$
consisting of a Hilbert space $\mathcal{K}$, a normal
state $\sigma$ on $\textrm{\boldmath $B$}(\mathcal{K})$,
a spectral measure $F:\mathcal{F}\rightarrow \textrm{\boldmath $B$}(\mathcal{K})$,
and a unitary operator $U$ on $\mathcal{H}\otimes\mathcal{K}$
satisfying
\begin{equation}
\{\mathcal{I}_\mathbb{M}(M,\Delta)\;|\;M\in\mathcal{M},\Delta\in\mathcal{F}\}\subset\mathcal{M},
\end{equation}
where $\mathcal{I}_\mathbb{M}$ is a CP instrument for $(\textrm{\boldmath $B$}(\mathcal{H}),S)$
defined by $\mathcal{I}_\mathbb{M}(X,\Delta)=(\mathrm{id}\otimes\sigma)[U^\ast(X\otimes F(\Delta))U]$
for all $X\in\textrm{\boldmath $B$}(\mathcal{H})$ and $\Delta\in\mathcal{F}$.
\end{definition}
\begin{definition}[Statistical equivalence class of measuring processes \cite{84QC}]
Two measuring processes $\mathbb{M}_1=(\mathcal{K}_1,\sigma_1,F_1,U_1)$
and $\mathbb{M}_2=(\mathcal{K}_2,\sigma_2,F_2,U_2)$ for $(\mathcal{M},S)$
are said to be statistically equivalent if
$\mathcal{I}_{\mathbb{M}_1}(M,\Delta)=\mathcal{I}_{\mathbb{M}_2}(M,\Delta)$
for all $M\in\mathcal{M}$ and $\Delta\in\mathcal{F}$.
\end{definition}
MO established the following one-to-one correspondence for the case of
$\mathcal{M}=\textrm{\boldmath $B$}(\mathcal{H})$.
\begin{theorem}[\text{\cite[Theorem 5.1]{84QC}}]
Let $\mathcal{H}$ be a Hilbert space and $(S,\mathcal{F})$ a measurable space.
Then there is a one-to-one correspondence
between statistical equivalence classes of measuring processes $\mathbb{M}=(\mathcal{K},\sigma,F,U)$
for $(\textrm{\boldmath $B$}(\mathcal{H}),S)$
and CP instruments $\mathcal{I}$ for $(\textrm{\boldmath $B$}(\mathcal{H}),S)$,
which is given by the relation $\mathcal{I}(X,\Delta)=\mathcal{I}_\mathbb{M}(X,\Delta)$
for all $\Delta\in\mathcal{F}$ and $X\in\textrm{\boldmath $B$}(\mathcal{H})$.
\end{theorem}
This theorem states that every CP instrument is modeled by a measuring process and
every measuring process defines a CP instrument.
To generalize the above theorem to general $\sigma$-finite von Neumann algebras,
we define the following property for CP instruments.
\begin{definition}[Normal extension property \text{\cite[Definition 3.4]{16A1}}]
Let $\mathcal{I}$ be a CP instrument for $(\mathcal{M},S)$.
$\mathcal{I}$ has the normal extension property (NEP) if there exists a unital normal CP map
$\widetilde{\Psi_\mathcal{I}}:\mathcal{M}\;\overline{\otimes}\; L^\infty(S,\mathcal{I})\rightarrow
\textrm{\boldmath $B$}(\mathcal{H})$ such that
$\widetilde{\Psi_\mathcal{I}}|_{\mathcal{M}\otimes_{\mathrm{min}}L^\infty(S,\mathcal{I})}=\Psi_\mathcal{I}$.
\end{definition}
The authors established the following theorem in \cite{16A1}.
\begin{theorem}[\text{\cite[Theorem 3.2]{16A1}}]\label{mainse4}
Let $\mathcal{M}$ be a von Neumann algebra on a Hilbert space
$\mathcal{H}$ and $(S,\mathcal{F})$ a measurable space.
Then there is a one-to-one correspondence
between statistical equivalence classes of measuring processes $\mathbb{M}=(\mathcal{K},\sigma,F,U)$
for $(\mathcal{M},S)$ and CP instruments $\mathcal{I}$ for $(\mathcal{M},S)$ \textbf{with the NEP},
which is given by the relation $\mathcal{I}(M,\Delta)=\mathcal{I}_{\mathbb{M}}(M,\Delta)$
for all $\Delta\in\mathcal{F}$ and $M\in\mathcal{M}$.
\end{theorem}
Thus every CP instrument defined on general $\sigma$-finite von Neumann algebras
is not always realized by a measuring process.
The followings are examples of CP instruments without the NEP given in \cite[Section 4]{16A1}.
\begin{example}[\textrm{\cite[Example 5.1]{16A1}}]\label{NoNEP1}
Let $m$ be Lebesgue measure on $[0,1]$. A CP instrument $\mathcal{I}_m$ for
$(L^{\infty}([0,1],m), [0,1])$ is defined by $\mathcal{I}_m(f,\Delta)=[\chi_\Delta] f$
for all $\Delta\in\mathcal{B}([0,1])$ and $f\in L^{\infty}([0,1],m)$.
\end{example}
\begin{example}[\textrm{\cite[Example 5.2]{16A1}}]\label{NoNEP2}
Let $\mathcal{R}$ be an approximately finite-dimensional (AFD) von Neumann algebra of type $\mathrm{II}_1$
on a separable Hilbert space $\mathcal{H}$.
Let $A$ be a self-adjoint operator with continuous spectrum affiliated with $\mathcal{R}$
and $\mathcal{E}$ a (normal) conditional expectation of $\mathcal{R}$ onto
$\{A\}^\prime\cap\mathcal{R}$,
where $\{A\}^\prime=\{E^{A}(\Delta)\;|\; \Delta\in \mathcal{B}(\mathbb{R})\}^\prime$.
A CP instrument $\mathcal{I}_A$ for $(\mathcal{R},\mathbb{R})$ is defined by
\begin{equation}\label{repeatable}
\mathcal{I}_A(M,\Delta)=\mathcal{E}(M)E^A(\Delta)
\end{equation}
for all $M\in\mathcal{R}$ and $\Delta\in\mathcal{B}(\mathbb{R})$.
\end{example}
It also is shown in \cite[Section 4]{16A1}
that every CP instrument defined on atomic von Neumann algebras has the NEP.
By contrast, CP instruments without the NEP are defined on non-atomic (injective) von Neumann algebras.
Let $\mathcal{M}$ be an injective von Neumann algebra on a Hilbert space $\mathcal{H}$
and $(S,\mathcal{F})$ a measurable space.
For every CP instrument $\mathcal{I}$ for $(\mathcal{M},S)$
there exists a net $\{\mathcal{I}_\alpha\}_{\alpha\in A}$
of CP instruments for $(\mathcal{M},S)$ with the NEP such that $\mathcal{I}_\alpha$
ultraweakly converges to $\mathcal{I}$ and that $\mathcal{I}(1,\Delta)=\mathcal{I}_\alpha(1,\Delta)$
for all $\alpha$ and $\Delta\in\mathcal{F}$ \cite[Section 4]{16A1}.
Since local algebras in algebraic quantum field theory are injective,
we can apply these results to the characterization of local measurements of quantum fields
\cite[Section 6]{16A1}.
Lastly, we shall introduce the theory of standard forms of
von Neumann algebras.
Let $\mathcal{H}$ be a Hilbert space.
For every subset $S$ of $\textrm{\boldmath $B$}(\mathcal{H})$,
let $S^\prime$ denote the commutant of $S$, i.e.,
$S^\prime=\{A\in\textrm{\boldmath $B$}(\mathcal{H})\;|\;AB=BA\;
\mathrm{for}\;\mathrm{all}\;B\in S\}$.
We call a convex subset $\mathcal{P}$ of $\mathcal{H}$
a cone of $\mathcal{H}$. For every cone $\mathcal{P}$ of $\mathcal{H}$, we define
the dual cone $\mathcal{P}^\vee$ of $\mathcal{P}$ by
\begin{equation}
\mathcal{P}^\vee=\{\xi\in\mathcal{H}\;|\;\langle \xi| \eta\rangle\geq 0\;
\mathrm{for}\;\mathrm{all}\;\eta\in\mathcal{P}\}.
\end{equation}
A cone $\mathcal{P}$ of $\mathcal{H}$ is said to be self-dual if $\mathcal{P}^\vee=\mathcal{P}$.
For a von Neumann algebra $\mathcal{M}$ on a Hilbert space $\mathcal{H}$,
a quadruple $(\mathcal{M},\mathcal{H},\mathcal{P},J)$ of an anti-linear isometry $J$ with $J^2=1$,
called the modular conjugation $J$ of $\mathcal{M}$,
and a self-dual cone $\mathcal{P}$ of $\mathcal{H}$ is called a standard form of $\mathcal{M}$
if it satisfies the following four conditions:\\
$(1)$ $J\mathcal{M}J=\mathcal{M}^\prime$;\\
$(2)$ $J\xi=\xi$ for all $\xi\in\mathcal{P}$;\\
$(3)$ $MJMJ\mathcal{P}\subset\mathcal{P}$ for all $M\in\mathcal{M}$;\\
$(4)$ $JZJ=Z^\ast$ for all $Z\in\mathcal{Z}(\mathcal{M})=\mathcal{M}\cap\mathcal{M}^\prime$.\\
In fact, it is shown in \cite[Lemma 3.19]{AH14} that the fourth condition is redundant.
Namely, a quadruple $(\mathcal{M},\mathcal{H},\mathcal{P},J)$
satisfying the conditions $(1)$, $(2)$ and $(3)$ is a standard form of $\mathcal{M}$.
For every von Neumann algebra $\mathcal{N}$ on a Hilbert space $\mathcal{K}$,
there exists a standard form $(\mathcal{M},\mathcal{H},\mathcal{P},J)$ such that
$\mathcal{N}$ is $W^\ast$-isomorphic to $\mathcal{M}$.
Therefore, we assume that von Neumann algebras appearing in the paper
are in standard forms without loss of generality.
\section{Error and Disturbance}\label{3}
In this section, we define an error and a disturbance in quantum measurement theory,
which are introduced and applied in past investigations \cite{88MS,89RS,03UVR,03UPQ,04URN}.
The concept of error is fundamental for measuring how accurately a measuring apparatus can measure an observable,
and is defined (or characterized) as the root-mean-square of
the ``difference" between an observable to be measured
and an output variable of the measuring apparatus as an observable actually measured.
On the other hand, the concept of disturbance is essential for
estimating the effect of measurement on the system to be measured,
and is defined as the root-mean-square of the ``difference" between an observable
before the measurement and the identical one after the measurement.
It should be noted that the disturbance of an observable $B$ caused by a measurement
of $A$ can be defined as the error of such a measurement of $B$ in the state just before
the $A$-measurement that is actually carried out by the precise $B$-measuring
apparatus just after the $A$-measurement \cite{03UVR,03UPQ}.
These quantities have been studied from various perspectives,
therefore, we will mention them minimally herein.
Let $\mathcal{M}$ be a von Neumann algebra on a Hilbert space $\mathcal{H}$
and $A$, $B$ self-adjoint elements of $\mathcal{M}$.
Let $\mathcal{I}$ be a CP instrument for $(\mathcal{M},\mathbb{R})$.
$\mathcal{I}$ physically corresponds to a measuring apparatus with output variable ${\bf x}$ taking values in
$(\mathbb{R},\mathcal{B}(\mathbb{R}))$.
We then define an error $\varepsilon(A)=
\varepsilon(A,\rho;\mathcal{I})$ of measurement of $A$ in $\rho$
and a disturbance $\eta(B)=\eta(B,\rho;\mathcal{I})$ of $B$ in $\rho$ by
\begin{align}
\varepsilon(A,\rho;\mathcal{I}) &=
\langle \rho,\Pi_\mathcal{I}^{(2)}-A\Pi_\mathcal{I}^{(1)}-\Pi_\mathcal{I}^{(1)}A+A^2 \rangle^{\frac{1}{2}},\\
\eta(B,\rho;\mathcal{I}) &=
\langle \rho, \mathcal{I}(B^2,\mathbb{R})-B\mathcal{I}(B,\mathbb{R})-
\mathcal{I}(B,\mathbb{R})B+B^2\rangle^{\frac{1}{2}},
\end{align}
respectively.
If $\mathcal{I}$ is realized by a measuring process
$(\mathcal{K},\sigma,F,U)$ for $(\mathcal{M},\mathbb{R})$,
it holds that
\begin{align}
\varepsilon(A,\rho;\mathcal{I}) &=\langle
\tilde{\rho}\otimes\sigma,(U^\ast(1\otimes F^{(1)})U-A\otimes1)^2 \rangle^{\frac{1}{2}},\\
\eta(B,\rho;\mathcal{I}) &=\langle
\tilde{\rho}\otimes\sigma,(U^\ast(B\otimes 1)U-B\otimes1)^2 \rangle^{\frac{1}{2}},
\end{align}
where $\tilde{\rho}$ is a normal state on $\textrm{\boldmath $B$}(\mathcal{H})$
such that $\rho(M)=\tilde{\rho}(M)$ for all $M\in\mathcal{M}$.
The operator $U^\ast(1\otimes F^{(1)})U-A\otimes1$ is often called the noise
operator of the measuring process $(\mathcal{K},\sigma,F,U)$ in measuring $A$ and the error
measure $\varepsilon$ is often called the noise-operator-based error.
Both $\varepsilon(A,\rho;\mathcal{I})$ and $\eta(B,\rho;\mathcal{I})$
are considered as
a natural generalization of
those in classical probability theory.
It should also be remarked that the definitions of error and disturbance
are independent of the existence and the choice of measuring processes
which realize the given CP instrument.
A justification of the use of the noise-operator-based error was discussed
extensively in Ref.~\cite{18PPT}; it will be briefly discussed also in the last
section.
Let $\mathcal{I}$ be a CP instrument for $(\mathcal{M},\mathbb{R})$ and
$(\mathcal{K},E,\pi,V)$ the minimal dilation of $\mathcal{I}$.
We assume that $E^{(1)}$ is a bounded operator on $\mathcal{K}$,
so that $E^{(2)}=(E^{(1)})^2$.
Then we have
\begin{align}
\varepsilon(A,\rho;\mathcal{I}) &=
\langle \rho,(E^{(1)}V-VA)^\ast (E^{(1)}V-VA)\rangle^{\frac{1}{2}},\\
\eta(B,\rho;\mathcal{I}) &=
\langle \rho,(\pi(B)V-VB)^\ast (\pi(B)V-VB) \rangle^{\frac{1}{2}}.
\end{align}
\section{Main Theorem}\label{4}
We are now ready to state the main theorem.
Let $(\mathcal{M},\mathcal{H},\mathcal{P},J)$ be a standard form
with a $\sigma$-finite von Neumann algebra $\mathcal{M}$.
\begin{theorem}[]\label{EDUR}
Let $A,B$ be elements of $\mathcal{M}_{s.a.}$, $\rho\in\mathcal{S}_n(\mathcal{M})$,
$\mathcal{I}$ a CP instrument for $(\mathcal{M},\mathbb{R})$ and
$(\mathcal{K},E,\pi,V)$ the minimal dilation of $\mathcal{I}$.
Assume that $E^{(1)}$ is a bounded operator on $\mathcal{K}$. Then we have
\begin{equation}
\varepsilon(A)^2\sigma(B)^2+\sigma(A)^2\eta(B)^2+2\varepsilon(A)\eta(B)
\sqrt{\sigma(A)^2\sigma(B)^2-D_{AB}^2}
\geq D_{AB}^2.
\end{equation}
\end{theorem}
Branciard \cite{Bra13} proved that
the following relation, called Branciard's geometric inequality, holds.
\begin{proposition}{}\label{Bra}
Let $\mathcal{L}$ be a real vector space with real inner product $(\cdot,\cdot)$.
For any vectors
$\mathbf{a},\mathbf{b},\mathbf{m},\mathbf{n}\in\mathcal{L}$ with $\mathbf{m}\;\bot\;\mathbf{n}$,
\begin{equation} \label{Branciard}
\Vert \mathbf{a}-\mathbf{m}\Vert^2 \Vert \mathbf{b}\Vert^2
+\Vert \mathbf{a}\Vert^2 \Vert \mathbf{b}-\mathbf{n}\Vert^2
+2\Vert \mathbf{a}-\mathbf{m}\Vert \Vert \mathbf{b}-\mathbf{n}\Vert
\sqrt{\Vert \mathbf{a}\Vert^2 \Vert \mathbf{b}\Vert^2-(\mathbf{a},\mathbf{b})^2}
\geq (\mathbf{a},\mathbf{b})^2.
\end{equation}
\end{proposition}
The proof is given in \cite[SI Text, section B]{Bra13}.
For every linear functional $\omega$ on $\mathcal{M}$, the adjoint functional $\omega^\ast$ of $\omega$
is defined by
\begin{equation}
\omega^\ast(M)=\overline{\omega(M^\ast)}
\end{equation}
for all $M\in\mathcal{M}$ \cite{Tak79}. We say that a linear functional $\omega$ on $\mathcal{M}$ is hermitian
if $\omega^\ast=\omega$. For every normal functional $\omega$ on $\mathcal{M}$,
there exist the smallest projections $E$ and $F$ such that
\begin{equation}
\langle \omega, M \rangle=\langle \omega, M E\rangle,
\hspace{5mm}\langle \omega, N \rangle=\langle \omega, FN \rangle
\end{equation}
for all $M,N\in\mathcal{M}$. The projections $E$ and $F$ are called, respectively,
the left and right support projections of $\omega$ and denoted by $S_l(\omega)$ and $S_r(\omega)$.
If $\omega$ is hermitian, then $S_r(\omega)=S_l(\omega)$, so that they are written as $S(\omega)$ \cite{Tak79}.
\begin{proposition}[\text{\cite[Chapter III, Theorem 4.2 and its proof]{Tak79}}]
Let $\mathcal{M}$ be a von Neumann algebra.
For every normal linear functional $\omega$ on $\mathcal{M}$,
there exist a partial isometry $V\in\mathcal{M}$ and
a positive normal linear functional $\varphi$ on $\mathcal{M}$ such that
\begin{equation}\label{polar}
\langle \omega,M \rangle=\langle \varphi,MV \rangle
\hspace{5mm}V^\ast V=S(\varphi).
\end{equation}
Furthermore, we have $S(\varphi)=S_r(\omega)$ and $VV^\ast=S_l(\omega)$.
If $\omega$ is hermitian, $V$ is self-adjoint.
\end{proposition}
The expression of $\omega$ in Eq.(\ref{polar}) is called the polar decomposition of $\omega$
and $\varphi$ is called the absolute value of $\omega$ and denoted by $|\omega|$.
We use the above propositions to prove Theorem \ref{EDUR}.
\begin{proof}[Proof of Theorem \ref{EDUR}]
The norm $\Vert \omega_{A,B,\rho}\Vert$ of $\omega_{A,B,\rho}$ satisfies the equality
\begin{equation}
\Vert \omega_{A,B,\rho}\Vert=\langle |\omega_{A,B,\rho}|,1\rangle.
\end{equation}
Since $\omega_{A,B,\rho}$ is hermitian,
there exists a self-adjoint partial isometry $W\in\mathcal{M}$ such that
\begin{equation}
\langle\omega_{A,B,\rho},M\rangle=\langle|\omega_{A,B,\rho}|,MW\rangle
\end{equation}
for all $M\in\mathcal{M}$, and that $W^\ast W=S(|\omega_{A,B,\rho}|)$.
Thus we have
\begin{equation}
\Vert \omega_{A,B,\rho}\Vert=\langle \omega_{A,B,\rho},W\rangle=\langle \xi_\rho|WJ(-i[A,B])J\xi_\rho \rangle.
\end{equation}
To use Eq. (\ref{Branciard}), we define
a real inner product $\langle \cdot|\cdot \rangle_\mathbb{R}$ on $\mathcal{K}$ by
\begin{equation}
\langle x|y \rangle_\mathbb{R} =\mathrm{Re}\langle x|y \rangle
\end{equation}
for all $x,y\in\mathcal{K}$,
and put
\begin{align}
a &= V(A-\rho(A))JW\xi_\rho,\\
b &= -iV(B-\rho(B))J\xi_\rho,\\
m &= (E^{(1)}-\rho(A))VJW\xi_\rho,\\
n &= -i(\pi(B)-\rho(B))VJ\xi_\rho.
\end{align}
Let $\Vert\cdot\Vert_\mathbb{R}$ denote the norm of $\mathcal{K}$ as a real Hilbert space
induced by $\langle \cdot|\cdot \rangle_\mathbb{R}$.
Since $WJV^\ast(E^{(1)}-\rho(A))(\pi(B)-\rho(B))VJ$ is self-adjoint,
$\langle \xi_\rho|WJV^\ast(E^{(1)}-\rho(A))(\pi(B)-\rho(B))VJ\xi_\rho \rangle$ is real.
Thus we have
\begin{align}
\langle m|n \rangle_\mathbb{R} &=\mathrm{Re}\langle (E^{(1)}-\rho(A))VJW\xi_\rho
|-i(\pi(B)-\rho(B))VJ\xi_\rho \rangle \nonumber\\
&= \mathrm{Re}\; (-i) \langle \xi_\rho|WJV^\ast(E^{(1)}-\rho(A))(\pi(B)-\rho(B))VJ\xi_\rho \rangle=0.
\end{align}
Next, $\langle a|b \rangle_\mathbb{R}$ and $\Vert a\Vert_\mathbb{R}^2$, respectively,
satisfy the following relations:
\begin{align}
\langle a|b \rangle_\mathbb{R} &=\mathrm{Re}\langle V(A-\rho(A))JW\xi_\rho
|-iV(B-\rho(B))J\xi_\rho \rangle \nonumber\\
&= \mathrm{Re}\langle (A-\rho(A))JW\xi_\rho|-i(B-\rho(B))J\xi_\rho \rangle\nonumber\\
&= \frac{\langle (A-\rho(A))JW\xi_\rho|-i(B-\rho(B))J\xi_\rho \rangle
+\langle -i(B-\rho(B))J\xi_\rho|(A-\rho(A))JW\xi_\rho \rangle}{2} \nonumber\\
&= \frac{\langle J(-i)(A-\rho(A))(B-\rho(B))JW\xi_\rho|\xi_\rho \rangle
+\langle Ji(B-\rho(B))(A-\rho(A))JW\xi_\rho|\xi_\rho \rangle}{2}\nonumber\\
&=\frac{1}{2}\langle \xi_\rho|WJ(-i[A,B])J\xi_\rho \rangle=\frac{1}{2}\Vert\omega_{A,B,\rho}\Vert
=D_{AB}.
\end{align}
\begin{align}
\Vert a\Vert_\mathbb{R}^2=\Vert a\Vert^2 &=\Vert V(A-\rho(A))JW\xi_\rho\Vert^2 \nonumber\\
&= \langle \xi_\rho|J(A-\rho(A))^2JW^2\xi_\rho\rangle \nonumber\\
&\leq \langle\xi_\rho|J(A-\rho(A))^2J\xi_\rho\rangle= \rho((A-\rho(A))^2)=\sigma(A)^2.
\end{align}
Since $\Pi_\mathcal{I}^{(1)}=V^\ast E^{(1)}V$ and $\Pi_\mathcal{I}^{(2)}=V^\ast E^{(2)}V$
are bounded by assumption, we have
\begin{align}
\Vert a-m\Vert_\mathbb{R}^2=\Vert a-m\Vert^2 &=\Vert (VA-E^{(1)}V)JW\xi_\rho\Vert^2 \nonumber\\
&= \langle \xi_\rho|J(\Pi_\mathcal{I}^{(2)}-A\Pi_\mathcal{I}^{(1)}-\Pi_\mathcal{I}^{(1)}A+A^2)J
W^2\xi_\rho\rangle \nonumber\\
&\leq \langle \xi_\rho|
J(\Pi_\mathcal{I}^{(2)}-A\Pi_\mathcal{I}^{(1)}-\Pi_\mathcal{I}^{(1)}A+A^2)J\xi_\rho\rangle\nonumber\\
&=\rho(\Pi_\mathcal{I}^{(2)}-A\Pi_\mathcal{I}^{(1)}-\Pi_\mathcal{I}^{(1)}A+A^2)=\varepsilon(A)^2.
\end{align}
Lastly, we have
\begin{align}
\Vert b\Vert_\mathbb{R}^2&=\Vert b\Vert^2=\Vert -iV(B-\rho(B))J\xi_\rho\Vert^2=\sigma(B)^2,\\
\Vert b-n\Vert_\mathbb{R}^2&=\Vert b-n\Vert^2=\Vert -i(VB-\pi(B)V)J\xi_\rho\Vert^2=\eta(B)^2.
\end{align}
Therefore, Eq. (\ref{Branciard}) implies
\begin{align}
&\; \hspace{4mm}\varepsilon(A)^2\sigma(B)^2+\sigma(A)^2\eta(B)^2+2\varepsilon(A)\eta(B)
\sqrt{\sigma(A)^2\sigma(B)^2-D_{AB}^2}\nonumber\\
&\geq \Vert (VA-E^{(1)}V)JW\xi_\rho\Vert^2\sigma(B)^2+
\Vert V(A-\rho(A))JW\xi_\rho\Vert^2\eta(B)^2 \nonumber\\
&\;\hspace{3mm}+2\Vert (VA-E^{(1)}V)JW\xi_\rho\Vert \eta(B)
\sqrt{\Vert V(A-\rho(A))JW\xi_\rho\Vert^2 \sigma(B)^2-D_{AB}^2}\geq D_{AB}^2,
\end{align}
which completes the proof of Theorem \ref{EDUR}.
\end{proof}
\section{Uncertainty Relation for Simultaneous Measurement}\label{5}
We shall define errors of $A$ and $B$ for simultaneous measurements of $A$ and $B$ in this section.
Let $\mathcal{I}$ be a CP instrument for $(\mathcal{M},\mathbb{R}^2)$
and $(\mathcal{K},\pi,E,V)$ the minimal dilation of $\mathcal{I}$.
We define spectral measures $E_{\bf x}$, $E_{\bf y}$ on $\mathcal{K}$ by
\begin{align}
E_{\bf x}(\Delta) &= E(\Delta\times\mathbb{R}),\hspace{5mm}\Delta\in\mathcal{B}(\mathbb{R}),\\
E_{\bf y}(\Gamma) &= E(\mathbb{R}\times\Gamma),\hspace{5mm}\Gamma\in\mathcal{B}(\mathbb{R}),
\end{align}
respectively.
It is natural (owing to the discussion in Section \ref{2}) to consider
that $\mathcal{I}$ corresponds to a measuring apparatus with two output variables ${\bf x}$ and ${\bf y}$,
both of which take values in $(\mathbb{R},\mathcal{B}(\mathbb{R}))$.
We then assume that we use ${\bf x}$ and ${\bf y}$ to measure $A$ and $B$, respectively.
Moreover, we assume that $E_{\bf x}^{(1)}$ and $E_{\bf y}^{(1)}$ are bounded operators on
$\mathcal{K}$ for simplicity, so that
$E_{\bf x}^{(2)}=(E_{\bf x}^{(1)})^2$ and $E_{\bf y}^{(2)}=(E_{\bf y}^{(1)})^2$.
We can naturally define errors of $A$ and $B$ in terms of $\mathcal{I}$ by
\begin{align}
\varepsilon(A,\rho;\mathcal{I}) &=
\langle \rho,V^\ast E_{\bf x}^{(2)}V-AV^\ast E_{\bf x}^{(1)}V
-V^\ast E_{\bf x}^{(1)}VA+A^2 \rangle^{\frac{1}{2}}\\
&=\langle \rho,(E^{(1)}_{\bf x}V-VA)^\ast (E^{(1)}_{\bf x}V-VA)\rangle^{\frac{1}{2}},\\
\varepsilon(B,\rho;\mathcal{I}) &=
\langle \rho, V^\ast E_{\bf y}^{(2)}V-BV^\ast E_{\bf y}^{(1)}V
-V^\ast E_{\bf y}^{(1)}VB+B^2\rangle^{\frac{1}{2}}\\
&=\langle \rho,(E^{(1)}_{\bf y}V-VB)^\ast (E^{(1)}_{\bf y}V-VB) \rangle^{\frac{1}{2}},
\end{align}
respectively.
If $\mathcal{I}$ is realized by a measuring process $(\mathcal{K},\sigma,F,U)$ for $(\mathcal{M},\mathbb{R}^2)$,
we have
\begin{align}
\varepsilon(A,\rho;\mathcal{I}) &=\langle
\tilde{\rho}\otimes\sigma,(U^\ast(1\otimes F^{(1)}_{\bf x})U-A\otimes1)^2 \rangle^{\frac{1}{2}},\\
\varepsilon(B,\rho;\mathcal{I}) &=\langle
\tilde{\rho}\otimes\sigma,(U^\ast(1\otimes F^{(1)}_{\bf y})U-B\otimes1)^2 \rangle^{\frac{1}{2}},
\end{align}
respectively,
where $\tilde{\rho}$ is a normal state on $\textrm{\boldmath $B$}(\mathcal{H})$
such that $\rho(M)=\tilde{\rho}(M)$ for all $M\in\mathcal{M}$,
and $F_{\bf x}$ and $F_{\bf y}$ are spectral measures on $\mathcal{K}$ defined by
$F_{\bf x}(\Delta)= F(\Delta\times\mathbb{R})$ for all $\Delta\in\mathcal{B}(\mathbb{R})$
and $F_{\bf y}(\Gamma) = F(\mathbb{R}\times\Gamma)$ for all $\Gamma\in\mathcal{B}(\mathbb{R})$,
respectively.
We have the following uncertainty relation for simultaneous measurements of two different observables.
Let $(\mathcal{M},\mathcal{H},\mathcal{P},J)$ be a standard form
with a $\sigma$-finite von Neumann algebra $\mathcal{M}$.
\begin{theorem}[]\label{EEUR}
Let $A,B$ be elements of $\mathcal{M}_{s.a.}$, $\rho\in\mathcal{S}_n(\mathcal{M})$,
$\mathcal{I}$ a CP instrument for $(\mathcal{M},\mathbb{R}^2)$ and
$(\mathcal{K},E,\pi,V)$ the minimal dilation of $\mathcal{I}$.
Assume that $E_{\bf x}^{(1)}$ and $E_{\bf y}^{(1)}$ are bounded operators on $\mathcal{K}$. Then we have
\begin{equation}
\varepsilon(A)^2\sigma(B)^2+\sigma(A)^2\varepsilon(B)^2+2\varepsilon(A)\varepsilon(B)
\sqrt{\sigma(A)^2\sigma(B)^2-D_{AB}^2}
\geq D_{AB}^2.
\end{equation}
\end{theorem}
\begin{proof}
We define a real inner product $\langle \cdot|\cdot \rangle_\mathbb{R}$ on $\mathcal{K}$ by
\begin{equation}
\langle x|y \rangle_\mathbb{R} =\mathrm{Re}\langle x|y \rangle
\end{equation}
for all $x,y\in\mathcal{K}$,
and put
\begin{align}
a &= V(A-\rho(A))JW\xi_\rho,\\
b &= -iV(B-\rho(B))J\xi_\rho,\\
m &= (E_{\bf x}^{(1)}-\rho(A))VJW\xi_\rho,\\
n &= -i(E_{\bf y}^{(1)}-\rho(B))VJ\xi_\rho.
\end{align}
Then we have
$\langle m|n \rangle_\mathbb{R} = 0$,
$\langle a|b \rangle_\mathbb{R} =D_{AB}$,
$\Vert a\Vert_\mathbb{R}^2 \leq \sigma(A)^2$,
$\Vert a-m\Vert_\mathbb{R}^2 \leq \varepsilon(A)^2$,
$\Vert b\Vert_\mathbb{R}^2 = \sigma(B)^2$ and
$\Vert b-n\Vert_\mathbb{R}^2=\varepsilon(B)^2$.
By Eq. (\ref{Branciard}), we get the desired inequality.
\end{proof}
\section{Concluding Remarks}\label{6}
Recently, Busch, Lahti, and Werner \cite{BLW14RMP} raised
a reliability problem for quantum generalizations of the classical
root-mean-square (rms) error, comparing the noise-operator based error,
which we adopted here,
with the Wasserstein 2-distance, another error measure
based on the distance between probability measures.
They pointed out several discrepancies between those two error measures
in favor of the latter, and claimed that a state-dependent formulation
for measurement uncertainty relations is not tenable.
In order to resolve the conflict, one of the authors \cite{18PPT}
introduced the following requirements for
any sensible error measure generalizing the classical rms error:
(I) The operational definability: The error measure be defined by the POVM of the measurement,
the observable to be measured, and the state of the system to be measured.
(II) The correspondence principle: The error measure coincide with the classical rms
error in the case where there exists the joint probability distribution
for the observable to be measured
just before the measurement and the meter observable just after the measurement.
(III) The soundness: The error measure take the value 0 if the measurement
is precise in the sense that the observable to be measured
just before the measurement and the meter observable just after the measurement are
perfectly correlated \cite{05PCN,06QPC}.
It was shown that the noise-operator based error satisfies all the requirements, (I)--(III),
whereas the Wasserstein 2-distance does not satisfy (II).
Thus, the Busch-Lahti-Werner criticism based on the comparison with
the Wasserstein 2-distance is not relevant, and their opinion against
the state-dependent formulation is unfounded.
In fact, in the classical case, where the observable algebra $\mathcal{M}$ is abelian,
the noise-operator-based error takes the value 0 if and only if
the measured observable just before the measurement and the meter observable just after the measurement take the same value with probability 1,
whereas the Wasserstein 2-distance takes the value 0 if and only if they only have
the same probability distribution. For more detailed discussions, we refer the reader to
Ref.~\cite{18PPT}.
In contrast to the violation of Heisenberg's relation \eq{Hei27} for the noise-operator-based error $\ep$, a state-dependent error measure,
Appleby \cite{App98c} considered the state-independent error $\ep_{\rm Appleby}$
defined by
\begin{equation}
\ep_{{\rm Appleby}}(A;\mathcal{I})=\sup_{\rho}\ep(A,\rho;\mathcal{I}),
\end{equation}
where the supremum is taken over all pure states $\rho$,
and derived the relation
\begin{equation}
\ep_{{\rm Appleby}}(Q;\mathcal{I})\,\ep_{{\rm Appleby}}(P;\mathcal{I})\ge\frac{\hbar}{2}
\end{equation}
for canonically conjugate observables $Q,P$ and for any joint measurement
$\mathcal{I}$ of $Q,P$,
except for the case where $\ep_{{\rm Appleby}}(Q;\mathcal{I})= 0$ or
$\ep_{{\rm Appleby}}(P;\mathcal{I})=0$.
It should be noted that in the state-independent formulation as above
the error measures
$\ep_{\rm Appleby}(Q;\mathcal{I})$ and $\ep_{\rm Appleby}(P;\mathcal{I})$
often diverge. In fact, they diverges for almost every linear measurement
\cite{13DHE}.
Even in the original $\gamma$-ray thought experiment, the error measure
$\ep_{\rm Appleby}(Q;\mathcal{I})$ obviously diverges.
The notion of the resolution power of a microscope is well-defined only in the case
where the object is well-localized in the scope of the microscope, and it cannot be
captured by the state-independent formulation.
Recently, Busch-Lahti-Werner \cite{BLW13,BLW14JMP} revived the state-independent
formulation similar to Appleby's \cite{App98c}, but the same criticisms
apply to their approach. In fact, the Busch-Lahti-Werner formulation in \cite{BLW13}
is equivalent to Appleby's formulation \cite{App98c}
for any linear measurements \cite{13DHE}.
For detailed discussions, we refer the reader to Ref.~\cite{13DHE}.
\section*{Acknowledgment}
The authors acknowledge the supports of the JSPS KAKENHI, No.~26247016, No.~17K19970, and of the IRI-NU collaboration.
KO is supported by the JSPS KAKENHI, No. 16K17641, No. 17H01277,
by Research Origin for Dressed Photon,
by Grant for Basic Science Research Projects from The Sumitomo Foundation
and by the JST CREST, Grant Number JPMJCR17N2, Japan.
|
\section{Introduction}\label{sec:intro}
Let $P$ denote a finite set of points in $3$-dimensional Euclidean (or real projective) space.
An \emph{ordinary line} of $P$ is a line passing through exactly two points of $P$.
The classical Sylvester-Gallai theorem states that any finite non-collinear point set in the plane has an ordinary line.
Green and Tao~\cite{GT13} proved a structure theorem for sets with few ordinary lines, and used it to prove the so-called Dirac-Motzkin conjecture:
any sufficiently large non-collinear $n$-point set in the plane has at least $n/2$ ordinary lines.
It is natural to ask the same question of ordinary planes, where an \emph{ordinary plane} of $P$ is a plane passing through exactly three points of $P$.
However, as pointed out by Motzkin~\cite{M51}, there are finite sets of points spanning $3$-space that do not have any ordinary planes.
His one example consists of the ten intersection points of triples of five planes in general position, and another consists of points chosen from two skew lines.
Motzkin~\cite{M51} suggested we consider instead planes $\Pi$ with all but one of the points of $P \cap \Pi$ lying on a line.
He then recovers an analogue of the Sylvester-Gallai theorem:
any finite non-coplanar set in $3$-space spans such a plane.
Purdy and Smith \cite{PS10} considered instead non-coplanar point sets in $3$-space that are in general position in the sense that no three points are collinear, proving a quadratic lower bound on the number of ordinary planes of the set.
Recently, Ball~\cite{B17} proved a structure theorem for sets with few ordinary planes by examining the so-called tetra-grid structure of the projective dual of such sets,
which is a $3$-dimensional analogue of the triangular grid used by Green and Tao~\cite{GT13}.
For a $4$-dimensional generalisation of Ball's ideas, see the very recent paper by Ball and Jimenez \cite{BJ18}.
The first aim of this paper is to give an alternative proof to a slightly more refined version of Ball's structure theorem in $3$-space (but with a stronger condition).
We do this by using results on non-generic projections of arbitrary space curves (see Section~\ref{sec:tools}), and by detailed considerations of space quartic curves, including notions from classical invariant theory (see Section~\ref{sec:quartics}).
We avoid the use of Ball's tetra-grids.
Our main result is Theorem~\ref{thm:strong} below.
A \emph{prism} is any set projectively equivalent to the vertex set of a prism over a regular polygon, which is a subset of $\mathbb{R}^3$ of $n=2m$ points of the form
\[ \setbuilder{\left( \cos \left( \frac{2k\pi}{m} \right), \sin \left( \frac{2k\pi}{m} \right), \pm 1 \right)}{k=0, \dotsc, m-1}. \]
An \emph{antiprism} is any set projectively equivalent to the vertex set of an antiprism over a regular polygon, that is, a subset of $\mathbb{R}^3$ of $n=2m$ points of the form
\begin{align*}
&\quad \setbuilder{\left( \cos \left( \frac{2k\pi}{m} \right), \sin \left( \frac{2k\pi}{m} \right), 1 \right)}{k=0, \dotsc, m-1}\\
&\cup \setbuilder{\left( \cos \left( \frac{2(k+1)\pi}{m} \right), \sin \left( \frac{2(k+1)\pi}{m} \right), -1 \right)}{k=0, \dotsc, m-1}.
\end{align*}
See also~\cite{B17}*{Section~2}.
In this theorem we refer to certain space quartic curves that have a natural group structure as described in Section~\ref{sec:quartics}.
We prove this theorem in Section~\ref{sec:proof}.
\begin{theorem}[Full structure theorem]\label{thm:strong}
Let $K>0$ and suppose $n \geqslant C\max\{K^8,1\}$ for some sufficiently large constant $C>0$.
Let $P$ be a set of $n$ points in real projective $3$-space with no three collinear. If $P$ spans at most $Kn^2$ ordinary planes,
then up to a projective transformation, $P$ differs in at most $O(K)$ points from a configuration of one of the following types:
\begin{enumerate}[label=\rom]
\item A subset of a plane;
\item A prism or an antiprism;
\item A coset of a subgroup of an elliptic or acnodal space quartic curve.
\end{enumerate}
\end{theorem}
Note that it is simple to show that conversely, any of the three types of sets described in Theorem~\ref{thm:strong} span $O(Kn^2)$ ordinary planes.
Theorem~\ref{thm:strong} forms the basis for proving its higher-dimensional analogue, which we do in a subsequent paper \cite{LS}.
By stereographic projection we immediately obtain the following strengthening of~\cite{LMMSSZ17}*{Theorem~1.5}.
An \emph{ordinary circle} of a finite set of points in the plane is a circle passing through exactly three points of the set.
For definitions of the group on an ellipse, of a circular cubic curve, and of a double polygon, see \cite{LMMSSZ17}.
In \cite{LMMSSZ17}*{Theorem~1.5}, we need $n \geqslant \exp\exp(CK^C)$; here we only assume $n \geqslant CK^8$.
\begin{corollary}
Let $K>0$ and suppose $n \geqslant C\max\{K^8,1\}$ for some sufficiently large constant $C>0$.
Let $P$ be a set of $n$ points in the Euclidean plane.
If $P$ spans at most $Kn^2$ ordinary circles,
then up to inversions and similarities of the plane, $P$ differs in at most $O(K)$ points from a configuration of one of the following types:
\begin{enumerate}[label=\rom]
\item A subset of a line;
\item A coset of a subgroup of an ellipse;
\item A coset of a subgroup of a smooth circular cubic curve;
\item A double polygon that is `aligned' or `offset'.
\end{enumerate}
\end{corollary}
Theorem~\ref{thm:strong} also gives us the means to determine the maximum number of \emph{$4$-point planes}, which are planes passing through exactly four points of a given point set.
The following corollary is an analogue of the orchard problem for circles \cite{LMMSSZ17}*{Theorem~1.3}.
It also immediately implies \cite{LMMSSZ17}*{Theorem~1.3} via stereographic projection.
We omit the proof, which goes along the same lines as the proof of \cite{LMMSSZ17}*{Theorem~1.3}.
(Note that the exact number in the case $n\equiv 0\pmod{8}$ was calculated incorrectly, but corrected in the arXiv version of \cite{LMMSSZ17}.)
\begin{corollary}
\mbox{}
\begin{enumerate}[label=\rom]
\item If $n$ is sufficiently large, the maximum number of $4$-point planes determined by a set of $n$ points in real projective $3$-space with no three collinear is equal to
\begin{equation*}
\begin{cases}
\frac{1}{24}n^3 - \frac14 n^2 + \frac{5}{6}n & \text{if } n \equiv 0 \pmod{8},\\
\frac{1}{24}n^3 - \frac14 n^2 + \frac{11}{24}n - \frac{1}{4} & \text{if } n \equiv 1, 3, 5, 7 \pmod{8},\\
\frac{1}{24}n^3 - \frac14 n^2 + \frac{7}{12}n - \frac12 & \text{if } n \equiv 2, 6 \pmod{8},\\
\frac{1}{24}n^3 - \frac14 n^2 + \frac56n - 1 & \text{if } n \equiv 4 \pmod{8}.
\end{cases}
\end{equation*}
\item Let $C$ be a sufficiently large constant.
If a set $P$ of $n$ points in real projective $3$-space with no three collinear determines more than $n^3/24 - 7n^2/24 + Cn$ $4$-point planes, then up to a projective transformation, $P$ lies on an elliptic or acnodal space quartic curve.
\end{enumerate}
\end{corollary}
The second aim of this paper is to consider the number of \emph{coplanar quadruples} (four distinct coplanar points) of an $n$-point set on quartic curves in complex $3$-space.
Raz, Sharir, and De Zeeuw~\cite{RSZ16} showed that such a set spans $O(n^{8/3})$ coplanar quadruples unless the curve contains a planar or a quartic component.
They left it as an open problem whether there always exist configurations on rational space quartic curves spanning $\Theta(n^3)$ coplanar quadruples.
As mentioned in \cite{RSZ16}, certain space quartic curves (elliptic normal curves) contain sets of $n$ points with $\Theta(n^3)$ coplanar quadruples.
The properties of space quartic curves that we need to prove Theorem~\ref{thm:strong} also enable us to solve this problem.
We prove the following theorem in Section~\ref{sec:quad}.
Any space quartic curve is contained in a quadric surface (see Section~\ref{sec:quartics}).
If the quadric is unique, the curve is called a \emph{space quartic of the second species}, otherwise it is a \emph{space quartic of the first species}.
Elliptic normal curves are always of the first species, while space quartics of the second species are always rational (see Section~\ref{sec:quartics}).
\begin{theorem}\label{thm:quad}
Let $\delta$ be a rational space quartic curve in $\mathbb{C}\mathbb{P}^3$.
If $\delta$ is of the first species, then there exist $n$ points on $\delta$ that span $\Theta(n^3)$ coplanar quadruples.
If $\delta$ is of the second species, then any $n$ points on $\delta$ span $O(n^{8/3})$ coplanar quadruples.
\end{theorem}
\subsection*{A note on our methods}
Although all our results can be formulated for affine space, it is both more natural and more convenient to work in projective space, and this is what we will do.
Also, although we will mostly work with real varieties (mostly curves), we will always consider them to be the real points of a variety in complex projective space.
In classical algebraic geometry, many results are formulated for smooth varieties and for generic points.
For example, the projection of a curve of degree $d$ in complex projective $3$-space from a generic point not on the curve onto a plane is again a curve of degree $d$.
Since we will project from given points not of our choosing, we cannot always assume that a projection point is generic, and will have to use more subtle results on projections (see Section~\ref{ssec:projections}).
Also, since we are working with an extremal problem, there is no guarantee that the curves on which we will find the points to lie are smooth; on the contrary, we should not be surprised that singularities occur in extremal objects.
Although it turns out that the curves that we consider will in the generic case be smooth quartics, quartic curves with singularities will also appear.
Thus we will have to use detailed classical results on rational space quartics.
\subsection*{Notation}
Let $\mathbb{F}$ denote the field of real or complex numbers, and let $\mathbb{F}\mathbb{P}^m$ denote the $m$-dimensional projective space over $\mathbb{F}$.
We only consider dimensions $m=2, 3$.
We denote the homogeneous coordinates of a point in $3$-dimensional projective space by a $4$-dimensional vector $[x_0, x_1, x_2, x_3]$.
We denote the algebraic surface where a homogeneous polynomial $f\in\mathbb{F}[x_0, x_1, x_2, x_3]$ vanish by $Z_{\mathbb{F}}(f)$.
More generally, we consider a \emph{variety} to be any intersection of algebraic surfaces.
We say that a variety is pure-dimensional if each of its irreducible components has the same dimension.
We consider a \emph{curve} of degree $d$ in $\mathbb{C}\mathbb{P}^3$ to be a variety $\delta$ of pure dimension $1$ such that a generic hyperplane of $\mathbb{C}\mathbb{P}^3$ intersects $\delta$ in $d$ distinct points.
We denote the Zariski closure of a set $S \subseteq \mathbb{C}\mathbb{P}^3$ by $\overline{S}$.
\section{Tools from classical algebraic geometry}\label{sec:tools}
\subsection{B\'ezout's theorem}\label{ssec:bezout}
We will use
a form of B\'ezout's theorem \cite{H92}*{Theorem~18.4}.
Two pure-dimensional varieties $X$ and $Y$ in $\mathbb{C}\mathbb{P}^m$
are said to \emph{intersect properly} if $\dim(X\cap Y) = \dim(X) + \dim(Y) - m$.
\begin{theorem}[B\'ezout's theorem]\label{thm:Bezout}
Let $X,Y\subset\mathbb{C}\mathbb{P}^m$
be pure-dimensional varieties that intersect properly.
If the intersection $X\cap Y$ has irreducible components $\delta_1, \delta_2, \dotsc, \delta_k$, then their total degree $\sum_{i=1}^k\deg(\delta_i)$ is at most $\deg(X)\deg(Y)$.
\end{theorem}
By counting the multiplicity of each irreducible component $\delta_i$, it is possible to formulate a version of this theorem where equality holds \cite{H92}*{Theorem~18.4}, but we do not need this strengthening.
We will instead use the following corollaries of Theorem~\ref{thm:Bezout}.
\begin{corollary}\label{cor:Bezout1}
Let $X$ and $Y$ be two distinct surfaces in $\mathbb{C}\mathbb{P}^3$.
Then the irreducible components $\delta_1, \delta_2, \dotsc, \delta_k$ of $X\cap Y$ are all curves of total degree $\sum_{i=1}^k\deg(\delta_i)$ at most $\deg(X)\deg(Y)$.
\end{corollary}
\begin{proof}
Since the surfaces are distinct, their intersection will have lower dimension.
It then follows by the projective dimension theorem \cite{H77}*{Theorem~7.2} that $\dim(X\cap Y)=1$, so $X$ and $Y$ intersect properly.
The corollary now follows from Theorem~\ref{thm:Bezout}.
\end{proof}
The following corollary is an immediate consequence of the previous one.
\begin{corollary}\label{cor:Bezout2}
Let $\delta$ be a curve of degree $d_1$
and $X$ a surface of degree $d_2$
in $\mathbb{C}\mathbb{P}^3$ not containing any irreducible component of $\delta$.
Then $\delta \cap X$ has at most $d_1d_2$ points.
\end{corollary}
We present a proof of the following well-known result after we have discussed projections in Section~\ref{ssec:projections}.
\begin{corollary}\label{cor:Bezout3}
Two distinct curves $\delta_1$ and $\delta_2$ in $\mathbb{C}\mathbb{P}^3$ with no common irreducible component and of degrees $d_1$ and $d_2$, respectively, intersect in at most $d_1d_2$ points.
\end{corollary}
\subsection{Projections of curves}\label{ssec:projections}
We consider the projection from a point $p \in \mathbb{F}\mathbb{P}^3$ to be the mapping $\pi_p$ from $\mathbb{F}\mathbb{P}^3 \setminus \{p\}$ onto a plane $\Pi$ of $\mathbb{F}\mathbb{P}^3$ not containing $p$, where a point $q$ is mapped to the intersection of the line $pq$ with~$\Pi$ \cite{H92}*{Example~3.4}.
The exact choice of $\Pi$ does not matter, although certain choices might be more convenient.
Let $P$ be a finite point set in $\mathbb{R}\mathbb{P}^3$ with no three points collinear.
If we project $P \setminus \{p\}$ from a point $p \in P$, all ordinary planes of $P$ through $p$ map to ordinary lines of $\pi_p(P \setminus \{p\})$ in $\mathbb{R}\mathbb{P}^2$.
Green and Tao showed that sets with few ordinary lines lie on (possibly reducible) planar cubic curves \cite{GT13}*{Theorem 1.5}.
It turns out that we need to understand projections of curves so that we can apply their result to the structure of the set $P$.
Let $\delta$ be an irreducible non-planar curve of degree $d$ and $p$ a point in $\mathbb{C}\mathbb{P}^3$.
We say that $\pi_p$ is \emph{generically one-to-one on $\delta$} if there is a finite subset $S$ of $\delta$ such that $\pi_p$ restricted to $\delta\setminus S$ is one-to-one.
(This is equivalent to the birationality of $\pi_p$ restricted to $\delta\setminus\{p\}$ \cite{H92}*{p.~77}.)
It is a well-known classical fact that, if $\pi_p$ generically one-to-one, the degree of the curve $\overline{\pi_p(\delta \setminus \{p\})}$ is $d-1$ if $p$ lies on $\delta$, and is $d$ if $p$ does not lie on~$\delta$~\cite{H92}*{Example 18.16}, \cite{Kollar}*{Section 1.15}.
We can now prove Corollary~\ref{cor:Bezout3}, which is essentially a consequence of B\'ezout's theorem (Theorem~\ref{thm:Bezout}).
\begin{proof}[Proof of Corollary~\ref{cor:Bezout3}]
Since $\delta_1$ and $\delta_2$ are distinct with no common irreducible component, their intersection is finite.
Projecting from a generic point $p \in \mathbb{C}\mathbb{P}^3\setminus{(\delta_1 \cup \delta_2)}$, we thus have two planar curves $\delta_1' := \pi_p(\delta_1)$ and $\delta_2' := \pi_p(\delta_2)$ of degrees $d_1$ and $d_2$, respectively, which intersect properly.
The corollary then follows from B\'ezout's theorem (Theorem~\ref{thm:Bezout}).
\end{proof}
In the projections that we will make, we will not have complete freedom in choosing a projection point, and therefore we cannot guarantee that $\pi_p$ is generically one-to-one on $\delta$.
For this reason, we will need more sophisticated results on the projection of curves from a point.
We start with the following more elementary proposition, which is a restatement of \cite{B17}*{Lemma~6.6}.
\begin{prop}\label{prop:cones}
Let $\sigma_1$ and $\sigma_2$ be two irreducible conics given by the intersection of two distinct planes and a quadric surface in $\mathbb{C}\mathbb{P}^3$.
Then there are at most two quadric cones containing both $\sigma_1$ and $\sigma_2$.
\end{prop}
We define a \emph{trisecant} of an irreducible non-planar curve $\delta$ in $\mathbb{C}\mathbb{P}^3$ to be a line that intersects $\delta$ in at least three distinct points, or that can be approximated in the Zariski topology by such lines.
More precisely, note that lines in $\mathbb{C}\mathbb{P}^3$ can be parametrised using Pl\"ucker coordinates by points on the Klein quadric in $\mathbb{C}\mathbb{P}^5$ \cite{H92}*{Example~6.3}.
A \emph{trisecant} of $\delta$ is a line that corresponds to a point in the Zariski closure of the set of points on the Klein quadric that correspond to lines that intersect $\delta$ in at least three distinct points \cite{H92}*{Example~8.9}.
The classical trisecant lemma states that the number of points on an irreducible non-planar curve in $\mathbb{C}\mathbb{P}^3$ that lie on infinitely many trisecants is finite \cite{ACGH}*{pp.~109--111}, \cite{Severi}*{p.~85, footnote}.
We first state a generalisation (Lemma~\ref{lem:trisecant}) of the trisecant lemma to curves that are not necessarily irreducible, which is a simple consequence of \cite{KKT08}*{Theorem 2}.
We then prove three quantitative versions of the trisecant lemma (Lemmas \ref{lem:projection2}, \ref{lem:projection3}, and \ref{lem:projection1}).
\begin{lemma}[Trisecant lemma]\label{lem:trisecant}
Let $\delta$ be a non-planar curve in $\mathbb{C}\mathbb{P}^3$.
Then the number of points on $\delta$ that lie on infinitely many trisecants of $\delta$ is finite.
\end{lemma}
Note that a point $p$ on a non-planar curve $\delta$ lies on infinitely many trisecants of $\delta$ if and only if the projection $\pi_p$ is not generically one-to-one on $\delta$.
Thus, according to the trisecant lemma there are finitely many such projection points on $\delta$.
The following special case of a theorem of Segre's \cite{S36} shows that there are also finitely many such projection points not on $\delta$.
\begin{prop}[Segre \cite{S36}]\label{prop:trisecant}
Let $\delta$ be an irreducible non-planar curve in $\mathbb{C}\mathbb{P}^3$.
Then the set of points
\[ X = \setbuilder{x\in\mathbb{C}\mathbb{P}^3\setminus \delta}{\text{$\pi_x$ is not generically one-to-one on $\delta$}}\]
is finite.
\end{prop}
For a curve $\delta$ and a point $p$ in $\mathbb{C}\mathbb{P}^3$, denote the cone over $\delta$ with vertex $p$ by $C_p(\delta)$, that is,
\[ C_p(\delta) := \overline{\pi_p^{-1}(\pi_p(\delta \setminus \{p\}))}. \]
Note that if $p \notin \delta$, then \cite{H92}*{Example~3.10}
\[ C_p(\delta) = \setbuilder{q\in\mathbb{C}\mathbb{P}^3\setminus\{p\}}{\text{the line $pq$ intersects $\delta$}}\cup\{p\},\]
and if $p\in\delta$, then
\[ C_p(\delta) = \setbuilder{q\in\mathbb{C}\mathbb{P}^3\setminus\{p\}}{\text{the line $pq$ intersects $\delta$ with multiplicity at least $2$}}\cup\{p\}.\]
\begin{lemma}\label{lem:projection2}
Let $\delta_1$ and $\delta_2$ be two irreducible curves in $\mathbb{C}\mathbb{P}^3$ of degree $d_1$ and $d_2$, respectively.
Suppose $\delta_1$ is not a line, and $\delta_1 \cup \delta_2$ is non-planar.
Then there are at most $O(d_1d_2)$ points $x$ on $\delta_1$ such that $\overline{\pi_x(\delta_1\setminus\{x\})}$ and $\overline{\pi_x(\delta_2\setminus\{x\})}$ coincide, or equivalently, for which $\delta_2\subset C_x(\delta_1)$.
\end{lemma}
\begin{proof}
Let $X = \setbuilder{x\in\delta_1}{\overline{\pi_x(\delta_1\setminus\{x\})} = \overline{\pi_x(\delta_2\setminus\{x\})}}$, and let
\[S = \delta_1\cap\bigcap_{p\in\delta_1\setminus\delta_2} C_p(\delta_2).\]
We claim that $X\setminus\delta_2 = S\setminus\delta_2$.
First, let $x\in X\setminus\delta_2$ and $p\in\delta_1\setminus\delta_2$.
If $x=p$, then clearly $x\in C_p(\delta_2)$.
Otherwise, $\pi_x(p)\in\pi_x(\delta_1\setminus\{x\})$.
Since $x\in X$, $\overline{\pi_x(\delta_1\setminus\{x\})} = \overline{\pi_x(\delta_2\setminus\{x\})}$, and since $x\notin\delta_2$, $\overline{\pi_x(\delta_2\setminus\{x\})} = \pi_x(\delta_2\setminus\{x\})$.
Therefore, $\pi_x(p)\in\pi_x(\delta_2\setminus\{x\})$, and it follows that the line $px$ intersects $\delta_2$, hence $x\in C_p(\delta_2)$.
Since $X\subseteq\delta_1$, we conclude that $x\in S\setminus\delta_2$.
Conversely, let $x\in S\setminus\delta_2$.
Then $x\in\delta_1$, and for all $p\in\delta_1\setminus\delta_2$, we have $x\in C_p(\delta_2)$.
Thus, if $x\neq p$, then the line $px$ intersects $\delta_2$.
Therefore, $\pi_x(\delta_1\setminus\{x\})\subseteq\pi_x(\delta_2\setminus\{x\})$.
Since $\delta_2$ is irreducible, the curve $\overline{\pi_x(\delta_2\setminus\{x\})}$ is irreducible.
Since $\delta_1$ is not a line, $\overline{\pi_x(\delta_1\setminus\{x\})}$ does not degenerate to a point.
Therefore, $\overline{\pi_x(\delta_1\setminus\{x\})} = \overline{\pi_x(\delta_2\setminus\{x\})}$, and $x\in X$.
Next, note that each $x\in S$ lies on infinitely many trisecants of the curve $\delta_1\cup\delta_2$.
Since $\delta_1 \cup \delta_2$ is non-planar, $S$ is finite by the trisecant lemma (Lemma~\ref{lem:trisecant}).
Therefore, $\delta_1\not\subset C_p(\delta_2)$ for some $p\in\delta_1\setminus\delta_2$.
By B\'ezout's theorem (Corollary~\ref{cor:Bezout2}), $|S|\leqslant|\delta_1\cap C_p(\delta_2)| \leqslant d_1\deg(C_p(\delta_2))\leqslant d_1d_2$.
Again by B\'ezout's theorem (Corollary~\ref{cor:Bezout3}), $|\delta_1\cap\delta_2|\leqslant d_1d_2$.
It then follows that $|X| \leqslant |X\setminus\delta_2| + |\delta_1\cap\delta_2| = |S\setminus\delta_2| + |\delta_1\cap\delta_2| \leqslant 2d_1d_2$.
\end{proof}
The following result is the $1$-dimensional case of a result from Ballico \cite{Ballico2003}; see also \cite{Ballico2004}*{Remark~1}.
For convenience we provide the proof of this special case.
\begin{lemma}\label{lem:projection3}
Let $\delta$ be an irreducible
non-planar curve in $\mathbb{C}\mathbb{P}^3$ of degree $d$.
Then there are at most $O(d^3)$ points $x\in\mathbb{C}\mathbb{P}^3\setminus\delta$ such that $\pi_x$ restricted to $\delta$ is not generically one-to-one.
\end{lemma}
\begin{proof}
By Proposition~\ref{prop:trisecant}, the set
\[ X = \setbuilder{x\in\mathbb{C}\mathbb{P}^3\setminus \delta}{\text{$\pi_x$ is not generically one-to-one on $\delta$}}\]
is finite, and we want to show that $|X|=O(d^3)$.
Let $x\in X$.
Since $\delta$ has finitely many singularities and there are only finitely many lines through $x$ that are tangent to $\delta$, we have that for all but finitely many points $p\in\delta$, the line $px$ intersects $\delta$ in a third point, that is, $x\in C_p(\delta)$ for all $p\in\delta\setminus E_x$, for some finite subset $E_x$ of $\delta$.
Let $\delta'=\delta\setminus\bigcup_{x\in X}E_x$ and $S=\bigcap_{p\in\delta'} C_p(\delta)$.
Then clearly $X\subseteq S\setminus\delta$.
Conversely, if $x\in S\setminus\delta$, then for any $p\in\delta'$, the line $px$ intersects $\delta$ with multiplicity at least $2$.
Since only finitely many lines through $x$ can be tangent to $\delta$, it follows that for all but finitely many points $p\in\delta$, the line $px$ intersects $\delta$ in a third point, hence $x\in X$.
This shows that $X=S\setminus\delta$.
Fix distinct $p,p'\in\delta'$.
Then $X\subseteq C_p(\delta)\cap C_{p'}(\delta)$.
This intersection consists of $\delta$, the line $pp'$, and some further irreducible curves $\delta_1, \dots, \delta_k$ of total degree at most $d^2-d-1$, by B\'ezout's theorem (Corollary~\ref{cor:Bezout1}).
If some $\delta_i\subset C_p(\delta)$ for all $p\in\delta'$, then $\delta_i\subseteq S$, and since $\delta_i\cap\delta$ is finite by B\'ezout's theorem (Corollary~\ref{cor:Bezout3}), we obtain that $X$ is infinite, a contradiction.
Therefore, for each $\delta_i$ there is a point $p_i\in\delta'$ such that $\delta_i\not\subset C_{p_i}(\delta)$.
By B\'ezout's theorem (Corollary~\ref{cor:Bezout2}), $|X\cap\delta_i| \leqslant |C_{p_i}(\delta)\cap\delta_i| \leqslant d\deg(\delta_i)$.
It follows that $|X\setminus pp'| \leqslant \sum_{i=1}^k|X\cap\delta_i| \leqslant \sum_{i=1}^kd\deg(\delta_i) = O(d^3)$.
Now find a third point $p''\in\delta'$ such that $p,p',p''$ are not collinear.
As before, $|X\setminus pp''| = O(d^3)$.
Since $pp'\cap pp''$ is a singleton, it follows that $|X|=O(d^3)$.
\end{proof}
If an irreducible non-planar curve $\delta$ of degree $d$ is smooth, then by a well-known result going back to Cayley (see \cites{Bertin, LeBarz82, GP82}), the trisecant variety of $\delta$ (the Zariski closure in $\mathbb{C}\mathbb{P}^3$ of the union of all trisecants of $\delta$) has degree $O(d^3)$.
For $p \in \delta$, if $\pi_p$ restricted to $\delta \setminus \{p\}$ is not generically one-to-one, then $C_p(\delta)$ is a component of the trisecant variety and has degree at least $2$.
It follows that there can be at most $O(d^3)$ points $p\in\delta$ such that $\pi_p$ is not generically one-to-one on $\delta$.
However, if $\delta$ is not smooth, we are not aware of any estimates of the degree of the trisecant variety, and we thus include the proof of the weaker bound $O(d^4)$ below in Lemma~\ref{lem:projection1}, based on an argument of Furukawa \cite{F11}.
This result answers the $1$-dimensional case of a question of Ballico \cite{Ballico2004}*{Question~1}.
We say that a point $z\in\mathbb{C}\mathbb{P}^3$ is a \emph{vertex} of a surface $Z$ in $\mathbb{C}\mathbb{P}^3$ if the projection $\overline{\pi_z(Z \setminus \{z\})}$ is a planar curve with $Z$ equal to the cone $C_z(\overline{\pi_z(Z\setminus\{z\})})$.
In \cite{F11}*{Lemma~2.3}, the vertices of a surface is characterised in terms of partial derivatives.
For any $4$-tuple of non-negative integers $\bm{i}=(i_0,i_1,i_2,i_3)$, we define $|\bm{i}| = i_0+i_1+i_2+i_3$.
For any homogeneous polynomial $f\in\mathbb{C}[x_0, x_1, x_2, x_3]$ of degree $e$, we define
\[ D_{\bm{i}} f =
\frac{\partial^{i_0}}{\partial x_0^{i_0}}
\frac{\partial^{i_1}}{\partial x_1^{i_1}}
\frac{\partial^{i_2}}{\partial x_2^{i_2}}
\frac{\partial^{i_3}}{\partial x_3^{i_3}} f. \]
Let $D f$ be the column vector $[D_{\bm{i}} f]_{\bm{i}}$, where $\bm{i}$ varies over all $4$-tuples such that $|\bm{i}| = e-1$.
Then $D f$ is an $\binom{e+2}{3}$-dimensional vector of linear forms in $x_0,x_1,x_2,x_3$.
According to \cite{F11}*{Lemma~2.3}, $z$ is a vertex of the surface $Z$ defined by the homogeneous polynomial $f$ of degree $e$ if and only if $(D f)(z)$ is the zero vector.
\begin{lemma}\label{lem:projection1}
Let $\delta$ be an irreducible non-planar curve of degree $d$ in $\mathbb{C}\mathbb{P}^3$.
Then there are at most $O(d^4)$ points $x$ on $\delta$ such that $\pi_x$ restricted to $\delta \setminus \{x\}$ is not generically one-to-one.
\end{lemma}
\begin{proof}
Let $X$ be the set of points $x$ on $\delta$ such that $\pi_x$ restricted to $\delta \setminus \{x\}$ is not generically one-to-one.
Let $V \subseteq \mathbb{C}[x_0, x_1, x_2, x_3]$ be the vector space of homogeneous polynomials of degree $d-2$ that vanish on $\delta$, and let $h_1,\dots,h_r$ be a basis of $V$.
Consider the matrix $A = [Dh_1, \dots, Dh_r]$.
Suppose first that $x \in X$.
Then $\deg(\overline{\pi_x(\delta\setminus\{x\}}) \leqslant d-2$, and there exists a cone of degree $\leqslant d-2$ with vertex $x$ containing $\delta$.
It follows that there is a polynomial $f\in V$ such that $Z_\mathbb{C}(f)$ contains $\delta$.
By \cite{F11}*{Lemma~2.3}, the rank of $A(x) = [Dh_1(x), \dots, Dh_r(x)]$ is less than $r$, so each $r \times r$ minor of $A$ vanishes at $x$.
Note that each such minor defines a surface of degree at most $r$.
Conversely, if $x$ lies on all of the surfaces defined by the $r\times r$ minors of $A$, then $A(x)$ has rank less than $r$.
There then exists $f\in V$ such that $(Df)(x)$ is the zero vector.
By \cite{F11}*{Lemma~2.3}, $x$ is a vertex of $Z_\mathbb{C}(f)$, which is a surface of degree at most $d-2$ and contains $\pi_x(\delta\setminus\{x\})$, so either $x$ is a singular point of $\delta$ or $x \in X$.
Since $\delta$ has at most $O(d^2$) singular points, it will follow that $X$ has at most $O(d^4)$ points if we can show that there are at most $O(d^4)$ points in
\[ \delta \cap \setbuilder{x\in\mathbb{C}\mathbb{P}^3}{\rank(A(x)) < r}.\]
Now $X$ is finite by the trisecant lemma (Lemma~\ref{lem:trisecant}), so $\delta$ is not a subset of all of the surfaces defined by the $r \times r$ minors of $A(x)$.
Fix one such surface $Z$ not containing $\delta$.
It has degree at most $r$, so by B\'ezout's theorem (Corollary~\ref{cor:Bezout2}), $\delta\cap Z$ has at most $dr$ points.
Since $r=O(d^3)$, the lemma follows.
\end{proof}
We have no reason to believe that the estimate $O(d^4)$ in the above lemma is best possible.
\section{Space quartics}\label{sec:quartics}
\subsection{Classification of space quartics}\label{ssec:classification}
By a \emph{space quartic}, we mean an irreducible non-planar curve of degree $4$ in $\mathbb{C}\mathbb{P}^3$.
Such a curve meets a generic plane in four points, and a generic quadric surface in eight points.
Since the dimension of the vector space of degree $2$ homogeneous polynomials in three variables is $10$, we can fit a quadric $Q$ through any nine points on a space quartic.
It follows by B\'ezout's theorem (Corollary~\ref{cor:Bezout2}) that the space quartic is contained in $Q$.
A space quartic $\delta$ is said to be of the \emph{first species} if more than one quadric contains $\delta$, in which case all the quadrics containing $\delta$ form a pencil (see below).
Otherwise it is said to be of the \emph{second species}, where $\delta$ is contained in a unique quadric.
The facts collected here are well known in the sense that they were discovered in the 19th century, see \cites{S1851, F1895, B1871, W1871, R1900, T36}, but it is not easy to find recent references, and so we include some of the proofs.
The homogeneous quadratic polynomial defining a quadric $Q$ in $\mathbb{F}\mathbb{P}^3$ can be written as $q(x) = x^TA_Qx$, where $A_Q$ is a symmetric $4\times 4$ matrix with entries in $\mathbb{F}$.
A pencil of quadrics is a collection of quadrics defined by linear combinations of two linearly independent quadratic polynomials $p$ and $q$:
$\setbuilder{Z_\mathbb{F}(\lambda p+\mu q)}{\lambda,\mu\in\mathbb{F}}$.
If $\delta$ is a smooth space quartic of the first species, then it is well known to be an elliptic curve \cite{S39}*{Sections~14.713 and 14.714}.
Otherwise $\delta$ is rational~\cite{S39}*{Sections~14.717 and~14.723}, \cite{T36}*{Chapter II} and can be parametrised as $[f_1(t), f_2(t), f_3(t), f_4(t)]$ for polynomials $f_1, f_2, f_3, f_4 \in \mathbb{F}[t]$ with no common factor and maximum degree $4$, with $t \in \mathbb{F}$.
Note that while $\delta$ is a curve in $\mathbb{F}\mathbb{P}^3$ and should be parametrised by the projective line $\mathbb{F}\mathbb{P}^1$, we consider $\delta$ to be parametrised by $t \in \mathbb{F}$ for simplicity, where $t$ corresponds to $[t,1]\in\mathbb{F}\mathbb{P}^1$.
We omit the simple proof of the following simplified parametrisation of an arbitrary rational space quartic.
\begin{prop}\label{prop:param}
Let $\delta$ be a rational space quartic in $\mathbb{F}\mathbb{P}^3$ parametrised by $t \in \mathbb{F}$.
After a linear fractional transformation on $\mathbb{F}$ and a projective transformation on $\mathbb{F}\mathbb{P}^3$ if necessary, we can parametrise $\delta$ as
\begin{equation}\label{eqn:param}
[t^4 - p, t^3 + q, t^2 - r, t + s],
\end{equation}
for some $p,q,r,s \in \mathbb{F}$ such that $r \ne s^2$, $q \ne s^3$, or $p \ne s^4$.
\end{prop}
With this parametrisation, we get the following condition on when four points, counting multiplicity, on a space quartic are coplanar.
\begin{lemma}\label{lem:coplanar}
Let $\delta$ be a space quartic in $\mathbb{F}\mathbb{P}^3$ given by the parametrisation~\eqref{eqn:param}.
A plane intersects $\delta$ in four points parametrised by $t_1, t_2, t_3, t_4$, counting multiplicity, if and only if
\begin{equation}\label{eqn:coplanar}
F(t_1,t_2,t_3,t_4) := t_1t_2t_3t_4 + s \sum_{i<j<k} t_it_jt_k + r \sum_{i < j} t_it_j + q \sum_i t_i + p = 0.
\end{equation}
In particular, if $t_1,t_2,t_3$ are distinct, then $F(t_1,t_1,t_2,t_3)=0$ if and only if the plane through $t_1$, $t_2$, $t_3$ intersects $\delta$ only in $t_1, t_2, t_3$ and contains the tangent of $\delta$ at $t_1$.
\end{lemma}
\begin{proof}
The plane in $\mathbb{F}\mathbb{P}^3$ with equation $a_1 x + a_2 y + a_3 z + a_4 w = 0$ intersects $\delta$ in $t_1, t_2, t_3, t_4$ if and only if we have
\begin{equation*}
a_1 (t^4 - p) + a_2 (t^3 + q) + a_3 (t^2 - r) + a_4 (t + s) \equiv a_1(t - t_1)(t - t_2)(t - t_3)(t - t_4).
\end{equation*}
Equating coefficients, the two polynomials in $t$ are then equal identically if and only if $F = 0$, as desired.
\end{proof}
Let $f(t) := F(t,t,t,t)$ be the so-called restitution of the multilinear form $F$ \cite{D03}*{Section~1.2}, and let $g$ be the binary quartic form obtained from homogenising $f$:
\begin{equation}\label{eqn:quartic}
g(\lambda, \mu) = \lambda^4 + 4s\lambda^3 \mu + 6r\lambda^2 \mu^2 + 4q\lambda \mu^3 + p\mu^4.
\end{equation}
We call $g$ the \emph{fundamental quartic} of $\delta$. The \emph{catalecticant} of $g$ is defined to be
\begin{equation*}
\cat(g) :=
\begin{vmatrix}
1 & s & r\\
s & r & q\\
r & q & p\\
\end{vmatrix}
= pr - q^2 - ps^2 + 2qrs - r^3.
\end{equation*}
This is an invariant of $g$, in the sense that it remains unchanged under a linear change of variables
\begin{equation*}
\begin{pmatrix}
\lambda' \\ \mu'
\end{pmatrix}
=
\begin{pmatrix}
a & b \\ c & d
\end{pmatrix}
\begin{pmatrix}
\lambda \\ \mu
\end{pmatrix}
,
\end{equation*}
where $ad - bc = 1$ \cite{D03}*{Example~1.4}.
The catalecticant of a binary quartic form was discovered by Boole \cite{W08}, and generalised to binary forms of even degree by Sylvester \cite{S1851} (who coined the term).
Sylvester \cite{S1851} also showed that a generic binary form of degree $d$ is the sum of two $d$-th powers of linear forms if and only if a certain matrix does not have full rank.
We need a version of this statement that is valid for all binary forms, not only generic ones, as can be found in Kanev \cite{K99}.
The following version, formulated only for quartic forms, will be used in Section~\ref{sec:quad}.
\begin{theorem}[Sylvester]\label{thm:sylvester}
A non-zero binary quartic form
\[g(\lambda, \mu) =
\lambda^4 + 4c_1\lambda^3 \mu + 6c_2\lambda^2 \mu^2 + 4c_3\lambda \mu^3 + c_4\mu^4 \in \mathbb{C}[\lambda, \mu]\]
can be expressed as one of
\[ (a \lambda + b \mu)^4, (a_1 \lambda + b_1 \mu)^4 + (a_2 \lambda + b_2 \mu)^4, (a_1 \lambda + b_1 \mu)(a_2 \lambda + b_2 \mu)^3,\]
for some $a, b, a_1, b_1, a_2, b_2 \in \mathbb{C}$, where $a_1b_2 \ne b_1a_2$, if and only if $\cat(g)$ vanishes.
\end{theorem}
A more immediate application of the catalecticant is the following condition on when a rational space quartic is of the first species \cite{F1895}*{Section~6}.
\begin{lemma}\label{lem:first}
Let $\delta$ be a space quartic in $\mathbb{F}\mathbb{P}^3$ given by the parametrisation~\eqref{eqn:param}.
Then $\delta$ is of the first species if and only if the catalecticant of its fundamental quartic vanishes.
\end{lemma}
\begin{proof}
We prove this by considering the equations of quadrics $Q$ that contain $\delta$.
If $Q$ contains $\delta$, substituting the four polynomials $t^4-p, t^3+q, t^2-r, t+s$ for the homogeneous coordinates of $x$ into the equation $x^T A_Q x = 0$ gives a degree $8$ polynomial in $t$ that has to be identically zero.
This gives nine equations in ten variables (the entries of the symmetric $4\times 4$ matrix $A_Q=(a_{i,j})$).
The first few equations, corresponding to the coefficients of $t^8, t^7, t^6, t^5$, are $a_{11}=0, a_{12}=0, 2a_{13}+a_{22}=0, a_{14}+a_{23}=0$.
So in fact we only have five equations in six variables:
\begin{equation*}
\begin{pmatrix}
-2r & 2s & 2 & 1 & 0 &0 \\
-2q & r & s & 0 & 1 & 0 \\
-2p & -2q & 0 & -2r & 2s & 1 \\
0 & -p & q & 0 & -r & s \\
2(pr-q^2) & -2(ps-qr) & 2qs & r^2 & -2rs & s^2
\end{pmatrix}
\begin{pmatrix}
a_{13}\\a_{14}\\a_{24}\\a_{33}\\a_{34}\\a_{44}
\end{pmatrix}
= 0.
\end{equation*}
There is always a non-trivial solution to this system, but we want to show that there are always at least two linearly independent solutions if and only if $\cat(g)=pr - q^2 - ps^2 + 2qrs - r^3 = 0$.
The nullity of the matrix is at least $2$ if and only if its rank is at most $4$, which in turn happens if and only if the six $5 \times 5$ minors all vanish.
These six minors are
\begin{align*}
-4\cat(g)&(q^2 - pr), & 2\cat(g)&(qr - ps), & -4\cat(g)&(qs - p),\\
2\cat(g)&(r^2 - 2qs + p), & 2\cat(g)&(rs - q), & -2\cat(g)&(s^2-r),
\end{align*}
and it is impossible for all of the last factors to be equal to zero, otherwise we have $r = s^2$, $q = s^3$, and $p = s^4$.
Thus the six minors all vanish if and only if $\cat(g) = 0$, as desired.
\end{proof}
\subsection{Groups on space quartics}\label{ssec:groups}
The extremal configurations in Theorem~\ref{thm:strong} are based on group laws on certain space quartics of the first species.
We include here the reducible quartic consisting of two disjoint conic sections of a quadric, which can be viewed as the intersection of a quadric with the union of two planes (which is a degenerate quadric).
Geometrically, the group laws relate to when four points on the curve are coplanar.
A smooth space quartic of the first species in $\mathbb{R}\mathbb{P}^3$ is an elliptic normal curve as described in \cite{M10}*{Section~4.4.5}.
There is a group law on the points of the quartic such that four points, counting multiplicity, are coplanar if and only if they sum to the identity (see for example~\cite{S09}*{Exercise~3.10}).
Analogous to real elliptic planar cubics, a real elliptic normal curve has one or two real connected components.
If there is one connected component, this group is isomorphic to the circle $\mathbb{R}/\mathbb{Z}$; if there are two connected components, it is isomorphic to $\mathbb{R}/\mathbb{Z} \times \mathbb{Z}_2$~\cite{M10}*{Section~4.4.5}.
A space quartic of the first species that is not smooth is rational with a single singular point~\cite{M10}*{Section~4.4.6}, which can be either a cusp, a crunode, or an acnode.
As discussed in \cite{M10}*{Section~4.4.6}, we can define a group law on the smooth points on such quartics.
The groups obtained are completely analogous to the groups on the smooth points of singular planar cubics.
In the case of a cuspidal space quartic, the group is isomorphic to $(\mathbb{R},+)$; for a space quartic with a crunode, it is isomorphic to the non-zero real numbers $(\mathbb{R}^*,\cdot)\cong\mathbb{R} \times \mathbb{Z}_2$; and an acnodal space quartic has group isomorphic to the circle group $\mathbb{R}/\mathbb{Z}$.
Geometrically, in the cuspidal and acnodal cases, four smooth points, counting multiplicity, are coplanar if and only if they sum to the identity;
in the crunodal case, four smooth points, counting multiplicity, are coplanar if and only if they sum to $(0,0)$ or $(0,1)$, depending on the curve.
Lastly, given two disjoint conic sections of a quadric, we can apply a projective transformation to obtain two circles that lie on a sphere.
Four points are coplanar if and only if they are concyclic on the sphere.
We can then define a group law on the union of the two circles completely analogous to the two concentric circles case in~\cite{LMMSSZ17}*{Section~3.2}.
Two points on each circle, counting multiplicity, are coplanar if and only if they sum to the identity, and in this case the group is isomorphic to $\mathbb{R}/\mathbb{Z} \times \mathbb{Z}_2$.
We also need to know the groups on a space quartic of the first species over the complex numbers, and here the situation is analogous to complex singular planar cubics.
The singular point is either a cusp or a node (a crunode and an acnode being indistinguishable over $\mathbb{C}$).
For a complex cuspidal space quartic, the group is isomorphic to $(\mathbb{C},+)$, and for a complex space quartic with a node, it is isomorphic to the non-zero complex numbers $(\mathbb{C}^*,\cdot)$.
\section{Ordinary planes}\label{sec:proof}
We prove Theorem~\ref{thm:strong} in this section.
First, in Section~\ref{ssec:intermediate}, we prove the weaker Lemma~\ref{lem:intermediate}, using results from Section~\ref{sec:tools} together with a result of Green and Tao's.
This provides an alternative to Ball's tetra-grid.
We then refine Lemma~\ref{lem:intermediate} in Section~\ref{ssec:proof}, replacing the polynomial error terms by linear error terms in Lemma~\ref{lem:intermediate2}.
Finally, using the properties of space quartics from Section~\ref{sec:quartics}, we determine the precise characterisation of the possible configurations of sets with few ordinary planes as described in Theorem~\ref{thm:strong}.
\subsection{Intermediate structure theorem}\label{ssec:intermediate}
We start with the following statement of Green and Tao's intermediate structure theorem for sets with few ordinary lines~\cite{GT13}.
\begin{theorem}[Green--Tao \cite{GT13}]\label{thm:GT}
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^2$, spanning at most $Kn$ ordinary lines, for some $K \geqslant 1$.
Then we have one of the following:
\begin{enumerate}[label=\rom]
\item $P$ is contained in the union of $O(K)$ lines and an additional $O(K^6)$ points;
\item $P$ lies on the union of an irreducible conic $\sigma$ and an additional $O(K^4)$ lines, with $|P \cap \sigma| = \frac{n}{2} \pm O(K^5)$;
\item $P$ is contained in the union of an irreducible cubic and an additional $O(K^5)$ points.
\end{enumerate}
\end{theorem}
With the results in Section~\ref{sec:tools} and Theorem~\ref{thm:GT}, we can now prove the following intermediate structure theorem for ordinary planes.
\begin{lemma}[Intermediate structure theorem]\label{lem:intermediate}
Let $K\geqslant 1$ and suppose $n \geqslant CK^8$ for some sufficiently large constant $C>0$.
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^3$ with no three collinear.
If $P$ spans at most $Kn^2$ ordinary planes,
then we have one of the following:
\begin{enumerate}[label=\rom]
\item $P$ is contained in the union of a plane and an additional $O(K^6)$ points;\label{case:1.1}
\item $P$ is contained in the union of two irreducible conics lying on distinct planes and an additional $O(K^8)$ points, with each conic containing $\frac{n}{2} \pm O(K^8)$ points of $P$;\label{case:1.2}
\item $P$ is contained in the union of a space quartic curve and an additional $O(K^5)$ points.\label{case:1.3}
\end{enumerate}
\end{lemma}
\begin{proof}
Let $P'$ denote the set of all points $p \in P$ such that there are at most $9Kn$ ordinary planes through $p$.
Then $|P'| \geqslant 2n/3$, and for any $p \in P'$, the projection $\pi_p(P \setminus \{p\})$ spans at most $9Kn$ ordinary lines.
Applying
Theorem~\ref{thm:GT} to $\pi_p(P\setminus\{p\})$ for any $p \in P'$, we have one of three cases:
\begin{enumerate}
\item $\pi_p(P\setminus\{p\})$ is contained in the union of $O(K)$ lines and an additional $O(K^6)$ points;
\item $\pi_p(P\setminus\{p\})$ lies on the union of a conic $\sigma$ and an additional $O(K^4)$ lines with $|\pi_p(P\setminus\{p\}) \cap \sigma| = n/2 \pm O(K^5)$;
\item $\pi_p(P\setminus\{p\})$ is contained in the union of an irreducible cubic and an additional $O(K^5)$ points.
\end{enumerate}
This partitions $P'$ into a disjoint union $P_1' \cup P_2' \cup P_3'$, depending on which of the above cases we obtain.
If $|P_1'| \geqslant 3$, let $p_1, p_2, p_3$ be three distinct points in $P_1'$.
Then apart from $O(K^6)$ points, $P$ is contained in the intersection of the union of $O(K)$ planes through $p_1$, the union of $O(K)$ planes through $p_2$, and the union of $O(K)$ planes through $p_3$.
Since $p_1, p_2, p_3$ are not collinear, if $\Pi_i$ is a plane through $p_i$, then $P \cap \Pi_1 \cap \Pi_2 \cap \Pi_3$ is contained in a line, which contains at most two points of $P$ except when $\Pi_1 = \Pi_2 = \Pi_3$ is the plane through $p_1, p_2, p_3$.
Thus we have $P$ lying in a plane except for at most $O(K^6) + O(K^2) = O(K^6)$ points, giving Case~\ref{case:1.1}.
Next suppose $|P_2'| \geqslant 3n/5$, and let $p_1, p_2, p_3$ be three distinct points in $P_2'$.
Then for each $i = 1, 2, 3$, there exist a quadric cone $C_i$ with vertex $p_i$ and planes $\{\Pi_{i,j} : j \in J_i\}$ through $p_i$ with $|J_i| = O(K^4)$,
such that $P \subset C_i \cup \bigcup_{j \in J_i} \Pi_{i,j}$ with $|P \cap C_i| = n/2 \pm O(K^5)$.
So all but at most $O(K^8)$ points of $P$ lie either on the intersection $C_1 \cap C_2 \cap C_3$, one of the $O(K^4)$ conics $C_i \cap \Pi_{i',j}$ for $i \ne i'$, or the plane $\Pi$ through $p_1, p_2, p_3$.
It is well known (and easy to deduce from B\'ezout's theorem (Corollary~\ref{cor:Bezout1})) that the intersection of two quadrics is either an irreducible space quartic, a twisted cubic and a line, or conics and lines.
We claim that any component $\delta$ of the intersection $C_1 \cap C_2 \cap C_3$ that is a twisted cubic or a space quartic cannot contain more than $O(K^4)$ points of $P$.
Choose a point $p \in P_2' \setminus (C_1 \cap C_2 \cap C_3)$ such that the projection $\pi_p$ restricted to $\delta$ is generically one-to-one.
Such a $p$ exists since $|P_2' \setminus (C_1 \cap C_2 \cap C_3)| \geqslant 3n/5 - (n/2 + O(K^5))$ and by Lemma~\ref{lem:projection3} there are only $O(1)$ exceptional points.
Then $\pi_p(\delta)$ is an irreducible planar cubic or quartic containing more than $O(K^4)$ points of $P$, contradicting $p \in P_2'$.
So the components of $C_1 \cap C_2 \cap C_3$ which contain more than $O(K^4)$ points of $P$ must all be conics.
No plane $\Pi'$ can contain more than $n/2 + O(K^5)$ points of $P$, otherwise projecting from $p' \in P_2' \cap \Pi'$ would give a line containing more than $n/2 + O(K^5)$ points in the projection, contradicting $p' \in P_2'$.
So choose a fourth point $p_4 \in P_2' \setminus \Pi$.
As before, $P$ is contained in the union of a quadric cone $C_4$ with vertex $p_4$ and $O(K^4)$ planes through $p_4$.
Since $p_4 \notin \Pi$, if $\Pi$ contains more than $O(K^4)$ points of $P$, all but at most $O(K^4)$ points of $P \cap \Pi$ must lie on the conic $C_4 \cap \Pi$.
We then have that all but at most $O(K^8)$ points of $P$ lie on $O(K^4)$ conics, and without loss of generality we can assume each conic contains more than $O(K^4)$ points.
Let $\Sigma$ be the set of such conics.
(Note that the same argument shows that if a plane $\Pi'$ contains a conic $\sigma \in \Sigma$, then all but at most $O(K^4)$ points of $P \cap \Pi'$ lie on $\sigma$.)
We show that $|\Sigma| = 2$, and that for each $\sigma \in \Sigma$, $|P \cap \sigma| = n/2 \pm O(K^8)$, thus giving Case~\ref{case:1.2}.
Let $\sigma_1$ be the conic in $\Sigma$ with the most points of $P$, and let $\Pi_1$ be the plane in which $\sigma_1$ lies.
Since no plane contains more than $n/2 + O(K^5)$ points of $P$, we have that $|P \setminus \Pi_1| \geqslant n/2 - O(K^5)$.
Let $\sigma_2$ be the conic in $\Sigma_1 \setminus \{\sigma_1\}$ with the most points of $P \setminus \Pi_1$, and let $\Pi_2$ be the plane in which $\sigma_2$ lies.
Then $|P \cap \sigma_1| \geqslant |P \cap \sigma_2| \geqslant \Omega(n/K^4)$.
Note that $\Pi_1\neq\Pi_2$, as no plane contains more than $n/2+O(K^5)$ points of $P$.
Suppose there exists $q \in P_2' \setminus (\Pi_1 \cup \Pi_2)$, so that (by B\'ezout's theorem (Corollary~\ref{cor:Bezout2})) $\sigma_1$ and $\sigma_2$ must both lie on the same quadric cone $C$ with vertex $q$.
Since $\sigma_1 \cup \sigma_2$ is the intersection of $\Pi_1 \cup \Pi_2$ and $C$, there can only be at most two such points by Proposition~\ref{prop:cones}.
Therefore, all but at most $O(K^4)$ points of $P_2'$ are contained in $\sigma_1 \cup \sigma_2$.
Without loss of generality, suppose $|P_2' \cap \sigma_1| \geqslant 3n/10 - O(K^4) = \Omega(n)$.
By Proposition~\ref{prop:cones} again, there exists at most two points in $P_2' \cap \sigma_1$ such that their quadric cones intersect in $\sigma_2$ and $\sigma'$ for some $\sigma' \in \Sigma \setminus \{\sigma_1, \sigma_2\}$.
We can then choose a $q' \in P_2' \cap \sigma_1$ such that the only conic in $\Sigma$ the quadric cone with vertex $q'$ contains is $\sigma_2$.
In particular, this means that $|P \cap \sigma_2| = n/2 \pm O(K^8)$.
But then we also have $|P_2' \cap \sigma_2| = \Omega(n)$.
Repeating the argument on $\sigma_2$ shows that $|P \cap \sigma_1| = n/2 \pm O(K^8)$ as well.
The remaining case is when $|P_3'| > 2n/3 - 3n/5 - 3 = \Omega(n)$.
Let $p$ and $p'$ be two distinct points in $P_3'$.
Then apart from $O(K^5)$ points, we have $P$ lying mostly on the intersection $\delta$ of two cubic cones, which is a curve with irreducible components $\delta_i$ of total degree $9$ by B\'ezout's theorem (Corollary~\ref{cor:Bezout1}).
Let $\delta_1$ be a component for which $|P_3' \cap \delta_i|$ is maximal.
Then $|P_3' \cap \delta_1| = \Omega(n)$.
Projecting from any $q \in P_3' \cap \delta_1$, we get that $\overline{\pi_q(\delta_1 \setminus \{q\})}$ is an irreducible cubic, and so $\delta_1$ must be non-planar.
By Lemma~\ref{lem:projection1}, all but $O(1)$ points $q'$ on $\delta_1$ are such that the projection $\pi_{q'}$ restricted to $\delta_1 \setminus \{q'\}$ is generically one-to-one.
We can thus choose such a $q' \in P_3' \cap \delta_1$ so that $\pi_{q'}$ projects $\delta_1 \setminus \{p_1\}$ generically one-to-one onto an irreducible cubic.
The component $\delta_1$ must then be a space quartic.
Now suppose there exists a component $\delta_2$ containing more than $O(K^5)$ points of $P$.
For any $q \in P_3' \cap \delta_1$, the cone $C_q(\delta_1)$ over $\delta_1$ has to contain $\delta_2$ by B\'ezout's theorem (Corollary~\ref{cor:Bezout2}).
Since $\delta_1$ is non-planar, this contradicts Lemma~\ref{lem:projection2}.
So we have all but at most $O(K^5)$ points of $P$ lying on a single space quartic, giving Case~\ref{case:1.3}.
\end{proof}
\subsection{Proof of the full structure theorem}\label{ssec:proof}
To get a more precise description of the structure of sets spanning few ordinary planes, we need a more precise description of sets that lie on certain cubic curves and span few ordinary lines.
The following two lemmas are~\cite{GT13}*{Lemma~7.4} and~\cite{GT13}*{Lemma~7.2}, respectively.
\begin{lemma}[Green--Tao \cite{GT13}]\label{lem:7.4}
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^2$ spanning at most $Kn$ ordinary lines, and suppose $n \geqslant CK$ for some sufficiently large absolute constant $C$.
Suppose all but $K$ points of $P$ lie on the union of an irreducible conic $\sigma$ and a line $\ell$, with $n/2 \pm O(K)$ points of $P$ on each of $\sigma$ and $\ell$.
Then up to a projective transformation, $P$ differs in at most $O(K)$ points from the vertices of a regular $m$-gon and the $m$ points at infinity corresponding to the diagonals of the $m$-gon, for some $m = n/2 \pm O(K)$.
\end{lemma}
\begin{lemma}[Green--Tao \cite{GT13}]\label{lem:7.2}
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^2$ spanning at most $Kn$ ordinary lines, and suppose $n \geqslant CK$ for some sufficiently large absolute constant $C$.
Suppose all but $K$ points of $P$ lie on an irreducible cubic $\gamma$.
Then $P$ differs in at most $O(K)$ points from a coset of a subgroup of $\gamma^*$, the smooth points of $\gamma$.
In particular, $\gamma$ is either an elliptic curve or an acnodal cubic.
\end{lemma}
We also need two further technical lemmas.
The following are \cite{GT13}*{Corollary~7.6} and \cite{GT13}*{Lemma~7.7}, respectively.
\begin{lemma}[Green--Tao \cite{GT13}]\label{cor:7.6}
Let $X_{2m}$ be the vertex set of a regular $m$-gon centred at the origin of the Euclidean plane, together with the $m$ points at infinity corresponding to the diagonals of the $m$-gon.
Let $p$ be a point not on the line at infinity, not the origin, and not a vertex of the $m$-gon.
Then at least $2m - O(1)$ of the $2m$ lines joining $p$ to a point of $X_{2m}$ do not pass through any further point of $X_{2m}$.
\end{lemma}
\begin{lemma}[Green--Tao \cite{GT13}]\label{lem:7.7}
Let $\gamma^*$ be an elliptic curve or the smooth points of an acnodal cubic curve.
Let $X$ be a coset of a finite subgroup of $\gamma^*$ of size $n$, where $n$ is greater than a sufficiently large absolute constant.
If $p\in\mathbb{R}\mathbb{P}^2\setminus\gamma^*$, then there are at least $n/1000$ lines through $p$ that pass through exactly one point in $X$.
\end{lemma}
Using the above four lemmas, we reduce the polynomial error terms in Lemma~\ref{lem:intermediate} to linear errors $O(K)$.
Since this refinement relies only on Green and Tao's results in~\cite{GT13}*{Section 7}, our proof will be similar to Ball's proof in~\cite{B17}.
\begin{lemma}\label{lem:intermediate2}
Let $K\geqslant 1$ and suppose $n \geqslant CK^8$ for some sufficiently large constant $C>0$.
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^3$ with no three collinear.
If $P$ spans at most $Kn^2$ ordinary planes,
then $P$ differs in at most $O(K)$ points from one of the following:
\begin{enumerate}[label=\rom]
\item A subset of a plane;\label{case:2.1}
\item A subset of two disjoint irreducible conics lying on distinct planes, each containing $\frac{n}{2} \pm O(K)$ points;\label{case:2.2}
\item A subset of a space quartic.\label{case:2.3}
\end{enumerate}
\end{lemma}
\begin{proof}
Let $P'$, $P_1'$, $P_2'$, and $P_3'$ be as in the proof of Lemma~\ref{lem:intermediate}.
If $|P_1'| \geqslant 3$, we are in Case~\ref{case:1.1} of Lemma~\ref{lem:intermediate}, and all but at most $O(K^6)$ points of $P$ lie in a plane $\Pi$.
Let $k:=|P\setminus\Pi|$.
Then for a fixed point $p\in P\setminus\Pi$, there are $\binom{n-k}{2}$ planes through $p$ and two points of $P\cap\Pi$, of which at most $k-1$ are not ordinary.
Therefore, there are at least $k(\binom{n-k}{2}-k+1)$ ordinary planes.
Since this is at most $Kn^2$ and $n \geqslant CK^8 > k = O(K^6)$ for sufficiently large $C$, we obtain that $k=O(K)$.
If $|P_2'| \geqslant 3n/5$, we are in Case~\ref{case:1.2} of Lemma~\ref{lem:intermediate}, and all but at most $O(K^8)$ points of $P$ lie on the union of two conics $\sigma_1 \cup \sigma_2$.
Let $S=P\setminus(\sigma_1\cup\sigma_2)$.
Let $\Pi_i$ be the plane supporting $\sigma_i$, $i=1,2$.
For any $p\in S\cap\Pi_1$ except at most two points also on $\Pi_2$, the projection $\pi_p$ maps $\sigma_1$ to a line and $\sigma_2$ to a conic.
For any $p\in S\setminus\Pi_1$ except at most two points, the projection $\pi_p$ maps $\sigma_1$ and $\sigma_2$ to distinct conics (by Proposition~\ref{prop:cones}).
Hence, for all but at most four points $p\in S$, there are at most four points $x\in P\cap\sigma_1$ for which $\pi_p(x)\in\pi_p(\sigma_2)$ and at most four points $x\in P\cap\sigma_2$ for which $\pi_p(x)\in\pi_p(\sigma_1)$.
Therefore, for any $q\in P_2' \cap (\sigma_1 \cap \sigma_2)$ except at most $4|S|+4|S|$ points, we have that $\pi_q(S)$ is disjoint from the line and the conic onto which all but at most $O(K^8)$ points of $P$ map.
Such a $q$ exists as $|P_2'|\geqslant 3n/5$.
By Lemma~\ref{lem:7.4}, the set $\pi_q(P\setminus\{q\})$ differs from a set $X$ projectively equivalent to a regular $m$-gon and the $m$ points at infinity corresponding to the diagonals of the $m$-gon in at most $O(K^8)$ points, where $m=n/2\pm O(K^8)$.
By Lemma~\ref{cor:7.6}, there are at least $n/2-O(K^8)$ ordinary lines through a fixed point of $\pi_q(S)$ and a point of $\pi_q(P\setminus(S\cup\{q\}))$.
Thus, there at least $|S|(n/2-O(K^8))$ ordinary lines.
Since there are at most $9Kn$ ordinary lines and $n \geqslant CK^8 > |S|=O(K^8)$ for sufficiently large $C$, we obtain that $|S|=O(K)$.
The same argument shows that $|\pi_q(P\setminus\{q\})\setminus X| =O(K)$, hence $|P\cap\sigma_i|\leqslant m+O(K)$, $i=1,2$.
It remains to show that $|X\setminus\pi_q(P\setminus\{q\})|=O(K)$.
Note that through any point $y \in X$, there are at least $m/2 - 1$ lines through $y$ and two more points of $X$.
By removing a point from $X$, we thus create at least $m/2-O(K)$ ordinary lines.
Therefore, there are at least $|X\setminus\pi_q(P\setminus\{q\})|(n/4-O(K^8))$ ordinary lines.
Since this is at most $9Kn$ and $n \geqslant CK^8$ for sufficiently large $C$, we obtain that $|X\setminus\pi_q(P\setminus\{q\})|=O(K)$.
Finally, if $|P_3'| > 2n/3 - 3n/5 - 3 = \Omega(n)$, we are in Case~\ref{case:1.3} of Lemma~\ref{lem:intermediate}, and all but at most $O(K^5)$ points of $P$ lie on a space quartic $\delta$.
By Lemma~\ref{lem:projection3}, the projection from all but finitely many points $p\in P\setminus\delta$ maps $\delta$ generically one-to-one onto a planar quartic, which has at most three singular points.
Thus, there are at most six points $x\in\delta$ such that $px$ intersects $\delta$ again.
Choose a point $q\in P_3' \cap \delta$ that is not one of these at most $6|P\setminus\delta|$ points.
Such a $q$ exists as $|P_3'| = \Omega(n)$.
Then, if we project from $q$, each point in $P\setminus\delta$ is projected onto a point not on the cubic $\overline{\pi_q(\delta\setminus\{q\})}$.
By Lemma~\ref{lem:7.2}, $\pi_q(P\setminus\{q\})$ differs in at most $O(K^5)$ points from a coset of a subgroup of an elliptic curve or the smooth points of an acnodal cubic.
By Lemma~\ref{lem:7.7}, there are at least $|P\setminus\delta|(n/1000-|P\setminus\delta|)$ ordinary lines.
Since this is at most $9Kn$ and $n \geqslant CK^8 > |P\setminus\delta|=O(K^5)$ for sufficiently large $C$, we obtain that $|P\setminus\delta|=O(K)$.
\end{proof}
We now show that if we are in Case~\ref{case:2.2} of Lemma~\ref{lem:intermediate2}, then there is a quadric containing both conics.
\begin{lemma}\label{lem:conics}
Let $K\geqslant 1$ and suppose $n \geqslant CK^8$ for some sufficiently large constant $C>0$.
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^3$ with no three collinear, spanning at most $Kn^2$ ordinary planes,
Suppose $P$ has $n/2 \pm O(K)$ points on each of two disjoint conics $\sigma_1$ and $\sigma_2$ lying on two distinct
planes $\Pi_1$ and $\Pi_2$, respectively.
Then there exists an irreducible quadric that contains $\sigma_1\cup\sigma_2$.
\end{lemma}
\begin{proof}
Let $P'$ be as in the proofs of Lemmas~\ref{lem:intermediate} and~\ref{lem:intermediate2} above.
We convert the problem to one in Euclidean geometry, by identifying $\mathbb{R}\mathbb{P}^3$ with the Euclidean affine space $\mathbb{R}^3$ together with a projective plane at infinity.
We apply a projective transformation such that the planes $\Pi_1$ and $\Pi_2$ become parallel, and such that $\sigma_1$ is a circle.
It then suffices to show that $\sigma_2$ is a circle as well, as $\sigma_1\cup\sigma_2$ is then contained in a circular cylinder.
Choose $p_i \in P' \cap \sigma_i$, $i=1,2$,
and consider the projection $\pi_i:=\pi_{p_i}$ to be onto the plane $\Pi_{3-i}$.
Then by Lemma~\ref{lem:intermediate2}, $\pi_1$ projects all but at most $O(K)$ points of $P$ onto the line at infinity and a disjoint conic on $\Pi_2$.
Since the conic $\sigma_2=\pi_1(\sigma_2)$ is disjoint from the line at infinity, it is an ellipse.
Now let $p_2 \in P' \cap \sigma_2$, and consider the projection $\pi_2$.
By Lemma~\ref{lem:7.4}, $\pi_2(P \setminus \{p_2\})$ differs in at most $O(K)$ points from the vertices of a regular $m$-gon and the $m$ points at infinity corresponding to the diagonals of the $m$-gon, for some $m = n/2 \pm O(K)$.
In particular, $P \cap \sigma_1$ differs in at most $O(K)$ points from a regular $m$-gon.
Therefore, $\pi_1(P \cap \sigma_1 \setminus \{p_1\})$ differs in at most $O(K)$ points from the points at infinity corresponding to the diagonals of the regular $m$-gon on $\sigma_1$,
which are also the points at infinity corresponding to the tangents to $\sigma_1$ at the vertices of the $m$-gon.
By Lemma~\ref{lem:7.4} again, $\pi_1(P \cap \sigma_2)$ differs in at most $O(K)$ points from an $m$-gon on the ellipse $\sigma_2$, projectively equivalent to a regular $m$-gon.
It easily follows that all but at most $O(K)$ of the tangent lines to $\sigma_2$ at the vertices of the $m$-gon are ordinary lines and so all but $O(K)$ must be points at infinity of $\pi_1(P \cap \sigma_1 \setminus \{p_1\})$.
Let $a,b,c$ be three consecutive vertices of the $m$-gon on $\sigma_2$ such that $a,b,c\in\pi_1(P \cap \sigma_2)$.
Then the point $d$ where the tangents at $a$ and $c$ intersect, forms an isosceles triangle with $a$ and $c$, and we have $|ad| = |cd|$.
But this can only happen if $d$ lies on one of the axes of symmetry of $\sigma_2$.
Since $n$ is sufficiently large depending on $K$, we can find many triples of consecutive vertices of the $m$-gon on $\sigma_2$, and we get a contradiction unless $\sigma_2$ is a circle.
\end{proof}
The following lemma shows that if we are in Case~\ref{case:2.3} of Lemma~\ref{lem:intermediate2}, then the space quartic is either elliptic or rational of the first species.
\begin{lemma}\label{lem:quartics}
Let $K\geqslant 1$ and supose $n\geqslant CK^8$ for some sufficiently large constant $C>0$.
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^3$ with no three collinear, spanning at most $Kn^2$ ordinary planes.
Suppose all but at most $O(K)$ points of $P$ lie on a rational space quartic $\delta$ given by the parametrisation \eqref{eqn:param}.
Then $\delta$ is of the first species.
\end{lemma}
\begin{proof}
Let $p_\alpha \in \delta$ be parametrised by $\alpha\in\mathbb{R}$ in such a way that the point of $\delta$ not corresponding to any $\alpha\in\mathbb{R}$ is not in $P$.
As before, let $P'$ denote the set of all points $p\in P\cap\delta$ with at most $9Kn$ ordinary planes through $p$.
Since $P$ spans at most $Kn^2$ ordinary planes, we have $|P'|\geqslant 2n/3-O(K)$.
Let $\pi_\alpha$ be the projection map from $p_\alpha$ onto a plane not containing $p_\alpha$.
By Lemma~\ref{lem:projection1}, we can choose $p_\alpha \in P'$ such that all but at most $O(K)$ points of $\pi_\alpha(P \setminus \{p_\alpha\})$ lie on a cubic curve $\gamma_\alpha$.
By Lemma~\ref{lem:7.2}, this set differs in at most $O(K)$ points from a coset of $\gamma_\alpha$, and $\gamma_\alpha$ is acnodal (since $\delta$ is rational).
Since $n$ is sufficiently large, there exist three distinct points $p_A, p_B, p_C \in P'$ such that for $\Omega(n)$ many $p_\alpha \in P'$, the projected points $\pi_\alpha(p_A),\pi_\alpha(p_B),\pi_\alpha(p_C)$ are consecutive elements in the coset given by Lemma \ref{lem:7.2}.
Let $\oplus_\alpha$ denote the group operation on $\gamma_\alpha$ so that we have $\pi_\alpha(p_A) \oplus_\alpha \pi_\alpha(p_C) = 2\pi_\alpha(p_B)$.
By considering the geometric definition of $\oplus_\alpha$ we obtain that if $p_\beta$ is the fourth point of intersection between $\delta$ and the plane through $p_A, p_C, p_\alpha$, and $p_{\beta'}$ the fourth point of intersection between $\delta$ and the plane through $p_B, p_B, p_\alpha$ (that is, containing the tangent line of $\delta$ at $p_B$ and passing through $p_\alpha$), then $\beta = \beta'$.
Equivalently, by Lemma~\ref{lem:coplanar}, we have
\begin{equation*}
F(A, C, \alpha, \beta) = 0 = F(B, B, \alpha, \beta).
\end{equation*}
Since $F$ is a polynomial and the above is true for sufficiently many $\alpha$ (since $n$ is sufficiently large), it holds for all $\alpha \in \mathbb{R}$.
Note that
\begin{equation*}
F(A, C, \alpha, \beta) =
\begin{pmatrix}
\alpha \beta, & \alpha + \beta, & 1
\end{pmatrix}
\begin{pmatrix}
1 & s & r\\
s & r & q\\
r & q & p\\
\end{pmatrix}
\begin{pmatrix}
AC\\
A+C\\
1\\
\end{pmatrix}
,
\end{equation*}
with a similar expression for $F(B,B,\alpha,\beta)$.
Since $\begin{pmatrix} AC, & A+C, & 1 \end{pmatrix}$ and $\begin{pmatrix} B^2, & 2B, & 1 \end{pmatrix}$ are linearly independent, the set of vectors
\begin{equation*}
\left\lbrace
\begin{pmatrix}
\alpha \beta, & \alpha + \beta, & 1
\end{pmatrix}
\begin{pmatrix}
1 & s & r\\
s & r & q\\
r & q & p\\
\end{pmatrix}
: \alpha \in \mathbb{R}
\right\rbrace
\end{equation*}
lie in a $1$-dimensional linear subspace of $\mathbb{R}^3$.
If the catalecticant of the fundamental quartic of $\delta$, which is the determinant $pr - q^2 - ps^2 + 2qrs -r^3$ of the above $3\times 3$ matrix, is non-zero, then
\begin{equation*}
\left\lbrace
\begin{pmatrix}
\alpha \beta, & \alpha + \beta, & 1
\end{pmatrix}
: \alpha \in \mathbb{R}
\right\rbrace
\end{equation*}
also lies in a $1$-dimensional subspace of $\mathbb{R}^3$, in which case both $\alpha \beta$ and $\alpha + \beta$ are constants depending only on $\delta$, say $\alpha \beta = c_1$ and $\alpha + \beta = c_2$.
But then $\alpha$ is a root of the fixed quadratic equation $x^2 - c_2x + c_1 = 0$, a contradiction.
Hence, the catalecticant vanishes, and $\delta$ is of the first species by Lemma~\ref{lem:first}.
\end{proof}
From Lemmas~\ref{lem:intermediate2},~\ref{lem:conics}, and~\ref{lem:quartics}, we see that up to at most $O(K)$ points, the set $P$ lies on a plane, two disjoint conic sections of an irreducible quadric (which by applying a projective transformation if necessary we can assume to be two disjoint circles on a sphere), or a space quartic of the first species.
Noting from Section~\ref{ssec:groups} that only elliptic and acnodal space quartics admit arbitrarily large finite subgroups,
Lemmas~\ref{lem:conics2} and~\ref{lem:quartics2}, which are proved in a similar way as~\cite{LMMSSZ17}*{Lemmas~5.5 and~5.6}, respectively, complete the proof of Theorem~\ref{thm:strong}.
Note that these are analogues of Lemmas~\ref{lem:7.4} and~\ref{lem:7.2} for ordinary planes.
\begin{lemma}\label{lem:conics2}
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^3$ with no three collinear, spanning at most $Kn^2$ ordinary planes, and suppose $n \geqslant CK$ for some sufficiently large absolute constant $C$.
Suppose all but at most $K$ points of $P$ lie on two disjoint circles, with $n/2 \pm O(K)$ points of $P$ on each circle.
Then, up to a projective transformation, $P$ differs in at most $O(K)$ points from a prism or an antiprism.
\end{lemma}
\begin{lemma}\label{lem:quartics2}
Let $P$ be a set of $n$ points in $\mathbb{R}\mathbb{P}^3$ with no three collinear, spanning at most $Kn^2$ ordinary planes, and suppose $n \geqslant CK$ for some sufficiently large absolute constant $C$.
Suppose all but at most $K$ points of $P$ lie on a space quartic $\delta$ of the first species.
Then $P$ differs in at most $O(K)$ points from a coset of a subgroup of $\delta^*$, the smooth points of $\delta$.
In particular, $\delta$ is either an elliptic or acnodal space quartic.
\end{lemma}
\section{Coplanar quadruples}\label{sec:quad}
The following is a special case of Raz, Sharir, and De Zeeuw's $4$-dimensional generalisation of the Elekes-Szab\'o theorem~\cite{RSZ16}, which we use to prove Theorem~\ref{thm:quad}.
\begin{theorem}\label{thm:ES}
Let $F \in \mathbb{C}[t_1, t_2, t_3, t_4]$ be irreducible of degree $d$, with no $\partial F / \partial t_i$ identically zero.
Then either for all $A \subset \mathbb{C}$ with $|A| = n$, we have $|Z_\mathbb{C}(F) \cap A^4| = O(n^{8/3})$,
or there exists a $2$-dimensional subvariety $Z_0 \subset Z_\mathbb{C}(F)$ such that for any $(s_1, s_2, s_3, s_4) \in Z_\mathbb{C}(F) \setminus V_0$, there exist open neighbourhoods $U_i$ of $s_i$ and injective analytic functions $\varphi_i: U_i \rightarrow \mathbb{C}$ such that
\begin{equation*}
F(t_1, t_2, t_3, t_4) = 0 \Longleftrightarrow \varphi_1(t_1) + \varphi_2(t_2) + \varphi_3(t_3) + \varphi_4(t_4) = 0,
\end{equation*}
for all $(t_1, t_2, t_3, t_4) \in U_1 \times U_2 \times U_3 \times U_4$.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{thm:quad}]
We first show that if an $n$-point set $P$ on a rational space quartic $\delta$ in $\mathbb{C}\mathbb{P}^3$ does not span $O(n^{8/3})$ coplanar quadruples, then $\delta$ is of the first species.
By Proposition~\ref{prop:param}, we may assume $\delta$ is given by the parametrisation~\eqref{eqn:param}.
Lemma~\ref{lem:coplanar} says four points parametrised by $t_1, t_2, t_3, t_4$ are coplanar if and only if \eqref{eqn:coplanar} holds.
It is clear that $F$ is not independent of any $t_i$ (otherwise $\delta$ would be planar).
Since $P$ does not span $O(n^{8/3})$ coplanar quadruples, Theorem~\ref{thm:ES} gives the existence of injective analytic $\varphi_1, \varphi_2, \varphi_3, \varphi_4$ such that $F = 0$ if and only if $\varphi_1(t_1) + \varphi_2(t_2) + \varphi_3(t_3) + \varphi_4(t_4) = 0$.
In particular, on the hypersurface $F=0$, we can express $t_4 = t_4 (t_1, t_2, t_3)$ as a function of $t_1, t_2, t_3$:
\begin{equation}\label{t4}
t_4(t_1,t_2,t_3) = -\dfrac{s t_{1} t_2t_3 + r(t_{1}t_2 + t_{1}t_3 + t_{2}t_3) + q(t_{1} + t_{2} + t_3) + p}{ t_{1} t_{2}t_3 + s(t_{1}t_2 + t_{1}t_3 + t_2t_3) + r(t_{1} + t_2 + t_3) + q}.
\end{equation}
We thus have for all $(t_1,t_2,t_3)\in U_1\times U_2\times U_3$ that
\begin{equation*}
\varphi_1(t_1) + \varphi_2(t_2) + \varphi_3(t_3) + \varphi_4(t_4(t_1,t_2,t_3)) = 0.
\end{equation*}
Partial differentiation with respect to $t_i$ for $i = 2,3$ gives
\begin{equation*}
\varphi_i'(t_i) + \varphi_4'(t_4) \frac{\partial t_4}{\partial t_i} = 0,
\end{equation*}
and so the quotient $\left( \partial t_4 / \partial t_2 \right) / \left( \partial t_4 / \partial t_3 \right)$ is independent of $t_1$.
The numerator of the partial derivative of this quotient with respect to $t_1$ is thus identically zero.
If we substitute \eqref{t4} into \[ \frac{\partial}{\partial t_1}\left(\frac{\partial t_4}{\partial t_2} \middle/ \frac{\partial t_4}{\partial t_3}\right) = 0,\]
we obtain (with the help of a computer algebra system such as SageMath)
\begin{equation*}
\frac{(pr - q^2 - ps^2 + 2qrs - r^3)F(t_1, t_1, t_2, t_3)(t_3 - t_2)}{h(t_1,t_2)^2} = 0,
\end{equation*}
where
\begin{align*}
h(t_1, t_2) &= (s^{2} - r) t_{1}^{2} t_{2}^{2} + (r s - q) t_1 t_2(t_{1} + t_{2}) + (r^{2} -qs) (t_{1}^{2} + t_2^2)\\
&\mathrel{\phantom{=}} \mbox{} + (r^{2} - p) t_{1} t_{2} + (q r - p s) (t_{1} + t_{2}) + q^{2} - p r.
\end{align*}
Since $t_1, t_2, t_3$ are arbitrary in $(U_1\times U_2\times U_3)\setminus (Z_{\mathbb{C}}(h)\times\mathbb{C})$, we obtain that the catalecticant $pr - q^2 - ps^2 + 2qrs - r^3$ vanishes.
By Lemma \ref{lem:first}, the rational space quartic $\delta$ is thus of the first species, as desired.
For the converse, suppose that $\delta$ is of the first species.
Then, as is well known (and explicitly demonstrated in Lemma~\ref{lem:group} below), the smooth points $\delta^*$ carry a group structure such that four points are coplanar if and only if their sum in the group is the identity.
If $\delta$ is nodal, then the group is isomorphic to the non-zero complex numbers under multiplication $(\mathbb{C}^*,\cdot)$.
If $\delta$ is cuspidal, then the group is isomorphic to the complex numbers under addition $(\mathbb{C},+)$.
In both groups it is trivial to find $n$ elements such that there are $\Theta(n^3)$ quadruples of distinct elements that sum to $0$.
In the multiplicative case, we can take the $n$-th roots of unity, and in the additive case, we can take the $n$ integers closest to $0$.
\end{proof}
If a rational space quartic $\delta$ of the first species is given by the parametrisation~\eqref{eqn:param}, we can also find the $\varphi_i$'s in Theorem~\ref{thm:ES} explicitly.
For convenience, in the next lemma we identify $\mathbb{C}\mathbb{P}^1$ with the affine line $\mathbb{C}$ together with a point $\infty$ at infinity.
\begin{lemma}\label{lem:group}
Let $\delta$ be a rational space quartic of the first species in $\mathbb{C}\mathbb{P}^3$ given by the parametrisation~\eqref{eqn:param}.
If $\delta$ is nodal, then there exists a parametrisation $\varphi\colon\mathbb{C}\mathbb{P}^1\to\delta$ such that $\varphi(0)=\varphi(\infty)$ is the node of $\delta$, and any four points $\varphi(t_1), \varphi(t_2),\varphi(t_3),\varphi(t_4)$ on $\delta\setminus\{\varphi(0)\}$ are coplanar if and only if $t_1t_2t_3t_4=1$.
If $\delta$ is cuspidal, then there exists a parametrisation $\varphi\colon\mathbb{C}\mathbb{P}^1\to\delta$ such that $\varphi(\infty)$ is the node of $\delta$, and any four points $\varphi(t_1), \varphi(t_2),\varphi(t_3),\varphi(t_4)$ on $\delta\setminus\{\varphi(\infty)\}$ are coplanar if and only if $t_1+t_2+t_3+t_4=0$.
\end{lemma}
\begin{proof}
Since $\delta$ is of the first species, the catalecticant of its fundamental quartic given by \eqref{eqn:quartic} vanishes by Lemma~\ref{lem:first}.
Applying Theorem~\ref{thm:sylvester} and dehomogenising, we can write $f(t)$ as one of
\[(a t + b)^4, (a_1 t + b_1)^4 + (a_2 t + b_2)^4, (a_1 t + b_1)(a_2 t + b_2)^3,\]
for some $a, b, a_1,b_1,a_2,b_2 \in \mathbb{C}$, where $a_1b_2 \ne b_1a_2$.
We cannot have $f(t) = (at + b)^4$, otherwise we have $r = s^2$, $q = s^3$, and $p = s^4$, contradicting the parametrisation~\eqref{eqn:param}.
If $f(t) = (a_1 t + b_1)^4 + (a_2 t + b_2)^4$, then its so-called polarisation \cite{D03}*{Section~1.2} will be equal to
\begin{equation*}
F(t_1,t_2,t_3,t_4) = \prod_{i=1}^4 (a_1 t_i + b_1) + \prod_{i=1}^4 (a_2 t_i + b_2).
\end{equation*}
So $F = 0$ if and only if
\begin{equation*}
\frac{a_1 t_1 + b_1}{a_2 t_1 + b_2} \cdot \frac{a_1 t_2 + b_1}{a_2 t_2 + b_2} \cdot \frac{a_1 t_3 + b_1}{a_2 t_3 + b_2} \cdot \frac{a_1 t_4 + b_1}{a_2 t_4 + b_2} = -1.
\end{equation*}
This is the nodal case, with the node corresponding to $\varphi(t)$ where the fractional linear transformation $(a_1t+b_1)/(a_2t+b_2) = 0,\infty$.
We can make a linear change of variables so that this becomes $t_1t_2t_3t_4=1$, and the node is $\varphi(0)=\varphi(\infty)$.
Otherwise $f(t) = (a_1 t + b_1)(a_2 t + b_2)^3$.
Since the catalecticant is an invariant of the fundamental quartic of $\delta$, we can use a linear change of variables to assume $f(s) = (cs-d)s^3$, for some $c,d \in \mathbb{C}$.
We then obtain the polarisation
\begin{equation*}
F(s_1, s_2, s_3, s_4) = cs_1s_2s_3s_4 - d \sum_{i < j <k} s_is_js_k = 0,
\end{equation*}
which can be written as $\varphi(s_1) + \varphi(s_2) + \varphi(s_3) + \varphi(s_4) = 0$, where
\begin{equation*}
\varphi(s) = \frac{c}{4} - \frac{d}{s}.
\end{equation*}
This is the case where $\delta$ is cuspidal.
\end{proof}
\begin{bibdiv}
\begin{biblist}
\bib{ACGH}{book}{
author={Arbarello, E.},
author={Cornalba, M.},
author={Griffiths, P. A.},
author={Harris, J.},
title={Geometry of Algebraic Curves. Vol. I},
series={Grundlehren der Mathematischen Wissenschaften},
volume={267},
publisher={Springer},
date={1985},
}
\bib{B17}{article}{
author={Ball, Simeon},
title={On sets defining few ordinary planes},
journal={Discrete Comput.\ Geom.},
volume={60},
date={2018},
pages={220--253},
}
\bib{BJ18}{article}{
author={Ball, Simeon},
author={Jimenez, Enrique},
title={On sets defining few ordinary solids},
note={arXiv:1808.06388},
}
\bib{Ballico2003}{article}{
author={Ballico, E.},
title={Special projections of projective varieties},
journal={Int. Math. J.},
volume={3},
date={2003},
pages={11--12},
issn={1311-6797},
}
\bib{Ballico2004}{article}{
author={Ballico, E.},
title={Special inner projections of projective varieties},
journal={Ann. Univ. Ferrara Sez. VII (N.S.)},
volume={50},
date={2004},
pages={23--26},
issn={0430-3202},
}
\bib{LeBarz82}{article}{
author={Le Barz, Patrick},
title={Formules multis\'ecantes pour les courbes gauches quelconques},
language={French},
conference={
title={Enumerative geometry and classical algebraic geometry (Nice, 1981)},
},
book={
series={Progr. Math.},
volume={24},
publisher={Birkh\"auser, Boston, Mass.},
},
date={1982},
pages={165--197},
}
\bib{Bertin}{article}{
author={Bertin, Marie-Am\'elie},
title={On the singularities of the trisecant surface to a space curve},
journal={Matematiche (Catania)},
volume={53},
date={1998},
pages={15--22 (1999)},
}
\bib{B1871}{article}{
author={Brill, A.},
title={Ueber die Doppelpunkte von Curven im Raume, deren Geschlecht Null ist},
date={1871},
journal={Math. Ann.},
volume={3},
pages={456--458},
}
\bib{D03}{book}{
author={Dolgachev, Igor},
title={Lectures on Invariant Theory},
publisher={Cambridge University Press},
date={2003},
}
\bib{F1895}{article}{
author={Forsyth, A. R.},
title={On twisted quartics of the second species},
journal={The Quarterly Journal of Pure and Applied Mathematics},
volume={27},
date={1895},
pages={247--269},
}
\bib{F11}{article}{
author={Furukawa, Katsuhisa},
title={Defining ideal of the Segre locus in arbitrary characteristic},
journal={J. Algebra},
volume={336},
date={2011},
pages={84--98},
}
\bib{GT13}{article}{
author={Green, Ben},
author={Tao, Terence},
title={On sets defining few ordinary lines},
date={2013},
journal={Discrete Comput.\ Geom.},
volume={50},
pages={409--468},
}
\bib{GP82}{article}{
author={Gruson, Laurent},
author={Peskine, Christian},
title={Courbes de l'espace projectif: vari\'et\'es de s\'ecantes},
language={French},
conference={
title={Enumerative geometry and classical algebraic geometry (Nice, 1981)},
},
book={
series={Progr. Math.},
volume={24},
publisher={Birkh\"auser, Boston, Mass.},
},
date={1982},
pages={1--31},
}
\bib{H92}{book}{
author={Harris, J.},
title={Algebraic Geometry: A First Course},
publisher={Springer},
date={1992},
}
\bib{H77}{book}{
author={Hartshorne, R.},
title={Algebraic Geometry},
publisher={Springer},
date={1977},
}
\bib{K99}{article}{
author={Kanev, Vassil},
title={Chordal varieties of {V}eronese varieties and catalecticant matrices},
journal={J. Math. Sci.},
volume={94},
year={1999},
pages={1114--1125},
}
\bib{KKT08}{article}{
author={Kaminski, J. Y.},
author={Kanel-Belov, A.},
author={Teicher, M.},
title={Trisecant lemma for nonequidimensional varieties},
journal={J. Math. Sci.},
volume={149},
year={2008},
pages={1087--1097},
}
\bib{Kollar}{book}{
author={Koll\'ar, J\'anos},
title={Lectures on Resolution of Singularities},
series={Annals of Mathematics Studies},
volume={166},
publisher={Princeton University Press}
date={2007},
}
\bib{LMMSSZ17}{article}{
author={Lin, Aaron},
author={Makhul, Mehdi},
author={Mojarrad, Hossein Nassajian},
author={Schicho, Josef},
author={Swanepoel, Konrad},
author={de Zeeuw, Frank},
title={On sets defining few ordinary circles},
journal={Discrete Comput. Geom.},
volume={59},
date={2018},
pages={59--87},
note={arXiv:1607.06597},
}
\bib{LS}{article}{
author={Lin, Aaron},
author={Swanepoel, Konrad},
title={On sets defining few ordinary hyperplanes},
note={arXiv:1808.10849},
}
\bib{M51}{article}{
author={Motzkin, T.},
title={The lines and planes connecting the points of a finite set},
date={1951},
journal={Trans. Amer. Math. Soc.},
volume={70},
pages={451--464},
}
\bib{M10}{thesis}{
author={Muntingh, Georg},
title={Topics in polynomial interpolation theory},
type={Ph.D. Dissertation},
address={University of Oslo},
date={2010},
}
\bib{PS10}{article}{
author={Purdy, George B.},
author={Smith, Justin W.},
title={Lines, circles, planes and spheres},
journal={Discrete Comput. Geom.},
volume={44},
date={2010},
pages={860--882},
}
\bib{R1900}{article}{
author={Richmond, H.},
title={Rational space-curves of the fourth order},
year={1900},
journal={Trans. Camb. Phil. Soc.},
volume={19},
pages={132--150},
}
\bib{RSZ16}{article}{
author={Raz, Orit E.},
author={Sharir, Micha},
author={de Zeeuw, Frank},
title={The {E}lekes-{S}zab{\'o} {T}heorem in four dimensions},
journal={Israel J. Math.},
year={2018},
volume={227},
pages={663--690},
}
\bib{S36}{article}{
author={Segre, Beniamino},
title={On the locus of points from which an algebraic variety is projected multiplly},
journal={Proc. Phys.-Math. Soc. Japan (3)},
year={1936},
volume={18},
pages={425--426},
}
\bib{Severi}{book}{
author={Severi, Francesco},
title={Vorlesungen \"uber algebraische Geometrie: Geometrie auf einer
Kurve, Riemannsche Fl\"achen, Abelsche Integrale},
series=
Bibliotheca Mathematica Teubneriana, Band 32},
publisher={Teubner, Lepizig},
date={1921},
}
\bib{S09}{book}{
author={Silverman, J. H.},
title={The Arithmetic of Elliptic Curves},
edition={Second Edition},
publisher={Springer},
date={2009},
}
\bib{S39}{book}{
author={Sommerville, D. M. Y.},
title={Analytic Geometry of Three Dimensions},
publisher={Cambridge University Press},
date={1939},
}
\bib{S1851}{article}{
author={Sylvester, J. J.},
title={On a remarkable discovery in the theory of canonical forms and of hyperdeterminants},
date={1851},
journal={Philosophical Magazine},
volume={2},
pages={391--410},
note={Paper 41 in The Collected Mathematical Papers of James Joseph Sylvester, Cambridge University Press, 1904},
}
\bib{T36}{book}{
author={Telling, H. G.},
title={The Rational Quartic Curve in Space of Three and Four Dimensions},
publisher={Cambridge University Press},
date={1936},
note={Re-issued 2015}
}
\bib{W1871}{article}{
author={Weyr, E.},
title={Ueber rationale Curven vierter Ordnung},
date={1871},
journal={Math. Ann.},
volume={4},
pages={243--244},
}
\bib{W08}{article}{
author={Wolfson, P.R.},
date={2008},
title={George Boole and the origins of invariant theory},
journal={Historia Mathematica},
volume={35},
pages={37--46},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{#1}}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{conjecture}[theorem]{Conjecture}
\newtheorem{fact}[theorem]{Fact}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{claim}[theorem]{Claim}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{observation}{Observation}[theorem]
\newtheorem{definition}[theorem]{Definition}
\newtheorem{problem}[theorem]{Problem}
\newtheorem{remark}[theorem]{Remark}
\newcommand {\proof}{{\bf Proof:\newline \indent}}
\newcommand{\noindent{\it Proof}}{\noindent{\it Proof}}
\newcommand{\noindent{\it Proof of Claim:}}{\noindent{\it Proof of Claim:}}
\newcommand{\noindent{\it Proof of Theorem \ref{idm-th}.}}{\noindent{\it Proof of Theorem \ref{idm-th}.}}
\newcommand{\noindent{\it Proof of Theorem \ref{pit-id-th}.}}{\noindent{\it Proof of Theorem \ref{pit-id-th}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{main-thm-4}.}}{\noindent{\it Proof of Theorem ~\ref{main-thm-4}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{lowVC}.}}{\noindent{\it Proof of Theorem ~\ref{lowVC}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{degparhard}.}}{\noindent{\it Proof of Theorem ~\ref{degparhard}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{degparalgo}.}}{\noindent{\it Proof of Theorem ~\ref{degparalgo}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{genparhard}.}}{\noindent{\it Proof of Theorem ~\ref{genparhard}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{genparuna}.}}{\noindent{\it Proof of Theorem ~\ref{genparuna}.}}
\newcommand{\noindent{\it Proof of Theorem ~\ref{main-thm-1}.}}{\noindent{\it Proof of Theorem ~\ref{main-thm-1}.}}
\newcommand{\noindent{\it Proof of Lemma \ref{partial}.}}{\noindent{\it Proof of Lemma \ref{partial}.}}
\newcommand{\noindent{\it Proof of Lemma \ref{inv}.}}{\noindent{\it Proof of Lemma \ref{inv}.}}
\newcommand{\hspace*{\fill}$\rule{2mm}{2mm}$~~~~~\bigskip}{\hspace*{\fill}$\rule{2mm}{2mm}$~~~~~\bigskip}
\newcommand{\hspace*{\fill}$\rule{2mm}{2mm}$~~~~~\bigskip}{\hspace*{\fill}$\rule{2mm}{2mm}$~~~~~\bigskip}
\renewenvironment{proof}{\noindent{\it Proof}. }{\hspace*{\fill}$\rule{2mm}{2mm}$~~~~~\bigskip}
\newenvironment{psketch}{\noindent{\it Proof}~{\it Sketch}. }{\hspace*{\fill}$\rule{2mm}{2mm}$~~~~~\bigskip}
\newenvironment{claimproof}{\noindent{\it Proof of Claim:}. }{\hspace*{\fill}\vspace{5mm}}
\usepackage{thm-restate}
\usepackage{hyperref}
\usepackage{cleveref}
\declaretheorem[name=Lemma]{lem}
\newcommand{\op}[1]{\ensuremath{\operatorname{#1}}}
\newcommand{\op{mini}}{\op{mini}}
\newcommand{\mbox{\small\rm 3-CNF}}{\mbox{\small\rm 3-CNF}}
\newcommand{\mbox{\small\rm PIT}}{\mbox{\small\rm PIT}}
\newcommand{\mbox{\small\rm PCP}}{\mbox{\small\rm PCP}}
\newcommand{\mbox{\small\rm IP}}{\mbox{\small\rm IP}}
\newcommand{\mbox{\small\rm PSPACE}}{\mbox{\small\rm PSPACE}}
\newcommand{\mbox{\small\rm LIM}}{\mbox{\small\rm LIM}}
\newcommand{\mbox{\small\rm co{-}RP}}{\mbox{\small\rm co{-}RP}}
\newcommand{\mbox{\small\rm NEXP}}{\mbox{\small\rm NEXP}}
\newcommand{\mbox{\small\rm XP}}{\mbox{\small\rm XP}}
\newcommand{\mbox{\small\rm FPT}}{\mbox{\small\rm FPT}}
\newcommand{\mbox{\small\rm MINI{[}1{]}}}{\mbox{\small\rm MINI{[}1{]}}}
\newcommand{\mbox{\small\rm MINI{-}3SAT}}{\mbox{\small\rm MINI{-}3SAT}}
\newcommand{\mbox{\small\rm MINI-1-in-3 POSITIVE 3-SAT}}{\mbox{\small\rm MINI-1-in-3 POSITIVE 3-SAT}}
\newcommand{\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}}{\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}}
\newcommand{\mbox{\small\rm 1-in-3 POSITIVE 3-SAT}}{\mbox{\small\rm 1-in-3 POSITIVE 3-SAT}}
\newcommand{\mbox{\small\rm REM}}{\mbox{\small\rm REM}}
\newcommand{\mbox{\small\rm CRT{-}CANON}}{\mbox{\small\rm CRT{-}CANON}}
\newcommand{\mbox{\rm Phase}}{\mbox{\rm Phase}}
\newcommand{\mbox{\rm Sub{-}phase}}{\mbox{\rm Sub{-}phase}}
\newcommand{\mbox{\rm Step}}{\mbox{\rm Step}}
\newcommand{\mbox{\rm T}}{\mbox{\rm T}}
\newcommand{\mbox{\rm UP}}{\mbox{\rm UP}}
\newcommand{\ensuremath{\rm PP}}{\ensuremath{\rm PP}}
\newcommand{\mbox{\rm GapP}}{\mbox{\rm GapP}}
\newcommand{\mbox{\rm NP}}{\mbox{\rm NP}}
\newcommand{\mbox{\rm FP}}{\mbox{\rm FP}}
\newcommand{\mbox{\rm coNP}}{\mbox{\rm coNP}}
\newcommand{\mbox{$\oplus$\rm P}}{\mbox{$\oplus$\rm P}}
\newcommand{\mbox{\rm Mod$_k$P}}{\mbox{\rm Mod$_k$P}}
\newcommand{\mbox{\rm Mod$_p$P}}{\mbox{\rm Mod$_p$P}}
\newcommand{\mbox{\rm \#P}}{\mbox{\rm \#P}}
\newcommand{\mbox{\rm LWPP}}{\mbox{\rm LWPP}}
\newcommand{\mbox{\rm UPSV}}{\mbox{\rm UPSV}}
\newcommand{\mbox{\rm SPP}}{\mbox{\rm SPP}}
\newcommand{\mbox{\rm AM}}{\mbox{\rm AM}}
\newcommand{\mbox{\rm coAM}}{\mbox{\rm coAM}}
\newcommand{\mbox{\rm TRUE}}{\mbox{\rm TRUE}}
\newcommand{\mbox{\rm FALSE}}{\mbox{\rm FALSE}}
\newcommand{\mbox{\rm coMA}}{\mbox{\rm coMA}}
\newcommand{\mbox{\rm ModP}}{\mbox{\rm ModP}}
\newcommand{\mbox{\rm DSPACE}}{\mbox{\rm DSPACE}}
\newcommand{\mbox{\rm BPL}}{\mbox{\rm BPL}}
\newcommand{\CeqP}{\mbox{\rm C$_=$P}}
\newcommand{\mbox{\it gap}}{\mbox{\it gap}}
\newcommand{\mbox{\rm BQP}}{\mbox{\rm BQP}}
\newcommand{\mbox{\rm BPP}}{\mbox{\rm BPP}}
\newcommand{\mbox{\rm EXPSPACE}}{\mbox{\rm EXPSPACE}}
\newcommand{\mbox{\rm AWPP}}{\mbox{\rm AWPP}}
\renewcommand{\L}{\mbox{\rm L}}
\newcommand{\mbox{\rm NL}}{\mbox{\rm NL}}
\newcommand{\mbox{\rm RL}}{\mbox{\rm RL}}
\newcommand{\mbox{\rm ZPL}}{\mbox{\rm ZPL}}
\newcommand{\mbox{\rm SL}}{\mbox{\rm SL}}
\newcommand{\mbox{\rm coRL}}{\mbox{\rm coRL}}
\newcommand{\mbox{\rm GapL}}{\mbox{\rm GapL}}
\newcommand{\mbox{\rm Mod$_k$L}}{\mbox{\rm Mod$_k$L}}
\newcommand{\mbox{\rm FL}}{\mbox{\rm FL}}
\newcommand{\mbox{\rm NC}}{\mbox{\rm NC}}
\newcommand{\mbox{\rm TC}}{\mbox{\rm TC}}
\newcommand{\mbox{\rm HN}}{\mbox{\rm HN}}
\newcommand{\mbox{$\#$\rm L}}{\mbox{$\#$\rm L}}
\newcommand{\ensuremath{\NC^1}}{\ensuremath{\mbox{\rm NC}^1}}
\newcommand{\ensuremath{\NC^2}}{\ensuremath{\mbox{\rm NC}^2}}
\newcommand{\ensuremath{\rm RNC}}{\ensuremath{\rm RNC}}
\newcommand{\ensuremath{\rm C_{=}P}}{\ensuremath{\rm C_{=}P}}
\newcommand{\mbox{\rm\small poly}}{\mbox{\rm\small poly}}
\newcommand{\mbox{\rm End}}{\mbox{\rm End}}
\newcommand{\ensuremath{\cong} }{\ensuremath{\cong} }
\newcommand{\ensuremath{\ncong}}{\ensuremath{\ncong}}
\newcommand{\ord}[1]{\ensuremath{{\it ord}_p(#1)}}
\newcommand{\ensuremath{\it diag}}{\ensuremath{\it diag}}
\newcommand{\ensuremath{\it cont}}{\ensuremath{\it cont}}
\newcommand{\mathbb{Q}}{\mathbb{Q}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{\mathbb{Z}_{p^e}}{\mathbb{Z}_{p^e}}
\newcommand{\mathcal{M}}{\mathcal{M}}
\newcommand{\ensuremath{\mathbb{F}}}{\ensuremath{\mathbb{F}}}
\newcommand{\ensuremath{\mathcal{G}}}{\ensuremath{\mathcal{G}}}
\newcommand{\ensuremath{\mathcal{G}_n^{(t)}}}{\ensuremath{\mathcal{G}_n^{(t)}}}
\newcommand{\mbox{\bf x}}{\mbox{\bf x}}
\newcommand{\mbox{\bf y}}{\mbox{\bf y}}
\newcommand{\mbox{\bf z}}{\mbox{\bf z}}
\renewcommand{\b}{\mbox{\bf b}}
\renewcommand{\v}{\mbox{\bf v}}
\newcommand{\sym}[1]{{\rm Sym(}#1{\bf )}}
\renewcommand{\angle}[1]{\langle #1\rangle}
\newcommand{\mbox{\it acc}}{\mbox{\it acc}}
\newcommand{\mbox{\it rej}}{\mbox{\it rej}}
\newcommand{\p}[1]{\ensuremath{p_{#1}^{\alpha_#1}}}
\newcommand{\ensuremath{q=\prod_{j=1}^n p_j^{\alpha_j}}}{\ensuremath{q=\prod_{j=1}^n p_j^{\alpha_j}}}
\newcommand{\mbox{\rm Ker}}{\mbox{\rm Ker}}
\newcommand{\mbox{\rm Prob}}{\mbox{\rm Prob}}
\newcommand{\mbox{\small\rm GRAPH-ISO}}{\mbox{\small\rm GRAPH-ISO}}
\newcommand{\mbox{\small\rm CGI}}{\mbox{\small\rm C{-}GRAPH-ISO}}
\newcommand{\mbox{\small\rm k{-}LIM}}{\mbox{\small\rm k{-}LIM}}
\newcommand{\mbox{\small\rm TOUR-ISO}}{\mbox{\small\rm TOUR-ISO}}
\newcommand{\mbox{\small\rm TOUR-AUTO}}{\mbox{\small\rm TOUR-AUTO}}
\newcommand{\mbox{\small\rm AUT}}{\mbox{\small\rm AUT}}
\newcommand{\mbox{\small\rm AND}}{\mbox{\small\rm AND}}
\newcommand{\mbox{\small\rm GA}}{\mbox{\small\rm GA}}
\newcommand{\mbox{\small\rm CF}}{\mbox{\small\rm CF}}
\newcommand{\mbox{\small\rm HYPER-TOUR-ISO}}{\mbox{\small\rm HYPER-TOUR-ISO}}
\newcommand{\mbox{\small\rm PASTE}}{\mbox{\small\rm PASTE}}
\newcommand{\mbox{\small\rm T{-}CANON}}{\mbox{\small\rm T{-}CANON}}
\newcommand{\mbox{\small\rm RT{-}CANON}}{\mbox{\small\rm RT{-}CANON}}
\newcommand{\mbox{\small\rm ID}}{\mbox{\small\rm ID}}
\newcommand{\mbox{\small\rm RING-ISO}}{\mbox{\small\rm RING-ISO}}
\newcommand{\mbox{\small\rm GRAPH-ISO}}{\mbox{\small\rm GRAPH-ISO}}
\newcommand{\mbox{\rm Sym}}{\mbox{\rm Sym}}
\newcommand{\mbox{\small\rm RING-NONISO}}{\mbox{\small\rm RING-NONISO}}
\newcommand{\mbox{\small\rm IS\mbox{-}FIELD}}{\mbox{\small\rm IS\mbox{-}FIELD}}
\newcommand{\mbox{\small\rm SQUARE-FREE}}{\mbox{\small\rm SQUARE-FREE}}
\newcommand{\mbox{\small\rm NILRAD}}{\mbox{\small\rm NILRAD}}
\newcommand{\mbox{\small\rm ORDER}}{\mbox{\small\rm ORDER}}
\newcommand{\mbox{\small\rm MEMBERSHIP}}{\mbox{\small\rm MEMBERSHIP}}
\newcommand{\mbox{\small\rm ID-MEMB}}{\mbox{\small\rm ID-MEMB}}
\newcommand{\mbox{\small\rm LDC}}{\mbox{\small\rm LDC}}
\newcommand{\mbox{\small\rm\textsc{Ideal{-}Memb}}}{\mbox{\small\rm\textsc{Ideal{-}Memb}}}
\newcommand{\mbox{\small\rm MAXIMAL IDEAL}}{\mbox{\small\rm MAXIMAL IDEAL}}
\newcommand{\mbox{\small\rm PRIMARY IDEAL}}{\mbox{\small\rm PRIMARY IDEAL}}
\newcommand{{R}}{{R}}
\newcommand{\mbox{\rm I}}{\mbox{\rm I}}
\newcommand{\ensuremath{\mathbb{N}}}{\ensuremath{\mathbb{N}}}
\newcommand{\mbox{\rm A}}{\mbox{\rm A}}
\newcommand{\mbox{\rm E}}{\mbox{\rm E}}
\newcommand{\mbox{\rm Y}}{\mbox{\rm Y}}
\newcommand{\mbox{\rm X}}{\mbox{\rm X}}
\renewcommand{\mathcal{M}}{\mbox{\rm M}}
\newcommand{\mathbb{C}}{\mathbb{C}}
\renewcommand{\H}{\mbox{\rm H}}
\renewcommand{\P}{\mbox{\rm P}}
\renewcommand{\S}{\mbox{\rm S}}
\renewcommand{\O}{\mbox{\rm O}}
\newcommand{\mbox{\rm W}}{\mbox{\rm W}}
\newcommand{\ensuremath{N^{(i)}}}{\ensuremath{N^{(i)}}}
\newcommand{\mbox{\rm Aut}}{\mbox{\rm Aut}}
\newcommand{\mbox{\rm CL}}{\mbox{\rm CL}}
\newcommand{\mbox{\textrm addinv}}{\mbox{\textrm addinv}}
\newcommand{\mbox{\textrm addid}}{\mbox{\textrm addid}}
\newcommand{\mbox{\textrm Perm}}{\mbox{\textrm Perm}}
\newcommand{\mbox{\textrm bin}}{\mbox{\textrm bin}}
\newcommand{\mbox{\textrm Divide}}{\mbox{\textrm Divide}}
\newcommand{\mbox{\textrm Var}}{\mbox{\textrm Var}}
\renewcommand{\angle}[1]{\langle #1\rangle}
\newcommand{\mbox{\small\rm \textsc{Ideal Membership}}}{\mbox{\small\rm \textsc{Ideal Membership}}}
\newcommand{\mbox{\small\rm \textsc{Subset Sum}}}{\mbox{\small\rm \textsc{Subset Sum}}}
\newcommand{\mbox{\textrm{rank}}}{\mbox{\textrm{rank}}}
\title{Univariate Ideal Membership Parameterized by Rank, Degree, and Number of Generators}
\author{V. Arvind\thanks{Institute of Mathematical Sciences (HBNI), Chennai,
India, \texttt{email: <EMAIL>}} \and Abhranil Chatterjee\thanks{Institute of Mathematical Sciences (HBNI), Chennai,
India, \texttt{email: <EMAIL>}} \and Rajit Datta\thanks{Chennai Mathematical Institute, Chennai, India, \texttt{email: <EMAIL>}} \and Partha
Mukhopadhyay\thanks{Chennai Mathematical Institute, Chennai, India,
\texttt{email: <EMAIL>}}
}
\begin{document}
\maketitle
\begin{abstract}
Let $\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ be the polynomial ring over the variables $\mbox{\rm X}=\{x_1,x_2,
\ldots, x_n\}$. An ideal $I=\angle{p_1(x_1), \ldots, p_n(x_n)}$
generated by univariate polynomials $\{p_i(x_i)\}_{i=1}^n$ is a
\emph{univariate ideal}. We study the ideal membership problem for
the univariate ideals and show the following results.
\begin{itemize}
\item Let $f(\mbox{\rm X})\in\ensuremath{\mathbb{F}}[\ell_1, \ldots, \ell_r]$ be a (low rank)
polynomial given by an arithmetic circuit where $\ell_i : 1\leq
i\leq r$ are linear forms, and $I=\angle{p_1(x_1), \ldots,
p_n(x_n)}$ be a univariate ideal. Given $\vec{\alpha}\in \ensuremath{\mathbb{F}}^n$,
the (unique) remainder $f(\mbox{\rm X}) \pmod I$ can be evaluated at
$\vec{\alpha}$ in deterministic time $d^{O(r)}\cdot \mbox{\rm\small poly}(n)$,
where $d=\max\{\deg(f),\deg(p_1)\ldots,\deg(p_n)\}$. This yields an
$n^{O(r)}$ algorithm for minimum vertex cover in graphs with
rank-$r$ adjacency matrices. It also yields an $n^{O(r)}$ algorithm
for evaluating the permanent of a $n\times n$ matrix of rank $r$,
over any field $\ensuremath{\mathbb{F}}$. Over $\mathbb{Q}$, an algorithm of similar run time
for low rank permanent is due to Barvinok~\cite{Bar96} via a
different technique.
\item Let $f(\mbox{\rm X})\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ be given by an arithmetic circuit of degree
$k$ ($k$ treated as fixed parameter) and $I=\angle{p_1(x_1), \ldots,
p_n(x_n)}$. We show that
in the special case when $I=\angle{x_1^{e_1}, \ldots,
x_n^{e_n}}$, we obtain a randomized $O^*(4.08^k)$ algorithm that
uses $\mbox{\rm\small poly}(n,k)$ space.
\item Given $f(\mbox{\rm X})\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ by an arithmetic circuit and
$I=\angle{p_1(x_1), \ldots, p_k(x_k)}$, membership testing is
$\mbox{\rm W}[1]$-hard, parameterized by $k$. The problem is $\mbox{\small\rm MINI{[}1{]}}$-hard in
the special case when $I=\angle{x_1^{e_1}, \ldots, x_k^{e_k}}$.
\end{itemize}
\end{abstract}
\section{Introduction}\label{intro}
Let $R=\ensuremath{\mathbb{F}}[x_1, x_2, \ldots, x_n]$ \footnote{We often use the
shorthand notation $\ensuremath{\mathbb{F}}[\mbox{\rm X}]$. } be the ring of polynomials over the
variables $\mbox{\rm X}=\{x_1, x_2, \ldots, x_n\}$. A subring $I\subseteq R$ is
an ideal if $ I R\subseteq I$. Computationally, an ideal $I$ is often
given by generators : $I = \angle{f_1,f_2,\ldots,f_{\ell}}$. Given
$f\in R$ and $I=\angle{f_1, \ldots, f_{\ell}}$, the \emph{Ideal
Membership problem} is to decide whether $f\in I$ or not. In
general, this is computationally highly intractable. In fact, it is
$\mbox{\rm EXPSPACE}$-complete even if $f$ and the generators $f_i,i\in[\ell]$
are given explicitly by sum of monomials \cite{MM82}. Nevertheless,
special cases of ideal membership problem have played important roles
in several results in arithmetic complexity. For example, the
polynomial identity testing algorithm for depth three
$\Sigma\Pi\Sigma$ circuits with bounded top fan-in; the structure
theorem for $\Sigma\Pi\Sigma(k,d)$ identities use ideal membership
very crucially \cite{AM10, KS07, SS13}.
In this paper, our study of ideal membership is motivated by a basic
result in algebraic complexity : the Combinatorial Nullstellensatz of
Alon \cite{alon99}, and we recall a basic result in that paper.
\begin{theorem}\label{alon}
Let $\ensuremath{\mathbb{F}}$ be any field, and $f(\mbox{\rm X})\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$. Define polynomials
$g_i(x_i) = \prod_{s\in S_i}(x_i - s)$ for nonempty subsets $S_i, 1\le
i\le n$ of $\ensuremath{\mathbb{F}}$. If $f$ vanishes on all the common zeros of $g_1,
\ldots, g_n$, then there are polynomials $h_1, \ldots, h_n$ satisfying
$\deg(h_i)\leq \deg(f) - \deg(g_i)$ such that $f=\sum_{i=1}^n h_ig_i$.
\end{theorem}
The theorem can be restated in terms of ideal membership: Let $f(\mbox{\rm X})
\in\ensuremath{\mathbb{F}}[\mbox{\rm X}] $ be a given polynomial, and
$I=\angle{g_1(x_1),g_2(x_2),\ldots,g_n(x_n)}$ be an ideal generated by
univariate polynomials $g_i$ \emph{without repeated roots}. Let
$Z(g_i)$ denote the zero set of $g_i, 1\leq i \leq n$. By
Theorem~\ref{alon}, if $f\not\in I$ then there is a
$\vec{\alpha}=(\alpha_1, \ldots, \alpha_n)\in Z(g_1)\times \cdots
\times Z(g_n)$ such that $f(\vec{\alpha})\neq 0$. Of course, if $f\in
I$ then $f|_{Z(g_1)\times\cdots\times Z(g_n)}=0$.
Ideals $I$ generated by univariate polynomials are called
\emph{univariate ideals}. For any univariate ideal $I$ and any
polynomial $f$, by repeated application of the division algorithm, we
can write $f(\mbox{\rm X})=\sum_{i=1}^n h_i(\mbox{\rm X}) g_i(x_i) + R(\mbox{\rm X})$ where $R$ is
unique and for each $i\in [n] : \deg_{x_i}(R) < \deg(g_i(x_i))$. Since
the remainder is unique, it is convenient to write $R = f \mbox{\textrm mod} I$. By
Alon's theorem, if $f\not\in I$ then there is a $\vec{\alpha}\in
Z(g_1) \times \cdots \times Z(g_n)$ such that $R(\vec{\alpha})\neq 0$.
As an application of the theorem, Alon and Tarsi showed that checking
$k$-colorability of a graph $G$ is polynomial-time equivalent to
testing whether the graph polynomial $f_{G}$ is in the ideal
$\angle{x_1^{k}-1, \ldots, x_n^{k}-1}$ \cite{alon99}. It follows that
univariate ideal membership problem $\mbox{\rm coNP}$-hard.
Univariate ideal membership is further motivated by its connection
with two well-studied problems. Computing the permanent of a $n\times
n$ matrix over any field $\ensuremath{\mathbb{F}}$ can be cast in terms of univariate ideal
membership. Given a matrix $A=(a_{i,j})_{1\leq i,j\leq
n}\in\ensuremath{\mathbb{F}}^{n\times n}$, consider the product of linear forms $P_A(\mbox{\rm X})
= \prod_{i=1}^{n} (\sum_{j=1}^n a_{ij} x_j)$. The following
observation is well known.
\begin{fact}\label{perm}
The permanent of the matrix $A$ is given by the coefficient of the
monomial $x_1x_2\cdots x_n$ in $P_A$.
\end{fact}
It follows immediately that $P_A(\mbox{\rm X}) \pmod{\angle{x_1^2, \ldots, x_n^2}}=
\mbox{\textrm Perm}(A) ~x_1 x_2 \cdots x_n$. I.e., the remainder $P_A
\pmod{\angle{x_1^2, \ldots, x_n^2}}$ evaluates to $\mbox{\textrm Perm}(A)$ at the
point $\vec{1}\in\ensuremath{\mathbb{F}}^n$.
Next, we briefly mention the connection of univariate ideal membership
with the multilinear monomial detection problem, a benchmark problem
that is useful in designing fast parameterized algorithms for a host
of problems~\cite{Kou08, Kou12,KW16,RW09}.
Notice that, given an arithmetic circuit $C$ computing a polynomial
$f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ of degree $k$, checking if $f$ has a nonzero multilinear
monomial of degree $k$ is equivalent to checking if $f
\pmod{\angle{x_1^2, \ldots, x_n^2}}$ is nonzero. Moreover, the
constrained multilinear detection problem studied in~\cite{BKK16,
Kou12} can also be viewed as a problem of deciding membership in a
univariate ideal.
\vspace{-.25cm}
\subsection{Our Results}
A contribution of this paper is to consider several parameterized
problems in arithmetic complexity as instances of univariate ideal
membership. One parameter of interest is the rank of a multivariate
polynomial: We say $f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ is a \emph{rank $r$} polynomial if $f
\in \ensuremath{\mathbb{F}}[\ell_1 , \ell_2 ,\ldots,\ell_r]$ for linear forms $\ell_j :
1\leq j\leq r$. This concept has found application in algorithms for
depth-3 polynomial identity testing \cite{SS13}. Given a univariate
ideal $I$, a point $\vec{\alpha}\in\ensuremath{\mathbb{F}}^n$, and an arithmetic circuit
computing a polynomial $f$ of rank $r$, we obtain an efficient
algorithm to compute $f\pmod{I}$ at $\vec{\alpha}$.
\begin{theorem}\label{main-thm-4}
Let $\ensuremath{\mathbb{F}}$ be an arbitrary field where the field arithmetic can be done
efficiently, and $C$ be a polynomial-size arithmetic circuit computing
a polynomial $f$ in $\ensuremath{\mathbb{F}}[\ell_1,\ell_2,\ldots,\ell_r]$, where $\ell_1,
\ell_2, \ldots, \ell_r$ are given linear forms in $\{x_1,x_2,\ldots,x_n\}$.
Let $I=\angle{p_1, \dots, p_n}$ be a univariate ideal generated by
$p_i(x_i)\in\ensuremath{\mathbb{F}}[x_i], 1\le i\le n$. Given $\vec{\alpha}\in \ensuremath{\mathbb{F}}^n$, we
can evaluate the remainder $f \pmod{I}$ at the point $\vec{\alpha}$ in
time $d^{O(r)} \mbox{\rm\small poly}(n)$, where $d=\max\{\deg(f),\deg(p_i): 1\le i\le
n\}$.
\end{theorem}
This also allows us to check whether $f\in I$ by picking a point
$\vec{\alpha}$ at random and checking whether $f\pmod{I}$ evaluated at
$\vec{\alpha}$ is zero or not. The intuitive idea behind the proof of
Theorem \ref{main-thm-4} is as follows.
Given a polynomial $f(\mbox{\rm X}) \in \ensuremath{\mathbb{F}}[\ell_1, \ldots, \ell_r]$, a
univariate ideal $I=\angle{p_1(x_1), \ldots, p_n(x_n)}$, and a point
$\vec{\alpha}\in\ensuremath{\mathbb{F}}^n$, we first find an invertible linear
transformation $T$ such that the polynomial $T(f)$ becomes a
polynomial over at most $2r$ variables. Additionally $T$ has the
property that $T$ fixes the variables $x_1, \ldots, x_r$. Then we
recover the polynomial (call it $\tilde{f}$) over at most $2r$
variables explicitly and perform division algorithm with respect to
the ideal $I_{[r]}=\angle{p_1(x_1), \ldots, p_r(x_r)}$. For notational convenience, call
$\tilde{f}$ be the polynomial obtained over at most $2r$ variables.
It turns out $T^{-1}(\tilde{f})$ is the \emph{true remainder} $f
\pmod{I_{[r]}}$. Since the variables $x_1, \ldots, x_r$ do not play
role in the subsequent stages of division, we can eliminate them by
substituting $x_i \leftarrow \alpha_i$ for each $1\leq i\leq r$. Then
we apply the division algorithm on $T^{-1}(\tilde{f})|_{x_i \leftarrow
\alpha_i : 1\leq i\leq r}$ recursively with respect to the ideal
$I_{[n]\setminus[r]}$ to compute the final remainder at the point
$\vec\alpha$.
Our next result is an efficient algorithm to detect vertex cover in
low rank graphs. A graph $G$ is said to be of rank $r$ if the rank of
the adjacency matrix $A_G$ is of rank $r$. Graphs of low rank were
studied by Lovasz and Kotlov~\cite{KL96,Kot97} in the context of graph
coloring. Our idea is to construct a low rank polynomial from the
graph and check its membership in an appropriate univariate ideal.
\begin{theorem}\label{lowVC}
Given a graph $G=(V,E)$ on $n$ vertices such that the rank of the
adjacency matrix $A_{G}$ is at most $r$, and a parameter $k$, there
is a randomized $n^{O(r)}$ algorithm to decide if the graph $G$
has vertex cover of size $k$ or not.
\end{theorem}
Theorem~\ref{main-thm-4} also yields an $n^{O(r)}$ algorithm to
compute the permanent of rank-$r$ matrices over any field. Barvinok
had given~\cite{Bar96} an algorithm of same running time for the
permanent of low rank matrices (over $\mathbb{Q}$) using apolar bilinear
forms. By Fact~\ref{perm}, if matrix $A$ is rank $r$ then $P_A$ is a
rank-$r$ polynomial, and for the univariate ideal $I=\angle{x_1^2,
\ldots, x_n^2}$ computing $P_A \pmod{I}$ at the point ${\vec{1}}$
yields the permanent. Theorem~\ref{main-thm-4} works more generally
for all univariate ideals. In particular, the ideal in the proof of
Theorem \ref{lowVC} is generated by polynomials that are not powers of
variables. Thus, Theorem~\ref{main-thm-4} can potentially have more
algorithmic consequences than the technique in \cite{Bar96}.
\begin{comment}
-------
Next we give a parametric hardness result. If $k$ is the degree of the
input polynomial then univariate ideal membership is $\mbox{\small\rm XP}$-hard with
$k$ as the fixed parameter.
\begin{theorem}\label{degparhard}
Given an arithmetic circuit $C$ computing a degree $k$ polynomial $f
\in \ensuremath{\mathbb{F}}[\mbox{\rm X}]$ and $p_1(x_1),p_2(x_2),\ldots,p_n(x_n)$ be univariates,
checking if $f \in\angle{p_1(x_1),p_2(x_2),\ldots,p_n(x_n)}$ is
$\mbox{\small\rm XP}$-hard with $k$ as the fixed parameter.
\end{theorem}
----------
\end{comment}
If $k$ is the degree of the
input polynomial and the ideal is given by the powers of variables as
generators, we have a randomized $\mbox{\small\rm FPT}$ algorithm for the problem.
\begin{theorem}\label{degparalgo}
Given an arithmetic circuit $C$ computing a polynomial $f(\mbox{\rm X}) \in
\mathbb{Z}[\mbox{\rm X}]$ of degree $k$ and integers $e_1,e_2,\ldots,e_n$,
there is a randomized algorithm to decide whether $f \not\in
\angle{x^{e_1}_1,x^{e_2}_2,\ldots,x^{e_n}_n}$ in $O^*(4.08^k)$
time.
\end{theorem}
Note that this generalizes the well-known problem of \emph{multilinear
monomial detection} for which the ideal of interest would be
$I=\angle{x^2_1,x^2_2,\ldots,x^2_n}$. Surprisingly, the run time of
the algorithm in Theorem~\ref{degparalgo} is independent of the
$e_i$. Brand et al.\ have given the first FPT algorithm for
multilinear monomial detection in the case of general circuit with run
time randomized $O^*(4.32^k)$~\cite{BDH18}. Recently, this problem has
also been studied using the Hadamard product~\cite{ACDM18} of the
given polynomial with the elementary symmetric polynomial (and
differently using apolar bilinear forms~\cite{Pra18}). Our proof of
Theorem~\ref{degparalgo} shows that checking membership of $f$ in the
ideal $\angle{x^{e_1}_1,\ldots,x^{e_n}_n}$ is efficiently reducible to
computing the (scaled) Hadamard product of $f$ with a modified
elementary symmetric polynomial.
When the number of generators in the ideal is treated as the fixed
parameter, the problem is $\mbox{\rm W}[1]$-hard.
\begin{theorem}\label{genparhard}
Given a polynomial $f(\mbox{\rm X}) \in \mathbb{F}[\mbox{\rm X}]$ by an arithmetic
circuit $C$ and univariate polynomials
$p_1(x_1),p_2(x_2),\ldots,p_k(x_k)$, checking if $f \not \in
\angle{p_1(x_1),p_2(x_2),\dots,p_k(x_k)}$ is $\mbox{\rm W}[1]$-hard with $k$ as
the parameter.
\end{theorem}
Theorem~\ref{genparhard} is shown by a suitable reduction from
independent set problem to ideal membership. To find an independent
set of size $k$, the reduction produces an ideal with $k$ univariates
and the polynomial created from the graph has $k$ variables. Unlike
Theorem \ref{degparalgo}, the above parameterization of the problem
remains $\mbox{\small\rm MINI{[}1{]}}$-hard even if the ideal is generated by powers of
variables. More precisely, we show the following result.
\begin{theorem}\label{genparuna}
Let $C$ be a polynomial-size arithmetic circuit computing a polynomial
$f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$. Let $I =
\angle{{x_1}^{e_1},{x_2}^{e_2},\ldots,{x_k}^{e_k}}$ be the given ideal
where $e_1, \ldots, e_k$ are given in unary. Checking if $f\not\in I$
is $\mbox{\small\rm MINI{[}1{]}}$-hard with $k$ as parameter.
\end{theorem}
It turns out that the complement of the ideal membership problem can
be easily reduced from $\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}$ problem which asks if there is a
$\vec{x}\in\{0,1\}^n$ satisfying $A\vec{x} = \vec{b}$, where
$A\in\ensuremath{\mathbb{F}}^{k\times n}$ and $\vec{b}\in\ensuremath{\mathbb{F}}^k$.
We can show $\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}$ is hard for the parameterized complexity class
$\mbox{\small\rm MINI{[}1{]}}$ by reducing the miniature version of $\mbox{\small\rm 1-in-3 POSITIVE 3-SAT}$ to it.
As already mentioned, the result of Alon and Tarsi \cite{alon99} shows
that the membership of $f_G$ in $\angle{x_1^k-1, \ldots, x_n^k-1}$ is
$\mbox{\rm coNP}$-hard and the proof crucially uses the fact that the roots of
the generator polynomials are all distinct. This naturally raises the
question if univariate ideal membership is in $\mbox{\rm coNP}$ when each
generator polynomial has distinct roots. We show membership in
$\mbox{\rm coNP}$.
\begin{theorem}\label{main-thm-1}
Let $f\in\mathbb{Q}[\mbox{\rm X}]$ be a polynomial of degree at most $d$ given by a
black-box. Let $I=\angle{p_1(x_1), \ldots, p_n(x_n)}$ be an ideal
given explicitly by a set of univariate polynomials $p_1, p_2,
\ldots, p_n$ as generators of maximum degree bounded by $d$. Let $L$
be the bit-size upper bound for any coefficient in $f, p_1, p_2,
\ldots, p_n$. Moreover, assume that $p_i$s have distinct roots over
$\mathbb{C}$. Then there is a non-deterministic algorithm running in time
$\mbox{\rm\small poly}(n, d, L)$ that decides the non-membership of $f $ in the ideal
$I$.
\end{theorem}
\begin{remark}
The distinct roots case discussed in Theorem~\ref{main-thm-1} is in
stark contrast to the complexity of testing membership of $P_A(\mbox{\rm X})$ in
the ideal $\angle{x_1^2, \ldots, x_n^2}$. That problem is equivalent
to checking if $\mbox{\textrm Perm}(A)$ is nonzero for a rational matrix $A$, which
is hard for the exact counting class C$_{=}$P. Hence it cannot be in
$\mbox{\rm coNP}$ unless the polynomial-time hierarchy collapses.
\end{remark}
Recall from Alon's Nullstellensatz that if $f\not\in I$, then there
is always a point $\vec{\alpha}\in Z(p_1)\times\ldots\times Z(p_n)$
such that $f(\vec\alpha)\neq 0$. Notice that in general the roots
$\alpha_i \in \mathbb{C}$ and in the standard \emph{Turing Machine} model the
$\mbox{\rm NP}$ machine can not guess the roots directly with only finite
precision. But we are able to prove that the $\mbox{\rm NP}$ machine can guess
the tuple of roots $\vec{\tilde{\alpha}}\in\mathbb{Q}^n$ using only
polynomial bits of precision and still can decide the
non-membership. The main technical idea is to compute efficiently a
parameter $M$ only from the input parameters such that
$|f(\vec{\tilde{\alpha}})|\leq M$ if $f\in I$, and
$|f(\vec{\tilde{\alpha}})|\geq 2M$ if $f\not\in I$. The $\mbox{\rm NP}$ machine
decides the non-membership according to the final value of
$|f(\vec{\tilde{\alpha}})|$. We remark that Koiran has considered
the weak version of Hilbert Nullstellensatz ($\mbox{\rm HN}$) problem
\cite{koi96}. The input is a set of multivariate polynomials $f_1,
f_2, \ldots, f_m \in \mathbb{Z}[\mbox{\rm X}]$ and the problem is to decide whether
$1\in\angle{f_1, \ldots, f_m}$. The result of Koiran shows that
$\overline{\mbox{\rm HN}}\in \mbox{\rm AM}$ (under GRH), and it is an outstanding open
problem problem to decide whether $\overline{\mbox{\rm HN}}\in \mbox{\rm NP}$.
\vspace{-0.25 cm}
\subsubsection*{Organization}
In Section \ref{prelim} we give some background results. We prove
Theorem \ref{main-thm-4} and Theorem \ref{lowVC} in Section
\ref{barvinokgeneral}.
In Section~\ref{parameterized-section}, we explore the parameterized
complexity of univariate ideal membership. In the first subsection, we
prove
\ref{degparalgo}, and in the
second subsection we prove Theorems~\ref{genparhard} and
\ref{genparuna}. Finally, in Section~\ref{main-thm1-sec}, we prove
Theorem \ref{main-thm-1}. Several proofs are given in the appendix.
\section{Preliminaries}\label{prelim}
\subsubsection*{Basics of Ideal Membership}\vspace{-0.15cm}
Let $\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ be the ring of polynomials $\ensuremath{\mathbb{F}}[x_1,x_2,\ldots,x_n]$. Let
$I\subseteq \ensuremath{\mathbb{F}}[\mbox{\rm X}]$ be an ideal given by a set of generators
$I=\angle{g_1, \ldots, g_{\ell}}$. Then for any polynomial
$f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$, it is a member of the ideal if and only if
$f=\sum_{i=1}^{\ell} h_i g_i$ where $\forall i :
h_i\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$. Dividing $f$ by the $g_i$ by applying the standard
division algorithm does not work in general to check if $f\in
I$. Indeed, the remainder is not even uniquely defined. However, if
the leading monomials of the generators are already pairwise
relatively prime, then we can apply the division algorithm to compute
the unique remainder.
\begin{theorem}[See\cite{CLO07}, Theorem 3, proposition 4, pp.101]
\label{GB-Syz}
Let $I$ be a polynomial ideal given by a basis $G = \{g_1, g_2,
\cdots, g_s\}$ such that all pairs $i\neq j$ $LM(g_i)$ and $LM(g_j)$
are relatively prime. Then $G$ is a Gr\"obner basis for $I$.
\end{theorem}
In particular, if the ideal $I$ is a univariate ideal given by
$I=\angle{p_1(x_1), \ldots, p_n(x_n)}$, we can apply the division
algorithm to compute the unique remainder $f \pmod{I}$. To bound the
run time of this procedure we note the following: Let $\bar{p}$ denote
the ordered list $\{p_1,p_2,\ldots,p_n\}$. Let $\mbox{\textrm Divide}(f ; \bar{p})$
be the procedure that divides $f$ by $p_1$ to obtain remainder $f_1$,
then divides $f_1$ by $p_2$ to obtain remainder $f_2$, and so on to
obtain the final remainder $f_n$ after dividing by $p_n$. We note the
following time bound for $\mbox{\textrm Divide}(f ; \bar{p})$.
\begin{fact}[See \cite{Sudan98}, Section 6, pp.5-12]\label{div-run-time}
Let $f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ be given by a size $s$ arithmetic circuit and
$p_i(x_i)\in\ensuremath{\mathbb{F}}[x_i]$ be given univariate polynomials. The running time
of $\mbox{\textrm Divide}(f ; \bar{p})$ is bounded by $O(s\cdot \prod_{i=1}^n (d_i +
1)^{O(1)})$, where $d_i=\max\{\deg_{x_i}(f),\deg(p_i(x_i))\}$.
\end{fact}
\vspace{-0.35cm}
\subsubsection*{On Roots of Univariate Polynomials}\vspace{-0.15cm}
The following lemma shows that the absolute value of any root of a univariate polynomial can be bounded in terms of the degree and the coefficients. The result is folklore.
\begin{lemma}\label{lemma-1}
Let $f(x) = \sum_{i=0}^d a_i x^i\in\mathbb{Q}[x]$ be a univariate polynomial and $\alpha$ be a root of $f$. Then, either $\frac{|a_0|}{\sum_{i=1}^d |a_i|}\leq |\alpha|<1$ or $1\leq |\alpha|\leq d \cdot \frac{\max_i |a_i|}{|a_d|}$.
\end{lemma}
\begin{proof}
Since $\alpha$ is a root of $f$, we have that, $0=f(\alpha)=\sum_{i=0}^d a_i \alpha^i=0$, and
$\sum_{i=1}^d a_i \alpha^i = -a_0$. Then by an application of triangle inequality, we get that
$\sum_{i=1}^d |a_i| |\alpha|^i \geq |a_0|$. Now we analyse two different cases.
In the first case assume that $|\alpha| < 1$.
Observe that $|\alpha| \cdot (\sum_{i=1}^d |a_i|) \geq |a_0|$, and hence $|\alpha|\geq \frac{|a_0|}{\sum_{i=1}^d |a_i|}$.
In the second case $|\alpha|\geq 1$.
Observe that $-a_d \alpha^d = \sum_{i=0}^{d-1} a_i \alpha^i$. Then use triangle inequality to get that
$|a_d| |\alpha|^d \leq |\alpha|^{d-1} \cdot (\sum_{i=0}^{d-1} |a_i|)$. Now we get the following,
$
|\alpha| \leq \frac{\sum_{i=0}^{d-1} |a_i|} {|a_d|}\leq d \cdot \frac{\max_i |a_i|}{|a_d|}.
$
The lemma follows by combining the two cases.
\end{proof}
The next lemma shows that the separation between two distinct roots of any univariate polynomial can be lower bounded in terms of degree and the size of the coefficients. This was shown by Mahler \cite{Mah64}.
\begin{lemma}\label{lemma-2}
Let $g(x) = \sum_{i=0}^d a_i x^i \in \mathbb{Q}[x$ and $2^{-L}\leq |a_i|\leq 2^L$ (if $a_i \neq 0$). Let $\alpha, \beta$ are two distinct roots of $g$. Then $|\alpha-\beta| \geq \frac{1}{2^{O(d L)}}$.
\end{lemma}
The following lemma states that any univariate polynomial can not get a very small value (in absolute sense) on any point which is far from every root.
\begin{lemma}\label{npguess}
Let $f= \sum_{i=1}^d a_i x^i$ be a univariate polynomial with $2^{-L} \leq |a_i| \leq 2^{L}$ (if $a_i \neq 0$).
Let $\tilde{\alpha}$ be a point such that $|\tilde{\alpha} - \beta_i| \geq \delta$ for every root $\beta_i$ of $f$ then $|f(\tilde{\alpha})| \geq 2^{-L} \delta^d$.
\end{lemma}
\begin{proof}
We observe that,
$f(\tilde{\alpha}) = c \prod_{i=1}^{d} (\tilde{\alpha} - \beta_i)$.
Since $|\tilde{\alpha} - \beta_i| \geq \delta$ we get,
$|f(\tilde{\alpha})| = |c| \prod_{i=1}^{d} |\tilde{\alpha} - \beta_i| \geq 2^{-L} \delta^d$.
This completes the proof.
\end{proof}
\vspace{-0.3cm}
\subsubsection*{Parameterized Complexity Classes}\vspace{-0.15cm}
We recall some standard definitions in parameterized Complexity
\cite[ch.1,pp. 7-14]{FKLMPPS15}. We only state them informally. For a
parameterized input problem $(x,k)$ with $k$ be the parameter of
interest, we say that the problem is in $\mbox{\small\rm FPT}$ if it has an algorithm
with run time $f(k) |(x,k)|^{O(1)}$ for some computable function $f$.
A parameterized reduction \cite[def. 13.1]{FKLMPPS15} between two
problems should be computable in time $f(k) |(x,k)|^{O(1)}$, and if
the reduction outputs $(x',k')$ then $k' \leq f(k)$. A parameterized
problem is in the class $\mbox{\small\rm XP}$ if it has an algorithm with run time
$|x|^{f(k)}$ for some computable function $f$.
For the purpose of this paper, it suffices to note that a
parameterized problem $L$ is in the class $\mbox{\rm W}[1]$ if there is a
parameterized reduction from $L$ to some standard $\mbox{\rm W}[1]$-complete
problem like, e.g., the $k$-Independent set problem (more details
can be found in, e.g, \cite[def. 13.16]{FKLMPPS15}).
The complexity class $\mbox{\small\rm MINI{[}1{]}}$ consists of parameterized problems that
are miniature versions of $\mbox{\rm NP}$ problems: For $L\in\mbox{\rm NP}$, its miniature
version $\op{mini}(L)$ has instances of the form $(0^n,x)$, where $|x|\le
k\log n$, $k$ is the fixed parameter, and $x$ is an instance of
$L$. Showing $\op{mini}(L)$ to be $\mbox{\small\rm MINI{[}1{]}}$-hard under parameterized
reductions is evidence of its parameterized intractability, for it
cannot be in $\mbox{\small\rm FPT}$ assuming the Exponential Time
Hypothesis~\cite{DEFPR03}.
\section{Ideal Membership for Low Rank Polynomials}\label{barvinokgeneral}
In this section we prove Theorem \ref{main-thm-4}. Given a $r$-rank
polynomial $f$ by an arithmetic circuit, a univariate ideal $I$, and a
point $\vec{\alpha}\in\ensuremath{\mathbb{F}}^n$, we give an $n^{O(r)}$ time algorithm to
evaluate the remainder polynomial $f\pmod I$ at $\vec{\alpha}$. As
mentioned in Section \ref{intro}, an application of our result yields
an $n^{O(r)}$ time algorithm for computing the permanent of rank-$r$
matrices over any field. Barvinok \cite{Bar96}, via a different
method, had obtained an $n^{O(r)}$ time algorithm for this problem
over $\mathbb{Q}$. We also obtain an $n^{O(r)}$ time algorithm for minimum
vertex cover of low rank graphs. We first define the notion
\emph{rank} of a polynomial in $\ensuremath{\mathbb{F}}[\mbox{\rm X}]$.
\begin{definition}\label{low-rank}
A polynomial $f(\mbox{\rm X})\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ is a \emph{rank-$r$ polynomial} if there
are linear forms $\ell_1, \ell_2, \ldots, \ell_r$ such that $f(\mbox{\rm X})$ is
in the sub-algebra $\ensuremath{\mathbb{F}}[\ell_1, \ldots, \ell_r]$.
\end{definition}
For an unspecified fixed parameter $r$, we refer to rank-$r$
polynomials as \emph{low rank polynomials}.
Given $\vec{\alpha}\in\ensuremath{\mathbb{F}}^n$, a univariate ideal $I=\angle{p_1(x_1),
\ldots, p_n(x_n)}$, and a rank $r$ polynomial $f(\ell_1, \ldots,
\ell_r)$ we show how to compute $f(\ell_1, \ldots, \ell_r) \pmod I$ at
$\vec\alpha$ using a recursive procedure $\mbox{\small\rm REM}(f(\ell_1, \ldots,
\ell_r), I, \vec\alpha)$ efficiently. We introduce the following
notation. For $S\subseteq [n]$, the ideal $I_S =\angle{p_i(x_i) : i\in
[S]}$.
We first observe the following lemma which shows how to remove the
redundant variables from a low rank polynomial.
\begin{lemma}\label{varsep}
Given a polynomial $f(\ell_1,\ldots,\ell_r)$ where
$\ell_1,\ldots,\ell_r$ are linear forms in $\ensuremath{\mathbb{F}}[\mbox{\rm X}]$, there is an
invertible linear transform $T: \mathbb{F}^{n} \mapsto
\mathbb{F}^{n}$ that fixes $x_1, \ldots, x_r$ and the transformed
polynomial $T(f)$ is over at most $2r$ variables.
\end{lemma}
\begin{proof}
Write each linear form $\ell_i$ in two parts: $\ell_i = \ell_{i,1}
+ \ell_{i,2}$, where $\ell_{i,1}$ is the part over variables
$x_1,\ldots,x_r$ and $\ell_{i,2}$ is over variables
$x_{r+1},\ldots,x_n$. W.l.o.g, assume that
$\{\ell_{i,2}\}^{r'}_{i=1}$ is a maximum linearly independent
subset of linear forms in $\{\ell_{i,2}\}^{r}_{i=1}$. Let
$T:\ensuremath{\mathbb{F}}^n\rightarrow\ensuremath{\mathbb{F}}^n$ be the invertible linear map that fixes
$x_1, \ldots, x_r$, maps the independent linear forms $\{\ell_{i,2}
\}^{r'}_{i=1}$ to variables $x_{r+1},\ldots,x_{r+r'}$, and suitably
extends $T$ to an invertible map. This completes the proof.
\end{proof}
The following lemma shows that the univariate division and evaluating
the remainder at the end can be achieved by division and evaluation
partially.
\begin{lemma} \label{partial}
Let $f(\mbox{\rm X})\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ and $I = \angle{p_1(x_1), \ldots, p_n(x_n)}$ be a
univariate ideal. Let $R(\mbox{\rm X})$ be the unique remainder $f \pmod I$. Let
$\vec{\alpha}\in\ensuremath{\mathbb{F}}^r, r\leq n$ and $R_r(\mbox{\rm X}) = f \pmod{I_{[r]}}$. Then
$R(\alpha_1, \ldots, \alpha_r, x_{r+1}, \ldots, x_n) = R_r(\alpha_1,
\ldots, \alpha_r, x_{r+1}, \ldots, x_n) \pmod{I_{[n]\setminus[r]}}$.
\end{lemma}
We require the following lemma in the proof of the main result of this
section.
\begin{lemma} \label{inv}
Let $f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$, and $T : \ensuremath{\mathbb{F}}^n\rightarrow \ensuremath{\mathbb{F}}^n$ be an invertible
linear transformation fixing $x_1, \ldots, x_r$ and mapping $x_{r+1},
\ldots, x_n$ to linearly independent linear forms over $x_{r+1},
\ldots, x_n$. Write $R = f \pmod{I_{[r]}}$ and $R' = T(f)
\pmod{I_{[r]}}$. Then $R' = T(R)$.
\end{lemma}
The proofs of Lemmas \ref{partial} and \ref{inv} are given
in Section~\ref{app1} of the appendix.
\subsubsection{Proof of Theorem~\ref{main-thm-4}}
\begin{toneproof}
We now describe a recursive procedure $\mbox{\small\rm REM}$ to solve the problem. The
initial call to it is $\mbox{\small\rm REM}(f(\ell_1, \ldots, \ell_r),
I_{[n]},\vec\alpha)$. We apply the invertible linear transformation
obtained in Lemma \ref{varsep} to get the polynomial $T(f)$ over the
variables $x_1, \ldots, x_r, x_{r+1}, \ldots, x_{r+r'}$ where $r'\leq
r$.\footnote{We use $f$ to denote $f(\ell_1, \ldots, \ell_r)$.} The
polynomial $T(f)$ can be explicitly computed in time
$\mbox{\rm\small poly}(L,s,n,d^{O(r)})$. Then we compute the remainder polynomial
$f'(x_1, \ldots, x_{r + r'}) = T(f) \pmod{I_{[r]}}$ by applying the
division algorithm which runs in time $\mbox{\rm\small poly}(L,s,n,d^{O(r)})$. Next
we compute the polynomial $g=f'(\alpha_1, \ldots, \alpha_r, x_{r+1},
\ldots, x_{r+r'})$. Notice from Lemma \ref{varsep} that $T^{-1}(x_{r
+ i}) = \ell_{i,2}$ for $1\leq i\leq r'$, thus we are interested in
the polynomial $g(\ell_{1,2}, \ldots, \ell_{r',2})$. Now we
recursively compute $\mbox{\small\rm REM}(g(\ell_{1,2}, \ldots, \ell_{r',2}),
I_{[n]\setminus[r]},\vec\alpha')$ where $\vec\alpha' =
(\alpha_{r+1},\ldots,\alpha_n)$.
\subsubsection*{Correctness of the algorithm.}
Let $R(\mbox{\rm X}) = f \pmod{I_{[n]}}$ be the unique remainder polynomial. Let
$R_r(\mbox{\rm X}) = f\pmod{I_{[r]}}$ and we know that $R_r
\pmod{I_{[n]\setminus [r]}} = R$. So by Lemma \ref{partial}, to show
the correctness of the algorithm, it is enough to show that
$g(\ell_{1,2}, \ldots, \ell_{r',2}) = R_r(\alpha_1, \ldots, \alpha_r,
x_{r+1}, \ldots, x_n)$.
Following Lemma \ref{inv}, write $R'=f'(x_1,\ldots, x_r, x_{r+1},
\ldots, x_n) = T(f)\pmod{I_{[r]}}$. Then, by Lemma \ref{inv} we
conclude that $R' = T(R_r)$. It immediately follows that $R_r=
T^{-1}(R')=f'(x_1,\ldots, x_r, T^{-1}(x_{r+1}), \ldots,
T^{-1}(x_n))$. Now by definition the polynomial $g(\ell_{1,2}, \ldots,
\ell_{r',2})$ is $f'(\alpha_1, \ldots, \alpha_r, T^{-1}(x_{r+1}),
\ldots, T^{-1}(x_{r+r'}))$ which is simply $R_r(\alpha_1, \ldots,
\alpha_r, x_{r+1}, \ldots, x_n)$.
\vspace{-0.35 cm}
\subsubsection{Time complexity.}
First, suppose that the field arithmetic over $\ensuremath{\mathbb{F}}$ can be implemented
using polynomial bits. This covers all the finite fields where the
field is given by an explicit irreducible polynomial. Also, over any
such field the polynomial $T(f)$ can be explicitly computed from the
input arithmetic circuit deterministically in time
$\mbox{\rm\small poly}(L,s,n,d^{O(r)})$.
Notice that in each recursive application the number of generators in
the ideal is reduced by at least one. Furthermore, in each recursive
step we need time $\mbox{\rm\small poly}(L,s,n,d^{O(r)})$ to run the division
algorithm. This gives us a recurrence of $t(n) \leq t(n-1) +
\mbox{\rm\small poly}(L,s,n,d^{{O}(r)})$ which solves to $t(n) \leq
\mbox{\rm\small poly}(L,s,n,d^{{O}(r)})$. Over $\mathbb{Q}$, we only need to argue that the
intermediate bit-size complexity growth is only polynomial in the
input size. The proof is given in the appendix (Section \ref{app1})
which involves fairly standard argument. The rest of the argument is
exactly same.
\end{toneproof}
\subsection{Vertex Cover Detection in Low Rank Graphs}
In the Vertex Cover problem, we are given a graph $G=(V,E)$ on $n$
vertices and an integer $k$ and the question is to decide whether there is a Vertex Cover of size $k$ in $G$.
This is a classical $\mbox{\rm NP}$-complete problem. In this section we show an efficient algorithm to detect vertex cover in a graph whose adjacency matrix is of low rank.
\begin{ttwoproof}
We present a reduction from Vertex Cover problem to Univariate Ideal Membership problem that produces a polynomial whose
rank is almost same as the rank of $A_G$.
Consider the ideal $I=\angle{x^2_1 - x_1,x^2_2 - x_2,\ldots, x^2_n - x_n}$ and the polynomial
\[f = \prod^{\binom{n}{2}}_{s=1} (\vec{x} A_G \vec{x}^T - s) \cdot \prod^{n-k-1}_{t=0}\left(\sum^{n}_{i=1} x_i - t\right),\]
where $A_G$ is the adjacency matrix of the graph $G$ and $\vec{x}=(x_1,x_2,\ldots,x_n)$ is row-vector.
\begin{lemma}\label{rank-f}
The rank of the polynomial $f$ is at most $r+1$.
\end{lemma}
\begin{proof}
We note that $A_G$ is symmetric since it encodes an undirected graph. Let $Q$ be an invertible $n \times n$ matrix that diagonalizes $A_G$.
So we have $Q A_G Q^T = D$ where $D$ is a diagonal matrix with only the first $r$ diagonal elements being non-zero.
Let $\vec{y}=(y_1,y_2,\ldots,y_n)$ be another row-vector of variables. Now, we show the effect of the transform $\vec{x}\mapsto \vec{y}Q$ on the polynomial
$\vec{x}A_G \vec{x}^T$.
Clearly, $\vec{y}Q A_G Q^T \vec{y}^T = \vec{y}D\vec{y}^T$ and since there are only $r$ non-zero entries on the diagonal, the polynomial $\vec{y}D\vec{y}^T$ is over the
variables $y_1,y_2,\ldots,y_r$. Thus $g = \prod^{\binom{n}{2}}_{s=1} (\vec{x}A_G \vec{x}^T - s)$ is a rank $r$ polynomial. Also $h=\prod^{n-k-1}_{t=0}(\sum^{n}_{i=1} x_i - t)$
is a rank $1$ polynomial as there is only one linear form $\sum^{n}_{i=1} x_i$. Since $f=gh$, we conclude that $f$ is a rank $r+1$ polynomial.
\end{proof}
Now the proof of Theorem \ref{lowVC} follows from the next claim.
\begin{claim}\label{VC}
The graph $G$ has a Vertex Cover of size $k$ if and only if $f \not \in I$.
\end{claim}
\begin{claimproof}
First, observe that the set of common zeroes of the generators of the ideal $I$ is the set
$\{ 0 ,1 \}^n$. Let $S$ be a vertex cover in $G$ such that $|S|\leq k$. We will exhibit a point $\vec{\alpha}\in\{0,1\}^n$ such that $f(\vec{\alpha})\neq 0$. This will imply that
$f\not\in I$. Identify the vertices of $G$ with $\{1,2,\ldots,n\}$. Define $\vec{\alpha}(i)=0$ if and only if $i\in S$. Since $\vec{x} A_G \vec{x}^T = \sum_{(i,j)\in E_G} x_i x_j$ and $S$ is a vertex cover for $G$, it is clear that $\vec{x} A_G \vec{x}^T(\vec{\alpha})=0$. Also $(\sum_{i=1}^n x_i)(\vec{\alpha})\geq n-k$. Then clearly $f(\vec{\alpha})\neq 0$.
For the other direction, suppose that $f \not \in I$. Then by Theorem \ref{alon}, there exists
$\vec{\alpha}\in\{0,1\}^n$ such that $f(\vec{\alpha})\neq 0$. Define the set $S\subseteq [n]$ as follows. Include $i\in S$ if and only if $\vec{\alpha}(i)=0$.
Since $f(\vec{\alpha})\neq 0$, and the range of values that $\vec{x} A_G \vec{x}^T$ can take is $\{0,1,\ldots, |E|\}$, it must be the case that $\vec{x} A_G \vec{x}^T(\vec{\alpha})=0$. It implies that the set $S$ is a vertex cover for $G$. Moreover, $\prod^{n-k-1}_{t=0}(\sum^{n}_{i=1} x_i - t)(\vec{\alpha})\neq 0$ implies that $|S|\leq k$.
\end{claimproof}
The degree of the polynomial $f$ is bounded by $n^2 + n$ and from Claim~\ref{VC} we know that $f \pmod I$ is a non-zero polynomial. By Schwarz-Zippel-Demillo-Lipton~\cite{DL78, Zip79, Sch80} lemma $(f \pmod I)(\vec{\beta})$ is non-zero with high probability when $\vec{\beta}$ is chosen randomly from a small domain. Now using Theorem~\ref{main-thm-4}, we need to just compute $(f \pmod I)(\vec{\beta})$ which can be performed in $(n,k)^{O(r)}$ time.
\end{ttwoproof}
\vspace{0cm}
\section{Parameterized Complexity of Univariate Ideals} \label{parameterized-section}\vspace{-.25 cm}
We have already mentioned in Fact~\ref{perm}, that checking if the
integer permanent is zero is reducible to testing membership of a
polynomial $f(\mbox{\rm X})$ in the ideal $\angle{x_1^2, \ldots, x_n^2}$. So
univariate ideal membership is hard for the complexity class
$\ensuremath{\rm C_{=}P}$ even when the ideal is generated by powers of variables
\cite{Sal92}. In this section we study the univariate ideal
membership with the lens of parametrized complexity. The parameters we
consider are either polynomial degree or number of the generators for
the ideal.
\vspace{-.25 cm}
\subsection{Parameterized by the Degree of the Polynomial}\vspace{-.15 cm}
We consider the following: Let $I$ be a univarite ideal given by
generators and $f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$ a degree $k$ polynomial. Is checking
whether $f$ is in $I$ fixed parameter tractable (with $k$ as
the fixed parameter)?
We show that
it admits an $\mbox{\small\rm FPT}$ algorithm for the special case when
$I=\angle{x_1^{e_1},x_2^{e_2},\ldots,x_n^{e_n}}$.
\begin{comment}
-----------------
\begin{tthreeproof}
To show hardness for $\mbox{\small\rm XP}$, we will reduce $\mbox{\rm 3-CNF-TAUT}$ to the above
parameterization of ideal membership and the degree of the
polynomial turns out to be a constant. We recall that $\mbox{\rm 3-CNF-TAUT}$
is the language of all tautological formulas i.e.\ every
assignment is a satisfying assignment.
\begin{claim}\label{claim-degparhard}
If there is an $h(k)\cdot n^{g(k)}$ time algorithm for the
univariate ideal membership problem then there is a $\mbox{\rm\small poly}(n)$
time algorithm for $\mbox{\rm 3-CNF-TAUT}$.
\end{claim}
The proof is given in Section \ref{app2} of the appendix.
\end{tthreeproof}
\vspace{-.25cm}
-------------
\end{comment}
\subsubsection{Proof of Theorem \ref{degparalgo}}
The proof uses the Hadamard product of polynomials and a
connection to noncommutative computation. This builds on our
recent work \cite{ACDM18}. We include Section \ref{app2} in
the appendix to provide the background. Here, we
recall the Hadamard product of polynomials. Let
$[m]f$ denote the coefficient of the monomial $m$ in the polynomial
$f$. For $f,g \in \ensuremath{\mathbb{F}}[\mbox{\rm X}]$, their Hadamard product is defined
as $f \circ g = \sum_{m} [m]f \cdot [m]g \cdot m$. We also
need a slight variant that we call the scaled Hadamard product.
For $f,g \in\ensuremath{\mathbb{F}}[X]$, their scaled Hadamard Product is
$f \circ^{s} g = \sum_{m} m! \cdot [m]f \cdot [m]g \cdot m$, where
$m=x^{e_1}_{i_1}x^{e_2}_{i_2} \ldots x^{e_r}_{i_r}$ and $m! =e_1!\cdot
e_2!\cdots e_r!$ abusing the notation.
If one of $f,g\in\ensuremath{\mathbb{F}}[X]$ is multilinear then the scaled Hadamard
product $f\circ^s g$ coincides with the Hadamard product $f\circ g$.
\vspace{.25cm}
\begin{tfourproof}
The proof consists of following three lemmas. Firstly, given
an input instance a degree-$k$ $f(\mbox{\rm X})$ and ideal
$I=\angle{x^{e_1}_1,x^{e_2}_2,\ldots,x^{e_n}_n}$ of ideal membership,
we reduce it to computing the (scaled) Hadamard product of
$f(\mbox{\rm X})$ and a polynomial $g(\mbox{\rm X})$, where $g(\mbox{\rm X})$ is a
weighted sum of all degree $k$ monomials that are not in $I$.
Then we show that we can compute Hadamard product of any two
polynomials in time roughly linear in the product of the size
of the circuits when one of the polynomials is given by a
diagonal circuit as input. Finally the last part of the proof is
a randomized construction of a homogeneous degree $k$ diagonal
circuit of top fain-in roughly $O^{*}(4.08^k)$ that computes a
polynomial weakly equivalent \footnote{Two polynomials $f$ and $g$
are said to be weakly equivalent if they share the same set of
monomials.} to the polynomial $g$ with constant probability.
To define the polynomial $g(\mbox{\rm X})$, let $S_{m,k}$ be the elementary
symmetric polynomial of degree $k$ over $m$ variables. Set $m =
\sum_{i=1}^n (e_i - 1)$. Let $S_{m,k}$ is defined over the variable
set $\{z_{1,1},\ldots,z_{1,e_1-1},\ldots,z_{n,1},\ldots,z_{n,e_n-1}\}$. We
define $g(\mbox{\rm X})$ as the polynomial obtained from $S_{m,k}$ replacing
each $z_{i,j}$ by $x_i$.
\begin{lemma}\label{lem1-4}
Given integers $e_1,e_2,\ldots,e_n$, and a polynomial $f(\mbox{\rm X})$ of degree $k$, $f\in \angle{x^{e_1}_1,x^{e_2}_2,\ldots,x^{e_n}_n}$ if and only if $f\circ^{s} g\equiv 0$.
\end{lemma}
\begin{proof}
Suppose, $f\not\in \angle{x^{e_1}_1,x^{e_2}_2,\ldots,x^{e_n}_n}$, then $f$ must contain a degree $k$ monomial $m = x^{f_1}_{1} x^{f_2}_{2}\ldots x^{f_n}_{n}$ such that $f_i<e_i$ for each $1\leq i \leq n$. From the construction, it is clear that $g(\mbox{\rm X})$ contains $m$. Therefore, the polynomial $f\circ^{s} g$ is not identically zero. The converse is also true for the similar reason.
\end{proof}\vspace{-.5 cm}
\begin{lemma}\label{lem2-4}
Given a circuit $C$ of size $s$ computing a polynomial $g \in \ensuremath{\mathbb{F}}[\mbox{\rm X}]$ and a homogeneous degree $k$ diagonal circuit $\Sigma\wedge^{[k]}\Sigma$ circuit $D$ of size $s'$ computing $f\in\ensuremath{\mathbb{F}}[\mbox{\rm X}]$, we can obtain a circuit computing a polynomial $f\circ^s g$ in deterministic $ss'\cdot\mbox{\rm\small poly}(n,k)$ time. Furthermore, for a scalar input $\vec{a}\in\ensuremath{\mathbb{F}}^n$, we can evaluate $(f\circ^s g)(\vec{a})$ using $\mbox{\rm\small poly}(n,k)$ space.
\end{lemma}
The proof easily follows from our recent work \cite{ACDM18}. We include a self-contained proof in the appendix (Section \ref{app2}).
\begin{lemma}\label{lem3-4}
There is an efficient randomized algorithm that constructs
with constant probability a homogeneous degree $k$ diagonal
circuit $D$ of top fan-in $O^{*}(4.08^k)$ which computes a
polynomial weakly equivalent to $g$ (defined before
Lemma~\ref{lem1-4}).
\end{lemma}
\begin{proof}
To construct such a diagonal circuit $D$, we use the idea of
\cite{Pra18}. We pick a collection of colourings
$\{\zeta:[m]\to[1.5\cdot k]\}$ of size roughly $O^{*}( (\frac{e}{\sqrt{3}})^k)$ uniformly at random. For each such colouring
$\zeta_i$, we define a $\Pi^{[1.5\cdot k]}\Sigma$ formula
$P_i = \prod_{j=1}^{1.5k} (L_j+ 1)$, where $L_j = \sum_{\ell : \zeta_i(\ell) = j} x_\ell$. We say that a monomial is \emph{covered}
by a coloring $\zeta_i$ if the monomial is in $P_i$. It is easy
to see that, given any multilinear monomial of degree $k$, the probability that a random coloring will cover the monomial is roughly $(\frac{\sqrt{3}}{e})^k$.
Hence, going over such a collection of colorings of size $O^{*}((\frac{e}{\sqrt{3}})^k)$ chosen uniformly at random, with a constant probability all the multilinear terms of degree $k$ will be covered. To take the Hadamard product with a polynomial of degree $k$, we need to extract out the degree $k$ homogeneous part (say $P'_i$) from each $P_i$. Notice that, using elementary symmetric polynomial over $1.5k$ many variables $S_{1.5k,k}$, we can write $P'_i = S_{1.5k,k}(L_1,\ldots,L_{1.5k})$.
Now we use Lemma \ref{Lee} to get a diagonal $\Sigma\wedge^{[k]}\Sigma$ circuit of top fan-in roughly $\binom{1.5k}{0.5k}$ for each $P'_i$. Define $D = \sum_{i=1}^{O^{*}((\frac{e}{\sqrt{3}})^k)}P'_i$. By a direct calculation, one can obtain a diagonal circuit $D$ of top fan-in $O^{*}(4.08^k)$
which is weakly equivalent to the polynomial $S_{m,k}$.
The construction of the polynomial $g(\mbox{\rm X})$ from $S_{m,k}$ is already explained before Lemma \ref{lem1-4}.
\end{proof}
\vspace{-.2cm}
Now, given a circuit $C$ computing $f\in \ensuremath{\mathbb{F}}[\mbox{\rm X}]$ and integers $e_1,\ldots,e_n$, to decide the membership of $f$ in the ideal $I=\angle{x^{e_1}_1,\ldots,x^{e_n}_n}$, we construct a diagonal circuit $D$ from Lemma~\ref{lem3-4} and take (scaled) Hadamard product with $C$ using Lemma~\ref{lem2-4}. Following Lemma~\ref{lem1-4}, we can decide the membership of $f$ in the ideal checking the polynomial $C\circ^s D$ is identically zero or not which can be performed by random substitution using Schwartz-Zippel Lemma \cite{Sch80, Zip79}. Over $\mathbb{Z}$ the given circuit can compute numbers as large as $2^{2^{n^{O(1)}}}$. To handle this while we evaluate the circuit, we do the evaluation modulo a random polynomial bit prime. This is a standard idea.
\end{tfourproof}
\vspace{-.2cm}
\subsection{Parameterized by Number of Generators}\label{hardness}
In this section, we consider the univariate ideal membership
parameterized on the number of generators of the ideal. More
precisely, given a polynomial $f(\mbox{\rm X})$, can we obtain an $\mbox{\small\rm FPT}$
algorithm for testing membership in the univariate ideal
$\angle{p_1(x_1), \ldots, p_k(x_k)}$ parameterized by $k$? We show
that the problem is $\mbox{\rm W}[1]$-hard. Moreover, in contrast to the
previous case, we obtain $\mbox{\small\rm MINI{[}1{]}}$-hardness for a special case of the
problem when the univariate generators are just power of variables.
\begin{tfiveproof}
We show a reduction from \emph{$k$-independent set}, a well known $\mbox{\rm W}[1]$-hard problem \cite{FKLMPPS15}, to this problem. Let $G=(V,E)$ be a graph on $n$ vertices and $k$ be the size of the independent set. We identify its vertex set with the numbers $\{1,2,\ldots,n \}$ and the edges are tuples over
$[n] \times [n]$. Define the univariate ideal $I=\angle{p_1(x_1), \ldots, p_k(x_k)}$ where for each $1\leq i \leq k$, we define $p_i(x_i) = \prod^{n}_{j=1}(x_i - j)$. Now we are going to define a polynomial $f$ that uses only $k$ variables which will be used for the ideal membership problem.
First consider the polynomial $D= \prod_{1\leq i \neq j\leq k}(x_i - x_j)$.
Now we define the polynomial,
\[f= \prod_{1\leq i \neq j\leq k} \prod_{(u,v) \in E\subseteq [n]\times [n]} [(x_i - u)^2 + (x_j - v)^2] \cdot [(x_j - u)^2 + (x_i - v)^2 ].\]
The proof follows from the following claim.
\begin{claim}
$f\cdot D \not \in \angle{p_1(x_1),p_2(x_2),\dots,p_k(x_k)}$ if and only if $G$ has an independent set of size $k$.
\end{claim}
\begin{claimproof}
We use Theorem \ref{alon} to prove the claim. Let $\{ j_1,j_2,\ldots,j_k \}$ be an independent set in $G$. Notice that $(j_1, \ldots, j_k)$ is a common zero of the generators $p_1, \ldots, p_k$.
Now notice that $f \cdot D$ does not vanish at the point $(j_1, \ldots, j_k)$ as all the edges
$(j_{\ell},j_{\ell'}) : 1\leq \ell, \ell'\leq k$ are absent in the edge set $E$.
Thus there is a common root of the ideal on which $f \cdot D$ does not vanish and hence $f\cdot D \not \in \angle{p_1(x_1),p_2(x_2),\dots,p_k(x_k)}$.
Now if $f\cdot D \not \in \angle{p_1(x_1),p_2(x_2),\dots,p_k(x_k)}$ then there is a common zero
$(j_1, \ldots,j_k)$ of the ideal on which $f \cdot D$ does not vanish. Using the same argument one can easily see that $\{j_1, \ldots, j_k\}$ is an independent set in $G$.
\end{claimproof}
\end{tfiveproof}
\vspace{0cm}
\subsubsection{Proof of Theorem \ref{genparuna}}
We first show a reduction from the linear algebraic problem $\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}$
to our univariate ideal membership problem.
\begin{definition}\textbf{$\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}$}\\
\textit{Input:} Integers $k,n$ in unary, a $k\times n$ matrix $A$ with
all the entries given in unary and a $k$ dimensional vector $\vec{b}$
with all entries in unary.\\ \textit{Parameter:
k}.\\ \textit{Question:} Does there exist an $\vec{x}\in \{0,1\}^n$
such that $A\vec{x} = \vec{b}$?
\end{definition}
It turns out that $\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}$ problem is more amenable to the
$\mbox{\small\rm MINI{[}1{]}}$-hardness proof. Finally we show a reduction from $\mbox{\small\rm MINI-1-in-3 POSITIVE 3-SAT}$ to
$\mbox{\small\rm \textit{k}{-}\textsc{Lin{-}Eq}}$ to complete the proof. It is easy to observe from the standard \textit{Schaefer Reduction} \cite{Sch78} that $\mbox{\small\rm MINI-1-in-3 POSITIVE 3-SAT}$ is $\mbox{\small\rm MINI{[}1{]}}$-hard. The full proof is given in the appendix (Section \ref{app2}).
\section{Non-deterministic Algorithm for Univariate Ideal Membership}\label{main-thm1-sec}
In this section we prove Theorem \ref{main-thm-1}.
Given a polynomial $f(\mbox{\rm X})\in\mathbb{Q}[\mbox{\rm X}]$ and a univariate ideal $I=\angle{p_1(x_1), \ldots, p_n(x_n)}$ where the generators are $p_1, \ldots, p_n$, we show a non-deterministic algorithm to decide the
(non)-membership of $f$ in $I$. By Theorem \ref{alon}, it suffices to show that there is a common zero $\vec{\alpha}$ of the generators $p_1, p_2, \ldots, p_n$ such that $f(\alpha)\neq 0$.
Since in general $\vec{\alpha}\in\mathbb{C}^n$, it is not immediately clear how to guess such a common zero by a $\mbox{\rm NP}$ machine. However, we are able to show that for the $\mbox{\rm NP}$ machine it suffices to guess such an $\vec{\alpha}$ upto polynomially many bits of approximation.
We begin by proving a few technical facts which are useful for the main proof.
Write $f(\mbox{\rm X})= \sum_{i=1}^n h_i(\mbox{\rm X}) ~p_i(x_i) + R(\mbox{\rm X})$ where for all $i\in [n]$, $\deg_{x_i} (R) < \deg(p_i)$. For any polynomial $g$, let $|c(g)|$ be the maximum coefficient (in absolute value) appearing in $g$. The following lemma gives an estimate for the coefficients of the polynomials $h_1, \ldots, h_n, R$.
\vspace{-.1cm}
\begin{lemma}\label{lemma-3}
Let $2^{-L}\leq |c(f)|, |c(p_i)|\leq 2^L$. Then there is $L'=\mbox{\rm\small poly}(L, d, n)$ such that $2^{-L'}\leq |c(h_i)|, |c(R)|\leq 2^{L'}$ where $d$ is the degree upper bound for $f$, and $\{p_i : 1\leq i\leq n\}$.
\end{lemma}
\begin{proof}
The estimate on $L'$ follows implicitly from the known results~\cite{coll67}.
It can be also seen by direct computation.
Write $f(\mbox{\rm X}) = \sum_{i} f_i(x_2, \ldots, x_n) ~x_1^i$ and then divide $x_1 ^i \pmod{p_1(x_1)}$ for each $i$. The modulo computation can be done by writing $x_1^i = q_1(x_1) p(x_1) + r_1(x_1)$ with the coefficients of $q_1$ and $r_1$ are unknown. We can then solve it using standard linear algebra. In particular, one can use the Cramer's rule for system of linear equation solution. The growth of the bit-size is only $\mbox{\rm\small poly}(L,d)$. More precisely, if $c_{\max}$ is the maximum among
$|c(f)|, |c(p_1)|$, any final coefficient is at most $c_{\max}\cdot 2^{\mbox{\rm\small poly}(L,d)}$. We repeat the procedure for the other univariate polynomials one by one. The final growth on the coefficients size is
at most $\mbox{\rm\small poly}(n,L,d)$.
\end{proof}
\vspace{-.05cm}
Let $\vec \alpha = (\alpha_1, \ldots, \alpha_n) \in \mathbb{C}^n$ be such that $p_i(\alpha_i) = 0$, $1\leq i\leq n$.
From Lemma~\ref{lemma-1}, we get that $\frac{1}{2^{\hat{L}}} \leq |\alpha_i|\leq 2^{\hat{L}}$ where
$\hat{L} = \mbox{\rm\small poly}(L,d)$. Let $\tilde{\alpha}_i\in\mathbb{Q}[i]$ be an $\epsilon$-approximation of $\alpha_i$, e.g. $|\alpha_i-\tilde{\alpha}_i|\leq \epsilon$. Then we show that the absolute value of $p_i(\tilde{\alpha}_i)$ is not too far from zero.
\vspace{-.05cm}
\begin{observation}\label{obs-1}
For $1\leq i\leq n$ we have that $|p_i(\tilde{\alpha}_i)|\leq \epsilon \cdot 2^{(d L)^{O(1)}}$.
\end{observation}
\begin{proof}
Let $p_i(x_i) = c\cdot \prod_{j=1}^d (x_i - \beta_{i,j})$ and w.l.o.g assume that $\tilde{\alpha}_i$ is the approximation of the root $\beta_{i,1}$. Then $|p_i(\tilde{\alpha}_i)| \leq \epsilon \cdot |c|\cdot \prod_{j=2}^d
|\tilde{\alpha}_i - \beta_{i,j}| \leq \epsilon \cdot |c|\cdot \prod_{j=2}^d
(|\beta_{i,1} - \beta_{i,j}| + \epsilon)\leq \epsilon \cdot 2^{\mbox{\rm\small poly}(d, L)}$. The final bound follows from the bound on the roots given in Lemma \ref{lemma-1}.
\end{proof}
Since we have an upper bound on the coefficients of the polynomials $\{h_i : 1\leq i\leq n\}$ from Lemma ~\ref{lemma-3}, it follows that for $1\leq i\leq n$ we have that $|h_i(\tilde{\alpha})|\leq 2^{(n d L)^{O(1)}}$. Here we use the fact that the approximate root ${\alpha}_i$ can be trivially bounded by $2^{\hat{L}+1}$.
\vspace{-0.05cm}
\begin{tsevenproof}
If $f$ is not in the ideal $I$, by Alon's Nullstellensatz, we know that there exists a tuple $\vec\alpha = (\alpha_1, \ldots, \alpha_n) \in Z(p_1) \times \ldots \times Z(p_n)$ such that
$R(\vec\alpha) \neq 0$.
Suppose that the $\mbox{\rm NP}$ Machine guess the tuple $\vec{\tilde{\alpha}}=(\tilde{\alpha}_1, \ldots, \tilde{\alpha}_n)$ which is the $\epsilon$-approximation of the tuple $\vec \alpha=(\alpha_1, \ldots, \alpha_n)$. Using the black-box for $f$, obtain the value for $f(\vec{\tilde{\alpha}})$. Next, we show that
the value $|f(\vec{\tilde{\alpha}})|$ distinguishes between the cases $f\in I$ and $f\not\in I$. The full proof is given in the appendix (Section \ref{app3}). The proof uses Lemma \ref{lemma-3} and Observation \ref{obs-1}.
If $f\in I$, we show that $|f(\vec{\tilde{\alpha}})| \leq \epsilon\cdot 2^{(n d L)^{c_2}}$. where the constant $c_2$ is fixed by Observation \ref{obs-1} and the bounds on $|h_i(\vec{\tilde{\alpha}})|$.
If $f\not\in I$, we show that $|f(\vec{\tilde{\alpha}})| \geq \frac{1}{2^{(n d L)^{c_3}}} - \epsilon \cdot (2^{(n d L)^{c_4}} + 2^{(n d L)^{c_2}})$, for some constant $c_3$ and $c_4$.
To make the calculation precise, let $3M = \frac{1}{2^{(n d L)^{c_3}}}$ and choose
$\epsilon$ such that $\epsilon \cdot (2^{(n d L)^{c_4}} + 2^{(n d L)^{c_2}}) \leq M$.
The final implication will be $|f(\vec{\tilde{\alpha}})|\leq M$ when $f\in I$ and $|f(\vec{\tilde{\alpha}})|\geq 2M$ when $f\not\in I$. It is important to note that the parameter $M$ can be pre-computed from the input parameters efficiently.
\end{tsevenproof}
\bibliographystyle{alpha}
|
\section{Proofs: GLS estimator}
\label{S:app-GLS}
This section contains the proof of Lemma \ref{lem:trivial'} and Corollary \ref{cor_glsVH}.
\subsection{Proof of Lemma \ref{lem:trivial'}}
\label{subsec-lem-trivial'}
For a given $n$, there exists $t$ such that $n_{t-1} \leq n < n_t$. Throughout this proof, $t$ is determined by the corresponding $n$ in this way.
We consider two cases. First, when $n< n_{t-1}+m^{t-1}$, in base $m$, $n_t-n$ is represented as
\begin{equation}
\label{eq:mbase1}
n_t-n=a_{t-1} m^{t-1}+\cdots+a_{1} m +a_0,
\end{equation}
where $a_i\in\{0,1,\ldots,m-1\}$ for $0\leq i \leq t-1$, $a_{t-1}\geq 1$.
And $\hat{\mu}_n^{(\nu)}$ can be represented as
\begin{equation}
\label{eq:lemb2_1}
\hat{\mu}_n^{(\bm{\nu})}=\frac{n_{t}{\hat{\mu}_t^{(\bm{\nu})}}-\sum_{k=n+1}^{n_t}y(X_k^{(\bm{\nu})})}{n}.
\end{equation}
Note that $\{X_k^{(\bm{\nu})}:n_t-m^{t-1}+1\leq k \leq n_t\}$ form the $(t-1)$-st generation of a subtree of $\mathbb{T}$ (rooted at a child of the root $\mathbb{T}$) and let $W_{t-1}^1=\sum_{k=n_t-m^{t-1}+1}^{n_t} y(X_k^{(\nu)})$. Similarly we can determine $a_{t-1}$ such subtrees by scanning the nodes from right to left in the $t$-th generation of $\mathbb{T}$ and define accordingly $W_{t-1}^2,\ldots,W_{t-1}^{a_{t-1}}$. Next we can determine a subtree of $\mathbb{T}$ where the next $m^{t-2}$ nodes in the $t$-th generation of $\mathbb{T}$ form its $(t-2)$-nd generation. We can determine $a_{t-2}$ such subtrees by continuing to scan the nodes from right to left in the $t$-th generation of $\mathbb{T}$ and define $W_{t-2}^1,\ldots,W_{t-2}^{a_{t-2}}$.
And so on.
By \eqref{eq:mbase1},
\begin{equation*}
\sum_{k=n+1}^{n_t} y(X_k^{(\bm{\nu})}) = \sum_{k=1}^{a_{t-1}}W_{t-1}^{k}+\sum_{k=1}^{a_{t-2}}W_{t-2}^{k}+\cdots+\sum_{k=1}^{a_{0}}W_{0}^{k}.
\end{equation*}
To proceed, we need the following concentration bounds for $m^{-t}W_t^{(\bm{\nu})}$ and $\hat{\mu}_t^{(\bm{\nu})}$ (proof below).
\begin{claim}
\label{lem:trivial}
For any initial distribution $\bm{\nu}$ of $X_0$, $m^{-t}W_t^{(\bm{\nu})}\to \mathbb{E}_{\bm{\pi}}(y)$ and $\hat{\mu}_t^{(\bm{\nu})} \to \mathbb{E}_{\bm{\pi}}(y)$ in $L^2$. For any $0<\delta<1$, there exists $C>0$ such that
\begin{equation}
\label{eq:lemb1rate}
\mathbb{E}[(m^{-t}W_t^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}}(y))^2] \leq Cm^{-(1-\delta)t}, \quad
\mathbb{E}[(\hat{\mu}_t^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}} (y))^2]\leq Ctm^{-(1-\delta)t}.
\end{equation}
The constant $C$ does not depend on the initial distribution $\nu$.
\end{claim}
Then by Claim \ref{lem:trivial}, the triangle inequality, $a_{t-1}\geq 1$ and $a_l\leq m-1$ for $0\leq l \leq t-1$, one has
\begin{equation}
\label{eq:lemb2_big}
\begin{aligned}
\left\Vert\frac{\sum_{k=n+1}^{n_t} y(X_k^{(\bm{\nu})})}{n_t-n}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}&=\left\Vert\frac{\sum_{l=0}^{t-1}m^{l}\sum_{k=1}^{a_l}m^{-l}(W_{l}^{k}-\mathbb{E}_{\bm{\pi}}(y))}{a_{t-1} m^{t-1}+\cdots+a_{1} m +a_0}\right\Vert_{L^2} \\
&\leq \sum_{l=0}^{t-1}\frac{a_l m^l}{a_{t-1} m^{t-1}+\cdots+a_{1} m +a_0} Cm^{-\frac{1-\delta}{2}l}\\
&\leq C \sum_{l=0}^{t-1}\frac{(m-1) m^l}{m^{t-1}} m^{-\frac{1-\delta}{2}l} = O(m^{-\frac{1-\delta}{2}t}).
\end{aligned}
\end{equation}
For any subsequence such that $n<n_{t-1}+m^{t-1}$, $n\to\infty$ implies $t\to \infty$. As a result,
\begin{equation*}
\frac{\sum_{k=n+1}^{n_t} y(X_k^{(\bm{\nu})})}{n_t-n}\xrightarrow{L^2}\mathbb{E}_{\bm{\pi}}(y).
\end{equation*}
From Claim \ref{lem:trivial}, $\hat{\mu}_t^{(\bm{\nu})} \xrightarrow{L^2} \mathbb{E}_{\bm{\pi}}(y)$. By \eqref{eq:lemb2_1}, the triangle inequality, the fact that $n_t/n=O(1)$ and $(n_t-n)/n=O(1)$,
\begin{equation*}
\left\Vert\hat{\mu}_n^{(\nu)}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}\leq \frac{n_t}{n}\left\Vert \hat{\mu}_t^{(\nu)}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}+\frac{n_t-n}{n}\left\Vert \frac{\sum_{k=n+1}^{n_t} y(X_k^{(\nu)})}{n_t-n}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}\xrightarrow{n\to\infty} 0.
\end{equation*}
In the second case, when $n\geq n_{t-1}+m^{t-1}$, in base $m$, $n-n_{t-1}$ is represented as
\begin{equation}
\label{eq:mbase2}
n-n_{t-1}=a_{t-1} m^{t-1}+\cdots+a_{1} m +a_0,
\end{equation}
where $a_i\in\{0,1,\ldots,m-1\}$ for $0\leq i \leq t-1$, $a_{t-1}\geq 1$.
And $\hat{\mu}_n^{(\bm{\nu})}$ can be represented as
\begin{equation}
\label{eq:lemb2_2}
\hat{\mu}_n^{(\bm{\nu})}=\frac{n_{t-1}{\hat{\mu}_{t-1}^{(\bm{\nu})}}+\sum_{k=n_{t-1}+1}^{n}y(X_k^{(\bm{\nu})})}{n}.
\end{equation}
Arguing as above, we can write
\begin{equation*}
\sum_{k=n_{t-1}+1}^{n} y(X_k^{(\bm{\nu})}) = \sum_{k=1}^{a_{t-1}}W_{t-1}^{k}+\sum_{k=1}^{a_{t-2}}W_{t-2}^{k}+\cdots+\sum_{k=1}^{a_{0}}W_{0}^{k}.
\end{equation*}
Similarly to the previous case, we can prove that for any subsequence such that $n\geq n_{t-1}+m^{t-1}$, when $n\to\infty$,
\begin{equation*}
\frac{\sum_{k=n_{t-1}+1}^{n} y(X_k^{(\bm{\nu})})}{n-n_{t-1}}\xrightarrow{L^2}\mathbb{E}_{\bm{\pi}}(y),
\end{equation*}
and
\begin{equation*}
\left\Vert\hat{\mu}_n^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}\leq \frac{n_{t-1}}{n}\left\Vert \hat{\mu}_{t-1}^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}+\frac{n-n_{t-1}}{n}\left\Vert \frac{\sum_{k=n_{t-1}+1}^{n} y(X_k^{(\nu)})}{n-n_{t-1}}-\mathbb{E}_{\bm{\pi}}(y)\right\Vert_{L^2}\xrightarrow{n\to\infty} 0.
\end{equation*}
Since $\hat{\mu}_n^{(\bm{\nu})}\xrightarrow{L^2}\mathbb{E}_{\bm{\pi}}(y)$ holds for both $n<n_{t-1}+m^{t-1}$ and $n\geq n_{t-1}+m^{t-1}$ as $n\to\infty$, one finally arrives at $\hat{\mu}_n^{(\bm{\nu})}\xrightarrow{L^2}\mathbb{E}_{\bm{\pi}}(y)$, which completes the proof.
\begin{proof}[Proof of Claim \ref{lem:trivial}]
The proof is similar to the proof of Theorem \ref{thm_mut}. First,
\begin{equation*}
\mathbb{E} [(\lambda_2^t Y_{t,2}^{(i)})^2]=\mathsf{Var}(\lambda_2^t Y_{t,2}^{(i)})+(\lambda_2^t\mathbb{E} Y_{t,2}^{(i)})^2=m^{-2t}\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_2 \rangle)+(\lambda_2^tY_{0,2}^{(i)})^2.
\end{equation*}
From \eqref{eq_var}, for any $0<\delta<1$, $\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_2 \rangle)=O(m^{(1+\delta)t})$ holds for all $i\in G$. As a result,
\begin{equation}
\mathbb{E} [(\lambda_2^t Y_{t,2}^{(i)})^2]=O(m^{-(1-\delta)t}).
\end{equation}
From \eqref{eq_Wt}, one has $m^{-t}W_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)=\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} \lambda_2^t Y_{t,2}^{(i)}$ and as a result,
\begin{equation}
\label{eq:Wl2}
\mathbb{E}[(m^{-t}W_t^{(i)}-\mathbb{E}_{\bm{\pi}}(y))^2]\leq \langle \bm{y},\bm{f}_2\rangle_{\bm{\pi}}^2 \mathbb{E} [(\lambda_2^t Y_{t,2}^{(i)})^2]=O(m^{-(1-\delta)t}).
\end{equation}
Recall from \eqref{eq_as_converge} that
\begin{equation*}
\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)=\frac{m^t}{n_t} \sum_{l=0}^{t}\frac{W_l^{(i)}-m^l \mathbb{E}_{\bm{\pi}}(y)}{m^t}.
\end{equation*}
By the Cauchy-Schwarz inequality,
\begin{equation}
\label{eq:mul2}
\begin{aligned}
\mathbb{E}\left[(\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y))^2\right]&\leq \left(\frac{m^t}{n_t}\right)^2 (t+1)\sum_{l=0}^{t} m^{2(l-t)}\mathbb{E}[(m^{-l}W_t^{(i)}-\mathbb{E}_{\bm{\pi}}(y))^2]\\
&=\left(\frac{m^t}{n_t}\right)^2 (t+1) \sum_{l=0}^{t} O(m^{-2t+(1+\delta)l}) \\
&=O(tm^{-(1-\delta)t}).
\end{aligned}
\end{equation}
So there exists $C>0$ such that for all $i\in G$, $$\mathbb{E}[(m^{-t}W_t^{(i)}-\mathbb{E}_{\bm{\pi}}(y))^2]\leq Cm^{-(1-\delta)t}, \quad \mathbb{E}[(\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y))^2]\leq Ctm^{-(1-\delta)t}.$$
So for any initial distribution $\bm{\nu}$ of $X_0$, since $\sum_{i\in G}\nu_i=1$,
\begin{equation*}
\mathbb{E}[(\hat{\mu}_t^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}} (y))^2]=\sum_{i\in G} \nu_i \mathbb{E}[(\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y))^2]\leq Ctm^{(1-\delta)t},
\end{equation*}
\begin{equation*}
\mathbb{E}[(m^{-t}W_t^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}}(y))^2] =\sum_{i\in G} \nu_i \mathbb{E}[(m^{-t}W_t^{(i)}-\mathbb{E}_{\bm{\pi}}(y))^2]\leq Cm^{-(1-\delta)t}.
\end{equation*}
So $m^{-t}W_t^{(\bm{\nu})}\to \mathbb{E}_{\bm{\pi}}(y)$ and $\hat{\mu}_t^{(\bm{\nu})} \to \mathbb{E}_{\bm{\pi}}(y)$ in $L^2$.
\end{proof}
\subsection{Proof of Corollary \ref{cor_glsVH}}
By the definition of the GLS estimator with VH adjustment in Section \ref{sec_ipw},
$$\hat{\mu}_{\mathsf{GLS,VH},t}^{(\bm{\nu})}= H_t\cdot \sum_{\sigma \in \mathbb{T}, \abs{\sigma} \leq t} w^*_{\sigma,t} \frac{y(X_\sigma^{(\nu)})}{\mathrm{deg}(X_\sigma^{(\nu)})}.$$
$H_t^{-1}$ is the GLS estimator of $\mathbb{E}_{\bm{\pi}}(y^\prime)$ where $y^\prime(i)=\mathrm{deg}(i)^{-1}$. So $H_t^{-1}$ converges to $\mathbb{E}_{\bm{\pi}} (y^\prime)$ in distribution (thus in probability).
Additionally,
$$\hat{\mu}^{\prime\prime}_{\mathsf{GLS},t}= \sum_{\sigma \in \mathbb{T}, \abs{\sigma} \leq t} w^*_{\sigma,t} \frac{y(X_\sigma^{(\bm{\nu})})}{\mathrm{deg}(X_\sigma^{(\bm{\nu})})}$$
is the GLS estimator of $\mathbb{E}_{\bm{\pi}} (y^{\prime\prime})$ where $y^{\prime\prime}(i)=y(i)/\mathrm{deg}(i)$. Then
$$\sqrt{n_t} \left[\hat{\mu}^{\prime\prime}_{\mathsf{GLS},t}-\mathbb{E}_{\bm{\pi}}(y^{\prime\prime})\right]\xrightarrow{d} \mathcal{N}\left(0,\frac{1+\lambda_2}{1-\lambda_2}\mathsf{Var}_{\bm{\pi}}(y^{\prime\prime})\right).$$
By Slutsky's theorem,
$$\sqrt{n_t}\left[\hat{\mu}_{\mathsf{GLS,VH},t}^{(\bm{\nu})}-\frac{\mathbb{E}_{\bm{\pi}}(y^{\prime\prime})}{\mathbb{E}_{\bm{\pi}}(y^\prime)}\right] \xrightarrow{d} \mathcal{N}\left(0,\frac{1+\lambda_2}{1-\lambda_2}\mathbb{E}_{\bm{\pi}}(y^\prime)^{-2}\mathsf{Var}_{\bm{\pi}}(y^{\prime\prime})\right).$$
Notice that $\mathbb{E}_{\bm{\pi}}(y^\prime)=N/\mathrm{vol}(G)$ and $\mathbb{E}_{\bm{\pi}}(y^{\prime\prime})=\sum_i y(i)/\mathrm{vol}(G)$, this gives the result
$$\sqrt{n_t}\left[\hat{\mu}_{\mathsf{GLS,VH},t}^{(\bm{\nu})}-\mu_{\mathsf{true}}\right] \xrightarrow{d} \mathcal{N}\left(0,\frac{1+\lambda_2}{1-\lambda_2}\mathbb{E}_{\bm{\pi}}(y^\prime)^{-2}\mathsf{Var}_{\bm{\pi}}(y^{\prime\prime})\right).$$
\section{Proof of Equation \eqref{eq:blockmodel}}
\label{section_proof_eq2}
In this section, we use $Z$ and $z$ in place of $B$ and $b$.
First we use mathematical induction to show that, for every $n\in\mathbb{Z}^+$, the following statement $P(n)$ holds:
For any given referral tree $\mathbb{T}$ with $n$ vertices $\{\sigma_1,\ldots,\sigma_n\}$, for any initial distribution $\bm{\nu}$ of $X_0$ and $z_1,\ldots,z_n\in\{1,\ldots,k\}$, the following holds
\begin{equation}
\label{eq:math_induction}
\mathbb{P}(Z_{\sigma_1}^{(\bm{\mu})}=z_1,\ldots,Z_{\sigma_n}^{(\bm{\mu})}=z_n)=\mathbb{P}(z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_n}^{(\bm{\nu})})=z_n)
\end{equation}
with $\mu_j=\sum_{i\in G:z(i)=j} \nu_i$ for $j=1,\ldots,k$.
Base case: We prove that $P(1)$ holds. Since $\mathbb{T}$ only contains the seed vertex $0$, it suffices to show that $\mathbb{P}(Z_0^{(\bm{\mu})}=z_0)=\mathbb{P}(z(X_0^{(\bm{\nu})})=z_0)$ for any $z_0\in\{1,\ldots,k\}$. However,
\begin{equation*}
\mathbb{P}(z(X_0^{(\bm{\nu})})=z_0)=\sum_{i\in G: z(i)=z_0} \nu_i=\mu_{z_1}=\mathbb{P}(Z_0^{(\bm{\mu})}=z_0).
\end{equation*}
So $P(0)$ is true.
Inductive step: We prove that if $P(n-1)$ holds for some unspecified value of $n\geq 2$, then $P(n)$ also holds. Assume $\sigma_{n}$ is a leaf node (i.e. $\sigma_{n}$ has no descendant) and $\sigma_{n-1}$ is the parent of $\sigma_{n}$. Then $\mathbb{T}\setminus\{\sigma_{n}\}$ is a referral tree with $n-1$ vertex. By the Markov property,
\begin{equation*}
\begin{aligned}
&\mathbb{P}(z(X_{\sigma_n}^{(\bm{\nu})})=z_n\mid z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_{n-1}}^{(\bm{\nu})})=z_{n-1})\\
= &\frac{\sum_{i\in G:z(i)=z_{n-1}} \mathbb{P}(z(X_{\sigma_n}^{(\bm{\nu})})=z_n\mid X_{\sigma_{n-1}}^{(\bm{\nu})}=i)\mathbb{P}(X_{\sigma_{n-1}}^{(\bm{\nu})}=i\mid z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_{n-2}}^{(\bm{\nu})})=z_{n-2})}{\mathbb{P}(z(X_{\sigma_{n-1}}^{(\bm{\nu})})
=z_{n-1}\mid z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_{n-2}}^{(\bm{\nu})})=z_{n-2})}\\
=&\frac{\sum_{i\in G:z(i)=z_{n-1}}\mathcal{P}_{z(i)z_n}\mathbb{P}(X_{\sigma_{n-1}}^{(\bm{\nu})}=i\mid z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_{n-2}}^{(\bm{\nu})})=z_{n-2})}{\mathbb{P}(z(X_{\sigma_{n-1}}^{(\bm{\nu})})
=z_{n-1}\mid z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_{n-2}}^{(\bm{\nu})})=z_{n-2})}
=\mathcal{P}_{z_{n-1}z_n}\\
=&\mathbb{P}(Z_{\sigma_n}^{(\bm{\mu})}=z_n\mid Z_{\sigma_1}^{(\bm{\mu})}=z_1,\ldots,Z_{\sigma_{n-1}}^{(\bm{\mu})}=z_{n-1}).
\end{aligned}
\end{equation*}
Additionally, the induction hypothesis that $P(n-1)$ holds gives
\begin{equation*}
\mathbb{P}(Z_{\sigma_1}^{(\bm{\mu})}=z_1,\ldots,Z_{\sigma_{n-1}}^{(\bm{\mu})}=z_{n-1})=\mathbb{P}(z(X_{\sigma_1}^{(\bm{\nu})})=z_1,\ldots,z(X_{\sigma_{n-1}}^{(\bm{\nu})})=z_{n-1}).
\end{equation*}
The above two equations give \eqref{eq:math_induction}, thereby showing $P(n)$ is true.
Since both the base case and the inductive step have been performed, by mathematical induction the statement $P(n)$ holds for all $n\in\mathbb{Z}^+$.
Finally we prove \eqref{eq:blockmodel} based on the above result. Assume $\mathbb{T}$ has $n$ vertices. For any $\{\sigma_{i_1}, \ldots, \sigma_{i_s}\}\subset\mathbb{T}$ and $z_{i_1},\ldots,z_{i_s}\in\{1,\ldots,k\}$, let $\{\sigma_{j_1}, \ldots, \sigma_{j_{n-s}}\}=\mathbb{T}\setminus\{\sigma_{i_1}, \ldots, \sigma_{i_s}\}$. Then
\begin{equation*}
\begin{aligned}
\mathbb{P}(Z_{\sigma_{i_1}}^{(\cdot)}=
z_{i_1},\ldots,Z_{\sigma_{i_s}}^{(\cdot)}=z_{i_s})&=\sum_{z_{j_1}=1}^{k}\cdots\sum_{z_{j_{n-s}}=1}^{k} \mathbb{P}(Z_{\sigma_1}^{(\cdot)}=z_1,\ldots,Z_{\sigma_n}^{(\cdot)}=z_n)
\\
&=\sum_{z_{j_1}=1}^{k}\cdots\sum_{z_{j_{n-s}}=1}^{k} \mathbb{P}(z(X_{\sigma_1}^{(\cdot)})=z_1,\ldots,z(X_{\sigma_n}^{(\cdot)})=z_n)\\
&=\mathbb{P}(z(X_{\sigma_{i_1}}^{(\cdot)})=z_{i_1},\ldots,z(X_{\sigma_{i_s}}^{(\cdot)})=z_{i_s}).
\end{aligned}
\end{equation*}
\section{Proofs: sample average}
\label{S:app-sample}
Define the mean matrix $\bm{M}\in\mathbb{R}^{N\times N}$ as
$$\bm{M}=\{\mathbb{E} Z_{1,j}^{(i)}:i,j=1,\ldots,N\}.$$
Let $\bm{V}_i$ denote the variance-covariance matrix of $\bm{Z}_1^{(i)}$, and define $$\bm{C}_t^{(i)}=\{\mathbb{E} Z_{t,j}^{(i)} Z_{t,k}^{(i)}: j,k=1,\ldots,N\}.$$
All components of $\bm{M}$ and $\bm{C}_t^{(i)}$ are finite. The following lemma is a standard result of multitype Galton-Watson process, see e.g. \cite{harris2002theory} or \cite{athreya2004branching}.
\begin{lemma}
The expectation of $\bm{Z}_t^{(i)}$ and $\bm{C}_t^{(i)}$ can be calculated from
\begin{equation}
\label{E_general}
\mathbb{E} (\bm{Z}_t^{(i)})=\bm{Z}_0^{(i)} \bm{M}^t, \quad \Rm{and}
\end{equation}
\begin{equation}
\label{var_iteration}
\bm{C}_t^{(i)}=(\bm{M}^\top)^t \bm{C}_0^{(i)} \bm{M}^t+\sum_{l=1}^{t} (\bm{M}^\top)^{t-l} \left(\sum_{k=1}^{N}\bm{V}_k\mathbb{E} Z_{l-1,k}^{(i)}\right) \bm{M}^{t-l}.
\end{equation}
\end{lemma}
\subsection{Proof of Lemma \ref{lem:martingale}}
\label{subsec-lem-martingale}
For the Markov model, $\bm{M}=m\bm{P}$ so $\bm{f}_j$ is the eigenvector of $\bm{M}$ corresponding to the eigenvalue $m\lambda_j$. The following lemma comes from the well-established theory of multitype Galton-Watson process.
\begin{lemma}
\label{lem-branching}
Let $\bm{\xi}$ be a right eigenvector of $\bm{M}$ and $\lambda$ be the corresponding eigenvalue. Then
$$\lambda^{-t}\langle \bm{Z}_t^{(i)}, \bm{\xi} \rangle$$
is a (complex-valued) martingale adapted to $\mathcal{F}_t=\sigma(\bm{Z}_l^{(i)}:1\leq l \leq t)$.
\end{lemma}
\begin{proof}
See Theorem 4' on Page 196 of \cite{athreya2004branching}.
\end{proof}
According to Lemma \ref{lem_eigendecomposition}, all $\lambda_j$ and $\bm{f}_j$ are real. One then applies Lemma \ref{lem-branching} to the Markov model with $\lambda=m\lambda_j$ and $\bm{\xi}=\bm{f}_j$ to complete the proof.
\subsection{Proof of Lemma \ref{lem:martingale-convergence}}
\label{subsec-lem-martingale-converence}
The next theorem is the martingale $L^p$ convergence theorem (see e.g. \cite{durrett2019probability}).
\begin{theorem}
\label{martingale_convergence_thm}
If $X_n$ is a martingale with $\sup_n \mathbb{E} \abs{X_n}^p<\infty$ where $p>1$, then $X_n \to X$ almost surely and in $L^p$.
\end{theorem}
It is essential to derive the variance of $\langle \bm{Z}_t^{(i)}, f_j \rangle$ before applying Theorem \ref{martingale_convergence_thm} to the martingales $Y_{t,j}^{(i)}, j\geq 2$. We conclude the result in the following claim and defer the proof to the end of this section.
\begin{claim}
\label{var_thm}
The variance of $\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle$ is
\begin{equation}
\label{eq_var}
\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle)=\begin{cases}
O((m\lambda_j)^{2t}) & \Rm{if} \ m\lambda_j^2 >1, \\
O(t(m\lambda_j)^{2t}) & \Rm{if} \ m\lambda_j^2=1, \\
O(m^t) & \Rm{if} \ m\lambda_j^2 <1. \\
\end{cases}
\end{equation}
\end{claim}
We begin with $Y_2^{(i)}$. By Theorem \ref{martingale_convergence_thm}, we only need to show $\sup_t \mathbb{E} (Y_{t,2}^{(i)}) ^2<\infty$. However,
\begin{equation*}
\mathbb{E} (Y_{t,2}^{(i)})^2=\mathsf{Var}(Y_{t,2}^{(i)})+(\mathbb{E} Y_{t,2}^{(i)})^2=(m\lambda_2)^{-2t}\,\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_2 \rangle)+(Y_{0,2}^{(i)})^2.
\end{equation*}
Since $m>\lambda_2^{-2}$, by Claim \ref{var_thm}, $\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_2 \rangle)=O((m\lambda_2)^{2t})$. This gives $\sup_t \mathbb{E} (Y_{t,2}^{(i)}) ^2<\infty$.
We move on to $Y_j^{(i)}$ for $j\geq 3$. By Theorem \ref{var_thm},
\begin{equation*}
\begin{aligned}
\mathbb{E} [(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)}]^2
&=(\lambda_2^{-1}\lambda_j)^{2t}[\mathsf{Var}(Y_{t,2}^{(i)})+(\mathbb{E} Y_{t,2}^{(i)})^2]\\
&=(m\lambda_2)^{-2t}\,\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle)+(\lambda_2^{-1}\lambda_j)^{2t} (Y_{0,2}^{(i)})^2.
\end{aligned}
\end{equation*}
Since $\lambda_2^{-1}\lambda_j<1$, $(\lambda_2^{-1}\lambda_j)^{2t} (Y_{0,2}^{(i)})^2\to 0$. Additionally, for $j\geq 3$,
\begin{equation*}
(m\lambda_2)^{-2t}\,\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle)=\begin{cases}
O((\lambda_2^{-1}\lambda_j)^{2t}) & \Rm{if} \ m\lambda_j^2 >1 \ \\
O(t(\lambda_2^{-1}\lambda_j)^{2t}) & \Rm{if} \ m\lambda_j^2=1 \ \\
O((m\lambda_2^2)^{-t}) & \Rm{if} \ m\lambda_j^2 <1 \ \\
\end{cases}
\end{equation*}
which converges to $0$ in all cases.
Thus $\mathbb{E} [(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)}]^2 \to 0$, which leads to $(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)} \to 0$ in $L^2$.
To prove almost sure convergence, let $\delta=\max\{(\lambda_2^{-1} \lambda_j)^2,m\lambda_2^{-2}\}\in(0,1)$. There exists $C>0$ such that
\begin{equation*}
\mathbb{E} [(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)}]^2\leq Ct\delta^t
\end{equation*}
always holds. Then $\forall\epsilon>0$,
\begin{equation*}
\mathbb{P}(|(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)}|>\epsilon)\leq \epsilon^{-2}\mathbb{E} [(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)}]^2\leq \epsilon^{-2} Ct\delta^t.
\end{equation*}
So
\begin{equation*}
\sum_{t=1}^{\infty} \mathbb{P}(|(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)}|>\epsilon)\leq\epsilon^{-2} C \sum_{t=1}^{\infty} t\delta^t<\infty.
\end{equation*}
By the Borel-Cantelli lemma, $(\lambda_2^{-1}\lambda_j)^tY_{t,j}^{(i)} \to 0$ almost surely.
\begin{proof}[Proof of Claim \ref{var_thm}]
From Lemma \ref{E_general} and the fact that $\bm{C}_0^{(i)}=(\bm{Z}_0^{(i)})^\top \bm{Z}_0^{(i)}$,
\begin{equation*}
\mathsf{Var}(\bm{Z}_t^{(i)})=\bm{C}_t^{(i)}-(\bm{M}^\top)^t (\bm{Z}_0^{(i)})^\top \bm{Z}_0^{(i)} \bm{M}^t=\sum_{l=1}^{t} (\bm{M}^\top)^{t-l} \left(\sum_{k=1}^{N}\bm{V}_k\mathbb{E} Z_{l-1,k}^{(i)}\right) \bm{M}^{t-l}.
\end{equation*}
As a result,
\begin{equation*}
\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle)=\sum_{l=1}^{t} \bm{f}_j^\top (\bm{M}^\top)^{t-l} \left(\sum_{k=1}^{N}\bm{V}_k\mathbb{E} Z_{l-1,k}^{(i)}\right) \bm{M}^{t-l} \bm{f}_j.
\end{equation*}
Since $\bm{f}_j$ is the eigenvector of $\bm{M}$ corresponding to the eigenvalue $m\lambda_j$, for every $n\in\mathbb{Z}^+$, $\bm{M}^n \bm{f}_j=(m\lambda_j)^n \bm{f}_j$. This yields
\begin{equation}
\label{eq_innerproduct}
\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle)=\sum_{l=1}^{t} (m\lambda_j)^{2t-2l} \sum_{k=1}^{N} (\bm{f}_j^\top \bm{V}_k \bm{f}_j) \mathbb{E} Z_{l-1,k}^{(i)}.
\end{equation}
Notice that $\sum_{k=1}^{N} \mathbb{E} Z_{l-1,k}=m^{l-1}$,
\begin{equation*}
\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_j \rangle) \leq c\sum_{l=1}^{t} (m\lambda_j)^{2t-2k} m^k=c(m\lambda_j)^{2t} \sum_{l=1}^{t} (m\lambda_j^2)^{-l},
\end{equation*}
where $c=\max\{\bm{f}_j^\top \bm{V}_k \bm{f}_j:1\leq j,k\leq N\}$.
This gives \eqref{eq_var}.
\end{proof}
\subsection{Proof of Lemma \ref{lem:nondegenerate-Y2}}
\label{subsec-lem-Y2}
By Lemma \ref{lem:martingale} and \ref{lem:martingale-convergence}, $\mathbb{E} Y_2^{(i)}=Y_{0,2}^{(i)}=f_2(i)$ and
\begin{equation*}
\mathsf{Var}(Y_2^{(i)})=\lim_{t\to\infty}\mathsf{Var}(Y_{t,2}^{(i)})=\lim_{t\to\infty}(m\lambda_2)^{-2t}\mathsf{Var}(\langle \bm{Z}_t^{(i)},\bm{f}_2 \rangle).
\end{equation*}
By \eqref{eq_innerproduct}
\begin{equation*}
(m\lambda_2)^{-2t}\,\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_2 \rangle)=\sum_{l=1}^{t} (m\lambda_2)^{-2l} \sum_{k=1}^{N} (\bm{f}_2^\top \bm{V}_k \bm{f}_2) \mathbb{E} Z_{l-1,k}^{(i)}.
\end{equation*}
Notice that $\bm{V}_k=m \left(\mathsf{diag}\{P_{k1},\ldots,P_{kN}\}- \bm{P}_k \bm{P}_k^T\right)$, where $\bm{P}_k^\top=(P_{k1},\ldots,P_{kN})$ is the $k$-th row of $\bm{P}$. Notice that $\sum_{j=1}^{N}P_{kj}=1$, by the Jensen's inequality
\begin{equation}
\label{eq:use_jensen}
\bm{f}_2^\top \bm{V}_k \bm{f}_2= m \sum_{j=1}^{N} P_{kj}f_2(j)^2- m \left(\sum_{j=1}^{N} P_{kj} f_2(j)\right)^2\geq 0
\end{equation}
for any $k=1,\ldots,N$. The assumptions $\langle \bm{f}_1,\bm{f}_2\rangle_\pi=0$ and $\bm{f}_1=\textbf{1}$ imply that $f_2$ is not a constant vector, thus the equality in \eqref{eq:use_jensen} does not hold. Similar to the proof of Theorem \ref{var_thm},
\begin{equation*}
(m\lambda_2)^{-2t}\,\mathsf{Var}(\langle \bm{Z}_t^{(i)}, \bm{f}_2 \rangle) \geq c\sum_{l=1}^{t} (m\lambda_2)^{-2l} m^k=c\sum_{l=1}^{t} (m\lambda_2^2)^{-l}
\end{equation*}
where $c=\min\{\bm{f}_2^\top \bm{V}_k \bm{f}_2:1\leq k\leq N\}>0$. Since $m\lambda_2^{2}>1$, this yields $\mathsf{Var}(Y_2^{(i)})>0$ for any $i=1,\ldots,N$.
\subsection{Proof of Corollary \ref{cor_decay}}
\label{subsec-proof-cor-decay}
The $L^2$ convergence in Theorem \ref{thm_mut} implies $L^1$ convergence. If a sequence of random variables $X_n\xrightarrow{L^1} X$, then $\abs{\mathbb{E}(X_n-X)}\leq \mathbb{E}\abs{X_n-X}$ implies $\mathbb{E} X_n \to \mathbb{E} X$. So
\begin{equation*}
\lim_{t\to\infty}\mathbb{E}\left(\lambda_2^{-t}[\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)]\right) = \mathbb{E} X^{(i)}=\frac{(m-1)\lambda_2}{m\lambda_2-1} \langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} f_2(i).
\end{equation*}
Since $\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} f_2(i)\neq 0$, the bias term decays like
\begin{equation*}
\left(\mathbb{E} (\hat{\mu}_t^{(i)})-\mathbb{E}_{\bm{\pi}} (y)\right)^2=\Theta(\lambda_2^{2t}).
\end{equation*}
Additionally, the $L^2$ convergence in Theorem \ref{thm_mut} also yields
$$\lambda_2^{-2t}\,\mathsf{Var}(\hat{\mu}_t^{(i)})=\mathsf{Var} (\lambda_2^{-t}[\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)]) \xrightarrow{t\to\infty} \mathsf{Var}(X^{(i)})>0.$$
So the variance term decays like $\mathsf{Var}(\hat{\mu}_t^{(i)})=\Theta(\lambda_2^{2t})$.
\subsection{Proof of Corollary \ref{cor_VH}}
By the definition of the VH estimator in Section \ref{sec_ipw},
$$\hat{\mu}^{(i)}_{\mathsf{VH},t}= H_t \cdot \frac{1}{n_t} \sum_{\sigma \in \mathbb{T}, \abs{\sigma}\leq t} \frac{y(X^{(i)}_\sigma)}{\mathrm{deg}(X^{(i)}_\sigma)}, \quad \text{where} \ H_t^{-1} = \frac{1}{n_t} \sum_{\sigma \in \mathbb{T}} \frac{1}{\mathrm{deg}(X_{\sigma}^{(i)})}.$$
$H_t^{-1}$ is the sample average of $y^\prime(X^{(i)}_\sigma)$'s up to generation $t$, where $y^\prime(j)=\mathrm{deg}(j)^{-1}$. In view of Theorem \ref{thm_mut}, $H_t^{-1}$ converges to $\mathbb{E}_{\bm{\pi}} (y^\prime)>0$ almost surely. Additionally,
$$\hat{\mu}_t^{\prime\prime} =\frac{1}{n_t} \sum_{\sigma \in \mathbb{T}, \abs{\sigma}\leq t} \frac{y(X_\sigma^{(i)})}{\mathrm{deg}(X_\sigma^{(i)})}$$
is the sample average of $y^{\prime\prime}(X_\sigma^{(i)})$'s up to generation $t$, where $y^{\prime\prime}(j)=y(j)/\mathrm{deg}(j)$. By Theorem \ref{thm_mut}, there exists some random variable $\bar{X}^{(i)}\in L^2$ such that $\lambda_2^{-t} [\hat{\mu}_t ^{\prime\prime}-\mathbb{E}_{\bm{\pi}} (y^{\prime\prime})]\to \bar{X}^{(i)}$ almost surely and in $L^2$. So
$$\lambda_2^{-t}\left[\hat{\mu}^{(i)}_{\mathsf{VH},t}-\frac{\mathbb{E}_{\bm{\pi}}(y^{\prime\prime})}{\mathbb{E}_{\bm{\pi}}(y^\prime)}\right] \to \mathbb{E}_{\bm{\pi}} (y^\prime)^{-1} \bar{X}^{(i)}\triangleq \tilde{X}^{(i)}$$
almost surely. Notice that $\mathbb{E}_{\bm{\pi}}(y^\prime)=N/\mathrm{vol}(G)$ and $\mathbb{E}_{\bm{\pi}}(y^{\prime\prime})=\sum_i y(i)/\mathrm{vol}(G)$, this gives the result that
$$\lambda_2^{-t}\left[\hat{\mu}^{(i)}_{\mathsf{VH},t}-\mu_{\mathsf{true}}\right]\to \tilde{X}^{(i)}$$
almost surely. The mean and variance of $\tilde{X}^{(i)}$ comes directly from Theorem \ref{thm_mut}.
\section{Background and notation}
\label{section_background}
This section (i) defines the Markov model, (ii) illustrates how this model is particularly tractable when the underlying network is a Blockmodel \citep{White1976Social}, and (iii) defines the IPW, VH, and GLS estimators.
\subsection{Markov model}
\label{subsec_markovmodel}
The Markov model consists of (1) a social network represented as a graph, (2) a Markov transition matrix on the nodes of the graph, (3) a referral tree to index the Markov process on the graph, and finally, (4) a node feature defined for each node in the graph. Each of these are defined below.
The results in this paper allow for an undirected, weighted graph. Let $G=(V,E)$ be a graph with vertex set $V=\{1,\ldots,N\}$ containing the people and edge set $E=\{(i,j):i,j\in V \ \Rm{are} \ \Rm{connected} \}$ containing the friendships.
Let $w_{ij}$ be the weight of the edge $(i,j)\in E$. For notational convenience, define $w_{ij}=0$ if $(i,j)\notin E$. If the graph is unweighted, define $w_{ij}=1$ for all $(i,j)\in E$. Throughout this paper, the graph is undirected (i.e. $w_{ij}=w_{ji}$ for all pairs $(i,j)$). Define the degree of node $i$ as $\mathrm{deg}(i)=\sum_{j} w_{ij}$ and the volume of the graph as $\mathrm{vol}(G)=\sum_{i} \mathrm{deg}(i)$. For simplicity, $i\in G$ is used synonymously with $i \in V$. Define the Markov transition matrix $\bm{P}\in\mathbb{R}^{N\times N}$ as
\begin{equation}
\label{eq_tran_matrix}
P_{ij}=\frac{w_{ij}}{\mathrm{deg}(i)}.
\end{equation}
Since $G$ is undirected, $\bm{P}$ is a reversible Markov transition matrix with a stationary distribution $\bm{\pi}: G\rightarrow \mathbb{R}$ with $\pi(i)=\mathrm{deg}(i)/\mathrm{vol}(G)$.
The referral tree is a rooted tree, i.e. a connected graph with $n$ nodes, no cycles, and a vertex $0$. This tree, $\mathbb{T}$, can be random (a Galton-Watson tree with expected offspring number $m$) or nonrandom (an $m$-tree, where each node has exactly $m$ offspring). If $\mathbb{T}$ is randomly generated, then the Markov process is conditioned on the tree. For simplicity, $\sigma\in\mathbb{T}$ is used synonymously with $\sigma$ belonging to the vertex set of $\mathbb{T}$. The seed participant is the root vertex $0$ in $\mathbb{T}$. For each non-root node $\sigma\in\mathbb{T}$, denote $p(\sigma)\in\mathbb{T}$ as the parent of $\sigma$ (i.e. the node one step closer to the root).
Assume that the nodes are sampled with a Markov process that is indexed by $\mathbb{T}$: each node $\sigma\in\mathbb{T}$ corresponds to an individual $X_\sigma$ sampled from the population $G$, and an edge $(\sigma,\tau)$ of $\mathbb{T}$ denotes that the sampled individual $X_\sigma$ referred the individual $X_\tau$ into the sample. Mathematically, let $\{X_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$ be a tree-indexed Markov process on the individuals from the social network $G$:
\begin{equation*}
\mathbb{P}(X_\sigma^{(\cdot)}=j \mid X_{p(\sigma)}^{(\cdot)}=i, X_\tau^{(\cdot)}:\tau \in \mathscr{D}(\sigma)^c)=\mathbb{P}(X_\sigma^{(\cdot)}=j\mid X_{p(\sigma)}=i)=P_{ij},
\end{equation*}
where $\mathscr{D}(\sigma) \subset \mathbb{T}$ denotes the set of $\sigma$ and all its descendants in $\mathbb{T}$. The superscript $(\cdot)$ indicates the initial condition: if the superscript is some $i\in G$, $X_0$ is initialized from $i$; if the superscript is some distribution $\bm{\nu}: G\to \mathbb{R}$ (e.g. the stationary distribution $\bm{\pi}$ of $\bm{P}$), $X_0$ is initialized from $\bm{\nu}$. When the initial state does not matter, we leave off the superscript. Following \cite{benjamini1994markov}, we call this process a $(\mathbb{T},\bm{P})$-walk on $G$.
In a special case, $\mathbb{T}$ can be the chain graph ($ 0 - 1 -2 -3- \dots $); this results in the model being a Markov chain. Just as a chain graph indexes a Markov chain, the graph $\mathbb{T}$ provides the indexing in this model. For simplicity, $\sigma\in\mathbb{T}$ is used synonymously with $\sigma$ belonging to the vertex set of $\mathbb{T}$. The seed participant is root vertex $0$ in $\mathbb{T}$. For each non-root node $\sigma\in\mathbb{T}$, denote $p(\sigma)\in\mathbb{T}$ as the parent of $\sigma$ (i.e. the node one step closer to the root). Assume that the nodes are sampled with a Markov process that is indexed by $\mathbb{T}$.
For each node $i\in G$, let $y(i)$ denote some characteristic of this node, for example whether $i$ is HIV+ or HIV-. Sometimes we regard $\bm{y}$ as a vector in $\mathbb{R}^N$, where $N$ is the number of nodes in $G$. We want to estimate the population average $\mu_{\mathsf{true}}=\sum_{i \in G} y(i)/N$
by the RDS sample $\{y(X_\sigma):\sigma\in\mathbb{T}\}$.
\subsection{A special case: Blockmodel}
\label{section_rank_k}
Consider $G$ as coming from a Blockmodel with $k$ blocks \citep{White1976Social}. That is, each node $i\in G$ is assigned to a block with $b(i) \in \{1,\ldots,k\}$, where each block $j$ contains $N/k$ nodes. If $b(i) = b(j)$, then $w_{i\ell} = w_{j \ell}$ for all $\ell \in \{1, \dots N\}$. Further suppose that if $b(i)=b(j)$, then $y(i)=y(j)$. The Stochastic Blockmodel \citep{Holland1983Stochastic} is derived from this model.
The idea behind a Blockmodel with $k$ blocks is clear: people in the same block share the same feature and the same friendship patterns. \cite{goel2009respondent} studied RDS with this model. The motivating example in Section \ref{sec_introduction} also uses a Blockmodel with $2$ blocks.
Let $\mathcal{W}\in\mathbb{R}^{k\times k}$ denote the weight matrix between blocks, where $\mathcal{W}_{b(i),b(j)}=w_{ij}$.
Define the corresponding Markov transition matrix between blocks $\mathcal{P}\in \mathbb{R}^{k\times k}$ from $\mathcal{W}$ similarly to \eqref{eq_tran_matrix}. Since $\mathcal{W}$ is symmetric, $\mathcal{P}$ is reversible.
Let $\{B_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$ denote a Markov process indexed by $\mathbb{T}$, where the state space is the block labels $\{1,\ldots,k\}$ and the transition matrix is $\mathcal{P}$.
The superscript of $B_\sigma^{(\cdot)}$ indicates the initial state of $B_0$ and is in correspondence with the initial state $X_0$ of the Markov process over $G$: if $X_0$ is initialized at $i\in G$, $B_0$ is initialized at $z(i)$ and the superscript is $z(i)$; if $X_0$ is initialized from any distribution $\bm{\nu}: G\to\mathbb{R}$, $B_0$ is initialized from the distribution $\bm{\mu}: \{1,\ldots,k\}\to\mathbb{R}$ with $\mu_j=\sum_{i\in G:z(i)=j} \nu_i$. For any $\{\sigma_{i_1}, \ldots, \sigma_{i_s}\}\subset\mathbb{T}$ and $b_{i_1},\ldots,b_{i_s}\in\{1,\ldots,k\}$,
\begin{equation}
\label{eq:blockmodel}
\mathbb{P}(B_{\sigma_{i_1}}^{(\cdot)}=b_{i_1},\ldots,B_{\sigma_{i_s}}^{(\cdot)}=b_{i_s})=\mathbb{P}(b(X_{\sigma_{i_1}}^{(\cdot)})=b_{i_1},\ldots,b(X_{\sigma_{i_s}}^{(\cdot)})=b_{i_s}).
\end{equation}
The proof of \eqref{eq:blockmodel} is in Appendix \ref{section_proof_eq2}. So $\{B_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$ is equal in distribution to $\{b(X_\sigma^{(\cdot)}): \sigma\in\mathbb{T}\}$. Instead of studying the Markov process $\{X_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$ in Section \ref{subsec_markovmodel}, we study the Markov process $\{B_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$. Intuitively, the original process $\{X_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$ keeps track of the individuals while $\{B_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$ keeps track of some feature of the individuals.
This time the node feature $\bm{y}\in\mathbb{R}^N$ is replaced by the block feature $\bm{b}\in \mathbb{R}^k$ and the Markov transition matrix is replaced by the Markov transition matrix between blocks $\mathcal{P}\in\mathbb{R}^{k\times k}$.
The Blockmodel is a special case of the Markov model in Section \ref{subsec_markovmodel}. In this paper, Theorem \ref{thm_mut}, Corollary \ref{cor_decay} and \ref{cor_VH} apply to the Markov model. Theorem \ref{thm_gls} and Corollary \ref{cor_glsVH} only apply to the Blockmodel with $2$ blocks.
\subsection{Estimators}
\label{sec_estimators}
Denote $\mathbb{E}_{\bm{\pi}} (y)=\sum_{i}\pi(i) y(i)$. The theoretical results in this paper study two estimators defined in this section. They are unbiased estimators of $\mathbb{E}_{\bm{\pi}} (y)$. When applying inverse probability weighting (in Section \ref{sec_ipw}), these estimators become unbiased estimators of $\mu_{\mathsf{true}}$ instead. Further, the VH adjustment provides a way to estimate the inverse probability weights.
\paragraph{Sample average}
\label{subsec_sampleaverage}
The RDS sample average is
\begin{equation}
\hat{\mu}^{(\cdot)}=\frac{1}{n} \sum_{\sigma\in\mathbb{T}} y(X_\sigma^{(\cdot)}).
\end{equation}
When $X_0$ is initialized from $\bm{\pi}$, $\hat{\mu}^{(\bm{\pi})}$ is an unbiased estimator of $\mathbb{E}_{\bm{\pi}} (y)$. When $X_0$ is initialized from $i\in G$, $\hat{\mu}^{(i)}$ is an asymptotically unbiased estimator of $\mathbb{E}_{\bm{\pi}} (y)$ (see Claim \ref{lem:trivial}).
\paragraph{GLS estimator}
\label{subsec_glsestimator}
\cite{roch2018generalized} proposed generalize least squares (GLS) in RDS to reduce the variance, particularly in the high variance regime. The GLS estimator is the weighted average
\begin{equation}
\hat{\mu}^{(\cdot)}_{\mathsf{GLS}}=\sum_{\sigma\in\mathbb{T}} w_\sigma^\ast y(X^{(\cdot)}_\sigma)
\end{equation}
where $\bm{w}^\ast$ minimizes the variance of the weighted average initialized from $\bm{\pi}$
\begin{equation}
\label{eq:glsdef}
\bm{w}^\ast=\arg\min_{\bm{w}} \,\mathsf{Var}\left(\sum_{\sigma\in\mathbb{T}} w_\sigma y(X^{(\bm{\pi})}_\sigma)\right) \quad s.t. \quad \sum_{\sigma\in\mathbb{T}} w_\sigma=1.
\end{equation}
When $X_0$ is initialized from $\bm{\pi}$, $\hat{\mu}^{(\bm{\pi})}_{\mathsf{GLS}}$ is an unbiased estimator of $\mathbb{E}_{\bm{\pi}} (y)$. When $X_0$ is initialized from $i\in G$, $\hat{\mu}^{(i)}_{\mathsf{GLS}}$ is an \emph{asymptotically} unbiased estimator of $\mathbb{E}_{\bm{\pi}} (y)$ (see Theorem \ref{thm_gls}).
\subsection{Inverse probability weighting}
\label{sec_ipw}
In general $\mu_{\mathsf{true}}\neq\mathbb{E}_{\bm{\pi}} (y)$. So $\hat{\mu}$ and $\hat{\mu}_{\mathsf{GLS}}$ are biased estimators for $\mu_{\mathsf{true}}$.
Inverse probability weighting can adjust for this bias. Define
$y^{\bm{\pi}}(i)=y(i)/(N\pi(i))$. The
IPW estimator and GLS estimator with IPW adjustment are the sample average and the GLS estimator of $y^{\bm{\pi}}(X_\sigma)$'s:
$$\hat{\mu}_{\mathsf{IPW}}=\frac{1}{n} \sum_{\sigma\in\mathbb{T}} y^{\bm{\pi}}(X_\sigma)= \frac{1}{n} \frac{\mathrm{vol}(G)}{N} \sum_{\sigma \in \mathbb{T}} \frac{y(X_\sigma)}{\mathrm{deg}(X_\sigma)}, \ \Rm{and}$$
$$\hat{\mu}_{\mathsf{IPW,GLS}}=\sum_{\sigma\in\mathbb{T}} w^{\bm{\pi}}_\sigma y^{\bm{\pi}}(X_\sigma)= \frac{\mathrm{vol}(G)}{N} \sum_{\sigma \in \mathbb{T}} w^{\bm{\pi}}_\sigma \frac{y(X_\sigma)}{\mathrm{deg}(X_\sigma)}.$$
When $X_0$ is initialized from the stationary distribution $\bm{\pi}$, they are unbiased estimates of $\mu_{\mathsf{true}}$.
However, computing these two estimators requires the average node degree $\mathrm{vol}(G)/N$, which is typically not available in practice.
The popular VH estimator replaces $\mathrm{vol}(G)/N$ in the IPW estimator with the harmonic mean of the degrees of the RDS samples \citep{volz2008probability}. Define
$$ H^{-1}=\frac{1}{n} \sum_{\sigma \in \mathbb{T}} \frac{1}{\mathrm{deg}(X_{\sigma})}, \qquad \hat{\pi}(i)=H^{-1} \mathrm{deg}(i), \qquad y^{\hat{\bm{\pi}}}(i)=\frac{y(i)}{\hat{\pi}(i)}.$$
The VH estimator is the sample average of $y^{\hat{\bm{\pi}}}(X_\sigma)$'s. The GLS estimator with VH adjustment uses a similar reweighting, but replaces $\mathrm{vol}(G)/N$ with a GLS estimate of $\mathbb{E}_{\bm{\pi}} (1/\mathrm{deg}(i))$ \citep{roch2018generalized}.
The VH estimator and GLS estimator with VH adjustment are two asymptotically unbiased estimators of $\mu_{\mathsf{true}}$ under the $(\mathbb{T},\bm{P})$-walk on $G$. Theorem \ref{thm_mut} and \ref{thm_gls} study the limit distribution of the sample average and GLS estimator. By a simple transformation (defining a new node function $y^{\bm{\pi}}(i)=y(i)/(N\pi(i))$), these results can also be applied to the IPW estimator and the GLS estimator with IPW adjustment. Corollary \ref{cor_VH} and \ref{cor_glsVH} extend these results to the VH estimator and GLS estimator with VH adjustment.
\subsection{Additional notation}
For two sequences $a_n$ and $b_n$, define the following notation: (i) $a_n=O(b_n)$ if and only if $\abs{a_n}$ is bounded above by $b_n$ (up to constant factor) asymptotically, i.e. $\exists k>0, \ \exists n_0, \ \forall n>n_0, \abs{a_n}\leq kb_n$.
(ii) $a_n=\Theta(b_n)$ if and only if $a_n$ is bounded both above and below by $b_n$ (up to constant factors) asymptotically, i.e.
$\exists k_1>0, \ \exists k_2>0, \ \exists n_0, \ \forall n>n_0, \ k_1 b_n\leq a_n \leq k_2 b_n$.
\section{Discussion}\label{sec:disc}
We prove the existence of a limit distribution for the IPW estimator under the Markov model of respondent-driven sampling and show that this limit distribution depends on the seed node---thus the limit distribution is a non-trivial mixture distribution when the seed is randomized. This result also shows that the ``seed bias'' of IPW is non-negligible. We also establish the asymptotic normality of the GLS estimator under certain conditions and show that this limiting normal does not depend on the seed node. This implies that the ``seed bias'' of GLS is negligible. Both results allow for the VH adjustment. Our empirical study on social networks as well as on simulated data illustrate that these theoretical results appear to hold beyond the technical conditions given in the theorems.
\section{Introduction}
\label{sec_introduction}
Network sampling techniques, including web crawling, snowball sampling, and respondent-driven sampling (RDS), contact individuals in hard-to-reach populations by following edges in a social network. This paper uses RDS as a motivating example \citep{heckathorn1997respondent}. It is used by the Centers for Disease Control (CDC) and the Joint United Nations Programme on HIV/AIDS (UN-AIDS) to sample populations most at risk for HIV (injection drug users, sex workers, and men who have sex with men) \citep{HIVbehavioralSurveilance, johnston2013introduction}. In the most recent survey of the literature \citep{white2015strengthening}, RDS had been applied in over 460 different studies, in 69 different countries.
An RDS sample is initialized with one or more ``seed individuals'' selected by convenience from the population. These individuals participate in the survey and are incentivized to refer additional participants (often up to 3 or 5 participants) into the sample. This process iterates until reaching the target sample size or there are no referrals. All participants are incentivized to take a survey and an HIV test. With this sample, we wish to estimate the proportion of individuals in the population that are HIV+.
\begin{table}[b]
\centering
\caption{Summary of properties of IPW and GLS estimators. In the columns, $m$ refers to the number of participants that the typical participant refers into the study and $\lambda_2$ is the second eigenvalue of the Markov transition matrix. }
\label{table_previous}
\begin{tabular}{|c|c|c|c|}
\hline
Result & Estimator & Low variance, i.e. $m<\lambda_2^{-2}$ & High variance, i.e. $m>\lambda_2^{-2}$ \\ \hline
\multirow{2}{*}{Variance} & IPW & \begin{tabular}[c]{@{}c@{}}$O(n^{-1})$\\ \citep{rohe2019critical}\end{tabular} & \begin{tabular}[c]{@{}c@{}}$O(n^{2 \log_m \lambda_2})$\\ \citep{rohe2019critical}\end{tabular} \\ \cline{2-4}
& GLS & \multicolumn{2}{c|}{$O(n^{-1})$ \citep{roch2018generalized}} \\ \hline
\multirow{2}{*}{Distribution} & IPW\&VH & \begin{tabular}[c]{@{}c@{}}Asymptotically normal\\ \citep{li2017central}\end{tabular} & \begin{tabular}[c]{@{}c@{}}Non-trivial mixture \\ $[$Current paper$]$ \end{tabular} \\ \cline{2-4}
& GLS & \multicolumn{2}{c|}{Asymptotically normal $[$Current paper$]$} \\ \hline
\end{tabular}
\end{table}
The Markov model for the RDS process has provided fundamental insight into RDS sampling \citep{salganik2004sampling, goel2009respondent, rohe2019critical}. For example, nodes with more connections are more likely to be sampled \citep{levin2009markov}. This creates bias and there are ways to adjust for it \citep{salganik2004sampling, volz2008probability}. While the inverse probability weighted (IPW) estimator requires a normalizing constant that is unknown in practice, the Volz-Heckathorn (VH) estimator provides a way to estimate this normalizing constant \citep{volz2008probability}. More recently, \cite{rohe2019critical} studied the variability of the IPW estimators and showed that there are two regimes (low variance and high variance). This regime is determined by two parameters of the Markov process that is described in Section \ref{subsec_markovmodel}. In brief, let $\lambda_2$ be the second eigenvalue of the Markov transition matrix on the social network and let $m$ be the average number of referrals provided by each node. When $m < \lambda_2^{-2}$, the variance of the IPW estimator decays at rate $n^{-1}$, where $n$ is the sample size. However, when $m > \lambda_2^{-2}$, the variance of IPW decays at a slower rate. Later, \cite{li2017central} showed that the VH and IPW estimators are asymptotically normal under the Markov model in the low variance regime. More recently, \cite{roch2018generalized} proposed a generalized least squares (GLS) estimator for the high variance regime and showed that the variance of this estimator is $O(n^{-1})$, even when $m>\lambda_2^{-2}$. These previous results are summarized in Table \ref{table_previous}.
This paper studies the limit distribution of (i) the GLS estimator and (ii) the IPW estimator in the high variance regime. These results also allow for the Volz-Heckathorn adjustment. For technical reasons, our analysis of the GLS estimator is restricted to a special case of the Markov model that was first used to study RDS in \cite{goel2009respondent}.
These technical results make many unrealistic assumptions which we discuss below. In particular, the Markov model allows for resampling of individuals. The results are asymptotic in the sample size, while the population size is fixed. This creates extensive resampling. Nevertheless, this model provides fundamental insights into the properties of the estimators and these properties continue to hold under more realistic simulation models in Sections \ref{sec_sumulation} and \ref{section_analysis}.
\begin{figure}[t]
\centering
\includegraphics[height=11cm]{Figure5new.pdf}
\caption{The model for this simulation is described in Section \ref{sec:motivate}. The two left panels show the distribution of sample proportion (i.e. the IPW estimator in this model). The two right panels show the distribution of GLS estimator.
Each panel in the top row has two curves corresponding to whether or not the seed node is HIV+. The solid line gives the distribution of the estimator when the process is initialized with an HIV+ node. The dashed line is initialized with an HIV- node. In the bottom row, the seed participant is selected uniformly at random. This figure demonstrates how the limit distribution of the IPW estimator can have two modes which correspond to whether the seed is HIV+ or HIV-. Moreover, the figure suggests that the GLS estimator is asymptotically normal and the dependence on the seed node is negligible.}
\label{fig_distributions_mean_gls}
\end{figure}
\subsection{A simple motivating example}\label{sec:motivate}
Here we consider a model studied in \cite{goel2009respondent}, which we refer to as the Blockmodel with $2$ blocks. In this example, the population that we wish to sample is equally divided into two groups: HIV+ and HIV-. The seed participant is selected from one of the two groups with equal probability. Each participant refers an iid number of offspring, generated from some offspring distribution. With probability $p$, the referred participant matches the HIV status of the participant that referred them. With probability $1-p$, their statuses differ. Each referral is independent, conditional on the status of the referring participant. Using a sample generated in this way, we wish to estimate the proportion of the population that is HIV+ (in this case, the true proportion is $0.5$).
Figure \ref{fig_distributions_mean_gls} displays a motivating simulation from this Blockmodel with $2$ blocks. Each sample size is 1000 individuals, sampled from the Blockmodel with $p=.95$ and offspring distribution $1+\Rm{Binomial}(2,0.5)$. For each sample of 1000, we construct both sample proportion (equivalent to the IPW estimator, see Section \ref{sec_ipw}) and GLS estimator. This process is repeated 10000 times. Figure \ref{fig_distributions_mean_gls} displays a kernel density estimate of the resulting distribution.
\subsection{Main contributions}
Many RDS papers discuss the ``bias from seed selection''. Section \ref{subsec_sample_mean} shows that the IPW and VH estimators have a limit distribution and this limit distribution depends on where the process is initialized (i.e. the ``seed'' node). If the seed node is randomized, then in simulations, the limit distribution of the IPW and VH estimators can have multiple modes, where each mode corresponds to a different set of initial conditions. The limit results for the IPW and VH estimators highlight how, conditioned on the seed node, the bias of these estimators decays at the same rate as the variance. So, unconditional on the seed node, this can create multiple modes in the limit distributions of the IPW and VH estimators. Similarly to classical results in multitype branching process theory \citep{kesten1966additional}, the exact limit distribution does not appear to have a concise and easily interpretable closed form.
While the IPW and VH estimators are not asymptotically normal in the high variance regime, Section \ref{subsec_gls} shows that the GLS estimator is asymptotically normal in this regime and this limit distribution does not depend on where the process is initialized. This pair of results provides additional insight into the notions of ``bias'' and ``variance'' for network sampling. In particular, the GLS estimator is the linear estimator with the smallest variance and that measure of variance includes the variability that comes from selecting the seed node (i.e. from the stationary distribution of the Markov process). Hence, it adjusts for the seed selection. Another way of saying this is that the GLS estimator reduces ``the bias from seed selection''. This blurring of the divide between ``variance" and ``bias from seed selection'' highlights one potential problem of conditioning on the seed node in a bootstrap resampling procedure \citep{baraff2016estimating}; in the high variance regime, conditioning on the seed node removes a large source of variability in the VH estimator.
\section{Main results}
\label{sec_main}
This section shows that, after proper scaling, the GLS estimator and the sample average both have a limit distribution. For GLS, the limit distribution is a normal distribution. For the sample average, on the other hand, the limit distribution is a non-trivial mixture distribution, where the mixture component is determined by the seed node. This mixture distribution can be multi-modal as illustrated in Figure \ref{fig_distributions_mean_gls}. These results can be further extended to the GLS estimator with VH adjustment and to the VH estimator respectively.
We will need the following standard lemma (e.g.~\cite[Lemma 12.2]{levin2009markov}) which provides the eigendecomposition of the Markov transition matrix $\bm{P}$.
\begin{lemma
\label{lem_eigendecomposition}
Let $\bm{P}$ be a reversible Markov transition matrix on the nodes in $G$ with respect to the stationary distribution $\bm{\pi}$. The eigenvectors of $\bm{P}$, denoted as $\bm{f}_1, \dots, \bm{f}_{N}$, are real valued functions of the nodes $i \in G$ and orthonormal with respect to the inner product
\begin{equation} \label{def:inner}
\langle \bm{f}_a, \bm{f}_b \rangle_{\bm{\pi}} = \sum_{i \in G} f_a(i) f_b(i) \pi(i).
\end{equation}
If $\lambda$ is an eigenvalue of $\bm{P}$, then $|\lambda|\le 1$. The eigenfunction $\bm{f}_1$ corresponding to the eigenvalue $1$ can be taken to be the constant vector $\bm{1}$.
\end{lemma}
Assume that the eigenvalues of $\bm{P}$ are $$\abs{\lambda_1}\geq\abs{\lambda_2}\geq\cdots\geq\abs{\lambda_N}.$$
Since it is a Markov transition matrix, its largest eigenvalue is $\lambda_1=1$.
Let $\bm{f}_i$ be the eigenvector corresponding to $\lambda_i$, normalized as in Lemma \ref{lem_eigendecomposition}. The eigenvector $\bm{f}_1$ corresponding to $\lambda_1$ is taken to be the constant vector $\bm{1}$. Expanding the node feature $y\in\mathbb{R}^N$ in the eigenbasis yields
\begin{equation}
\label{eq_eigendecompose}
\bm{y}=\sum_{j=1}^{N} \langle \bm{y},\bm{f}_j \rangle_{\bm{\pi}} \bm{f}_j.
\end{equation}
\subsection{Results for the sample average and the IPW and VH estimators}
\label{subsec_sample_mean}
This section shows that the sample average, IPW and VH estimators have a limit distribution and that this limit distribution in fact depends on where the process is initialized (i.e. the ``seed'' node).
For each node $\sigma\in\mathbb{T}$, let $\abs{\sigma}$ be the distance of $\sigma$ from the root $0$. Define $\{X_\sigma:\sigma\in\mathbb{T},\abs{\sigma}=t\}$ as the individuals in the $t$-th generation of the sample. Denote the sample average up to generation $t$ as $\hat{\mu}_t$. Superscripts on $\hat \mu$ will denote how $X_0$ is initialized.
Theorem \ref{thm_mut} studies the limit distribution of the sample average $\hat{\mu}_t^{(i)}$. Recall that the sample average of RDS samples is $\hat{\mu}^{(\cdot)}=n^{-1} \sum_{\sigma\in\mathbb{T}} y(X_\sigma^{(\cdot)}).$
\begin{theorem}
\label{thm_mut}
Assume the eigenvalues of the transition matrix $\bm{P}$ are
\begin{equation}
\label{eq_eigenvalue}
1=\lambda_1>\lambda_2>\abs{\lambda_3}\geq\cdots\geq\abs{\lambda_N}.
\end{equation}
Assume $\mathbb{T}$ is an $m$-tree. When $m>\lambda_2^{-2}$, there exist a random variable $X^{(i)}\in L^2$ such that
\begin{equation}
\label{eq_mu_converge}
\lambda_2^{-t}\left[\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)\right] \rightarrow X^{(i)}
\end{equation}
almost surely and in $L^2$ as $t\to\infty$, and
\begin{equation}
\label{eq_different_expectation}
\mathbb{E} X^{(i)}=\frac{(m-1)\lambda_2}{m\lambda_2-1}
\,\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}}
\,f_2(i).
\end{equation}
Moreover, if $\langle \bm{y}, \bm{f}_2 \rangle_{\bm{\pi}}\neq 0$, then $\mathsf{Var}(X^{(i)})>0$ for any $i=1,\ldots,N$.
\end{theorem}
Note that the result is based on the technicial condition that $\mathbb{T}$ is an $m$-tree. The simulations in Section \ref{sec_sumulation} suggest that the result still holds when $\mathbb{T}$ is a Galton-Watson tree. Condition \eqref{eq_eigenvalue} in Theorem \ref{thm_mut} can be weakened to
\begin{equation*}
1=\lambda_1>\lambda_2=\cdots=\lambda_k>\abs{\lambda_{k+1}}\geq\cdots\geq\abs{\lambda_N},
\end{equation*}
but the statement of the conclusion becomes more involved.
See Remark \ref{remark:generalize_eig} for a complete statement.
Using the above result, we can study how the bias and variance of the sample average decays, conditioned on the seed node.
\begin{corollary}
\label{cor_decay}
Assume the conditions of Theorem \ref{thm_mut} hold.
\begin{enumerate}
\item When $\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}}\neq 0$ and $f_2(i)\neq 0$, the bias of $\hat{\mu}_t^{(i)}$ decays like
\begin{equation}
\left[\mathbb{E} (\hat{\mu}_t^{(i)})-\mathbb{E}_{\bm{\pi}} (y)\right]^2 = \Theta(\lambda_2^{2t}).
\end{equation}
\item When $\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}}\neq 0$, the variance of $\hat{\mu}_t^{(i)}$ decays like
\begin{equation}
\mathsf{Var} (\hat{\mu}_t^{(i)}) = \Theta(\lambda_2^{2t}).
\end{equation}
\end{enumerate}
\end{corollary}
When $X_0$ is initialized from $\bm{\pi}$, $\hat{\mu}_t^{(\bm{\pi})}$ is an unbiased estimator of $\mu_{\mathsf{true}}$. By \eqref{eq_different_expectation}, for $i,j$ such that $f_2(i)\neq f_2(j)$, the limit distributions of $\lambda_2^{-t}\hat{\mu}_t^{(i)}$ and $\lambda_2^{-t}\hat{\mu}_t^{(j)}$ are different because $X^{(i)}$ and $X^{(j)}$ have different expectations. Thus the limit distribution of $\lambda_2^{-t}\hat{\mu}_t^{(\bm{\pi})}$ is a non-trivial mixture. The motivating example in the introduction illustrates this mixture. It is further explored with the simulation in Section \ref{sec_sumulation}.
Theorem \ref{thm_mut} studies the limit distribution of the sample average. Using the transformation discussed in Section \ref{sec_ipw}, the result also applies to the IPW estimator. Denote the VH estimator up to generation $t$ as $\hat{\mu}_{\mathsf{VH},t}$. The following corollary extends the result to the VH estimator.
\begin{corollary}
\label{cor_VH}
Under the conditions of Theorem \ref{thm_mut}, there exists a random variable $\tilde{X}^{(i)}\in L^2$ such that $$\lambda_2^{-t}\left[\hat{\mu}^{(i)}_{\mathsf{VH},t}-\mu_{\mathsf{true}}\right]\to \tilde{X}^{(i)}$$
almost surely, and
\begin{equation*}
\mathbb{E} \tilde{X}^{(i)}=\mathbb{E}_{\bm{\pi}}(y^\prime)^{-1}\frac{(m-1)\lambda_2}{m\lambda_2-1}
\,\langle \bm{y}^{\prime\prime},\bm{f}_2 \rangle_{\bm{\pi}}\,
f_2(i),
\end{equation*}
where $y'(j)=\mathrm{deg}(j)^{-1}$ and $y''(j)=y(j)/\mathrm{deg}(j)$. Moreover, if $\langle \bm{y}^{\prime\prime},\bm{f}_2 \rangle_{\bm{\pi}} \neq 0$, then $\mathsf{Var}(\tilde{X}^{(i)})>0$ for any $i=1,\ldots,N$.
\end{corollary}
Similarly, when $X_0$ is initialized from $\bm{\pi}$, the limit distirbution of $\hat{\mu}_{\mathsf{VH},t}^{(\bm{\pi})}$ is a non-trivial mixture of the limit distributions of $\hat{\mu}_{\mathsf{VH},t}^{(i)}$ for all $i\in G$.
\subsection{Results for the GLS estimator}
\label{subsec_gls}
For the GLS estimator, the two right panels of Figure \ref{fig_distributions_mean_gls} suggest that the estimator is not sensitive to the initial distribuiton of $X_0$. This section shows that the GLS estimator is asymptotically normal with parameters that do not depend on the initial distribution of $X_0$.
Given the referral tree $\mathbb{T}$, define the covariance matrix $\bm{\Sigma}\in\mathbb{R}^{n\times n}$ as
\begin{equation*}
\bm{\Sigma}_{\sigma,\tau}=\mathsf{Cov} (y(X_\sigma),y(X_\tau))
\end{equation*}
for any $\sigma, \tau\in\mathbb{T}$, where $n$ is the number of nodes in $\mathbb{T}$. According to \cite{roch2018generalized}, $\bm{w}^\ast$ in \eqref{eq:glsdef} is given by
\begin{equation}
\label{eq_gls_weight}
\bm{w}^\ast=(\bm{x}^\top \textbf{1})^{-1} \bm{x}^\top, \quad \text{where} \quad \bm{\Sigma} \bm{x}=\bm{1}.
\end{equation}
Here $\bm{x}$ is the vectorization of the RDS sample $\{X_\sigma^{(\cdot)}: \sigma\in\mathbb{T}\}$.
For the Blockmodel with $2$ blocks, the GLS estimator admits a closed-form expression:
\begin{equation}
\label{eq_gls_expression}
\hat{\mu}_{\mathsf{GLS}}=\sum_{\sigma\in\mathbb{T}} \frac{1-\lambda_2 (\mathrm{deg}(\sigma)-1)}{n(1-\lambda_2(1-\frac{2}{n}))} y(X_\sigma),
\end{equation}
where $\lambda_2$ is the second eigenvalue of the Markov transition matrix between blocks and $\mathrm{deg}(\sigma)$ is the degree of $\sigma \in \mathbb{T}$.
Let $\hat{\mu}_{\mathsf{GLS},t}$ be the GLS estimator of the RDS samples up to generation $t$. Based on \eqref{eq_gls_expression}, the following theorem establishes the asymptotic normality of the GLS estimator.
\begin{theorem}
\label{thm_gls}
Consider the Blockmodel with $2$ blocks on
an $m$-tree $\mathbb{T}$. Assume $\abs{\lambda_2}<1$. Then, for any initial distribution $\bm{\nu}$ of $X_0$,
\begin{equation}
\sqrt{n_t}\left[\hat{\mu}_{\mathsf{GLS},t}^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}} (y)\right] \to \mathcal{N}\left(0,\frac{1+\lambda_2}{1-\lambda_2}\,\mathsf{Var}_{\bm{\pi}}(y)\right).
\end{equation}
in distribution as $t \to \infty$, where $\mathsf{Var}_{\bm{\pi}}(y)=\mathbb{E}_{\bm{\pi}}(y^2)-(\mathbb{E}_{\bm{\pi}}(y))^2$ and $n_t=1+m+\cdots+m^t$ is the number of RDS samples up to generation $t$.
\end{theorem}
Theorem \ref{thm_gls} shows that the GLS estimator is asymptotically normal \emph{both in the low variance and high variance regimes}. Note that the result is based on \eqref{eq_gls_expression} and the technical condition that $\mathbb{T}$ is an $m$-tree. The simulations in Section \ref{sec_sumulation} suggest that the asymptotic normality of the GLS estimator still holds when $\mathbb{T}$ is a Galton-Watson tree, or the model is no longer a Blockmodel with $2$ blocks.
Theorem \ref{thm_gls} studies the limit distribution of the GLS estimator. Using the transformation discussed in Section \ref{sec_ipw}, the result also applies to the GLS estimator with IPW adjustment. Denote the GLS estimator with VH adjustment of RDS samples up to generation $t$ as $\hat{\mu}^{(\cdot)}_{\mathsf{GLS,VH},t}$. The following corollary extends the result to the GLS estimator with VH adjustment.
\begin{corollary}
\label{cor_glsVH}
Under the conditions in Theorem \ref{thm_gls}, for any initial distribution $\bm{\nu}$ of $X_0$,
$$\sqrt{n_t}\left[\hat{\mu}_{\mathsf{GLS,VH},t}^{(\bm{\nu})}-\mu_{\mathsf{true}}\right] \xrightarrow{d} \mathcal{N}\left(0,\frac{1+\lambda_2}{1-\lambda_2}\,\mathbb{E}_{\bm{\pi}}(y')^{-2}
\,\mathsf{Var}_{\bm{\pi}}(y^{\prime\prime})\right).$$
where $y'(i)=\mathrm{deg}(i)^{-1}$ and $y''(i)=y(i)/\mathrm{deg}(i)$.
\end{corollary}
\section*{Acknowledgements}
Yan is partially supported by the elite undergraduate training program of School of Mathematical Sciences in Peking University. Roch is supported by NSF grants DMS-1614242 CCF-1740707 (TRIPODS) and DMS-1916378, and a Simons Fellowship. Rohe is supported by NSF grant DMS-1612456 and ARO grant W911NF-15-1-0423.
\section{Proof outlines for the main results}
This section outlines the proofs for Theorems \ref{thm_mut} and~\ref{thm_gls}. Well-established theory for multi-type branching processes and martingale limit theorems play an important role. For each proof, the main idea is to extract the underlying martingale structure for the estimator; it is this structure that determines the asymptotic behavior. The proofs of Corollary~\ref{cor_decay}, \ref{cor_VH} and \ref{cor_glsVH} are relegated to Appendices~\ref{S:app-sample}
and~\ref{S:app-GLS}.
\subsection{Analysis of the sample average}
Denote $\bm{Z}_{t,j}$ as the number of $j\in G$ in the $t$-th generation and define $\bm{Z}_t=(Z_{t,1},\ldots,Z_{t,N})$. When $\mathbb{T}$ is an $m$-tree and $X_0$ is initialized from $i\in G$,
$$\bm{Z}_t^{(i)}=(Z_{t,1}^{(i)},\ldots,Z_{t,N}^{(i)})$$
is a multitype Galton-Watson process \citep{harris2002theory,athreya2004branching}. The next lemma can be derived from a standard result in the literature of multitype Galton-Watson processes.
\begin{lemma}
\label{lem:martingale}
Assume the conditions of Theorem \ref{thm_mut}. For any $j=1,\ldots,N$, $$Y_{t,j}^{(i)}=(m\lambda_j)^{-t} \langle \bm{Z}_t^{(i)}, \bm{f}_j\rangle$$ is a real-valued martingale adapted to $\mathcal{F}_t=\sigma\{\bm{Z}_l^{(i)}:1\leq l\leq t\}$.
\end{lemma}
\begin{proof}
See Appendix \ref{subsec-lem-martingale}.
\end{proof}
Let $W_t$ denote the summation of the $t$-th generation RDS samples,
\[W_t = \sum_{\sigma\in\mathbb{T}:\abs{\sigma}=t} y(X_\sigma),\]
and let $S_t = \sum_{j = 0}^t W_j$ denote the summation up to generation $t$. Recall that $n_t$ is the number of nodes in $\mathbb{T}$ between the root $0$ and generation $t$ (inclusive), i.e. $n_t = |\{\sigma\in\mathbb{T},\abs{\sigma}\le t\}|$. Thus the sample average up to generation $t$ is $\hat{\mu}_t = S_t / n_t$. Superscripts on $\bm{Z}, S$ and $W$ will denote how $X_0$ is initialized if necessary.
Recall from \eqref{eq_eigendecompose} and $\bm{f}_1=\bm{1}$ that
\begin{equation*}
\bm{y}=\sum_{j=1}^{N} \langle \bm{y},\bm{f}_j \rangle_{\bm{\pi}} \bm{f}_j=\mathbb{E}_{\bm{\pi}}(y)\bm{1}+\sum_{j=2}^{N} \langle \bm{y},\bm{f}_j \rangle_{\bm{\pi}} \bm{f}_j.
\end{equation*}
As a result, one obtains the following decomposition of $W_t^{(i)}$
\begin{equation}
\label{eq:Wt-decomposition}
W_t^{(i)}=\sum_{\sigma\in\mathbb{T},\abs{\sigma}=t} y(X_\sigma^{(i)})=\bm{y}^\top \bm{Z}_t^{(i)}=m^t\mathbb{E}_{\bm{\pi}} (y)+\sum_{j=2}^{N} \langle \bm{y},\bm{f}_j \rangle_{\bm{\pi}} (m\lambda_j)^t Y_{t,j}^{(i)}.
\end{equation}
The last step utilizes the simple fact that $\bm{1}^\top\bm{Z}_t^{(i)}=m^t$. This motivates us to study the limit distribution of $Y_{t,j}^{(i)}$.
\begin{lemma}
\label{lem:martingale-convergence}
Assume the conditions of Theorem \ref{thm_mut}. Then there exists a random variable $Y_2^{(i)}$ such that $$Y_{t,2}^{(i)} \to Y_2^{(i)}$$ almost surely and in $L^2$.
For $j\geq 3$,
$$(\lambda_2^{-1}\lambda_j)^t\,Y_{t,j}^{(i)} \to 0$$
almost surely and in $L^2$.
\end{lemma}
\begin{proof}
See Appendix \ref{subsec-lem-martingale-converence}.
\end{proof}
Lemma \ref{lem:martingale-convergence} informally reveals that, under proper scaling, the asymptotic distributional characterization of $W_t^{(i)}$ is determined by $Y_{t,2}^{(i)}$. The next lemma derives the first and second moments of $Y_2^{(i)}$.
\begin{lemma}
\label{lem:nondegenerate-Y2}
Assume the conditions of Theorem \ref{thm_mut}. Then $\mathbb{E} Y_2^{(i)}=f_2(i)$, and $\mathsf{Var}(Y_2^{(i)})>0$ for any $i=1,\ldots,N$ if we further assume that $\langle \bm{y}, \bm{f}_2 \rangle_{\bm{\pi}}\neq 0$ holds.
\end{lemma}
\begin{proof}
See Appendix \ref{subsec-lem-Y2}.
\end{proof}
\subsection{Proof of Theorem \ref{thm_mut}}
Apply \eqref{eq:Wt-decomposition} and Lemma \ref{lem:martingale-convergence} collectively to obtain
\begin{equation}
\label{eq_Wt}
\frac{W_t^{(i)}-m^t\mathbb{E}_{\bm{\pi}} (y)}{(m\lambda_2)^t}=\sum_{j=2}^{N} \langle \bm{y},\bm{f}_j \rangle_{\bm{\pi}} (\lambda_2^{-1}\lambda_j)^t Y_{t,j}^{(i)} \rightarrow \langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} Y_{2}^{(i)}
\end{equation}
almost surely and in $L^2$. Recalling that $S_t=\sum_{j=0}^t W_j$, $\hat{\mu}_t=S_t/n_t$,
and the number of samples between $0$ and generation $t$ is
$n_t=\sum_{l=0}^{t}m^l$, one arrives at
\begin{equation}
\label{eq_as_converge}
\lambda_2^{-t}\left[\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)\right]=\frac{S_t^{(i)}-n_t \mathbb{E}_{\bm{\pi}} (y)}{n_t\lambda_2^t}=\frac{m^t}{n_t} \sum_{l=0}^{t} (m\lambda_2)^{l-t}\frac{W_l^{(i)}-m^l \mathbb{E}_{\bm{\pi}} (y)}{(m\lambda_2)^l}.
\end{equation}
Since $\lim_{t\to\infty}m^t/n_t=(m-1)/m$, from \eqref{eq_Wt}
\begin{equation}
\label{eq:mu_converge}
\lambda_2^{-t}\left[\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)\right] \to \frac{m-1}{m}\sum_{r=0}^{\infty} (m\lambda_2)^{-r}\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} Y_{2}^{(i)}=\frac{(m-1)\lambda_2}{m\lambda_2-1}\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} Y_{2}^{(i)}\triangleq X^{(i)}
\end{equation}
almost surely as $t\to\infty$.
To prove $L^2$ convergence, recall that if a sequence of random variables $X_n \to X$ in probability, and $\Vert X_n \Vert_{L^2} \to \Vert X \Vert_{L^2}$, then $X_n \to X$ in $L^2$. Observe, similarly to the above limits, that
\begin{align*}
\left\Vert \lambda_2^{-t}(\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)) \right\Vert_{L^2}^2&=\frac{m^{2t}}{n_t^2}\mathbb{E} \left[\left(\sum_{l=0}^{t} \frac{W_l^{(i)}-m^l \mathbb{E}_{\bm{\pi}} (y)}{(m\lambda_2)^t}\right)^2\right]\\
&=\frac{m^{2t}}{n_t^2}\sum_{k=1}^{t}\sum_{l=1}^{t} \mathbb{E}\left\{(m\lambda_2)^{-2t}\left[W_k^{(i)}-m^k \mathbb{E}_{\bm{\pi}} (y)\right] \left[W_l^{(i)}-m^l \mathbb{E}_{\bm{\pi}} (y)\right]\right\} \\
&\xrightarrow{t\to\infty}\lim_{r\to\infty}
\left(\frac{m-1}{m}\right)^2\mathbb{E}\left[\left(\langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} Y_{2}^{(i)}\right)^2\right]\sum_{k=1}^{r}\sum_{l=1}^{r} (m\lambda_2)^{k+l-2r} \\
&=\mathbb{E}\left[\left(\frac{(m-1)\lambda_2}{m\lambda_2-1} \langle \bm{y},\bm{f}_2 \rangle_{\bm{\pi}} Y_{2}^{(i)}\right)^2\right].
\end{align*}
So the convergence in \eqref{eq:mu_converge} is also in $L_2$.
Finally, in view of Lemma \ref{lem:nondegenerate-Y2} and the definition of $X^{(i)}$ in \eqref{eq:mu_converge}, it is straightforward to check that \eqref{eq_different_expectation} holds. Moreover, if $\langle \bm{y},\bm{f}_2\rangle_{\bm{\pi}}\neq0$, then $\mathsf{Var}(X^{(i)})>0$ for any $i=1,\ldots,N$.
\begin{remark}
\label{remark:generalize_eig}
If condition \eqref{eq_eigenvalue} in Theorem \ref{thm_mut} is weakened to
\begin{equation*}
1=\lambda_1>\lambda_2=\cdots=\lambda_k>\abs{\lambda_{k+1}}\geq\cdots\geq\abs{\lambda_N},
\end{equation*}
then \eqref{eq_mu_converge} becomes
\begin{equation*}
\lambda_2^{-t} (\hat{\mu}_t^{(i)}-\mathbb{E}_{\bm{\pi}} (y)) \rightarrow \frac{(m-1)\lambda_2}{m\lambda_2-1} \sum_{j=2}^{k} \langle \bm{y},\bm{f}_j \rangle_{\bm{\pi}}Y_{j}^{(i)}.
\end{equation*}
\end{remark}
\subsection{Analysis of the GLS estimator}
In previous sections, the subscript of the estimators is $t$ or $l$, which denotes the generation. This section requires us to study each node in a generation. Accordingly we order the nodes of the $m$-tree $\mathbb{T}$ by scanning each level from the root down. For example, for a $2$-tree, the root node is $1$, its offsprings are $2$ and $3$, the offsprings of $2$ are $4$ and $5$, the offsprings of $3$ are $6$ and $7$, etc. In a change of notation from the previous sections, when the subscript is $n$, $\hat{\mu}^{(\cdot)}_{n}$ now denotes the sample mean up to node $n$, i.e.
$\hat{\mu}_{n}=n^{-1}\sum_{k=1}^{n} y(X_k)$.
Assume the Markov transition matrix between blocks is
$$\mathcal{P}=\Big(\begin{matrix}
p & 1-p \\ 1-q & q
\end{matrix}\Big).$$
The second eigenvalue is $\lambda_2=p+q-1$ and the stationary
distribution is $\bm{\pi} = (\frac{1-q}{1-\lambda_2},\frac{1-p}{1-\lambda_2})$. For $k\geq 1$, define
\begin{equation}
\label{eq:Mn-martingale}
M_n=\sum_{k=1}^{n} \left[y(X_k^{(\bm{\nu})})-\lambda_2 y(X_{p(k)}^{(\bm{\nu})})\right]-n(1-\lambda_2)\mathbb{E}_{\bm{\pi}} (y),
\end{equation}
where $p(k)$ is the parent node of $k$ in the ordering
defined above.
In view of \eqref{eq_gls_expression}, the relation between $M_n$ and $\hat{\mu}_{\mathsf{GLS,t}}$ is
\begin{equation}
\label{eq:mtg-gls-relationship}
\sqrt{n_t}\left[\hat{\mu}_{\mathsf{GLS},t}^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}} (y)\right]=\frac{M_{n_t}}{\sqrt{n_t}(1-\lambda_2)}+O_{\mathbb{P}}\left(\frac{1}{\sqrt{n_t}}\right).
\end{equation}
Thus it suffices to study the asymptotic behavior of $M_n$. One can check that $M_n$ is a martingale adapted to the filtration
$\mathcal{F}_n=\sigma(X_k^{(\nu)}: k\leq n)$:
\begin{equation*}
\begin{aligned}
\mathbb{E}\left(M_n-M_{n-1}\mid \mathcal{F}_{n-1}\right)&=\begin{cases}
p y_1+(1-p) y_2-\lambda_2 y_1-(1-\lambda_2)\mathbb{E}_{\bm{\pi}} (y) \ &\text{if} \ X_{p(n)}^{(\bm{\nu})}=1\\
(1-q) y_1+q y_2-\lambda_2 y_2-(1-\lambda_2)\mathbb{E}_{\bm{\pi}} (y) \ &\text{if} \ X_{p(n)}^{(\bm{\nu})}=2
\end{cases} =0,
\end{aligned}
\end{equation*}
both of which are $0$, as can be seen from
$(1-\lambda_2)\mathbb{E}_{\bm{\pi}} (y)=(1-q) y_1+(1-p) y_2$ and the expression for $\lambda_2$ above.
It is necessary to introduce a martingale central limit theorem (see e.g. \cite[Fifth Edition, Theorem 8.2.8]{durrett2019probability}).
\begin{theorem}[Martingale CLT]
\label{thm_martingale_CLT}
Let a martingale $M_n$ satisfy $\mathbb{E}(M_n)=0$, and
\begin{enumerate}
\item $\frac{1}{n} \sum_{k=1}^{n} \mathbb{E}((M_k-M_{k-1})^2\mid M_1,\ldots,M_{k-1})\rightarrow \sigma^2>0$ in probability as $n\to \infty$, and
\item for every $\epsilon>0$, $\frac{1}{n} \sum_{k=1}^{n} \mathbb{E}((M_k-M_{k-1})^2; \abs{M_k-M_{k-1}}>\epsilon\sqrt{n})\rightarrow 0$ as $n\to\infty$,
\end{enumerate}
then $M_n/\sqrt{n}\to \mathcal{N}(0,\sigma^2)$ in distribution as $n\to\infty$.
\end{theorem}
It then boils down to showing that $M_n$ defined in \eqref{eq:Mn-martingale} satisfies the conditions in the above theorem. The detailed proof is provided in the next section. We will need a technical lemma which states that, although the limit distribution of the sample average $\hat{\mu}_n^{(i)}$ differs in the high and low variance regimes under appropriate scalings, $\hat{\mu}_n^{(i)}$ itself always converges to $\mathbb{E}_{\bm{\pi}}(y)$ in $L^2$.
\begin{lemma}
\label{lem:trivial'}
Assume the conditions of Theorem \ref{thm_gls}. Then for any initial distribution $\bm{\nu}$ of $X_0$, $\hat{\mu}_n^{(\bm{\nu})} \to \mathbb{E}_{\bm{\pi}}(y)$ in $L^2$.
\end{lemma}
\begin{proof}
See Appendix \ref{subsec-lem-trivial'}
\end{proof}
\subsection{Proof of Theorem \ref{thm_gls}}
\label{subsec-theorem-gls}
Without loss of generality, assume $\mathbb{E}_{\bm{\pi}}(y)=0$ and $\mathsf{Var}_{\bm{\pi}}(y)=1$, which can be equivalently viewed as applying the same linear transformation to each entry of $\bm{y}$ as well as $\hat{\mu}_{\mathsf{GLS},t}$.
We begin by showing that $M_n$ defined in \eqref{eq:Mn-martingale} satisfies the first condition in Theorem \ref{thm_martingale_CLT}.
By invoking the martingale property $\mathbb{E}(M_k-M_{k-1} \mid \mathcal{F}_{k-1})=0$, one obtains
\begin{equation*}
\mathbb{E}\left[(M_k-M_{k-1})^2\mid \mathcal{F}_{k-1}\right]=\mathsf{Var}\left(M_k-M_{k-1}\mid \mathcal{F}_{k-1}\right)=\mathsf{Var}\big(y(X_k^{(\bm{\nu})})\mid X_{p(k)}^{(\bm{\nu})}\big).
\end{equation*}
Notice that $\mathbb{E}_{\bm{\pi}}(y)=0$ and $\mathsf{Var}_{\bm{\pi}} (y)=1$ imply $(1-q)y_1^2+(1-p)y_2^2=2-p-q$, which yields
\begin{equation*}
\begin{aligned}
\mathsf{Var}\big(y(X_k^{(\bm{\nu})})\mid X_{p(k)}^{(\bm{\nu})}\big)&=\begin{cases}
py_1^2+(1-p)y_2^2-\lambda_2^2 y_1^2 &\text{if} \ X_{p(n)}^{(\bm{\nu})}=1\\
(1-q)y_1^2+qy_2^2-\lambda_2^2 y_2^2 &\text{if} \ X_{p(n)}^{(\bm{\nu})}=2
\end{cases}\\
&=(1-\lambda_2)(1+\lambda_2 y(X_{p(k)}^{(\bm{\nu})})^2).
\end{aligned}
\end{equation*}
Thus
$\mathbb{E}[(M_k-M_{k-1})^2\mid \mathcal{F}_{k-1}]=(1-\lambda_2)(1+\lambda_2 y(X_{p(k)}^{(\bm{\nu})})^2)$.
For notational simplicity, denote
$$V_n=\frac{1}{n} \sum_{k=1}^{n} \mathbb{E}\left[(M_k-M_{k-1})^2\mid \mathcal{F}_{k-1}\right]=\frac{1-\lambda_2}{n} \sum_{k=1}^{n} (1+\lambda_2 y(X_{p(k)}^{(\bm{\nu})})^2).$$
When $\mathbb{T}$ is an $m$-tree, each node from level $0$ to $t-1$ is counted $m$ times as a parent. Define a new node feature $\bm{y}^\prime=(y(1)^2,y(2)^2)^\top$. Let $\hat{\gamma}_{n}^{(\bm{\nu})}$ be the sample average of $y^\prime(X_\sigma^{(\bm{\nu})})$'s up to node $n$. By Lemma \ref{lem:trivial'} applied to $\hat{\gamma}_{n}^{(\bm{\nu})}$,
\begin{equation*}
\begin{aligned}
V_n&=1-\lambda_2+\lambda_2(1-\lambda_2)\frac{m\sum_{k=1}^{p(n)}y'(X_k^{(\nu)})+O(1)}{n}=1-\lambda_2+\lambda_2(1-\lambda_2)\hat{\gamma}_{p(n)}^{(\bm{\nu})}+O(1/n)\\
&\xrightarrow{L^2} 1-\lambda_2+\lambda_2(1-\lambda_2) \mathbb{E}_{\bm{\pi}} (y')=1-\lambda_2^2
\end{aligned}
\end{equation*}
as $n\to\infty$. Here $O(1)$ in the first line comes from the fact that $y'(X_{p(n)}^{(\bm{\nu})})$ might be counted less than $m$ times, which results in a remainder term bounded by $m\Vert \bm{y}^\prime\Vert_\infty$, and the second line uses $\mathbb{E}_{\bm{\pi}}(y^\prime)=\mathsf{Var}_{\bm{\pi}}(y)=1$. Since $L^2$ convergence implies convergence in probability, the first condition is verified.
We now move on to the second condition. Notice that $$\abs{M_k-M_{k-1}}=\abs{y(X_k^{(\bm{\nu})})-\lambda_2 y(X_{p(k)}^{(\bm{\nu})})}\leq (1+\lambda_2) \inftynorm{\bm{y}}.$$
So $\mathbb{P}(\abs{M_k-M_{k-1}}>\epsilon\sqrt{n})=0$ when $n>\epsilon^{-2}(1+\lambda_2)^2 \inftynorm{\bm{y}}^2$.
This gives
\begin{equation*}
\frac{1}{n} \sum_{k=1}^{n} \mathbb{E}((M_k-M_{k-1})^2; \abs{M_k-M_{k-1}}>\epsilon\sqrt{n})=0
\end{equation*}
for sufficiently large $n$. Thus the second condition is verified.
As a result, we obtain from Theorem \ref{thm_martingale_CLT} that $M_n/\sqrt{n} \to \mathcal{N}(0,1-\lambda_2^2)$ in distribution. Combined with \eqref{eq:mtg-gls-relationship} and Slutsky's theorem, one finally arrives at
\begin{equation*}
\sqrt{n_t}\left[\hat{\mu}_{\mathsf{GLS},t}^{(\bm{\nu})}-\mathbb{E}_{\bm{\pi}}(y)\right] \xrightarrow{d}\mathcal{N}\left(0,\frac{1+\lambda_2}{1-\lambda_2}\,\mathsf{Var}_{\bm{\pi}}(y)\right).
\end{equation*}
\section{Simulation studies}\label{sec_sumulation}
In this section, data are simulated from a
Blockmodel with $2$ or $3$ blocks. As stated in Section \ref{section_rank_k}, a Blockmodel with $k$ blocks consists of a reversible transition matrix $\mathcal{P}\in\mathbb{R}^{k\times k}$ between blocks, block feature $\bm{y}\in\mathbb{R}^k$, and a referral tree $\mathbb{T}$. In this specification, the block feature $\bm{y}$ is assumed to be centralized, so that $\mathbb{E}_{\bm{\pi}} (y)=0$. For a Blockmodel with $2$ blocks, let
\begin{equation*}
\mathcal{P}=\Big(\begin{matrix}
p & 1-p \\ 1-q & q
\end{matrix}\Big).
\end{equation*}
denote the transition matrix between 2 blocks.
The second eigenvalue of $\mathcal{P}$ is $\lambda_2=p+q-1$.
In the simulation settings below, the block feature is given prior to centralization. In fact, all of the $2$-Blockmodels use $\bm{y}=(1,0)^\top$ and the $3$-Blockmodels use $\bm{y}=(0,1,2)^\top$ . All of the experiments are based on 5000 simulated datasets.
\subsection{Sample average} \label{sec:sim_average}
Here we consider the behavior of the sample average $\hat{\mu}_t$ in the high variance regime $m>\lambda_2^{-2}$. In this setting, the asymptotic distribution of $\lambda_2^{-t}\hat{\mu}_t^{(\bm{\pi})}$ is no longer normal, unlike the low variance regime. Instead, its asymptotic distribution is a mixture of the distributions of $\lambda_2^{-t}\hat{\mu}_t^{(i)}$ for all $i\in G$.
The simulation is performed on two different Blockmodels with $2$ blocks. We consider a balanced model with $p=q=.95$ and an unbalanced model with $p=0.95$ and $q=0.85$. For both models, $\mathbb{T}$ is a Galton-Watson tree with offspring distribution $1+\Rm{Binomial}(2,1/2)$. Under these settings, $m>\lambda_2^{-2}$ for both models. Figure \ref{fig_sample mean for two rank-2 models} displays the results of the experiment with $t=50$.
\begin{figure}[t]
\centering
\includegraphics[height=11cm]{Figure3new.pdf}
\caption{Kernel density estimates of $\lambda_2^{-t}\hat{\mu}_t$ for balanced (the left panels) and unbalanced (the right panels) Blockmodel with $2$ blocks over $5000$ replicates. For each scenario, the top panel corresponds to the case when $X_0$ is initialized from group 1 (the solid curve) and group 2 (the dashed curve), the lower panel corresponds to the case when $X_0$ is initialized from the stationary distribution.}
\label{fig_sample mean for two rank-2 models}
\end{figure}
\subsection{GLS estimator}\label{sec:sim_gls}
Here we consider the behavior of the GLS estimator in both the low and high variance regimes. The first experiment corroborates the result of Theorem \ref{thm_gls}, namely that the GLS estimator is asymptotically normal in both variance regimes. The simulation is performed on two different Blockmodels with $2$ blocks. In the first model $(p,q) = (0.95, 0.85)$; in the second model $(p,q) = (0.8, 0.7)$. For both models, $\mathbb{T}$ is a $2$-tree. Under these settings, $m>\lambda_2^{-2}$ for the first model and $m<\lambda_2^{-2}$ for the second model. The two quantile-quantile plots in Figure \ref{fig_gls_2tree} correspond to the two models. It appears that the distribution of the GLS estimator gets closer to the normal distribution as the sample size increases.
The second experiment suggests that the asymptotic normality of GLS estimator extends beyond the conditions in Theorem \ref{thm_gls}. We consider a two-block model with $(p,q) = (0.8, 0.7)$ and a three-block model, where the transition matrix between the blocks is
\begin{equation*}
\mathcal{P}=\left(\begin{matrix}
0.8 & 0.1 & 0.1 \\
0.2 & 0.6 & 0.2 \\
0.2 & 0.2 & 0.6 \\
\end{matrix}\right).
\end{equation*}
For both models, $\mathbb{T}$ is a Galton-Watson tree with offspring distribution $1+\Rm{Binomial}(2,1/2)$. Results for this experiment are displayed in Figure \ref{fig_gls_extend}.
\begin{figure}[t]
\centering
\includegraphics[height=7cm]{Figure2new.pdf}
\caption{Q-Q plot of $\mu_{t,GLS}$ for the Blockmodels with $2$ blocks, with $m>\lambda_2^{-2}$ (left panel) and $m<\lambda_2^{-2}$ (right panel). $\mathbb{T}$ is a $2$-tree. For each scenario, the Q-Q plot is created over $5000$ replicates. The six dashed Q-Q lines with different colors correspond to $\mathbb{T}$ with 5, 6, 7, 8, 9 or 10 levels. The red solid line is $y=x$. .}
\label{fig_gls_2tree}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=7cm]{Figure4new.pdf}
\caption{Q-Q plot of $\hat{\mu}_{t,GLS}$ for the Blockmodels with $2$ blocks (left panel) and $3$ blocks (right panel), where $\mathbb{T}$ is a Galton-Watson tree. For each scenario, the Q-Q plot is created over $5000$ replicates. The six dashed Q-Q lines with different colors correspond to $\mathbb{T}$ with 5, 6, 7, 8, 9 or 10 levels. The red solid line is $y=x$. }
\label{fig_gls_extend}
\end{figure}
\section{Analysis of Adolescent Health Data}\label{section_analysis}
\begin{figure}[t]
\centering
\includegraphics[height=14cm]{SampleMeandensity.pdf}
\caption{Simulation results based on the Add Health Study described in Section \ref{section_analysis}. The figures display kernel density estimates of the sample average. The 25 subplots correspond to the Comm 17, 75, 42, 15, 28, 39, 40, 41, 50, 34, 45, 48, 36, 43, 61, 54, 59, 73, 44, 68, 60, 58, 84, 57, 49 networks. The red solid line is $x=\mu_{true}$. This figure suggests that VH estimator has multiple modes.}
\label{fig_adolescent_vh_density}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=14cm]{GLSdensity.pdf}
\caption{Simulation results based on the Add Health Study described in Section \ref{section_analysis}. The figures display kernel density estimates of the GLS estimator (solid line) and the SBM-fGLS estimator (dashed line). The 25 subplots correspond to the Comm 17, 75, 42, 15, 28, 39, 40, 41, 50, 34, 45, 48, 36, 43, 61, 54, 59, 73, 44, 68, 60, 58, 84, 57, 49 networks. The red solid line is $x=\mu_{true}$. This figure shows that when the bottleneck of the network is not too strong, both estimators have only one mode.}
\label{fig_adolescent_gls_density}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=14cm]{GLSQQ.pdf}
\caption{Simulation results based on the Add Health Study described in Section \ref{section_analysis}. The figures display Q-Q plots of the GLS estimator (solid line) and the SBM-fGLS estimator (dashed line). The 25 subplots correspond to the Comm 17, 75, 42, 15, 28, 39, 40, 41, 50, 34, 45, 48, 36, 43, 61, 54, 59, 73, 44, 68, 60, 58, 84, 57, 49 networks. This figure illustrates that when the bottleneck of the network is not too strong, both estimators appear approximately normal (even under without replacement sampling).}
\label{fig_adolescent_qq}
\end{figure}
In this section, we consider numerical experiments where the RDS samples are simulated without replacement from empirically derived social networks. Specifically, we use social networks collected in the National Longitudinal Study of Adolescent Health (Add Health). In the 1994-95 school year, the Add Health study collected a nationally representative sample of adolescents in grades seven through twelve. The sample covers 84 pairs of middle and high schools in which students nominated up to five male and five female friends in their middle or high school network \citep{harris2011national}.
In this analysis, we consider 25 networks with at least $1000$ nodes. All contacts are symmetrized and all graphs are restricted to the largest connected component. The RDS sampling process is initialized from a seed node which is selected with probability proportional to node degree (i.e. the stationary distribution). Then, each participant recruits $\xi\sim 1+\Rm{Binomial}(2,1/2)$ participants uniformly at random from their contacts whom have not yet been recruited. If the participant has fewer than $\xi$ contacts eligible to recruit, then the participant recruits all of their eligible contacts. The RDS process stops when there are $500$ participants. If the process terminates before collecting $500$ participants, then the process is restarted. For each network, we collect $500$ different RDS samples. We generate $2000$ such simulated data sets.
We use school-status as the binary node feature and focus on estimating the proportion of the population in high school. We construct a sample average, a GLS estimator and a SBM-fGLS estimator for the proportion of students in high school. The GLS estimator requires an estimate of the covariance matrix $\Sigma$, which can be calculated from the Markov transition matrix of the network (typically not available in practice) and equation (6) in \cite{rohe2019critical}. The SBM-fGLS estimator proposed in \cite{roch2018generalized} estimates $\Sigma$ using the RDS samples.
Consider a measure of the network bottleneck. Let $\bm{A}\in\mathbb{R}^{N\times N}$ denote the adjacency matrix of the network. Define the diagonal matrix $\bm{D}\in\mathbb{R}^{N\times N}$ and the matrix $\bm{L}\in\mathbb{R}^{N\times N}$ so that
\begin{equation*}
D_{ii}=\sum_{k=1}^{N} W_{ik}, \quad \bm{L}=\bm{D}^{-1/2} \bm{A} \bm{D}^{-1/2}.
\end{equation*}
Then $\tilde{\lambda}$ is defined as
\begin{equation}\label{equ:bottle}
\tilde{\lambda}=\tilde{\bm{y}}^\top \bm{L} \tilde{\bm{y}},
\end{equation}
where $\tilde{\bm{y}}$ is the standardized form of the node feature $\bm{y}$, so that $\sum_{i=1}^{N} \tilde{y}_i=0$ and $\twonorm{\tilde{\bm{y}}}=1$. $\tilde{\lambda}$ provides a measure of the network bottleneck; as long as the second eigenvalue $\lambda_2$ is not too close to $1$, then this quantity will not be close to $1$. Table \ref{table_lambda} displays the $\tilde{\lambda}$ of the $25$ networks.
\begin{table}[htpb]
\centering
\caption{Network characteristics for the 25 networks in the Add Health study used in the numerical experiments in Section \ref{section_analysis}. ID gives the network ID (school ID) from the study listed in increasing order by $\tilde{\lambda}$, a measure of the strength of bottleneck in the network, see \eqref{equ:bottle}.}
\label{table_lambda}
\begin{tabular}{|cc|cc|cc|cc|cc|}
ID & $\tilde{\lambda}$ & ID & $\tilde{\lambda}$ & ID & $\tilde{\lambda}$ & ID & $\tilde{\lambda}$ & ID & $\tilde{\lambda}$ \\ \hline
17 & 0.739 & 39 & 0.842 & 45 & 0.869 & 54 & 0.879 & 60 & 0.911 \\
75 & 0.744 & 40 & 0.844 & 48 & 0.869 & 59 & 0.881 & 58 & 0.917 \\
42 & 0.771 & 41 & 0.847 & 36 & 0.874 & 73 & 0.886 & 84 & 0.923 \\
15 & 0.818 & 50 & 0.867 & 43 & 0.874 & 44 & 0.889 & 57 & 0.925 \\
28 & 0.839 & 34 & 0.868 & 61 & 0.878 & 68 & 0.897 & 49 & 0.944 \\
\end{tabular}
\end{table}
In Figure \ref{fig_adolescent_vh_density}, the 25 subplots show the kernel density estimation of VH estimator corresponding to the 25 networks. In Figure \ref{fig_adolescent_gls_density} and \ref{fig_adolescent_qq}, the 25 subplots show the kernel density estimation and quantile-quantile plots of GLS and SBM-fGLS estimator with VH adjustment corresponding to the 25 networks. We plot these results over $2000$ replicates. The 25 subplots are in order of descending $\tilde{\lambda}$. It is clear that the VH estimator has two modes, so these networks are all beyond the critical threshold. Except for networks with extremely strong bottleneck (i.e. with large $\tilde{\lambda}$), the GLS estimators with VH adjustment are approximately normally distributed. The distribution of SBM-fGLS estimator with VH adjustment are not enough close to the normal distribution for some networks, which means that our results for the GLS estimator might not always hold for the SBM-fGLS estimator. It is possible for the GLS estimator to exceed one. In practice, one would provide a modified estimate capped at one. |
\section{Introduction}
In this work we continue our investigations of infinite horizon optimal control problems with nonconvex cost functionals which we started in \cite{KKR17}. We focus on optimal control of nonlinear dynamical systems which are affine in the control. The input control is a vector-valued function $u=(u_1,\ldots,u_m)$ in the space $L^\infty(0,\infty;\mathbb{R}^m)$ under control constraints. The focus rests on that part of the cost functional which involves the control. It is given as follows:
\begin{equation}\label{CFpq}
\int^\infty_0 \left(\sum^m_{i=1}|u_i(t)|^p\right)^{q/p}dt,
\end{equation}
where $0<p<1$ and $p\leq q \leq 1$. This functional is nonsmooth and nonconvex, leading to a challenging optimal control problem with interesting properties for the optimal control laws, in particular sparsity and switching. It appears that the terminology "sparse" is not rigorously defined in the literature, but generally it is used to describe the property of the optimal control to be identically zero over nontrivial subsets of the temporal domain. Here, by sparsity we refer to the situation that the whole vector $u(t)$ is zero. Switching control, is related to coordinate-wise sparsity, and is used to describe the property
\[
u_i(t)u_j(t)=0 \ \text{for }\ i,j\in\{1,\ldots,m\},\ i\neq j,\ t\geq 0,
\]
which is equivalent to saying that at most one coordinate of $u(t)$ is non-zero at $t$.
While the use of the control penalty \eqref{CFpq} does not guarantee sparsity or switching properties, it enhances them. This is illustrated in Figure \ref{uballs}, where unit balls for different $q/p$ ratios are shown. For a fixed $q$ decreasing $p$ (column-wise in the sub-figure) one direction becomes dominant over the other.
\begin{figure}[!h]
\includegraphics[width=0.32\textwidth]{p1q1}\hfill
\includegraphics[width=0.32\textwidth]{p06q06}\hfill
\includegraphics[width=0.32\textwidth]{p02q02.png}\\
\includegraphics[width=0.32\textwidth]{p06q1.png}\hfill
\includegraphics[width=0.32\textwidth]{p04q06.png}\hfill
\includegraphics[width=0.32\textwidth]{p01q02.png}
\caption{Contour levels ($0.1$ to $1$) of different balls $\|u\|_p^q$.}\label{uballs}
\end{figure}
To further illustrate the effect of \eqref{CFpq} let us consider
the case $p=1/2$ and $q=1$. Then the running cost for the control is given by
\[
\sum^m_{i=1} |u_i(t)|+2\sum_{i,j\in\{1,\ldots,m\},i\neq j}|u_i(t)u_j(t)|^{1/2},
\]
where the $L^1$-penalization on $u_i$ will support sparsity in the control and the product penalization enhances switching phenomena. More generally, if $\frac{q}{p}=j \in \mathbb{N}$ is an integer, then
the running cost is is combination of an $L^q$-penalization on each control coordinate $u_i$, and it further contains weighted summands of (up to) $j-$ tuples of fractional powers of $|u_i|$, with the sum of the powers for each tuple summing to $q$. Fixing $q$, and decreasing $p$ we expect that the control cost \eqref{CFpq} increases the switching nature of the optimal controls, since the weights on the tuples compared to those on the singletons increase. Moreover, decreasing $q$ we expect that the subdomain over which the optimal control vanishes (in all coordinates) increases. These properties will be illustrated by numerical experiments.
The case with $p=q$ and $0<p\leq 1$ has been studied in \cite{KKR17}. Existence and sparsity properties of optimal controls have been analyzed for this case, and these properties have been observed in the numerical simulations in the case with $0<p=q<1$. In the present work, the analysis is made for more general nonconvex problems with the control cost \eqref{CFpq}. Concerning the question of existence of optimal solutions, which is not guaranteed in general, we follow the ideas from \cite{KKR17} to reformulate the problem in infinite-dimensional sequence spaces by descretizing the controls, and extending an important result on weakly sequentially continuous mappings from \cite{IK14} to obtain the existence result for our purposes.
The analysis of the sparsity and switching structure is based on the optimality conditions. For this purpose we derive the necessary first order optimality conditions of the original problem, which follow from general results which are available in the literature. We also derive sufficient optimality condition for the reformulated problems. Subsequently, we investigate the sparsity and switching properties of the optimal controls under box constraints. Finally, by using dynamic programming techniques, optimal control laws are approximated globally in the state space for linear and nonlinear dynamical systems.
Let us mention previous related work on sparse and switching control. Closed-loop infinite horizon sparse optimal control problems with $L^p$ ($0<p\leq 1$) functionals were analyzed in \cite{KKR17}. Open-loop, finite horizon $L^1$ sparse optimal control for dynamical systems have been studied in e.g. \cite{HAJ79,VM06,ALT15,CFPT}. Open-loop, finite horizon sparse optimal control for partial differential equations was studied in e.g. \cite{HSW12,CCK12,PV13}. The Hamilton-Jacobi-Bellman equation for impulse and switching controls was discussed in \cite{BC97,Y89}. The synthesis of sparse feedback laws via dynamic programming has been studied in \cite{Falcone14,KKK16,Albi17}. In the context of partial differential equations optimal control of systems switching among different modes were analysed in \cite{HLS09,HS13}, problems with convex switching enhancing functionals were investigated in \cite{CIK16}, and problems with nonconvex switching penalization in \cite{CKR17}. In \cite{Z11} switching controls based on functionals suggested
by controllability considerations were investigated.
Mixed (quasi-)norms as
in \eqref{CFpq} with $p \neq q$ have been used earlier, though typically in convex situations with
$ p\ge1, q\ge 1$. These investigations were carried out in the context of machine learning, regression analysis, and mathematical imaging, with
the goal of achieving group sparsity or structured parsimony, see e.g. \cite{BC97, Fornasier09, Kowa, WNF09, ZRY09},
and the references given there.
The structure of the paper is the following. The short section 2 contains the precise problem formulation. Existence of optimal controls, which are discretized in time, is obtained in section 3. The sparsity and switching structure of the optimal controls is analyzed on the basis of the optimality conditions for the time-continuous as well as the time discrete problems in sections 4 and 5, respectively, and section 6 contains numerical results.
\section{Optimal control problem}
Let $U\subset\mathbb{R}^m$ be a closed set and let $f_i:\mathbb{R}^d\rightarrow\mathbb{R}^d$ be continuous functions for $i=0,\ldots,m$. We consider the following control system: given $x\in\mathbb{R}^d$,
\begin{equation}\label{odey}
\left\{
\begin{array}{ll}
\dot y(t)=f_0(y(t))+\sum^m_{i=1}f_i(y(t))u_i(t) & \text{in}\ ]0,\infty[,\\
y(0)=x.
\end{array}
\right.
\end{equation}
Here $y(t)\in\mathbb{R}^d$ is the state variable and $u(t)=(u_1(t),\ldots,u_m(t))\in\mathbb{R}^m$ is the input control.
Given $p\in]0,1[$, we set for the vector $u=(u_1,\ldots,u_m)\in\mathbb{R}^m$
\[
\|u\|_p=\left(\sum^m_{i=1}|u_i|^p\right)^{1/p}.
\]
Let $q\in[p,1]$, $\lambda>0$, $\gamma>0$ and $y_d\in\mathbb{R}^d$. For any $x\in\mathbb{R}^d$, consider the cost functional
\begin{equation}\label{costfunctional}
J(x,u):=\int^{\infty}_0 e^{-\lambda t}\left(\frac{1}{2}\|y(t)-y_d\|^2_2+\gamma\|u(t)\|_p^q\right)dt,
\end{equation}
where $(y,u)$ satisfies the state equation \eqref{odey}, and the infinite horizon optimal control problem
\begin{equation}\label{ocpb}
\inf\left\{ J(x,u)\,:\,u\in L^{\infty}(0,\infty;U)\right\}.
\end{equation}
In \eqref{costfunctional}, $\lambda$ is called the discount factor, $\gamma$ is the weight of control cost and $\|\cdot\|_2$ is the Euclidean norm in $\mathbb{R}^d$. The following assumptions are made.
\begin{enumerate}
\item[(H1)] The control set $U$ is compact and convex.
\item[(H2)] There exists $L>0$ such that $\|f_i(x_1)-f_i(x_2)\|_2\leq L \|x_1-x_2\|_2$ for all $x_1,x_2\in\mathbb{R}^d$, and $i=0,\ldots,m$.
\item[(H3)] For each $x\in\mathbb{R}^d$, there exists $u\in L^{\infty}(0,\infty;U)$ such that $J(x,u)<\infty$.
\end{enumerate}
Let us mention that the cost functional $J$ is convex in the state variable and nonconvex in the control. The case $q=p$ has been discussed in \cite{KKR17}.
\section{Time-discretized model}
Since the cost functional $J$ is not convex in $u$,
existence of optimal controllers for problem \eqref{ocpb} does not hold in general. For this purpose we analyse the existence in the case of a time-discretized approximation to \eqref{ocpb}.
We introduce the temporal grid $(t_k)_{k\in\mathbb{N}}$:
\[
0=t_0<t_1<\cdots<t_k<t_{k+1}<\cdots,
\]
and denote by $I_k=[t_k,t_{k+1}[$ for $k\in\mathbb{N}$. The control is then restricted to the following set of piecewise constant functions:
\[
U^\Delta=\{ u=(u_1,\ldots,u_m)\in L^{\infty}(0,\infty;U)\,:\,u_i(t)=u_{i,k}\ \text{for}\ t\in I_k,\ i=1,\ldots,m,\ k\in\mathbb{N}\}.
\]
Consider the following optimal control problem
\begin{equation}\label{ocpdtd}
\inf_{u\in U^\Delta} J^\Delta(x,u):=\int^\infty_0 e^{-\lambda t}\left(\frac{1}{2}\|y(t)-y_d\|^2_2+\gamma
\left(\sum^m_{i=1}\sum^\infty_{k=0}|u_{i,k}|^p\mathbbm{1}_{I_k}(t)\right)^{q/p}\right)dt,
\end{equation}
where $y$ solves \eqref{odey}. A direct computation shows that
\[
J^\Delta (x,u)=\int^\infty_0 e^{-\lambda t}\frac{1}{2}\|y(t)-y_d\|^2_2 dt+ \gamma\sum^\infty_{k=0}b_k\left(\sum^m_{i=1}|u_{i,k}|^p\right)^{q/p},
\]
where
\[
b_k=\int_{I_k} e^{-\lambda t}dt=\frac{1}{\lambda}(e^{-\lambda t_k}-e^{-\lambda t_{k+1}}).
\]
For any $r>0$, the infinite dimensional sequence space $\ell^r=\{u\in\ell^\infty\,:\sum^\infty_{k=1}|u_k|^r<\infty\}$ is endowed with
\[
\|u\|_r=\left(\sum^{\infty}_{k=0}|u_k|^r\right)^{1/r}.
\]
For convenience we recall that $\ell^r$, with $1<r<\infty$, are reflexive Banach spaces and $\ell^{r_1}\subset\ell^{r_2}$ if $1\leq r_1<r_2\leq\infty$.
To investigate the existence of optimal controls, we follow the idea introduced in \cite{IK14} by defining the following reparametrization $\psi:\ell^{q/p}\rightarrow\ell^q$ with
\[
\psi(z)_k=|z_k|^{\frac{1}{p}}\mathop{sgn}(z_k),\ \text{for}\ z=(z_1,z_2,\ldots)\in\ell^{q/p},\ k=1,2,\ldots.
\]
Using the fact that $\psi$ is an isomorphism, \eqref{ocpdtd} is equivalent to
\begin{equation}\label{ocpdtdw}
\inf_{b_k^{-1/q}\psi(w_i)_k\in U} \int^{\infty}_0 e^{-\lambda t}\frac{1}{2}\|y(t)-y_d\|^2_2 dt+ \gamma\sum^\infty_{k=0}\left(\sum^m_{i=1}|w_{i,k}|\right)^{q/p},
\end{equation}
where $y(\cdot)$ satisfies
\begin{equation}\label{DSy}
\left\{
\begin{array}{ll}
\dot y(t)=f_0(y(t))+\sum^m_{i=1}f_i(y(t))b_k^{-1/p}|w_{i,k}|^{1/p}\mathop{sgn}(w_{i,k}) & \text{for}\ t\in[t_k,t_{k+1}),\ k=0,1,\ldots,\\
y(0)=x.
\end{array}
\right.
\end{equation}
To obtain existence for problem \eqref{ocpdtd}, the following lemma is needed which gives some important properties of $\psi$. The idea of proof is inspired by \cite[Lemma 2.1]{IK14}.
\begin{lem}\label{lemell2c}
Let $q> p$ and let $\beta$ denote the conjugate exponent of $q/p$. The mapping $\psi:\ell^{q/p}\rightarrow\ell^{\beta}$ is weakly (sequentially) continuous, i.e. $z^n\rightarrow \bar z$ weakly in $\ell^{q/p}$ implies that $\psi(z^n)\rightarrow\psi(\bar z)$ weakly in $\ell^{\beta}$.
\end{lem}
\begin{proof}
Let $r=\frac{1}{p}+1$ and let $r^*$ denote the conjugate exponent of $r$ given by $r^*=p+1$. Then
\[
1<\frac{q}{p}\leq \frac{1}{p}<r,
\]
which implies $r^*<\beta$. For any $z\in\ell^{q/p}$, we have
\[
\|z\|^r_r=\sum^\infty_{k=1}|z_k|^r,\ \|\psi(z)\|^{r^*}_{r^*}=\sum^\infty_{k=1}|z_k|^{r^*/p}=\sum^\infty_{k=1}|z_k|^r,
\]
and
\[
(\psi(z),z)_{\ell^{r^*},\ell^r}=\sum^{\infty}_{k=1} \psi(z)_k\cdot z_k=\sum^\infty_{k=1}|z_k|^{1/p+1}=\sum^\infty_{k=1}|z_k|^r.
\]
The above computations imply that
\[
(\psi(z),z)_{\ell^{r^*},\ell^r}=\|\psi(z)\|_{r^*}\|z\|_r, \text{ and } \|\psi(z)\|^{r^*}_{r^*}=\|z\|^r_r,
\]
which means that $\psi$ is the duality mapping from $\ell^r$ to $\ell^{r^*}$ and is weakly sequentially continuous. If $z^n\rightarrow\bar z$ weakly in $\ell^{q/p}$, then $z^n\rightarrow\bar z$ weakly in $\ell^r$ since $1<q/p<r$. Therefore, $\psi(z^n)\rightarrow\psi(\bar z)$ weakly in $\ell^{r^*}$. Using that $r^*<\beta$, this implies that $\psi(z^n)\rightarrow\psi(\bar z)$ weakly in $\ell^{\beta}$.
\end{proof}
\begin{thm}\label{thmexistence}
There exists a minimizer $\bar w$ to \eqref{ocpdtdw}, and hence a minimizer $\bar u$ to \eqref{ocpdtd}.
\end{thm}
\begin{proof}
The case $q=p$ has been dealt with in \cite{KKR17}. Therefore, we focus on the case $q> p$.
Let $w^n=(w^n_1,\ldots,w^n_m)$ be a minimizing sequence for the problem \eqref{ocpdtdw}, we set $u^n=(u^n_1,\ldots,u^n_m)$ with
\[
b_k^{1/p}u^n_{i,k}=\psi(w^n_i)_k,\ i=1,\ldots,m,\ k=0,1,\ldots.
\]
Note that $b_k$ and $u^n_{i,k}$ are uniformly bounded with respect to $k,i,n$, and therefore $w^n_{i,k}$ are uniformly bounded.
For each $i \in \{1,\ldots,m\}$,
\[
\sum^{\infty}_{k=0}|w^n_{i,k}|^{q/p}\leq \sum^{\infty}_{k=0}\left(\sum^m_{i=1}|w^n_{i,k}|\right)^{q/p},
\]
which implies that $w^n_i$ is bounded in $\ell^{q/p}$. On the other hand,
\[
\sum^\infty_{k=0}|\psi(w^n_i)_k|^\beta
=\sum^\infty_{k=0}|w^n_{i,k}|^{\beta/p}.
\]
Noting that $\ell^{q/p}\subset\ell^{\beta/p}$ since $\beta=\frac{q}{q-p}>q$, we deduce that
\[
w^n_i\in \ell^{\beta/p}\ \text{and}\ \psi(w^n_i)=\left(b_k^{1/p}u^n_{i,k}\right)_{k\in\mathbb{N}}\in\ell^\beta.
\]
It follows that $\{(\psi(w^n_i), w^n_i)\}_{n=1}^\infty$ is a bounded sequence in $\ell^\beta\times\ell^{q/p}$. Therefore, a subsequence of $\{w^n_i\}_{n=1}^\infty$ converges weakly to some $\bar w_i$ in $\ell^{q/p}$ (see \cite[pp. 73]{Ci90}). By the same reason, a subsequence of $\{\psi(w^n_i)\}_{n=1}^\infty$ converges weakly to some $\xi_i$ in $\ell^\beta$. From Lemma \ref{lemell2c} one deduces that
\[
\xi_i=\psi(\bar w_i), \text{ for each } i \in \{1, \dots, n\}.
\]
Let $y_n$ be the solution to \eqref{DSy} with control $w^n$. Then on each interval $I_k$, we can deduce by the Arzel\`a-Ascoli theorem that there exists $\bar y_k:I_k\rightarrow\mathbb{R}^d$ such that
\[
y_n\rightarrow\bar y_k \ \text{uniformly in}\ I_k,\ \text{as}\ n\rightarrow\infty.
\]
For $\bar y:[0,\infty)\rightarrow\mathbb{R}^d$ defined by $\bar y|_{I_k}=\bar y_k$ for $k\in\mathbb{N}$, it follows that for any $T>0$
\[
y_n\rightarrow \bar y\ \text{uniformly in}\ [0,T),\ \text{as}\ n\rightarrow\infty.
\]
Therefore, $\bar y$ is the solution to \eqref{DSy} corresponding to $\bar w:=(\bar w_1,\ldots,\bar w_m)$. Here we use that the dynamics $f$ is affine in $\psi(w_i)$, $i=1,\ldots,m$. Using the fact that $y_n\rightarrow\bar y$ pointwise in $[0,\infty)$ and $w^n_{i,k}\rightarrow \bar w_{i,k}$ for any $i=1,\ldots,m$, $k\in\mathbb{N}$, we obtain by Fatou's lemma that
\begin{eqnarray*}
&& \int^\infty_0 e^{-\lambda t}\frac{1}{2}\|\bar y(t)-y_d\|^2_2 dt +\sum^{\infty}_{k=0}\left(\sum^m_{i=1}|\bar w_{i,k}|\right)^{q/p} \\
&\leq& \mathop{\lim\,\inf}_{n\rightarrow\infty}\int^\infty_0 e^{-\lambda t}\frac{1}{2}\|y_n(t)-y_d\|^2_2dt +\sum^{\infty}_{k=0}\left(\sum^m_{i=1}|w^n_{i,k}|\right)^{q/p},
\end{eqnarray*}
which implies that $\bar w$ is a minimizer for problem \eqref{ocpdtdw}. Hence a minimizer $\bar u\in U^\Delta$ for problem \eqref{ocpdtd} is given by
\[
\bar u=(\bar u_1,\ldots,\bar u_m),\text{ with } \ \bar u_{i,k}=b_k^{-1/p}\psi(\bar w_i)_k,\ i=1,\ldots,m,\ k\in\mathbb{N}.
\]
\end{proof}
\section{Sparsity and switching properties: the time-continuous problem}
For the time-continuous problem \eqref{ocpb}, the necessary optimality conditions are known from the literature and are next recalled for convenience.
\begin{lem}\label{le:k1}
Assume that $f_i$ are $C^1$ for $i=0,\ldots,m$. For each $x\in\mathbb{R}^d$, if $\bar u$ is a locally optimal control for problem \eqref{ocpb} and $\bar y$ is the associated optimal trajectory, then there exists an adjoint state $\varphi:[0,\infty[\rightarrow\mathbb{R}^d$ such that
\[
\left\{
\begin{array}{ll}
-\dot \varphi(t)=Df_0(\bar y(t))\varphi(t)+\sum^m_{i=1}Df_i(\bar y(t))\bar u_i(t)+e^{-\lambda t}(\bar y(t)-y_d) & \text{for}\ t>0,\\
\lim_{t\rightarrow\infty}\varphi(t)=0,
\end{array}
\right.
\]
and $\varphi$ satisfies, for $t\in]0,\infty[$ a.e.,
\begin{eqnarray}\label{optimality}
&&\left\langle f_0(\bar y(t))+\sum^m_{i=1}f_i(\bar y(t))\bar u_i(t),\varphi(t)\right\rangle+e^{-\lambda t}\left(\frac{1}{2}\|\bar y(t)-y_d\|^2_2+\gamma\|\bar u(t)\|_p^q\right) \nonumber\\
&\leq&\left\langle f_0(\bar y(t))+\sum^m_{i=1}f_i(\bar y(t))u_i,\varphi(t)\right\rangle+e^{-\lambda t}\left(\frac{1}{2}\|\bar y(t)-y_d\|^2_2+\gamma\|u\|_p^q\right)
\end{eqnarray}
for all $u\in U$.
\end{lem}
According to \eqref{optimality}, for $t\in]0,\infty[$ a.e. we look for the minimizer of the following function
\[
G_t(u):=\sum^m_{i=1}\langle f_i(\bar y(t)), \varphi(t)\rangle u_i+\gamma e^{-\lambda t}\|u\|^q_p,\ \forall\,u\in U.
\]
Assume that the set of control constraints $U$ has the form of box constraints:
\begin{equation}\label{eq:k1}
U_\infty:=\{u=(u_1,\ldots,u_m)\in\mathbb{R}^m\,:\, -\rho_i\leq u_i\leq \rho_i,\ i=1,\ldots,m\},
\end{equation}
where $\rho_i>0$. In this case the optimality condition can be used to derive the following structural properties of a minimizer.
\begin{thm}\label{THMswitching}
Let $\bar u$ be an optimal control for problem \eqref{ocpb} with $U_\infty$ given in \eqref{eq:k1}, let $\bar y$ be the associated optimal trajectory and $\varphi$ the associated adjoint state. For $t\in]0,\infty[$ a.e., we define the following index sets:
\[
I^-(t)=\{i\in\{1,\ldots,m\}\,:\, |\langle f_i(\bar y(t)),\varphi(t)\rangle|\rho_i^{1-q}<\gamma e^{-\lambda t}\},
\]
\[
I^0(t)=\{i\in\{1,\ldots,m\}\,:\, |\langle f_i(\bar y(t)),\varphi(t)\rangle|\rho_i^{1-q}=\gamma e^{-\lambda t}\},
\]
\[
I^+(t)=\{i\in\{1,\ldots,m\}\,:\, |\langle f_i(\bar y(t)),\varphi(t)\rangle|\rho_i^{1-q}>\gamma e^{-\lambda t}\},
\]
Then the following properties hold:
\begin{enumerate}[(i)]
\item
For $t\in ]0,\infty[$ a.e. and $i\in I^-(t)$,
\[
\bar u_i(t)=0.
\]
\item
For $t\in ]0,\infty[$ a.e. and $i\in I^0(t)$,
\[
\left\{
\begin{array}{lllll}
\bar u_i(t)=0, & \text{if}\ I^+(t)\neq\emptyset,\\[1.7ex]
\bar u_i(t)\in\{0,-\rho_i\mathop{sgn}(\langle f_i(\bar y(t)),\varphi(t)\rangle)\}, & \\
\bar u_i(t) \bar u_{j}(t)=0,\ i,j\in I^0(t),\ i\neq j, & \text{if}\ I^+(t)=\emptyset,\ q\in[p,1[,\\[1.7ex]
\bar u_i(t)\in [0,-\rho_i\mathop{sgn}(\langle f_i(\bar y(t)),\varphi(t)\rangle)], & \\
\bar u_i(t) \bar u_{j}(t)=0,\ i,j\in I^0(t),\ i\neq j, & \text{if}\ I^+(t)=\emptyset,\ q=1.
\end{array}
\right.
\]
\item
For $t\in ]0,\infty[$ a.e. and $i\in I^+(t)$, we have
\[
\bar u_i(t)\in\{0,-\rho_i\mathop{sgn}(\langle f_i(\bar y(t)),\varphi(t)\rangle)\},
\]
with $\max_{i\in I^+(t)}|\bar u_i(t)| \neq 0$.
\end{enumerate}
\end{thm}
Let us briefly comment on sparsity and switching properties which follow from Theorem \ref{THMswitching}. For the coordinates in the index set $I^-(t)$, the controllers are zero. We refer to these coordinates as the sparse control coordinates at the time $t$. If $I^+(t)= \emptyset$, then $i\in I^0(t) \cup I^-(t)$ for all $i=1,\dots, m$, and hence $u$ is switching or sparse at time $t$. If $I^+(t) \neq \emptyset$ then the coordinates in $I^0(t)$ behave like those in $I^-(t)$, they are $0$.
The coordinates of the optimal control in the index set $I^+(t)$ are not completely determined by (iii). They are either active, or zero and thus they join the set of sparse control coordinates.
Comparing to the case $p=q$ which was treated in \cite[Proposition 5.2]{KKR17}, the case (iii) is such that the control is necessarily active. Thus $p<q$ enhances additional sparsity compared to $p=q$. Finally, as a consequence of the box constraints, the optimal control is of bang-off-bang type, except for case (ii) with $q=1$.
\begin{proof}
We shall use that by Lemma \ref{le:k1} we know that $\bar u(t)$ minimizes $G_t$ in $U_\infty$ for a.e. $t\in (0,\infty)$. For convenience of notations, let us set
\[
\varphi_{t,i}=\langle f_i(\bar y(t)),\varphi(t)\rangle,\ \gamma_t=\gamma e^{-\lambda t}.
\]
In Step 1 below we verify (i) and (ii). The claims in (iii) are proved in Step 2.
\smallskip
{\bf Step 1: proof of (i) and (ii)}.\\
At first, let us focus on the case $q>p$. Consider further the case $\varphi_{t,i}\leq 0$ for $i=1,\ldots,m$. Then $\bar u_i(t)\geq 0$ for $i=1,\ldots,m$. We introduce
\[
\Omega:=\{u\in U_\infty\,:\,0\leq u_i\leq \rho_i,\ i=1,\ldots,m\}.
\]
Let us decompose $G_t$ in $\Omega$ as follows:
\[
G_t(u)=G_1(u)+\gamma_t G_2(u),
\]
where
\[
G_1(u)=\sum^m_{i=1}\varphi_{t,i}u_i+\gamma_t\sum^m_{i=1} u_i^q,\ G_2(u)=\left(\sum^m_{i=1}u_i^p\right)^{q/p}-\sum^m_{i=1}u_i^q.
\]
$G_1$ is a concave function in $\Omega$, $G_2\geq 0$, and $G_2=0$ if and only if $\sum_{i,j=1,\ldots,m,i\neq j}|u_iu_j|=0$. Here we use that $\frac{q}{p}\in[1,\frac{1}{p}]$ and the fact that
\[
\left(\sum^m_{i=1}a_i\right)^r\geq \sum^m_{i=1}a_i^r,
\]
for each $a_i\geq 0$, $r>1$, and equality holds if and only if $a_ia_j=0$ for all $i,j=1,\ldots,m$, $i\neq j$. Then we deduce that
\[
G_t(u)\geq G_1(u),
\]
and equality holds if and only if $\sum_{i,j=1,\ldots,m,i\neq j}|u_iu_j|=0$.
If $I^0(t)=\emptyset$ and $I^+(t)=\emptyset$, i.e. $\rho^{1-q}|\varphi_{t,i}|<\gamma_t$ for $i=1,\ldots,m$, we have
\[
\varphi_{t,i}u_i+\gamma_t u_i^q>0\ \text{for}\ u_i\in]0,\rho_i],
\]
where $u = u(t)$. Therefore $G_1$ attains its unique minimum at $(0,\ldots,0)$ and $G_2(0,\ldots,0)=0$. Consequently $\bar u_i(t)=0$ for $i=1,\ldots,m$.
If $I^+(t)=\emptyset$, we have
\[
\varphi_{t_i}u_i+\gamma_t u_i^q>0\ \text{for}\ u_i\in]0,\rho],\ i\in I^-(t),\ \text{and}\ \varphi_{t,j}u_j+\gamma_t u_j^q\geq 0\ \text{for}\ u_j\in[0,\rho_j],\ j\in I^0(t).
\]
Moreover for $j\in I^0(t)$, the expression $\varphi_{t,j}u_j+\gamma_t u_j^q$ attains its minimum in $[0,\rho_j]$ at $0$ and $\rho_j$ if $q<1$, and $\varphi_{t,j}u_j+\gamma_t u_j^q\equiv 0$ if $q=1$. Therefore,
\[
\bar u_i(t)=0\ \text{for}\ i\in I^-(t),\ \bar u_j(t)\in \{0,\rho_j\}\ \text{for}\ j\in I^0(t),\ q<1,\ \text{and}\ \sum_{j,j'\in I^0(t),j\neq j'}|u_j u_{j'}|=0,
\]
and
\[
\bar u_i(t)=0\ \text{for}\ i\in I^-(t),\ \bar u_j(t) \in [0,\rho_j]\ \text{for}\ j\in I^0(t),\ q=1,\ \text{and}\ \sum_{j,j'\in I^0(t),j\neq j'}|u_j u_{j'}|=0.
\]
If $I^+(t)\neq\emptyset$, we have
\[
\varphi_{t,i}u_i+\gamma u_i^q\geq 0\ \text{for}\ u_i\in[0,\rho_i],\ i\in I^-(t)\cup I^0(t),
\]
and $\varphi_{t,j}u_j+\gamma u_j^q$ attains its unique minimum in $[0,\rho_j]$ at $\rho_j$ for $j\in I^+(t)$. Thus, for any $u\in\Omega$, we define $\tilde u\in\Omega$ as follows:
\[
\tilde u_i=0\ \text{for}\ i\in I^-(t)\cup I^0(t),\ \text{and}\ \tilde u_i=u_i\ \text{for}\ i\in I^+(t).
\]
If $\tilde u=(0,\ldots,0)$, then for any $j\in I^+(t)$ we set $\hat u\in \Omega$ with
\[
\hat u_j=\rho\ \text{and}\ \hat u_i=0\ \text{for}\ i\neq j,\ i=1,\ldots,m.
\]
Thus we have
\[
G(u)\geq G_1(u)\geq G_1(\tilde u)>G_1(\hat u)=G(\hat u).
\]
Otherwise if $\tilde u\neq (0,\ldots,0)$ and $u\neq\tilde u$,
\[
G(u)=G_1(u)+\gamma_t G_2(u)> G_1(u)+\gamma_t G_2(\tilde u)\geq G_1(\tilde u)+\gamma_t G_2(\tilde u)=G(\tilde u).
\]
We then deduce that
\[
\bar u_i(t)=0\ \text{for}\ i\in I^-(t)\cup I^0(t).
\]
The proof for the case when $\varphi_{t,i}\leq 0$ for $i=1,\ldots,m$ is thus concluded. The other cases when $\varphi_{t,i}$ have different signs can be treated analogously.
Now we proceed to look at the case $q=p$. In this situation, $G_2\equiv 0$ and $G\equiv G_1$. The minimizers of $G_1$ have been analyzed in the previous arguments, and we therefore arrive at the conclusion.
\smallskip
{\bf Step 2: proof of (iii)}.\\
We turn to analyze the behavior of the coordinates with indices in
$I^+(t)$. In particular in this case $I^+(t)\neq\emptyset$, and consequently by (i) and (ii)
\[
\bar u_i(t)=0,\ \text{for}\ i\in I^-(t)\cup I^0(t).
\]
Therefore,
\begin{equation}\label{defGt+}
G_t(\bar u(t))=\sum^\ell_{\tau=1}\varphi_{t,i_\tau}\bar u_{i_\tau}(t)+\gamma_t\left(\sum^\ell_{\tau=1}|\bar u_{i_\tau}(t)|^p\right)^{q/p},
\end{equation}
where $\{i_1,\ldots,i_\ell\}\subset\{1,\ldots,m\}$ is such that $I^+(t)=\{i_1,\ldots,i_\ell\}$.
Then the problem consists in finding the minimizer of the function
\begin{equation}\label{EQtildeG}
\tilde G(w):=\sum^\ell_{\tau=1}\psi_\tau w_\tau + \gamma_t\left(\sum^\ell_{\tau=1}\rho_\tau^p |w_{\tau}|^p\right)^{q/p},\ \text{for}\ w=(w_1,\ldots,w_\ell)\in [-1,1]^\ell,
\end{equation}
where, to simplify notation, we set for $\tau=1,\ldots,\ell$
\begin{equation}\label{EQwpsi}
w_\tau=\frac{u_{i_\tau}}{\rho_{i_\tau}},\ \psi_\tau=\varphi_{t,i_\tau}\rho_{i_\tau},\ \text{and}\ \rho_\tau=\rho_{i_\tau}.
\end{equation}
Following the definition of $I^+(t)$, we have
\begin{equation}\label{EQpsi}
|\psi_\tau|\rho_\tau^{-q}>\gamma_t,\ \text{for}\ \tau=1,\ldots,\ell.
\end{equation}
Let $\bar w$ be the minimizer and let us start by considering the case
\[
\psi_\tau<0,\ \text{for all }\ \tau=1,\ldots,\ell.
\]
Then it is trivial to see that
\[
\bar w_\tau\geq 0,\ \text{for all}\ \tau=1,\ldots,\ell.
\]
We aim to prove that the minimizer $\bar w$ is not in the interior of $[0,1]^\ell$. Without loss of generality, we assume that
\[
1\geq \bar w_1\geq \bar w_2\geq\cdots\geq \bar w_\ell\geq 0.
\]
We can therefore limit our attention to the subset
\begin{equation}\label{EQuorder}
\{(w_1,\ldots,w_\ell)\,:\, 1\geq w_1\geq w_2\geq\cdots w_\ell\geq 0\}.
\end{equation}
Note that $\bar w$ can be expressed as $\bar w= (\beta_0 \bar w_1,\beta_1\beta_0 \bar w_1,\ldots,\beta_{\ell-1}\ldots\beta_0 \bar w_1)$ where $\beta_0=1$, and $\beta_\tau\in[0,1],\ \tau=1,\dots,\ell-1$. Moreover $\bar w_1 \in [0,1]$ is a minimizer of the functional
\[
G_{\beta,1}(w_1)=\sum^\ell_{\tau=1}\psi_\tau\beta_{\tau-1}\cdots\beta_0 w_1+\gamma_t\left(\sum^\ell_{\tau=1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_0^p\right)^{q/p}w_1^q.
\]
We will exclude the case that $ w_1 \to G_{\beta,1}(w_1)$ assumes a minimum in the interior of $[0,1]$.
Indeed, if such a minimum $w_1^*$ is attained in the interior of $[0,1]$, then
\[
0=G_{\beta,1}'(w_1^*)=\sum^\ell_{\tau=1}\psi_\tau\beta_{\tau-1}\cdots\beta_0+\gamma_t\left(\sum^\ell_{\tau=1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_0^p\right)^{q/p}q(w_1^*)^{q-1}.
\]
Therefore,
\[
G_{\beta,1}(w_1^*)=(1-q)\gamma_t\left(\sum^\ell_{\tau=1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_0^p\right)^{q/p}q(w_1^*)^{q}\geq 0.
\]
Note that
\[
G_{\beta,1}(1,0,\ldots,0)=\psi_1+\gamma_t\rho_1^q=\rho_1^q(-|\psi_1|\rho_1^{-q}+\gamma_t)<0,
\]
where \eqref{EQpsi} is applied. Thus,
\[
G_{\beta,1}(w_1^*)>G_{\beta,1}(1,0,\ldots,0),
\]
which contradicts the assumption that $w_1^*$ is the minimizer.
Consequently, the minimum can not be attained in the interior of $[0,1]$ and thus $\bar w_1\in\{0,1\}$. Moreover $G_{\beta,1}(1,0,\ldots,0)<0$ and $G_{\beta,1}(0,\ldots,0)=0$, and thus
\begin{equation}\label{EQw1}
\bar w_1=1.
\end{equation}
We next claim the following: for $j\in\{2,\ldots,\ell-1\}$, if $\bar w_{j-1}\in\{0,1\}$, then
\begin{equation}\label{EQClaim}
\bar w_{j}\in\{0,1\},
\end{equation}
and verify this statement by induction. If $\bar w_{j-1}=0$, by \eqref{EQuorder} we have
\[
\bar w_{j}=0
\]
as claimed. If $\bar w_{j-1}=1$, then
$
\bar w_\tau=1,\ \text{ for all } \,\tau=1,\ldots,j-1.
$
To characterize further $\bar w_j$, we apply the same idea as for determining $\bar w_1$. This time we restrict our attention to the subset
\[
\{(w_j,w_{j+1},\ldots, w_\ell)\,:\,1\geq w_j\geq w_{j+1}\geq \cdots\geq w_\ell\geq 0\},
\]
and note that for the optimal $(\bar w_j,\dots, \bar w_\ell)= (\bar w_j,\beta_j \bar w_j,\ldots, \beta_{\ell-1}\dots\beta_j \bar w_j)$,
where $\beta_\tau\in[0,1]$ for $\tau=1,\ldots,\ell-1$. We denote for any $w_j\in[0,1]$
\begin{eqnarray*}
G_{\beta,j}(w_j)&=&\sum^{j-1}_{\tau=1}\psi_\tau+\psi_jw_j+\sum^\ell_{\tau=j+1}\psi_\tau\beta_{\tau-1}\cdots\beta_jw_j\\
&&+\gamma_t\left[\sum^{j-1}_{\tau=1}\rho_\tau^p+\rho_j^p w_j^p+\sum^\ell_{\tau=j+1}\rho_\tau^p\beta^p_{\tau-1}\cdots\beta_j^p w_j^p\right]^{q/p},
\end{eqnarray*}
and note that $\bar w_j$ is a minimizer of $G_{\beta,j}$ on $[0,1]$.
If a minimum $w_j^*$ is attained in the interior of $[0,1]$, then
\[
G_{\beta,j}'(w_j^*)=0.
\]
This yields that,
\[
\psi_j+\sum^\ell_{\tau=j+1}\psi_\tau\beta_{\tau-1}\cdots\beta_j+\gamma_t\frac{q}{p}S_j^{q/p-1}\left(\rho_j^p+\sum^\ell_{\tau=j+1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_j^p\right)p(w_j^*)^{p-1}=0,
\]
where
\[
S_j=\sum^{j-1}_{\tau=1}\rho_\tau^p+\rho_j^p (w_j^*)^p+\sum^\ell_{\tau=j+1}\rho_\tau^p\beta^p_{\tau-1}\cdots\beta_j^p (w_j^*)^p.
\]
Therefore,
\[
\psi_j+\sum^\ell_{\tau=j+1}\psi_\tau\beta_{\tau-1}\cdots\beta_j=-\gamma_t S_j^{q/p-1}\left(\rho_j^p+\sum^\ell_{\tau=j+1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_j^p\right)q(w_j^*)^{p-1}.
\]
By applying the above equality to compute $G_{\beta,j}(w_j^*)$, we obtain
\begin{eqnarray*}
&& G_{\beta,j}(w_j^*)\\
&=& \sum^{j-1}_{\tau=1}\psi_\tau+\left(\psi_j+\sum^\ell_{\tau=j+1}\psi_\tau\beta_{\tau-1}\cdots\beta_j\right)w_j^*+\gamma_t S_j^{q/p}\\
&=& \sum^{j-1}_{\tau=1}\psi_\tau -\gamma_t S_j^{q/p-1}\left(\rho_j^p+\sum^\ell_{\tau=j+1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_j^p\right)q(w_j^*)^p+\gamma_t S_j^{q/p}\\
&=& \sum^{j-1}_{\tau=1}\psi_\tau+\gamma_t S_j^{q/p-1}\left[\sum^{j-1}_{\tau=1}\rho_
\tau^p + (1-q)\left(\rho_j^p+\sum^\ell_{\tau=j+1}\rho_\tau^p\beta_{\tau-1}^p\cdots\beta_j^p\right)(w_j^*)^p\right].
\end{eqnarray*}
Using the fact that $q\leq 1$ and $w_j^*>0$, it holds that
\begin{eqnarray*}
G_{\beta,j}(w_j^*)&\geq &\sum^{j-1}_{\tau=1}\psi_\tau+\gamma_t S_j^{q/p-1}\sum_{\tau=1}^{j-1}\rho_\tau^p\\
&>& \sum^{j-1}_{\tau=1}\psi_\tau+\gamma_t \left(\sum^{j-1}_{\tau=1}\rho_\tau^p\right)^{q/p-1}\sum_{\tau=1}^{j-1}\rho_\tau^p\\
&=& \sum^{j-1}_{\tau=1}\psi_\tau+\gamma_t \left(\sum^{j-1}_{\tau=1}\rho_\tau^p\right)^{q/p}\\
&=& G_{\beta,j}(0),
\end{eqnarray*}
which contradicts the assumption that $w_j^*$ is the minimizer.
Consequently, the minimum can not be attained in the interior of $[0,1]$. We then deduce that
\[
\bar w_j\in\{0,1\},
\]
which completes the proof for the claim \eqref{EQClaim}. Together with \eqref{EQw1}, it is deduced that
\[
\bar w_\tau\in\{0,1\},\ \text{for}\ \tau=1,\ldots,\ell,
\]
which concludes the case where $\psi_\tau<0$ for all $\tau=1,\ldots,\ell$.
For the other cases where $\psi_\tau$ is positive for some $\tau\in \{1,\ldots,\ell\}$, $\psi_\tau$ and $w_\tau$ can be replaced by $-\psi_\tau$ and $-w_\tau$ in \eqref{EQtildeG}. Then by following the same arguments as in the previously we can obtain that $-\bar w_\tau\in \{0,1\}$. Therefore we conclude that
\[
\bar w_\tau\in\{0,-\mathop{sgn}(\psi_\tau)\},\ \text{for}\ \tau=1,\ldots,\ell,
\]
with the additional information that $|\bar w_1|=1$.
The definition of $w_\tau$ and $\psi_\tau$ in \eqref{EQwpsi} implies that
\[
\bar u_{i_\tau}\in\{0,-\rho_{i_\tau}\mathop{sgn}(\varphi_{t,i_{\tau}})\},\ \text{for}\ i_\tau\in I^+(t),\ \tau=1,\ldots,\ell,
\]
with the additional information that $\max_{i\in I^+} \frac{|\bar u_i|}{\rho_i} \neq 0$. This completes the proof of (iii).
\end{proof}
In Theorem \ref{THMswitching} the study has been made for the case of box constraints. Next we briefly consider the problem under Euclidean norm constraints. In this case, due to the coupling of the coordinates which is inherent to the Euclidean norm, it appears to be more complicated
to achieve explicit information on the structure of the minimizers compared to that which was obtained for box constraints.
We define for $\rho>0$
\begin{equation}\label{eq:k2}
U_2:=\{u=(u_1,\ldots,u_m)\in\mathbb{R}^m\,:\,\sum^m_{i=1}u^2_i\leq \rho^2\}.
\end{equation}
\begin{thm}\label{THMswitching2}
Let $\bar u$ be an optimal control for problem \eqref{ocpb} with $U$ given in \eqref{eq:k2}, let $\bar y$ be the associated optimal trajectory, and $\varphi$ its associated adjoint state.
Let $I^-(t), I^0(t)$ and $I^+(t)$ be as defined in Theorem \ref{THMswitching}. If for some $t\in ]0,T[$ the cardinality of $I^+(t)$ is less or equal to 1,
then (i), (ii), and (iii) of that theorem remain valid. Otherwise we have
\begin{equation}\label{eqkk20}
\sum_{i=1}^m |\bar u_i(t)|^2 = \rho^2, \text{ for a.e. } t \in ]0,\infty[.
\end{equation}
\end{thm}
\begin{proof}
{ \em{Step 1.}} From \eqref{optimality} we know that for $t\in ]0,\infty[$ a.e., $\bar u(t)$ is the minimizer of the following function
\[
\bar G_t(u):=\sum^m_{i=1} \alpha_i(t) u_i +\gamma e^{-\lambda t}\|u\|^q_p,\ \text{ for all } \,u\in U_2,
\]
where $\alpha_i(t)=\langle f_i(\bar y(t)),\varphi(t)\rangle $.
At first we note that $U_2$ is a subset of $U_\infty$, if $\rho_i= \rho$ for all $i$, and hence $\min_{u\in U_\infty} \bar G_t(u) \le \min_{u\in U_2} \bar G_t(u)$. Moreover, if a minimizer of $\bar G_t$ over $U_\infty$ is contained in $U_2$, then this minimizer is also a minimizer of $\bar G_t$ over $U_2$.
Following this observation, let
$\bar u(t)$ be a minimizer of $\bar G_t$ over $U_\infty$ with cardinality
of $I^+(t) \le 1$. Then by Theorem \ref{THMswitching} all components of $\bar u(t)$ are 0 except for at most one. In case the cardinality of $I^+(t)$ equals one, then there is one non-trivial coordinate of the control at time $t$ whose norm then equals $\rho$.
{ \em{Step 2.}} Now we turn to the general case (assuming that $I^+$ is nonempty) and prove that the optimal control is necessarily active. Since $I^+(t)$ is non-empty there exists at least one index $\tau$ such that $\gamma_t-|\alpha_\tau(t)|\rho^{1-q}<0$. Setting the value of this coordinate equal to $\rho$ we obtain
\[
G(( 0,\ldots0,\rho,0,\ldots,0))=\alpha_\tau(t)\rho+\gamma_t \rho^q=\rho^q(\gamma_t+\alpha_\tau(t)\rho^{1-q})=\rho^q(\gamma_t-|\alpha_\tau(t)|\rho^{1-q})<0,
\]
which implies that at least one coordinate of $\bar u$ is nontrivial and $G(\bar u(t)) < 0$. Let $\tilde \ell$ denote the number of nontrivial coordinates of $\bar u$ and without loss of generality assume that these are the $\tilde \ell$ first ones of $\bar{u}(t)$.
Let us start with the case where $\alpha_i(t)\le 0$ for all $i=1,\ldots,\tilde \ell$. It is trivial to see that $\bar u_i(t)\geq 0$, for $i=1,\ldots, \tilde \ell$ in this case. We set
\[
\Omega:=\left\{(u_1,\ldots,u_{\tilde \ell})\in\mathbb{R}^{\tilde \ell}\, :\, \sum^{\tilde \ell}_{i=1}u_i^2\leq\rho^2,\ u_i\geq 0,\ i=1,\ldots,\tilde \ell \right\}.
\]
Thus $\bar u(t)\in\Omega$. We prove by contradiction that $\bar u\in\partial\Omega$. If this is not the case, i.e. $\bar u$ is in the interior of $\Omega$, then
\[
\frac{\partial G}{\partial u_i}(\bar u(t))=0,\ i=1,\ldots,\tilde \ell,
\]
from which we deduce that
\[
\alpha_i(t)+\gamma_t\frac{q}{p}\left(\sum^{\tilde \ell}_{j=1}|\bar u_j(t)|^p\right)^{q/p-1} p|\bar u_i(t)|^{p-1}=0,\ i=1,\ldots,{\tilde \ell}.
\]
It follows that
\[
\alpha_i(t) \bar u_i(t)=-\gamma_t q\left(\sum^{\tilde \ell}_{j=1}|\bar u_j(t)|^p\right)^{q/p-1}|\bar u_i(t)|^{p}=0,\ i=1,\ldots,{\tilde \ell}.
\]
Therefore,
\begin{eqnarray*}
G(\bar u(t))&=& \sum^{\tilde \ell}_{i=1}\alpha_i(t) \bar u_i(t) +\gamma_t\left(\sum^{\tilde \ell}_{i=1}|\bar u_i(t)|^p\right)^{q/p}\\
&=& -\gamma_t q\left(\sum^{\tilde \ell}_{i=1}|\bar u_i(t)|^p\right)^{q/p-1}|\bar u_i(t)|^{p}+\gamma_t\left(\sum^{\tilde \ell}_{i=1}|\bar u_i(t)|^p\right)^{q/p}\\
&=& \gamma_t (1-q) \left(\sum^{\tilde \ell}_{i=1}|\bar u_i(t)|^p\right)^{q/p} \; \geq \;0.
\end{eqnarray*}
Since we already know that $G(\bar u(t)) <0$ this gives a contradiction. Consequently $\bar u(t)\in\partial\Omega$. Since $\bar u_i \neq 0$ for $i=1,\dots,{\tilde \ell}$ this implies that
\begin{equation*}
\sum^{\tilde \ell}_{i=1}\bar u_i^2(t)=\rho^2\ \text{and}\ \bar u_i(t)> 0, \text{ for }\ i=1,\ldots,{\tilde \ell}.
\end{equation*}
If some of the coordinates of $\alpha$ are such that $\alpha_i(t) \geq 0$, then necessarily $\bar u_i(t) \leq 0$ and, adapting $\Omega$ accordingly, it can again be verified that $\sum_{1}^{l} \bar u_i^2(t) = \rho^2$.
\end{proof}
\section{Sparsity and switching properties: the time-discretized problem}
In this subsection we consider the following linear dynamical system: for $x\in\mathbb{R}^d$,
\begin{equation}\label{dslinear}
\left\{
\begin{array}{ll}
\dot y(t)=Ay(t)+Bu(t),\\
y(0)=x,
\end{array}
\right.
\end{equation}
where $A\in\mathbb{R}^{d\times d}$ and $B\in\mathbb{R}^{d\times m}$. Let us recall the optimal control problem: given $x\in\mathbb{R}^d$, consider
\[
\inf\left\{J^\Delta (x,u)\, :\, (y,u)\ \text{satisfies}\ \eqref{dslinear},\ u\in U^\Delta\right\}.
\]
The cost functional is recalled as follows:
\[
J^\Delta(x,u)=\int^\infty_0 \frac{1}{2}e^{-\lambda t}\|y(t)-y_d\|^2_2dt+\gamma R(u),
\]
where
\[
R(u)=\sum^\infty_{k=0}b_k\left(\sum^m_{i=1}|u_{i,k}|^p\right)^{q/p}\ \text{for}\ u\in U^\Delta.
\]
To investigate the optimality conditions satisfied by the optimal controllers, we introduce firstly the adjoint equation associated to $(y,u)$ satisfying \eqref{dslinear}:
\begin{equation}\label{Eqadjoint}
\left\{
\begin{array}{ll}
-\dot \varphi(t)=A^T\varphi(t)+e^{-\lambda t}(y(t)-y_d) & \text{for}\ t>0,\\
\lim_{t\rightarrow\infty} \varphi(t)=0.
\end{array}
\right.
\end{equation}
Here $\varphi$ is called the adjoint state of $y$. Since the controls in $U^\Delta$ are piecewise constant functions, we consider at first the optimal control on each time interval $I_k$, $k\in\mathbb{N}$.
\begin{prop}\label{propminIk}
{\color{black}Let $\tilde u\in U^\Delta$ satisfy the following: for any $k\in\mathbb{N}$, $\tilde u(\cdot)\equiv\tilde u_k$ in $I_k$ and
\begin{equation}\label{minIk}
\tilde u_k\in\mathop{arg\,min}_{u=(u_1,\ldots,u_m)\in U}\left\{\int_{I_k}\langle\tilde \varphi(t), B u \rangle dt +\gamma b_k\|u\|_p^q\right\},
\end{equation}
where $\tilde y$ is the corresponding trajectory and $\tilde \varphi$ is the adjoint state associated to $(\tilde y, \tilde u)$. }
Further for any arbitrary $\omega\in\mathbb{R}^m$ such that $\omega+\tilde u_k\in U$, we define the perturbed control
\[
u^\omega(t):=\left\{
\begin{array}{ll}
\tilde u_k+\omega & \text{if}\ t\in I_k,\\
\tilde u(t) & \text{otherwise}.
\end{array}
\right.
\]
Then it holds that $J^\Delta(x,u^\omega)\geq J^\Delta (x,\tilde u)$.
\end{prop}
\begin{proof}
Let $y^\omega$ be the trajectory associated with $u^\omega$. Then $y^\omega-\tilde y$ satisfies
\[
\left\{
\begin{array}{ll}
\dot y^\omega(t)-\dot{\tilde y}(t)=A(y^\omega(t)-y(t))+B\omega\mathbbm{1}_{I_k}(t) & \text{for}\ t\in(0,\infty),\\
y^\omega(0)-\tilde y(0)=0.
\end{array}
\right.
\]
Assumption \eqref{minIk} implies that
\begin{equation}\label{EqminIk}
\int_{I_k}\langle\tilde \varphi(t), B (\tilde u_k+\omega) \rangle dt +\gamma b_k\|\tilde u_k+\omega\|_p^q \geq \int_{I_k}\langle\tilde \varphi(t), B \tilde u_k \rangle dt +\gamma b_k\|\tilde u_k\|_p^q.
\end{equation}
By the definition of $u^\omega$, \eqref{EqminIk} is equivalent to
\begin{equation}\label{Eqvpomega}
\int_{I_k} \langle \tilde \varphi(t),B\omega\rangle dt+\gamma R(u^\omega)-\gamma R(\tilde u)\geq 0.
\end{equation}
For almost all $t>0$ we obtain,
\begin{eqnarray*}
&&\frac{d}{dt}\langle \tilde \varphi(t), y^\omega(t)-\tilde y(t)\rangle\\
&=& \langle \dot{\tilde \varphi}(t),y^\omega(t)-\tilde y(t)\rangle + \langle \tilde \varphi(t),\dot y^\omega(t)-\dot{\tilde y}(t)\rangle\\
&=& \langle -A^T\tilde \varphi(t)-e^{-\lambda t}(\tilde y(t)-y_d), y^\omega(t)-\tilde y(t)\rangle+\langle \tilde \varphi(t), A(y^\omega (t)-\tilde y(t))+B\omega \mathbbm{1}_{I_k}(t)\rangle \\
&=& -e^{-\lambda t}\langle \tilde y(t)-y_d, y^\omega(t)-\tilde y(t)\rangle +\langle \tilde \varphi(t), B\omega \mathbbm{1}_{I_k}(t)\rangle.
\end{eqnarray*}
Note that $\lim_{t\rightarrow\infty}\tilde \varphi(t)=0$ and $y^\omega(0)-\tilde y(0)=0$, and therefore
\[
\int^\infty_0 \frac{d}{dt}\langle \tilde \varphi(t), y^\omega(t)-\tilde y(t)\rangle=0.
\]
Consequently we obtain
\[
\int^\infty_0 \left[-e^{-\lambda t}\langle \tilde y(t)-y_d, y^\omega(t)-\tilde y(t)\rangle +\langle \tilde \varphi(t), B\omega \mathbbm{1}_{I_k}(t)\rangle\right]dt=0,
\]
i.e.,
\begin{equation}\label{EqBomega}
\int^\infty_0 e^{-\lambda t}\langle \tilde y(t)-y_d, y^\omega(t)-\tilde y(t)\rangle dt=\int_{I_k} \langle \tilde \varphi(t), B\omega\rangle.
\end{equation}
To compute the left-hand side of \eqref{EqBomega}, we have for every $t>0$
\begin{equation}\label{Eqsquare}
\|y^\omega(t)-y_d\|^2_2-\|\tilde y(t)-y_d\|^2_2=\|y^\omega(t)-\tilde y(t)\|^2_2+2\langle y^\omega(t)-\tilde y(t),\tilde y(t)-y_d\rangle,
\end{equation}
and now \eqref{Eqvpomega}, \eqref{EqBomega} and \eqref{Eqsquare} imply that
\begin{align*}
&&\int^\infty_0 \frac{1}{2}e^{-\lambda t}\|y^\omega(t)-y_d\|^2_2dt- \int^\infty_0 \frac{1}{2}e^{-\lambda t}\|\tilde y(t)-y_d\|^2_2dt \\ &&-\int^\infty_0 \frac{1}{2}e^{-\lambda t}\|y^\omega(t)-\tilde y(t)\|^2_2dt+\gamma R(u^\omega)-\gamma R(\tilde u)\geq 0.
\end{align*}
Then we deduce tha
\[
J^\Delta (x,u^\omega)-J^\Delta (x,\tilde u)\geq \int^\infty_0 \frac{1}{2}e^{-\lambda t}\|y^\omega(t)-\tilde y(t)\|^2_2dt \geq 0,
\]
which ends the proof.
\end{proof}
Proposition \ref{propminIk} provides the way to construct optimal controls on each $I_k$, and this procedure can be naturally extended to construct globally optimal controls.
\begin{thm}\label{thmminimizer}
{\color{black}Let $\bar u\in U^\Delta$ satisfy the following: for any $k\in \mathbb{N}$ and $t\in I_k$,
\begin{equation}\label{umin}
\bar u(t)\in\mathop{arg\,min}_{u=(u_1,\ldots,u_m)\in U}\left\{\int_{I_k}\langle\bar \varphi(t), B u \rangle dt +\gamma b_k\|u\|_p^q\right\},
\end{equation}
where $\bar y$ is the corresponding trajectory and $\bar \varphi$ is the adjoint state associated to $(\bar y, \bar u)$.}
Then $\bar u \in U^\Delta$ is a minimizer of problem \eqref{ocpdtd}, i.e.
\[
J^\Delta (x,u)\geq J^\Delta (x,\bar u),\ \forall\,u\in U^\Delta.
\]
\end{thm}
\begin{proof}
For any $u\in U^\Delta$, we define a sequence $(u^n)_{n\in\mathbb{N}}$ by
\[
u^n(t):=\left\{
\begin{array}{ll}
\bar u(t) & \text{if}\ t\in I_k,\ k=0,\ldots,n,\\
u(t) & \text{otherwise}.
\end{array}
\right.
\]
Therefore, $u^n\rightarrow \bar u$ pointwise in $[0,\infty[$. Let $y^n$ be the trajectory associated with $u^n$. By the same argument as in Theorem \ref{thmexistence}, we deduce that
\[
y^n\rightarrow \bar y\ \text{pointwise in}\ [0,\infty[.
\]
Assumption \eqref{umin} and Proposition \ref{propminIk} imply that
\[
J^\Delta (x,u)\geq J^\Delta (x,u^0)\geq J^\Delta (x,u^1)\geq \cdots \geq J^\Delta (x,u^n),\ \forall\,n\in\mathbb{N}.
\]
By Fatou's Lemma,
\begin{eqnarray*}
J^\Delta (x,u)&\geq& \mathop{\lim\,\inf}_{n\rightarrow\infty} \left\{\int^\infty_0 \frac{1}{2} e^{-\lambda t}\|y^n(t)-y_d\|^2_2 dt +\gamma R(u^n)\right\}\\
&\geq& \int^\infty_0 \frac{1}{2} e^{-\lambda t}\|\bar y(t)-y_d\|^2_2 dt +\gamma R(\bar u)\\
&=& J^\Delta (x,\bar u),
\end{eqnarray*}
and we conclude that $\bar u$ is a minimizer of problem \eqref{ocpdtd}.
\end{proof}
Based on the optimality conditions \eqref{umin}, similar results on sparsity and switching properties as Theorem \ref{THMswitching} can be deduced by the same arguments as in the proof of Theorem \ref{THMswitching}.
\begin{thm}
Following the same assumptions and notations in Theorem \ref{thmminimizer}, we set
\[
\varphi_k=\int_{I_k}B^T{\color{black}\bar \varphi(t)}dt,\ \gamma_k=\gamma b_k.
\]
For each $k\in\mathbb{N}$, we define the following index sets:
\[
I^-_k=\{i\in\{1,\ldots,m\}\,:\, |\varphi_{k,i}|\rho_i^{1-q}<\gamma_k \},
\]
\[
I^0_k=\{i\in\{1,\ldots,m\}\,:\, |\varphi_{k,i}|\rho_i^{1-q}=\gamma_k \},
\]
\[
I^+_k=\{i\in\{1,\ldots,m\}\,:\, |\varphi_{k,i}|\rho_i^{1-q}>\gamma_k \},
\]
The following properties hold:
\begin{enumerate}[(i)]
\item
For $k\in\mathbb{N}$, $t\in I_k$ and $i\in I^-_k$,
\[
\bar u_i(t)=0.
\]
\item
For $k\in\mathbb{N}$, $t\in I_k$ and $i\in I^0_k$,
\[
\left\{
\begin{array}{lllll}
\bar u_i(t)=0, & \text{if}\ I^+_k\neq\emptyset,\\[1.7ex]
\bar u_i(t)\in\{0,-\rho_i\mathop{sgn}(\varphi_{k,i})\}, & \\
\bar u_i(t) \bar u_{j}(t)=0,\ i,j\in I^0_k,\ i\neq j, & \text{if}\ I^+_k=\emptyset,\ q\in[p,1[,\\[1.7ex]
\bar u_i(t)\in [0,-\rho_i\mathop{sgn}(\varphi_{k,i})], & \\
\bar u_i(t) \bar u_{j}(t)=0,\ i,j\in I^0_k,\ i\neq j, & \text{if}\ I^+_k=\emptyset,\ q=1.
\end{array}
\right.
\]
\item
For $k\in\mathbb{N}$, $t\in I_k$ and $i\in I^+_k$, we have
\[
\bar u_i(t)\in\{0,-\rho_i\mathop{sgn}(\varphi_{k,i})\},
\]
and $\max_{i I_k^+}|\bar u_i(t)| \neq 0$.
\end{enumerate}
\end{thm}
\section{Numerical experiments}
In this section we present numerical experiments for the computation of optimal control laws for the problem
\begin{equation*}
\underset{u(\cdot)\in L^{\infty}(0,\infty;U_\infty)}{\inf}J(x,u):=\int^{\infty}_0 e^{-\lambda t}\left(\frac{1}{2}\|y(t)-y_d\|^2_2+\gamma\|u(t)\|_p^q\right)dt,
\end{equation*}
constrained to the nonlinear dynamical system
\begin{equation*}
\left\{
\begin{array}{ll}
\dot y(s)=f(y(s),u(s)):=f_0(y(s))+\sum^m_{i=1}f_i(y(s))u_i(s) & \text{in}\ ]0,\infty[,\\
y(0)=x.
\end{array}
\right.
\end{equation*}
For the realization of globally optimal control laws we proceed as in \cite{KKR17}, i.e. by following a dynamic programming approach. The value function $V(x):=\inf J(x,u)$ associated to this infinite horizon optimal control problem satisfies the following first order Hamilton-Jacobi-Bellman equation
\begin{equation*}
\lambda V(x)+\underset{u\in U_\infty}{\sup}\{-f(x,u)\cdot\nabla V(x)-\frac{1}{2}\|x-y_d\|^2_2-\gamma\|u\|_p^q\}=0\,,
\end{equation*}
which leads to the optimal feedback map
\begin{equation} \label{eq6.30}
\bar{u}(x):=\underset{u\in U_\infty}{\text{argmin}}\left\{f(x,u)\cdot\nabla V(x)+\gamma\|u\|_p^q \right\}\,.
\end{equation}
The solution of the Hamilton-Jacobi-Bellman equation and of the optimal feedback mapping are numerically approximated by a first-order semi-Lagrangian scheme with policy iteration as discussed in \cite{Alla15}. The well-posedness of this numerical scheme is guaranteed under boundedness and continuity assumptions for the dynamics $f(x,u)$ and the cost. Convergence of controls, however, is only guaranteed for convex running costs. Nevertheless, the results we report indicate that the semi-Lagrangian scheme converges to optimal controls exhibiting the expected sparsity and switching properties. This scheme has also been applied to the solution of sparse optimal feedback control problems in \cite{Albi17,Falcone14}. In the case $p=q=1$ the minimization operation in \eqref{eq6.30} can be realized by means of semismooth Newton methods as \cite{KKK16}. For different values of $p$ and $q$, the minimizer is chosen by discretizing the control set $U_\infty$ into a finite number of values and making a pointwise evaluation of the Hamiltonian.
\subsubsection*{Eikonal dynamics}
We begin by considering eikonal-type dynamics for planar motion of the form
\begin{align*}
\dot x_1(s)&=u_1(s)\\
\dot x_2(s)&=u_2(s)\,,
\end{align*}
where $|u_i(s)|\leq 0.5$ for $i=1,2.$ The state space is set to be $\Omega=[-1,1]^2$, the discount factor $\lambda=0.2$, and $\gamma=1$. The goal is to drive the state to the origin, and therefore $y_d=(0,0)$. The optimal control fields in the state space for different $p,q$ values are shown in Figure \ref{eikcon}.
\begin{figure}[!ht]
\centering
\subfloat[$p=1, q=1$]{\label{fig:t1b}\includegraphics[width=0.33\textwidth]{controlp1q1}}\hfill
\subfloat[$p=0.8, q=1$]{\label{fig:t1c}\includegraphics[width=0.33\textwidth]{controlp08q1}}\hfill
\subfloat[$p=0.6, q=1$]{\label{fig:t1d}\includegraphics[width=0.33\textwidth]{controlp06q1}}\\
\subfloat[$p=0.2, q=0.2$]{\label{fig:t1e}\includegraphics[width=0.33\textwidth]{controlp02q02}}\hfill
\subfloat[$p=0.2, q=0.3$]{\label{fig:t1f}\includegraphics[width=0.33\textwidth]{controlp02q03}}\hfill
\subfloat[$p=0.2, q=0.8$]{\label{fig:t1g}\includegraphics[width=0.33\textwidth]{controlp02q08}}
\caption{Eikonal dynamics, optimal control fields for different control penalizations$\|u\|_p^q$.}\label{eikcon}
\end{figure}
We observe the following:
\begin{itemize}
\item[a)] The case $p=q=1$ has been already reported in \cite{KKR17}. There exists a switching band of width $\gamma\lambda$, where the optimal control points unidirectionally towards the origin, and $\bar{u}=0$ for $\|u\|_{\infty}\leq\gamma\lambda$.
\item[b,c)] Departing from $p=q=1$ and reducing the value of $p$, a switching region with only one active control component arises. It increases as the ratio $q/p$ increases. Note that for $q=1$, the region where $\bar{u}=0$ remains unchanged.
\item[d)] The switching and the sparsity regions are larger for $p=q=0.2$ than for $p=q=1$. Only in the particular case $\rho=1$ these regions would remain the same.
\item[e,f)] Increasing the $q/p$ ratio by departing from smaller values of $q$ generates a larger switching region, leading to a fully switching controller for a ratio of $q/p$ sufficiently large. Note that increasing $q/p$ for $q\neq 1$ also leads to a decrease of the sparsity region.
\end{itemize}
\subsubsection*{Nonlinear dynamics of a double-well potential}
We now address the synthesis of optimal controllers for nonlinear dynamics. We consider a system corresponding to a single one-dimensional particle moving in a double-well potential, subject to a controlled damping, and a direct external forcing via
\begin{align*}
\dot x(s)&=v(s)\\
\dot v(s)&= -(1+u_1(s))v(s)+(x(s)-x^3(s)) + u_2(s).
\end{align*}
In the absence of control action ($u_1=u_2=0$), the damped particle has two stable equilibrium positions, namely $x=\pm 1,v=0$ (we drop the state-space notation $(x_1,x_2)$ for $(x,v)$), with their corresponding basins of attraction. Here our goal is to steer the particle to the equilibrium $y_d=(1,0)$. We consider a set of initial conditions in $\Omega=[-2,2]^2$, and set $\gamma=0.1$, $\rho=1$, and $\lambda=0.01$. Optimal controls are shown in Figure \ref{dw1}.
\begin{figure}[!ht]
\subfloat[$u_1 ,p=1, q=1$]{\label{fig:t2a}\includegraphics[width=0.33\textwidth]{control1dwp1q1}}\hfill
\subfloat[$u_1, p=0.6, q=1$]{\label{fig:t2b}\includegraphics[width=0.33\textwidth]{control1dwp06q1}}\hfill
\subfloat[$u_1, p=0.2, q=1$]{\label{fig:t2c}\includegraphics[width=0.33\textwidth]{control1dwp02q1}}\\
\subfloat[$u_2, p=1, q=1$]{\label{fig:t2d}\includegraphics[width=0.33\textwidth]{control2dwp1q1}}\hfill
\subfloat[$u_2, p=0.6, q=1$]{\label{fig:t2e}\includegraphics[width=0.33\textwidth]{control2dwp06q1}}\hfill
\subfloat[$u_2, p=0.2, q=1$]{\label{fig:t2f}\includegraphics[width=0.33\textwidth]{control2dwp02q1}}\\
\subfloat[$\|u\|_0, p=1, q=1$]{\label{fig:t2g}\includegraphics[width=0.33\textwidth]{activedwp1q1}}\hfill
\subfloat[$\|u\|_0, p=0.6, q=1$]{\label{fig:t2h}\includegraphics[width=0.33\textwidth]{activedwp06q1}}\hfill
\subfloat[$\|u\|_0, p=0.2, q=1$]{\label{fig:t2i}\includegraphics[width=0.33\textwidth]{activedwp02q1}}
\caption{Optimal controls for the double-well nonlinear control problem. The first two rows show the control variables $u_1$ and $u_2$ for different values of $p$ and $q$.}\label{dw1}
\end{figure}
\noindent We observe:
\begin{itemize}
\item [a,b,c)] By reducing the value of $p$ with $q=1$, the region where the control $u_1$ is active decreases.
\item [d,e,f)] Reducing $q$ does not affect the sparsity pattern of $u_2$. The linear control action via $u_2$ is more relevant for the stabilization goal than the bilinear control term $u_1v$. As expected it becomes insignificant as $v$ becomes small.
\item[g,h,i)] Overall, the reduction of $p$ has a significant effect on the increase of the switching region.
\end{itemize}
\noindent In order to investigate a setting with a richer interplay between the control variables and the switching structure, we consider a modified version of the double-well control system given by
\begin{align*}
\dot x(s)&=v(s)\\
\dot v(s)&= -(1+u_1(s))v(s)+(x(s)-x^3(s)) + u_2(s)x(s),
\end{align*}
where $u_2$ enters now in a bilinear fashion. The optimal controllers are significantly different compared to the previous setting, as shown in Figure \ref{dw2}.
\begin{figure}[!ht]
\subfloat[$u_1 ,p=1, q=1$]{\label{fig:t3a}\includegraphics[width=0.33\textwidth]{control1dw4p1q1}}\hfill
\subfloat[$u_1 ,p=0.2, q=0.6$]{\label{fig:t3b}\includegraphics[width=0.33\textwidth]{control1dw4p02q06}}\hfill
\subfloat[$u_1 ,p=0.2, q=1$]{\label{fig:t3c}\includegraphics[width=0.33\textwidth]{control1dw4p02q1}}\\
\subfloat[$u_2,p=1, q=1$]{\label{fig:t3d}\includegraphics[width=0.33\textwidth]{control2dw4p1q1}}\hfill
\subfloat[$u_2 ,p=0.2, q=0.6$]{\label{fig:t3e}\includegraphics[width=0.33\textwidth]{control2dw4p02q06}}\hfill
\subfloat[$u_2 ,p=0.2, q=1$]{\label{fig:t3f}\includegraphics[width=0.33\textwidth]{control2dw4p02q1}}\\
\subfloat[$\|u\|_0 ,p=1, q=1$]{\label{fig:t3g}\includegraphics[width=0.33\textwidth]{activedw4p1q1}}\hfill
\subfloat[$\|u\|_0 ,p=0.2, q=0.6$]{\label{fig:t3h}\includegraphics[width=0.33\textwidth]{activedw4p02q06}}\hfill
\subfloat[$\|u\|_0 ,p=0.2, q=1$]{\label{fig:t3i}\includegraphics[width=0.33\textwidth]{activedw4p02q1}}
\caption{Optimal controls for the double-well control problem with two bilinear controls. The first two rows show the control variables $u_1$ and $u_2$ for different values of $p$ and $q$.}\label{dw2}
\end{figure}
\noindent We note that:
\begin{itemize}
\item [a,b,c)] The sparsity region of $u_1$ increases as the ratio $q/p$ increases.
\item [d,e, f)] The sparsity region of $u_2$ also increases as $q/p$ increases.
\item [g,h,i)] Overall, the switching pattern of the two control variables becomes dominant as the ratio $q/p$ becomes large. Only a reduced region of the state space requires the simultaneous action of two control variables.
\end{itemize}
\paragraph{Concluding remarks.} In this paper we have studied infinite horizon optimal control problems with a control cost of the form $\| u \|_p^q$, where $0<p\leq q\leq 1$, leading to a non-convex, non-smooth optimization problem. From the analysis of the associated optimality conditions, we have shown that such control penalizations induce not only sparsity, but also a switching structure in the optimal control field. The switching pattern is determined by the different parameters of the control problem, but most notably, by the value of $q$ and the ratio $q/p$. By means of dynamic programming techniques, we have shown numerically that, for an increased $q/p$ ratio the optimal control has a dominant switching pattern, tending to minimize a counting $\|\cdot\|_0$ measure over an enlarged region of the state space. We believe that an important direction for future research is a thorough study of the interplay between the underlying dynamical structure of the control system and the switching pattern. More concretely, it would be desirable to know whether the sparse/switching control does benefit from the basin of attraction of a given equilibrium point, or whether the inclusion of $\|\cdot\|_p^q$ norms could lead to minimum time-type controllers.
|
\section{Introduction}
In this paper, we first investigate the Liouville property of nonnegative solutions to the following non-critical higher order Lane-Emden-Hardy equations
\begin{equation}\label{PDE}\\\begin{cases}
(-\Delta)^{m}u(x)=\frac{u^{p}(x)}{|x|^{a}} \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \mathbb{R}^{n}, \\
u(x)\geq0, \,\,\,\,\,\,\,\, x\in\mathbb{R}^{n},
\end{cases}\end{equation}
where $u\in C^{2m}(\mathbb{R}^{n})$ if $-\infty<a\leq0$, $u\in C^{2m}(\mathbb{R}^{n}\setminus\{0\})\cap C^{2m-2}(\mathbb{R}^{n})$ if $0<a<2m$, $n\geq3$, $1<p<\frac{n+2m-2a}{n-2m}$ if $0\leq a<2$, and $1<p<+\infty$ if $2\leq a<2m$.
For $0<\alpha\leq n$, PDEs of the form
\begin{equation}\label{GPDE}
(-\Delta)^{\frac{\alpha}{2}}u(x)=\frac{u^{p}(x)}{|x|^{a}}
\end{equation}
are called the fractional order or higher order Hardy (Lane-Emden, H\'{e}non) equations for $a>0$ ($a=0$, $a<0$, respectively), which have many important applications in conformal geometry and Sobolev inequalities. We say equations \eqref{GPDE} have critical order if $\alpha=n$ and non-critical order if $0<\alpha<n$. Liouville type theorems for equations \eqref{GPDE} (i.e., nonexistence of nontrivial nonnegative solutions) have been quite extensively studied (see \cite{BG,CD,CDQ,CFY,CL,CL1,CLiu,DQ1,DQ,GS,Lei,Lin,MP,P,PS,WX} and the references therein). It is crucial in establishing a priori estimates and existence of positive solutions for non-variational boundary value problems of a class of elliptic equations (see \cite{BM,CDQ,CL3,CL4,GS1,PQS}).
In the special case $a=0$, equation \eqref{GPDE} becomes the well-known Lane-Emden equation, which also arises as a model in astrophysics. For $\alpha=2$ and $1<p<p_{s}:=\frac{n+2}{n-2}$ ($:=\infty$ if $n=2$), Liouville type theorem was established by Gidas and Spruck in their celebrated article \cite{GS}. Later, the proof was simplified to a large extent by Chen and Li in \cite{CL} using the Kelvin transform and the method of moving planes (see also \cite{CL1}). For $n>\alpha=4$ and $1<p<\frac{n+4}{n-4}$, Lin \cite{Lin} proved the Liouville type theorem for all the nonnegative $C^{4}(\mathbb{R}^{n})$ smooth solutions of \eqref{GPDE}. When $\alpha\in(0,n)$ is an even integer and $1<p<\frac{n+\alpha}{n-\alpha}$, Wei and Xu established Liouville type theorem for all the nonnegative $C^{\alpha}(\mathbb{R}^{n})$ smooth solutions of \eqref{GPDE} in \cite{WX}. For general $a\in\mathbb{R}$, $0<\alpha\leq n$, $0<p<\min\{\frac{n+\alpha-2a}{n-\alpha},\frac{n+\alpha-a}{n-\alpha}\}$ ($1<p<+\infty$ if $\alpha=n$), there are also lots of literatures on Liouville type theorems for general fractional order or higher order Hardy-H\'{e}non equations \eqref{GPDE}, for instance, Bidaut-V\'{e}ron and Giacomini \cite{BG}, Chen, Dai and Qin \cite{CDQ}, Chen and Fang \cite{CF}, Cheng and Liu \cite{CLiu}, Dai and Qin \cite{DQ1}, Gidas and Spruck \cite{GS}, Lei \cite{Lei}, Mitidieri and Pohozaev \cite{MP}, Phan \cite{P}, Phan and Souplet \cite{PS} and many others. For Liouville type theorems on systems of PDEs of type \eqref{GPDE} with respect to various types of solutions (e.g., stable, radial, nonnegative, sign-changing, $\cdots$), please refer to \cite{BG,DQ,FG,M,P,PQS,S,SZ} and the references therein.
For the critical nonlinearity cases $p=\frac{n+\alpha}{n-\alpha}$ with $a=0$ and $0<\alpha<n$, the quantitative and qualitative properties of solutions to fractional order or higher order conformally invariant equations \eqref{GPDE} have also been widely studied. In the special case $n>\alpha=2$, equation \eqref{GPDE} becomes the well-known Yamabe problem (for related results, please see Gidas, Ni and Nirenberg \cite{GNN1,GNN}, Caffarelli, Gidas and Spruck \cite{CGS} and the references therein). For $n>\alpha=4$, Lin \cite{Lin} classified all the positive $C^{4}$ smooth solutions of \eqref{GPDE}. In \cite{WX}, among other things, Wei and Xu proved the classification results for all the positive $C^{\alpha}$ smooth solutions of \eqref{GPDE} when $\alpha\in(0,n)$ is an even integer. For $n>\alpha=3$, Dai and Qin \cite{DQ1} classified the positive $C^{3,\epsilon}_{loc}\cap\mathcal{L}_{1}$ classical solutions of \eqref{GPDE}. In \cite{CLO}, by developing the method of moving planes in integral forms, Chen, Li and Ou classified all the positive $L^{\frac{2n}{n-\alpha}}_{loc}$ solutions to the equivalent integral equation of the PDE \eqref{GPDE} for general $\alpha\in(0,n)$, as a consequence, they obtained the classification results for positive weak solutions to PDE \eqref{GPDE}. Subsequently, Chen, Li and Li \cite{CLL} developed a direct method of moving planes for fractional Laplacians $(-\Delta)^{\frac{\alpha}{2}}$ with $0<\alpha<2$ and classified all the $C^{1,1}_{loc}\cap\mathcal{L}_{\alpha}$ positive solutions to the PDE \eqref{GPDE} directly as an application, where the function space
\begin{equation}\label{0-0}
\mathcal{L}_{\alpha}(\mathbb{R}^{n}):=\Big\{f: \mathbb{R}^{n}\rightarrow\mathbb{R}\,\big|\,\int_{\mathbb{R}^{n}}\frac{|f(x)|}{1+|x|^{n+\alpha}}dx<\infty\Big\}.
\end{equation}
In the limiting (i.e., critical order) cases $\alpha=n$, there are also a large amount of literatures on classification results for positive solutions to the following critical order conformally invariant equations with exponential nonlinearities
\begin{equation}\label{0-3}
(-\Delta)^{\frac{n}{2}}u=(n-1)!e^{nu},
\end{equation}
for instance, Chen and Li \cite{CL1}, Chang and Yang \cite{CY}, Chen and Zhang \cite{CZ}, Lin \cite{Lin}, Wei and Xu \cite{WX} and Zhu \cite{Zhu}. For more literatures on the quantitative and qualitative properties of solutions to fractional order or higher order conformally invariant PDE and IE problems, please refer to \cite{CD,CL1,CZ,DFHQW,DFQ,Zhu} and the references therein.
In this paper, we will establish Liouville type theorem for nonnegative classical solutions of \eqref{PDE} in the cases $1<p<\frac{n+2m-2a}{n-2m}$ if $0\leq a<2$ and $1<p<+\infty$ if $2\leq a<2m$. Lei \cite{Lei} has proved the nonexistence of positive solutions to \eqref{PDE} for $0\leq a<2$ and $1<p<\frac{n-a}{n-2m}$. One should note that, our results extend the range $p\in(1,\frac{n-a}{n-2m})$ and $0\leq a<2$ in \cite{Lei} to the full range $1<p<\frac{n+2m-2a}{n-2m}$ if $0\leq a<2$ and $1<p<+\infty$ if $2\leq a<2m$.
Our Liouville type result for \eqref{PDE} is the following theorem.
\begin{thm}\label{Thm0}
Assume $n\geq3$, $1\leq m<\frac{n}{2}$, $0\leq a<2m$, $1<p<\frac{n+2m-2a}{n-2m}$ if $0\leq a<2$, $1<p<+\infty$ if $2\leq a<2m$, and $u$ is a nonnegative solution of \eqref{PDE}. If one of the following two assumptions
\begin{equation*}
0\leq a\leq2+2p \,\,\,\,\,\,\,\,\,\,\,\, \text{or} \,\,\,\,\,\,\,\,\,\,\,\, u(x)=o(|x|^{2}) \,\,\,\, \text{as} \,\, |x|\rightarrow+\infty
\end{equation*}
holds, then $u\equiv0$ in $\mathbb{R}^{n}$.
\end{thm}
\begin{rem}\label{remark1}
In \cite{CLiu}, Cheng and Liu proved Liouville type theorem for \eqref{PDE} in the cases $a<0$ and $1<p<\frac{n+2m-a}{n-2m}$ (there is actually an extra assumption $p>\frac{n}{n-2m}$ in \cite{CLiu}, but it is clear from their proof that the assumption $p>\frac{n}{n-2m}$ is redundant and unnecessary). Among other things, Lei \cite{Lei} established the nonexistence of positive solutions to \eqref{PDE} for $0\leq a<2$ and $1<p<\frac{n-a}{n-2m}$. However, we found a few technical mistakes in their proof, more precisely, in their proof of super poly-harmonic properties (see Theorem 2 in \cite{CLiu} and Theorem 2.1 in \cite{Lei}). For instance, the possibility that constant $C_{\ast}=0$ have to be ruled out in the proof of Theorem 2 in \cite{CLiu}, and a factor $R^{-a}$ should be added to the last inequality in the proof of Theorem 2.1 in \cite{Lei} since $R$ is sufficiently large (thus the assumption $a<2$ is needed therein). In this paper, we will prove the super poly-harmonic properties in Theorem \ref{lemma0} via a unified approach for both $a<0$ and $a\geq0$, as a consequence, we repair the proof in \cite{CLiu} and extend the results in \cite{Lei}.
\end{rem}
\begin{rem}\label{remark3}
For $0<a<2m$, if we consider the nonnegative solutions $u\in C^{2m}(\mathbb{R}^{n}\setminus\{0\})\cap C(\mathbb{R}^{n})$, then it is clear from our proof of Theorem \ref{Thm0} that Liouville theorem as Theorem \ref{Thm0} also holds for $1<p<\frac{n+2m-2a}{n-2m}$ (see Section 2). The main difference is, instead of Theorem \ref{lemma0}, we will show super poly-harmonic properties except the origin $0\in\mathbb{R}^{n}$, that is, $(-\Delta)^{i}u\geq0$ in $\mathbb{R}^{n}\setminus\{0\}$ for $i=1,\cdots,m-1$ (see remark \ref{remark4}).
\end{rem}
\begin{rem}\label{remark0}
In Theorem \ref{Thm0} and Remark \ref{remark3}, the smoothness assumption on $u$ at $x=0$ is necessary. Equation \eqref{PDE} admits a distributional solution of the form $u(x)=C|x|^{-\sigma}$ with $\sigma=\frac{2m-a}{p-1}>0$.
\end{rem}
We also consider the following higher order Navier problem
\begin{equation}\label{tNavier}\\\begin{cases}
(-\Delta)^{m}u(x)=u^{p}(x)+t \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \Omega, \\
u(x)=-\Delta u(x)=\cdots=(-\Delta)^{m-1}u(x)=0 \,\,\,\,\,\,\,\, \text{on} \,\,\, \partial\Omega,
\end{cases}\end{equation}
where $n\geq3$, $1\leq m<\frac{n}{2}$, $\Omega\subset\mathbb{R}^{n}$ is a bounded domain with $C^{2m-2}$ boundary $\partial\Omega$ and $t\geq0$.
Theorem 6 in Chen, Fang and Li \cite{CFL} implies immediately the following a priori estimates for any positive solution $u$ to \eqref{tNavier}.
\begin{thm}\label{Thm-CFL} (\cite{CFL})
Assume $\frac{n}{n-2m}<p<\frac{n+2m}{n-2m}$. Then, for any positive solution $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ to the higher order Navier problem \eqref{tNavier}, we have
\begin{equation*}
\|u\|_{L^{\infty}(\overline{\Omega})}\leq C(n,m,p,\Omega).
\end{equation*}
\end{thm}
As an application of the Liouville theorems (Theorem \ref{Thm0}), we can prove the following a priori estimates for any positive solution $u$ to \eqref{tNavier} via the method of moving planes in local way and blowing-up methods (for related literatures on these methods, please see \cite{BC,BM,CDQ,CL3,CL4,CY0,Li,SZ1}). Our a priori estimates extend the range of $p$ in Theorem \ref{Thm-CFL} remarkably.
\begin{thm}\label{Thm1}
Assume $1<p<\frac{n+2m}{n-2m}$. If one of the following two assumptions
\begin{equation*}
\text{i)} \,\,\, \Omega \,\, \text{is strictly convex}, \,\, 1<p<\frac{n+2m}{n-2m}, \quad\quad\quad\, \text{or} \quad\quad\quad\, \text{ii)} \,\,\, 1<p\leq\frac{n+2}{n-2}
\end{equation*}
holds. Then, for any positive solution $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ to the higher order Navier problem \eqref{tNavier}, we have
\begin{equation*}
\|u\|_{L^{\infty}(\overline{\Omega})}\leq C(n,m,p,t,\lambda_{1},\Omega),
\end{equation*}
where $\lambda_{1}$ is the first eigenvalue for $(-\Delta)^{m}$ in $\Omega$ with Navier boundary conditions.
\end{thm}
As a consequence of the a priori estimates (Theorem \ref{Thm-CFL} and Theorem \ref{Thm1}), by applying the Leray-Schauder fixed point theorem, we can derive the following existence result for positive solution to the following Navier problem for higher order Lane-Emden equations
\begin{equation}\label{Navier}\\\begin{cases}
(-\Delta)^{m}u(x)=u^{p}(x) \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \Omega, \\
u(x)=-\Delta u(x)=\cdots=(-\Delta)^{m-1}u(x)=0 \,\,\,\,\,\,\,\, \text{on} \,\,\, \partial\Omega,
\end{cases}\end{equation}
where $n\geq3$, $1\leq m<\frac{n}{2}$ and $\Omega\subset\mathbb{R}^{n}$ is a bounded domain with $C^{2m-2}$ boundary $\partial\Omega$.
\begin{thm}\label{Thm2}
Assume $1<p<\frac{n+2m}{n-2m}$. If one of the following two assumptions
\begin{equation*}
\text{i)} \,\, \Omega \,\, \text{is strictly convex}, \,\, 1<p<\frac{n+2m}{n-2m}, \quad\, \text{or} \quad\, \text{ii)} \,\, p\in\left(1,\frac{n+2}{n-2}\right]\bigcup\left(\frac{n}{n-2m},\frac{n+2m}{n-2m}\right)
\end{equation*}
holds. Then, the higher order Navier problem \eqref{Navier} possesses at least one positive solution $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$. Moreover, the positive solution $u$ satisfies
\begin{equation*}
\|u\|_{L^{\infty}(\overline{\Omega})}\geq\left(\frac{\sqrt{2n}}{diam\,\Omega}\right)^{\frac{2m}{p-1}}.
\end{equation*}
\end{thm}
It's well known that the super poly-harmonic properties of solutions are crucial in establishing Liouville type theorems and the representation formulae for higher order or fractional order PDEs (see e.g. \cite{CDQ,CF,CFL,CL2,WX}). In Section 2, we will first prove the super poly-harmonic properties of solutions for both $a<0$ and $a\geq0$ via a unified approach (see Theorem \ref{lemma0}). As a consequence, we can show the equivalence between the PDE \eqref{PDE} and the corresponding integral equation \eqref{IE}. Then, by applying the method of moving planes in integral forms and Pohozaev identity, we prove the Liouville theorem (Theorem \ref{Thm0}) for \eqref{PDE}. In Sections 3 and 4, we will prove a priori estimates and existence of positive solutions to non-critical higher order Lane-Emden equations in bounded domains $\Omega$, using the arguments from Chen, Dai and Qin \cite{CDQ} for critical order Hardy-H\'{e}non equations and results from Chen, Fang and Li \cite{CFL}. In Section 3, we will derive a priori estimates for any positive solutions to the higher order Naiver problem \eqref{tNavier} (Theorem \ref{Thm1}) by applying the method of moving planes in local way and Kelvin transforms. We will first establish the boundary layer estimates (Theorem \ref{Boundary}), in which the properties of the boundary $\partial\Omega$ play a crucial role. The global a priori estimates follows from the boundary layer estimates, blowing-up analysis and the Liouville theorem (Theorem \ref{Thm0}). Section 4 is devoted to the proof of Theorem \ref{Thm2}. The existence of positive solutions to the higher order Lane-Emden equations \eqref{Navier} with Navier boundary conditions will be established via the a priori estimates (Theorem \ref{Thm-CFL} and Theorem \ref{Thm1}) and the Leray-Schauder fixed point theorem (Theorem \ref{L-S}).
\section{Proof of Theorem \ref{Thm0}}
In this section, we will prove Theorem \ref{Thm0} by using contradiction arguments. Now suppose on the contrary that $u\geq0$ satisfies equation \eqref{PDE} but $u$ is not identically zero, then there exists some $\bar{x}\in\mathbb{R}^{n}$ such that $u(\bar{x})>0$.
In the following, we will use $C$ to denote a general positive constant that may depend on $n$, $m$, $a$, $p$ and $u$, and whose value may differ from line to line.
\subsection{Super poly-harmonic properties}
The super poly-harmonic properties of solutions are closely related to the representation formulae and Liouville type theorems (see \cite{CDQ,CF,CFL,CL2,WX} and the references therein). Therefore, in order to prove Theorem \ref{Thm0}, we need the following theorem about the super poly-harmonicity.
\begin{thm}\label{lemma0}(Super poly-harmonic properties). Assume $n\geq3$, $1\leq m<\frac{n}{2}$, $-\infty<a<2m$, $1<p<+\infty$ and $u$ is a nonnegative solution of \eqref{PDE}. If one of the following two assumptions
\begin{equation*}
-\infty<a\leq2+2p \,\,\,\,\,\,\,\,\,\,\,\, \text{or} \,\,\,\,\,\,\,\,\,\,\,\, u(x)=o(|x|^{2}) \,\,\,\, \text{as} \,\, |x|\rightarrow+\infty
\end{equation*}
holds, then
\begin{equation*}
(-\Delta)^{i}u(x)\geq0
\end{equation*}
for every $i=1,2,\cdots,m-1$ and all $x\in\mathbb{R}^{n}$.
\end{thm}
\begin{proof}
Let $u_{i}:=(- \Delta)^{i}u$. We want to show that $u_{i}\geq0$ for $i=1,2,\cdots, m-1$. Our proof will be divided into two steps.
\textbf{\emph{Step 1.}} We first show that
\begin{equation}\label{2-1}
u_{ m-1}=(-\Delta)^{ m- 1}u\geq0.
\end{equation}
If not, then there exists $0\neq x^{1}\in\mathbb{R}^n$, such that
\begin{equation}\label{2-2}
u_{ m-1}(x^{1})<0.
\end{equation}
Now, let
\begin{equation}\label{2-3}
\bar{f}(r)=\bar{f}\big(|x-x^1|\big):=\frac{1}{|\partial B_{r}(x^{1})|}\int_{\partial B_{r}(x^{1})}f(x)d\sigma
\end{equation}
be the spherical average of $f$ with respect to the center $x^1$. Then, by the well-known property $\overline{\Delta u}=\Delta\bar{u}$ and $-\infty<a<2m<n$, we have, for any $r\geq0$ and $r\neq|x^{1}|$,
\begin{equation}\label{2-4}
\left\{{\begin{array}{l} {-\Delta\overline{u_{ m-1}}(r)=\overline{\frac{u^{p}(x)}{|x|^{a}}}(r)}, \\ {} \\ {-\Delta\overline{u_{ m-2}}(r)=\overline{u_{ m-1}}(r)}, \\ \cdots\cdots \\ {-\Delta\overline u(r)=\overline{u_1}(r)}. \\ \end{array}}\right.
\end{equation}
From the first equation in \eqref{2-4}, by Jensen's inequality, we get, for any $r\geq0$ and $r\neq|x^{1}|$,
\begin{align}\label{2-5}
-\Delta\overline{u_{ m-1}}(r)&=\frac{1}{{| {\partial
B_{r}({x^{1}})}| }}\int_{\partial B_{r}(
{x^{1}})}\frac{{u^{p}(x)}}{|x|
^{a}}d\sigma\nonumber\\
& \geq({r+| {x^{1}}| })^{-a}\frac
{1}{{| {\partial B_{r}({x^{1}})}| }}
\int_{\partial B_{r}({x^{1}})}{u^{p}( x)}d\sigma\\
& \geq({r+| {x^{1}}| })^{-a}\left(
{\frac{1}{{| {\partial B_{r}({x^{1}})}| }
}\int_{\partial B_{r}({x^{1}})}{u(x)
}d\sigma}\right)^{p}\nonumber\\
& =(r+|x^{1}|)^{-a}\bar{u}^{p}(r)\geq0 \quad\quad \text{if} \,\,\, 0\leq a<2m,\nonumber
\end{align}
and
\begin{equation}\label{2-5'}
-\Delta\overline{u_{ m-1}}(r)\geq\big|r-|x^{1}|\big|^{-a}\bar{u}^{p}(r)\geq0 \quad\quad \text{if} \,\, -\infty<a<0.
\end{equation}
From \eqref{2-5} and \eqref{2-5'}, one has
\begin{equation}\label{2-6}
-\frac{1}{r^{n-1}}\Big(r^{n-1}\overline{u_{ m-1}}\,'(r)\Big)'\geq0.
\end{equation}
Since $-\infty<a<2m<n$, we can integrate both sides of \eqref{2-6} from $0$ to $r$ and derive
\begin{equation}\label{2-7}
\overline{u_{ m-1}}\,'(r)\leq0, \,\,\,\,\,\, \overline{u_{ m-1}}(r)\leq\overline{u_{ m-1}}(0)=u_{ m-1}(x^{1})=:-c_{0}<0
\end{equation}
for any $r\geq0$. From the second equation in \eqref{2-4}, we deduce that
\begin{equation}\label{2-8}
-\frac{1}{{r^{n-1}}}\Big({r^{n-1}\overline{u_{ m-2}}\,'}(r)\Big)'=\overline{u_{ m-1}}(r)\leq-c_{0}, \,\,\,\,\,\, \forall \,\, r\geq0,
\end{equation}
integrating from $0$ to $r$ yields
\begin{equation}\label{2-9}
\overline{u_{ m - 2}}\,'(r)\geq\frac{c_{0}}{n}r, \,\,\,\,\,\, \overline{u_{ m-2}}(r)\geq\overline{u_{ m-2}}(0)+\frac{c_{0}}{2n}r^{2}, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
Hence, there exists $r_{1} > 0$ such that
\begin{equation}\label{2-10}
\overline{u_{ m-2}}(r_{1})>0.
\end{equation}
Next, take a point $x^{2}$ with $|x^{2}-x^{1}|=r_{1}$ as the new center, and make average of $\bar{f}$ at the new center $x^{2}$, i.e.,
\begin{equation}\label{2-11}
\overline{\overline{f}}(r)=\overline{\overline{f}}\big(|x-x^{2}|\big):=\frac{1}{|\partial B_{r}(x^{2})|}\int_{\partial B_{r}(x^{2})}\bar f(x)d\sigma.
\end{equation}
One can easily verify that
\begin{equation}\label{2-12}
\overline{\overline{u_{ m-2}}}(0)=\overline{u _{ m - 2}}(x^{2})=:c_{1}>0.
\end{equation}
Then, from \eqref{2-5} and Jensen's inequality, we deduce that $(\overline{\overline{u}},\overline{\overline{u_{1}}},\cdots,\overline{\overline{u_{ m-1}}})$ satisfies
\begin{equation}\label{2-13}
\left\{{\begin{array}{l} {-\Delta\overline{\overline{u_{ m-1}}}(r)=\overline{\overline{\frac{u^{p}(x)}{|x|^{a}}}}(r)\geq0}, \\
{} \\
{-\Delta\overline{\overline{u_{ m-2}}}(r)=\overline{\overline{u_{ m-1}}}(r)}, \\ \cdots\cdots \\ {-\Delta\overline{\overline{u}}(r)=\overline{\overline{u _1}}(r)} \\ \end{array}}\right.
\end{equation}
for any $r\geq0$. Using the same method as obtaining the estimate \eqref{2-9}, we conclude that
\begin{equation}\label{2-14}
\overline{\overline{u_{ m-2}}}(r)\geq\overline{\overline{u_{ m-2}}}(0)+\frac{{c_{0}}}{{2n}}r^{2}, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
Thus we infer from \eqref{2-7}, \eqref{2-12}, \eqref{2-13} and \eqref{2-14} that
\begin{equation}\label{2-15}
\overline{\overline{u_{ m-1}}}(r)\leq\overline{\overline{u_{ m-1}}}(0)<0, \,\,\,\,\,\,\,\,\, \overline{\overline{u_{ m-2}}}(r)\geq\overline{\overline{u_{ m-2}}}(0)>0, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
From the third equation in \eqref{2-13} and integrating, we infer that
\begin{equation}\label{2-16}
\overline{\overline{u_{ m-3}}}\,'(r)\leq-\frac{c_{1}}{n}r \,\,\,\,\,\, \text{and} \,\,\,\,\,\, \overline{\overline{u_{ m-3}}}(r)\leq\overline{\overline{u_{ m-3}}}(0)-\frac{{c_{1}}}{{2n}}r^{2}, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
Hence, there exists $r_{2}>0$ such that
\begin{equation}\label{2-17}
\overline{\overline{u_{ m-3}}}(r_{2})<0.
\end{equation}
Next, we take a point $x^{3}$ with $|x^{3}-x^{2}|=r_{2}$ as the new center and make average of $\bar{\bar{f}}$ at the new center $x^{3}$, i.e.,
\begin{equation}\label{2-18}
\overline{\overline{\overline{f}}}(r)=\overline{\overline{\overline{f}}}\big(|x-x^{3}|\big):=\frac{1}{|\partial B_{r}(x^{3})|}\int_{\partial B_{r}(x^{3})}\overline{\overline{f}}(x)d\sigma.
\end{equation}
It follows that
\begin{equation}\label{2-19}
\overline{\overline{\overline{u_{ m-3}}}}(0)=\overline{\overline{u _{ m-3}}}(x^{3})=:-c_{2}<0.
\end{equation}
One can easily verify that $\overline{\overline{\overline{u}}}$ and $\overline{\overline{\overline{u_{i}}}}$ ($i=1,\cdots, m-1$) satisfy entirely similar equations as $(\overline{\overline{u}},\overline{\overline{u_{1}}},\cdots,\overline{\overline{u_{ m-1}}})$ (see \eqref{2-13}). Using the same method as deriving \eqref{2-15}, we arrive at
\begin{equation}\label{2-20}
\overline{\overline{\overline{u_{ m-1}}}}(r)\leq\overline{\overline{\overline{u_{ m-1}}}}(0)<0, \,\,\,\,\,\,\, \overline{\overline{\overline{u_{ m-2}}}}(r)\geq\overline{\overline{\overline{u_{ m-2}}}}(0)>0, \,\,\,\,\,\,\,
\overline{\overline{\overline{u_{ m-3}}}}(r)\leq\overline{\overline{\overline{u_{ m-3}}}}(0)<0
\end{equation}
for any $r\geq0$. Continuing this way, after $ m$ steps of re-centers (denotes the centers by $x^{1},x^{2},\cdots,x^{ m}$, the $ m$ times averages of $f$ by $\widetilde{f}$ and the resulting functions coming from taking $ m$ times averages by $\widetilde{u}$ and $\widetilde{u_{i}}$ for $i=1,2,\cdots, m-1$), we finally obtain that
\begin{equation}\label{2-21}
-\Delta\widetilde{u_{ m-1}}(r)\geq\widetilde{\frac{u^{p}(x)}{|x|^{a}}}(r)\geq0,
\end{equation}
and for every $i=1,\cdots, m-1$,
\begin{equation}\label{2-22}
(-1)^{i}\widetilde{u_{ m-i}}(r)\geq(-1)^{i}\widetilde{u_{ m-i}}(0)>0, \,\,\, \,\,\, (-1)^{ m}\widetilde{u}(r)\geq(-1)^{ m}\widetilde{u}(0)>0, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
Moreover, in the above process, we may choose $|x^{ m}|$ sufficiently large, such that
\begin{equation}\label{2-100}
|x^{ m}-x^{ m-1}|\geq|x^{ m-1}-x^{ m-2}|+\cdots+|x^{2}-x^{1}|+|x^{1}|+2.
\end{equation}
Now, if $ m$ is odd, estimate \eqref{2-22} implies immediately that
\begin{equation}\label{2-23}
\widetilde{u}(r)\leq\widetilde{u}(0)<0,
\end{equation}
which contradicts the fact that $u\geq0$. Therefore, we only need to deal with the cases that $ m$ is an even integer hereafter.
Since $ m$ is even, we have $\widetilde{u}(r)\geq\widetilde{u}(0)>0$ for any $r\geq0$, furthermore, one can actually observe from the above ``re-centers and iteration" process that
\begin{equation}\label{2-101}
\widetilde{u}(0)\geq\frac{c}{2n}|x^{ m}-x^{ m-1}|^{2}
\end{equation}
for some constant $c>0$. Thus we may choose $|x^{ m}|$ larger, such that both \eqref{2-100} and the following
\begin{equation}\label{2-102}
\widetilde{u}(0)\geq(2p)^{\frac{2mp}{(p-1)^{2}}}\left(1+\frac{2n}{p}\right)^{\frac{2m}{p-1}}
\end{equation}
hold.
For arbitrary $\lambda>0$, define the re-scaling of $u$ by
\begin{equation}\label{2-24}
u_{\lambda}(x):=\lambda^{\frac{2m-a}{p-1}}u(\lambda x).
\end{equation}
Then one can easily verify that equation \eqref{PDE} is invariant under this re-scaling. After $ m$ steps of re-centers for $u_{\lambda}$, we denote the centers for $u_{\lambda}$ by $x_{\lambda}^{1},x_{\lambda}^{2},\cdots,x_{\lambda}^{ m}$ and the resulting function coming from taking $ m$ times averages by $\widetilde{u_{\lambda}}$ and $\widetilde{u_{\lambda,i}}$ for $i=1,2,\cdots, m-1$. Then \eqref{2-21} and \eqref{2-22} still hold for $(\widetilde{u_{\lambda}},\widetilde{u_{\lambda,1}},\cdots,\widetilde{u_{\lambda, m-1}})$ and $x_{\lambda}^{k}=\frac{1}{\lambda}x_{k}$ for $k=1,\cdots, m$, thus one has the following estimate
\begin{equation}\label{2-25}
|x_{\lambda}^{ m}-x_{\lambda}^{ m-1}|+\cdots+|x_{\lambda}^{2}-x_{\lambda}^{1}|+|x_{\lambda}^{1}|\leq
|x^{ m}-x^{ m-1}|+\cdots+|x^{2}-x^{1}|+|x^{1}|=:M
\end{equation}
holds uniformly for every $\lambda\geq1$.
Since we have \eqref{2-22} and $m$ is even, it follows that
\begin{equation}\label{2-26}
\widetilde{u}(r)\geq\widetilde{u}(0)\geq(2p)^{\frac{2mp}{(p-1)^{2}}}\left(1+\frac{2n}{p}\right)^{\frac{2m}{p-1}}>0, \,\,\,\,\,\, \forall \,\, r\geq0,
\end{equation}
and hence
\begin{equation}\label{2-28}
\widetilde{u_{\lambda}}(r)\geq\widetilde{u_{\lambda}}(0)=\lambda^{\frac{2m-a}{p-1}}\widetilde{u}(0)
\geq\lambda^{\frac{2m-a}{p-1}}(2p)^{\frac{2mp}{(p-1)^{2}}}\left(1+\frac{2n}{p}\right)^{\frac{2m}{p-1}}>0, \,\,\,\,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
For $0\leq a<2m$, by the estimate \eqref{2-28}, we may assume that, we already have
\begin{equation}\label{2-28'}
\widetilde{u}(0)\geq(1+M)^{\frac{a}{p-1}}(2p)^{\frac{2mp}{(p-1)^{2}}}\left(1+\frac{2n}{p}\right)^{\frac{2m}{p-1}},
\end{equation}
or else we may replace $u$ by $u_{\lambda}$ with $\lambda=(1+M)^{\frac{a}{2m-a}}$ (still denoted by $u$).
For any $0\leq r\leq1$, we have
\begin{equation}\label{2-27}
\widetilde{u}(r)\geq\widetilde{u}(0)\geq l_{0}\,r^{\alpha_{0}},
\end{equation}
where
\begin{equation}\label{2-29}
l_{0}:=\widetilde{u}(0)\geq\max\left\{(1+M)^{\frac{a}{p-1}},1\right\}(2p)^{\frac{2mp}{(p-1)^{2}}}\alpha_{0}^{\frac{2m}{p-1}}, \,\quad\, \alpha_{0}:=\max\Big\{1,\frac{2n}{p}\Big\}\geq1.
\end{equation}
As a consequence, we infer from \eqref{2-21}, \eqref{2-100}, \eqref{2-25} and \eqref{2-27} that, for any $0\leq r\leq1$,
\begin{align}\label{2-30}
-\Delta\widetilde{u_{ m-1}}(r)&\geq\Big(r+|x^{ m}-x^{ m-1}|+\cdots+|x^{2}-x^{1}|+|x^{1}|\Big)^{-a}\widetilde{u}^{p}(r) \nonumber \\
&\geq\big(1+M\big)^{-a}\,l_{0}^{p}\,r^{\alpha_{0}p}\\
&\geq C_{0}\,l_{0}^{p}\,r^{\alpha_{0}p} \,\,\,\,\,\,\,\,\,\,\,\,\quad \text{if} \,\, 0\leq a<2m, \nonumber
\end{align}
and
\begin{align}\label{2-30'}
-\Delta\widetilde{u_{ m-1}}(r)&\geq\Big(|x^{ m}-x^{ m-1}|-|x^{ m-1}-x^{ m-2}|-\cdots-|x^{2}-x^{1}|-|x^{1}|-r\Big)^{-a}
\widetilde{u}^{p}(r) \nonumber \\
&\geq l_{0}^{p}\,r^{\alpha_{0}p}\\
&\geq C_{0}\,l_{0}^{p}\,r^{\alpha_{0}p}, \,\,\,\,\,\,\,\,\,\,\,\,\quad \text{if} \,\, -\infty<a<0, \nonumber
\end{align}
where
\begin{equation}\label{2-99}
C_{0}:=\min\left\{(1+M)^{-a},1\right\}\in(0,1].
\end{equation}
Integrating both sides of \eqref{2-30} and \eqref{2-30'} from $0$ to $r$ twice and taking into account of \eqref{2-22} yield
\begin{equation}\label{2-31}
\widetilde{u_{ m-1}}(r)<-\frac{C_{0}l_{0}^{p}}{(\alpha_{0}p+n)(\alpha_{0}p+2)}r^{\alpha_{0}p+2}, \,\,\,\,\,\, \forall \,\, 0\leq r\leq1.
\end{equation}
This implies
\begin{equation}\label{2-32}
-\frac{1}{r^{n-1}}\left(r^{n-1}\widetilde{u_{ m-2}}\,'(r)\right)'<-\frac{C_{0}l_{0}^{p}}{(\alpha_{0}p+n)(\alpha_{0}p+2)}r^{\alpha_{0}p+2},
\end{equation}
and consequently,
\begin{equation}\label{2-33}
\widetilde{u_{ m-2}}(r)>\frac{C_{0}l_{0}^{p}}{(\alpha_{0}p+n)(\alpha_{0}p+2)(\alpha_{0}p+n+2)(\alpha_{0}p+4)}r^{\alpha_{0}p+4}, \,\,\,\,\,\, \forall \,\, 0\leq r\leq1.
\end{equation}
Continuing this way, since $m$ is an even integer, by iteration, we can finally arrive at
\begin{equation}\label{2-34}
\widetilde{u}(r)>\frac{C_{0}l_{0}^{p}}{(\alpha_{0}p+2n)^{2m}}r^{\alpha_{0}p+2m}, \,\,\,\,\,\, \forall \,\, 0\leq r\leq1.
\end{equation}
Now, define
\begin{equation}\label{2-35}
\alpha_{k+1}:=2\alpha_{k}p\geq\alpha_{k}p+2n \,\,\,\,\,\, \text{and} \,\,\,\,\,\, l_{k+1}:=\frac{C_{0}l_{k}^{p}}{(2\alpha_{k}p)^{2m}}
\end{equation}
for $k=0,1,\cdots$. Then \eqref{2-34} implies
\begin{equation}\label{2-36}
\widetilde{u}(r)>\frac{C_{0}l_{0}^{p}}{(2\alpha_{0}p)^{2m}}r^{2\alpha_{0}p}=l_{1}r^{\alpha_{1}}, \,\,\,\,\, \forall \,\, r\in[0,1].
\end{equation}
Suppose we have $\widetilde{u}(r)\geq l_{k}r^{\alpha_{k}}$, then go through the entire process as above, we can derive $\widetilde{u}(r)\geq l_{k+1}r^{\alpha_{k+1}}$ for any $0\leq r\leq1$. Therefore, one can prove by induction that
\begin{equation}\label{2-37}
\widetilde{u}(r)\geq l_{k}r^{\alpha_{k}}, \,\,\,\,\,\, \forall \,\, r\in[0,1], \,\,\,\,\,\, \forall \,\, k\in\mathbb{N}.
\end{equation}
Through direct calculations, we have
\begin{eqnarray}\label{2-38}
l_{k}&=&\frac{C_{0}^{\frac{p^{k}-1}{p-1}}l_{0}^{p^{k}}}{(2p)^{2m(k+(k-1)p+(k-2)p^{2}+\cdots+p^{k-1})}\alpha_{0}^{\frac{2m(p^{k}-1)}{p-1}}} \\
\nonumber &=& \frac{C_{0}^{\frac{p^{k}-1}{p-1}}l_{0}^{p^{k}}(2p)^{\frac{2mk}{p-1}}}{(2p)^{\frac{2m(p^{k+1}-p)}{(p-1)^{2}}}\alpha_{0}^{\frac{2m(p^{k}-1)}{p-1}}}
\geq (2p)^{\frac{2mk}{p-1}}\left(\frac{C_{0}^{\frac{1}{p-1}}l_{0}}{(2p)^{\frac{2mp}{(p-1)^{2}}}\alpha_{0}^{\frac{2m}{p-1}}}\right)^{p^{k}}
\end{eqnarray}
for $k=0,1,2,\cdots$. From \eqref{2-29}, \eqref{2-99}, \eqref{2-37} and \eqref{2-38}, we deduce that
\begin{equation}\label{2-40}
\widetilde{u}(1)\geq(2p)^{\frac{2mk}{p-1}}\rightarrow+\infty, \,\,\,\,\,\, \text{as} \,\, k\rightarrow\infty.
\end{equation}
This is absurd. Therefore, \eqref{2-1} must hold, that is, $u_{ m-1}=(-\Delta)^{ m- 1}u\geq0$.
\textbf{\emph{Step 2.}} Next, we will show that all the other $u_{i}$ ($i=1,\cdots, m-2$) must be nonnegative, that is,
\begin{equation}\label{2-41}
u_{ m-i}(x)\geq0, \,\,\,\,\,\,\,\,\,\,\, \forall \,\, i=2,3,\cdots, m-1, \,\,\,\,\,\, \forall \,\, x\in\mathbb{R}^{n}.
\end{equation}
Suppose on the contrary that, there exists some $2\leq i\leq m-1$ and some $x^{0}\in\mathbb{R}^{n}$ such that
\begin{equation}\label{2-42}
u_{ m-1}(x)\geq0, \,\,\,\,\, u_{ m-2}(x)\geq0, \,\,\,\, \cdots, \,\,\,\, u_{ m-i+1}(x)\geq0, \,\,\,\,\,\, \forall \,\, x\in\mathbb{R}^{n},
\end{equation}
\begin{equation}\label{2-43}
u_{ m-i}(x^{0})<0.
\end{equation}
Then, repeating the similar ``re-centers and iteration" arguments as in Step 1, after $ m-i+1$ steps of re-centers (denotes the centers by $\bar{x}^{1},\bar{x}^{2},\cdots,\bar{x}^{ m-i+1}$), the signs of the resulting functions $\widetilde{u_{ m-j}}$ ($j=i,\cdots, m-1$) and $\widetilde{u}$ satisfy
\begin{equation}\label{2-44}
(-1)^{j-i+1}\widetilde{u_{ m-j}}(r)\geq(-1)^{j-i+1}\widetilde{u_{ m-j}}(0)>0, \,\,\,\,\,\,
(-1)^{ m-i+1}\widetilde{u}(r)\geq(-1)^{ m-i+1}\widetilde{u}(0)>0
\end{equation}
for any $r\geq0$. Since $u\geq0$, it follows immediately from \eqref{2-44} that $ m-i+1$ is even and
\begin{equation}\label{2-45}
\widetilde{u}(r)\geq\widetilde{u}(0)>0, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
Furthermore, since $ m-i$ is odd, we infer from \eqref{2-44} that
\begin{equation}\label{2-48}
-\Delta\widetilde{u}(r)=\widetilde{u_{1}}(r)\leq\widetilde{u_{1}}(0)=:-\widetilde{c}<0, \,\,\,\,\,\, \forall \,\, r\geq0,
\end{equation}
and hence, by integrating, one has
\begin{equation}\label{2-49}
\widetilde{u}(r)\geq\widetilde{u}(0)+\frac{\widetilde{c}}{2n}r^{2}>\frac{\widetilde{c}}{2n}r^{2}, \,\,\,\,\,\, \forall \,\, r\geq0.
\end{equation}
Therefore, if we assume that $u(x)=o(|x|^{2})$ as $|x|\rightarrow+\infty$, we will get a contradiction from \eqref{2-49}.
Or, if we assume that $-\infty<a\leq2+2p$, combining \eqref{2-49} with the estimate \eqref{2-21}, we get that, for $r\geq r_{0}$ sufficiently large,
\begin{eqnarray}\label{2-46}
-\Delta\widetilde{u_{ m-1}}(r)&\geq&\left(r+|\bar{x}^{ m-i+1}-\bar{x}^{ m-i}|+\cdots+|\bar{x}^{2}-\bar{x}^{1}|+|\bar{x}^{1}|
\right)^{-a}\widetilde{u}^{p}(r) \\
\nonumber &\geq&\left(\frac{\widetilde{c}}{4n}\right)^{p}r^{2p-a} \quad\quad\quad \text{if} \,\, 0\leq a\leq2+2p,
\end{eqnarray}
and
\begin{eqnarray}\label{2-46'}
-\Delta\widetilde{u_{ m-1}}(r)&\geq&\left(r-|\bar{x}^{ m-i+1}-\bar{x}^{ m-i}|-\cdots-|\bar{x}^{2}-\bar{x}^{1}|-|\bar{x}^{1}|
\right)^{-a}\widetilde{u}^{p}(r) \\
\nonumber &\geq&\left(\frac{\widetilde{c}}{4n}\right)^{p}r^{2p-a} \quad\quad\quad \text{if} \,\, -\infty<a<0.
\end{eqnarray}
Now, by a direct integration on \eqref{2-46} and \eqref{2-46'}, we get, if $-\infty<a<2+2p$, then
\begin{equation}\label{2-47}
\widetilde{u_{ m-1}}(r)\leq\widetilde{u_{ m-1}}(r_{0})-\left(\frac{\widetilde{c}}{4n}\right)^{p}
\frac{r^{2+2p-a}-r_{0}^{2+2p-a}}{(n+2p-a)(2+2p-a)}\rightarrow-\infty, \,\,\,\,\,\, \text{as} \,\,\, r\rightarrow\infty;
\end{equation}
if $a=2+2p$, then
\begin{equation}\label{2-47'}
\widetilde{u_{ m-1}}(r)\leq\widetilde{u_{ m-1}}(r_{0})-\left(\frac{\widetilde{c}}{4n}\right)^{p}\frac{\ln r-\ln r_{0}}{n-2}\rightarrow-\infty, \,\,\,\,\,\, \text{as} \,\,\, r\rightarrow\infty.
\end{equation}
This contradicts $u_{ m-1}\geq0$ and thus \eqref{2-41} must hold. This concludes the proof of Theorem \ref{lemma0}.
\end{proof}
\begin{rem}\label{remark4}
For $0<a<2m$, if we consider the nonnegative solutions $u\in C^{2m}(\mathbb{R}^{n}\setminus\{0\})\cap C(\mathbb{R}^{n})$, then it is clear from our proof of Theorem \ref{lemma0} that we can show super poly-harmonic properties except the origin $0\in\mathbb{R}^{n}$, that is, $(-\Delta)^{i}u\geq0$ in $\mathbb{R}^{n}\setminus\{0\}$ for $i=1,\cdots,m-1$.
\end{rem}
\subsection{Equivalance between PDE and IE}
By applying Theorem \ref{lemma0} for $a\geq0$, we can deduce from $-\Delta u\geq0$, $u\geq0$, $u(\bar{x})>0$ and maximum principle that
\begin{equation}\label{2-50}
u(x)>0, \,\,\,\,\,\,\, \forall \,\, x\in\mathbb{R}^{n}.
\end{equation}
Then, by maximum principle, Lemma 2.1 from Chen and Lin \cite{CLin} and induction, we can also infer further from $(-\Delta)^{i} u\geq0$ ($i=1,\cdots, m-1$), $u>0$ and equation \eqref{PDE} that
\begin{equation}\label{2-51}
(-\Delta)^{i}u(x)>0, \,\,\,\,\,\,\,\, \forall \,\, i=1,\cdots, m-1, \,\,\,\, \forall \,\, x\in\mathbb{R}^{n}.
\end{equation}
Next, we will show that the positive solution $u$ to \eqref{PDE} also satisfies the following integral equation
\begin{equation}\label{IE}
u(x)=\int_{\mathbb{R}^{n}}\frac{C}{|x-\xi|^{n-2m}}\cdot\frac{u^{p}(\xi)}{|\xi|^{a}}d\xi.
\end{equation}
Indeed, we have the following theorem on the equivalence between PDE \eqref{PDE} and IE \eqref{IE}.
\begin{thm}\label{equivalence}
Assume $n\geq3$, $1\leq m<\frac{n}{2}$, $0\leq a<2m$ and $1<p<\infty$. Suppose $u$ is nonnegative classical solution to \eqref{PDE}, then it also solves the integral equation \eqref{IE}, and vice versa.
\end{thm}
\begin{proof}
Let ${\delta}(x-{\xi})$ be the Dirac Delta function and $\phi_{r}(x-{\xi})$ be the solution of the following equation
\begin{equation}\begin{cases}\label{2e1}
(-\Delta)^{m}\phi_{r}(x-{\xi})={\delta}(x-{\xi}),\ \ \ \ {\xi}\in B_{r}(x),\\
\phi_{r}(x-{\xi})=(-\Delta)\phi_{r}(x-{\xi})=...=(-\Delta)^{m-1}\phi_{r}(x-{\xi}), \,\,\,\,\, {\xi}\in {\partial}B_{r}(x).
\end{cases}\end{equation}
One can easily verify that, $(-\Delta)^{i}\phi_{r}(x-{\xi})$ must take the following form
\begin{equation}\label{2e2}
(-\Delta)^{i}\phi_{r}(x-{\xi})=\frac{c_{i}}{|x-{\xi}|^{n+2i-2m}}+\sum_{k=1}^{m-i}c_{i,k}\frac{|x-{\xi}|^{2m-2i-2k}}{{r}^{n-2k}}
\end{equation}
for $i=0,1,\cdots,m-1$, where the coefficients satisfy $c_{i}+\sum_{k=1}^{m-i}c_{i,k}=0$ ($i=0,1,\cdots,m-1$). In particular, when $i=m-1$, by \eqref{2e2}, we have
\begin{equation}\label{2e3}
(-\Delta)^{m-1}\phi_{r}(x-{\xi})=\frac{C_{m-1}}{|x-{\xi}|^{n-2}}-\frac{C_{m-1}}{{r}^{n-2}}, \,\,\,\,\,\,\, {\xi}\in \overline{B_{r}(x)},
\end{equation}
and hence
\begin{equation}\label{2e4}
\frac{\partial\left[(-\Delta)^{m-1}\phi_{r}(x-{\xi})\right]}{\partial v_{\xi}}\leqslant 0, \,\,\,\,\,\,\,\, {\xi}\in {\partial}B_{r}(x),
\end{equation}
where $v_{\xi}$ denotes the unit outer normal vector at $\xi\in\partial B_{r}(x)$. Next we define function $f(\xi)$ by
\begin{equation}\label{2e5}
(-\Delta)^{m-1}\phi_{r}(x-{\xi})=\frac{C_{m-1}}{|x-{\xi}|^{n-2}}-\frac{C_{m-1}}{{r}^{n-2}}=:f({\xi})\geqslant 0.
\end{equation}
It is obvious that $f\in {L}^{1}(B_{r}(x))$, thus $(-\Delta)^{m-2}\phi_{r}(x-{\xi})$ is super-harmonic in the sense of distribution in ${B_{r}(x)}$, and hence we derive
\begin{equation}\label{2e6}
\inf_{\xi\in B_{r}(x)}(-\Delta)^{m-2}\phi_{r}(x-{\xi})\geq\inf_{\xi\in\partial B_{r}(x)}(-\Delta)^{m-2}\phi_{r}(x-{\xi})=0
\end{equation}
and
\begin{equation}\label{2e7}
\frac{\partial{(-\Delta)^{m-2}\phi_{r}(x-{\xi})} }{\partial {v}_{\xi}}\leqslant 0, \,\,\,\,\,\,\,\, {\xi}\in {\partial}B_{r}(x).
\end{equation}
Continuing this way, we conclude that, for $i=0,1,2...m-1$,
\begin{equation}\label{2e8}
\inf_{\xi\in B_{r}(x)}(-\Delta)^{i}\phi_{r}(x-{\xi})\geq\inf_{\xi\in\partial B_{r}(x)}(-\Delta)^{i}\phi_{r}(x-{\xi})=0
\end{equation}
and
\begin{equation}\label{2e9}
\frac{\partial\left[(-\Delta)^{i}\phi_{r}(x-{\xi})\right]}{\partial {v}_{\xi}}\leqslant 0, \,\,\,\,\,\,\,\, {\xi}\in {\partial}B_{r}(x).
\end{equation}
From \eqref{2e3}, we can get $(-\Delta)^{m-1}\phi_{r}(x-{\xi})$ monotone increases about $r$ and tends to ${\frac{C_{m-1}}{|x-{\xi}|^{n-2}}}$ as $r\rightarrow+\infty$. As a consequence, we arrive at, for any ${r}_{2}>{r}_{1}>0$,
\begin{equation}\label{2e10}
(-\Delta)[(-\Delta)^{m-2}\phi_{{r}_{2}}(x-{\xi})-(-\Delta)^{m-2}\phi_{{r}_{1}}(x-{\xi})]\geq0,\, \, \, \, {\xi}\in B_{r_{1}}(x),
\end{equation}
and
\begin{equation}\label{2e11}
0=(-\Delta)^{m-2}\phi_{{r}_{1}}(x-{\xi})\leq(-\Delta)^{m-2}\phi_{{r}_{2}}(x-{\xi}),\, \, \,\, {\xi}\in {\partial}B_{r_{1}}(x).
\end{equation}
By maximum principle, we deduce that
\begin{equation}\label{2e12}
(-\Delta)^{m-2}\phi_{{r}_{2}}(x-{\xi})\geq(-\Delta)^{m-2}\phi_{{r}_{1}}(x-{\xi}), \,\,\,\,\,\, \forall \,\, \xi\in\mathbb{R}^{n}.
\end{equation}
So $(-\Delta)^{m-2}\phi_{r}(x-{\xi})$ also monotone increases about $r$ and tends to ${\frac{C_{m-2}}{|x-{\xi}|^{n-4}}}$ as $r\rightarrow+\infty$. Continuing this way, we can derive
\begin{equation}\label{2e13}
(-\Delta)^{i}\phi_{r}(x-{\xi})\,\,\big\uparrow\,\,\frac{C_{i}}{|x-{\xi}|^{n-2m+2i}}, \,\,\,\,\,\,\, \text{as} \,\, r\rightarrow+\infty.
\end{equation}
By Lemma 1 in \cite{LiC} and equation \eqref{PDE}, we have $(-\Delta)^{m-1}u$ solves the following equation
\begin{equation}\label{bocher}
(-\Delta)^{m}u=\frac{u^{p}(x)}{|x|^{a}}+m\delta(0) \quad\quad \text{in} \,\,\, B_{\rho}(0)
\end{equation}
in the sense of distributions for arbitrary $\rho>0$, where $m\geq0$ and $\delta(0)$ is the Delta distribution concentrated at the origin. Since $u\in C(\mathbb{R}^{n})$, it follows that $m=0$. Therefore, multiplying both sides of \eqref{bocher} by $\phi_{r}(x-{\xi})$ and integrating by parts on $B_{r}(x)$, by Theorem \ref{lemma0} and \eqref{2e9}, one has
\begin{eqnarray}\label{2e14}
&&\int_{B_{r}(x)}{\phi_{r}(x-{\xi})}\frac{u^{p}(\xi)}{|\xi|^{a}}d{\xi} \\
\nonumber &=& u(x)+\sum_{i=0}^{m-1}\int_{{\partial}B_{r}(x)}(-\Delta)^{i}u(\xi)\cdot\frac{\partial\left[(-\Delta)^{m-i-1}\phi_{r}(x-{\xi})\right]}{\partial {v}_{\xi}}d{\sigma}
\\
\nonumber &{\leqslant}& u(x)
\end{eqnarray}
for any $x\neq0$. At the same time, multiplying $(-\Delta)^{i}u$ by $(-\Delta)^{m-i}\phi_{r}(x-{\xi})$ ($i=1,\cdots,m-1$) and integrating by parts on $B_{r}(x)$, by Theorem \ref{lemma0} and \eqref{2e9}, one also has
\begin{eqnarray}\label{2e15}
&& \int_{B_{r}(x)}{(-\Delta)^{m-i}\phi_{r}(x-{\xi})}\cdot(-\Delta)^{i}u({\xi})d{\xi} \\
\nonumber &=& u(x)+\sum_{j=0}^{i-1}\int_{{\partial}B_{r}(x)}(-\Delta)^{j}u(\xi)\cdot\frac{\partial\left[(-\Delta)^{m-j-1}\phi_{r}(x-{\xi})\right]}{\partial {v}_{\xi}}d{\sigma}
\\
\nonumber &{\leqslant}& u(x).
\end{eqnarray}
Thus, by letting $r\rightarrow+\infty$ in \eqref{2e14}, \eqref{2e15} and using Levi's monotone convergence theorem, we obtain
\begin{equation}\label{2e16}
\int_{\mathbb{R}^{n}}\frac{1}{|x-\xi|^{n-2m}}\cdot\frac{u^{p}(\xi)}{|\xi|^{a}}d{\xi}<\infty
\end{equation}
and
\begin{equation}\label{2e17}
\int_{\mathbb{R}^{n}}\frac{(-\Delta)^{i}u(\xi)}{|x-{\xi}|^{n-2i}}<\infty
\end{equation}
for $i=1,\cdots,m-1$. Therefore, there exists a sequence $\{{r}_{k}\}$ such that, as ${r}_{k}\rightarrow\infty$,
\begin{equation}\label{2e18}
\frac{1}{{{r}_{k}}^{n-2m-1}}\int_{{\partial}B_{{r}_{k}}(x)}\frac{u^{p}(\xi)}{|\xi|^{a}}d{\sigma}\rightarrow0,
\end{equation}
and
\begin{equation}\label{2e19}
\frac{1}{{{r}_{k}}^{n-2i-1}}\int_{{\partial}B_{{r}_{k}}(x)}(-\Delta)^{i}{u}(\xi)d{\sigma}\rightarrow0\,\,\,\,\,\,\,\, \text{for} \,\, i=1,2,\cdots,m-1.
\end{equation}
From \eqref{2e18}, it follows that, as $r_{k}\rightarrow+\infty$,
\begin{equation}\label{2e20}
\frac{1}{{{r}_{k}}^{n-2m-1+a}}\int_{{\partial}B_{{r}_{k}}(x)}{{u}^{p}(\xi)}d{\sigma}=\frac{1}{{{r}_{k}}^{n-1-(2m-a)}}\int_{{\partial}B_{{r}_{k}}(x)}{{u}^{p}(\xi)}d{\sigma}\rightarrow0
\end{equation}
Then, by Jensen's inequality, we have
\begin{equation}\label{2e21}
\left({\frac{1}{{{r}_{k}}^{n-1-(2m-a)}}\int_{{\partial}B_{{r}_{k}}(x)}{{u}^{p}(\xi)}}d{\sigma}\right)^{\frac{1}{p}}\frac{1}{{{r}_{k}}^{\frac{2m-a}{p}}}
\geq\frac{1}{{{r}_{k}}^{n-1}}\int_{{\partial}B_{{r}_{k}}(x)}{{u}(\xi)}d{\sigma},
\end{equation}
and hence
\begin{equation}\label{2e22}
\frac{1}{{{r}_{k}}^{n-1}}\int_{{\partial}B_{{r}_{k}}(x)}{{u}(\xi)}d{\sigma}\rightarrow0.
\end{equation}
Combining this with \eqref{2e2} and \eqref{2e19} implies
\begin{equation}\label{2e23}
\sum_{i=0}^{m-1}\int_{{\partial}B_{{r}_{k}}(x)}(-\Delta)^{i}u(\xi)\cdot\frac{\partial\left[(-\Delta)^{m-i-1}\phi_{r_{k}}(x-{\xi})\right]}{\partial {v}_{\xi}}d{\sigma}\rightarrow0,
\end{equation}
inserting \eqref{2e23} into \eqref{2e14} and letting $r_{k}\rightarrow+\infty$, we derive immediately
\begin{equation}\label{2e24}
u(x)=\int_{\mathbb{R}^{n}}\frac{C}{|x-{\xi}|^{n-2m}}\cdot\frac{u^{p}(\xi)}{|\xi|^{a}}d\xi,
\end{equation}
that is, $u$ satisfies the integral equation \eqref{IE}.
Conversely, assume that $u$ is a nonnegative classical solution of integral equation \eqref{IE}, then
\begin{eqnarray}\label{2e25}
(-\Delta)^{m}u(x)
\nonumber &=& \int_{\mathbb{R}^{n}}{\left[(-\Delta)^{m}\left(\frac{C}{|x-\xi|^{n-2m}}\right)\right]}\frac{u^{p}(\xi)}{|\xi|^{a}}d{\xi}
\\
\nonumber &=& \int_{\mathbb{R}^{n}}\delta(x-\xi)\frac{u^{p}(\xi)}{|\xi|^{a}}d{\xi}=\frac{u^{p}(x)}{|x|^{a}},
\end{eqnarray}
that is, $u$ also solves the PDE \eqref{PDE}. This completes the proof of equivalence between PDE \eqref{PDE} and IE \eqref{IE}.
\end{proof}
For $2\leq a<2m$ and $1<p<\infty$, one can easily observe that the regularity at $0$ of $u$ indicated by the integral equation \eqref{IE} contradicts with $u\in C^{2m-2}(\mathbb{R}^{n})$, thus we must have $u\equiv0$ in $\mathbb{R}^{n}$.
In the following, we will also obtain a contradiction for $1<p<\frac{n+2m-2a}{n-2m}$ and $0\leq a<2m$ by applying the method of moving planes and Pohozaev identity to the equivalent integral equation \eqref{IE} (see subsection 2.3 and 2.4). The proof still works for $u\in C^{2m}(\mathbb{R}^{n}\setminus\{0\})\cap C(\mathbb{R}^{n})$.
\subsection{Radial symmetry of positive solution}
From Theorem \ref{equivalence}, we know that the positive classical solution $u$ to PDE \eqref{PDE} is also a positive solution to the equivalent integral equation \eqref{IE}.
If $u$ is a nonnegative solution to IE \eqref{IE}, we must have either $u\equiv0$ or $u>0$ in $\mathbb{R}^{n}$. The next Theorem says that all the locally integrable positive solutions to IE \eqref{IE} must be radially symmetric and monotone decreasing about the origin.
\begin{thm}\label{symmetry}
Assume $n\geq3$, $1\leq m<\frac{n}{2}$, $0\leq a<2m$ and $1<p<\frac{n+2m-a}{n-2m}$. Suppose $u$ is a positive solution to IE \eqref{IE} satisfying $\frac{u^{p-1}}{|x|^{a}}\in L^{\frac{n}{2m}}_{loc}(\mathbb{R}^{n})$, then $u$ is radially symmetric and monotone decreasing about the origin.
\end{thm}
\begin{proof}
We define the Kelvin transform of $u$ by
\begin{equation}\label{s1}
\bar{u}(x)=\frac{1}{|x|^{n-2m}}u\left(\frac{x}{|x|^{2}}\right), \,\,\,\,\,\,\,\, x\neq0.
\end{equation}
Since $u$ satisfies the integral equation
\begin{equation}\label{s2}
u(x)=C\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})}{|x-y|^{n-2m}|y|^{a}}d{y},
\end{equation}
it follows that, for $x\neq0$,
\begin{eqnarray}\label{s3}
\bar{u}(x)&=&\frac{C}{|x|^{n-2m}}\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})}{\left|\frac{x}{|x|^{2}}-y\right|^{n-2m}|y|^{a}}d{y}
\\
\nonumber &=& \frac{C}{|x|^{n-2m}}\int_{\mathbb{R}^{n}}\frac{u^{p}\left(\frac{y}{|y|^{2}}\right)}{\left|\frac{x}{|x|^{2}}-\frac{y}{|y|^{2}}\right|^{n-2m}\frac{1}{|y|^{a}}}
\cdot\frac{1}{|y|^{2n}}d{y}
\\
\nonumber &=& \frac{C}{|x|^{n-2m}}\int_{\mathbb{R}^{n}}\frac{|x|^{n-2m}|y|^{n-2m}}{|x-y|^{n-2m}}\cdot\frac{u^{p}\left(\frac{y}{|y|^{2}}\right)}{|y|^{2n-a}}dy
\\
\nonumber &=& C\int_{\mathbb{R}^{n}}\frac{1}{|x-y|^{n-2m}}\cdot\frac{u^{p}\left(\frac{y}{|y|^{2}}\right)}{|y|^{n+2m-a}}dy\\
\nonumber &=& C\int_{\mathbb{R}^{n}}\frac{1}{|x-y|^{n-2m}}\cdot\frac{\overline{u}^{p}(y)}{|y|^{\tau}}dy,
\end{eqnarray}
where $\tau:=n+2m-a-p(n-2m)>0$.
We will apply the method of moving planes in integral forms to the integral equation \eqref{s3} and carry out the process of moving plane in the $x_{1}$ direction. For this purpose, we need some definitions.
Let $\lambda\leq0$ be an arbitrary non-positive real number and let the moving plane be
\begin{equation}\label{D1}
T_{\lambda}:=\{x\in\mathbb{R}^{n}:\,x_{1}=\lambda\}.
\end{equation}
We denote
\begin{equation}\label{D2}
\Sigma_{\lambda}:=\{x=(x_{1},x_{2},\cdots,x_{n})\in\mathbb{R}^{n}:\,x_{1}<\lambda\},
\end{equation}
and let
\begin{equation}\label{D3}
x^{\lambda}:=(2\lambda-x_{1},x_{2},\cdots,x_{n})
\end{equation}
be the reflection of $x$ about the plane $T_{\lambda}$, and define
\begin{equation}\label{D4}
\bar{u}_{\lambda}(x):=\bar{u}(x^{\lambda}), \,\,\,\,\,\,\,\, \omega_{\lambda}(x):=\bar{u}_{\lambda}(x)-\bar{u}(x).
\end{equation}
By properly exploiting some global properties of the integral equations, we will show that, for $\lambda$ sufficiently negative,
\begin{equation}\label{s4}
\omega_{\lambda}(x)\geq0, \,\,\,\,\,\,\,\,\,\, \forall \,\, x\in\Sigma_{\lambda}\setminus\{0^{\lambda}\}.
\end{equation}
Then, we start moving the plane $T_{\lambda}$ from near $x_{1}=-\infty$ to the right as long as \eqref{s4} holds, until its limiting position and finally derive symmetry and monotonicity. Therefore, the moving plane process can be divided into two steps.
\emph{Step 1. Start moving the plane from near $x_{1}=-\infty$.} Define the set
\begin{equation}\label{s5}
\Sigma^{-}_{\lambda}:=\{x\in\Sigma_{\lambda}\setminus\{0^{\lambda}\} \, | \, \omega_{\lambda}(x)<0\}.
\end{equation}
We can deduce from (2.3) that, for $x\in{\Sigma_{\lambda}}\setminus\{0^{\lambda}\}$,
\begin{eqnarray}\label{s6}
&&\omega_{\lambda}(x)=\bar{u}_{\lambda}(x)-\bar{u}(x)\\
\nonumber &=& C\int_{\mathbb{R}^{n}}\frac{1}{|x^{\lambda}-y|^{n-2m}}\frac{\bar{u}^{p}(y)}{|y|^{\tau}}dy-C\int_{\mathbb{R}^{n}}\frac{1}{|x-y|^{n-2m}}\frac{\bar{u}^{p}(y)}{|y|^{\tau}}dy
\\
\nonumber &=&C\int_{\Sigma_{\lambda}}\frac{1}{|x^{\lambda}-y|^{n-2m}}\frac{\bar{u}^{p}(y)}{|y|^{\tau}}dy+C\int_{\Sigma_{\lambda}}\frac{1}{|x^{\lambda}-y^{\lambda}|^{n-2m}}\frac{\bar{u}^{p}(y^{\lambda})}{|y^{\lambda}|^{\tau}}dy
\\
\nonumber & & -C\int_{\Sigma_{\lambda}}\frac{1}{|x-y|^{n-2m}}\frac{\bar{u}^{p}(y)}{|y|^{\tau}}dy-C\int_{\Sigma_{\lambda}}\frac{1}{|x-y^{\lambda}|^{n-2m}}\frac{\bar{u}^{p}(y^{\lambda})}{|y^{\lambda}|^{\tau}}dy
\\
\nonumber &=& C\int_{\Sigma_{\lambda}}\left(\frac{1}{|x-y|^{n-2m}}-\frac{1}{|x-y^{\lambda}|^{n-2m}}\right)\left(\frac{\bar{u}^{p}(y^{\lambda})}{|y^{\lambda}|^{\tau}}-\frac{\bar{u}^{p}(y)}{|y|^{\tau}}\right)dy\\
\nonumber &\geq& C\int_{\Sigma_{\lambda}}\left(\frac{1}{|x-y|^{n-2m}}-\frac{1}{|x-y^{\lambda}|^{n-2m}}\right)\frac{\bar{u}^{p}(y^{\lambda})-\bar{u}^{p}(y)}{|y|^{\tau}}dy\\
\nonumber &\geq& C\int_{\Sigma^{-}_{\lambda}}\frac{p\bar{u}^{p-1}(y)}{|x-y|^{n-2m}}\cdot\frac{\omega_{\lambda}(y)}{|y|^{\tau}}dy
\end{eqnarray}
In particular, for $x\in{\Sigma^{-}_{\lambda}}$, we have
\begin{equation}\label{s7}
0>\omega_{\lambda}(x)\geq{C}\int_{\Sigma^{-}_{\lambda}}\frac{p\bar{u}^{p-1}(y)}{|x-y|^{n-2m}}\cdot\frac{\omega_{\lambda}(y)}{|y|^{\tau}}dy.
\end{equation}
By Hardy-Littlewood-Sobolev inequality, one gets, for arbitrary $\frac{n}{n-2m}<q<\infty$,
\begin{eqnarray}\label{s8}
\|\omega_{\lambda}\|_{{L^{q}(\Sigma^{-}_{\lambda})}}&\leq& C\left\|\int_{\Sigma^{-}_{\lambda}}\frac{p\bar{u}^{p-1}(y)}{|x-y|^{n-2m}}\cdot\frac{\omega_{\lambda}(y)}{|y|^{\tau}}dy\right\|_{{L^{q}(\Sigma^{-}_{\lambda})}}
\\
\nonumber &\leq&C\left\|\frac{\bar{u}^{p-1}(x)}{|x|^{\tau}}\cdot\omega_{\lambda}(x)\right\|_{{L^{\frac{nq}{n+2mq}}(\Sigma^{-}_{\lambda})}}
\\
\nonumber &\leq& C\left\|\frac{\bar{u}^{p-1}}{|x|^{\tau}}\right\|_{L^{\frac{n}{2m}}(\Sigma^{-}_{\lambda})}\cdot\|\omega_{\lambda}\|_{{L^{q}(\Sigma^{-}_{\lambda})}}
\end{eqnarray}
Since $\frac{u^{p-1}}{|x|^{a}}\in L^{\frac{n}{2m}}_{loc}(\mathbb{R}^{n})$, we have, for any $r>0$,
\begin{eqnarray}\label{s9}
\int_{|x|\geq{r}}\frac{\bar{u}^{(p-1)\frac{n}{2m}}(x)}{|x|^{\tau\frac{n}{2m}}}dx&=&
\int_{|x|\geq{r}}\frac{1}{|x|^{(\tau+(p-1)(n-2m))\frac{n}{2m}}}u^{(p-1)\frac{n}{2m}}\Big(\frac{x}{|x|^{2}}\Big)dx
\\
\nonumber &=&\int_{|x|\leq{\frac{1}{r}}}\frac{1}{|x|^{2n-(4m-a)\frac{n}{2m}}}u^{(p-1)\frac{n}{2m}}(x)dx
\\
\nonumber &=&\int_{|x|\leq{\frac{1}{r}}}\frac{u^{(p-1)\frac{n}{2m}}(x)}{|x|^{\frac{an}{2m}}}dx<+\infty.
\end{eqnarray}
Therefore, there exists a $\Lambda_{0}$ sufficiently large, such that, for any $\lambda\leq-\Lambda_{0}$,
\begin{equation}\label{s10}
C\left\|\frac{\bar{u}^{p-1}(x)}{|x|^{\tau}}\right\|_{L^{\frac{n}{2m}}(\Sigma^{-}_{\lambda})}\leq\frac{1}{2}.
\end{equation}
Thus, we must have, for any $\frac{n}{n-2m}<q<\infty$,
\begin{equation}\label{s11}
\|\omega_{\lambda}\|_{{L^{q}(\Sigma^{-}_{\lambda})}}=0,
\end{equation}
Combining this with \eqref{s6} implies $\Sigma^{-}_{\lambda}=\emptyset$, and hence
\begin{equation}\label{s12}
\omega_{\lambda}(x)\geq0, \,\,\,\,\,\, \,\,\,\, \forall \, x\in\Sigma_{\lambda}\setminus\{0^{\lambda}\}.
\end{equation}
\emph{Step 2. Move the plane to the limiting position to derive symmetry and monotonicity.} Now we move the plane $T_{\lambda}$ to the right as long as \eqref{s4} holds. Define
\begin{equation}\label{s13}
\lambda_{0}:=\sup\{\lambda\in{\mathbb{R}}\, | \, \omega_{\rho}\geq0 \,\,\, \text{in} \,\,\, \Sigma_{\rho}\setminus\{0^{\rho}\}, \,\, \forall \, \rho\leq\lambda\}.
\end{equation}
By applying a entirely similar argument as in Step1, we can also start moving the plane from near $x_{1}=+\infty$ to the left, thus we must have $\lambda_{0}<+\infty$. Now, we will show that $\lambda_{0}=0$.
Suppose on the contrary that $\lambda_{0}<0$, we will show that
\begin{equation}\label{s14}
\omega_{\lambda_{0}}(x)\equiv0,\,\,\,\,\,\,\, \forall \,\, {x}\in\Sigma_{\lambda_{0}}\setminus\{0^{\lambda_{0}}\}.
\end{equation}
We prove \eqref{s14} by contradiction arguments. Suppose on the contrary that $\omega_{\lambda_{0}}(x)\geq0$, but $\omega_{\lambda_{0}}(x)$ is not identically zero in $\Sigma_{\lambda_{0}}\setminus\{0^{\lambda_{0}}\}$. We will obtain a contradiction with \eqref{s13} via showing that the plane $T_{\lambda}$ can be moved a little bit further to the right, more precisely, there exist an $0<\varepsilon<|\lambda_{0}|$ small enough, such that $w_{\lambda}\geq0$ in $\Sigma_{\lambda}\setminus\{0^{\lambda}\}$ for all $\lambda\in[\lambda_{0},\lambda_{0}+\varepsilon)$.
It can be clearly seen from \eqref{s8} and \eqref{s10} in \emph{Step 1} that, our goal is to prove that, one can choose $\varepsilon>0$ sufficiently small such that, for all $\lambda\in[\lambda_{0},\lambda_{0}+\varepsilon)$,
\begin{equation}\label{s15}
\left\|\frac{\bar{u}^{p-1}(x)}{|x|^{\tau}}\right\|_{L^{\frac{n}{2m}}(\Sigma^{-}_{\lambda})}\leq\frac{1}{2C},
\end{equation}
where the constant $C$ is the same as in \eqref{s8} and \eqref{s10}.
In fact, by \eqref{s9}, we can choose $R>0$ large enough, such that
\begin{equation}\label{s16}
\left(\int_{|x|\geq{R}}\frac{\bar{u}^{(p-1)\frac{n}{2m}}(x)}{|x|^{\tau\frac{n}{2m}}}dx\right)^{\frac{2m}{n}}<\frac{1}{4C}.
\end{equation}
Now fix this $R$, in order to derive \eqref{s15}, we only need to show
\begin{equation}\label{s17}
\lim_{\lambda\rightarrow\lambda_{0}+}\mu\left({\Sigma^{-}_{\lambda}}\cap{B_{R}(0)}\right)=0.
\end{equation}
To this end, we define $E_{\delta}:=\{x\in(\Sigma_{\lambda_{0}}\setminus\{0^{\lambda_{0}}\})\cap{B_{R}(0)}\,|\,w_{\lambda_{0}}(x)>\delta\}$
and $F_{\delta}:=({\Sigma_{\lambda_{0}}}\cap{B_{R}(0)})\setminus{E_{\delta}}$ for any $\delta>0$, and let $D_{\lambda}:=({\Sigma_{\lambda}}\setminus{\Sigma_{\lambda_{0}}})\cap{B_{R}(0)}$ for any $\lambda>\lambda_{0}$. Then, one can easily verify that
\begin{equation}\label{s18}
\lim_{\delta\rightarrow{0}^{+}}\mu(F_{\delta})=0, \,\,\,\,\,\,\, \lim_{\lambda\rightarrow\lambda_{0}+}\mu(D_{\lambda})=0,
\end{equation}
\begin{equation}\label{s19}
{\Sigma^{-}_{\lambda}}\cap{B_{R}(0)}={\Sigma^{-}_{\lambda}}\cap(E_{\delta}\cup{F_{\delta}}\cup{D_{\lambda}})
\subset({\Sigma^{-}_{\lambda}}\cap E_{\delta})\cup{F_{\delta}}\cup{D_{\lambda}}.
\end{equation}
For an arbitrary fixed $\eta>0$, one can choose a $\delta>0$ small enough, such that $\mu(F_{\delta})\leq\eta$. For this fixed $\delta$, we are to prove
\begin{equation}\label{s20}
\lim_{\lambda\rightarrow\lambda_{0}+}\mu(\Sigma_{\lambda}^{-}\cap{E_{\delta}})=0.
\end{equation}
Indeed, one can observe that $\bar{u}(x^{\lambda_{0}})-\bar{u}(x^{\lambda})=\omega_{\lambda_{0}}(x)-\omega_{\lambda}(x)>\delta$
for all $x\in{\Sigma_{\lambda}^{-}\cap{E_{\delta}}}$. It follows that
\begin{equation}\label{s21}
(\Sigma_{\lambda}^{-}\cap{E_{\delta}})\subset{G_{\delta}^{\lambda}}
:=\{x\in{B_{R}(0)}\cap\big((\Sigma_{\lambda_{0}}\setminus\{0^{\lambda_{0}}\})\setminus\{0^{\lambda}\}\big)\,|\,\bar{u}(x^{\lambda_{0}})-\bar{u}(x^{\lambda})>\delta\}.
\end{equation}
By Chebyshev's inequality, we get
\begin{eqnarray}\label{s22}
\mu(G_{\delta}^{\lambda})&\leq&\frac{1}{\delta^{r}}\int_{G_{\delta}^{\lambda}}\left|\bar{u}(x^{\lambda_{0}})-\bar{u}(x^{\lambda})\right|^{r}dx \\
\nonumber &\leq&\frac{1}{\delta^{r}}\int_{B_{R}(0)}\left|\bar{u}(x)-\bar{u}\left(x+2(\lambda-\lambda_{0})e_{1}\right)\right|^{r}dx
\end{eqnarray}
for any $1\leq{r}<\frac{n}{n-2m}$, where $e_{1}=(1,0......,0)\in{\mathbb{R}^{n}}$, and hence
\begin{equation}\label{s23}
\lim_{\lambda\rightarrow\lambda_{0}+}\mu(G^{\lambda}_{\delta})=0,
\end{equation}
from which \eqref{s20} follows immediately.
Therefore, by \eqref{s18}, \eqref{s19} and \eqref{s20}, we have
\begin{equation}\label{s24}
\lim_{\lambda\rightarrow\lambda_{0}+}\mu(\Sigma_{\lambda}^{-}\cap{B_{R}(0)})\leq\mu(F_{\delta})\leq\eta.
\end{equation}
Since $\eta>0$ is arbitrarily chosen, \eqref{s17} follows immediately from \eqref{s24}. Combining \eqref{s16} and \eqref{s17}, we finally arrive at \eqref{s15}.
From the last inequality of \eqref{s8}, we have, for any $\frac{n}{n-2m}<q<\infty$,
\begin{equation}\label{s25}
\|\omega_{\lambda}\|_{L^{q}(\Sigma^{-}_{\lambda})}\leq C\left\|\frac{\bar{u}^{p-1}}{|x|^{\tau}}\right\|_{L^{\frac{n}{2m}}(\Sigma^{-}_{\lambda})}
\cdot\|\omega_{\lambda}\|_{L^{q}(\Sigma^{-}_{\lambda})}.
\end{equation}
By \eqref{s15} and the above estimate, we deduce that, there exists an $\varepsilon>0$ sufficiently small, such that, for all $\lambda\in[\lambda_{0},\lambda_{0}+\varepsilon)$, $\|\omega_{\lambda}\|_{{L^{q}(\Sigma^{-}_{\lambda})}}=0$, thus $\mu(\Sigma^{-}_{\lambda})=0$. Furthermore, by \eqref{s6}, we have $\Sigma^{-}_{\lambda}=\emptyset$, and hence $\omega_{\lambda}(x)\geq0$ in $\Sigma_{\lambda}\setminus\{0^{\lambda}\}$ for all $\lambda\in[\lambda_{0},\lambda_{0}+\varepsilon)$. This contradicts with the definition of ${\lambda_{0}}$. Therefore, \eqref{s14} must hold. By \eqref{s6} and \eqref{s14}, we get, for any $x\in{\Sigma_{\lambda_{0}}}$,
\begin{eqnarray}\label{s26}
\nonumber 0=\omega_{\lambda_{0}}(x)&=&\bar{u}(x^{\lambda_{0}})-\bar{u}(x)
\\
&=&C\int_{\Sigma_{\lambda_{0}}}\left(\frac{1}{|x-y|^{n-2m}}-\frac{1}{|x-y^{\lambda_{0}}|^{n-2m}}\right)
\left(\frac{1}{|y^{\lambda_{0}}|^{\tau}}-\frac{1}{|y|^{\tau}}\right)\bar{u}^{p}(y)dy
\\
\nonumber &>&0.
\end{eqnarray}
That is a contradiction! Thus we must have $\lambda_{0}=0$, and hence
\begin{equation}\label{s27}
u(-x_{1},x_{2},....,x_{n})\geq{u(x_{1},x_{2},....,x_{n})},\,\,\,\,\, \forall \,\, {x}\in{\Sigma_{0}}.
\end{equation}
We can also move the plane from $x_{1}=+\infty$ to the left and the limiting position is also $\lambda_{0}=0$, so one has
\begin{equation}\label{s28}
u(-x_{1},x_{2},....,x_{n})\leq{u(x_{1},x_{2},....,x_{n})},\,\,\, \,\,\,\,\, \forall \,\, {x}\in{\Sigma_{0}}.
\end{equation}
Therefore,
\begin{equation}\label{s29}
u(-x_{1},x_{2},....,x_{n})\equiv{u(x_{1},x_{2},....,x_{n})},\,\,\, \,\,\,\,\, \forall \,\, {x}\in{R^{n}},
\end{equation}
that is, $u(x)$ is symmetric with respect to the plane $T_{0}=\{x\in\mathbb{R}^{n}\,|\,x_{1}=0\}$.
Since the equation is invariant under rotation, the $x_{1}$ direction can be chosen arbitrarily. We conclude that the positive solution $u(x)$ must be radially symmetric and monotone decreasing about the origin $0\in{\mathbb{R}^{n}}$. This finishes our proof of Theorem \ref{symmetry}.
\end{proof}
\subsection{Pohozaev identity and nonexistence of positive radially symmetric solutions}
By Theorem \ref{symmetry}, we deduce that the positive classical solution $u$ to PDE \eqref{PDE} is a positive radially symmetric solution to IE \eqref{IE}, i.e., $u(x)=u(|x|)>0$. Next, we will show that there is no positive radially symmetric classical solutions to \eqref{IE}, which leads to a contradiction.
\begin{thm}\label{nonexistence}
Assume $n\geq3$, $1\leq m<\frac{n}{2}$, $0\leq a<2m$ and $1<p<\frac{n+2m-2a}{n-2m}$, then \eqref{IE} has no positive radially symmetric classical solutions.
\end{thm}
\begin{proof}
Suppose $u(x)=u(|x|)>0$ is a positive radially symmetric classical solution to \eqref{IE}, that is,
\begin{equation}\label{P1}
u(x)=C\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})}{|x-y|^{n-2m}|y|^{a}}d{y}.
\end{equation}
Then, for any $\mu>0$,
\begin{equation}\label{P2}
u({\mu}x)=C\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})}{|{\mu}x-y|^{n-2m}|y|^{a}}d{y}.
\end{equation}
Take the derivatives on both sides of \eqref{P2} with respect to ${\mu}$ and let ${\mu}=1$, we get
\begin{eqnarray}\label{P3}
x\cdot\nabla{u(x)}&=& C\frac{\mathrm{d} }{\mathrm{d}\mu}\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})}{|{\mu}x-y|^{n-2m}|y|^{a}}d{y}\bigg|_{{\mu}=1}
\\
\nonumber &=& -(n-2m)C\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})({\mu}x-y)\cdot x}{|{\mu}x-y|^{n-2m-2}|y|^{a}}d{y}\bigg|_{{\mu}=1}
\\
\nonumber &=& -(n-2m)C\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})(x-y)\cdot x}{|x-y|^{n-2m-2}|y|^{a}}d{y}.
\end{eqnarray}
Multiply both sides of \eqref{P3} by $\frac{{u^{p}({x}})}{|x|^{a}}$ and integrate on $B_{r}(0)$ for any $r>0$, one has
\begin{eqnarray}\label{P4}
LHS &=& \int_{B_{r}(0)}(x\cdot\nabla{u(x)}){\frac{{u^{p}({x}})}{|x|^{a}}}dx
\\
\nonumber &=& \int_{0}^{r}\int_{{\partial}B_{s}(0)}s\frac{\mathrm{d}(u(s)) }{\mathrm{d} s}\cdot{\frac{{u^{p}({s}})}{s^{a}}}d{\sigma}ds
\\
\nonumber &=& \int_{0}^{r}\frac{w_{n}{s^{n-a}}}{p+1}d\big(u^{p+1}(s)\big)
\\
\nonumber &=& \frac{w_{n}}{p+1}r^{n-a}u^{p+1}(r)-\frac{(n-a)w_{n}}{p+1}\int_{0}^{r}u^{p+1}(s)s^{n-a-1}ds
\\
\nonumber &=& \frac{r^{1-a}}{p+1}\int_{{\partial}B_{r}(0)}u^{p+1}(x)d\sigma-\frac{n-a}{p+1}\int_{B_{r}(0)}\frac{u^{p+1}(x)}{|x|^{a}}dx,
\end{eqnarray}
and
\begin{equation}\label{P5}
RHS=-(n-2m)C\int_{B_{r}(0)}\frac{u^{p}(x)}{|x|^{a}}\int_{\mathbb{R}^{n}}\frac{(x-y)\cdot x{u}^{p}(y)}{|x-y|^{n-2m-2}|y|^{a}}dy,
\end{equation}
where $w_{n}$ denotes the area of the unit sphere. Since $u$ is a positive radially symmetric classical solution to \eqref{IE}, we have $u(r)$ monotone decreases about $r\geq0$ and
\begin{eqnarray}\label{P6}
u(x)=u(r)&=& C\int_{\mathbb{R}^{n}}\frac{u^{p}(y)}{|x-y|^{n-2m}|y|^{a}}dy
\\
\nonumber &=& \frac{C}{r^{n-2m}}\int_{\mathbb{R}^{n}}\frac{u^{p}(y)}{\left|\frac{x}{|x|}-\frac{y}{r}\right|^{n-2m}|y|^{a}}dy
\\
\nonumber &\geq& \frac{C}{r^{n-2m}}\int_{0}^{r}\int_{{\partial}B_{s}(0)}\frac{u^{p}(s)}{\left|\frac{x}{|x|}-\frac{s\sigma}{r}\right|^{n-2m}s^{a}}d(s\sigma)ds,
\end{eqnarray}
where $r=|x|$ and $\sigma$ is an arbitrary unit vector on $\partial B_{1}(0)$. Observe that
\begin{equation}\label{P7}
\frac{1}{\left|\frac{x}{|x|}-\frac{s\sigma}{r}\right|^{n-2m}}\geq\frac{1}{2^{n-2m}},\,\,\,\,\,\,\,\,\,\,\, \forall \,\,s\in[0,r] \,\,\,\text{and} \,\,\, \forall \,\, \sigma\in\partial B_{1}(0),
\end{equation}
thus we infer from \eqref{P6} that
\begin{eqnarray}\label{P8}
\nonumber
u(x)=u(r)
\nonumber &\geq& \frac{C}{r^{n-2m}}\int_{0}^{r}w_{n}s^{n-1-a}\frac{u^{p}(s)}{2^{n-2m}}ds
\\
\nonumber &\geq& \frac{Cw_{n}}{{(2r)}^{n-2m}}u^{p}(r)\int_{0}^{r}s^{n-1-a}ds
\\
\nonumber &=& \frac{Cw_{n}}{{2}^{n-2m}(n-a)}\cdot\frac{u^{p}(r)}{r^{n-2m}}r^{n-a}
\\
\nonumber &=& \frac{Cw_{n}r^{2m-a}u^{p}(r)}{{2}^{n-2m}(n-a)}.
\end{eqnarray}
Therefore,
\begin{equation}\label{P9}
u^{p-1}(r)\leq\frac{{2}^{n-2m}(n-a)}{Cw_{n}r^{2m-a}}.
\end{equation}
Let $\tilde{C}=\Big({\frac{{2}^{n-2m}(n-a)}{Cw_{n}}}\Big)^{\frac{1}{p-1}}>0$, then one has the following decay estimate
\begin{equation}\label{P10}
u(r)\leq\frac{\tilde{C}}{r^{\frac{2m-a}{p-1}}}, \quad\quad\, \forall \,\, r>0.
\end{equation}
Note that $1<p<\frac{n+2m-2a}{n-2m}$, we derive from the decay estimate \eqref{P10} that
\begin{equation}\label{P11}
\int_{\mathbb{R}^{n}}\frac{u^{p+1}(x)}{|x|^{a}}dx<\infty,
\end{equation}
and hence
\begin{equation}\label{P12}
\int_{0}^{\infty}r^{-a}\left(\int_{{\partial}B_{r}(0)}u^{p+1}(x)d{\sigma}\right)dr<\infty.
\end{equation}
Thus there exists a sequence $\{r_{j}\}$, such that ${r_{j}}\rightarrow+\infty$ as $j\rightarrow\infty$ and
\begin{equation}\label{P13}
r_{j}^{1-a}\left(\int_{{\partial}B_{r_{j}}(0)}u^{p+1}(x)d{\sigma}\right)\rightarrow0,\,\,\,\,\,\quad \text{as} \,\, j\rightarrow\infty.
\end{equation}
By letting $r=r_{j}\rightarrow+\infty$ in \eqref{P4} and \eqref{P5}, we conclude from \eqref{P11} and \eqref{P13} that
\begin{equation}\label{P14}
-\frac{n-a}{p+1}\int_{\mathbb{R}^{n}}\frac{u^{p+1}(x)}{|x|^{a}}dx
=-(n-2m)C\int_{\mathbb{R}^{n}}\frac{u^{p}(x)}{|x|^{a}}\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})x\cdot(x-y)}{|x-y|^{n-2m+2}|y|^{a}}d{y}dx.
\end{equation}
At the same time, by direct calculations, we have
\begin{eqnarray}\label{P15}
&& \frac{2m-n}{2}\int_{\mathbb{R}^{n}}\frac{u^{p+1}(x)}{|x|^{a}}dx \\
\nonumber &=&\frac{2m-n}{2}\int_{\mathbb{R}^{n}}\frac{u^{p}(x)}{|x|^{a}}C\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})}{|x-y|^{n-2m}|y|^{a}}d{y}dx
\\
\nonumber &=& \frac{(2m-n)C}{2}\int_{\mathbb{R}^{n}}\int_{\mathbb{R}^{n}}\frac{|x-y|^{2}u^{p}(x)u^{p}(y)}{|x|^{a}|y|^{a}|x-y|^{n-2m+2}}dydx
\\
\nonumber &=& \frac{(2m-n)C}{2}\int_{\mathbb{R}^{n}}\int_{\mathbb{R}^{n}}\frac{[(x-y)\cdot x+(y-x)\cdot y]u^{p}(x)u^{p}(y)}{|x|^{a}|y|^{a}|x-y|^{n-2m+2}}dydx
\\
\nonumber &=& -(n-2m)C\int_{\mathbb{R}^{n}}\frac{u^{p}(x)}{|x|^{a}}\int_{\mathbb{R}^{n}}\frac{{u^{p}({y}})x\cdot (x-y)}{|x-y|^{n-2m+2}|y|^{a}}d{y}dx.
\end{eqnarray}
Combining \eqref{P14} and \eqref{P15}, we deduce further that
\begin{equation}\label{P16}
\left(\frac{2m-n}{2}+\frac{n-a}{p+1}\right)\int_{\mathbb{R}^{n}}\frac{u^{p+1}(x)}{|x|^{a}}dx=0.
\end{equation}
Since $1<p<\frac{n+2m-2a}{n-2m}$, it is easy to see that
\begin{equation}\label{P17}
\frac{2m-n}{2}+\frac{n-a}{p+1}>0,
\end{equation}
thus we must have
\begin{equation}\label{P18}
\int_{\mathbb{R}^{n}}\frac{u^{p+1}(x)}{|x|^{a}}dx=0,
\end{equation}
which is a contradiction with $u>0$! Therefore, \eqref{IE} does not have any positive radially symmetric classical solutions.
\end{proof}
Since we have proved the positive classical solution $u$ to PDE \eqref{PDE} is also a positive radially symmetric solution to IE \eqref{IE}, Theorem \ref{nonexistence} leads to a contradiction. Therefore, we must have $u\equiv0$ in $\mathbb{R}^{n}$, that is, the unique nonnegative solution to PDE \eqref{PDE} is $u\equiv0$ in $\mathbb{R}^{n}$.
This concludes the proof of Theorem \ref{Thm0}.
\section{Proof of Theorem \ref{Thm1}}
In this section, we will prove Theorem \ref{Thm1} via the method of moving planes in local way and blowing-up techniques.
\subsection{Boundary layer estimates}
In this subsection, we will first establish the following boundary layer estimates by applying Kelvin transform and the method of moving planes in local way. The properties of the boundary $\partial\Omega$ will play a crucial role in our discussions.
\begin{thm}\label{Boundary}
Assume one of the following two assumptions
\begin{equation*}
\text{i)} \,\,\, \Omega \,\, \text{is strictly convex}, \,\, 1<p<\frac{n+2m}{n-2m}, \quad\quad\quad\, \text{or} \quad\quad\quad\, \text{ii)} \,\,\, 1<p\leq\frac{n+2}{n-2}
\end{equation*}
holds. Then, there exists a $\bar{\delta}>0$ depending only on $\Omega$ such that, for any positive solution $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ to the higher order Navier problem \eqref{tNavier}, we have
\begin{equation*}
\|u\|_{L^{\infty}(\overline{\Omega}_{\bar{\delta}})}\leq C(n,m,p,\lambda_{1},\Omega),
\end{equation*}
where the boundary layer $\overline{\Omega}_{\bar{\delta}}:=\{x\in\overline{\Omega}\,|\,dist(x,\partial\Omega)\leq\bar{\delta}\}$.
\end{thm}
\begin{rem}\label{remark2}
When $m=1$, Theorem \ref{Boundary} still holds for $p=\frac{n+2}{n-2}$.
\end{rem}
\begin{proof}
We will carry out our proof of Theorem \ref{Boundary} by discussing the two different assumptions i) and ii) separately.
\emph{Case i)} $\Omega$ is strictly convex and $1<p<\frac{n+2m}{n-2m}$. For any $x^{0}\in\partial\Omega$, let $\nu^{0}$ be the unit internal normal vector of $\partial\Omega$ at $x^{0}$, we will show that $u(x)$ is monotone increasing along the internal normal direction in the region
\begin{equation}\label{3-1}
\overline{\Sigma_{\delta_{0}}}=\left\{x\in\overline{\Omega}\,|\,0\leq(x-x^{0})\cdot\nu^{0}\leq\delta_{0}\right\},
\end{equation}
where $\delta_{0}>0$ depends only on $x^{0}$ and $\Omega$.
To this end, we define the moving plane by
\begin{equation}\label{3-2}
T_{\lambda}:=\{x\in\mathbb{R}^n\,|\,(x-x^{0})\cdot\nu^{0}=\lambda\},
\end{equation}
and denote
\begin{equation}\label{3-2'}
\Sigma_{\lambda}:=\{x\in\Omega\,|\,0<(x-x^{0})\cdot\nu^{0}<\lambda\}
\end{equation}
for $\lambda>0$, and let $x^{\lambda}$ be the reflection of the point $x$ about the plane $T_{\lambda}$.
Let $u_{i}:=(-\Delta)^{i}u$ for $1\leq i\leq m-1$. By maximum principle, we have
\begin{equation}\label{3-3}
u_{i}(x)>0 \quad\quad\, \text{in} \,\, \Omega
\end{equation}
for $1\leq i\leq m-1$. Define
\begin{equation}\label{3-4}
W^{\lambda}(x):=u(x^{\lambda})-u(x) \quad\quad\, \text{and} \quad\quad\, W^{\lambda}_{i}(x):=u_{i}(x^{\lambda})-u_{i}(x)
\end{equation}
for $1\leq i\leq m-1$. Then we can deduce from \eqref{tNavier} that, for any $\lambda$ satisfying the reflection of $\Sigma_{\lambda}$ is contained in $\Omega$,
\begin{equation}\label{3-5}
\left\{{\begin{array}{l} {-\Delta W^{\lambda}_{m-1}(x)=u^{p}(x^{\lambda})-u^{p}(x)=p\eta^{p-1}_{\lambda}(x)W^{\lambda}(x), \,\,\,\,\,\, x\in\Sigma_{\lambda},}\\ {} \\ {-\Delta W^{\lambda}_{m-2}(x)=W^{\lambda}_{m-1}(x), \,\,\,\,\,\, x\in\Sigma_{\lambda},} \\ \cdots\cdots \\ {-\Delta W^{\lambda}(x)=W^{\lambda}_1(x), \,\,\,\,\,\, x\in\Sigma_{\lambda},} \\ {} \\
{W^{\lambda}(x)\geq0, \, W^{\lambda}_{1}(x)\geq0, \cdots, W^{\lambda}_{m-1}(x)\geq0, \,\,\,\,\,\, x\in\partial\Sigma_{\lambda},} \\ \end{array}}\right.
\end{equation}
where $\eta_{\lambda}(x)$ is valued between $u(x^{\lambda})$ and $u(x)$ by mean value theorem. Now, we will prove that there exists some $\delta>0$ sufficiently small (depending on $m$, $p$, $\|u\|_{L^{\infty}(\overline{\Omega})}$ and $\Omega$), such that
\begin{equation}\label{3-6}
W^{\lambda}(x)\geq 0 \quad\quad\, \text{in} \,\, \Sigma_{\lambda}
\end{equation}
for all $0<\lambda\leq\delta$. This provides a starting point to move the plane $T_{\lambda}$.
Indeed, suppose on the contrary that there exists a $0<\lambda\leq\delta$ such that
\begin{equation}\label{3-7}
W^{\lambda}(x)<0 \quad\quad\, \text{somewhere in} \,\, \Sigma_{\lambda}.
\end{equation}
Let
\begin{equation}\label{3-8}
\zeta(x):=\cos\frac{(x-x^{0})\cdot\nu^{0}}{\delta},
\end{equation}
then it follows that $\zeta(x)\in[\cos1,1]$ for any $x\in\Sigma_{\lambda}$ and $-\frac{\Delta\zeta(x)}{\zeta(x)}=\frac{1}{\delta^2}$. Define
\begin{equation}\label{3-9}
\overline{W^{\lambda}}(x):=\frac{W^{\lambda}(x)}{\zeta(x)} \quad\quad \text{and} \quad\quad \overline{W^{\lambda}_{i}}(x):=\frac{W^{\lambda}_{i}(x)}{\zeta(x)}
\end{equation}
for $i=1,\cdots, m-1$ and $x\in\Sigma_{\lambda}$. Then there exists a $x_{0}\in\Sigma_{\lambda}$ such that
\begin{equation}\label{3-10}
\overline{W^{\lambda}}(x_{0})=\min_{\overline{\Sigma_{\lambda}}}\overline{W^{\lambda}}(x)<0.
\end{equation}
Since
\begin{equation}\label{3-11}
-\Delta W^{\lambda}(x_{0})=-\Delta\overline{W^{\lambda}}(x_{0})\zeta(x_{0})-2\nabla\overline{W^{\lambda}}(x_{0})\cdot\nabla\zeta(x_{0})
-\overline{W^{\lambda}}(x_{0})\Delta\zeta(x_{0}),
\end{equation}
one immediately has
\begin{equation}\label{3-12}
W^{\lambda}_{1}(x_{0})=-\Delta W^{\lambda}(x_{0})\leq\frac{1}{\delta^2}W^{\lambda}(x_{0})<0.
\end{equation}
Thus there exists a $x_{1}\in\Sigma_{\lambda}$ such that
\begin{equation}\label{3-13}
\overline{W^{\lambda}_{1}}(x_{1})=\min_{\overline{\Sigma_{\lambda}}}\overline{W^{\lambda}_{1}}(x)<0.
\end{equation}
Similarly, it follows that
\begin{equation}\label{3-14}
W^{\lambda}_{2}(x_{1})=-\Delta W^{\lambda}_{1}(x_{1})\leq\frac{1}{\delta^2}W^{\lambda}_{1}(x_{1})<0.
\end{equation}
Continuing this way, we get $\{x_{i}\}_{i=1}^{ m-1}\subset\Sigma_{\lambda}$ such that
\begin{equation}\label{3-15}
\overline{W^{\lambda}_{i}}(x_{i})=\min_{\overline{\Sigma_{\lambda}}}\overline{W^{\lambda}_{i}}(x)<0,
\end{equation}
\begin{equation}\label{3-16}
W^{\lambda}_{i+1}(x_{i})=-\Delta W^{\lambda}_{i}(x_{i})\leq\frac{1}{\delta^2}W^{\lambda}_{i}(x_{i})<0
\end{equation}
for $i=1,2,\cdots, m-2$, and
\begin{equation}\label{3-17}
\overline{W^{\lambda}_{ m-1}}(x_{ m-1})=\min_{\overline{\Sigma_{\lambda}}}\overline{W^{\lambda}_{ m-1}}(x)<0,
\end{equation}
\begin{equation}\label{3-18}
p\eta^{p-1}_{\lambda}(x_{ m-1})W^{\lambda}(x_{ m-1})=-\Delta W^{\lambda}_{ m-1}(x_{ m-1})\leq\frac{1}{\delta^2}W^{\lambda}_{ m-1}(x_{ m-1})<0.
\end{equation}
Therefore, we have
\begin{eqnarray}\label{3-19}
W^{\lambda}(x_{0}) &\geq& \delta^2W^{\lambda}_{1}(x_{0})\geq \delta^{2}W^{\lambda}_{1}(x_{1})\frac{\zeta(x_{0})}{\zeta(x_{1})}
\geq\delta^{4}W^{\lambda}_{2}(x_{1})\frac{\zeta(x_{0})}{\zeta(x_{1})} \\
\nonumber &\geq& \delta^{4}W^{\lambda}_{2}(x_{2})\frac{\zeta(x_{0})}{\zeta(x_{2})}\geq\delta^{6}W^{\lambda}_{3}(x_{2})\frac{\zeta(x_{0})}{\zeta(x_{2})}
\geq\delta^{6}W^{\lambda}_{3}(x_{3})\frac{\zeta(x_{0})}{\zeta(x_{3})} \\
\nonumber &\geq& \cdots\cdots\geq\delta^{2m-2}W^{\lambda}_{ m-1}(x_{ m-1})\frac{\zeta(x_{0})}{\zeta(x_{ m-1})} \\
\nonumber &\geq& p\delta^{2m}\eta^{p-1}_{\lambda}(x_{ m-1})W^{\lambda}(x_{ m-1})\frac{\zeta(x_{0})}{\zeta(x_{ m-1})} \\
\nonumber &\geq& p\delta^{2m}\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})}W^{\lambda}(x_{0}),
\end{eqnarray}
that is,
\begin{equation}\label{3-20}
1\leq p\delta^{2m}\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})},
\end{equation}
which is absurd if we choose $\delta>0$ small enough such that
\begin{equation}\label{3-21}
0<\delta<\left(p\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})}\right)^{-\frac{1}{2m}}.
\end{equation}
So far, our conclusion is: the method of moving planes can be carried on up to $\lambda=\delta$.
Next, we will move the plane $T_{\lambda}$ further along the internal normal direction at $x^{0}$ as long as the property
\begin{equation}\label{3-22}
W^{\lambda}(x)\geq0 \quad\quad\, \text{in} \,\, \Sigma_{\lambda}
\end{equation}
holds. One can conclude that the moving planes process can be carried on (with the property \eqref{3-22}) as long as the reflection of $\overline{\Sigma_{\lambda}}$ is still contained in $\Omega$.
In fact, let $T_{\lambda_{0}}$ be a plane such that \eqref{3-22} holds and the reflection of $\overline{\Sigma_{\lambda_{0}}}$ about $T_{\lambda_{0}}$ is contained in $\Omega$. Then there exists a $\kappa>0$ such that, the reflection of $\overline{\Sigma_{\lambda_{0}+\kappa}}$ about $T_{\lambda_{0}+\kappa}$ is still contained in $\Omega$. By \eqref{3-5}, \eqref{3-22} and strong maximum principles, one actually has
\begin{equation}\label{3-23}
W^{\lambda_{0}}(x)>0, \quad\quad\, W_{i}^{\lambda_{0}}(x)>0 \quad\quad \text{in} \,\, \Sigma_{\lambda_{0}},
\end{equation}
thus there exists a constant $c_{\delta}>0$ such that
\begin{equation}\label{3-24}
W^{\lambda_{0}}(x)\geq c_{\delta}>0, \quad\quad\, W_{i}^{\lambda_{0}}(x)\geq c_{\delta}>0 \quad\quad \text{in} \,\, \overline{\Sigma_{\lambda_{0}-\frac{\delta}{2}}}.
\end{equation}
By the continuity of $u$, we infer that, there exists a $0<\epsilon<\min\{\kappa,\frac{\delta}{2}\}$ such that, for any $\lambda\in(\lambda_{0},\lambda_{0}+\epsilon]$,
\begin{equation}\label{3-25}
W^{\lambda}(x)>0, \quad\quad\, W_{i}^{\lambda}(x)>0 \quad\quad \text{in} \,\, \overline{\Sigma_{\lambda_{0}-\frac{\delta}{2}}}.
\end{equation}
Suppose there exists a $\lambda_{0}<\lambda\leq\lambda_{0}+\epsilon$ such that
\begin{equation}\label{3-26}
W^{\lambda}(x)<0 \quad\quad\, \text{somewhere in} \,\, \Sigma_{\lambda}\setminus\overline{\Sigma_{\lambda_{0}-\frac{\delta}{2}}}.
\end{equation}
Let
\begin{equation}\label{3-27}
\overline{\zeta}(x):=\cos\frac{\left(x-x^{0}-(\lambda_{0}-\frac{\delta}{2})\nu^{0}\right)\cdot\nu^{0}}{\delta} \quad\quad \text{and} \quad\quad \widetilde{W^{\lambda}}(x):=\frac{W^{\lambda}(x)}{\overline{\zeta}(x)}
\end{equation}
for $x\in\Sigma_{\lambda}\setminus\overline{\Sigma_{\lambda_{0}-\frac{\delta}{2}}}$. Then there exists a $x_{0}\in\Sigma_{\lambda}\setminus\overline{\Sigma_{\lambda_{0}-\frac{\delta}{2}}}$ such that
\begin{equation}\label{3-28}
\widetilde{W^{\lambda}}(x_{0})=\min_{\overline{\Sigma_{\lambda}\setminus\overline{\Sigma_{\lambda_{0}-\frac{\delta}{2}}}}}\widetilde{W^{\lambda}}(x)<0,
\end{equation}
by using similar arguments as proving \eqref{3-19}, one can also arrive at
\begin{equation}\label{3-29}
W^{\lambda}(x_{0})\geq p\delta^{2m}\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})}W^{\lambda}(x_{0}),
\end{equation}
which contradicts with the choice of $\delta$. Therefore, we have proved that
\begin{equation}\label{3-30}
W^{\lambda}(x)\geq0 \quad\quad\, \text{in} \,\, \Sigma_{\lambda}
\end{equation}
for any $\lambda\in(\lambda_{0},\lambda_{0}+\epsilon]$, that is, the plane $T_{\lambda}$ can be moved forward a little bit from $T_{\lambda_{0}}$.
Therefore, there exists a $\delta_{0}>0$ depending only on $x^{0}$ and $\Omega$ such that, $u(x)$ is monotone increasing along the internal normal direction in the region
\begin{equation}\label{3-31}
\overline{\Sigma_{\delta_{0}}}:=\left\{x\in\overline{\Omega}\,|\,0\leq(x-x^{0})\cdot\nu^{0}\leq\delta_{0}\right\}.
\end{equation}
Since $\partial\Omega$ is $C^{2m-2}$, there exists a small $0<r_{0}<\frac{\delta_{0}}{8}$ depending on $x^{0}$ and $\Omega$ such that, for any $x\in B_{r_{0}}(x^{0})\cap\partial\Omega$, $u(x)$ is monotone increasing along the internal normal direction at $x$ in the region
\begin{equation}\label{3-32}
\overline{\Sigma_{x}}:=\left\{z\in\overline{\Omega}\,\Big|\,0\leq(z-x)\cdot\nu_{x}\leq\frac{3}{4}\delta_{0}\right\}.
\end{equation}
where $\nu_{x}$ denotes the unit internal normal vector at the point $x$ ($\nu_{x^{0}}:=\nu^{0}$). Since $\Omega$ is strictly convex, there also exists a $\theta>0$ depending on $x^{0}$ and $\Omega$ such that
\begin{equation}\label{3-33}
I:=\left\{\nu\in\mathbb{R}^n\,|\,|\nu|=1, \, \nu\cdot\nu^{0}\geq\cos\theta\right\}\subset\left\{\nu_{x}\,|\,x\in B_{r_{0}}(x^{0})\cap\partial\Omega\right\},
\end{equation}
and hence, we have, for any $x\in B_{r_{0}}(x^{0})\cap\partial\Omega$ and $\nu\in I$,
\begin{equation}\label{3-34}
u(x+s\nu) \quad \text{is monotone increasing with respect to} \,\,\, s\in\left[0,\frac{\delta_{0}}{2}\right].
\end{equation}
Let
\begin{equation}\label{3-35}
D:=\{x+r_{0}\nu^{0}\,|\,x\in B_{r_{0}}(x^{0})\cap\partial\Omega\},
\end{equation}
one can easily verify that
\begin{equation}\label{3-36}
\max_{\overline{B_{r_{0}}(x^{0})\cap\Omega}}u(x)\leq\max_{\overline{D}}u(x).
\end{equation}
For any $x\in\overline{D}$, let
\begin{equation}\label{3-37}
\overline{V_{x}}:=\left\{x+\nu\,\Big|\,\nu\cdot\nu^{0}\geq|\nu|\cos\theta, \, |\nu|\leq\frac{\delta_{0}}{4}\right\}
\end{equation}
be a piece of cone with vertex at $x$, then it is easy to see that
\begin{equation}\label{3-38}
u(x)=\min_{z\in\overline{V_{x}}}u(z).
\end{equation}
Now we need the following Lemma to control the integral of $u$ on $\overline{V_{x}}$.
\begin{lem}\label{lemma2}
Let $\lambda_{1}$ be the first eigenvalue for $(-\Delta)^{m}$ in $\Omega$ with Navier boundary condition, and $0<\phi\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ be the corresponding eigenfunction (without loss of generality, we may assume $\|\phi\|_{L^{\infty}(\overline{\Omega})}=1$), i.e.,
\begin{equation*}\\\begin{cases}
(-\Delta)^{ m}\phi(x)=\lambda_{1}\phi(x) \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \Omega, \\
\phi(x)=-\Delta \phi(x)=\cdots=(-\Delta)^{ m-1}\phi(x)=0 \,\,\,\,\,\,\,\, \text{on} \,\,\, \partial\Omega.
\end{cases}\end{equation*}
Then, we have
\begin{equation*}
\int_{\Omega}u^{p}(x)\phi(x)dx\leq C(\lambda_{1},p,|\Omega|).
\end{equation*}
\end{lem}
\begin{proof}
Multiply both side of \eqref{tNavier} by the eigenfunction $\phi(x)$ and integrate by parts, one gets
\begin{eqnarray}\label{3-39}
\int_{\Omega}u^{p}(x)\phi(x)dx&\leq&\int_{\Omega}(u^{p}(x)+t)\phi(x)dx=\int_{\Omega}(-\Delta)^{ m}u(x)\cdot\phi(x)dx \\
\nonumber &=&\int_{\Omega}u(x)\cdot(-\Delta)^{ m}\phi(x)dx=\lambda_{1}\int_{\Omega}u(x)\phi(x)dx.
\end{eqnarray}
By H\"{o}lder's inequality, we have
\begin{equation}\label{3-40}
\int_{\Omega}u^p(x)\phi(x)dx\leq\lambda_{1}\left(\int_{\Omega}u^{p}(x)\phi(x)dx\right)^{\frac{1}{p}}\left(\int_{\Omega}\phi(x)dx\right)^{\frac{1}{p'}},
\end{equation}
and hence
\begin{equation}\label{3-41}
\int_{\Omega}u^{p}(x)\phi(x)dx\leq\lambda^{p'}_{1}\int_{\Omega}\phi(x)dx\leq\lambda^{p'}_{1}|\Omega|.
\end{equation}
This completes the proof of Lemma \ref{lemma2}.
\end{proof}
By \eqref{3-38} and Lemma \ref{lemma2}, we see that, for any $x\in\overline{D}$,
\begin{eqnarray}\label{3-42}
C(\lambda_{1},p,\Omega) &\geq& \int_{\Omega}u^{p}(x)\phi(x)dx\geq\int_{\overline{V_{x}}}u^{p}(z)\phi(z)dz \\
\nonumber &\geq& u^p(x)|\overline{V_{x}}|\cdot\min_{\overline{\Omega^{r_{0}}}}\phi=:u^{p}(x)\cdot C(n,m,x^{0},\Omega),
\end{eqnarray}
where $\overline{\Omega^{r_{0}}}:=\{x\in\Omega\,|\,dist(x,\partial\Omega)\geq r_{0}\}$, and hence
\begin{equation}\label{3-43}
u(x)\leq C(n,m,p,x^{0},\lambda_{1},\Omega), \quad\quad \forall x\in\overline{D}.
\end{equation}
Therefore, we arrive at
\begin{equation}\label{3-44}
\max_{\overline{B_{r_{0}}(x^{0})\cap\Omega}}u(x)\leq\max_{\overline{D}}u(x)\leq C(n,m,p,x^{0},\lambda_{1},\Omega).
\end{equation}
Since $x^{0}\in\partial\Omega$ is arbitrary and $\partial\Omega$ is compact, we can cover $\partial\Omega$ by finite balls $\{B_{r_{k}}(x^{k})\}_{k=0}^{K}$ with centers $\{x^{k}\}_{k=0}^{K}\subset\partial\Omega$ ($K$ depends only on $\Omega$). Therefore, there exists a $\bar{\delta}>0$ depending only on $\Omega$ such that
\begin{equation}\label{3-45}
\|u\|_{L^{\infty}(\overline{\Omega}_{\bar{\delta}})}\leq\max_{0\leq k\leq K}\max_{\overline{B_{r_{k}}(x^{k})\cap\Omega}}u(x)\leq\max_{0\leq k\leq K}C(n,m,p,x^{k},\lambda_{1},\Omega)=:C(n,m,p,\lambda_{1},\Omega),
\end{equation}
where the boundary layer $\overline{\Omega}_{\bar{\delta}}:=\{x\in\overline{\Omega}\,|\,dist(x,\partial\Omega)\leq\bar{\delta}\}$. This completes the proof of boundary layer estimates under assumption i).
\emph{Case ii)} $1<p\leq\frac{n+2}{n-2}$. Under this assumption, we do not require the convexity of $\Omega$ anymore. Since $\partial\Omega$ is $C^{2m-2}$, there exists a $R_{0}>0$ depending only on $\Omega$ such that, for any $x^{0}\in\partial\Omega$, there exists a $\overline{x^{0}}$ satisfying $\overline{B_{R_{0}}(\overline{x^{0}})}\cap\overline{\Omega}=\{x^{0}\}$. For any $x^{0}\in\partial\Omega$, we define the Kelvin transform centered at $\overline{x^{0}}$ by
\begin{equation}\label{3-46}
x\mapsto x^{\ast}:=\frac{x-\overline{x^{0}}}{|x-\overline{x^{0}}|^{2}}+\overline{x^{0}}, \quad\quad \Omega\rightarrow\Omega^{\ast}\subset B_{\frac{1}{R_{0}}}(\overline{x^{0}}),
\end{equation}
and hence there exists a small $0<\varepsilon_{0}<\frac{1}{100R_{0}}$ depending on $x^{0}$ and $\Omega$ such that $B_{\varepsilon_{0}}\big((\overline{x^{0}})^{\ast}\big)\cap\partial\Omega^{\ast}$ is strictly convex.
Now we define
\begin{equation}\label{3-47}
\overline{u}(x^{\ast}):=\frac{1}{|x^{\ast}-\overline{x^{0}}|^{n-2}}u\left(\frac{x^{\ast}-\overline{x^{0}}}{|x^{\ast}-\overline{x^{0}}|^{2}}+\overline{x^{0}}\right),
\end{equation}
\begin{equation}\label{3-48}
\overline{u_{i}}(x^{\ast}):=\frac{1}{|x^{\ast}-\overline{x^{0}}|^{n-2}}u_{i}\left(\frac{x^{\ast}-\overline{x^{0}}}{|x^{\ast}-\overline{x^{0}}|^{2}}+\overline{x^{0}}\right)
\end{equation}
for $i=1,\cdots, m-1$. Then, we have
\begin{equation}\label{3-48'}
\overline{u}(x^{\ast})>0, \quad\quad \overline{u_{i}}(x^{\ast})>0 \quad\quad \text{in} \,\,\, \Omega^{\ast},
\end{equation}
and from \eqref{tNavier}, we infer that $\overline{u}(x^{\ast})$ and $\overline{u_{i}}(x^{\ast})$ satisfy
\begin{equation}\label{3-49}
\left\{{\begin{array}{l} {-\Delta \overline{u_{ m-1}}(x^{\ast})=\frac{1}{|x^{\ast}-\overline{x^{0}}|^{\tau}}\overline{u}^{p}(x^{\ast})+\frac{t}{|x^{\ast}-\overline{x^{0}}|^{n+2}}, \,\,\,\,\,\, x^{\ast}\in\Omega^{\ast},}\\ {} \\ {-\Delta \overline{u_{ m-2}}(x^{\ast})=\frac{1}{|x^{\ast}-\overline{x^{0}}|^{4}}\overline{u_{ m-1}}(x^{\ast}), \,\,\,\,\,\, x^{\ast}\in\Omega^{\ast},} \\ \cdots\cdots \\ {-\Delta \overline{u}(x^{\ast})=\frac{1}{|x^{\ast}-\overline{x^{0}}|^{4}}\overline{u_{1}}(x^{\ast}), \,\,\,\,\,\, x^{\ast}\in\Omega^{\ast},} \\ {} \\
{\overline{u}(x^{\ast})=\overline{u_{1}}(x^{\ast})=\cdots=\overline{u_{ m-1}}(x^{\ast})=0, \,\,\,\,\,\, x^{\ast}\in\partial\Omega^{\ast},} \\ \end{array}}\right.
\end{equation}
where $\tau:=n+2-p(n-2)\geq0$. Let $\nu^{0}$ be the unit internal normal vector of $\partial\Omega^{\ast}$ at $(x^{0})^{\ast}$, we will show that $\overline{u}(x^{\ast})$ is monotone increasing along the internal normal direction in the region
\begin{equation}\label{3-50}
\overline{\Sigma_{\delta_{\ast}}}=\left\{x^{\ast}\in\overline{\Omega^{\ast}}\,|\,0\leq(x^{\ast}-(x^{0})^{\ast})\cdot\nu^{0}\leq\delta_{\ast}\right\},
\end{equation}
where $\delta_{\ast}>0$ depends only on $x^{0}$ and $\Omega$.
For this purpose, we define the moving plane by
\begin{equation}\label{3-51}
T^{\ast}_{\lambda}:=\{x^{\ast}\in\mathbb{R}^n\,|\,(x^{\ast}-(x^{0})^{\ast})\cdot\nu^{0}=\lambda\},
\end{equation}
and denote
\begin{equation}\label{3-52}
\Sigma^{\ast}_{\lambda}:=\{x^{\ast}\in\Omega^{\ast}\,|\,0<(x^{\ast}-(x^{0})^{\ast})\cdot\nu^{0}<\lambda\}
\end{equation}
for $\lambda>0$, and let $x^{\ast}_{\lambda}$ be the reflection of the point $x^{\ast}$ about the plane $T^{\ast}_{\lambda}$.
Define
\begin{equation}\label{3-53}
U^{\lambda}(x^{\ast}):=\overline{u}(x^{\ast}_{\lambda})-\overline{u}(x^{\ast}) \quad\quad\, \text{and} \quad\quad\, U^{\lambda}_{i}(x^{\ast}):=\overline{u_{i}}(x^{\ast}_{\lambda})-\overline{u_{i}}(x^{\ast})
\end{equation}
for $1\leq i\leq m-1$. Then we can deduce from \eqref{3-49} that, for any $\lambda$ satisfying the reflection of $\Sigma^{\ast}_{\lambda}$ is contained in $\Omega^{\ast}$,
\begin{equation}\label{3-54}
\left\{{\begin{array}{l} {-\Delta U^{\lambda}_{ m-1}(x^{\ast})=\frac{\overline{u}^{p}(x^{\ast}_{\lambda})}{|x^{\ast}_{\lambda}-\overline{x^{0}}|^{\tau}}
-\frac{\overline{u}^{p}(x^{\ast})}{|x^{\ast}-\overline{x^{0}}|^{\tau}}
+\frac{t}{|x^{\ast}_{\lambda}-\overline{x^{0}}|^{n+2}}-\frac{t}{|x^{\ast}-\overline{x^{0}}|^{n+2}}, \,\,\,\,\,\, x^{\ast}\in\Sigma^{\ast}_{\lambda},}\\ {} \\ {-\Delta U^{\lambda}_{ m-2}(x^{\ast})=\frac{\overline{u_{ m-1}}(x^{\ast}_{\lambda})}{|x^{\ast}_{\lambda}-\overline{x^{0}}|^{4}}
-\frac{\overline{u_{ m-1}}(x^{\ast})}{|x^{\ast}-\overline{x^{0}}|^{4}}, \,\,\,\,\,\, x^{\ast}\in\Sigma^{\ast}_{\lambda},} \\ \cdots\cdots \\ {-\Delta U^{\lambda}(x^{\ast})=\frac{\overline{u_{1}}(x^{\ast}_{\lambda})}{|x^{\ast}_{\lambda}-\overline{x^{0}}|^{4}}
-\frac{\overline{u_{1}}(x^{\ast})}{|x^{\ast}-\overline{x^{0}}|^{4}}, \,\,\,\,\,\, x^{\ast}\in\Sigma^{\ast}_{\lambda},} \\ {} \\
{U^{\lambda}(x^{\ast})\geq0, \, U^{\lambda}_{1}(x^{\ast})\geq0, \cdots, U^{\lambda}_{ m-1}(x^{\ast})\geq0, \,\,\,\,\,\, x^{\ast}\in\partial\Sigma^{\ast}_{\lambda}.} \\ \end{array}}\right.
\end{equation}
Notice that for any $x^{\ast}\in\Sigma^{\ast}_{\lambda}$ with $\lambda<\frac{1}{R_{0}}$, one has
\begin{equation}\label{3-55}
0<|x^{\ast}_{\lambda}-\overline{x^{0}}|<|x^{\ast}-\overline{x^{0}}|<\frac{1}{R_{0}},
\end{equation}
and hence, by direct calculations, it follows from \eqref{3-54} and $t\geq0$ that
\begin{equation}\label{3-56}
\left\{{\begin{array}{l} {-\Delta U^{\lambda}_{ m-1}(x^{\ast})\geq\frac{p\xi^{p-1}_{\lambda}(x^{\ast})}{|x^{\ast}-\overline{x^{0}}|^{\tau}}U^{\lambda}(x^{\ast})\geq pR^{\tau}_{0}\xi^{p-1}_{\lambda}(x^{\ast})U^{\lambda}(x^{\ast}), \,\,\,\,\,\, x^{\ast}\in\Sigma^{\ast}_{\lambda},}\\ {} \\ {-\Delta U^{\lambda}_{ m-2}(x^{\ast})\geq R^{4}_{0}\,U^{\lambda}_{ m-1}(x^{\ast}), \,\,\,\,\,\, x^{\ast}\in\Sigma^{\ast}_{\lambda},} \\ \cdots\cdots \\ {-\Delta U^{\lambda}(x^{\ast})\geq R^{4}_{0}\,U^{\lambda}_{1}(x^{\ast}), \,\,\,\,\,\, x^{\ast}\in\Sigma^{\ast}_{\lambda},} \\ {} \\
{U^{\lambda}(x^{\ast})\geq0, \, U^{\lambda}_{1}(x^{\ast})\geq0, \cdots, U^{\lambda}_{ m-1}(x^{\ast})\geq0, \,\,\,\,\,\, x^{\ast}\in\partial\Sigma^{\ast}_{\lambda}.} \\ \end{array}}\right.
\end{equation}
where $\xi_{\lambda}(x^{\ast})$ is valued between $\overline{u}(x^{\ast}_{\lambda})$ and $\overline{u}(x^{\ast})$ by mean value theorem, and thus
\begin{equation}\label{3-57}
\|\xi_{\lambda}\|_{L^{\infty}(\overline{\Sigma^{\ast}_{\lambda}})}\leq\left(diam\,\Omega+R_{0}\right)^{n-2}\|u\|_{L^{\infty}(\overline{\Omega})}.
\end{equation}
Now, we will prove that there exists some $\delta>0$ sufficiently small (depending on $m$, $p$, $\|u\|_{L^{\infty}(\overline{\Omega})}$ and $\Omega$), such that
\begin{equation}\label{3-58}
U^{\lambda}(x^{\ast})\geq 0 \quad\quad\, \text{in} \,\, \Sigma^{\ast}_{\lambda}
\end{equation}
for all $0<\lambda\leq\delta$. This provides a starting point to move the plane $T^{\ast}_{\lambda}$.
In fact, suppose on the contrary that there exists a $0<\lambda\leq\delta$ such that
\begin{equation}\label{3-59}
U^{\lambda}(x^{\ast})<0 \quad\quad\, \text{somewhere in} \,\, \Sigma^{\ast}_{\lambda}.
\end{equation}
Let
\begin{equation}\label{3-60}
\psi(x^{\ast}):=\cos\frac{(x^{\ast}-(x^{0})^{\ast})\cdot\nu^{0}}{\delta},
\end{equation}
then $\psi(x^{\ast})\in[\cos1,1]$ for any $x^{\ast}\in\Sigma^{\ast}_{\lambda}$ and $-\frac{\Delta\psi}{\psi}=\frac{1}{\delta^2}$. Define
\begin{equation}\label{3-61}
\overline{U^{\lambda}}(x^{\ast}):=\frac{U^{\lambda}(x^{\ast})}{\psi(x^{\ast})} \quad\quad \text{and} \quad\quad \overline{U^{\lambda}_{i}}(x^{\ast}):=\frac{U^{\lambda}_{i}(x^{\ast})}{\psi(x^{\ast})}
\end{equation}
for $i=1,\cdots, m-1$ and $x^{\ast}\in\Sigma^{\ast}_{\lambda}$. Then there exists a $x^{\ast}_{0}\in\Sigma^{\ast}_{\lambda}$ such that
\begin{equation}\label{3-62}
\overline{U^{\lambda}}(x^{\ast}_{0})=\min_{\overline{\Sigma^{\ast}_{\lambda}}}\overline{U^{\lambda}}(x^{\ast})<0.
\end{equation}
Since
\begin{equation}\label{3-63}
-\Delta U^{\lambda}(x^{\ast}_{0})=-\Delta\overline{U^{\lambda}}(x^{\ast}_{0})\psi(x^{\ast}_{0})-2\nabla\overline{U^{\lambda}}(x^{\ast}_{0})\cdot\nabla\psi(x^{\ast}_{0})
-\overline{U^{\lambda}}(x^{\ast}_{0})\Delta\psi(x^{\ast}_{0}),
\end{equation}
one immediately has
\begin{equation}\label{3-64}
R^{4}_{0}\,U^{\lambda}_{1}(x^{\ast}_{0})\leq-\Delta U^{\lambda}(x^{\ast}_{0})\leq\frac{1}{\delta^2}U^{\lambda}(x^{\ast}_{0})<0.
\end{equation}
Thus there exists a $x^{\ast}_{1}\in\Sigma^{\ast}_{\lambda}$ such that
\begin{equation}\label{3-65}
\overline{U^{\lambda}_{1}}(x^{\ast}_{1})=\min_{\overline{\Sigma^{\ast}_{\lambda}}}\overline{U^{\lambda}_{1}}(x^{\ast})<0.
\end{equation}
Similarly, it follows that
\begin{equation}\label{3-66}
R^{4}_{0}\,U^{\lambda}_{2}(x^{\ast}_{1})\leq-\Delta U^{\lambda}_{1}(x^{\ast}_{1})\leq\frac{1}{\delta^2}U^{\lambda}_{1}(x^{\ast}_{1})<0.
\end{equation}
Continuing this way, we get $\{x^{\ast}_{i}\}_{i=1}^{ m-1}\subset\Sigma^{\ast}_{\lambda}$ such that
\begin{equation}\label{3-67}
\overline{U^{\lambda}_{i}}(x^{\ast}_{i})=\min_{\overline{\Sigma^{\ast}_{\lambda}}}\overline{U^{\lambda}_{i}}(x^{\ast})<0,
\end{equation}
\begin{equation}\label{3-68}
R^{4}_{0}\,U^{\lambda}_{i+1}(x^{\ast}_{i})\leq-\Delta U^{\lambda}_{i}(x^{\ast}_{i})\leq\frac{1}{\delta^2}U^{\lambda}_{i}(x^{\ast}_{i})<0
\end{equation}
for $i=1,2,\cdots, m-2$, and
\begin{equation}\label{3-69}
\overline{U^{\lambda}_{ m-1}}(x^{\ast}_{ m-1})=\min_{\overline{\Sigma^{\ast}_{\lambda}}}\overline{U^{\lambda}_{ m-1}}(x^{\ast})<0,
\end{equation}
\begin{equation}\label{3-70}
pR^{\tau}_{0}\xi^{p-1}_{\lambda}(x^{\ast}_{ m-1})U^{\lambda}(x^{\ast}_{ m-1})\leq-\Delta U^{\lambda}_{ m-1}(x^{\ast}_{ m-1})\leq\frac{1}{\delta^2}U^{\lambda}_{ m-1}(x^{\ast}_{ m-1})<0.
\end{equation}
Therefore, we have
\begin{eqnarray}\label{3-71}
U^{\lambda}(x^{\ast}_{0}) &\geq& (\delta R^{2}_{0})^{2}U^{\lambda}_{1}(x^{\ast}_{0})\geq (\delta R^{2}_{0})^{2}U^{\lambda}_{1}(x^{\ast}_{1})\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{1})} \\
\nonumber &\geq& (\delta R^{2}_{0})^{4}U^{\lambda}_{2}(x^{\ast}_{1})\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{1})}
\geq(\delta R^{2}_{0})^{4}U^{\lambda}_{2}(x^{\ast}_{2})\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{2})} \\
\nonumber &\geq& (\delta R^{2}_{0})^{6}U^{\lambda}_{3}(x^{\ast}_{2})\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{2})}
\geq(\delta R^{2}_{0})^{6}U^{\lambda}_{3}(x^{\ast}_{3})\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{3})} \\
\nonumber &\geq& \cdots\cdots\geq(\delta R^{2}_{0})^{2m-2}U^{\lambda}_{ m-1}(x^{\ast}_{ m-1})\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{ m-1})} \\
\nonumber &\geq& p\delta^{2m}R^{4m-(p-1)(n-2)}_{0}\xi^{p-1}_{\lambda}(x^{\ast}_{ m-1})U^{\lambda}(x^{\ast}_{ m-1})
\frac{\psi(x^{\ast}_{0})}{\psi(x^{\ast}_{ m-1})} \\
\nonumber &\geq& p\delta^{2m}R^{4m-(p-1)(n-2)}_{0}\left(diam\,\Omega+R_{0}\right)^{(p-1)(n-2)}\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})}U^{\lambda}(x^{\ast}_{0}),
\end{eqnarray}
that means,
\begin{equation}\label{3-72}
1\leq p\delta^{2m}\left(diam\,\Omega+R_{0}\right)^{4m}\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})},
\end{equation}
which is absurd if we choose $\delta>0$ small enough such that
\begin{equation}\label{3-73}
0<\delta<\left(diam\,\Omega+R_{0}\right)^{-2}\left(p\|u\|^{p-1}_{L^{\infty}(\overline{\Omega})}\right)^{-\frac{1}{2m}}.
\end{equation}
So far, we have proved that the plane $T^{\ast}_{\lambda}$ can be moved on up to $\lambda=\delta$.
Next, we will move the plane $T^{\ast}_{\lambda}$ further along the internal normal direction at $(x^{0})^{\ast}$ as long as the property
\begin{equation}\label{3-74}
U^{\lambda}(x^{\ast})\geq0 \quad\quad\, \text{in} \,\, \Sigma^{\ast}_{\lambda}
\end{equation}
holds. Completely similar to the proof of \emph{Case i)}, one can actually show that the method of moving planes can be carried on (with the property \eqref{3-74}) as long as the reflection of $\overline{\Sigma^{\ast}_{\lambda}}$ is still contained in $\Omega^{\ast}$. We omit the details here.
Therefore, there exists a $\delta_{\ast}>0$ depending only on $x^{0}$ and $\Omega$ such that, $\overline{u}(x^{\ast})$ is monotone increasing along the internal normal direction in the region
\begin{equation}\label{3-75}
\overline{\Sigma_{\delta_{\ast}}}:=\left\{x^{\ast}\in\overline{\Omega^{\ast}}\,|\,0\leq\left(x^{\ast}-(x^{0})^{\ast}\right)\cdot\nu^{0}\leq\delta_{\ast}\right\}.
\end{equation}
Since $\partial\Omega^{\ast}$ is $C^{2m-2}$, there exists a small $0<\varepsilon_{1}<\min\{\frac{\delta_{\ast}}{8},\varepsilon_{0}\}$ depending on $x^{0}$ and $\Omega$ such that, for any $x^{\ast}\in B_{\varepsilon_{1}}\big((x^{0})^{\ast}\big)\cap\partial\Omega^{\ast}$, $\overline{u}(x^{\ast})$ is monotone increasing along the internal normal direction at $x^{\ast}$ in the region
\begin{equation}\label{3-76}
\overline{\Sigma_{x^{\ast}}}:=\left\{z^{\ast}\in\overline{\Omega^{\ast}}\,\Big|\,0\leq(z^{\ast}-x^{\ast})\cdot\nu_{x^{\ast}}\leq\frac{3}{4}\delta_{\ast}\right\}.
\end{equation}
where $\nu_{x^{\ast}}$ denotes the unit internal normal vector at the point $x^{\ast}$ ($\nu_{(x^{0})^{\ast}}:=\nu^{0}$). Since $B_{\varepsilon_{1}}\big((x^{0})^{\ast}\big)\cap\partial\Omega^{\ast}$ is strictly convex, there exists a $\theta>0$ depending on $x^{0}$ and $\Omega$ such that
\begin{equation}\label{3-77}
S:=\left\{\nu^{\ast}\in\mathbb{R}^n\,|\,|\nu^{\ast}|=1, \, \nu^{\ast}\cdot\nu^{0}\geq\cos\theta\right\}\subset\left\{\nu_{x^{\ast}}\,|\,x^{\ast}\in B_{\varepsilon_{1}}\big((x^{0})^{\ast}\big)\cap\partial\Omega^{\ast}\right\},
\end{equation}
and hence, it follows that, for any $x^{\ast}\in B_{\varepsilon_{1}}\big((x^{0})^{\ast}\big)\cap\partial\Omega^{\ast}$ and $\nu^{\ast}\in S$,
\begin{equation}\label{3-78}
\overline{u}(x^{\ast}+s\nu^{\ast}) \quad \text{is monotone increasing with respect to} \,\,\, s\in\left[0,\frac{\delta_{\ast}}{2}\right].
\end{equation}
Now, let
\begin{equation}\label{3-79}
D^{\ast}:=\left\{x^{\ast}+\varepsilon_{1}\nu^{0}\,|\,x^{\ast}\in B_{\varepsilon_{1}}\big((x^{0})^{\ast}\big)\cap\partial\Omega^{\ast}\right\},
\end{equation}
one immediately has
\begin{equation}\label{3-80}
\max_{\overline{B_{\varepsilon_{1}}((x^{0})^{\ast})\cap\Omega^{\ast}}}\overline{u}(x^{\ast})\leq\max_{\overline{D^{\ast}}}\overline{u}(x^{\ast}).
\end{equation}
For any $x^{\ast}\in\overline{D^{\ast}}$, let
\begin{equation}\label{3-81}
\overline{V_{x^{\ast}}}:=\left\{x^{\ast}+\nu^{\ast}\,\Big|\,\nu^{\ast}\cdot\nu^{0}\geq|\nu^{\ast}|\cos\theta, \, |\nu^{\ast}|\leq\frac{\delta_{\ast}}{4}\right\}
\end{equation}
be a piece of cone with vertex at $x^{\ast}$, then it is obvious that
\begin{equation}\label{3-82}
\overline{u}(x^{\ast})=\min_{z^{\ast}\in\overline{V_{x^{\ast}}}}\overline{u}(z^{\ast}).
\end{equation}
Therefore, by \eqref{3-82} and Lemma \ref{lemma2}, we get, for any $x^{\ast}\in\overline{D^{\ast}}$,
\begin{eqnarray}\label{3-83}
&& C(\lambda_{1},p,\Omega)\geq\int_{\Omega}u^{p}(x)\phi(x)dx \\
\nonumber &=&\int_{\Omega^{\ast}}\frac{\overline{u}^{p}(x^{\ast})}{|x^{\ast}-\overline{x^{0}}|^{2n-p(n-2)}}
\phi\left(\frac{x^{\ast}-\overline{x^{0}}}{|x^{\ast}-\overline{x^{0}}|^2}+\overline{x^{0}}\right)dx^{\ast} \\
\nonumber &\geq& \int_{\overline{V_{x^{\ast}}}}\frac{\overline{u}^{p}(z^{\ast})}{|z^{\ast}-\overline{x^{0}}|^{2n-p(n-2)}}
\phi\left(\frac{z^{\ast}-\overline{x^{0}}}{|z^{\ast}-\overline{x^{0}}|^2}+\overline{x^{0}}\right)dz^{\ast}\\
\nonumber &\geq& \overline{u}^p(x^{\ast})R^{2n-p(n-2)}_{0}|\overline{V_{x^{\ast}}}|\cdot\min_{\overline{\Omega^{r_{1}}}}\phi
=:\overline{u}^{p}(x^{\ast})\cdot C(n,m,p,x^{0},\Omega),
\end{eqnarray}
where $\overline{\Omega^{r_{1}}}:=\{x\in\Omega\,|\,dist(x,\partial\Omega)\geq r_{1}\}$ with $r_{1}=\varepsilon_{1}R^{2}_{0}$, and hence
\begin{equation}\label{3-84}
\overline{u}(x^{\ast})\leq C(n,m,p,x^{0},\lambda_{1},\Omega), \quad\quad \forall x\in\overline{D^{\ast}}.
\end{equation}
As a consequence, we derive that
\begin{equation}\label{3-85}
\max_{\overline{B_{\varepsilon_{1}}((x^{0})^{\ast})\cap\Omega^{\ast}}}\overline{u}(x^{\ast})\leq\max_{\overline{D^{\ast}}}\overline{u}(x^{\ast})\leq C(n,m,p,x^{0},\lambda_{1},\Omega).
\end{equation}
There exists a small $r_{0}>0$ depending only on $x^{0}$ and $\Omega$ such that, for each $x\in\overline{B_{r_{0}}(x^{0})\cap\Omega}$, one has $x^{\ast}\in\overline{B_{\varepsilon_{1}}\big((x^{0})^{\ast}\big)\cap\Omega^{\ast}}$. Therefore, \eqref{3-85} yields
\begin{eqnarray}\label{3-86}
\max_{\overline{B_{r_{0}}(x^{0})\cap\Omega}}u(x)&=&\max_{x\in\overline{B_{r_{0}}(x^{0})\cap\Omega}}|x^{\ast}-\overline{x^{0}}|^{n-2}\overline{u}(x^{\ast}) \\
\nonumber &\leq& \frac{1}{R^{n-2}_{0}}\max_{\overline{B_{\varepsilon_{1}}((x^{0})^{\ast})\cap\Omega^{\ast}}}\overline{u}(x^{\ast})\leq C(n,m,p,x^{0},\lambda_{1},\Omega).
\end{eqnarray}
Since $x^{0}\in\partial\Omega$ is arbitrary and $\partial\Omega$ is compact, we can cover $\partial\Omega$ by finite balls $\{B_{r_{k}}(x^{k})\}_{k=0}^{K}$ with centers $\{x^{k}\}_{k=0}^{K}\subset\partial\Omega$ ($K$ depends only on $\Omega$). Therefore, there exists a $\bar{\delta}>0$ depending only on $\Omega$ such that
\begin{equation}\label{3-87}
\|u\|_{L^{\infty}(\overline{\Omega}_{\bar{\delta}})}\leq\max_{0\leq k\leq K}\max_{\overline{B_{r_{k}}(x^{k})\cap\Omega}}u(x)\leq\max_{0\leq k\leq K}C(n,m,p,x^{k},\lambda_{1},\Omega)=:C(n,m,p,\lambda_{1},\Omega),
\end{equation}
where the boundary layer $\overline{\Omega}_{\bar{\delta}}:=\{x\in\overline{\Omega}\,|\,dist(x,\partial\Omega)\leq\bar{\delta}\}$. This completes the proof of boundary layer estimates under assumption ii).
This concludes our proof of Theorem \ref{Boundary}.
\end{proof}
\subsection{Blowing-up analysis and interior estimates}
In this subsection, we will obtain the interior estimates (and hence, global a priori estimates) via the blowing-up analysis arguments (for related literatures on blowing-up methods, please refer to \cite{BC,BM,CDQ,CL3,CL4,CY0,Li,SZ1}).
Suppose on the contrary that Theorem \ref{Thm1} does not hold. By the boundary layer estimates (Theorem \ref{Boundary}), there exists a sequence of positive solutions $\{u_{k}\}\subset C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ to the higher order Navier problem \eqref{tNavier} and a sequence of interior points $\{x^{k}\}\subset\Omega\setminus\overline{\Omega}_{\bar{\delta}}$ such that
\begin{equation}\label{32-1}
m_{k}:=u_{k}(x^{k})=\|u_{k}\|_{L^{\infty}(\overline{\Omega})}\rightarrow+\infty \quad \text{as} \,\, k\rightarrow\infty.
\end{equation}
For $x\in\Omega_{k}:=\{x\in\mathbb{R}^{n}\,|\,\lambda_{k}x+x^{k}\in\Omega\}$, we define
\begin{equation}\label{32-2}
v_{k}(x):=\frac{1}{m_{k}}u_{k}(\lambda_{k}x+x^{k}) \quad \text{with} \,\, \lambda_{k}:=m_{k}^{\frac{1-p}{2m}}\rightarrow0 \quad \text{as} \,\, k\rightarrow\infty.
\end{equation}
Then $v_{k}(x)$ satisfies $\|v_{k}\|_{L^{\infty}(\overline{\Omega_{k}})}=v_{k}(0)=1$ and
\begin{eqnarray}\label{32-3}
(-\Delta)^{m}v_{k}(x)&=&\frac{1}{m_{k}}\lambda^{2m}_{k}(-\Delta)^{m}u_{k}(\lambda_{k}x+x^{k}) \\
\nonumber &=& \frac{1}{m_{k}}\lambda^{2m}_{k}\left(u^{p}_{k}(\lambda_{k}x+x^{k})+t\right)=v^{p}_{k}(x)+\frac{t}{m^{p}_{k}}
\end{eqnarray}
for any $x\in\Omega_{k}$. Since $dist(x^{k},\partial\Omega)>\bar{\delta}$, one has
\begin{equation}\label{32-4}
\Omega_{k}\supset\left\{x\in\mathbb{R}^{n}\,|\,|\lambda_{k}x|\leq\bar{\delta}\right\}=\overline{B_{\frac{\bar{\delta}}{\lambda_{k}}}(0)},
\end{equation}
and hence
\begin{equation}\label{32-5}
\Omega_{k}\rightarrow\mathbb{R}^{n} \quad \text{as} \,\, k\rightarrow\infty.
\end{equation}
For arbitrary $x^{0}\in\mathbb{R}^{n}$, there exists a $N_{1}>0$, such that $\overline{B_{1}(x^{0})}\subset\Omega_{k}$ for any $k\geq N_{1}$. By \eqref{32-3} and $\|v_{k}\|_{L^{\infty}(\overline{\Omega_{k}})}\leq1$, we can infer from regularity theory and Sobolev embedding that
\begin{equation}\label{32-6}
\|v_{k}\|_{C^{2m-1,\gamma}(\overline{B_{1}(0)})}\leq C(1+t),
\end{equation}
and further that
\begin{equation}\label{32-7}
\|v_{k}\|_{C^{2(2m-1),\gamma}(\overline{B_{1}(0)})}\leq C(1+t)
\end{equation}
for $k\geq N_{1}$, where $0\leq\gamma<1$. As a consequence, by Arzel\`{a}-Ascoli Theorem, there exists a subsequence $\{v^{(1)}_{k}\}\subset\{v_{k}\}$ and a function $v\in C^{2m}(\overline{B_{1}(x^{0})})$ such that
\begin{equation}\label{32-8}
v^{(1)}_{k}\rightrightarrows v \quad \text{and} \quad (-\Delta)^{ m}v^{(1)}_{k}\rightrightarrows(-\Delta)^{ m}v \quad\quad \text{in} \,\, \overline{B_{1}(x^{0})}.
\end{equation}
There also exists a $N_{2}>0$ such that $\overline{B_{2}(x^{0})}\subset\Omega_{k}$ for any $k\geq N_{2}$. By \eqref{32-3} and $\|v_{k}\|_{L^{\infty}(\overline{\Omega_{k}})}\leq1$, we can deduce that
\begin{equation}\label{32-9}
\|v^{(1)}_{k}\|_{C^{2(2m-1),\gamma}(\overline{B_{2}(0)})}\leq C(1+t)
\end{equation}
for $k\geq N_{2}$, where $0\leq\gamma<1$. Therefore, by Arzel\`{a}-Ascoli Theorem again, there exists a subsequence $\{v^{(2)}_{k}\}\subset\{v^{(1)}_{k}\}$ and $v\in C^{2m}(\overline{B_{2}(x^{0})})$ such that
\begin{equation}\label{32-10}
v^{(2)}_{k}\rightrightarrows v \quad \text{and} \quad (-\Delta)^{ m}v^{(2)}_{k}\rightrightarrows(-\Delta)^{ m}v \quad\quad \text{in} \,\, \overline{B_{2}(x^{0})}.
\end{equation}
Continuing this way, for any $j\in\mathbb{N}^{+}$, we can extract a subsequence $\{v^{(j)}_{k}\}\subset\{v^{(j-1)}_{k}\}$ and find a function $v\in C^{2m}(\overline{B_{j}(x^{0})})$ such that
\begin{equation}\label{32-11}
v^{(j)}_{k}\rightrightarrows v \quad \text{and} \quad (-\Delta)^{ m}v^{(j)}_{k}\rightrightarrows(-\Delta)^{ m}v \quad\quad \text{in} \,\, \overline{B_{j}(x^{0})}.
\end{equation}
By extracting the diagonal sequence, we finally obtain that the subsequence $\{v^{(k)}_{k}\}$ satisfies
\begin{equation}\label{32-12}
v^{(k)}_{k}\rightrightarrows v \quad \text{and} \quad (-\Delta)^{ m}v^{(k)}_{k}\rightrightarrows(-\Delta)^{ m}v \quad\quad \text{in} \,\, \overline{B_{j}(x^{0})}
\end{equation}
for any $j\geq1$. Therefore, we get from \eqref{32-3} that $0\leq v\in C^{2m}(\mathbb{R}^{n})$ satisfies
\begin{equation}\label{32-13}
(-\Delta)^{ m}v(x)=v^{p}(x) \quad\quad \text{in} \,\, \mathbb{R}^{n}.
\end{equation}
By the Liouville theorem (Theorem \ref{Thm0}), we must have $v\equiv0$ in $\mathbb{R}^{n}$, which is a contradiction with
\begin{equation}\label{32-14}
v(0)=\lim_{k\rightarrow\infty}v^{(k)}_{k}(0)=1.
\end{equation}
This concludes our proof of Theorem \ref{Thm1}.
\section{Proof of Theorem \ref{Thm2}}
In this section, by applying the a priori estimates (Theorem \ref{Thm-CFL} and Theorem \ref{Thm1}) and the following Leray-Schauder fixed point theorem (see e.g. \cite{CLM,CDQ}), we will prove the existence of positive solutions to the higher order Lane-Emden equations \eqref{Navier} with Navier boundary conditions.
\begin{thm}\label{L-S}
Suppose that $X$ is a real Banach space with a closed positive cone $P$, $U\subset P$ is bounded open and contains $0$. Assume that there exists $\rho>0$ such that $B_{\rho}(0)\cap P\subset U$ and that $T:\,\overline{U}\rightarrow P$ is compact and satisfies
\vskip 5pt
\noindent i) For any $x\in P$ with $|x|=\rho$ and any $\lambda\in[0,1)$, $x\neq\lambda Tx$;
\vskip 3pt
\noindent ii) There exists some $y\in P\setminus\{0\}$ such that $x-Tx\neq ty$ for any $t\geq0$ and $x\in\partial U$.
\vskip 5pt
\noindent Then, $T$ possesses a fixed point in $\overline{U_{\rho}}$, where $U_{\rho}:=U\setminus B_{\rho}(0)$.
\end{thm}
Now we let
\begin{equation}\label{4-1}
X:=C^{0}(\overline{\Omega}) \quad\quad \text{and} \quad\quad P:=\{u\in X \,|\, u\geq0\}.
\end{equation}
Define
\begin{equation}\label{4-2}
T(u)(x):=\int_{\Omega}G_{2}(x,y^{ m})\int_{\Omega}G_{2}(y^{ m},y^{ m-1})\int_{\Omega}\cdots\int_{\Omega}G_{2}(y^{2},y^{1})u^{p}(y^{1})
dy^{1}dy^{2}\cdots dy^{ m},
\end{equation}
where $G_{2}(x,y)$ is the Green's function for $-\Delta$ with Dirichlet boundary condition in $\Omega$. Suppose $u\in C^{0}(\overline{\Omega})$ is a fixed point of $T$, i.e., $u=Tu$, then it is easy to see that $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ and satisfies the Navier problem
\begin{equation}\label{4-3}\\\begin{cases}
(-\Delta)^{ m}u(x)=u^{p}(x) \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \Omega, \\
u(x)=-\Delta u(x)=\cdots=(-\Delta)^{ m-1}u(x)=0 \,\,\,\,\,\,\,\, \text{on} \,\,\, \partial\Omega.
\end{cases}\end{equation}
Our goal is to show the existence of a fixed point for $T$ in $P\setminus B_{\rho}(0)$ for some $\rho>0$ (to be determined later) by using Theorem \ref{L-S}. To this end, we need to verify the two conditions i) and ii) in Theorem \ref{L-S} separately.
\emph{i)} First, we show that there exists $\rho>0$ such that for any $u\in\partial B_{\rho}(0)\cap P$ and $0\leq\lambda<1$,
\begin{equation}\label{4-4}
u-\lambda T(u)\neq0.
\end{equation}
For any $x\in\overline{\Omega}$, it holds that
\begin{eqnarray}\label{4-5}
\nonumber |T(u)(x)|&=&\left|\int_{\Omega}G_{2}(x,y^{ m})\int_{\Omega}G_{2}(y^{ m},y^{ m-1})\cdots\int_{\Omega}G_{2}(y^{2},y^{1})u^{p}(y^{1})
dy^{1}\cdots dy^{ m}\right| \\
&\leq& \int_{\Omega}G_{2}(x,y^{ m})\int_{\Omega}G_{2}(y^{ m},y^{ m-1})\cdots\int_{\Omega}G_{2}(y^{2},y^{1})dy^{1}\cdots dy^{ m}\cdot\|u\|^{p}_{C^{0}(\overline{\Omega})} \\
\nonumber &\leq& \rho^{p-1}\left\|\int_{\Omega}G_{2}(x,y)dy\right\|^{ m}_{C^{0}(\overline{\Omega})}\cdot\|u\|_{C^{0}(\overline{\Omega})}.
\end{eqnarray}
Let $g(x):=\int_{\Omega}G_{2}(x,y)dy$, then it solves
\begin{equation}\label{4-6}\\\begin{cases}
-\Delta_{x}g(x)=1 \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \Omega, \\
g(x)=0 \,\,\,\,\,\,\,\, \text{on} \,\,\, \partial\Omega.
\end{cases}\end{equation}
For a fixed point $x^{0}\in\Omega$, we define the function
\begin{equation}\label{4-7}
\beta(x):=\frac{(diam\,\Omega)^{2}}{2n}\left(1-\frac{|x-x^{0}|^{2}}{(diam\,\Omega)^{2}}\right)_{+},
\end{equation}
then it satisfies
\begin{equation}\label{4-8}\\\begin{cases}
-\Delta_{x}\beta(x)=1 \,\,\,\,\,\,\,\,\,\, \text{in} \,\,\, \Omega, \\
\beta(x)>0 \,\,\,\,\,\,\,\, \text{on} \,\,\, \partial\Omega.
\end{cases}\end{equation}
By maximum principle, we get
\begin{equation}\label{4-9}
0\leq g(x)<\beta(x)\leq\frac{(diam\,\Omega)^{2}}{2n}, \quad\quad \forall \,\, x\in\overline{\Omega}.
\end{equation}
Therefore, we infer from \eqref{4-5} and \eqref{4-9} that
\begin{equation}\label{4-10}
\|T(u)\|_{C^{0}(\overline{\Omega})}<\rho^{p-1}\frac{(diam\,\Omega)^{2m}}{(2n)^{m}}\|u\|_{C^{0}(\overline{\Omega})}=\|u\|_{C^{0}(\overline{\Omega})}
\end{equation}
if we take
\begin{equation}\label{4-11}
\rho=\left(\frac{\sqrt{2n}}{diam\,\Omega}\right)^{\frac{2m}{p-1}}>0.
\end{equation}
This implies that $u\neq\lambda T(u)$ for any $u\in\partial B_{\rho}(0)\cap P$ and $0\leq\lambda<1$.
\emph{ii)} Now, let $\varphi\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ be the unique positive solution of
\begin{equation}\label{4-12}\\\begin{cases}
(-\Delta)^{ m}\varphi(x)=1, \,\,\,\,\,\,\,\,\,\, \,\,\, x\in\Omega, \\
\varphi(x)=-\Delta\varphi(x)=\cdots=(-\Delta)^{m-1}\varphi(x)=0, \,\,\,\,\,\,\,\,\,\,\,\, x\in\partial\Omega.
\end{cases}\end{equation}
We will show that
\begin{equation}\label{4-13}
u-T(u)\neq t\varphi \quad\quad \forall \,\, t\geq0, \quad \forall u\in\partial U,
\end{equation}
where $U:=B_{R}(0)\cap P$ with sufficiently large $R>\rho$ (to be determined later). First, observe that for any $u\in\overline{U}$,
\begin{equation}\label{4-14}
\left\|(-\Delta)^{ m}T(u)\right\|_{C^{0}(\overline{\Omega})}=\|u\|^{p}_{C^{0}(\overline{\Omega})}\leq R^{p},
\end{equation}
and hence
\begin{equation}\label{4-15}
\|T(u)\|_{C^{0,\alpha}(\Omega)}\leq CR^{p} \quad\quad \forall \,\, 0<\alpha<1,
\end{equation}
thus $T:\, \overline{U}\rightarrow P$ is compact.
We use contradiction arguments to prove \eqref{4-13}. Suppose on the contrary that, there exists some $u\in\partial U$ and $t\geq0$ such that
\begin{equation}\label{4-16}
u-T(u)=t\varphi,
\end{equation}
then one has $\|u\|_{C^{0}(\overline{\Omega})}=R>\rho>0$, $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$ and satisfies the Navier problem
\begin{equation}\label{4-17}\\\begin{cases}
(-\Delta)^{ m}u(x)=u^{p}(x)+t, \,\,\,\,\,\,\,\,\, u(x)>0, \,\,\,\,\,\,\,\,\,\, x\in\Omega, \\
u(x)=-\Delta u(x)=\cdots=(-\Delta)^{ m-1}u(x)=0, \,\,\,\,\,\,\,\,\,\,\,\, x\in\partial\Omega.
\end{cases}\end{equation}
Choose a constant $C_{1}>\lambda_{1}$. Since $u(x)>0$ in $\Omega$ and $p>1$, it is easy to see that, there exists another constant $C_{2}>0$ (e.g., take $C_{2}=C_{1}^{\frac{p}{p-1}}$), such that
\begin{equation}\label{4-18}
u^{p}(x)\geq C_{1}u(x)-C_{2}.
\end{equation}
If $t\geq C_{2}$, then we have
\begin{equation}\label{4-19}
(-\Delta)^{ m}u(x)=u^{p}(x)+t\geq C_{1}u(x)-C_{2}+t\geq C_{1}u(x) \quad\quad \text{in} \,\, \Omega.
\end{equation}
Multiplying both side of \eqref{4-19} by the eigenfunction $\phi(x)$, and integrating by parts yield
\begin{eqnarray}\label{4-20}
C_{1}\int_{\Omega}u(x)\phi(x)dx&\leq&\int_{\Omega}(-\Delta)^{ m}u(x)\cdot\phi(x)dx=\int_{\Omega}u(x)\cdot(-\Delta)^{ m}\phi(x)dx \\
\nonumber &=&\lambda_{1}\int_{\Omega}u(x)\phi(x)dx,
\end{eqnarray}
and hence
\begin{equation}\label{4-21}
0<(C_{1}-\lambda_{1})\int_{\Omega}u(x)\phi(x)dx\leq0,
\end{equation}
which is absurd. Thus, we must have $0\leq t<C_{2}$. Next, we carry on our proof by discussing two different assumptions.
If $\frac{n}{n-2m}<p<\frac{n+2m}{n-2m}$, by the a priori estimates (Theorem \ref{Thm-CFL}), we derive that
\begin{equation}\label{4-CFL}
\|u\|_{L^{\infty}(\overline{\Omega})}\leq C(n,m,p,\Omega)=:C'_{0}.
\end{equation}
If $\Omega$ is strictly convex, $1<p<\frac{n+2m}{n-2m}$, or if $1<p\leq\frac{n+2}{n-2}$, by the a priori estimates (Theorem \ref{Thm1}), we know that
\begin{equation}\label{4-22}
\|u\|_{L^{\infty}(\overline{\Omega})}\leq C(n,m,p,t,\lambda_{1},\Omega).
\end{equation}
We will show that the above a priori estimates \eqref{4-22} are uniform with respect to $0\leq t<C_{2}$, i.e., for $0\leq t<C_{2}$,
\begin{equation}\label{4-23}
\|u\|_{L^{\infty}(\overline{\Omega})}\leq C(n,m,p,C_{2},\lambda_{1},\Omega)=:C''_{0}.
\end{equation}
Indeed, it is clear from Theorem \ref{Boundary} that, the thickness $\bar{\delta}$ of the boundary layer and the boundary layer estimates are uniform with respect to $t$. Therefore, if \eqref{4-23} does not hold, there exist sequences $\{t_{k}\}\subset[0,C_{2})$, $\{x^{k}\}\subset\Omega\setminus\overline{\Omega}_{\bar{\delta}}$ and $\{u_{k}\}$ satisfying
\begin{equation}\label{4-24}\\\begin{cases}
(-\Delta)^{ m}u_{k}(x)=u_{k}^{p}(x)+t_{k}, \,\,\,\,\,\,\,\,\,\, \,\,\, x\in\Omega, \\
u_{k}(x)=-\Delta u_{k}(x)=\cdots=(-\Delta)^{ m-1}u_{k}(x)=0, \,\,\,\,\,\,\,\,\,\,\,\, x\in\partial\Omega,
\end{cases}\end{equation}
but $m_{k}:=u_{k}(x^{k})=\|u_{k}\|_{L^{\infty}(\overline{\Omega})}\rightarrow+\infty$ as $k\rightarrow\infty$. For $x\in\Omega_{k}:=\{x\in\mathbb{R}^{n}\,|\,\lambda_{k}x+x^{k}\in\Omega\}$, we define $v_{k}(x):=\frac{1}{m_{k}}u_{k}(\lambda_{k}x+x^{k})$ with $\lambda_{k}:=m^{\frac{1-p}{2m}}_{k}\rightarrow0$ as $k\rightarrow\infty$. Then $v_{k}(x)$ satisfies $\|v_{k}\|_{L^{\infty}(\overline{\Omega_{k}})}=v_{k}(0)=1$ and
\begin{equation}\label{4-25}
(-\Delta)^{ m}v_{k}(x)=v^{p}_{k}(x)+\frac{t_{k}}{m^{p}_{k}}
\end{equation}
for any $x\in\Omega_{k}$. Since $0\leq t<C_{2}$ and $m_{k}\rightarrow+\infty$, by completely similar blowing-up methods as in the proof of Theorem \ref{Thm1} in subsection 3.2, we can also derive a subsequence $\{v^{(k)}_{k}\}\subset\{v_{k}\}$ and a function $v\in C^{2m}(\mathbb{R}^{n})$ such that
\begin{equation}\label{4-26}
v^{(k)}_{k}\rightrightarrows v \quad\quad \text{and} \quad\quad (-\Delta)^{m}v^{(k)}_{k}\rightrightarrows(-\Delta)^{m}v \quad\quad \text{in} \,\, \overline{B_{j}(x^{0})}
\end{equation}
for arbitrary $j\geq1$, and hence $0\leq v\in C^{2m}(\mathbb{R}^{n})$ solves
\begin{equation}\label{4-27}
(-\Delta)^{m}v(x)=v^{p}(x) \quad\quad \text{in} \,\, \mathbb{R}^{n}.
\end{equation}
By Theorem \ref{Thm0}, one immediately has $v\equiv0$, which contradicts with $v(0)=1$. Therefore, the uniform estimates \eqref{4-23} must hold.
Now we let $C_{0}:=\max\{C'_{0},C''_{0}\}>0$ and $R:=C_{0}+\rho$ and $U:=B_{C_{0}+\rho}(0)\cap P$, then \eqref{4-CFL} and \eqref{4-23} implies
\begin{equation}\label{4-28}
\|u\|_{L^{\infty}(\overline{\Omega})}\leq C_{0}<C_{0}+\rho,
\end{equation}
which contradicts with $u\in\partial U$. This implies that
\begin{equation}\label{4-29}
u-T(u)\neq t\varphi
\end{equation}
for any $t\geq0$ and $u\in\partial U$ with $U=B_{C_{0}+\rho}(0)\cap P$.
From Theorem \ref{L-S}, we deduce that there exists a $u\in\overline{\big(B_{C_{0}+\rho}(0)\cap P\big)\setminus B_{\rho}(0)}$ satisfies
\begin{equation}\label{4-30}
u=T(u),
\end{equation}
and hence $\rho\leq\|u\|_{L^{\infty}(\overline{\Omega})}\leq C_{0}+\rho$ solves the higher order Navier problem
\begin{equation}\label{4-31}\\\begin{cases}
(-\Delta)^{ m}u(x)=u^{p}(x), \,\,\,\,\,\,\, u(x)>0, \,\,\,\,\,\,\,\,\, x\in\Omega, \\
u(x)=-\Delta u(x)=\cdots=(-\Delta)^{ m-1}u(x)=0, \,\,\,\,\,\,\,\,\,\,\,\, x\in\partial\Omega.
\end{cases}\end{equation}
By regularity theory, we can see that $u\in C^{2m}(\Omega)\cap C^{2m-2}(\overline{\Omega})$.
This concludes our proof of Theorem \ref{Thm2}.
|
\section*{Acknowledgements}
We thank the reviewers, Samira Abnar, and Willem Zuidema for their useful and constructive feedback.
DH is funded by the Netherlands Organization for Scientific Research (NWO), through a Gravitation Grant 024.001.006 to the Language in Interaction Consortium.
\section{Experimental Setup}\label{sec:data}
Our experimental setup consists of 2 phases: first we extract the relevant sentences and NPI constructions from a corpus, and then, after passing the sentences through an LM, we apply several diagnostic tasks to them.
\subsection{NPI extraction}
For extraction we used the parsed Google Books corpus \cite{googlebooks}.
We focus on the most common NPI pairs, in which the NPI \textit{any} (or any variation thereon) is licensed by a negative operator (\textit{not}, \textit{n\textquotesingle t}, \textit{never}, or \textit{nobody}), as they can reliably be extracted from a parsed corpus.
As variations of \textit{any} we consider \textit{anybody}, \textit{anyone}, \textit{anymore}, \textit{anything}, \textit{anytime}, and \textit{anywhere} (7 in total including \textit{any}).
We first identify candidate NPI-LC relations looking only at the surface form of the sentence, by selecting sentences that contain the appropriate lexical items.
We use this as a pre-filtering step for our second method, in which we extract specific subtrees given the parse tree of the sentence.
We consider 6 different subtrees, that are shown in Table \ref{trees}.
An example of such a subtree that licenses an NPI is the following:
\begin{center}
\begin{tikzpicture}
\Tree
[.\texttt{VP}
[.\texttt{VBD} {did} ]
[.\texttt{RB} {not} ]
[.\texttt{VP} \edge[roof]; {$\cdots$ \textit{any} $\cdots$} ]
]
\end{tikzpicture}
\end{center}
\noindent which could, for instance, be a subtree of the parse tree of \textit{Bill did not buy any books}.
In this subtree, the scope of the licensor \textit{not} encompasses the \texttt{VP} of the sentence.
We use this scope to pinpoint the exact range in which an NPI can reside.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.95\columnwidth, trim=3mm 3mm 3mm 2mm, clip]{lc_npi_dist}
\caption{Distribution of distances between NPI and licensing context. Note the log scale on the y-axis.}
\label{fig:pp_distances}
\end{figure}
Once all NPI constructions have been extracted, we are able to gain more insight in the distance between the licensing operator and an NPI, which we plot in Figure \ref{fig:pp_distances}.
Note the use of a log scale on the y-axis: in the majority of the constructions (47.2\%) the LC and NPI are situated only 2 positions from each other.
\begin{table*}[h]
\begin{center}
\begin{tabular}{|l|r l|}\hline
Construction & \# & (\% / corpus)\\
\hline
All corpus sentences & 11.213.916 & \\
Containing any variation of \textit{any} & 301.836 & (2.69\%)\\
Licensed by negative operator & 123.683 & (1.10\%) \\
Detected by subtree extractor & 112.299 & (1.00\%) \\\hline\hline
1. \texttt{(VP (VP \textit{RB} [VP]))} & 70.017 & \\
\multicolumn{1}{|l|}{ \textit{He did \underline{n't}} [ \textit{have \textbf{any} trouble going along} ] \textit{.}}&&\\[3pt]
2. \texttt{(VP (MD \textit{RB} [VP]))} & 27.698 & \\
\multicolumn{1}{|l|}{ \textit{I could \underline{not}} [ \textit{let \textbf{anything} happen to either of them} ] \textit{.}}&&\\[3pt]
3. \texttt{(VP (VP \textit{RB} [NP/PP/ADJP]))} & 8708 & \\
\multicolumn{1}{|l|}{ \textit{"There was \underline{n't}} [ \textit{\textbf{any} doubt in his mind who was preeminent} ] \textit{."}}&&\\[3pt]
4. \texttt{(VP (NP \textit{RB} [VP]))} & 3564 & \\
\multicolumn{1}{|l|}{ \textit{Those words \underline{never}} [ \textit{lead to \textbf{anything} good} ] \textit{.}}&&\\[3pt]
5. \texttt{(S (\textit{RB} [S/SBAR]))} & 1347 & \\
\multicolumn{1}{|l|}{ \textit{The trick is \underline{not}} [ \textit{to process \textbf{any} of the information I encounter} ] \textit{.}}&&\\[3pt]
6. \texttt{(\textit{RB} [NP/PP ADVP])} & 930 & \\
\multicolumn{1}{|l|}{ \textit{There was \underline{not}} [ \textit{a trace of water \textbf{anywhere}} ] \textit{.}}&&\\\hline
\end{tabular}
\end{center}
\caption{Various sentence constructions and their counts that were extracted from the corpus. Similar verb POS tags are grouped under \texttt{VP}, except for modal verbs (\texttt{MD}). LC scope is denoted by square brackets.}
\label{trees}
\end{table*}
\subsection{Model}
For all our experiments, we use a pretrained 2-layer LSTM language model with 650 hidden units made available by \citet{gulordava2018colorless}.\footnote{\url{github.com/facebookresearch/colorlessgreenRNNs/tree/master/data}}
For all tests we used an average hidden final state as initialization, which is computed by passing all sentences in our corpus to the LM, and averaging the hidden states that are returned at the end of each sentence.
We use two different methods to assess the LSTMs ability to handle NPI constructions, which we will discuss in the next two sections: one that is based on the probabilities that are returned by the LM, and one based on its internal activations.
\section{Conclusion}\label{sec:discussion}
We ran several diagnostic tasks to investigate the ability of a neural language model to handle NPIs.
From the results on the perplexity task we conclude that the model is capable to detect the relationship between an NPI and the licensing contexts that we considered.
We showed that the language model is able to pick up a distinct signal that indicates a strong relationship between a negative polarity item and its licensing context.
By comparing the perplexities of the NPI constructions to those of the equivalent PPIs, it follows that removing the licensing operator has a remarkably different effect on the NPIs than on the PPIs.
This effect, however, does seem to vanish when the distance between the NPI and licensing context is increased.
From our scope detection task it followed that the licensing signal that the LM detects can in fact be extracted from the hidden representations, providing further evidence of the ability of the model in handling NPIs.
There are many other natural language phenomena related to language scope, and we hope that our methods presented here can provide an inspiration for future research, trying to link linguistics theory to neural models.
The setup of our second experiment, for example, would translate easily to the detection of the nuclear scope of quantifiers.
In particular, we believe it would be interesting to look at a wider typological range of NPI constructions, and investigate how our diagnostic tasks translate to other types of such constructions.
Furthermore, the findings of our experiments could be compared to those of human judgments syntactic gap filling task. These judgments could also provide more insight into the grammaticality of the rewritten sentences.
The hypotheses that are described in Section 2 and several others that are mentioned in the literature on NPIs are strongly based on a specific kind of entailment relation that should hold for the contexts in which NPIs reside.
An interesting follow-up experiment that would provide a stronger link with the literature in formal linguistics on the subject matter, would be based on devising several entailment tasks that are based on the various hypotheses that exists for NPI licensing contexts.
It would be interesting to see whether the model is able to detect whether a context is downward entailing, for example, or if it has more difficulty identifying non-veridical contexts.
This would then also create a stronger insight in the semantic information that is stored in the encodings of the model.
Such experiments would, however, require the creation of a rich artificial dataset, which would give much more control in determining the inner workings of the LSTM, and is perhaps a necessary step to gain a thorough insight in the LM encodings from a linguistic perspective.
\section{Sentence Grammaticality}\label{sec:perplexity}
In our first series of experiments, we focus on the probabilities that are assigned by the model to different sequences.
More specifically, we compare the exponent of the normalized negative log probability (also referred to as \textit{perplexity}) of different sentences.
The lower the perplexity score of a sentence is, the better a model was able to predict its tokens.
\subsection{Rewriting sentences}
While studying perplexity scores of individual sentences is not very informative, comparing perplexity scores of similar sentences can provide information about which sentence is preferred by the model.
We exploit this by comparing the negative polarity sentences in our corpus with an ungrammatical counterpart, that is created by removing or rewriting the licensing context.\footnote{\textit{Not} and \textit{never} are removed, \textit{nobody} is rewritten to \textit{everybody}.}
To account for the potential effect of rewriting the sentence, we also consider the sentences that originate from replacing the NPI in the original and rewritten sentence with its positive counterpart.
In other words, we replace the variations of \textit{any} by those of \textit{some}: \textit{anything} becomes \textit{something}, \textit{anywhere} becomes \textit{somewhere}, etc.
We refer to these 4 conditions with the terms \ss{\textbf{NPI}$_{neg}$}, \ss{\textbf{NPI}$_{pos}$}, \ss{\textbf{PPI}$_{neg}$} and \ss{\textbf{PPI}$_{pos}$}:
\begin{center}
\begin{tabular}{l l}
\ss{\textbf{NPI}$_{neg}$}:&\textit{Bill did not buy any books}\\
\ss{\textbf{NPI}$_{pos}$}:&* \textit{Bill did buy any books} \\
\ss{\textbf{PPI}$_{neg}$}:&\# \textit{Bill did not buy some books}\\
\ss{\textbf{PPI}$_{pos}$}:&\textit{Bill did buy some books}
\end{tabular}
\end{center}
\ss{\textbf{PPI}$_{neg}$} would be correct when interpreting \textit{some} as indefinite article (\textit{non-emphatic some}).
In our setup, \ss{\textbf{NPI}$_{neg}$} always refers to the original sentence, as we always use a sentence containing an NPI in a negative context as starting point.
Of the 7 \textit{any} variations, \textit{anymore} is the only one without a PPI counterpart, and these sentences are therefore not considered for this comparison.
\subsection{Comparing sentences}
For all sentences, we compute the perplexity of the original sentence, as well as the perplexity of the 3 rewritten versions of it.
To discard any influence that the removal of the licensing operator might have on its continuation after the occurrence of the NPI, we compute the perplexity of the sentence up to and including the position of the NPI. I.e., in the example of \textit{Bill did not buy any books} the word \textit{books} would not be taken into account when computing the perplexity.
In addition to perplexity, we also consider the conditional probabilities of the PPIs and NPIs, given the preceding sentence.\footnote{We also considered the \textsc{slor} score \cite{pauls2012large}, that was shown in \cite{lau2017grammaticality} to have a strong correlation with human grammaticality judgments.
The \textsc{slor} score can be seen as a perplexity score that is normalized by the average unigram probability of the sentence.
It turned out, however, that this score had such a strong correlation with the perplexity scores (Spearman's $\rho$ of -0.66, Kendall's $\tau$ of -0.54), that we omitted a further analysis of the outcome.
}
For example, for \ss{\textbf{NPI}$_{neg}$} we would then compute \textit{P(any $|$ Bill did not buy)}.
\subsection{Expectations} We posit the following hypotheses about the outcome of the experiments.
\begin{itemize}
\item $PP($\ss{\textbf{NPI}$_{neg}$}$) < PP($\ss{\textbf{NPI}$_{pos}$}$)$:
We expect an NPI construction to have a lower perplexity than the rewritten sentence in which the licensing operator has been removed.
\item $PP($\ss{\textbf{PPI}$_{pos}$}$) < PP($\ss{\textbf{PPI}$_{neg}$}$)$:
Similarly, we expect a PPI to be preferred in the positive counterpart of the sentence, in which no licensing operator occurs.
\item $PP($\ss{\textbf{NPI}$_{neg}$}$) < PP($\ss{\textbf{PPI}$_{neg}$}$)$:
We expect an NPI to be preferred to a PPI inside a negative context.
\item $PP($\ss{\textbf{PPI}$_{pos}$}$) < PP($\ss{\textbf{NPI}$_{pos}$}$)$:
We expect the opposite once the licensor for this context has been removed.
\end{itemize}
\subsection{Results}
In Figure \ref{fig:pp_hist}, we plot the distribution of the perplexity scores for each sentence type.
The perplexities of the original and rewritten sentence without the NPI are indicated by \ss{\textbf{SEN}$_{neg}$} and \ss{\textbf{SEN}$_{pos}$}, respectively.
This figure shows that the original sentences have the lowest perplexity, whereas the NPIs in a positive context are deemed most improbable by the model.
\begin{figure}[]
\centering
\includegraphics[width=0.95\columnwidth, trim=3mm 2mm 3mm 3mm, clip]{pp_hist}
\caption{Distribution of perplexity scores for all the sentences.}
\label{fig:pp_hist}
\end{figure}
More insightful we consider Figure \ref{perps}, in which we plot the distribution of the relative differences of the perplexity scores and conditional probabilities for each of the above mentioned comparisons, and we report the percentage of sentences that complied with our hypotheses.
The relative difference between two values $a$ and $b$, given by $(a-b)/((a+b)/2)$, neatly maps each value pair in a window between -2 $(a\ll b)$ and 2 $(a\gg b)$, thereby providing a better insight in the difference between two arrays of scores.
We highlight some of the previously mentioned comparisons below.
\begin{figure*}[h!]
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\columnwidth]{pp_dists}\\
\begin{tabular}{|r|c|c|c|}\hline
\multicolumn{4}{|c|}{\textsc{Perplexity}}\\\hline
$\mathbf{<}$ & \ss{NPI$_{pos}$} & \ss{PPI$_{neg}$} & \ss{PPI$_{pos}$}\\\hline
\ss{NPI$_{neg}$} & 99.2\% & 88.7\% & 95.8\%\\
\ss{NPI$_{pos}$} & - & 3.6\% & 17.3\% \\
\ss{PPI$_{neg}$} & - & - & 91.0\%
\\\hline
\end{tabular}
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\columnwidth]{p_dists}
\begin{tabular}{|r|c|c|c|}\hline
\multicolumn{4}{|c|}{$P(w|c)$}\\\hline
$\mathbf{>}$ & \ss{NPI$_{pos}$} & \ss{PPI$_{neg}$} & \ss{PPI$_{pos}$}\\\hline
\ss{NPI$_{neg}$} & 99.3\% & 94.8\% & 93.4\%\\
\ss{NPI$_{pos}$} & - & 34.0\% & 19.1\% \\
\ss{PPI$_{neg}$} & - & - & 30.1\%
\\\hline
\end{tabular}
\end{minipage}
\caption{Results of perplexity and conditional probability tests. For perplexity a lower score is better, for probability a higher score is better. The plots denote the distribution of the relative differences between the scores of the 6 sentence pairs that are considered.}
\label{perps}
\end{figure*}
\paragraph{$PP($\ss{NPI$_{neg}$}$) < PP($\ss{NPI$_{pos}$}$)$}
From Figure \ref{perps} it is clear that the model has a very strong preference for NPIs to reside inside the negative scope, an observation that is supported by both the perplexity and probability scores.
While observable in both plots, this preference is most clearly visible when considering conditional probabilities: the high peak shows that the difference between the probabilities is the most defined of all comparisons that we made.
\paragraph{$PP($\ss{NPI$_{neg}$}$) < PP($\ss{PPI$_{neg}$}$)$}
The model has a strong preference for NPIs over PPIs inside negative scope, although this effect is slightly less prevalent in the perplexity scores. This might be partly due to the fact that there exist interpretations for \textit{some} inside negative scope that are correct (the non-emphatic \textit{some}, as described in Section \ref{sec:npi}).
When looking solely at the conditional probabilities the preference becomes clearer, showing similar behavior to the difference between \ss{\textbf{NPI}$_{neg}$} and \ss{\textbf{NPI}$_{pos}$}.
\paragraph{$PP($\ss{NPI$_{neg}$}$) < PP($\ss{PPI$_{pos}$}$)$}
The original sentences with NPIs are strongly preferred over the rewritten sentences with PPIs, which indicates that the rewriting in general leads to less probable sentences.
This finding is confirmed by comparing the perplexities of the original and rewritten sentence \textit{without} the NPI or PPI (dotted line in the left plot in Figure \ref{perps}): the original sentence containing the licensing context has a lower perplexity than the rewritten sentence in 92.7\% of the cases.
The profile of the differences between the 2 sentences is somewhat similar to the other comparisons in which the negative context is preferred.
Given that the considered sentences were taken from natural data, it is not entirely unsurprising that removing or rewriting a scope operator has a negative impact on the probability of the rest of the sentence.
This observation, however, does urge care when running experiments like this.
\paragraph{$PP($\ss{PPI$_{pos}$}$) < PP($\ss{NPI$_{pos}$}$)$}
When comparing NPIs and PPIs in the rewritten sentences, it turns out that the model does show a clear preference that is not entirely due to a less probable rewriting step.
Both the perplexity (17.3\%) and probability (19.1\%) show that the NPI did in fact strongly depend on the presence of the licensing operator, and not on other words that it was surrounded with.
The model is thus able to pick up a signal that makes it prefer a PPI to an NPI in a positive context, even if that positive context was obtained by rewriting it from a negative context.
\paragraph{$PP($\ss{PPI$_{neg}$}$) < PP($\ss{NPI$_{pos}$}$)$}
PPIs in a negative context are strongly preferred to NPIs in a faulty positive context: a lower perplexity was assigned to \ss{\textbf{NPI}$_{pos}$} in only 3.6\% of the cases.
This shows that the model is less strict on the allowed context for PPIs, which might be related to the non-emphatic variant of \textit{some}, as mentioned before.
\paragraph{$PP($\ss{PPI$_{neg}$}$) < PP($\ss{PPI$_{pos}$}$)$}
A surprising result is the higher perplexity that is assigned to PPIs inside the original negative context compared to PPIs in the rewritten sentence, which is opposite to what we hypothesized. It is especially remarkable considering the fact that the conditional probability indicates an opposite result (at only 30.1\% preference for the original sentence). Once more the outcome of the perplexity comparison might partly be due to the rewriting resulting in a less probable sentence. When solely looking at the conditional probability score, however, we can conclude that the model has a preference for PPIs to reside in positive contexts.
\begin{figure*}[h!]
\centering
\includegraphics[width=0.8\textwidth]{p_split}\\
\caption{Outcomes for the conditional probability task, split out on the distance between licensing context and NPI. The averages that are reported in Figure \ref{perps} are denoted by the dotted lines.}
\label{fig:longdist}
\end{figure*}
\paragraph{Long distances}
As shown in Figure \ref{fig:pp_distances}, most distances between the LC and the NPI are rather short. It might therefore be useful to look at the performance of the model on sentences that contain longer distance dependencies. In Figure \ref{fig:longdist} the outcomes of the conditional probability task are split out on the distance between the LC and the NPI.
From this plot it follows that the shorter dependencies were mostly responsible for the outcome of our hypotheses. The significant differences between the original sentence and the rewritten sentences \ss{\textbf{NPI}$_{pos}$} and \ss{\textbf{PPI}$_{neg}$} becomes less defined when the distance is increased.
This might be partly due to the lower occurrence of these constructions: 47.2\% of the sentences in our corpus are situated only 2 positions from each other. Moreover, it would be interesting to see how this behavior matches with that of human judgments.
\paragraph{Conclusion}
We conclude that the LM is able to detect a signal that indicates a strong relationship between an NPI and its licensing context. By comparing the scores between equivalent sentence constructions we were able to account for possible biases of the model, and showed that the output of the model complied with our own hypotheses in almost all cases.
\section{Scope detection} \label{sec:scope}
In the previous section, we assessed the ability of a neural LM to handle NPI constructions, based on the probabilities returned by the LM.
In the current section, we focus on the hidden states that the LM uses to arrive at a probability distribution over the vocabulary.
In particular, we focus on the \textit{scope} of the licensing operator, which determines where an NPI can occur.
\subsubsection*{Setup}
Using the parse tree extraction method described in Section 3, we annotate all sentences in our corpus with the scope of the licensing operator.
Following \citet{hupkes2018visualisation}, we then train \textit{diagnostic classifiers} to predict for each word in the sentence whether it is inside the licensing scope.
This is done on the basis of the hidden representation of the LM that is obtained after it just processed this word.
We differentiate between 5 different labels: pre-licensing scope words (1), the licensing operator (2), words inside the scope (3), the NPI itself (4), and post-licensing scope words (5). The sentence \textit{The man that died didn't have any relatives, but he died peacefully.}, for example, is annotated as follows:\\
\textit{The$_1$ man$_1$ that$_1$ died$_1$ did$_1$ n't$_2$ have$_3$ any$_4$ relatives$_3$ ,$_5$ but$_5$ he$_5$ died$_5$ peacefully$_5$ .$_5$}\\
The main positions of interest are the transition from within the licensing scope to the post-scope range, and the actual classification of the NPI and LC.
Of lesser interest are the pre- and post-licensing scope, as these are both diverse embeddings that do not depend directly on the licensing context itself.
We train our model on the intermediate hidden states of the final layer of the LSTM, using a logistic regression classifier.
The decoder of the LM computes the probability distribution over the vocabulary by a linear projection layer from the final hidden state. By using a linear model for classification (such as logistic regression) we can investigate the expressiveness of the hidden state: if the linear model is able to fulfill a classification task, it could be done by the linear decoding layer too.
As a baseline test, we also train a logistic regression model on representations that were acquired by an additive model using GloVe word embeddings \cite{pennington2014glove}.
Using these embeddings as a baseline we are able to determine the importance of the language model: if it turns out that the LM does not outperform a simple additive model, this indicates that the LM did not add much syntactic information to the word embeddings themselves (or that no syntactic information is required to solve this task).
We used 300-dimensional word embeddings that were trained on the English Wikipedia corpus (as is our own LM).
For both tasks (LM and GloVe) we use a subset of 32k NPI sentences which resulted in a total of 250k data points.
We use a split of 90\% of the data for training, and the other 10\% for testing classification accuracy.
\subsubsection*{Results}
The classifier trained on the hidden states of the LM achieved an accuracy of \textbf{89.7\%} on the test set. The model that was trained on the same dataset using the GloVe baseline scored \textbf{72.5\%}, showing that the information that is encoded by the LM does in fact contribute significantly to this task.
To provide a more qualitative insight into the power of this classifier, we provide 3 remarkable sentences that were classified accurately by the model. Note the correct transition from licensing scope to post-scope, and the correct classification of the NPI and LC in all sentences here.
\begin{enumerate}
\item I$_1$ 'd$_1$ \underline{never$_2$} seen$_3$ \textbf{anything}$_4$ like$_3$ it$_3$ and$_5$ it$_5$ ...$_5$ was$_5$ ...$_5$ beautiful$_5$ .$_5$
\item ``$_1$ I$_1$ do$_1$ \underline{n't$_2$} think$_3$ I$_3$ 'm$_3$ going$_3$ to$_3$ come$_3$ to$_3$ you$_3$ for$_3$ reassurance$_3$ \textbf{anymore}$_4$ ,$_5$ ''$_5$ Sibyl$_5$ grumbled$_5$ .$_5$
\item But$_1$ when$_1$ it$_1$ comes$_1$ to$_1$ you$_1$ ,$_1$ I$_1$ 'm$_1$ \underline{not$_2$} taking$_3$ \textbf{any}$_4$ more$_3$ risks$_3$ than$_3$ we$_3$ have$_3$ to$_3$ .$_5$
\end{enumerate}
We ran a small evaluation on a set of 3000 sentences (47020 tokens), of which 56.8\% were classified completely correctly. Using the GloVe classifier only 22.1\% of the sentences are classified flawlessly.
We describe the classification results in the confusion matrices that are displayed in Figure \ref{conf}.
\begin{figure*}[h]
\begin{tabular}{|r|ccccc|}
\multicolumn{6}{c}{\textit{LSTM Embeddings}}\\\hline
& \multicolumn{5}{c|}{Correct label} \\
Pred. & 1 & 2 & 3 & 4 & 5 \\\hline
1 & \textbf{14891} & 83 & 408 & 2 & 760 \\
2 & 203 & \textbf{2870} & 42 & 0 & 59 \\
3 & 850 & 42 & \textbf{14555} & 15 & 1286 \\
4 & 13 & 1 & 32 & \textbf{3005} & 44 \\
5 & 520 & 11 & 821 & 0 & \textbf{6507} \\\hline
Total & 16477 & 3007 & 15858 & 3022 & 8656 \\\hline
\end{tabular}
\quad
\begin{tabular}{|r|ccccc|}
\multicolumn{6}{c}{\textit{GloVe embeddings}}\\\hline
& \multicolumn{5}{c|}{Correct label} \\
Pred. & 1 & 2 & 3 & 4 & 5 \\\hline
1 & \textbf{11166} & 87 & 1077 & 0 & 249 \\
2 & 178 & \textbf{1847} & 82 & 0 & 0 \\
3 & 4708 & 1072 & \textbf{14166} & 353 & 4003 \\
4 & 17 & 0 & 84 & \textbf{2669} & 36 \\
5 & 408 & 1 & 449 & 0 & \textbf{4368} \\\hline
Total & 16477 & 3007 & 15858 & 3022 & 8656 \\\hline
\end{tabular}
\caption{Confusion matrices for the scope detection task trained on the embeddings of an LSTM and the averages of GloVe embeddings.}
\label{conf}
\end{figure*}
Looking at the results on the LSTM embeddings, it appears that the post-licensing scope tokens (5) were misclassified most frequently: only 75.2\% of those data points were classified correctly.
The most common misclassification for this class is class 3: an item inside the licensing scope.
This shows that for some sentences it is hard to distinguish the actual border of the licensing scope, although 90.3\% of the first post-scope embeddings (i.e.\ the first embedding after the scope has ended) were classified correctly.
The lower performance of the model on this class is mostly due to longer sentences in which a large part of the post-licensing scope was classified incorrectly.
This causes the model to pick up a noisy signal that trips up the predictions for these tokens.
It is promising, however, that the NPIs (4) and licensing operator items (2) themselves are classified with a very high accuracy, as well as the tokens inside the licensing scope (3).
When comparing this to the performance on the GloVe embeddings, it turns out that that classifier has a strong bias towards the licensing scope class (3).
This highlights the power of the LSTM embeddings, revealing that is not a trivial task at all to correctly classify the boundaries of the context scope.
We therefore conclude that the information that is relevant to NPI constructions can be accurately extracted from the sentence representations, and furthermore that our neural LM has a significant positive influence on encoding that structural information.
\section{Introduction}
In the past decade, we have seen a surge in the development of neural language models (LMs).
As they are more capable of detecting long distance dependencies than traditional n-gram models, they serve as a stronger model for natural language.
However, it is unclear what kind of properties of language these models encode.
This does not only hinder further progress in the development of new models, but also prevents us from using models as explanatory models and relating them to formal linguistic knowledge of natural language, an aspect we are particularly interested in in the current paper.
Recently, there has been an increasing interest in investigating what kind of linguistic information is represented by neural models, \citep[see, e.g.,][]{conneau2018you,DBLP:journals/corr/LinzenDG16,tran2018importance}, with a strong focus on their \textit{syntactic} abilities.
In particular, \cite{gulordava2018colorless} used the ability of neural LMs to detect noun-verb congruence pairs as a proxy for their awareness of syntactic structure, yielding promising results.
In this paper, we follow up on this research by studying a phenomenon that has received much attention by linguists and for which the model requires -- besides knowledge of syntactic structure -- also a \textit{semantic} understanding of the sentence: negative polarity items (NPIs).
In short, NPIs are a class of words that bear the special feature that they need to be \textit{licensed} by a specific licensing context (LC) (a more elaborate linguistic account of NPIs can be found in the next section).
A common example of an NPI and LC in English are \textit{any} and \textit{not}, respectively: The sentence \textit{He didn't buy any books} is correct, whereas \textit{He did buy any books} is not.
To properly process an NPI construction, a language model must be able to detect a relationship between a licensing context and an NPI.
Following \citet{DBLP:journals/corr/LinzenDG16} and \citet{gulordava2018colorless}, we devise several tasks to assess whether neural LMs (focusing in particular on LSTMs) can handle NPI constructions, and obtain initial positive results.
Additionally, we use diagnostic classifiers \cite{hupkes2018visualisation} to increase our insight in how NPIs are processed by neural LMs, where we look in particular at their understanding of the \textit{scope} of an LCs, an aspect which is also relevant for many other natural language related phenomena.
We obtain positive results focusing on a subset of NPIs that is easily extractable from a parsed corpus but also argue that a more extensive investigation is needed to get a complete view on how NPIs -- whose distribution is highly diverse -- are processed by neural LMs.
With this research and the methods presented in this paper, we hope to pave the way for other studies linking neural language models to linguistic theory.
In the next section, we will first briefly discuss NPIs from a linguistic perspective.
Then, in Section \ref{sec:data}, we provide the setup of our experiments and describe how we extracted NPI sentences from a parsed corpus.
In Section \ref{sec:perplexity}, we describe the setup and results of an experiment in which we compare the grammaticality of NPI sentences with and without a licensing context, using the probabilities assigned by the LM.
Our second experiment is outlined in Section \ref{sec:scope}, in which we describe a method for scope detection on the basis of the intermediate sentence embeddings.
We conclude our findings in Section \ref{sec:discussion}.
\section{Negative Polarity Items}\label{sec:npi}
NPIs are a complex yet very common linguistic phenomenon, reported to be found in at least 40 different languages \cite{haspelmath1997indefinite}.
The complexity of NPIs lies mostly in the highly idiosyncratic nature of the different types of items and licensing contexts.
Commonly, NPIs occur in contexts that are related to negation and modalities, but they can also appear in imperatives, questions and other types of contexts and sentences.
This broad range of context types makes it challenging to find a common feature of these contexts, and no overarching theory that describes when NPIs can or cannot occur yet exists \cite{Barker2018}.
In this section, we provide a brief overview of several hypotheses about the different contexts in which NPIs can occur, as well as examples that illustrate that none of these theories are complete in their own regard.
An extensive description of these theories can be found in \citet{giannakidou2008negative}, \citet{hoeksema2008natural}, and \citet{Barker2018}, from which most of the example sentences were taken.
These sentences are also collected in Table \ref{npis}.
\paragraph{Entailment}
A downward entailing context is a context that licenses entailment to a subset of the initial clause.
For example, \textit{Every} is downward entailing, as \textit{Every} [ \textit{student} ] \textit{left} entails that \textit{Every} \mbox{[ \textit{tall student} ]} \textit{left}.
\citet{ladusaw1980polarity} hypothesize that NPIs are licensed by downward entailing contexts.
Rewriting the previous example to \textit{Every} \mbox{[\textit{ student with any sense} ]} \textit{left} yields a valid expression, contrary to the same sentence with the upward entailing context \textit{some}: \textit{Some} \mbox{[\textit{student with any sense} ]} \textit{left}.
An example of a non-downward entailing context that is a valid NPI licensor is \textit{most}.
\paragraph{Non-veridicality}
A context is non-veridical when the truth value of a proposition (\textit{veridicality}) that occurs inside its scope cannot be inferred.
An example is the word \textit{doubt}: the sentence \textit{Ann doubts that Bill ate some fish} does not entail \textit{Bill ate some fish}.
\citet{giannakidou1994semantic} hypothesizes that NPIs are licensed only in non-veridical contexts, which correctly predicts that \textit{doubt} is a valid licensing context: \textit{Ann doubts that Bill ate any fish}.
A counterexample to this hypothesis is the context that is raised by the veridical operator \textit{only}: \textit{Only Bob ate fish} entails \textit{Bob ate fish}, but also licenses \textit{Only Bob ate any fish} \cite{Barker2018}.
\subsection{Related constructions}
Two grammatical constructions that are closely related to NPIs are Free Choice Items (FCIs) and Positive Polarity Items (PPIs).
\paragraph{Free Choice Items} FCIs inhibit a property called \textit{freedom of choice} \cite{vendler1967linguistics}, and are licensed in contexts of generic or habitual sentences and modal verbs. An example of such a construction is the generic sentence \textit{Any cat hunts mice}, in which \textit{any} is an FCI. Note that \textit{any} in this case is not licensed by negation, modality, or any of the other licensing contexts for NPIs. English is one of several languages in which a word can be both an FCI and NPI, such as the most common example \textit{any}. Although this research does not focus on FCIs, it is important to note that the somewhat similar distributions of NPIs and FCIs can severely complicate the diagnosis whether we are dealing with an NPI or an FCI.
\paragraph{Positive Polarity Items}
PPIs are a class of words that are thought to bear the property of scoping above negation \cite{giannakidou2008negative}. Similar to NPIs their contexts are highly idiosyncratic, and the exact nature of their distribution is hard to define. PPIs need to be situated in a veridical (often affirmative) context, and can therefore be considered a counterpart to the class of NPIs. A common example of a PPI is \textit{some}, and the variations thereon. It is shown by \citet{giannakidou2008negative} that there exist multiple interpretations of \textit{some}, influenced by its intonation. The emphatic variant is considered to be a PPI that scopes above negation, while the non-emphatic \textit{some} is interpreted as a regular indefinite article (such as \textit{a}).
|
\section{introduction}
The $B$-meson decays are important for precision test of standard model (SM) and for seeking of new physics beyond the SM. Within the framework of SM, they can be used to fix the masses and couplings of the basic particles, research the CP-violation phenomena, determine more precise values for the Cabibbo-Kobayashi-Maskawa (CKM) matrix elements, and etc, cf. Refs.\cite{Kennedy:1990ib, Shifman:1978bx, Cabibbo:1963, Kobayashi:1973, Buras:1983ap, Chetyrkin:1996vx, Ball:1998kk, Deshpande:1997rr}.
For the $B$-meson decays, one has to deal with the one-particle, the two-particle, and the three or more particle matrix elements. Those hadronic matrix elements are key components for extracting useful information on the underlying flavor transitions and studying the decay constants, the transition form factors (TFFs), the mixings and decay amplitudes. The $\gamma$-structures of those non-perturbative hadronic matrix elements can be decomposed into Lorentz-invariant structures by using covariant decomposition, leading to basic TFFs for various decay channels.
\begin{table}[htb]
\begin{center}
\begin{tabular}{ccl}
\hline
~~Matrix element~~ & ~~TFFs~~ & ~~~~~~HFFs~~ \\
\hline
$\begin{array}{c}\langle V|\bar{q}\gamma^\mu b|B\rangle
\\ \langle V|\bar{q}\gamma^\mu\gamma^5 b|B\rangle\end{array}$ &
$\begin{array}{c}V\\ A_0, A_1, A_2\end{array}$ &
$\left\}\begin{array}{c} \mathcal{H}_{\mathcal{V},0},~\mathcal{H}_{\mathcal{V},t} \\
\mathcal{H}_{\mathcal{V},1},~\mathcal{H}_{\mathcal{V},2}\end{array}\right.$ \\[5mm]
$\begin{array}{c}\langle V|\bar{q}\sigma^{\mu\nu}q_\nu b|B\rangle \\
\langle V|\bar{q}\sigma^{\mu\nu}\gamma^5 q_\nu b|B\rangle\end{array}$ &
$\begin{array}{c}T_1\\T_2, T_3\end{array}$ &
$\left\}\begin{array}{c}\mathcal{H}_{\mathcal{T},0} \\
\mathcal{H}_{\mathcal{T},1},~\mathcal{H}_{\mathcal{T},2}\end{array}\right.$ \\ \hline
\end{tabular}
\caption{The relations among the $B\to$ vector meson transition form factors (TFFs), the helicity form factors (HFFs), and the hadronic matrix elements.}
\label{tab:formfact}
\end{center}
\end{table}
Specifically, for the $B \to$ light vector meson decays, we need to deal with seven TFFs for the hadronic matrix elements~\cite{Fu:2014uea, Bharucha:2010im}, which are shown in Table \ref{tab:formfact}. For convenience, we also present the relations among the $B\to$ vector meson helicity form factors (HFFs) and the hadronic matrix elements in Table~\ref{tab:formfact}.
The $B\to$ light vector meson decays have been analyzed by various experimental groups, such as the BaBar collaboration~\cite{Lees:2012tva, delAmoSanchez:2010af}, the Belle collaboration~\cite{Sibidanov:2013rkk}, the LHCb collaboration~\cite{Aaij:2013qta, Aaij:2012cq}, the ATLAS collaboration~\cite{ATLAS:2013ola}, the CLEO collaboration~\cite{Behrens:1999vv}. On the other hand, the TFFs/HFFs for the $B \to$ light vector meson decays have been calculated under various approaches, such as the light-cone sum rules (LCSR)~\cite{svz, Ball:1997rj, Huang:1998gp, Ball:2004rg, Huang:2008zg, AKhodjamirian:2010, Fu:2014pba, Fu:2014cna, Cheng:2017bzz, Ahmady:2013cga, Straub:2015ica}, the lattice QCD (LQCD)~\cite{Lattice96:1, Lattice96:2, Lattice98, DelDebbio:1997nu, Lattice04, Horgan:2013hoa, Horgan:2013pva, Agadjanov:2016fbd}, the perturbative QCD (pQCD)~\cite{Kurimoto:2001zj, Chen:2002bq, Kurimoto:2002sb, Keum:2004is, Fan:2013qz}, or some Phenomenological model~\cite{Cheng:2017smj, Cheng:2017sfk}. Those approaches are complementary to each other, which are applicable for different $q^2$-region. The pQCD approach is valid in low $q^2$-region, the LCSR is applicable in small and intermediate $q^2$-region around $m_b^2-2m_b\chi$ ($\chi\sim 500$ MeV is the typical hadronic scale of the decay) and the LQCD is applicable in high $q^2$-region. Among them, the LCSR prediction can be extrapolated to whole $q^2$-region, thus providing an important bridge for connecting various approaches.
There are large differences for the predicted and measured $B\to\rho$ decay widths at the large $q^2$-region, c.f. Refs.\cite{delAmoSanchez:2010af, Ahmady:2013cga, DelDebbio:1997nu}. In the paper, we shall adopt the LCSR approach to recalculate the $B\to\rho$ hadronic matrix elements. In different to previous LCSR treatment~\cite{Fu:2014cna, Fu:2014pba}, we shall express the hadronic matrix elements by using the HFF with the help of the covariant helicity projection approach~\cite{Korner:1989qb}. The HFFs are also Lorentz-invariant functions which can be formally expressed as the linear combination of the usually adopted TFFs.
\begin{table}[htb]
\begin{center}
\begin{tabular}{c c c c c}
\hline
~Transition~& ~$J^P$~ & ~Mass (GeV) ~ & ~~~~HFFs~~ \\
\hline
& $0^-$ & 5.28 & $\mathcal{H}_{\mathcal{V},t}$ \\
$b\to d$ & $1^-$ & 5.33 & $\mathcal{H}_{\mathcal{V},1}$ \\
& $1^+$ & 5.72 & $\mathcal{H}_{\mathcal{V},0}$,~$\mathcal{H}_{\mathcal{V},2}$\\
\hline
\end{tabular}
\caption{The masses of low-lying $B_d$ resonances~\cite{Bharucha:2010im} and their relations to the HFFs, which are obtained by relating the dominant poles in the LCSRs to those low-lying resonances.}
\label{tab:reson}
\end{center}
\end{table}
There are some advantages for the use of HFF~\cite{Bharucha:2010im}: I) Dispersive bounds on the HFF parameterization can be achieved via the diagonalizable unitarity relations; II) There are relations between the HFFs and the spin-parity quantum numbers, especially when taking the heavy-quark and/or large-energy limit. Thus, they can be conveniently adopted for considering the contributions from the excited states. The relations among the HFFs and the low-lying states can be obtained by relating the dominant poles in the LCSRs to those low-lying resonances. We present the masses of low-lying $B_d$ resonances with explicit quantum numbers $J^P$ in Table~\ref{tab:reson}, which shall be used in our numerical calculations; III) The LCSRs for the $B\to V$ HFFs can be conveniently used for studying the polarized decay widths.
The remaining parts of the paper are organized as follows. In Sec.II, we give the calculation technology for the $B \to \rho$ HFFs within the LCSR approach. In Sec.III, we present the numerical results. By extrapolating those HFFs to the whole $q^2$-region, we study the properties of the $B$-meson semileptonic decay $B\to\rho\ell\nu_\ell$. Sec.IV is reserved for a summary.
\section{Calculation technology for the $B\to\rho$ HFFs}
As for the $B\to\rho\ell\nu_\ell$ semileptonic decays, we need to deal with the hadronic matrix element:
\begin{equation}
\sum\limits_{\alpha=0,\pm,t}\langle \rho(k,\varepsilon_\alpha (k))|\bar q \, \gamma_\mu(1-\gamma^5) \, b |B(p)\rangle. \label{HME}
\end{equation}
where $k=(k^0,0,0,|\vec{k}|)$, $\varepsilon_\alpha(k)$ are $\rho$-meson longitudinal ($\alpha=0$) and transverse ($\pm$) polarization vectors. In the $B$-meson rest frame with the $z$ axis along the $\rho$-meson moving direction, and we have
\begin{eqnarray}
\varepsilon _0(k) &=& \frac{1}{m_{\rho}} (|\vec{k}| ,0,0,k^0), \\
\varepsilon _\pm(k) &=& \mp \frac{1}{{\sqrt 2 }} (0,1, \mp i,0),
\end{eqnarray}
where $|\vec{k}|=\sqrt{\lambda}/2m_B$, $k^0 = {(M_B^2 + m_\rho ^2-q^2)}/{2m_\rho}$ with $q=p-k$, $\lambda = (t_{-} - q^2)(t_{+} - q^2)$ with $t_\pm=(m_B\pm m_\rho)^2$. The polarization vectors satisfy $k\cdot\varepsilon_\alpha(k)=0$.
As proposed by Ref.\cite{Korner:1989qb}, one can adopt the covariant helicity projection approach to study those hadronic matrix element (\ref{HME}). The off-shell $W$-boson has similar polarization vectors as those of $\rho$-meson, e.g. the off-shell $W$-boson with momentum $q=(q^0,0,0,-|\vec q\,|)$ are
\begin{eqnarray}
\varepsilon _0(q) &=& \frac{1}{{\sqrt {q^2 } }} \, (|\vec q| ,0,0, -q^0 ), \\
\varepsilon _\pm(q) &=& \mp \frac{1}{{\sqrt 2 }} \, (0,1, \mp i,0), \\
\varepsilon _t(q) &=& \frac{1}{{\sqrt {q^2 } }} \, q,
\end{eqnarray}
where $|\vec q|=|\vec k|$, $q^0 = {(M_B^2 - m_\rho ^2 + q^2)}/{2m_\rho}$, and the extra vector $\varepsilon _t(q)$ is the time-like polarization vector. The linear combinations of the transverse helicity projection vector $\varepsilon _ {\pm} (q)$ give
\begin{eqnarray}
\varepsilon _1 (q) &=& \frac{{\varepsilon _ - (q) - \varepsilon _ + (q)}}{{\sqrt 2 }} = (0,1,0,0), \\
\varepsilon _2 (q) &=& \frac{{\varepsilon _- (q) + \varepsilon _ +(q)}}{{\sqrt 2 }} = (0,0,i,0).
\end{eqnarray}
Using the off-shell $W$-boson polarization vectors, one can project out the relevant HFFs from the hadronic matrix elements~\cite{Bharucha:2010im}
\begin{eqnarray}
\mathcal{H}_{\rho,\sigma}(q^2 ) &=& \sqrt{\frac{q^2}{\lambda}} \, \sum\limits_{\alpha=0,\pm,t}
{\varepsilon_\sigma^{*\mu}(q)} \times \nonumber\\
&& \langle \rho(k,\varepsilon_\alpha(k)) |\bar q \, \gamma_\mu(1-\gamma^5) \, b |B(p)\rangle,
\label{HFF:Definition}
\end{eqnarray}
where $q=p-k$. In the following, we shall not consider the time-like HFF ($t$), which can be treated by using the same way and has no contribution to semileptonic decay width due to chiral suppression.
Following the standard LCSR procedures~\cite{Ball:1998kk, Ball:2004rg, Khodjamirian:2006st}, we can derive the LCSRs for the $B\to\rho$ HFFs. We first define a two-point correlation function as
\begin{eqnarray}
\Pi_{\sigma}(p,q) &=& -i \sqrt{\frac{q^2}{\lambda}} \, \sum\limits_{\alpha=0,\pm,t} {\varepsilon_\sigma^{*\mu}(q)}\int d^4 x e^{iq\cdot x} \nonumber \\
&&\times \langle \rho(k,\varepsilon_\alpha(k))|T\{j_{V-A,\mu}(x),j_B^\dag (0)\}|0\rangle , \label{correlators}
\end{eqnarray}
where the currents $j_{V-A,\mu}(x) = \bar d (x){\gamma _\mu }(1 - {\gamma _5})b(x)$ and $j_B^\dag (0)=i m_b \bar b(0) \gamma_5 q(0)$ which has the same quantum state of the $B$-meson with $J^{P}=0^-$, and $\sigma=(0,1,2)$.
In the time-like $q^2$-region, one can insert a complete series of the intermediate hadronic states in the correlator (\ref{correlators}) and single out the pole term of the $B$-meson lowest pseudoscalar,
\begin{widetext}
\begin{eqnarray}
\Pi _{\sigma}^{\rm H }&=&\sqrt{\frac{q^2}{\lambda}} \, \sum\limits_{\alpha=0,\pm,t}
{\varepsilon_\sigma^{*\mu}(q)}\frac{\langle \rho(k,\varepsilon_\alpha(k))
|\bar q \, \gamma_\mu(1-\gamma^5) \, b |B\rangle \langle B|\bar b i \gamma_5 q|0\rangle }{m_b[m_B^2 - (p + q)^2]}
\nonumber \\
&& + \sqrt{\frac{q^2}{\lambda}} \, \sum\limits_{\alpha=0,\pm,t} \sum\limits_{\rm H}
{\varepsilon_\sigma^{*\mu}(q)} \frac{\langle \rho(k,\varepsilon_\alpha(k))
|\bar q \, \gamma_\mu(1-\gamma^5) \, b |B^H\rangle \langle B^H|\bar b i \gamma _5 q |0\rangle }{m_b[m_{B^H}^2 - (p + q)^2]},
\end{eqnarray}
\end{widetext}
where $\langle B|\bar b i \gamma_5 q|0\rangle={m_B^2 f_B}/{m_b}$ with $f_B$ being the $B$-meson decay constant. By replacing the contributions from the higher-level resonances and continuum states with the dispersion relations, the invariant amplitudes can be rewritten as
\begin{eqnarray}
\Pi _{\sigma}^{\rm H } &=& \frac{m_B^2 f_B}{m_b[m_B^2 - (p + q)^2]} \mathcal{H}_{\rho,\sigma}(q^2) \nonumber\\
&& +\int_{s_0}^\infty \frac{\rho_\sigma^{\rm H}}{s - (p + q)^2}ds + \cdots,
\end{eqnarray}
where $s_0$ stands for the continuum threshold parameter and the ellipsis is the subtraction constant or the finite $q^2$-polynomial, which has no contribution to the final sum rules. The spectral densities $\rho^{\rm H}_{\sigma}(s)$ can be approximated by using the ansatz of the quark-hadron duality~\cite{Shifman:1978by}, i.e. $\rho^{\rm H}_{\sigma}(s)= \rho^{\rm QCD}_{\sigma}(s)\theta (s-s_0)$.
In the space-like $q^2$-region, i.e. $(p+q)^2 - m_b^2 \ll 0$ and $q^2 \ll m^2_b$ for the momentum transfer, which correspond to small light-cone distance $x^2 \approx 0$, the correlator (\ref{correlators}) can be calculated by using the operator product expansion (OPE). By using the $b$-quark propagator given by Ref.\cite{Huang:1998gp}, we obtain
\begin{eqnarray}
\Pi_{\sigma} ^{\rm OPE }(p,q) &=& -i \sqrt{\frac{q^2}{\lambda}} \, \sum\limits_{\alpha=0,\pm,t} {\varepsilon_\sigma^{*\mu}(q)} \int \frac{d^4 x d^4 k}{(2\pi )^4} \frac{e^{i(q-k)\cdot x}}{m_b^2 - k^2}\nonumber\\
&& \big\{k^\nu \langle \rho(k,\varepsilon_\alpha(k))|{\rm T} \{ \bar d(x)\gamma _\mu \gamma _\nu \gamma _5 q(0) \}|0\rangle \nonumber \\
&& + {k^\nu }\langle \rho(k,\varepsilon_\alpha(k))|{\rm T}\{\bar d(x)\gamma _\mu \gamma _\nu {q}(0)\}|0\rangle \nonumber\\
&& + m_b \langle \rho(k,\varepsilon_\alpha(k))|{\rm T}\{ \bar d(x) \gamma_\mu \gamma_5 q(0)\}|0\rangle \nonumber\\
&& - m_b\langle \rho(k,\varepsilon_\alpha(k))|{\rm T}\{ \bar d(x) \gamma_\mu q(0) \}|0\rangle +\cdots \big\}.
\end{eqnarray}
The nonlocal matrix elements can be expressed in terms of the $\rho$-meson LCDAs of various twists~\cite{Ball:2004rg, Ball:2007zt}, which are put in the Appendix.
The LCSRs for the $B \to \rho$ HFFs are then ready to be derived by equating the correlator in the time-like and space-like regions due to analytic property of the correlator in different $q^2$-regions. After applying the Borel transformation, which removes the subtraction term in the dispersion relation and exponentially suppresses the contributions from unknown excited resonances, we get the required LCSRs for the HFFs:
\begin{widetext}
\begin{eqnarray}
\mathcal{H}_{\rho,0} &=& \frac{m_\rho m_b(m_B^2-m_\rho^2-q^2)}{2\sqrt{\lambda} m_\rho f_B m_B^2}\int_0^1 du e^{\left( {m_{B}^2 - s(u)} \right) / M^2} \bigg\{ \frac{ m_\rho f_\rho^\bot \cal C}{2u^2 m_\rho ^2}\Theta(c(u,s_0))\phi_{2;\rho}^\bot(u) +\frac{m_\rho f_\rho^\bot}{2u} \Theta(c(u,s_0))\nonumber\\
&&\times \psi_{3;\rho}^\|(u)+ \frac{ m_b f_\rho^\parallel}{u} \Theta \left(c \left( u,s_0 \right) \right) \phi_{3;\rho}^ \bot \left( u \right) - m_\rho f_\rho^\bot\bigg[ \frac{m_b^2{\cal C}}{8u^4M^4} \widetilde{\widetilde\Theta}(c(u,s_0)) + \frac{{\cal C}-2m_b^2}{8u^3M^2} \widetilde\Theta(c(u,s_0))\nonumber\\
&&- \frac{1}{8u^2}\Theta(c(u,s_0))\bigg]\phi_{4;\rho}^\bot(u) - \frac{m_b m_\rho^2 f_\rho^\parallel }{u^2 M^2}\widetilde \Theta \left( c\left( u,s_0 \right) \right)C_\rho(u) - m_\rho f_\rho^\bot\bigg[\frac{\cal C}{u^3M^2}\widetilde\Theta(c(u,s_0))-\frac{1}{u^2} \nonumber\\
&&\times \Theta(c(u,s_0))\bigg]I_L(u) - m_\rho f_\rho^\bot \bigg[\frac{2m_b^2}{2u^2M^2}\widetilde \Theta(c(u,s_0))+\frac{1}{2u} \Theta(c(u,s_0)) \bigg] H_3(u)\bigg\}+\int_0^1 dv \int_0^1 du \int_0^1 d {\mathcal D}\nonumber\\
&&\times e^{(m_{B}^2 - s(u)) / M^2}\frac{\widetilde\Theta(c(u,s_0))}{u^2 M^2} \frac{m_b m_\rho^2(m_B^2-m_\rho^2-q^2)} {24 \sqrt{\lambda}m_\rho f_B m_B^2} \bigg\{f_\rho^\bot \bigg[\widetilde {\Psi} _{4;\rho}^\bot (\underline \alpha ) - 12\bigg(\Psi _{4;\rho}^\bot (\underline \alpha )-2v \Psi _{4;\rho}^\bot (\underline \alpha )\nonumber\\
&&+2\Phi _{4;\rho}^{\bot(1)} (\underline \alpha ) -2\Phi _{4;\rho}^{\bot(2)} (\underline \alpha )+4v \Phi _{4;\rho}^{\bot(2)} (\underline \alpha )\bigg)\bigg]\bigg(m_B^2-m_\rho^2+2 u m_\rho^2\bigg)+ 2 m_b m_{K^*} f_\rho^\parallel \bigg(\widetilde {\Phi} _{3;\rho}^\parallel (\underline \alpha)\nonumber\\
&& + 12 \Phi _{3;\rho}^\parallel (\underline \alpha )\bigg)\bigg\}-\frac{\sqrt{\lambda}m_\rho m_b}{4 m_\rho f_B m_B^2}\int_0^1 du e^{\left( {m_{B}^2 - s(u)} \right) / M^2} \bigg\{ \frac{m_\rho f_\rho^\bot}{u m_\rho ^2}\Theta(c(u,s_0))\phi_{2;\rho}^\bot(u)- \frac{m_b f_\rho^\bot}{u M^2} \widetilde\Theta (c(u,s_0)) \nonumber \\
&&\times\psi_{3;\rho}^{\|}(u) - \frac{m_\rho f_\rho^\bot}{4}\bigg[\frac{m_b^2}{u^3M^4} \widetilde{\widetilde\Theta}(c(u,s_0)) + \frac{1}{u^2M^2} \widetilde\Theta(c(u,s_0))\bigg]\phi_{4;\rho}^\bot(u)+ \frac{2 m_b f_\rho^\parallel }{u^2 M^2}\widetilde \Theta \left( c\left( {u,{s_0}} \right) \right)
\nonumber \\
&&\times A_\rho(u) - \frac{m_\rho^2 m_b^3 f_\rho^\parallel}{2u^4 M^6}\widetilde {\widetilde{\widetilde \Theta }}\left( c\left( u,s_0 \right) \right)B_\rho(u) + \frac{2 m_b m_\rho^2 f_\rho^\parallel }{u^2 M^4}\widetilde {\widetilde \Theta }\left( c\left( u,s_0 \right)\right)C_\rho(u) + 2 m_\rho f_\rho^\bot \nonumber\\
&&\times \bigg[\frac{{\cal C} - 2m_b^2}{u^3 M^4}\widetilde{\widetilde\Theta}(c(u,s_0)) - \frac{1}{u^2M^2}\widetilde\Theta (c(u,s_0))\bigg]I_L(u)- \frac{m_\rho f_\rho^\bot}{u M^2}\widetilde \Theta (c(u,s_0))H_3(u)\bigg\} - \int_0^1 dv \int_0^1 du\nonumber\\
&&\times \int_0^1 d {\mathcal D} e^{\left( {m_{B}^2 - s(u)} \right) / M^2} \frac{ \sqrt{\lambda} m_b m_\rho^2 f_\rho^\bot} {24m_\rho f_B m_B^2(m_B+m_\rho)} \frac{ m_B + m_\rho}{u^2 M^2}\widetilde\Theta(c(u,s_0))\bigg[\widetilde {\Psi} _{4;\rho}^\bot (\underline \alpha ) + 12\bigg( 2v \Psi _{4;\rho}^\bot (\underline \alpha )\nonumber\\
&&-\Psi _{4;\rho}^\bot (\underline \alpha )+(4v-2) \Phi _{4;\rho}^{\bot(1)} (\underline \alpha ) + 2\Phi _{4;\rho}^{\bot(2)} (\underline \alpha ) \bigg)\bigg],
\label{HFF BV0}
\end{eqnarray}
\end{widetext}
\begin{widetext}
\begin{eqnarray}
\mathcal{H}_{\rho,1} &=& \frac{\sqrt{2q^2} m_b }{2f_B m_B^2}\int_0^1 du e^{\left( {m_{B}^2 - s(u)} \right) / M^2} \bigg\{f_\rho^\bot \Theta(c(u,s_0))\phi_{2;\rho}^\bot(u)+\frac{m_\rho m_b f_\rho^\parallel}{2u^2 M^2}\widetilde \Theta \left( c\left( u,s_0 \right) \right) \psi _{3;\rho}^ \bot (u) \nonumber\\
&& - \bigg[\frac{m_b^2}{u^2M^4}\widetilde {\widetilde\Theta}(c(u,s_0))+\frac{1}{uM^2}\widetilde\Theta(c(u,s_0))\bigg]
\frac{ m_\rho^2 f_\rho^\bot}{4}\phi_{4;\rho}^\bot (u)\bigg\}+ \int_0^1 dv \int_0^1 du \int_0^1 d {\mathcal D} e^{(m_{B}^2 - s(u)) / M^2} \nonumber\\
&& \times \frac{ \sqrt{2q^2} m_\rho^2 f_\rho^\bot} {6(m_B+m_\rho)} \frac{\widetilde\Theta(c(u,s_0))}{u^2 M^2}\bigg[(2v-1)\widetilde {\Psi} _{4;\rho}^\bot (\underline \alpha ) + 12\bigg(\Psi _{4;\rho}^\bot (\underline \alpha )-2 (v-1)(\Phi _{4;\rho}^{\bot(1)} (\underline \alpha )-\Phi _{4;\rho}^{\bot(2)} (\underline \alpha ))\bigg)\bigg],
\label{HFF BV1}
\end{eqnarray}
\end{widetext}
\begin{widetext}
\begin{eqnarray}
\mathcal{H}_{\rho,2} &=& \frac{\sqrt{2q^2}m_\rho m_b}{\sqrt{\lambda}f_B m_B^2}\int_0^1 du e^{\left( {m_{B}^2 - s(u)} \right) / M^2} \bigg\{ \frac{ m_\rho f_\rho^\bot \cal C}{2u^2 m_\rho ^2}\Theta(c(u,s_0))\phi_{2;\rho}^\bot(u) +\frac{m_\rho f_\rho^\bot}{2u} \Theta(c(u,s_0))\nonumber\\
&&\times \psi_{3;\rho}^\|(u)+ \frac{ m_b f_\rho^\parallel}{u} \Theta \left(c \left( u,s_0 \right) \right) \phi_{3;\rho}^ \bot \left( u \right) - m_\rho f_\rho^\bot\bigg[ \frac{m_b^2{\cal C}}{8u^4M^4} \widetilde{\widetilde\Theta}(c(u,s_0)) + \frac{{\cal C}-2m_b^2}{8u^3M^2} \widetilde\Theta(c(u,s_0))\nonumber\\
&&- \frac{1}{8u^2}\Theta(c(u,s_0))\bigg]\phi_{4;\rho}^\bot(u) - \frac{m_b m_\rho^2 f_\rho^\parallel }{u^2 M^2}\widetilde \Theta \left( c\left( u,s_0 \right) \right)C_\rho(u) - m_\rho f_\rho^\bot\bigg[\frac{\cal C}{u^3M^2}\widetilde\Theta(c(u,s_0))-\frac{1}{u^2} \nonumber\\
&&\times \Theta(c(u,s_0))\bigg]I_L(u) - m_\rho f_\rho^\bot \bigg[\frac{2m_b^2}{2u^2M^2}\widetilde \Theta(c(u,s_0))+\frac{1}{2u} \Theta(c(u,s_0)) \bigg] H_3(u)\bigg\}+\int_0^1 dv \int_0^1 du \int_0^1 d {\mathcal D}\nonumber\\
&&\times e^{\left( {m_{B}^2 - s(u)} \right) / M^2}\frac{\widetilde\Theta(c(u,s_0))}{u^2 M^2} \frac{\sqrt{2q^2}m_b m_\rho^2} {12\sqrt{\lambda} f_B m_B^2} \bigg\{f_\rho^\bot \bigg[\widetilde {\Psi} _{4;\rho}^\bot (\underline \alpha ) - 12\bigg(\Psi _{4;\rho}^\bot (\underline \alpha )-2v \Psi _{4;\rho}^\bot (\underline \alpha )\nonumber\\
&&+2\Phi _{4;\rho}^{\bot(1)} (\underline \alpha ) -2\Phi _{4;\rho}^{\bot(2)} (\underline \alpha )+4v \Phi _{4;\rho}^{\bot(2)} (\underline \alpha )\bigg)\bigg]\bigg(m_B^2-m_\rho^2+2 u m_\rho^2\bigg)+ 2 m_b m_\rho f_\rho^\parallel \bigg(\widetilde {\Phi} _{3;\rho}^\parallel (\underline \alpha)+ 12 \Phi _{3;\rho}^\parallel (\underline \alpha )\bigg)\bigg\}, \label{HFF BV2}
\end{eqnarray}
\end{widetext}
where we have implicitly set the factorization scale as $\mu$. $\int d{\cal D}=\int d\alpha_1 d\alpha_2 d\alpha_3 \delta(1 - \sum \limits_{i \rm = 1}^{\rm{3}} {\alpha_i})$. ${\mathcal C}=m_b^2+u^2m_\rho^2-q^2$, ${\mathcal E} = m_b^2 - u^2 m_\rho^2 + q^2$, $\mathcal F = m_b^2 - u^2 m_\rho^2 - q^2$, ${\mathcal H} = q^2/(m_B^2 - m_\rho^2)$, $\mathcal Q = m_B^2 - m_\rho^2 - q^2$, $c(\varrho,s_0) = \varrho s_0 - m_b^2 + \bar \varrho q^2 - \varrho \bar \varrho m_\rho^2$ and $s(\varrho)=[ m_b^2 - \bar \varrho(q^2 - \varrho m_\rho^2)] / \varrho $ ($\varrho = u$) with $\bar \varrho = 1 - \varrho$. $\Theta(c(u,s_0))$ denotes the usual step function. $\widetilde \Theta(c(u,s_0))$ and $\widetilde {\widetilde \Theta}(c(u,s_0))$ can be obtained from the surface terms $\delta(c({u_0},{s_0}))$ and $\Delta (c({u_0},{s_0}))$, whose explicit forms have been given in Ref.\cite{Fu:2014uea}. The functions $A_\rho(u)$, $B_\rho(u)$, $C_\rho(u)$, $H_3(u)$ and $I_L(u)$ are defined as:
\begin{eqnarray}
A_\rho(u) =&&\int_0^u dv \left[ \phi _{2;\rho}^\| (v) - \phi _{3;\rho}^ \bot (v) \right], \\
B_\rho(u) =&&\int_0^u dv \phi _{4;\rho}^\| (v), \\
C_\rho(u) =&&\int_0^u dv \int_0^v {dw} \left[\psi _{4;\rho}^\|(w) + \phi _{2;\rho}^\|(w)\right. \nonumber\\
&&\left.- 2 \phi_{3;\rho}^\bot(w) \right], \\
H_3(u)=&& \int_0^u dv \left[\psi_{4;\rho}^{\bot}(v)-\phi_{2;\rho}^\bot(v)\right]
\end{eqnarray}
and
\begin{eqnarray}
I_L(u) =&& \int_0^u dv \int_0^v dw \big[\phi_{3;\rho}^\|(w) -\frac{1}{2} \phi_{2;\rho}^\bot(w) \nonumber\\
&& -\frac{1}{2} \psi_{4;\rho}^\bot(w)\big].
\end{eqnarray}
\section{Numerical results}
\subsection{Input parameters and the HFFs}
We take the $\rho$-meson decay constants~\cite{Ball:2007zt}, $f_{\rho}^\bot=0.145(9)~{\rm GeV}$ and $f_{\rho}^\|=0.216(9)~{\rm GeV}$, the $b$-quark pole mass $m_b=4.80\pm0.05~{\rm GeV}$, the $\rho$-meson mass ${m_{\rho}} = 0.775~{\rm GeV}$, the $B$-meson mass $m_B = 5.279~{\rm GeV}$~\cite{Agashe:2014kda} and the $B$-meson decay constant $f_B=0.160\pm0.019{\rm GeV}$~\cite{Fu:2014pba}. The factorization scale $\mu$ is set as the typical momentum transfer of $B\to \rho$, i.e. $\mu\simeq (m_{B}^2-m_b^2)^{1/2} \sim 2~{\rm GeV}$, and we set its error as $\Delta\mu=\pm1~{\rm GeV}$~\cite{Ball:2004rg}.
\begin{table}[htb]
\centering
\caption{The $\rho$-meson LCDAs with different twist-structures, where $\delta \simeq m_{\rho}/m_b$ \cite{Ball:2004rg}.} \label{DA_delta}
\begin{tabular}{ cc c c }
\hline
& ~~twist-2~~ & ~~twist-3~~ & ~~twist-4~~ \\
\hline
~~$\delta^0$~~ & $\phi_{2;\rho}^\bot$ & / & $\Phi _{4;\rho}^{\bot(1)}$, $\Phi _{4;\rho}^{\bot(2)}$ \\
$\delta^1$ & $\phi_{2;\rho}^\|$ & $\phi_{3;\rho}^\bot$, $\psi_{3;\rho}^\bot$, $\Phi_{3;\rho}^\|$, $\widetilde\Phi_{3;\rho}^\|$ & / \\
$\delta^2$ & / & $\phi_{3;\rho}^\|$, $\psi_{3;\rho}^\|$ & $\phi_{4;\rho}^\bot$, $\psi_{4;\rho}^\bot$, $\Psi_{4;\rho}^\bot$, $\widetilde{\Psi} _{4;\rho}^\bot$ \\
$\delta^3$ & / & / & $\phi_{4;\rho}^\|$, $\psi_{4;\rho}^\|$ \\
\hline
\end{tabular}
\end{table}
Up to twist-4 accuracy, the needed $\rho$-meson light-cone distribution amplitudes (LCDAs) are grouped in Table~\ref{DA_delta}, in which $\delta=m_{\rho}/m_b \sim 0.16$. Since the contributions from the twist-4 terms themselves are numerically small, we thus directly adopt the twist-4 LCDA model derived from the conformal expansion of the matrix element to do the numerical calculation~\cite{Ball:2007zt}. Contributions from the twist-3 LCDAs $\phi_{3;\rho}^\bot$, $\psi_{3;\rho}^\bot$, $\Phi_{3;\rho}^\|$ and $\widetilde\Phi_{3;\rho}^\|$ are suppressed by $\delta^1$ and the twist-3 contributions from the LCDAs $\phi_{3;\rho}^\|$ and $\psi_{3;\rho}^\|$ are suppressed by $\delta^2$. The 2-particle twist-3 LCDAs, i.e. $\phi_{3;\rho}^\bot$, $\psi_{3;\rho}^\bot$, $\phi_{3;\rho}^\|$ and $\psi_{3;\rho}^\|$, can be related to the twist-2 LCDAs $\phi_{2;\rho}^\|$ and $\phi_{2;\rho}^\bot$ via the Wandzura-Wilczek approximation~\cite{Wandzura:1977qf,Ball:1997rj}. The 3-particle twist-3 LCDAs are also numerically small and we shall adopt the models of Ref.\cite{Ball:2007zt} to do the calculation. The twist-2 LCDAs, $\phi_{2;\rho}^\|$ and $\phi_{2;\rho}^\bot$, can be derived by integrating out the transverse momentum dependence of the twist-2 light-cone wavefunction model constructed in Refs.\cite{BHL, Wu:2010zc, Fu:2014cna, Fu:2014pba, Wu:2013lga, Huang:1994dy, Cao:1997hw, Huang:2004su}. For convenience, we call it as the WH-DA model, which states
\begin{eqnarray}
&& \phi _{2;\rho }^\lambda (x,\mu_0) = \frac{{A_{2;\rho}^\lambda \sqrt {3x\bar x} {m_q}}}{{8{\pi ^{3/2}}\widetilde f_\rho ^\lambda b_{2;\rho}^\lambda }}[1 + {B_{2;\rho}^\lambda }C_2^{3/2}(\varsigma )] \nonumber\\
&& \times \left[ {{\rm{Erf}}\left( {b_{2;\rho}^\lambda \sqrt {\frac{{{\mu^2_0} + m_q^2}}{{x\bar x}}} } \right) - {\rm{Erf}}\left( {b_{2;\rho}^\lambda \sqrt {\frac{{m_q^2}}{{x\bar x}}} } \right)} \right], \label{DA:WH}
\end{eqnarray}
where $\lambda=\|$ or $\perp$, respectively. The reduced decay constants $\widetilde f_{\rho}^\perp=f_{\rho}^\perp/\sqrt{3}$ and $\widetilde f_{\rho}^\|=f_{\rho}^\|/\sqrt{5}$, $\varsigma=2x-1$, and the error function ${\rm Erf}(x) = \frac2{\sqrt \pi}\int_0^x e^{ - t^2}dt$. The lepton quark mass $m_q$ is usually taken as $0.3$ GeV and we vary it within the region of $[0.2, 0.4]$ GeV for its uncertainty. The parameters $A_{2;{\rho}}^\lambda$, $B_{2;{\rho}}^\lambda$ and $b_{2;{\rho}}^\lambda$ can be determined by using the usual constraints:
\begin{itemize}
\item The normalization condition, $\int \phi_{2;\rho}^{\lambda}(x)dx=1$;
\item The average of the squared transverse momentum, $\langle{\bf k}_\bot^2 \rangle_{\rho}^{1/2} = 0.37~{\rm GeV}$~\cite{Wu:2010zc, Huang:2013yya}.
\item The second Gegenbauer moments of the twist-2 LCDAs $\phi^{\perp}_{2;\rho}$ and $\phi^{\|}_{2;\rho}$, $a_2^\bot (1~{\rm GeV}) = 0.14(6)$ and $a_2^\| (1~{\rm GeV}) = 0.15(7)$~\cite{Ball:2007zt}.
\end{itemize}
\begin{figure}[htb]
\centering
\includegraphics[width=0.45\textwidth]{phiLc.eps}
\includegraphics[width=0.45\textwidth]{phiTc.eps}
\caption{The leading-twist LCDA $\phi_{2;{\rho}}^\lambda (x,\mu_0=1~{\rm GeV})$, where $\lambda$ stands for the transverse ($\lambda=\bot$) and longitudinal ($\lambda=\|$) components, respectively. $m_q=0.3$ GeV.}
\label{figDA}
\end{figure}
\begin{table}[htb]
\centering
\caption{Parameters of the ${\rho}$-meson transverse leading-twist LCDA for some typical choices of $a_{2}^{\bot}(1~{\rm GeV})$. $m_q=0.3$ GeV.}
\label{DA_parameter_1}
\begin{tabular}{c c c c}
\hline
~~$a_{2}^{\bot}$~~ & ~~$A_{2;\rho}^{\bot}$ ~~&~~ $B_{2;\rho}^{\bot}$ ~~&~~ $b_{2;\rho}^{\bot}$~~\\
\hline
$0.20$ & $22.679$& $0.151$ & $0.555$ \\
$0.14 $ & $23.808$ & $0.100$ & $0.572$ \\
$0.08$ & $25.213$& $0.050$ & $0.595$ \\
\hline
\end{tabular}
\end{table}
\begin{table}[htb]
\centering
\caption{Parameters of the ${\rho}$-meson longitudinal leading-twist LCDA for some typical choices of $a_{2}^{\|}(1~{\rm GeV})$. $m_q=0.3$ GeV.}
\label{DA_parameter_2}
\begin{tabular}{c c c c}
\hline
~~$a_{2}^{\|}$ ~~&~~ $A_{2;\rho}^{\|}$ ~~&~~$B_{2;\rho}^{\|}$ ~~&~~ $b_{2;\rho}^{\|}$~~\\
\hline
$0.22$ & $22.620$ & $0.168$ & $0.549$ \\
$0.15$ & $23.951$ & $0.109$ & $0.569$ \\
$0.08$ & $25.275$ & $0.048$ & $0.590$ \\
\hline
\end{tabular}
\end{table}
Using those constraints, we can obtain the LCDA at the scale of $1$ GeV, whose behavior at any other scales can be achieved via the renormalization group evolution~\cite{Ball:2006nr}. The LCDA at any other scales can be obtained by using the conventional evolution equation. We present the parameters of $\phi_{2;\rho}^{\perp}$ and $\phi_{2;\rho}^{\|}$ in Table~\ref{DA_parameter_1} and \ref{DA_parameter_2}, and the corresponding curves in Fig.\ref{figDA}. Those two LCDAs are close in shape, both of which change from a convex behavior to a doubly humped behavior with the increment of the second Gegenbauer moment.
\begin{figure}[htb]
\centering
\includegraphics[width=0.45\textwidth]{phiLc1mq.eps}
\includegraphics[width=0.45\textwidth]{phiTc1mq.eps}
\caption{The leading-twist LCDA $\phi_{2;{\rho}}^\lambda (x,\mu_0=1~{\rm GeV})$ for $m_q\in[0.2, 0.4]$ GeV, where $\lambda$ stands for the transverse ($\lambda=\bot$) and longitudinal ($\lambda=\|$) components, respectively. $a_2^\bot (1~{\rm GeV}) = 0.14$ and $a_2^\| (1~{\rm GeV}) = 0.15$. }
\label{figDAmass}
\end{figure}
Fig.\ref{figDAmass} shows how the LCDA $\phi_{2;{\rho}}^\lambda$ changes with $m_q$. It is drawn by fixing all other input parameters to be their central values, and the LCDA parameters are refitted by fixing the second Gegenbauer moments $a_2^\bot (1~{\rm GeV}) = 0.14$ and $a_2^\| (1~{\rm GeV}) = 0.15$. As shown by Fig.(\ref{figDAmass}), different choices of light constitute quark $m_q$ can make sizable effects to the LCDA. Thus when discussing the uncertainties, the LCDA uncertainties from different choice of $m_q$ shall also be included.
\begin{table}[htb]
\caption{The Borel parameter $M^2$ for the HFFs $\mathcal{H}_{\rho,\sigma}$ at the continuum threshold $s_0=34.0$ $GeV^2$.}
\label{Borel}
\begin{tabular}{c c c c c c}
\hline
&$\mathcal{H}_{\rho,0}$ & $\mathcal{H}_{\rho,1}$ & $\mathcal{H}_{\rho,2}$
\\ \hline
$M^2$ & $25^{+0.5}_{-0.7}$ & $34.1^{+12.5}_{-7.8}$ & $21.8^{+3.7}_{-2.0}$
\\ \hline
\end{tabular}
\end{table}
As for the LCSRs of the HFFs, we also need to know the continuum threshold $s_0$ and the allowable range of the Borel parameter $M^2$, i.e. the so-called Borel window. The continuum threshold $s_0$, being as the demarcation of the $B$-meson ground state and higher mass contributions, is usually set as the one that is close to the first known resonance of the $B$-meson ground state. For the purpose, we set $s_0$ as $34.0\pm1.0$ ${\rm GeV}^2$, which indicates that the excitation energy is around 0.45 GeV to 0.65 GeV. The correlator is expanded over $1/M^2$, when we calculate it to all-power series, it shall be independent to the choice of $1/M^2$. However we only know its first several terms, and we have to set a proper range for $M^2$. As a conservative prediction, we require the continuum contribution to be less than $65\%$ of the total LCSR to set the upper limit of $M^2$, e.g.
\begin{eqnarray}
\frac{\int_{s_0}^\propto ds \rho^{\rm tot}(s)e^{-s/M^2}} {\int_{m_b^2}^\propto ds\rho^{\rm tot}(s)e^{-s/M^2}} \le 65\%. \label{con65}
\end{eqnarray}
Generally, the net contributions from the highest-twist terms increase with the decrement of $M^2$, and the lower limit of $M^2$ is usually fixed by requiring the highest-twist contributions to be small so as to ensure the convergence of the twist expansion. For the present considered three HFFs $\mathcal{H}_{\rho,\sigma}$, the twist-4 contributions behave quite differently. As a unified criteria for those HFFs, we adopt the flatness of the HFFs over $M^2$ to set the lower limit of $M^2$, e.g., we require the HFFs to be changed less than $1\%$ within the Borel window. The determined Borel window $M^2$ are listed in the Table~\ref{Borel}.
\begin{table}[htb]
\caption{Uncertainties of the LCSR predictions on the HFFs $\mathcal{H}_{\rho,\sigma}$ at the $q^2=10$ caused by the errors of the input parameters, e.g. $\Delta${DA} shows the uncertainty caused by varying the leading-twist LCDAs with the parameters listed in Tables \ref{DA_parameter_1} and \ref{DA_parameter_2}, in which the uncertainties caused by varying $m_q$ from $0.2 {\rm{GeV}} \to 0.4{\rm{GeV}}$ are also included. }
\label{HFF uncertainties}
\begin{tabular}{ c c c c c c c c }
\hline
&~~${\rm Central}$~~&~~$\Delta${DA}~~&~~$\Delta{\mu}$~~&~~$\Delta{M^2}$~~&~~$\Delta{s_0}$~~&~~$\Delta({m_b;f_B})$
\\ \hline
$\mathcal{H}_{\rho,0}$ & $0.688$ & $^{+0.003}_{-0.003}$ & $^{+0.000}_{-0.005}$ &$^{+0.006}_{-0.004}$& $^{+0.027}_{-0.027}$ & $^{+0.076}_{-0.062}$
\\
$\mathcal{H}_{\rho,1}$ & $0.314$ & $^{+0.002}_{-0.002}$ & $^{+0.000}_{-0.002}$ &$^{+0.000}_{-0.000}$& $^{+0.015}_{-0.018}$ & $^{+0.020}_{-0.016}$
\\
$\mathcal{H}_{\rho,2}$ & $0.408$ & $^{+0.003}_{-0.003}$ & $^{+0.000}_{-0.003}$ &$^{+0.003}_{-0.003}$& $^{+0.024}_{-0.026}$ & $^{+0.032}_{-0.025}$
\\
\hline
\end{tabular}
\end{table}
We take the HFFs $\mathcal{H}_{\rho,\sigma}(q^2=10)$ as explicit examples to show how the HFFs change with the input parameters. The results are collected in Table \ref{HFF uncertainties}, where errors from the $B$-meson decay constant $f_B$, the $b$-quark pole mass $m_b$, the $\rho$-meson mass ${m_{\rho}}$, the factorization scale $\mu$, the Borel parameter $M^2$ and the continuum threshold $s_0$. Table \ref{HFF uncertainties} shows that the main errors of those HFFs come from the parameters $m_b$, $f_B$, and $s_0$, whose effects could be up to $\sim 10\% -20\%$ accordingly.
\subsection{Extrapolation of the HFFs to all $q^2$-region}
The LCSR method is only valid for large energy of the final-state vector meson, e.g. $E_\rho \gg \Lambda_{\rm QCD}$. It implies a not too large $q^2$ via the relation $q^2 = m_B^2 - 2 m_B E_\rho$, e.g.
\begin{displaymath}
0\leq q^2\leq q^2_{\rm LCSR, max} \simeq 14~{\rm GeV}^2.
\end{displaymath}
On the other hand, the allowable physical range for $q^2$ is about $[0, 20.3]~{\rm GeV}^2$, in which the upper limit is fixed by $q^2_{\rm max}=(m_B-m_\rho)^2$~\cite{Ball:2004rg}. We adopt the method suggested by Ref.\cite{Bharucha:2010im} to do the extrapolation of the HFFs, i.e. the HFFs $\mathcal{H}_{\rho,\sigma}$ shall be extrapolated as a simplified series expansion as follows:
\begin{eqnarray}
\mathcal{H}_{\rho,0}(t) &=&\frac{1}{B(t) \sqrt{z(t,t_-)} \phi_T^{V-A}(t)} \sum_{k=0,1} a_k^{\rho,0} z^k , \\
\mathcal{H}_{\rho,1}(t) &=&\frac{\sqrt{-z(t,0)}}{B(t) \phi_T^{V-A}(t)} \sum_{k=0,1} a_k^{\rho,1} z^k , \\
\mathcal{H}_{\rho,2}(t) &=&\frac{\sqrt{-z(t,0)}}{B(t) \sqrt{z(t,t_-)} \phi_T^{V-A}(t)} \sum_{k=0,1} a_k^{\rho,2} z^k ,
\end{eqnarray}
where $\phi_I^X(t)=1$, $\sqrt{-z(t,0)}=\sqrt{q^2}/m_B$, $B(t)=1- q^2/m_{\rho,\sigma}^2$, $\sqrt{z(t,t_-)}=\sqrt{\lambda}/m_B^2$, and
\begin{eqnarray}
z(t)=\frac{\sqrt{t_+ - t}-\sqrt{t_+ - t_0}}{\sqrt{t_+ - t}+\sqrt{t_+ - t_0}}
\end{eqnarray}
with $t_\pm=(m_B\pm m_\rho)^2$ and $t_0=t_+(1-\sqrt{1-t_-/t_+})$.
\begin{table}[htb]
\centering
\caption{The fitted parameters $a^{\rho,\sigma}_{k}$ for the HFFs $\mathcal{H}_{\rho,\sigma}$, where all input parameters are set to be their central values.}
\label{analytic}
\begin{tabular}{ c c c c c c c c }
\hline
&$\mathcal{H}_{\rho,0}$ & $\mathcal{H}_{\rho,1}$ & $\mathcal{H}_{\rho,2}$
\\ \hline
$a_0^{\rho,\sigma}$ & 0.257 & $0.386$ & $0.354$
\\
$a_1^{\rho,\sigma}$ & 1.511 & $-1.020$ & $-0.310$
\\
$\Delta$ & 0.238 & $0.045$ & $0.128$
\\ \hline
\end{tabular}
\end{table}
The parameters $a_k^{\rho,\sigma}$ can be determined by requiring the ``quality'' of fit ($\Delta$) to be less than one, where $\Delta$ is defined as
\begin{equation}
\Delta=\frac{\sum_t\left|\mathcal{H}_{\rho,\sigma}(t)-\mathcal{H}_{\rho,\sigma}^{\rm fit}(t)\right|} {\sum_t\left|\mathcal{H}_{\rho,\sigma}(t)\right|}\times 100, \label{delta}
\end{equation}
where $t\in[0,\frac{1}{2},\cdots,\frac{27}{2},14]~{\rm GeV}^2$. We put the determined parameters $a_{k}^{\rho,\sigma}$ in Table~\ref{analytic}, in which all the input parameters are set to be their central values.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45 \textwidth, height=0.25 \textwidth]{BV0lihezLmq.eps}
\includegraphics[width=0.45 \textwidth, height=0.25 \textwidth]{BV1lihezLmq.eps}
\includegraphics[width=0.45 \textwidth, height=0.25 \textwidth]{BV2lihezLmq.eps}
\caption{The extrapolated LCSR predictions on the $B\to \rho$ HFFs $\mathcal{H}_{\rho,{(0,1,2)}}(q^2)$. The solid lines are center values and the shaded bands represent their uncertainties. }
\label{HFF:H012}
\end{center}
\end{figure}
We put the extrapolated $B\to \rho$ HFFs $\mathcal{H}_{\rho,\sigma}(q^2)$ in Fig.(\ref{HFF:H012}), where the shaded band stands for the squared average of all the mentioned uncertainties. All the HFFs are monotonically increase with the increment of $q^2$, and at the large recoil point, we have $\mathcal{H}_{\rho,0}(0)=0.435^{+0.055}_{-0.045}$ and $\mathcal{H}_{\rho,\{1,2\}}(0)\equiv 0$.
\section{The $B\to \rho$ semileptonic decay and the CKM matrix element $|V_{\rm ub}|$}
In this subsection, we apply the HFFs $\mathcal{H}_{\rho,\sigma}(q^2)$ to study the semileptonic decay $B\to\rho\ell\nu_\ell$, which is frequently used for precision test the SM and for searching of new physics beyond SM.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{dGmqall.eps}
\caption{The LCSR prediction for the differential decay width $1/|V_{ub}|^{2} \times d \Gamma/{d q^2}$. The LQCD prediction~\cite{Lattice96:1} and the extrapolated prediction of UKQCD group by using of the LQCD result~\cite{DelDebbio:1997nu} are presented as a comparison. The shaded bands are their theoretical errors. } \label{dGamma}
\end{center}
\end{figure}
Within the SM, the total differential decay width of $B\to\rho\ell\nu_\ell$ can be written as
\begin{equation}
\frac{1}{|V_{\rm ub}|^2}\frac{d\Gamma}{dq^2}={\cal G} \lambda(q^2)^{3/2} [\mathcal{H}_{\rho,0}^2(q^2) +\mathcal{H}_{\rho,1}^2(q^2)+\mathcal{H}_{\rho,2}^2(q^2)], \label{difftot}
\end{equation}
where the terms proportional $m_{\ell}^2$ have been suppressed due to the large chiral suppression for the light leptons with negligible masses, the parameter ${\cal G}={G_F^2}/{(192\pi^3 m_B^3)}$ with the fermi coupling constant $G_F=1.166\times10^{-5}~{\rm GeV}^{-2}$~\cite{Agashe:2014kda}, and the phase-space factor $\lambda(q^2)=(m_B^2 + m_\rho^2 - q^2)^2-4 m_B^2 m_\rho^2$. Our LCSR prediction for the differential decay width $1/|V_{\rm ub}|^{2} \times d \Gamma/{d q^2}$ is presented in Fig.(\ref{dGamma}), where the uncertainties from all error sources are added in quadrature. As a comparison, the UKQCD group LQCD prediction~\cite{Lattice96:1} and their extrapolated LQCD prediction (with the help of the heavy quark symmetry, kinematic constraints and the LCSR scaling relations)~\cite{DelDebbio:1997nu} are presented as a comparison. Our LCSR prediction is consistent with the LQCD prediction within the intermediate $q^2$-region; however our LCSR prediction prefer a larger $1/|V_{\rm ub}|^{2} \times d \Gamma/{d q^2}$ in low $q^2$-region and a smaller $1/|V_{\rm ub}|^{2} \times d \Gamma/{d q^2}$ in high $q^2$-region.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{dGmq.eps}
\caption{The LCSR prediction for the differential decay width $1/|V_{ub}|^{2} \times d \Gamma/{d q^2}$ for $m_q\in[0.2, 0.4]$ GeV, where other input parameters are set to be their central values.} \label{dGammamq}
\end{center}
\end{figure}
As a minor point, we pick out the uncertainty caused by varying $m_q\in[0.2, 0.4]$ GeV from the above uncertainty, and present the LCSR prediction for the differential decay width $1/|V_{ub}|^{2} \times d \Gamma/{d q^2}$ in Fig.(\ref{dGammamq}). It shows the uncertainty caused by $m_q$ is small, which agree with the observation of Table~\ref{HFF uncertainties} that the dominant uncertainties are from the parameters $m_b$, $f_B$, and $s_0$.
\begin{table}[htb]
\caption{The LCSR predictions and the extrapolated LQCD predictions of the UKQCD group~\cite{DelDebbio:1997nu} for the total decay width $\Gamma/|V_{\rm ub}|^{2}$ and the ratio $\Gamma_{\|}/\Gamma_{\bot}$. }
\begin{tabular}{ c c c c c }
\hline
&$\Gamma/|V_{ub}|^{2}$ & $\Gamma_{\|}/\Gamma_{\bot}$
\\ \hline
LCSR & $12.1^{+2.6}_{-2.5}$ & $1.14^{+0.35}_{-0.34}$
\\
UKQCD & $10.9^{+2.3}_{-1.5}$ & $0.80^{+0.04}_{-0.03}$
\\ \hline
\end{tabular}
\label{Ta:gammaTLS}
\end{table}
We present the total decay width $\Gamma/|V_{ub}|^{2}$ in Table \ref{Ta:gammaTLS}, in which we also present the ratio $\Gamma_{\|}/\Gamma_{\bot}$ as a useful reference. The total decay width, $\Gamma=\Gamma^\|+\Gamma^\bot$, where the decay width for the $\rho$-meson longitudinal components $\Gamma^\|$ is defined as
\begin{displaymath}
\Gamma^\|= {\cal G} |V_{\rm ub}|^2 \int_0^{q^2_{\rm max}} dq^2 \lambda(q^2)^{3/2} \mathcal{H}_{\rho,0}^2(q^2)
\end{displaymath}
and the decay width for the $\rho$-meson transverse components $\Gamma^\bot$ is defined as
\begin{displaymath}
\Gamma^\bot={\cal G} |V_{\rm ub}|^2 \int_0^{q^2_{\rm max}} dq^2 \lambda(q^2)^{3/2} [\mathcal{H}_{\rho,1}^2(q^2)+ \mathcal{H}_{\rho,2}^2(q^2)].
\end{displaymath}
Table \ref{Ta:gammaTLS} shows that, due to the large cancelation of the differences among different $q^2$-regions, the difference for the total decay width $\Gamma$ between the integrated LCSR and LQCD predictions shall be greatly suppressed.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{dGnewLT.eps}
\caption{The LCSR predictions for the polarized differential decay widths $1/|V_{ub}|^{2} \times d \Gamma^{\|}/{d q^2}$ and $1/|V_{ub}|^{2} \times d \Gamma^{\bot}/{d q^2}$. The LQCD result for total differential decay width~\cite{Lattice96:1} is presented as a comparison. } \label{dGammapol}
\end{center}
\end{figure}
We present the LCSR predictions for the polarized differential decay widths $1/|V_{ub}|^{2} \times d \Gamma^{\|}/{d q^2}$ and $1/|V_{ub}|^{2} \times d \Gamma^{\bot}/{d q^2}$ in Fig.(\ref{dGammapol}), in which all the input parameters are set to be their central values. Fig.(\ref{dGammapol}) shows that the differential decay widths for the final-state $\rho$-meson transverse and longitudinal components behave quite differently. The longitudinal differential decay width $d \Gamma^{\|}/{d q^2}$ monotonously deceases with the increment of $q^2$, and the transverse differential decay width $d \Gamma^{\bot}/{d q^2}$ shall first increase and then decrease with the increment of $q^2$. Both of them tend to zero for $q^2\to q^2_{\rm max}$ due to the phase-space suppression. As a result, the $\rho$-meson longitudinal component dominates low $q^2$-region, and its transverse component dominates high $q^2$-region~\footnote{Such dominance could be explained as a consequence of Lorentz invariance~\cite{Hiller:2013cza}.}.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{lowandhighrlow.eps}
\includegraphics[width=0.45\textwidth]{lowandhighrhigh.eps}
\caption{The LCSR predictions for the ratios $R_{\rm low}$ and $R_{\rm high}$. The BaBar~\cite{delAmoSanchez:2010af} results and the values by using extrapolated LQCD predictions~\cite{DelDebbio:1997nu} are also presented. }
\label{Rlowhigh}
\end{center}
\end{figure}
Experimentally, the BaBar collaboration measured the partial decay widths in three different $q^2$-regions~\cite{delAmoSanchez:2010af}
\begin{eqnarray}
\Delta \Gamma_{\rm low} &=& \int_0^8 \frac{d \Gamma}{d q^2} d q^2 = (0.747 \pm 0.234) \times 10^{-4}, \\
\Delta \Gamma_{\rm mid} &=& \int_8^{16} \frac{d \Gamma}{d q^2} d q^2 = (0.980 \pm 0.187) \times 10^{-4}, \\
\Delta \Gamma_{\rm high}&=& \int_{16}^{20.3} \frac{d \Gamma}{d q^2} d q^2 = (0.256 \pm 0.072) \times 10^{-4},
\end{eqnarray}
which lead to
\begin{eqnarray}
R_{\rm low} &=& \frac{\Gamma_{\rm low}}{\Gamma_{\rm mid}} = 0.762 \pm 0.280, \label{EQlow} \\
R_{\rm high} &=& \frac{\Gamma_{\rm high}}{\Gamma_{\rm mid}} = 0.216 \pm 0.089. \label{EQhigh}
\end{eqnarray}
Our LCSR calculation gives, $R_{\rm low}=0.967^{+0.308}_{-0.285}$ and $R_{\rm high}=0.219^{+0.058}_{-0.070}$; and the extrapolated LQCD calculation gives, $R_{\rm low}=0.668^{+0.283}_{-0.154}$ and $R_{\rm high}=0.409^{+0.032}_{-0.051}$. A comparison of those two ratios is presented in Fig.(\ref{Rlowhigh}). The LCSR predictions agree with the BaBar measurement with errors, while the extrapolated LQCD prefers a larger $R_{\rm high}$, which is about $1.6\,\sigma$ deviation from the BaBar measurement. Because the (middle) partial decay widths $\Delta \Gamma_{\rm mid}$ for the LCSR and LQCD approaches are close to each other, by comparing $R_{\rm low}$ and $R_{\rm high}$ with the experimental data, one can get the correct decay widths in different $q^2$-region and thus confirm which theoretical prediction is more reliable.
As a final remark, with the help of the branching ratio ${\cal B}(B^0\to\rho^-\ell^+\nu_\ell)=(2.45\pm 0.32)\times 10^{-4}$ and the lifetime $\tau(B^0)= 1.520\pm 0.004 {\rm ps}$~\cite{Beringer:1900}, we obtain $|V_{\rm ub}|=(2.96^{+0.52}_{-0.51})\times10^{-3}$, where the error is weighted average of all the mentioned error sources. This value agrees with the BaBar predictions~\cite{Wulsin:2010fp}, $(2.75\pm 0.24)\times10^{-3}$ and $(2.83\pm 0.24)\times10^{-3}$, and the CLEO predictions~\cite{Behrens:1999vv}, $(3.23\pm 0.24^{+0.23}_{-0.26}\pm0.58)\times10^{-3}$ and $(3.25\pm 0.14^{+0.21}_{-0.29}\pm0.55)\times10^{-3}$, within errors.
\section{Summary} \label{summary}
We have studied the HFFs for the $B$-meson semileptonic decay $B\to\rho\ell\nu_\ell$ within the LCSR approach. Fig.(\ref{HFF:H012}) shows that the extrapolated HFFs within the whole $q^2$-region. At the large recoil point, only the $\rho$-meson longitudinal component contributes, e.g. $\mathcal{H}_{\rho,0}(0)=0.435^{+0.055}_{-0.045}$ and $\mathcal{H}_{\rho,\{1,2\}}(0)\equiv 0$, where the errors are squared averages of the considered error sources. By applying the extrapolated HFFs to the semileptonic decay $B\to\rho\ell\nu_\ell$, we observe that the differential decay width $1/|V_{\rm ub}|^{2} \times d \Gamma/{d q^2}$, as shown by Fig.(\ref{dGamma}), is consistent with the Lattice QCD prediction within the intermediate $q^2$-region. However our LCSR prediction prefer a larger $1/|V_{\rm ub}|^{2} \times d \Gamma/{d q^2}$ in low $q^2$-region and a smaller $1/|V_{\rm ub}|^{2} \times d \Gamma/{d q^2}$ in high $q^2$-region. More explicitly, Fig.(\ref{dGammapol}) shows that the longitudinal decay width dominates the lower $q^2$-region and the transverse one dominates the higher $q^2$-region. Two typical ratios $R_{\rm low}$ and $R_{\rm high}$ can be used to test those properties. Our LCSR calculation shows that $R_{\rm low}=0.967^{+0.308}_{-0.285}$ and $R_{\rm high}=0.219^{+0.058}_{-0.070}$. Fig.(\ref{Rlowhigh}) shows that those predictions agree with the BaBar measurements within errors. Thus by using the HFFs with definite polarizations, some useful information can be achieved. A more precise measurement of those ratios shall be helpful for testing various calculation approaches.
\hspace{2cm}
{\bf Acknowledgements}: This work was supported in part by the Natural Science Foundation of China under Grant No.11625520 and No.11765007; by the Fundamental Research Funds for the Central Universities under the Grant No.2018CDPTCG0001/3; by the Project of Guizhou Provincial Department of Science and Technology under Grant No.[2017]1089; by the Project for Young Talents Growth of Guizhou Provincial Department of Education under Grant No.KY[2016]156; the Key Project for Innovation Research Groups of Guizhou Provincial Department of Education under Grant No.KY[2016]028.
|
\section*{Laboratory-Numerical Convection Experiments}
We study Rayleigh-B\'enard convection (RBC), in which a fluid layer is heated from below and cooled from above. The system is defined by three non-dimensional control parameters: the Rayleigh number $Ra = \alpha g \Delta H^3 / (\kappa \nu)$, which describes the strength of the buoyancy forcing relative to thermal and viscous dissipative effects; the Prandtl number $Pr = \nu/\kappa$, which discribes the thermophysical fluid properties; and the container's aspect ratio $\Gamma = D/H$. Here, $\alpha$ is the isobaric thermal expansion coefficient, $\nu$ is the kinematic viscosity, $\kappa$ is the thermal diffusivity, $g$ us the gravitational acceleration, $\Delta$ is the temperature drop across the fluid layer, $D$ is the diameter and $H$ is the height of convection tank.
We have carried out combined laboratory-numerical RBC experiments using gallium, a low $Pr \simeq 0.027$ liquid metal, as the working fluid.
The fluid is contained within a $\Gamma = 2$ cylinder, the largest container in which a single LSC will form and in which the highest convective velocities are attained \cite{Bailon2010, Sakievich2016, Stevens2018, Pandey2018}. This experimental design, which elaborates on the liquid mercury investigation of Tsuji et al. \cite{Tsuji2005} and departs from canonical $Pr \sim 1$, $\Gamma \sim 1$ studies, allows a sole LSC to develop in a maximally unconfined, strongly turbulent environment. The large amplitude velocity and temperature signatures in this system enable us to detect and quantify novel modes of the LSC that may not have been easily observed in the canonical set-up. In fact, we find a fundamentally new mode of large-scale turbulent convection with a three-dimensional oscillation that deviates from the quasi-planar description of LSC motion (Fig. \ref{fig:velocity}). Instead of sloshing or twisting side-to-side motions, our results show a flow that circulates around a crescent-shaped vortex, which in turn orbits the tank in the direction opposite the fluid velocity. As seen in Supplementary Movie S3, this vortex looks like a twirling jump rope.
\begin{figure}[t]
\centering
\begin{overpic}[width=\linewidth]{fig2.png}
\put(1,96.9){\small \textsf{A}}
\put(1,72.7){\small \textsf{B}}
\put(1,49.3){\small \textsf{C}}
\put(1,26.3){\small \textsf{D}}
\end{overpic}
\caption{Spatio-temporal evolution of laboratory convection velocities. Simultaneous ultrasonic Doppler measurements for the case at $Pr = 0.027$, $Ra~=~1.03 \times 10^6$ and $\Gamma = 2$. The particular measuring lines are indicated as a dashed red line in the schematics left of each data panel. Negative (positive) velocities represent flow towards (away from) the transducer. The measurements are non-dimensionalised using the free-fall velocity $u_{f\!f}=\sqrt{\alpha g \Delta H}$ and the free-fall time $\tau_{f\!f}= H/u_{f\!f}$. The ordinate corresponds to the measuring depth along the tank height (A, B), diameter (C), and chord (D), respectively. The measurements in (A--C) lie in the symmetry plane of the LSC; the chord probe measurements in (D) lie perpendicular to the LSC symmetry plane. The axial velocity in (A) and (B) show a mean up- and downflow, respectively, and relatively weak periodic fluctuations. While the mean velocity is zero in (C) and (D), strong oscillations are observed that span the tank. The flow along the chord in (D) shows a periodic double cell structure whereby the oscillation is in phase to (A). The strong periodic oscillation in (C) and (D) cannot be explained via the current LSC paradigm. (The white horizontal stripe in (C) is due to the standing echo from the \SI{1}{mm} diameter center point thermistor.)}
\label{fig:Doppler}
\end{figure}
Our $\Gamma = 2$ liquid gallium laboratory experiments are performed on the RoMag device (see Methods) and span a Rayleigh number range of $7.1 \times 10^4 \leq Ra \leq 5.1 \times 10^6$. Ultrasonic Doppler velocimetry \citep{Vogt2018} is utilised to measure the instantaneous flow distribution along four different measuring lines (Fig.~\ref{fig:Doppler}). Two ultrasonic transducers are attached antipodaly to the upper end block at cylindrical radius $r/R = 2/3$ to measure the vertical velocity field (Fig.~\ref{fig:Doppler}\textit{A, B}). Another two transducers are fixed to the sidewall horizontally at height $z/H = 1/2$ to measure midplane velocities along the diameter (Fig.~\ref{fig:Doppler}\textit{C}) and along a chord (Fig.~\ref{fig:Doppler}\textit{D}). The transducers are all oriented to align approximately with the symmetry plane of the large-scale flow, except for the chord probe that is perpendicular to it. Additionally, 29 thermistors are used to measure the experimental temperature field, including the central temperature of the bulk fluid, the vertical temperature difference across the fluid layer, and along one-third of the mid-plane sidewall. With this set-up, we diagnose the 3D dynamics of the liquid metal LSC.
\begin{figure*}
\centering
\begin{overpic}[width=17.8cm]{fig3.png}
\put(7,34){\small \textsf{A}}
\put(34.5,34){\small \textsf{B}}
\put(65.5,34){\small \textsf{E}}
\put(7,16){\small \textsf{C}}
\put(34.5,16){\small \textsf{D}}
\put(65.5,16){\small \textsf{F}}
\end{overpic}
\caption{Characteristic frequency and its scaling. Measured velocity and temperature frequency spectra in (A) and (B), respectively, in laboratory data at $Ra~=~1.03 \times 10^6$ and in (C) and (D) from DNS with $Ra = 1.12 \times 10^6$. All spectra are calculated using data obtained near the center of the fluid domain.
The four dashed lines in (A)--(D) indicate $f_0$ taken from (A).
(E) shows the dominant frequency as a function of $Ra$, normalised by the thermal diffusion time, $H^2/\kappa$, for experiments (open circles) and DNS (blue star). The best fit of the experimental data yields a $Ra^{0.419}$ scaling. (F) shows the frequency normalised by the estimated LSC turnover time, $6H/u_{r, rms}$, where we use the maximum possible path length of $2H + 2D = 6H$.}
\label{fig:freq}
\end{figure*}
All of our ultrasonic Doppler results exhibit a distinctive velocity pattern, visualised in Fig.~\ref{fig:Doppler} for $Ra = 1.03 \times 10^6$. The vertical velocities in Fig.~\ref{fig:Doppler}\textit{A, B} show flow near the axial plane of the LSC, with \textit{A} representing the cold downwelling flow (blue) and \textit{B} representing the warm upwelling motions (red) of the LSC. In addition, our measurements reveal both high and low frequency oscillations within the vertical velocity fields. The higher frequency oscillations correspond to small-scale plumes, whereas the lower frequency signals correspond to the fundamental oscillatory modes of the LSC. These vertical velocity measurements are in agreement with the quasi-planar model of the LSC. However, we find that the low frequency oscillation is strongest in the horizontal direction, aligned along the LSC's horizontal mid-plane (Fig.~\ref{fig:Doppler}\textit{C}). Further, the mid-plane chord probe measurements (Fig.~\ref{fig:Doppler}\textit{D}) show that the horizontal velocity switches sign across the midpoint of the chord, which lies in the symmetry plane of the LSC. These data indicate that the fluid is periodically diverging from the axial LSC plane and then converging back towards it. The measured velocities approach the free-fall velocity \citep{Glazier1999} $u_{f\!f} = \sqrt{\alpha g \Delta H}$. These velocities are well within our measurement capabilities and are thus detected as robust features of the flow. Significantly, these diverging-converging chord-probe flows indicate the presence of a strongly 3D flow pattern that is inconsistent with the previously proposed horizontal sloshing and torsional modes \citep{Brown2009, Xi2009}, and, thus, requires a novel physical explanation.
To diagnose the modes of behaviour of the inertial LSC flow within the opaque liquid metal, we carried out high resolution direct numerical simulations (DNS), using the fourth order finite volume code \textsc{goldfish} to provide detailed information on the spatially and temporally fully-resolved three-dimensional temperature and velocity fields; see Methods. The main DNS employs parameter values of $\Gamma = 2$; $Ra = 1.12 \times 10^6$ and $Pr = 0.025$, and is run for 1000 free fall time units after reaching statistical equilibrium.
We compare the outputs from the DNS and the laboratory experiments in Fig.~\ref{fig:freq}. Measured near the central point of the fluid bulk, the horizontal velocity (Fig.~\ref{fig:freq}\textit{A, C}) and temperature (Fig.~\ref{fig:freq}\textit{B, D}) spectra of the two systems quantitatively agree.
In fact, the peak oscillation frequency $f_0$ (dotted vertical line) is identical in all four spectra. This quantitative agreement demonstrates that our DNS captures the essential behaviours of the laboratory experiments, and is well-suited as a diagnostic tool to interpret the flows existing in the opaque liquid metal. In addition, the agreement shows that the idealised boundary conditions available in the DNS are sufficiently replicated in the laboratory experiments.
As a further verification of our coupled laboratory-numerical system, we compare the $f_0$ values (Fig.~\ref{fig:freq}\textit{A}) from our entire suite of experiments to the Grossmann--Lohse model for the bulk properties of LSC-dominated convective flows \citep{Grossmann2000, Stevens2013}. The value of $f_0$ is connected to the characteristic velocity of the LSC (Fig.~\ref{fig:freq}\textit{E, F}) and, thus, to the momentum transport \citep{Grossmann2002}. This transport is expressed by the Reynolds number $Re$, i.e. $f_0 H^2/\kappa = c Re Pr$ with a constant $c$ determined by the geometry, where $\kappa$ is the fluid's thermal diffusivity.
The best fit to the Grossmann-Lohse model \cite{Grossmann2002, Stevens2013} predicts an effective scaling in our parameter range of $Re \propto Ra^{0.435 \pm 0.002}$, which corresponds here to $f_0 \propto Ra^{0.422 \pm 0.002}$. By linear regression of our data set, we find that $f_0 \propto Ra^{0.419 \pm 0.006}$, or, equivalently, $Re \propto Ra^{0.433 \pm 0.006 }$ (Fig.~\ref{fig:freq}\textit{E}). Thus our DNS and experimental results are in agreement with time-averaged, kinematic models of LSC-dominated convective flows\citep{Grossmann2000}, but broaden the understanding of the time-varying LSC dynamics.
Similar to previous studies, we characterise the LSC via temperature measurements acquired along the midplane circumference of the cylinder \citep{Brown2005, Stevens2011, Zhou2009, Xi2009}. These sidewall temperature measurements provide information about the large-scale convective flows in the interior of the convection cell. Because of the high thermal diffusivity of liquid metals, the large-scale temperature signal is exceptionally clear as the small-scale temperatures are damped by diffusion. In addition, the high velocities in inertial liquid metal convective flows strongly advect the large-scale temperature field, producing midplane temperatures that almost reach the imposed maximal temperatures that exist on the top and bottom boundaries. Thus, temperature signals provide a strong and clear window into the LSC dynamics in liquid metal flows.
\begin{figure}[bh!]
\centering
\includegraphics[width=.9\linewidth]{fig4}
\caption{Midplane temperature signal on the sidewall. (A) Numerically and experimentally obtained temperature $(T-T_m)/\Delta$ at half height shown for 200 free-fall time units $t/\tau_{f\!f}$. The black dash-dotted line indicates the azimuthal $\phi$-width covered by the thermocouple array in the laboratory experiment.
The instantaneous position of the LSC $\xi_{LSC}$ is demarcated by the green line, which exhibits relatively small meanderings.
In contrast, the jump roping of the LSC causes the thermal pattern to fluctuate strongly, with warmer fluid (then colder fluid) occupying between $1/3$ and $2/3$ of the circumference, creating what looks like the baffles of an accordion. (B) Conditionally averaged DNS profiles of the temperature sidewall for the 12 phases of the jump rope oscillation. They reveal the splitting of the cold LSC downflow at $t=0$ (dark blue) and hot LSC upflow at $t=\tau_0/2$ (dark magenta) by clear double minima and maxima, respectively. The disparity of the profiles at $t=\tau_0/4$ (light blue), $t=3\tau_0/4$ (light magenta) suggests that the motion is three-dimensional. The grey lines are the profiles for the remaining 8 phases of the oscillation. The total mean (black dashed line) results in a simple cosine function.}
\label{fig:temperature}
\end{figure}
\section*{Jump Rope Vortex Cycle}
Fig.~\ref{fig:temperature}\textit{A} shows the midplane sidewall temperatures in the DNS (left) at $Ra = 1.12 \times 10^6$ and in the laboratory experiment (right) at $Ra = 1.03 \times 10^6$. The distinct blue and pink regions (cold and hot, respectively) reveal strong temperature alternations within the fluid, which we call an accordion pattern.
At $t/\tau_{f\!f} = 0$, the cold fluid covers roughly two-thirds of the sidewall circumference; by $t/\tau_{f\!f} \simeq 10$, the cold fluid covers only one-third of the circumference and then subsequently expands in azimuth again.
Averaging the sidewall temperatures over the entire circumference, $T_{a\!v\!g} = \langle T(r=R, z=H/2) \rangle_\phi$, consequentially yields quasi-periodic oscillations in $T_{a\!v\!g}$; see Methods. Unlike cases dominated by torsion and sloshing modes \citep{Funfschilling2004, Brown2009, Xi2009, Zhou2009}, our sidewall measurements have the same frequency $f_0$ as found in the central measurements shown in Fig.~\ref{fig:freq}\textit{A}--\textit{D} (see Supplementary Movies S2 and S3). The accordion thermal pattern in the sidewall temperature pattern (Fig.~\ref{fig:temperature}\textit{A} and Supplementary Fig.~S1) and the corresponding oscillation in $T_{a\!v\!g}$ (Fig.~\ref{fig:conditional}) provide the simplest means by which to identify the jump rope mode in convection data.
To fully diagnose the 3D characteristics of the LSC in the DNS, we first determined the position of the LSC within the fluid domain for each snapshot. Extending the single-sinusoidal-fitting method of Cioni et al. \cite{Cioni1997} and Brown et al. \cite{Brown2005}, each snapshot's midplane sidewall temperature distribution is fit to the function
\begin{equation*}
T_{fit}(\phi) = a \cos(\phi - \xi_{LSC}) + b \cos(2 (\phi - \xi_{LSC})) + T_{a\!v\!g} \, ,
\end{equation*}
where $\xi_{LSC}$ denotes the symmetry plane of the LSC (i.e., the green line in Fig.~\ref{fig:temperature}\textit{A} and also in Supplementary Movie S1). The values of $a$ and $b$ give the relative amplitudes of the hot and cold sidewall signals.
We then decompose the oscillation period $\tau_0 = 1/f_0$ into 12 characteristic phases based on the rms value of the sidewall mid-plane temperature (see Methods). Using these 12 phases, $10\,000$ temporally equidistant snapshots are conditionally-averaged over a total time span of $t/\tau_{f\!f} = 1000$. In this operation, we first orient the solutions so that the LSC symmetry plane remains fixed in azimuth and then average the temperature and velocity fields at the same temporal phase values, as shown in Supplementary Movie S2.
Our 12 conditionally-averaged midplane, sidewall temperature profiles are shown in Fig.~\ref{fig:temperature}\textit{B}. In agreement with previous quasi-planar LSC studies, the average of all phases yields a cosine function (dashed black line). However, our conditional averaging also reveals a more complex thermal structure with three extrema in the $t=0$ and $t = \tau_0/2$ profiles, with either a double-maxima-single-minimum or double-minima-single-maximum structure. Thus our conditional averaging extracts new information about the presence of additional complexity within the LSC dynamics. Further, a pure cosine function is not present at $t=\tau_0/4$ and $t=3\tau_0/4$, requiring the existence of spatiotemporal complexity in the LSC flow field.
The convective flow can only be understood by considering the LSC as a fully 3D vortical structure, whose vortex core traces a path similar to that of a jump rope that precesses around the tank in the direction opposite that of the LSC flow itself (Supplementary Movie S2). The motion of the jump rope vortex is illustrated by the streamlines that circumscribe the vortex core in Fig.~\ref{fig:velocity}. At the centre of the vortex (pink h\"ornchen-like structure), the velocity magnitude is zero.
As viewed in Supplementary Movie S3 and Supplementary Fig.~S2, the LSC fluid motions are in the clockwise direction, whereas the motion of the LSC core is counter-clockwise, akin to a planetary gear.
Looking in detail at the jump rope vortex cycle, we find at $t = 0$ that the LSC core is restricted to the horizontal midplane $z = H/2$ (Fig.~\ref{fig:velocity}\textit{B}).
At this time, the vortex core nears the midplane sidewall, impinging on the cold downwelling flow that exists there. This splits the downwelling fluid into two branches, creating the two distinct minima in the $t=0$ midplane temperature profile (Fig.~\ref{fig:temperature}\textit{B}). The splitting motions generate horizontally divergent flows, which explain the chord probe measurements of the experimental velocity field (Fig.~\ref{fig:Doppler}\textit{D}). By necessity, on the other side of the tank, the warm upwelling flow gets collimated, creating a high pressure wave along the bottom boundary layer. We hypothesise this collimation promotes the generation of instabilities, and thereby the detachment of warm convective plumes that drive the LSC more vigorously in the lower half of the cylinder, mainly along the $\xi_{LSC}$ plane, and push the centre of the LSC upward, initiating a broadening of the warm upwelling.
We find at $t=\tau_0/4$ (Fig.~\ref{fig:velocity}\textit{A, C}) that the vortex core no longer impinges on the midplane, but has moved to the upper half of the tank and the LSC has stretched to its longest elliptical path length. At this point, the cold downflow is still split, but the mid-plane temperature extrema are less pronounced. We find mirrored motions at $\tau = 0$ and $\tau = \tau_0/2$, and similarly at $\tau = \tau_0/4$ and $\tau = 3\tau_0/4$, thereby producing a symmetrical jump rope cycle.
\section*{Summary and Discussion}
We have verified that the observed jump rope behaviour is not unique to thermal convection in small $Pr$ fluids by simulating water with $Pr = 4.38$ (see SI and Fig.~S1). A coarse conditional-averaging scheme shows that the fundamental jump rope mode is detectable for $\Gamma =2$. The oscillation frequency is much lower in this fluid and the thermal and kinematic flow fields are both equally turbulent so that the sidewall temperature signals are far less pronounced (see SI). Further, such sidewall signals have also been detected in recent $Pr \simeq 12$, $\Gamma = 2$ oil experiments (Ping Wei, private communication).
Our results significantly alter the fundamental view of large-scale circulations. Based on our combined laboratory-numerical experiments, we find that the LSC in a $\Gamma > 1$ container is not confined to a quasi-2D circulation plane perturbed by horizontal sloshing and twisting modes. Instead, we find the LSC has a dominant 3D vortex core that precesses in a jump rope-like motion in the direction opposite to that of the LSC flow field.
We hypothesise that additional 3D modes exist within the LSC and that advanced techniques (e.g., dynamic mode decomposition \cite{Horn2017}, Koopman filtering \cite{Giannakis2018}) will be required to elucidate the full dynamics underlying turbulent convection in effectively unconfined natural systems.
\matmethods{
\subsection*{Laboratory Setup}
The experiments were performed with the RoMag device \citep{King2012b, King2013} using the liquid metal gallium confined in a right cylinder of aspect ratio $\Gamma=D/H=2$ with diameter $D=\SI{196.8}{mm}$ and height $H = \SI{98.4}{mm}$. The container's sidewall is made of stainless steel while the endwalls are made of copper. A non-inductively wound heater provides a heating power between \SI{6}{W} and \SI{1600}{W} at the bottom copper endwall. This heat is removed by a thermostated bath that circulates water through a double wound heat exchanger located above the top endwall. The sidewalls are wrapped by a \SI{20}{mm} layer of closed-cell foam insulation, followed by \SI{30}{mm} of Insulfrax fibrous thermal blanketing, and finally a \SI{30}{mm} layer of closed-cell foam insulation
in order to minimise radial heat losses.
Twenty-three experiments were conducted where the range of mean fluid temperatures varied between $\SI{35}{^\circ C} \leq T_{m} \leq \SI{47}{^\circ C}$ and the temperature drop across the fluid layer between $\SI{0.28} \leq \Delta \leq \SI{19.36}{K}$. Using the material properties for gallium \citep{Aurnou2018}, the Prandtl number ranges between $0.026 \leq Pr \leq 0.028$ and the Rayleigh number between $7.1 \times10^4 \leq Ra \leq 5.1 \times10^6$. Ultrasound Doppler velocimetry is used to measure the instantaneous velocity distribution along four different measuring lines, as shown in Fig.~\ref{fig:Doppler}. This technique is a useful tool to measure the velocities in opaque fluids such as liquid metals non-invasively \citep{Vogt2018, Vogt2013, Vogt2014}.
The four transducers (TR0805SS, Signalprocessing SA) capture the velocity component parallel to their ultrasound beam with a spatial resolution of about \SI{1}{mm} in beam direction and a temporal resolution of about \SI{1}{Hz}. All transducers are in direct contact with the liquid metal. They are approximately oriented in the LSC symmetry plane, except for the chord probe that is perpendicular to it.
A total number of 29 thermocouples is used to monitor the temperatures in the experiment. Six thermistors are embedded in each of the copper endwalls \SI{2}{mm} away from the fluid layer and are used to determine the mean fluid temperature, $T_m$, and temperature drop, $\Delta$, across the fluid layer. Seven thermistors are distributed inside the fluid layer while fifteen thermocouples are placed around the perimeter outside the fluid volume at midheight $z/H=0.5$. Thirteen of those thermocouples are positioned in an array ten degrees apart and used in the experimental array in Fig.~\ref{fig:temperature}\textit{A}. The temporal resolution of the thermal measurements is \SI{10}{Hz}. Experiments are conducted until equilibration is reached, when the thermal signals vary by less than 1\% over thirty minutes. Data is then saved for between three and six thermal diffusion times. In post processing, thermocouples placed between the insulation layers provide an estimate for sidewall heat losses. Additionally, the heat losses through vertical conduction in the stainless steel sidewall are also accounted for and the top and bottom fluid temperatures are corrected to include the conduction in the copper endways.
\subsection*{Direct Numerical Simulations}
The direct numerical simulations (DNS) have been conducted with the fourth order accurate finite volume code \textsc{goldfish} \cite{Shishkina2015, Shishkina2016, Horn2017}. It numerically solves the non-dimensional Navier-Stokes equations in the Oberbeck-Boussinesq approximation augmented by the temperature equation in a cylindrical $(r, \, \phi, \, z)$ domain:
\begin{eqnarray}
\vec{\nabla} \cdot {\vec{u}} &=& 0, \label{eq:NS1}\\
D_{{t}} {\vec{u}} &=& Ra^{-\frac{1}{2}}
Pr^{\frac{1}{2}} \gamma^{-\frac{3}{2}} \vec{\nabla}^2 {\vec{u}} -
\vec{\nabla} {p} + {T} {\hat{\vec{e}}}_z, \label{eq:NS2}\\
D_{{t}} {T} &=& Ra^{-\frac{1}{2}} Pr^{-\frac{1}{2}}
\gamma^{-\frac{3}{2}} \vec{\nabla}^2 {T},\label{eq:NS3}
\end{eqnarray}
where $\gamma$ is the radius-to-height aspect ratio $R/H = \Gamma/2$. The radius $R$, the buoyancy velocity $(g \alpha R \Delta)^{1/2}$, the temperature difference $\Delta$ and the material properties at the mean temperature are used as the reference scales. The mechanical boundary conditions are no-slip on all solid walls, the temperature boundary conditions are isothermal for the top and bottom and perfectly insulating for the sidewalls.
The numerical resolution for the main DNS with $Pr = 0.025$, $Ra = 1.12 \times 10^6$, $\Gamma = 2$ is $N_r \times N_\phi \times N_z = 168 \times 171 \times 168$; the total run-time was 1000 free-fall time units after reaching a statistical steady-state. The obtained results were verified on a finer $280 \times 256 \times 280$ mesh.
In addition, DNS for the same $Pr$ and $Ra$ were also carried out with a smaller aspect ratio of $\Gamma = 1$ for $t = 500 \tau_{f\!f}$. In this case, the dominant LSC motion is the commonly known combined sloshing and twisting.
The change of behaviour of the dominant LSC mode depending on the aspect ratio was also confirmed in a high-$Pr$ fluid by means of DNS with $Pr = 4.38$ and $Ra = 10^8$ and both aspect ratios $\Gamma = 1$ and $2$, with a total run-time of $ t = 1414\tau_{f\!f}$ and
$ t = 612 \tau_{f\!f}$, respectively (see SI Fig.~S1).
\subsection*{Conditional averaging} The temperature signal $T_{a\!v\!g} = \langle T(r=R, z=H/2) \rangle_\phi$, i.e.~the sidewall temperature at midheight averaged in azimuthal direction, shows a distinct oscillation with frequency $f_0$, shown in Fig.~\ref{fig:conditional}. To extract the characteristic behaviour during one cycle, we have sampled 10 snapshots per time unit, resulting in a total of 10,000 snapshots. We then defined seven intervals based on the standard deviation $\sigma$ of $T_{a\!v\!g}$ by the boundaries $-\frac{5}{4}\sigma,\, -\frac{3}{4}\sigma,\, -\frac{1}{4}\sigma, \,\frac{1}{4}\sigma,\, \frac{3}{4}\sigma, \, \frac{5}{4}\sigma$.
For $|T_{a\!v\!g}| > \frac{5}{4} \sigma$, all snapshots in those intervals were averaged.
In the remaining intervals we additionally considered whether the signal was in a phase where the temperature increases or decreases, respectively. This was algorithmically achieved by determining if the snapshot was located between a maximum and minimum or between a maximum and minimum. However, due to the possible occurrence of several multiple local extrema during one cycle, we had to hand-select the maxima and minima as shown in Fig.~\ref{fig:conditional}.
The result were twelve averaging intervals.
\begin{figure}[h!]
\begin{overpic}[width=\linewidth]{conditional}
\put(1,16.3){\small $T_{a\!v\!g}$}
\put(51,-2){\small $t/\tau_{f\!f}$}
\end{overpic}\\
\caption{Circumferentially-averaged temperature signal at mid-height $T_{a\!v\!g}$ from DNS with $Pr = 0.025$, $Ra = 1.12 \times 10^6$ and $\Gamma = 2$. The magenta (blue) downward (upward) triangles mark the hand-selected maxima (minima).
The horizontal dashed lines correspond to multiples of the standard deviation $\sigma$ of $T_{a\!v\!g}$, given by $\pm \frac{n}{4} \sigma, \, n \in \{1,3,5\}$, by which we separated the conditional averaging intervals.
The oscillation frequency of the signal matches $f_0$ as obtained at the centre point, shown in Fig.~\ref{fig:freq}, and is one of the most accessible identification schemes for the jump rope vortex.
\label{fig:conditional}}
\end{figure}
}
\showmatmethods{}
\acknow{This work was supported by the NSF Geophysics Program, award no. 1547269. T.V. acknowledges the Helmholtz Association for financial support within the framework of the Helmholtz-Alliance LIMTECH and for supporting his stay in Los Angeles. S.H. acknowledges funding by the DFG under grant HO 5890/1-1 and the Leibniz-Rechenzentrum for providing computational resources on SuperMUC.}
\showacknow{}
\bibstyle{pnas-new}
|
\section{Introduction}
The CRs' spectrum, and their compositions, contains abundant information for probing their origin, acceleration and propagation mechanisms. The spectral break of the CR called "knee" was first discovered in 1959 \cite{1959JETP...35....8K}, but its origin still remains a puzzle over nearly 60 years. Although current measurements about the knee are different among the ground-based experiments under different conditions, they can come into an agreement assuming the 15\% uncertainty of the energy reconstruction \cite{2003JPhG...29..809H}. And it is summarized that the knee is described by a resembled double-power law, with the spectral index -2.7 below 4 PeV, and steepening to -3.1 above this energy \cite{2013AIPC.1516..185H}.
\par
In order to explore the origin of the knee, precise measurement about the spectra of individual compositions is important. Due to the low-flux level of the CR knee about only 1 $m^{-2} yr^{-1}$, current measurements are settled ground-based to obtain the larger detecting aperture. And they are all performed through the indirect method, i.e. the Monte Carlo simulation about the extensive air shower (EAS) induced by the CRs impinging on the atmosphere. In lack of the knee energy data from the accelerator, the adopted hadronic interaction model extrapolated from the lower energy may bring unknown systematic errors in determining the spectrum and the compositions. There have been much efforts spent on the component measurement, but results make large discrepancy. Measurement from KASCADE favours that the knee of the all-particle spectrum is contributed by the steepening of the light species \cite{2005APh....24....1A}, and a heavy knee (mainly by the iron) is observed around 80 PeV \cite{2011PhRvL.107q1104A}. But the ARGO+WFCT finds the spectral break of the light species appears at $\sim$ 700 TeV \cite{2015PhRvD..92i2005B}. Meanwhile, in the analysis with various experiments, it is found the dominant specie at the knee is likely to be the Helium \cite{2006JPhG...32....1E}, and measurement from GAMMA attribute the knee to the light species (P and He) \cite{2014PhRvD..89l3003T}. Different from those, the $AS\gamma$ experiment indicates the knee originates from the nuclei higher than the Helium \cite{2006PhLB..632...58T}.
\par
With the poor information about the individual components, various explanations about the the knee's origin have been proposed (see \cite{2004APh....21..241H} and references therein). From the point of view about the astrophysical origin, many proposals attribute it to the change of the acceleration mechanism \cite{2007ApJ...661L.175B, 2002PhRvD..66h3004K, 2004A&A...417..807V}, the single source contribution \cite{2009arXiv0906.3949E, 2014PhRvD..89l3003T}, or the propagation effect in the Galaxy \cite{1995ICRC....2..697S, 1993A&A...268..726P, 2001ICRC....5.1889L, 2014PhRvD..90d1302G, 2015PhRvD..91h3009G}. In the consideration of the interaction effect, diverse models are proposed including the new channel of the hadronic interaction model \cite{2010EPJC...68..573D, 2001ICRC....5.1760K} and collision with exotic particles \cite{2003GReGr..35.1117K, 2008JCAP...12..003M, 2009JCAP...06..027B, 2003APh....19..379W, 2016arXiv161108384J} (relic neutrinos, the Dark Matter, SUSY, graviton, etc.). Besides, the regular process such as pair-production and disintegration of photons \cite{2009ApJ...700L.170H, 2001ICRC....5.1979T, 2002APh....17...23C} also belong to such kind of model. By investigating these proposals, it is found that most of them can be divided into two categories, the Z-dependent knee and the A-dependent knee (A, Z correspond to the CR nuclei's atomic number and charge respectively). The former one mainly relates to the CR's acceleration and propagation mechanisms, while the other one indicates many of them associate with the new physics processes. Thus, distinguishing between these two model is essential for exploring both the fundamental problems of CR and the new physics.
\par
The LHAASO experiment \cite{LHAASO_review} is the next generation of the ground-based experiment located at high altitude of 4410 m, at which the EAS induced by nuclei around the knee develops to the maximum and is expected to have less dependence on the hadronic interaction model. On the other hand, the LHAASO experiment combines the hybrid detection method, including detecting the charged particles, muons, as well as the Chrenkov/fluorescence photons. The charged particles construct the major part of the EAS' lateral distribution which is useful in determining the arrival directions, core locations, and primary energies, while the collected Cherenkov image is a good estimator about the CR energy and also sensitive to the CR component. The muons, generated by the decay of the charged pions, depend on the primary mass of CRs and have the ability in recognition of the primary CR species as well.
\par
Benefit from those advantages, the forthcoming LHAASO experiment will bring us an opportunity on the precise measurement about individual CR compositions. In this work, we investigate the capability of LHAASO in distinguishing the Z-dependent and A-dependent knee models. The contents are organized as follows: the section II contains the brief information about LHAASO, the section III contains the detail procedure of the analysis and the results of both the Z-dependent and the A-dependent knee models. In the last section, a conclusion and discussion is delivered.
\section{The LHAASO experiment}
The LHAASO experiment will be located at high altitude (4410 m a.s.l.) in the Daocheng site, Sichuan Province, China. It consists of an EAS array KM2A covering 1.3 $km^2$ area, 78000 $m^2$ closed packed water Cherenkov detector array (WCDA), and 12 wide-field Cherenkov/fluorescence telescopes (WFCTA). The KM2A is composed by two sub-arrays, including 1 $km^2$ array of 5195 electromagnetic particle detectors (ED) and the overlapping 1 $km^2$ array of 1171 underground water Cherenkov ranks for muon detectors (MD). The WCDA contains three water ponds with the effective depth about 4 m. Each pond is divided by 5 m $\times$ 5 m cells with an 8-inch PMT located at the bottom center to watch the Cherenkov light generated by the EAS secondary particles in the water. And the focal plane camera in each telescope of WFCTA has 32 $\times$ 32 pixels with every pixel size $0.5 \ \times \ 0.5$. The layout of each component of LHAASO is illustrated in the Fig. \ref{fig:lhaaso}.
\par
The LHAASO experiment aims at measuring precise primary CR spectrum through the hybrid detection method from 10 TeV to EeV, together with a sensitivity of 1\% Crab unity in order to survey the northern hemisphere to identify the gamma ray sources with full duty cycle, and measures the spectra of all the sources simultaneously within a wide energy range between $10^{11}$ and $10^{15}$ eV. The WFCTA will effectively measure the primary energy of different mass groups and will be operated in three modes covering wide energy range. The first and second are the Cherenkov mode focusing on the low energy (30 TeV - 10 PeV) and middle energy (10 PeV - 100 PeV) respectively, and the third mode is the fluorescence light detecting with the rearranged array allowing the energy spectrum and composition measurement above 100 PeV.
\par
The CR composition measurement is a challenging task, and there are four parameters from the hybrid detectors above-mentioned crucial in identifying the primary components. One parameter is offered by the WCDA calculating the energy flow near the core, another parameter is offered by the MD counting the total muons, the other two are extracted from WFCTA on behalf of the morphology of the Cherenkov image and the shower maximum $X_{max}$. These parameters utilize both the lateral and longitudinal features of the shower from the EAS secondary particles and Cherenkov photons, and are sensitive to the primary mass group. The detail description of the hybrid detecting method is shown in \cite{lhaaso_yin}. In this work, we focus on the first mode of WFCTA in combination with other arrays in the energy range 100 TeV $\sim$ 10 PeV. In order to insure the purity of the primary component, we test the expected observation about the proton (P) and the light species (P + He), which is sufficient to test the ability of LHAASO to distinguish whether the knee is A-dependent or Z-dependent. And it has been justified that the purity can reach up to 90 \% for P and 95 \% for P + He \cite{lhaaso_yin}.
\begin{center}
\begin{figure*}
\centering
\includegraphics[width=.45\textwidth]{lhaaso.eps}
\caption{Layout of the LHAASO experiment. The insets show the details of one pond of the WCDA, and the EDs (red points) and MDs (blue points) of the KM2A. the WFCTA, located at the edge of the WCDA, is also shown}
\label{fig:lhaaso}
\end{figure*}
\end{center}
\section{Statistical Test Of The CR Knee Models}
\subsection{Statistics Analysis Algorithm}
In the analysis, the hypothesis test is implemented to investigate the ability in distinguishing between the CR knee models. There are two hypotheses need to be tested, including the Z-dependent model versus the A-dependent model under the hypothesis of the A-dependent knee, and the A-dependent model versus the Z-dependent model under the hypothesis of the Z-dependent knee. And the chi-square test-statistics $\triangle \chi^2$ \cite{2018PhLB..780..181F} is formulated as
\begin{equation}
\label{test_statistics}
\triangle \chi^2 = min \chi^2 (H_0) \ - \ min \chi^2 (H_1)
\end{equation}
\par
As the confusing issue is whether the CR knee is Z-dependent or A-dependent, the $H_0$ and $H_1$ in Eq. \ref{test_statistics} represent these two hypotheses. The $H_1$ is the real explanation about the knee, and $H_0$ is the imaginary one. Noted that a better fitting result tends to the lower-value of the chi-square, the test-statistics $\triangle \chi^2$ will have a positive mean value only if the $H_1$ hypothesis is preferred. The chi-square of each model H is constructed by combining both the P and light components (P + He) measurements through
\begin{equation}
\label{chi_square}
\chi^2 (H) = \sum_i \left( \frac{(\mu_i(P|H) - N_i(P))^2}{\sigma_{stat}^2(N_i(P))} + \frac{(\mu_i(P+He|H) - N_i(P+He))^2}{\sigma_{stat}^2(N_i(P+He))} \right)
\end{equation}
Where $N_i(P)$ and $N_i(P+He)$ are the observed event counts of each bin, and $\mu_i(P|H)$ and $\mu_i(P+He|H)$ are fitted corresponding to the expectations of each model. Currently only the statistical error is considered, while the influence from the systematic uncertainty will be discussed in the last section. To minimise the chi-square with respect to each model, the MINUIT algorithm from the Root package \cite{Root_Minuit} is implemented.
\par
In the evaluation for the expectations of each bin, detection responses are adopted through
\begin{equation}
\label{bin_counts}
\mu_i = \Phi_i \cdot \triangle E_i \cdot T \cdot Aper_i
\end{equation}
where $\Phi_i$ is flux from the model prediction, $Aper_i$ is the aperture of the hybrid detection from the LHAASO and is adopted from \cite{lhaaso_yin}. T is the total exposure time and it is set to 3 years with 10\% duty cycle of the WFCTA. As mentioned above, the hybrid detection from LHAASO results in a low contamination rate, which is only around 5\% about the light components and 10\% about the proton. Assuming the contamination can be subtracted appropriately, it is expected to have little influence and can be neglected in the analysis.
\par
Assuming the event count $N_i$ falling in each bin obeys the Gaussian distribution, it is sampled using the Monte Carlo method, whose expectation shape is determined by the empirical function or explicit interaction calculation. In order to determine the distribution of the test-statistics $\triangle \chi^2$, 10000 pseudo experiments under the $H_0$ hypothesis are sampled according to directly fitting to the observation with the $H_0$ model, and the same algorithm in Eq. \ref{test_statistics} is performed for each pseudo experiment. We derive the exclusion significance of the $H_0$ hypothesis by comparing the observation $\triangle \chi^2_{obs}$ with the distribution of $\triangle \chi^2$. Totally 500 observation is generated to obtain an averaged exclusion significance as the evaluation of the LHAASO's capability in discriminating different knee models.
\subsection{A-Dependent Model Analysis}
The mechanism resulting in the A-dependent knee often relates to the threshold interaction. Such an process occurs when the nuclei $m_A$ impinges on the target, and the energy of each nucleon reaches the same threshold. As a result, the spectral break corresponds the threshold energy, and increase with the atomic number A. Our previous work \cite{2016arXiv161108384J} found the similar Lorentz factor $\sim 10^6$ for both the CR knee and electron's TeV cutoff as reported by latest measurement \cite{2015ICRC...34..411S, 2011ICRC....6...47B, 2009A&A...508..561A}, and tried to explore the latent mechanism behind this similarity. An unknown light particle is introduced which is abundant in the Galaxy with the mass less than 1 eV. The CRs will suffer the energy loss when the momentum transfer is sufficient to convert the particle to the heavier one, i.e. $10^6$ times heavier. As a result the thresholds of the CR proton and Helium derived are about 1 and 4 PeV respectively. Considering that the threshold interpretations of the knee often relate to the similar mechanism, we adopt our model as representative of them and denote it as the X model.
\par
The threshold interaction can generate the sharp knee structure by tuning the interaction cross section across the knee. And we depict the sharp knee shape with the double power-law function, where
\begin{equation}
\label{power_law}
\Phi(E) = \begin{cases}
\Phi_0 \left( \frac{E_{cut}}{1 \ TeV} \right)^{\gamma_1} \left( \frac{E}{E_{cut}} \right)^{\gamma_1}, \ E < E_{cut} \\
\Phi_0 \left( \frac{E_{cut}}{1 \ TeV} \right)^{\gamma_1} \left( \frac{E}{E_{cut}} \right)^{\gamma_2}, \ E > E_{cut}
\end{cases}
\end{equation}
The parameter $\Phi_0$ is the normalized flux. And the knee of each component $E_{cut}$ is $Z \cdot E_0$ under the Z-dependent model ($H_0$) and $A \cdot E_0$ under the A-dependent model ($H_1$).
\par
One sampled observation together with the fitting results of different hypotheses are illustrated in the Fig. \ref{fig:X_model_obs}. Both of model fittings have no difference about the proton spectrum, while the divergence is mainly produced by the light component spectrum at the last three bins by eye, whose energy is above the knee of the He. As the model predicts the knee of He is located at about 4 PeV. The Z-dependent fitting generates the He's knee at about 2 PeV, which is insufficient to explain the higher-energy spectral break of the light components.
\begin{center}
\begin{figure*}
\centering
\includegraphics[width=.45\textwidth]{obs_Xmodel.eps}
\caption{One sampled observation and the corresponding fitting results under the X model. The blue dots are the sampled light component spectrum, and the red dots are the sampled proton spectrum. Solid lines are the Z-dependent model ($H_0$) fitting results, and dashed lines are the A-dependent model ($H_1$) fitting result.}
\label{fig:X_model_obs}
\end{figure*}
\end{center}
\par
The results from one of the pseudo experiment under the Z-dependent fitting are shown in the left panel Fig. \ref{fig:X_model_H0}. Different from the observation, the pseudo experiment generates the knee of the He at about 2 PeV and the A-dependent fitting at the knee of He exceeds this energy. The collect of the $\triangle \chi^2$ from all the pseudo experiments and the observation $\triangle \chi^2_{obs}$ are shown in the right panel of Fig. \ref{fig:X_model_H0}. And $\triangle \chi^2$'s distribution shows a well-defined Gaussian shape. The test-statistcs of the observation separates far from the pseudo experiments, which means the actual model is easy to be recognized. The significance is calculated by dividing the bias $(\triangle \chi^2_{obs} - <\triangle \chi^2>)$ by the root-mean square (RMS) of the distribution of the $\triangle \chi^2$. Averaging the significance over the total 500 sampled observations, the mean significance under the X model is about 10.7 $\sigma$.
\begin{center}
\begin{figure*}
\centering
\includegraphics[width=.45\textwidth]{pseudo_Xmodel.eps}
\includegraphics[width=.45\textwidth]{sample_Xmodel.eps}
\caption{left: One sampled pseudo experiment and the corresponding fitting results under the Z-dependent fitting of the X model. The blue dots are the sampled light component spectrum, and the red dots are the sampled proton spectrum. Solid lines are the Z-dependent model ($H_0$) fitting results, and dashed lines are the A-dependent model ($H_1$) fitting result. right: The distribution of the test-statistics $\triangle \chi^2$ (blue) and the observation $\triangle \chi^2_{obs}$ (red dash-doted line) from one sampled observation}
\label{fig:X_model_H0}
\end{figure*}
\end{center}
\par
Although a prominent significance is obtained under the X model, an essential part is attributed to the intrinsic divergence from the empirical function in Eq. \ref{power_law} of the fine structures in the X model. As shown in Fig. \ref{fig:X_model_obs}, the bump features appear at the break points of both the P and Helium knees, which are not afford to described by the empirical function. In order to eliminate these divergences, we generalize the prediction by parameterizing the X model with the double power-law function in Eq. \ref{power_law}, and denote this new prediction as the GX model. The fitting parameters are listed in Tab. \ref{tab:GX_model}.
\begin{table}
\centering
\caption{\label{tab:GX_model} The fitting parameters of the GX model.}
\begin{tabular}{ccccccc}
\toprule
\hline
parameters & P & He \\
\hline
$\Phi_0$ ( $m^{-2} \cdot s^{-1} \cdot sr^{-1} \cdot TeV^{-1}$ ) & 0.086 & 0.058 \\
$E_0$ (TeV) & 807 & 4$\times$807 \\
$\gamma_1$ & -2.7 & -2.64 \\
$\gamma_2$ & -3.04 & -3.33 \\
\hline
\bottomrule
\end{tabular}
\end{table}
\par
Implementing the same analysis procedure under the GX model, the averaged exclusion significance of the imaginary model is obtained about 6.6 $\sigma$, which is less than the X model by 4.1 $\sigma$. The difference is accordingly caused by the elimination from the fine structures. One of the sampled observation is plotted in the Fig. \ref{fig:GX_model_obs}. It should be pointed that the GX model only describes the A-dependent knee structure, and is independent from different dynamics implied among such kind of interpretations. Thus the result derived from the GX model is suitable to represent all the A-dependent models with He-dominant knee.
\begin{center}
\begin{figure*}
\centering
\includegraphics[width=.45\textwidth]{obs_GXmodel.eps}
\caption{One pseudo experiment and the corresponding fitting results under the GX model.}
\label{fig:GX_model_obs}
\end{figure*}
\end{center}
\subsection{Z-Dependent Model Analysis}
From the astrophysical point of view, the origin of the knee, whether induced by the acceleration limit or the leakage from the Galaxy, is compactly associated with the ambient magnetic field. In those mechanisms, the charged part of the nuclei is the participant and leaves the neutral part as the spectator. The Z-dependent knee is commonly adopted and supported by some experiments, such as KASCADE \cite{2012MSAIS..19...49K}, GAMMA \cite{2007APh....28..169G} and so on. On the other hand, it is also found that the spectral knee is a sharp structure \cite{2008ApJ...678.1165A}, which is not favoured by the smooth-knee prediction from the traditional diffusion model \cite{2009arXiv0906.3949E}. In order to account the sharp knee structure under the Z-dependent prediction, we adopt the fitting result from the GAMMA experiment \cite{2014PhRvD..89l3003T} as a representative, which extracts the knee of the P and He at 3.2 and 6.4 PeV respectively.
\par
The double power-law function in Eq. \ref{power_law} is also used to depict the model hypotheses. One observation and relevant fittings are shown in the Fig \ref{fig:GAMMA_model_obs}. It should be emphasized that the prediction at the knee from the GAMMA model is proton-dominant, while the analysis about the former A-dependent knee based on the He-dominant assumption. This uncertainty about the dominant component causes a significant variance of the testing results. In this scenario, the knee of the He under the $H_0$ fitting is beyond the concerned energy range, which is expected to be $\sim$ 12 PeV.
\begin{center}
\begin{figure*}
\centering
\includegraphics[width=.45\textwidth]{obs_GAMMA.eps}
\caption{One pseudo experiment and the corresponding fitting results under the GAMMA model. The blue dots are the sampled light component spectrum, and the red dots are the sampled proton spectrum. Solid lines are the Z-dependent model ($H_1$) fitting results, and dashed lines are the A-dependent model ($H_0$) fitting result.}
\label{fig:GAMMA_model_obs}
\end{figure*}
\end{center}
\par
We find that the corresponding distribution of $\triangle \chi^2$ is not the Gaussian shape and most fitting results from the two model are same while only 5\% of the samples show a slight difference. As mentioned above, the pseudo experiments can exhibit only one spectral break for the light components, which is contributed by the proton. The statistical information can be extracted by integrating the portion of $\triangle \chi^2$ larger than the observation $\triangle \chi^2_{obs}$, which is defined as
\begin{equation}
p-value = P(\triangle \chi^2 \geq \triangle \chi^2_{obs}|H_0)
\end{equation}
And we derive that the averaged p-value is about 0.023, corresponding to 2 $\sigma$.
\par
Besides, another model described by Horandel \cite{2003APh....19..193H} also depicts the Z-dependent knee but a smooth structure. This model is adopted in the analysis as well. Notice that the double power-law function is hard to describe a smooth knee, the poly-gonato model \cite{2003APh....19..193H} with an extra sharpness parameter $\epsilon_c$ is used instead, which is formed as
\begin{equation}
\Phi(E) = \Phi_0 \left( \frac{E}{1 \ TeV} \right)^{\gamma_1} \left[ 1 + \left( \frac{E}{E_{cut}} \right)^{\epsilon_c} \right]^{(\gamma_2 - \gamma_1)/\epsilon_c}
\label{poly_gonato}
\end{equation}
\par
In the Eq. \ref{poly_gonato}, the spectral index is $\gamma_1$ below the knee energy $E_{cut}$ and $\gamma_2$ above, where the intermediate region around the knee depends on the sharpness parameter $\epsilon_c$, which determines the spectral transition rate. Similar to the results of the GAMMA model, the predictions between the two models are found barely no difference. One common feature of them is the P-dominant assumption around 3$ \sim$ 5 PeV, so the fitted knee of the He under the $H_0$ hypothesis is out of energy range as well. As a result, poor exclusion capability is obtained only 1 $\sigma$. In general, due to the limitation from the concerned energy range, our analysis losses the significance for classifying the P-dominant model, whether it is Z-dependent or A-dependent. The measurement extending to higher energies is thus essential in discriminating the knee models.
\section{Conclusion And Discussion}
Realizing that different interpretations about the origin of the knee correspond to different spectral shapes, where the acceleration or the propagation origin result in the Z-dependent knee, and many of the interaction models with new physics results in the A-dependent knee, precise measurement for the individual component is important. Benefit from the merit of the high altitude and the hybrid detection method, the forthcoming LHAASO experiment is expected sensitive to the individual components. In this work, we investigate the capability of LHAASO in distinguishing these knee models. In the consideration of the energy range $10^5 \ \sim \ 10^7$ GeV with 3-year observation statistics, we find the the Z-dependent hypothesis can be excluded at the significance of 6.6 $\sigma$ under the A-dependent knee (He-dominant), while the A-dependent hypothesis mixes with the Z-dependent knee and is harder to be excluded with the significance only 2 $\sigma$.
\par
The influence of the systematic uncertainty can be addressed if we attribute the major part to the energy calibration and the detecting efficiency. These concerned systematic uncertainties lead to the integral shift of the spectrum along the axes. Due to the only acting factor in this analysis is the the ratio of the knee energy of the He and P, which is split into 4 and 2 with respect to different knee models, the integral shift of the spectrum is not expected to influence the ratio significantly. Thus, this test is insensitive to the variance of the systematic uncertainties, which is a unique advantage.
\par
The lack of significant in recognition of the Z-dependent knee in this analysis is mainly due to the relative narrow energy band around the knee and the P-dominant assumption. Other modes of LHAASO that focus the higher energies is required to further determine the spectral index above the He's knee. Meanwhile latest measurement from ARGO \cite{2014arXiv1408.6739D, 2016NPPP..279....7M}, ARGO+WFCT \cite{2015PhRvD..92i2005B} and $AS\gamma$ \cite{2006PhLB..632...58T} also find some hints that the knee of the light components occurs at hundreds TeV, which corresponds to higher flux at the knees of the P and He. If this observation is confirmed by further LHAASO experiment, the analysis will fall into the concerned energy range naturally and can be performed much more conveniently with higher statistics.
\section{Acknowledgements}
This work is supported by the National Key R\&D Program of China (number 2018YFA0404200), the Natural Sciences Foundation of China (numbers 11575203, 11635011).
|
\section{Introduction}
\noindent The objective of this paper is to characterize Leavitt path algebras
with bounded index of nilpotence. A ring $R$ is said to have \textit{bounded
index of} \textit{nilpotence} if there is a positive integer $n$ such that
$x^{n}=0$ for all nilpotent elements $x$ in $R$. If $n$ is the least such
integer then $R$ is said to have \textit{index of nilpotence $n$}. We show
that the Leavitt path algebra $L:=L_{K}(E)$ of a directed graph $E$ over a
field $K$ has index of nilpotence at most $n$ if and only if no cycle in the
graph $E$ has an exit and there is a fixed positive integer $n$ such that the
number of distinct paths that end at any given vertex $v$ (including $v$, but
not including the cycle $c$ in case $v$ lies on $c$) is less than or equal to
$n$. In this case, $L$ becomes a subdirect product of matrix rings $M_{t}(K)$
or $M_{t}(K[x,x^{-1}])$ of finite order $t\leq n$. Examples are constructed
showing that $L$ need not decompose as a direct sum of these matrix rings
$M_{t}(K)$ or $M_{t}(K[x,x^{-1}])$, though the decomposition is possible when
$E$ is row-finite. We show that a Leavitt path algebra $L$ with bounded index
of nilpotence is always directly-finite and that $L$ is a $%
\mathbb{Z}
$-graded $\Sigma$-$V$ ring, that is, each graded simple left/right $L$-module
is graded $\Sigma$-injective. Examples show that the converse of these
statements do not hold. Interestingly, it turns out that the graphical
conditions on $E$ that ensure $L$ has a bounded index of nilpotence are
exactly the same graphical conditions on $E$ that were shown in \cite{BLR} to
imply that $L$ satisfies a polynomial identity. When $E$ is a finite graph,
these graphical conditions also imply that $L$ has GK-dimension $\leq1$. Such
statements illustrate a unique phenomenon in the study of Leavitt path
algebras where a single graph property of $E$ often implies different
ring-theoretic properties for $L$ and these ring-theoretic properties are
usually independent of each other for general rings (see \cite{R1} for several
illustrations of this phenomenon of Leavitt path algebras). This feature of
Leavitt path algebras makes them really useful tools in constructing examples
of rings of various desired ring-theoretic properties. Finally, as an
application of our results, we answer a question raised in \cite{JST}, whether
an exchange $\Sigma$-$V$ ring has bounded index of nilpotence.
For the general notation, terminology and results in Leavitt path algebras, we
refer the reader to \cite{AArS}.\ We give below an outline of some of the
needed basic concepts and results.
A (directed) graph $E=(E^{0},E^{1},r,s)$ consists of two sets $E^{0}$ and
$E^{1}$ together with maps $r,s:E^{1}\rightarrow E^{0}$. The elements of
$E^{0}$ are called \textit{vertices} and the elements of $E^{1}$
\textit{edges}.
A vertex $v$ is called a \textit{sink} if it emits no edges and a vertex $v$
is called a \textit{regular} \textit{vertex} if it emits a non-empty finite
set of edges. An \textit{infinite emitter} is a vertex which emits infinitely
many edges. For each $e\in E^{1}$, we call $e^{\ast}$ a ghost edge. We let
$r(e^{\ast})$ denote $s(e)$, and we let $s(e^{\ast})$ denote $r(e)$.
A\textit{\ path} $\mu$ of length $n>0$ is a finite sequence of edges
$\mu=e_{1}e_{2}\cdot\cdot\cdot e_{n}$ with $r(e_{i})=s(e_{i+1})$ for all
$i=1,\cdot\cdot\cdot,n-1$. In this case $\mu^{\ast}=e_{n}^{\ast}\cdot
\cdot\cdot e_{2}^{\ast}e_{1}^{\ast}$ is the corresponding ghost path. A vertex
is considered a path of length $0$.
A path $\mu$ $=e_{1}\dots e_{n}$ in $E$ is \textit{closed} if $r(e_{n}%
)=s(e_{1})$, in which case $\mu$ is said to be \textit{based at the vertex
}$s(e_{1})$. A closed path $\mu$ as above is called \textit{simple} provided
it does not pass through its base more than once, i.e., $s(e_{i})\neq
s(e_{1})$ for all $i=2,...,n$. The closed path $\mu$ is called a
\textit{cycle} if it does not pass through any of its vertices twice, that is,
if $s(e_{i})\neq s(e_{j})$ for every $i\neq j$. An \textit{exit }for a path
$\mu=e_{1}\dots e_{n}$ is an edge $e$ such that $s(e)=s(e_{i})$ for some $i$
and $e\neq e_{i}$.
An \textit{infinite rational path} $p$ is an infinite path of the form
$p=x_{1}x_{2} \ldots x_{n} \ldots$ where there is an $m \geq1$ such that
$x_{k} = g$, a fixed closed path for all $k \geq m$ and that $x_{k}$ is an
edge $e_{k}$ if $k <m$. Thus $p$ will be of the form $p=e_{1}e_{2}\cdot
\cdot\cdot e_{m-1}gggg\cdot\cdot\cdot$ where $g$ is a closed path and the
$e_{i}$ are edges. An infinite path which is not rational is called an
\textit{irrational path}.
A graph $E$ is said to satisfy \textit{Condition (K)}, if every vertex $v$ on
a closed path $c$ is also the base of a another closed path $c^{\prime}%
$different from $c$. A graph $E$ is said to satisfy \textit{Condition (L)} if
every cycle has an exit.
If there is a path from vertex $u$ to a vertex $v$, we write $u\geq v$. A
subset $D$ of vertices is said to be \textit{downward directed }\ if for any
$u,v\in D$, there exists a $w\in D$ such that $u\geq w$ and $v\geq w$. A
subset $H$ of $E^{0}$ is called \textit{hereditary} if, whenever $v\in H$ and
$w\in E^{0}$ satisfy $v\geq w$, then $w\in H$. A hereditary set is
\textit{saturated} if, for any regular vertex $v$, $r(s^{-1}(v))\subseteq H$
implies $v\in H$.
Given an arbitrary graph $E$ and a field $K$, the \textit{Leavitt path algebra
}$L_{K}(E)$ is defined to be the $K$-algebra generated by a set $\{v:v\in
E^{0}\}$ of pair-wise orthogonal idempotents together with a set of variables
$\{e,e^{\ast}:e\in E^{1}\}$ which satisfy the following conditions:
(1) \ $s(e)e=e=er(e)$ for all $e\in E^{1}$.
(2) $r(e)e^{\ast}=e^{\ast}=e^{\ast}s(e)$\ for all $e\in E^{1}$.
(3) (The ``CK-1 relations") For all $e,f\in E^{1}$, $e^{\ast}e=r(e)$ and
$e^{\ast}f=0$ if $e\neq f$.
(4) (The ``CK-2 relations") For every regular vertex $v\in E^{0}$,
\[
v=\sum_{e\in E^{1},s(e)=v}ee^{\ast}.
\]
Every Leavitt path algebra $L_{K}(E)$ is a $\mathbb{Z}$\textit{-graded
algebra}, namely, $L_{K}(E)={\displaystyle\bigoplus\limits_{n\in\mathbb{Z}}}
L_{n}$ induced by defining, for all $v\in E^{0}$ and $e\in E^{1}$, $\deg
(v)=0$, $\deg(e)=1$, $\deg(e^{\ast})=-1$. Here the $L_{n}$ are abelian
subgroups satisfying $L_{m}L_{n}\subseteq L_{m+n}$ for all $m,n\in\mathbb{Z}
$. Further, for each $n\in\mathbb{Z}$, the \textit{homogeneous component
}$L_{n}$ is given by
\[
L_{n}=\{ {\textstyle\sum} k_{i}\alpha_{i}\beta_{i}^{\ast}\in L:\text{ }%
|\alpha_{i}|-|\beta_{i}|=n\}.
\]
Elements of $L_{n}$ are called \textit{homogeneous elements}. An ideal $I$ of
$L_{K}(E)$ is said to be a \textit{graded ideal} if $I=$
${\displaystyle\bigoplus\limits_{n\in\mathbb{Z}}} (I\cap L_{n})$. If $A,B$ are
graded modules over a graded ring $R$, we write $A\cong_{gr}B$ if $A$ and $B$
are graded isomorphic and we write $A\oplus_{gr}B$ to denote a graded direct
sum. We will also be using the usual grading of a matrix of finite order. For
this and for the various properties of graded rings and graded modules, we
refer to \cite{H-1}, \cite{HR} and \cite{NV}.
A \textit{breaking vertex }of a hereditary saturated subset $H$ is an infinite
emitter $w\in E^{0}\backslash H$ with the property that $0<|s^{-1}(w)\cap
r^{-1}(E^{0}\backslash H)|<\infty$. The set of all breaking vertices of $H$ is
denoted by $B_{H}$. For any $v\in B_{H}$, $v^{H}$ denotes the element
$v-\sum_{s(e)=v,r(e)\notin H}ee^{\ast}$. Given a hereditary saturated subset
$H$ and a subset $S\subseteq B_{H}$, $(H,S)$ is called an \textit{admissible
pair.} Given an admissible pair $(H,S)$, the ideal generated by $H\cup
\{v^{H}:v\in S\}$ is denoted by $I(H,S)$. It was shown in \cite{T} that the
graded ideals of $L_{K}(E)$ are precisely the ideals of the form $I(H,S)$ for
some admissible pair $(H,S)$. Moreover, $L_{K}(E)/I(H,S)\cong L_{K}%
(E\backslash(H,S))$. Here $E\backslash(H,S)$ is a \textit{Quotient graph of
}$E$ where $(E\backslash(H,S))^{0}=(E^{0}\backslash H)\cup\{v^{\prime}:v\in
B_{H}\backslash S\}$ and $(E\backslash(H,S))^{1}=\{e\in E^{1}:r(e)\notin
H\}\cup\{e^{\prime}:e\in E^{1} $ with $r(e)\in B_{H}\backslash S\}$ and $r,s$
are extended to $(E\backslash(H,S))^{0}$ by setting $s(e^{\prime})=s(e)$ and
$r(e^{\prime})=r(e)^{\prime}$. It is known (see \cite{R}) that if $P$ is a
prime ideal of $L$ with $P\cap E^{0}=H$, then $E^{0}\backslash H$ is downward directed.
Let $\Lambda$ be an infinite set and $R$, a ring. Then $M_{\Lambda}(R)$
denotes the ring of $\Lambda\times\Lambda$ matrices in which all except at
most finitely many entries are non-zero.
\bigskip
\section{Results}
\noindent In this section, we characterize Leavitt path algebras having
bounded index of nilpotence. We begin with the following useful proposition
some part of which might be implicit in earlier works on Leavitt path algebras.
\begin{prop}
\label{matrixCreation} Let $E$ be an arbitrary graph and let $L:=L_{K}(E)$.
\begin{enumerate}
[(a)]
\item Let $v$ be a vertex in $E$ which does not lie on a closed path. If, for
some $n\geq1$, there are $n$ distinct paths $p_{1},\cdot\cdot\cdot,p_{n}$ in
$E$ that end at $v$, then the set $T_{n}=\{{\displaystyle\sum\limits_{i=1}%
^{n}}{\displaystyle\sum\limits_{j=1}^{n}}k_{ij}p_{i}p_{j}^{\ast}:k_{ij}\in
K\}$ is a subring of $L$ isomorphic to the matrix ring $M_{n}(K)$.
\item Let $v$ be a vertex in $E$ lying on a cycle $c$ and let $f$ be an exit
for $c$ at $v$. Then, for every integer $n\geq1$, the subset
$S_{n}=\{{\displaystyle\sum\limits_{i=1}^{n}}$ ${\displaystyle\sum
\limits_{j=1}^{n}}k_{ij}c^{i}ff^{\ast}(c^{\ast})^{j}:k_{ij}\in K\}$ is a
subring of $L$ isomorphic to the matrix ring $M_{n}(K)$.
\item Let $v$ be a vertex lying on a cycle $c$ without exits in $E$. If, for
some $n\geq1$, there are $n$ distinct paths $p_{1},\cdot\cdot\cdot,p_{n}$ in
$E$ that end at $v$ and do not go through the entire cycle $c$, then again the
set $T_{n}=\{{\displaystyle\sum\limits_{i=1}^{n}}{\displaystyle\sum
\limits_{j=1}^{n}}k_{ij}p_{i}p_{j}^{\ast}:k_{ij}\in K\}$ is a subring of $L$
isomorphic to the matrix ring $M_{n}(K)$.
\end{enumerate}
\end{prop}
\begin{proof}
(a) First observe that $p_{j}^{\ast}p_{k}\neq0$ if and only if $p_{j}=p_{k}$.
Because, if $p_{j}^{\ast}p_{k}\neq0$, then either $p_{j}=p_{k}p^{\prime}$ or
$p_{k}=p_{j}q^{\prime}$ for some paths $p^{\prime},q^{\prime}$. Since
$r(p_{j})=r(p_{k})=v$, we get $s(p^{\prime})=v=r(p^{\prime})$ and
$s(q^{\prime})=v=r(q^{\prime})$. Since $v$ does not lie on a closed path, we
conclude that $p^{\prime}=v=q^{\prime}$. So $p_{j}=p_{k}$. Conversely, if
$p_{j}=p_{k}$, then clearly $p_{j}^{\ast}p_{k}=$ $p_{j}^{\ast}p_{j}=v\neq0$.
For all $i,j$, let $\varepsilon_{ij}=p_{i}p_{j}^{\ast}$. Clearly,
$(\varepsilon_{ii})^{2}=\varepsilon_{ii}$ and $\varepsilon_{ij}\varepsilon
_{kl}=\varepsilon_{il}$ or $0$ according as $j=k$ or not. Thus the
$\varepsilon_{ij}$ form a set of matrix units and it is readily seen that the
set $T_{n}=\{{\displaystyle\sum\limits_{i=1}^{n}}{\displaystyle\sum
\limits_{j=1}^{n}}k_{ij}p_{i}p_{j}^{\ast}:k_{ij}\in K\}$ is a subring of $L$
isomorphic to the matrix ring $M_{n}(K)$.
(b) Suppose $c$ is a cycle in $E$ with an exit $f$ at a vertex $v$. Consider
the set $\{\varepsilon_{ij}=c^{i}ff^{\ast}(c^{\ast})^{j}:1\leq i,j\leq n\}$.
Clearly, the $\varepsilon_{ij}$ form a set of matrix units as $(\varepsilon
_{ii})^{2}=\varepsilon_{ii}$ and $\varepsilon_{ij}\varepsilon_{kl}%
=\varepsilon_{il}$ or $0$ according as $j=k$ or not. It is then easy to check
that the set $S_{n}=\{{\displaystyle\sum\limits_{i=1}^{n}}$
${\displaystyle\sum\limits_{j=1}^{n}}k_{ij}c^{i}ff^{\ast}(c^{\ast})^{j}%
:k_{ij}\in K\}$ is a subring of $L$ isomorphic to the matrix ring $M_{n}(K)$.
(c) Let $v$ be the base of a cycle $c$ without exits and $p_{1},\cdot
\cdot\cdot,p_{n}$ be $n$ distinct paths that end at $v$ and not go through the
entire cycle $c$. Using the fact that $c$ is a cycle without exits and
repeating the arguments as in (a), it follows that $p_{j}^{\ast}p_{k}\neq0$ if
and only if $p_{j}=p_{k}$. As before let $\varepsilon_{ij}=p_{i}p_{j}^{\ast}$
with $1\leq i,j\leq n$. Clearly, $(\varepsilon_{ii})^{2}=\varepsilon_{ii}$ and
$\varepsilon_{ij}\varepsilon_{kl}=\varepsilon_{il}$ or $0$ according as $j=k$
or not. Thus the $\varepsilon_{ij}$ form a set of matrix units and it is
readily seen that $T_{n}=\{{\displaystyle\sum\limits_{i=1}^{n}}%
{\displaystyle\sum\limits_{j=1}^{n}}k_{ij}p_{i}p_{j}^{\ast}:k_{ij}\in K\}$ is
a subring of $L$ isomorphic to the matrix ring $M_{n}(K)$.
\end{proof}
We are now ready to describe all the Leavitt path algebras with bounded index
of nilpotence. It is interesting to note that the Leavitt path algebras having
bounded index of nilpotence are precisely those that satisfy a polynomial identity.
Recall that an algebra $A$ over a field $K$ is said to satisfy a polynomial
identity if there exists a non-zero element $f$ in $K[x_{1},\ldots,x_{n}]$
such that $f(a_{1},\ldots,a_{n})=0$ for all $a_{i}$ in $A$. Clearly every
commutative ring satisfies a polynomial identity but there are many
interesting classes of noncommutative rings too that satisfy a polynomial
identity.\ For instance, the Amitsur-Levitzky theorem (see \cite{P}) states
that, for any $n\geq1$, the matrix ring $M_{n}(R)$ over a commutative ring $R$
satisfies a polynomial identity of degree $2n$. In \cite{BLR} it is shown
that the Leavitt path algebra $L_{K}(E)$ of an arbitrary graph $E$ over a
field $K$ satisfies a polynomial identity if and only if no cycle in $E$ has
an exit and there is a fixed positive integer $d$ such that the number of
distinct paths that end at any given vertex $v$ (including $v$, but not
including the entire cycle $c$ in case $v$ lies on $c$) is less than or equal
to $d$. When $E$ is a finite graph, then the Leavitt path algebra $L_{K}(E)$
satisfying a polynomial identity is known to be equivalent to the
Gelfand-Kirillov dimension of $L_{K}(E)$ being at most one \cite{BLR}.
\begin{theorem}
\label{bdd Index} Let $E$ be an arbitrary graph. Then the following properties
are equivalent for $L:=L_{K}(E)$:
\begin{enumerate}
\item $L$ has index of nilpotence less than or equal to $n$;
\item No cycle in $E$ has an exit and there is a fixed positive integer $n$
such that the number of distinct paths that end at any given vertex $v$
(including $v$, but not including the entire cycle $c$ in case $v$ lies on
$c$) is less than or equal to $n$;
\item For any graded prime ideal $P$ of $L$, $L/P\cong_{gr}M_{t}(K)$ or
$M_{t}(K[x,x^{-1}])$ where $t\leq n$ with appropriate matrix gradings;
\item $L$ is a graded subdirect product of graded rings $\{A_{i}:i\in I\}$
where, for each $i$, $A_{i}\cong_{gr}M_{t_{i}}(K)$ or $M_{t_{i}}(K[x,x^{-1}])
$ with appropriate matrix gradings where, for each $i$, $t_{i}\leq n$, a fixed
positive integer.
\item $L$ satisfies a polynomial identity.
\end{enumerate}
\end{theorem}
\begin{proof}
Assume (i), that is, assume that the index of nilpotence of $L$ is $\leq n$.
We claim that no cycle in $E$ can have an exit. Because, otherwise, by
Proposition \ref{matrixCreation} (b), $L$ will contain subrings of matrices of
arbitrary finite size and this will give rise to unbounded index of nilpotence
for $L$. Thus every vertex $v$ in $E$ either does not lie on a closed path or
lies on a cycle without exits. If there are more than $n$ distinct paths
ending at $v$, then again, by Proposition \ref{matrixCreation} (a) and (c),
$L$ will contain a copy of a matrix ring of order greater than $n$ over $K$
which will imply that the index of nilpotence of $L$ is greater than $n$, a
contradiction. This proves (ii).
Assume (ii). Let $P=I(H,S)$ be a graded prime ideal of $L$. Our hypothesis
implies that no cycle in $E\backslash(H,S)$ has an exit and that $n$ is also
the upper bound for the number of distinct paths ending at any vertex in
$E\backslash(H,S)$. So $E\backslash(H,S)$ contains no infinite irrational
paths. This means that every path ends at a sink or at a cycle without exits.
Also, as $I(H,S)$ is a graded prime ideal, Theorem 3.12 of \cite{R} implies
that $(E\backslash(H,S))^{0}$ is downward directed. Consequently,
$E\backslash(H,S)$ contains either (a) exactly one sink $w$ or (b) exactly one
cycle $c$ without exits based at a vertex $v$. Now in case (a), there are no
more than $n$ distinct paths ending at $w$ and, in case (b), there are no
more than $n$ paths which end at $v$ and do not go through the cycle $c$. We
then appeal to Corollary 2.6.5 and Lemma 2.7.1 of \cite{AArS} to conclude that
$L/P\cong L_{K}(E\backslash(H,S))\cong M_{t}(K)$ or $M_{t}(K[x,x^{-1}])$
according as $E\backslash(H,S)$ contains a sink or a cycle without exits. This
proves (iii).
Assume (iii). Now, for any graded prime ideal $P$, $L/P\cong_{gr}M_{t}(K)$ or
$M_{t}(K[x,x^{-1}])$ with appropriate matrix gradings where $t\leq n$, a fixed
positive integer. It is known that the intersection $I$ of all graded prime
ideals of $L$ is zero. For the sake of completeness, we shall outline the
argument. If $I\neq0$, being a graded ideal, $I$ the will contain vertices.
But, given any vertex $v$, a graded ideal $M$ maximal among graded ideals with
respect to $v\notin M$ is a graded prime ideal, because, for any two
homogeneous elements $a,b$, if $a\notin M$ and $b\notin M$, then $v\in M+LaL$
and $v\in M+LbL$. Consequently, $v=v^{2}\in(M+LaL)(M+LbL)=M+LaLbL$. Since
$v\notin M$, $aLb\nsubseteqq M$. Thus $M$ is a graded prime ideal. But then
$v\notin M$ implies $v\notin I$, a contradiction. Thus $\cap\{P:P$ graded
prime ideal$\}=0$. Consequently, $L$ is a graded subdirect product of the
graded rings $L/P$ graded isomorphic to $M_{t}(K)$ or $M_{t}(K[x,x^{-1}])$
under appropriate matrix gradings, where $t\leq n$, a fixed positive integer.
This proves (iv).
Assume (iv). If $t\leq n$, then the matrix rings $M_{t}(K)$ and $M_{t}%
(K[x,x^{-1}])$ will each have index of nilpotence $\leq n$. Consequently, a
subdirect product of such rings will also have nilpotence index $\leq n$. This
proves (i).
Assume (iv). By the Amitsur-Levitzky theorem \cite{P}, both $M_{t}(K)$ and
$M_{t}(K[x,x^{-1}])$ with $t\leq n$ are polynomial identity rings satisfying a
polynomial identity of degree $\leq2n$. From this, it is clear that the
subdirect product $L$ also satisfies a polynomial identity of degree $\leq n$.
This proves (v).
The implication (v) $\implies$ (ii) has been established in \cite{BLR}.
\end{proof}
\noindent The Leavitt path algebra in Theorem \ref{bdd Index} need not
decompose as a direct sum of matrix rings, as the following example shows.
\begin{example}
\textrm{\label{InfiniteClock} Consider the following \textquotedblleft
infinite clock" graph $E$:%
\[%
\begin{array}
[c]{ccccc}
& & \bullet_{w_{1}} & & \bullet_{w_{2}}\\
& \ddots & \uparrow & \nearrow & \\
& \cdots & \bullet_{v} & \longrightarrow & \bullet_{w_{3}}\\
& \swarrow & \vdots & \ddots & \\
_{w_{n}} & & & &
\end{array}
\]
\noindent Thus $E^{0}=\{v\}\cup\{w_{1},w_{2},\cdot\cdot\cdot,w_{n},\cdot
\cdot\cdot\}$ where the $w_{i}$ are all sinks. For each $n\geq1$, let $e_{n}$
denote the single edge connecting $v$ to $w_{n}$. The graph $E$ is acyclic and
so every ideal of $L$ is graded (\cite{HR}). The number of distinct paths
ending at any given sink \ (including the sink) is $\leq2$. For each $n\geq1$,
$H_{n}=\{w_{i}:i\neq n\}$ is a hereditary saturated set, $B_{H_{n}}=\{v\}$ and
($E\backslash(H_{n},B_{H_{n}}))^{0}=\{v,w_{n}\}$ is downward directed. Also E%
$\backslash$%
(H$_{n}$,B$_{H_{n}}$) trivially satisfies Condition (L). Hence the ideal
$P_{n}$ generated by $H_{n}\cup\{v-e_{n}e_{n}^{\ast}\}$ is a\ graded primitive
ideal by Theorem 4.3(iii) of \cite{R} and $L_{K}(E)/P\cong M_{2}(K)$.
Moreover, every graded primitive (equivalently, prime) ideal $P$ of $L_{K}(E)$
is equal to $P_{n}$ for some $n$. By \cite[Theorem 4.12]{HRS-1}, $L_{K}(E)$ is
a graded $\Sigma$-$V$ ring. }
\textrm{But $L_{K}(E)$ cannot decompose as a direct sum of the matrix rings
$M_{2}(K) $. Because, otherwise, $v$ would lie in a direct sum of finitely
many copies of $M_{2}(K)$. Since the ideal generated by $v$ is $L_{K}(E)$,
$L_{K}(E)$ will then be a direct sum of finitely many copies of $M_{2}(K)$.
This is impossible since $L_{K}(E)$ contains an infinite set of orthogonal
idempotents $\{e_{n}e_{n}^{\ast}:n\geq1\}$. }
\textrm{We can also describe the internal structure of this ring $L_{K}(E)$.
The socle $S$ of $L_{K}(E)$ is the ideal generated by the sinks $\{w_{i}%
:i\geq1\}$, $S\cong\bigoplus\limits_{\aleph_{0}}M_{2}(K)$ and $L_{K}(E)/S\cong
K$. }
\end{example}
But the decomposition is possible if the graph is row-finite, as shown in the
following theorem.
\begin{theorem}
\label{row-finite bddIndex}Let $E$ be a row-finite graph. Then the following
properties are equivalent for $L:=L_{K}(E)$:
\begin{enumerate}
[(i)]
\item $L$ has bounded index of nilpotence $\leq n$;
\item There is a fixed positive integer $n$ and a graded isomorphism%
\[
L\cong_{gr}{\displaystyle\bigoplus\limits_{i\in I}}M_{n_{i}}(K)\oplus
{\displaystyle\bigoplus\limits_{j\in J}}M_{n_{j}}(K[x,x^{-1}])
\]
where $I,J$ are arbitrary index sets and, for all $i\in I$ and $j\in J$,
$n_{i}\leq n$ and $n_{j}\leq n$ . In particular, $L$ is graded semi-simple
(that is, a direct sum of graded simple left/right ideals).
\end{enumerate}
\end{theorem}
\begin{proof}
Assume (i). By Theorem \ref{bdd Index}, no cycle in $E$ has an exit and the
number of distinct paths that end at any vertex $v$ is $\leq n$, with the
proviso that if $v$ sits on a cycle $c$, then these paths do not include the
entire cycle $c$. If $A$ is the graded ideal generated by all the sinks in $E$
and all the vertices on cycles without exits, then, by Corollary 2.6.5 and
Lemma 2.7.1 of \cite{AArS}, $A\cong{\displaystyle\bigoplus\limits_{i\in I}%
}M_{n_{i}}(K)\oplus{\displaystyle\bigoplus\limits_{j\in J}}M_{n_{j}%
}(K[x,x^{-1}])$. By giving appropriate matrix gradings, this isomorphism
becomes a graded isomorphism. We claim that $L=A$. Let $H\subseteq A$ be the
set consisting of all the sinks and all the vertices on cycles in $E$. By
hypothesis, every path in $E$ that does not include an entire cycle has length
$\leq n$ and ends at a vertex in $H$. So if $u$ is any vertex in $E$, using
the fact that all the vertices in $E$ are regular and by a simple induction on
the length of the longest path from $u$, we can conclude that $u$ belongs to
the saturated closure of $H$. This implies that $L=A\cong_{gr}%
{\displaystyle\bigoplus\limits_{i\in I}}M_{n_{i}}(K)\oplus
{\displaystyle\bigoplus\limits_{j\in J}}M_{n_{j}}(K[x,x^{-1}])$. This proves (ii).
(ii) $\implies$ (i) follows from the fact that the matrix rings $M_{n_{i}}(K)
$ and $M_{n_{j}}(K[x,x^{-1}])$ with $n_{i},n_{j}\leq n$ have index of
nilpotence $\leq n$.
\end{proof}
One consequence of Theorem \ref{bdd Index} is the following.
\begin{prop}
\label{bdx => SigmaV} Let $E$ be an arbitrary graph. If $L:=L_{K}(E)$ has
bounded index of nilpotence $n$, then $L$ is a graded $\Sigma$-$V$ ring.
\end{prop}
\begin{proof}
If $L$ has bounded index of nilpotence $n$, then for any graded primitive
ideal $P$ of $L$ (since it is also graded prime), we have, by Theorem
\ref{bdd Index}(iii), $L/P\cong_{gr}M_{t}(K)$ or $M_{t}(K[x,x^{-1}])$ with
appropriate matrix gradings, where $t\leq n$. By \cite[Theorem 4.12]{HRS-1},
we then conclude that $L$ is a graded $\Sigma$-$V$ ring.
\end{proof}
The converse of the above result does not hold, as can be seen in the two
examples below.
\begin{example}
\textrm{\label{Inverse infinite clock} Let $E$ be the \textquotedblleft
inverse infinite clock" graph consisting of a sink $w$ and countably infinite
edges $\{e_{n}:n\geq1\}$ with $r(e_{n})=w$ and s(e$_{n}$) = w}$_{n}$ \textrm{
for all $n$. }
\textrm{%
\[%
\begin{array}
[c]{ccccc}
& & \bullet_{w_{1}} & & \bullet_{w_{2}}\\
& \ddots & \downarrow & \swarrow & \\
& \cdots & \bullet_{w} & \longleftarrow & \bullet_{w_{3}}\\
& \nearrow & \vdots & \ddots & \\
\bullet_{w_{n}} & & & &
\end{array}
\]
} \textrm{\noindent Then $L_{K}(E)\cong M_{\infty}(K)$, the infinite
$\omega\times\omega$ matrix with finitely many non-zero entries. Now, under
appropriate matrix grading, $M_{\infty}(K)$ is graded semisimple (that is, a
graded direct sum of \ graded simple modules) and so all graded left/right
$M_{\infty}(K)$-modules are graded injective and hence $L$ is a graded
$\Sigma$-$V$ ring. But $L$ does not have bounded index of nilpotence by
Theorem \ref{bdd Index}(ii). }
\end{example}
\begin{example}
\textrm{\label{bdd idx no sigmaV} Consider the following graph $F$ consisting
of two cycles $g$ and $c$ connected by an edge:
\[%
\begin{array}
[c]{ccccccc}%
\bullet & \longrightarrow & \bullet & & \bullet & \longrightarrow & \bullet\\
\uparrow & g & \downarrow & & \uparrow & c & \downarrow\\
\bullet & \longleftarrow & \bullet & \longrightarrow & \bullet_{v} &
\longleftarrow & \bullet
\end{array}
\]
} \textrm{\noindent Now $F$ is downward directed, $c$ is a cycle without exits
and the various powers of the cycle $g$ give rise to infinitely many distinct
paths that end at the base $v$ of the cycle $c$. Hence $L_{K}(F)\cong
M_{\infty}(K[x,x^{-1}])$ (by Lemma 2.7.1 of \cite{AArS}) and is graded
semisimple. Hence each graded simple module over $L_{K}(F)$ is graded $\Sigma
$-injective. But $L_{K}(F)$ does not have bounded index of nilpotence, as
$M_{\infty}(K[x,x^{-1}])$ contains subrings isomorphic to $M_{n}(K[x,x^{-1}])$
for every positive integer $n$. }
\end{example}
In the monograph \cite{JST}, the following open question (6.33: Problem 3,
Chapter 6) was raised:
\begin{question}
Does every exchange right $\Sigma$-$V$ ring have bounded index of nilpotence?
\end{question}
We answer this question in the negative in the following remark.
\begin{remark}
\textrm{Consider the graph $E$ of Example \ref{Inverse infinite clock}. As
noted there, $L_{K}(E)\cong M_{\infty}(K)$ $\mathrm{\ }$which is semisimple
and hence is a $\Sigma$-$V$ ring. Since $E$ is acyclic, $L_{K}(E)$ is von
Neumann regular \cite{AR} and hence is an exchange ring. But $L_{K}(E)$ does
not have bounded index of nilpotence by Theorem \ref{bdd Index}(ii). }
\end{remark}
\begin{remark}
\textrm{It was shown in \cite{HRS-1} that a Leavitt path algebra $L_{K}(E)$ is
directly-finite (equivalently, graded directly-finite with respect to vertices) if and only if no cycle in $E$ has an exit. In view of Theorem
\ref{bdd Index}, it is clear that a Leavitt path algebra of bounded index is
always directly-finite. But, for arbitrary rings with bounded index of
nilpotence, this is not the case. Let $S=\prod R_{k}$, where each $R_{k}%
\cong\mathbb{Z}(p^{n})$, the ring of integers modulo a fixed integer $n\geq2$.
Now $(pS)^{n}=0$ and if $a\in S$ is nilpotent, then $a\in pS$ and $a^{n}=0$.
Consequently, $S$ has bounded index of nilpotence. In fact, the index of
nilpotence of $S$ is $n$. But $S$ is not directly-finite, since $S\cong%
\prod\limits_{k\geq2}R_{k}\cong\prod\limits_{k\geq3}R_{k}\cong\cdot\cdot\cdot
$. }
\textrm{Conversely, a directly-finite Leavitt path algebra need not have
bounded index of nilpotence. If $E$ is the graph consisting of an infinite
line segment
\[
\bullet\longrightarrow\bullet\longrightarrow\bullet\longrightarrow\cdot
\cdot\cdot\bullet\longrightarrow\cdot\cdot\cdot
\]
then clearly $L_{K}(E)$ is directly-finite, but, by Theorem \ref{bdd Index}
(ii), $L_{K}(E)\cong M_{\infty}(K)$ does not have bounded index of nilpotence.
}
\end{remark}
\bigskip
\bigskip
\bigskip
|
\section{Introduction}
Combinatorial optimization problems that have many applications can be translated into problems to find ground states of the Ising model~\cite{A.Lucas2014}.
This formulation motivates us to develope machines speciallizing in the search for the ground states.
A well-known example of such machines is the hardware devices provided by D-Wave Systems Inc.~\cite{M.Johnson2011}.
The devices implement quantum annealing~\cite{T.Kadowaki1998, J.Brooke1999, G.Santoro2002, A.Das2008}
(or adiabatic quantum computation~\cite{E.Farhi2000, T.Albash2018}),
a heuristic which harnesses quantum effects,
instead of themal effects in simulated annealing~\cite{S.Kirkpatrick1983},
to search for the ground states.
It has had an impact to impelement quantum annealing directly using physics of artificial spins.
Other machines for such a purpose,
solving problems in terms of the Ising model,
have also been proposed and actually developed,
which utilize, or are inspired by, interaction and dynamics in underlying physical phenomena~\cite{J.Britton2012, A.Marandi2014, T.Inagaki2016Oct, P.McMahon2016, I.Mahboob2016, M.Yamaoka2016, H.Goto2016, S.Nigg2017, S.Puri2017, B.Sutton2017, S.Tsukamoto2017, O.Kyriienko2018arXiv}.
The coherent Ising machine (CIM) is such a machine
based on network of degenerate optical parametric oscillators (DOPOs)~\cite{Z.Wang2013, A.Marandi2014, T.Inagaki2016Oct, P.McMahon2016}.
The degree of freedom we utilize as an Ising spin is phase of the signal field of a DOPO.
The signal field is amplified by the pump field via interaction in a nonlinear optical crystal in a cavity~\cite{A.Yariv2006, H.Carmichael2002, *H.Carmichael2008}.
Above the threshold of the pump field,
the phase of the signal field bifurcates.
The phase difference from the pump field takes either 0 or $\pi$ at random,
which encodes an Ising spin for the CIM.
Interaction of DOPOs makes correlation in their phases~\cite{A.Marandi2012Mar} as interaction of Ising spins.
Network composed of configurable interactions of DOPOs with the pump field around the threshold is thus expected to represent the lowest energy states of the corresponding Ising model~\cite{Z.Wang2013}.
Such network in the CIM was experimentally constructed with a system of time-multiplexed DOPO pulses in a ring cavity~\cite{A.Marandi2014, T.Inagaki2016Jun, K.Takata2016}.
Optical coupling of the pulses is realized by delay lines connected to the main ring.
The length of each delay line is tuned to be an integer multiple of the pulse-repetition period so that a fraction of a pulse taking a detour via a delay line interacts to another one running after it.
This system almost surely found the ground states of some Ising models~\cite{A.Marandi2014, T.Inagaki2016Jun, K.Takata2016},
but connection of spins is limited because of difficulty in making arbitrary graph structure under this scheme.
For instance,
the regular graphs of degree-$k$ need $k$ delay lines.
To deal with this issue,
the CIM with aid of field-programmable gate array (FPGA) has been developed,
where FPGA manages the coupling of the pulses running in a ring cavity~\cite{T.Inagaki2016Oct, P.McMahon2016}.
This type of CIM can treat up to 2000 Ising spins with all-to-all couplings,
exhibiting faster convergence to states for comparable or lower Ising energy than finely tuned simulated annealing running on a CPU.
Exploiting FPGA,
the CIM does not show extreme decrease in the performance even for problems with fully-connected graphs.
This feature contributes to the CIM's advantage
when we compare the CIM with the machines of D-Wave systems Inc.~\cite{R.Hamerly2018arXiv, C.McGeoch2018arXiv}
adopting the so-called ``Chimera graph''~\cite{P.Bunyk2014, V.Choi2008, V.Choi2011, J.Cai2014arXiv}.
On the other hand,
under the use of FPGA,
the effect of dynamics of the DOPOs for the performance is murky~\cite{A.King2018arXiv}.
It is also obscure whether properties of DOPO, in particular, as quantum light are exploited to accelerate finding the solution.
We have needed theoretical description of the CIM to estimate its efficiency to solve Ising problems and also to clarify the dominant property for the efficiency.
Dynamics in the CIM based on quantum mechanics has been investigated~\cite{K.Takata2015, D.Maruo2016, T.Shoji2017, A.Yamamura2017}.
The detailed analysis, however, sufferred from huge comutational cost,
and the system size was limited.
The CIM for large-size problems has been numerically simulated by the semiclassical counterpart~\cite{T.Inagaki2016Oct, P.McMahon2016, Y.Haribara2017}.
The numerical simulation is a powerful tool to gain insight to the CIM.
We, however, cannot conclude the performance of the CIM in general on the basis of behavior for particular instances observed with the numerical simulations.
For estimating the efficiency of the CIM without the dynamics simulations,
it is a good strategy to find a factor
that deteremines whether the CIM could find the correct solutions of a problem and
that scales the computaion time required to solve it.
For quantum annealing, such a factor is the minimal energy gap between the ground state and the first excited one of problem Hamiltonian in an annealing passage~\cite{T.Albash2018}.
The problem-size dependence of the minimal energy gap is often used to represent the difficulty of the problem for quantum annealing.
This formulation is generalized into open quantum systems,
where the gap of eigenvalues of Liouvillian, instead of Hamiltonian, takes the role~\cite{L.Venuti2016}.
Considering the CIM as an annealing machine,
we would obtain the factor along the argument similar to that for other annealing algorithms~\cite{S.Geman1984, H.Nishimori1998, T.Albash2018, L.Venuti2016}.
To this end,
we first need the instantaneous steady state or its distributions of the CIM for the parameters.
We additionally have to check that there exist values of the parameters
for which the steady state gives ground states of the target Ising model,
since the cost function in the CIM,
presented below,
does not agree with the Hamiltonian of the Ising model.
Then we will move to a stage at which
we discuss the factor and also the schedule to surely reach the optimal state for the problem by adiabatical evolution.
In this paper,
we explore the steady state distributions of a model of the CIM.
Our aim is to clarify properties of a large number of DOPOs in solving combinatorial optimization problems, in particular, in the long time limit,
where the steady state distributions are possibly realized.
The model does not include any component corresponding to the FPGA and the delay lines
but is just simple network of DOPOs optically interacting with each other.
We deduce the Fokker-Planck equation
describing the dynamics of DOPOs in the network.
Under some ansatz,
we show the approximate steady state distribution.
Statistical-mechanical approach enables us to investigate the distributions for large-size problems.
We then examine the distributions and the most probable states for rather simple problems.
\section{Model}
We investigate network of DOPOs interacting with each other via mutual-injection paths as a theoretical model of the CIMs.
Our model is a generalization of the model for two DOPOs previously proposed~\cite{K.Takata2015},
which is also a generalization of the single DOPO model~\cite{P.Drummond1981} by adding the interaction of the DOPOs via a mutual injection path.
In the two-DOPO model the signal fields of a frequency $\omega$,
which are used to represent Ising variables later,
are assumed to be highly confined in each cavity and the path.
The spatial-phase factor $e^{ikd}$ of the bosonic operator in the injection path is considered,
where $k$ is the wave number for the signal mode,
and $d$ is the length of the path.
The mutual injection leads to in-phase couplings, i.e., the ferromagnetic coupling in terms of spin systems, if $e^{ikd} = 1$ and out of phase, i.e., the antiferromagnetic one, if $e^{ikd} = -1$.
The signal field is amplified in a nonlinear crystal via the interaction with the pump mode of a frequency $\omega_p = 2\omega$.
The pump mode is excited by the classical, driving field entering each DOPO.
The driving field is also used as the phase reference.
Above the threshold of the pump strength,
the bifurcation of quadrature amplitude of each signal field is observed.
The sign of the quadrature amplitude encodes an Ising spin.
The DOPOs output the configuration of Ising spins according to their coupling in which the target Ising Hamiltonian is embedded.
We generalize the two-DOPO model into a system of $N$ DOPOs.
We assume that each pair of two DOPOs in the system interacts as in the above two-DOPO model.
While there exists similar generalization to the one-dimensional ring network consisting of $N$ DOPOs~\cite{D.Maruo2016},
we here treat almost arbitrary network as well as the implementation of the Zeeman term in the Ising Hamiltonian.
The Hamiltonian for our model is written as
\begin{equation}
H = \sum_{j=1}^N H^{(j)}_\text{DOPO} + \sum_{j<l} H^{(jl)}_\text{int} + \sum_{j=1}^N H^{(j)}_{Z} + H_\text{res}.
\end{equation}
The first sum includes~\cite{P.Drummond1981}
\begin{equation}
\begin{split}
H^{(j)}_\text{DOPO}
=& \hbar\omega \hat{a}^{\dagger}_{sj} \hat{a}_{sj} + 2\hbar\omega \hat{a}^{\dagger}_{pj} \hat{a}_{pj}
+ i\hbar \frac{\kappa}{2} \left( \hat{a}^{\dagger 2}_{sj} \hat{a}_{pj} - \hat{a}^{\dagger}_{pj} \hat{a}^2_{sj}\right) \\
& + i\hbar \left( \epsilon_p \hat{a}^{\dagger}_{pj} e^{-2i\omega t} - \epsilon_p \hat{a}_{pj} e^{2i\omega t} \right),
\end{split}
\end{equation}
where $\hat{a}_{sj}$ and $\hat{a}^{\dagger}_{sj}$ are the bosonic annihilation and creation operators, respectively,
for the signal modes $j = 1, 2, \dots , N$,
and $\hat{a}_{pj}$ and $\hat{a}^{\dagger}_{pj}$ are for the pump modes $j$.
The coupling constant of quadratic nonlinear interaction of the signal and pump modes is denoted by $\kappa$.
The pump mode is excited by the real driving field $\epsilon_p$ of a frequency $\omega_d = \omega_p = 2\omega$.
The Ising variable for discrete optimization problems is encoded in the sign of the in-phase amplitude of the signal mode $(\hat{a}_{sj}+\hat{a}_{sj}^\dagger)/2$~\cite{T.Inagaki2016Oct, P.McMahon2016, Z.Wang2013, A.Marandi2014, K.Takata2015},
which can be observed via homodyne detection.
We consider the beam-splitter interaction Hamiltonian between the signal modes in the cavity and the injection path
to tune the interactions between different DOPOs,
\begin{equation}
\begin{split}
H^{(jl)}_\text{int}
= \hbar\omega \hat{a}^{\dagger}_{cjl} \hat{a}_{cjl}
+ & i\hbar \zeta_{jl}\Big( \hat{a}_{cjl}\hat{a}_{sj}^\dagger - \hat{a}_{cjl}^\dagger\hat{a}_{sj} \\
&+ \hat{a}_{sl}\hat{a}_{cjl}^\dagger e^{-i\theta_{jl}} - \hat{a}_{sl}^\dagger\hat{a}_{cjl}e^{i\theta_{jl}} \Big),
\end{split}
\end{equation}
where the signal modes in the injection paths for $j$ and $l$ are denoted by $\hat{a}_{cjl}$ and $\hat{a}^{\dagger}_{cjl}$,
and $\zeta_{jl}$ denotes the interaction coefficient of the signal modes and the injection-path mode for the path between cavities $j$ and $l$.
Phase $\theta_{jl}$ is equal to $kd_{jl}$,
where $d_{jl}$ is the path length between the cavities.
Hamiltonian $H^{(j)}_Z$ is for excitation of the signal mode by the real field $\epsilon_s$ of a frequency $\omega$ to tune the effect for the Zeeman term in the Ising Hamiltonian,
\begin{equation}
H^{(j)}_Z = i\hbar \left(\epsilon_{sj}\hat{a}_{sj}^\dagger e^{-i\omega t} - \epsilon_{sj}\hat{a}_{sj}e^{i\omega t}\right).
\end{equation}
We do not explicitly show $H_\text{res}$,
which is a standard one for interaction with surroundings (reserviors)~\cite{H.Carmichael2002, P.Drummond1981, K.Takata2015}.
The master equation for the density operator $\hat{\rho}$ for the system,
where the degrees of freedom of the reseviors are traced out,
is obtained under standard approximations introduced to treat the reserviors~\cite{H.Carmichael2002, *H.Carmichael2008, H.Breuer2007};
the Born-Markov approximation and neglecting implicit interactions of the reserviors through the internal couplings in the system.
We set the reservior at zero temperature to eliminate thermal effects.
Accoridingly, noises in dynamics will be derived only from quantum fluctuations.
The resulting master equation is
\begin{equation}
\begin{split}
\frac{d\hat{\rho}}{dt}
= & \frac{1}{i\hbar}\left[ \sum_{j=1}^N H^{(j)}_\text{DOPO} + \sum_{j<l} H^{(jl)}_\text{int} + \sum_{j=1}^N H^{(j)}_{Z}, \hat{\rho} \right] \\
&+ \sum_{j=1}^N 2\gamma_s \left( \hat{a}_{sj}\hat{\rho}\hat{a}_{sj}^\dagger -\frac{1}{2} \left\{ \hat{a}_{sj}^\dagger\hat{a}_{sj}, \hat{\rho} \right\} \right) \\
&+ \sum_{j=1}^N 2\gamma_p \left( \hat{a}_{pj}\hat{\rho}\hat{a}_{pj}^\dagger -\frac{1}{2} \left\{ \hat{a}_{pj}^\dagger\hat{a}_{pj}, \hat{\rho} \right\} \right) \\
&+ \sum_{j<l}^N 2\gamma_c \left( \hat{a}_{cjl}\hat{\rho}\hat{a}_{cjl}^\dagger -\frac{1}{2} \left\{ \hat{a}_{cjl}^\dagger\hat{a}_{cjl}, \hat{\rho} \right\} \right),
\end{split} \label{eq:master}
\end{equation}
where $\gamma_s$, $\gamma_p$, and $\gamma_c$ are coefficients for the decay of the signal, pump, and injection-path modes through dissipation, respectively.
We here utilize the positive $P$ representation~\cite{P.Drummond1980Jan} to analyze the master equation.
The density operator in the positive $P$ representaiton is expanded in terms of the coherent product states and a distribution function $P(\bm{\alpha}, \bm{\beta})$ as
\begin{equation}
\hat{\rho} = \int d^{4N}\bm{\alpha}d^{4N}\bm{\beta} P(\bm{\alpha},\bm{\beta}) \frac{\ket{\bm{\alpha}}\bra{\bm{\beta}^*}}{\langle\bm{\beta}^*|\bm{\alpha}\rangle}. \label{eq:def_positive-P}
\end{equation}
The $c$-number vector $\bm{\alpha}$ composed of $\alpha_{sj}$, $\alpha_{pj}$, and $\alpha_{cjl}$ for $j, l=1, \dots , N$ and $j<l$ represents the coherent product state $\ket{\bm{\alpha}} = \prod_{j=1}^N \ket{\alpha_{sj}}\ket{\alpha_{pj}}\prod_{j<l}\ket{\alpha_{cjl}}$,
and $\bm{\beta}$ describes another one $\bra{\bm{\beta}^*} = \prod_{j=1}^N \bra{\beta^*_{sj}}\bra{\beta^*_{pj}}\prod_{j<l}\bra{\beta^*_{cjl}}$.
The distribution function $P(\bm{\alpha}, \bm{\beta})$ itself in the positive $P$ representation is not uniquely determined,
but the normal-ordered average is calculated with any distribution $P(\bm{\alpha}, \bm{\beta})$ that satisfies Eq.~(\ref{eq:def_positive-P}).
The nonuniqueness allows the distribution to be real and positive~\cite{P.Drummond1980Jan, H.Carmichael2002, *H.Carmichael2008},
even when the density operator is composed of the superposition of different coherent states.
In this expression the average of in-phase amplitude $(\hat{a}_{sj} + \hat{a}_{sj}^\dagger)/2$ is computed as the average of $(\alpha_j + \beta_j)/2$ over the distribution $P(\bm{\alpha},\bm{\beta})$.
Substituting Eq.~(\ref{eq:def_positive-P}) into Eq.~(\ref{eq:master}),
we obtain the Fokker-Planck equation for the distribution $P(\bm{\alpha},\bm{\beta})$
through typical calculations for this representation~\cite{H.Carmichael2002, *H.Carmichael2008, K.Takata2015},
\begin{widetext}
\begin{equation}
\begin{split}
\frac{dP(\bm{\alpha},\bm{\beta})}{dt}
= \Bigg\{ \sum_{j=1}^N & \Bigg[ \frac{\partial}{\partial \alpha_{sj}} \left(\gamma_s \alpha_{sj} - \kappa\beta_{sj}\alpha_{pj} -\sum_{l(>j)}\zeta_{jl}\alpha_{cjl} + \sum_{l(<j)}\zeta_{lj}\alpha_{clj}e^{i\theta_{lj}} -\epsilon_{sj} \right) \\
&+ \frac{\partial}{\partial \beta_{sj}} \left(\gamma_s \beta_{sj} - \kappa\alpha_{sj}\beta_{pj} -\sum_{l(>j)}\zeta_{jl}\beta_{cjl} + \sum_{l(<j)}\zeta_{lj}\beta_{clj}e^{-i\theta_{lj}} -\epsilon_{sj} \right) \\
& + \frac{1}{2}\frac{\partial^2}{\partial\alpha_{sj}^2} \kappa \alpha_{pj} + \frac{1}{2}\frac{\partial^2}{\partial\beta_{sj}^2} \kappa \beta_{pj}
+ \frac{\partial}{\partial\alpha_{pj}} \left(\gamma_p\alpha_{pj}-\epsilon_p+\frac{\kappa}{2}\alpha_{sj}^2\right)
+ \frac{\partial}{\partial\beta_{pj}} \left(\gamma_p\beta_{pj}-\epsilon_p+\frac{\kappa}{2}\beta_{sj}^2\right) \Bigg] \\
+ \sum_{j<l} & \left\{ \frac{\partial}{\partial\alpha_{cjl}} \left[ \gamma_c\alpha_{cjl} + \zeta_{jl}\left(\alpha_{sj}-\alpha_{sl}e^{-i\theta_{jl}}\right)\right]+ \frac{\partial}{\partial\beta_{cjl}} \left[\gamma_c\beta_{cjl} + \zeta_{jl}\left(\beta_{sj}-\beta_{sl}e^{i\theta_{jl}}\right)\right] \right\} \Bigg\} P(\bm{\alpha},\bm{\beta})
\end{split}
\end{equation}
\end{widetext}
where we have taken the rotating frame with $\omega$ for the signal modes and $2\omega$ for the pump modes.
The Ito rule leads to the corresponding stochastic differential equations:
\begin{equation}
\begin{split}
d\alpha_{sj}
=& \Bigg(-\gamma_s \alpha_{sj} + \kappa\beta_{sj}\alpha_{pj} \\
&+\sum_{l(>j)}\zeta_{jl}\alpha_{cjl} - \sum_{l(<j)}\zeta_{lj}\alpha_{clj}e^{i\theta_{lj}} +\epsilon_{sj} \Bigg)dt \\
&+\sqrt{\kappa\alpha_{pj}} dW_{\alpha_{sj}}(t),
\end{split} \label{eq:dalphasj}
\end{equation}
\begin{equation}
\begin{split}
d\beta_{sj}
=&\Bigg(-\gamma_s \beta_{sj} + \kappa\alpha_{sj}\beta_{pj} \\
&+\sum_{l(>j)}\zeta_{jl}\beta_{cjl} - \sum_{l(<j)}\zeta_{lj}\beta_{clj}e^{-i\theta_{lj}} +\epsilon_{sj} \Bigg)dt \\
&+\sqrt{\kappa\beta_{pj}} dW_{\beta_{sj}}(t),
\end{split} \label{eq:dbetasj}
\end{equation}
where $dW_{x}(t)$ is the standard Wiener increment for variable $x$.
Similarly, we obtain those for the pump modes,
\begin{align}
d\alpha_{pj}
=&\left(-\gamma_p\alpha_{pj}+\epsilon_p-\frac{\kappa}{2}\alpha_{sj}^2\right)dt, \label{eq:dalpha_pj} \\
d\beta_{pj}
=&\left(-\gamma_p\beta_{pj}+\epsilon_p-\frac{\kappa}{2}\beta_{sj}^2\right)dt
\end{align}
and for the injection-path modes,
\begin{align}
d\alpha_{cjl}
=& \left[-\gamma_c\alpha_{cjl} - \zeta_{jl}\left(\alpha_{sj}-\alpha_{sl}e^{-i\theta_{jl}}\right)\right]dt, \\
d\beta_{cjl}
=& \left[-\gamma_c\beta_{cjl} - \zeta_{jl}\left(\beta_{sj}-\beta_{sl}e^{i\theta_{jl}}\right)\right]dt \label{eq:dbeta_cjl}
\end{align}
for $j<l$.
We assume that the pump and injection-path modes decay much faster than the signal modes, i.e., $\gamma_p, \gamma_c \gg \gamma_s$.
The pump and injection-path modes are thus adiabatically eliminated.
Note that we have simplified the model more than the two-DOPO model previously investigated~\cite{K.Takata2015}
in which the injection-path mode was not eliminated.
Substituting the values for instantaneous steady states of the modes into Eqs.~(\ref{eq:dalphasj}) and (\ref{eq:dbetasj}),
we obtain
\begin{align}
\begin{split}
d\mu_j
= \Bigg[ & -\mu_j + p \nu_j\left(1-\mu_j^2\right) -\xi \left(-\sum_{l(\neq j)}J_{jl}\mu_l - h_j\right) \Bigg]d\tau \\
&+g\sqrt{1-\mu_j^2}dW_{\mu_j}(\tau)
\end{split} \label{eq:dmu_j} \\
\begin{split}
d\nu_j
= \Bigg[ & -\nu_j + p\mu_j\left(1-\nu_j^2\right) - \xi \left(-\sum_{l(\neq j)}J_{jl}\nu_l - h_j\right) \Bigg]d\tau \\
&+g\sqrt{1-\nu_j^2}dW_{\nu_j}(\tau).
\end{split} \label{eq:dnu_j}
\end{align}
Here we introduced the normalized variables $\mu_j = g\alpha_{sj}/\sqrt{p}$ and $\nu_j=g\beta_{sj}/\sqrt{p}$,
where $g = \kappa^2/(2\gamma'_s\gamma_p)$ controls the strength of the noise,
and $p = \kappa\epsilon_p/(\gamma'_s\gamma_p)$ is the pump rate.
We set $p>0$ and do not vary $p$ in time.
The strength of the injection is controlled by $\xi=\xi_0/(\gamma'_s\gamma_c)$.
The parameters $J_{jl} = \zeta_{jl}^2 e^{-i\theta_{jl}}/\xi_0$ and $h_j = g\gamma_c\epsilon_{sj}/(\sqrt{p}\xi_0)$ represent the coupling constant and the longitudinal field, respectively,
in the Ising Hamiltonian for problems
which the CIM tries to solve.
We have set $\zeta_{lj} = \zeta_{jl}$ and $e^{i\theta_{lj}} = e^{-i\theta_{jl}}$.
To guarantee $J_{jl}$ is real, $e^{-i\theta_{jl}}$ is usually set to $1$ or $-1$.
The parameter $\gamma'_s = \gamma_s + \sum_{l(\neq j)}\zeta_{jl}^2/\gamma_c$ characterizes the effective signal loss and specifies the time scale $\tau = \gamma'_s t$.
To uniquely determine $\gamma'_s$,
we restrict the setting to satisfy that $\sum_{l(\neq j)}\zeta_{jl}^2$ does not depend on $j$.
By this restriction,
$J_{jl}$ lies mainly in two classes.
One contains the couplings for the regular graph with uniform magnitude.
This class includes the ferromagnetic Ising model on a lattice.
The system with the coupling $J_{jl} = J$ or $-J$ on a lattice is also included.
The sum $\sum_{j(\neq l)}|J_{jl}|$ in this class is equal to $zJ$,
where $z$ denotes the coordination number.
The other class is that a site $j$ connects a large number $O(N)$ of sites,
and $J_{jl}$ is determined by an independent, identical distribution.
An exmple is the fully-connected Ising spin-glass model, the so-called Sherrington-Kirkpatrick model~\cite{D.Sherrington1975},
where $J_{jl}$ is extracted from the Gaussian distribution, independently, identically.
The sum $\sum_{j(\neq l)}|J_{jl}|$ in this class in the large-$N$ limit is almost surely equal to some constant that does not depend on $j$.
We obtain the Fokker-Planck equation of the reduced distribution $\tilde{P}(\bm{\mu},\bm{\nu})$ for the signal modes from the stochastic differential equations under the adiabatical elimination of the other modes,
\begin{equation}
\begin{split}
\frac{d\tilde{P}(\bm{\mu},\bm{\nu})}{dt}
= \mathcal{L} \tilde{P} & (\bm{\mu},\bm{\nu}) \\
= \sum_{j=1}^N & \bigg\{ \frac{\partial}{\partial \mu_j} \left[ \mu_j - p\nu_j\left( 1-\mu_j^2\right) +\xi V_{\mu,j} \right] \\
&+ \frac{\partial}{\partial \nu_j} \left[ \nu_j - p\mu_j\left( 1-\nu_j^2\right) +\xi V_{\nu,j} \right] \\
& + \frac{1}{2}\frac{\partial^2}{\partial \mu_j^2}g^2\left(1-\mu_j^2\right) \\
& + \frac{1}{2}\frac{\partial^2}{\partial \nu_j^2}g^2\left(1-\nu_j^2\right) \bigg\} \tilde{P}(\bm{\mu},\bm{\nu}),
\end{split} \label{eq:FP}
\end{equation}
where $V_{\mu,j} = -\sum_{l(\neq j)}J_{jl}\mu_l - h_j$ and $V_{\nu,j} = -\sum_{l(\neq j)}J_{jl}\nu_l - h_j$.
This equation is rewritten as
\begin{equation}
\frac{d\tilde{P}(\bm{\mu},\bm{\nu})}{dt}
= -\sum_{j=1}^N\left( \frac{\partial S_{\mu_j}}{\partial \mu_j} + \frac{\partial S_{\nu_j}}{\partial \nu_j}\right).
\end{equation}
Here $S_{\mu_j}$ and $S_{\nu_j}$ are given as
\begin{align}
\begin{split}
S_{\mu_j}
= \bigg[ -\mu_j +p\nu_j & \left( 1-\mu_j^2\right) - \xi V_{\mu,j} \\
&- \frac{1}{2}\frac{\partial}{\partial \mu_j}g^2\left(1-\mu_j^2\right)\bigg] \tilde{P}(\bm{\mu},\bm{\nu}),
\end{split} \\
\begin{split}
S_{\nu_j}
= \bigg[ -\nu_j +p\mu_j & \left( 1-\nu_j^2\right) - \xi V_{\nu,j} \\
& - \frac{1}{2}\frac{\partial}{\partial \nu_j}g^2\left(1-\nu_j^2\right)\bigg] \tilde{P}(\bm{\mu},\bm{\nu}),
\end{split}
\end{align}
which compose the probability current.
\section{Steady state distribution}
We derive the stationary solution $\tilde{P}_\text{SS}(\bm{\mu},\bm{\nu})$ of Eq.~(\ref{eq:FP}) that satisfies $\mathcal{L}\tilde{P}_\text{SS}(\bm{\mu},\bm{\nu})=0$.
The fact that the distribution for the positive $P$ representation can be real and positive~\cite{P.Drummond1980Jan, H.Carmichael2002, H.Carmichael2008} allows us to introduce a potential function $\Phi(\bm{\mu},\bm{\nu})$ as
\begin{equation}
\tilde{P}_\text{SS}(\bm{\mu},\bm{\nu}) = Z_N^{-1} e^{-\Phi(\bm{\mu},\bm{\nu})},
\end{equation}
where $Z_N^{-1}$ is a constant for normalization, referred to as the partition function later.
The probability current for $\tilde{P}_\text{SS}(\bm{\mu},\bm{\nu})$ is expressed as
\begin{align}
\begin{split}
S_{\mu_j}
= \bigg[ &-\left(1-g^2\right)\mu_j +p\nu_j\left( 1-\mu_j^2\right) - \xi V_{\mu,j} \\
&+ \frac{1}{2}g^2\left(1-\mu_j^2\right)\frac{\partial \Phi}{\partial \mu_j} \bigg]\tilde{P}_\text{SS}(\bm{\mu},\bm{\nu}),
\end{split} \\
\begin{split}
S_{\nu_j}
= \bigg[ &-\left(1-g^2\right)\nu_j +p\mu_j\left( 1-\nu_j^2\right) - \xi V_{\nu,j} \\
& + \frac{1}{2}g^2\left(1-\nu_j^2\right)\frac{\partial\Phi}{\partial \nu_j}\bigg]\tilde{P}_\text{SS}(\bm{\mu},\bm{\nu}).
\end{split}
\end{align}
A simple strategy to find a solution is to assume detailed balance
that guarantees the existence of the equilibrium distribution as a stationary distribution~\cite{H.Risken1989}.
The detailed balance condition in the Fokker-Planck equation is equivalent to the absence of the probability current~\cite{H.Risken1989},
$S_{\mu_j} = S_{\nu_j} = 0 \ \forall j$.
The potential function under the detailed balance condition, say $\Phi_\text{DB}(\bm{\mu},\bm{\nu})$, satisfies
\begin{align}
\frac{\partial \Phi_\text{DB}}{\partial \mu_j}
&= \frac{2}{g^2\left(1-\mu_j^2\right)}\left[\left(1-g^2\right)\mu_j - p\nu_j\left( 1-\mu_j^2\right) + \xi V_{\mu,j}\right], \label{eq:dPhi_DBdmu_j} \\
\frac{\partial \Phi_\text{DB}}{\partial \nu_j}
&= \frac{2}{g^2\left(1-\nu_j^2\right)}\left[\left(1-g^2\right)\nu_j - p\mu_j\left( 1-\nu_j^2\right) + \xi V_{\nu,j}\right]. \label{eq:dPhi_DBdnu_j}
\end{align}
For $\xi \neq 0$, however, no function satisfies the above eqautions,
since the above equations lead to
\begin{align}
\frac{\partial}{\partial\mu_l}\frac{\partial\Phi_\text{DB}}{\partial \mu_j}
&= -\frac{2\xi J_{jl}}{g^2\left(1-\mu_j^2\right)} \label{eq:second-derivative_Phi_DB_jl} \\
\frac{\partial}{\partial\mu_j}\frac{\partial\Phi_\text{DB}}{\partial \mu_l}
&= -\frac{2\xi J_{jl}}{g^2\left(1-\mu_l^2\right)}, \label{eq:second-derivative_Phi_DB_lj}
\end{align}
where we have used $J_{lj} = J_{jl}$.
The differentiation of the function with respect to $\mu_j$ and $\mu_l$ thus cannot be interchanged for $\xi \neq 0$.
This fact demonstrates that the state distribution governed by Eq.~(\ref{eq:FP}) for $\xi \neq 0$ does not exhibit detailed balance.
Hence, there exists a finite probability current,
and the system has no equilibrium state distributions.
When $\xi = 0$,
where the DOPOs do not interact with each other,
the detailed balance can be holded.
On the other hand,
if the noises in the stochastic differential equations [Eqs.~(\ref{eq:dmu_j}) and (\ref{eq:dnu_j})] were governed by the Gaussian distribution,
i.e., $gdW_{\mu_j}(\tau)$ and $gdW_{\mu_j}(\tau)$,
the denominators in the right hand side of Eqs.~(\ref{eq:second-derivative_Phi_DB_jl}) and (\ref{eq:second-derivative_Phi_DB_lj}) would turn to $g^2$,
and the detailed balance could be recovered.
The breakdown of detailed balance thus stems from combination of the interactions and multiplicative noises.
The violation of the detailed balance condition has been discussed in several studies,
in particular, in the context of efficient sampling methods,
where it has been confirmed that faster convergence to steady states is realized by violating the detailed balance condition~\cite{H.Suwa2010, K.Turitsyn2011, H.Fernandes2011, Y.Sakai2013}.
In other words,
the stochastic dynamics without detailed balance can reach the steady states in a shorter time
than the corresponding dynamics
that obeys the detailed balance condition~\cite{A.Ichiki2013, M.Ohzeki2015Jul, M.Ohzeki2015Sep}.
The violation of the detailed balance condition of the dynamics in the CIM is hence expected to accelerate the convergence to the steady states.
Note that the origin of the noises is quantum fluctuations.
While the breakdown of the detailed balance condition suggests a nontrivial character of the CIM in the relaxation to steady states,
it is an obstacle that makes it difficult to derive steady state distributions.
We here focus on a part of the steady state distribution in which
the difference in the magnitude $\mu_i^2$ ($\nu_j^2$) of different DOPOs is small.
We represent
\begin{equation}
\mu_j^2 = q^\mu + \delta_j^\mu, \ \ \
\nu_j^2 = q^\nu + \delta_j^\nu,
\end{equation}
where
\begin{equation}
q^\mu = \frac{1}{N}\sum_{j=1}^N \mu_j^2, \ \ \
q^\nu = \frac{1}{N}\sum_{j=1}^N \nu_j^2.
\end{equation}
In a region,
where $\delta_j^\mu$ and $\delta_j^\nu$ are small,
there can be a potential function that approximately satisfies Eqs.~(\ref{eq:dPhi_DBdmu_j}) and (\ref{eq:dPhi_DBdnu_j}).
Note again that
the potential function of the CIM in general never exhibits detailed balance,
but there could be a part in which Eqs.~(\ref{eq:dPhi_DBdmu_j}) and (\ref{eq:dPhi_DBdnu_j}) hold.
If the other part of the potential function is nearly equal to zero,
we do not have to consider that part.
The expansion of the potential function in terms of $\delta$,
obtained from Eqs.~(\ref{eq:dPhi_DBdmu_j}) and (\ref{eq:dPhi_DBdnu_j}),
is
\begin{equation}
\begin{split}
g^2\Phi
&= g^2\Phi_0 -\frac{2\xi}{1-q^\mu}\sum_{j<l}J_{jl}\mu_j \mu_l -\frac{2\xi}{1-q^\nu}\sum_{j<l}J_{jl}\nu_j \nu_l \\
& -\frac{2\xi}{1-q^\mu}\sum_{j=1}^N h_j \mu_j -\frac{2\xi}{1-q^\nu}\sum_{j=1}^N h_j \nu_j + O(\delta^2), \label{eq:g^2Phi}
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
g^2\Phi_0
= -N\left(1-g^2\right) [ \ln \left(1-q^\mu \right) + & \ln \left( 1-q^\nu \right) ] \\
& - 2p\sum_{j=1}^N \mu_j \nu_j.
\end{split} \label{eq:Phi_0}
\end{equation}
The function $\Phi_0$ provides the terms for the independent DOPOs,
reproducing the known result for a single DOPO~\cite{M.Wolinsky1988},
which with large $p$ form a double-well in the potential function.
The double-well corresponds to the bifurcation of the in-phase amplitude above the threshold of pump rate
to make pseudo Ising variables.
The other terms undertake the coupling of DOPOs embedded for a target optimization problem.
When we neglect terms $O(\delta^2)$,
the terms for the coupling compose of the Hamiltonian or the cost function for the continuous relaxation of a target discrete optimization problem.
In the subspace that satisfies $\delta = 0$ the Ising Hamiltonian is effectively reproduced~\cite{T.Leleu2017}.
In the view of steady state distributions,
the CIM thus finds solutions with the combination of the Ising-like double-wells and the continuous relaxation of the problem.
It is not obvious, however, that the global minimum of the potential function given by Eq.~(\ref{eq:g^2Phi}) agrees with the ground state of the Ising Hamiltonian for the discrete optimization problem.
Note that the presence of the double-well does not directly indicate the superposition of the two coherent states for the wells, namely our up-spin and down-spin.
It rather leads to the classical mixture of them in the single- and two-DOPO cases at least~\cite{M.Reid1992, K.Takata2015}.
In time for transient evolution, a sign of the superposition was found in numerical simulations of a single DOPO case~\cite{L.Krippner1994} and in the two-DOPO model~\cite{K.Takata2015}.
This feature might be a characteristic property of DOPOs even for solving combinatorial optimization problems,
but the transient time scale is out of our scope in the present study.
\section{Typical solutions with the potential function in the large-size limit}
If the higher order terms $O(\delta^2)$ in Eq.~(\ref{eq:g^2Phi}) are negligible,
the potential function under the detailed balance condition can be a good approximation of the true one.
It is available to analytically evaluate the property of the approximate potential function without those terms.
We here examine the potential function without the higher order terms,
applying it to two simple examples.
We consider only the real part of $\mu_j$ and $\nu_j$.
This simplification is based on the standard initial condition of the dynamics in which all signal fields is set to the vacuum state
and on real $J_{jl}$ with $e^{-i\theta_{jl}} = 1, -1$.
Equations~(\ref{eq:dmu_j}) and (\ref{eq:dnu_j}) show that
if all $\mu_j$ and $\nu_j$ have no imaginary part at an instantaneous time,
they remain real for all time~\cite{M.Wolinsky1988}.
What we like to know is the configuration of Ising spins yielded from the CIM according to the potential function.
The Ising spins are encoded in the sign of the in-phase amplitude of the signal fields~\cite{T.Inagaki2016Oct, P.McMahon2016, Z.Wang2013, A.Marandi2014, K.Takata2015}.
We define an operator $\hat{\sigma}_j$ by $\hat{\sigma}_j \ket{x_j} = \text{sign}(x_j)\ket{x_j}$,
where $\ket{x_j}$ is the eigenstate of operator $\hat{x}_j = (\hat{a}_{sj} + \hat{a}_{sj}^\dagger)/2$,
and sign($x_j$) is 1 if $x_j>0$ and $-1$ if $x_j<0$.
Its expectation value is
\begin{equation}
\begin{split}
\text{tr} & \left( \hat{\rho}\hat{\sigma}_j \right) \\
&= \int d^N\bm{x} d^N\bm{\mu}d^N\bm{\nu} \text{sign}\left(x_j\right) \frac{\langle\bm{x}|\bm{\alpha}\rangle\langle\bm{\beta}^*|\bm{x}\rangle}{\langle\bm{\beta}^*|\bm{\alpha}\rangle} \tilde{P}\left(\bm{\mu}, \bm{\nu} \right) \\
&= \int d^N\bm{\mu}d^N\bm{\nu} \left\{ 1 - 2H\left[\frac{\sqrt{p}}{g}\left(\mu_j + \nu_j \right)\right]\right\} \tilde{P}\left(\bm{\mu}, \bm{\nu} \right) \\
&\simeq \int d^N\bm{\mu}d^N\bm{\nu} \text{sign}\left(\mu_j + \nu_j \right) \tilde{P}\left(\bm{\mu}, \bm{\nu} \right),
\end{split}
\end{equation}
where $H(x) = \int_x^\infty dt e^{-t^2/2}/\sqrt{2\pi}$.
The last line is obtained from the second line by ignoring the fluctuation in the coherent state.
This is a rather good approximation for small $g$.
For estimating the efficiency of the CIM to solve combinatorial optimization problems,
it is important to invesigate its behavior for large-size problems.
The method of statistical mechanics is suitable for this situation~\cite{T.Aonishi2017}.
We define the partition function and free energy for the potential function by
\begin{align}
Z_N(\eta)
&= \int d^N\bm{\mu}d^N\bm{\nu} e^{-\Phi(\bm{\mu},\bm{\nu})-g^{-2}\eta M(\bm{\sigma})}, \\
f(\eta)
&= -\lim_{N\to\infty} \frac{g^2}{N}\ln Z_N(\eta),
\end{align}
repectively.
Here the term $g^2\eta M(\bm{\sigma})$ is introduced to evaluate the expectation value of order parameter $M(\bm{\sigma})$, e.g., $M(\bm{\sigma}) = \sum_{j=1}^N \sigma_j$, where $\sigma_j = \text{sign}(\mu_j + \nu_j)$.
If we had the true solution $\bm{\sigma}^{0}$ of the problem,
$M(\bm{\sigma})$ could be overlap between the Ising spins in the CIM and the solution, i.e.,
$M(\bm{\sigma}) = \sum_{j=1}^N \sigma_j\sigma_j^0$,
which estimates how correct the answer of the CIM is.
The free energy chracterizes the macroscopic property of the system,
giving the expectation values of macroscopic quantities, e.g., $M(\bm{\sigma})$.
\subsection{Fully-connected ferromagnetic coupling}
We first investigate, as the simplest example, an optimization problem
that is mapped onto the fully-connected ferromagnetic Ising model without the Zeeman terms.
All the coupling constants $J_{jl}$ are equal to $J/(2N)$, $J>0$,
and the longitudinal field $h_{j}$ vanishes.
The correct ground states of the corresponding Ising model are all-up and all-down.
The potential function without the higher order terms $O(\delta^2)$ for this problem is
\begin{equation}
g^2\Phi
= g^2\Phi_0 -\frac{N\xi J}{2\left(1-q^\mu \right)} \left(m^\mu\right)^2 -\frac{N\xi J}{2\left(1-q^\nu \right)} \left( m^\nu \right)^2, \label{eq:Phi_ferro}
\end{equation}
where $\Phi_0$ is given by Eq.~(\ref{eq:Phi_0}), and
\begin{equation}
m^\mu = \frac{1}{N} \sum_{j=1}^N \mu_j, \ \ \
m^\nu = \frac{1}{N} \sum_{j=1}^N \nu_j. \label{eq:m^mu-m^nu}
\end{equation}
The order paramter is $M(\bm{\sigma}) = \sum_{j=1}^N \sigma_j$.
The partition function is written as
\begin{equation}
\begin{split}
Z(\eta)
=& \int d^N\bm{\mu} d^N\bm{\nu} d^2\bm{m}d^2\bm{q} \delta\left(Nm^\mu - \sum_{j=1}^N\mu_j \right) \\
&\times \delta\left(Nm^\nu - \sum_{j=1}^N\nu_j \right) \delta\left(Nq^\mu - \sum_{j=1}^N\mu_j^2 \right) \\
&\times \delta\left(Nq^\nu - \sum_{j=1}^N\nu_j^2 \right) e^{-\Phi-g^{-2}\eta M(\bm{\sigma})} \\
=& \int d^2\bm{m}d^2\bm{q}d^2\bm{\tilde{m}}d^2\bm{\tilde{q}} \exp \bigg( g^{-2}N \\
& \times \bigg\{ \left(1-g^2\right) \left[ \ln \left(1-q^\mu\right) + \ln \left(1-q^\nu\right) \right] \\
& \hspace{20pt} + \bm{\tilde{m}}^\text{T} \bm{m} + \bm{\tilde{q}}^\text{T}\bm{q} \\
& \hspace{20pt} + \frac{\xi J}{2\left(1-q^\mu \right)}\left(m^\mu\right)^2 + \frac{\xi J}{2\left(1-q^\nu \right)}\left(m^\nu\right)^2 \\
& \hspace{20pt} + g^2 \ln \int d\mu d\nu e^{-g^{-2}\phi}
\bigg\} \bigg).
\end{split} \label{eq:Z_ferro}
\end{equation}
Here $\bm{m} = (m^\mu,m^\nu)^\text{T}$, $\bm{q} = (q^\mu,q^\nu)^\text{T}$, $\bm{\tilde{m}} = (\tilde{m}^\mu,\tilde{m}^\nu)^\text{T}$, and $\bm{\tilde{q}} = (\tilde{q}^\mu,\tilde{q}^\nu)^\text{T}$,
where $\bm{x}^\text{T}$ denotes the tranpose of a column vector $\bm{x}$.
The variables $\bm{\tilde{m}}$ and $\bm{\tilde{q}}$ are introduced for the integral expression of the delta function.
In addition, we have
\begin{equation}
\phi = \bm{z}^\text{T}\tilde{Q}\bm{z}+ \bm{\tilde{m}}^\text{T}\bm{z} + \eta \sigma, \label{eq:phi}
\end{equation}
where $\bm{z} = (\mu, \nu)^\text{T}$,
and elements of matrix $\tilde{Q}$ are $\tilde{Q}_{11} = \tilde{q}^\mu$, $\tilde{Q}_{12} = \tilde{Q}_{21} = -p$, and $\tilde{Q}_{22} = \tilde{q}^\nu$,
and $\sigma = \text{sign}(\mu + \nu)$.
The integral is calculated as
\begin{equation}
\begin{split}
&\int d\mu d\nu e^{-g^{-2}\phi} \\
&= \frac{\pi g^2}{\sqrt{\text{det}\tilde{Q}}} \exp\left(\frac{g^{-2}}{4}\bm{\tilde{m}}^\text{T}\tilde{Q}^{-1}\bm{\tilde{m}} \right) G\left(\bm{\tilde{m}}, \bm{\tilde{q}}, \eta \right).
\end{split} \label{eq:integral_phi}
\end{equation}
We do not explicitly show $G\left(\bm{\tilde{m}}, \bm{\tilde{q}}, \eta \right)$,
but the function for the symmetric case discussed later is given in App.~\ref{app:G}.
It should be noted that $G\left(\bm{\tilde{m}}, \bm{\tilde{q}}, 0 \right) = 1$.
Since the exponent of the integrand of $Z(\eta)$ is propotional to $N$,
in the large $N$ limit the method of steepest descent gives
\begin{equation}
\begin{split}
f(\eta)
= & \underset{\bm{m},\bm{q},\bm{\tilde{m}},\bm{\tilde{q}}}{\text{extr}}
\bigg\{ -\left(1-g^2\right)\left[ \ln \left(1-q^\mu \right) + \ln \left(1-q^\nu \right) \right] \\
& - \bm{\tilde{m}}^\text{T}\bm{m} - \bm{\tilde{q}}^\text{T}\bm{q} \\
&- \frac{\xi J}{2\left(1-q^\mu \right)}\left(m^\mu\right)^2 - \frac{\xi J}{2\left(1-q^\nu \right)}\left(m^\nu\right)^2 \\
& - g^2\ln\left(\pi g^2\right) + \frac{g^2}{2}\ln\text{det}\tilde{Q} - \frac{1}{4}\bm{\tilde{m}}^\text{T}\tilde{Q}^{-1}\bm{\tilde{m}} \\
& -g^2\ln G\left(\bm{\tilde{m}}, \bm{\tilde{q}}, \eta \right) \bigg\},
\end{split}
\end{equation}
where $\text{extr}_{\bm{m},\bm{q},\bm{\tilde{m}},\bm{\tilde{q}}}$ represents taking an extremum with respect to $\bm{m},\bm{q},\bm{\tilde{m}},\bm{\tilde{q}}$.
The terms in the forth line contribute to the entropic part of the free energy.
Since the term in the last line is only used to compute the average of $\sigma$,
$\eta$ is set to zero
when searching for saddle points of the free energy.
The saddle points are determined as
\begin{align}
\tilde{m}^\mu
&= -\frac{\xi J}{1-q^\mu}m^\mu, \\
\tilde{q}^\mu
&= \frac{1-g^2}{1-q^\mu} - \frac{\xi J}{2\left(1-q^\mu \right)^2}\left(m^\mu \right)^2, \\
m^\mu
&= -\frac{\tilde{m}^\mu \tilde{q}^\nu + \tilde{m}^\nu p}{2\left( \tilde{q}^\mu \tilde{q}^\nu - p^2 \right)}, \\
q^\mu
&= \frac{g^2\tilde{q}^\nu}{2\left( \tilde{q}^\mu \tilde{q}^\nu - p^2 \right)} + \left[ \frac{\tilde{m}^\mu \tilde{q}^\nu + \tilde{m}^\nu p}{2\left( \tilde{q}^\mu \tilde{q}^\nu - p^2 \right)} \right]^2.
\end{align}
We also have the equations obtained by interchanging $\mu$ and $\nu$ in superscripts in the above equations.
In particular, $\bm{m}$ and $\bm{q}$ for the saddle points are the expectation values of them under the distribution governed by the potential function in Eq.~(\ref{eq:Phi_ferro}).
We have not found any solutions that satisfy $m^\mu \neq m^\nu$ by numerical calculations.
We thus restrict ourselves to consider symmetric solutions for which the parameters do not depend on $\mu$ and $\nu$, i.e.,
$m^\mu = m^\nu = m$, $q^\mu = q^\nu = q$, $\tilde{m}^\mu = \tilde{m}^\nu = \tilde{m}$, $\tilde{q}^\mu = \tilde{q}^\nu = \tilde{q}$.
Accordingly, the saddle point equations reduce to
\begin{align}
\tilde{m}
&= -\frac{\xi J}{1-q} m, \label{eq:tildem^sp} \\
\tilde{q}
&= \frac{1-g^2}{1-q} - \frac{\xi J}{2\left(1-q \right)^2} m^2, \\
m
&= -\frac{\tilde{m}}{2\left( \tilde{q} - p \right)}, \\
q
&= \frac{g^2\tilde{q}}{2\left( \tilde{q}^2 - p^2 \right)} + \left[ \frac{\tilde{m}}{2\left( \tilde{q} - p \right)} \right]^2. \label{eq:q^sp}
\end{align}
Intuitively,
the term $-2g^{-2}p\sum_{j}\mu_j \nu_j$ in the potential function $\Phi_0$ [Eq.~\ref{eq:Phi_0}] enhances the overlap between $\mu$ and $\nu$,
and then the symmetric solution is realized.
It should be noted that the restriction that
the macroscopic parameters above do not depend on $\mu$ and $\nu$
does not mean that
we assume $\mu_j = \nu_j$.
The condition $\mu_j = \nu_j$ would restrict our analysis into a smaller subspace,
where the density operator is represented as a classical mixture of the coherent states.
The condition, consequently, leads to a different entropic part from the above one we actually obtained.
This difference indicates that our analysis including approximations still reflects some quantum effects.
What we like to calculate is the expectation value of the Ising spins $m_\sigma = \langle N^{-1} \sum_{j=1}^N \sigma_j \rangle$,
where $\langle X \rangle$ denotes the average of $X$ over the distribution $e^{-\Phi}Z(\eta=0)^{-1}$.
Using the function $G$ in the symmetric case shown in App.~\ref{app:G},
we obtain
\begin{equation}
m_\sigma
= \frac{df(\eta)}{d\eta}\bigg|_{\eta=0}
= -1 + 2H\left( \frac{\tilde{m}}{g\sqrt{\tilde{q}-p}} \right), \label{eq:m_sigma}
\end{equation}
where the values of parameters $\tilde{m}$ and $\tilde{q}$ are for the saddle point.
We first examine the solution for $\xi = 0$,
where the system has no interactions between different DOPOs.
In this case,
the saddle point equations give $m= m_\sigma = 0$
that is consistent with the fact that the system has no bias.
The solution $q$ has a positive value for $g>0$.
In the limit $g \to 0$,
$q$ shows not the fluctuation but just the square of the amplitude of $\mu$ and $\nu$ frozen at a basin of the potential function.
Thus we can find the character of the shape of the potential function in the behavior of $q$.
There is a threshold $p = 1$.
Below the threshold, $p < 1$,
the solution of $q$ is equal to zero,
which means that
the potential function has the unique minimum at $\mu = \nu= 0$.
Above the threshold, $p > 1$,
another solution appears with a finite value, $q = 1 - 1/p$, in addition to $q=0$.
The potential function then have the minima at $\mu = \nu = \pm \sqrt{1-1/p}$ and the unstable extremum at $\mu = \nu = 0$.
This behavior agrees with the known bifurcation for a single DOPO~\cite{M.Wolinsky1988}.
We move to investigation of the system for finite $\xi$,
where the DOPOs interact with each other.
To gain insight into this case,
we consider the limit $g \to 0$,
where Eq.~(\ref{eq:q^sp}) reduces to $q = m^2$.
Accordingly,
the saddle-point equation for $m$ results in
\begin{equation}
\begin{split}
m =& \frac{1}{2}\left[\frac{1}{1-m^2} - \frac{\xi J}{2\left(1-m^2\right)^2}m^2 - p\right]^{-1} \\
& \times \frac{\xi J}{1-m^2}m
\ \ \ (g=0).
\end{split}
\end{equation}
This equation has three (five) possible solutions of $m^2$ ($m$);
$m_0 = 0$ and $m_\pm^2 = 1-(1\pm\sqrt{1-2p\xi J})/2p$.
To choose physically reasonable solutions,
we examine the stability of the possible ones.
The first candidate $m_0$ is stable only when $p+\xi J/2<1$,
and the others are unstable or complex in this condition.
The second one $m_+^2$ is stable only when $p > 1/2$ and $p+\xi J/2>1$,
but the region $2p\xi J>1$ is excluded,
where $m_+$ becomes complex.
The third one $m_-^2$ is always unstable or complex.
Summarizing,
we have a finite real solution, $m_+$, only for $p > 1/2$ and $p+\xi J/2>1$ except for $2p\xi J>1$.
When $p+\xi J/2<1$, the solution is $m_0 (=0)$.
In the other region our approach does not yield any real solution.
These solutions determine $m_\sigma$ via Eq.~(\ref{eq:m_sigma});
negative (positive) $\tilde{m}$, i.e., positive (negative) $m$, leads to $m_\sigma = 1$ ($m_\sigma = -1$),
which is the correct ground state of the corresponding Ising model.
The boundary, therefore, is $p + \xi J/2 = 1$,
and larger $p$ and $\xi$ under $p >1/2$ give $m_\sigma = 1$ or $-1$~\cite{T.Aonishi2017}.
When $g=0$, $p + \xi J_{ij}N$ is an effective pump rate,
if all the DOPOs display the same $\mu$ and $\nu$.
In this example, where $J_{ij} = J/2N$,
$p + \xi J/2$ is the effective pump rate.
The boundary obtained here is given by the effective pump rate equal to unity.
The emergence of the finite solution is identified with a phase transtion in the $p$--$\xi J$ phase space.
If we carry out the annealing approach by controling $p$ or $\xi$ with keeping the above instantaneous steady states,
the system undergoes the phase transition.
It is interesting that
$m_\sigma$ exhibits the discontinuous change at the boundary,
namely, the first-order phase transition,
while $m$ continuously changes as the second-order one.
This definite difference is only for the case $g \to 0$,
but this finding suggests a feature of the scheme,
in which the problem is solved by the continous variables, i.e., $\mu$ and $\nu$,
encoding the discrete ones.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig1_v0.22-.eps}
\end{center}
\caption{Heatmap for the solutions of the saddle-point equations [Eqs.~(\ref{eq:tildem^sp})--(\ref{eq:q^sp})] for different $g$ in the $p$--$\xi J$ plane.
(a) $m$, (b) $q$, and (c) $m_\sigma$ for $g=0.01$,
and (d) $m$, (e) $q$, and (f) $m_\sigma$ for $g=0.4$.
The dashed straight line is the boundary above which $m$ for $g \to 0$ has a finite real value,
while the dashed curve is the boundary above which there is no real, stable solution of $m$ for $g \to 0$.
Initial conditions in solving the equations are set for simplicity so that $m$ tends not to be negative.
The white region is where the real, stable solutions are not found.
}
\label{fig:solutions^sp_ferro}
\end{figure}
We numerically find the stable solutions of the saddle-point equations for $g>0$.
For small finite $g$ ($=0.01$) [Fig.~\ref{fig:solutions^sp_ferro}~(a)--(c)],
the solution is similar to that for $g=0$.
Note that $g$ for the actual CIM is smaller than 0.01.
We find a rather sharp transition of $m_\sigma$ from 0 to a finite value very close to unity,
at the almost same location as that for $g = 0$.
The solution $m_\sigma = -1$ is also obtained,
but we set an initial condition in solving the equations so that $m$ tends not to be negative for simplicity.
The result for a rather large $g$ ($=0.4$) [Fig.~\ref{fig:solutions^sp_ferro}~(d)--(f)] displays different behavior due to noises.
The boundary for $m$ to take a finite value becomes a curve instead of the straight line as for $g=0$.
A stronger pump is thus needed to have finite $m$.
This is because $g$ partially plays a role of temperature in the dynamics of our system.
Hence, the system with large $g$ has large fluctuations.
The value of $m$ thus tends to vanish,
whereas $q$ takes a finite value.
However,
it is only when $p$ and $\xi$ is small that
the system is completely disturbed by the fluctuations.
Stronger pump and interactions make basins in the potential function deeper.
The state of the system is captured by the deep basin.
A finite $m$ is observed, consequently.
Whereas $m_\sigma$ for $g=0.01$ suddenly changes at the boundary,
its change for $g=0.4$ is smoother due to the noises.
We nevertheless are able to obtain almost all-up (or all-down) state with larger $p$ and $\xi$.
\subsection{Fully-connected ferromagnetic coupling with random fields}
We next examine the fully-connected ferromagnetic coupling with random fields.
The couplings are the same as those in the previous example,
$J_{ij} = J/2N$, $J>0$,
while a longitudinal field for each site is randomly taken from some distribution identically, independently.
The potential function without the higher order terms $O(\delta^2)$ for this problem is
\begin{equation}
\begin{split}
g^2\Phi
=& g^2\Phi_0 -\frac{N\xi J}{2\left(1-q^\mu \right)} \left(m^\mu\right)^2 -\frac{N\xi J}{2\left(1-q^\nu \right)} \left( m^\nu \right)^2 \\
& -\frac{2\xi}{1-q^\mu}\sum_{j=1}^N h_j \mu_j -\frac{2\xi}{1-q^\nu}\sum_{j=1}^N h_j \nu_j,
\end{split}
\end{equation}
where $\Phi_0$ and both $m^\mu$ and $m^\nu$ are given in Eqs.~(\ref{eq:Phi_0}) and (\ref{eq:m^mu-m^nu}), respectively.
We also set $M(\bm{\sigma}) = \sum_{j=1}^N \sigma_j$ in this problem.
We examine the case in which
the fields take binary values $h_0$ or $-h_0$ at random.
The corresponding Ising model governed by the Gibbs distribution has been investigated well~\cite{H.Nishimori2011}.
It is known that the system at zero temperature undergoes a first-order phase transition
with increasing the amplitude $h_0$ of the fields.
The field for the transition is $h_0/J = 1/2$,
below which the spins are all-up or all-down,
whereas each spin is parallel to the field above the critical point.
Our aim here is to clarify whether our approach captures this transition.
It is not difficult,
as shown in App.~\ref{app:random-field},
to extend the partition function and free energy for the no-field model to the case with fields,
but the randomness in the fields has to be carefully treated.
We here exploit the self-averaging property,
where the free energy for an instance of random fields is almost surely equal to the averaged one in the large $N$ limit~\cite{H.Nishimori2011}.
We can derive the corresponding saddle-point equations for the symmetric solutions from the averaged free energy over the configuration of random fields,
\begin{align}
\tilde{m}
&= -\frac{\xi J}{1-q} m, \label{eq:tildem^sp_random-field} \\
\tilde{q}
&= \frac{1-g^2}{1-q} - \frac{\xi J}{2\left(1-q \right)^2} m^2 - \frac{2\left(\xi h_0 \right)^2}{\left(\tilde{q} - p\right)\left(1-q\right)^3}, \label{eq:tildeq^sp_random-field} \\
m
&= -\frac{\tilde{m}}{2\left( \tilde{q} - p \right)}, \label{eq:m^sp_random-field} \\
q
&= \frac{g^2\tilde{q}}{2\left( \tilde{q}^2 - p^2 \right)} + \frac{\tilde{m}^2 + \left[2\xi h_0/\left(1-q\right)\right]^2}{4\left( \tilde{q} - p \right)^2}. \label{eq:q^sp_random-field}
\end{align}
The expectation value of Ising spins is
\begin{equation}
\begin{split}
m_\sigma =& -1 + H\left[\frac{\tilde{m}-2\xi h_0/(1-q)}{g\sqrt{\tilde{q}-p}}\right] \\
& + H\left[\frac{\tilde{m}+2\xi h_0/(1-q)}{g\sqrt{\tilde{q}-p}}\right].
\end{split} \label{eq:m_sigma_random-field}
\end{equation}
In the limit $g\to 0$,
the saddle-point equations for finite $m$ lead to $q = m^2 + q_h$,
where $q_h = (2h_0/J)^2$ indicates the variance of $\mu$ and $\nu$ purely driven by the random fields.
Accordingly, Eq.~(\ref{eq:m^sp_random-field}) with Eq.~(\ref{eq:tildeq^sp_random-field}) turns to
\begin{equation}
\begin{split}
m
= \frac{1}{2} & \Bigg[\frac{1}{1-m^2-q_h} \\
& - \frac{\xi J}{2\left(1-m^2-q_h\right)^2}\left(m^2+2q_h\right) -p \Bigg]^{-1} \\
& \times \frac{\xi J}{1-m^2-q_h} m \ \ \ (g=0).
\end{split}
\end{equation}
This equation has three (five) possible solutions of $m^2$ ($m$);
$m_0=0$ and $m_\pm^2 = (1-q_h)[1-(1\pm \sqrt{1-2p'\xi'J})/(2p')]$,
where $p' = (1-q_h)p$ and $\xi' = (1+q_h)\xi/(1-q_h)$.
We can find the physical solutions,
which are real and stable ones,
through the same argument as in the no-field case,
but $p$ and $\xi$ in that case are replaced with $p'$ and $\xi'$ here.
We hence have a finite solution ($m_+$)
only when $p' > 1/2 $ and $ p' + \xi' J/2 > 1$,
but the region $2p'\xi'J>1$,
where $m_+$ becomes complex,
is excluded.
For $p' + \xi'J/2 <1$,
the solution is $m_0(=0)$.
In the other region,
$p'<1/2 $ and $ p' + \xi'J/2 >1$,
there is no real solution.
These solutions determine $m_\sigma$ via Eq.~(\ref{eq:m_sigma_random-field}) with Eq.~(\ref{eq:tildem^sp_random-field});
if $m^2 > q_h$, $|m_\sigma| = 1$,
otherwise $m_\sigma = 0$.
The discontinuous change of $|m_\sigma|$ from 0 to $1$ thus occurs at $h_0/(m_+ J) = 1/2$ in the region for $p' > 1/2 $ and $ p' + \xi' J/2 > 1$.
The quantity $h_0/m_+$ represents the magnitude of the effective field in the Ising model
which the CIM actually solves,
since being divided by $m_+$ relaxes the discrepancy,
in the balance between the two-body and one-body interactions,
of the model represented with $\mu$ and $\nu$ from the Ising model.
This finding demonstrates that
the CIM detects the first-order phase transition at which
the ratio of the effective field to the coupling constant is equal to 1/2.
This boundary agrees with that for the corresponding Ising system at zero temperature governed by the conventional Gibbs distribution~\cite{H.Nishimori2011}.
When we compare the condition $h_0/(m_+J) = 1/2$ with the transition point in the corresponding Ising model,
it is considered as the boundary for $h_0$ with fixed $p$ and $\xi$ in the region for finite $m$.
The condition is also interpreted as the boundary for $p$ or $\xi$ with fixed $h_0$.
For the latter we discuss the transition which the system undergoes in the annealing approach with controling $p$ or $\xi$.
As in the no-field case,
$m$ continuously changes with increasing $p$ or $\xi$,
while $m_\sigma$ jumps at the boundary.
There is a difference, however, in the mechanism for the jump.
The discontinous change in the random-field case is caused by the first-order phase transition of the genuine Ising model,
whereas that observed in the no-field case is just due to the bifurcation of the DOPOs.
The first-order phase transition is in general owed to the presence of an energy barrier between multiple minima~\cite{H.Nishimori2011},
which makes it difficult to search for the ground states in the energy landscape.
Hence, that transition should be avoided in the scheme.
For instance,
such a transition in quantum annealing often concerns an exponentially small energy gap~\cite{T.Jorg2008, T.Jorg2010Feb, T.Jorg2010May, C.Laumann2012}
and thus inefficienty of the method.
The continuous change of $m$ in our approach demonstrates the absence of local minima,
despite the target Ising model has the first-order phase transition.
We find the solutions of the saddle-point equations for $g>0$ by numerically solving them.
For $g=0.01$ the resulting $m_\sigma$ steeply decreases from 1 to 0 with increasing $h_0/(mJ)$ with fixed $p$ and $\xi$ as shown in Fig.~\ref{fig:m_sigma_random-field}~(a).
The sudden change takes place around $h_0/(mJ) = 0.5$,
which agrees with the transition point of the corresponding Ising model at zero temperature.
The transition yielded by our model, however, is not the first-order one,
while the corresponding Ising model exhibits the first-order one even for finite low temperatures~\cite{H.Nishimori2011}.
This difference probably originates from the fact that
our model is governed by the continuous degrees of freedom, $\mu$ and $\nu$,
rather than the discrete ones in the genuine Ising model.
Larger $p$ and $\xi$ enhance Ising-like behavior.
As a result, steeper change at the transition point is found.
Except around the transition point,
we obtain the correct ground state of the target Ising model.
For larger $g$ [Fig.~\ref{fig:m_sigma_random-field}~(b)],
we do not find any sudden change of $m_\sigma$,
but it monotonically decreases with increasing the effective field.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig2_v0.23-.eps}
\end{center}
\caption{Obtained $m_\sigma$ from Eq.~(\ref{eq:m_sigma_random-field})
with the solution of the saddle-point equations [Eqs.~(\ref{eq:tildem^sp_random-field})--(\ref{eq:q^sp_random-field})]
as a function of $h_0/(mJ)$
for (a) $g= 0.01$ and (b) $g=0.1$
at $(p, \xi) = (1.0,0.1)$, $(1.1, 0.1)$, $(1.2, 0.1)$, $(1.0, 0.2)$, $(1.1, 0.2)$, $(1.2, 0.2)$, $(1.0, 0.3)$, $(1.1, 0.3)$, and $(1.2, 0.3)$
from bottom to top.
The plots are shifted vertically by 0.5 for clarity.
Real, stable solutions at large $h_0/mJ$ are not obtained for some sets of $(p,\xi)$.
Initial conditions in solving the saddle-point equations are set for simplicity so that $m$ tends not to be negative.
}
\label{fig:m_sigma_random-field}
\end{figure}
\section{Summary and discussion}
We investigated network of degenerate optical parametric oscillators (DOPOs)
as a model of the coherent Ising machine (CIM)~\cite{Z.Wang2013, A.Marandi2014, T.Inagaki2016Oct, P.McMahon2016},
an architecture for solving problems expressed with the Ising models.
The network is composed of the optical coupling of DOPOs,
representing the Ising model with the parameters $J_{jl}$ and $h_j$.
Motivated by the annealing approach~\cite{S.Geman1984, H.Nishimori1998, T.Albash2018, L.Venuti2016},
we tried to derive the steady state distributioins for the dynamics described with the positive $P$ representation~\cite{P.Drummond1980Jan}.
The distribution is expected to yield answers of the CIM on the problems in the long time limit.
We obtained approximate steady state distributions for arbitrary Ising problems under the ansatz that
the difference in $\mu_i^2$ and $\nu_i^2$ from other DOPOs is small.
Using the method of statistical mechanics in the large problem-size limit,
we showed a prescription to obtain the most probable states in the distributions
in which higher order terms for the inhomogeneity of $\mu_i^2$ and $\nu_i^2$ are neglected.
For two rather simple problems,
i.e., fully-connected ferromagnetic coupling without/with binary random fields,
we derived phase diagrams in the $p$--$\xi J$ plane.
The phase diagrams demonstrate that
the most probable states in a particular range of the parameters correspond to the true optimal states.
In particular, in the random-field problem,
the distribution correctly detects the phase transition
that occurs in the genuine Ising model with varying the magnitude of the fields.
We found through this analysis an interesting feature of our system that
despite the nature of the first-order phase transition in terms of the Ising spins is effectively detected,
the order parameter $m$ for the quadrature amplitude does not show discontinuous change.
Our analysis is based on the approximate steady state distributions,
but it is probably valid for the no-field problem,
since the problem has no effect
that disturbs the uniformity in the magnitudes.
On the other hand,
the random fields contribute to the growth of the difference in the magnitude.
Hence,
our result of the random-field problem, in particular, for the fields with large amplitude might be modified
when including the higher order terms for the inhomogeneity of the magnitude.
Numerical simulations for this problem should be done to examine this issue.
In addition, it is interesting to estimate effects of the higher order terms to the most probable states.
The obstacle that makes it difficult to derive the exact steady state distributions is the violation of the detailed-balance condition in the dynamics.
This is not just an obstacle but a property
that has us expect faster convegence to steady states than
the conventional dynamics fluctuated by the simple Gaussian noise~\cite{H.Suwa2010, K.Turitsyn2011, H.Fernandes2011, Y.Sakai2013, A.Ichiki2013, M.Ohzeki2015Jul, M.Ohzeki2015Sep}.
The violation stems from the multiple facts;
the multiplicative noise and the coupling of the DOPOs.
The former, generated by the coupling of the signal and pump fields,
also appears in a single DOPO,
but it is not enough for the breakdown of the detailed balance.
The coupling of the DOPOs is thus essential to the peculiar dynamics.
The dynamics of the DOPOs with this property might provide an advantage of the CIM in solving problems.
Investigation on this subject is left as a future work.
|
\section{Introduction}
Purposeful memory is one of the essential ingredients that serves us to distinguish the behavior of higher organisms from that of simpler (living or non-living) entities; this is an idea which has been recognized for very long and has represented a matter of debate in science, psychology and epistemology for decades. Nowadays, quantitative approaches to animal and human behavior have progressively become a field of active research in biological physics. The study of neuronal patterns in the brain with the help of imaging techniques and its modelization through mathematical tools from network theory and/or population dynamics represent a prime example \cite{chialvo10,papo17}, but many other could be cited.
In particular, for those behavioral processes related to dispersal and/or navigation of living beings, it is clear that random walks represent a convenient description in order to condense and/or account for the properties of real trajectories and individual/collective space-use \cite{codling08,mendez14,bartumeus16}. Simpler approaches within this area just neglect any effects from spatial memory, so the position of the individual walker $X(t)$ is then a Markovian random variable. However, a rich bibliography on random-walks with memory rules of different sort has been developed throughout the years and connected to biological movement, too. Some examples of this include models in which walkers tend to exhibit locally some kind of directional memory or persistence (these could be loosely termed as \textit{annealed} memory models); this is the case of Persistent random walks \cite{holmes93,selmeczi05,codling08}, Continuous-Time random walks (CTRW) \cite{shlesinger18}, L\'{e}vy Walks \cite{viswanathan96,ariel15}, etc. On the other side, \textit{quenched} memory models represent a more complex situation in which local information about the sites visited is stored by the walker somehow, and so future rules of advance will explicitly depend on it, which in general makes the process highly non-Markovian. Some well-known frameworks falling within this class are the different versions and generalizations of the self-avoiding random walk (though originally this model was proposed to describe polymer growth, not biological movement) \cite{freund92}, elephant random walks (which were probably the first class of solvable models proposed with long-range memory effects) \cite{schutz04}, or, more recently, models with preferential relocation to already visited sites \cite{boyer14,falcon17} or preferential persistence for familiar paths \cite{abramson14,kazimierski15}.
Though a large amount of realism can be gained by introducing memory in the description of animal and human movement, it is clear that mathematical treatment becomes then cumbersome due to its non-Markovian character. First-passage and coverage properties, for instance, of non-Markovian processes represent a formidable problem for which it is very difficult to extract analytical results unless additional assumptions are considered \cite{guerin16}. For the case of \textit{annealed} memory some exceptions can be found, like those works where non-stationary random walk patterns are considered \cite{campos16}, but for \textit{quenched} memory it is very difficult to find references in the literature where this has been even addressed; we can cite the recent work by Kearney and Martin \cite{kearney16} on the first-passage properties of P\'{o}lya urns and their connection to random walks as one of the few exceptions.
Taking all this context into account, our aim is to propose the study of first-passage and recurrence times statistics for \textit{quenched} memory random walks roughly describing the movement of biological organisms with different memory (cognitive) abilities. A whole analytical treatment of such situations, as mentioned above, is in general unattainable but still some general properties of interest can be derived in many situations numerically. If $\Theta_n$ represents a random variable describing the time at which the individual hits a site for $n$-th time, then we will focus here on the random variables
\begin{eqnarray}
\nonumber T_1 &=& \Theta_1 \\
\nonumber T_n &=& \Theta_n-\Theta_{n-1}, \quad n>1
\end{eqnarray}
representing the random times between consecutive hittings. So, a consequence of introducing memory effects in the trajectories could be the emergence of correlations between $T_1,T_2,\ldots$. Note that all these variables would be independent for the simplest case of Markovian random walks, since memory is lost after hittings. So, characterizing the correlations between these recurrence times could provide a way to classify memory random walks.
As a first step towards this aim, we here study the correlation properties between first-passage and successive recurrence times in a generalized (continuous-time) version of the random-walk with preferential relocation to visited sites \cite{boyer14,boyer14b}. This preferential relocation process, as reported in the original work, tries to capture some basic properties of foraging in higher animals driven by a tendency to revisit sites where resources (e.g. food,...) have been successfully detected previously (so assuming that these resources are never depleted and/or can be replenished in a relatively short time). Hence, this process is obviously expected to yield a positive correlation between the variables $T_n$, though this idea has never been explored previously as far as we know.
The present article is structured as follows. In Section \ref{model} we derive the Continuous-Time Random Walk master equation for random walks with preferential relocation to visited sites, and justify the interest of such generalization if compared to the original (time-discrete) version. In Section \ref{dispersal}, we study the dispersal properties through the Mean Square Displacement (MSD) of the corresponding random walkers to check that our results are in agreement with those from the original model. In section \ref{passage} we formally present the hitting and recurrence problem for this case and provide analytical justification to support the existence of positive correlations between first-passages and sussequent recurrence times for a particularly simplified case. Also, we show results from Monte Carlo simulations in order to understand the main properties of the process. Finally, the conclusions from our study are presented in Section \ref{conclusions}.
\section{Continuous-Time model} \label{model}
\subsection{CTRW framework}
We start by revising briefly the classical CTRW formalism to facilitate understanding of the model presented below. The CTRW is based on the idea that the walker performs jumps of random (i.i.d.) lengths separated by random (i.i.d.) waiting times. A possible mathematical derivation is based on the combination of two balance equations. The first one states that the probability of reaching a position $x$ at the $m$-th step, denoted by $j_m(x)$, satisfies the mesoscopic balance equation
\begin{equation}
j_m(x)= \int_{-\infty}^{\infty} j_{m-1}(x-z) \Phi(z) dz + \delta_{m,0} \delta (t) j_0,
\end{equation}
where $\Phi(x)$ is the jump kernel, which determines the probability distribution function of the jump lengths, and the last term stands for the initial condition (where $\delta_{m,0}$ is a Kronecker Delta function, $\delta (t)$ is a Dirac Delta function, and $j_0=\delta(x-x_0)$ with $x_0$ the position of the site occupied at $t=0$).
One can also include time explicitly within this expression to write
\begin{eqnarray}
j_m(x,t)&=& \int_{0}^{t} \int_{-\infty}^{\infty} j_{m-1}(x-z,t-\tau) \varphi(\tau) \Phi(z) dz d\tau \nonumber \\
&+& \delta_{m,0} \delta (t) j_0,
\label{eq1}
\end{eqnarray}
where $j_m(x,t)$ corresponds to the probability that the $m$-th step is done to position $x$ at time $t$ (so $j_m(x)=\int_{0}^{\infty} j_m(x,t)dt$) and $\varphi(t)$ is the probability distribution function of waiting times between consecutive jumps.
A second equation is introduced through the probability of being at position $x$ at time $t$ after $m$ steps have been made, $p_m(x,t)$; this expression reads
\begin{equation}
p_m(x,t) = \int_0^{t} j_m(x,t-\tau) \phi(\tau) d\tau ,
\label{eq2}
\end{equation}
where $\phi(\tau)$ is the probability that the walker has not jumped in a time $\tau$ since it arrived to $x$, so this satisfies $\phi(t)=\int_{t}^{\infty} \varphi(\tau) d \tau$. So, by combining (\ref{eq1}) and (\ref{eq2}) one obtains the master equation which contains the statistical properties of the CTRW. In particular, if we sum the equations (\ref{eq1}) and (\ref{eq2}) for $m$ from $0$ to $\infty $ (with $j_m=0$ for $m<0$) we recover the well-known master equation of the CTRW as it appears in the books on the subject \cite{mendez14}:
\begin{equation}
j(x,t)= \int_{0}^{t} \int_{-\infty}^{\infty} j(x-z,t-\tau) \varphi(\tau) \Phi(z) dz d\tau + \delta (t) j_0
\end{equation}
\begin{equation}
p(x,t) = \int_0^{t} j(x,t-\tau) \phi(\tau) d\tau
\label{prop}
\end{equation}
\subsection{Preferential relocation to visited sites}
\label{preferential}
Now we include in the CTRW framework the possibility that the walker can use memory to return to previously visited sites. While the original model with preferential relocation rules was discrete both in time and space \cite{boyer14}, a continuous generalization has already been proposed in a different context \cite{boyer17}, though a mesoscopic derivation was not provided there as we do in the following. In particular, a generalization within the CTRW framework allows one the possibility to consider in a natural way different waiting time distributions, including those with power-law tails (leading to a L\'{e}vy statistics) or with a combination of different dispersal modes, for example.
We will consider that at the end of each waiting time the particle can either (i) decide to do a random jump governed by the kernel $\Phi(x)$ with probability $\alpha$ (in the following, we denote this as the \textit{normal} transport mode), or (ii) use its memory and then \textit{fly} instantaneously to a previously visited site with probability $1-\alpha$ (in the following, this is termed as \textit{memory} transport mode).
If we denote the memory kernel (it is, the probability to remember a site visited $i$ jumps ago) as $K_m(i)$, with $\sum_{i=1}^{m} K_m(i) =1$ for any $m$, then we can write
\begin{equation}
j_m(x,t)= \alpha \int_{0}^{t} \int_{-\infty}^{\infty} j_{m-1}(x-z,t-\tau) \varphi(\tau) \Phi(z) dz d\tau + (1-\alpha) \sum_{i=1}^{m} K_m(i) \left[ j_{m-i}(x,t) \ast \varphi(t) ^{\ast i} \right] ,
\end{equation}
where the asterisk symbol $\ast$ denotes the time-convolution operator, and $\varphi(t) ^{\ast i}$ denotes the time-convolution of the distribution $\varphi(t)$ with itself $i$ times.
From now on, for the sake of simplicity we will use a memory kernel that gives the same weight to all previously visited sites, so $K_n(i)=1/m$, in agreement with the original model \cite{boyer14}. For that case, the previous equation becomes
\begin{equation}
j_m(x,t)= \alpha \int_{0}^{t} \int_{-\infty}^{\infty} j_{m-1}(x-z,t-\tau) \varphi(\tau) \Phi(z) dz d\tau + \frac{1-\alpha}{m} \sum_{i=1}^{m} \left[ j_{m-i}(x,t) \ast \varphi(t) ^{\ast i} \right] .
\end{equation}
We can apply Fourier-Laplace transforms to the previous equation to take advantage of the renewal property of the process in time and space,
\begin{equation}
j_m(k,s)= \alpha j_{m-1}(k,s) \varphi(s) \Phi(k) + \frac{1-\alpha}{m} \sum_{i=1}^{m} j_{m-i}(k,s) \varphi(s)^{i}.
\label{jnks}
\end{equation}
Note that we use $k$ and $s$ as the Fourier and Laplace arguments. So that, $j_m(k,s)$ represents the Fourier-Laplace transform of $j_m(x,t)$, and we will distinguish them just by explicitly writing their variables.
Now, due to the explicit dependence of the memory factor on $m$, it is also convenient to carry out a Z-transform on the jump index $m$ such that
$$\hat{j}(\lambda,k,s)=\sum_{m=0}^{\infty} \lambda^{m} j_m(k,s),$$
with $0<\lambda <1$; by applying this on Eq. (\ref{jnks}) we obtain
\begin{equation}
\hat{j}(\lambda,k,s)= \alpha \lambda \hat{j}(\lambda,k,s) \varphi(s) \Phi(k) + (1-\alpha) \int_{0}^{\lambda} \frac{ \varphi(s) \hat{j}(u,k,s)}{1-u \varphi(s)} du.
\label{jlks}
\end{equation}
Due to the integral that appears now in the equation one needs to rewrite this expression as a differential equation by differentiating with respect to $\lambda$ and replacing the integral in the resulting equation with the help of Eq. (\ref{jlks}). This procedure leads to
\begin{equation}
\frac{d\hat{j}(\lambda,k,s)}{d\lambda} = \hat{j}(\lambda,k,s)\left[ \frac{\alpha \varphi(s) \Phi(k)}{1-\alpha \lambda \varphi(s) \Phi(k)} + \frac{(1-\alpha) \varphi(s)}{(1-\lambda \varphi(s)) (1-\alpha \lambda \varphi(s) \Phi(k))} \right] .
\end{equation}
The solution of this first-order differential equation reads
\begin{equation}
\hat{j}(\lambda,k,s)= \left[ \frac{1}{1-\alpha \lambda \varphi(s) \Phi(k) } \right]^{\frac{\alpha[1-\Phi(k)]}{1-\alpha \Phi(k)}} \left[ \frac{1}{1- \lambda \varphi(s)} \right] ^{\frac{1-\alpha}{1-\alpha \Phi(k)}},
\label{jlks2}
\end{equation}
where we have taken into account the boundary condition $\hat{j}(\lambda \rightarrow 0^{+},k,s)=1$.
Finally, taking into account that we are interested in the behavior of $j(k,s)$ (independent of $m$) we can remove the dependence on the jump index by using $j(k,s) \equiv \hat{j} (\lambda=1,k,s)=\sum_{m=0}^{\infty} j_m(k,s) $. On setting $\lambda=1$ into Eq. (\ref{jlks2}) we have
\begin{equation}
j(k,s)= \left[\frac{1}{1-\alpha \varphi(s) \Phi(k) } \right]^{\frac{\alpha[1-\Phi(k)]}{1-\alpha \Phi(k)}} \left[\frac{1}{1- \varphi(s)}\right] ^{\frac{1-\alpha}{1-\alpha \Phi(k)}}.
\label{jks}
\end{equation}
This expression, together with (\ref{prop}) (which is still valid for the model with memory), represent the CTRW generalization of the random walk with preferential relocation to visited places. Next, we will use this result to derive several properties of the model.
\section{Dispersal properties}
\label{dispersal}
First we will explore the dispersal properties of the model in order to check that they agree with those found for the time-discrete version \cite{boyer14}. For this, we will compute the MSD $\langle x^2(t)\rangle$, which is nothing but the second moment of $p(x,t)$ in space. So that, working again in Fourier-Laplace space
\begin{equation}
\langle x^2(s)\rangle =-\lim _{k\rightarrow 0}\frac{\partial ^2 p(k,s)}{\partial k^2}=-\frac{1-\varphi(s)}{s}\lim _{k\rightarrow 0}\frac{\partial ^2 j(k,s)}{\partial k^2},
\label{msd1}
\end{equation}
where we have made use of Eq. (\ref{prop}). Performing the second derivative of $j(k,s)$ we find, after some tedious calculations,
\begin{equation}
\langle x^2(s)\rangle =- \frac{\alpha}{1- \alpha} \frac{\Phi''(0)}{s} \ln \left[\frac{1-\alpha\varphi(s))}{1- \varphi(s)} \right],
\end{equation}
where $\Phi''(0)$ stands for the second derivative of $\Phi(k)$ evaluated at $k=0$. If the waiting time distribution has finite moments then we can make use of the expansion $\varphi(s) \simeq 1-\tau s+...$ for large times, with $\tau$ the mean waiting time. Hence, using this expansion and assuming $s\rightarrow 0$ we obtain
\begin{equation}
\langle x^2(s)\rangle =\frac{\alpha}{1- \alpha} \Phi''(0)\frac{\ln (\tau s)}{s},
\label{msdf}
\end{equation}
which yields
\begin{equation}
\langle x^2(t)\rangle =-\frac{\alpha}{1- \alpha} \Phi''(0)\ln (t/\tau)\quad \text{for}\quad t\rightarrow \infty
\label{msdf2}
\end{equation}
after inverting by Laplace. This result is very general and holds for any dispersal kernel and waiting time distributions with finite moments and predicts a ultra-slow diffusion due to the logarithmic growth of the MSD \cite{MeIo,iomin16}. If for example the dispersal kernel is Gaussian $\Phi(x)=(2\pi\sigma^2)^{-1/2}\exp (-x^2/2\sigma ^2)$
the asymptotic form of the MSD is given by
\begin{equation}
\langle x^2(t)\rangle =\frac{\alpha}{1- \alpha} \sigma ^2\ln (t/\tau)\quad \text{for}\quad t\rightarrow \infty.
\label{msdf3}
\end{equation}
In Figure \ref{fig1} we confirm this result via direct comparison with Monte Carlo simulations of the random walk process described in Section \ref{preferential} (points) for different $\alpha$ values. We find that not only the ultra-slow (logarithmic) character of the dispersal is observed, but the prefactor $\alpha \sigma^2 / (1-\alpha)$ predicted in (\ref{msdf}) (governing the slope of the solid lines in the plot) fits almost perfectly the behavior of the simulated process. This serves as a checking of the validity of our approach prior to the analysis of the recurrence statistics we carry out in the next Section.
\begin{figure}[htbp]
\includegraphics[scale=1.0]{figure1_new.eps}
\caption{MSD for the CTRW model with preferential relocation to visited places. Symbols represent the results from Monte Carlo simulations averaged over $10^5$ realizations, while lines correspond to the asymptotic prediction for a Gaussian kernel from the model in Eq. (\ref{msdf3}) (appropriately shifted to facilitate visualization). For the waiting time distribution $\varphi(t)$ an exponential distribution with $\tau =1$ has been used.}
\label{fig1}
\end{figure}
\section{Recurrence statistics}
\label{passage}
\subsection{Theoretical framework}
\label{theory}
First-passage times and the statistics of recurrence times for Markovian random walks can be derived in general for finite domains with the help of renewal properties \cite{condamin05}, but this no longer holds for non-Markovian walks \cite{guerin16}. While the problem is analytically unattainable for the model with preferential relocation presented in Section \ref{model}, some results can still be derived by simplifying the model considerably.
In particular, we will use the assumption that the system is finite (with $N$ denoting the number of accessible sites) and it is spatially unconstrained (this is, the kernel $\Phi(x)$ is such that any site is accessible from any other with the same probability $1/N$). This is not unreasonable in the context of the preferential relocation model we are considering here: since the \textit{memory} mode can lead the individual to any visited site without restrictions, then it may be plausible to consider that the \textit{normal} transport mode has this spatial capacity, too. In Section \ref{numerical}, however, we will relax this assumption to check how our numerical results change when only jumps of a given maximum length are permitted.
\subsubsection{Case $n=1$}
To illustrate our method, we first derive a formal expression for the first-passage distribution, following a similar strategy to that in Section \ref{preferential} based on discretizing the process according to the jump index $m$. We denote by $S_{1}(t)$ the survival probability through a target site $x^{*}$ (this is, the probability that after time $t$ that site has not been visited yet; note that we do not consider an explicit dependence on $x^{*}$ due to the assumption of unconstrained space). Then, this probability can be decomposed through
\begin{equation}
S_{1}(t) = \sum_{m=0}^{\infty} S_{1}^{(m)} P(m \vert t) = \sum_{m=0}^{\infty} S_{1}^{(m)} \left[ \varphi(t) ^{\ast m} \ast \phi(t) \right] ,
\label{fpt1}
\end{equation}
where $S_{1}^{(m)}$ is the survival probability after $m$ jumps, and $P(m \vert t)$ is the probability of having performed $m$ jumps at time $t$. The latter is given by the convolution of $m$ times the waiting time distribution (plus the convolution with $\phi(t)$, which is necessary to assert that the $(m+1)$-th jump has not been done yet).
Since the \textit{memory} transport mode does not contribute to the first-passage time (as the target site has not been visited yet), the survival probability after $m$ jumps is
\begin{equation}
S_{1}^{(m)}=(1- \alpha /N)^{m},
\label{fpt1b}
\end{equation}
making use of the unconstrained space assumption. This result follows since the probability to find the target at any jump is just $\alpha /N$. Transforming (\ref{fpt1}) to the Laplace space one finds
\begin{equation}
S_{1}(s) = \frac{\phi(s)}{1- \left( 1- \alpha /N \right) \varphi(s)}.
\label{fp2}
\end{equation}
So that, one can write as usual the first-passage distribution $f_1(t)$ as the time derivative of $S_{1}(t)$ or, alternatively, $f_1(s)=sS_{1}(s)-1$. The first moment of $f_1(t)$, $\langle T_1 \rangle = \int_0^{\infty} t f_1(t) \text{d}t$, provides the mean first-passage time, which reads then
\begin{equation}
\langle T_1 \rangle = \lim_{s \rightarrow 0} \frac{\text{d}f_1(s)}{\text{d}s} = \lim_{s \rightarrow 0} \left[ s \frac{\text{d}S_{1}(s)}{\text{d}s} + S_{1}(s) \right]
\label{t1}
\end{equation}
So, if we assume again that $\varphi(t)$ has finite moments (with $\varphi(s) \approx 1-\tau s+ \ldots$), then we reach from (\ref{fp2}) and (\ref{t1})
\begin{equation}
\langle T_1 \rangle = \frac{N \tau}{\alpha}
\label{t1b}
\end{equation}
and, similarly, for the second order moment
\begin{equation}
\langle T_1^2 \rangle = 2 \left( \frac{N \tau}{\alpha} \right) ^2.
\label{secondt1}
\end{equation}
These results are to be expected. In the absence of contributions from the \textit{memory} mode the process is Markovian, so (\ref{t1b}) simply represents the Wald's identity \cite{redner01} for a stochastic process with constant rate $\alpha /\tau N$ (which is nothing but the rate at which the target site will be reached through the \textit{normal} mode).
\subsubsection{Case $n > 1$}
Now we can extend the procedure above to study subsequent recurrence times ($T_2,T_3,\ldots$, as we defined them in the Introduction) through a site. For this, we introduce the distribution function $f_n(t)$ for the random variable $T_n$ and, equivalently, the joint distribution $f_n(t;m_{n-1})$ for $T_n$ and the random variable $m_{n-1}$, which is the number of jumps done by the individual (counting since $t=0$) when it hits the site for the $(n-1)$-th time (this is, the number of jumps performed when $t=\Theta_{n-1}$). Then we can write
\begin{eqnarray}
\nonumber \langle T_n \rangle &=& \sum_{m_{n-1}=0}^{\infty} \int_{0}^{\infty} t f_n (t;m_{n-1}) \text{d}t \nonumber \\
&=&
\sum_{m_{n-1}=0}^{\infty} \int_{0}^{\infty} t f_n (t \vert m_{n-1}) q_{n-1}(m_{n-1}) \text{d}t \nonumber \\
&=&-\sum_{m_{n-1}=0}^{\infty} q_{n-1}(m_{n-1}) \int_{0}^{\infty} t \frac{\text{d} S_n (t \vert m_{n-1})}{\text{d}t} \text{d}t, \nonumber \\
\label{npt}
\end{eqnarray}
where in the second step we have introduced, through the Bayes theorem, the conditional probability $f_n (t \vert m_{n-1})$ and the probability distribution function $q_{n-1}(m_{n-1})$ for the variable $m_{n-1}$. Finally, in the last step we have rewritten the conditional distribution $f_n (t \vert m_{n-1})$ in terms of the conditional survival probability $S_n (t \vert m_{n-1})$.
The reason for writing the mean recurrence time in that way is because for each recurrence process one can provide explicit expressions equivalent to (\ref{fpt1}) and (\ref{fpt1b}) for the first-passage, respectively:
\begin{equation}
S_n(t \vert m_{n-1}) = \sum_{m=0}^{\infty} S_{n \vert m_{n-1}}^{(m)} P(m \vert t) = \sum_{m=0}^{\infty} S_{n \vert m_{n-1}}^{(m)} \left[ \varphi(t) ^{\ast m} \ast \phi(t) \right]
\label{npt1}
\end{equation}
and
\begin{equation}
S_{n \vert m_{n-1}}^{(m)} = \prod_{i=0}^{m-1} \left( 1- \frac{\alpha}{N} - \frac{(1- \alpha) (n-1)}{m_{n-1}+i} \right).
\label{npt1b}
\end{equation}
Note first of all that both expressions for the survival probabilities reduce to (\ref{fpt1}) and (\ref{fpt1b}) for $n=1$ (since $m_0=0$ and so the conditional probability is unnecessary in that case). Now, we observe that the \textit{memory} mode explicitly contributes to the survival probability through the last term within the parenthesis of (\ref{npt1b}). So, the probability that the \textit{memory} mode leads the individual to the target is given by $(n-1)/(m_{n-1}+i)$, where $(m_{n-1}+i)$ is the total number of jumps performed up to date (it is, those done up to the $(n-1)$-th hitting, $m_{n-1}$, plus those done afterwards, $i$).
As a whole, the expressions (\ref{npt}-\ref{npt1b}) provide a recurrent method to determine the statistics of recurrence times as follows. First, once we know the properties of the first-passage time, we can use them to determine $q_1(m_1)$, which by definition satisfies
\begin{equation}
q_1(m_1)=S_1^{(m_1)}-S_1^{(m_1-1)}.
\label{q1}
\end{equation}
This, in combination with (\ref{npt}-\ref{npt1b}) for $n=2$ will be used to determine $f_2 (t;m_{1})$ and its mean value $\langle T_2 \rangle$. Then we will be able to determine $q_2(m_2)$ from
\begin{equation}
q_n(m_n)= \sum_{m=0}^{m_n-1} q_{n-1}(m) \left[ S_{n \vert m}^{(m_n-m)} - S_{n \vert m}^{(m_n-m-1)} \right].
\label{qn}
\end{equation}
and the same idea can be applied recurrently for $n=3,4,\ldots$
The previous method, obviously, will become increasingly cumbersome as $n$ increases, so at practice we can only expect it to be of practical utility for $n$ small. In the next subsection we will illustrate its use for $n=2$, which is enough for the specific objectives we pursue in this paper.
\subsection{Recurrence time for $n=2$}
\label{approx}
Although the method described above could be applied to any waiting time distribution, to keep notation and results manageable we will focus here on the case of exponential waiting times, $\varphi(t)= \tau ^{-1} e^{-t / \tau}$. Note that in this case the random walk would become Markovian in absence of the preferential relocation rule (this is, for $\alpha=1$).
We already know that the mean first-passage time is determined by (\ref{t1b}). Also, introducing (\ref{fpt1b}) into (\ref{q1}) one has
\begin{equation}
q_1(m_1)= \frac{\alpha}{N} \left( 1- \frac{\alpha}{N} \right) ^{m_1-1}
\label{n21}
\end{equation}
On the other side, the combination of Eqs. (\ref{npt1}) and (\ref{npt1b}) yields for this case
\begin{equation}
S_2(t \vert m_{1}) = \sum_{m=0}^{\infty} B_{m,m_1} \left( \frac{t}{\tau} \right) ^{m} \frac{ e^{-t/ \tau}}{\Gamma(m+1)},
\label{n22}
\end{equation}
where we have defined
\begin{equation}
B_{m,m_1} \equiv \prod_{i=0}^{m-1} \left( 1- \frac{\alpha}{N} - \frac{1- \alpha }{m_{1}+i} \right).
\label{n23}
\end{equation}
So, we finally insert (\ref{n21}-\ref{n22}) into (\ref{npt}) to determine the mean recurrence time after first hitting, $T_2$. By doing this and performing the integral in $t$ we get
\begin{equation}
\langle T_2 \rangle = \sum_{m_{1}=0}^{\infty} \frac{\alpha \tau}{N} \left( 1- \frac{\alpha}{N} \right) ^{m_1-1} \sum_{m=0}^{\infty} B_{m,m_1}.
\label{t2}
\end{equation}
This expression cannot be further simplified due to the product within the definition of $B_{m,m_1}$, but we can reach useful approximations in the large-domain limit $N \gg 1$, so the number of jumps required for the first-passage, $m_1$, is also large in average. Using such approximation, one can provide an analytical approximation for $\langle T_2 \rangle$ which can be expressed as a combination of exponential integral (or, alternatively, confluent hypergeometric) functions (see the discussion in the Appendix for the details).
Also, note that the same idea can be applied to higher order or cross moments (e.g. $\langle T_2^{2} \rangle$ or $\langle T_1 T_2 \rangle$) (see Appendix). This allows us to provide an estimation for the correlation coefficient between first-passage and subsequent recurrence times
\begin{equation}
\text{Cor} (T_1,T_2) \equiv \frac{\langle T_1 T_2 \rangle - \langle T_1 \rangle \langle T_2 \rangle}{ \sigma_1 \sigma_2},
\label{cor}
\end{equation}
where $\sigma_i$ denotes the standard deviation of $T_i$. The goodness of that estimation will be checked through the comparison with results from Monte Carlo simulations in the next section.
\subsection{Numerical results}
\label{numerical}
The first results we show from our Monte Carlo simulations try to verify the validity of the approximations carried out in the previous Section.
\begin{figure}
\includegraphics[scale=1.0]{figure2_new.eps}
\caption{Comparison for the (a) mean recurrence time after first hitting, and (b) correlation coefficient between first hitting and first recurrence times as a function of the memory parameter $\alpha$, between results obtained from Monte Carlo simulations (symbols) averaged over $10^5$ realizations, and the analytical approximation reported in Section \ref{approx} (solid line). Different Results are shown for the case without spatial constraints and an exponential waiting time distribution $\varphi(t)=\tau^{-1}e^{-t/ \tau}$ with $\tau=1$. The dotted line represents the heuristic approximation $\langle T_2 \rangle = \tau N / \alpha (2-\alpha)$.}
\label{fig2}
\end{figure}
In Figure 2 we see that our analytical approximations for $\langle T_2 \rangle$ and $\text{Cor} (T_1,T_2)$ (solid lines) fits rather well the results for $\alpha$ small but the approximation fails clearly as long as we approach the limit without memory, $\alpha=1$. Surprisingly, we find that a simple heuristic approximation is able to fit the behavior for $\langle T_2 \rangle$ accurately (dotted lines in Figure 2). Such approximation follows from assuming that the probability of reaching the target at each particular jump is constant and equals $\alpha /N + (1-\alpha) \alpha / N = \alpha (2-\alpha)/N$. Here, the two terms in the sum represent the contribution from the \textit{normal} and \textit{memory} modes, respectively. For the latter we are assuming that the individual revisits the target with probability $(1-\alpha)/\langle m \rangle $, where $\langle m \rangle$ is the mean number of jumps done previously, $N/ \alpha$ (as discussed in Section \ref{theory}). The rate at which the target is found will be then $\tau$ times the inverse of the overall probability, leading to the estimation $\langle T_2 \rangle = \tau N / \alpha (2-\alpha)$. The agreement found in Figure 2 between this approximation (dotted line) and the numerical results is clear. However, this kind of approximation cannot be extended to cross moments (e.g., $\langle T_1 T_2 \rangle$) and so it is not helpful to obtain an estimation of the correlation coefficient, as we pursue here.
Regarding the correlation coefficient, Figure 2 shows that the tendency predicted by our analytical approach is approximately correct (except, again, in the limit $\alpha \rightarrow 1$), but the accumulated error in the estimation of $\langle T_2 \rangle$, $\langle T_2^2 \rangle$ and $\langle T_1 T_2 \rangle$ makes that the quantitative agreement is not completely satisfactory. In any case, this plot confirms the main idea of the present work, which is the fact that consecutive hitting times become positively correlated as a consequence of the preferential relocation rule.
\begin{figure}
\includegraphics[scale=1.0]{figure3_new.eps}
\caption{Results for the (a) mean recurrence time after first hitting and (b) correlation coefficient between first-passage and subsequent recurrence times as a function of $n$, obtained from Monte Carlo simulations averaged over $10^5$ realizations. Different values of the memory parameter $\alpha$ are reported (see legends). The solid line in the lower panel is just a visual cue to emphasize the logarithmic character of the decay.}
\label{fig3}
\end{figure}
Most of the conclusions above for $n=2$ can be extended to subsequent recurrence times ($n>2$), as reported in Figure 3. There we observe that the mean value of $T_n$ becomes progressively reduced as a function of $n$ due to the accumulated effect of memory, and also it can be checked that its behavior as a function of the memory parameter $\alpha$ is qualitatively the same as for $T_2$ (not shown). Furthermore, Figure 3b yields a very interesting result, as is the fact that correlations between first-passage and subsequent recurrence times $T_n$ show an ultra-slow (logarithmic, as for the MSD) relaxation to zero as a function of $n$. This tells us that the signature of strong memory induced by the preferential relocation rule does not only emerge at the level of dispersal (as was already known from previous works \cite{boyer14,boyer17}) but also on the recurrence statistics.
\begin{figure}
\includegraphics[scale=1.0]{figure4.eps}
\caption{Results for the (a,b) mean recurrence time after first-hitting and (c,d) correlation coefficient between first-passage and subsequent recurrence times as a function of $n$, obtained from Monte Carlo simulations, for the case where the \textit{normal} mode imposes a constraint given by a maximum jump length of size $L$. Two values of $L$ and different values of the memory parameter $\alpha$ are reported (see legends) for the sake of completeness.}
\label{fig4}
\end{figure}
Finally, to complete the numerical analysis we provide results for the case where individuals are not allowed to jump freely from any site to another, but only short jumps are to be expected.
So that, we reformulate the dispersal process by assuming that the individual can only do jumps up to a maximum distance $L$ (this rule only applies to the \textit{normal} mode, while the \textit{memory} mode is kept unchanged and without spatial constraints). The results for $L=N/4$ and $L=N/8$ are provided in Figure 4, where the initial position is chosen at random at each realization of the Monte Carlo simulation.
Note that the spatial constraint increases considerably the values of the first-passage and recurrence times (since now further sites become increasingly difficult to be reached due to the ultra-slow dispersal properties of the model). So, low values of $L$ become computationally very costly. Apart from that, we observe that the spatial constraint does not modify qualitatively the picture found in the Figures 2 and 3. In particular, positive correlations between first-passage and recurrence times are still present and actually become inforced.
\section{Conclusions}
\label{conclusions}
Summarizing the ideas reported in the work, we have confirmed that one of the essential signatures of random walks with memory (in this case, we have just focused on the model with preferential relocation to visited places) is the existence of correlations between first-passage and subsequent recurrence times through a site or, equivalently, between consecutive recurrence times. Furthermore, we have been able to characterize such correlations not only numerically but also through an approximated analytical study, at least for the case $n=2$.
Anyway, the most interesting result we obtain is probably (FIgure 3) that the decay of these correlations with time for the case of walks with preferential relocation exhibits an ultra-slow (logarithmic) behavior, in accordance with the dispersal properties of the model which were already known from previous works \cite{boyer14,boyer14b}. Since memory effects persist in the model for arbitrarily long times, such long-range dynamics is also present in the hitting and recurrence statistics. This suggests that such correlations capture adequately the memory dynamics in the model and so can be used as a proxy to identify what kind of memory rules govern the process. This could be of great interest, for example, in the analysis of trajectories of real organisms as a method to understand how memory has been employed during the process.
\section*{Acknowledgements} This research has been supported by the Spanish government through Grants No. CGL2016-78156-C2-2-R and FIS2015-72434-EXP.
\section{APPENDIX. Derivation of the moments of $T_2$}
\label{appendix}
We start from the definition of the coefficients $B_{m,m_1}$, given implicitly in (\ref{n22}), and express them in a more convenient form:
\begin{equation}
B_{m,m_1} = \left[ \gamma (m_1) \right] ^m \frac{\left( 1+ \frac{1}{m_1-\beta} \right) \ldots \left( 1+ \frac{n-1}{m_1-\beta} \right) }{\left( 1+ \frac{1}{m_1} \right) \ldots \left( 1+ \frac{n-1}{m_1} \right)},
\end{equation}
where we have defined $\beta \equiv (1-\alpha)/(1-\alpha/N)$ and $\gamma (m_1) \equiv (1-\alpha /N)(1-\beta/m_1)$. In this way it is clear that we can propose an expansion for $m_1 \gg 1$ in the form
\begin{eqnarray}
B_{m,m_1} &=& \left[ \gamma (m_1) \right] ^m \left( 1+\sum_{i=1}^{m-1} \frac{v_i(m)}{(m_1-\beta)^i} \right) \nonumber\\
&\times &\left( 1-\sum_{i=1}^{m-1} \frac{w_i(m)}{(m_1)^i} \right),
\end{eqnarray}
with the first coefficients $v_i$, $w_i$ given by
\begin{eqnarray}
\nonumber v_1(m) &=& w_1(m) = \frac{m (m-1)}{2} \\
\nonumber v_2(m) &=& \frac{m (m-1) (m-2) (3m-1)}{72} \\
\nonumber w_2(m) &=& v_2(m)-v_1^{2}(m) \\
\ldots
\end{eqnarray}
Leading this expansion up to second order (this is, up to powers of order $m_1^{-2}$) and summing over all values of $m=0$ we obtain, after some lengthy algebra,
\begin{eqnarray}
\nonumber \sum_{m=0}^{\infty} B_{m,m_1} &=& \frac{1}{1-\gamma(m_1)} + \frac{ (1-\alpha) \gamma (m_1)}{m_1^2 (1-\gamma(m_1))^3} - \\
&-& \frac{\beta \left( 1+ \gamma (m_1) \right) }{m_1^3 (1- \gamma(m_1))^4} + \frac{\beta^2 (1+2 \gamma (m_1)) }{m_1^4 (1-\gamma (m_1))^5}\nonumber \\
&+& \ldots
\label{bm}
\end{eqnarray}
In order to apply (\ref{t2}) we use the fact that in the limit of large media size ($N \gg 1$, $m_1 \gg 1$) the sum over $m_1$ can be adequately approximated by an integral
\begin{eqnarray}
\langle T_2 \rangle &=& \sum_{m_{1}=0}^{\infty} \frac{\alpha \tau}{N} \left( 1- \frac{\alpha}{N} \right) ^{m_1-1} \left( \sum_{m=0}^{\infty} B_{m,m_1} \right) \nonumber \\
&\approx& \frac{\alpha \tau}{N} \int_{0}^{\infty} e^{-\frac{\alpha}{N} u} \left( \sum_{m=0}^{\infty} B_{m,u} \right) \text{d} u.
\label{t2aprox}
\end{eqnarray}
So, by introducing (\ref{bm}) into (\ref{t2aprox}) one finally obtains an approximated expression for $\langle T_2 \rangle$. This result can be expressed as a combination of algebraic and exponential integral functions (the resulting expression is too long to be reproduced here).
Similarly, the second order moment of $\langle T_2 \rangle$ can be approximated using exactly the same procedure and leading to
\begin{equation}
\langle T_2^2 \rangle \approx \frac{2 \alpha \tau^2}{N} \int_{0}^{\infty} e^{-\frac{\alpha}{N} u} \left( \sum_{m=0}^{\infty} (1+m) B_{m,u} \right) \text{d} u.
\label{t2aproxb}
\end{equation}
and the same for the cross moment,
\begin{equation}
\langle T_1 T_2 \rangle \approx \frac{\alpha \tau^2}{N} \int_{0}^{\infty} u e^{-\frac{\alpha}{N} u} \left( \sum_{m=0}^{\infty} B_{m,u} \right) \text{d} u
\end{equation}
so the value of the correlation coefficient between $T_1$ and $T_2$, as defined in (\ref{cor}), can be estimated.
|
\section{Introduction}
\IEEEPARstart{T}{he} developments in Internet of Things (IoT) applications led to the demand to develop the near-sensor edge computation architectures \cite{shi2016edge}. The edge computing provides motivation to develop near-sensor data analysis that support non-Von Neumann computing architectures such as neuromorphic computing architectures \cite{vermesan2017internet,hoang2015will}. In such architectures, implementing the on-chip neural network learning remain as an important task that determines its overall effectiveness and use.
{There are several works proposing the implementation of memristive neural network with backpropagation algorithm in digital and mixed-signal domain domain \cite{8030206,8060399,burr2017neuromorphic,nn3,7966300,kim2017analog,tsai2018recent}. However, the analog learning circuits based on conventional backpropagation learning algorithm \cite{zhang2017memristor,negrov2017approximate,7966300,nn1,7867846} in memristive crossbars have not been fully implemented. } The implementation of such learning algorithm opens up an opportunity to create an analog hardware-based learning architecture. This would transfer the learning algorithms from the separate software and FPGA-based units to on-chip analog learning circuits, which can simplify and speed up the learning process.
Extending our previous work on analog circuits for implementing backpropagation learning algorithm \cite{ISCASs}, we present a system level integration of the analog learning circuits with that of traditional neuro-memristive crossbar array. We illustrate how this learning circuit can be used in different biologically inspired learning architectures, such as three layer Artificial Neural Networks (ANN), Deep Neural Networks (DNN) \cite{8060399, 7966300}, Binary Neural Network (BNN) \cite{8297384}, Multiple Neural Network (MNN) \cite{363444}, Hierarchical Temporal Memory (HTM) \cite{25} and Long-Short Term Memory (LSTM) \cite{6795963}.
{The algebraic and integro-differential operations of backprogation learning algorithm, which are difficult to accurately implement on a digital system, are available inherently on the analog computing system. Further, modern edge-AI computing
solutions warrant intelligent data processing at sensor levels, and analog system can reduce the demands for having high speed
data converters and interface circuits. The proposed analog backpropagation learning circuit
enables a natural on-chip analog neural network architecture implementation which is beneficial in terms of processing speed, reducing overall power and lesser complexity, comparing to digital counterparts.
}
{
The main contributions of this paper include the following.
\begin{itemize}
\item We introduce the complete design of analog backpropagation learning circuit proposed in \cite{ISCASs} with control switches sign control circuit and weight update unit.
\item We illustrate how the proposed backpropagation learning circuit can be integrated into different neuromorphic architectures, like DNN, BNN, MNN, LSTM and HTM.
\item We show the implementation of additional activation functions that are useful for various neuromorphic architectures.
\item We verified the proposed architecture for XOR problem, handwritten digits recognition and face recognition, and illustrate the effect of non-ideal behavior of memristors on the performance of the system.
\end{itemize}
}
{
This paper is organized into following sections:
Section \ref{s2} introduces the relevant background of the learning architectures and backpropagation algorithm. Section \ref{s3} describes the proposed architecture of the backpropagation and the circuit level design. Section \ref{s4} illustrates how the proposed backpropagation circuits can be integrated into different learning architectures. Section \ref{s5} contains the circuit and system level simulation results. Section \ref{s6} discusses advantages and limitations of the proposed circuits and introduces the aspects of the design that should be investigated in future, and Section \ref{s7} concludes the paper. There is also a supplementary Material that includes the expanded background information, explicit explanation of the proposed circuit, the device parameters of the main backpropagation circuit proposed in Section \ref{s3} and simulation results for the learning circuit performance.
}
\section{Background}
\label{s2}
\subsection{Learning algorithms and biologically inspired learning architectures}
Three main brain inspired learning architectures that we consider in this work are neural networks \cite{8060399, 8297384, 363444,krestinskaya2018neuro}, HTM \cite{25} and LSTM \cite{6795963}.
\subsubsection{Neural Networks}
There is a variety of the architectures and learning algorithms for the neural networks. In this work, we integrate analog learning circuits to three different types of artificial neural network \cite{jain1996artificial}: DNN \cite{8060399, 7966300}, BNN \cite{8297384} and MNN \cite{363444}. DNNs consist of many hidden layers and can have various combination of activation functions between the layers. Deep learning neural networks are useful for classification \cite{els1}, regression \cite{6932438}, clustering \cite{7010902} and prediction tasks \cite{8290807}. A neural network which uses any combination of binary weight or hard threshold activation functions is typically known as BNN. There have been several successful implementations of BNN algorithms in software \cite{557699, 125859, 6797577,5260357} and an attempt to implement it on hardware \cite{secco2018supervised, 8052915,bnn}. The analog hardware implementation of BNN system with learning remains as an open problem \cite{hubara2016binarized,bnn}. MNN is a systematic arrangement of the artificial neural networks that can process information from different data sources to perform data fusion and classification. Multiple neural network implies that the data from different data sources, such as various sensors, is applied to separate neural networks, and the output from each network is fetched into the decision network. This approach allows simplifying the complex computation processes, especially when there is a large number of data sources \cite{363444}. The analog hardware implementation of MNN is another new idea proposed in this paper.
\subsubsection{Hierarchical Temporal Memory}
HTM is a neuromorphic machine learning algorithm that mimics the information processing mechanism of neocortex in the human brain. HTM architecture is hierarchical and modular, and it enables sparse processing of information. HTM is divided into two parts: (1) Spatial Pooler (SP) and (2) Temporal Memory (TM) \cite{hawkinsintelligence,krestinskaya2018hierarchical}. The main purpose of the HTM SP is to encode the input data and produce its sparse distributed representation that finds application in various data classification problems. The HTM TM is primarily known for contextual analysis, sequence learning \cite{7727380} and prediction tasks \cite{bami2016, george2005hierarchical}. The HTM SP consists of four main phases: (1) initialization, (2) overlap, (3) inhibition, and (4) learning. There are several hardware implementations proposed for the HTM SP, such as conventional HTM SP \cite{fedorova2016htm} and modified HTM SP \cite{tcad}. Both architectures are based on memristive devices located in the initialization and overlap stages. The hardware implementation of the learning stage for HTM SP has not been proposed yet. According to \cite{25}, the backpropagation algorithm can be one of the approaches to updating weight in the HTM SP.
\subsubsection{Long-Short Term Memory}
LSTM is a cognitive architecture that is based on the sequential learning and temporal memory processing mechanisms in the human brain. LSTM processing relies on state change and time dependency of processed events. The LSTM algorithm is a modification of recurrent neural network that takes into account history of processed data and controls the information flow through gates \cite{olstm3,7508408}. LSTM is used in a wide range of applications in the contextual data processing based on prediction making and natural language processing. Hardware implementation of LSTM is a new topic studied in \cite{kam,smagulova2018design}.
\subsection{Backpropagation with gradient descent}
In this paper, an analog implementation of gradient descent backpropagation algorithm \cite{chauvin1995backpropagation,mehrotra1997elements} is proposed for different neural network configurations.
The algorithm consists of four steps: forward propagation, backpropagation to the output layer, backpropagation to the hidden layer, and weight updating process. In this section, we present the main equations of backpropagation algorithm with gradient descent for a three-layer neural network with sigmoid activation function to relate it to the proposed hardware implementation.
In the forward propagation step, the dot product of input matrix $X$ and weighted connections between input layer and hidden layer $w_{12}$ is calculated and passed through the sigmoid activation function: $Y_h=\sigma(X\cdot w_{12})$, where $Y_h$ is an output of the hidden layer \cite{7458870,8030932}. The forward propagation step is repeated in all the neural network layers. The output of the three-layer network $Y_o$ is calculated as $Y_o=\sigma(Y_h\cdot w_{23})$, where $w_{23}$ is the matrix representing the weighted connections between the hidden and output layers.
The backpropagation algorithm uses the cost function defined in Eq.\ref{eq3} for the calculation of derivative of error with respect to the change in weight. In Eq.\ref{eq3}, $E$ is an error, $N$ is a number of neurons in the layer, $y_{target}$ is an ideal output and $y_{real}$ is the obtained output after the forward propagation \cite{leonard1990improvement}.
\begin{equation}
E=\frac{1}{2}\sum_{i=1}^{N} (y_{target}-y_{real})^2
\label{eq3}
\end{equation}
Equation \ref{eq4} shows the calculation of the derivative of output layer error $E_o$, where $\delta$ denotes the rate of change of the error with respect to the weight $w_{23}$ \cite{7458870,8030932}. The error for the output layer $e$ is calculated as a difference between the expected neural network output $Y$ and real output of the network $Y_o$: $e=Y-Y_o$. The derivative of the sigmoid function is the following: $\frac{\partial Y_o}{\partial w_{23}} = Y_o(1-Y_o)$.
\begin{equation}
\frac{\partial E_o}{\partial w_{23}}=Y_h\cdot \delta _2 = Y_h \cdot (e \odot \frac{\partial Y_o}{\partial w_{23}})
\label{eq4}
\end{equation}
The derivative of error for the hidden layer is shown in Eq. \ref{eq5}, where $X'$ is an inverted input matrix and $e_h$ is the error of the hidden layer. The error $e_h$ is calculated propagating back $\delta_2$ as following: $e_h=\delta _2 \cdot w_{23}'$. And the derivative of the hidden layer output $E_h$ is the same as in the output layer: $\frac{\partial Y_h}{\partial w_{12}} = Y_h(1-Y_h)$ \cite{golden1996mathematical}.
\begin{equation}
\frac{\partial E_h}{\partial w_{12}}=X'\cdot \delta _1 = X' \cdot (e_h \odot \frac{\partial Y_h}{\partial w_{12}})
\label{eq5}
\end{equation}
In the final stage, the weight update calculation is performed using Eq. \ref{ee1} and Eq. \ref{ee2}, where $\eta$ is the learning rate responsible for the speed of convergence. The optimized learning rate depends on the type and number of inputs and number of the neurons in the hidden layers \cite{8030932, golden1996mathematical}.
{
\begin{equation}
\Delta w_{23} =\frac{\partial E_o}{\partial w_{23}} \times \eta
\label{ee1}
\end{equation}
\begin{equation}
\Delta w_{12} =\frac{\partial E_h}{\partial w_{12}} \times \eta
\label{ee2}
\end{equation}
}
The weight matrices are updated considering the calculated change in weight: $w_{23\_new}=w_{23}+\Delta w_{23}$ and $w_{12\_new}=w_{21}+\Delta w_{12}$.
\section{Backpropagation with memristive circuits}
\label{s3}
\subsection{Overall architecture}
This subsection illustrates the overall implementation of the backpropagation algorithm on hardware, while the details of the implementation of the activation functions and particular blocks are shown in Section \ref{sec3b}.
The proposed hardware implementation of the learning algorithm is illustrated in Fig. \ref{f1}. {Depending on the application requirements and limitations of the memristive devices, the inputs to the system can be either binary or non-binary. For example, for the HTM applications, the inputs are non-binary \cite{tcad}, whereas, for the binary neural network, inputs can be binary \cite{hubara2016binarized}. The outputs of the neural network can also be binary or non-binary, depending on the activation function.
Memristive crossbar arrays emulate the set of synapses between the neurons in the neural network layers. The synapses can also be binary or non-binary depending on the applications and practical limitations of programming states of memristor device. } While an ideal non-volatile memristor device can store and be programmed to any particular value between $R_{ON}$ and $R_{OFF}$, the real memristor devices can have problems with switching to the intermediate resistive values. It is easier and simpler to switch the memristor to either $R_{ON}$ and $R_{OFF}$ state. The implementation of the analog weights is also possible using 16-level $\textrm{Ge}_2\textrm{Sb}_2\textrm{Te}_5$ (GST) memristors \cite{xiao2017gst}. However, the memristor technology is not mature like CMOS, and even if the memristor can be precisely programmed and work accurately under the controlled environment in the lab, the behavior of the memristor in the multi-level large-scale simulation still needed to be verified.
Therefore, the binary synapses are the easiest to be implemented.
\begin{figure}
\centering
\includegraphics[width=90mm]{Untitled}
\caption{Overall architecture of the proposed analog backpropagation learning circuits for memristive crossbar neural networks. In the forward propagation process, MAIN BLOCK 1 (MB1) is involved. The backpropagation through the output layer is performed by MAIN BLOCK 2 (MB2) and MAIN BLOCK 4 (MB4). The backpropagation through the hidden layer is performed by MAIN BLOCK 3 (MB3). The weight update process of the output layer and the hidden layer is performed by MB4 and MB3, respectively. {The blocks with the notation (o) correspond to output layer and the block with notation (h) correspond to hidden layer.}}
\label{f1}
\end{figure}
{
The example shown in Fig. \ref{f1} demonstrates the basic three-layer ANN with the proposed backpropagation architecture, control circuit and weight update circuits. The neural network has three input neurons, two output neurons and five neurons in a hidden layer. The operation of the crossbar and switching between forward propagation, backpropagation and weight update operations is controlled by the switching transistors $M_{in}$, $M_{u}$ and $M_{r}$, which in turn are controlled by the sequence control block. The CROSSBAR 1 corresponds to the set of synaptic connections between the input layer and the hidden layer, and the CROSSBAR 2 represents the synapses connecting the hidden layer to the output layer. The three input signals are shown as $V_{in1}$, $V_{in2}$ and $V_{in3}$, and the corresponding normalized input signals are shown as $V_{i1}$, $V_{i2}$ and $V_{i3}$, respectively. The range of output signals from the normalization circuit depends on the application, limitations of memristors and linearity of the switch transistors. The inputs are fetched to the rows of the CROSSBAR 1 i.e. to the input switching transistors $M_{in}$.} Each memristor in a single column of the crossbar corresponds to the connections of all inputs to a particular single output. The crossbar performs dot product multiplication of inputs and the weights of a single column. The output of the multiplication for feed-forward propagation is read from the NMOS read transistor $M_{r}$ connected to a crossbar column. In this work, we investigate the approach, when the output of the crossbar is represented by a current flowing through the read transistor. However, the configuration can be changed to the voltage output from the crossbar column if required. The voltage-based approach is more complicated than current-based approach because the amplifier or OpAmp-based buffer is required to read the voltages without the loading effect from the following interfacing circuits. The current-based approach requires only the use of a current mirror ensuring the reduction of the on-chip area, {and also compatible with simple current driven sigmoid implementation}.
{
The parameters of input transistors $M_{in}$, read transistors $W_{r}$ and corresponding control signals $V_c$ should be selected carefully, considering the range of input signal. When the signals are propagated back through in the CROSSBAR 2, the propagated error can be both positive and negative. Therefore, if it is important to make sure that the size of the transistors and $V_c$ are set to eliminate the current when transistor in OFF state and conduct the current in linear region when transistor is ON. These parameters should be adjusted depending on the technology.}
{
The outputs from the crossbar columns are read sequentially one at a time to avoid the interference with the currents from the other columns. The read-transistors $M_{r}$ at the end of each column are used to switch ON and OFF the columns and maintain the order of the reading sequence for the forward propagation process.
As the negative resistance is not practical to implement in the memristive device, the negative weights are implemented using either input controller, which changes the input sign according to the sign of the synapse weight, or by the sign control circuit and additional crossbar that stores the sign of the weights. Fig. \ref{f1} illustrates the approach with sign control circuit and additional SIGN CROSSBAR 1 and SIGN CROSSBAR 2.
}
{
MAIN BLOCK 1 (MB1) in Fig. \ref{f1} performs the forward propagation and is responsible for the calculation of the activation function. MB1 (h) and MB1 (o) correspond to hidden layer and output layer, respectively. Depending on the application, the activation function can include the sigmoid, derivative of the sigmoid, tangent, derivative of the tangent, approximate sigmoid and approximate tangent functions. The approximate functions represented here refer to "hard" logical threshold sigmoid and tangent as shown in \cite{golden1996mathematical}. To implement the conventional backpropagation algorithm with gradient descent, MB1 performs the calculation of the sigmoid and the sigmoid derivative functions. The outputs of MB1 are stored in MEMORY UNIT (MU) 1(h) and used as the inputs to CROSSBAR 2. The outputs of MU1(h) can be normalized by normalization circuit. The output currents from the second crossbar are fetched to the second MB1 and the final output of the feed-forward propagation are obtained from MB1(o) and stored in MU1(o) for further application in backpropagation. The final outputs depend on the activation function. In addition, MB1 produces the outputs of the activation function, stored in MU1(h) and MU1(o), and derivative of the activation function, stored in MU2(h) and MU2(o) that are useful for backpropagation process.
}
{
After the forward propagation process, the backpropagation process is implemented. The sequence control block switches off the column read transistors $M_r$ of both crossbars and switches on the row read transistors $V_r$ of CROSSBAR 2.
The column transistors $M_{in}$ of CROSSBAR 2 are switched ON to propagate the inputs MU3, which stores the outputs of MAIN BLOCK 2 (MB2) corresponding to the propagation through the output layer. It ensures that the propagation is performed in the opposite direction. If the neural network contains more than three layers, all crossbars except the first crossbar corresponding to the synaptic weights between the input and the first hidden layer are reconnected to perform backpropagation operation. The backpropagation process through the output layer is implemented using MB2 and MAIN BLOCK 4 (MB4), and the backpropagation through the hidden layer corresponds to the MAIN BLOCK 3 (MB3). The possible architecture of analog memory unit is illustrated in \cite{tcad,aidana,aidanaconf}.
}
The final stage in the backpropagation algorithm is the weight update stage, where the values of the memristors are updated based on the specific rules. As the crossbar values are read and processed sequentially, MEMORY UNIT 4 and 5 store the update value before the update process starts. The weight update process is implemented by applying the voltage pulse of a particular duration and amplitude across each memristor. {The update pulse depends on the required change of the weights, calculated gradient of error, and the memristor type and technology. The amplitude of the update pulse depends on the outputs of MB2 and MB4 and calculated by weight update circuit. While the duration of the pulses are controlled by the sequence control circuit.
The update process of the memristor is controlled by transistors $M_{in}$ and $M_{r}$. To update the memristor, corresponding row transistors $M_{in}$ and column transistor $M_r$ are switched ON.
As the main architecture for the synapses that we consider in this work is 1M, each memristor is updated either one at a time. However, this process is slow, and in particular cases memristor weight can be updated in several cycles as illustrated in \cite{cam}. If the two state memristors are used in the crossbar ($R_{ON}$ and $R_{OFF}$, which is useful for binary neural network), the update process can be performed in two cycles: (1) update of all $R_{ON}$ memristors, which should be switched to $R_{OFF}$ and (2) update of all $R_{OFF}$ memristors, which should be switched to $R_{ON}$. This method is useful, when the other states between $R_{ON}$ and $R_{OFF}$ are not important for processing. Such method increases the speed of update process, however it is useful for neural network with binary synapses.}
\subsection{Circuit-level implementation of main blocks in backpropagation algorithm}
\label{sec3b}
{
This section briefly introduces the proposed architecture for backpropagation circuits shown in \cite{ISCASs}, while the detailed explanation of the circuit and all circuit parameters are provided in Supplementary Material (Section III).
In this work, the analog circuits for the proposed backpropagation implementation are designed for 180nm CMOS process.
The circuit level implementation of all backpropagation blocks is illustrated in Fig. \ref{f2}, while the components used in the circuits are shown in Fig. \ref{cnew}.
MB1 performs the forward propagation for the conventional backpropagation architecture. MB2 performs the backpropagation process through the output layer, which is finished by MB4. MB3 performs the backpropagation through the hidden layer.
The circuit level implementations of components from the main backpropagation blocks are shown in Fig. \ref{cnew}. Fig. \ref{cnew}(a)
illustrates the implementation of current buffer. Fig. \ref{cnew}(b) shown the implementation of adjusted sigmoid function proposed in \cite{5719144}. Fig. \ref{cnew}(c) shows the OpAmp circuit. Fig. \ref{cnew} (d) illustrated the multiplication circuit based on the current difference in transistors $M_{31}$ and $M_{32}$. Fig. \ref{cnew} demonstrated the implementation of analog switch circuit.
}
\begin{figure*}[h]
\centering
\includegraphics[width=140mm]{f2}
\caption{The circuit level architecture of the proposed backpropagation implementation. The separate implementation of the MB1, MB2, MB3 and MB4 is illustrated. In addition, the involved circuit components, such as DA, IA and IVC, are shown.}
\label{f2}
\end{figure*}
\begin{figure*}[h]
\centering
\includegraphics[width=180mm]{circuits}
\caption{Circuit components: (a) The current buffer circuit which is connected to the read transistor $M_r$ in Fig. \ref{f1}. The circuit is used in MB1 and MB3 to eliminate the loading effect of the activation function to the performance of the crossbar. (b) Sigmoid activation function used in MB1 inspired from \cite{5719144}. (c) Two stage OpAmp design used for all OpAmp based components in the proposed analog memristive learning circuit. (d) Multiplication circuit based on the Hilbert multiplier principle. The circuit is used in MB1 and MB2. (e) Analog switch design used in MB2, MB3 and MB4.}
\label{cnew}
\end{figure*}
{
\subsection{Sign control circuit}
}
\label{sec3c}
As the neural network weights can be both positive and negative, and the negative weight cannot be practically implemented by the memristor, the implementation of the additional weight control circuit is required. For each negative weight, the sign of the input voltage is changed. There are two possible ways to implement the sign. One of the solutions is to store the sign for each sequence in the external storage unit and apply it to the circuit with the weight normalization circuit. The other solution is to store the sign of each weight in the additional memristive crossbar elements. A memristive weight sign control circuit shown in Fig. \ref{f8} is proposed.
\begin{figure}
\centering
\includegraphics[width=60mm]{signcontrol}
\caption{Memristive weight sign control circuit that can be integrated to the crossbar to control the weight of the synapse or applied as an external circuit with a separate memristors to store the sign of the weight.}
\label{f8}
\end{figure}
The sign of each memristor in the crossbar is stored in the memristive crossbar or separate memristors as $R_{ON}$ or $R_{OFF}$. The analog sign read circuit follows the memristor storing the sign of the weight. There are two possible solutions. The first is to implement a single analog sign read circuit and switch it between the memristors in the crossbar, which requires additional on-chip area and storage. And the second and more effective solution is to implement the number of sign read circuits equivalent to the number of rows in a crossbar, which allows reading the sign of all the memristors in a single column. This allows achieving the trade-off between the required area, power and processing time. In Fig. \ref{f8}, the sign of the memristor representing the weight is stored in the memristor $R_{sign}$. When the sign is read, $V_{c}=1.25V$ is applied. If $R_{sign}$ is set to $R_{ON}$, the output of the analog switch $V_{sign}$ is positive and vice versa. The weight sign read circuit acts as a switch. If $R_{sign}=R_{ON}$, the voltage $V_{c}$ is above the switching threshold, it selects, and outputs the positive voltage $V_{in}$, which is the input voltage to the crossbar. If $R_{sign}=R_{OFF}$, the voltage drops and $V_{c}$ is below the switching threshold, and the switch outputs the voltage $-V_{in}$. The parameters of the transistors are the following: $M_{61}=0.18\mu/0.72\mu$, $M_{62}=0.18\mu/10.36\mu$ and $M_{63}=0.18\mu/0.36\mu$. The transistors in the circuit have an underdrive voltage $V_{DD}=1V$.
Comparing to the existing implementations of negative voltages in the crossbar array \cite{hu2016dot,truong2014new,hu2012hardware}, the method for sign control reduces the complexity of the implementation and ensures the stability of the output. For example, the crossbar in \cite{truong2014new} can perform dot product multiplication for both positive and negative signals. However, the system is complex because of the amplifiers that perform subtraction of the voltages. To ensure the amplification is not affected by the following circuits, such amplifier should include the capacitor, which increases the on-chip area of the circuit.
Also, this way of implementing negative signals require the accuracy preprocessing stage and additional adjustment of the input signals. A similar method of implementing negative sign is shown in \cite{hu2012hardware}. It involves a set of summing amplifiers with resistors, which also consume a large amount of area.
{
\subsection{Weight update circuit}
The implementation of memristive weight update circuit is illustrated in Fig. \ref{wuc}. The weight update circuit determined the pulse amplitude required to program the memristor in a crossbar array, depending on the calculated weight update value by MB2 and MB4.
The circuit is adaptable for different learning the due to the application of memristive devices $R_{40}$ and $R_{41}$ in the amplifiers. All the resistors in weight update circuit are $1k\Omega$ and the memristors are programmed considering the required learning rate. As the negative (to switch from $R_{OFF}$ to $R_{ON}$) and positive (to switch from $R_{ON}$ to $R_{OFF}$) programming amplitudes are not of the same amplitude for memristive devices, the analog switch selects the amplitude of the signal based on sign of the input voltage from MB1 and MB4. The implementation of analog switch is shown in Fig. \ref{cnew} (e) The shifted input signal is applied to analog switch control $V_c$ that determines, which input to the switch should be selected $V_{SW1}$ or $V_{SW2}$. The input to $V_{SW1}$ corresponds to the positive input voltage, while $V_{SW2}$ corresponds to the negative input voltage.
}
\begin{figure}
\centering
\includegraphics[width=40mm]{wuc}
\caption{{Memristive weight update circuit that converts the weight update value from MB3 and MB4 to the pulse of particular amplitude. The used analog switch is shown in Fig. \ref{cnew} (e).}}
\label{wuc}
\end{figure}
{
\subsection{Modular approach}
As large scale crossbars usually suffer from leakage currents, the most widely used architecture for the crossbar synapses is 1 transistor 1 memristor $1T1M$ \cite{krestinskaya2018neuro,yao2017face,wang2014ferroelectric}. Different variants of transistors and selector devices are used in the literature for the crossbar architecture for improving the crossbar performance. Architecture based in 1T1M synapses allows to remove the leakages which cause the reduction of output current in read transistors. However, this architecture of the synapses has significantly larger on-chip area and power consumption, comparing to single memristor (1M) crossbar architectures. In this paper, we avoid the application of $1T1M$ synapses and investigate the application of $1M$ synapses to maintain small on-chip area and low power consumption, and use the modular approach to reduce the leakage current problem and make the programming of the memristive arrays less complicated.
As illustrated in \cite{dasha}, modular approach allows to reduce the leakage currents in the crossbar.
In this approach, the large crossbar is divided into smaller crossbars as shown in Fig. \ref{mod}, and the current from all modular crossbars is summed up to process through the activation function in MB1. As illustrated in simulation results, this approach allow to achieve similar performance accuracy, as single crossbar approach.
}
{
In addition, if the network is scaled, the sequential processing can introduce the limitation to the system in the form of reduced processing speed. In this case, the parallel processing can be introduced, which involves the concurrent computation and simultaneous execution of the output computations. The modular approach can be useful as well, however, each modular crossbar should have corresponding processing blocks for backpropagation algorithm (MB1, MB2, MB3 and MB4). This introduces additional complexity for the system and increases on-chip area and power but reduces the processing time.
Modular approach may also allow to remove the analog storage units. As the size of the crossbar will be reduced, a time delayed signal produced by a signal delay circuit can be used instead of analog storage unit.
}
\begin{figure}
\centering
\includegraphics[width=90mm]{modular}
\caption{{Modular approach to reduce the leakage current and complexities in programming of 1M array.}}
\label{mod}
\end{figure}
\section{Learning architectures}
\label{s4}
\begin{figure*}
\centering
\includegraphics[width=180mm]{actFunc}
\caption{{Implementation of various activation functions for DNN: (a) Tangent function based on the sigmoid circuit from Fig. \ref{cnew} (b). This architecture allows to implement a single circuit for both sigmoid and tangent functions in a multilayer neural network or another learning architecture and switch between these two functions. (b) Approximate sigmoid and approximate tangent functions driven by input current. To implement the sigmoid and tangent, the voltage levels $V_{DD1}$ and $V_{SS1}$ are varied. (c) Approximate sigmoid and approximate tangent functions driven by input voltage. (d) Linear activation function based on analog switch. (e) Linear activation function based on diode. (f) Additional thresholding circuit to normalize the neural network output for binary outputs. }}
\label{f9}
\end{figure*}
The proposed analog memristive backpropagation learning circuits can be used for various applications and learning architectures, such as neural networks, HTM and LSTM hardware implementations. To apply the proposed backpropagation circuits for various architectures, the implementation of additional functional blocks and activation functions is required. { In this section, we illustrate the implementations of tangent, current and voltage driven approximate sigmoid and tangent, and linear activation functions and thresholding circuit to normalize the output of the neural network.}
To implement the tangent function, the sigmoid function can be adjusted. The use of the sigmoid circuit (Fig. 3(b)) allows building a single circuit for both of the functions and switch between the sigmoid and tangent implementations when it is required. The implementation of the tangent function is shown in Fig. \ref{f9}(a). The sigmoid and buffer part remain the same as in the sigmoid implementation and the voltage shift circuit based on the difference amplifier is added. The difference amplifier is based on the same OpAmp shown in Fig. \ref{cnew}(c) with $R_{16}=10k\Omega$, $R_{17}=1k\Omega$, $R_{18}=2.5k\Omega$, $R_{19}=15k\Omega$ and $R_{20}=1k\Omega$. The circuit shifts the voltage level of the sigmoid and allows to implement a tangent function with the same circuit.
The implementation of an approximate sigmoid and tangent functions can be done with a simple thresholding circuit shown in Fig.\ref{f9}(b) and Fig.\ref{f9}(c). There are options: current-control and voltage-control approximate functions. The current-control circuit is shown in Fig. \ref{f9}(b). The input current $I_{in}$ is applied to the current-to-voltage converter based on the OpAmp with $R_{22}=20k\Omega$ and inverted by the inverter with $M64=0.18\mu/0.36\mu$ and $M65=0.18\mu/1.72\mu$. The $W/L$ ratio of $M64$ and $M65$ can be adjusted depending on the required transition part between the high and low value of approximate sigmoid and tangent functions. The voltages $V_{DD1}$ and $V_{SS1}$ are different for sigmoid and tangent implementations. For the approximate sigmoid $V_{DD1}=1V$ and $V_{SS1}=0V$, which means that the transistors have an under-drive voltage level for TSMS 180nm CMOS technology. In the approximate tangent implementation, $V_{DD1}=1V$ and $V_{SS1}=-1V$. The simple thresholding circuits can implement the voltage-controlled sigmoid and tangent with two inverters shown in Fig. \ref{f9}(c). The $W/L$ transistor ratios of $W_{66}-W_{69}$ and voltage levels of $V_{DD1}$ and $V_{SS1}$ can be adjusted to obtain a required amplitude, range and transition region for the sigmoid and tangent.
{
The implementation of linear activation functions are shown in Fig. \ref{f9}(d) and Fig. \ref{f9}(e). Both units are driven by voltage and the OpAmp circuits are shown in Fig. \ref{cnew}(c). We propose the implementation of linear activation function based on analog switch shown in Fig. \ref{f9}(d). The analog switch selects the $0V$ output for negative input signal, and $V_{in}$ for positive input. The switch is controlled by the shifted inverted input signal which is fed to the switch control $V_c$. All values of resistances are set to $1k\Omega$. The possibility to implement linear activation function is to use diode (Fig. \ref{f9}(e)). The diode based linear activation function has smaller on-chip area and lower power consumption, however the output range is smaller comparing to linear activation function with switch. To implement a current controlled linear activation function, IVC can be used in both circuits. In linear activation unit based on the analog switch, OpAmp $R_{22}-R_{23}$ can be replaced by IVC. In diode based linear activation function. additional IVC component before OpAmp is required.
}
{
As we verified with the simulation results, if the ideal outputs is required to be binary, additional thresholding circuit is required in the output layer to normalize the outputs and achieve high accuracy. This was demonstrated using XOR problem in simulation results. The thresholding circuit that has been used for simulations is shown in Fig. \ref{f9}(f), where the parameters of the transistors are $M66=M68=0.36\mu /0.18\mu$ and $M67=M69=0.72\mu /0.18\mu$. The thresholding function is connected to the output layer after the training process and allows to increase the accuracy significantly during the testing stage.
}
\subsection{Neural Networks}
There is a number of neural network architectures, where the proposed learning circuit can be used. The complete analog learning and training circuits for most of the architectures and networks covered in Section \ref{s4} has not been implemented in analog hardware. There are different architectures and types of the neural networks that can use the proposed learning circuit without making a significant modification to the proposed design, such as DNN, BNN, and MNN.
\subsubsection{Deep Neural Network}
The proposed configuration for DNN with memristive analog learning circuits is shown in Fig. \ref{f11}. The DNN configuration contains N+1 layers, and N crossbars correspond to the synapses between the layers. In the forward propagation process, MB1 is used. MB1 can be modified to implement various activation functions. MB2 and MB4 perform the backpropagation process through the output layer of DNN, and MB3 does the backpropagation through hidden layers. In the update process, MB4 and MB3 are applied. The blocks MB2, MB3 and MB4 in each layer can be modified depending on the activation function applied in forward propagation in MB1 for each layer.
\begin{figure}
\centering
\includegraphics[width=90mm]{f11}
\caption{Deep neural network implementation with the backpropagation learning. Red arrows correspond to forward propagation process. black arrows refer to backpropagation process. Green arrows show the weight update process.}
\label{f11}
\end{figure}
\subsubsection{Binary Neural Network}
BNN can be implemented with the proposed circuit using two-stage memristors in the memristive crossbar representing the weights. The implementation of the BNN is shown in Fig. \ref{f12}. In BNN, the forward propagation process and backpropagation process are the same as in a three-layer neural network shown in Section \ref{s3}. However, due to the limitations of the binary weights, the direct update of the weights after the error calculation will not provide high accuracy results. We suggest to store the value of the change in error in the external storage and training units in time and, after a certain period of the training, update the weights. This method can improve the accuracy results for the classification problems using BNNs.
\begin{figure}
\centering
\includegraphics[width=70mm]{f12}
\caption{Three layer BNN with backpropagation learning. The crossbars contain memristors that can be programmed only for $R_{ON}$ and $R_{OFF}$ stages. To improve the accuracy and the performance of the network, the change in error is stored in the external storage and update unit. The crossbar weights are updated based on several iterations in time.}
\label{f12}
\end{figure}
\subsubsection{Multiple Neural Network}
The proposed memristive analog learning circuits can be used for the MNN approach. This is useful when several sources of input data are used, and the decision on the output depends on the fusion of the results from each data source. The output from the different data sources are normalized, and outputs of each data source are fetched to separate crossbars, and the outputs of all the crossbars are fed into the decision layer. The decision layer crossbar is the same as the other crossbars in the system. As MNN consists of several networks and the decision layer can be treated as a separate network, all the networks can be trained either separately \cite{363444} or as a single network. The architecture of MNN with backpropagation learning that is trained as a single network is illustrated in Fig. \ref{f13}. We recommend using this approach, when the neural network inputs are taken from different data sources, such as various sensors in the system. One of the examples of the use of such system is gender recognition, where voice signals and face images can be used as inputs to MNN. The activation functions of the layer are different for separate crossbars and the decision layer and depend on the data that is used for the processing. The number of the required backpropagation blocks is equaled to the number of the crossbars that a network contains.
\begin{figure}
\centering
\includegraphics[width=80mm]{f13}
\caption{Multiple neural network with backpropagation learning. The inputs from different data sources are fetched into different crossbars and the outputs from the crossbars are used as the inputs to the decision layer containing the memristive synapses.}
\label{f13}
\end{figure}
\subsection{Hierarchical Temporal Memory}
The other learning architecture, where the proposed circuit can be used is HTM. There are several hardware implementations proposed for the HTM SP and HTM TM \cite{ibrayev2017chip,fedorova2016htm,tcad}. However, the learning stage of the HTM SP has not been implemented on hardware yet. This stage can include either update process based on Hebb`s rule or the backpropagation update of the HTM SP weights \cite{25}.
There are two main analog architectures for the HTM SP: conventional HTM SP \cite{fedorova2016htm} and modified HTM SP \cite{tcad}. The application of the proposed learning circuits for both architectures is shown in Fig. \ref{f14}. Fig. \ref{f14a} illustrates the application of the proposed learning architecture for the conventional HTM circuits. After the forward propagation through the HTM SP, the HTM SP output is compared to the ideal HTM output. In the conventional HTM SP circuit, the calculated error from the comparison circuit or MB2 is fetched back to the memristive crossbar to calculate the error in the weights, and the weights are updated. Fig. \ref{f14b} shows the application of the proposed circuits for the modified HTM SP architecture. The conventional HTM SP architecture consists of the receptor and inhibition blocks. The weights of the synapses are located in the receptor block. After the comparison of the HTM SP output to the ideal output, the error is propagated back through the receptor block and MB3, and the memristive weights are updated.
\begin{figure}
\centering
\subfigure[]
{
\includegraphics[width=87mm]{f14a}
\label{f14a}
}
\subfigure[]
{
\includegraphics[width=87mm]{f14b}
\label{f14b}
}
\caption{Analog hardware implementation of the (a) conventional HTM SP algorithm and (b) modified HTM SP algorithm with backpropagation learning stage.}
\label{f14}
\end{figure}
\subsection{Long Short Term Memory}
LSTM architecture can also be implemented using the proposed memristive analog backpropagation circuits. The full implementation of LSTM with analog circuits has not been proposed yet. However, the analog implementation of the separate LSTM components has been shown in \cite{kam,smagulova2018design}. Fig. \ref{f15} illustrates the full system level LSTM architecture consisting of the output gate, input gate, write gate and forget gate. The weights of LSTM gates $W_i$, $W_o$, $W_f$ and $W_c$ can be stored in the memristive crossbars. The activation functions in the LSTM architecture can be replaced with different variations of MB1. While, the weight update process of the crossbar $W_o$ is performed by MB4 as the update of the output layer, while MB3 performs the update of $W_i$, $W_f$ and $W_c$.
\begin{figure}
\centering
\includegraphics[width=87mm]{f15}
\caption{Analog memristive hardware implementation of the LSTM algorithm.}
\label{f15}
\end{figure}
\section{Simulation results}
\label{s5}
The circuit simulations were performed in SPICE, and the verification of the ideal backpropagation algorithm is done in MATLAB.
\subsection{Circuit performance}
The memristor model used in the crossbar simulations is Biolek`s modified S-model \cite{7527252} for HP $\textrm{TiO}_2$ memristor with the threshold voltage $V_{th}$ of 1V \cite{strukov2008missing}. This memristor model is developed for large-scale simulations to simplify the computation and processing \cite{7527252}. The memristor characteristics and switching time for $R_{ON}=3k\Omega$ and $R_{OFF}=62k\Omega$ are shown in Fig. \ref{memristor}. {Fig. \ref{memristor} (b) and Fig. \ref{memristor} (c) illustrate memristor updated process applying pulse of 1s with different amplitudes.} The switching time is large, and speed of learning process with the memristive elements is slow. However, the learning and training process is a one-time process in the neural network. After the training during the testing stage, the reading time is small, and the data processing is fast.
\begin{figure}[!t]
\centering
\subfigure[]
{
\includegraphics[width=80mm]{Graph1}
}
\subfigure[]
{
\includegraphics[width=80mm]{Graph2}
}
\subfigure[]
{
\includegraphics[width=80mm]{Graph3}
}
\caption{Memristor characteristics: (a) hysteresis for different frequencies for $R_{initial}=10k\Omega$, (b) switching time from $R_{ON}=3k\Omega$ to $R_{OFF}=62k\Omega$ for different applied voltage amplitudes, and (c) switching time from $R_{OFF}=62k\Omega$ to $R_{ON}=3k\Omega$ for different applied voltage amplitudes.}
\label{memristor}
\end{figure}
{
Simulation results illustrating the performance of the circuits in terms of amplitude are shown in Supplementary Material (Section IV).
}
The simulation results for additional activation functions are shown in Fig. \ref{add1}. Fig. \ref{add1} (a) represents the simulation of the proposed tangent function. Fig. \ref{add1} (b) and \ref{add1} (c) illustrate the simulation of current driven approximate sigmoid and tangent, respectively.
{Fig. \ref{add1} (d) and \ref{add1} (e) show the simulation of linear activation functions with diode and switch, respectively. Fig. \ref{add1} shows that the activation function with switch is more linear. }
The timing diagram for the memristive weight sign control circuit is shown in Fig. \ref{add2}. Fig. \ref{add2} (a) represents the positive input voltage. Fig. \ref{add2} (b) illustrates the ideal output and real memristive weight sign control circuit output for $R_{ON}$, when the weight is positive. Fig. \ref{add2} (c) illustrates the ideal output and real output of the proposed weight sign control circuit for $R_{OFF}$, when the weight is negative.
{The output of weight update circuit is shown in Fig. \ref{add3}. The memristors is the circuit are programmed for high negative update amplitude and low positive update amplitude, as shown in Fig. \ref{add3} (c).}
\begin{figure*}[!ht]
\centering
\includegraphics[width=180mm]{outputActivation}
\caption{{Simulation of additional activation functions versus current: (a) tangent, (b) approximate current driven sigmoid and (c) approximate current driven tangent, (d) linear activation with diode, (e) linear activation with switch.}}
\label{add1}
\end{figure*}
\begin{figure}[!ht]
\centering
\includegraphics[width=90mm]{grWSC}
\caption{Timing diagram for memristive weight sign control circuit implementation: (a) input to the circuit, (b) ideal output and real output for $R_{ON}$ (when the weight is positive) and (c) ideal and real outputs for $R_{OFF}$ (when the weight is negative).}
\label{add2}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=90mm]{grWUC}
\caption{{Output of weight update circuit: (a) input voltage, (b) switch control voltage, (c) output of weight update circuit programmed to high amplitude voltage for negative update voltages (update from $R_{OFF}$ to $R_{ON}$) and low amplitude voltage for positive update voltages ($R_{ON}$ to $R_{OFF}$).}}
\label{add3}
\end{figure}
Table \ref{tableaa} represents the calculation of the on-chip area and maximum power dissipation for separate components for the analog learning circuit implementation and additional components and activation functions. Also, the example of the area and power dissipation for a small crossbar is shown.
\begin{table}[h]
{
\centering
\caption{Power consumption and on-chip area calculation for the separate circuit components.}
\label{tableaa}
\begin{tabular}{|p{3cm}|p{2cm}| p{2cm}|}
\hline
\textbf{Circuit component} & \textbf{Power consumption} & \textbf{On-chip area} \\ \hline
Crossbar (4 input neurons and 10 output neurons) & $5 \mu W $ & $1.36 \mu m^2 $ \\ \hline
Crossbar with control switches & $1200 \mu W $ & $ 115.3 \mu m^2 $ \\ \hline
Weight sign control circuit & $195.1 \mu W$& $16.64 \mu m^2$ \\ \hline
Sigmoid & $11.4 \mu W$& $184.00 \mu m^2$ \\ \hline
Current buffer & $149.0 \mu W $ & $280.00 \mu m^2 $ \\ \hline
OpAmp (maximum) & $39.8 mW $ & $2801.76 \mu m^2 $ \\ \hline
Analog switch & $162.3 \mu W$ & $1.55\mu m^2 $\\ \hline
Approximate current driven sigmoid/tangent & $52.9 mW $ & $2118.00 \mu m^2 $ \\ \hline
Approximate voltage driven sigmoid/tangent & $ 41.2 pW$ & $0.40 \mu m^2 $ \\ \hline
Linear activation units with diode & $ 963.7 \mu W$ & $244 \mu m^2 $ \\ \hline
Linear activation unit with switch & $23.214 mW$ & $951.06 \mu m^2 $ \\ \hline
Weight update circuit & $ 14.34 mW$ & $1269.63 \mu m^2 $ \\ \hline
\end{tabular}
}
\end{table}
Table \ref{table11} shows the on-chip area and maximum power dissipation for separate components for the main blocks of the proposed analog backpropagation learning circuit implementation.
\begin{table}[h]
\centering
\caption{Area and power calculations for the main blocks of the proposed design and total area and power for three layer network.}
\label{table11}
\begin{tabular}{|p{3.5cm}|p{1.7cm}|p{1.7cm}|}
\hline
{Configuration}& {Area ($\mu m^2)$} & {Maximum Power($mW$)} \\
\hline
\hline
MB1 (hidden layer) & 4885.86 & 3.70 \\
\hline
MB2 + MB1 (output layer) & 8264.88 & 10.64 \\
\hline
MB3 & 15238.69 & 61.78 \\
\hline
MB4 & 9734.33 & 39.53 \\
\hline
\hline
\textbf{Total} & \textbf{38123.76} & \textbf{115.65} \\
\hline
\end{tabular}
\end{table}
\subsection{System level simulations}
The system level simulations have been performed for XOR problem and handwritten digits recognition for ANN and face recognition for DNN. For setup in XOR problem, there are 2 neurons in the input layers, 4 neurons in the hidden layer and 1 neuron in the output layer. {During the training, input was selected randomly out of four possible inputs. } The simulation results for XOR problem for different learning rates for ideal simulations and backpropagation circuit are shown in Fig. \ref{xor} for number of iterations $n=50,000$. In the real circuit with non-ideal behavior, the error after $50,000$ iterations is 15\% higher than in ideal simulations. We verified that it is caused by the non-ideal behavior of analog multiplication circuits, which will be improved further in the future work. The accuracy results for XOR simulation for the cases with and without thresholding circuits are shown in Table III.
\begin{figure}[!ht]
\centering
\includegraphics[width=70mm]{xorfinalnew}
\caption{{Error rate versus number of iterations for (a) simulations ideal algorithm and (b) simulation of proposed backpropagation circuit.}}
\label{xor}
\end{figure}
{
The variability analysis for random offsets in memristor programming value is shown in Fig. \ref{xor1} for learning rates $\eta=0.15$, $\eta=0.3$ and $\eta=0.5$. The offset in the weight update value is represented as following: $w=w+(\Delta w + \Delta w \cdot x)$, where $x$ is a random variation of the weight of particular percentage. This variation can be caused by non-ideal behavior of processing, update circuit and control circuits for update pulse duration. The architecture was tested for the variation of 50\%, 100\%, 200\% and 300 \%. The simulation results on Fig. \ref{xor1} show that the variation in the update value does not have a significant effect on the performance of the architecture. Also, the offsets in update value affect the system with smaller learning rate ($\eta=0.15$) more than the system with larger learning rate ($\eta=0.5$). However, the value of error converges to small error in all the cases. Therefore, even the significant error in the memristor update value, does not effect the performance of the learning process. The final accuracy for 100,000 iterations is illustrated in Table III.
}
\begin{figure}[!ht]
\centering
\includegraphics[width=90mm]{offset}
\caption{{ Effect of the offset of the memristor update value on the performance of the architecture for (a) $\eta=0.15$, (b) $\eta=0.3$ and (c) $\eta=0.5$ .}}
\label{xor1}
\end{figure}
{
The performance analysis for the case of random mismatch in final memristor values after update is shown Fig. \ref{xor2}. The performance accuracy after 100,000 iterations for all cases is shown in Table III. The mismatch is defined as following: $w=(w+\Delta w)+ (w+ \Delta w )\cdot x$, where $x$ is the percentage of variation, shown in Fig. \ref{xor2} as 1\%, 2\%, 4\% and 5\%. This mismatch has more significant effect on the performance of the architecture. For the small learning rate (Fig. \ref{xor2} (a)), the mismatch in the memristor values does not allow system to converge. For the larger learning rates, the system converge slower that in ideal case for 1-2\% of mismatch and does not converge for larger mismatches. However, such case is the effect of the non-linear behavior and instability of memristive device, which should be investigated further at the device level.
}
\begin{figure}[!ht]
\centering
\includegraphics[width=90mm]{mismatch}
\caption{{ Mismatch in the final weight of memristor of 1\%, 2\%, 4\% and 5\% for (a) $\eta=0.15$ (b) $\eta=0.3$ (c) $\eta=0.5$.}}
\label{xor2}
\end{figure}
\begin{table}[]
\label{xoracc}
{
\caption{Accuracy for XOR simulations (2 input neurons, 4 hidden layer neurons and 1 output neuron).)}
\begin{tabular}{|l|l|c|c|}
\hline
\multicolumn{2}{|l|}{\textbf{\begin{tabular}[c]{@{}l@{}}Configuration\\ for XOR simulation\end{tabular}}} & \textbf{\begin{tabular}[c]{@{}c@{}}ANN accuracy \\ (without \\ thresholding\\ circuit)\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}ANN accuracy\\ (with \\ thresholding\\ circuit, $\theta=0.5$)\end{tabular}} \\ \hline \hline
\multicolumn{2}{|l|}{\begin{tabular}[c]{@{}l@{}}Ideal memristors\\ $\eta=0.15$, $n=50,000$\end{tabular}} & 84.8\% & 100\% \\ \hline
\multicolumn{2}{|l|}{\begin{tabular}[c]{@{}l@{}}Ideal memristors\\ $\eta=0.15$, $n=100,000$\end{tabular}} & 96.26\% & 100\% \\ \hline
\multicolumn{2}{|l|}{\begin{tabular}[c]{@{}l@{}}Ideal memristors\\ $\eta=0.3$, $n=100,000$\end{tabular}} & 97.76\% & 100\% \\ \hline
\multicolumn{2}{|l|}{\begin{tabular}[c]{@{}l@{}}Ideal memristors\\ $\eta=0.5$, $n=100,000$\end{tabular}} & 98.31\% & 100\% \\ \hline
\hline
\begin{tabular}[c]{@{}l@{}}Offset in memristors\\ programming value\\ $\eta=0.15$, \\ $n=100,000$\end{tabular} & \begin{tabular}[c]{@{}l@{}}50\%\\ 100\%\\ 200\%\\ 300\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}96.10\%\\ 96.10\%\\ 96.07\%\\ 96.10\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}100\%\\ 100\%\\ 100\%\\ 100\%\end{tabular} \\ \hline
\begin{tabular}[c]{@{}l@{}}Offset in memristors\\ programming value\\ $\eta=0.3$,\\ $n=100,000$\end{tabular} & \begin{tabular}[c]{@{}l@{}}50\%\\ 100\%\\ 200\%\\ 300\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}96.48\%\\ 96.48\%\\ 96.41\%\\ 95.72\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}100\%\\ 100\%\\ 100\%\\ 100\%\end{tabular} \\ \hline
\begin{tabular}[c]{@{}l@{}}Offset in memristors\\ programming value\\ $\eta=0.5$,\\ $n=100,000$\end{tabular} & \begin{tabular}[c]{@{}l@{}}50\%\\ 100\%\\ 200\%\\ 300\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}98.56\%\\ 98.58\%\\ 98.56\%\\ 98.33\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}100\%\\ 100\%\\ 100\%\\ 100\%\end{tabular} \\ \hline
\hline
\begin{tabular}[c]{@{}l@{}}Random mismatches\\ in memristor value\\ $\eta=0.15$,\\ $n=100,000$\end{tabular} & \begin{tabular}[c]{@{}l@{}}1\%\\ 2\%\\ 4\%\\ 5\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}50.02\%\\ 50.82\%\\ 50\%\\ 50\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}50\%\\ 50\%\\ 50\%\\ 50\%\end{tabular} \\ \hline
\begin{tabular}[c]{@{}l@{}}Random mismatches\\ in memristor value\\ $\eta=0.3$,\\ $n=100,000$\end{tabular} & \begin{tabular}[c]{@{}l@{}}1\%\\ 2\%\\ 4\%\\ 5\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}99.7\%\\ 99.89\%\\ 56.73\%\\ 50\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}100\%\\ 100\%\\ 62.5\%\\ 50\%\end{tabular} \\ \hline
\begin{tabular}[c]{@{}l@{}}Random mismatches \\ in memristor value\\ $\eta=0.5$,\\ $n=100,000$\end{tabular} & \begin{tabular}[c]{@{}l@{}}1\%\\ 2\%\\ 4\%\\ 5\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}91.77\%\\ 80.29\%\\ 99.22\%\\ 63.23\%\end{tabular} & \begin{tabular}[c]{@{}c@{}}100\%\\ 100\%\\ 100\%\\ 87.5\%\end{tabular} \\ \hline
\end{tabular}
}
\end{table}
{ To verify the performance of the proposed approaches for real pattern recognition problems, we tested ANN for handwritten digits recognition and DNN for face recognition for 2 approaches: single crossbar (shown in Fig. \ref{f1}) and modular crossbar (shown in Fig. \ref{mod}) using $Ge_2Sb_2Te_5$ (GST) memristors with 16 resistive levels \cite{kuzum2011nanoelectronic,xiao2017gst}.
In ANN simulation, MNIST database \cite{mnist} with $70,000$ images of the size of $28\times 28$ was used, where 86\% of images was selected for testing and 14\% for testing. The setup for ANN consisted of $28\times 28$ input layer neurons, 42 hidden layer neurons and 10 output neurons (corresponding to 10 classes of digits). In the modular approach, 16 crossbars with 49 input neurons, 8 crossbars with 98 input neurons and 4 crossbars with 196 input neurons were tested. For DNN verification, we performed face recognition task using Yale database for face recognition with 165 images of 15 people \cite{georghiades1997yale}. The images were rescaled by the size of $32\times 32$, and 45\% of the dataset was used for training and 55\% for testing. The DNN configuration consisted of 6 layers of 1024, 800, 500, 100, 30 and 15 neurons. The simulation results are shown in Table IV. As the accuracy for all modular configurations is approximately the same, the modular crossbar approach is presented by a single value in the table. The simulation results show that the performance accuracy for both real ANN and DNN is reduced slightly, comparing to ideal case. As the obtained accuracy is the same and the research work \cite{dasha} shows that the leakage currents are reduced in modular approach, the crossbar with 1M synapses can be divided into modular crossbars to avoid 1T1M synapses and reduce the on-chip area of the crossbar.
}
\begin{table}[]
\label{anndnn}
{
\centering
\caption{ANN accuracy for handwritten digits recognition application and DNN accuracy for face recognition application.}
\begin{center}
\begin{tabular}{|l|c|c|}
\hline
\textbf{Configuration} & \textbf{\begin{tabular}[c]{@{}c@{}}ANN accuracy \\ (MNIST,\\ handwritten digits)\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}DNN accuracy \\ (Yale, \\ face recognition)\end{tabular}} \\ \hline
Ideal simulations & 93\% & 78.9\% \\ \hline
Single crossbar & 92\% & 73.3\% \\ \hline
Modular crossbar & 92\% & 75.5\% \\ \hline
\end{tabular}
\end{center}
}
\end{table}
\section{Discussion}
\label{s6}
The proposed analog hardware implementation of the backpropagation algorithm can be used to implement the online training of different learning architectures, {which can be used for near-sensor processing.} The analog memristive learning architecture allows removing the additional software based or digital offline training and learning process. This can increase the processing speed and reduce the processing time, comparing to digital analogies, where the number of components to achieve high sampling rates in analog-to-digital converters (ADC) and digital-to-analog converters (DAC) are large. The possible errors in the training caused by leakage currents and parasitic effects can be mitigated by the increase of the number of iterations in the learning stage.
If the sneak path problems occur during the training state and the memristor update value is not accurate, this can be fixed in the following learning stages, but more update iterations are required for error to converge and reach high accuracy in this case.
As demonstrated in the XOR simulation, the problem of the non-ideal variation of the update value of the memristor due to non-ideal performance of the circuit and other device instabilities can be eliminated by increasing number of training iterations.
The limitations of the proposed architecture include the scalability of the memristive crossbar arrays and limitations of the current memristive devices. The problems of the parasitics, leakage current and sneak paths in the memristive crossbar have to be investigated further. The other drawback is a limitation of the memristive devices in terms of the number of resistance levels that can be achieved for particular memristive devices.
The future work will include the implementation of the crossbar with a physically realizable memristor \cite{messaris2018data}, evaluation of the performance of the architecture with different memristive devices, evaluation of the abilities of different memristive devices and adjustment of circuit parameters for particular devices.
In addition, the limitations of the memristive devices, electromagnetic effects, frequency effects and their effect on the accuracy and the performance of the proposed learning architecture have to be studied. Also, the endurance of the memristive devices should be studied, especially for the case of several iterations in the learning process.
The testing of the complete systems for large scale problems has to be performed and the limitations, such as loading effects and parasitics, have to be identified from the physical design constraints perspective. The effect of the additional components of the overall system performance and processing speed has to be determined under such conditions that become technology specific issues.
The future work will include the full circuit implementation of the proposed HTM, LSTM and MNN architectures and verification of their performance for large scale problems.
\section{Conclusion}
\label{s7}
In this paper, we presented the circuit design of an analog CMOS-memristive backpropagation learning circuit and its integration to different neural network architectures. The circuit architectures are presented for a three-layer neural network, DNN, BNN, MNN, conventional and modified HTM SP and LSTM. We presented the analog circuit implementation of interfacing circuits and activation functions that can be used to implement various learning architectures. The implementation of backpropagation with analog circuits offers simplicity of building differential operations combined with a dot-product operator as memristor crossbar that is useful of building neural networks. Using databases of MNIST (character recognition) and Yale (face recognition) an application level validation of the proposed learning circuits for ANN and DNN architectures is successfully demonstrated.
The presented design of crossbar does not take into account physical design issues of memristive devices, while sneak path problem of crossbar arrays is accounted in the simulations by including conductance variability of real memristor devices and wire resistors in the crossbar. However, the signal integrity issues is a topic to investigate further, when the memristor technology is mature and is suitable for a fabricating reliable large-scale arrays. The area and power of the proposed circuit design need to be further optimized a fully parallel implementation for real-time applications.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{Intro}
Percolation is an intensely studied model of statistical mechanics and has been widely applied to
interpret and describe numerous physical, natural and social systems \cite{book}. The popularity of the
model can be attributed to the coexistence of simplicity in its proposition and richness in its outcome \cite{review}.
The mathematical model of percolation was first proposed by Broadbent and Hammersley in 1957 \cite{Broadbent}. It gradually
became widely accepted by physicists \cite{Isichenko} and was successfully applied to study and understand the properties of
metal-insulator transition \cite{Ball}, magnetic materials \cite{Dotsenko}, spin quantum Hall effect \cite{Gruzberg},
growth models \cite{Saberi, Dsouza} and networks \cite{Derenyi, Callaway, Kalisky}.
Percolation is also frequently used in subjects like chemistry, geophysics, environmental sciences, medical
sciences and social sciences to analyze issues such as polymer gelation \cite{Coniglio}, colloids \cite{Anekal, Gnan},
flow of oil through
porous media \cite{King}, fractality of coast lines \cite{Sapoval, Saberi_coast}, spreading of forest fire
\cite{Albano, Albano2}, epidemic outbursts \cite{Grassberger}, neuron communication \cite{Zhou}, tumor induced
angiogenesis \cite{Paul} and numerous other systems. It is a highly active field of research with many open
problems \cite{Araujo}.
In a typical site percolation problem, the sites of a regular empty lattice are occupied randomly with a probability $p$,
called occupation probability. A cluster is formed when two neighboring sites are occupied. If any nearest
neighbor of any of the sites in the cluster gets occupied, it is also included in the cluster. For small
values of $p$, many small isolated clusters are formed in the lattice. As $p$ is gradually increased, these
clusters start to merge and at a certain value of $p$ ($=p_c$, called the percolation threshold), a single cluster
spans the lattice. This sudden appearance of a spanning cluster marks a phase transition (continuous in this case)
when the cluster-size and the correlation length diverge. The percolation transition possesses a number of
remarkable characteristic features and exhibits interesting critical behavior to form an important
universality class.
Another primitive and widely used model is bond percolation where the empty bonds between the preoccupied sites
are occupied randomly.
The value of percolation threshold for bond percolation in infinite square lattice has been analytically
calculated to be $p_c=\frac{1}{2}$, unlike site percolation, for which, best numerical estimate is $p_c\approx0.592746$.
An interesting extension of these two basic models is the site-bond percolation \cite{Hoshen}, where the sites are occupied with
probability $p_s$ and the bonds between neighboring occupied sites are filled with probability $p_b$.
Percolation models like explosive percolation \cite{Achlioptas, Ziff_explo, Cho, Dsouza}, bootstrap percolation \cite{Adler},
directed percolation \cite{Broadbent, Grassberger, Albano},
correlated percolation \cite{Coniglio, Makse} and a lot of other variants are available in the literature.
These models have been proposed and
studied not only for meeting the requirement of different systems but also out of pure mathematical interest.
Several percolation studies have addressed the geometric and transport properties of disordered systems
\cite{Coniglio, Makse, Araujo2}. A model \cite{Kundu}
has been introduced with an additional source of disorder, in which the sites are occupied randomly with discs of
random radii. The bonds are considered occupied if the discs satisfy certain predefined conditions. The critical
behavior indicates that this model belongs to the same universality class as ordinary percolation. Another
interesting model \cite{Hassan} deals with a weighted planar stochastic lattice (WPSL); and from the calculated
values of the critical
exponents the authors conclude that percolation on WPSL belongs to a different universality class.
\begin{figure*}[t]
\centering
\subfigure[]{\includegraphics[scale=0.4]{distlat1.eps}}\hspace{0.2cm}%
\subfigure[]{\includegraphics[scale=0.4]{distlat2.eps}}\hspace{0.2cm}%
\subfigure[]{\includegraphics[scale=0.4]{distbox.eps}}
\caption{(Color online)Two typical representations of $5 \times 5$ distorted lattices are shown with (a) $\alpha=0.1$ and
(b) $\alpha=0.2$. The intersection points of the dotted lines are the undistorted lattice positions and blue/red circles
represent the distorted positions. A possible configuration of two neighboring lattice points (magenta) is shown in (c).
These two points may take any position within their respective squares (in gray) of length $2\alpha$ centered at
the undistorted positions (open circles). The undistorted distance is taken to be $1$, whereas, their distance
$\delta$ in the distorted lattice may vary between $\delta_m$ and $\delta_M$.}
\label{distlat}
\end{figure*}
In the present work, we introduce a new model of percolation in a distorted square lattice. To begin with, a regular empty
square lattice has been considered. The positions of the sites are then shifted to create a distorted
lattice. The amount of shifts are not same for all the sites but are controlled by a tunable distortion parameter.
The percolation properties are studied for low to moderate distortion with a vision to work on a lattice that
is not perfectly ordered but not too much disorganized either. In section \ref{lattice} we give the details about
the preparation of lattice. The percolation logic on this lattice is explained in section \ref{process}. We present
our results in section \ref{results}. In \ref{pcvar}, we show the variation of percolation threshold and in \ref{univ}
we explore the scaling behavior of the order parameter and calculate the approximate values of the critical
exponents in order to identify the universality class of the present model. Finally we summarize in section \ref{sum}.
\section{The model}
\subsection{The distorted lattice}\label{lattice}
$L \times L$ distorted lattice is created by slightly ruffling the sites of a regular $L \times L$ lattice.
This has been done systematically by setting a distortion parameter $\alpha$ which denotes the maximum amount of
dislocation along $x$ or $y$ axis. Such a lattice can be generated using the following steps:\\
\begin{itemize}
\item Initially an empty square lattice with equidistant nearest neighbors is considered. The
lattice constant is set to unity.
\item A suitable value for the distortion parameter $\alpha$ is fixed. Since the undistorted distance
is set to $1$, $\alpha$ may be varied within the range: $0<\alpha<0.5$.
\item A lattice site at position $(i,j)$ is chosen. Two random numbers, $r_x$ and $r_y$, are generated
for $x$ and $y$ direction respectively, in the range $\{-\alpha,\alpha\}$. The position of this site
is modified to $(X_{ij},Y_{ij})$, where, $X_{ij}=i+r_x$ and $Y_{ij}=j+r_y$.
\item The above step is repeated for every site of the lattice. A distorted lattice is thereby created.
\end{itemize}
Two suggestive representations of distorted lattice are shown in figure \ref{distlat} for two different values of
$\alpha$. As shown therein, each site can now be located at any point $(i+r_x,j+r_y)$ within a square of length $2\alpha$ with
the undistorted position $(i,j)$ at the center of the square. The distance between any two neighboring sites is denoted
in general by $\delta$. The minimum and maximum limits of $\delta$ are therefore
\begin{equation}
\delta_m=1-2\alpha
\end{equation}
and
\begin{equation}
\delta_M=\sqrt{(1+2\alpha)^2+(2\alpha)^2}
\end{equation}
respectively. Note that for $\alpha> 0.5$, the lattice is over-distorted: the squares of possible occupancy
of two neighboring sites in figure \ref{distlat}(c) would overlap and the notion of identifying a site with reference
to the regular lattice points $(i,j)$ would lose its meaning. We restrict this study from low to moderate $\alpha$.
\begin{figure*}[t]
\centering
\subfigure[]{\includegraphics[scale=0.45]{alphapc.eps}}\hspace{0.3cm}
\subfigure[]{\includegraphics[scale=0.45]{dpc.eps}}
\caption{(Color online) (a) Variation of percolation threshold $p_c$ with distortion $\alpha$ for different
fixed values of the connection threshold $d$ (each curve corresponds to a fixed value of $d$, in the range $1.0025$
to $1.25$). Distortion is varied within the range $\alpha=0 - 0.3$. The (blue) curve at the bottom corresponds
to $d=1.25$ and the one (brown) with a sharp rise is for $d=1.0025$.
(b) Variation of percolation threshold $p_c$ with the connection threshold $d$ for different
fixed values of distortion $\alpha$ (each curve corresponds to a fixed value of $\alpha$, in the range $0.05$
to $0.3$). Connection threshold is varied within the range $d=1.0 - 1.3$. The (blue) curve at the top-right
corresponds to $\alpha=0.3$ and the left-bottom one (red) is for $\alpha=0.05$.
All the plots of both the figures are generated for a system size $L=2^{10}$ and each $p_c$-value is obtained by averaging
over $200$ independent realizations of the lattice for a given set of values of $\alpha$ and $d$.}
\label{alpc}
\end{figure*}
\subsection{Percolation process}\label{process}
In usual site percolation, some of the empty sites are occupied randomly corresponding to an occupation
probability $p$. In the present model, the following process has been adopted after generating the distorted empty lattice:
\begin{itemize}
\item The empty sites of the distorted lattice are occupied randomly as per a given $p$.
\item A connection threshold ($d$) is set.
\item Distance ($\delta$) between any pair of occupied neighbors is calculated. These two neighbors are
considered to be connected if $\delta\le d$; otherwise, the link is broken.
\item The above step is repeated for each pair of occupied neighbors and the clusters are identified
accordingly to determine the possibility of percolation.
\end{itemize}
It is clear from the above scheme that even if two nearest neighbors are occupied, they may {\it not} belong
to the same cluster. This is the main difference of the present model with usual site percolation.
The connectivity of any two occupied neighbors depends on $\delta$ (and therefore, on $\alpha$) and on the value
of $d$. Since $\delta_m\le \delta \le \delta_M$, the range of interest for the connection threshold $d$ is also
$\delta_m\le d \le \delta_M$. If $d < \delta_m$, no cluster formation is possible; whereas, for $d > \delta_M$,
every occupied neighbor is connected and the case is similar to usual (undistorted) site percolation. However,
as we shall see in the next section, when $d\le 1$, the system suddenly ceases to percolate even for $p=1.0$,
making the range $\delta_m < d \le 1$ uneventful.
In this work, the cluster identification and numbering has been done by the well known Hoshen-Kopelman (HK)
algorithm \cite{HK}. The connectivity criterion ($\delta\le d$) has been incorporated into the HK algorithm
to appropriately reflect the properties of distorted lattice. We emphasize that this is a controlled site
percolation model and is clearly distinct from bond percolation (where every site is occupied and bonds are
occupied randomly) and site-bond percolation (where both the sites and the bonds between the occupied neighbors
are occupied randomly) models.
\section{results and discussions}\label{results}
\subsection{Variation of percolation threshold}\label{pcvar}
Let us first study the effect of this distortion on the percolation threshold($p_c$). For an undistorted square
lattice, this value is known to be $p_c\approx 0.592746=p_{cu}$ as the lattice size tends to infinity. It is not
hard to realize that $p_c$ depends on the relative strengths of $\alpha$ and $d$ in a distorted lattice.
To demonstrate these dependences
clearly, we first calculate $p_c(\alpha)$ for different fixed values of $d$. Fig. \ref{alpc}(a) shows eight curves,
one each for a value of $d$ ranging between $1.0$ and $1.25$. All the curves stay at the value $p_{cu}$ as long as
$\alpha$ is low enough, so that $\delta_M<d$. For example, the (blue) curve for $d=1.25$ at the bottom
(see fig. \ref{alpc}(a)) remains at $p_{cu}$ up to an appreciable value of $\alpha$. This is expected since this
situation is similar to undistorted percolation (even if $\alpha$ is non-zero) due to large value of $d$. The maximum
distance $\delta_M$ between the neighboring points has to exceed the connection threshold $d$ for the manifestation
of any effect of distortion. For lesser values of $d$, percolation threshold is affected by less distortion.
It may be concluded from these plots that when the connection threshold is held fixed, $p_c$ increases with $\alpha$.
This means more distortion makes it more difficult for the system to percolate. This result can be attributed to
the fact that the average distance between two neighboring points increases with $\alpha$ as
$\delta_{min}(\alpha)\approx 1+0.337629\alpha^2$. For lower values of the connection
threshold ($d=1.05, 1.025, 1.0125, 1.0025$ here), $p_c$ reaches very close to $1$ at a certain value of $\alpha$.
As $d$ becomes smaller, this situation occurs with smaller $\alpha$. This indicates that
all the sites need to be occupied to span the lattice. Moreover, if $\alpha$ is further increased the system is no longer
guaranteed to percolate. In fig. \ref{alpc}(a), each $p_c$ is calculated by generating $200$ independent representations
for specific set of $\alpha$ and $d$. For each value of $d$, $p_c$ is shown up to the value of $\alpha$ for which all the $200$
representations percolate. At $d=1.0$, the system ceases to percolate for any non-zero value of $\alpha$. We have also
checked that this situation persists for any $d<1.0$. Thus, for any value of connection threshold which is less than
or equal to the lattice constant (or, the undistorted nearest neighbor distance, here taken to be $1.0$),
the system can not percolate if any distortion, be it very small, is present.
This is an important observation in the context of realistic applications of percolation, particularly since
a perfectly ordered natural system can hardly be found. Consider, for example the simulation of
the forest-fire model. Here, $d$ can be interpreted as the fire-spreading threshold. A distorted array
will make it difficult for the fire to percolate. Moreover, depending on the relative values of $\alpha$ and $d$,
distortion can even completely stop fire-percolation in a forest where the fire would have definitely percolated
for an undistorted array of same number of trees.
Variation of the percolation threshold($p_c$) with the connection threshold($d$) is shown in fig. \ref{alpc}(b)
for six different values of the distortion parameter ($\alpha$). A higher value of $d$ is expected to favor percolation and this
is evident from fig. \ref{alpc}(b), which shows that $p_c$ decreases with $d$ when $\alpha$ remains fixed. If $d$
is large enough (and $\alpha$ is small enough) to ensure that $d>\delta_M$, the effect of distortion disappears and
consequently, the percolation threshold stays at $p_c=p_{cu}$. In the other extreme, spanning becomes more difficult
with a low connection threshold and beyond a certain value, system can not percolate even after occupying all the
sites. For each $\alpha$, the displayed data start with the minimum value of $d$ which ensures that all the $200$ independent
realizations do necessarily percolate.
All the plots of fig.\ref{alpc}(a) and fig.\ref{alpc}(b) can be regarded as separation curves between percolating and
non-percolating phases. This is shown in fig.\ref{phase}. For example, if $d=1.025$ and $\alpha=0.05$, spanning
is guaranteed if $85\%$ of the sites are occupied, since the point ($0.05,0.85$) in fig.\ref{phase}(a) falls in
the percolating (green) region. Similarly, if $\alpha=0.15$ and $d=1.05$, a $70\%$ occupancy is not sufficient
for spanning (see fig.\ref{phase}(b)).
The density plot (fig.\ref{phase}(c)) shows the variation of the percolation threshold with both connection
threshold and distortion. Higher $p_c$ is obtained for high $\alpha$ and low $d$. The blank portion on the left side
indicates that the system can not percolate for those values $\alpha$ and $d$.
\begin{figure}[h]
\centering
\subfigure[]{\includegraphics[width=0.45\columnwidth]{Phase_a.eps}}
\subfigure[]{\includegraphics[width=0.45\columnwidth]{Phase_b.eps}}
\subfigure[]{\includegraphics[width=0.8\columnwidth]{densplot.eps}}
\caption{(Color online) Plots of (a) $p_c(\alpha)$ with $d=1.025$ and (b) $p_c(d)$ with $\alpha=0.15$. The curves
(and all the other curves in fig. \ref{alpc}(a) and fig. \ref{alpc}(b)) separate the spaces in percolating and
non-percolating phases. (c) A density plot for $p_c(d,\alpha)$.}
\label{phase}
\end{figure}
\begin{figure*}[t]
\centering
\subfigure[]{\includegraphics[width=\columnwidth]{alphasmax.eps}}
\subfigure[]{\includegraphics[width=\columnwidth]{dsmax.eps}}
\caption{(Color online) Plots of the order parameter $\Omega_L(p)$ with $p$ (a) for different values of $\alpha$
keeping $d=1.1$ and (b) for different values of $d$ keeping $\alpha=0.1$. The curves shift towards left as the
effect of distortion ceases when $\alpha$ decreases and $d$ increases. Each data point of both the plots are generated
by averaging over $1000$ independent realizations of the distorted lattice.}
\label{smax}
\end{figure*}
\subsection{Order parameter and universality class}\label{univ}
\begin{figure*}[t]
\centering
\subfigure[]{\includegraphics[width=0.95\columnwidth]{smax.eps}}
\subfigure[]{\includegraphics[width=0.9\columnwidth]{collapse.eps}}
\caption{(Color online) (a) Plots of $\Omega_L(p)$ for different system sizes $L=2^6,2^7,2^8,2^9,2^{10}$ with
$\alpha=0.2$ and $d=1.1$ for all the points. [Inset] A log-log plot $\Omega_L(p)$ with $L$ at a particular value
of $(p-p_c)L^{1/\nu}$ along with a straight line fit of the data gives $\beta/\nu=0.1143$. (b) A nice data
collapse obtained when $\Omega_L(p) L^{\beta/\nu}$ is plotted against $(p-p_c)L^{1/\nu}$. [Inset] Plot of
$\log_2\langle S_{max}\rangle$ at $p=p_c$ with $\log_2L$ and a straight line fit gives $d_f=1.8857$, consistent with
$d_f=2-\beta/\nu$.}
\label{scaling}
\end{figure*}
The percolation order parameter is usually defined as \begin{equation}
\Omega_L(p)=\langle S_{max}\rangle/L^2,
\end{equation}
where $S_{max}$ stands for the number of occupied sites in the largest cluster and $\langle\rangle$ denotes
configuration-average. In order to calculate $\Omega_L(p)$ for an $L\times L$ system , one needs to occupy the sites as per
the occupation probability $p$, count the number of sites in the largest cluster, average over many such configurations
and finally divide it by $L^2$. In fig. \ref{smax}(a), $\Omega_L(p)$ is plotted for four values of the distortion
parameter $\alpha=0,0.1,0.15,0.2$ from left to right for a fixed $d=1.1$. As the effect of distortion diminishes,
the curve shifts towards left. Similar situation is displayed in fig. \ref{smax}(b); here the influence of distortion
is reduced by increment in connection threshold ($d$), although $\alpha=0.1$ remains unchanged.
In fig. \ref{smax}(b), the two curves on the left are very close to each other.
This indicates that the undistorted scenario is being approached and the curves with still higher $d$ would be identical.
Figure \ref{scaling}(a) shows $\Omega_L(p)$ for different system sizes: $L=128,256,512,1024$, with the steepest one
corresponding to the largest $L$. The parameters of $\alpha=0.2$ and $d=1.1$ have been held fixed. For this set of
values we find
$p_c=0.8025$. Using the value of the critical exponent $\nu=0.75$ for standard percolation, when the horizontal
axis is scaled as $(p-p_c)L^{1/\nu}$, the curves of $\Omega_L(p)$ separate from each other
with the $L=128$ curve being on top and the $L=1024$ one at the bottom. The values of $\Omega_L(p)$ are collected
at a fixed value of $(p-p_c)L^{1/\nu}$ and using these values a plot of $\ln \Omega_L(p)$ with $\ln L$ is generated
(see inset of figure \ref{scaling}(a)). The slope of the straight line fit of this data gives $\beta/\nu=0.1143$.
As known from standard percolation criticality, if the vertical axis is now scaled as $\Omega_L(p) L^{\beta/\nu}$ and
plotted against $(p-p_c)L^{1/\nu}$, a data collapse should be obtained. We do get a nice data collapse in
figure \ref{scaling}(b) using the obtained value of $\beta/\nu$.
It is known that the percolating cluster at $p=p_c$ is a fractal whose fractal-dimension $d_f$ is given by
\begin{equation}
\langle S_{max}\rangle \sim L^{d_f}.
\label{dfeq}
\end{equation}
We calculate $\langle S_{max}\rangle$
for $L=2^7,2^8,2^9,2^{10}$ with $\alpha=0.2$ and $d=1.1$ at $p=0.8025=p_c$. Eq. \ref{dfeq} suggests that a log-log plot
of $\langle S_{max}\rangle$ and $L$ would fit into a straight line with its slope$=d_f$. In the inset of figure
\ref{scaling}(b), we show this plot along with a linear fit that gives $d_f=1.8857$. This confirms the well known
relation between the fractal dimension of the percolating cluster at $p_c$ and the relevant critical exponents in
two dimensional systems
\begin{equation}
d_f=2-\beta/\nu.
\end{equation}
The whole process has been repeated for two other sets of values of $\alpha$ and $d$ (plots are not shown as they look
very similar to those given in figure \ref{scaling}). For $\alpha=0.1, d=1.1$, we get
$p_c=0.6617, \beta/\nu=0.1150$ and for $\alpha=0.2, d=1.2$, we get $p_c=0.6665, \beta/\nu=0.1165$. These values
should be compared to those for the ordinary percolation, for which, $\beta/\nu=0.1012$ and $d_f=1.8958$.
These results tend to indicate that the present variant of percolation in a distorted
lattice may belong to the same universality class as the ordinary percolation. It has to be understood that within
the scope of the present work, the values of the exponents,
percolation threshold and fractal dimension are suggestive. More precise values, rigorous analysis
on their dependence on $\alpha$ and $d$ and a conclusive decision on universality class require further detailed
study and extensive numerical calculation involving averages over larger number of configurations with
(preferably) bigger lattice sizes. We plan for such a detailed study in our future endeavor.
\section{summary}\label{sum}
To summarize, we have proposed a new model of percolation in which the empty sites of a regular square lattice
are distorted. Distortion is incorporated into the system through a parameter $\alpha$, which randomly shifts
the position of each site within a square of length $2\alpha$ centered at the regular location of the site.
Thus the nearest neighbor distance may be more, equal or less than that of the undistorted
lattice. Two adjacent occupied sites are considered connected only when their distance is less
than a predefined value, called connection threshold $d$. In a Monte-Carlo study via HK algorithm, we find that
spanning becomes difficult for higher values of $\alpha$ and lower values of $d$. The value of the percolation
threshold $p_c$ depends on its interplay with the two parameters $\alpha$ and $d$ and varies within $p_{cu}\le p_c\le 1$.
Interestingly, if $d$ is less or equal to its value in the undistorted lattice (we take this value to be unity),
the system fails to percolate with any non-zero $\alpha$ (be it very small, meaning slight distortion) even when
all the sites are occupied. From the obtained values of the fractal dimension of spanning cluster at $p_c$ and the critical
exponents related to it, we predict with caution that this model may belong to the same universality class as usual
site percolation.
\begin{acknowledgments}
The computation facilities availed at the Department of Physics, University of Gour Banga, Malda is gratefully
acknowledged. One of us (AS) would like to thank Raja Paul of IACS, Kolkata for an illuminating discussion.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:Introduction}
\modif{A growing number of time series applications address training from orders of magnitude more series than the largest in the benchmark UCR repository--- the 8,926 training series ElectricDevices data set. In contrast, the phoneme dataset \cite{hamooni2014phoneme} contains 370,000 series. The satellite dataset \cite{tan2017indexing} contains 1,000,000 series.}
The prior state-of-the-art in time series classification does not scale
to such quantities.
In 2017, a meticulous study was conducted to compare the behaviour of the state-of-the-art \cite{bagnall2017great}. The authors draw the following conclusions:
\begin{enumerate}
\item The state-of-the-art is led by four classifiers that are: Collection of Transformation Ensembles (COTE) \cite{bagnall2015cote}, Elastic Ensembles (EE) \cite{lines2015ee}, Shapelet Transform (ST) \cite{hills2014shapelets} and Bag of SFA Symbols (BOSS) \cite{schafer2015boss}.
\item COTE is a special case in that it subsumes two of the other classifiers: it is a large ensemble classifier that includes EE and ST as sub-classifiers; COTE is on average, ``clearly superior to other published techniques.''
\item COTE's runtime complexity is bounded by (a) Shapelet Transform, which is $O(n^2\cdot l^4)$ \cite{bagnall2015cote} for $n$ time series of length $l$, and (b) the parameter searches for EE, some of which are $O(n^2\cdot l^3)$. The authors conclude ``An algorithm that is faster than COTE but not significantly less accurate would be a genuine advance in the field.''
\end{enumerate}
This is the challenge we tackle in this paper: developing an algorithm that is competitive with the accuracy of the state-of-the-art, but can learn from datasets with millions of time series.
We call our algorithm \emph{Proximity Forest}. It is a tree based ensemble that makes the most of the decades of research into developing consistent similarity measures for time series.
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=.48\linewidth]{Fig1a}}
\subfigure[]{\includegraphics[width=.48\linewidth]{Fig1b}}
\caption{\label{fig:intro-train-time}Comparison of Proximity Forest (in blue) with Elastic Ensemble (in red). (a) Classification accuracy as a function of the size of the training data. (b) Training time as a function of the size of the training data. Note that we take Elastic Ensemble because it is the classifier that prevents scalability of the state-of-the-art COTE ensemble, which includes EE as one of its classifiers. }
\end{figure}
Typical decision trees branch on the value of an attribute. Treating the values at each time stamp as belonging to a single attribute does not work well on time series because the relevant signals are not necessarily aligned by time stamp.
Instead, Proximity Forest branches on the \emph{proximity} of a query time series to a set of reference series.
`Proximity' is defined by a given (time series) similarity measure and a set of parameters (most time series measures have parameters that are critical to their proper function).
Our trees define separating hyperplanes for which the position is supported by time series themselves (whereas a traditional tree would split using a threshold on the value of an attribute).
Proximity Forest, as opposed to nearest neighbour approaches, truly abstracts a model from data, which makes it possible to classify with time that is logarithmic with respect to training set size, as opposed to linear time for EE and COTE.
Moreover, we will show that we specifically designed its training to scale linearly with the quantity of data, as opposed to at least quadratically for EE, COTE and ST.
\modiff{
Figure~\ref{fig:intro-train-time} shows the accuracy and the training time required by Proximity Forest and EE on our satellite dataset with increasing training set size. Two important elements are illustrated: (1) Proximity Forest scales linearly with training set size while EE scales quadratically; and, (2) Proximity Forest's classification accuracy on this dataset is substantially better than EE, even when they train on the same quantity of data.
For our application, Proximity Forest can learn from 1M time series in 17 hours (on 1 CPU) while it would take over 200 years for EE --~a 103,000x speedup. Furthermore, ST and COTE learn slower than EE and thus will have \emph{even larger} training times. Note that EE is a component of COTE and hence sets a lower bound on COTE's training time. Our discussion will often focus on EE because EE is very similar to our algorithm in that it is trying to leverage existing time series similarity measures.
The virtues of Proximity Forest are not limited to large datasets, however. Our experiments show that it also outperforms EE in classification accuracy on the majority of the datasets of the UCR Archive.}
The remainder of this paper is structured as follows: in Section~\ref{sec:Related work}, we review the state-of-the-art in time series classification, with a particular focus on scalability.
We then introduce Proximity Forest in Section~\ref{sec:Proximity Forest}.
Our experiments (Section~\ref{sec:Experiments}) show that Proximity Forest (1) outperforms all other scalable algorithms on our case study in both accuracy and training time; and (2) is competitive with the state-of-the-art on UCR data in terms of accuracy. The section ends with a study of Proximity Forest's parameters and a discussion of how their values vary the results.
\section{Time series classification -- related work}
\label{sec:Related work}
We present here a non-exhaustive review of the state-of-the-art in time series classification and similar decision tree-based algorithms. We focus on our particular interest in this paper: scalable training and classification.
\subsection{Distance-based classification}
\subsubsection{Distances}
Time series have particular properties that have led to the development of specific similarity measures: they are often auto-correlated (the value of the time series at a timestamp is likely to be close to the ones just before and after), and often include non-linear distortions in the time axis (for example, because the start of the phenomenon of interest is delayed, or because sections of the phenomena are faster or slower).
This has rendered typical similarity measures severely flawed and led to the development of specific similarities, of which most have an ability to re-align the series along a common intrinsic time-line.
Important measures
include Dynamic Time Warping (DTW) \cite{sakoe1971dynamic,sakoe1978dynamic}, Derivative DTW (DDTW) \cite{keogh2001derivative,gorecki2013using}, Weighted DTW (WDTW) \cite{jeong2011weighted}, Longest Common Subsequence (LCSS) \cite{vlachos2006indexing}, Edit Distance with Real penalty (ERP) \cite{chen2004marriage,chen2005robust}, Time Warp Edit distance (TWE) \cite{marteau2009time} and Move-Split-Merge (MSM) \cite{stefan2013move}.
A more complete description and comparison of these distance measures can be found in \cite{bagnall2017great,lines2015ee,wang2013experimental}.
Note also that most of these distances have parameters of which tuning is critical to their functioning.
\subsubsection{Nearest-neighbor approaches}
It is most common to
classify time series data using Nearest Neighbour classification based on a relative distance (such as those returned from the measures above) \cite{Haghiri2017}. In fact, for more than a decade, the NN algorithm combined with the DTW measure was extremely difficult to beat \cite{wang2013experimental}.
It is important to note here that when researchers mention the use of NN with a time series measure, the measure is not directly applied with a default parameterization, but rather its parameters are first learned on the data, usually by cross-validation.
There are two main issues with NN approaches: (1) the tuning of the measures' parameters is usually quadratic with the size of the training data, and (2) the classification is at least linear with the size of the training data. Both of these issues are further compounded by the fact that most measures have a computational complexity that sits between linear and quadratic with the length of the series.
To alleviate the second issue of scalable classification, targeted techniques have been developed. Data reduction techniques aim at simplifying the training database without penalizing the classification quality; either by directly removing objects from the original database \cite{pkekalska2006prototype,ueno2006anytime} or by summarizing the database and replacing sets of time series with representatives using average time series \cite{petitjean2012summarizing,petitjean2014dynamic,petitjean2016faster,marteau2016averaging}.
Indexing is more difficult on time series than it is on traditional data, mostly because time series measures do not obey the triangular equality (at most obeying a relaxed $p$-triangular inequality \cite{lemire2009faster}), which makes exact pruning very inefficient (however, a general approach to improving the efficiency of NN searches in non-Euclidean space is available in \cite{Lifshits2010}).
To perform exact indexing, the main research effort has been put onto developing lower bounds (and mostly for DTW) \cite{keogh2006lbkeogh,lemire2009faster}.
Recently,
impelled by the motivating application
\modif{of earth observation data analytics}, we have developed an algorithm for approximate and efficient NN search under DTW \cite{tan2017indexing}, an algorithm using the idea of a hierarchical $k$-means clustering
As mentioned previously, Elastic Ensemble (EE) \cite{lines2015ee} is a recent state-of-the-art time series classifier. It is an ensemble of 11 NN classifiers, each learned with a different time series measure (with their parameters tuned accordingly). The EE algorithm has played a significant role in the design of Proximity Forest and will be discussed at greater depth in Sections~\ref{novelPF} and ~\ref{learnPF}.
\subsection{Approaches that learn features}
The following approaches construct an abstraction of the training dataset by learning features that represent the classes in the time series.
\subsubsection{Shapelets} \label{shapelets}
The aim of shapelet algorithms is to find subseries (or consecutive subsets of time series) that can help discriminate between the different classes.
To classify a time series, the learned shapelet is placed at the best position in the time series (usually under Euclidean distance), and the `matching' of the shapelet to the time series correspond to its distance at this best position. The original shapelet-classifier \cite{ye2011shapelets} inserted this algorithm at the node of a decision tree as a splitting criterion.
This algorithm has a high training complexity ($O(n^4)$) due to the large number of candidate shapelets and the repeated scanning of the data. Subsequent research has focused on optimising the original algorithm to address both classification accuracy and scalability, notably Fast Shapelets \cite{rakthanmanon2013fast}, Learning Time Series Shapelets \cite{grabocka2016fastshape} and Shapelet Transforms \cite{hills2014shapelets}.
Shapelet Transforms (ST) is a current state-of-the-art classifier that identifies the best $k$ shapelets in a single scan of the data (the number of shapelets can be reduced afterwards). The data is then transformed by defining an attribute to represent each shapelet with the value being the (usually Euclidean) distance between the shapelet and the best position in the time series. The transformed dataset can now be used with any classifier or ensemble of classifiers (such as in \cite{bagnall2015cote}). While ST is considered a state-of-the-art classifier, it has little potential to scale to large datasets given its training complexity of $O(n^2\cdot l^4)$.
\subsubsection{Bag of Words approaches}
Bag of Words algorithms are similar to Shapelets in that they start by identifying exemplar subseries in the data to discriminate between classes. However rather than finding the similarity to the relative best positions in a time series, bag of words approaches differentiate classes by the relative frequency of the subseries. To calculate these frequencies, the algorithms discretise the values into a series of symbols, assigning letters to each subseries, and thus representing the original time series as `words'. Notable approaches are the Bag of Patterns \cite{Lin2012bop}; the Symbolic Aggregate Approximation-Vector Space Model (SAX-VSM) \cite{senin2013sax}; and the Bag of SFA Symbols (BOSS) \cite{schafer2015boss}, which is currently considered state-of-the-art.
The BOSS algorithm transforms the time series into a word using a Symbolic Fourier Approximation (SFA) \cite{Schafer2012SFA} thus making it robust to noise and delivering a high classification accuracy. It is however of limited use on large datasets as it has a high training complexity $O(n^2)$ \cite{bagnall2017great}. The authors identified this as a weakness and subsequently produced similar approaches with improved scalability, the Bag of SFA Symbols in Vector Space (BOSS-VS) \cite{schafer2015scalable}. The same authors recently proposed WEASEL \cite{Schafer2017WEASEL}, which improves on the computation time of BOSS and on the accuracy of BOSS-VS, but has a very high memory complexity (our experiments will show that it doesn't scale beyond 10,000 time series). In this way, WEASEL is more optimised for speed on small datasets than for scalability.
\subsection{Ensemble approaches}
Ensemble approaches are combinations of multiple classifiers. Each contributing algorithm can be weighted to maximize classification accuracy, while the time complexity is that of the slowest constituent. Some of these approaches have been discussed above as they are based around one main type of classifier, for example EE and ST.
The Collection of Transformation Ensembles (COTE) \cite{bagnall2015cote} is an ensemble comprising 35 classifiers across four time series domains: time, frequency, change and shapelet transformation.
For the time domain, COTE uses the 11 distance measures of EE, while in the other three domains, classifiers are recruited from outside time series classification -- $k$-nearest neighbours, naive Bayes, decision trees, random forest, rotation forest, support vector machines (two models) and a Bayesian network approach. On the benchmark UCR datasets, COTE has the highest average classification accuracy of all current approaches. However, its time complexity is bound by that of the Shapelet Transform, which is $O(n^2\cdot l^4)$ and the parameter searches for the elastic distance measures (EE), some of which are $O(n^2\cdot l^3)$.
\modif{
\subsection{Decision Tree approaches}\label{sec:treeapproaches}
A number of decision tree approaches have been developed for time series classification.
Time Series Forest (TSF) \cite{deng2013TimeSeriesForest} first derives summary features for all time series by dividing them into intervals and summarising each interval by its mean, standard deviation and gradient.
Then a Random Forest-like strategy is employed to select between a random subset of these features at each node in each of an ensemble of trees. A novel selection criterion is used that considers both entropy gain and the margin by which a feature separates the classes.
This continues until the entropy gain ceases to improve, at which stage the node is defined as a leaf. TSF has been shown to be a reasonably accurate classifier: its accuracy ranks behind EE and DTW without being significantly worse \cite{bagnall2017great}. \modiff{However, its main virtue is computational efficiency. TSF learns in $O(n \log(n)\cdot l \cdot k)$ for a forest of $k$ trees built from $n$ series of length $l$, which is a much lower complexity than the current state-of-the-art.}
Generalized Random Shapelet Forest (gRSF) \cite{karlsson2016ShapeletForest} extracts a shapelet from
a randomly chosen time series and finds the distance between this time series and each other time series. The data is then split according to whether it is above or below a threshold distance to the representative shapelet. This is applied recursively until the node is either pure on the number of instances remaining at a node is less than 3. As mentioned in section \ref{shapelets}, the main pitfall of shapelet-based methods is the high computational cost of finding candidate shapelets and comparing shapelets to other time series. The gRSF minimises this issue by randomising many of the model choices---a candidate shapelet is generated from a randomly chosen time series by choosing a random starting point and random length, this is repeated $r$ times and the best candidate is chosen for a given split. The resulting algorithm has accuracy competitive with Learning Time Series Shapelets and better than DTW.
\modif{A number of approaches have been developed that form decision trees where splits are based on similarity to chosen exemplars \cite{Balakrishnan2006,Douzal2012treetimeseries,Yamada2003decision}. One strategy is to select a single exemplar and then choose a cut point on a similarity measure with respect to that exemplar. Series with similarity scores lower than the cut point follow one branch and the remaining examples follow the other. The other strategy is to select multiple exemplars, one associated with each branch. Series follow the branch with whose exemplar they are most similar.
These approaches are hampered by the high computational complexity of their search for exemplars at each node.
Similarity Forests \cite{Sathe2017SimForest} and Comparison-based Random Forests \cite{Haghiri2018} generalise this idea to attribute-value data with random selection of exemplars and developed forests of such trees. Similarity Forests add a cutoff value on the difference in the distance between the two exemplars, and optimizes that cutoff value based on weighted Gini.}
The idea of using similarity as the splitting criterion in tree structures has also been successfully used for \emph{indexing} of regular tabular data \cite{flann,annoy,Sathe2017SimForest} and of time series with DTW \cite{tan2017indexing}.
}
\section{Proximity Forest}
\label{sec:Proximity Forest}
In this section, we present our novel algorithm for time series classification: Proximity Forest. We start by highlighting why there is a need for a new time series classifier. We then present our model and the two key algorithms (1) how to learn a Proximity Forest and (2) classifying with a Proximity Forest. We conclude this section with some comments about its complexity.
\subsection{Why do we need a novel time series classifier?}
\label{novelPF}
The previous section highlighted that the last decade has seen numerous classifiers and distance measures specifically designed for time series classification. Based on this, one could wonder why there is a need for a novel algorithm; the answer is simple: most state-of-the-art algorithms do not scale to large time series datasets. We have seen that some do not scale in the learning phase (ST, EE, COTE). Others require a scan of the training database to perform each classification (EE, COTE). Those that do scale to medium-size datasets, such as BOSS-VS, compromise accuracy in order to do so (as we will show for both our case study and for UCR datasets). Throughout the development and advancement of much of the current state-of-the-art, scalability has usually been secondary to accuracy. This is because time is not a significant issue when considering data with only few time series. However
\modif{a growing number of modern applications
consist of hundreds of thousands to millions of time series.}
These applications require a classifier that is both accurate \emph{and} scalable in both learning and classification.
BOSS-VS is a classifier that appears to have developed with a focus on scalability. However, as we will see in Section~\ref{sec:Experiments}, its accuracy ranks some 30 percentage points lower in our case study, and therefore is not competitive with the accuracy of the state-of-the-art.
While COTE is currently \emph{the} state-of-the-art \modif{in terms of accuracy}, its learning phase is bound by the runtime complexity of both ST and EE.
On our 1M dataset --~and as depicted in Figure~\ref{fig:intro-train-time}~-- the sole learning phase of COTE associated to training EE would require \textbf{73 thousand} days, or 200 years. This is even more startling knowing that the series in this dataset are very short with only 46 timestamps.
\modif{The large runtime complexity of COTE is largely due to the fact that EE does not abstract much information during the learning phase, and therefore has a significantly greater number of processes to complete during testing. A corollary of this is that a distance-based classifier that learns faster than EE for the same level of accuracy would also present an improvement to COTE. It is for this reason that our design of algorithm incorporates many elements of EE---11 distance measures and similar parametrisation---and why our experiments provide a direct comparison of Proximity Forest against EE.}
We therefore argue that the need for a scalable and accurate classifier has not yet been met.
We incorporate three critical elements into the design of our novel approach:
\begin{enumerate}
\item We make the most of over 30 years of research into designing consistent measures for time series.
\item We specifically design our ensemble to have a high variability between the different individual classifiers. This results in an improved overall classification accuracy over a single classifier, based on the principle of ensemble methods. In general, averaging the predictions of multiple models each having high variance and low bias results in an ensemble classifier with a lower total error than any single classifier. This is analogous to how a Random Forest model, another ensemble of decision trees, will only learn from a fraction of the available features for each individual node in order to introduce variability between the trees \cite{Breiman2001RF,Ho1995Rdf}. This observation is important, because we did not design the learning of an individual tree to maximize its accuracy; if we had wanted to design a single tree model, we would have made different design choices. We designed the learning of individual trees so that the overall classification performance is maximised.
\item We design Proximity Forest to be extremely scalable with an \modif{average-case learning complexity of $O(n \log\left( n\right) \cdot l^2)$ and a classification complexity of $O(\log \left(n\right)\cdot l^2)$ per tree for $n$ training time series of length $l$. This contrasts with the state of the art learning in $O(n^2\cdot l^3)$ (Elastic Ensemble) or $O(n^2\cdot l^4)$ (Shapelet Transform, COTE). Again here, we might have made different choices if scalability wasn't a design objective.}
\end{enumerate}
\modif{To achieve scalability we employ tree-based classifiers. These are scalable due to their use of a divide-and-conquer strategy. At each level the data are divided into multiple subsets, as result of which the trees are on average of depth $O(\log n)$, hence increasing sublinearly in depth relative to training set size.}
\modif{Our use of decision trees for time series classification is not novel in itself. Trees are attractive due to their divide and conquer methodology and resulting potential for efficient learning and classification. Previous implementations, however, have lacked competitiveness in accuracy \cite{deng2013TimeSeriesForest,Douzal2012treetimeseries,Yamada2003decision} or time \cite{Balakrishnan2006,Douzal2012treetimeseries,Yamada2003decision}.
}
\modif{To achieve scalability we merge the strategy of learning decision trees where splits are based on similarity to chosen time series exemplars \cite{Balakrishnan2006,Douzal2012treetimeseries,Yamada2003decision} with the strategy of forming forests of such trees in which the exemplars are chosen at random \cite{Sathe2017SimForest}. To this amalgam we add the critical ingredient of stochastic selection between a large range of similarity measures, which both reduces bias and provides a beneficial increase in variance between ensemble members.}
\subsection{How to learn a Proximity Forest?}
\label{learnPF}
We seek to learn a Proximity Forest from a training set comprising $n$ labeled time series, each of which is of length $l$, where the labels are integers from $1$ to $c$.
A Proximity Forest is an ensemble of $k$ Proximity Trees.
A Proximity Tree is similar to a regular decision tree, but differs in the tests applied at internal nodes. Whereas a regular decision tree applies a test based on the value of an attribute (e.g.\ if height > 160 cm, follow the left branch, otherwise follow the right branch), each \emph{branch} of an internal node of a Proximity Tree has an associated exemplar
and an object follows the branch corresponding to the exemplar to which it is \emph{closest} according to a parameterized similarity measure. We will see later how the exemplars and measures are chosen.
A~tree is either a leaf or an internal node.
An internal node has two fields, \texttt{measure}, a function \modif{$\mathit{object}\times \mathit{object} \rightarrow \mathbb{R}$}, and \texttt{branches}, a vector of branches. Each branch has two fields, a time series (\texttt{exemplar}) and a tree to which an object is passed if it is nearer to the branch's exemplar than any other (\texttt{subtree}).
\modiff{
If all data reaching a node has the same class, i.e. is pure, the $create\_leaf$ function creates a new leaf node and assigns this class label to its field \texttt{class}. This label is then assigned to any query time series reaching this leaf during the testing phase.
}
\paragraph{How do we choose the splitting criteria?} A Proximity Tree creates, at each node, one branch for each class that exists in the data it receives from its parent. These exemplars are chosen uniformly at random among each class. The parameterized similarity measures are also chosen uniformly at random among a pool that will be described below after we have given the main overview of the algorithm. We will detail, after the main algorithm, how it is possible to \emph{learn} with randomized trees.
Algorithm~\ref{build} presents the algorithm for learning a single tree. Each node is constructed recursively from the root node down to the leaves. If the data at the node is pure --~ie. all data belongs to the same class~-- then the node becomes a leaf and the recursion finishes.
\modif{At each node, a pool of $r$ candidate splits are evaluated (Algorithm~\ref{newsplitter}). For each candidate, a parameterised measure is chosen uniformly at random among a pool
of such measures. We then select an exemplar for each class represented at the node and pass the data down the branches by finding the closest exemplar (one per class) for each time series in the data using the split's distance measure.
Once each candidate split has been created, we then select the candidate that maximizes the difference between the Gini impurity of the parent node and the weighted sum of Gini impurity of the child nodes. We then call the construction of the tree recursively on each branch for the successful candidate; this constructs all subtrees. When this is done, the tree is constructed.
Increasing the number of candidate splits per node will lead to an improvement of the quality of each split. However, it will also lead to an increase of the training time. The choice for the value of $r$ will be discussed later in Section~\ref{subsubsec:candidates_eval}.
}
\modiff{
\begin{algorithm}
\caption{$\mathrm{build\_tree}(D,\Delta,R)$}
\label{build}
\DontPrintSemicolon
\KwIn{$D$: a time series dataset}
\KwIn{$\Delta$: a set of parameterized distance measures}
\KwIn{$R$: number of candidate splits to consider at each node}
\KwOut{$T$: a Proximity Tree}
\vspace*{5pt}
\uIf{$\mathrm{is\_pure}(D)$}{\Return $\mathrm{create\_leaf}(D)$\;
}
\vspace*{5pt}
\tcp{create tree, to be returned, represented as its root node}
$T \leftarrow \mathrm{create\_node}()$\;
\vspace*{5pt}
\tcp{Creating $R$ candidate splitters}
$\mathcal{R} \leftarrow \emptyset$\;
\For{$i=1$ \KwTo $R$}{
$r\leftarrow gen\_candidate\_splitter\left(D,\Delta\right)$ \tcp*{generate random splitter}
Add splitter $r$ to $\mathcal{R}$\;
}
\vspace*{5pt}
\tcp{select best splitter; it splits using measure $\delta^\star$ and exemplars $E^\star$}
$(\delta^\star,E^\star)\leftarrow \argmax\limits_{r\in \mathcal{R}} \mathrm{Gini}\left(r\right)$\;
\vspace*{5pt}
$T_\delta \leftarrow \delta^\star$ \tcp*{retain measure for root node of $T$}
$T_B\leftarrow \emptyset$\tcp*{$T_B$ will store the branches under root node of $T$}
\ForEach{\emph{exemplar} $e \in E^\star$}{
\tcp{$D^\star_e$ is the subset of $D$ that are the closest to $e$ based on $\delta^\star$}
$D^\star_e \leftarrow \left\{d\in D \;\raisebox{-4pt}{\resizebox{2.5pt}{20pt}{$\mid$}}\; \argmin\limits_{e'\in E^*}\delta^\star(d, e') = e\right\}$\;
$t\leftarrow \mathrm{build\_tree}(D^\star_e,\Delta,R)$ \tcp*{build subtree for that branch}
Add branch $\left(e, t\right)$ to $T_B$\tcp*{a branch is a pair $(\text{exemplar,sub-tree})$}
}
\vspace*{5pt}
\Return $T$
\end{algorithm}
}
\modiff{
\begin{algorithm}
\caption{$\mathrm{gen\_candidate\_splitter}(D,\Delta)$}
\label{newsplitter}
\DontPrintSemicolon
\KwIn{$D$: a time series dataset.}
\KwIn{$\Delta$: a set of parameterized distance measures to sample from}
\KwOut{$(\delta, E)$: a parameterized distance measure and a set of exemplars}
\vspace*{5pt}
$\delta \xleftarrow[]{\sim}\Delta$ \tcp*{sample a parameterized measure $\delta$ uniformly at random from $\Delta$}
\vspace*{5pt}
\tcp{Select one exemplar per class to constitute the set $E$}
$E \leftarrow$ $\emptyset$\;
\ForEach{\emph{class} $c$ \emph{present in} $D$}{
$D_c \leftarrow \left\{d\in D \mid class(d) = c\right\}$ \tcp*{$D_c$ is the data for class $c$}
$e \xleftarrow[]{\sim}D_c$ \tcp*{sample an exemplar $e$ uniformly at random from $D_c$}
Add $e$ to $E$
}
\vspace*{5pt}
\Return $(\delta, E)$
\end{algorithm}
}
\paragraph{The case of $R=1$: is selecting at random still `learning'?} One might wonder what the tree is actually learning when one only considers a single candidate ($R=1$). In that case, no selection of `the best possible split' is performed. It is interesting to note that choosing splitting criteria independently of the output value has been studied before, a key example being Extremely Randomized Trees \cite{geurts2006extremely}. In that work, they showed that splitting completely at random still ensures consistency (tending to Bayes Optimal error as the data tends to infinity).
The main reason is that the exemplars are not random points in the input space. They are sampled from the data distribution of each class. In~consequence, the trees \emph{are} still learning an abstraction of the data, using the trees as a density estimator \cite{ting2016overcoming}.
We depict in Figure~\ref{fig:splitting} a graphical representation of a simple split obtained on the \texttt{Trace} dataset. It is interesting to see that \modiff{
in Euclidean space, the splitting criterion is actually forming a hyperplane that is equidistant to the exemplars. Note that this intuition is more complex for time series measures, because most of them do not have properties of a metric \cite{lemire2009faster}. The scatter plot depicts each time series as a dot in this space, with the x-axis representing distance to the first exemplar and the y-axis distance to the second.}
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{Fig2}
\caption{Visual depiction of the root node for the `Trace' dataset (simplified to 2 classes). The top chart represents the data at the root node (one colour per class) while the data at the bottom left and right represent the data once split by the tree. The two time series in the middle left and right are the exemplars on which the tree is splitting. The scatter plot at the center represents the distance of each time series at the root node with respect to the left and right exemplars (resp. x- and y-axes). }
\label{fig:splitting}
\end{figure}
\paragraph{How to choose the parametrised measure on which to split?}
The parametrised distance measure gives a measure of the similarity between the exemplar time series. \modif{For each candidate split at each node}, the algorithm chooses a distance measure at random from the following 11 distance measures used by the Elastic Ensemble (EE) learner that we described above: Euclidean Distance (ED); Dynamic Time Warping using the full window (DTW); Dynamic Time Warping with a restricted warping window (DTW-R); Weighted Dynamic Time Warping (WDTW); Derivative Dynamic Time Warping using the full window (DDTW); Derivative Dynamic Time Warping with a restricted warping window (DDTW-R); Weighted Derivative Dynamic Time Warping (WDDTW); Longest Common Subsequence (LCSS); Edit Distance with Real Penalty (ERP); Time Warp Edit Distance (TWE); and, Move-Split-Merge (MSM). Randomising the choice of distance measure is a deliberate decision to introduce variability between each tree, for the reasons stated earlier.
Once a distance measure is chosen at random, it is then parametrised.
\modif{The parametrisations are addressed in turn. They are deliberately chosen to mimic as closely as possible the EE algorithm. Even though better values might be chosen here, we mimic EE's parameterization to allow direct comparison. Euclidean distance, full DTW, and full DDTW distances have no parameters to select. DTW-R and DDTW-R only require a warping window parameter that is chosen uniformly at random in $[\!\![0,\lfloor\frac{l+1}{4}\rfloor]\!\!]$ (thus allowing a warping of elements at most $\frac{l}{2}$ apart). WDTW and WDDTW requires also one parameter to select that it is used into the weighted value $g$ to control the level of penalization between two different time stamps -- we use $g\sim U(0,1)$. The parametrisation of ERP is a distance threshold that controls for how close elements have to be to be considered similar; we sample it uniformly at random in $[\frac{\sigma}{5},\sigma]$, with $\sigma$ being the standard deviation of the data. LCSS has as first parameter the same distance threshold value (which is sampled in the same way), and has a second parameter --~the warping window size~-- which is chosen in the same way as for DTW-R. TWE has two parameters $\gamma$ and $\lambda$ which respectively control for the stiffness and penalty value in the alignment. Following \cite{marteau2009time}, $\lambda$ is sampled at random from $\cup_{i=0}^{9} \frac{i}{9}$ and $\gamma$ following at random from the exponentially growing sequence $\{10^{-5},10^{-4},5\dot 10^{-4},10^{-3},5\dot 10^{-3},\cdots,1\}$, resulting in 100 possible parameterizations. The final measure, MSM, has a single parameter which is sampled from an exponential sequence similar to the one for $\gamma$ in TWE with 100 values ranging from $10^-2$ to $10^2$, as recommended in \cite{stefan2013move}.
Choosing the parameter at random has a twofold effect: 1) it skips the cross-validation step which has a quadratic complexity; and 2) it introduces variability between trees, which provides superior learning through lower-biased trees and ensembling.
}
In the following experiments we will show that Proximity Forest is not only orders of magnitude faster than EE, but that its accuracy also ranks higher than EE.
\subsection{Classifying with a Proximity Forest}
The process of classification for a single Proximity Tree is detailed in Algorithm~\ref{alg:classification}: a query time series begins at the root node and the distance from the query to each of the exemplar time series is calculated, by using the node's distance measure and exemplars selected when constructing the tree. The query time series is then passed down the branch of the exemplar to which it is nearest. The query time series then traverses down the tree by repeating this process until it reaches a leaf, where it is assigned the class represented by that leaf. This process is repeated for each tree constructed as part of the forest. A~Proximity Forest then uses majority voting between its constituent Proximity Trees.
\begin{algorithm}
\caption{classification$\left(Q, T\right)$}
\label{alg:classification}
\DontPrintSemicolon
\KwIn{$Q$: Query Time Series}
\KwIn{$T$: Proximity Tree}
\vspace*{5pt}
\uIf{$\mathrm{is\_leaf}(T)$}{\Return majority class of $T$}
\vspace*{5pt}
\tcp{find the branch with exemplar closest to $Q$ using measure $T_\delta$}
$(e, T^\star) \leftarrow \argmin\limits_{(e',T')\in T_B} T_\delta(Q,e')$\;
\vspace*{5pt}
\Return classification$\left(Q, T^\star\right)$ \tcp*{recursive call on subtree $T^\star$}
\end{algorithm}
\subsection{Comparative complexity analysis}
\label{subsec:computation_analysis}
\modif{
During the training phase, at each node, let us assume that $n'$ data points are present at the node. We first scan it once taking $O(n')$ time to split the data into $c$ groups, one for each class $c$ present at the node
}
\modif{We then generate $r$ candidate splits, \textit{i.e.}, {}$r$ sets of exemplar time series. For each such candidate set, we sample $c'\leqslant c$ exemplar time series, \textit{i.e.}, {}one time series for class represented among the $n'$ time series available at the node -- this is done in $O(1)$ given that the data is already organised by class. For the candidate split to be operational, we also require a parameterized measure to use to compare against these $c'$ exemplars. Most of the parametrized measures can be chosen in $O(1)$, except for LCSS and ERP which calculate the standard deviation in $O(n'\cdot l)$ for data at the node while selecting the parameter.\footnote{\modif{Note that these parametrisations can be performed in constant time also if the data are $z$-normalized, which is the case for all UCR datasets.}}}
\modif{We now have $r$ candidate splits that are ready to be evaluated. We now wash the $n'$ time series down the branches for all candidate splits. This is done by comparing each time series to the $c'$ exemplars, each comparison taking from $O(c\cdot l)$ to $O(c\cdot l^2)$. Overall, this takes $O(n'\cdot c'\cdot l^2)$. If $r=1$, the training process at this node is finished and we call the training function recursively for each of the $c'$ children nodes. If $r>1$, we calculate the Gini coefficient for each of the $r$ candidate splits in $O(c^2)$, keep the best one, and delete other candidate splits.}
\modif{As the total number of examples that reach any of the nodes at a single given level cannot be greater than the total number of examples, $n$, the total computation per level of the tree is thus $O(n\cdot r\cdot c\cdot l^2)$. In the worst case, the majority of the training data at each level will pass down a single branch and the depth of the tree will be $O(n)$, resulting in a worst training time complexity of $O(n^2\cdot r\cdot c\cdot l^2)$. However, as the exemplars are following the class distribution, unless the data are in some way degenerate (for example if one class comprises only outliers), the average tree depth can be expected to be $O(\log n)$. In practice it will often be much smaller, because, unlike typical divide and conquer approaches, the tree terminates as soon as a node is pure rather than having to separate each individual object. Thus, for non-degenerate data we can expect average case training time complexity of $O\left(n\log\left( n\right) \cdot r\cdot c\cdot l^2\right)$ for a single tree and thus $O\left(k\cdot n\log\left(n\right) \cdot r\cdot c\cdot l^2\right)$ for a full Proximity Forest comprising $k$ Proximity Trees.}
The experiments presented in the next section will include runtimes and comparison to current state-of-the-art algorithms. These confirm that this expected average case quasi-linear complexity with respect to data quantity is borne out in practice.
During classification, a time series of length $l$ will pass through an average of $\log n$ nodes on each of the $k$ trees. At each node, the distance to at most $c$ exemplars must be computed. For each of these distance computations, the complexity will again depend upon the chosen distance measure; the fastest being $O(l)$ and the slowest $O(l^2)$. Thus, the resulting average case complexity is $O(k\cdot \log n \cdot c \cdot l^2)$.
\section{Experiments}
\label{sec:Experiments}
This section describes the experiments that evaluate our Proximity Forest.
\modif{We start with} the Satellite Image Time Series (SITS) dataset, a (very) large time series dataset describing the evolution of the Earth as pictured every five days by a high-resolution satellite. This dataset \modif{is an example of the large time series datasets that motivate} the need for a new time series classification algorithm, as no current state-of-the-art approach scales to this magnitude. Conversely, there are classifiers designed for scalability, namely BOSS-VS, that compromise classification accuracy to do so. The first experiments presented in this section use the SITS dataset to demonstrate the ability of Proximity Forest to be both scalable \emph{and} accurate. The second section assesses the Proximity Forest on the datasets of the UCR time series classification repository \cite{UCR}, the benchmark in the field. It demonstrates that the classification accuracy of Proximity Forest is competitive with the current state-of-the-art. The final section discusses other considerations surrounding Proximity Forest, such as the effect of varying the number of trees, and the standard deviation of the results.
It should be mentioned that throughout the following experiments we have emphasized a comparison with EE. This is because it is viewed as the closest relative to Proximity Forest amongst the current state-of-the-art, given that neither method includes data transforms or shapelets. It is also the constituent of COTE that bounds its learning time and therefore any improvement over the runtime of EE, for the same classification accuracy, would also equate to an improvement on COTE, the current leader in the field.
To facilitate others to build on our work, as well as to ensure reproducibility, we have made our code and the full raw results available at \url{https://github.com/fpetitjean/ProximityForest/}{}.
\subsection{Case study: Satellite Image Time Series Dataset}
\label{subsec:sits_expe}
The SITS dataset contains approximately 1 million time series with a train-test split of approximately 90\%-10\%.\footnote{The split ensures that no 2 times series come from the same plot of land.} Each time series has a length of 46 and is labeled as one of 24 possible land-use classes (e.g.\ `wheat', `corn', `plantation', `urban'). \modif{Here the labeled data has been extracted from three sources: 1) ground field campaigns for most of the vegetation classes, 2) farmer's declaration to complete the data for some crop classes, and 3) existing map for the urban areas.}
The experiments presented in this section were performed on this dataset, comparing the performance of Proximity Forest against three competitors: BOSS-VS (designed for scalability), WEASEL (designed for speed and quality), and EE (designed for quality). We use 5 runs for each experiment of Proximity Forest and 1 run of each of the competitors -- as their results are deterministic.
Throughout this experiment, we use 100 trees; we will see in Section~\ref{subsec:pf_variations} that this gives a good tradeoff between accuracy and computational time/memory. \modif{Although we are mainly assessing the scalability, we will also have a quick look to the accuracy}.
\subsubsection{Training scalability}
To assess scalability, we train and test each algorithm on subsample data with increasing training set size, allowing training time, testing time and accuracy to be measured as a function of training size. Figure~\ref{fig:SITS-train-test}(a) shows training time against training size for each of the 4 algorithms.
\paragraph{Versus EE.}First, it is evident that Proximity Forest presents a notable saving in training time over EE, confirming that it trains in linear time rather than the quadratic time for EE. Even for a small training set of about 2,000 time series, learning an EE model took about 10 hours, compared to Proximity Forest's 79 seconds. Fitting a quadratic curve through both EE and Proximity Forest is quite informative: EE returns a quadratic component of 6.3 while Proximity Forest only \num{-8.10e-6}, clearly highlighting both the quadratic complexity of EE, and also that Proximity Forest is in practice very close to its theoretical average complexity and scales quasi-linearly with $n$.
\paragraph{Versus WEASEL.}
WEASEL is very fast but its memory footprint did not allow it to scale beyond 8,000 time series even when given 64 GB of RAM. This clearly highlights the difference with BOSS-VS: we can see that WEASEL was not developed for scalability, but rather for speed on small datasets.
\paragraph{Versus BOSS-VS}Proximity Forest trains slower than BOSS-VS for a given training size, however this is counteracted by \modif{the low accuracy of BOSS-VS} discussed below.
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=.46\linewidth]{Fig3a}}
\subfigure[]{\includegraphics[width=.46\linewidth]{Fig3b}}
\caption{Training time (a) and testing time per query (b) as a function of training set size for Proximity Forest, EE, WEASEL and BOSS-VS.}
\label{fig:SITS-train-test}
\end{figure}
\subsubsection{Testing scalability}
\modif{Figure~\ref{fig:SITS-train-test}(b) shows testing time against training size for each of the 4 algorithms. }
The story here is very similar to that of training: it confirms the way Proximity Forest scales logarithmically with training set size, while EE must scan the full database many times. Here again, WEASEL becomes infeasible to apply with relatively small quantities of training data. Proximity Forest and BOSS-VS require respectively 0.0679 ms and 0.0077 ms to classify a time series with a model trained on 1M time series.
\subsubsection{Is Proximity Forest accurate and scalable?}
We have now seen that Proximity Forest is highly scalable and only beaten by BOSS-VS in terms of training time. We will now study how its accuracy scales with training set size.
The main results are presented in Figure~\ref{fig:SITS-accuracy} which plots the accuracy as a function of training set size for Proximity Forest, EE, WEASEL and BOSS-VS.
\begin{figure}
\centering
\includegraphics[width=.65\linewidth]{Fig4}
\caption{Accuracy as a function of training set size for Proximity Forest, EE, WEASEL and BOSS-VS.}
\label{fig:SITS-accuracy}
\end{figure}
The first element to observe is that Proximity Forest obtains \modif{greater accuracy than the competitors} for large training sets. WEASEL and EE become infeasible to apply at relatively small data quantities and BOSS-VS --~which is faster than Proximity Forest~-- does not learn effective classifiers on this dataset. With
63.8\% accuracy at 3,400 training set size, this is
26.3 percentage points more accurate than BOSS-VS, and
4.6 and 4.7 percentage points more accurate than WEASEL and EE, respectively. \modif{Such differences are substantial in a problem comprising 24 classes.}
\modif{Moreover, Proximity Forest is more accurate than the other algorithms from 500 training instances upwards}. This is not surprising, as trees usually have a better control over variance than NN algorithms, because of their higher bias and abstraction capabilities. Proximity Forest thus appears to be \modif{both accurate and highly scalable. We will show in the next subsection that this result holds also on the benchmark UCR archive.}
\subsection{Experiments on the UCR Archive}
In this section, we study the behavior of Proximity Forest on the 85 datasets of the traditional UCR archive \cite{UCR}.
It is useful to remember here that our aim is not to show that Proximity Forest is more accurate than the state-of-the-art, but only that it is competitive while being highly scalable.
We compare the mean error-rate of Proximity Forest to the error-rates on the standard train/test split for the state-of-the-art, as tested in \cite{bagnall2017great}. \modif{We average Proximity Forest results over 10 runs for each experiment.}
We compare Proximity Forest to five classifiers currently representing the state-of-the-art -- DTW-R, COTE, EE, ST and BOSS\footnote{It should be highlighted that the results presented here are for the original BOSS algorithm, and not the BOSS-VS discussed above in the SITS experiments. BOSS-VS is a scalable variation of BOSS, where concessions are made to accuracy in favor of training time. The original BOSS is therefore more competitive in this section.}.
\modif{The Proximity Forest results are obtained for 100 trees with selection between 5 candidates per node. A detailed discussion about the Proximity Forest parameters will be performed in Section \ref{subsec:pf_variations}.}
We first show the comparison with Proximity Forest's closest relative, EE. Figure \ref{fig:acc_EE_PF_UCR}%
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=.48\linewidth]{Fig5a}}
\subfigure[]{\includegraphics[width=.48\linewidth]{Fig5b}}
\caption{Comparison of Proximity Forest and Elastic Ensemble classifiers on UCR datasets in terms of a) accuracy and b) training and testing times in log scale.}
\label{fig:acc_EE_PF_UCR}
\end{figure}
provides scatter plots of the relative accuracy, total training time and total testing time of each of these classifiers. Each point represents a different UCR dataset.
\modif
Figure~\ref{fig:acc_EE_PF_UCR}(a) shows that Proximity Forest is more accurate on
\modif{60} datasets and less accurate on
only \modif{11} datasets, with \modif{14} ties. Moreover, for many datasets Proximity Forest is substantially more accurate than EE.}
Figure~\ref{fig:acc_EE_PF_UCR}(b) demonstrates that Proximity Forest has several orders of magnitude advantage in training time. When considering testing time, Proximity Forest has greater test time per query than EE for
\modif{12} datasets, the majority of which are small datasets (\textit{i.e.} less than \modif{50} training instances). The largest such difference is observed for the Phonemes dataset for which Proximity Forest takes about
\modif{17 seconds} per query compared to 13 seconds per query for EE. In contrast, the test time for Proximity Forest is much smaller than EE for the biggest datasets (\textit{i.e.} more than 800 training instances). For example, the biggest test time difference is for the HandOutlines dataset for which Proximity Forest takes about
\modif{19 seconds} per query compared to 286 seconds per query for EE.
The commonly accepted method to compare multiple classifiers over multiple datasets is by average ranks. For each dataset, we rank the classifiers and then calculate the average of each classifier's ranks across all datasets. When comparing 6 algorithms over 85 datasets, \cite{demvsar2006statistical} shows that for the rankings to be significantly different at level $\alpha=0.05$, the critical difference (CD) between the average ranks has to be greater than:
\begin{equation}
CD = q_{0.05}(A)\cdot \sqrt{\frac{A(A+1)}{6\cdot N_d}} = 2.850\cdot\sqrt{\frac{42}{510}} \approx 0.82
\end{equation}
The average ranks and critical difference are presented in Figure~\ref{fig:CriticalDiffDiagram}; the critical difference of 0.82 is displayed by the black line.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{Fig6}
\caption{Critical difference diagram for five state-of-the-art classifiers and Proximity Forest (PF) with 5 candidates.}
\label{fig:CriticalDiffDiagram}
\end{figure}%
\modif{It can be seen that COTE ranks highest (average rank of
2.28), which is to be expected considering it incorporates the other state-of-the-art algorithms. However, COTE is not ranked significantly higher than Proximity Forest (average rank of 2.88) or ST (average rank of 3.08).
Proximity Forest is ranked second. Its rank is not significantly different to COTE, ST or BOSS, but it is ranked significantly higher than both EE and DTW.} This affirms Proximity Forest as a classifier with accuracy competitive with the state-of-the-art.
Proximity Forest is the most accurate classifier for 22 of the 85 datasets. However, there is no obvious commonality between these datasets to suggest conditions under which the algorithm is likely to excel. \modif{The detailed accuracy results for Proximity Forest and the five state-of-the-art algorithms are shown in Appendix~\ref{ucr_results}.}
\modif{\subsection{Parameters of Proximity Forest}}
\label{subsec:pf_variations}
\modif{Proximity Forest has two main parameters that merit further investigation. We first explore the sensitivity of accuracy to the number of trees in each ensemble. Then, we discuss the influence of the number of candidates assessed at each node.
A third design choice, random selection of similarity measure per tree as opposed to per node, is explored in Appendix~\ref{ann:on_tree}.}
\subsubsection{On the choice of the number of trees}
The number of trees is the \modif{first} parameter of the Proximity Forest algorithm with the optimal value being large enough to provide competitive accuracy, yet small enough not to create excessive computational expense. The UCR datasets experiments outlined above were repeated with values of 5, 10, 50, and 100 trees to analyse how many trees were required to meet our needs. \modif{Here, the number of candidates $r$ has been fixed to 1. The Proximity Forest results are averaged over 50 runs.}
Figure~\ref{fig:CD-n-Trees} presents the critical difference diagram for accuracy and different number of trees. As expected, the more trees the higher the average accuracy: models with 100 trees had an average rank of
1.19 compared to
1.93, 2.98 and 3.89 for models with 50, 10, and 5 trees respectively. The difference between the highest ranked models
are large enough to say that models with 100 trees are significantly better than models with 50 trees at the level of alpha equals 0.05.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{Fig7}
\caption{Critical difference diagram for Proximity Forest with 5, 10, 50 or 100 trees.}
\label{fig:CD-n-Trees}
\end{figure}
Figure~\ref{fig:NoTreesPlot} compares the classification accuracy for 100 trees against 10 and 50 trees by representing them as a ratio of their error rates. Each point represents a single dataset. This shows that having 100 trees is better on most datasets. Moreover, the fact that the data is gathered close to the line with equation $x=1$ shows that it is unlikely that more trees would provide a very significant improvement, because the ratio of error-rates between 100 and 50 is already close to 1 (ie the errors are only slightly reduced).
We have not experimented with forests comprising more than 100 trees as we felt the computational demands outweighed the expected benefits for our large set of experiments. Memory, training time and testing time all scale linearly with the number of trees, which means that
doubling the number of trees doubles the required memory and time.
However,
where computational resources are not an issue, the take home message is that the more trees the better.
\begin{figure}
\centering
\includegraphics[width=.6\linewidth]{Fig8}
\caption{Ratio of the error rates of Proximity Forest models: 100 trees over 10 trees (x-axis) against 100 trees over 50 trees (y-axis). A value of less than 1 on either axis indicates that the model with 100 trees has higher accuracy.}
\label{fig:NoTreesPlot}
\end{figure}
As a randomized algorithm, it is finally interesting to study the standard deviation of the errors for Proximity Forest and how it varies with the number of trees. This is what we present in Figure~\ref{fig:stddev} where the y-axis represents the standard deviation on error-rate for 100 trees as a function of the standard deviation on $k$ equals to 5, 10, and 50 trees. Each point represents a single dataset. One can see that the standard deviation reduces as we increase the number of trees, and that the magnitude of this improvement reduces when increasing $k$. Results for 50 trees are starting to be relatively close to the $y=x$ line, showing that only marginal improvements could be expected when going to $k>100$.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{Fig9}
\caption{Standard deviations $\sigma$ of error rates on the 85 datasets of the UCR archive for Proximity Forest models: 100 trees against 50, 10 and 5 trees.}
\label{fig:stddev}
\end{figure}
\modif{\subsubsection{Split selection using the Gini index}
\label{subsubsec:candidates_eval}
This section explores the influence of the number of candidates $r$ that are randomly selected at each node. As a reminder, a set of $r$ candidates -- exemplars and parametrized distance measures -- are evaluated at each node based on the Gini index. The one maximizing the Gini index is retained. To evaluate the influence of $r$, the UCR experiments were repeated for 1, 2, and 5 candidates on 100 trees. The results are averaged over 10 runs.
Figure~\ref{fig:gini_influence} compares the classification accuracy for 5 candidates against 1 and 2 candidates. Each point represents the ratio of the error for 5 candidates to that for the alternative on an UCR dataset. %
\begin{figure}
\centering
\includegraphics[width=.6\linewidth]{Fig10}
\caption{Ratio of the error rates of Proximity Forest models: 5 candidates over 1 candidate (x-axis) against 5 candidates over 2 candidates (y-axis). A value of less than 1 on either axis suggests that the model with 5 candidates has superior accuracy}
\label{fig:gini_influence}
\end{figure}
Choosing between 5 candidates results in higher accuracy for most datasets. More precisely, selecting between 5 candidates results in greater accuracy than either 1 or 2 candidates on 61 datasets. Increasing the number of candidates lead to a reduction of the randomness on each node by discarding the worse splitters. Accordingly, the overall Proximity Forest accuracies are improved.
}
\modif{Increasing the number of candidates to more than 5 may further improve the classification accuracy. However, increasing the number of candidates per node has substantial impact on training time. Indeed, the analysis of the Proximity Forest's computational complexity in section~\ref{subsec:computation_analysis} shows that the training time scales linearly with the number of candidates. To verify this analysis, we compare both training and testing time of Proximity Forest for 1 and 5 candidates in Figure~\ref{fig:PF_c1_c5_time}. The testing time is displayed per query. Each point represents a dataset.}
\begin{figure}
\centering
\includegraphics[width=.65\linewidth]{Fig11}
\caption{Training and testing time of Proximity Forest for 1 and 5 candidates on UCR datasets.}
\label{fig:PF_c1_c5_time}
\end{figure}%
\modif{The results show a mean increase of 4.6 times in training time between 1~and 5 candidates, and a mean decrease of 0.93 times in testing time.
It is notable that selection between multiple alternatives both reduces testing time and increases the training time by slightly less than the expected multiple of 5 times. This is because it results in slightly shallower trees. Selection of better splits better separates the classes, requiring fewer splits to obtain pure nodes that are made into leaves.}
\modif{The tuning of the number of candidates is therefore driven by a trade-off between accuracy and time.}
\section{Conclusion}
\label{sec:Conclusion}
We introduced Proximity Forest: a novel, scalable algorithm for accurate time series classification. Motivated by a need for an algorithm that could learn from millions of time series, Proximity Forest is an ensemble of trees with a novel splitting criterion that makes it possible to make the most of decades of work in designing time series measures. In our case study, we demonstrated that Proximity Forest
scales quasi-linearly with the quantity of training data, whereas most state-of-the-art algorithms scale quadratically.
\modiff{Our experiments on the UCR datasets
show that Proximity Forest is not only very fast. It also has highly competitive accuracy relative to the current state-of-the-art, and is significantly more accurate than EE.}
We believe that there are a number of improvements that can be explored to increase the accuracy of Proximity Forest while maintaining its quasi-linear complexity\modif{, such as improving the randomized selection of parameters for the distance measures -- the current strategy was designed primarily to emulate EE as directly as possible}. We would also like to investigate to what extent this novel algorithm might shed new light on the task of time series indexing.
\section*{Supplementary material}
To ensure reproducibility, we make available the results of the experiments as well as our source code at \url{https://github.com/fpetitjean/ProximityForest/}{}.
\section*{Acknowledgements}
This research was supported by the Australian Research Council under grant DE170100037.
This material is based upon work supported by the Air Force Office of Scientific Research, Asian Office of Aerospace Research and Development (AOARD) under award number FA2386-17-1-4036. \modiff{We are grateful to the editor and anonymous reviewers whose suggestions and comments have greatly strengthened the paper.}
|
\section{Introduction}
A major open problem in the study of the traveling salesman problem is to determine the integrality ratio of
the standard LP relaxation. The same question can be asked for the variant in which start and end of the tour are
given and distinct.
For this LP (see \eqref{eq:subtour_lp_with_degree} below) the conjectured integrality ratio is $\frac{3}{2}$,
which is asymptotically attained by simple examples.
Better and better upper bounds have been shown \cite{AnKS15, Seb13, Vyg16, GotV16, SebvZ16}.
The previously best-known upper bound by Seb\H{o} and van Zuylen \cite{SebvZ16} is $\frac{3}{2}+\frac{1}{34}>1.5294$.
We improve the analysis of their algorithm and show that the integrality ratio is smaller than $1.5284$.
Even better approximation algorithms, with ratios $\frac{3}{2}+\epsilon$ \cite{TraV18} and $\frac{3}{2}$ \cite{Zen18}, have been found recently,
but these do not imply an upper bound on the integrality ratio of the LP.
In the $s$-$t$-path TSP we are given a finite metric space $(V,c)$ and vertices $s,t\in V$ with $s\ne t$.
The task is to compute a path $(V,H)$ with endpoints $s$ and $t$
that contains all elements of $V$ and minimizes $c(H) \coloneqq \sum_{\{v,w\}\in H} c(v,w)$.
An equivalent formulation asks for a minimum-cost multi-subset $H$ of $E= {V \choose 2}$ such that the graph
$(V,H)$ is connected and $s$ and $t$ are its only odd-degree vertices.
(Such a graph contains an Eulerian $s$-$t$-walk and we can shortcut whenever we revisit a vertex.)
Christofides' algorithm, adapted to the $s$-$t$-path TSP by Hoogeveen \cite{Hoo91}, computes a minimum-cost spanning tree $(V,S)$
and adds a minimum-cost matching on the set $T =\{v \in V : |S\cap \delta(v)| \text{ odd }\}\triangle \{s,t\}$
of vertices whose degree has the wrong parity.
Adding such a matching is called parity correction.
This algorithm yields a tour of cost at most $\frac{5}{3}$ times the LP value.
\section{Best-of-many Christofides with lonely edge deletion}
An, Kleinberg and Shmoys \cite{AnKS15} proposed and analyzed the best-of-many Christofides algorithm.
It starts by computing an optimum solution $x^*$ to the LP
\begin{equation} \label{eq:subtour_lp_with_degree}
\begin{aligned}
&\min c(x) \\
&\text{s.t.} & x(\delta(U)) &\geq 2 & & \text{for } \emptyset \subset U \subseteq V\setminus\{s,t\},\\
& & x(\delta(U)) &\geq 1 & & \text{for } \{s\} \subseteq U \subseteq V\setminus\{t\}, \\
& & x(\delta(v)) &= 2 & & \text{for } v\in V\setminus\{s,t\}, \\
& & x(\delta(v)) &= 1 & & \text{for } v\in \{s,t\}, \\
& & x(e) &\geq 0 & & \text{for } e\in E.
\end{aligned}\hspace{-4mm}
\end{equation}
Here and in the following we abbreviate $x(F) = \sum_{e\in F} x(e)$ and $c(x) \coloneqq \sum_{e\in E} c(e)x(e)$.
We write $x^*$ as a convex combination of incidence vectors of spanning trees, i.e.
$x^* = \sum_{j=1}^k p_{j} \chi^{S_j}$ for spanning trees $(V,S_1), \dots, (V,S_k)$ and nonnegative coefficients
$p_{1}, \dots, p_{k}$ with $\sum _{j=1}^k p_{j} =1$.
Then parity correction as in Christofides' algorithm is applied to each of the $k$ spanning trees; finally
the best of the resulting $s$-$t$-tours is selected.
A key observation of \cite{AnKS15}, used in all subsequent works, was that the set
$\mathcal{N} \coloneqq \{ \delta(U) : \{s\} \subseteq U \subseteq V\setminus\{t\}, x^*(\delta(U)) < 2\}$
of \emph{narrow} is induced by a chain.
The analysis of the best-of-many Christofides algorithm was improved by Seb\H{o} \cite{Seb13}.
The algorithm can be further improved by using a convex combination with certain properties \cite{Vyg16,GotV16}.
In particular, Gottschalk and Vygen \cite{GotV16} showed:
\begin{theorem}\label{thm:gao_tree_decomposition}
Let $x^*$ be an optimum solution to the LP \eqref{eq:subtour_lp_with_degree} and $\mathcal{N}$ the set of narrow cuts.
Then there exist
spanning trees $(V,S_1), \dots, (V,S_k)$ and nonnegative coefficients $p_{1}, \dots, p_{k}$ with $\sum _{j=1}^k p_{j} =1$
such that $x^* = \sum_{j=1}^k p_{j} \chi^{S_j}$ and for every $C\in\mathcal{N}$ there exists an $r\in\{1,\ldots,k\}$ with
$\sum_{j=1}^r p_{j} = 2-x^*(C)$ and $|C\cap S_j|=1$ for all $j=1,\ldots,r$.
\end{theorem}
Schalekamp et al.\ \cite{SchXX18} found a simpler proof of this theorem.
We will work with such a convex combination henceforth.
Seb\H{o} and van Zuylen \cite{SebvZ16} had the brilliant idea to delete some of the edges in each spanning tree and do parity correction
on the resulting forest.
This can save cost because parity correction will often reconnect the connected components of the forest anyway.
Call an edge $e$ and a cut $C\in\mathcal{N}$ \emph{lonely} in tree $S_j$
if $\{e\}=C\cap S_j$ and $\sum_{i=1}^j p_{i} \le 2-x^*(C)$.
Then we also say that $e$ is lonely at $C$.
We denote the lonely cuts in $S_j$ by $\mathcal{L}(S_j)$.
Let $F_j$ be the edge set of the forest that results from $S_j$ by deleting its lonely edges.
The algorithm by Seb\H{o} and van Zuylen \cite{SebvZ16} does parity correction on each forest $(V,F_j)$.
Let $T_j\coloneqq\{v\in V:|F_j\cap \delta(v)| \text{ odd}\}\triangle\{s,t\}$
denote the set of vertices whose degree in $F_j$ has the wrong parity.
Instead of adding a matching on $T_j$, we can add an arbitrary $T_j$-join $J$ (an edge set such that $T_j$ is the set odd-degree vertices
of $(V,J)$). Although this is equivalent, it will come handy.
Every $T_j$-join $J$ must contain an edge (in fact, an odd number of edges) in every lonely cut of $S_j$ (because they are all $T_j$-cuts,
i.e.\ cuts $\delta(U)$ for a vertex set $U$ with $|U\cap T_j|$ odd).
However, this does not imply that $F_j\stackrel{.}{\cup} J$ is connected, because an edge of $J$ can belong to several lonely cuts of $S_j$.
In this case we can, for all but one of these cuts, add two copies of the lonely edge of $S_j$ in this cut
(to ensure connectivity without changing parities).
If we choose a $T_j$-join $J$ for parity correction, we will pay a total of at most $\sum_{e\in J} c^j(e)$, where
$$c^j(e) \ \coloneqq \ c(e) \ + \sum_{C\in\mathcal{L}(S_j),\, e\in C} 2c(C\cap S_j) \ - \ \max\left\{0,\, \max_{C\in\mathcal{L}(S_j),\, e\in C} 2c(C\cap S_j) \right\};$$
here the second and third terms account for the reconnection cost.
We now describe formally the best-of-many Christofides algorithm with lonely edge deletion due to Seb\H{o} and van Zuylen.
This is the algorithm that we will analyze.
\begin{enumerate}
\item Compute an optimum solution $x^*$ to the LP \eqref{eq:subtour_lp_with_degree}.
\item Compute $x^* = \sum_{j=1}^k p_{j} \chi^{S_j}$ as in Theorem \ref{thm:gao_tree_decomposition}
\item Do the following for each $j=1,\ldots,k$:
\begin{enumerate}
\item Compute a $T_j$-join $J_j$ with minimum $c^j$-cost.
\item Compute a minimum $c$-cost subset $R_j \subseteq S_j \setminus F_j$
of the lonely edges such that $F_j \stackrel{.}{\cup} J_j \stackrel{.}{\cup} R_j$ is connected.
\item Find an Eulerian $s$-$t$-walk in $H_j \coloneqq F_j \stackrel{.}{\cup} J_j \stackrel{.}{\cup} R_j\stackrel{.}{\cup} R_j$ and shortcut whenever a vertex is visited more than once.
\end{enumerate}
\item Return the cheapest of these $k$ tours.
\end{enumerate}
We remark that Seb\H{o} and van Zuylen \cite{SebvZ16}
also consider the result of the normal best-of-many Christofides algorithm and output the better of the solutions,
but this is not necessary as our analysis will reveal.
\section{Outline of the new analysis}
By definition of $c^j$, the cost of the tour $H_j$ is at most $c(F_j) + c^j(J_j)$.
The cost of the $T_j$-join $J_j$ is the minimum cost of a vector $y$ in the $T_j$-join polyhedron \cite{EdmJ73}
\begin{equation}\label{eq:Tjoinpolyhedron}
\bigl\{y\in\mathbb{R}^E_{\ge 0} : y(\delta(U))\ge 1 \text{ for $ U\subset V \text{ with } |U\cap T_j|$ odd} \bigr\}.
\end{equation}
We call a vector $y$ in \eqref{eq:Tjoinpolyhedron} a parity correction vector.
Note that every parity correction vector yields an upper bound on the cost of $J_j$.
A first attempt to design a parity correction vector could be the vector $\beta x^* + (1-2\beta) \chi^{S_j}$ for some $0\le \beta \le \frac{1}{2}$.
This vector has value at least one on all cuts except the narrow cuts.
The narrow cuts can be repaired by adding fractions of incidence vectors of lonely edges (not necessarily from the same tree).
We will pay all this and the reconnection cost by what we gain by deleting the lonely edges.
Then our total cost is
\begin{equation}\label{eq:uniform_weights}
\min_{j=1}^k c(H_j) \le \sum_{j=1}^k p_j c(H_j) \le \sum_{j=1}^k p_j \bigl(c(S_j) + \beta c(x^*) + (1-2\beta) c(S_j)\bigr) = (2-\beta) c(x^*).
\end{equation}
Hence we would like to choose $\beta$ as large as possible.
Unfortunately, for $\beta = \frac{1}{2}$ we need too much from the lonely edges.
By reducing $\beta$, we can increase the value of $\beta x^* + (1-2\beta) \chi^{S_j}$ on the narrow cuts and thus decrease the
required amount of lonely edges.
Choosing $\beta = \frac{8}{17}$ is sufficient and this is essentially what Seb\H{o} and van Zuylen did.
In our parity correction vector for a forest $F_j$ we will use lonely edges of $S_j$ and of earlier trees.
If we increase $\beta$ for the early trees and decrease $\beta$ for the late trees, we need more from the lonely edges
in the early trees, but less in the late trees.
This will improve our bound if the late trees are cheaper (and this is indeed true in the worst case).
The algorithm computes $k$ tours $H_1,\ldots,H_k$.
All previous analyses, like \eqref{eq:uniform_weights}, computed an upper bound on $\sum_{j=1}^k p_j c(H_j)$.
Instead, we will compute a weighted average with different weights, giving a higher weight to tours resulting from early trees.
We choose $\beta_j$ and weights $q_j >0$ with $\sum_{j} q_j = 1$ and such that $q_j \cdot (2-2\beta_j) = M \cdot p_j$ for some constant $M>0$.
Such a choice allows to bound the cost of our tour against the LP value $c(x^*)$:
\[\sum_{j=1}^{k} q_j \cdot c(H_j) \le \sum_{j=1}^{k} q_j \left(\beta_j c(x^*)
+ (2-2\beta_j) c(S_j)\right) = \left(\sum_{j=1}^{k} q_j \beta_j + M\right) c(x^*). \]
Intuitively, choosing $\frac{q_j}{p_j}$ (and thus $\beta_j$) larger for the early trees is good, because for the early trees
we delete more lonely edges (cf.\ Theorem \ref{thm:gao_tree_decomposition}).
This allows us to choose the average value of $\beta$ larger and thus improves our upper bound.
We first analyze the cost of a tour resulting from a single tree $S_j$.
Later, we will take a weighted average.
\section{Analyzing one tree}
Let $j\in\{1,\ldots,k\}$.
To bound the cost of parity correction of the forest $F_j$, we follow Wolsey's approach \cite{Wol80} and
use a vector in the $T_j$-join polyhedron \eqref{eq:Tjoinpolyhedron}.
Let $0 \le \beta \le \frac{1}{2}$ and $\alpha \coloneqq 1- 2\beta \ge 0$.
Moreover, for $C\in\mathcal{N}$, let $v^C\in\mathbb{R}_{\ge 0}^{E}$ be a vector with $v^C(C) = 1$ and
$v^C(e) = 0$ unless $e$ is lonely at $C$ in some tree (not necesarily in $S_j$).
We will choose $v^C$ later.
We define
\begin{equation*}
y^j_{\beta} \ \coloneqq \beta x^* + \alpha \chi^{S_j}
+\! \displaystyle\sum_{C\in\mathcal{L}(S_j)} \beta (2-x^*(C)) \chi^{S_j\cap C}
+\! \displaystyle\sum_{C\in\mathcal{N}\setminus\mathcal{L}(S_j)}\!\! \max\left\{0,\, \beta (2-x^*(C)) - \alpha \big. \right\} v^C.
\end{equation*}
The first sum is the contribution from lonely edges of the tree $S_j$ itself, in oder to repair the lonely cuts of $S_j$.
The second sum is the contribution from lonely edges of earlier trees, in order to repair the other narrow cuts.
Obviously, $y^j_{\beta}$ is a nonnegative vector.
We show that $y^j_{\beta}$ is a parity correction vector, i.e.\ a vector in \eqref{eq:Tjoinpolyhedron}.
\begin{lemma}
For every $T_j$-cut $C$ we have $y^j_{\beta}(C)\ge 1$.
\end{lemma}
\noindent \hbox{\textbf{Proof:}}\quad
Let $C=\delta(U)$ be a $T_j$-cut.
Since $|\{ v\in U : |F_j\cap\delta(v)| \text{ odd}\}|$ is odd if and only if
$|F_j\cap \delta(U)|$ is odd, we conclude that
$|U\cap \{s,t\}|+ |F_j\cap C|$ is odd.
We now distinguish several cases.
\\[2mm]
\underline{Case 1:} $|U\cap \{s,t\}|$ is odd (i.e., $C$ is an $s$-$t$-cut). \\[1mm]
Then $|F_j\cap C|$ is even. We now consider two subcases.
\\[1mm]
\underline{Case 1a:} $C\in\mathcal{L}(S_j)$. \\[1mm]
Then $y^j_{\beta}(C)\ge \beta x^*(C)+\alpha+ \beta(2-x^*(C)) = \alpha+2\beta = 1$.
\\[1mm]
\underline{Case 1b:} $C\notin\mathcal{L}(S_j)$. \\[1mm]
Since $|F_j\cap C|$ is even, we have $|S_j\cap C| \ge |F_j\cap C| \ge 2$ or $|F_j\cap C| = 0$.
Since $C\notin\mathcal{L}(S_j)$, if $F_j\cap C$ is empty, the cut $C$ must contain at least two edges that are lonely in $S_j$.
So we have also in this case $|S_j\cap C| \ge 2$.
Thus $y^j_{\beta}(C)\ge \beta x^*(C) +2\alpha + \max\{0,\beta(2-x^*(C))-\alpha\}$
(note that the last term is zero if $C\notin\mathcal{N}$).
We conclude
$y^j_{\beta}(C)\ge \beta x^*(C) +2\alpha + \beta(2-x^*(C))-\alpha
= \alpha+2\beta = 1$.
\\[2mm]
\underline{Case 2:} $|U\cap \{s,t\}|$ is even. \\[1mm]
Then $x^*(C)\ge 2$.
Hence, $y^j_{\beta}(C)\ge \beta x^*(C)+ \alpha \ge 2 \beta + \alpha = 1$.
\hspace*{\fill} {\boldmath $\Box$} \par \vskip0.5em
Moreover,we have $y^j_{\beta} \ge 0$. Thus $y^j_{\beta}$ is contained in the $T_j$-join polyhedron \eqref{eq:Tjoinpolyhedron}, and so
$\min \{c^j(J) : J\text{ a } T_j\text{-join} \} \le c^j(y^j_{\beta})$.
A key observation of Seb\H{o} and van Zuylen \cite{SebvZ16} was that the need for reconnection is unlikely.
Only bad edges can result in reconnection, where an edge is called \emph{bad} (for $S_j$) if it is contained in more than one lonely cut.
The edges in $S_j$ are never bad for $S_j$, nor are the lonely edges of trees that come earlier in the list $S_1, \dots, S_r$.
Therefore, an edge $e$ with $v^C(e) > 0$ for some $C\in \mathcal{N}\setminus\mathcal{L}(S_j)$ is not bad for $S_j$.
At this point one uses the particular choice of the decomposition of $x^*$ into incidence vectors of spanning trees.
For every edge $e$ that is not bad we have $c^j(e) =c(e)$.
Hence
\begin{eqnarray*}
c^j(y^j_{\beta})
&=& \beta \, c^j(x^*) \ +\ \alpha \, c(S_j) + \sum_{C\in\mathcal{L}(S_j)} \beta(2-x^*(C)) \, c(S_j \cap C) \\
&& + \sum_{C\in\mathcal{N}\setminus\mathcal{L}(S_j)} \max\left\{0,\, \beta(2-x^*(C)) - \alpha \big. \right\} c(v^C).
\end{eqnarray*}
\bigskip
Moreover, Seb\H{o} and van Zuylen \cite{SebvZ16} showed:
\bigskip
\begin{lemma}
$$c^j(x^*) \ \le \ c(x^*) + \sum_{C\in\mathcal{L}(S_j)} 2(x^*(C)-1) c(S_j\cap C).$$
\end{lemma}
Therefore, the cost of the tour that results from the tree $S_j$ is at most
\begin{equation}\label{eq:cost_single_tour}
\begin{aligned}
&c(F_{j}) + \min \{c^{j}(J) : J\text{ a } T_{j}\text{-join} \} \\
\le \ &c(S_{j}) - \!\sum_{C\in\mathcal{L}(S_{j})} c(S_{j}\cap C) + c^{j}(y^j_{\beta}) \\
\le \ &(1+\alpha)c(S_j) + \beta c(x^*) + \sum_{C\in\mathcal{L}(S_{j})}\bigl(2\beta(x^*(C)-1)- 1 + \beta(2-x^*(C)) \bigr) \, c(S_{j}\cap C)\\
& + \sum_{C\in\mathcal{N}\setminus\mathcal{L}(S_j)} \max\left\{0,\, \beta (2-x^*(C)) - \alpha \big. \right\} c(v^C) \\
=\ &(1+\alpha)c(S_j) + \beta c(x^*) - \sum_{C\in\mathcal{L}(S_{j})}\bigl(\alpha+ \beta (2-x^*(C)) \bigr) \, c(S_{j}\cap C)\\
& + \sum_{C\in\mathcal{N}\setminus\mathcal{L}(S_j)} \max\left\{0,\, \beta(2-x^*(C)) - \alpha \big. \right\} c(v^C),
\end{aligned}
\end{equation}
since $\alpha = 1 - 2\beta$.
\section{Average cost}
It will be useful to index the trees by a continuum and define $S_{\sigma}$ for all $0<\sigma\le 1$, where
$S_{\sigma}=S_j$ if $\sum_{i=1}^{j-1} p_i <\sigma \le \sum_{i=1}^j p_i$.
Let $h\colon [0,1]\to[0,1]$ be an integrable function to be chosen later.
The weight of the tour resulting from $S_{\sigma}$ will be proportional to $1+h(\sigma)$.
Also $\alpha$ and $\beta$ depend on $\sigma$, namely as follows:
$$\alpha_{\sigma} = \frac{1-h(\sigma)}{1+h(\sigma)} \qquad \text{ and } \qquad \beta_{\sigma} = \frac{h(\sigma)}{1+h(\sigma)}.$$
Note that indeed $0\le \beta_{\sigma}\le\frac{1}{2}$ and $\alpha_{\sigma}+2\beta_{\sigma}=1$ for all $\sigma$.
Moreover, we set
\[ v^C \ \coloneqq \ \frac{1}{\int_0^z \bigl( 1 - h(\sigma) + z h(\sigma) \bigr) \,\text{d}\sigma} \,
\int_{0}^{z} \bigl( 1 - h(\sigma) + z h(\sigma) \bigr) \cdot \chi^{S_\sigma\cap C} \,\text{d}\sigma, \]
where we abbreviated $z\coloneqq 2-x^*(C)$.
Then indeed $v^C(C) = 1$ for all $C\in\mathcal{N}$, and $v^C(e) = 0$ unless $e$ is lonely at $C$.
We will now show under which condition the last two terms in \eqref{eq:cost_single_tour} vanish:
\begin{lemma}\label{lemma:lonely_edge_usage}
Suppose
\begin{equation}
\label{eq:requirementforh}
\int_z^1 \max \bigl\{0, h(\sigma)-1+zh(\sigma) \bigr\} \,\text{d}\sigma + \int_0^z \bigl( h(\sigma)-1-zh(\sigma) \bigr) \,\text{d}\sigma \ \le \ 0
\end{equation}
for all $z\in[0,1]$.
Then
\begin{equation}
\label{eq:vanish}
\begin{aligned}
\int_0^1 (1+h(\sigma)) \Biggl( \,
& - \sum_{C\in \mathcal{L}(S_{\sigma})} \!
\bigl(
\alpha_{\sigma} + \beta_{\sigma} (2-x^*(C)) \bigr)\, c(S_{\sigma}\cap C) \\
+&\!\! \sum_{C\in\mathcal{N}\setminus\mathcal{L}(S_j)} \!\!
\max\left\{0,\, \beta_{\sigma} (2-x^*(C)) - \alpha_{\sigma} \big. \right\} c(v^C)
\Biggr) \,\textnormal{d}\sigma
\end{aligned}
\end{equation}
is nonpositive.
\end{lemma}
\noindent \hbox{\textbf{Proof:}}\quad
Again writing $z\coloneqq 2-x^*(C)$, using
$$ (1+h(\sigma)) \bigl( \alpha_{\sigma} + \beta_{\sigma} (2- x^*(C)) \bigr) \ = \
1 - h(\sigma) + z h(\sigma) $$
and
$$ (1+h(\sigma)) \max\left\{0,\, \beta_{\sigma} (2-x^*(C)) - \alpha_{\sigma} \big. \right\}
\ = \ \max\left\{0,\ h(\sigma) - 1 + z h(\sigma) \big. \right\},$$
and changing the order of summation,
we can rewrite \eqref{eq:vanish} as
\begin{align*}
- \sum_{C\in\mathcal{N}} \int_0^z \bigl( 1 - h(\sigma) + z h(\sigma) \bigr)\, c(S_{\sigma}\cap C) \,\textnormal{d}\sigma
+ \sum_{C\in\mathcal{N}} \int_z^1 \max\left\{0,\, h(\sigma) - 1 + z h(\sigma) \big. \right\} c(v^C) \,\textnormal{d}\sigma.
\end{align*}
Hence (plugging in the definition of $v^C$) and using $1-h(\sigma) +z \cdot h(\sigma)>0$, it suffices to show that, for every $z\in(0,1]$,
$$-1 + \frac{1}{\int_0^z \bigl( 1 - h(\sigma) + z h(\sigma) \bigr)
\,\text{d}\sigma} \int_z^1 \max\left\{0,\, h(\sigma) - 1 + z h(\sigma) \big. \right\} d\sigma\ \le \ 0. $$
which follows directly from \eqref{eq:requirementforh}.
\hspace*{\fill} {\boldmath $\Box$} \par \vskip0.5em
\begin{lemma}
Let $h\colon [0,1]\to[0,1]$ be an integrable function with \eqref{eq:requirementforh} for all $z\in[0,1]$.
Then the best-of-many Christofides algorithm with lonely edge deletion computes a solution of cost
at most $\rho^* c(x^*)$, where
$$\rho^* \ = \ 1 + \frac{1}{1 + \int_0^1 h(\sigma) \,\textnormal{d}\sigma}.$$
\end{lemma}
\noindent \hbox{\textbf{Proof:}}\quad
Combining \eqref{eq:cost_single_tour} and Lemma \ref{lemma:lonely_edge_usage},
we get the following upper bound on the total cost of the best-of-many Christofides algorithm with lonely edge deletion:
\begin{align*}
& \sfrac{1}{\int_0^1 (1+h(\sigma)) \,\text{d}\sigma} \int_0^1 (1+h(\sigma))
\left( \Big. \beta_{\sigma} c(x^*) + (1+\alpha_{\sigma}) c(S_{\sigma}) \right) \,\text{d}\sigma \\
= \ & \sfrac{1}{\int_0^1 (1+h(\sigma)) \,\text{d}\sigma} \int_0^1
\left( \Big. h(\sigma) c(x^*) + 2 c(S_{\sigma}) \right) \,\text{d}\sigma \\
= \ & \sfrac{1}{\int_0^1 (1+h(\sigma)) \,\text{d}\sigma} \, \left( \int_0^1 h(\sigma) \,\text{d}\sigma + 2 \right) c(x^*) \\
= \ & \sfrac{1}{\int_0^1 (1+h(\sigma)) \,\text{d}\sigma} \, \left( \int_0^1 (1+h(\sigma)) \,\text{d}\sigma + 1 \right) c(x^*) \\
= \ & \left( 1 + \sfrac{1}{1 + \int_0^1 h(\sigma) \,\text{d}\sigma} \right) c(x^*)
\end{align*}
\hspace*{\fill} {\boldmath $\Box$} \par \vskip0.5em
Now we can prove the main result:
\begin{theorem}\label{theorem:upper_bound_rho}
Let
$$\rho^* \ \coloneqq \ 1 + \frac{1}{1 + 4 \ln(\frac{5}{4})}.$$
Then the best-of-many Christofides algorithm with lonely edge deletion computes a solution of cost
at most $\rho^* c(x^*)$.
\end{theorem}
\noindent \hbox{\textbf{Proof:}}\quad
We set $h(\sigma)=\frac{4}{4+\sigma}$ for $0\le\sigma\le 1$.
Then $\int_0^1 \sfrac{4}{4+\sigma} \,\text{d}\sigma = 4\ln(\frac{5}{4})$.
We need to check \eqref{eq:requirementforh}.
Note that $h(\sigma)-1+z h(\sigma) >0$ if and only if $\sfrac{4}{4+\sigma} = h(\sigma)> \frac{1}{1+z}$, i.e., $\sigma<4z$.
Hence to prove \eqref{eq:requirementforh} it suffices to show
$$\int_z^{4z} \bigl( h(\sigma)-1+zh(\sigma) \bigr) \,\text{d}\sigma + \int_0^z \bigl( h(\sigma)-1-zh(\sigma) \bigr) \,\text{d}\sigma \ \le \ 0$$
The left-hand side is
$$4(1+z) ( \ln (4z+4) - \ln(z+4) ) + 4(1-z) ( \ln (z+4) - \ln(4) ) -4z,$$
so (dividing by 4) we need to check
$$(1+z) \ln \sfrac{4z+4}{z+4} + (1-z) \ln \sfrac{z+4}{4} - z \ \le \ 0.$$
This is true for $z=0$, moreover the derivative of the left-hand side is
$$\ln \sfrac{16(z+1)}{(z+4)^2} -\sfrac{2z}{z+4}.$$
Using $\ln x \le x-1$ for all $x>0$ this is at most
$$ \sfrac{16(z+1)}{(z+4)^2} - 1 -\sfrac{2z}{z+4}
\ = \ \sfrac{16(z+1)-(z+4)^2 -2z(z+4)}{(z+4)^2}
\ = \ \sfrac{-3z^2}{(z+4)^2}
\ \le \ 0.$$
\hspace*{\fill} {\boldmath $\Box$} \par \vskip0.5em
Theorem \ref{theorem:upper_bound_rho} immediately implies that the integrality ratio is at most $\rho^*$.
Note that $\rho^* < 1.5284$.
We see that \eqref{eq:requirementforh} is tight only for $z=0$ with our choice of $h$. A better choice would lead to a better
upper bound on the integrality ratio.
However, we do not know how to find the best $h$.
Numerical computations indicate that the best value that can be obtained in this way is approximately $1.5273$.
\subsubsection*{Acknowledgment}
We thank the anonymous referees for their useful remarks that helped to improve the presentation.
Moreover, one referee suggested a Python-Gurobi script to compute
the best bound on the integrality ratio that can be obtained by our approach numerically.
|
\section*{Introduction}
The many applications of Beilinson-Bernsteins's localization theory to representation theory are well known. Classically, the focus lies on representations on ${\rm\bf C}$-vector spaces. More recently, arithmetic applications such as in \cite{harris2013,harder2014,zagier2010,januszewskirationality,januszewskiperiods1,januszewskiperiods2} have underlined the interest in considering more general bases than the complex numbers.
The present paper grew out of an attempt to understand Michael Harris' sketch of a Beilinson-Bernstein localization theory over ${\rm\bf Q}$ in \cite{harris2013}, which since then has been amended by \cite{harris2013erratum}. That such a theory exists over arbitrary algebraically closed field of characteristic $0$ has been known to Beilinson, Bernstein and others since a long time (cf.\ Beilinson 1983 ICM address \cite{beilinson1983}). However, a proper discussion of such a theory even over ${\rm\bf Q}$ is still lacking in the literature, and we attempt to fill this gap here.
Due to its geometric nature, the general algebro-geometric language of schemes provides the proper framework to formulate a general theory of $\mathcal D$-modules, which allows for even further generalizations. Consequently, we aim at the largest reasonable generality which still allows us to give essentially full proofs or refer to appropriate standard references. For large parts we work over an arbitrary noetherian or an arbitrary locally finitely presented base scheme $S.$
Harder constructed in \cite{harder2014} models for certain irreducible $({\mathfrak {g}},K)$-modules over ${\rm\bf Z}$. Subsequently Hayashi gave in \cite{hayashi2018} a general construction of models of irreducible $({\mathfrak {g}},K)$-modules over the integers. The general theory we develop here yields also such models and it remains an interesting problem to relate the different constructions.
While we were preparing this manuscript, Bernstein, Higson and Subag worked out an elementary example of a contraction family for forms of $\SL_2({\rm\bf C})$ over ${\rm\bf P}^1$ in \cite{bernsteinetal1,bernsteinetal2}. Tan, Yao and Yu studied in \cite{tanetal2016} deformations of $D$-modules in the case of $\SL(2,{\rm\bf R})$ without providing formal justifications for the formalism they apply.
We hope that our work helps to unify these different approaches. In the notion of families of Harish-Chandra pairs from Bernstein, Higson and Subag corresponds to modules over $S={\rm\bf A}^1$ and $S={\rm\bf P}^1$ respectively. However, we need not assume that the ground field is ${\rm\bf C}.$ With the theory we develop here, we may even work over the affine or projective line over ${\rm\bf Z}.$
We suppose a fluent understanding of modern scheme theoretic language. Throughout the paper we make frequent use of results from Grothendieck-Dieudonn\'e's EGA I--IV \cite{ega1,ega2,ega31,ega32,ega42,ega44} and occasionally to SGA \cite{sga1,sga3}. On the $\mathcal D$-module side we content us mostly with \cite{hottaetal2008}.
The outline of the paper is as follows. In the first section we recall and summarize basic statements on Sheaves of differential operators, connections, and $\mathcal D$-modules. For most statements, the only reference available seems to be EGA IV \cite{ega42,ega44}, where the results are scattered through a number of technical statements. The second section is devoted to the study of $\mathcal D$-modules, $\mathcal D$-affineness, different notions of coherence and existence of flat resolutions. In the third section we study functorial properties of $\mathcal D$-modules: We establish a flat base change theorem (Theorem \ref{thm:flatDmodulebasechange}) and show that the property of being $\mathcal D$-affine is local in the $fpqc$-topology (Theorem \ref{thm:Daffinefpqclocal}). Subsequently we prove the faithfully flat base change theorem for $\mathcal D$-modules (Theorem \ref{thm:fpqcDmoduledescent}). derived inverse and direct images and their relation to base change. In section 4 we discuss twisted sheaves of differential operators, group actions and equivariant sheaves. In section 5 we summarize our results in the special case where the base is a field, and which is sufficient for the applications in \cite{harris2013}.
{\em Acknowledgements.} The author thanks Leticia Barchini, Michael Harris and Takuma Hayashi for helpful remarks on a preliminary version of this manuscript.
\subsection*{Notation}
We work over an arbitrary base scheme $S.$ For us a scheme is simply a \lq{}pr\'esch\'ema\rq{} in \cite{ega1}. For applications we will specialize to $S=\Spec{\rm\bf Z}$, $S=\Spec{\rm\bf Q}$ or $S=\Spec{\rm\bf C}$, etc. We write $\mathcal O_S$ for the structure sheaf of $S$, and also $S$ for the underlying topological space.
For a morphism $s:S\to S'$ of schemes, we adopt the notation of \cite{ega1}, i.e.\ $s$ is a pair $(\psi,\theta)$, where $\psi:S\to S'$ is a continuous map of topological spaces and $\theta:\mathcal O_{S'}\to \psi_*\mathcal O_S$ is a morphism of sheaves of rings subject to the conditition that locally $(\psi,\theta)$ arises from classical maps between affine spaces (induced by ring homomorphisms). In particular, $s$ is a morphism of locally ringed spaces but not every such morphism is a morphism of schemes. We write $\theta^\#:\psi^*\mathcal O_{S'}\to\mathcal O_S$ for the morphism corresponding to $s$ under the standard adjunction.
While $\theta$ turns $\psi_*\mathcal O_S$ into an $\mathcal O_{S'}$-module, we remark that $\psi^*\mathcal O_{S'}$ is not an $\mathcal O_S$-module. Hence the pushforward $\psi_*\mathcal F$ of any $\mathcal O_S$-module $\mathcal F$ is automatically an $\mathcal O_{S'}$-module. However if we depart from an $\mathcal O_{S'}$-module $\mathcal G$, we need to consider $\mathcal O_S\otimes_{\psi^*\mathcal O_{S'}}\psi^*\mathcal G$ in order to obtain an $\mathcal O_S$-module.
To add to the confusion, we denote this $\mathcal O_S$-module by $s^*\mathcal G$, and also adopt the notation $s_*\mathcal G$ for $\psi_*\mathcal G$. It will be clear from the context when which notion of pull back is at work.
A scheme $X$ over $S$ is a morphism of schemes $x:X\to S$ and a morphism of schemes $x:X\to S$, $y:Y\to S$ over $S$ is a morphism of schemes $f:X\to Y$ with the property that the obvious triangle commutes (i.e.\ $x=y\circ f$).
The functor $\mathcal F\to x_* \mathcal F$ of sheaves on $X$ to sheaves on $S$ is left exact and its right derived functors $R^*x_*$ are the appropriate generalization for ordinary sheaf cohomology of $\mathcal F$ in the relative situation. We naturally obtain cohomology {\em sheaves} ($R^qx_*\mathcal F$ is a sheaf of $\mathcal O_S$-modules on $S$), and it makes conceptually sense to consider all categories of sheaves {\em enriched} over themselves, i.e.\ for sheaves $\mathcal F,\mathcal G$ on $X$ we consider the internal hom $\iHom(\mathcal F,\mathcal G)$ which is again a sheaf on $X$. Likewise, if $\mathcal F,\mathcal G$ are $\mathcal O_X$-modules, we consider $\iHom(\mathcal F,\mathcal G)$ as an $\mathcal O_X$-module via the $\mathcal O_X$-module structure on $\mathcal G$, and we also have the sheaf $\iHom_{\mathcal O_X}(\mathcal F,\mathcal G)$ of $\mathcal O_X$-module homomorphisms.
For $X$ a scheme and an $\mathcal O_X$-algebra $\mathcal A$, we write $\Mod(\mathcal A)$ for the category of left $\mathcal A$-modules on $X$. We let $\Mod_{\rm qc}(\D_{X/S})$ denote the full subcategory of $\mathcal O_X$-quasi-coherent $\mathcal A$-modules.
The following definition is borrowed from \cite{hottaetal2008}.
\begin{definition}
An $\mathcal A$-module $\mathcal M$ is called $\mathcal A$-{\em coherent} if $\mathcal M$ is a locally finitely generated $\mathcal A$-module and if for {\em every} open $U\subseteq X$, every locally finitely generated submodule of $\mathcal M|_U$ is a locally finitely presented $\mathcal A_U$-module.
\end{definition}
\section{Generalities on Sheaves of differential operators}
We begin with a review of sheaves of differential operators in the scheme theoretic language.
\subsection{Sheaves of differential operators}
We follow \cite[\textsection 16]{ega44}. For a scheme $x:X\to S$ over $S$ we consider the diagonal embedding
$$
\Delta_X:\quad X\to X\times_S X.
$$
Writing $\Delta_X=(\psi,\theta)$ we have a short exact sequence
$$
0\to\mathcal I_X\to\psi^*\mathcal O_{X\times_S X}\to\mathcal O_X\to 0
$$
of sheaves on $X$.
\subsection{Differentials}
We have the sheaf
$$
\Omega_{X/S}\;:=\;\mathcal I_X/\mathcal I_X^2,
$$
of (relative) $1$-differentials on $X$ (with respect to $S$). It has a natural $\mathcal O_X$-module structure induced by the inverse of the natural isomorphism
$$
\psi^*(\mathcal O_{X\times_S X})/\mathcal I_X\;\cong\;\mathcal O_X.
$$
As usual, $\Omega_{X/S}$ represents derivations, externally and internally:
\begin{equation}
\iHom_{\mathcal O_X}(\Omega_{X/S},\mathcal F)\;\cong\;\Der_S(\mathcal O_X,\mathcal F),
\label{eq:omegaxderivations}
\end{equation}
$$
u\;\mapsto\;u\circ d_{X/S},
$$
where $d_{X/S}$ is defined below (cf. Corollaire (16.5.5) in \cite{ega44}). Taking global sections gives the external representability.
The (relative) tangent sheaf of $X\to S$ is by definition
$$
{\Der}_S(\mathcal O_X,\mathcal O_X),
$$
canonically identified with the $\mathcal O_X$-dual of $\Omega_{X/S}$ via \eqref{eq:omegaxderivations}.
We set $\Omega_{X/S}^k:=\bigwedge^k\Omega_{X/S}$ and write $d:\Omega_{X/S}^k\to\Omega_{X/S}^{k+1}$ for the exterior differential, which extends the differential $d_{X/S}$ recalled below (cf.\ Th\'eor\`eme (16.6.2) in loc.\ cit.).
\subsection{Sheaves of principal parts}
Following D\'efinition (16.3.1) in \cite{ega44} we set for every $n\geq 0$
$$
\Pp_{X/S}^n\;:=\;\psi^*(\mathcal O_{X\times_S X})/\mathcal I_X^{n+1}.
$$
For $n>0$ this is not a sheaf of $\mathcal O_X$-modules in a canonical way (if $\Omega_{X/S}\neq0$, we need to make a choice, see below) but a sheaf of algebras {\em augmented} over $\mathcal O_X$. Then $\Pp_{X/S}^n$ is the {\em sheaf of principal parts of order $n$}. It comes with a natural increasing filtration induced by the quotients
$$
\mathcal I_X^{m}/\mathcal I_X^{1+n},
$$
for $0\leq m\leq n$. The associated graded sheaf $\Gr(\Pp^n_{X/S})$ admits the sheaf $\Omega_{X/S}^1$ of $1$-differentials as $\Gr_1(\Pp^n_{X/S})$. The sheaves of principal parts form a projective system of sheaves with projective limit denoted $\Pp_{X/S}^\infty$.
Now the projections $p_i:X\otimes_S X\to X$ induce two $\mathcal O_X$-module (and a fortiori two $\mathcal O_X$-algebra) structures on $\psi^*(\mathcal O_{X\times_S X})$ which descend to two different $\mathcal O_X$-module structures on $\Pp_{X/S}^n$. The standard $\mathcal O_X$-module structure is the one resulting from $p_1$. To distinguish the second module structure, we let
$$
d_{X/S}^n:\quad\mathcal O_X\to\Pp_{X/S}^n
$$
denote the $\mathcal O_X$-algebra structure resulting from $p_2$. Then, for any section $t\in\Gamma(U,\mathcal O_X)$, we have its {\em differential}
$$
dt\;:=\;d_{X/S}t\;:=\;d_{X/S}^1t - t\;\in\;\Gamma(U,\Omega_{X/S}).
$$
This notion generalizes to a map
$$
\delta^n:\quad \mathcal O_X(U)\to\Gamma(U,\Pp_{X/S}^n),
\quad t\;\mapsto\;d^nt-t.
$$
We remark that these maps are compatible with the canonical projections $\Pp_{X/S}^m\to\Pp_{X/S}^n$, $m\geq n$.
Finally, we have for any $\mathcal O_X$-module $\mathcal F$ the $\Pp_{X/S}^n$-module
$$
\Pp_{X/S}^n(\mathcal F)\;:=\;\Pp_{X/S}^n\otimes_{\mathcal O_X}\mathcal F.
$$
\subsection{Differential operators}\label{sec:differentialoperators}
For each pair of $\mathcal O_X$-modules $\mathcal F,\mathcal G$ and each $n\in{\rm\bf N}_{0}\cup\{\infty\}$, we have a morphism of sheaves
$$
\delta_{X/S}^n:\quad \iHom_{\mathcal O_X}(\Pp_{X/S}^n(\mathcal F),\mathcal G)\;\to\;\iHom(\mathcal F,\mathcal G),
$$
$$
u\;\mapsto\;u\circ d_{X/S}^n
$$
whose (sheaf theoretic image) we denote by $\Diff_{X/S}^n(\mathcal F,\mathcal G)$. We emphasize that in the codomain of $\delta_{X/S}^n$ we are considering additive morphisms of sheaves. Then $\Diff_{X/S}^n(\mathcal F,\mathcal G)$ is the sheaf of differential operators $D:\mathcal F\to\mathcal G$ of order $n$. By construction, those are $\psi^*(\mathcal O_S)$-linear.
Another characterization of $\Diff_{X/S}^n(\mathcal F,\mathcal G)$ for $n<\infty$ is the following. We have on $\iHom_{\psi^*\mathcal O_S}(\mathcal F,\mathcal G)$ an $\psi^*(\mathcal O_{X\times_S X})$-module structure coming from the two $\mathcal O_X$-module structures on $\iHom_{\psi^*\mathcal O_S}(\mathcal F,\mathcal G)$ (we may act on the argument and on the range). Then $\Diff_{X/S}^n(\mathcal F,\mathcal G)$ is (by definition) the subsheaf of $\iHom_{\psi^*\mathcal O_S}(\mathcal F,\mathcal G)$ annihilated by $\mathcal J_X^{n+1}$.
We refer to Proposition (16.8.8) of \cite{ega44} for yet another explicit characterization of differential operators.
The map $\delta_{X/S}^n$ induces a canonical isomorphism
\begin{equation}
\iHom_{\mathcal O_X}(\Pp_{X/S}^n(\mathcal F),\mathcal G)\;\cong\;\Diff_{X/S}^n(\mathcal F,\mathcal G).
\label{eq:Phomdiffiso}
\end{equation}
of sheaves of abelian groups (cf.\ Proposition (16.8.4) of loc.\ cit.).
We set $\Diff_{X/S}(\mathcal F,\mathcal G):=\Diff_{X/S}^\infty(\mathcal F,\mathcal G)$. Then {\em locally} every section $D$ of $\Diff_{X/S}(\mathcal F,\mathcal G)$ has an order, which, by definition, is the least integer $n$ such that $D$ occurs (locally) in the image of
$$
\Diff_{X/S}^n(\mathcal F,\mathcal G)\;\to\;
\Diff_{X/S}(\mathcal F,\mathcal G).
$$
By definition the (global) order of $D$ is the supremum of all local orders. It may be infinite if $X$ is not quasi-compact (which never happens in our applications). For quasi-compact $X$, $\Diff_{X/S}(\mathcal F,\mathcal G)$ is the presheaf colimit of the $\Diff_{X/S}^n(\mathcal F,\mathcal G)$, $n\in{\rm\bf N}_0$. In general, $\Diff_{X/S}(\mathcal F,\mathcal G)$ is the corresponding colimit in the category of sheaves.
By Proposition (16.8.9) \cite{ega44}, composition of morphisms of sheaves induces a well defined morphism
$$
\Diff_{X/S}^m(\mathcal G,\mathcal H)\otimes_{\mathcal O_X} \Diff_{X/S}^n(\mathcal F,\mathcal G)\;\to\;\Diff_{X/S}^{m+n}(\mathcal F,\mathcal H),
$$
$$
D\otimes D'\;\mapsto\;D\circ D'.
$$
In particular,
$$
\D_{X/S}\;:=\;
\Diff_{X/S}(\mathcal O_X,\mathcal O_X)
$$
carries a natural ring structure. It comes with a natural left and a natural right $\mathcal O_X$-module structure, and an increasing filtration given by
$$
\D_S^n(X)\;:=\;
\Diff_{X/S}^n(\mathcal O_X,\mathcal O_X).
$$
If $X\to S$ is locally of finite presentation, then $\D_S^n(X)$ for all $n\in{\rm\bf N}_0\cup\{\infty\}$ is a quasi-coherent $\mathcal O_X$-module for both module structures (cf.\ Proposition (16.8.6) in loc.\ cit.).
\subsection{The smooth case}\label{sec:smooth}
Recall that for $n\geq1$, the graded sheaf $\Gr(\Pp^n_{X/S})$ admits the sheaf $\Omega_{X/S}^1$ of $1$-differentials as $\Gr_1(\Pp^n_{X/S})$. The aim of this section is to recall how $\Omega_{X/S}^1$ \lq{}generates\rq{} $\D_{X/S}$. Recall the canonical isomorphism \eqref{eq:omegaxderivations}. Then if $x:X\to S$ is locally of finite presentation, and if $\mathcal F$ is quasi-coherent, so is $\Der_S(\mathcal O_X,\mathcal F)$. If, additionally, $S$ is locally noetherian and $\mathcal F$ is coherent, then $\Der_S(\mathcal O_X,\mathcal F)$ is coherent \cite[Corollaire (16.5.6)]{ega44}. We obtain the important special case
\begin{proposition}
If $S$ is locally noetherian and $X\to S$ locally of finite type\footnote{hence also locally of finite presentation}, then
$$
\Der_S(X)\;:=\;
\Der_{S}(\mathcal O_X,\mathcal O_X)
$$
is a coherent $\mathcal O_X$-module.
\end{proposition}
With the usual Lie bracket $[-,-]$, $\Der_S(X)$ carries the structure of an $x^*\mathcal O_S$-Lie algebra.
Recall that a morphism $x:X\to S$ is {\em smooth}, if it is locally of finite presentation and formally smooth. Furthermore, every smooth morphism is differentially smooth, cf.\ Proposition (17.12.4) in \cite{ega44}. An important consequence of the differential smoothness of a smooth $X/S$ is that the canonical morphism
$$
{\bf S}^\bullet_{\mathcal O_X}(\Omega_{X/S}^1)\;\to\;\Gr_\bullet(\Pp_{X/S})
$$
is an isomorphism. Furthermore, we have
\begin{proposition}[Propositions (17.2.3) and (17.12.4) in \cite{ega44}]
Let $x:X\to S$ be a smooth morphism. Then $\Omega_{X/S}^1$ is locally free of (locally) finite rank. Furthermore, the same conclusion applies to $\Pp_{X/S}^n$ and $\Gr_n(\Pp_{X/S})$.
\end{proposition}
\begin{corollary}\label{cor:smoothnesstoflatness}
If $X\to S$ is smooth\footnote{we need not assume $S$ to be locally noetherian in this particular case. Finite presentation is implicit in smooth and suffices to guarantee that the sheaf under consideration is quasi-coherent, which with the previous Proposition is enough.}, then $\Der_{S}(X)$ is locally free of finite rank, and for each $n\geq 0$, $\D_{X/S}^n$ is locally free of finite rank. In particular, $\D_{X/S}$ is locally free and therefore a flat $\mathcal O_X$-module.
\end{corollary}
\begin{remark}\label{rmk:smoothdifferentialoperators}
Assume $X/S$ smooth. Then, for any sufficiently small open affine $U\subseteq X$, $\Omega_{X/S}^1|_U$ is a free $\mathcal O_X|_U$-module, of rank $n$ say. For any $p\in U$, we find sections $s_1,\dots,s_n\in\Gamma(\mathcal O_X,U)$ with the property that the canonical images of the collection $ds_1,\dots,ds_n$ form a $k(p)$-basis of $\Omega_{X/S}^1\otimes_{\mathcal O_X}k(p)$. Therefore, by Nakayama's Lemma, the stalks $(ds_1)_p,\dots,(ds_n)_p$ form a basis of $\Omega_{X/S,p}^1$. Hence, if we assume $U$ small enough, $ds_1,\dots,ds_n$ form a basis of $\Gamma(U,\Omega_{X/S}^1)$. Let
$
\frac{\partial}{\partial s_1},\dots,\frac{\partial}{\partial s_n},
$
denote the dual basis in $\Gamma(U,\Der_S(X))$. Then any $\Gamma(S,\mathcal O_S)$-derivation $D$ of $\Gamma(U,\mathcal O_X)$ may be written uniquely as
$$
D\;=\;\sum_{i=1}^n a_i\cdot \frac{\partial}{\partial s_i},\quad a_1,\dots, a_n\in\Gamma(U,\mathcal O_X).
$$
For a multi-exponent ${\bf n}\in{\rm\bf N}_0^n$, we set
$$
D^{\bf n}\;:=\;\left(\frac{\partial}{\partial s_1}\right)^{{\bf n}_1}\cdots\;\;\left(\frac{\partial}{\partial s_n}\right)^{{\bf n}_n}.
$$
Then the collection of the $D^{\bf n}$, $|{\bf n}|\leq m$, is a basis of the $\mathcal O_X(U)$-module $\Gamma(U,\D_{X/S}^m)$, cf.\ Th\'eor\`eme (16.11.2) in \cite{ega44}.
In particular, we conclude that $\Gamma(U,\D_{X/S})$ is the $x^*\mathcal O_S(U)$-algebra generated by elements $\tilde{f},\tilde{\theta}$, where $f\in\mathcal O_X(U)$ and $\theta\in\Gamma(U,\Der_S(X))$, subject to the relations
\begin{itemize}
\item[(1)] $\tilde{f}_1+\tilde{f}_2\;=\;\widetilde{f_1+f_2}$,
\item[(2)] $\tilde{f}_1\tilde{f}_2\;=\;\widetilde{f_1f_2}$,
\item[(3)] $\tilde{\theta}_1+\tilde{\theta}_2\;=\;\widetilde{\theta_1+\theta_2}$,
\item[(4)] $[\tilde{\theta}_1,\tilde{\theta}_2]\;=\;\widetilde{[\theta_1,\theta_2]}$,
\item[(5)] $\tilde{f}\tilde{\theta}\;=\;\widetilde{f\theta}$,
\item[(6)] $[\tilde{\theta},\tilde{f}]\;=\;\widetilde{\theta(f)}$,
\end{itemize}
with elements drawn from the obvious sets of sections.
\end{remark}
\begin{remark}
In general, without any smoothness assumptions, it is easy to see that, if $s_1,\dots,s_n$ generate $\Omega_{X/S}|_U$, then dually the products of all $\delta^ks_1,\dots,\delta^ks_n$, of degree at most $m$, generate $\Pp_{X/S}^m|_U$, and that a differential operator $D$ on $U$ of degree $\leq m$ is determined by its values on the collection of these generators.
\end{remark}
\subsection{Flat connections}
\begin{definition}
Let $x:X\to S$ be a scheme over $S$ and $\mathcal M$ an $\mathcal O_X$-module. A {\em flat connection} on $\mathcal M$ is an $x^*\mathcal O_S$-linear morphism
$$
\nabla:\quad\Der_S(X)\to\iEnd_{x^*\mathcal O_S}(\mathcal M),\quad \theta\mapsto\nabla_\theta,
$$
satisfying for every\footnote{A collection of open subschemes $U$ which constitutes a base of the Zariski topology on $X$ is sufficient, in particular a collection of affine open $U\subseteq X$ covering $X$ is already enough.} open $U\subseteq X$ the following three conditions:
\begin{itemize}
\item[(i)] $\nabla_{f\theta}\;=\;f\nabla_\theta,\quad f\in\mathcal O_X(U),\,\theta\in\Gamma(U,\Der_S(X))$,
\item[(ii)] $\nabla_{\theta}(fs)\;=\;\theta(f)s+f\nabla_\theta(s),\;\; f\in\mathcal O_X(U),\,\theta\in\Gamma(U,\Der_S(X)),\,s\in\Gamma(U,\mathcal M)$,
\item[(iii)] $\nabla_{[\theta_1,\theta_2]}\;=\;[\nabla_{\theta_1},\nabla_{\theta_2}],\quad \theta_1,\theta_2\in\Gamma(U,\Der_S(X))$.
\end{itemize}
\end{definition}
Without any smoothness assumptions on $X/S$, each $\D_{X/S}$-module structure on $\mathcal M$ gives rise to a flat connection $\nabla$. Conversely, for $X\to S$ smooth, remark \ref{rmk:smoothdifferentialoperators} shows that each flat connection $\nabla$ as above gives rise to a unique left-$\D_X$-module structure on $\mathcal M$ via the rule
$$
\theta\cdot s\;:=\;\nabla_\theta(s),\quad\theta\in\Gamma(U,\Der_S(X)),\,s\in\Gamma(U,\mathcal M).
$$
This induces a natural $1$-to-$1$ correspondence between flat connections and $\D_{X/S}$-module structures on $\mathcal M$.
In the non-smooth case, a flat connection need not extend to an action of $\D_{X/S}$.
\subsection{Left and right $\mathcal D$-modules}\label{sec:leftright}
For smooth $X/S$, we define the canonical line bundle $\omega_{X/S}$ as the invertible sheaf on $X$ given locally on sufficiently small $U\subseteq X$ by $\bigwedge^{\dim_S U}\Omega_{U/S}^1$. For equidimensional $X$ (for example connected $X$), we have $\omega_{X/S}=\bigwedge^n\Omega_{X/S}$ for $n=\dim_SX$.
Now $\omega_{X/S}$ carries a natural right $\D_{X/S}$-module structure, and therefore, $\omega_{X/S}^{\otimes-1}=\Hom_{\mathcal O_X}(\omega_{X/S},\mathcal O_X)$ is a left $\D_{X/S}$-module. Writing $\D_{X/S}^{\rm op}$ for the $\mathcal O_X$-algebra opposite to $\D_{X/S}$, we have a canonical isomorphism
$$
\D_{X/S}^{\rm op}\;\cong\;\omega_{X/S}\otimes_{\mathcal O_X}\D_{X/S}\otimes_{\mathcal O_X}\omega_{X/S}^{\otimes-1}.
$$
of $\mathcal O_X$-algebras, as in the classical case. We remark that the category $\Mod(\D_{X/S}^{\rm op})$ of left $\D_{X/S}^{\rm op}$-modules, and the category of right $\D_{X/S}$-modules, and likewise $\Mod(\D_{X/S})$ and the category of right $\D_{X/S}^{\rm op}$-modules, are canonically isomorphic. Additionally, we have an equivalence
$$
\omega_{X/S}\otimes_{\mathcal O_X}(-):\quad \Mod(\D_{X/S})\;\to\;
\Mod(\D_{X/S}^{\rm op}),
$$
with quasi-inverse
$$
\omega_{X/S}^{\otimes-1}\otimes_{\mathcal O_X}(-):\quad \Mod(\D_{X/S}^{\rm op})\;\to\;
\Mod(\D_{X/S}).
$$
Therefore, the categories of left $\D_{X/S}$-modules and right $\D_{X/S}$-modules are equivalent. All these functors preserve $\mathcal O_X$-(quasi-)coherence.
\subsection{Base Change properties}
By Proposition (16.4.5) of \cite{ega44}, we have for any morphism of schemes $S'\to S$ a canonical isomorphism
$$
\psi^*\Omega_{X/S}\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}\;\cong\;\Omega_{X'/S'}
$$
where $(\psi,\theta):X':=X\times_S S'\to X$. Furthermore, without any assumption on $S'\to S$,
\begin{equation}
\psi^*\Pp_{X/S}^n\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}\;\to\;\Pp_{X'/S'}^n,
\label{eq:Pnbasechange}
\end{equation}
is always an isomorphism by Proposition (16.4.5) of loc.\ cit.
More generally, for every $n\in{\rm\bf N}_0$ the natural map
\begin{equation}
\psi^*\Gr_n(\Pp_{X/S}^\infty)\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}\;\to\;\Gr_n(\Pp_{X'/S'}^\infty),
\label{eq:grbasechange}
\end{equation}
is surjective. If $S'\to S$ is flat or $\Gr_n(\Pp_{X/S}^\infty)$ is flat, then \eqref{eq:grbasechange} is known to be an isomorphism by Corollaire (16.4.6) in loc.\ cit. The latter is the case if $X\to S$ is smooth, since then $\Gr_n(\Pp_{X/S}^\infty)$ (and also $\Pp_{X/S}^n$, $n<\infty$) is locally free of finite type (cf.\ Proposition (17.12.4) in loc.\ cit.).
We show
\begin{proposition}
Let $X$ be locally of finite presentation over an arbitrary base $S$. Then for every flat\footnote{It seems to me that one may replace the hypothesis on $S'/S$ by smoothness of $X/S$ and a suitable hypothesis on $\mathcal F$ and $\mathcal G$, i.e.\ coherent and locally free for $\mathcal F$ and quasi-coherent for $\mathcal G$ should be enough, cf.\ \cite{ega1}, Chap.\ 0, (4.3.3.1) and Bourbaki, Alg\`ebre, Chap.\ II, \S 4.2, Prop.\ 2, (ii).} morphism $S'\to S$ of schemes we have for every $n\in{\rm\bf N}_0$, every $\mathcal O_X$-module $\mathcal F$ of finite presentation and every $\mathcal O_X$-module $\mathcal G$ a natural isomorphism
$$
\psi^*\Diff_{X/S}^n(\mathcal F,\mathcal G)\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}\;\to\;
\Diff_{X'/S'}^n(\psi^*\mathcal F\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'},\mathcal G\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}),
$$
where $(\psi,\theta):X':=X\times_S S'\to X$.
The same conclusions also hold for $n=\infty$.
\end{proposition}
\begin{proof}
We assume first that $n<\infty$. For readability we set
$$
\mathcal F'\;:=\;\psi^*\mathcal F\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}
$$
and likewise for $\mathcal G$.
Since $X/S$ is locally of finite presentation, Corollaire (16.4.22) of loc.\ cit.\ tells us that $\Pp_{X/S}^n$ is quasi-coherent of finite presentation. Therefore, by \cite{ega1}, Chap.\ 0, (6.7.6.1), (4.3.3.1), and \eqref{eq:Pnbasechange} we have a natural isomorphism
\begin{equation}
\psi^*\iHom_{\mathcal O_X}(\Pp_{X/S}^n\otimes_{\mathcal O_X}\mathcal F,\mathcal G)\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}
\;\cong\;
\iHom_{\mathcal O_{X'}}(\Pp_{X'/S'}^n\otimes_{\mathcal O_{X'}}\mathcal F',\mathcal G').
\label{eq:firsthomtensoriso}
\end{equation}
The claim for finite $n$ follows with \eqref{eq:Phomdiffiso}.
For infinite $n$,
we have
$$
\varinjlim_m\Diff_{X/S}^m(\mathcal F,\mathcal G)
\;=\;\Diff_{X/S}^\infty(\mathcal F,\mathcal G),
$$
and likewise for $X'/S'$. Now the functor $-\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}$ commutes with direct limits (cf.\ \cite{ega1}, Chap.\ 0, (4.3.2)), hence the claim follows.
\end{proof}
\begin{corollary}\label{cor:dbasechange}
If $X/S$ is of locally of finite presentation and $S'\to S$ is a flat morphism, we have for every $n\in{\rm\bf N}_0\cup\{\infty\}$ a canonical isomorphism
\begin{equation}
\psi^*\D_{X/S}^n\otimes_{\psi^*\mathcal O_X}\mathcal O_{X'}
\;\to\;
\D_{X'/S'}^n.
\label{eq:Dbasechange}
\end{equation}
\end{corollary}
\section{Generalities on $\mathcal D$-modules}\label{sec:Dmodules}
Let $S$ be any scheme and $X\to S$ locally of finite presentation. Recall that we write $\Mod_{\rm qc}(\D_{X/S})$ for the category of $\mathcal O_X$-quasi-coherent (left) $\D_{X/S}$-modules on $X$.
\subsection{$\mathcal D$-affine schemes}
\begin{definition}
We say that a scheme $x:X\to S$ is {\em $\mathcal D$-affine}, if it is locally of finite presentation and if the following two conditions are satisfied:
\begin{itemize}
\item[(i)] $x_*:\Mod(\mathcal O_X)\to\Mod(\mathcal O_S)$ sends quasi-coherent modules to quasi-coherent modules.
\item[(ii)] For every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$
\begin{equation}
\forall q>0:\quad R^qx_*(\mathcal M)\;=\;0.
\label{eq:affinecohomologyvanishing}
\end{equation}
\item[(iii)] For every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ we have an equivalence
$$
\mathcal M\;=\;0\quad\Leftrightarrow\quad x_*\mathcal M\;=\;0.
$$
\end{itemize}
\end{definition}
\begin{remark}
Obviously ${\bf1}_X:X\to X$ is $\mathcal D$-affine. More generally, any affine morphism $x:X\to S$ of finite presentation is $\mathcal D$-affine by Corollaire (1.3.2) in \cite{ega31}.
\end{remark}
\begin{remark}\label{rmk:onDaffinei}
Sufficient conditions on $x:X\to S$ to satisfy condition (i) are (cf.\ \cite[Corollaire 9.2.2]{ega1} for (b), (c) and (d)):
\begin{itemize}
\item[(a)] $S$ is affine.
\item[(b)] $x$ is (quasi-)separated and quasi-compact.
\item[(c)] $x$ is separated and of finite type.
\item[(d)] $x$ is quasi-compact and $X$ is locally noetherian.
\item[(e)] $x$ is of finite type and $S$ is locally noetherian\footnote{this implies that $X$ is locally noetherian and that $x$ is of finite presentation. In our basic case $S=\Spec k$, $S$ is even noetherian and we may study any $X$ of finite type over $S$.}.
\end{itemize}
\end{remark}
\begin{remark}
In the case of $S=\Spec k$ for $k$ a field (or more generally a noetherian ring), any scheme $X\to S$ of finite type is of finite presentation and quasi-compact. Furthermore, in that case, and more generally if $S$ is affine, the pushforward functor $x_*(-)$ may be identified with the functor $\Gamma(X,-)$ of global sections, hence (i) is automatically satisfied in this case and the derived functors in (ii) correspond to sheaf cohomology $H^q(X,-)$.
\end{remark}
\begin{proposition}\label{prop:Daffinezariskilocal}
Let $S$ be a scheme. The property of being a $\mathcal D$-affine $S$-scheme is local in the Zariski topology: If $x:X\to S$ is a scheme over $S$, then $x$ is $\mathcal D$-affine if and only if for some (equivalently, all) open cover $(V_j)_j$ of $S$, the collection of the schemes $x^{-1}(V_j)\to V_j$ is $\mathcal D$-affine.
\end{proposition}
We will see in Theorem \ref{thm:Daffinefpqclocal} below that being $\mathcal D$-affine is local in the $fpqc$ topology.
\begin{proof}
Properties (i) and (iii) are Zariski local by definition. As for (ii), it suffices to observe that $R^qx_*(\mathcal M)$ is the sheaf associated to the presheaf $
V\mapsto H^q(x^{-1}(V),\mathcal M).
$
\end{proof}
\subsection{Quasi-coherent $\mathcal D$-modules}
\begin{proposition}\label{prop:firstequivalenceovergeneralbase}
Assume that $X$ is a $\mathcal D$-affine scheme over a scheme $S$. Then
\begin{itemize}
\item[(i)] Every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ is generated by its relative sections over $S$ as a $\D_{X/S}$-module.
\item[(ii)] The functor
$$
x_*:\quad
\Mod_{\rm qc}(\D_{X/S})\to
\Mod_{\rm qc}(x_*\D_{X/S})
$$
is an equivalence of categories.
\end{itemize}
\end{proposition}
\begin{proof}
The proof proceeds as in the classical case with only minor modifications. We sketch the argument.
For (i), consider the image $\mathcal N$ of the canonical map
$$
\D_{X/S}\otimes_{\mathcal O_X}x^*x_*\mathcal M
\;\to\;
\mathcal M,
$$
induced by the counit $x^*x_*\to{\bf1}$, which, by definition, is the submodule generated by the relative sections. We obtain a short exact sequence
$$
\begin{CD}
0@>>>\mathcal N@>>>\mathcal M@>>>\mathcal M/\mathcal N@>>> 0.
\end{CD}
$$
Since $X$ is $\mathcal D$-affine, application of $x_*(-)$ yields a short exact sequence
$$
\begin{CD}
0@>>>x_*\mathcal N@>i>>x_*\mathcal M@>>>x_*(\mathcal M/\mathcal N)@>>> 0.
\end{CD}
$$
Now $i$ is surjective, since the canonical map
$$
x_*x^*x_*\mathcal M\to x_*\mathcal M
$$
is already surjective for formal reasons. Whence the cokernel of $i$ vanishes, and consequently $\mathcal M/\mathcal N$ vanishes since $X$ is $\mathcal D$-affine.
As for statement (ii), observe first that $\mathcal C:=x_*\D_{X/S}$ is quasi-coherent, because $\D_{X/S}$ is quasi-coherent and $X$ is $\mathcal D$-affine. Now $\D_{X/S}$ is a right $x^*\mathcal C$-module, and every $\mathcal O_S$-quasi-coherent left $\mathcal C$-module $M$ gives rise to an $\mathcal O_X$-quasi-coherent left $x^*\mathcal C$-module $x^*M$ (cf.\ \cite[Prop.\ (9.1.4)]{ega1}). Therefore, we have a functor
$$
\D_{X/S}\otimes_{x^*\mathcal C}x^*(-):\quad
\Mod_{\rm qc}(x_*\D_{X/S})\to
\Mod_{\rm qc}(\D_{X/S}).
$$
We claim that this is a quasi-inverse of $x_*$. By construction, $\D_{X/S}\otimes_{x^*\mathcal C}x^*(-)$ is a left adjoint of $x_*$. Whence it suffices to prove that the unit and counit of the adjunction are isomorphisms.
Since $\mathcal C$ is a quasi-coherent $\mathcal O_S$-module, locally any $\mathcal O_S$-quasi-coherent $\mathcal C$-module $\mathcal M$ admits a presentation
$$
\begin{CD}
\mathcal C^{(n)}|_U@>>> \mathcal C^{(m)}|_U@>>> \mathcal M|_U@>>> 0,
\end{CD}
$$
where $m$ and $n$ are cardinal numbers and $\mathcal C^{(n)}$ denotes the direct sum of $n$ copies of $\mathcal C$ (any affine $U\subseteq S$ does the job). We write $x|_U:x^{-1}U\to U$ for the restriction of $x:X\to S$ to $x^{-1}U$. We remark that
$$
\D_{X/S}|_{x^{-1}U}\;=\;\D_U(x^{-1}U).
$$
Therefore, $x|_U:x^{-1}U\to U$ is $\mathcal D$-affine. Since $(x|_U)_*$ is exact, $\D_U(x^{-1})\otimes_{x^*\mathcal C|_U}((x|_U)^*(-))$ is right exact, and we obtain with the canonical isomorphism
$$
(x_*\D_{X/S}\otimes_{x^*\mathcal C}x^* {\mathcal M})|_U\;=\;
\D_U(x^{-1}U)\otimes_{x^*\mathcal C|_{x^{-1}U}} {\mathcal M}|_U.
$$
the diagram
$$
\begin{CD}
\mathcal C^{(n)}|_U@>>> \mathcal C^{(m)}|_U@>>> \mathcal M|_U@>>> 0\\
@| @| @VV\varepsilon_{\mathcal M}|_U V @|\\
\mathcal C^{(n)}|_U@>>> \mathcal C^{(m)}|_U@>>> x_*(\D_{X/S}\otimes_{x^*\mathcal C}x^*{\mathcal M})|_U
@>>> 0
\end{CD}
$$
with exact rows. We conclude that locally, and hence a fortiori globally, the unit $\varepsilon_\bullet$ of the adjunction is an isomorphism.
As for the counit
$$
\delta_{\mathcal M}:\quad
\D_{X/S}\otimes_{x^*\mathcal C}x^*x_*\mathcal M\to\mathcal M,
$$
we already know by (i) that it is surjective, i.e.\ we obtain a short exact sequence
$$
\begin{CD}
0@>>>\mathcal N@>>>
\D_{X/S}\otimes_{x^*\mathcal C}x^*x_*\mathcal M@>>>\mathcal M@>>>0,
\end{CD}
$$
where $\mathcal N$ denotes the kernel of the counit. Since $\varepsilon_{x_*\mathcal M}$ is an isomorphism, application of $x_*$ yields a short exact sequence
$$
\begin{CD}
0@>>>x_*\mathcal N@>>>
x_*\mathcal M@>>>x_*\mathcal M@>>>0,
\end{CD}
$$
whence $x_*\mathcal N$ vanishes. Since $\mathcal N$ is quasi-coherent, the fact that $X$ is $\mathcal D$-affine shows $\mathcal N=0$.
\end{proof}
\begin{proposition}\label{prop:injectiveobjects}
For $X/S$ arbitrary with $X$ quasi-compact, and $\mathcal A$ an $\mathcal O_X$-flat $\mathcal O_X$-quasi-coherent $\mathcal O_X$-algebra. Then each $\mathcal M\in\Mod_{\rm qc}(\mathcal A)$ embeds into an injective object $\mathcal I\in\Mod_{\rm qc}(\mathcal A)$, which we may assume to be injective in $\Mod(\mathcal O_X)$.
\end{proposition}
We emphasize that $\mathcal I$ needs not be injective within the category of all $\mathcal A$-modules.
\begin{proof}
The proof proceeds mutatis mutandis as the proof of Proposition 1.4.14 in \cite{hottaetal2008}.
\end{proof}
\begin{corollary}\label{cor:injectiveobjects}
For $X/S$ smooth, $S$ quasi-compact, each $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ embeds into an injective object $\mathcal I\in\Mod_{\rm qc}(\D_{X/S})$, which we may assume to be flabby.
\end{corollary}
\begin{proof}
By Corollary \ref{cor:smoothnesstoflatness}, $\D_{X/S}$ is a flat $\mathcal O_X$-module, whence the claim.
\end{proof}
\begin{remark}\label{rmk:sufficientcondition2}
Corollary \ref{cor:injectiveobjects} tells us that for $x:X\to S$ smooth and $S$ quasi-compact the category $\Mod_{\rm qc}(\D_{X/S})$ contains enough injectives. They are $x_*(-)$-acyclic and in the case at hand, condition (ii) in definition of $\mathcal D$-affineness becomes equivalent to $x_*(-)$ being exact on $\mathcal O_X$-quasi-coherent $\D_{X/S}$-modules.
\end{remark}
\subsection{Coherent $\mathcal D$-modules}
We begin with an easy observation in
\begin{proposition}\label{prop:coherentDgenerators}
Assume $X\to S$ of finite type and $S$ noetherian. Then a coherent $\D_{X/S}$-module $\mathcal M$ is generated by an $\mathcal O_X$-coherent $\mathcal O_X$-submodule.
\end{proposition}
The last statement is to be understood that there is a coherent $\mathcal O_X$-submodule $\mathcal F$ of $\mathcal M$ with the property that the canonical map
$$
\D_{X/S}\otimes_{\mathcal O_X}\mathcal F\to\mathcal M
$$
be an epimorphism of sheaves.
\begin{proof}
Since $\mathcal M$ is locally finitely generated, we find a finite affine covering $U_i$, $i\in I$, of $X$ with the property that on each $U_i$, $\mathcal M|_{U_i}$ is a finitely generated $\D_{U_i/S}$-module. This is the same to say that there is a coherent $\mathcal O_{U_i}$-submodule $\mathcal F_i$ in $\mathcal M|_{U_i}$ which generates $\mathcal M|_{U_i}$ as $\D_{U_i/S}$-module. Now each $\mathcal F_i$ canonically extends to a coherent $\mathcal O_X$-submodule $\mathcal G_i$ of $\mathcal M$ (cf.\ Th\'eor\`eme (9.4.7) in \cite{ega1}), and the finite sum $\mathcal G$ of the modules $\mathcal G_i$, $i\in I$, is a coherent $\mathcal O_X$-submodule of $\mathcal M$ which generates $\mathcal M$.
\end{proof}
If $X\to S$ is locally of finite presentation, we saw in section \ref{sec:differentialoperators} that $\D_{X/S}=\Diff_{X/S}^\infty(\mathcal O_X,\mathcal O_X)$ is naturally equipped with the structure of a filtered $\mathcal O_X$-algebra. The finite filtration steps $\D_{X/S}^n$ are quasi-coherent $\mathcal O_X$-modules.
Later in this section, we will restrict our attention to the case where $X/S$ is smooth. Then, by Corollary \ref{cor:smoothnesstoflatness}, $\D_{X/S}$ is locally free and all filtration steps are locally free of finite rank. We will also assume that $S$ be noetherian.
Before we proceed to this case, we recall and generalize some essential facts about {\em good filtrations}. In the case of rings, all of the following statements are contained in Appendix D.1 of \cite{hottaetal2008}.
\begin{proposition}\label{prop:affinegoodfiltration}
For a filtered (not necessarily commutative) ring $A$ and an $A$-module $M$ the following are equivalent:
\begin{itemize}
\item[(i)] $M$ is a finitely generated $A$-module.
\item[(ii)] $M$ admits a filtration such that $\Gr(M)$ is a finitely generated $\Gr(A)$-module.
\end{itemize}
\end{proposition}
If $M$ is finitely generated, we call a filtration as in (ii) {\em good}. Every finitely generated module admits good filtrations by the above Proposition.
\begin{proposition}\label{prop:affinenoetheriancriterion}
If for a filtered ring $A$, $\Gr(A)$ is left / right noetherian, then so is $A$.
\end{proposition}
The the sheaf theoretic analogues of these notions are the following. If $\mathcal M$ is a quasi-coherent $\mathcal O_X$-module over a scheme $X$, then a filtration $\mathcal F_i$, $i\in I$, of $\mathcal M$ is always assumed to consist of quasi-coherent $\mathcal O_X$-modules $\mathcal F_i$. Then $\Gr_\bullet(\mathcal M)$ is quasi-coherent as well. Furthermore, if $\mathcal M$ is an $\mathcal O_X$-algbera (with unit), we assume that locally on affine opens $U\subseteq X$ all filtrations statisfy the conditions (a)-(e) from Appendix D.1 in \cite{hottaetal2008}.
We remark that we only suppose that
$$
\varinjlim_{i\in I}\mathcal F_i\;=\;\mathcal M
$$
as {\em sheaves}. In particular, we only know locally for open affine\footnote{or more generally quasi-compact} $U\subseteq X$, that
$$
\bigcup_{i\in I}\Gamma(U,\mathcal F_i)\;=\;\Gamma(U,\mathcal M).
$$
\begin{proposition}\label{prop:schemegoodfiltration}
Let $X$ be a noetherian scheme and $\mathcal A$ a filtered quasi-coherent $\mathcal O_X$-algebra (always with unit). Assume that $\Gr_\bullet(\mathcal M)$ is quasi-coherent. Let $\mathcal M$ be a quasi-coherent $\mathcal A$-module. Then the following are equivalent:
\begin{itemize}
\item[(i)] $\mathcal M$ is a (locally) finitely generated $\mathcal A$-module.
\item[(ii)] $\mathcal M$ admits a filtration such that $\Gr_\bullet(\mathcal M)$ is quasi-coherent and $\Gr(\mathcal M)$ is a finitely generated $\Gr(\mathcal A)$-module.
\end{itemize}
\end{proposition}
\begin{proof}
Cover $X$ with finitely many affine open subsets $U_1,\dots,U_r$. For every such affine open $U_j$, statements (i) and (ii) are equivalent to statements (i) and (ii) in Proposition \ref{prop:affinegoodfiltration} for $B_j=\Gamma(U_i,\mathcal A)$ and $M_j=\Gamma(U_j,\mathcal M)$. Therefore, $\mathcal M$ is locally finitely generated, and by Th\'eor\`eme (9.4.7) in \cite{ega1} it is also globally finitely generated. The implication (ii) $\Rightarrow$ (i) follows.
Assuming (i), i.e.\ that $\mathcal M$ is locally finitely generated, we first conclude with Th\'eor\`eme (9.4.7) of loc.\ cit.\ that we find an $\mathcal O_X$-coherent submodule $\mathcal M_0\subseteq\mathcal M$ together with an epimorphism
$$
s:\quad\mathcal A\otimes_{\mathcal O_X}\mathcal M_0\to\mathcal M.
$$
The canonical images of $\mathcal A_i\otimes\mathcal M_0$ under $s$ then define a filtration $\mathcal M_i$ on $\mathcal M$ with the desired properties.
\end{proof}
Again we call a filtration of $\mathcal M$ satisfying (ii) {\em good}.
\begin{proposition}\label{prop:schemenoetheriancriterion}
Assume $X$ is a noetherian scheme. A filtered quasi-coherent $\mathcal O_X$-algebra $\mathcal A$ is locally left / right noetherian if $\Gr(\mathcal A)$ is locally left / right noetherian.
\end{proposition}
We remark that this is only really meaningful if $\mathcal O_X$ itself is locally noetherian.
\begin{proof}
The claim reduces to Proposition \ref{prop:affinenoetheriancriterion} in the same way like Proposition \ref{prop:schemegoodfiltration} reduces to Proposition \ref{prop:affinegoodfiltration}.
\end{proof}
\begin{proposition}\label{prop:Dnoetherian}
Assume $X/S$ smooth and $S$ locally noetherian.
Then $\D_{X/S}$ is locally left and right noetherian. In particular, for every affine open $U\subseteq X$, the $\mathcal O_X|_U$-algebra $\D_{X/S}|_U$ is left and right noetherian. Furthermore, for any $p\in X$, the $\mathcal O_{X,p}$-algebra $\D_{X/S,p}$ is left and right noetherian.
\end{proposition}
\begin{proof}
The hypothesis on $X\to S$ implies that that $X$ is locally noetherian and also that $\D_{X/S}$ is a quasi-coherent (even locally free) filtered $\mathcal O_X$-module with each filtration step locally free of finite rank by Corollary \ref{cor:smoothnesstoflatness}. More precisely, we know by Remark \ref{rmk:smoothdifferentialoperators} that locally, for sufficient affine open $U\subseteq X$, we obtain a polynomial algebra
$$
\Gr(\D_{U/S})\;=\;\mathcal O_X(U)[D_1,\dots,D_n].
$$
Since $\mathcal O_X(U)$ is noetherian, this algebra is noetherian, and whence by Proposition \ref{prop:schemenoetheriancriterion}, $\D_{U/S}$ is locally left and right noetherian. A foriori, we conclude that for every affine open $U\subseteq X$, $\D_{X/S}|_U=\D_{U/S}$ is left and right noetherian. The claim about the stalks follows by a similar argument.
\end{proof}
\begin{corollary}\label{cor:Dnoetherian}
If $X\to S$ is smooth and $S$ noetherian, then $\D_{X/S}$ is a left and right noetherian $\mathcal O_X$-algebra.
\end{corollary}
\begin{proposition}\label{prop:coherentDmodules}
For smooth $X\to S$ and $S$ locally noetherian, a $\D_{X/S}$-module $\mathcal M$ is $\D_{X/S}$-coherent if and only if $\mathcal M$ is a $\mathcal O_X$-quasi-coherent and locally finitely generated $\D_{X/S}$-module.
\end{proposition}
\begin{proof}
The proof proceeds as in the classical case. Assume $\mathcal M$ to be $\D_{X/S}$-coherent. Then, by definition, $\mathcal M$ is locally finitely generated as $\D_{X/S}$-module. We need to show that $\mathcal M$ is $\mathcal O_X$-quasi-coherent.
For every open $U\subseteq X$ such that $\mathcal M|_U$ is finitely generated, we obtain an exact sequence
$$
\begin{CD}
(\D_{U/S})^n@>>>
(\D_{U/S})^m@>>>\mathcal M|_U@>>>0
\end{CD}
$$
with finite $m,n\geq 0$, by the definition of $\D_{X/S}$-coherence. Now we already know that $\D_{U/S}$ is $\mathcal O_U$-quasi-coherent, and the claim follows.
Assume converely that $\mathcal M$ is an $\mathcal O_X$-quasi-coherent locally finitely generated $\D_{X/S}$-module. Choose any open $U\subseteq X$ and consider for any $m\geq 0$ the exact sequence
$$
\begin{CD}
0@>>>\mathcal K@>>>
(\D_{U/S})^m@>>>\mathcal M|_U
\end{CD}
$$
of $\mathcal O_U$-modules. We need to show that $\mathcal K$ is a locally finitely generated $\D_{U/S}$-module. Hence, without loss of generality, we may assume $U$ affine. Then, because $\mathcal M$ is $\mathcal O_X$-quasi-coherent, $\mathcal K$ is quasi-coherent as well, and the above exact sequence is the same as the exact squence
$$
\begin{CD}
0@>>>\Gamma(U,\mathcal K)@>>>
\Gamma(U,\D_{X/S})^m@>>> \Gamma(U,\mathcal M)
\end{CD}
$$
of $\mathcal O_X(U)$-modules. Now by Proposition \ref{prop:Dnoetherian}, $\Gamma(U,\D_{X/S})$ is a left noetherian ring and therefore $\Gamma(U,\mathcal K)$ is a finitely generated $\Gamma(U,\D_{X/S})$-module. We conclude that every locally finitely generated submodule of $\mathcal M|_U$ is a locally finitely presentated $\D_{U/S}$-module. This concludes the proof.
\end{proof}
\begin{corollary}
For $X\to S$ smooth and $S$ locally noetherian, $\D_{X/S}$ is a $\D_{X/S}$-coherent $\D_{X/S}$-module.
\end{corollary}
\begin{theorem}
Let $X\to S$ be smooth and $S$ locally noetherian. Then a $\D_{X/S}$-module $\mathcal M$ is $\mathcal O_X$-coherent if and only if $\mathcal M$ is a locally free $\mathcal O_X$-module of finite rank.
\end{theorem}
\begin{proof}
If $\mathcal M$ is locally free, then $\mathcal M$ is quasi-coherent and $\mathcal O_X$-coherent because $X$ is locally noetherian.
For the other implication, assume that $\mathcal M$ is coherent as $\mathcal O_X$-module. It suffices to see that for each $x\in X$, the stalk $\mathcal M_x$ is free. The proof of this claim proceeds as the proof of Theorem 1.4.10 in \cite{hottaetal2008}.
\end{proof}
\begin{corollary}
For $X/S$ smooth and $S$ locally noetherian, the category $\Mod_{\rm c}(\D_{X/S})$ of $\D_{X/S}$-coherent $\D_{X/S}$-modules is a thick abelian subcategory of $\Mod_{\rm qc}(\D_{X/S})$. In particular, the category of integrable connections on $X$ (i.e.\ the category of pairs $(\mathcal M,\nabla)$, with $\mathcal M$ a locally free $\mathcal O_X$-module of finite rank and $\nabla$ a flat connection on $\mathcal M$), is abelian.
\end{corollary}
\begin{proposition}\label{prop:coherentequivalenceovergeneralbase}
Assume $X/S$ smooth and $\mathcal D$-affine and $S$ locally noetherian. Then, under the equivalence in Proposition \ref{prop:firstequivalenceovergeneralbase}, $\Mod_{\rm c}(\D_{X/S})$ corresponds to locally finitely generated $x_*(\D_{X/S})$-modules.
\end{proposition}
\begin{proof}
By restricting to an affine open $V\subseteq S$, we may assume that $S$ is affine without loss of generality (cf.\ Proposition \ref{prop:Daffinezariskilocal}). In this case the proof proceeds as the proof of Proposition 1.4.13 in \cite{hottaetal2008}.
\end{proof}
\subsection{Locally free resolutions}
Recall that if $S$ is noetherian, an $S$-scheme $X$ is {\em projective} if the structural morphism $X\to S$ factors as $X\to\mathbb P_S^n\to S$, where the first map is assumed to be a closed immersion, and the second is the canonical structural morphism. With this terminology $X$ is called {\em quasi-projective}, if $X\to S$ factors into an open immersion $X\to Y$ of finite type and a projective map $Y\to S$ (cf.\ Corollaire (5.3.3) in \cite{ega2}).
\begin{theorem}\label{thm:freeresolutions}
Assume that $X\to S$ is smooth and quasi-projective and $S$ noetherian. Then:
\begin{itemize}
\item[(i)] For any $\mathcal O_X$-quasi-coherent $\D_{X/S}$-module $\mathcal M$ we find a locally free\footnote{both as $\mathcal O_X$ and as $\D_{X/S}$-module at once since we are supposing $X/S$ smooth; without this assumption only locally free as $\D_{X/S}$-module.} $\D_{X/S}$-module $\mathcal F$ which admits $\mathcal M$ as homomorphic image.
\item[(ii)] If
$\mathcal M$ is $\D_{X/S}$-coherent, then $\mathcal F$ in (i) may be chosen of finite rank over $\D_{X/S}$.
\end{itemize}
\end{theorem}
\begin{proof}
Choose a quasi-coherent $\mathcal O_X$-submodule $\mathcal F_0$ of $\mathcal M$ which generates $\mathcal M$. If $\mathcal M$ is $\D_{X/S}$-coherent, we may suppose $\mathcal F_0$ locally free of finite rank by Proposition \ref{prop:coherentDgenerators}, because $X$ is quasi-compact.
It is a standard fact that $\mathcal F_0$ is an epimorphic image of a locally free $\mathcal O_X$-module $\mathcal G$, of finite rank if $\mathcal F_0$ is of finite rank.
Indeed, fix an immersion $i:X\to\mathbb P_S^n$, which realizes $X$ as a quasi-projective scheme. Since $S$ is noetherian, so is $X$, and we recall that therefore every $\mathcal O_X$-quasi-coherent $\mathcal O_X$-module $\mathcal F_0$ is an inductive limit of coherent submodules (cf.\ Corollaire (9.4.9) in \cite{ega1}. Therefore, we may assume without loss of generality that $\mathcal F_0$ is coherent.
By Proposition (4.5.5) d) in \cite{ega2}, there are $n,r\geq 0$ and an epimorphism
$$
\mathcal O(-n)^{r}\;\to\;i_*(\mathcal F_0),
$$
where the domain is locally free of finite rank. Therefore, the composition
$$
\mathcal F:=i^*(\mathcal O(-n)^{r})\;\to\;i^*i_*(\mathcal F_0)\;\to\;\mathcal F_0
$$
is surjective, with domain again locally free of finite rank, and the standard fact follows.
From this we deduce an epimorphism $
\D_{X/S}\otimes_{\mathcal O_X}\mathcal F\to\mathcal M
$
of $\D_{X/S}$-modules, and the claim follows because $\D_{X/S}$ is locally free by Corollary \ref{cor:smoothnesstoflatness}.
\end{proof}
We conclude that for $X/S$ smooth and quasi-projective and $S$ noetherian, every quasi-coherent $\D_{X/S}$-module admits a locally free resolution. In order to bound the resolution, we need to recall a few facts about dimensions.
The relative dimension of $x:X\to S$ at a $p\in X$ is by definition the dimension $\dim_p X_p$ of the fiber at $p$, where the latter is the infimum of the dimensions of open subschemes $U\subseteq X_p$ containing $p$. In general the dimension is only upper semi-continuous in $p\in X$.
If $x:X\to S$ is smooth, then for each $p\in X$, we have
$$
\dim_p x\;=\;\rank_{\mathcal O_{X,p}}\Omega_{X/S,p}^1
$$
and in particular the dimension is a continuous function in $p\in X$ (cf.\ Proposition (17.10.2) \cite{ega44}). The relative dimension of $X/S$ is by definition the supremum of all relative dimensions at all $p\in X$.
In remark \ref{rmk:smoothdifferentialoperators} we have seen that $\rank_{\mathcal O_{X,p}}\Omega_{X/S,p}^1$ equals the relative dimension of the ring $\Gr(\D_{U/S})$ over $\mathcal O_{U}$ for a sufficiently small affine open $U\subseteq X$ containing $p$.
If $S$ is noetherian we therefore obtain for every $p\in X$ for the absolute dimension the formula
$$
\dim \Gr(\D_{X/S,p})\;=\;2\dim_p X+\dim_{x(p)}S.
$$
This function is only upper semi-continuous in general. Likewise, we obtain for any affine open $U\subseteq X$,
$$
\dim \Gr(\D_{U/S})\;\leq\;2\dim U+\dim S.
$$
With Theorem D.2.6 in \cite{hottaetal2008} we conclude
\begin{theorem}\label{thm:existenceofboundedcomplexes}
Assume $X/S$ smooth and $S$ noetherian. Then the left (resp.\ right) global dimension of $\D_{X/S}$ is bounded by $2\dim X+\dim S$.
\end{theorem}
\begin{remark}\label{rmk:locallyprojectiveresolutions}
In particular, under the hypothesis of the Theorem \ref{thm:freeresolutions}, we see that every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ admits a locally projective resolution of length $\leq 2\dim X+\dim S$, which in all degrees may be chosen finitely generated if $\mathcal M$ is $\D_{X/S}$-quasi-coherent.
\end{remark}
\begin{remark}\label{rmk:flatresolutions}
For a not necessarily $\mathcal O_X$-quasi-coherent $\D_{X/S}$-module, Proposition \ref{prop:Dnoetherian} implies the existence of bounded flat resolutions (under the assumptions of Theorem \ref{thm:existenceofboundedcomplexes}).
\end{remark}
\section{Operations with $\mathcal D$-modules}\label{sec:operationswithDmodules}
We first recall several statements about base change for quasi-coherent sheaves and then discuss applications.
\subsection{Review of base change}
We recall that if $x:X\to S$ is a morphism of schemes and $\mathcal F$ is an $\mathcal O_X$-module, then $\mathcal F$ is called flat over $S$ if for every $p\in X$ the module $\mathcal F_p$ is a flat $\mathcal O_{S,x(p)}$-module. If $S=\Spec k$ is the spectrum of a field, then every $\mathcal F$ is flat over $S$.
We recall the following results from \cite{ega31, ega32}.
\begin{theorem}[{Flat base change I, \cite[Proposition (1.4.15)]{ega31}}]\label{thm:flatbasechange1}
Let $x:X\to S$ be separated of finite type, and $\mathcal F$ a quasi-coherent $\mathcal O_X$-module. Then we have for every flat morphism $s:S'\to S$ and every degree $q$ a canonical isomorphism
$$
R^qx'_*(y^*\mathcal F)
\;\cong\;
s^*R^qx_*(\mathcal F)
$$
where $x':X'=X\times_S S'\to S'$ and $y:X'\to X$ is the canonical projection.
\end{theorem}
\begin{theorem}[Flat base change II]\label{thm:flatbasechange2}
Let $x:X\to S$ be separated quasi-compact, and $\mathcal F$ a quasi-coherent $\mathcal O_X$-module, flat over $S$, and assume that for all degrees $q$ the modules $R^qx_*(\mathcal F)$ are flat over $S$. Then we have for every $s:S'\to S$ and every degree $q$ a canonical isomorphism
$$
R^qx'_*(y^*\mathcal F)
\;\cong\;
s^*R^qx_*(\mathcal F)
$$
with the notation as in Theorem \ref{thm:flatbasechange1}.
\end{theorem}
\begin{proof}
This is a special case of Corollaire (6.9.9), \'equation (6.9.9.2), \cite{ega32}, where the complex $\Pp^\bullet$ in the statement of loc.\ cit.\ is given by $\mathcal F$ concentrated in degree $0$.
\end{proof}
\subsection{Flat base change for $\mathcal D$-modules}
We put ourselves in the following general situation. Let $S$ be an arbitrary scheme, $x:X\to S$ locally of finite presentation, and $s:S'\to S$ flat. We let $x'=X\times_S S'$ and write $x':X'\to S'$ and $y:X'\to X$ for the canonical projections.
These are precisely the hypotheses of Corollary \ref{cor:dbasechange}, and we see that the canonical map
\begin{equation}
y^*\D_{X/S}\;\to\;\D_{X'/S'}.
\label{eq:Dringbasechange}
\end{equation}
is an isomorphism. Therefore, for any $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$, the quasi-coherent $\mathcal O_{X'}$-module $y^*\mathcal M$ carries a natural $\D_{X'/S'}$-module structure. More precisely we obtain
\begin{theorem}\label{thm:flatDmodulebasechange}
Let $S$ be a scheme, $x:X\to S$ locally of finite presentation, $s:S'\to S$ flat, and $X'=X\times_S S'$, with the remaining notation as before. Then
\begin{itemize}
\item[(i)] Pullback along $y$ induces an exact functor
$$
y^*:\quad
\Mod_{\rm qc}(\D_{X/S})\to \Mod_{\rm qc}(\D_{X'/S'}).
$$
\item[(ii)] The functor $y^*$ preserves the following properties of $\mathcal M$:
\begin{itemize}
\item[(a)] (local) finite generation,
\item[(b)] (local) finite presentation,
\item[(c)] $\mathcal O_X$-coherence,
\item[(d)] $\mathcal O_X$-local freeness.
\end{itemize}
\item[(iii)] We have in each degree $q$ for every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ a canonical map
$$
s^*(R^qx_*\mathcal M)\;\to\;(R^qx'_*)y^*\mathcal M,
$$
which is an isomorphism whenever $x:X\to S$ is separated of finite type.
\item[(iv)] For $x:X\to S$ is separated of finite presentation, the following diagram commutes stricly:
$$
\begin{CD}
\Mod_{\rm qc}(\D_{X'/S'})@>x'_*>>
\Mod(x'_*\D_{X'/S'})\\
@Ay^*AA @AAs^*A\\
\Mod_{\rm qc}(\D_{X/S})@>x_*>>
\Mod(x_*\D_{X/S})
\end{CD}
$$
Here $s^*$ is exact as well.
\end{itemize}
\item[(v)] If $x:X\to S$ is separated of finite presentation and $s:S'\to S$ faithfully flat, then $y^*$ is faithful and we have for every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ and every degree $q$,
$$
R^qx_*\mathcal M=0\quad\Leftrightarrow\quad(R^qx'_*)y^*\mathcal M=0.
$$
\end{theorem}
\begin{proof}
Statement (i) is clear. With the isomorphism \eqref{eq:Dringbasechange} and the exactness of $y^*$, statements (ii), (a), (b) follow readily. Statements (c) and (d) follow from $y^*\mathcal O_X=\mathcal O_{X'}$ and the fact that $y^*$ commutes with arbitrary direct sums. Statement (iii) is well known. Statement (iv) follows from Theorem \ref{thm:flatbasechange1}. Statement (v) is a consequence of (iv) and the faithfulness of $s^*$.
\end{proof}
\begin{corollary}\label{cor:flatDmodulebasechange}
Under the general hypotheses of Theorem \ref{thm:flatDmodulebasechange}, if additionally $X\to S$ is smooth and $S$ locally noetherian, then the exact functor $y^*$ preserves $\mathcal D$-coherence.
\end{corollary}
\begin{proof}
This follows from Theorem \ref{thm:flatDmodulebasechange} (ii), (a) together with Proposition \ref{prop:coherentDmodules}.
\end{proof}
The first elementary consequence of faithfully flat descent is
\begin{corollary}\label{cor:Daffinenessdescent}
Let $x:X\to S$ be separated of finite presentation and $s:S'\to S$ faithfully flat and quasi-compact. Then, if $X'/S'$ is $\mathcal D$-affine, then $X/S$ is $\mathcal D$-affine.
\end{corollary}
\begin{proof}
Let $\mathcal M$ be in $\Mod_{\rm qc}(\D_{X/S})$. By faithfully flat descent, $x_*\mathcal M$ is quasi-coherent, because $s^*x_*\mathcal M=x'_*y^*\mathcal M$ is quasi-coherent. The higher direct images
$$
R^qx_*\mathcal M\;=\;0
$$
vanish for $q>0$, because they do so after application of the faithful functor $s^*$. Likewise, if $x_*\mathcal M$ vanishes, then $x'_*y^*\mathcal M$ vanishes, hence $y^*\mathcal M$ vanishes and since $y^*$ is faithful, this implies the vanishing of $\mathcal M$.
\end{proof}
We remark that if $x:X\to S$ is additionally quasi-compact, then all higher direct images $R^qx_*$, $q\geq 0$, always preserve quasi-coherence. This holds more generally for quasi-separated quasi-compact $X/S$, cf.\ Remark \ref{rmk:onDaffinei}.
\begin{theorem}\label{thm:Daffinefpqclocal}
For $x:S\to S$ separated of finite presentation being $\mathcal D$-affine is local in the $fpqc$ topology.
\end{theorem}
\begin{proof}
Using Corollary \ref{cor:Daffinenessdescent}, the proof proceeds as in the case of the Zariski topology (cf.\ Proposition \ref{prop:Daffinezariskilocal}).
\end{proof}
\subsection{Faithfully flat descent for $\mathcal D$-modules}
Here are more interesting consequences of faithfully flat descent. Set $S'':=S'\times_S S'$, and consider the commutative square
$$
\begin{CD}
S'' @>\pi_2>> S'\\
@V\pi_1VV @VVsV\\
S' @>>x> S\\
\end{CD}
$$
where $\pi_i:S''\to S'$ denotes the $i$-th projection. We set $X''=X\times_S S''$ and denote $z:X''\to X$ the canonical projection.
By \eqref{eq:Dbasechange}, we may define a descent datum for $\mathcal D$-modules is defined mutatis mutandis like a descent datum for $\mathcal O_X$-modules, replacing all $\mathcal O_X$-linear maps by $\D_{X/S}$-linear maps.
\begin{theorem}\label{thm:fpqcDmoduledescent}
Let $x:X\to S$ be separated of finite presentation and $s:S'\to S$ faithfully flat and quasi-compact. For any quasi-coherent $\D_{X/S}$-modules $\mathcal M,\mathcal N$ we have:
\begin{itemize}
\item[(i)] A $\mathcal O_X$-linear morphism $f:\mathcal M\to\mathcal N$ is $\D_{X/S}$-linear if and only if the induced morphism
$$
y^*f:\quad y^*\mathcal M\to y^*\mathcal N
$$
is $\D_{X'/S'}$-linear.
\item[(ii)] For $S''=S'\times_S S'$, $X''=X\times_SS''$, the sequence of abelian groups
\begin{equation}
\begin{CD}
\Hom_{\D_{X/S}}(\mathcal M,\mathcal N)@>y^*>>
\Hom_{\D_{X'/S'}}(y^*\mathcal M,y^*\mathcal N)@>\pi_1^*>>
\Hom_{\D_{X''/S''}}(z^*\mathcal M,z^*\mathcal N)\vspace*{-1.5em}
\\
@. @>>\pi_2^*>
\end{CD}
\label{eq:fpqcmorphismdescent}
\end{equation}
is exact.
\item[(iii)] $y^*$ induces a natural equivalence between the category $\Mod_{\rm qc}(\D_{X/S})$ and the category of quasi-coherent $\D_{X'/S'}$-modules with descent data.
\item[(iv)] The following properties are local in the $fpqc$ topology, i.e.
$$
\mathcal M\;\text{satisfies property (P)}\quad\Leftrightarrow\quad
y^*\mathcal M\;\text{satisfies property (P)},
$$
where (P) is:
\begin{itemize}
\item[(a)] (local) finite generation,
\item[(b)] (local) finite presentation,
\item[(c)] $\mathcal O_X$-flatness,
\item[(d)] $\mathcal O_X$-local freeness of finite type,
\item[(e)] $\mathcal O_X$-local freeness of finite rank $n$,
\item[(f)] $\mathcal D$-coherence.
\end{itemize}
\end{itemize}
\end{theorem}
\begin{remark}
Statement (ii) (\lq{}$fpqc$ descent for morphisms of $\mathcal D$-modules\rq{}) may be restated as follows. A morphism $f:y^*\mathcal M\to y^*\mathcal N$ of $\D_{X'/S'}$-modules is of the form $f=y^*g$ for some morphism $g:\mathcal M\to \mathcal N$ of $\D_{X/S}$-modules if and only if
$$
\pi_1^*f\;=\;
\pi_2^*f.
$$
Statement (iii) has a similar but more involved explicit formulation. A special case of (ii) and (iii) is Galois descent for $\mathcal D$-modules in the case $S=\Spec k$ and $S'=\Spec k'$, $k'/k$ a Galois extension of fields. In particular, properties (a) - (f) over $k$ may be checked over an algebraic closure $k'$.
\end{remark}
\begin{remark}
In the situation of a finite Galois extension $k'/k$ with Galois group $\Gal(k'/k)$, statements (ii) and (iii) are equivalent to saying that the category of $\D_{X/S}$-modules is equivalent to the category of $\D_{X'/S'}$-modules with a semi-linear action of $\Gal(k'/k)$.
From this perspective, the proof of Theorem 5.12 in \cite{januszewskirationality} amounts to establishing that there is such a Galois action in the case at hand if and only if said symmetric bilinear forms exist.
\end{remark}
\begin{proof}
Statement (i) is an elementary consequence of (ii).
For statement (ii), we observe that, by $fpqc$ descent for morphisms, we know that
$$
\begin{CD}
\Hom_{\mathcal O_{X}}(\mathcal M,\mathcal N)@>y^*>>
\Hom_{\mathcal O_{X'}}(y^*\mathcal M,y^*\mathcal N)@>\pi_1^*>>
\Hom_{\mathcal O_{X''}}(z^*\mathcal M,z^*\mathcal N)\vspace*{-1.5em}
\\
@. @>>\pi_2^*>
\end{CD}
$$
is exact. By (i) we deduce from this the diagram \eqref{eq:fpqcmorphismdescent}, which remains exact because the functor of $\D_{X/S}$-invariants is left exact.
Statement (iii) follows from $fpqc$ descent for quasi-coherent sheaves, observing that the functor $\D_{-/-}$ is an $fpqc$ sheaf by means of \eqref{eq:Dbasechange}.
Let us prove statement (iv).
First we observe that properties (a), (b) and (f) are all local properties in the Zariski topology and that $U\subseteq X$ is open if and only if $y^{-1}U\subseteq X'$ is open, and likewise on $S$. Therefore, we may assume without loss of generality that $X$ and $S$ are affine. By Corollaire (2.2.12) in \cite{ega42} we may assume without loss of generality that also $S'$ and hence $X'$ are both affine. This reduces us to the situation of modules over the rings $\Gamma(X,\D_{X/S})$ and $\Gamma(X,\D_{X'/S'})$, where the second is faithfully flat over the first. Therefore, properties (a) and (b) are $fpqc$ local by \cite[Proposition 11, Chap.\ I, \S 3, no.\ 6]{bourbaki}. The statement for property (f) reduces to properties (a) and (b) by the definition of $\mathcal D$-coherence.
Properties (c), (d) and (e), may be verified within the category of quasi-coherent $\mathcal O_X$-modules, where the statements follow from Proposition (2.5.1) and Proposition (2.5.2), (iii) and (iv), in \cite{ega42}.
\end{proof}
\subsection{Inverse and direct images}
Let $f:Y\to X$ be a morphism of $S$-schemes, $X$ and $Y$ both smooth over $S$. We recall that then, for each $n\geq 0$, $\Pp^n_{X/S}$ and $\Pp^n_{Y/S}$ are both locally free of (locally) finite rank. Therefore, we have
\begin{equation}
f^*\iHom_{\mathcal O_X}(\Pp_{X/S}^n,\mathcal O_X)\;=\;
\iHom_{\mathcal O_Y}(f^*\Pp_{X/S}^n,\mathcal O_Y),
\label{eq:Dnsmoothbigbasechange}
\end{equation}
which follows readily stalk wise.
From $f$ we obtain an induced morphism
of augmented $\mathcal O_Y$-algebras
$
f^*\Pp_{X/S}^n\to\Pp^n_{Y/S},
$
cf.\ Proposition (16.4.18) \cite{ega44}. This in turn induces by \eqref{eq:Dnsmoothbigbasechange} a canonical map
$
\D_{Y/S}^n\to
f^*\D_{X/S}^n
$
of sheaves of rings, whose colimit gives us a canonical map
\begin{equation}
\D_{Y/S}\to
f^*\D_{X/S}
\label{eq:canonicalpullbackmap}
\end{equation}
of filtered sheaves of rings. Therefore, we obtain a right exact functor
$$
f^*:\quad\Mod_{\rm qc}(\D_{X/S})\to
\Mod_{\rm qc}(\D_{Y/S}),
$$
which is classically called the {\em inverse image functor}. In the literature $f^*\D_{X/S}$ is often denoted $\D_{Y\to X}$.
We also have an induced exact sequence
$$
f^*\Omega_{X/S}^1\to\Omega_{Y/S}^1\to\Omega_{Y/X}^1\to 0
$$
of quasi-coherent $\mathcal O_Y$-modules (cf.\ Corollaire (16.4.9) in \cite{ega44}), which, if $f$ is also smooth, extends to a short exact sequence
\begin{equation}
0\to f^*\Omega_{X/S}^1\to\Omega_{Y/S}^1\to\Omega_{Y/X}^1\to 0,
\label{eq:smoothexactsequence}
\end{equation}
cf.\ Proposition (17.2.3) \cite{ega44}.
Back in the general situation of an $S$-morphism $f:Y\to X$, we deduce from the canonical map $f^*\Omega_{X/S}^1\to\Omega_{Y/S}^1$ a morphism
$
\Der_S(Y)\to f^*\Der_S(X),
$
and the operation $f^*$ is compatible with the passage from $\mathcal D$-modules to connections.
The {\em direct image} of a {\em right} $\D_{Y/S}$-module $\mathcal M$ is defined as
$$
f_*(\mathcal M\otimes_{\D_{Y/S}}f^*\D_{X/S}).
$$
For left modules $\mathcal M$ this translates to
\begin{equation}
\omega_{X/S}^{\otimes-1}\otimes_{\mathcal O_X}f_*\left(f^*\D_{X/S}\otimes_{\D_{Y/S}}\left(\omega_{Y/S}\otimes_{\mathcal O_Y}\mathcal M\right)\right),
\label{eq:explicitdirectimage}
\end{equation}
cf.\ section \ref{sec:leftright}. This formula simplifies if we define the {\em
transfer bimodule}
$$
\D_{X\leftarrow Y/S}\;:=\;\omega_{Y/S}\otimes_{\mathcal O_Y}f^*\D_{X/S}\otimes_{f^*\mathcal O_X}f^*\left(\omega_{X/S}^{\otimes-1}\right).
$$
This is an $(f^*\D_{X/S},\D_{Y/S})$-bimodule, and \eqref{eq:explicitdirectimage} takes the simpler form
$
f_*(\D_{X\leftarrow Y/S}\otimes_{\D_{Y/S}}\mathcal M).
$
\subsection{Derived inverse and direct images}
We assume as before that $f:Y\to X$ is a morphism of smooth $S$-schemes, with $S$ locally noetherian.
For a sheaf of rings $\mathcal R$ on $X$ (or $Y$), we write $D^{\rm b}(\mathcal R)$ for the (locally) bounded derived category of $\mathcal R$-modules, i.e.\ we assume that for each $p\in X$ exists an open neighborhood $U\subseteq X$ of $p$ with the property that the restriction of the complex under consideration is bounded in the classical sense. If $X$ is quasi-compact, then this notion reduces to the classical notion of bounded complexes.
For $\bullet\in\{-,\rm qc,\rm c\}$, we let $D^{\rm b}_{\bullet}(\mathcal R)$ denote the full (triangulated) subcategory of complexes whose cohomology lies degree-wise in $\Mod_\bullet(\mathcal R)$.
Remark \ref{rmk:flatresolutions} translates to the fact that every $\mathcal K\in D^{\rm b}(\D_{X/S})$ may be represented by a (locally) bounded complex of flat $\D_{X/S}$-modules. If $\mathcal K$ lies in $D^{\rm b}_{\rm qc}(\D_{X/S})$, then we always find bounded locally projective resolutions (cf.\ Remark \ref{rmk:locallyprojectiveresolutions}).
Define the derived preimage as the left derived functor of $f^*$, i.e.\ it is given by
$$
Lf^*:\quad D^{\rm b}_{\bullet}(\D_{X/S})\to D_{\bullet}^{\rm b}(\D_{Y/S}),\quad \bullet\in\{-,\rm qc\}.
$$
Explicitly, if $f=(\psi,\theta)$, then $Lf^*(\mathcal M^\bullet)$ is given by
$$
\mathcal O_Y\otimes_{\psi^*\mathcal O_X}\psi^*\mathcal M^\bullet\;=\;
f^*\D_{X/S}\otimes_{\psi^*\D_{X/S}}\psi^{*}\mathcal M^\bullet.
$$
It is clear that $Lf^*$ preserves complexes of $\mathcal O_X$-quasi-coherent modules. It does not necessarily perserve $\mathcal D$-coherent modules.
\begin{proposition}\label{prop:smoothpreimage}
Assume $f:Y\to X$ is a smooth morphism between smooth $S$-schemes, $S$ locally noetherian. Then
\begin{itemize}
\item[(i)] For any degree $q> 0$, and any $\mathcal M\in\Mod(\D_{X/S})$, we have
$$
L_qf^*(\mathcal M)\;=\;0.
$$
\item[(ii)] $f^*$ (and $Lf^*$) preserves $\mathcal D$-coherent modules.
\end{itemize}
\end{proposition}
\begin{proof}
As for (i), observe that smooth morphisms are flat, whence the claim follows.
By (i), it suffices to observe that the canonical map \eqref{eq:canonicalpullbackmap} is surjective. By the exactness of the sequence \eqref{eq:smoothexactsequence}, this follows with Remark \ref{rmk:smoothdifferentialoperators}.
\end{proof}
The definition of the derived direct image for left $\D_{Y/S}$-modules is explicitly given by
$$
\int_f: \quad D^{\rm b}(\D_{Y/S})\to D^{\rm b}(\D_{X/S}),
$$
$$
\mathcal M^\bullet\mapsto
Rf_*(\D_{X\leftarrow Y/S}\otimes^L_{\D_{Y/S}}\mathcal M^\bullet).
$$
Here $Rf_*$ is understood as the restriction of the derived functor
$$
Rf_*:\quad D^{\rm b}(\mathcal O_{Y})\to
D^{\rm b}(\mathcal O_{X}),
$$
to
$$
Rf_*:\quad
D^{\rm b}(f^*\D_{X/S})\to
D^{\rm b}(\D_{X/S}),
$$
and the inner functor is a left derived functor, which on the level of $\mathcal D$-modules is given by
\begin{equation}
\D_{X\leftarrow Y/S}\otimes^L_{\D_{Y/S}}(-):
\quad D^{\rm b}(\D_{Y/S})\to
D^{\rm b}(f^*\D_{X/S}).
\label{eq:leftderiveddirectimage}
\end{equation}
\begin{proposition}\label{prop:basechangefordirectimages1}
Let $f:Y\to X$ be a morphism of smooth $S$-schemes, $S$ locally noetherian. Assume that $s:S'\to S$ is flat. Denote by $f':Y'\to X'$ the base change of $f$ to $S'$. Then we have a canonical commuting square
$$
\begin{CD}
D^{\rm b}(\D_{Y'/S'})
@>\D_{X'\leftarrow Y'/S'}\otimes^L_{\D_{Y'/S'}}(-)>>
D^{\rm b}({f'}^*\D_{X'/S'})\\
@Ay^*AA @AAy^*A\\
D^{\rm b}(\D_{Y/S})
@>>\D_{X\leftarrow Y/S}\otimes^L_{\D_{Y/S}}(-)>
D^{\rm b}(f^*\D_{X/S})
\end{CD}
$$
\end{proposition}
\begin{proof}
The proof is an elementary degenerate instance of the Grothendieck spectral sequence. Since the passage from left $\mathcal D$-modules to right $\mathcal D$-modules commutes with flat base change, we may consider the equivalent diagram
$$
\begin{CD}
D^{\rm b}(\D_{Y'/S'}^{\rm op})
@>(-)\otimes^L_{\D_{Y'/S'}}{f'}^*\D_{X'/S'}>>
D^{\rm b}({f'}^*\D_{X'/S'}^{\rm op})\\
@Ay^*AA @AAy^*A\\
D^{\rm b}(\D_{Y/S}^{\rm op})
@>>(-)\otimes^L_{\D_{Y/S}}f^*\D_{X/S}>
D^{\rm b}(f^*\D_{X/S}^{\rm op})
\end{CD}
$$
Then by the flatness of $s:S'\to S$, the canonical maps $y^*\D_{Y/S}\to \D_{Y'/S'}$ and $y^*\D_{X/S}\to \D_{X'/S'}$ are isomorphisms (cf.\ Corollary \ref{cor:dbasechange}). In particular, we conclude (with \cite[(4.3.3.1), Chap.\ 0]{ega1}), that for every right $\D_{Y/S}$-module $\mathcal M$ we have a canonical isomorphism
\begin{equation}
y^*(\mathcal M)
\otimes_{\D_{Y'/S'}}
{f'}^*\D_{X'/S'}
=
y^*\left(
\mathcal M
\otimes_{\D_{Y'/S'}}
f^*\D_{X/S}
\right).
\label{eq:preimagecommutativity}
\end{equation}
For every object $\mathcal K$ in $D^{\rm b}(\D_{Y/S}^{\rm op})$ we find a bounded complex $\mathcal M^\bullet$ of $\D_{Y/S}$-flat right modules representing it. Furthermore, the base change of a $\D_{Y/S}$-flat right module is $\D_{Y'/S'}$-flat, hence $y^*(\mathcal M^\bullet)$ is a flat complex representing $y^*\mathcal K$. This shows that
$
y^*(\mathcal K)
\otimes^L_{\D_{Y'/S'}}
{f'}^*\D_{X'/S'}
$
is represented by the complex
\begin{equation}
y^*(\mathcal M^\bullet).
\otimes_{\D_{Y'/S'}}
{f'}^*\D_{X'/S'}
\label{eq:fprimeycomplex}
\end{equation}
The functor
$
(-)\otimes_{\D_{Y/S}}f^*\D_{X/S}
$
preserves flatness as well, and we concude that
$
y^*\left(
\mathcal K
\otimes^L_{\D_{Y/S}}
{f}^*\D_{X/S}
\right)
$
is represented by the complex
\begin{equation}
y^*\left(
\mathcal M^\bullet
\otimes_{\D_{Y/S}}
f^*\D_{X/S}
\right).
\label{eq:yfcomplex}
\end{equation}
The canonical isomorphism \eqref{eq:preimagecommutativity} shows that the complexes \eqref{eq:yfcomplex} and \eqref{eq:fprimeycomplex} are canonically isomorphic.
\end{proof}
\begin{proposition}\label{prop:basechangefordirectimages2}
Let $f:Y\to X$ be a morphism of smooth $S$-schemes, $S$ noetherian. Assume that $s:S'\to S$ is flat. Denote by $f':Y'\to X'$ the base change of $f$ to $S'$. Then we have a canonical commuting square
$$
\begin{CD}
D_{\rm qc}^{\rm b}(\D_{X'\leftarrow Y'/S'})
@>Rf'_*>>
D^{\rm b}_{\rm qc}(\D_{X/S})\\
@Ay^*AA @AAy^*A\\
D_{\rm qc}^{\rm b}(\D_{X\leftarrow Y/S})
@>>Rf_*>
D_{\rm qc}^{\rm b}(\D_{X/S})
\end{CD}
$$
\end{proposition}
\begin{proof}
This is a special case of Theorem \ref{thm:flatbasechange1}.
\end{proof}
\begin{theorem}[Base change for direct images]\label{thm:basechangefordirectimages}
Let $f:Y\to X$ be a morphism of smooth $S$-schemes, $X\to S$ separated and $S$ locally noetherian. Assume that $s:S'\to S$ is flat. Denote by $f':Y'\to X'$ the base change of $f$ to $S'$. Then we have a canonical commuting square
$$
\begin{CD}
D^{\rm b}_{\rm qc}(\D_{Y'/S'})
@>\int_{f'}>>
D^{\rm b}_{\rm qc}(\D_{X'/S'})\\
@Ay^*AA @AAy^*A\\
D^{\rm b}_{\rm qc}(\D_{Y/S})
@>>\int_f>
D^{\rm b}_{\rm qc}(\D_{X/S})
\end{CD}
$$
\end{theorem}
\begin{proof}
By Propositions \ref{prop:basechangefordirectimages1} and \ref{prop:basechangefordirectimages2} it suffices to prove that the functor \eqref{eq:leftderiveddirectimage} preserves $\mathcal O_X$-quasi-coherent modules. This will be achieved in Proposition \ref{prop:preservationofqcmodules} below.
\end{proof}
\begin{lemma}\label{lem:directimageforclosedimmersions}
Assume $S$ locally noetherian and $j:Y\to X$ a closed immersion of smooth $S$-schemes. Then the functor \eqref{eq:leftderiveddirectimage} for $f=j$ preserves $\mathcal O_X$-quasi-coherent modules, i.e.\ it induces a functor
$$
\D_{X\leftarrow Y/S}\otimes^L_{\D_{Y/S}}(-):
\quad D^{\rm b}_{\rm qc}(\D_{Y/S})\to
D^{\rm b}_{\rm qc}(j^*\D_{X/S}),
$$
and satisfies
$$
\forall q\neq 0:\quad H^q\left(\D_{X\leftarrow Y/S}\otimes^L_{\D_{Y/S}}(\mathcal M)\right)\;=\;0
$$
for any complex $\mathcal M$ of $\D_{Y/S}$-modules concentrated in degree $0$.
\end{lemma}
\begin{proof}
We pass again to right $\mathcal D$-modules, for which it suffices to show that we obtain a well defined functor.
$$
(-)\otimes^L_{\D_{Y/S}}j^*\D_{X/S}:
\quad D^{\rm b}_{\rm qc}(\D_{Y/S}^{\rm op})\to
D^{\rm b}_{\rm qc}(j^*\D_{X/S}^{\rm op}).
$$
This is implied by the second claim, translated to right modules.
We first remark that the morphism
$$
j^*\Pp_{X/S}^n\to\Pp^n_{Y/S},
$$
is known to be surjective (cf.\ Proposition (16.4.20) \cite{ega44}), and hence, dually
$$
\D_{Y/S}\to
j^*\D_{X/S}
$$
is a monomorphism.
Furthermore, we have a short exact sequence
$$
0\to \mathcal N_{Y/X}\to j^*(\Omega_{X/S}^1)\to\Omega_{Y/S}^1\to 0,
$$
which is locally split (cf.\ Proposition (17.2.5) \cite{ega44}). Assume $p\in Y$ arbitrary and $U\subseteq Y$ a sufficiently small open containing $p$, such that we obtain a splitting
\begin{equation}
j^*(\Omega_{X/S}^1)|_U = \mathcal N_{Y/X}|_U\oplus\Omega_{Y/S}^1|_U,
\label{eq:Omegasplitting}
\end{equation}
and such that $j^*(\Omega_{X/S}^1)$ and $\Omega_{Y/S}^1$ are both free. We conclude that $\mathcal N_{Y/X}|_U$ is a finitely generated projective $\mathcal O_Y$-module, hence locally free, and, after possibly shrinking $U$ further, finite free.
Hence, locally, we have an isomorphism
\begin{equation}
j^*\D_{X/S}|_U\;=\;
\D_{Y/S}|_U\otimes_{\mathcal O_U}{\rm S}_{\mathcal O_U}(\mathcal N_{Y/X}|_U),
\end{equation}\label{eq:dsymmetriciso}
where ${\rm S}_{\mathcal O_U}(\mathcal N_{Y/X}|_U)$ denotes the symmetric $\mathcal O_U$-algebra generated by $\mathcal N_{Y/X}|_U$. Therefore, $j^*\D_{X/S}|_U$ is a locally free left $\D_{Y/S}$-module, and the claim follows.
\end{proof}
\begin{remark}
The isomorphism \eqref{eq:dsymmetriciso} has the following explicit description. For sufficiently small affine open $U\subseteq Y$ and $V\subseteq X$ containing $p$ and $f(p)$, we find sections $s_1,\dots,s_n$ in $\Gamma(V,\mathcal O_X)$ such that $ds_1,\dots,ds_n$ form a basis of $\Gamma(V,\Omega_{X/S}^1)$, and for which the first $m$ components $ds_1,\dots,ds_m$ originate from a basis of $\mathcal N_{Y/X}|_U$ and the last $n-m$ components from a basis of $\Omega_{Y/S}^1|_U$ respectively via the splitting \eqref{eq:Omegasplitting}. Following Remark \ref{rmk:smoothdifferentialoperators}, we obtain, locally on $U$,
$$
j^*\D_{X/S}|_U\;=\;
\D_{Y/S}|_U\otimes_{\mathcal O_U}\mathcal O_U\left[\left(\frac{\partial}{\partial s_i}\right)_{1\leq i\leq n}\right].
$$
\end{remark}
\begin{lemma}\label{lem:directimageforprojections}
Assume $S$ locally noetherian and $X$ and $Y$ smooth $S$-schemes, and write $p:Y\times_S X\to Y$ for the canonical projection. Then the functor \eqref{eq:leftderiveddirectimage} for $f=p$ preserves $\mathcal O_X$-quasi-coherent modules, i.e.\ it induces a functor
$$
\D_{Y\leftarrow Y\times_S X/S}\otimes^L_{\D_{Y/S}}(-):
\quad D^{\rm b}_{\rm qc}(\D_{Y/S})\to
D^{\rm b}_{\rm qc}(p^*\D_{X/S}).
$$
\end{lemma}
\begin{proof}
We consider the Spencer resolution
$$
\cdots
\to\Omega_{X/S}^{{\bf n}-1}\otimes_{\mathcal O_X}\D_{X/S}
\to\Omega_{X/S}^{\bf n}\otimes_{\mathcal O_X}\D_{X/S}
\to\omega_{X/S}\to 0
$$
of the right $\D_{X/S}$-module $\omega_{X/S}$, which is concentrated in degree $0$, and where locally ${\bf n}$ is the dimension, i.e.\ for equidimensional $U\subseteq X$ of relative dimension $n$, the Spencer resolution is explicitly given by
$$
\cdots
\to\bigwedge^{n-1}\Omega_{X/S}|_U\otimes_{\mathcal O_U}\D_{X/S}|_U
\to\bigwedge^{n}\Omega_{X/S}|_U\otimes_{\mathcal O_U}\D_{X/S}|_U
\to\omega_{X/S}|_U\to 0.
$$
Hence the Spencer resolution is a (locally) bounded complex $\mathcal S_{X/S}^\bullet$: Its restriction to $U$ vanishes in degrees $<-(n+1)=-(\dim_S U+1)$.
Its differential is explicitly given for local coordinates in \cite[Lemma 1.5.27]{hottaetal2008}. The proof in loc.\ cit.\ applies to our general situation as well and shows that the Spencer resolution is acyclic. It is obviously locally free over $\mathcal O_X$.
Now consider the outer tensor product
$
\D_{Y/S}\boxtimes\,\mathcal S_{X/S}^\bullet
$
on $Y\times_S X$, which is a resolution of $\D_{Y\leftarrow Y\times_S X/S}\boxtimes\,\omega_{X/S}$, and a straightforward computation with the Spencer complex shows that for every $\mathcal M\in\Mod_{\rm qc}(\D_{Y\times_SX/S})$, the $p^*\D_{Y/S}$-module
$
\D_{Y\leftarrow Y\times_SX/S}\otimes_{\D_{Y\times_SX/S}}^L\mathcal M
$
is represented by the quasi-coherent relative de Rham complex
$
\Omega_{Y\times_SX/Y}^{\bullet-{\bf n}}\otimes_{\mathcal O_{Y\times_SX}}\mathcal M,
$
with ${\bf n}$ as before. This concludes the proof.
\end{proof}
\begin{proposition}\label{prop:preservationofqcmodules}
Assume $S$ noetherian and $f:Y\to X$ a morphism of smooth $S$-schemes, with $X\to S$ separated. Then the functor \eqref{eq:leftderiveddirectimage} preserves $\mathcal O_X$-quasi-coherent modules, i.e.\ it induces a functor
$$
f^*\D_{X/S}\otimes^L_{\D_{Y/S}}(-):
\quad D^{\rm b}_{\rm qc}(\D_{Y/S})\to
D^{\rm b}_{\rm qc}(f^*\D_{X/S}).
$$
\end{proposition}
\begin{proof}
Since $X/S$ is separated, the graph
$
F:Y\to Y\times_SX
$
is a closed immersion, and we have $f=p\circ F$ in the notation of Lemma \ref{lem:directimageforprojections}. The claim follows from Lemmas \ref{lem:directimageforclosedimmersions} and \ref{lem:directimageforprojections}.
\end{proof}
\section{Twisted sheaves of differential operators}
We briefly recall relevant notions and properties of twisted sheaves of differential operators.
\subsection{Sheaves of differential operators of higher rank}
Let $\mathcal M$ denote a quasi-coherent $\mathcal O_X$-module on $X$. Attached to $\mathcal M$ is an $\mathcal O_X$-quasi-coherent ring of differential operators
$$
\D_{X/S}^{\mathcal M}\;:=\;\Diff_{X/S}({\mathcal M},{\mathcal M}).
$$
For $\mathcal M$ locally free of finite rank we obtain for every $n\in{\rm\bf N}_0$ canonical isomorphism
\begin{eqnarray*}
\Diff_{X/S}^n({\mathcal M},{\mathcal M})
&=&
\iHom_{\mathcal O_X}(\Pp_{X/S}^n({\mathcal M}),{\mathcal M})\\
&=&
\iHom_{\mathcal O_X}(\Pp_{X/S}^n\otimes_{\mathcal O_X}{\mathcal M},\mathcal O_X)\otimes_{\mathcal O_X}{\mathcal M}\\
&\cong&
{\mathcal M}^\vee\otimes_{\mathcal O_X}
\iHom_{\mathcal O_X}(\Pp_{X/S}^n,\mathcal O_X)\otimes_{\mathcal O_X}{\mathcal M}\\
&=&
{\mathcal M}^\vee\otimes_{\mathcal O_X}
\D_{X/S}^n\otimes_{\mathcal O_X}{\mathcal M},
\end{eqnarray*}
where
$
{\mathcal M}^\vee:=
\iHom_{\mathcal O_X}({\mathcal M},\mathcal O_X)
$
denotes the dual module. Passing to the inductive limit shows
\begin{equation}
\D_{X/S}^{\mathcal M}\;\cong\;
{\mathcal M}^\vee\otimes_{\mathcal O_X}
\D_{X/S}\otimes_{\mathcal O_X}{\mathcal M},
\label{eq:twistedDtountwistedD}
\end{equation}
canonically as $\mathcal O_X$-algebras. In particular, we see that $\D_{X/S}^{\mathcal M}$ is locally isomorphic to a finite direct sum of copies of $\D_{X/S}$ and comes with a canonical filtration. We remark that Corollary \ref{cor:dbasechange} remains valid for $\D_{X/S}^{\mathcal M}$ replacing $\mathcal D_{X/S}$.
\subsection{Twisted sheaves of differential operators}
Since in our applications we are working in the context of (families of) algebraic groups, we content us with the following definition. For an approach to a more general definition (in the rank $1$ case) we refer to \S2 of \cite{kashiwara1989}.
\begin{definition}
Let $X/S$ be an $S$-scheme locally of finite presentation. A {\em twisted sheaf of differential operators on $X$} is an $\mathcal O_X$-module $\mathcal A$ on $X$ which is locally isomorphic to $\mathcal D_{X/S}^{\mathcal M}$ for a locally free $\mathcal O_X$-module $\mathcal M$ of finite rank. We call the (local) rank of $\mathcal M$ the {\em rank} of $\mathcal A$.
\end{definition}
We remark that being a twisted sheaf of differential operators is a local property on $X$ but not on the base. Furthermore, in the above definition we may replace the Zariski topology by a finer topology to obtain different notions of twisting.
The results from sections \ref{sec:Dmodules} and \ref{sec:operationswithDmodules} carry over without modification to rings of twisted differential operators $\mathcal A$. In particular we have a notion of $\mathcal A$-affine schemes, which is again a local property in the $fpqc$ topology, and likewise the statements on $\mathcal D$-modules, base change properties, etc.\ generalize to $\mathcal A$-modules.
\subsection{Group actions and quotients}
Let $G$ be a group scheme over a base scheme $S$ and $X$ a scheme over $S$. Then a right action of $G$ on $X$ is a morphism $\mu:X\times_S G\to X$ of $S$-schemes satisfying the categorification of the usual two axioms of a group action (associativity and trivial action of the unit section of $G$), which are enforced by the commutativity of the the obvious corresponding diagrams in the category of $S$-schemes.
A notorious difficulty in computing quotients in the category of schemes is that for arbitrary $S$-schemes $A$ we do in general {\em not} have an equality
\begin{equation}
(X/G)(A)\;=\;X(A)/G(A).
\label{eq:quotients}
\end{equation}
The reason is the following. If we embed the category of $S$ schemes into the category of contravariant functors of $S$-schemes to sets, associating to an $S$-scheme $Y$ the functor $Y:A\mapsto Y(A):=\Hom_S(A,Y)$ of $A$-valued points, then the right hand side of \eqref{eq:quotients} corresponds to the the quotient of $X$ by the action of $G$ in this larger category of functors, which is too large for our purpose.
Still one may construct the left hand side \eqref{eq:quotients} by embedding $S$-schemes into a larger category. A reasonable choice is the category of $fppf$ sheaves\footnote{The category of sheaves for every subcanonical topology does the job; for smooth groups and smooth actions the \'etale topology is often also useful to compute the left hand side of \eqref{eq:quotients}; over a field $S=\Spec k$, if $A$ is given by a separable extension of $k$, the computation of $X(A)/G(A)$ directly relates to Galois cohomology; however the coarser the underlying topology, the fewer quotients do exist, because the canonical map to the quotient must be a covering in the considered topology.} over $S$.
The $fppf$ sheaf associated to the quotient one wants to study is (under suitable assumptions) the sheaf associated to the presheave given by the right hand side of \eqref{eq:quotients} for $A$ running through the (elements of) $fppf$ coverings. The notion of {\em $fppf$ quotient} commutes with arbitrary base change.
An axiomatic definition of a {\em (universal) categorical quotient} $X/G$ is the following. We have the action $\mu$ and also the canonical projection
$$
\pi_1:\quad X\times_S G\to X.
$$
Then a map $f:X\to Y$ of $S$-schemes is {\em constant on $G$-orbits}, if
\begin{equation}
f\circ\mu\;=\;f\circ \pi_1
\label{eq:constantorbitcondition}
\end{equation}
Then the quotient $X/G$ is characterized by the fact that the canonical projection $\pi:X\to X/G$ that comes along with it, is universal among all maps $f:X\to Y$ which are constant on $G$-orbits: each such map is supposed to factor uniquely over $\pi$.
In other words $\pi:X\to X/G$ is universal among all $G$-equivariant maps where the codomain has trivial $G$-action.
In general the quotient $X/G$ does not exist in the category of $S$-schemes. Worse, if it exists, it needs not commute with arbitrary base change in general.
If $X/G$ exists as an $fppf$ quotient and is representable by an $S$-scheme $X/G$, then $X/G$ is a (universal) geometric\footnote{A {\em geometric quotient} is a quotient in the category of ringed spaces.} and in particular a (universal) categorical quotient \cite{raynaud1967}. There is also a converse to this statement.
If $G$ acts freely on $X$, and if furthermore the $fppf$ quotient $(X/G)_{fppf}$ is representable by an $S$-scheme $X/G$, then the canonical projection $X\to X/G$ is an $fppf$ covering and turns $X$ into a $G$-torsor over $Y$ in the $fppf$ topology, trivialized by the covering $X\to X/G$.
Left actions and their quotients are defined similarly.
For the existence of quotients we quote
\begin{theorem}[{\cite[Expos\'e 5, \S 7]{sga3}}]
Let $S$ be locally notherian and $G$ be a proper flat $S$-group scheme of finite type. Then
\begin{itemize}
\item[(i)] If $\mu_X:X\times_SG\to X$ is a strictly free\footnote{A free action is one for which $X\times_SG \to X\times_S X, (x,g)\mapsto (x,xg)$ is a monomorphism in the category of $S$-schemes; the action is strictly free if this map is an immersion.} $G$-action on a quasi-projective $S$-scheme $X$, then the $fppf$ quotient $X/G$ is representable by an $S$-scheme.
\item[(ii)]
If $G$ is quasi-projective over $S$ and $H\subseteq G$ is a closed subgroup scheme which is proper and flat, then the $fppf$ quotients $G/H$ and $H\backslash G$ are representable. If $H$ is normal in $G$, then $G/H$ is an $S$-group and the canonical projection $G\to G/H$ is a morphism of $S$-groups.
\end{itemize}
\end{theorem}
Remark that statement (ii) is a consequence of statement (i).
\subsection{Group actions on sheaves}
Assume as before that $G$ is an $S$-group scheme. Let $\mu_X:X\times_S G\to X$ be a right action of $G$ on an $S$-scheme $X$. For a coherent $\mathcal O_X$-module $\mathcal F$ we define a {\em right $G$-action} on $\mathcal F$ (compatible with $\mu_X$) as an isomorphism $\mu_{\mathcal F}:\pi_1^*\mathcal F\to\mu_X^*\mathcal F$, where $\pi_1:X\times_S G\to X$ is the canonical projection, subject to the commutativity of the diagrams (4.4.2) and (4.4.3) in \cite{kashiwara1989} (appropriately modified for right actions). We also say that $\mathcal F$ is a {\em (right) $G$-equivariant sheaf}.
A consequence of faithfully flat descent \cite[Expos\'e VIII, Corollaire 1.3]{sga1} is
\begin{proposition}\label{prop:equivariantsheaves}
Let $G$ be an $S$-group scheme, acting via $\mu_X:X\times_S G\to X$ on an $S$-scheme $X$. Assume that the $fppf$ quotient $X/G$ exists and is representable by an $S$-scheme $Y=X/G$, let $\pi:X\to Y$ denote the canonical projection and $\mathcal G$ be a (quasi-)coherent $\mathcal O_Y$-module.
Then the canonical isomorphism $\mu_{\pi^*\mathcal G}:\pi_1^*\pi^*\mathcal G\to\mu_X^*\pi^*\mathcal G$ defines a right $G$-action on the (quasi-)coherent $\mathcal O_X$-module $\pi^*\mathcal G$.
This correspondence induces an equivalence between the categories of (quasi-)coherent (resp.\ finite locally free) $\mathcal O_Y$-modules $\mathcal G$ and (quasi-)coherent (resp.\ finite locally free) $\mathcal O_X$-modules $\mathcal F$ with right $G$-actions (and $G$-equivariant morphisms of sheaves).
\end{proposition}
Geometrically, we have the following interpretation of a $G$-action on a quasi-coherent $\mathcal O_X$-module $\mathcal F$. If $v_{\mathcal F}:{\mathbb V}(\mathcal F^\vee)\to X$ denotes the \lq{}algebraic\rq{} vector bundle (cf.\ \cite[D\'efinition (1.7.8)]{ega2}) associated to $\mathcal F^\vee$, i.e.\ the $S$-scheme ${\mathbb V}(\mathcal F^\vee)$ over $X$ such that $\mathcal F$ coincides with the sheaf of sections of $v_{\mathcal F}$, a $G$-action $\mu_{\mathcal F}$ on $\mathcal F$ is the same as a $G$-action on ${\mathbb V}(\mathcal F^\vee)$, for which the morphism $v_{\mathcal F}$ is $G$-equivariant, and which is linear in each fibre, i.e.\ for each $S$-scheme $A$, every $g\in G(A)$, and every $x\in X(A)$, the induced isomorphism $g:{\mathbb V}(\mathcal F^\vee)_x\to {\mathbb V}(\mathcal F^\vee)_{xg}$ is linear.
Another correspondence between sheaves and modules arises as follows. Assume $\mathcal M$ is a rational representation of $G$ over $S$, i.e.\ $\mathcal M$ is a (quasi-)coherent sheaf of $\mathcal O_S$-modules, together with a right $\mathcal O_G$-comodule structure
$$
\mathcal M\to \mathcal M\otimes_{\mathcal O_S}\psi_*\mathcal O_G
$$
in the category of sheaves on $S$. As above we may associate to $\mathcal M$ an algebraic vector bundle ${\mathbb V}(\mathcal M)$ over $S$ with a scheme theoretic right action of $G$. Assume that $G$ acts freely on $X$ on the right, and that the $fppf$ quotient $X/G$ exists as a scheme.
If $\pi:X\to X/G$ denotes the canonical projection, we may set for each Zariski open $U\subseteq X/G$,
$$
\mathcal L_{\mathcal M}(U)\;:=\;\{f\in\Hom_S(\pi^{-1}U,{\mathbb V}(\mathcal M))\mid f\;\text{is $G$-equivariant}\}
$$
In the case of a locally noetherian base $S$, a direct computation shows that this is a (quasi-)coherent sheaf of $\mathcal O_{X/G}$-modules, and $\mathcal M\to\mathcal L_{\mathcal M}$ is from exact functor from the category of quasi-coherent $G$-modules on $S$ to the category of quasi-coherent sheaves on $X/G$, which preserves coherence.
\subsection{$G$-equivariant sheaves of differential operators}
As before $G/S$ denotes a group scheme. Write $e:S\to G$ for the unit section. We identify the Lie algebra of $G$ with the $\mathcal O_S$-module
$$
\underline{{\mathfrak {g}}}_S\;=\;e^*\Der_{S}(G).
$$
$G$ acts on $\underline{{\mathfrak {g}}}_S$ via the adjoint representation.
Assume that a smooth $S$-group scheme $G$ acts on $X$ from the left. The differential of the action of $G$ on $X$ induces a morphism
\begin{equation}
\underline{{\mathfrak {g}}}_S\;\to\;x_*\Der_{S}(X)
\label{eq:liegtotangentX}
\end{equation}
of $\mathcal O_S$-Lie algebras. We denote by $U(\underline{{\mathfrak {g}}}_S)$ the sheaf of universal enveloping algebras of $\underline{{\mathfrak {g}}}_S$ on $S$, and set
$$
\underline{U}_X(\underline{{\mathfrak {g}}})\;:=\;x^*U(\underline{{\mathfrak {g}}}_S),
$$
and furnish it with the following ring structure. Let $x=(\psi,\theta)$. Then since $G$ acts on $X$, we have a canonical action of $\psi^*\underline{{\mathfrak {g}}}_S$ on $\mathcal O_X$, which induces a diagonal action of $\psi^*\underline{{\mathfrak {g}}}_S$ on
$$
\underline{U}_X(\underline{{\mathfrak {g}}})\;=\;\mathcal O_X\otimes_{\psi^*\mathcal O_S}\psi^*U(\underline{{\mathfrak {g}}}_S),
$$
which in turn extends to a diagonal action of $\psi^*U(\underline{{\mathfrak {g}}}_S)$. We let $\mathcal O_X$ act on $\underline{U}_X(\underline{{\mathfrak {g}}})$ via the first factor. This induces on $\underline{U}_X(\underline{{\mathfrak {g}}})$ the structure of an $\mathcal O_X$-algebra satisfying the three compatibility conditions (4.2.1)-(4.2.3) in \cite{kashiwara1989}. In particular, this action is compatible with \eqref{eq:liegtotangentX}.
Then any $G$-equivariant $\mathcal O_X$-module $\mathcal M$ comes with a canonical action of $\underline{U}_X(\underline{{\mathfrak {g}}})$.
We say that a twisted sheaf of differential operators $\mathcal A$ on $X/S$ is $G$-equivariant if it is equipped with a $G$-action that satisfies the conditions (4.6.1) and (4.6.2) in \cite{kashiwara1989}. This allows us to extend the notion of $G$-equivariance to $\mathcal A$-modules as in \S4.7 in loc.\ cit.
\section{The case of fields}
In this section we assume that $k\subseteq k'$ is an extension of fields\footnote{or more generally a faithfully flat extension of rings.} (not necessarily of characteristic $0$) and we specialize the previous situation to $S:=\Spec k$ and $S':=\Spec k'$. Then $s:S'\to S$ is fixed as well and is a flat morphism of schemes.
In the sequel $x:X\to S$ is a scheme of finite type, $X'=X\times_S S'$ denotes its base change to $S'$, $y:X'\to X$ the canonical projection. We remark that $X\to S$ is automatically of finite presentation\footnote{since $k$ is a field and in particular noetherian.}, and $X$ is quasi-compact and quasi-separated because $S$ enjoys these properties as a (noetherian\footnote{only necessary for quasi-separatedness}) affine scheme.
\subsection{$\mathcal D$-modules over fields}
We remark that since $S$ is affine, pushforward $x_*$ may be identified with the global sections functor, that we interpret as a left exact functor
$$
\Gamma(X,-):\quad
\Mod_{\rm qc}(\D_{X/S})\to
\Mod(\Gamma(X,\D_{X/S})).
$$
By Corollary \ref{cor:dbasechange} we see\footnote{Recall that $X$ was of finite presentation over an affine scheme and therefore quasi-compact} that we have a canonical isomorphism $y^*\D_{X/S}\cong\D_{X'/S'}$, and since we know that $\D_{X/S}$ is $\mathcal O_X$-quasi-coherent, and every quasi-coherent $\mathcal O_S$-module is flat over $S$, we obtain with the flat base change Theorem \ref{thm:flatbasechange2} a canonical isomorphism
\begin{equation}
\Gamma(X,\D_{X/S})\otimes_k k'\cong \Gamma(X',\D_{X'/S'}).
\label{eq:Dglobalsectionsbasechange}
\end{equation}
In particular, the base change $y^*\mathcal M$ of any $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ carries a natural $\D_{X'/S'}$-module structure. To be more precise, we have
\begin{theorem}\label{thm:Dmodulebasechangek}
Let $s:S'=\Spec k'\to S=\Spec k$ be an extension of fields\footnote{or more generally a faithfully flat extension of rings.}, $x:X\to S$ is an $S$-scheme of finite type, $X'=X\times_S S'$ its base change and $y:X'\to X$ the canonical projection. Then we have a well defined functor\footnote{In general it seems that one needs to consider the composite of $y^*(-)$ with $\D_{X'/S'}\otimes_{y^*\D_{X/S}}(-)$.}
$$
y^*(-):\quad \Mod_{\rm qc}(\D_{X/S})\to\Mod_{\rm qc}(\D_{X'/S'}).
$$
It enjoys the following properties:
\begin{itemize}
\item[(i)] The following diagram of functors commutes (strictly):
$$
\begin{CD}
\Mod_{\rm qc}(\D_{X/S})@>\Gamma(X,-)>>
\Mod(\Gamma(X,\D_{X/S}))\\
@Vy^*VV @VVs^*V\\
\Mod_{\rm qc}(\D_{X'/S'})@>>\Gamma(X',-)>
\Mod(\Gamma(X',\D_{X'/S'}))
\end{CD}
$$
\item[(ii)] $s^*(-)$ and $y^*(-)$ are exact and faithful.
\item[(iii)] A module $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ is $\mathcal O_X$-coherent, if and only if $y^*\mathcal M$ is $\mathcal O_{X'}$-coherent.
\item[(iv)] For every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ and every degree $q\geq 0$ a natural isomorphism
\begin{equation}
H^q(X,\mathcal M)\otimes_k k'\;\cong\;H^q(X',y^*\mathcal M).
\label{eq:cohomologyiso}
\end{equation}
\item[(v)] For every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ and every degree $q\geq 0$ we have an equivalence
\begin{equation}
H^q(X,\mathcal M)\;=\;0\quad\Leftrightarrow\quad H^q(X',y^*\mathcal M)\;=\;0.
\label{eq:vanishingequivalence}
\end{equation}
\end{itemize}
\end{theorem}
\begin{proof}
We already saw that $y^*$ is well defined and that the diagram in (i) commutes. Since $s:S'\to S$ is faithfully flat, $y:X'\to X$ is faithfully flat, and therefore $y^*$ is exact and faithful, this shows (ii) and (iii). By the flat base change Theorem \ref{thm:flatbasechange2} (iv) follows. Statement (v) is a consequence of (iv) and the faithfulness of $S'\to S$.
\end{proof}
\begin{proposition}\label{prop:daffinegeometric}
For any scheme $X$ of finite type over a field $k$, being $\mathcal D$-affine is a {\em geometric property} of $X$, in the sense that $X$ is $\mathcal D$-affine, if and only if $X'$ is $\mathcal D$-affine for one (or for all) field extensions $k'/k$.
\end{proposition}
\begin{proof}
By Theorem \ref{thm:Dmodulebasechangek}, (v), condition (i) for being $\mathcal D$-affine is independent of the base field $k$. By statement (iv) of the same Proposition the vanishing of global sections in condition (ii) is independent of the base field, and since field extensions yield faithfully flat morphisms of schemes, the vanishing condition on the left hand side is also independent of the base field.
\end{proof}
\begin{proposition}
Let $G$ be a linear reductive group over a field $k$ of characteristic $0$, and $\mathcal B$ its flag variety of Borel subalgebras of $G$, considered as $k$-scheme. Then $\mathcal B$ is $\mathcal D$-affine over $k.$
\end{proposition}
\begin{proof}
By our previous discussion this reduces to the case of algebraically closed $k$, where the statement is known.
\end{proof}
We recall that for $X/S$ smooth, $\D_{X/S}$ is a left and right noetherian module by Corollary \ref{cor:Dnoetherian}.
A special case of Propositions \ref{prop:firstequivalenceovergeneralbase} and \ref{prop:coherentequivalenceovergeneralbase} is
\begin{proposition}\label{prop:firstequivalenceoverfields}
Assume that $X$ is a $\mathcal D$-affine scheme over a field $S=\Spec k$. Then
\begin{itemize}
\item[(i)] Every $\mathcal M\in\Mod_{\rm qc}(\D_{X/S})$ is generated by its global sections as $\D_{X/S}$-module.
\item[(ii)] The functor
$$
\Gamma(X,-):\quad
\Mod_{\rm qc}(\D_{X/S})\to
\Mod(\Gamma(X,\D_{X/S}))
$$
is an equivalence of categories.
\item[(iii)] If $X/S$ is smooth, then $\Gamma(X,-)$ induces an equivalence between $\D_{X/S}$-coherent modules and finitely generated $\Gamma(X,\D_{X/S})$-modules.
\end{itemize}
\end{proposition}
Proposition \ref{prop:firstequivalenceoverfields} provides another proof for the statement of Corollary \ref{cor:injectiveobjects} in the particular case at hand.
\addcontentsline{toc}{section}{References}
\bibliographystyle{plain}
|
\section{Introduction}
Kirchdoerfer and Ortiz \cite{Kirchdoerfer:2016, Kirchdoerfer:2017, Kirchdoerfer:2018} and Conti {\sl et al.} \cite{Conti:2018} have recently proposed a new class of problems in static and dynamic elasticity, referred to as {\sl Data-Driven problems}, defined on the space of strain-stress field pairs, or phase space. The problems consist of minimizing the distance between a given material data set and the subspace of compatible strain fields and stress fields in equilibrium. They find that the classical solutions are recovered in the case of linear elasticity and identify conditions for convergence of Data-Driven solutions corresponding to sequences of material data sets. Data-Driven elasticity effectively reformulates the classical initial-boundary-value problem of elasticity directly from material data, thus bypassing the empirical material modelling step altogether. By eschewing empirical models, material modelling empiricism, modelling error and uncertainty are eliminated entirely and no loss of experimental information is incurred.
It should be noted that the use of material data as a basis for constitutive modeling is classical and remains the subject of extensive ongoing research. There is a vast body of literature devoted to that subject, including recent developments based on statistical learning, model and data reduction, nonlinear regression, and others, which would be too lengthy to enumerate here. It bears emphasis, that what sets the present approach apart from these other approaches is that we reformulate the classical boundary value problems of mechanics, including inelasticity and approximations thereof, directly on the basis of the material data, without any attempt at modeling the data or performing any form of data reduction or manipulation.
A natural extension of the Data-Driven paradigm concerns inelastic materials whose response is irreversible and history dependent. The theory of {\sl materials with memory} furnishes the most general representation of such materials. According to Rivlin \cite{Rivlin:1972}:
\begin{quotation}
{\sl "The characteristic property of inelastic solids which distinguishes them from elastic solids is the fact that the stress measured at time t depends not only on the instantaneous value of the deformation but also on the entire history of deformation."}
\end{quotation}
The origins of the theory may be traced to a series of papers by Green and Rivlin starting in 1957 \cite{Green:1957, Green:1959, Green:1960}, who proposed the use of hereditary constitutive laws, originally developed by Boltzman \cite{Boltzmann:1874} and Volterra \cite{Volterra:1909} in the linear case, for the description of non-linear viscoelastic materials as an alternative to models using constitutive equations of the rate type \cite{Rivlin:1955}. The hereditary functional approach to inelasticity was introduced into thermodynamics by Coleman \cite{Coleman:1964a}. A linearization of Green and Rivlin's theory was developed by Pipkin and Rivlin \cite{Pipkin:1961}. Rheological properties of solids often have a {\sl fading memory property}, enunciated by Truesdell \cite{Truesdell:1965} as:
\begin{quotation}
{\sl "Events which occurred in the distant past have less influence in determining the present response than those which occurred in the recent past"}.
\end{quotation}
The concept of fading memory was formalized by Coleman and Noll \cite{Coleman:1961a, Coleman:1961b} as a continuity property of the hereditary functional and subsequently extended by Wang \cite{Wang:1965a, Wang:1965b}, Perzyna \cite{Perzyna:1967} and others.
Other general representations of inelasticity are based on continuum thermodynamics with internal variables (cf., e.~g., \cite{Rice:1975}). These representations replace an explicit dependence on history by a dependence on the effects of history, i.~e., the current microstructure of the material element. The variables used to describe that microstructure are referred to as internal variables. Together with the state of stress or deformation and a thermodynamic variable such as temperature or entropy, they define the local state of a material element. Such models were introduced for viscoelastic deformation by Eckart \cite{Eckart:1948}, Meixner \cite{Meixner:1953}, Biot \cite{Biot:1954} and Ziegler \cite{Ziegler:1958}, and have been extensively studied since \cite{Coleman:1963, Schapery:1964, Valanis:1966, Coleman:1967, Lubliner:1972, Lubliner:1973}. The foundations underlying the memory-functional and the internal-variable formalisms were critically reviewed by Kestin and Rice \cite{Kestin:1970}. The correspondence and, in some cases, equivalence between the material-with-memory, internal variable and differential formulations of inelasticity have also been extensively investigated \cite{Coleman:1967, Valanis:1967, Lubliner:1969, Lubliner:1973}.
In the context of Data-Driven inelasticity, the representational paradigms just outlined translate into corresponding representational paradigms for the material data set. Specifically, we identify the material data set $D(t)$ at time $t$ with the collection of stress-strain pairs $(\epsilon(t),\sigma(t))$ that are attainable by the material at that time. For inelastic materials, $D(t)$ depends on the past history $\{(\epsilon(s),\sigma(s))\}_{s < t}$ of stress and strain. The central issue of Data-Driven inelasticity thus concerns the formulation of rigorous yet practical representational paradigms for the evolving material data set. The practicality of the representation revolves around the amount of data that needs to be carried, or generated, along with the calculations. By rigorous we specifically mean representations that result, albeit at increasing computational cost, in convergent approximations.
We specifically consider three representational paradigms: i) {\sl materials with memory}, i.~e., conditioning the material data set to the past history of deformation; ii) {\sl differential materials}, i.~e., conditioning the material data set to short histories of stress and strain; and iii) {\sl history variables}, i.~e., conditioning the material data set to {\sl ad hoc} variables encoding partial information about the history of stress and strain. We also consider combinations of the three paradigms thereof and investigate their ability to represent the evolving data sets of different classes of inelastic materials, including viscoelasticity, viscoplasticity and plasticity. The resulting Data-Driven inelasticity problems then consist of minimizing distance in phase space between the evolving data set and a time-dependent constraint set. We additionally concern ourselves with the numerical implementation and convergence characteristics of the resulting Data-Driven schemes.
We structure the paper as follows. In Section~\ref{L4cpYd} we succinctly summarize the Data-Driven approach to elasticity by way of background and in order to set essential notation. Extensions to inelasticity predicated on various representations of the material data set are put forth and developed in Section~\ref{4qR2FP}. In Section~\ref{IXEV5B} we present selected examples of application to viscoelastic solids that demonstrate the suitability of differential representations of the material data set and the performance of the resulting Data-Driven schemes. Further examples of application are presented in Section~\ref{X0aZjS} that demonstrate how hybrid differential/history variable representations of the material data set can be used to account for hardening plasticity. Finally, an extended discussion of possible extensions and alternative approaches is presented in Section~\ref{y1HJx6}.
\section{Background: Data-Driven elasticity}
\label{L4cpYd}
We begin by recalling the Data-Driven reformulation of elasticity \cite{Kirchdoerfer:2016, Kirchdoerfer:2017} as a basis for subsequent generalizations to inelasticity. For simplicity, we consider discrete, or discretized, systems consisting of $N$ nodes and $M$ material points. The system undergoes displacements $\bu = \{\bu_a\}_{a=1}^N$, with $\bu_a \in \mathbb{R}^{n_a}$ and $n_a$ the dimension of the displacement at node $a$, under the action of applied forces $\fb = \{\fb_a\}_{a=1}^N$, with $\fb_a \in \mathbb{R}^{n_a}$. The internal state of the system is characterized by local stress and strain pairs $\{(\mbs{\epsilon}_e, \mbs{\sigma}_e)\}_{e=1}^M$, with $\mbs{\epsilon}_e, \mbs{\sigma}_e \in \mathbb{R}^{m_e}$ and $m_e$ the dimension of stress and strain at material point $e$. We regard $\bz_e = (\mbs{\epsilon}_e, \mbs{\sigma}_e)$ as a point in a local phase space $Z_e = \mathbb{R}^{m_e} \times \mathbb{R}^{m_e}$ and $\bz = \{(\mbs{\epsilon}_e, \mbs{\sigma}_e)\}_{e=1}^M$ as a point in the global phase space $Z = Z_1 \times \cdots \times Z_M$.
The internal state of the system is subject to the compatibility and equilibrium constraints of the general form
\begin{subequations}\label{XUI9Eu}
\begin{align}
& \label{Bd2bYR}
\mbs{\epsilon}_e = \bB_e \bu ,
\quad e=1,\dots,M,
\\ & \label{j8dCjM}
\sum_{e=1}^M w_e \bB_e^T \mbs{\sigma}_e = \fb ,
\end{align}
\end{subequations}
where $\{w_e\}_{e=1}^M$ are elements of volume and $\bB_e$ is a discrete strain operator for material point $e$. We note that constraints (\ref{XUI9Eu}) are universal, or material-independent. They define a subspace, or constraint set,
\begin{equation}\label{DWdog7}
E = \{ \bz \in Z \, : \, (\ref{Bd2bYR}) \text{ and } (\ref{j8dCjM}) \} ,
\end{equation}
consisting of all compatible and equilibrated internal states. In (\ref{DWdog7}) and subsequently, the symbol $:$ is used to mean 'given' or 'subject to' or 'conditioned to'. Within this subspace, the internal state satisfies the power identity
\begin{equation}
\fb \cdot \bu
=
\sum_{e=1}^M w_e \, \mbs{\sigma}_e \cdot \mbs{\epsilon}_e .
\end{equation}
In classical elasticity, the problem (\ref{XUI9Eu}) is closed by appending local material laws, e.~g., functions of the general form
\begin{equation}\label{d6cdNV}
\mbs{\sigma}_e = \hat{\mbs{\sigma}}_e(\mbs{\epsilon}_e) ,
\quad
e = 1,\dots, m ,
\end{equation}
where $\hat{\mbs{\sigma}}_e : \mathbb{R}^{m_e} \to \mathbb{R}^{m_e}$. However, often material behavior is only known through a material data set $D_e$ of points $\bz_e = (\mbs{\epsilon}_e,\mbs{\sigma}_e) \in Z_e$ obtained experimentally or by some other means. Again, the conventional response to this situation is to deduce a material law $\hat{\mbs{\sigma}}_e$ from the data set $D_e$ by some appropriate means, thus reverting to the classical setting (\ref{d6cdNV}).
The Data-Driven reformulation of the classical problems of mechanics consists of formulating boundary-value problems directly in terms of the material data, thus entirely bypassing the material modeling step altogether \cite{Kirchdoerfer:2016}. A class of Data-Driven problems consists of finding the compatible and equilibrated internal state $\bz \in E$ that minimizes the distance to the global material data set $D = D_1 \times \cdots \times D_M$. To this end, we metrize the local phase spaces $Z_e$ by means of norms of the form
\begin{equation}\label{fou9Oe}
| \bz_e |_e
=
\left(
\mathbb{C}_e \mbs{\epsilon}_e \cdot \mbs{\epsilon}_e
+
\mathbb{C}_e^{-1} \mbs{\sigma}_e \cdot \mbs{\sigma}_e
\right)^{1/2} ,
\end{equation}
for some symmetric and positive-definite matrices $\{\mathbb{C}_e\}_{e=1}^M$, with corresponding distance
\begin{equation}
d_e(\bz_e,{\mbs{y}}} \def\bz{{\mbs{z}}_e) = | \bz_e - {\mbs{y}}} \def\bz{{\mbs{z}}_e |_e ,
\end{equation}
for ${\mbs{y}}} \def\bz{{\mbs{z}}_e, \bz_e \in Z_e$. The local norms induce a metrization of the global phase $Z$ by means of the global norm
\begin{equation}\label{9oakLa}
| \bz |
=
\Big( \sum_{e=1}^m w_e | \bz_e |_e^2 \Big)^{1/2} ,
\end{equation}
with associated distance
\begin{equation}
d(\bz,{\mbs{y}}} \def\bz{{\mbs{z}}) = |\bz - {\mbs{y}}} \def\bz{{\mbs{z}}|,
\end{equation}
for ${\mbs{y}}} \def\bz{{\mbs{z}}, \bz \in Z$. The distance-minimizing Data-Driven problem is, then,
\begin{equation}\label{K5b6z0}
\min_{{\mbs{y}}} \def\bz{{\mbs{z}} \in D} \min_{\bz \in E} d(\bz,{\mbs{y}}} \def\bz{{\mbs{z}})
=
\min_{\bz \in E} \min_{{\mbs{y}}} \def\bz{{\mbs{z}} \in D} d(\bz,{\mbs{y}}} \def\bz{{\mbs{z}}) ,
\end{equation}
i.~e., we wish to find the point ${\mbs{y}}} \def\bz{{\mbs{z}} \in D$ in the material data set that is closest to the constraint set $E$ of compatible and equilibrated internal states or, equivalently, we wish to find the compatible and equilibrated internal state $\bz \in E$ that is closest to the material data set $D$.
We emphasize that the local material data sets can be graphs, point sets, 'fat sets' and ranges, or any other arbitrary set in phase space. Evidently, the classical problem is recovered if the local material data sets are chosen as
\begin{equation}
D_e
=
\{
(\mbs{\epsilon}_e,\hat{\mbs{\sigma}}_e(\mbs{\epsilon}_e))
\} ,
\end{equation}
i.~e., as {\sl graphs} in $Z_e$ defined by the material law (\ref{d6cdNV}). Thus, the Data-Driven reformulation (\ref{K5b6z0}) extends---and subsumes as special cases---the classical problems of mechanics.
We note that, for fixed ${\mbs{y}}} \def\bz{{\mbs{z}} \in D$, the closest point projection $\bz = P_E {\mbs{y}}} \def\bz{{\mbs{z}}$ onto $E$ follows by minimizing the quadratic function $d^2(\cdot,{\mbs{y}}} \def\bz{{\mbs{z}})$ subject to the constraints (\ref{XUI9Eu}). The compatibility constraint (\ref{Bd2bYR}) can be enforced directly by introducing a displacement field $\bu$. The equilibrium constraint (\ref{j8dCjM}) can then be enforced by means of Lagrange multipliers $\mbs{\lambda}$ representing virtual displacements of the system. With ${\mbs{y}}} \def\bz{{\mbs{z}} \equiv \{(\mbs{\epsilon}'_e, \mbs{\sigma}'_e)\}_{e=1}^M$ given, e.~g., from a previous iteration, the corresponding Euler-Lagrange equations are \cite{Kirchdoerfer:2016}
\begin{subequations}
\begin{align}
&
\Big(\sum_{e=1}^M w_e \bB_e^T \mathbb{C}_e \bB_e \Big) \bu
=
\sum_{e=1}^M w_e \bB_e^T \mathbb{C}_e \mbs{\epsilon}'_e ,
\\ &
\Big(\sum_{e=1}^M w_e \bB_e^T \mathbb{C}_e \bB_e \Big) \mbs{\lambda}
=
\fb - \sum_{e=1}^M w_e \bB_e^T \mbs{\sigma}'_e ,
\end{align}
\end{subequations}
which define two standard linear displacement problems. The closest point $\bz = P_E {\mbs{y}}} \def\bz{{\mbs{z}} \in E$ then follows as
\begin{subequations}
\begin{align}
&
\mbs{\epsilon}_e = \bB_e \bu ,
\quad
e = 1,\dots,M ,
\\ &
\mbs{\sigma}_e = \mbs{\sigma}'_e + \mathbb{C}_e \bB_e \mbs{\lambda} ,
\quad
e = 1,\dots,M .
\end{align}
\end{subequations}
A simple Data-Driven solver consists of the fixed point iteration \cite{Kirchdoerfer:2016}
\begin{equation}\label{K1TK7z}
\bz_{j+1} = P_E P_D \bz_j ,
\end{equation}
for $j = 0,1,\dots$ and $\bz_0 \in Z$ arbitrary, where $P_D$ denotes the closest point projection in $Z$ onto $D$. Iteration (\ref{K1TK7z}) first finds the closest point $P_D \bz_j$ to $\bz_j$ on the material data set $D$ and then projects the result back to the constraint set $E$. The iteration is repeated until $P_D \bz_{j+1} = P_D \bz_j$, i.~e., until the data association to points in the material data set remains unchanged.
The convergence properties of the fixed-point solver (\ref{K1TK7z}) have been investigated in \cite{Kirchdoerfer:2016}. The Data-Driven paradigm has been extended to dynamics \cite{Kirchdoerfer:2018}, finite kinematics \cite{Nguyen:2018} and objective functions other than phase-space distance can be found in \cite{Kirchdoerfer:2017}. The well-posedness of Data-Driven problems and properties of convergence with respect to the data set have been investigated in \cite{Conti:2018}.
\section{Extension to inelasticity}
\label{4qR2FP}
A natural extension of the Data-Driven paradigm just described concerns inelastic materials whose response is irreversible and history dependent. The equilibrium boundary-value problem for these materials is, therefore, time dependent. For simplicity, we restrict attention to time-discrete formulations and seek to approximate solutions at times $t_0$, $t_1$, $\dots$, $t_k$, $t_{k+1}$, $\dots$. In this setting, the compatibility and equilibrium constraints (\ref{XUI9Eu}) become
\begin{subequations}\label{QB7839}
\begin{align}
& \label{OJiD0T}
\mbs{\epsilon}_{e,k+1} = \bB_e \bu_{k+1} ,
\quad e=1,\dots,M,
\\ & \label{R2Ku3B}
\sum_{e=1}^M w_e \bB_e^T \mbs{\sigma}_{e,k+1} = \fb_{k+1} ,
\end{align}
\end{subequations}
where $\bu_{k+1}$, $\fb_{k+1}$, $\mbs{\epsilon}_{k+1}$ and $\mbs{\sigma}_{k+1}$ are the displacements, forces, strains and stresses at time $t_{k+1}$, respectively. The constraints (\ref{QB7839}) define the constraint set
\begin{equation}
E_{k+1} = \{ \bz \in Z \, : \, (\ref{OJiD0T}) \text{ and } (\ref{R2Ku3B}) \} ,
\end{equation}
which is now time-dependent on account of the time-dependency of the applied loads.
In addition, the instantaneous response of inelastic materials is characterized by its dependence on the past history of deformation. By virtue of this history dependence, the set of stress-strain pairs attainable at a material point depends itself on time. We specifically define the instantaneous local material data set as
\begin{equation}\label{PS6guZ}
D_{e,k+1}
=
\{
(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1}) \, : \,
\text{past local history}
\} ,
\end{equation}
i.~e., the set of local stress-strain pairs $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ attainable at time $t_{k+1}$ at material point $e$ given the past history of the material point. We additionally define a global material data set at time $t_{k+1}$ as $D_{k+1} = D_{1,k+1}$ $\times$ $\cdots$ $\times$ $D_{M,k+1}$.
With these definitions, the Data-Driven problem of inelasticity is
\begin{equation}\label{yKEW6b}
\min_{{\mbs{y}}} \def\bz{{\mbs{z}} \in D_{k+1}} \min_{\bz \in E_{k+1}} d(\bz_{k+1},{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1})
=
\min_{\bz \in E_{k+1}} \min_{{\mbs{y}}} \def\bz{{\mbs{z}} \in D_{k+1}} d(\bz_{k+1},{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}) ,
\end{equation}
i.~e., we wish to find the point ${\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}$ in the material data set $D_{k+1}$ at time $t_{k+1}$ that is closest to the constraint set $E_{k+1}$ at time $t_{k+1}$ or, equivalently, we wish to find the internal state $\bz_{k+1}$ in the constraint set $E_{k+1}$ at time $t_{k+1}$ that is closest to the material data set $D_{k+1}$ at time $t_{k+1}$. Evidently, the inelastic Data-Driven problem (\ref{yKEW6b}) represents a natural extension of the elasticity Data-Driven problem (\ref{K5b6z0}) in which both the constraint set and the material data set are a function of time.
The central challenge now is to formulate rigorous yet practical means of characterizing the history dependence of the local material data sets $D_{e,k+1}$, eq.~(\ref{PS6guZ}). As noted in the introduction, inelastic material behavior can alternatively be described by means of hereditary laws, within the general framework of materials with memory, rheological and thermodynamical models based on internal variables, by means of so-called differential models and by other means. These constitutive formulations give rise to corresponding representational paradigms in the context of Data-Driven inelasticity, which we elucidate next.
\subsection{General materials with memory}
A general material with memory is a material whose state of stress is a function of the past history of strain, i.~e.,
\begin{equation}\label{9qwUSn}
\mbs{\sigma}_e(t)
=
\hat{\mbs{\sigma}}_e(\{\mbs{\epsilon}_e(s)\}_{s \leq t}) ,
\end{equation}
where $\mbs{\sigma}_{e}(t)$ is the stress at material point $e$ and time $t$, $\{\mbs{\epsilon}_e(s)\}_{s \leq t}$ is the corresponding history of strain prior to $t$ and $\hat{\mbs{\sigma}}_e$ is a hereditary functional. For linear rheological materials, $\hat{\mbs{\sigma}}_e$ takes the form of a hereditary or Duhamel integral expressed in terms of a relaxation kernel \cite{Flugge:1975}.
In a discrete setting, (\ref{9qwUSn}) can be approximated as
\begin{equation}\label{mKGT32}
\mbs{\sigma}_{e,k+1}
=
\hat{\mbs{\sigma}}_e(\{\mbs{\epsilon}_{e,l}\}_{l\leq k+1})
\end{equation}
where $\mbs{\sigma}_{e,k+1}$ is the stress at material point $e$ at time $t_{k+1}$, $\{\mbs{\epsilon}_{e,l}\}_{l\leq k+1}$ is the strain history of material point $e$ up to time $t_{k+1}$ and $\hat{\mbs{\sigma}}_e$ is a discrete hereditary function. In this representation, the local material data sets (\ref{PS6guZ}) take the form
\begin{equation}\label{2oKN63}
D_{e,k+1}
=
\{
(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1}) \, : \,
\{\mbs{\epsilon}_{e,l}\}_{l\leq k}
\} ,
\end{equation}
i.~e., they consist of pairs $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ of stress and strain known to be attainable at time $t_{k+1}$ given the past history $\{\mbs{\epsilon}_{e,l}\}_{l\leq k}$. In particular, we note that the material data set at time $t_{k+1}$ depends on the entire history of strain up to and including time $t_k$.
As noted in the introduction, materials often exhibit a fading memory property whereby their instantaneous behavior is a function primarily of the recent state history and is relatively insensitive to the distant past history. Examples include viscoelastic materials exhibiting relaxation and bounded creep. For those materials, the strain history in (\ref{mKGT32}) can be truncated beyond a certain decay time, which simplifies the parametrization of the local material data sets $D_{e,k+1}$. These simplifications notwithstanding, keeping track of long deformation histories, and sampling material behavior conditioned to them, may be challenging and onerous even for materials with fading memory.
\subsection{Internal variable formalism}
Thermodynamic models based on internal variables are often used to characterize inelasticity and history dependence. In these models, the state at a material point $e$ is described in terms of, e.~g., its strain, temperature and an additional array of auxiliary variables $\bq_e$ variables, or internal variables. Thermal processes are beyond the scope of this paper and we shall omit explicit reference to temperature and other thermodynamic variables for simplicity.
In order to describe the behavior of the material, we may assume a Helmholtz free energy $F_e(\mbs{\epsilon}_e, \bq_e)$, with corresponding equilibrium relations
\begin{subequations}\label{O4cXId}
\begin{align}
& \label{e36oQ2}
\mbs{\sigma}_e(t)
=
D_1F_e(\mbs{\epsilon}_e(t), \bq_e(t)) ,
\\ & \label{e33BDo}
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t)
=
-
D_2F_e(\mbs{\epsilon}_e(t), \bq_e(t)) ,
\end{align}
\end{subequations}
where ${\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e$ are thermodynamic driving forces conjugate to $\bq_e$ and $D_1F_e$ and $D_2F_e$ denote the derivatives of $F_e$ with respect to strain and internal variables, respectively. In addition, the evolution of the internal variables is governed by kinetic relations of the form
\begin{equation}\label{Lg7W5y}
D\psi_e(\dot{\bq}_e(t))
+
D_2F_e(\mbs{\epsilon}_e(t), \bq_e(t))
=
\mbs{0} ,
\end{equation}
where $\psi_e$ is a dissipation function and $D\psi_e$ its derivative.
In a time-discrete setting, the evolution of the internal variables is governed by incremental kinetic relations, e.~g., of the form \cite{Ortiz:1999}
\begin{equation}\label{ZnF8G1}
D\psi_e(\frac{\bq_{e,k+1}-\bq_{e,k}}{t_{k+1}-t_k})
+
D_2F_e(\mbs{\epsilon}_{e,k+1}, \bq_{e,k+1})
=
\mbs{0} ,
\end{equation}
and the stress-strain relations (\ref{e36oQ2}) specialize to
\begin{equation}\label{0BDNIT}
\mbs{\sigma}_{e,k+1}
=
D_1F_e(\mbs{\epsilon}_{e,k+1}, \bq_{e,k+1}) .
\end{equation}
Eqs.~(\ref{ZnF8G1}) and (\ref{0BDNIT}) define a close system of equations that can be solved for $\bq_{e,k+1}$ and $\mbs{\sigma}_{e,k+1}$ given $\mbs{\epsilon}_{e,k+1}$ and $\bq_{e,k}$. The corresponding material data set admits the representation
\begin{equation}\label{7cY9he}
D_{e,k+1}
=
\{
(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1}) \, : \,
\bq_{e,k}
\} ,
\end{equation}
i.~e., $D_{e,k+1}$ is the set of all stress and strain pairs $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ accessible to the material given the prior internal state $\bq_{e,k}$.
\subsection{Relation between the internal variable and hereditary representations}
\label{43VIZg}
The internal variable formalism, eqs.~(\ref{O4cXId}) and (\ref{Lg7W5y}), may be regarded as a convenient device for defining hereditary laws of the form (\ref{9qwUSn}). Thus, let \begin{equation}
\bq_e(t)
=
\hat{\bq}_e(\{\mbs{\epsilon}_e(s)\}_{s \leq t})
\end{equation}
denote the solution of (\ref{Lg7W5y}), regarded as a system of ordinary differential equations in $\bq_e(t)$. Inserting into (\ref{e36oQ2}), we obtain the hereditary law
\begin{equation}
\mbs{\sigma}_e(t)
=
D_1F_e(\mbs{\epsilon}_e(t), \hat{\bq}_e(\{\mbs{\epsilon}_e(s)\}_{s \leq t})) ,
\end{equation}
which, evidently, is a particular case of (\ref{9qwUSn}).
In the time-discrete setting, the internal variable formalism, eqs.~(\ref{ZnF8G1}) and (\ref{0BDNIT}), may also be regarded as a means of defining time-discrete hereditary laws of the form (\ref{mKGT32}). Thus, solving (\ref{ZnF8G1}) for the interval variables gives a relation
\begin{equation}\label{iQG1K4}
\bq_{e,k+1}
=
{\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e(\bq_{e,k}, \mbs{\epsilon}_{e,k+1}) ,
\end{equation}
where ${\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e$ plays the role of a {\sl propagator}. Inserting into (\ref{0BDNIT}), we further obtain the stress-strain relation
\begin{equation}\label{fYFK2S}
\mbs{\sigma}_{e,k+1}
=
D_1F_e(\mbs{\epsilon}_{e,k+1}, {\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e(\bq_{e,k}, \mbs{\epsilon}_{e,k+1})) ,
\end{equation}
conditioned to the prior internal state $\bq_{e,k}$. Iterating this relation, we obtain
\begin{equation}
\begin{split}
\mbs{\sigma}_{e,k+1}
& =
D_1F_e
(
\mbs{\epsilon}_{e,k+1},
{\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e({\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e({\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e(\cdots, \mbs{\epsilon}_{e,k-1}), \mbs{\epsilon}_{e,k}), \mbs{\epsilon}_{k+1})
)
\\ & \equiv
\hat{\mbs{\sigma}}_e(\{\mbs{\epsilon}_{e,l}\}_{l\leq k+1}) ,
\end{split}
\end{equation}
which defines a discrete hereditary law of the form (\ref{mKGT32}) for the stresses as a function of the past history of strain. However, instead of the general history parametrization (\ref{2oKN63}) the material data set now admits the more explicit representation (\ref{7cY9he}), which greatly reduces the complexity of the parametrization of the material data set relative to that based on a general hereditary framework.
\subsection{History variables}
Despite its appeal, the essential conceptual drawback of the internal variable formalism is that the internal variable set is often not known or is the result of modeling assumptions. The efficiency of the internal variable parametrization can be retained, while eschewing {\sl ad hoc} modeling assumptions, simply by reinterpreting internal variables as history variables. Contrary to internal variables, history variables need not have a specific physical meaning and their function is simply to record partial information about the history of the material.
By way of motivation, we may iterate the update (\ref{iQG1K4}) to obtain the relation
\begin{equation}
\bq_{e,k}
=
{\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e({\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e({\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e(\cdots, \mbs{\epsilon}_{e,k-2}), \mbs{\epsilon}_{e,k-1}), \mbs{\epsilon}_{k})
\equiv
\hat{\bq}_e(\{\mbs{\epsilon}_{e,l}\}_{l\leq k}) ,
\end{equation}
which gives the internal variables at $t_{k}$ as a function of the strain history up to and including $t_k$. More generally, we may consider history variables of the form
\begin{equation}\label{RTL6EL}
\bq_{e,k}
=
\hat{\bq}_e
(
\{\mbs{\epsilon}_{e,l}\}_{l\leq k} ,
\{\mbs{\sigma}_{e,l}\}_{l\leq k}
) ,
\end{equation}
i.~e., functions of the stress and strain histories up to and including $t_k$. Implicit in the internal variable framework is that the current material data set $D_{e,k+1}$ depends on the deformation history only through a reduced set of history-dependent internal variables $\bq_{e,k}$, eq.~(\ref{7cY9he}).
The paradigm shift now consists of regarding the variables $\bq_{e,k}$ not as physical variables but as {\sl ad hoc} history variables that record and store partial information about the past internal history of the material point. Thus, the history variables $\bq_{e,k}$ at time $t_k$ are the result of applying {\sl ad hoc} history functionals $\hat{\bq}_e$ to the prior history of stress and strain. The history functionals query that history and extract and record selected information. The history information is then used to condition and parametrize the material data sets as in (\ref{7cY9he}). However, in the new reinterpretation (\ref{7cY9he}) represents the set of all known stress and strain pairs $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ consistent with all past stress and strain histories for which the chosen history functionals $\hat{\bq}_e$ evaluate to $\bq_{e,k}$.
Importantly, the choice of history variables is no longer a matter of material modeling, as is the case for internal variables, but a question of approximation theory. Specifically, the aim is to produce sequences of history functionals that constrain arbitrary histories of stress and strain increasingly tightly, and exactly in the limit.
In particular, the sequence of Data-Driven solutions constrained by an increasing number of history variables should converge to the exact Data-Driven solution corresponding to (\ref{2oKN63}). In practice, the central representational challenge is to characterize general material histories to arbitrary accuracy with as few history variables as possible.
\subsection{Differential representations}
Differential models of inelasticity (cf., e.~g., \cite{Flugge:1975}) offer the advantage of reducing history dependence to short histories of stress and strain. Differential materials are characterized by a differential constraint of the form
\begin{equation}\label{k8exoU}
\mbs{f}_e
(
\{\mbs{\epsilon}_e^{(\alpha)}(t)\}_{\alpha=0}^p, \{\mbs{\sigma}_e^{(\beta)}(t)\}_{\beta=0}^q
)
=
\mbs{0} ,
\end{equation}
between the strain and its first $p$ time derivatives and stress and its first $q$ derivatives, for some material-specific function $\fb_e$ taking values in $\mathbb{R}^{m_e}$. In a time-discrete setting, the time derivatives are replaced by divided-difference formulas of the form
\begin{equation}\label{X21vSf}
\mbs{z}_{e,k+1}^{(\alpha)}
=
\sum_{l=0}^\alpha
c_{k+1,\alpha,l} \, \mbs{z}_{e,k+1-l} ,
\end{equation}
for some coefficients $\{c_{k+1,\alpha,l}\}_{l=0}^\alpha$ dependent on the choice of discrete times $\{t_{k+1-l}\}_{l=0}^\alpha$. For constant time step,
\begin{equation}
\mbs{z}_{e,k+1}^{(\alpha)}
=
\frac{1}{\Delta t^\alpha}
\sum_{l=0}^\alpha
(-1)^l
{\alpha \choose l}
\mbs{z}_{e,k+1-l} ,
\end{equation}
with coefficients independent of $k+1$ as expected. Inserting these formulas into (\ref{k8exoU}), we obtain a relation of the form
\begin{equation}\label{fYCd00}
\mbs{f}_{e}
(
\{\mbs{\epsilon}_{e,k+1-l}\}_{l=0}^p,
\{\mbs{\sigma}_{e,k+1-l}\}_{l=0}^q
)
=
\mbs{0} ,
\end{equation}
between the short histories of strain of length $p$ and short histories of stress of length $q$.
In this representation, the local material data sets (\ref{PS6guZ}) take the form
\begin{equation}\label{32DiI8}
D_{e,k+1}
=
\{
(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1}) \, : \,
(
\{\mbs{\epsilon}_{e,k-l}\}_{l=0}^{p-1},
\{\mbs{\sigma}_{e,k-l}\}_{l=0}^{q-1}
)
\} ,
\end{equation}
i.~e., consist of all pairs $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ of stress and strain at time $t_{k+1}$ that are attainable, or known to be attainable, to the material element given the past short histories of stress and strain $(\{\mbs{\epsilon}_{e,k-l}\}_{l=0}^{p-1}, \{\mbs{\sigma}_{e,k-l}\}_{l=0}^{q-1})$ .
We note from (\ref{32DiI8}) that, for differential models, the material data set (\ref{32DiI8}) indeed depends on history through short histories of stress and strain. This parametrization is in contrast with that obtained from general representations of materials with memory, eq.~(\ref{2oKN63}), in which the history dependence of the material data set is parameterized in terms of entire, or long, histories of strain only. We thus conclude that conditioning of material data sets by means of both stress and strain histories may result in smaller parameterizations than otherwise required when only strain histories are accounted for. It may also be reasonably expected that increasing the order of differential representations (\ref{32DiI8}) should lead to increasingly accurate, and in the limit exact, representations of broad classes of materials.
\subsection{Equivalence between the internal variable and differential formalisms}
\label{YF98LU}
The correspondence between the internal variable and differential formalisms can be established as follows. For simplicity, we specifically assume internal variables of the form $\bq = \{\bq_1, \dots, \bq_N\}$, with $\bq_i \in \mathbb{R}^{m_e}$. This assumption sets the tensorial character of the internal variables to be that of a collection of internal strains. Begin by writing (\ref{e36oQ2}) as
\begin{equation}\label{BJYKd4}
\mbs{\sigma}_e(t)
=
\fb_0(\mbs{\epsilon}_e(t), \bq_e(t)) .
\end{equation}
Assuming sufficient differentiability, we can differentiate this relation with respect to time and combine the result with the kinetic relations (\ref{Lg7W5y}) to obtain the identity
\begin{equation}
\begin{split}
\dot{\mbs{\sigma}}_e(t)
& =
D_1\fb_0(\mbs{\epsilon}_e(t), \bq_e(t)) \dot{\mbs{\epsilon}}_e(t)
\\ & +
D_2\fb_0(\mbs{\epsilon}_e(t), \bq_e(t))
D\psi_e^{-1}(- D_2F_e(\mbs{\epsilon}_e(t), \bq_e(t)))
\\ & \equiv
\fb_1(\mbs{\epsilon}_e(t), \dot{\mbs{\epsilon}}_e(t), \bq_e(t)) .
\end{split}
\end{equation}
Iterating this process, we obtain the system of equations
\begin{equation}\label{Vg0NTD}
\mbs{\sigma}_e^{(\alpha)}(t)
=
\fb_\alpha(\{\mbs{\epsilon}_e^{(\beta)}(t)\}_{\beta=0}^\alpha, \bq_e(t)) ,
\quad
\alpha = 1,\dots,N ,
\end{equation}
with the functions $\fb_\alpha$ defined recursively. Assuming solvability, system (\ref{Vg0NTD}) can be solved for the internal variables to obtain a hereditary relation of the form
\begin{equation}
\bq_e(t)
=
\hat{\bq}_e
(
\{\mbs{\epsilon}_e^{(\alpha)}(t)\}_{\alpha=0}^N,
\{\mbs{\sigma}_e^{(\alpha)}(t)\}_{\alpha=0}^N
) .
\end{equation}
Inserting this relation in (\ref{BJYKd4}), we obtain the differential constraint
\begin{equation}
\mbs{\sigma}_e(t)
-
\fb_0
(
\mbs{\epsilon}_e(t),
\hat{\bq}_e
(
\{\mbs{\epsilon}_e^{(\alpha)}(t)\}_{\alpha=0}^N,
\{\mbs{\sigma}_e^{(\alpha)}(t)\}_{\alpha=0}^N
)
)
=
0 ,
\end{equation}
which is of the general form (\ref{k8exoU}).
A similar connection can be forged directly in the time-discrete setting. Thus, iterating the propagator (\ref{iQG1K4}), we obtain the system of equations
\begin{equation}\label{IZI84d}
\begin{split}
&
\mbs{\sigma}_{e,k+1-l}
= \\ &
D_1F_e(\mbs{\epsilon}_{e,k+1-l}, {\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e \cdots
{\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e({\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e(\bq_{k+1-N},\mbs{\epsilon}_{k+1-N}),
\mbs{\epsilon}_{k-N},\cdots,\mbs{\epsilon}_{k+1-l})) ,
\end{split}
\end{equation}
for $l=1,\dots,N$. Assuming again solvability, the system (\ref{IZI84d}) can be solved to obtain
\begin{equation}
\bq_{k+1-N}
=
\hat{\bq}_{k+1-N}
(
\{\mbs{\epsilon}_{e,k-l}\}_{l=0}^{N-1},
\{\mbs{\sigma}_{e,k-l}\}_{l=0}^{N-1}
)
\end{equation}
and
\begin{equation}
\begin{split}
&
\mbs{\sigma}_{e,k+1}
= \\ &
D_1F_e(\mbs{\epsilon}_{e,k+1},
{\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e \cdots
{\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e({\mbs{P}}} \def\bQ{{\mbs{Q}}} \def\bR{{\mbs{R}}_e(\bq_{k+1-N},\mbs{\epsilon}_{k+1-N}),
\mbs{\epsilon}_{k-N},\cdots,\mbs{\epsilon}_{k+1})) ,
\end{split}
\end{equation}
which supplies a time-discrete differential representation of the form (\ref{fYCd00}).
We thus conclude that internal variable and differential representations of material behavior are equivalent when the constitutive relations are sufficiently differentiable and the material behavior is stable. As already noted, within a Data-Driven framework the key conceptual advantage of the differential representation is that it relies on fundamental data only, namely, stress and strain data, and the internal variable set, if any, need not be known.
\section{Numerical examples: Viscoelasticity}
\label{IXEV5B}
We proceed to illustrate the preceding representational paradigms, and the Data-Driven schemes that they engender, by means of selected examples of application. Viscoelasticity is characterized by the smoothness of the kinetic equations and the existence of a stable equilibrium manifold. The corresponding data sets of viscoelasticity therefore lend themselves ideally to a differential representation, eqs.~(\ref{k8exoU}) and (\ref{fYCd00}).
\subsection{Example: The Standard Linear Solid}
The Standard Linear Solid, consisting of a Maxwell unit in parallel with an elastic unit, provides a simple and convenient example. The Standard Linear Solid Helmholtz free energy is
\begin{equation}\label{FOav7e}
F_e(\mbs{\epsilon}_e, \bq_e)
=
\frac{1}{2}
\mathbb{E}_0 \, \mbs{\epsilon}_e \cdot \mbs{\epsilon}_e
+
\frac{1}{2}
\mathbb{E}_1
(\mbs{\epsilon}_e - \bq_e)
\cdot
(\mbs{\epsilon}_e - \bq_e)
\end{equation}
where $\bq_e \in \mathbb{R}^{m_e}$ is an internal inelastic strain and $\mathbb{E}_0$ and $\mathbb{E}_1$ are moduli. The corresponding equilibrium relations (\ref{O4cXId}) are
\begin{subequations}
\begin{align}
& \label{wrlU1l}
\mbs{\sigma}_e(t)
=
D_1 F_e(\mbs{\epsilon}_e(t), \bq_e(t))
=
\mathbb{E}_0 \, \mbs{\epsilon}_e(t)
+
\mathbb{E}_1
(\mbs{\epsilon}_e(t) - \bq_e(t))
\\ & \label{s1lakO}
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t)
=
-
D_2 F_e(\mbs{\epsilon}_e(t), \bq_e(t))
=
\mathbb{E}_1
(\mbs{\epsilon}_e(t) - \bq_e(t)) ,
\end{align}
\end{subequations}
where ${\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e$ is the thermodynamic driving force conjugate to $\bq_e$. Assuming linear kinetics, we further have
\begin{equation}\label{BoA7iu}
\mathbb{E}_1
\dot{\bq}_e(t)
=
\frac{{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t)}{\tau_1}
=
\frac{\mathbb{E}_1}{\tau_1}
(\mbs{\epsilon}_e(t) - \bq_e(t))
\end{equation}
where $\tau_1$ is a relaxation time.
A straightforward calculation shows that the inelastic strain $\bq_e(t)$ can be eliminated from the above equations, using the time-derivative of (\ref{wrlU1l}) in addition, and that the resulting differential constraint is
\begin{equation}\label{8ZP3cV}
\mbs{\sigma}_e(t)
+
\tau_1 \dot{\mbs{\sigma}}_e(t)
-
\mathbb{E}_0 \mbs{\epsilon}_e(t)
-
(\mathbb{E}_0 + \mathbb{E}_1) \tau_1 \dot{\mbs{\epsilon}}_e(t)
=
0 ,
\end{equation}
which is of the form (\ref{k8exoU}). A straightforward time discretization further gives
\begin{equation}\label{5jkpKD}
\mbs{\sigma}_{e,k+1}
+
\tau_1
\frac{\mbs{\sigma}_{e,k+1} - \mbs{\sigma}_{e,k}}{t_{k+1} - t_k}
-
\mathbb{E}_0
\mbs{\epsilon}_{e,k+1}
-
(\mathbb{E}_0 + \mathbb{E}_1) \tau_1
\frac{\mbs{\epsilon}_{e,k+1} - \mbs{\epsilon}_{e,k}}{t_{k+1} - t_k}
=
0 ,
\end{equation}
The corresponding differential representation (\ref{32DiI8}) of the data set is
\begin{equation}\label{E7DbKS}
D_{e,k+1}
=
\{
(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})
\, : \,
(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k})
\text{ and }
(\ref{5jkpKD})
\} ,
\end{equation}
which, for fixed $(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k})$, defines a linear subspace of phase space of dimension $\mathbb{R}^{m_e}$. We conclude that first-order differential representations of the data set of the form (\ref{32DiI8}), with $p=q=1$, suffice to represent the Standard Linear Solid exactly. More generally, first-order differential representations of the form (\ref{32DiI8}) can only be expected to furnish an approximation of the actual, and unknown, material behavior.
\begin{figure}[h]
\centering
\includegraphics[width=0.75\linewidth]{./Figures/DDSLS.pdf}
\caption{Schematic representation of the relaxation test for a Standard Linear Solid bar. Inlaid expressions shown in the limit $\Delta t \to 0$ for simplicity. The constraint set $E_{k+1}$, left, is fixed at a constant strain while the data set $D_{k+1}$ moves downward parallel to itself so as to trace the relaxation curve of the bar, right.} \label{7CBiWT}
\end{figure}
\subsection{Example: The relaxation test}
We illustrate the Data-Driven problem defined by the Standard Linear Solid by means of the simple example of relaxation test of a bar, Fig.~\ref{7CBiWT}. In this case, the solution consists of a single time-dependent stress and strain pair $(\epsilon(t),\sigma(t))$. The constraint set $E_{k+1}$ is then constant and simply restricts the strain to be constant and equal to a prescribed value $\bar{\epsilon}$, i.~e.,
\begin{equation}
E_{k+1} = \{ (\epsilon, \sigma) \, : \, \epsilon = \bar{\epsilon} \} .
\end{equation}
Inserting this condition into the differential constraint (\ref{5jkpKD}), gives the relation
\begin{equation}
{\sigma}_{k+1}
+
\tau_1
\frac{{\sigma}_{k+1} - {\sigma}_{k}}{t_{k+1} - t_k}
-
\mathbb{E}_0
\bar{{\epsilon}}
=
0 .
\end{equation}
A straightforward calculation gives the Data-Driven solution as
\begin{equation}\label{NYRK94}
\begin{split}
\epsilon_k
=
\bar{\epsilon},
\quad
{\sigma}_k
& =
\mathbb{E}_1\bar{\epsilon}
\left(\frac{\tau_1}{\Delta t + \tau_1}\right)^k
\\ & +
\left[
\sum_{n=1}^{k-1}
\left(
\frac{\tau_1}{\Delta t + \tau_1}
\right)^n
\frac{\Delta t}{\Delta t + \tau_1}
+
\left(
\frac{\tau_1}{\Delta t + \tau_1}
\right)^k
\right]
\mathbb{E}_0 \bar{\epsilon} ,
\end{split}
\end{equation}
where we assume $t_{k+1} - t_k = \Delta t = $ constant, for simplicity. Inserting (\ref{NYRK94}) into (\ref{5jkpKD}) defines the data set $D_{k+1}$ as a line in phase space of slope approximately equal to $\mathbb{E}_0 + \mathbb{E}_1$ intersecting the stress axis at approximately $\sigma_k - (\mathbb{E}_0 + \mathbb{E}_1) \epsilon_k$.
Thus, the initial material data set $D_0$ is a line of slope roughly $\mathbb{E}_0 + \mathbb{E}_1$ through the origin that intersects the constraint set $E_0$ at $\sigma_0 = (\mathbb{E}_0 + \mathbb{E}_1) \bar{\epsilon}$, which is the instantaneous response of the solid. Subsequent material data sets $D_{k+1}$ translate downwards in phase space and their intersection with the constraint set $E_{k+1}$ traces the relaxation curve of the bar. More general Data-Driven solutions can be obtained if the material data set $D_{k+1}$ is allowed to be a point set, e.~g., approximating the Standard Linear Solid data set just described. In this case, the Data-Driven solution is the point in the constraint set $E_{k+1}$ closest to the material data set $D_{k+1}$. With the passage of time, these points again trace a Data-Driven relaxation curve of the bar, Fig.~\ref{7CBiWT}.
\begin{figure}[h]
\centering
\includegraphics[width=0.65\linewidth]{./Figures/TrussBC_vector3.pdf}
\caption{Truss Geometry, load points, and output locations. Red arrows represent applied loads, black arrows prescribed displacements. Nodes highlighted in black are fixed. Vertical displacements are output and monitored at the node highlighted in red. Member forces are output and monitored at the bar highlighted in red. }
\label{4NZE8X}
\end{figure}
\subsection{Convergence analysis: Truss structures}
\label{BuKeb1}
We demonstrate the convergence properties of Data-Driven viscoelasticity with the aid of the three-dimensional truss structure shown in Fig.~\ref{4NZE8X}. The geometry of the truss, which comprises 1,246 bars, the boundary conditions and the applied loads are also shown in Fig.~\ref{4NZE8X}. The loads are linearly ramped up to $t=10$, subsequently held constant up to $t=50$, linearly ramped back to zero at $t=60$, and held again constant up to $t=100$. The data sets are generated on the fly by randomizing the Standard Linear Solid data set (\ref{E7DbKS}). The data points are assumed to be uniformly distributed within a band of width $\Delta \epsilon_{e,k+1} = 0.030$. A typical local material data set is shown in Fig.~\ref{7CBiWT}. The resulting material data sets converge uniformly to the Standard Linear Solid graph in the sense defined in \cite{Conti:2018}. The parameters of the reference Standard Linear Solid used in calculations are $\mathbb{E}_0 = 75,000$, $\mathbb{E}_1 = 100,000$ and $\tau_1 = 5$. In addition, a constant time step $\Delta t = 1$ is used in all calculations.
\begin{figure}[h]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=0.99\linewidth]{Figures/StandViscoelasticity_Disp.pdf}
\caption{}
\label{w8e4TM}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=0.99\linewidth]{Figures/StandViscoelasticity_Force.pdf}
\caption{}
\label{lF01xo}
\end{subfigure}
\caption{Viscoelastic truss problem. Time-history comparison for data solver at various data resolutions for a) deflections at a degree of freedom with an applied force and b) axial forces in output bar.}
\label{Z4K9kb}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.65\linewidth]{Figures/StandViscoelasticity_Convergence.pdf}
\caption{Viscoelastic truss problem. Data convergence of the Data Driven viscoelastic problem to the reference Standard Linear Solid solution.}
\label{m5VKwr}
\end{figure}
Fig.~\ref{w8e4TM} depicts displacement histories at the output node shown in Fig.~\ref{4NZE8X} and Fig.~\ref{lF01xo} shows the history of the resultant of the reaction forces at the kinematically constraint nodes, cf.~Fig.~\ref{4NZE8X}. The convergence of the time histories towards the solution of the reference Standard Linear Solid with increasing number of materials data points is evident in the figures. The rate of convergence can be monitored by means of the weighted $\ell^2$ error
\begin{equation}\label{c558Xj}
\text{Error}
=
\left(
\sum_{k=0}^{T-1}
|z_{k+1} - z_{k+1}^{\text{ref}}|^2 \,
{\rm e}^{-t_{k+1}/\tau_1} (t_{k+1} - t_k)
\right)^{1/2}
\end{equation}
where $T$ is the number of time steps, $| \cdot |$ is as in (\ref{9oakLa}) and $z_{e,k}^{\text{ref}} = (\epsilon_{e,k}^{\text{ref}}, \sigma_{e,k}^{\text{ref}} )$ is the solution for the reference Standard Linear Solid. Weighted norms such as (\ref{c558Xj}) arise naturally in the analysis of viscoelastic problems (cf., e.~g., \cite{Lions:1972}). Compiling statistics over $50$ independent runs, i.~e., with different randomizations of the data set, we arrive at the convergence plot shown in Fig.~\ref{m5VKwr}. Remarkably, the computed rate of convergence is quadratic, or twice the linear rate of convergence characteristic of elastic problems \cite{Kirchdoerfer:2016}.
\section{Numerical examples: Plasticity}
\label{X0aZjS}
\begin{figure}[h]
\centering
\includegraphics[width=0.75\linewidth]{./Figures/DDP2.pdf}
\caption{Schematic representation of the evolution of a typical data set for the isotropic-kinematic linear-hardening solid, left. Rheological model consisting of two elastic elements and a hardening slider, right.}
\label{1ZESqU}
\end{figure}
Plasticity (cf., e.~g., \cite{Lubliner:1990}) supplies an example of a class of material data sets that are not amenable to a strict differential representation and require the use of history variables in addition.
\subsection{Example: The isotropic-kinematic linear-hardening solid}
We illustrate this class of materials by means of the simple isotropic-kinematic linear-hardening solid, Fig.~\ref{1ZESqU}. In this case, the free energy is of the form
\begin{equation}
F_e(\mbs{\epsilon}_e, \bq_e, q_e)
=
\frac{1}{2}
\mathbb{E}_0 \, \mbs{\epsilon}_e \cdot \mbs{\epsilon}_e
+
\frac{1}{2}
\mathbb{E}_1
(\mbs{\epsilon}_e - \bq_e)
\cdot
(\mbs{\epsilon}_e - \bq_e)
+
W_e(q_e)
\end{equation}
where $\bq_e \in \mathbb{R}^{m_e}$ is a internal inelastic strain, $q_{e}$ is an effective accumulated plastic strain, $W_e$ is a stored energy of cold work and $\mathbb{E}_0$ and $\mathbb{E}_1$ are moduli. The equilibrium relations (\ref{O4cXId}) evaluate to
\begin{subequations}
\begin{align}
& \label{B85dgt}
\mbs{\sigma}_e(t)
=
\mathbb{E}_0 \mbs{\epsilon}_e(t)
+
\mathbb{E}_1 (\mbs{\epsilon}_e(t) - \bq_e(t))
\\ &
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t)
=
\mathbb{E}_1 (\mbs{\epsilon}_e(t) - \bq_e(t))
\\ &
- p_e(t)
=
W_e'(q_e)
\equiv
\sigma_e(q_e) ,
\end{align}
\end{subequations}
where $\sigma_e(q_e)$ is the yield stress. For the rate-independent solid, the dual kinetic potential is of the form
\begin{equation}\label{NDKmA2}
\psi^*({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e , p_e)
=
\left\{
\begin{array}{ll}
0, & \text{if }
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e , p_e) \leq 0 , \\
+\infty, & \text{otherwise} ,
\end{array}
\right.
\end{equation}
for some convex yield function $f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e , p_e)$, i.~e., $\psi^*$ vanishes within the elastic domain $f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e , p_e) \leq 0$ and equals $+\infty$ elsewhere in driving-force space. We note that $\psi^*({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e , p_e)$ is not differentiable and, therefore, the corresponding kinetic relations
\begin{equation}\label{Ie6NO0}
(\dot{\bq}_e(t), \dot{q}_e(t))
\in
\partial
\psi^*({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t))
\end{equation}
are set-valued and must be understood in the sense of subdifferentials \cite{Rockafellar:1970}. Equivalently, the kinetic relations (\ref{Ie6NO0}) can be expressed in terms of Drucker's principle of maximum dissipation
\begin{equation}\label{8DEmAC}
\max_{({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t))}
\big\{
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) \cdot \dot{\bq}_e(t)
+
p_e(t) \dot{q}_e(t)
-
\psi^*({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t))
\big\} ,
\end{equation}
where the rates $(\dot{\bq}_e(t),\dot{q}_e(t))$ are regarded as given. In view of (\ref{NDKmA2}), (\ref{8DEmAC}) is in turn equivalent to
\begin{equation}\label{5VnV1o}
\max_{({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t))}
\big\{
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) \cdot \dot{\bq}_e(t)
+
p_e(t) \dot{q}_e(t)
\, : \,
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t)) \leq 0
\big\} ,
\end{equation}
which defines a standard convex-optimization problem \cite{Rockafellar:1970}. Introducing a Lagrange multiplier $\lambda_e(t)$, the corresponding Euler-Lagrange equations are
\begin{subequations}
\begin{align}
&
\dot{\bq}_e(t)
=
\lambda_e(t)
\frac{\partial f}{\partial \bq_e}({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t)) ,
\\ &
\dot{q}_e(t)
=
\lambda_e(t)
\frac{\partial f}{\partial q_e}({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t)) ,
\end{align}
\end{subequations}
subject to the Kuhn-Tucker conditions
\begin{equation}
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t)) \leq 0 ,
\quad
\lambda_e(t) \geq 0 ,
\quad
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t) , p_e(t)) \lambda_e(t) = 0 ,
\end{equation}
which encode the yielding and loading-unloading conditions. A fully-implicit discretization of (\ref{5VnV1o}) gives the time-discrete maximum dissipation principle
\begin{equation}
\max_{({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1})}
\big\{
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} \cdot
(\bq_{e,k+1} - \bq_{e,k})
+
p_{e,k+1}
(q_{e,k+1} - q_{e,k})
\, : \,
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1}) \leq 0
\big\} ,
\end{equation}
where $(\bq_{e,k+1}, q_{e,k+1})$ are regarded as given. The corresponding Euler-Lagrange equations are
\begin{subequations}\label{9NntA7}
\begin{align}
& \label{D7BKjG}
\bq_{e,k+1} - \bq_{e,k}
=
\lambda_{e,k+1}
\frac{\partial f}{\partial \bq_e}({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1})
\\ & \label{0mDLWD}
q_{e,k+1} - q_{e,k}
=
\lambda_{e,k+1}
\frac{\partial f}{\partial q_e}({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1})
\end{align}
\end{subequations}
subject to the Kuhn-Tucker loading-unloading conditions
\begin{equation}
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1}) \leq 0 ,
\quad
\lambda_{e,k+1} \geq 0 ,
\quad
f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1}) \lambda_{e,k+1} = 0 .
\end{equation}
These equations are closed by the time-discrete equilibrium relations
\begin{subequations}\label{9r6HVL}
\begin{align}
& \label{F7BThA}
\mbs{\sigma}_{e,k+1}
=
\mathbb{E}_0 \mbs{\epsilon}_{e,k+1}
+
\mathbb{E}_1 (\mbs{\epsilon}_{e,k+1} - \bq_{e,k+1})
\\ &
{\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_e(t)
=
\mathbb{E}_1 (\mbs{\epsilon}_{e,k+1} - \bq_{e,k+1})
\\ &
-
p_{e,k+1}
=
W_e'(q_{e,k+1})
\equiv
\sigma_e(q_{e,k+1}) ,
\end{align}
\end{subequations}
and jointly define a convex problem for $(\mbs{\sigma}_{e,k+1}, \bq_{e,k+1}, q_{e,k+1})$ given $\mbs{\epsilon}_{e,k+1}$ and $(\mbs{\sigma}_{e,k}, \bq_{e,k}, q_{e,k})$. A solution of this problem can be conveniently obtained by means of an elastic predictor-plastic corrector split \cite{Ortiz:1999, Vladimirov:2008}.
We note that the material data set $D_{e,k+1}$ of points $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ attainable at time $t_{k+1}$ is fully characterized by $(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k})$ and $q_{e,k}$. The dependence of $D_{e,k+1}$ on $(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k})$ is consistent with a differential representation. However, the additional dependence on $q_{e,k}$ is typical of a history variable representation. Indeed, the history-variable character of $q_{e,k}$ can be revealed as follows. Taking a convenient seminorm $| \cdot |$ of (\ref{D7BKjG}) and eliminating $\lambda_{e,k+1}$ together with (\ref{0mDLWD}), we obtain
\begin{equation}\label{2B0VSn}
q_{e,k+1} - q_{e,k}
=
\frac{|\bq_{e,k+1} - \bq_{e,k}|}{|\partial f/\partial \bq_e({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1})|}
\frac{\partial f}{\partial q_e}({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e,k+1} , p_{e,k+1}) .
\end{equation}
At this point, we note that the choice of yield function $f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e} , p_{e})$ is arbitrary up to scaling by positive functions, since that operation leaves the elastic domain invariant. Therefore, we may choose a normalization of $f({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e} , p_{e})$ such that
\begin{equation}
\frac
{
\partial f/\partial q_e({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e} , p_{e})
}
{
| \partial f/\partial \bq_e({\mbs{p}}} \def\bq{{\mbs{q}}} \def\br{{\mbs{r}}_{e} , p_{e}) |
}
=
1 .
\end{equation}
With this normalization, (\ref{2B0VSn}) reduces to
\begin{equation}\label{il3XRm}
q_{e,k+1} - q_{e,k}
=
|\bq_{e,k+1} - \bq_{e,k}|
\end{equation}
From (\ref{F7BThA}), we additionally have
\begin{equation}
\bq_{e,k+1} - \bq_{e,k}
=
\mathbb{E}_1^{-1}
\big(
( \mathbb{E}_0 + \mathbb{E}_1 )
(\mbs{\epsilon}_{e,k+1} - \mbs{\epsilon}_{e,k})
-
\mbs{\sigma}_{e,k+1} + \mbs{\sigma}_{e,k}
\big) ,
\end{equation}
which, inserted into (\ref{il3XRm}), further gives the incremental relation
\begin{equation}
q_{e,k+1} - q_{e,k}
=
\big|
\mathbb{E}_1^{-1}
\big(
( \mathbb{E}_0 + \mathbb{E}_1 )
(\mbs{\epsilon}_{e,k+1} - \mbs{\epsilon}_{e,k})
-
\mbs{\sigma}_{e,k+1} + \mbs{\sigma}_{e,k}
\big)
\big| .
\end{equation}
Finally, summing over the history of the material, we obtain the relation
\begin{equation}\label{IH8SXh}
q_{e,k}
=
\sum_{h \leq k}
\big|
\mathbb{E}_1^{-1}
\big(
( \mathbb{E}_0 + \mathbb{E}_1 )
(\mbs{\epsilon}_{e,h} - \mbs{\epsilon}_{e,h-1})
-
\mbs{\sigma}_{e,h} + \mbs{\sigma}_{e,h-1}
\big)
\big| ,
\end{equation}
which is of the form (\ref{RTL6EL}).
It follows from the preceding analysis that the material data set of an isotropic-kinematic plastic solid admits the mixed differential-hereditary representation
\begin{equation}\label{qo5b8V}
D_{e,k+1}
=
\{
(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})
\, : \,
(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k}, q_{e,k}),
\text{ and }
(\ref{9NntA7}-\ref{9r6HVL})
\} ,
\end{equation}
which, for fixed $(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k}, q_{e,k})$, defines a linear subspace of phase space of dimension $\mathbb{R}^{m_e}$. Again, from a Data-Driven perspective, the right interpretation of this result is that a mixed differential-hereditary of the form (\ref{qo5b8V}) suffices to represent the isotropic-kinematic plastic solid exactly. However, for general plastic solids the history variable (\ref{IH8SXh}) represents an {\sl ad hoc} choice intended to record partial information about the history of the material. Then, representations of the form (\ref{qo5b8V}), with $D_{e,k+1}$ consisting of points $(\mbs{\epsilon}_{e,k+1}, \mbs{\sigma}_{e,k+1})$ in phase space known to be attainable from initial conditions $(\mbs{\epsilon}_{e,k}, \mbs{\sigma}_{e,k}, q_{e,k})$, cf.~Fig.~\ref{1ZESqU}, can only be expected to furnish an approximation of the actual material behavior.
\subsection{Convergence analysis: Truss structures}
We again demonstrate the convergence properties of Data-Driven plasticity with the aid of the three-dimensional truss structure shown in Fig.~\ref{4NZE8X}. The boundary conditions are as in the viscoelastic calculations of Section~\ref{BuKeb1}. The loads are linearly ramped up linearly from $0$ to $0.8$ at $t=20$, ramped down to $-0.9$ at $t=60$ and finally ramped up again to $1.0$ at $t=100$. The data sets are generated on the fly by randomizing the isotropic-kinematic linear-hardening data set (\ref{qo5b8V}). The data points are assumed to be uniformly distributed within a band of width $\Delta \epsilon_{e,k+1} = 0.04$. A typical local material data set is shown in Fig.~\ref{1ZESqU}. The resulting material data sets converge uniformly to the material data sets of isotropic-kinematic hardening solid, cf.~\cite{Conti:2018}. The parameters of the reference isotropic-kinematic hardening solid used in calculations are $\mathbb{E}_0 = 10,000$, $\mathbb{E}_1 = 100,000$ and initial yield stress $\sigma_1 = 500$. Finally, a constant time step $\Delta t = 1$ is used in all calculations.
\begin{figure}[h]
\centering
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=0.99\linewidth]{Figures/IsoPlasticity_Disp.pdf}
\caption{}
\label{Z19SwM}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=0.99\linewidth]{Figures/IsoPlasticity_Force.pdf}
\caption{}
\label{7EDnGV}
\end{subfigure}
\caption{Plastic truss problem. Time-history comparison for data solver at various data resolutions for a) deflections at a degree of freedom with an applied force and b) axial forces measured in output bar.}
\label{2HAVL2}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.65\linewidth]{Figures/IsoPlasticity_Convergence.pdf}
\caption{Plastic truss problem. Data convergence of the Data Driven viscoelastic problem to the reference isotropic-kinematic hardening solution.}
\label{3DFGSu}
\end{figure}
Fig.~\ref{Z19SwM} depicts displacement histories at the output node shown in Fig.~\ref{4NZE8X} and Fig.~\ref{7EDnGV} shows the history of the resultant of the reaction forces at the kinematically constraint nodes, cf.~Fig.~\ref{4NZE8X}. The convergence of the time histories towards the solution of the reference isotropic-kinematic hardening solid with increasing number of materials data points is evident in the figures. The rate of convergence can be monitored by means of the rate-independent error
\begin{equation}\label{sZRRM4}
\text{Error}
=
\sum_{k=0}^{T-1}
| (\bz_{k+1} - \bz_k) - (\bz_{k+1}^{\text{ref}} - \bz_k^{\text{ref}}) |
\end{equation}
where $T$ is the number of time steps, $| \cdot |$ is as in (\ref{9oakLa}) and $z_{e+k}^{\text{ref}} = (\epsilon_{e+k}^{\text{ref}}, \sigma_{e+k}^{\text{ref}} )$ is the solution for the reference elastic-plastic Solid. Bounded-variation norms such as (\ref{sZRRM4}) arise naturally in the analysis of plasticity problems (cf., e.~g., \cite{Mielke:2008}). Compiling statistics over $50$ independent runs, we arrive at the convergence plot shown in Fig.~\ref{3DFGSu}. The computed rate of convergence is roughly linear, which coincides with the linear rate of convergence characteristic of elastic problems \cite{Kirchdoerfer:2016}.
\section{Summary and concluding remarks}
\label{y1HJx6}
We have extended the Data-Driven formulation of problems in elasticity of Kirchdoerfer and Ortiz \cite{Kirchdoerfer:2016} to inelasticity. This extension differs fundamentally from Data-Driven problems in elasticity in that the material data set evolves in time as a consequence of the history dependence of the material. Therefore, the central issue of Data-Driven inelasticity concerns the practical representation of evolving, history-dependent material data sets. In this regard, we have investigated three representational paradigms: i) {\sl materials with memory}, i.~e., conditioning the material data set to the past history of deformation; ii) {\sl differential materials}, i.~e., conditioning the material data set to short histories of stress and strain; and iii) {\sl history variables}, i.~e., conditioning the material data set to {\sl ad hoc} variables encoding partial information about the history of stress and strain. We have also considered combinations of these three paradigms thereof. We find that many classical models of viscoelasticity and plasticity can be represented by means of material data sets of the differential and/or history variable type. Evidently, such representations only afford approximations of actual, often complex, material behavior. The central approximation question therefore concerns the formulation of material set representations capable of accounting increasingly accurately for arbitrary inelastic behavior as further information is added to the representation. A rigorous analysis of this question is beyond the scope of this paper and, instead, we have presented selected numerical examples that demonstrate the range of Data-Driven inelasticity and the numerical performance of implementations thereof.
A number of additional considerations and possible extensions of Data-Driven inelasticity suggest themselves.
\underline{\sl Connection to machine learning}.
We note that the closest-point projection $P_D$ in (\ref{K1TK7z}) entails a search over the entire material data set $D$. This search can be carried out, e.~g., by recourse to range-search algorithms \cite{Kirchdoerfer:2016}, tree-search algorithms, or similar fast search algorithms. Interestingly, search algorithms rely on spatial data structures, such as quadtrees and octrees, based on the principle of recursive subdivision. Such structures represent density, neighbor, clustering and other relations between the data points, which in turn may be regarded as a form of unsupervised machine learning (cf., e.~g., \cite{Duda:2001}). However, we emphasize that here the aim is to 'learn' the data set in its entirety, instead of replacing it by a model or some other reduced representation. In particular, the learning process does not entail any loss of information relative to the material data set.
\underline{\sl Multi-fidelity Data-Driven problems}.
A number of interrelated extensions and variations of the Data-Driven paradigm presented in this paper are noteworthy. We begin by noting that data enter the distance-minimizing Data-Driven problem (\ref{K5b6z0}) with uniform confidence, i.~e., all data are presumed to be equally reliable. However, in practice some data are of higher quality than others. The importance of keeping careful record of the pedigree, or ancestry, of each data point and of devising metrics for quantifying the level of confidence that can be placed on the data is well-recognized in Data Science \cite{Newman:2002, Raissi:2017}. A generalization of the distance-minimizing Data-Driven problem (\ref{yKEW6b}) that accounts for data fidelity is
\begin{equation}\label{97HKA5}
\min_{\bz_{k+1} \in E_{k+1}}
\min_{{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1} \in D_{k+1}}
\Big(
d^2(\bz_{k+1},{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1})
+
C({\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}) ,
\Big)
\end{equation}
where the fidelity cost $C({\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}) \geq 0$ measures the uncertainty, or lack or fidelity, of data point ${\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}$. Thus, $C({\mbs{y}}} \def\bz{{\mbs{z}}_{k+1})=0$ if ${\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}$ is absolutely certain and $C({\mbs{y}}} \def\bz{{\mbs{z}}'_{k+1})\geq C({\mbs{y}}} \def\bz{{\mbs{z}}_{k+1})$ if ${\mbs{y}}} \def\bz{{\mbs{z}}'_{k+1}$ is less certain, or of lesser fidelity, than ${\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}$. It is clear from (\ref{97HKA5}) that data points now influence the Data-Driven solution according to their fidelity, i.~e., high-fidelity data are given more weight in determining the solution than low-fidelity data.
A standard quantification of experimental data uncertainty consists of appending error bars to the data, corresponding to an estimate of the standard deviation of the measurements, and identifying the data points with the center of the distribution. If $s({\mbs{y}}} \def\bz{{\mbs{z}}_e)$ is the standard deviation of a local data point of mean value ${\mbs{y}}} \def\bz{{\mbs{z}}_e$ and assuming a Gaussian distribution, the expected distance between a local state $\bz_e$ and the measurement is
\begin{equation}
\int
\frac{|\bz_e - {\mbs{y}}} \def\bz{{\mbs{z}}'_e|^2}{(\sqrt{2\pi} s({\mbs{y}}} \def\bz{{\mbs{z}}))^{2m_e}}
\exp\Big( - \frac{|{\mbs{y}}} \def\bz{{\mbs{z}}_e-{\mbs{y}}} \def\bz{{\mbs{z}}'_e|_e^2}{2 s^2({\mbs{y}}} \def\bz{{\mbs{z}}_e)} \Big)
\, dy'_e
=
|\bz_e-{\mbs{y}}} \def\bz{{\mbs{z}}_e|_e^2
+
2 m_e s^2({\mbs{y}}} \def\bz{{\mbs{z}}_e) .
\end{equation}
Comparing this identity with (\ref{97HKA5}) affords the identification
\begin{equation}
C({\mbs{y}}} \def\bz{{\mbs{z}}) = \sum_{e=1}^M 2 m_e s^2({\mbs{y}}} \def\bz{{\mbs{z}}_e) ,
\end{equation}
which relates the fidelity cost $C({\mbs{y}}} \def\bz{{\mbs{z}})$ to the uncertainty of the data.
\underline{\sl History-matching Data-Driven problems}.
The extended Data-Driven problem (\ref{97HKA5}) suggests the following variation of the Data-Driven inelasticity paradigm. Suppose that it is possible to collect history data of material elements directly, i.~e., a local material history repository $H_e$ is available consisting of corresponding pairs of short histories $\{\bz_{e,k+1-l}\}_{l=0}^{N}$ $=$ $(\{\mbs{\epsilon}_{e,k+1-l}\}_{l=0}^{N}, \{\mbs{\sigma}_{e,k+1-l}\}_{l=0}^{N})$. For instance, for the Standard Linear Solid, data repositories of this type consist of two-time histories $(\bz_{e,k}, \bz_{e,k+1})$ $=$ $(\{\mbs{\epsilon}_{e,k}, \mbs{\epsilon}_{e,k+1}\}, \{\mbs{\sigma}_{e,k}, \mbs{\sigma}_{e,k+1}\})$ in the local material history space $H_e = \mathbb{R}^{4 m_e}$. We can metrize $H_e$ by means of the norm
\begin{equation}
| \{\bz_{e,k+1-l}\}_{l=0}^{N} |_e
=
\Big(
\sum_{l=0}^{N} C_{e,l} | \bz_{e,k+1-l} |_e^2
\Big)^{1/2} ,
\end{equation}
where $\{C_{e,l}\}_{l=0}^{N}$ are positive weights. We can further define a global material history set as $H = H_1 \times \cdots \times H_M$, with norm
\begin{equation}
| \{\bz_{k+1-l}\}_{l=0}^{N} |
=
\Big(
\sum_{e=1}^m w_e \, | \{\bz_{e,k+1-l}\}_{l=0}^{N} |_e^2
\Big)^{1/2} .
\end{equation}
A history-matching Data-Driven problem can now be defined as
\begin{equation}\label{dcWC0i}
\min_{\bz_{k+1} \in E_{k+1}}
\min_{\{{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1-l}\}_{l=0}^{N} \in H}
d^2(\{\bz_{k+1-l}\}_{l=0}^{N}, \{{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1-l}\}_{l=0}^{N}) ,
\end{equation}
i.~e., the Data-Driven solution at time $t_{k+1}$ is the admissible state $\bz_{k+1} \in E_{k+1}$ such that the history $\{\bz_{k+1-l}\}_{l=0}^{N}$ is closest to the material history set $H$. Thus, in this history-matching paradigm the prior history to ${\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}$ is no longer fixed to $\{\bz_{k+1-l}\}_{l=0}^{N-1}$ and all prior histories in $H$ are considered with weights depending on their distance to $\{\bz_{k+1-l}\}_{l=0}^{N-1}$. We further note that problem (\ref{dcWC0i}) is in fact of the form (\ref{97HKA5}) with cost
\begin{equation}
C({\mbs{y}}} \def\bz{{\mbs{z}}_{k+1})
=
d^2(\{\bz_{k+1-l}\}_{l=0}^{N-1}, \{{\mbs{y}}} \def\bz{{\mbs{z}}_{k+1-l}\}_{l=0}^{N-1}) .
\end{equation}
Thus, the history-matching reformulation of the Data-Driven problem simply collects into a data set $D_{k+1}$ all states ${\mbs{y}}} \def\bz{{\mbs{z}}_{k+1}$ in $H$ and assigns them confidence weights according to the distance of the corresponding prior histories to the actual prior history $\{\bz_{k+1-l}\}_{l=0}^{N-1}$.
We have repeated the Standard Linear Solid test calculations described in Section~\ref{BuKeb1} using history material data sets in $((\epsilon_{e,k}, \sigma_{e,k}), (\epsilon_{e,k+1}, \sigma_{e,k+1}))$ space and history matching. The results of the calculations are ostensibly identical to those of Section~\ref{BuKeb1} and are not plotted here in the interest of brevity. History repositories enjoy the advantage that prior histories can be sampled off-line and a data set $D_{k+1}$ need not be known for all possible prior histories. The disadvantage is that history data add to the dimensionality of the data set. Therefore, history matching is only practical when prior histories are short, e.~g., in the context of low-order differential representations.
\underline{\sl Goal-oriented self-consistent data acquisition}.
An issue of critical importance concerns the acquisition of material data sets with appropriate coverage of phase space for specific applications. For general materials, phase space is of a dimension such that it cannot be covered uniformly by data. High-dimensional spaces are encountered in other areas of physics such as statistical mechanics, where the high dimensionality of state space is usually handled by means of {\sl importance sampling} techniques. The main idea is to generate data that are highly relevant to the particular problem under consideration, while eschewing irrelevant areas of phase space. A method for generating such goal-oriented data sets is the self-consistent approach of Leygue {\sl et al.} \cite{Leygue:2018}. In that approach, from a collection of non-homogeneous strain fields, e.~g., measured through Digital Image Correlation (DIC), a self-consistent iteration builds a material data set of strain--stress pairs that cover the region of phase-space relevant to a particular problem. In effect, the self-consistent approach generates the material data set and solves for the corresponding Data-Driven solution simultaneously.
These extensions and generalizations of Data-Driven inelasticity suggest worthwhile directions for further research.
\section*{Acknowledgments}
MO gratefully acknowledges the support of the Deutsche Forschungsgemeinschaft (DFG) through the Sonderforschungsbereich 1060 {\sl ``The mathematics of emergent effects''}. SR and RE gratefully acknowledge the financial support of the Deutsche Forschungsgemeinschaft (DFG) through the project {\sl ``Model order reduction in space and parameter dimension – towards damage-based modeling of polymorphic uncertainty in the context of robustness and reliability''} within the priority programm SPP 1886 {\sl ``Polymorphic uncertainty modelling for the numerical design of structures''}.
|
\section{Introduction}
\label{sec:intro}
In multi-modal translation, the task is to translate from a source sentence
and the image that it describes, into a target sentence in another language.
As both automatic image captioning systems and crowd captioning efforts tend to mainly yield descriptions in English,
multi-modal translation can be useful for generating descriptions of images for languages other than English.
In the MeMAD project\footnote{\url{https://www.memad.eu/}}, multi-modal translation
is of interest for creating textual versions or descriptions of audio-visual content.
Conversion to text enables both indexing for multi-lingual image and video search,
and increased access to the audio-visual materials for visually impaired users.
We adapt%
\footnote{Our fork available from \url{https://github.com/Waino/OpenNMT-py/tree/develop_mmod}}
the Transformer \cite{vaswani2017attention} architecture
to use global image features extracted from Detectron,
a pre-trained object detection and localization neural network.
We use two additional training corpora: MS-COCO \cite{mscoco} and OpenSubtitles2018 \cite{opensubtitles}.
MS-COCO is multi-modal, but not multi-lingual.
We extended it to a synthetic multi-modal and multi-lingual training set.
OpenSubtitles is multi-lingual, but does not include associated images,
and was used as text-only training data.
This places our entry in the unconstrained category of the WMT shared task.
Details on the architecture used in this work can be found in Section \ref{sec:arch}.
Further details on the synthetic data are presented in Section \ref{sec:text_based}.
Data sets are summarized in Table \ref{tab:data}.
\begin{table}[t]
\begin{center}
{\small
\begin{tabular}{lccccr}
\toprule
Data set & images & en & de & fr & sentences \\
\midrule
Multi30k & \checkmark & \checkmark & \checkmark & \checkmark & 29k \\
MS-COCO & \checkmark & \checkmark & $+$ & $+$ & 616k \\
OpenSubtitles & & \checkmark & \checkmark & \checkmark & 23M/42M \\
& \multicolumn{5}{r}{1M, 3M, and 6M subsets used.} \\
\bottomrule
\end{tabular}}
\end{center}
\caption{Summary of data set sizes.
\checkmark means attribute is present in original data.
$+$ means data set augmented in this work.}
\label{tab:data}
\end{table}
\section{Experiment 1: Optimizing Text-Based Machine Translation}
\label{sec:text_based}
Our first aim was to select the text-based MT system
to base our multi-modal extensions on.
We tried a wide range of models,
but only include results with the two strongest systems:
Marian NMT with the \emph{amun} model \cite{mariannmt},
and OpenNMT \cite{opennmt} with the \emph{Transformer} model.
We also studied the effect of additional training data. Our initial experiments showed
that movie subtitles and their translations work rather well to augment the given training data.
Therefore, we included parallel subtitles from the OpenSubtitles2018 corpus to train better text-only MT models.
For these experiments, we apply the Marian amun model,
an attentional encoder-decoder model with bidirectional LSTM's on the encoder side.
In our first series of experiments, we observed that domain-tuning is very important when using Marian.
The domain-tuning was accomplished
by a second training step on in-domain data after training the model on the entire data set.
Table~\ref{tab:multidomain} shows the scores on development data.
We also tried decoding with an ensemble of three independent runs, which also pushed the performance a bit.
\begin{table}
\begin{center}
{\small
\begin{tabular}{lccc}
\toprule
\textsc{en-fr} & flickr16 & flickr17 & mscoco17 \\
\midrule
multi30k & 61.4 & 54.0 & 43.1 \\
\quad +{\sc subs}{$_{\txt{full}}$} & 53.7 & 48.9 & 47.0 \\
\quad\quad +domain-tuned & 66.1 & 59.7 & \bf 51.7 \\
\quad\quad\quad +ensemble-of-3 & \bf 66.5 & \bf 60.2 & 51.6 \\
\bottomrule
\toprule
\textsc{en-de} & flickr16 & flickr17 & mscoco17 \\
\midrule
multi30k & 38.9 & 32.0 & 27.7 \\
\quad +{\sc subs}{$_{\txt{full}}$} & 41.3 & 34.1 & 31.3 \\
\quad\quad +domain-tuned & 43.3 & 38.4 & 35.0 \\
\quad\quad\quad +ensemble-of-3 & \bf 43.9 & \bf 39.6 & \bf 37.0 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:multidomain} Adding subtitle data and domain tuning for image caption translation (BLEU\% scores). All results with Marian Amun.}
\end{table}
Furthermore, we tried to artificially increase the amount of in-domain data
by translating existing English image captions to German and French.
For this purpose, we used the large MS-COCO data set with its 100,000 images
that have five image captions each. We used our best multidomain model (see Table~\ref{tab:multidomain})
to translate all of those captions and used them as additional training data.
This procedure also transfers the knowledge learned by the multidomain model into the caption translations, which helps us to improve the coverage of the system with less out-of-domain data.
Hence, we filtered the large collection of translated movie subtitles
to a smaller portion of reliable sentence pairs (one million in the experiment we report)
and could train on a smaller data set with better results.
We experimented with two filtering methods.
Initially, we implemented a basic heuristic filter (\subsH{}),
and later we improved on this with a language model filter (\subsLM{}).
Both procedures consider each sentence pair, assign it a quality score,
and then select the highest scoring 1, 3, or 6 million pairs, discarding the rest.
The \subsH{} method counts terminal punctuation (\emph{`.', `...', `?', `!'}) in the source and target sentences,
initializing the score as the negative of the absolute value of the difference between these counts. Afterwards, it further
decrements the score by $1$ for each occurrence of terminal punctuation beyond the first in each of the sentences.
The \subsLM{} method first preprocesses the data by filtering samples by length and ratio of lengths,
applying a rule-based noise filter,
removing all characters not present in the Multi30k set,
and deduplicating samples.
Afterwards, target sentences in the
remaining pairs are scored using a character-based deep LSTM language model trained on the Multi30k data.
Both selection procedures are intended for noise filtering,
and \subsLM{} additionally acts as domain adaptation.
Table~\ref{tab:cocotrans} lists the scores we obtained on development data.
\setlength{\tabcolsep}{0.2em}
\begin{table}
\begin{center}
{\small
\begin{tabular}{llccc}
\toprule
& \textsc{en-fr} & flickr16 & flickr17 & mscoco17 \\
\midrule
A & \subsH{1M}$+$MS-COCO & 66.3 & 60.5 & 52.1 \\
A & \quad +domain-tuned & 66.8 & 60.6 & 52.0 \\
A & \quad\quad +labels & \bf 67.2 & 60.4 & 51.7 \\
T & \subsLM{1M}+MS-COCO & 66.9 & 60.3 & \bf 52.8 \\
T & \quad +labels & \bf 67.2 & \bf 60.9 & 52.7 \\
\bottomrule
\toprule
& \textsc{en-de} & flickr16 & flickr17 & mscoco17 \\
\midrule
A & \subsH{1M}$+$MS-COCO & 43.1 & 39.0 & 35.1 \\
A & \quad +domain-tuned & 43.9 & 39.4 & 35.8 \\
A & \quad\quad +labels & 43.2 & 39.3 & 34.3 \\
T & \subsLM{1M}+MS-COCO & \bf 44.4 & 39.4 & 35.0 \\
T & \quad +labels & 44.1 & \bf 39.8 & \bf 36.5 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:cocotrans} Using automatically translated image captions and domain labels (BLEU\% scores).
A is short for Amun, T for Transformer.}
\end{table}
\setlength{\tabcolsep}{0.5em}
To make a distinction between automatically translated captions,
subtitle translations and human-translated image captions, we also introduced domain labels
that we added as special tokens to the beginning of the input sequence. In this way,
the model can use explicit information about the domain when deciding how to translate given input.
However, the effect of such labels is not consistent between systems.
For Marian amun, the effect is negligible as we can see in Table~\ref{tab:cocotrans}.
For the Transformer, domain labels had little effect on BLEU
but were clearly beneficial according to chrF-1.0.
\subsection{Preprocessing of textual data}
The final preprocessing pipeline for the textual data consisted of
lowercasing, tokenizing using Moses,
fixing double-encoded entities and other encoding problems,
and normalizing punctuation.
For the OpenSubtitles data we additionally used the \subsLM{} subset selection.
Subword decoding has become popular in NMT.
Careful choice of translation units
is especially important as one of the target languages of our system is German,
a morphologically rich language.
We trained a shared 50k subword vocabulary using Byte Pair Encoding (BPE) \cite{sennrich2015neural}.
To produce a balanced multi-lingual segmentation, the following procedure was used:
First, word counts were calculated individually for
English and each of the 3 target languages Czech\footnote{Czech was later dropped as a target language due to time constraints.}, French and German.
The counts were normalized to equalize the sum of the counts for each language.
This avoided imbalance in the amount of data skewing the segmentation in favor of some language.
Segmentation boundaries around hyphens were forced, overriding the BPE.
Multi-lingual translation with target-language tag
was done following \newcite{johnson2016google}.
A special token, e.g. <TO\_DE> to mark German as the target language,
was prefixed to each paired English source sentence.
\section{Experiment 2: Adding Automatic Image Captions}
Our first attempt to add multi-modal information to the translation model includes the incorporation of automatically created image captions in a purely text-based translation engine. For this, we generated five English captions for each of the images in the provided training and test data. This was done by using our in-house captioning system \cite{Shetty2018}.
The image captioning system uses a 2-layer LSTM with residual connections to generate captions based on scene context and object location descriptors, in addition to standard CNN-based features.
The model was trained with the MS-COCO training data and used to be state of the art in the COCO leaderboard\footnote{\url{https://competitions.codalab.org/competitions/3221}} in Spring 2016.
The beam search size was set to five.
We tried two models for the integration of those captions: (1) a dual attention multi-source model
that adds another input sequence with its own decoder attention and (2) a concatenation model
that adds auto captions at the end of the original input string separated by a special token.
In the second model, attention takes care of learning how to use the additional information
and previous work has shown that this, indeed, is possible \cite{niehues2016pre,ostling2017helsinki}. For both models,
we applied Marian NMT that already includes a working implementation of dual attention translations.
Table~\ref{tab:autocap} summarizes the scores on the three development test sets for English-French and English-German.
\setlength{\tabcolsep}{0.22em}
\begin{table}
\small
\begin{center}
{\small
\begin{tabular}{lccc}
\toprule
\textsc{en-fr} & flickr16 & flickr17 & mscoco17 \\
\midrule
multi30k & 61.4 & 54.0 & 43.1 \\
\quad +autocap (dual attn.) & 60.9 & 52.9 & 43.3 \\
\quad +autocap 1 (concat) & 61.7 & 53.7 & 43.9 \\
\quad +autocap 1-5 (concat) & \bf 62.2 & \bf 54.4 & \bf 44.1 \\
\bottomrule
\toprule
\textsc{en-de} & flickr16 & flickr17 & mscoco17 \\
\midrule
multi30k & 38.9 & 32.0 & 27.7 \\
\quad +autocap (dual attn.) & 37.8 & 30.2 & 27.0 \\
\quad +autocap 1 (concat) & 39.7 & \bf 32.2 & \bf 28.8 \\
\quad +autocap 1-5 (concat) & \bf 39.9 & 32.0 & 28.7 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:autocap} Adding automatic image captions (only the best one or all 5). The table shows BLEU scores in \%. All results with Marian Amun.}
\end{table}
\setlength{\tabcolsep}{0.5em}
We can see that the dual attention model does not work at all and the scores slightly drop.
The concatenation approach works better probably because
the common attention model learns interactions between the different types of input.
However, the improvements are small if any and the model basically learns to ignore
the auto captions, which are often very different from the original input.
The attention pattern in the example of Figure~\ref{fig:autocaps} shows
one of the very rare cases where we observe at least some attention to the automatic captions.
\begin{figure}[h]
\flushright
\includegraphics[width= 1\hsize]{Images/autocaps_example.png}
\caption{Attention layer visualization for an example where at least one of the attention weights for the last part of the sentence, which corresponds to the automatically generated captions, obtains a value above 0.3}
\label{fig:autocaps}
\end{figure}
\section{Experiment 3: Multi-modal Transformer}
One benefit of NMT,
in addition to its strong performance,
is its flexibility in enabling different information sources to be merged.
Different strategies to include image features both on the encoder and decoder side have been explored.
We are inspired by the recent success of the Transformer architecture
to adapt some of these strategies for use with the Transformer.
Recurrent neural networks start their processing from some \textbf{initial hidden state}.
Normally, a zero vector or a learned parameter vector is used,
but the initial hidden state is also a natural location to introduce additional context e.g. from other modalities.
Initializing can be applied in either the encoder (IMG$_{E}$) or decoder (IMG$_{D}$) \cite{calixto2017dcu}.
These approaches are not directly applicable to the Transformer,
as it is not a recurrent model, and lacks a comparable initial hidden state.
\textbf{Double attention} is another popular choice, used by e.g. \newcite{caglayan2017lium}.
In this approach, two attention mechanisms are used, one for each modality.
The attentions can be separate or hierarchical.
While it would be possible to use double attention with the Transformer,
we did not explore it in this work.
The multiple multi-head attention mechanisms in the Transformer
leave open many challenges in how this integration would be done.
\textbf{Multi-task learning} has also been used,
e.g. in the Imagination model \cite{elliott2017imagination},
where the auxiliary task consists of reconstructing the visual features from the source encoding.
Imagination could also have been used with the Transformer,
but we did not explore it in this work.
The \textbf{source sequence} itself is also a possible location for including the visual information.
In the IMG$_{W}$ approach,
the visual features are encoded as a pseudo-word embedding
concatenated to the word embeddings of the source sentence.
When the encoder is a bidirectional recurrent network, as in \newcite{calixto2017dcu},
it is beneficial to add the pseudo-word
both at the beginning and the end
to make it available for both encoder directions.
This is unnecessary in the Transformer,
as it has equal access to all parts of the source in the deeper layers of the encoder.
Therefore, we add the pseudo-word only to the beginning of the sequence.
We use an affine projection of the image features $V \in \mathbb{R}^{80}$
into a pseudo-word embedding $x_{I} \in \mathbb{R}^{512}$
\[x_{I} = W_{src} \cdot V + b_{I}.\]
In the LIUM \emph{trg-mul} \cite{caglayan2017lium},
the \textbf{target embeddings} and visual features are interacted through elementwise multiplication.
\[y'_{j} = y_{j} \odot \tanh(W_{mul}^{dec} \cdot V)\]
Our initial gating approach resembles \emph{trg-mul}.
\subsection{Architecture}
\label{sec:arch}
The baseline NMT for this experiment is the OpenNMT implementation of the Transformer.
It is an encoder-decoder NMT system
using the Transformer architecture \cite{vaswani2017attention} for both the encoder and decoder side.
The Transformer is a deep, non-recurrent network for processing variable-length sequences.
A Transformer is a stack of layers, consisting of two types of sub-layer:
multi-head (MH) attention (Att) sub-layers and feed-forward (FF) sub-layers:
\begin{align}
\attention(Q, K, V) & = \softmax(\frac{QK^T}{\sqrt{d_{k}}})V \nonumber\\
a_{i} & = \attention(QW_{i}^{Q}, KW_{i}^{K}, VW_{i}^{V}) \nonumber\\
\multihead(Q, K, V) & = [a_{1};\mathellipsis;a_{h}]W^{O} \nonumber\\
\feedforward(x) & = \max(0, xW_{1} + b_{1})W_{2} + b_{2}
\end{align}
where $Q$ is the input query,
$K$ is the key,
and $V$ the attended values.
Each sub-layer is individually wrapped in a residual connection and layer normalization.
When used in translation, Transformer layers are stacked into an encoder-decoder structure.
In the encoder, the layer consists of a self-attention sub-layer followed by a FF sub-layer.
In self-attention, the output of the previous layer is used as queries, keys and values $Q = K = V$.
In the decoder, a third context attention sub-layer is inserted between the self-attention and the FF.
In context attention, $Q$ is again the output of the previous layer,
but $K = V$ is the output of the encoder stack.
The decoder self-attention is also masked to prevent access to future information.
Sinusoidal position encoding makes word order information available.
\textbf{Decoder gate}.
Our first approach is inspired by \emph{trg-mul}.
A gating layer is introduced to modify the pre-softmax prediction distribution.
This allows visual features to directly suppress a part of the output vocabulary.
The probability of correctly translating a source word with visually resolvable ambiguity
can be increased by suppressing the unwanted choices.
At each timestep the decoder output $s_{j}$ is projected to an unnormalized distribution over the target vocabulary.
\[ y_{j} = W \cdot s_{j} + b \]
Before normalizing the distribution using a softmax layer, a gating layer can be added.
\begin{align}
g &= \sigma(W_{gate}^{dec} \cdot V + b_{gate}^{dec}) \nonumber\\
y'_{j} &= y_{j} \odot g
\end{align}
Preliminary experiments showed that gating based on only the visual features did not work.
Suppressing the same subword units during the entire decoding of the sentence was too disruptive.
We addressed this by using the decoder hidden state as additional input to control the gate.
This causes the vocabulary suppression to be time dependent.
\begin{align}
g_{j} &= \sigma(U_{gate}^{dec} \cdot s_{j} + W_{gate}^{dec} \cdot V + b_{gate}^{dec}) \nonumber\\
\end{align}
\textbf{Encoder gate}.
The same gating procedure can also be applied to the output of the encoder.
When using the encoder gate,
the encoded source sentence is disambiguated,
instead of suppressing part of the output vocabulary.
\begin{align}
g_{i} &= \sigma(U_{gate}^{enc} \cdot h_{i} + W_{gate}^{enc} \cdot V + b_{gate}^{enc}) \nonumber\\
h'_{i} &= h_{i} \odot g_{i}
\end{align}
The gate biases $b_{gate}^{dec}$ and $b_{gate}^{enc}$ should be initialized to positive values,
to start training with the gates opened.
We also tried combining both forms of gating.
\begin{table}
\begin{center}
{\small
\begin{tabular}{lccc}
\toprule
\sc{en-fr} & flickr16 & flickr17 & mscoco17\\
\midrule
IMG$_{W}$ & \em 68.30 & \bf 62.45 & 52.86 \\
enc-gate & 68.01 & 61.38 & \bf 53.40 \\
dec-gate & 67.99 & 61.53 & 52.38 \\
enc-gate + dec-gate & \bf 68.58 & \em 62.14 & \em 52.98 \\
\bottomrule
\toprule
\sc{en-de} & flickr16 & flickr17 & mscoco17\\
\midrule
IMG$_{W}$ & \em 45.09 & 40.81 & 36.94 \\
enc-gate & 44.75 & \bf 41.44 & \bf 37.76 \\
dec-gate & \bf 45.21 & 40.79 & 36.47 \\
enc-gate + dec-gate & 44.91 & \em 41.06 & \em 37.40 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:strategies}
Comparison of strategies for integrating visual information (BLEU\% scores).
All results using Transformer, Multi30k+MS-COCO+\subsLM{3M}, Detectron mask surface, and domain labeling.
}
\end{table}
\setlength{\tabcolsep}{0.23em}
\begin{table}
\begin{center}
{\small
\begin{tabular}{lccc}
\toprule
\sc{en-fr} & flickr16 & flickr17 & mscoco17\\
\midrule
\subsLM{3M} detectron & 68.30 & 62.45 & 52.86 \\
\quad $+$ensemble-of-3 & 68.72 & 62.70 & 53.06 \\
\quad\quad $-$visual features & \bf 68.74 & \bf 62.71 & 53.14 \\
\quad $-$MS-COCO & 67.13 & 61.17 & \bf 53.34 \\
\quad $-$multi-lingual & 68.21 & 61.99 & 52.40 \\
\subsLM{6M} detectron & 68.29 & 61.73 & 53.05 \\
\subsLM{3M} gn2048 & 67.74 & 61.78 & 52.76 \\
\subsLM{3M} text-only & 67.72 & 61.75 & 53.02 \\
\bottomrule
\toprule
\sc{en-de} & flickr16 & flickr17 & mscoco17\\
\midrule
\subsLM{3M} detectron & 45.09 & 40.81 & 36.94 \\
\quad $+$ensemble-of-3 & 45.52 & \bf 41.84 & \bf 37.49 \\
\quad\quad $-$visual features & \bf 45.59 & 41.75 & 37.43 \\
\quad $-$MS-COCO & 45.11 & 40.52 & 36.47 \\
\quad $-$multi-lingual & 44.95 & 40.09 & 35.28 \\
\subsLM{6M} detectron & 45.50 & 41.01 & 36.81 \\
\subsLM{3M} gn2048 & 45.38 & 40.07 & 36.82 \\
\subsLM{3M} text-only & 44.87 & 41.27 & 36.59 \\
\quad $+$multi-modal finetune & 44.56 & 41.61 & 36.93 \\
\bottomrule
\end{tabular}}
\end{center}
\caption{\label{tab:multimodal}
Ablation experiments (BLEU\% scores).
The row \subsLM{3M} \emph{detectron} shows our best single model.
Individual components or data choices are varied one by one.
$+$ stands for adding a component, and
$-$ for removing a component or data set.
Multiple modifications are indicated by increasing the indentation.}
\end{table}
\setlength{\tabcolsep}{0.5em}
\subsection{Visual feature selection}
\label{sec:visual_features}
Image feature selection was performed using the LIUM-CVC translation system~\cite{caglayan2017lium} training on the WMT18 training data, and evaluating on the \emph{flickr16, flickr17} and \emph{mscoco17} data sets.
This setup is different from our final NMT architecture as the visual feature selection stage was performed at an earlier phase of our experiments.
However, the LIUM-CVC setup without training set expansion was also faster to train which enabled a more extensive feature selection process.
We experimented with a set of state-of-the-art visual features, described below.
\textbf{CNN-based features} are 2048-dimensional feature vectors produced by applying reverse spatial pyramid pooling on features extracted from the 5$^{th}$ Inception module of the pre-trained GoogLeNet~\cite{szegedy2015going}.
For a more detailed description, see~\cite{Shetty2018}.
These features are referred to as gn2048 in Table~\ref{tab:multimodal}.
\textbf{Scene-type features}
are 397-dimensional feature vectors representing the association score of an image to each of the scene types in SUN397~\cite{xiao2010sun}. Each association score is determined by a separate Radial Basis Function Support Vector Machine (RBF-SVM) classifier trained from pre-trained GoogLeNet CNN features~\cite{Shetty2018}.
\textbf{Action-type features} are 40-dimensional feature vectors created with RBF-SVM classifiers similarly to the scene-type features, but using the Stanford 40 Actions dataset~\cite{conf/iccv/YaoJKLGF11} for training the classifiers.
Pre-trained GoogLeNet CNN features~\cite{szegedy2015going} were again used as the first-stage visual descriptors.
\textbf{Object-type and location features} are generated using the Detectron software\footnote{\url{https://github.com/facebookresearch/Detectron}} which implements Mask R-CNN~\cite{he2017mask} with ResNeXt-152~\cite{xie2017aggregated} features.
Mask R-CNN is an extension of Faster R-CNN object detection and localization~\cite{ren2015faster} that also generates a segmentation mask for each of the detected objects.
We generated an 80-dimensional \emph{mask surface} feature vector by expressing the image surface area covered by each of the MS-COCO classes based on the detected masks.
We found that the Detectron mask surface resulted in the best BLEU scores in all evaluation data sets for improving the German translations.
Only for \emph{mscoco17} the results could be slightly improved with a fusion of mask surface and the SUN 397 scene-type feature.
For French, the results were more varied, but we focused on improving the German translation results as those were poorer overall.
We experimented with different ways of introducing the image features into the translation model implemented in LIUM-CVC, and found as in~\cite{caglayan2017lium}, that \emph{trg-mul} worked best overall.
Later we learned that the \emph{mscoco17} test set has some overlap with the COCO 2017 training set, which was used to train the Detectron models.
Thus, the results on that test set may not be entirely reliable.
However, we still feel confident in our conclusions as they are also confirmed by the \emph{flickr16} and \emph{flickr17} test sets.
\subsection{Training}
\label{sec:train}
\begin{figure}[h]
\centering
\includegraphics[width=0.93\hsize]{Images/15444421898_846c46ac5b.jpg} \caption{Image 117 was translated correctly as feminine ``eine besitzerin steht still und ihr brauner hund rennt auf sie zu .'' when not using the image features, but as masculine ``ein besitzer \textellipsis'' when using them. The English text contains the word ``her''. The person in the image has short hair and is wearing pants.}
\label{fig:blinded}
\end{figure}
We use the following parameters for the network:%
\footnote{Parameters were chosen following the OpenNMT FAQ
\url{http://opennmt.net/OpenNMT-py/FAQ.html\#how-do-i-use-the-transformer-model}}
6 Transformer layers in both encoder and decoder,
512-dimensional word embeddings and hidden states,
dropout 0.1, batch size 4096 tokens, label smoothing 0.1,
Adam with initial learning rate 2 and $\beta_{2}$ 0.998.
For decoding, we use an ensemble procedure,
in which
the predictions of 3 independently trained models
are combined by averaging after the softmax layer
to compute combined prediction.
We evaluate the systems using uncased BLEU using multibleu.
During tuning, we also used characterF \cite{popovic2015chrf} with $\beta$ set to 1.0.
There are no images paired with the sentences in OpenSubtitles.
When using OpenSubtitles in training multi-modal models,
we feed in the mean vector of all visual features in the training data
as a dummy visual feature.
\subsection{Results}
\label{sec:results}
Based on the previous experiments,
we chose the Transformer architecture,
Multi30k+MS-COCO+\subsLM{3M} data sets,
Detectron mask surface visual features,
and domain labeling.
Table~\ref{tab:strategies} shows the BLEU scores
for this configuration with different ways of integrating the visual features.
The results are inconclusive.
The ranking according to chrF-1.0 was not any clearer.
Considering the results as a whole and the simplicity of the method,
we chose IMG$_{W}$ going forward.
Table~\ref{tab:multimodal} shows results of ablation experiments
removing or modifying one component or data choice at a time,
and results when using ensemble decoding.
Using ensemble decoding gave a consistent but small improvement.
Multi-lingual models were clearly better than mono-lingual models.
For French, 6M sentences of subtitle data gave worse results than 3M.
We experimented with adding multi-modality to a pre-trained text-only system using a fine tuning approach.
In the fine tuning phase, a \emph{dec-gate} gating layer was added to the network.
The parameters of the main network were frozen, allowing only the added gating layer to be trained.
Despite the freezing, the network was still able to unlearn most of the benefits of the additional text-only data.
It appears that the output vocabulary was reduced back towards the vocabulary seen in the multi-modal training set.
When the experiment was repeated so that the finetuning phase included the text-only data,
the performance returned to approximately the same level as without tuning
(+multi-modal finetune row in Table~\ref{tab:multimodal}).
To explore the effect of the visual features on the translation of our final model,
we performed an experiment where we retranslated using the ensemble while ``blinding'' the model.
Instead of feeding in the actual visual features for the sentence,
we used the mean vector of all visual features in the training data.
The results are marked \emph{-visual features} in Table~\ref{tab:multimodal}.
The resulting differences in the translated sentences were small,
and mostly consisted of minor variations in word order.
BLEU scores for French were surprisingly slightly improved by this procedure.
We did not find clear examples of successful disambiguation.
Figure~\ref{fig:blinded} shows one example of a detrimental use of visual features.
It is possible that adding to the training data
forward translations of MS-COCO captions from a text-only translation system introduced a biasing effect.
If there is translational ambiguity that should be resolved using the image,
the text-only system will not be able to resolve it correctly,
instead likely yielding the word that is most frequent in that textual context.
Using such data for training a multi-modal system might bias it towards ignoring the image.
On this year's \emph{flickr18} test set,
our system scores 38.54 BLEU for English-to-German
and 44.11 BLEU for English-to-French.
\section{Conclusions}
\label{sec:concl}
Although we saw an improvement from incorporating multi-modal information,
the improvement is modest.
The largest differences in quality between the systems we experimented with
can be attributed to the quality of the underlying text-only NMT system.
We found the amount of in-domain training data
and multi-modal training data to be of great importance.
The synthetic MS-COCO data was still beneficial,
despite being forward translated,
and the visual features being over-confident due to being extracted from a part of the image classifier training data.
Even after expansion with synthetic data,
the available multi-modal data is dwarfed by the amount of text-only data.
We found that movie subtitles worked well for this purpose.
When adding text-only data, domain adaptation was important,
and increasing the size of the selection met with diminishing returns.
Current methods do not fully address the problem of how to efficiently learn
from both large text-only data and small multi-modal data simultaneously.
We experimented with a fine tuning approach to this problem, without success.
Although the effect of the multi-modal information was modest,
our system still had the highest performance of the task participants
for the English-to-German and English-to-French language pairs,
with absolute differences of +6.0 and +3.5 BLEU\%, respectively.
\section*{Acknowledgments}
\label{sec:ack}
This work has been supported by the European Union's Horizon 2020 Research and Innovation Programme under Grant Agreement No 780069, and by the Academy of Finland in the project 313988.
In addition the Finnish IT Center for Science (CSC) provided computational resources.
We would also like to acknowledge the support by NVIDIA and their GPU grant.
|
\section{Introduction}
\label{sec:intro}
Emotion recognition in text is the task of associating words, phrases
or documents with predefined emotions drawn from psychological models
\cite{Ekman1999,Plutchik2001}. In this paper, we phrase it as single
label classification of \emotionword{joy}, \emotionword{anger}, \emotionword{fear}, \emotionword{sadness}, \emotionword{surprise}, and
\emotionword{disgust}. It has been applied to, \textit{e.\,g.}\xspace, tales \cite{Alm2005}, blogs
\cite{Aman2007}, and as a very popular domain, microblogs on Twitter
\cite{dodds2011temporal}. The latter in particular provides a large
source of data in the form of user messages \cite{costa2014}, often
used with self-assigned classes by the authors, as this can lead to a
huge albeit noisy data set \cite{Wang2012}. This procedure is often
referred to as \textit{self-labeling}, or, in general, as distant
labeling.
Nowadays, state-of-the-art classification models for emotion
prediction typically take into account sequential information, for
instance with recurrent neural networks or convolutional neural
networks \cite{Felbo2017,Koper2017}. Clearly, these models are able to
capture information expressed in phrases, for instance modifications
of an emotion phrase, like in ``I am slightly unhappy.'' However, such
models do not allow for obtaining a better semantic and linguistic
understanding of the meaning of modifications of emotion expressions
\textit{per se}.
We aim in this paper at getting a better understanding of the impact
and use of modifications of emotion words in Twitter. We perform
modifier cue detection and subsequently identify their scope.
Modifiers are commonly divided into intensifiers (which assign an
intensity to a word) and negators (\textit{e.\,g.}\xspace, \textit{not}), amongst other
classes. Intensifiers are further separated into amplifiers
(\textit{very}, \textit{entirely}, we do not distinguish maximizers
and boosters) and downtoners (\textit{quite}, \textit{slightly})
\cite{Randolph1985}. We focus on these three modifiers:
\textit{negations, amplifiers}, and \textit{downtoners}. From these,
negations are most studied and most challenging in interpretation. For
instance, ``not sad'' might express \emotionword{joy}, \emotionword{fear}, or \emotionword{anger}, or none of
the above. We will argue later that it is closer to expressing \emotionword{joy}
than to \emotionword{anger} or \emotionword{fear}.
Similarly, downtoners might change the prior emotion (\textit{i.\,e.}\xspace, the emotion
of a word or phrase without considering context) of an
expression. However, we will see that for instance ``slightly sad''
most likely still expresses the prior emotion \emotionword{sadness} but also
changes the other emotions which can be expressed by the same sentence
at the same time.
Intensifications (\textit{e.\,g.}\xspace, ``very sad'') seem to be straight-forward in
interpretation. We will argue that such formulations separate the
prior emotion (\emotionword{sadness}) of the word more clearly from a secondary
emotion to be predicted (\textit{e.\,g.}\xspace, \emotionword{fear}).
This research is similar to analyses of the meaning of negations in
the context of sentiment
\cite{Wiegand2010,Kiritchenko2016,Ruppenhofer2015}. However, the
degree of freedom for interpretation is increased due to the greater
set of classes (emotion categories \textit{vs.}\xspace polarity). The only work in the
context of emotions with modifiers we are aware of is by Carillo
\textit{et al.} \cite{Carillo2013}. They focus on the classification
task of sentiment but treat modifiers emotion-specific. In contrast,
we aim at classifying emotions particularly to analyze the role of
modifiers. More specifically, we (1), select and evaluate an
appropriate \emph{modifier scope detection method in the context of
emotion words} on a manually annotated corpus which we make publicly
available\footnote{The data used in this study is available at
\url{http://www.ims.uni-stuttgart.de/data/modifieremotion}.}, and
(2), evaluate the impact of the best performing approach in a
\emph{bag-of-words model showing its value for emotion
classification}. Finally, (3), as the main contribution, we develop
a simple lexical model in which emotion words are weighted differently
based on their modifier scope and prior emotion for the purpose of
\emph{model introspection}: The weights serve as a tool to study the
\emph{meaning of modified emotion words}.
\section{Background and Related Work}
\label{sec:previous}
\subsection{Emotion Analysis}
Ekman defines \emotionword{joy}, \emotionword{anger}, \emotionword{fear}, \emotionword{sadness}, \emotionword{surprise} and \emotionword{disgust} as
the minimal set of six basic emotions that can be differentiated by
facial expressions, the set we use in this paper \cite{Ekman1992}.
Plutchik adds \emotionword{anticipation} and \emotionword{trust} and
the concepts of intensity, emotion mixtures and opposing classes to
the model, which we analyze empirically here \cite{Plutchik2001}.
The first text collection which is nowadays used for emotion
classification is the ISEAR corpus of descriptions of emotional events
\cite{Scherer1997}. Alm \textit{et al.}\ were the first discussing
issues of annotation and prediction of emotions in tales
\cite{Alm2005}. Aman \textit{et al.}\ built classifiers on top of
blog posts \cite{Aman2007}. Headlines were the subject of analysis in
the SemEval competition on affect recognition \cite{Strapparava2007}.
Next to these manually built corpora, Wang \textit{et al.}\ generated
a training corpus by using the so-called self-labeling information
provided by authors of tweets with their hashtags
\cite{Wang2012}. Their results show that the performance of an emotion
classification system can be significantly improved with a large
amount of data. Similarly, \cite{Purver2012} use self-labeling with
emoticons and hashtags. The first manually-annotated corpus of tweets
for emotion analysis made publicly available was provided by
\cite{Mohammad2015}, followed by a larger set with a focus on emotion
intensity prediction \cite{Mohammad2017}. The corpus by
\cite{Schuff2017} provides multiple annotations of each instance and
analyzes interactions between classes. It is a re-annotation of a
SemEval corpus for stance detection \cite{Mohammad2016}.
\subsection{Modifier Detection for Sentiment and Opinion Analysis}
\label{sec:mod_detect}
Negations have been extensively studied in different contexts.
Chapman \textit{et al.}\ use a list of negation cue phrases and assume
the scope to include all tokens up to the next punctuation mark or to
the next adversative conjunction \cite{Chapman2001}. Pang \textit{et
al.}\ include negation detection in a sentiment document
classification system \cite{Pang2002}.
On a more fine-grained level, Councill \textit{et al.} use a lexicon
for negation cue detection and a linear-chain conditional random field
for scope recognition, based on part of speech tags and dependency
relations \cite{Councill2010}. Reitan \textit{et al.} use a similar
approach on a tweet corpus \cite{Reitan2015}. Jia \textit{et al.}\ use
rules based on typed dependencies to determine the scope of a negation
cue \cite{Jia2009}.
A straight-forward approach to modify features in a machine
learning-based text classifier with negation information is to prepend
modified entries in the bag of words (\textit{i.\,e.}\xspace, create an additional bag of
modified words in addition to non-modified words, \textit{e.\,g.}\xspace,
\cite{Pang2002}). For a word-list-based classifier, Polanyi \textit{et
al.}\ propose to classify a document as positive or negative based
on the sum of weights of positive and negative words
\cite{Polanyi2006}. Positive words have a weighting of $+2$, while
negative words have a weighting of $-2$. If a word is negated, its
weight is multiplied with $-1$. If a word is amplified, its weighting
is modified additively (to $+3$ or $-3$, respectively) and, if it is
modified by a downtoner accordingly (to $+1$ or $-1$,
respectively). Kennedy \textit{et al.}\ showed an improvement with
this approach on movie review classification
\cite{kennedy2006sentiment}. Follow-up work investaged the use of
negations and modality in a linguistic experiment and also model
multiple negations in the same expression \cite{Benamara2012}.
Taboada et al. discuss lexicon-based methods for sentiment analysis in
a broader context \cite{Taboada2011}. More recent work developed
machine-learning-based classifiers to detect speculation and negations
particularly for sentiment analysis \cite{Cruz2015}.
We are not aware of any previous work on modifier detection for
emotion expressions with the goal of emotion prediction. However,
Carillo \textit{et al.}\ build a model for sentiment classification in
which they learn weights of modifications for an improved polarity
prediction \cite{Carillo2013}.
For a more comprehensive overview of previous work in negation and
modifier detection in sentiment analysis, we refer to surveys and
reviews previously published
\cite{Wiegand2010,Zhu2014,morante2012sem}.
\section{Methods}
\label{chap:methods}
We first aim at showing empirically that handling emotion words
specifically with negations, amplifications, and downtoners improves
the classification in contrast to a purely word-based model. We
describe our modifier cue detection methods (Section~\ref{sec:mcd}),
explain the modifier scope detection (Section~\ref{sec:MSD}) and
present a simple bag-of-words based method to evaluate the impact of
modifier detection (Section~\ref{sec:emo_class}).
\subsection{Modifier Cue Detection}
\label{sec:mcd}
We limit ourselves to modifications of emotions, in which the modifier
cue $t$ is explicitly mentioned and build on top of existing modifier
lists of negations (\textit{e.\,g.}\xspace, cannot, never, not), amplifiers (\textit{e.\,g.}\xspace,
extremely, very, lot), and downtoners (\textit{e.\,g.}\xspace, few, less, rarely, some)
and merge them
\cite{wikiintensifier,englishclub,thomson1986practical,Romero2012,benamara2007sentiment,Councill2010}. For
a discussion of implicit emotion detection, we refer the reader to our
recent work on the implicit emotion shared task \cite{Klinger2018x}. We
do not differentiate maximizers and boosters \cite{Ito2003}. To focus
our study to those terms which are predominantly used as modifier
instead of other meanings, we calculate
\[r^t_{\textrm{mod}}=\frac{\#t\textrm{ used as modifier}}{\#\textrm{
used}}\] with
$\textrm{mod}\in\{\textrm{downtoner},\textrm{amplifier},\textrm{negation}\}$
and $\#$ denoting the count. We estimate this value on a corpus
subsample of 100 tweets for each $t$. We accept $t$ as modifier iff
$r^t_{\textrm{mod}}>0.5$ to ensure the main role of a term to be a
modifier. For instance, we dismissed the amplifier \textit{too}, as it
is used more often in a non modifying context. The resulting
dictionaries have 39 negation terms, 69 amplifier terms and 36
downtoner terms
\begin{table}[t]
\centering\footnotesize
\caption{Features for modifier scope classification\hspace{\textwidth} (proposed by \cite{Councill2010} except for *).}
\label{features}
\begin{tabularx}{\linewidth}{lX}
\toprule
Feature & \multicolumn{1}{l}{Description} \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Word & Normalized string of a token. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
POS & Part of speech of a token. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Right Dist. & Token distance to the nearest explicit modifier cue in the sentence to the right of a token. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Left Dist. & Token distance to the nearest explicit modifier cue in the sentence to the left of a token. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Dep Dist. * & Minimum number of edges that must be traversed in the dependency tree from a token to an explicit modifier cue. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Dep1 POS & Part of speech of the the first order parent of a token. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Dep1 Dist. & Minimum number of edges that must be traversed in the dependency tree from the first order parent of a token to an explicit modifier cue. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Dep2 POS & Part of speech of the second order parent of a token. \\
\cmidrule(lr){1-1} \cmidrule(lr){2-2}
Dep2 Dist. & Minimum number of edges that must be traversed in the dependency tree from the second order parent of a token to an explicit modifier cue. \\
\bottomrule
\end{tabularx}
\phantom{x}\\[1mm
\end{table}
\subsection{Modifier Scope Detection}
\label{sec:MSD}
As we are specifically interested in the importance and meaning of
modifiers on emotion terms (and not on other words), we only take them
into account in the predictive models where appropriate. We therefore
compare three approaches for modifier scope detection and select the
best performing one.
\subsubsection{Next-$n$ Heuristic}
\label{subsec:nextn}
As a combination of previous work for modifier handling, we define
maximally $n$ tokens as the scope which follow the cue up to the next
punctuation mark or adversative conjunction
\cite{Pang2002,Chapman2001,Hu2004}. For example, in the tweet
``Happiness is not a goal; it is a by-product.''
the words ``a'' and ``goal'' would be in
the negation scope (with any $n \geq 2$), but not the words following
the semicolon.
\subsubsection{Dependency Tree Heuristic}
\label{sec:DepTree}
We extend the approach by \cite{Jia2009} to our set of modifiers
and specifically emotion words in a heuristic on dependency trees
(generated with Stanford CoreNLP 3.7.0, \cite{manning2014}): We flag
every parent as modified if its direct child corresponds to a modifier
cue. For instance, in Figure~\ref{dep_tree_1}, ``love'' is recognized
as negated because ``not'' is in our negation lexicon.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{dep_tree_1.pdf}
\caption{Dependency tree example.}
\label{dep_tree_1}
\end{figure}
To recognize ``hate'' as being in scope as well, we propagate the
modification information along conjunction edges. Adversative
conjunctions block this propagation.
\subsubsection{Binary SVM}
Similarly to a set of submissions to the shared task on negation scope
detection \cite{morante2012sem}, our third approach is a
classification of tokens with linear support vector machines
(SVM). For each modifier, we train one separate model to predict for a
candidate token if it is modified or not. We assume that a token
cannot be modified twice. The priority of our classifiers is negation
detection, then amplifier detection, followed by downtoner detection.
We use features previously proposed \cite{Councill2010} (\textit{cf.}\xspace
Table~\ref{features}). POS and dependencies are recognized with the
Stanford CoreNLP tools. As an example, the features for the word
``hate'' in Figure~\ref{dep_tree_1} are: \textit{Word = hate},
\textit{POS = VB}, \textit{Right Dist. = 0} (no modifier cue to the
right), \textit{Left Dist. = 3}, \textit{Dep Dist. = 0} (is leaf
node), \textit{Dep1 POS = VB}, \textit{Dep1 Dist. = 1}, \textit{Dep2
POS = null} (first order parent is root node), \textit{Dep2
Dist. = 0}.
\subsection{Emotion Classification}
\label{sec:emo_class}
The classification task is to assign a tweet to one of the emotions
from \emotionword{joy}, \emotionword{anger}, \emotionword{fear}, \emotionword{sadness}, \emotionword{surprise}, and \emotionword{disgust}. Note that
we opt for not using a model which can take into account sequential
information (\textit{e.\,g.}\xspace, a long short-term memory, a convolutional neural
net, an $n$-gram model, or non-linear kernels), because the impact of
the modifier detection would be ``hidden'' in the handling of
sequences in general. In contrast, we use a linear support vector
machine with only unigram features such that the SVM is not able to
capture modification effects itself. With this approach we might not
reach highest performance but obtain a model suitable to study
modification effects.
\section{Emotion Classification Experiments under Consideration of Modifiers}
\label{sec:experiments}
In the following Section~\ref{sec:corpora}, we discuss the corpora
used for our evaluation shown in Section \ref{sec:results}, which
shows and discusses the results of our experiments.
\begin{table}
\centering
\caption{Emotion classification corpora.}
\label{emo_corpora}
\begin{tabular}{lrrrrr}
\toprule
emotion & \textsc{trainRepr}\xspace & \textsc{testRepr}\xspace & \textsc{train}\scalebox{0.8}{\faBalanceScale}\xspace \\
\cmidrule(r){1-1} \cmidrule(lr){2-2} \cmidrule(lr){3-3}\cmidrule(l){4-4}
joy & 597.992& 299.028& 1.000\\
anger & 59.591& 29.501& 1.000\\
fear & 68.886& 34.504& 1.000\\
sadness & 207.026& 103.607& 1.000\\
surprise & 24.582& 12.483& 1.000\\
disgust& 1.923& 877& 1.000\\
\cmidrule(r){1-1} \cmidrule(lr){2-2} \cmidrule(lr){3-3}\cmidrule(l){4-4}
total & 960.000& 480.000& 6.000\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{Modifier scope detection corpora.}
\label{mod_corpora}
\begin{tabular}{lrrrr}
\toprule
Modifier & \multicolumn{1}{c}{\rt{\textsc{modEval}}} &
\multicolumn{1}{c}{\rt{\textsc{trainNeg}}} &
\multicolumn{1}{c}{\rt{\textsc{trainAmp}}} & \multicolumn{1}{c}{\rt{\textsc{trainDown}}} \\
\cmidrule(r){1-1} \cmidrule(lr){2-2} \cmidrule(lr){3-3} \cmidrule(lr){4-4} \cmidrule(lr){5-5}
negation & 315& 630& 0& 0\\
amplifier & 249& 0& 497& 0\\
downtoner & 74& 0& 0& 148\\
\cmidrule(r){1-1} \cmidrule(lr){2-2} \cmidrule(lr){3-3} \cmidrule(lr){4-4} \cmidrule(lr){5-5}
total & 638& 630& 497& 148\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Corpora}
\label{sec:corpora}
\subsubsection{Self-Labeling for Emotion Classification}
To generate corpora of substantial size, we use a self-labeling
approach: we retrieve tweets with specific hashtags for each emotion
using the REST and Streaming APIs provided by Twitter. The hashtags
are \#glad, \#happiness, \#happy, \#joy, \#lucky, \#luck, and
\#pleasure for \emotionword{joy}, \#anger, \#hate, \#hatred, and \#rage for \emotionword{anger},
\#afraid, \#angst, \#fear, \#panic, \#scare, and \#worry for \emotionword{fear},
\#bitter, \#grief, \#misery, \#sad, \#sadness, and \#sorrow for
\emotionword{sadness}, \#surprise and \#surprised for \emotionword{surprise}, and \#disgust for
\emotionword{disgust}.
We assume this hashtags to denote the label of the respective tweets
to create a large dataset. We replace hashtags, URLs, and usernames by
the same strings, respectively. Table~\ref{emo_corpora} shows our
separation of the crawled data into train and test sets for emotion
classification. The corpora \textsc{trainRepr}\xspace and \textsc{testRepr}\xspace are uniformly
sampled randomly. We use these two corpora to train our emotion
classifier and to evaluate the real world performance and modifier
impact. Additionally, we create the corpus \textsc{train}\scalebox{0.8}{\faBalanceScale}\xspace which will
be discussed in Section~\ref{sec:emotionlexicon}.
\subsubsection{Manually-annotated Corpora}
To select the best performing modifier scope detection method and to
estimate their performance, we manually annotate a corpus which is
also used for the SVM scope detection model training. The annotation
is performed by one author of the paper. The task is to categorize
pairs of an emotion-bearing word $z_e$ with a modifier word
$z_{\textrm{mod}}$ into ``$z_{\textrm{mod}}$ modifies $z_e$'' or
``not''. For instance, Figure~\ref{annotation_example} visualizes
that \textit{not} modifies \textit{love} and \textit{very} modifies
\textit{hate}. However, \textit{not} does not modify \textit{hate} and
\textit{very} does not modify \textit{love}. We therefore have four
instances with two positive and two negative annotations for two
different modifiers and two emotion words.
\begin{figure}[t]
\centering
\includegraphics[scale=0.2]{annotation.pdf}
\caption{Annotation example.}
\label{annotation_example}
\end{figure}
The resources we create should be valuable also outside of our
specific parameter setting. For instance, our selection of dictionary
entries cannot be complete. Therefore, in the annotation process, the
annotators do not see automatically detected modifiers or
automatically recognized emotion terms but need to mark them
themselves such that the corpus quality is not decreased by error
propagation from preprocessing steps.
Therefore, more specifically, we use three different sampling methods
to obtain a corpus densily populated with relevant instances, but not
limited to those detected with our resources: Equally-sized subsets
are sampled based on the occurrence of (1) both modifier cue and
emotion word, (2) only modifier cue, (3) only emotion word. Using
different sampling methods enables us to expand our emotion and
modifier lexicons with emotion-bearing words and modifier cues found
during annotation. We annotate 1,000 tweets resulting in 1,913
modifier-emotion word pairs. Table~\ref{mod_corpora} summarizes the
annotation, split into subcorpora for training the modifier detectors
and an evaluation set: The corpus \textsc{modEval} contains one-third
of the annotations from each modifier type. We use this corpus to
evaluate the performance of the different modifier scope detection
approaches. Furthermore, we create three corpora \textsc{trainNeg},
\textsc{trainAmp} and \textsc{trainDown} containing the remaining two
thirds of annotations for scope detection model training. The table
also shows that of all detected modifiers, downtoners are the least
common ones.
\begin{table*}[t]
\centering
\caption{Comparison of modifier detection methods on \textsc{modEval}
corpus. The results of the best method for each modifier and the average are highlighted in boldface
for precision, recall, and $\text{F}_1$\xspace, respectively.}
\label{mod_detect_compare}
\begin{tabular}{l|lccccccccc}
\toprule
\multicolumn{2}{l}{}& \multicolumn{3}{c}{Next-2} &
\multicolumn{3}{c}{DepTree} & \multicolumn{3}{c}{SVM} \\
\cmidrule(l){3-5}\cmidrule(rl){6-8}\cmidrule(l){9-11}
\multicolumn{1}{l}{}& Modifier & P & R & F$_{1}$ & P & R & F$_{1}$ & P& R & F$_{1}$ \\
\cmidrule(r){2-2}\cmidrule(lr){3-3}\cmidrule(r){4-4}\cmidrule(l){5-5}
\cmidrule(lr){6-6}\cmidrule(r){7-7}\cmidrule(l){8-8}
\cmidrule(lr){9-9}\cmidrule(r){10-10}\cmidrule(l){11-11}
& Negator & \textbf{93.6}& 87.9& \textbf{90.7}& 93.0& 80.4& 86.2 & 78.7 & \textbf{89.4} & 83.7\\
& Amplifier & \textbf{91.7}& \textbf{93.7}& \textbf{92.7}& 90.7& 83.0& 86.7 & 91.4 &89.4 & 90.4\\
& Downtoner& 72.8& \textbf{88.9}& \textbf{80.0}& \textbf{75.0}& 50.0& 60.0 & 66.7 & 55.6 & 60.7\\
& Macro-avg. & 86.0& \textbf{90.2}& \textbf{87.8}& \textbf{86.3}& 71.1& 77.7 & 78.9 & 78.2 & 78.3\\
\bottomrule
\end{tabular}
\end{table*}
\subsection{Results}
\label{sec:results}
\subsubsection{Modifier Scope Detection}
\label{sec:modifierscopedetection}
The results of the selection
of parameter $n$ in the \textit{next-$n$ method}
(Section~\ref{subsec:nextn}) on the training corpora are shown in
Figure~\ref{fig:nextn}. The best result is obtained for $n=2$.
\begin{figure}[t]
\centering
\begin{tikzpicture}[scale=0.9]
\begin{axis}[
axis lines = left,
xlabel = $n$,
ylabel = {$\text{F}_1$\xspace-measure},
legend style={draw=none}
]
\addplot+[sharp plot] coordinates
{(1,80.4) (2,88.3) (3,82.7) (4,79.0) (5,74.5) (6,71.7) (7,69.0) (8,66.4) (9,64.6) (10,63.7) (11,62.6) (12,61.9)};
\addlegendentry{All modifiers}
\addplot+[sharp plot] coordinates
{(1,73.5) (2,90.8) (3,83.8) (4,77.4) (5,73.5) (6,70.8) (7,66.9) (8,64.8) (9,63.9) (10,63.1) (11,61.2) (12,60.6)};
\addlegendentry{Negations}
\addplot+[sharp plot] coordinates
{(1,89.9) (2,92.3) (3,88.9) (4,86.4) (5,82.4) (6,79.7) (7,78.1) (8,75.8) (9,74.7) (10,74.0) (11,73.8) (12,73.3)};
\addlegendentry{Amplifiers}
\addplot+[sharp plot] coordinates
{(1,78.4) (2,85.0) (3,80.0) (4,78.8) (5,74.8) (6,73.3) (7,70.5) (8,67.3) (9,65.0) (10,63.8) (11,62.8) (12,62.2)};
\addlegendentry{Downtoners}
\end{axis}
\end{tikzpicture}
\caption{Different values of $n$ for next-$n$ modifier
detection, evaluated on the \textsc{trainNeg}, \textsc{trainAmp},
\textsc{trainDown} corpora.}
\label{fig:nextn}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.8]{tensor}
\caption{Visualization of our four weight matrices for
dictionary-based emotion recognition with modifiers. The orange
slice corresponds to occurrences of input words with prior emotion
fear, the blue slice to output emotion joy and the red
slice to input words in the scope of a negation. The cell in all
three slices contains the weight such word contributes to the
overall classification.}
\label{fig:tensor}
\end{figure}
This value goes against our expectations, as \cite{Reitan2015}
detected an average of $n = 3.8$ to work best for negations on Twitter. One reason
might be that we do not consider the full scope of a modifier but limit our analysis to emotion words only.
Table~\ref{mod_detect_compare} shows the results on
\emph{modEval} for emotion scope detection. The simplest method, the
next-$2$ heuristic, shows best results throughout all modifier
types. The performance for downtoners is substantially lower than for
negations and amplifiers.
The SVM method (83.7\,\% $\text{F}_1$\xspace for negations) achieves comparable results
to the approach by \cite{Councill2010} (80\,\% $\text{F}_1$\xspace on product
reviews). The main source of errors for the DepTree approach are errors in the
dependency trees because of missing punctuation.
The source of errors for the comparably low
performance on downtoner scope detection depends on the method. For
the next-$2$ approach, a challenge is that downtoner cues appear more
often after the scope. For the DepTree approach, we observe that
downtoner cues are more often not a direct child of associated emotion
words. Regarding the SVM approach, we presume two main reasons for the
limited performance: Firstly, we prioritize negations and amplifiers
and secondly, because we have a limited training set for downtoner.
\begin{table*}[t]
\centering
\caption{Results on \textsc{testRepr}\xspace corpus and subcorpora limited to instances
with respective modifiers, with and without modifier detection. The
SVM bag-of-words model is trained with unigram features on
\textsc{trainRepr}\xspace. The lexicon weight matrices are
trained on \protect\textsc{train}\scalebox{0.8}{\faBalanceScale}\xspace.}
\label{testAll}
\renewcommand*{\arraystretch}{1}
\setlength\tabcolsep{4.1mm}
\newcolumntype{P}{>{\centering\arraybackslash}p{10mm}}
\begin{tabular}{l|l|r|rrrrrr}
\toprule
\multicolumn{3}{c}{} & \multicolumn{6}{c}{SVM bag of words}\\
\cmidrule(lr){4-9}
\multicolumn{3}{c}{} & \multicolumn{3}{c}{w/o mod.\ det.} &\multicolumn{3}{c}{w/ next2 heur.}\\
\cmidrule(lr){4-6}\cmidrule(lr){7-9}
\multicolumn{1}{c}{}& emotion & size & P&R&$\text{F}_1$\xspace & P&R&$\text{F}_1$\xspace\\
\cmidrule(r){1-2}\cmidrule(lr){3-3}\cmidrule(lr){4-6}\cmidrule(lr){7-9}
\multirow{7}{5mm}{\rt{all data}}
& joy & 299,028 & 82.0 & 94.6 & 87.9 & \textbf{83.2} & 93.8 & \textbf{88.2} \\
& anger & 29,501 & 68.3 & 32.2 & 43.7 & 65.3 & \textbf{34.5} & \textbf{45.1} \\
& fear & 34,504 & 77.4 & 50.7 & 61.3 & 76.6 & \textbf{53.5} & \textbf{63.0} \\
& sadness & 103,607 & 74.1 & 66.6 & 70.1 & 72.6 & \textbf{68.2} & \textbf{70.3} \\
& surprise& 12,483 & 75.3 & 32.3 & 45.2 & 72.3 & \textbf{33.3} & \textbf{45.6} \\
& disgust & 877 & 18.8 & 3.5 & 5.8 & 17.3 & 3.2 & 5.4 \\
& Macro & 480,000 & 66.0 & 46.6 & 52.3 & 64.6 & \textbf{47.8} & \textbf{52.9} \\
\cmidrule(r){1-2}\cmidrule(lr){3-3}\cmidrule(lr){4-6}\cmidrule(lr){7-9}
\multirow{7}{5mm}{\rt{negations}}
& joy & 22,459 & 70.5 & 83.9 & 76.7 & \textbf{72.3} & \textbf{85.4} & \textbf{78.3} \\
& anger & 5,686 & 61.9 & 35.1 & 44.8 & \textbf{64.3} & \textbf{37.4} & \textbf{47.3} \\
& fear & 6,685 & 75.1 & 50.0 & 60.0 & 70.5 & \textbf{57.8} & \textbf{63.5} \\
& sadness & 24,299 & 75.0 & 79.0 & 77.0 & \textbf{77.6} & \textbf{79.1} & \textbf{78.3} \\
& surprise & 1,122 & 39.8 & 12.3 & 18.8 & \textbf{40.7} & \textbf{12.5} & \textbf{19.1} \\
& disgust & 165 & 31.3 & 3.1 & 5.6 & 22.8 & 3.1 & 5.4 \\
& Macro & 60,416 & 58.9 & 43.9 & 47.2 & 58.1 & \textbf{45.9} & \textbf{48.7} \\
\cmidrule(r){1-2}\cmidrule(lr){3-3}\cmidrule(lr){4-6}\cmidrule(lr){7-9}
\multirow{7}{5mm}{\rt{amplifiers}}
& joy & 23,622 & 79.6 & 90.8 & 84.9 & \textbf{80.6} & 90.0 & \textbf{85.0} \\
& anger & 3,300 & 61.1 & 29.9 & 40.2 & \textbf{64.0} & 29.2 & 40.1 \\
& fear & 3,017 & 72.6 & 48.7 & 58.3 & 67.9 & \textbf{55.8} & \textbf{61.3} \\
& sadness & 15,773 & 76.9 & 77.0 & 77.0 & 76.1 & \textbf{77.5} & 76.8 \\
& surprise & 872 & 50.4 & 17.5 & 25.9 & \textbf{51.4} & 16.9 & 25.4 \\
& disgust & 109 & 28.6 & 3.7 & 6.6 & 23.9 & \textbf{04.6} & \textbf{7.7} \\
& Macro & 46,693 & 61.5 & 44.6 & 48.8 & 60.7 & \textbf{45.7} & \textbf{49.4} \\
\cmidrule(r){1-2}\cmidrule(lr){3-3}\cmidrule(lr){4-6}\cmidrule(lr){7-9}
\multirow{7}{5mm}{\rt{downtoners}}
& joy & 7,900 & 78.2 & 91.1 & 84.1 & \textbf{79.8} & 90.7 & \textbf{84.9} \\
& anger & 979 & 51.9 & 22.0 & 30.9 & \textbf{62.6} & \textbf{35.0} & \textbf{44.9} \\
& fear & 980 & 71.6 & 43.3 & 54.0 & 63.4 & \textbf{48.6} & \textbf{55.0} \\
& sadness & 4,232 & 73.5 & 72.5 & 73.0 & \textbf{74.7} & 72.1 & \textbf{73.4} \\
& surprise & 370 & 54.3 & 15.5 & 24.0 & 46.0 & 15.2 & 22.8 \\
& disgust & 25 & 50.0 & 4.0 & 7.5 & \textbf{66.7} & \textbf{16.0} & \textbf{25.9} \\
& Macro & 14,486 & 63.2 & 41.4 & 45.6 & \textbf{65.5} & \textbf{46.3} & \textbf{51.1} \\
\bottomrule
\end{tabular}
\end{table*}
\subsubsection{Emotion Classification under Consideration of
Modifiers}
In this paper, we aim at analyzing the impact of negations,
amplifiers, downtoners and to understand their contribution to emotion
analysis, mainly as a justification to further inspect their role on
emotion-bearing words. Therefore, in this result section, we show that
our hypothesis that they affect the interpretation of emotion words
actually holds.
To achieve that, we test our systems on a uniform subsample from
Twitter, namely \textsc{trainRepr}\xspace/\textsc{testRepr}\xspace, which has a real-world
distribution of modifiers and non-modified emotions. For inclusion of
modifier detection, the bag-of-word features of tokens in the scope
are prefixed with respective abbreviations (amp, down, neg) and use
the next-$2$-heuristic.
Table~\ref{testAll} shows the results for SVM classification on four
different subsets of data, namely the full data set for training
(\textsc{trainRepr}\xspace) and testing (\textsc{testRepr}\xspace) (called ``all data'' in the table),
the subset of data which contains at least one negator, one
amplifier, or one downtoner, respectively. For these subsets, only
the respective modifier detection is applied.
Altogether, the classifier is best performing on \emotionword{joy}, followed by
\emotionword{sadness} and \emotionword{fear}. The modifier detection contributes consistently,
though partially only to a limited degree, to all class predictions (on
all data for \emotionword{joy} with $+.3$, \emotionword{anger} with $+1.4$, \emotionword{fear} with $+1.7$,
\emotionword{sadness} with $+.2$). Most of the improvement originates from an
increase in recall when training and testing on all data. When
limiting the experiment to different modifiers, we see that this is
likely a result of the negation detection, while amplifiers and
downtoners contribute partially to precision and partially to recall,
depending on the respective emotion.
Inspecting the contribution by modification, we observe the strongest
contribution over the model without handling modifications for
downtoners, with an improvement of $+5.5$ percentage points (pp). Here, 14
pp improvement originate from the emotion \emotionword{anger} and 18 from
\emotionword{disgust}.
Across all modifiers, most important is the special handling of fear,
with 3.5 pp in negations and 3 pp in amplifiers.
\section{Analysis of the Impact of Modifiers in the Context of Emotion
Words}
We showed in the previous section that modifier detection improves
classification in a bag-of-words model. Now we come to the main
contribution of this paper, a deeper analysis of the meaning of
negators, amplifiers, and downtoners on emotion words.
\subsection{Experimental Setting: Weighted Emotion Lexicon}
\label{sec:emotionlexicon}
For this analysis, we extend the work by Polanyi \textit{et al.} from
shifting values in one dimension of polarity according to different
modifiers to multiple dimensions, \textit{i.\,e.}\xspace, six fundamental emotions
\cite{Polanyi2006}. In addition, instead of proposing a fixed set of
weights, we estimate these from data. We use the NRC lexicon for
emotion word recognition, similar to lists of positive/negative words
\cite{Polanyi2006,Mohammad2013}.
The parameters of the model are represented in four $6\times 6$
matrices $W_{\text{no-mod}}$, $W_{\text{amp}}$, $W_{\text{down}}$,
$W_{\text{neg}}$. In each matrix, one cell $w_{ij}$ corresponds to
the weight which a word of emotion $e_i$ in the respective
modification scope contributes to the emotion $e_j$. This data
structure is visualized in Figure~\ref{fig:tensor}.
Input text is represented as four count vectors of length 6
($\vec{x}_{\text{no-mod}}$, $\vec{x}_{\text{amp}}$,
$\vec{x}_{\text{down}}$ $\vec{x}_{\text{neg}}$) of words whose scope
contains emotion words of the respective emotion. For instance,
$x_{\text{down},i}$ is the count of downtoned words which belong to
$e_i$.
The posterior emotion score vectors resulting from words of specific
modification scopes for an input text $x$ are then
\[
\vec{e}_{\text{mod}} = W^T_{\text{mod}}\times\vec{x}_{\text{mod}}
\]
with $\text{mod} \in
\{\text{no-mod},\text{amp},\text{down},\text{neg}\}$.
The overall emotion score is then the element-wise sum across rows
\[
\vec{e}(x) = \sum_{\textrm{mod}}\vec{e}_{\text{mod}}
\]
of these vectors.
Finally, the decision for an input text is
\[
e(x) = \argmax_{i} (e_i(x))\,,
\]
where $i$ corresponds to one of the basic emotions.
Based on this setting, we optimize the weights on a balanced corpus
\textsc{train}\scalebox{0.8}{\faBalanceScale}\xspace to further develop an understanding of the meaning of
modifiers by model inspection. The weights are not influenced by
different training set sizes which would make interpretation
difficult. It only includes tweets containing at least one emotion and a
modifier word. As optimization paradigm, we use hill climbing and $\text{F}_1$\xspace
as the objective function. We do random restarts with initialization
of $w \sim \mathcal{N}(0,1)$ and take the best matrix from the set of
optimization results. The slice $W_{\text{mod}}$ for each modifier is
optimized for $\approx 120$ hours, resulting in 28 optimization runs
with 2720 epochs on average for the neutral matrix, 49 optimization
runs with 1391 epochs for the negative matrix, 53 optimization runs
with 1248 epochs for the amplifier matrix and 64 optimization runs
with 990 epochs for the downtoner matrix. Weight updates are
performed as $w' = w + r$ with $r \sim \mathcal{N}(0,1)$. We stop
optimization if no improvement is observed in 500 epochs.%
\footnote{We do not report the results of the prediction of this model on
independent data as it is outperformed by the SVM
classification. Instead, we focus on the analysis of the model
parameters in the following.}
\begin{figure*}[t]
\centering
\includegraphics[page=1,scale=0.58]{heatmaps-crop}\hfill
\includegraphics[page=2,scale=0.58]{heatmaps-crop}\hfill
\includegraphics[page=3,scale=0.58]{heatmaps-crop}\hfill
\includegraphics[page=4,scale=0.58]{heatmaps-crop}
\caption{Weighting matrices for the lexical model. Columns correspond
to the predicted emotion, rows correspond to the prior emotion of
the observed words.}
\label{bestweightingmatrices}
\end{figure*}
\subsection{Analysis of Weighting Matrices in the Lexical Model}
The results of this optimization procedure are shown in
Figure~\ref{bestweightingmatrices}.
We discuss the results based on the following hypotheses:
\emph{Words outside of modifier scope} mainly contribute positively to
the emotion classification corresponding to their prior emotion and
negatively to emotions of opposing polarity. \emph{Words in negation
scope} contribute to emotions of their opposing polarity or express
no emotion. \emph{Words in amplifier scope} contribute more to
emotions of their prior emotion than words outside of modifier scope.
\emph{Words in downtoner scope} contribute less to emotions of their
prior emotion than words outside of modifier scope.
\paragraph{Emotion words outside of modification scope}
The hypothesis is supported by the matrix; \textit{i.\,e.}\xspace, each word of each
emotion mainly contributes to their prior emotion with the highest
weight for \emotionword{surprise}, followed by \emotionword{fear}, \emotionword{anger}/\emotionword{disgust}, and \emotionword{joy}
(\textit{i.\,e.}\xspace, in a text with \emotionword{joy} and \emotionword{fear} words, both outside of modifier
context, the classification output would be \emotionword{fear}). We observe a
positive contribution of emotion words to other than their prior
emotion for those of same polarity, namely anger to \emotionword{fear} (0.8), \emotionword{fear}
to \emotionword{sadness} (0.4), and \emotionword{disgust} to \emotionword{sadness}\footnote{Example: ``They
are 'terrorists' not 'Islamists', you {pathetic} excuse for a
journalist !!!! \#hate...''} (0.2). Contrasting our expectation,
\emotionword{surprise} contributes to \emotionword{sadness} (0.8), \emotionword{disgust} (0.6), and
\emotionword{joy}\footnote{Example: ``Still can't believe my cute baby shower
\#afternoontea \#surprise \#ourgirl''} (0.5), showing that \emotionword{surprise}
can be divided into positive and negative realizations. The negative
contribution of \emotionword{anger} words is striking for \emotionword{joy} ($-3.1$) and
\emotionword{surprise} ($-3.0$), supporting the second of the hypotheses.
\paragraph{Emotion words inside a negation scope}
The hypothesis that negated words mainly contribute to emotions of
opposite polarity holds for \emotionword{joy} to \emotionword{sadness}\footnote{Example: ``Not
sure how this happened but in two days I've somehow gained 5
lbs...so {not happy} about this. \#ugly \#fatty \#depressed \#sad''}
(1.2) and \emotionword{disgust} (0.7), and \emotionword{sadness} to \emotionword{joy}\footnote{Example: ``Yes!
I'm about to eat this piece of cheesecake and I {don't feel guilty}
about it. \#indulgingalittle \#cheesecake \#happy''}
(1.0). However, some emotions do not show this flip in polarity, for
instance in \emotionword{fear} and \emotionword{surprise}. For the class \emotionword{surprise}, a
reason is that tweets often use comparisons like the phrase ``\ldots
no party like\ldots'', with ``no party'' indicating negated
\emotionword{surprise}\footnote{Example: ``{Ain't no party like} a birthday party
when @LJ\_Rader shows up \#surprise''}. Examples for \emotionword{fear} appearing
in negated context include those whose authors encourage people not to
have fear but still use \emotionword{fear}-related hashtags\footnote{Examples:
``Don't worry, let God take control. \#worry'', ``"No fear is
stronger than you are." - Mark David Gerson \#fear \#quote
\#spirituality''}.
Altogether, the weights are lower than in the matrix for emotion words
outside of a modifier scope, backing our hypothesis that partially no
emotion is expressed with a negated emotion word.
\paragraph{Emotion words in the scope of amplifiers}
Most diagonal weights of the amplifier matrix show an increased value
in comparison to the matrix for emotion words outside of scope, as
hypothesized (for \emotionword{joy} by factor 2\footnote{Example: ``Wishing you a
very happy day! \#happiness \#positivity''}, \emotionword{fear} by factor 1.9,
\emotionword{sadness} by 1.6). For some emotions, in addition to the hypothesis, the
amplifier clearly strengthens the non-occurrence of another emotion:
an amplified \emotionword{joy} word is a clear signal for non-occurrence of \emotionword{anger}
($-3.8$), while it has nearly no contribution without modification
($-0.1$). This pattern can also be observed for \emotionword{joy} and \emotionword{fear} ($-1.6$
instead of $0.2$ without modification) but only to a lesser degree for
other emotions. For \emotionword{anger} words, the contribution to \emotionword{sadness} flips
from a negative to a positive contribution. Interestingly, amplified
words of \emotionword{fear} contribute positively to all emotions.
\paragraph{Emotion words in the scope of downtoners}
The weights on the diagonal for emotion words in the scope of
downtoners is lower than for words out of scope of a modifier,
however, higher than for negations. Therefore, downtoners can
partially be interpreted as ``light version'' of
negations.\footnote{Example for downtoned sadness with impact on joy:
``pray more and {worry less} \#pray \#faith \#love \#peace
\#happiness...'', and vice versa: ``Just a bit happy to be back in
Ibiza...''} However, as expected, they do not flip the
polarity. Counter examples are downtoned words associated with \emotionword{fear}
and their impact on \emotionword{surprise}. Most of such tweets contain a phrase
similar to ``little surprise'', which has a meaning similar to
negation. While on average the weights are lower than for other
modifications and no modifications, striking is the highest weight in
all matrices for \emotionword{sadness} contributing negatively to \emotionword{anger} ($-4.3$). A
reason could be that practically no tweet occurs in the corpus that
contains a downtoned word for \emotionword{sadness} and is labeled as \emotionword{anger}.
\section{Summary, Conclusion and Future Work}
In this paper, we showed that modifier detection and handling has an
impact on the prediction of emotions. This impact differs by emotion
and by modifier: the prediction of \emotionword{disgust} and \emotionword{anger} are most
affected by downtoners, while \emotionword{joy} and \emotionword{anger} are most affected by
negations. Amplifications are most relevant to \emotionword{fear}. Across all
emotions the prediction of \emotionword{surprise} and \emotionword{sadness} are not that
strongly affected.
A deeper look on the impact of negations, amplifiers, and downtoners
on separate emotions discloses results which are mostly in line with
the models by Plutchik and Russell \cite{Posner2005}. Interesting
results include that modifiers influence different pairs of emotions
to different degrees: highest weights ($-3.7$) can be observed for
disgust--surprise (\emph{observation}--\emph{prediction}) without
modifiers. Amplifying words denoting \emotionword{surprise}, however, does not
increase such weights but decreases them -- amplifiers separate (some)
emotions stronger from all than their prior emotions. This is
particularly the case for \emotionword{fear}, where the weight increases from 1.9
to 3.6 (without modifier to amplifier). For negations, which are
probably the most challenging modifiers to understand emotions, we see
the highest (negative) weights for \emotionword{disgust} and \emotionword{fear}, \emotionword{surprise} and
\emotionword{anger} -- ``not surprised'' definitely does not mean \emotionword{anger}, and ``not
disgusted'' definitely does not mean \emotionword{fear}. More intuitively are
positive weights which are, again, in line with psychological models.
Future work includes more detailed parameter tuning in our models. We
made the assumption that a maximal $\text{F}_1$\xspace of scope detection is optimal
for classification and therefore set $n=2$. However, a different ratio
of precision and recall might be beneficial. Therefore, jointly
optimizing parameters of emotion scope detection in the downstream
task might uncover a different parameter setting.
One source of error in the scope detection are mistakes in the parse
tree generation. An evaluation of different parsers and optimizing
them for the task at hand might lead to improved performance.
The weight matrices in our lexical model were optimized separately for
each modifier. However, we represent them as a 3D tensor
already. Therefore, a next step will be a joint optimization of all
parameters. We assume that interactions between them might lead to
improved results.
Our study is built on top of document-level classification. We propose
follow-up studies to investigate the word level and subword level with
the use of distributional semantics. In addition, we did not take into
account implicit modifications and modifying inflections and
derivations. This strain of work will connect our results in this
paper to the initiatives of predicting the intensities of whole
tweets, as shown by Mohammad \textit{et al.}\ in previous work
\cite{Mohammad2017}. In addition, the analysis and comparison with
sequence-based classifiers including attention mechanisms will allow
for a deeper analysis of end-to-end systems. We assume that it is more
challenging to obtain knowledge regarding modifiers from these
methods, however, given the work in this paper, we will analyze if our
hypotheses also manifest in these approaches.
\section*{Acknowledgements}
This research has been partially funded by the German Research Council
(DFG), project SEAT (Structured Multi-Domain Emotion Analysis from
Text, KL 2869/1-1). We thank Evgeny Kim and Laura Bostan for
proof-reading and fruitful discussions.
|
\section*{Introduction}
A \emph{real structure} on a complex algebraic variety $X$ is an antiregular involution $\mu$ on $X$, where \emph{antiregular} means that the following diagram commutes:
\[\xymatrix@R=4mm@C=2cm{
X \ar[rr]^{\mu} \ar[d] && X \ar[d] \\
\mathrm{Spec}(\mathbb{C}) \ar[rr]^{\mathrm{Spec}(z \mapsto \overline{z})} && \mathrm{Spec}(\mathbb{C})
}\]
Two real structures $\mu$ and $\mu'$ are called \emph{equivalent} if there exists $\varphi \in \Aut(X)$ such that $\mu'=\varphi \circ \mu \circ \varphi^{-1}$. To any real structure $\mu$ on $X$ one can associate the quotient $\mathcal{X}=X/\mu$, which is a real algebraic space satisfying $\mathcal{X} \times_{\mathrm{Spec}(\mathbb{R})} \mathrm{Spec}(\mathbb{C}) \simeq X$ as complex varieties. Moreover, if $X$ is quasi-projective, then $\mathcal{X}$ is actually a real variety. The quotient $\mathcal{X}$ is called a \emph{real form} of $X$. Two real forms $\mathcal{X}$ and $\mathcal{X}'$ are $\mathbb{R}$-isomorphic if and only if the corresponding real structures are equivalent.
Describing all the equivalence classes of real structures on a given complex variety is a classical problem in algebraic geometry.
We refer to \cite[Chp. 2]{Man17} and \cite[Chp. 3]{Ben16} for an exposition (in French) of the foundations of this theory.
When $X$ carries some extra structure, it is natural to ask for $\mathcal{X}$ to also carry this extra structure. For instance, if $X=G$ is a complex algebraic group, then it is particularly interesting to describe the real structures $\sigma$ on $G$ which are group involutions, so that the real form $\mathcal{G}=G/\sigma$ is a real algebraic group (and not just a real variety). Such real structures are called \emph{real group structures}; see \S\S~\ref{sec:generalities}-\ref{sec:quasi split and inner twists} for a recap of their classification.
Another class of complex varieties carrying extra structure are the complex varieties endowed with an algebraic group action, which yields the key notion of equivariant real structure. Let $G$ be a complex reductive algebraic group, let $\sigma$ be a real group structure on $G$, and let $X$ be a complex $G$-variety. A real structure $\mu$ on $X$ is called a $(G,\sigma)$-\emph{equivariant real structure} if $\mu(g \cdot x)=\sigma(g) \cdot \mu(x)$ for all $g \in G$ and all $x \in X$.
Two $(G,\sigma)$-equivariant real structures $\mu$ and $\mu'$ are called \emph{equivalent} if there exists $\varphi \in \Aut^G(X)$ such that $\mu'=\varphi \circ \mu \circ \varphi^{-1}$. Then the real form $\mathcal{X}=X/\mu$ is a real $\mathcal{G}$-variety, and two real forms $\mathcal{X}$ and $\mathcal{X}'$ are isomorphic as $\mathcal{G}$-varieties if and only if the corresponding equivariant real structures are equivalent.
A \emph{horospherical subgroup} $ H \subseteq G$ is a subgroup containing a maximal unipotent subgroup of $G$, and a \emph{horospherical $G$-variety} is a normal $G$-variety with an open orbit $G$-isomorphic to $G/H$ with $H$ a horospherical subgroup of $G$. Classical examples of horospherical varieties are given by flag varieties and toric varieties. Horospherical varieties form a subclass of spherical varieties (see \cite{Pau81, Kno91}) whose combinatorial description is much more accessible.
A presentation of the theory of horospherical varieties can be found in \cite{Pas08}. Also, the combinatorial description of horospherical subgroups of $G$ from Pasquier in terms of pairs ($I,M$), where $I$ is a subset of the set of simple roots of $G$ and $M$ is a certain lattice depending on $I$, is recalled in \S~\ref{sec:setting}.
The present article aims at studying equivariant real structures on horospherical varieties in a systematic way. Our main result is the following:
\begin{theorem} \label{th:1}
\emph{(Theorem~\ref{th:main results} and Proposition~\ref{prop:number of structures})}\
Let $G$ be a complex reductive algebraic group with a real group structure $\sigma$.
Let $H$ be a horospherical subgroup of $G$ with datum $(I,M)$.
Then a $(G,\sigma)$-equivariant real structure exists on $G/H$ if and only if $(I,M)$ is stable for the action of the Galois group $\Gal(\mathbb{C}/\mathbb{R})$ defined by $\sigma$ and $\Delta_H(\sigma)$ is trivial, where $\Delta_H$ is the map defined by \eqref{map delta} in \S~\ref{sec: useful map}.
Moreover, if such a structure exists, then there are exactly $2^n$ equivalence classes of $(G,\sigma)$-equivariant real structures on $G/H$, where $n$ is a non-negative integer that can be calculated explicitly (see \S~\ref{sec:number of equi real structures} for details on how to compute $n$).
\end{theorem}
\begin{remark}
A couple of months after the release of the present article, Borovoi and Gagliardi obtained in \cite{BGb} a criterion for the existence of equivariant real structures on general spherical homogeneous spaces generalizing our Theorem~\ref{th:main results}.
\end{remark}
For the sake of brevity, we do not recall the \emph{Luna-Vust theory of spherical embeddings} (see \cite{Kno91,Tim11} for a presentation), and how to describe such embeddings in terms of the combinatorial data called \emph{colored fans}: these are fans such as those for toric varieties but with additional information called \emph{colors}. The reader is not required to be familiar with this theory as it is used only in \S~\ref{sec:extension of real structures}, and in that section we describe explicitly the properties of the colored fans used.
As explained in \cite{Hur11}, if $\sigma$ is a real group structure on $G$, and if $G/H$ is a spherical homogeneous space endowed with a $(G,\sigma)$-equivariant real structure, then $\sigma$ defines an action of the Galois group $\Gal(\mathbb{C}/\mathbb{R})$ on the set of colored fans defining a $G$-equivariant embedding of $G/H$.
The next result is an immediate consequence of \cite[Theorem 2.23]{Hur11} and \cite[Theorem 9.1]{Wed} together with a quasi-projectivity criterion for spherical varieties due to Brion (see \S~\ref{sec:extension of real structures} for details).
\begin{theorem} \label{th:2} \emph{(Corollary~\ref{cor:extension})} \
Let $\mu$ be a $(G,\sigma)$-equivariant real structure on a horospherical homogeneous space $G/H$, and let $X$ be a horospherical $G$-variety with open orbit $G/H$. Then the real structure $\mu$ extends on $X$ if and only if the colored fan of the embedding $G/H \hookrightarrow X$ is invariant for the action of the Galois group $\Gal(\mathbb{C}/\mathbb{R})$ defined by $\sigma$, in which case the corresponding real form $X/\mu$ is a real variety.
\end{theorem}
To illustrate the effectiveness of our results, we then consider the equivariant real structures on smooth projective horospherical $G$-varieties of Picard rank $1$ (the odd symplectic grassmannians are examples of such varieties).
\begin{theorem} \label{th:3}
Let $G$ be a complex simply-connected semisimple algebraic group with a real group structure $\sigma$.
Let $X$ be a smooth projective horospherical $G$-variety of Picard rank $1$.
If a $(G,\sigma)$-equivariant real structure exists on $X$, then it is unique up to equivalence.
The cases where such a real structure exists are classified in Example \ref{ex:G/P 2} (when $X=G/P$ with $P$ a maximal parabolic subgroup of $G$) and in Theorem \ref{th:real forms of horo Picard 1} (when $X$ is non-homogeneous).
\end{theorem}
As mentioned before, horospherical varieties are a subclass of spherical varieties. Equivariant real structures on spherical varieties already appeared in the literature, but the scope was not the same as in this article. More precisely:
\begin{itemize}[leftmargin=*]
\item In \cite{Hur11,Wed} the authors consider the situation where a real structure on the open orbit is given and they determine in which cases this real structure extends to the whole spherical variety. They do not treat the case of equivariant real structures on homogeneous spaces.
(Note also that they work over an arbitrary field and not just over $\mathbb{R}$.)
\item In \cite{ACF14,Akh15,CF15} the authors study equivariant real structures on spherical homogeneous spaces $G/H$ and their equivariant embeddings when $N_G(H)/H$ is finite. Such varieties are never horospherical, except the flag varieties.
\item In \cite{BGa} the authors extend part of the results in \cite{ACF14,Akh15,CF15} and work over an arbitrary base field of characteristic zero.
\item In \cite{MJT} the authors obtain a criterion for the existence of equivariant real structures on symmetric spaces using the involution associated with the symmetric space instead of the homogeneous spherical data.
\end{itemize}
Besides their easy combinatorial description and their ubiquity in the world of algebraic group actions, horospherical varieties lend themselves very nicely to the study of equivariant real structures for several reasons: First of all, the group of $G$-equivariant automorphisms of $G/H$ is a torus (see \S~\ref{sec:setting}), which reduces the computation of the number of equivalence classes of equivariant real structures on $G/H$ to the case of tori (see \S~\ref{sec:number of equi real structures}). Secondly, if $\sigma$ is a quasi-split real group structure on $G$ such that $\sigma(H)$ is conjugate to $H$, then, by the classification of horospherical homogeneous spaces, it is easy to show that there exists a conjugate $H'$ of $H$ such that $\sigma(H')=H'$ (Proposition \ref{prop: datum of the conjugate}); this fact, which also holds for the spherical varieties considered in \cite{ACF14} as proved there, is essential in the proof of Theorem~\ref{th:main results}. Thirdly, the Galois descent is effective for horospherical varieties, i.e. the real form $X/\mu$ is always a real variety and not only a real algebraic space as can happen for spherical varieties (see Remark~\ref{rk:alg space}).
\smallskip
In \S\S~\ref{sec:generalities}-\ref{sec:quasi split and inner twists} we recall definitions and well-known facts about real group structures on complex algebraic groups. In \S~\ref{sec: useful map} we define the cohomologial invariant $\Delta_H(\sigma)$ that appears in the existence criterion of Theorem \ref{th:1}.
Then in \S~\ref{sec:equiv real structures} we recall the notion of equivariant real structures. In particular, we show how to determine if such a structure exists on a given homogeneous space, and if so, then how to use Galois cohomology to determine the set of equivalence classes of these equivariant real structures.
The main part of this article is \S~\ref{sec:equi real groups strcture for horo} in which we prove the results above. In \S~\ref{sec:setting} we recall the basic notions regarding the horospherical homogeneous spaces and their combinatorial classification. In \S\S~\ref{sec:quasi-split horo}-\ref{sec: non-quasi split case} we prove the existence criterion in Theorem \ref{th:1} (this is Theorem \ref{th:main results}), and in \S~\ref{sec:number of equi real structures} we prove the quantitative part in Theorem \ref{th:1} (this is Proposition \ref{prop:number of structures}). In \S~\ref{sec:extension of real structures} we recall the main result of \cite{Hur11,Wed} regarding the extension of equivariant real structures from a spherical homogeneous space to the whole spherical variety and we apply it to prove Theorem \ref{th:2} (which is Corollary \ref{cor:extension}). Then, we apply our results to classify the equivariant real structures on smooth projective horospherical varieties of Picard rank $1$ and prove Theorem \ref{th:3} (see \S~\ref{sec: smooth proj Picard rank one}).
Finally the list of real group structures on complex simply-connected simple algebraic groups together with the list of the corresponding Tits classes is recalled in Appendix \ref{sec: tables}. These cohomology classes are useful to compute the cohomological invariant $\Delta_H(\sigma)$ in examples.
\smallskip
\noindent \textbf{Notation.}
In this article we work over the field of real numbers $\mathbb{R}$ and the field of complex numbers $\mathbb{C}$.
We denote by $\mu_n$ the group of $n$-th roots of unity and by $\Gamma$ the Galois group $\Gal(\mathbb{C}/\mathbb{R})=\{1,\gamma\} \simeq \mu_2$.
A \emph{variety} over a field $\k$ is a geometrically reduced separated scheme of finite type over $\k$; in particular, varieties can be reducible.
An \emph{algebraic group} $G$ over $\k$ is a group scheme over $\k$. By an \emph{algebraic subgroup} of $G$ we mean a closed subgroup scheme of $G$. Reductive algebraic groups are always assumed to be connected for the Zariski topology. When we talk about a \emph{group involution} $\sigma$ we mean that $\sigma$ is an automorphism of algebraic groups (possibly over a subfield of $\k)$ such that $\sigma \circ \sigma=Id$. We refer the reader to \cite{Hum75} for the standard background on algebraic groups.
We always denote by $G$ a complex algebraic group, by $Z(G)$ its center, by $B$ a Borel subgroup of $G$, by $T$ a maximal torus of $B$, and by $U$ the unipotent radical of $B$ (which is also a maximal unipotent subgroup of $G$). If $H$ is a subgroup of $G$, then $N_G(H)$ denote the normalizer of $H$ in $G$. We write $\mathbb{X}=\mathbb{X}(T)=\mathrm{Hom}_{gr}(T,\mathbb{G}_m)$ for the character group of $T$ and $\mathbb{X}^\vee=\mathbb{X}^\vee(T)=\mathrm{Hom}_{gr}(\mathbb{G}_m,T)$ for the cocharacter group of $T$. When $G$ is semisimple we denote by $\mathrm{Dyn}(G)$ its Dynkin diagram.
\section{Real group structures}
In this section, we start by recalling definitions and well-known facts about real group structures on complex algebraic groups. We are mostly interested in the case of reductive groups. In particular, we show how to obtain all real group structures on complex reductive algebraic groups, by piecing together the structures on complex tori and on complex simply-connected simple algebraic groups.
The main references we use are \cite{Con14} for results concerning the structure of reductive algebraic groups, \cite{Ser02} for general results concerning Galois cohomology and \cite{Man17,Ben16} for generalities on real structures. In \textit{loc.cit.} the author treats the case of real structures on complex quasi-projective varieties; the corresponding results referred to here concern complex algebraic groups and can be treated in exactly the same way.
\subsection{Generalities and first classification results} \label{sec:generalities}
\begin{definition} (Real group structures on complex algebraic groups.)
\begin{enumerate}[(i),leftmargin=*]
\item Let $G$ be a complex algebraic group. A \emph{real group structure on $G$} is an antiregular group involution $\sigma: G \to G$, i.e., a group involution over $\mathrm{Spec}(\mathbb{R})$ which makes the following diagram commute:
\[\xymatrix@R=4mm@C=2cm{
G \ar[rr]^{\sigma} \ar[d] && G \ar[d] \\
\mathrm{Spec}(\mathbb{C}) \ar[rd] \ar[rr]^{\mathrm{Spec}(z \mapsto \overline{z})} && \mathrm{Spec}(\mathbb{C}) \ar[ld]\\
& \mathrm{Spec}(\mathbb{R})
}\]
\item Two real group structures $\sigma$ and $\sigma'$ on $G$ are \emph{equivalent} if there exists a (regular) group automorphism $\varphi \in \Aut_{gr}(G)$ such that $\sigma'=\varphi \circ \sigma \circ \varphi^{-1}$.
\end{enumerate}
\end{definition}
\begin{remark} \label{rk:any antiregular group automorphism is of the form}
If $\sigma$ is an antiregular group involution on $G$, then any antiregular group automorphism is of the form $\varphi\circ\sigma$ for some group automorphism $\varphi$. If $\varphi$ and $\sigma$ commute, then $\varphi\circ\sigma$ is an involution if and only if $\varphi$ is an involution.
\end{remark}
If $(G,\sigma)$ is a complex algebraic group with a real group structure, then the quotient scheme $\mathcal{G}=G/\sigma$ is a real algebraic group which satisfies $\mathcal{G} \times_{\mathrm{Spec}(\mathbb{R})} \mathrm{Spec}(\mathbb{C}) \simeq G$ as complex algebraic groups. The real group $\mathcal{G}$ is called a \emph{real form of $G$}. Two real forms are $\mathbb{R}$-isomorphic if and only if the corresponding real group structures are equivalent (see \cite[Corollary 3.13]{Ben16}).
\begin{definition} \label{def:real part}
If $(G,\sigma)$ is a complex algebraic group with a real group structure, then $G_0=G(\mathbb{C})^\sigma$ is called the \emph{real part} (or \emph{real locus}) of $(G,\sigma)$; it coincides with the set of $\mathbb{R}$-points of the real algebraic group $G/\sigma$ (see \cite[Proposition 3.14]{Ben16} for details).
\end{definition}
With the notation of Definition \ref{def:real part} the group of $\mathbb{C}$-points $G(\mathbb{C})$ is a complex Lie group and $G_0=G(\mathbb{C})^\sigma$ is a real Lie group.
In fact, the real part $G_0$ determines the real group structure for a connected complex algebraic group:
\begin{theorem} \label{th: real part determines real structure} \emph{(\cite[Theorem 3.41]{Ben16})}
Let $G$ be a connected complex algebraic group with two real group structures $\sigma$ and $\sigma'$.
Then $\sigma$ and $\sigma'$ are equivalent if and only if there is a (scheme) automorphism $\varphi \in \Aut(G)$ such that $\varphi(G(\mathbb{C})^{\sigma})=G(\mathbb{C})^{\sigma'}$.
\end{theorem}
Because of Theorem \ref{th: real part determines real structure} we will sometimes consider the real part $G_0=G(\mathbb{C})^\sigma$ instead of the real group structure $\sigma$ when describing all the possible equivalence classes of real group structures on connected complex algebraic groups.
\bigskip
Let $G$ be a complex reductive algebraic group, let $T=Z(G)^0$ be the neutral component of the center of $G$, and let $G'$ be the derived subgroup of $G$. Then the homomorphism $T \times G' \to G$, $(t,g') \mapsto t^{-1}g'$ is a central isogeny with kernel $T \cap G'$. Also, there is a 1-to-1 correspondence
\[ \left \{ \text{real group structures $\sigma$ on $G$} \right \} \leftrightarrow \left \{ \begin{tabular}{l}
\text{real group structures $(\sigma_1,\sigma_2)$ on $T \times G'$} \\
\text{ \ \ such that $\sigma_{1| T \cap G'}= \sigma_{2| T \cap G'}$}
\end{tabular} \right \} \]
given by $\sigma \mapsto (\sigma_{|T},\sigma_{|G'})$. Therefore, to determine real group structures on complex reductive algebraic groups, it suffices to determine real group structures on complex tori and on complex semisimple algebraic groups.
\begin{lemma} \label{lem: real form on tori} \emph{(Classification of real group structures on complex tori.)}\\
Let $T \simeq \mathbb{G}_m^n$ be an $n$-dimensional complex torus.
\begin{enumerate}[(i),leftmargin=*]
\item If $n=1$, then $T$ has exactly two inequivalent real group structures, defined by $\sigma_0: t \mapsto \overline{t}$ and $\sigma_1: t \mapsto \overline{t}^{-1}$.
\item If $n=2$, then $\sigma_2: (t_1,t_2) \mapsto (\overline{t_2},\overline{t_1})$ defines a real group structure on $T$.
\item \label{item: n at least 2} If $n\ge 2$, then every real group structure on $T$ is equivalent to exactly one real group structure of the form $\sigma_0^{\times n_0}\times\sigma_1^{\times n_1}\times\sigma_2^{\times n_2}$, where $n=n_0+n_2+2n_2$.
\end{enumerate}
\end{lemma}
\begin{proof}
This result is well-known to specialists (see for instance \cite[Chp.4, \S~10.1]{Vos98}) but we give a sketch of the proof for the sake of completeness.
Clearly, each $\sigma_i$ for $i=1,2$ or $3$ defines a real group structure on $T$. For $n=1$, the structures $\sigma_0$ and $\sigma_1$ are inequivalent, since the real parts are not diffeomorphic. Also since $\Aut_{gr}(\mathbb{G}_m) \simeq \mu_2$, these are the only two real group structures on a one-dimensional torus. For dimension two, $\sigma_2$ defines a new real group structure since it is an antiregular group involution.
Also, for $n\ge2$ all the real group structures on $T$ given in \ref{item: n at least 2} exist and are inequivalent since the corresponding real parts are $(\mathbb{R}^*)^{n_0} \times (S^1)^{n_1} \times (\mathbb{C}^*)^{n_2}$ which are pairwise non-diffeomorphic.
It remains to show that all real group structures are equivalent to one of the structures given in (iii). This is done by determining all the conjugacy classes of $\Aut_{gr}(\mathbb{G}_m^n)\simeq \GL_n(\mathbb{Z})$.
More precisely, any real group structure in dimension $n$ is equivalent to $\sigma=\varphi\circ (\sigma_0^{\times n})$ for some $\varphi\in \Aut_{gr}(\mathbb{G}_m^n)$ (see Remark \ref{rk:any antiregular group automorphism is of the form}). Since all group
automorphisms commute with $\sigma_0^{\times n}$, we see that $\sigma$ is an involution if and only if $\varphi$ is an involution. Also, two involutions define equivalent real group structures if and only if they are in the same conjugacy class.
Finally, note that the conjugacy classes of elements of order $2$ in $\GL_n(\mathbb{Z})$ are represented exactly by block diagonal matrices of the form
\[ \mathrm{diag} \left( 1,\ldots,1,-1,\ldots,-1, \begin{bmatrix}
0 & 1 \\1 & 0
\end{bmatrix},\ldots, \begin{bmatrix}
0 & 1 \\1 & 0
\end{bmatrix} \right).\]
Each block corresponds to $\sigma_0$, $\sigma_1$ and $\sigma_2$ respectively, which proves the result.
\end{proof}
\begin{remark}\label{rem:torus-variety}
If one forgets the group structure and considers $T \simeq \mathbb{G}_m^n$ as a complex variety, then there are other real structures on $T$. For instance $\tau_1: t \mapsto -\overline{t}^{-1}$ is a real structure on $\mathbb{G}_m$ but not a real group structure. In fact, one can show that any real structure on $T$ is equivalent to a product $\sigma_0^{\times n_0} \times \sigma_1^{\times n_1} \times \sigma_2^{\times n_2} \times \tau_1^{\times m}$ for some $n_0, n_1,n_2, m \in \mathbb{N}$.
\end{remark}
It remains to determine the real group structures on complex semisimple algebraic groups. For any complex semisimple algebraic group $G'$, there exists a central isogeny $\varphi: \widetilde{G'} \to G'$, where $\widetilde{G'}$ is a simply-connected semisimple algebraic group (uniquely defined by $G'$ up to isomorphism); see \cite[Exercise 1.6.13]{Con14}.
Then $\widetilde{G'}$ is isomorphic to a product of simply-connected simple algebraic groups \cite[\S~6.4]{Con14}; the later is uniquely defined up to permutation of the factors.
The next lemma is also well-known but for sake of completeness we recall the proof. It reduces the classification of real group structures on simply-connected semisimple groups to the classification of real group structures on simply-connected simple groups.
\begin{lemma} \label{lem:easy_lemma_reduction}
Let $\sigma$ be a real group structure on a complex simply-connected semisimple algebraic group $G' \simeq \prod_{i \in I} G_i$, where the $G_i$ are the simple factors of $G'$. Then, for a given $i \in I$, we have the following possibilities:
\begin{enumerate}[(i),leftmargin=*]
\item \label{item: Gi stable}$\sigma(G_i)=G_i$ and $\sigma_{|G_i}$ is a real group structure on $G_i$; or
\item \label{item:Gi and Gj swap} there exists $j \neq i$ such that $\sigma(G_i)=G_j$, then $G_i \simeq G_j$ and $\sigma_{| G_i \times G_j}$ is equivalent to $(g_1,g_2) \mapsto (\sigma_0(g_2),\sigma_0(g_1))$, where $\sigma_0$ is an arbitrary real group structure on $G_i \simeq G_j$.
\end{enumerate}
\end{lemma}
\begin{proof}
We use the fact that the factors $G_i$ are the unique simple normal subgroups of $G$ (see \cite[Theorem 5.1.19] {Con14}). In particular any group automorphism of $G$ permutes the factors. Since $\sigma$ is a group involution,
either $\sigma(G_i)=G_i$ and we get \ref{item: Gi stable}, or $\sigma(G_i)=G_j$ for some $j \neq i$. In the second case, $G_i$
and $G_j$ are $\mathbb{R}$-isomorphic, and since they are both simply-connected simple groups they must be $\mathbb{C}$-isomorphic (this
follows for instance from the classification of simply-connected simple algebraic groups in terms of Dynkin diagrams). Therefore
$G_i \times G_j \simeq H \times H$, for some simply-connected simple group $H$, and $\sigma_{|G_i \times G_j}$ identifies with $
\sigma_{H \times H}: (h_1,h_2) \mapsto (\sigma_1(h_2),\sigma_1^{-1}(h_1))$ for some antiregular automorphism $\sigma_1$ on $H$. But
then it suffices to conjugate $\sigma_{H \times H}$ with the group automorphism $\varphi$ defined by $(h_1,h_2) \mapsto (\sigma_1 \circ
\sigma_0(h_2),h_1)$ to get the real group structure $(g_1,g_2) \mapsto (\sigma_0(g_2),\sigma_0(g_1))$, where $\sigma_0$ is an arbitrary real group structure on $H$.
\end{proof}
The real group structures on complex simply-connected simple algebraic groups are well-known (a recap can be found in Appendix \ref{sec: tables}); they correspond to real Lie algebra structures on complex simple Lie algebras (see \cite[\S~V\!I.10]{Kna02} for the classification of those).
Therefore we can determine all the real group structures on complex simply-connected semisimple algebraic groups from Lemma \ref{lem:easy_lemma_reduction}. In the next subsection, we will give a brief outline of a way to classify them, using quasi-split structures and inner twists.
\begin{example}
Up to equivalence, there are two real group structures on $\SL_2$ given by $\sigma_0(g)=\overline{g}$ and $\sigma_1(g)=\leftexp{t}{\overline{g}}^{-1}$.
Up to equivalence, there are four real group structures on $\SL_2 \times \SL_2$ given by $\sigma_i \times \sigma_j$ with $(i,j) \in \{(0,0),(0,1),(1,1)\}$ and $\sigma_2: (g_1,g_2) \mapsto (\sigma_0(g_2),\sigma_0(g_1))$.
Similarly, we let the reader check that, up to equivalence, there are six real group structures on $\SL_2 \times \SL_2 \times \SL_2$ and nine real group structures on $\SL_2 \times \SL_2 \times \SL_2 \times \SL_2$.
\end{example}
\subsection{Quasi-split real group structures and inner twists} \label{sec:quasi split and inner twists}
Let $(G,\sigma)$ be a complex reductive algebraic group with a real group structure. Note that the set of Borel subgroups of $G$ and, for any $n \in \mathbb{N}$, the set of
$n$-dimensional
tori of $G$ are each preserved by $\sigma$ (this is a direct consequence of the definition of Borel subgroups and tori).
\begin{definition}\textcolor{white}{--}
\begin{enumerate}[(i),leftmargin=*]
\item If there exists a Borel subgroup $B \subseteq G$ such that $\sigma(B)=B$, then $\sigma$ is called \emph{quasi-split}.
Let $T \subseteq B$ be a maximal torus such that $\sigma(T)=T$ (such a torus always exists by Theorem \ref{th:ABC Galois theory} \ref{item: one qs in each fiber}). With the notation of Lemma \ref{lem: real form on tori}, if the restriction $\sigma_{|T}$ is equivalent to a product $\sigma_0^{\times \dim(T)}$, then $\sigma$ is called \emph{split}.
\item For $c \in G$ we denote by $\mathrm{inn}_c$ the inner automorphism of $G$ defined by
\[ \mathrm{inn}_c: G \to G, g \mapsto cgc^{-1}.\]
If $\sigma_1$ and $\sigma_2$ are two real group structures on $G$ such that $\sigma_2 = \mathrm{inn}_c \circ \sigma_1$, for some $c \in G$, then $\sigma_2$ is called an \emph{inner twist} of $\sigma_1$.
\end{enumerate}
\end{definition}
\begin{remark} \label{rk:trivial split action}
If $\sigma$ is a quasi-split real group structure on $G$, then for any pair $(B',T')$ with $B' \subseteq G$ a Borel subgroup and $T' \subseteq B'$ a maximal torus, there exists an equivalent real group structure $\sigma'=\mathrm{inn}_c \circ \sigma \circ \mathrm{inn}_c^{-1}$, for some $c \in G$, stabilizing both $B'$ and $T'$.
\end{remark}
There exists a unique split real group structure on $G$ up to equivalence (see \cite[Theorem 6.1.17]{Con14} or \cite[Chp.5, \S~4.4]{OV90} when $G$ is semisimple) that we will always denote by $\sigma_0$. There is also a unique compact real group structure $\sigma_c$ on $G$ up to equivalence (see \cite[ Chp. 5, \S\S~1.3-1.4]{OV90} when $G$ is semisimple and Lemma \ref{lem: real form on tori} when $G$ is a torus, the general case follows from these two cases).
In general, $G$ may have several inequivalent quasi-split real group structures. If $G$ is a simple algebraic group, then $G$ has at most two inequivalent quasi-split real group structures (this will follow from Theorem \ref{th:ABC Galois theory} \ref{item: dynkin conjugacy classes}). On the other hand, for tori all real group structures are quasi-split.
The next classical lemma yields a description of the set of real group structures obtained as inner twists of a given real group structure.
\begin{lemma} \label{lem: parametrization of the inner twists}
For a given $c \in G$, the antiregular group automorphism $\mathrm{inn}_c \circ \sigma$ is a real group structure on $G$ if and only if $c \sigma(c) \in Z(G)$ (and then $c \sigma(c)= \sigma(c)c)$. Also, $\mathrm{inn}_c \circ \sigma=\mathrm{inn}_{c'} \circ \sigma$ if and only if $c^{-1}c' \in Z(G)$.
\end{lemma}
\begin{proof}
Since $\mathrm{inn}_c \circ \sigma$ is an antiregular group automorphism, it is a real group structure if and only if it is an involution, i.e.
\begin{align*}
(\mathrm{inn}_c \circ \sigma) \circ (\mathrm{inn}_c \circ \sigma)=Id & \Leftrightarrow \forall g \in G, c \sigma(c \sigma(g) c^{-1}) c ^{-1}=g \\
& \Leftrightarrow \forall g \in G, c \sigma(c) g =g c \sigma(c) \\
& \Leftrightarrow c \sigma(c) \in Z(G).
\end{align*}
A similar computation yields the second equivalence stated in the lemma.
\end{proof}
Note that in the previous lemma, we do not give a general condition describing when $\mathrm{inn}_c\circ\sigma$ is equivalent to $\mathrm{inn}_{c'}\circ\sigma$.
\begin{example} \label{ex: forms of SL3}
Let $c=c^{-1}=\begin{bmatrix}
0 & 0 & -i\\ 0 & -1 & 0 \\ i & 0 & 0
\end{bmatrix}\in \SL_3$.
Up to equivalence, the group $\SL_3$ has three real group structures given by $\sigma_0(g)=\overline{g}$, which is split and whose real part is $\SL_3(\mathbb{R})$,
$\sigma_1(g)=c(\leftexp{t}{\overline{g}}^{-1})c^{-1}$, which is quasi-split and whose real part is $\SU(1,2)$, and $\sigma_2(g)=\leftexp{t}{\overline{g}}^{-1}=\mathrm{inn}_c \circ \sigma_1$, which is an inner twist of $\sigma_1$ and whose real part is $\SU(3)$.
\end{example}
We now recall a $\Gamma$-action that will play an important role in \S~\ref{sec:equi real groups strcture for horo} when studying the equivariant real structures on horospherical varieties.
\begin{definition} \label{def: Gamma-action on X(T)}
Let $(G,\sigma_{qs})$ be a complex reductive algebraic group with a quasi-split real group structure that preserves a Borel subgroup $B \subseteq G$ and a maximal torus $T \subseteq B$. It induces a $\Gamma$-action on $\mathbb{X}$ and on $\mathbb{X}^\vee$ as follows:
\[\forall \chi \in \mathbb{X},\ \,^\gamma\hskip-1pt \chi=\tau \circ \chi \circ \sigma_{qs} \ \ \ \ \text{ and } \ \ \ \ \forall \lambda \in \mathbb{X}^\vee,\ \,^\gamma\hskip-1pt \lambda=\sigma_{qs} \circ \lambda \circ \tau\, ,\]
where $\tau(t)=\overline{t}$ is complex conjugation.
\end{definition}
Let us note that the sets of roots, coroots, simple roots, and simple coroots associated with the triple $(G,B,T)$ are preserved by this $\Gamma$-action \cite[Remark 7.1.2]{Con14}, and so $\Gamma$ acts on the based root datum of $G$. Moreover, if $\sigma_{qs}=\sigma_0$ is a split real group structure, then the $\Gamma$-action on $\mathbb{X}$ and $\mathbb{X}^\vee$ is trivial.
\smallskip
We now recall the definition of the first Galois cohomology pointed set as it will appear several times in the rest of this article. Since we are concerned with real structures, we will restrict the presentation to Galois cohomology for $\Gamma$-groups. More details on Galois cohomology in a more general setting can be found in \cite{Ser02}.
\begin{definition} \label{def:Galois H1}
If $A$ is a $\Gamma$-group, then the first Galois cohomology pointed set is $H^1(\Gamma,A)=Z^1(\Gamma,A)/\sim$, where $Z^1(\Gamma,A)=\{ a \in A \ | \ a^{-1}= \,^\gamma\hskip-1pt a \}$ and two elements $a_1$, $a _2 \in Z^1(\Gamma,A)$ satisfy $a_1 \sim a_2$ if $a_2=b^{-1} a_1\,^\gamma\hskip-1pt b$ for some $b \in A$.
\end{definition}
\begin{remark}\label{rem:cohom-ab}
If $A$ is an abelian group, then $H^1(\Gamma, A)$ is an abelian group.
In this case, we can also define $H^2(\Gamma,A)$ which identifies with the group $A^\Gamma/\{a \,^\gamma\hskip-1pt a\ |\ a\in A\}$; see \cite[\S~I.2]{Ser02} for details.
\end{remark}
\begin{remark} \label{rk:2 torsion}
We have $a^2 =a (a^{-1})^{-1}=a(\,^\gamma\hskip-1pt a)^{-1} \sim 1$ for all $a \in Z^1(\Gamma,A)$. In the case where $H^1(\Gamma, A)$ is finite, this implies that its cardinal is a power of $2$.
\end{remark}
\begin{notation}
The Galois cohomology obviously depends on the $\Gamma$-action on the group $A$. In this article, we will sometimes consider different $\Gamma$-actions for the same group $A$. If the action is not clear from the context, then we will specify the action by writing: $H^i(\Gamma,A)$ for the $\Gamma$-action on $A$ induced by $\sigma$; this means that $\sigma$ is an involution on $A$, and that the non-trivial element $\gamma\in\Gamma$ acts on $A$ by applying $\sigma$.
For example, if $A$ an automorphism group of a variety $X$, and $\mu$ is an involution on $X$, then when we write $H^i(\Gamma,A)$ for the $\Gamma$-action on $A$ induced by $\mu$-conjugation, we mean that the non-trivial element $\gamma\in\Gamma$ acts on $A$ by conjugating automorphisms by $\mu$.
\end{notation}
We now calculate the first and second cohomology groups for the case where $A=\mathbb{T}$ is a torus, and the $\Gamma$-action on $\mathbb{T}$ is induced by a real group structure.
The result for the first cohomology group will be used later in the article, in Lemma \ref{lem:injective-chi} and in Proposition \ref{prop:number of structures}. We will use the result for the second cohomology group in Remark \ref{rem:n0}.
\begin{proposition}\label{prop:torus-cohom}
Let $\mathbb{T}$ be a torus endowed with a real group structure equivalent to a product $\sigma_0^{\times n_0} \times \sigma_1^{\times n_1} \times \sigma_2^{\times n_2}$ (with the notation of Lemma \ref{lem: real form on tori}), then
\begin{itemize}
\item[(i)] $H^1(\Gamma, \mathbb{T}) \simeq (\mu_2)^{n_1}$; and
\item[(ii)] $H^2(\Gamma, \mathbb{T}) \simeq (\mu_2)^{n_0}$.
\end{itemize}
\end{proposition}
\begin{proof} This result comes from a direct computation. More precisely, if $\mathbb{T} \simeq \mathbb{G}_m$ with the real structure $\sigma_0$, given $a\in \mathbb{T}^\Gamma$, there exists $b\in\mathbb{T}$ such that $a=\pm b\,^\gamma\hskip-1pt b$. If $\mathbb{T} \simeq \mathbb{G}_m$ with the real structure $\sigma_1$, any $a\in\mathbb{T}^\Gamma$ is of the form $b\,^\gamma\hskip-1pt b$ for a choice of $b\in T$, and the same holds for $\mathbb{T}\simeq\mathbb{G}_m^2$ endowed with the real structure $\sigma_2$. Now by using the definitions of the two cohomology groups, one finds that each $\sigma_0$-component of the real structure induces a non-trivial component of $H^2(\Gamma,\mathbb{T})$, isomorphic to $\mu_2$, and each $\sigma_1$-component of the real structure induces a non-trivial component of $H^1(\Gamma,\mathbb{T})$, isomorphic to $\mu_2$. The $\sigma_0$- and $\sigma_2$-components have no effect on the first cohomology group, and the $\sigma_1$- and $\sigma_2$-components have no effect on the second cohomology group.
\end{proof}
Let $\sigma$ be a real group structure on the complex reductive algebraic group $G$.
As $\sigma$ is a group involution, it preserves $Z(G)$.
Let $\mathrm{Inn}(G) \simeq G/Z(G)$ be the group of inner automorphisms of $G$ and let $\mathrm{Out}(G)$ be the quotient group $\Aut_{gr}(G)/\mathrm{Inn}(G)$.
The Galois group $\Gamma$ acts on $\Aut_{gr}(G)$ by $\sigma$-conjugation, i.e. $\,^\gamma\hskip-1pt \varphi=\sigma \circ \varphi \circ \sigma$ for all $\varphi \in \Aut_{gr}(G)$; this $\Gamma$-action stabilizes $\mathrm{Inn}(G)$ on which it coincides with the $\Gamma$-action induced by $\sigma$ on $G/Z(G)$.
The short exact sequence
\begin{equation*}
1 \to \mathrm{Inn}(G) \to \Aut_{gr}(G) \to \mathrm{Out}(G)\to 1
\end{equation*}
induces a long exact sequence in Galois cohomology (see \cite[\S~5.5]{Ser02}):
\[ \hspace{-9mm} 1 \to \mathrm{Inn}(G)^\Gamma \to \Aut_{gr}(G)^\Gamma \to \mathrm{Out}(G)^\Gamma \to H^1(\Gamma,\mathrm{Inn}(G)) \to H^1(\Gamma,\Aut_{gr}(G)) \xrightarrow[]{\kappa} H^1(\Gamma,\mathrm{Out}(G)).\]
The next theorem gathers the main results we will need regarding the classification of the real group structures on $G$ via Galois cohomology.
\begin{theorem} \label{th:ABC Galois theory}
We keep the previous notation, and we fix a Borel subgroup $B \subseteq G$ and a maximal torus $T \subseteq B$. The following statements hold:
\begin{enumerate}[(i),leftmargin=*]
\item \label{item:correspondance H1 and real structures} Let $\Gamma$ act on $ \Aut_{gr}(G)$ by $\sigma$-conjugation as above. Then the map
\begin{equation*}
H^1(\Gamma,\Aut_{gr}(G)) \to \{\text{real group structures on $G$}\}/\text{equiv} \ \ \text{induced by}\ \varphi \mapsto \varphi \circ \sigma
\end{equation*}
is a bijection that sends the identity element to the equivalence class of $\sigma$.
\item \label{item: fibers parametrized inner twists} We have $\kappa(\sigma_1)=\kappa(\sigma_2)$ if and only if $\sigma_2$ is equivalent to an inner twist of $\sigma_1$.
\item \label{item: one qs in each fiber}There is exactly one equivalence class of quasi-split real group structures in each non-empty fiber of the map $\kappa$. Moreover, in each of them there is a quasi-split real group structure that stabilizes $B$ and $T$.
\item \label{item: dynkin conjugacy classes} If moreover $G$ is simply-connected semisimple, then the map
\[ \{ \text{quasi-split real group structures on $G$ preserving $B$ and $T$}\} \to \Aut(\mathrm{Dyn}(G)) \]
of Definition \ref{def: Gamma-action on X(T)} induces a bijection between the set of equivalence classes of quasi-split real group structures and the set of conjugacy classes of elements of order $\leq 2$ in $\Aut(\mathrm{Dyn}(G))$.
\end{enumerate}
\end{theorem}
\begin{proof}
All the proofs and details of the statements can be found in \cite[\S~7]{Con14}.
\end{proof}
\begin{example} \item
\begin{enumerate}[leftmargin=*]
\item We keep the notation of Example \ref{ex: forms of SL3} and we write $\sigma_{ij}=\sigma_i \times \sigma_j$.
The group $G=\SL_3 \times \SL_3$ has seven inequivalent real group structures: $\sigma_{00}$ (split), $\sigma_{01}$ (quasi-split) and its inner twist $\sigma_{02}=\mathrm{inn}_{(1,c)} \circ \sigma_{01}$, $\sigma_{11}$ (quasi-split) and its two inner twists $\sigma_{12}=\mathrm{inn}_{(1,c)} \circ \sigma_{11}$ and $\sigma_{22}=\mathrm{inn}_{(c,c)} \circ \sigma_{11}$, and the quasi-split real group structure $\sigma:(g_1,g_2) \mapsto (\sigma_0(g_2),\sigma_0(g_1))$. Also, $\Aut(\mathrm{Dyn}(G))$ is the dihedral group $\left \langle r,s \ | \ r^4=s^2=(sr)^2=1 \right \rangle$ which is the union of five conjugacy classes $\{1\}$, $\{r^2\}$, $\{s,rsr^{-1}\}$, $\{ sr, rs \}$, and $\{r,r^{-1}\}$, four of which consist of elements of order $\leq 2$. We may check that the bijection in Theorem \ref{th:ABC Galois theory} \ref{item: dynkin conjugacy classes} is the following one:
$\sigma_{00} \leftrightarrow \{1\}$,
$\sigma_{01} \leftrightarrow \{sr,rs\}$,
$\sigma_{11} \leftrightarrow \{r^2\}$,
and $\sigma \leftrightarrow \{s,rsr^{-1}\}$.
\smallskip
\item Let $G=\Spin_8$. Then $G$ has a unique (up to equivalence) split real group structure denoted, as usual, by $\sigma_0$. Note that $\mathrm{Dyn}(G)=D_4$ and that $\Aut(\mathrm{Dyn}(G)) \simeq \mathfrak{S}_3$.
This group has exactly one conjugacy class of non-trivial elements of order $2$, which corresponds via the bijection in Theorem \ref{th:ABC Galois theory} \ref{item: dynkin conjugacy classes} to the unique (up to equivalence) quasi-split non-split real group structure $\sigma_1$. Moreover, there are two other inequivalent real group structures on $G$, which are both inner twists of $\sigma_0$, and one other inequivalent real group structure on $G$ which is an inner twist of $\sigma_1$. The real parts of these real groups structures are the real Lie groups $\Spin(8-m,m)$ with $0\le m\le 4$. The case $m=4$ corresponds to $\sigma_0$, and its inner twists correspond to $m=2$ and $m=0$ (the later is the real part of the compact real group structure on $G)$. The case $m=3$ corresponds to $\sigma_1$, and the case $m=1$ corresponds to the inner twist of $\sigma_0$.
\end{enumerate}
\end{example}
\subsection{A cohomological invariant} \label{sec: useful map}
Let $(G,\sigma_{qs})$ be a complex reductive algebraic group with a quasi-split real group structure.
We consider the short exact sequence of $\Gamma$-groups
\[1\to Z(G)\to G\to G/Z(G)\to 1,\]
where the $\Gamma$-action is induced by $\sigma_{qs}$. More precisely, the element $\gamma\in\Gamma$ acts on $G$ and $Z(G)$ by $\sigma_{qs}$, and on $G/Z(G)$ by the induced real group structure $\overline{\sigma_{qs}}$.
Since $Z(G)$ is an abelian group, there is a connecting map (see \cite[\S~I.5.7]{Ser02})
\begin{equation*}
\delta: H^1(\Gamma, G/Z(G))\to H^2(\Gamma,Z(G)).
\end{equation*}
It follows from Lemma \ref{lem: parametrization of the inner twists} and Definition \ref{def:Galois H1} that $Z^1(\Gamma, G/Z(G))$ identifies with the set of inner twists of $\sigma_{qs}$ and
\[H^1(\Gamma, G/Z(G)) \simeq \{ c \in G \ | \ c \sigma_{qs}(c) \in Z(G) \}/\sim \]
where $c \sim c'$ if $c^{-1}b^{-1}c' \sigma_{qs}(b) \in Z(G)$ for some $b \in G$.
Also, there is an isomorphism of abelian groups (see Remark \ref{rem:cohom-ab}):
\[ H^2(\Gamma,Z(G)) \simeq Z(G)^\Gamma/\{a \sigma_{qs}(a) \ | \ a \in Z(G)\}.\]
With these identifications, the connecting map $\delta$ is the map induced by
\[ \{ c \in G \ | \ c \sigma_{qs}(c) \in Z(G) \} \to Z(G)^\Gamma,\; c \mapsto c \sigma_{qs}(c).\]
If $\sigma$ is a real group structure on $G$ equivalent to $\mathrm{inn}_c \circ \sigma_{qs}$, then we will also write $\delta(\sigma)$ instead of $\delta(\overline{c})$.
\begin{definition}
When $G$ is a simply-connected semisimple algebraic group, the element $\delta(\sigma)$ is called the \emph{Tits class} of the real group structure $(G,\sigma)$.
\end{definition}
Tables where the Tits classes are determined for any $(G,\sigma)$, with $G$ a simply-connected simple algebraic group and $\sigma$ a real group structure on $G$, can be found in Appendix \ref{sec: tables}.
\smallskip
Consider now the case where $H$ is a subgroup of $G$ such that:
\begin{itemize}
\item the algebraic group $N_G(H)/H$ is abelian (this is the case, for example, if $H$ is a spherical subgroup of $G$ by \cite[Proposition 3.4.1]{Per14}); and
\item $H$ is conjugate to a subgroup $H'$ which is stable by $\sigma_{qs}$.
\end{itemize}
Replacing $H$ by $H'$, we may assume that $\sigma_{qs}(H)=H$ to simplify the situation.
Then $\sigma_{qs}$ induces a real group structure on $N_G(H)/H$, namely $\overline{\sigma_{qs}}(nH)=\sigma_{qs}(n)H$, and we can consider the second cohomology group $H^2(\Gamma, N_G(H)/H)$. The natural homomorphism $\chi_H:Z(G)\to N_G(H)/H$, induced by the inclusion $Z(G) \to N_G(H)$,
yields an homomorphism between the second cohomology groups
\[ \chi_H^*:H^2(\Gamma, Z(G)) \to H^2(\Gamma, N_G(H)/H).\]
In the rest of this article we will denote the composed map $\chi_H^* \circ \delta$ by
\begin{equation} \label{map delta} \tag{\textasteriskcentered}
\Delta_H: H^1(\Gamma, G/Z(G)) \to H^2(\Gamma, N_G(H)/H).
\end{equation}
The element $\Delta_H(\sigma) \in H^2(\Gamma, N_G(H)/H)$ is the \emph{cohomological invariant} referred to in the title of this subsection.
\begin{remark}\label{rem:n0}
A consequence of Proposition \ref{prop:torus-cohom} is that if $N_G(H)/H$ is a torus, and if the $\Gamma$-action is induced by a real group structure on this torus with $n_0=0$, then $H^2(\Gamma, N_G(H)/H)$ is trivial and so $\Delta_H$ is the trivial map.
\end{remark}
\begin{lemma} \label{lem:injective-chi}
With the previous notation, let $H$ be a maximal unipotent subgroup of $G$. Then the two conditions above are satisfied and $\Delta_H(\sigma)$ is trivial if and only if $\delta(\sigma)$ is trivial.
\end{lemma}
\begin{proof}
First, note that $\sigma_{qs}$ stabilizes a Borel subgroup $B$, and therefore also its maximal unipotent subgroup $U$. Also, $N_G(H)=B$, and therefore $N_G(H)/H=\mathbb{T}$ is a torus (isomorphic to a maximal torus of $G)$. Thus the two conditions above hold.
By applying an appropriate conjugation, we can assume that $H=U$ and then $\sigma_{qs}(H)=H$.
Now we will show that $\chi_U^*$ is injective, which will imply the result.
Consider the short exact sequence
\vspace{-2mm}
\[0\mapsto Z(G)\to \mathbb{T}\to\overline{\mathbb{T}}=\mathbb{T}/Z(G)\to 0.\]
This exact sequence induces an exact sequence of cohomology groups:
\[H^1(\Gamma,\overline{\mathbb{T}})\to H^2(\Gamma,Z(G))\to H^2(\Gamma,\mathbb{T}),\]
where the second map is simply $\chi_U^*$.
The torus $\overline{\mathbb{T}}$ is isomorphic to a maximal torus of the adjoint semi-simple group $G/Z(G)$ and the quasi-split real group structure $\overline{\sigma_{qs}}$ on $G/Z(G)$, induced by $\sigma_{qs}$ on $G$, acts on the character group of $\overline{\mathbb{T}}$ by permutations. Indeed, since $\overline{\sigma_{qs}}$ stabilizes the Borel subgroup $\overline{B}$ of $G/Z(G)$, it preserves the positive roots of $(G/Z(G),\overline{B},\overline{\mathbb{T}})$.
This means in particular that the restriction of $\overline{\sigma_{qs}}$ on $\overline{\mathbb{T}}$ is equivalent to $\sigma_0^{\times n_0}\times\sigma_2^{\times n_2}$ (that is, there are no factors of type $\sigma_1)$.
In the cohomology group $H^1(\Gamma,\overline{\mathbb{T}})$, the $\Gamma$-action on $\mathbb{T}$ is induced by $\sigma_{qs}$. Thus, by Proposition \ref{prop:torus-cohom}, the group $ H^1(\Gamma,\overline{\mathbb{T}})$ is trivial. This implies that $\chi_U^*$ is injective.
\end{proof}
The map $\Delta_H$ will be a key-ingredient in \S~\ref{sec: non-quasi split case} to determine the existence of a $(G,\sigma)$-equivariant real structure on a horospherical homogeneous space $G/H$ when $\sigma$ is a non quasi-split real group structure on $G$.
\section{Equivariant real structures} \label{sec:equiv real structures}
In this section, we recall the notion of equivariant real structures on $G$-varieties.
We show how to determine if such a structure exists on homogeneous spaces, and, if so, how to use Galois cohomology to determine the set of equivalence classes of these equivariant real structures. We also make some observations on the existence of extensions of these real structures on quasi-homogeneous spaces. These results will be used in the next section.
\begin{definition}
Let $(G,\sigma)$ be a complex algebraic group with a real group structure, and let $X$ be a $G$-variety.
\begin{enumerate}[(i),leftmargin=*]
\item A \emph{$(G,\sigma)$-equivariant real structure} on $X$ is an antiregular involution $\mu$ on $X$ such that
\vspace{-2mm}
\[ \forall g \in G, \; \forall x \in X, \;\; \mu(g \cdot x)=\sigma(g) \cdot \mu(x).\]
\item Two equivariant real structures $\mu$ and $\mu'$ on a $(G,\sigma)$-variety $X$ are \emph{equivalent} if there exists a $G$-equvariant automorphism $\varphi \in \Aut^G(X)$ such that $\mu'=\varphi \circ \mu\circ \varphi^{-1}$.
\end{enumerate}
\end{definition}
\begin{remark}
Let $(X,\mu)$ be a $(G,\sigma)$-equivariant real structure. Whenever the real part $X_0=X(\mathbb{C})^\mu$ of $X$ is non-empty, it defines a real manifold endowed with an action of the real Lie group $G_0=G(\mathbb{C})^\sigma$.
\end{remark}
\begin{remark}
If $X$ is a \emph{quasi-homogeneous} $G$-variety, i.e. a $G$-variety with an open orbit $G/H$, then a given equivariant real structure $\mu$ on $G/H$ need not extend to $X$ (as we will see in \S~\ref{sec:extension of real structures}). If $\mu$ extends to $X$, then this extension is unique.
\end{remark}
\begin{lemma} \label{lem: two conditions}
Let $(G,\sigma)$ be a complex algebraic group with a real group structure, and let $X=G/H$ be a homogeneous space.
Then $X$ has $(G,\sigma)$-equivariant real structure if and only if there exists $g \in G$ such that these two conditions hold:
\begin{enumerate}
\item \label{eq: sigma compatible}
\emph{$(G,\sigma)$-compatibility condition:} $\sigma(H)=gHg^{-1}$
\item \label{eq: involution}
\emph{involution condition:}\hspace{15mm} $\sigma(g)g \in H$
\end{enumerate}
in which case one is given by $\mu(kH)=\sigma(k)gH$ for all $k\in G$.
\end{lemma}
\begin{proof}
If $\mu$ defines a $(G,\sigma)$-equivariant real structure on $X$, then $\mu$ is determined by $\mu(eH)=gH$ for some $g \in G$.
Then $\mu$ must be compatible with $\sigma$, i.e. $\mu(eH)=\mu(hH)=\sigma(h)\mu(eH)$ for all $h \in H$, this yields condition \ref{eq: sigma compatible}. Also, $\mu$ must be an involution, this yields condition \ref{eq: involution}.
For the converse, if $g \in G$ satisfies the conditions \ref{eq: sigma compatible} and \ref{eq: involution}, then the map $\mu:G\to G$ defined by $\mu(kH)=\sigma(k)gH$ is clearly a $(G,\sigma)$-equivariant real structure on $X$.
\end{proof}
\begin{remark} \label{rk:conjugate}
If $H'$ is conjugate to $H$, then $G/H$ has a $(G,\sigma)$-equivariant real structure $\mu$ if and only if $G/H'$ has a $(G,\sigma)$-equivariant real structure $\mu'$.
Indeed, if $H'=kHk^{-1}$ and $\mu(eH)=gH$, then we can define $\mu'$ by $\mu'(eH')=g'H$ with $g'=\sigma(k)gk^{-1}$.
We check that $\sigma(H')=g'H'{g'}^{-1}$ if and only if $\sigma(H)=gH{g}^{-1}$, and that $\sigma(g')g' \in H'$ if and only if $\sigma(g) g \in H$.
\end{remark}
Let us note that, if $N_G(H)=H$, then the condition \ref{eq: involution} of Lemma~\ref{lem: two conditions} holds and we recover \cite[Theorem~2.1]{Akh15}. Also, with the notation of Lemma \ref{lem: two conditions}, we have $X(\mathbb{C})^\mu \neq \emptyset$ if and only if there exists $k \in G$ such that $k^{-1} \sigma(k)g \in H$.
\begin{example} \label{ex:Equivariant real structures on toric varieties}
(Equivariant real structures on toric varieties.)\\
For a complete account on toric varieties, we refer the interested reader to \cite{Ful93}.
Let $(T,\sigma)$ be a complex torus with a real group structure (those were described in Lemma \ref{lem: real form on tori}). Let $X$ be a complex toric variety with open orbit $X_0 \simeq T$.
We start with the case of a homogeneous toric variety, that is, we consider the case $X=X_0$.
It is easy to check, using
Lemma \ref{lem: two conditions} that
the homogeneous space $X_0$ always has a $(T,\sigma)$-equivariant real structure;
we can for instance choose $t_0=1$ and consider $\mu=\sigma$. Now we can find all the equivalence classes of $(T,\sigma)$-equivariant structures on $X_0$. We use the notation of Remark \ref{rem:torus-variety}. Suppose that $T$ is endowed with the real group structure $\sigma=\sigma_0^{\times n_0}\times\sigma_1^{\times n_1}\times\sigma_2^{\times n_2}$. Then each antiregular involution of the form
$\sigma_0^{\times n_0}\times\mu'_1\times\cdots\times\mu'_{n_1}\times\sigma_2^{\times n_2}$, where $\mu'_i=\sigma_1$ or $\tau_1$ for each $i=1,\ldots, n_1$,
defines a $(T,\sigma)$-equivariant real structure on $T$, and no two of these involutions are (equivariantly) equivalent. Moreover, any $(T,\sigma)$-equivariant real structure on $T$ is of this form.
As for the quasi-homogeneous case, by \cite[Theorem 1.25]{Hur11} the equivariant real structure $\mu$ on $X_0$ extends on $X$ if and only if the $\Gamma$-action on $\mathbb{X}^\vee \otimes_{\mathbb{Z}} \mathbb{Q}$, introduced in Definition \ref{def: Gamma-action on X(T)}, stabilizes the fan of $X$.
\end{example}
\begin{lemma} \label{lem:structure-exist}
Let $(G,\sigma_1)$ be a complex algebraic group with a real group structure, and let $X=G/H$ be a homogeneous space. Let $\sigma_2=inn(c)\circ \sigma_1$ be an inner twist of $\sigma_1$ (for some $c \in G$), and suppose that $\sigma_1(H)=H$. Then
\begin{enumerate}[(i),leftmargin=*]
\item $X$ has $(G,\sigma_1)$-equivariant real structure; and
\item $X$ has a $(G,\sigma_2)$-equivariant real structure if and only if there exists $n\in N_G(H)$ such that $c\sigma_1(c)\sigma_1(n)n\in H$.
\end{enumerate}
\end{lemma}
\begin{proof}
The first statement follows from Lemma \ref{lem: two conditions} (take $g=1)$.
For the second statement, note that $\sigma_2(H)=cHc^{-1}$. According to Lemma \ref{lem: two conditions}, the variety $X$ has a $(G,\sigma_2)$-equivariant real structure if and only if there exists $g\in G$ such that $gHg^{-1}=\sigma_2(H)=cHc^{-1}$, and $g\sigma_2(g)\in H$.
But $gHg^{-1}=cHc^{-1}$ if and only if there exists $n\in N_G(H)$ such that $g=cn$.
Then the second condition in Lemma \ref{lem: two conditions} yields $\sigma_2(g)g=c\sigma_1(g)c^{-1}g=c\sigma_1(c)\sigma_1(n)n \in H$. The converse is straightforward, it suffices to take $g=cn$.
\end{proof}
In the next proposition, we use Lemma \ref{lem:structure-exist} to give a cohomological condition to determine the existence of an equivariant real structure on $G/H$. Because of the well-known tables of structures on semisimple algebraic groups, this cohomological interpretation is particularly well-adapted to calculate examples.
\begin{proposition} \label{prop:coho condition}
Let $(G,\sigma_{qs})$ be a complex reductive algebraic group with a quasi-split real group structure, and let $\sigma=\mathrm{inn}_c \circ \sigma_{qs}$ be an inner twist of $\sigma_{qs}$ (for some $c \in G$). Let $X=G/H$ be a homogeneous space, and assume that $N_G(H)/H$ is abelian and $\sigma_{qs}(H)=H$. Then
\begin{enumerate}[(i),leftmargin=*]
\item $X$ has a $(G,\sigma_{qs})$-equivariant real structure; and
\item $X$ has a $(G,\sigma)$-equivariant real structure if and only if $\Delta_H(\sigma)$ is trivial.
\end{enumerate}
\end{proposition}
\begin{proof}
Unraveling the definition of $\Delta_H$ in \S~\ref{sec: useful map}, we verify that the cohomological condition $\Delta_H(\sigma)$ being trivial is equivalent to the second condition in Lemma \ref{lem:structure-exist}.
The result then follows directly from Lemma \ref{lem:structure-exist}.
\end{proof}
\begin{remark}
In \cite{Bor}, Borovoi considers the case where $X$ is a quasi-projective $G$-variety over an arbitrary field of characteristic zero which admits a $(G,\sigma_{qs})$-equivariant real structure. He obtains a similar cohomological criterion for the existence of a $(G,\sigma)$-equivariant real structure on $X$, but his point of view differs from ours in the fact that he uses exclusively Galois cohomology while we go through group-theoretical considerations (Lemmas \ref{lem: two conditions} and \ref{lem:structure-exist}).
\end{remark}
\begin{remark}
If $(G,\sigma)$ is a simply-connected simple algebraic group with a real group structure, then the Tits class of $(G,\sigma)$ is often trivial (see tables in Appendix \ref{sec: tables}) in which case by Proposition \ref{prop:coho condition} the existence of a $(G,\sigma)$-equivariant real structure on $G/H$ reduces to the study of the conjugacy class of $H$.
\end{remark}
The previous results provide conditions for the existence of an equivariant real structure on a homogeneous space. On the other hand, the classical way to determine the number of equivalence classes for such structures is via Galois cohomology.
\begin{lemma} \label{lem:Galois H1 to param eq real structures} \emph{(Equivariant real structures and Galois cohomology.)}\\
Let $(G,\sigma)$ be a complex algebraic group with a real group structure, and let $(X,\mu_0)$ be a homogeneous space with a $(G,\sigma)$-equivariant real structure. The Galois group $\Gamma$ acts on $\Aut^G(X)$ by $\mu_0$-conjugacy. Then the map
\[\begin{array}{ccl}
\hspace{-5mm} H^1(\Gamma,\Aut^G(X)) &\to &\{ \text{equivalence classes of $(G,\sigma)$-equivariant real structures on $X$}\}\\
\varphi &\mapsto &\ \ \ \varphi \circ \mu_0
\end{array}\]
is a bijection that sends the identity element to the equivalence class of $\mu_0$.
\end{lemma}
\begin{proof}
Note that any antiregular automorphism of $X$ is of the form $\mu=\varphi\circ\mu_0$, where $\varphi$ is a regular automorphism on $X$. Now $\mu$ defines a $(G,\sigma)$-equivariant real structure if and only if $\mu\circ\mu=id$ and $\varphi$ is $G$-equivariant.
With the notation of Definition \ref{def:Galois H1}, this is equivalent to requiring that $\varphi\in Z^1(\Gamma, \Aut^G(X))$. Finally, $\mu=\varphi\circ\mu_0$ is equivalent to $\mu'=\varphi'\circ\mu_0$, with $\varphi,\varphi'\in Z^1(\Gamma, \Aut^G(X))$, if and only if there exists $\psi\in \Aut^G(X)$ such that $\psi\circ\mu' \circ \psi^{-1}=\mu$, that is, $\psi\circ\varphi \circ (\,^\gamma\hskip-1pt \psi^{-1})=\varphi'$. This is precisely the equivalence condition defining $H^1(\Gamma, \Aut^G(X))$.
\end{proof}
The interested reader may also consult \cite[\S~I\!I\!I.1]{Ser02} for more general results on the classification of real structures via Galois cohomology.
If $X=G/H$ is a homogeneous space, then we recall that $\Aut^G(G/H) \simeq N_G(H)/H$ (see e.g. \cite[Proposition 1.8]{Tim11}). In particular, if $N_G(H)=H$, then $\Aut^G(X)$ is the trivial group.
\begin{corollary} \label{cor: H=N(H)}
\emph{(see also \cite[Theorem~4.12]{ACF14})} \
Let $(G,\sigma)$ be a complex reductive algebraic group with a real group structure, and let $X=G/H$ be a homogeneous space such that $N_G(H)=H$.
Then $X$ has a $(G,\sigma)$-equivariant real structure $\mu$ if and only if $\sigma(H)=cHc^{-1}$ for some $c \in G$. Moreover, if $\mu$ exists then it is equivalent to $\mu: gH \mapsto \sigma(g)cH$.
\end{corollary}
\begin{proof}
If $N_G(H)=H$, then Condition \ref{eq: sigma compatible} in Lemma \ref{lem: two conditions} implies Condition \ref{eq: involution}, and so $G/H$ has a $(G,\sigma)$-equivariant real structure if and only if $\sigma(H)$ is conjugate to $H$.
As $\Aut^G(X) \simeq N_G(H)/H= \{1\}$ the uniqueness part of the statement follows from Lemma \ref{lem:Galois H1 to param eq real structures}. The last statement is given by Lemma \ref{lem: two conditions}.
\end{proof}
\begin{example} \label{ex:G/P 1}
Let $(G,\sigma)$ be a complex reductive algebraic group with a real group structure, and let $X=G/P$ be a flag variety. Then $X$ has an equivariant real structure $\mu$ if and only if $\sigma(P)$ is conjugate to $P$. Moreover, if $\mu$ exists then it is equivalent to $\mu: gP \mapsto \sigma(g)cP$, where $c \in G$ satisfies $\sigma(P)=cPc^{-1}$.
\end{example}
\section{Horospherical varieties} \label{sec:equi real groups strcture for horo}
The main part of this section deals with the question of existence of equivariant real structures on horospherical homogeneous spaces.
Let $H$ be a horospherical subgroup of a reductive algebraic group $G$. The group $\mathbb{T}=N_G(H)/H$ is a torus and we apply the results of the previous sections (where $N_G(H)/H$ is assumed to be abelian) to determine whether there exists an equivariant real structure on $G/H$.
We explain also how to determine the number of equivalence classes of equivariant real structures on $G/H$. Then we recall when the equivariant real structures extend to a given horospherical variety. Finally, we end this section with the study of equivariant real structures on classical examples of horospherical varieties that arise in algebraic geometry.
\subsection{Setting and first definitions} \label{sec:setting}
We fix once and for all a triple $(G,B,T)$, where $G$ is a complex reductive algebraic group, $B \subseteq G$ is a Borel subgroup and $T \subseteq B$ is a maximal torus. Let $\SS=\SS(G,B,T)$ be the set of simple roots corresponding to the root system associated with the triple $(G,B,T)$. We denote by $\sigma_0$ a split real group structure on $G$ such that $\sigma_0(B)=B$ and $\sigma_0(T)=T$.
\begin{definition}
A subgroup $H$ of $G$ is \emph{horospherical} if it contains a maximal unipotent subgroup of $G$.
A homogeneous space $G/H$ is \emph{horospherical} if $H$ is a horospherical subgroup of $G$.
\end{definition}
\begin{remark}
Let $\sigma$ be a real group structure on $G$. Then $\sigma$ maps a maximal unipotent subgroup of $G$ to a maximal unipotent subgroup of $G$ and so the set of horospherical subgroups of $G$ is preserved by $\sigma$.
\end{remark}
\begin{example}
Tori and flag varieties are examples of horospherical homogeneous spaces.
Let $U$ be a maximal unipotent subgroup of $\SL_2$, then $\SL_2/U$ is a horospherical homogeneous space isomorphic to the affine plane minus the origin $\mathbb{A}^2 \setminus \{0\}$.
\end{example}
\begin{definition}
A \emph{horospherical $G$-variety} is a normal $G$-variety with an open horospherical $G$-orbit.
\end{definition}
\begin{example} \label{ex:SL2}
It follows from the combinatorial description of spherical embeddings (see e.g. \cite[\S~15.1]{Tim11}) that the horospherical $\SL_2$-varieties with open orbit $\SL_2$-isomorphic to $\SL_2/U$ are the following:
$\mathbb{A}^2 \setminus \{0\}$, $\mathbb{A}^2$, $\P^2$, $\P^2 \setminus \{0\}$, ${\mathrm{Bl}}_0(\mathbb{A}^2)$, and ${\mathrm{Bl}}_0(\P^2)$.
\end{example}
Other examples of horospherical varieties can be found in \cite{Pas06,Pas18b}.
\bigskip
We now recall the combinatorial description of the horospherical subgroups given in \cite{Pas08}.
For $I \subseteq \SS$, we denote by $P_I$ the \emph{standard parabolic subgroup} generated by $B$ and the unipotent subgroups of $G$ associated with the simple roots $\alpha \in I$ (this gives a 1-to-1 correspondence between the power set of $\SS$ and the set of conjugacy classes of parabolic subgroups of $G$). In particular, $P_{\emptyset}=B$.
Let $I \subseteq \SS$ and let $M$ be a sublattice of $\mathbb{X}$ such that $\left \langle \chi,\alpha^\vee \right \rangle =0$ for all $\alpha \in I$ and all $\chi \in M$, where $ \left \langle \cdot,\cdot \right \rangle: \mathbb{X} \times \mathbb{X}^\vee \to \mathbb{Z}$ is the duality bracket, and $\alpha^\vee \in \mathbb{X}^\vee$ is the coroot associated to the root $\alpha$; this condition simply means that $\chi$ extends to a character of the parabolic subgroup $P_I$.
Then
\begin{equation*}
H_{(I,M)}= \bigcap_{\chi \in M} \Ker(\chi)
\end{equation*}
is a horospherical subgroup of $G$ whose normalizer is $P_I$ (see \cite[\S~2]{Pas08} for details).
\begin{definition}
The horospherical subgroup $H_{(I,M)}$ defined above is called \emph{standard horospherical subgroup}.
If $H$ is a horospherical subgroup, then by \cite[Proposition 2.4]{Pas08} there exists a unique pair $(I,M)$ as above such that $H$ is conjugate to $H_{(I,M)}$; we write that $(I,M)$ is the (\emph{horospherical})\emph{datum} of $H$.
\end{definition}
\begin{example} \label{ex:flag varieties 1}
The datum of a parabolic subgroup conjugate to $P_I$ is $(I,\{0\})$. The datum of a maximal unipotent subgroup of $G$ is $(\emptyset ,\mathbb{X})$.
\end{example}
If $H$ is a horospherical subgroup of $G$, then $P:= N_G(H)$ is a parabolic subgroup of $G$ and the quotient group $\mathbb{T}= P/H \simeq \Aut^G(G/H)$ is a torus (see \cite[Proposition 2.2 and Remarque 2.2]{Pas08}).
\subsection{Quasi-split case} \label{sec:quasi-split horo}
We denote by $\sigma_{qs}$ a quasi-split real group structure on $G$ preserving $B$ and $T$ as before.
\begin{lemma} \label{lem:two Gamma actions on PI}
Let $\Gamma$ act on the set $\SS$ as in Definition \ref{def: Gamma-action on X(T)}.
If $I \subseteq \SS$, then $\sigma_{qs}(P_I)=P_{\,^\gamma\hskip-1pt{I}}$,
where $\,^\gamma\hskip-1pt I=\{ \,^\gamma\hskip-1pt s \ | \ s \in I\} \subseteq \SS$.
\end{lemma}
\begin{proof}
First note that $\sigma_{qs}(B)=B$ implies that $\sigma(P_I)$ is a parabolic subgroup of the form $P_J$ for some $J \subseteq \SS$ (see \cite[\S~29.4, Th.]{Hum75}).
Since $\Gamma$ acts on $\SS$ by permutation (see Definition \ref{def: Gamma-action on X(T)}), we have $\,^\gamma\hskip-1pt \alpha \in \SS$ for each $\alpha \in \SS$.
Also, if $\alpha\in\SS$ and $U_{-\alpha}$ is the corresponding unipotent subgroup of $G$, then $\sigma_{qs}(U_{-\alpha})=U_{-\,^\gamma\hskip-1pt\alpha}$ by definition of $U_{-\alpha}$ and of the $\Gamma$-action on $\SS$.
But $P_I=\left \langle B, U_{-\alpha} \text{ with } \alpha \in I \right \rangle $, and so the last sentence implies the result.
\end{proof}
\begin{proposition} \label{prop: datum of the conjugate}
Let $H$ be a horospherical subgroup of $G$ with datum $(I,M)$.
Then the horospherical datum of $\sigma_{qs}(H)$ is $\,^\gamma\hskip-1pt (I, M):=(\,^\gamma\hskip-1pt I,\,^\gamma\hskip-1pt M)$ with the $\Gamma$-action on $\mathbb{X}$ introduced in Definition \ref{def: Gamma-action on X(T)}.
Also, $\sigma_{qs}(H_{(I,M)})=H_{\,^\gamma\hskip-1pt (I,M)}$, and so $\sigma_{qs}(H_{(I,M)})$ is conjugate to $H_{(I,M)}$ if and only if $\sigma_{qs}(H_{(I,M)})=H_{(I,M)}$.
\end{proposition}
\begin{proof}
First of all, by Lemma \ref{lem:two Gamma actions on PI}, we have $\sigma_{qs}(P_I)=P_{\,^\gamma\hskip-1pt I}$. Thus, the normalizer of $\sigma_{qs}(H)$ is $\sigma_{qs}(P_I)=P_{\,^\gamma\hskip-1pt I}$.
Also, by definition of the $\Gamma$-action on the coroot lattice $\mathbb{X}^{\vee}$ (see Definition \ref{def: Gamma-action on X(T)}), we have $\left \langle \,^\gamma\hskip-1pt m, \,^\gamma\hskip-1pt (\alpha^\vee) \right \rangle =\left \langle m,\alpha^\vee \right \rangle=0$ for all $\alpha \in I$ and $m \in M$. Thus the sublattice $M'$ of $\mathbb{X}$ associated with $\sigma_{qs}(H)$ is $\,^\gamma\hskip-1pt M$. This proves the first part of the proposition.
Now we have $\sigma_{qs}(P_I)=P_{\,^\gamma\hskip-1pt I}$ by Lemma \ref{lem:two Gamma actions on PI}, and
\[ \sigma_{qs}(H_{(I,M)})=\bigcap_{m \in M} \sigma_{qs}\left( \Ker(m)\right)= \bigcap_{m \in M} \Ker(\,^\gamma\hskip-1pt m)=H_{\,^\gamma\hskip-1pt (I,M)}.\]
Therefore, if $\sigma_{qs}(H_{(I,M)})=H_{\,^\gamma\hskip-1pt (I,M)}$ is conjugate to $H_{(I,M)}$, they have the same horospherical data, that is, $\,^\gamma\hskip-1pt (I,M)=(I,M)$, which finishes the proof.
\end{proof}
\begin{remark}
The previous result can also be obtained as a consequence of \cite[Theorem 3]{CF15} which states that for a spherical subgroup $H \subseteq G$, the subgroups $H$ and $\sigma(H)$ are conjugate in $G$ if and only if the Luna-Vust invariants of $G/H$ are $\Gamma$-stable.
\end{remark}
\begin{remark}
Let $H$ be a spherical subgroup of $G$ such that $\sigma_{qs}(H)$ is conjugate to $H$. If $N_G(N_G(H))=N_G(H)$ (which holds when $H$ is horospherical), then, by \cite[Theorem~4.14]{ACF14} (see also \cite[Proposition~2.5]{MJT}), there exists a spherical subgroup $H'$ conjugate to $H$ such that $\sigma_{qs}(H')=H'$.
\end{remark}
\begin{corollary} \label{cor:equiv reals tructure if invs preserved}
Let $H$ be a horospherical subgroup of $G$ with datum $(I,M)$.
Then $X=G/H$ has a $(G,\sigma_{qs})$-equivariant real structure if and only if $\,^\gamma\hskip-1pt (I,M)=(I,M)$.
\end{corollary}
\begin{proof}
By remark \ref{rk:conjugate}, we may assume that $H=H_{(I,M)}$.
If $X=G/H$ has a $(G,\sigma_{qs})$-equivariant real structure, then $H$ and $\sigma_{qs}(H)$ are conjugate by Lemma \ref{lem: two conditions}, and so $\,^\gamma\hskip-1pt (I,M)=(I,M)$ by Proposition \ref{prop: datum of the conjugate}.
Conversely, if $\,^\gamma\hskip-1pt (I,M)=(I,M)$, then $\sigma_{qs}(H)=H$ by Proposition \ref{prop: datum of the conjugate}. Thus we see that the two conditions of Lemma \ref{lem: two conditions} are satisfied for $g=1$, which means that $X$ has a $(G,\sigma_{qs})$-equivariant real structure.
\end{proof}
\begin{remark}
In the quasi-split case, if there exists a $(G,\sigma_{qs})$-equivariant real structure on $X=G/H$, then there exists one, say $\mu$, such that $X(\mathbb{C})^\mu \neq \emptyset$. Indeed, as in the previous corollary, we may assume that $H=H_{(I,M)}$ and that $\sigma_{qs}(H)=H$. Then $\mu(eH)=eH$, and so $eH \in X(\mathbb{C})^\mu$. \end{remark}
\begin{remark} \label{rk:split case}
If $\sigma_{qs}=\sigma_0$ is a split real group structure on $G$, then $X=G/H$ has always a $(G,\sigma_{qs})$-equivariant real structure since $\Gamma$ acts trivially on $\mathbb{X}$.
\end{remark}
\begin{example} \label{ex:H=U}
If $H=U$ is a maximal unipotent subgroup of $G$, then its horospherical datum $(\emptyset, \mathbb{X})$ is $\Gamma$-stable, and so there always exists a $(G,\sigma_{qs})$-equivariant real structure on $G/U$, namely $\overline{\sigma_{qs}}(gU)=\sigma_{qs}(g)U$.
\end{example}
\begin{example}\label{ex:SL4 part2}
Let $T$ be the maximal torus of $G=\SL_4$ formed by diagonal matrices, and let $B$ be the Borel subgroup formed by upper-triangular matrices.
Let $L_i:T \to \mathbb{G}_m, (t_1,t_2,t_3,t_4) \mapsto t_i$, where $i \in \{1,2,3,4\}$.
Then the simple roots of $(G,B,T)$ are $\alpha_1=L_1-L_2$, $\alpha_2=L_2-L_3$, and $\alpha_3=L_3-L_4$.
Let $P \subseteq G$ be the standard parabolic subgroup
associated with $I=\{\alpha_2\}$, and let $H$ be the kernel of the character $\chi=L_1+L_4$ in $P$.
Then $H$ is a horospherical subgroup of $G$ with datum $(I,M)=(\{\alpha_2\},\mathbb{Z}\left\langle \chi \right\rangle)$.
The group $G$ has two inequivalent quasi-split real group structures, namely $\sigma_s(g)=\overline{g}$ (split) and $\sigma_{qs}(g)=\mathrm{inn}_h \circ \sigma_c(g)$ (non split quasi-split), where $\sigma_c(g)=\leftexp{t}{\overline{g}}^{-1}$ is the compact real group structure on $G$ and $h=\begin{bmatrix}
0 & 0 &0 &-i \\ 0& 0 & 1 &0\\ 0 & 1&0 &0\\ i & 0 & 0 & 0
\end{bmatrix}$. The $\Gamma$-action on $\mathbb{X}(T)$ induced by $\sigma_s$ is trivial (Remark \ref{rk:split case}), and so $\,^\gamma\hskip-1pt (I,M)=(I,M)$. On the other hand, the $\Gamma$-action induced by $\sigma_{qs}$ is determined by the relations $\,^\gamma\hskip-1pt L_1=-L_4$ and $\,^\gamma\hskip-1pt L_2=-L_3$. Thus, we still have $\,^\gamma\hskip-1pt (I,M)=(I,M)$, but the $\Gamma$-action on $M$ is non-trivial since $\,^\gamma\hskip-1pt \chi=-\chi$. By Corollary \ref{cor:equiv reals tructure if invs preserved}, the homogeneous space $X=G/H$ has a $(G,\sigma_s)$- and a $(G,\sigma_{qs})$-equivariant real structure.
\end{example}
\subsection{General case} \label{sec: non-quasi split case}
We now consider the case where $\sigma$ is any real group structure on the complex reductive algebraic group $G$.
By Theorem \ref{th:ABC Galois theory}, there exists a quasi-split real structure $\sigma_{qs}$ (uniquely defined by $\sigma$, up to equivalence) and an element $c \in G$ such that $\sigma$ is equivalent to $ \mathrm{inn}_c \circ \sigma_{qs}$. We may replace $\sigma$ and $\sigma_{qs}$ in their equivalence classes and assume that $\sigma_{qs}(B)=B$, $\sigma_{qs}(T)=T$, and $\sigma=\mathrm{inn}_c \circ \sigma_{qs}$ to avoid technicalities.
Then we still have a well-defined $\Gamma$-action on $\mathbb{X}$ (preserving $\SS)$ induced by $\sigma_{qs}$ (see Definition \ref{def: Gamma-action on X(T)}).
\begin{lemma} \label{lem: G1 form implies G0 form}
Let $H$ be a horospherical subgroup with datum $(I,M)$.
If $X=G/H$ has a $(G,\sigma)$-equivariant real structure, then $\,^\gamma\hskip-1pt (I,M)=(I,M)$.
\end{lemma}
\begin{proof}
By Remark \ref{rk:conjugate}, we may assume that $H=H_{(I,M)}$ is a standard horospherical subgroup. If $X$ has a $(G,\sigma)$-equivariant real structure, then Lemma \ref{lem: two conditions}~\ref{eq: sigma compatible} yields that $\sigma(H)$ and $H$ are conjugate. Thus $\sigma_{qs}(H)=\mathrm{inn}_{c^{-1}} \circ \sigma(H)=c^{-1} \sigma(H)c$ and $H$ are also conjugate, and so $\,^\gamma\hskip-1pt (I,M)=(I,M)$ by Proposition \ref{prop: datum of the conjugate}.
\end{proof}
The previous lemma means that the existence of a $(G,\sigma)$-equivariant real structure on $G/H$ implies the existence of a $(G,\sigma_{qs})$-equivariant real structure on $G/H$. However, we will see that these two conditions are not equivalent.
\begin{proposition} \label{prop:coho cond for horospherical space}
Let $\sigma=\mathrm{inn}_c \circ \sigma_{qs}$ be a real group structure on $G$ as above.
Let $H$ be a horospherical subgroup of $G$ with datum $(I,M)$, and assume that $X=G/H$ has a $(G,\sigma_{qs})$-equivariant real structure.
Then $X$ has a $(G,\sigma)$-equivariant real structure if and only if $\Delta_H(\sigma)$ is trivial, where $\Delta_H$ is the map defined in \S~\ref{sec: useful map}.
\end{proposition}
\begin{proof}
First, by Remark \ref{rk:conjugate}, we may assume that $H=H_{(I,M)}$ is a standard horospherical subgroup. By assumption, $X$ has a $(G,\sigma_{qs})$-equivariant real structure, and so $\sigma_{qs}(H)=H$ by Proposition \ref{prop: datum of the conjugate}.
Now the result follows from Proposition \ref{prop:coho condition}.
\end{proof}
We now summarize our main results in the following theorem:
\begin{theorem} \label{th:main results}
Let $H$ be a horospherical subgroup of $G$ with datum $(I,M)$.
Let $\sigma=\mathrm{inn}_c \circ \sigma_{qs}$ be a real group structure on $G$ where $\sigma_{qs}$ is the corresponding quasi-split real group structure preserving $B$ and $T$.
Then the following four conditions are equivalent:
\begin{enumerate}[(i),leftmargin=*]
\item \label{item main th i} $G/H$ has a $(G,\sigma_{qs})$-equivariant real structure;
\item \label{item main th ii} $\,^\gamma\hskip-1pt (I,M)=(I,M)$;
\item \label{item main th iii} $H$ is conjugate to $\sigma(H)$;
\item \label{item main th iv} $H$ is conjugate to $\sigma_{qs}(H)$.
\end{enumerate}
Moreover $G/H$ has a $(G,\sigma)$-equivariant real structure if and only if the (equivalent) conditions (i)-(iv) are satisfied and $\Delta_H(\sigma)$ is trivial, where $\Delta_H$ is the map defined in \S~\ref{sec: useful map}.
\end{theorem}
\begin{proof}
The equivalence of \ref{item main th i} and \ref{item main th ii} is Corollary \ref{cor:equiv reals tructure if invs preserved}. The equivalence of \ref{item main th iii} and \ref{item main th iv} is straightforward since $\sigma=\mathrm{inn}_c \circ \sigma_{qs}$. The equivalence of \ref{item main th i} and \ref{item main th iv} follows from Proposition \ref{prop: datum of the conjugate} and Lemma \ref{lem: two conditions}.
The second part of the theorem is Proposition \ref{prop:coho cond for horospherical space}.
\end{proof}
\begin{corollary}
With the notation of Theorem \ref{th:main results}, the horospherical homogeneous space $G/H$ has a $(G,\sigma)$-equivariant real structure if $\,^\gamma\hskip-1pt (I,M)=(I,M)$ and one of the following conditions holds:
\begin{enumerate}[(i),leftmargin=*]
\item $H^2(\Gamma,Z(G))=\{1\}$; or
\item $H^2(\Gamma,N_G(H)/H)=\{1\}$.
\end{enumerate}
\end{corollary}
\begin{proof}
Let us note that if (i) or (ii) holds, then the map $\Delta_H$ is trivial.
By Theorem \ref{th:main results}, if $\,^\gamma\hskip-1pt (I,M)=(I,M)$ and $\Delta_H$ is trivial, then the horospherical homogeneous space $G/H$ has a $(G,\sigma)$-equivariant real structure.
\end{proof}
\begin{example} \label{ex:G/P 2}
(Equivariant real structures on flag varieties.)\\
Let $X=G/P$ be a flag variety, and let $I \subseteq \SS$ be such that the standard parabolic subgroup $P_I$ is conjugate to $P$. By Remark \ref{rk:conjugate} we may assume that $P=P_I$.
We saw in Example \ref{ex:G/P 1} that $X$ has a $(G,\sigma)$-equivariant real structure if and only if $\sigma(P)$ is conjugate to $P$, which we also recover from Theorem \ref{th:main results} since $\Delta_P(\sigma)$ is always trivial.
\end{example}
\begin{example} \label{ex:H=U part 2}
Let $(G,\sigma)$ be a simply-connected simple algebraic group with a real group structure, and let $U$ be a maximal unipotent subgroup of $G$. By Example \ref{ex:H=U}, there always exists a $(G,\sigma_{qs})$-equivariant real structure on $G/U$. Thus, by Theorem \ref{th:main results} and Lemma \ref{lem:injective-chi}, there exists a $(G,\sigma)$-equivariant real structure on $G/U$ if and only if the Tits class $\delta(\sigma)$ is trivial.
\end{example}
\begin{example}\label{ex:SL4 part3}
We resume Example \ref{ex:SL4 part2}.
The complex algebraic group $G=\SL_4$ has five inequivalent real group structures (see Appendix \ref{sec: tables}): a split one $\sigma_s$ (with real part $\SL_4(\mathbb{R}))$ and an inner twist $\sigma_s'$ (with real part $\SL_2(\mathbb{H}))$, the non split quasi-split one $\sigma_{qs}$ (with real part $\SU(2,2))$ and two inner twists $\sigma_{qs}'$ and $\sigma_c$ (with real parts $\SU(3,1)$ and $\SU(4))$.
By Theorem \ref{th:main results} and Example \ref{ex:SL4 part2}, the homogeneous space $X=G/H$ has a $(G,\sigma)$-equivariant real structure if and only if $\Delta_H(\sigma)$ is trivial.
Let $\mathbb{T}=P/H \simeq \mathbb{G}_m$. If $\sigma=\sigma_s$ resp. $\sigma=\sigma_{qs}$, then the real group structure on $\mathbb{T}$ induced by $\sigma$ is equivalent to $\sigma_0$ resp. to $\sigma_1$ (this follows from a direct computation).
Therefore, by Proposition \ref{prop:torus-cohom}, the group $H^2(\Gamma,\mathbb{T})$ is trivial when the $\Gamma$-action on $\mathbb{T}$ comes from $\sigma_{qs}$, and so $\Delta_H(\sigma)$ is trivial for the inner twists of $\sigma_{qs}$. It remains to consider the case $\sigma=\sigma_s'$.
Note that for this case, there exists $c\in G$ such that $\sigma_s'=\mathrm{inn}_c\circ \sigma_s$, and that $c\sigma_s(c)\in Z(G)$ and is fixed by $\sigma_s$. This means that $c\sigma_s(c)=\pm 1\in H$. Thus $\Delta_H(\sigma_s')$ is trivial, and so by Theorem \ref{th:main results} there exists a $(G,\sigma_s')$-equivariant real structure on $X$.
\end{example}
\begin{remark}
We will see in the next sections that only $\sigma_{qs}$, the quasi-split inner twist of $\sigma$, matters to count the number of equivalence classes of $(G,\sigma)$-equivariant real structures on $G/H$ or to determine whether a given $(G,\sigma)$-equivariant real structure on $G/H$ extends to an equivariant embedding of $G/H$. More precisely, let $X$ be a horospherical variety with open $G$-orbit isomorphic to $X_0=G/H$. Suppose that $\sigma$ is an inner twist of $\sigma_{qs}$. Then if at least one $(G,\sigma)$-equivariant real structure exists on $X_0$, then the number of $(G,\sigma)$-equivariant real structures on $X_0$ (resp. on $X$) is exactly the same as the number of $(G,\sigma_{qs})$-equivariant real structures on $X_0$ (resp. on $X$).
\end{remark}
\subsection{Number of equivalence classes of equivariant real structures} \label{sec:number of equi real structures}
Let $X=G/H$ be a horospherical homogeneous space, and let $\sigma$ be a real group structure on $G$. As before, we may choose $\sigma$ such that that $\sigma=\mathrm{inn}_{c} \circ \sigma_{qs}$, for some $c \in G$, where $\sigma_{qs}$ is a quasi-split real group structure on $G$ stabilizing $B$ and $T$.
In this section \textbf{we suppose that there exists a $(G,\sigma)$-equivariant real structure $\mu_0$ on $X$.} By \cite[Chp.~I\!I\!I, \S~4.3, Th.~4]{Ser02} the group $H^1(\Gamma,\Aut^G(X))$ is finite. Moreover, since all elements of $H^1(\Gamma,\Aut^G(X))$ are $2$-torsion (see Remark \ref{rk:2 torsion}), we have $H^1(\Gamma,\Aut^G(X)) \simeq (\mu_2)^r$ for some non-negative integer $r$ that we want to determine.
By Remark \ref{rk:conjugate}, we may replace $H$ by a standard horospherical subgroup conjugate to $H$ and assume that $\sigma_{qs}(H)=H$. Then $\sigma_{qs}$ induces a real group structure $\overline{\sigma_{qs}}$ on the torus $\mathbb{T}=N_G(H)/H \simeq \Aut^G(X)$. By Lemma \ref{lem: real form on tori}, the real group structure $\overline{\sigma_{qs}}$ is equivalent to
\[\sigma_0^{\times n_0} \times \sigma_1^{\times n_1} \times \sigma_2^{\times n_2} \text{ for some } n_0,n_1,n_2 \in \mathbb{N} \text{ such that } n_0+n_1+2n_2=\dim(\mathbb{T}).\]
\begin{proposition} \label{prop:number of structures}
With the notation and assumptions above, there are exactly $2^{n_1}$ equivalence classes of $(G,\sigma)$-equivariant real structures on $X=G/H$.
\end{proposition}
\begin{proof}
We treat first the case when $\sigma=\sigma_{qs}$.
Note that for this case, there exists a natural equivariant real structure on $X$, namely $\mu_0=\overline{\sigma_{qs}}$ the real structure induced by $\sigma_{qs}$. Then, by Lemma \ref{lem:Galois H1 to param eq real structures},
there is a bijection between the set of equivalence classes of $(G,\sigma_{qs})$-equivariant real structures on $X$ and
$H^1(\Gamma, \Aut^G(X))$, where $\Gamma$ acts on $\Aut^G(X)$ by $\mu_0$-conjugation.
Identifying $\Aut^G(X)$ with $\mathbb{T}=N_G(H)/H$, the $\Gamma$-action on $\Aut^G(X)$ coincides with the $\Gamma$-action on $\mathbb{T}$ induced by $\sigma_{qs}$.
Thus, the number of equivalence classes of $(G,\sigma)$-equivariant real structures on $X$ equals the cardinality of $H^1(\Gamma, \mathbb{T})$, where the action of $\Gamma$ on $\mathbb{T}$ is given by $\,^\gamma\hskip-1pt(aH)=\sigma_{qs}(a)H$ for any $a\in N_G(H)$. Since $\overline{\sigma_{qs}}$ is equivalent to $\sigma_0^{\times n_0} \times \sigma_1^{\times n_1} \times \sigma_2^{\times n_2}$ by assumption, the cardinality of $H^1(\Gamma,\mathbb{T})$ is $2^{n_1}$ by Proposition \ref{prop:torus-cohom}.
We now consider the general case when $\sigma=\mathrm{inn}_c \circ \sigma_{qs}$.
First of all, there is no particular privileged choice of the equivariant real structure $\mu_0$ on $X$ to define the $\Gamma$-action on $\Aut^G(X)$, as there was in the quasi-split case, namely $\mu_0=\overline{\sigma_{qs}}$. However we assumed that such a $\mu_0$ exists and this will be sufficient for our purposes.
By Lemma \ref{lem:structure-exist}, this means there exists $n\in N_G(H)$ such that $\sigma_{qs}(c)c\sigma_{qs}(n)n\in H$, and then we can take $\mu_0$
defined by $\mu_0(kH)=\sigma(k)cnH$ for all $k \in N_G(H)$.
Let $\varphi \in \Aut^G(X)$ and $a \in N_G(H)$ such that $\varphi(kH)=kaH$, for all $k\in G$. Then
\begin{align*}
\hspace{-10mm} &\,^\gamma\hskip-1pt \varphi (kH)=\mu_0 \circ \varphi \circ \mu_0(kH)
= \mu_0 \circ \varphi(\sigma(k)cnH)\\
=& \mu_0 (\sigma(k)cnaH)
= k\sigma(cna)cnH
= kc \sigma_{qs}(cna)nH.
\end{align*}
Note that $c\sigma_{qs}(cna)nH=c\sigma_{qs}(cn)n(n^{-1}\sigma_{qs}(a)n)H$. Since $c\sigma_{qs}(cn)n\in H$, and $(n^{-1}\sigma_{qs}(a)n)\in N_G(H)$, we find
\[\,^\gamma\hskip-1pt \varphi (kH)=kn^{-1}\sigma_{qs}(a)nH.\]
But $\mathbb{T}$ is abelian, and so $\,^\gamma\hskip-1pt \varphi (kH)=k\sigma_{qs}(a)H$, and therefore the $\Gamma$-action on $\mathbb{T}$ corresponding to the $\mu_0$-conjugation is given by $\,^\gamma\hskip-1pt (aH)=\sigma_{qs}(a)H$ as in the quasi-split case that we treated first. Thus the number of equivalence classes of $(G,\sigma)$-equivariant real structures on $X$ is again equal to the cardinality of $H^1(\Gamma,\mathbb{T})$, where the $\Gamma$-action on $\mathbb{T}$ is the one induced by $\sigma_{qs}$. This finishes the proof.
\end{proof}
\begin{remark} \label{rk:split case number}
Suppose that $\mathrm{Out}(G)=\{1\}$. By Theorem \ref{th:ABC Galois theory} any real group structure on $G$ is an inner twist of the split one $\sigma_s$, and so the induced real group structure on $\mathbb{T}$ is equivalent to $\sigma_0^{\times n_0}$. Thus Proposition \ref{prop:number of structures} implies that when a $(G,\sigma)$-equivariant real structure exists on $G/H$, then it is unique up to equivalence.
\end{remark}
\begin{example} \label{ex:number of real structures on tori}
Let $G=T$ be a torus with a real group structure $\sigma$ equivalent to $\sigma_0^{\times n_0} \times \sigma_1^{\times n_1} \times \sigma_2^{\times n_2}$ for some $n_0,n_1,n_2 \in \mathbb{N}$. Then the number of equivalence classes of $(G,\sigma)$-equivariant real structures on $X=T$ is $2^{n_1}$. This could also be seen directly from Example \ref{ex:Equivariant real structures on toric varieties}.
\end{example}
\begin{corollary} \label{cor:number case G/U}
With the same notation and assumptions as in Proposition \ref{prop:number of structures}, if we moreover assume that $H=U$ is a maximal unipotent subgroup of $G$, then there is a unique equivalence class of $(G,\sigma)$-equivariant real structures on $X=G/U$.
\end{corollary}
\begin{proof}
If $H=U$, then the inclusion $T \hookrightarrow B$ yields an isomorphism $T \simeq B/U=\mathbb{T}$, and so the induced real group structure $\overline{\sigma_{qs}}$ on $\mathbb{T}$ coincides with the restriction $(\sigma_{qs})_{|T}$. Since $\sigma_{qs}(B)=B$, this structure must preserve the positive roots of $(G,B,T)$, which means that $(\sigma_{qs})_{|T}$ is equivalent to a product $\sigma_0^{n_0} \times \sigma_2^{n_2}$. Therefore $n_1=0$ and the result follows from Proposition \ref{prop:number of structures}.
\end{proof}
\begin{example}
Let $G=\SL_3$. In Example \ref{ex: forms of SL3} we saw that $G$ has three inequivalent real group structures, namely $\sigma_s$ (split), $\sigma_{qs}$ (non split quasi-split), and $\sigma_c$ (compact, inner twist of $\sigma_{qs})$. Since $\sigma_s$ and $\sigma_{qs}$ stabilize $B$ (up to conjugate), they stabilize also $U$ and so there exists an equivariant real structure on $X=G/U$ in these two cases. Moreover, as the Tits class of $(G,\sigma_c)$ is trivial (see Table 2 in Appendix \ref{sec: tables}), it follows from Theorem \ref{th:main results} and Lemma \ref{lem:injective-chi} that there exists also a $(G,\sigma_c)$-equivariant real structure on $X$. Finally, by Corollary \ref{cor:number case G/U}, each of these three equivariant real structures is unique up to equivalence.
\end{example}
\begin{example}\label{ex:SL4 part4}
We resume Example \ref{ex:SL4 part3}. We saw that if $\sigma=\sigma_s$ resp. $\sigma=\sigma_{qs}$, then the real group structure on $\mathbb{T}$ induced by $\sigma$ is equivalent to $\sigma_0$ resp. to $\sigma_1$. Thus, by Proposition \ref{prop:number of structures}, there is one equivalence class of $(G,\sigma)$-equivariant real structures on $G/H$ when $\sigma$ is equivalent to an inner twist of $\sigma_s$, and there are two equivalence classes of $(G,\sigma)$-equivariant real structures on $G/H$ when $\sigma$ is equivalent to an inner twist of $\sigma_{qs}$.
\end{example}
\subsection{Extension of the equivariant real structures} \label{sec:extension of real structures}
As before we fix a triple $(G,B,T)$, where $G$ is a complex reductive algebraic group, $B \subseteq G$ is a Borel subgroup, and $T \subseteq B$ is a maximal torus.
Let $\sigma$ be a real group structure on $G$. In this section we determine when a given $(G,\sigma)$-equivariant real structure on a horospherical homogeneous space $G/H$ extends to a $(G,\sigma)$-equivariant real structure on a horospherical variety whose open orbit is $G/H$.
For the next theorem, we will consider the $\Gamma$-action on the set of colored cones, defined by Huruguen in \cite{Hur11}. Let $\sigma$ be an inner twist of a quasi-split real group structure $\sigma_{qs}$ on $G$, and consider all the colored fans defining $G$-equivariant embeddings $G/H\hookrightarrow X$. Then the $\Gamma$-action on $\mathbb{X}^\vee \otimes_\mathbb{Z} \mathbb{Q}$ defined by $\sigma_{qs}$ (see Definition \ref{def: Gamma-action on X(T)}) induces a $\Gamma$-action on this set of colored fans. In particular, if $\sigma_{qs}=\sigma_0$ is split, then this $\Gamma$-action is trivial.
\begin{theorem} \label{th:Hur-Wed}\emph{(\cite[Theorem 2.23]{Hur11} and \cite[Theorem 9.1]{Wed}.)}\\
Let $\mu$ be a $(G,\sigma)$-equivariant real structure on a \textbf{spherical} homogeneous space $G/H$, and let $X$ be a spherical $G$-variety with open orbit $G/H$. Then the real structure $\mu$ extends on $X$ if and only if the colored fan of the spherical embedding $G/H \hookrightarrow X$ is $\Gamma$-invariant.
\end{theorem}
\begin{remark} \label{rk:alg space}
If the equivariant real structure $\mu$ on $G/H$ extends to $X$, then the corresponding real form $X/\Gamma$ always exists as a real algebraic space but not necessarily as a real variety; see \cite[\S~2.4]{Hur11} for such an example.
\end{remark}
\begin{corollary} \label{cor:extension}
Let $\mu$ be a $(G,\sigma)$-equivariant real structure on a \textbf{horospherical} homogeneous space $G/H$, and let $X$ be a horospherical $G$-variety with open orbit $G/H$. Then the real structure $\mu$ extends on $X$ if and only if the colored fan of the embedding $G/H \hookrightarrow X$ is $\Gamma$-invariant in which case the corresponding real form $X/\Gamma$ is a real variety.
\end{corollary}
\begin{proof}
The condition for the equivariant real structure $\mu$ on $G/H$ to extend to $X$ is given by Theorem~\ref{th:Hur-Wed}. It remains to observe that when $\mu$ extends to $X$, the real algebraic space $X/\Gamma$ is a real variety. This follows from the fact that horospherical $G$-varieties are covered with $\Gamma$-stable quasi-projective horospherical $G$-varieties. A proof of this fact is obtained by using the characterization due to M. Brion of quasi-projectivity for spherical varieties (see e.g. \cite[Corollary 3.2.12]{Per14}) together with the fact that $\Gamma$ is of order $2$.
\end{proof}
\begin{example}
Consider the equivariant embeddings of $SL_2/U$ of Example \ref{ex:SL2}. There are two inequivalent real group structures on $SL_2$: $\sigma_0$ which is split, and $\sigma_1$, whose real part is compact. Note that $\sigma_1$ is an inner twist of $\sigma_0$. We deduce from Example \ref{ex:H=U part 2} and Remark \ref{rk:split case number} (or Corollary \ref{cor:number case G/U}) that there exists a unique equivalence class of $(SL_2,\sigma_0)$-equivariant real structure on $SL_2/U$, but that there is no $(SL_2,\sigma_1)$-equivariant structure on $SL_2/U$ as the Tits class $\delta(\sigma_1)$ is non-trivial.
Thus, any $(SL_2,\sigma_0)$-equivariant real structure on $SL_2/U$ extends to $X$.
\end{example}
\begin{example}\label{ex:SL4 part5}
We resume Example \ref{ex:SL4 part4}. Let $\sigma$ be an inner twist of a non split quasi-split real group structure on $G=\SL_4$. We see from Example \ref{ex:SL4 part2} that the $\Gamma$-action on $N:=M^\vee=\mathbb{Z} \left \langle \chi^\vee \right \rangle $ induced by $\sigma$ satisfies $\,^\gamma\hskip-1pt \chi^\vee=-\chi^\vee$. Let $\mathcal{F} \subseteq N_\mathbb{Q}$ be a colored fan corresponding to a $G$-equivariant embedding $G/H \hookrightarrow Y$.
Then, by Corollary \ref{cor:extension}, a $(G,\sigma)$-equivariant real structure on $G/H$ extends to $Y$ if and only if the colored fan $\mathcal{F}$ is symmetric with respect to the origin of $N_\mathbb{Q}$. It follows from Luna-Vust theory that either $Y=G/H$ (case $\mathcal{F}=\{ (\{0\},\emptyset )\})$ or $Y$ is a $\P^1$-bundle over $G/P$ which is the union of two $G$-orbits of codimension $1$, the two $G$-invariant sections of the structure morphism $Y \to G/P$, and the open $G$-orbit (case $\mathcal{F}=\{(\mathbb{Z}_+ \left \langle \chi^\vee \right \rangle,\emptyset),(\mathbb{Z}_- \left \langle \chi^\vee \right \rangle,\emptyset)\})$.
\end{example}
\subsection{Smooth projective horospherical varieties of Picard rank 1} \label{sec: smooth proj Picard rank one}
In this section we apply the results obtained in the previous sections to classify the real structures on the smooth projective horospherical $G$-varieties of Picard rank $1$.
Examples of such varieties are given by the flag varieties $X=G/P$, with $P$ a maximal parabolic subgroup, and the odd symplectic grassmannians; these correspond to the case (3) in Theorem \ref{th:Pasquier classication Picard 1} and were studied for example in \cite{Mih07,Pec13}.
The smooth projective horospherical $G$-varieties of Picard rank $1$ were classified by Pasquier in \cite{Pas09} who proved the following result:
\begin{theorem} \label{th:Pasquier classication Picard 1} \emph{(\cite[Theorem 0.1]{Pas09})}
Let $X$ be a smooth projective horospherical $G$-variety of Picard rank $1$. Then either $X=G/P$ is a flag variety (with $P$ a maximal parabolic subgroup) or $X$ has three $G$-orbits and can be constructed in a uniform way from a triple $(\mathrm{Dyn}(G),\varpi_Y,\varpi_Z)$ belonging to the following list:
\begin{enumerate}
\item $(B_n,\varpi_{n-1},\varpi_n)$ with $n \geq 3$;
\item $(B_3,\varpi_1,\varpi_3)$;
\item $(C_n,\varpi_m,\varpi_{m-1})$ with $n \geq 2$ and $m \in [2,n]$;
\item $(F_4,\varpi_2,\varpi_3)$;
\item $(G_2,\varpi_1,\varpi_2)$,
\end{enumerate}
where $\varpi_Y$, $\varpi_Z$ are fundamental weights of $G$ such that the two closed orbits of $X$ are $G$-isomorphic to the flag varieties $G/P(\varpi_Y)$ and $G/P(\varpi_Z)$. (Here, if $\varpi$ is a fundamental root, $P(\varpi)$ is the parabolic subgroup $P_I$, where $I=\SS\setminus\{\varpi\}$.)
\end{theorem}
We have already looked at equivariant real structures on flag varieties in Examples \ref{ex:G/P 1} and \ref{ex:G/P 2}. Therefore, we will only consider equivariant real structures in the non-homogeneous cases.
\begin{theorem} \label{th:real forms of horo Picard 1}
We keep the notation of Theorem \ref{th:Pasquier classication Picard 1}. Let $\sigma$ be a real group structure on $G$, let $G_0$ be the corresponding real part, and let $X$ be a non-homogeneous smooth projective horospherical $G$-variety of Picard rank $1$ associated with a triple $(\mathrm{Dyn}(G),\varpi_Y,\varpi_Z)$. Then $X$ admits a $(G,\sigma)$-equivariant real structure if and only if $(\mathrm{Dyn}(G),G_0,\varpi_Y,\varpi_Z)$ belongs to the following list:
\begin{enumerate}
\item $(B_n,G_0, \varpi_{n-1},\varpi_n)$ with $G_0=\Spin_{n+4t,n+1-4t}(\mathbb{R})$ and $n \geq 3$, $t \in \mathbb{Z}$;
\item $(B_3, G_0, \varpi_1,\varpi_3)$ with $G_0=\Spin_7(\mathbb{R})$ or $\Spin_{3,4}(\mathbb{R})$;
\item $(C_n, \Sp(2n,\mathbb{R}),\varpi_m,\varpi_{m-1})$ with $n \geq 2$ and $m \in [2,n]$;
\item $(F_4, G_0,\varpi_2,\varpi_3)$ with $G_0$ the real part of one of the three inequivalent real group structures on $F_4$; or
\item $(G_2, G_0,\varpi_1,\varpi_2)$ with $G_0$ the real part of one of the two inequivalent real group structure on $G_2$ (the split one and the compact one).
\end{enumerate}
Moreover, when such a structure exists, then it is unique up to equivalence.
\end{theorem}
\begin{proof}
In cases (1)-(5) of Theorem \ref{th:Pasquier classication Picard 1}, we observe that $\Aut(\mathrm{Dyn}(G))=\{1\}$. Thus, by Remark \ref{rk:split case number}, if a $(G,\sigma)$-equivariant real structure exists on $G/H$, then it is unique up to equivalence. Also, by Theorem \ref{th:ABC Galois theory} \ref{item: dynkin conjugacy classes} any real group structure $\sigma$ on $G$ is an inner twist of the split real group structure $\sigma_0$ on $G$. Therefore the induced $\Gamma$-action on $\mathbb{X}(T)$ is trivial (Remark \ref{rk:trivial split action}). This has two important consequences:
First, by Theorem \ref{th:main results}, the open orbit $X_0=G/H$ always admits a $(G,\sigma_0)$-equivariant real structure.
Second, by Corollary \ref{cor:extension}, any $(G,\sigma)$-equivariant real structure on $X_0$ extends to $X$.
Moreover, again by Theorem \ref{th:main results}, the homogeneous space $X_0$ admits a $(G,\sigma)$-equivariant real structure if and only if $\Delta_H(\sigma)$ is trivial. (We recall that the Tits class $\delta(\sigma)$ of $(G,\sigma)$ can be found in the tables in Appendix \ref{sec: tables}.) One can check that the cases where $\delta(\sigma)$ is trivial are exactly the cases that appear in the statement of the theorem. Therefore, to finish the proof of this theorem, it suffices to prove that the homomorphism $\chi_H^*$ defined in \S~\ref{sec: useful map} is injective in the cases where $\delta(\sigma)$ is non-trivial. This will indeed imply that $\Delta_H(\sigma)$ is non-trivial.
The cases left to consider are those where $G$ is of type $B_n$, for $n\ge 3$, or of type $C_n$, for $n\ge 2$. Recall that $\mathbb{T}=N_G(H)/H$ is a quotient of the maximal torus $T$ obtained by composing the following homomorphisms:
\[ T \hookrightarrow B \twoheadrightarrow B/U \to N_G(H)/H=\mathbb{T},\]
where the map $B/U \to N_G(H)/H$ is the map induced by the inclusion $B \hookrightarrow N_G(H)$.
Note that $N_G(H)=BH=TUH=TH$, and so the homomorphism $T \to \mathbb{T}$ is indeed onto, with kernel $T \cap H$. Thus the induced real structure $\overline{\sigma_s}$ on $\mathbb{T}$ (which is a $1$-dimensional torus in cases (1)-(3)) is obtained from ${\sigma_s}_{|T} \sim \sigma_0 \times \sigma_0$, and so $\overline{\sigma_s}$ is equivalent to $\sigma_0$. By Proposition \ref{prop:torus-cohom}, this means that $H^2(\Gamma,\mathbb{T}) \simeq \mu_2$. A generator of this group is given by the class of the $\sigma_0$-invariant element $-1$.
We now consider the group $H^2(\Gamma,Z(G))$. In type $B_n$ and $C_n$, note that the center of the simply-connected simple group $G$ is $Z(G) \simeq \mu_2$. Since we are considering the cases where the Tits classes are non-trivial, the group $H^2(\Gamma,Z(G))$ is non-trivial, and, since $Z(G)\simeq \mu_2$, it is isomorphic to $Z(G)$ on which $\Gamma$ acts trivially.
Recall that the homomorphism $\chi_H^*: H^2(\Gamma,Z(G)) \simeq Z(G) \to H^2(\Gamma,\mathbb{T})$ is induced by the homomorphism $\psi: Z(G) \rightarrow \mathbb{T}$. Since $H^2(\Gamma,\mathbb{T})$ is generated by the class of $-1$, either $\psi$ is injective, and then $\chi_H^*$ is an isomorphism, or else $\psi$ is trivial, and then $\chi_H^*$ is trivial. We will show that $\psi$ is injective, i.e., that $Z(G)$ is not contained in $H$ and this will finish the proof of this theorem.
We are left to prove that $Z(G)$ is not contained in $H$. For this, use the following construction of $H$ from \cite[\S~1.3]{GPPS}. For each triple $(\mathrm{Dyn}(G),\varpi_1,\varpi_2)$ in Theorem \ref{th:Pasquier classication Picard 1}, consider the projective space $\P(V_1\oplus V_2)$, where $V_1$ and $V_2$ are the irreducible $G$-modules with highest weights $\varpi_1$ and $\varpi_2$ respectively.
Let $v_i$ be a highest weight vector of $V_i$ for $i=1,2$. Then $H$ is the stabilizer of the line generated by $v_1\oplus v_2\in\P(V_1\oplus V_2)$. With this description, we see that $Z(G)$ is contained in $H$ if and only if $Z(G)$ acts on $V_1 \oplus V_2$ by $\pm1$. However, in each case one can check
that $Z(G)$ acts trivially on $V_i$ and by -1 on $V_j$ with $\{i,j\}=\{1,2\}$. Hence, $Z(G)$ is not contained in $H$ and $\chi_H^*$ is an isomorphism.
\end{proof}
\begin{remark}
Let $(\mathrm{Dyn}(G),\varpi_1,\varpi_2)$ be a triple from Theorem \ref{th:Pasquier classication Picard 1}.
Let $\sigma=\sigma_s$ be a split real group structure on $G$, and let $\mu_s$ be a $(G,\sigma_s)$-equivariant real structure on $X$. Then the geometric construction of $X$ given in \cite[\S~1]{Pas09} (see also \cite[\S~1.3]{GPPS}) yields a construction of the real part of $(X,\mu_s)$ provided that one takes $\mathbb{R}$ as a base field in the construction instead of $\mathbb{C}$.
\end{remark}
\smallskip
\noindent \textbf{Acknowledgments.}
The authors are grateful to Mikhail Borovoi for stimulating discussions and e-mail exchanges about this project; in particular, the cohomological characterization for the existence part in Theorem \ref{th:1} was inspired by \cite{Bor}. We also thank the anonymous referees for their helpful comments.
\newpage
|
\section{Introduction}
Neutrino physics, like leptonic CP violation, is an interesting topic
\cite{neutrino} in the current research of particle physics. Among
other things, it might be the final place where experiments of particle
physics will give definite results in the near future. The results will
check various theoretical models about the fermion masses of the
Standard Model (SM).
We proposed that supersymmetry (SUSY) \cite{susy} can be the theory
underlying the fermion masses in Refs. \cite{1,2,3}. The basic idea is
the following. It assumes a flavor symmetry. The flavor symmetry is
broken after the sneutrinos obtain nonvanishing vacuum expectation
values (VEVs). (In this way, SUSY is motivated.) These VEVs result in
a nonvanishing neutrino mass. The empirical smallness of neutrino
masses needs very large SM super partner masses to be understood which
are about $10^{12}$ GeV. Thus, our SUSY is of high scale breaking
\cite{meta,hsusy,splitsusy}.
A further natural assumption is that the flavor symmetry breaks softly.
Namely the soft SUSY breaking masses of the sfermions do not obey the
flavor symmetry either. The theoretical reason is that the soft masses
are due to the supergravity effect which generically breaks any global
symmetry. Soft breaking of the flavor symmetry implies that the lepton
number violation due to sneutrino VEVs is explicit instead of being
spontaneous. Therefore there is no any massless Nambu-Goldstone boson
related to nonvanishing sneutrino VEVs. Actually the large masses of
the model make the low energy effective theory just the SM via Higgs
mass fine tuning, except for that we have an understanding of the
hierarchical pattern of the charged lepton masses, or that of the SM
Yukawa coupling constants.
To briefly review the model in a simple way, the SM is SUSY generalized.
The flavor symmetry is $Z_3$ cyclic among the three generation SU(2)$_L$
lepton doublets $L_1$, $L_2$ and $L_3$. The other fields are trivial
under $Z_3$. The $Z_3$ invariant combinations are $\sum_{i=1}^{3}L_i$
and
$\epsilon_{\alpha\beta}(L_1^{\alpha}L_2^{\beta}+L_2^{\alpha}L_3^{\beta}+L_3^{\alpha}L_1^{\beta})$
with $\alpha$ and $\beta$ denoting the SU(2)$_L$ indices.
In terms of the following redefined lepton superfields,
$L_e = \displaystyle \frac{1}{\sqrt 2}(L_1-L_2)$,
$L_\mu = \displaystyle \frac{1}{\sqrt 6}(L_1+L_2-2L_3)$,
$L_\tau = \displaystyle \frac{1}{\sqrt 3}(\sum_i L_i)$,
the above $Z_3$ invariant combinations are $L_\tau$ and
$\epsilon_{\alpha\beta}L_e^{\alpha}L_\mu^{\beta}$, respectively. The
superpotential is then
\begin{equation}
\label{1}
{\mathcal W} \supset y_\tau\epsilon_{\alpha\beta}
L_\tau^{\alpha}H_d^{\beta}E_\tau^c
+\epsilon_{\alpha\beta}L_e^{\alpha}L_\mu^{\beta}
(\lambda_\tau E_\tau^c + \lambda_\mu E_\mu^c)
+\bar{\mu}\epsilon_{\alpha\beta}H_u^{\alpha}H_d^{\beta} \,,
\end{equation}
where $H_u$ and $H_d$ are the two Higgs doublets, the right-handed
lepton singlet $E_\tau^c$ is defined as the one which couples to
$L_\tau$, and $E_\mu^c$ is that orthogonal to $E_\tau^c$ and with a
coupling to $L_eL_\mu$. $y_\tau$, $\lambda_\tau$ and $\lambda_\mu$ are
coupling constants. (Note that considering the mixing between $L_\tau$
and $H_d$ gives the same form of the above superpotential \cite{2}.) It
is seen that the electron is massless, because $E_e^c$ is always absent
in the Lagrangian. This is true whenever SUSY is conserved, the
nonvanishing electron mass is due to SUSY breaking (together with
electroweak gauge symmetry and flavor symmetry breaking via loops).
Note that all the coupling constants in our superpotential are assumed
to be natural values, say typically $\sim 0.01-1$, and the mass
parameter $\bar{\mu}$ is taken to be large $\sim 10^{12}$ GeV. The SM
fermion mass hierarchy is due to symmetries and their breaking.
In addition, a heavy vector-like SU(2)$_L$ triplet field $T(\bar{T})$
with hypercharge $2(-2)$ needs to be introduced so as to make the Higgs
mass realistic \cite{3,meta}. This triplet field also contributes to
neutrino masses. In terms of the redefined fields, the flavor symmetric
superpotential relevant to the triplet $T$ and $\bar{T}$ fields is
\begin{equation}
\label{2}
\begin{array}{lll}
{\mathcal W} & \supset & y^\nu \{L_\tau H_d\} T
+ \lambda^\nu_1 \{L_e L_e + L_\mu L_\mu\} T
+ \lambda^\nu_2 \{L_\tau L_\tau\} T \\
& & + \lambda^\nu_3 \{H_d H_d\} T
+ \lambda^\nu_4 \{H_u H_u\} \bar{T}
+ M_T T \bar{T} \,
\end{array}
\end{equation}
with $M_T$ the mass $\sim 10^{13}$ GeV. The braces denote that the two
doublets form an SU(2)$_L$ triplet representation.
The soft SUSY breaking terms in the Lagrangian are in general form which
also break the flavor symmetry \cite{1,2,3}. All the mass parameters of
the model are taken to be about $10^{12}-10^{13}$ GeV. The spontaneous
gauge symmetry breaking of the SM occurs. Through fine tuning, the
right electroweak vacuum is obtained. By including contribution due to
the triplet field, this model can give reasonable neutrino spectrum and
the mixing pattern, and predicted the right order of $\theta_{13}$
\cite{2,3}. (The quark sector was considered in Ref. \cite{2}.)
Roughly speaking about the electroweak symmetry breaking. There are
five scalar doublets, the mass parameters are all large $\sim 10^{12}$
GeV. Eigenvalues of their mass-squared matrix are generically large.
However, one of these values can be exceptional, because it is a
difference between two large parameters. It is this difference that
makes the fine-tuning possible. Whence the difference is tuned to be
about $-$(100 GeV)$^2$, correct electroweak symmetry breaking occurs.
The corresponding eigenstate field is one superposition of the five
doublets. It is the only light scalar doublet, and is just the SM
Higgs field from the point of view of the low energy effective field
theory. The SM Higgs gets a VEV is equivalent to that the original two
Higgses and sleptons get their VEVs \cite{2,3}.
\section{Complex couplings and sneutrino VEVs}
In this paper, we will carefully consider CP violation of the lepton
sector, and completely analyze the neutrino masses and mixing. In
general, the coupling constants are complex, however, because of the
flavor symmetry, many of them can be made real via field phase
rotation. In the superpotential Eq. (\ref{1}) for charged leptons, all
the couplings can be adjusted to be real. On the other hand, in the
superpotential Eq. (\ref{2}) for neutrino masses, the couplings cannot
be all taken real, as can be seen in the following way. The mass
parameters $\bar{\mu}$ and $M_T$ are taken real, thus $H_u$ and $H_d$
always have opposite phases, and so do $T$ and $\bar{T}$.
$\lambda^\nu_2$ is real via rotating the phase of $L_\tau$,
$\lambda^\nu_4$ is real via rotating $H_u$ (or $\bar{T}$), $y_\tau$ is
real via $E_\tau^c$, $\lambda_\tau$ real via $L_eL_\mu$ rotating, and
$\lambda_\mu$ real via $E_\mu^c$. In such a phase convention, only
$y^\nu$, $\lambda_1^\nu$ and $\lambda_3^\nu$ can be complex. The
$\lambda_1^\nu$ term will contribute to the neutrino masses, which was
omitted in our previous analysis \cite{3}.
In the soft SUSY breaking terms, the mass parameters and coupling
constants are generally complex, and there is no enough freedom to
rotate all of the phases away.
The scalar potential relevant to the electroweak symmetry breaking is
\begin{equation}
\begin{split}
V=&(|\bar{\mu}|^2+m_{h_u}^2)|h_u|^2+(|\bar{\mu}|^2+m_{h_d}^2)|h_d|^2
+\frac{g^2+g'^2}{8}
(|h_u|^2-|h_d|^2-{\tilde{l}_\alpha}^\dag{\tilde{l}_\alpha})^2\\
&+\frac{g^2}{4}[2|h_u^\dag h_d|^2
+2(h_u^\dag\tilde{l}_\alpha)(\tilde{l}_\alpha^\dag h_u)
+2(h_d^\dag\tilde{l}_\alpha)(\tilde{l}_\alpha^\dag h_d) \\
&-2|h_d|^2({\tilde{l}_\alpha}^\dag {\tilde{l}_\alpha})
+(\tilde{l}_\alpha^\dag\tilde{l}_\beta)(\tilde{l}_\beta^\dag\tilde{l}_\alpha)
-(\tilde{l}_\alpha^\dag\tilde{l}_\alpha)(\tilde{l}_\beta^\dag\tilde{l}_\beta)]
\\
&+(\frac{1}{2}m_{d\alpha}^2h_d^\dag\tilde{l}_\alpha
+\frac{1}{2}m_{\alpha\beta}^2\tilde{l}_\alpha^\dag\tilde{l}_\beta
+B_\mu h_u h_d+B_{\mu\alpha}h_u \tilde{l}_\alpha+{\rm h.c.})
\end{split}
\end{equation}
where $g$ and $g'$ are SM gauge coupling constants. $h_u$ and $h_d$
denote the scalar components of $H_u$ and $H_d$, respectively, and
$\tilde{l}_\alpha$'s left-handed sleptons. $m_{h_{(u,d)}}^2$,
$m_{d\alpha}^2$, $m_{\alpha\beta}^2$ and $B_\mu$, $B_{\mu\alpha}$ are
soft squared masses.
In considering CP violation of the scalar potential, the essential point
lies in the soft bilinear terms where the mass parameters are complex.
Field redefinition of $h_d$ and $\tilde{l}_\alpha$ may remove phases of
$B_\mu$ and $B_{\mu\alpha}$ respectively, however, the phases of
$m_{d\alpha}^2$ and off-diagonal terms of $m_{\alpha\beta}^2$ are still
there. This means that after the electroweak symmetry breaking, Higgs
and sneutrino VEVs are complex in general. (Previously we took all the
VEVs real.) In the analysis, we still have the freedom to choose the
VEV of Higgs field $h_u$ to be real, and VEVs of the Higgs and the
sneutrino fields are denoted as
($v_u$, $v_d e^{i\delta_{v_d}}$, $v_{l_e}e^{i\delta_{l_e}}$,
$v_{l_\mu}e^{i\delta_{l_\mu}}$, $v_{l_\tau}e^{i\delta_{l_\tau}}$) where
the phases have been explicitly written down. These VEVs enter the
lepton mass matrices and thus contribute to CP violation in the leptonic
mixing.
\section{Neutrino masses}
The sneutrino VEVs result in a nonvanishing neutrino mass,
\begin{equation}
\label{4}
M^\nu_0 =-\displaystyle\frac{a^2}{M_{\tilde{Z}}e^{i\delta_Z}}
\left(\begin{array}{ccc}
v_{l_e} v_{l_e}e^{2i\delta_{l_e}} &v_{l_e} v_{l_\mu} e^{i(\delta_{l_e}+\delta_{l_\mu})} &v_{l_e} v_{l_\tau}e^{i(\delta_{l_e}+\delta_{l_\tau})} \\
v_{l_\mu}v_{l_e}e^{i(\delta_{l_e}+\delta_{l_\mu})} &v_{l_\mu} v_{l_\mu}e^{2i\delta_{l_\mu}} &v_{l_\mu} v_{l_\tau}e^{i(\delta_{l_\mu}+\delta_{l_\tau})} \\
v_{l_\tau} v_{l_e}e^{i(\delta_{l_e}+\delta_{l_\tau})}&v_{l_\mu} v_{l_\tau}e^{i(\delta_{l_\mu}+\delta_{l_\tau})}&v_{l_\tau} v_{l_\tau}e^{2i\delta_\tau}
\end{array}
\right),
\end{equation}
where $a=\sqrt{(g^2+g'^2)/2}$, $M_{\tilde{Z}}$ is the Zino mass which is
the typical superpartner mass, and the phase of Zino mass term,
$\delta_Z$, is explicitly written. This is due to gauge interactions,
it is natural realization of the type-I seesaw mechanism \cite{type1}
where the role of right-handed neutrinos is replaced by the Zino. In
addition, the superpotential (\ref{2}) contributes following neutrino
masses \cite{3},
\begin{equation}
M_1^\nu=-\frac{\lambda_4^\nu v_u^2}{M_T}
\begin{pmatrix}
\lambda_1^\nu e^{\delta_{\lambda_1}} & 0 & 0\\
0 &\lambda_1^\nu e^{\delta_{\lambda_1}} & 0\\
0 & 0 & \lambda_2^\nu
\end{pmatrix}\,,
\end{equation}
where the phase of coupling $\lambda_{1}^\nu$ has been explicitly
written. This part of neutrino mass generation is realization of the
type-II seesaw mechanism \cite{type2}.
The full neutrino mass matrix is
\begin{equation}
M^\nu = M_0^\nu + M_1^\nu \,.
\end{equation}
Note this is the full neutrino mass matrix of the model. It is due to
tree level contribution of lepton number violation. The loop level
contribution due to R-parity violation is negligible \cite{2}, because
the sparticles in the loops are very heavy.
The physics analysis including $\lambda_1^\nu$ is different from our
previous one \cite{3}. We observe that it is natural to take that
$M_1^\nu$ is numerically dominant over $M_0^\nu$, then there appears a
degeneracy between the first two neutrinos. This roughly fits the
neutrino spectrum obtained from neutrino oscillation experiments. This
degeneracy is perturbed by $M_0^\nu$ which also contributes neutrino
mixing. Furthermore, it is interesting to note that inclusion of
$\lambda_1^\nu$ in certain cases does not really increase difficulty in
the analysis because $M_1^\nu$ is diagonal.
We rewrite $M^\nu$ by adjusting the diagonal part $M_1^\nu$ to be
proportional to identity matrix,
\begin{equation}
M^\nu = \tilde{M}_0^\nu + \tilde{M}_1^\nu \,,
\end{equation}
where
\begin{equation}
\begin{split}
\tilde{M}_0^\nu= & -\frac{a^2}{M_{\tilde{Z}}e^{i\delta_Z}}
\begin{pmatrix}
v_{l_e}v_{l_e}e^{i2\delta_{l_e}}&v_{l_e}v_{l_\mu}e^{i(\delta_{l_e}+\delta_{l_\mu})}&v_{l_e}v_{l_\tau}e^{i(\delta_{l_e}+\delta_{l_\tau})}\\
v_{l_\mu}v_{l_e}e^{i(\delta_{l_e}+\delta_{l_\mu})}&v_{l_\mu}v_{l_\mu}e^{2i\delta_{l_\mu}} &v_{l_\mu}v_{l_\tau}e^{i(\delta_{l_\tau}+\delta_{l_\mu})}\\
v_{l_\tau}v_{l_e}e^{i(\delta_{l_e}+\delta_{l_\tau})}&v_{l_\tau}v_{l_\mu}e^{i(\delta_{l_\tau}+\delta_{l_\mu})}&v_{l_\tau}v_{l_\tau}e^{2i\delta_{l_\tau}}+\Delta \lambda e^{i(\delta_{\lambda}+\delta_Z)}
\end{pmatrix},
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\tilde{M}_1^\nu= & -\frac{a^2}{M_{\tilde{Z}}}
\begin{pmatrix}
\lambda_1^\prime e^{i\delta_{\lambda_1}} & 0 & 0 \\
0 & \lambda_1^\prime e^{i\delta_{\lambda_1}} & 0 \\
0 & 0 & \lambda_1^\prime e^{i\delta_{\lambda_1}}
\end{pmatrix}\,,
\end{split}
\end{equation}
where
$\lambda_1^\prime=\displaystyle\frac{M_{\tilde{Z}}}{a^2}\frac{\lambda_1\lambda_4v^2_u}{M_T}$,
$\lambda_2^\prime=\displaystyle\frac{M_{\tilde{Z}}}{a^2}\frac{\lambda_2\lambda_4 v^2_u}{M_T}$,
and
$\Delta\lambda e^{i\delta_\lambda}=\lambda_2^\prime-\lambda^\prime_1 e^{i\delta_{\lambda_1}}$.
Generally, $M^\nu$ is complex, the phases make further analytical
calculation \cite{7} difficult. For illustration and an easy analysis,
and without losing generality about CP violation, we simply take
$\delta_{l_\alpha}=0$ and $\delta_\lambda=-\delta_Z$ in the following.
Then, up to an overall factor, $\tilde{M}^\nu_0$ is a real symmetric
matrix and can be diagonalized by an orthogonal matrix. It just needs
diagonalizing $\tilde{M}^\nu_0$, because $\tilde{M}_1^\nu$ is
essentially an unit matrix which does not affect this diagonalization.
By further assuming that $v_{l_e}^2+v_{l_\mu}^2 \ll v_{l_\tau}^2$ which
is reasonable because
$v_{l_\tau}=\displaystyle\frac{v_1+v_2+v_3}{\sqrt{3}}$ which does not
violate the $Z_3$ flavor symmetry, it is found that $\tilde{M}^\nu_0$
is diagonalized by,
\begin{equation}
O_\nu\simeq\\
\begin{pmatrix}
\displaystyle\frac{v_{l_\mu}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}&
\displaystyle\frac{v_{l_e}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}&
\displaystyle\frac{v_{l_e}v_{l_\tau}}{v_{l_\tau}^2+\Delta\lambda}\\
-\displaystyle\frac{v_{l_e}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}&
\displaystyle\frac{v_{l_\mu}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}&
\displaystyle\frac{v_{l_\mu}v_{l_\tau}}{v_{l_\tau}^2+\Delta\lambda}\\
0&-\displaystyle\frac{\sqrt{v_{l_e}^2+v_{l_\mu}^2}v_{l_\tau}}
{v_{l_\tau}^2+\Delta\lambda}&1
\end{pmatrix}
\end{equation}
with eigenvalues
\begin{equation}
\tilde{M}^{\nu{\rm ~ diag}}_0 \simeq \displaystyle
-\frac{a^2}{M_{\tilde{Z}}}e^{-i\delta_Z}
\begin{pmatrix}
0&0&0\\
0&\displaystyle(v_{l_e}^2+v_{l_\mu}^2)
\frac{\Delta\lambda}{v_{l_\tau}^2+\Delta\lambda}&0\\
0&0&\displaystyle v_{l_\tau}^2+\Delta\lambda
\end{pmatrix}\,.
\end{equation}
In fact, $O_\nu$ diagonalizes $M^\nu$,
\begin{equation}
{O_\nu}^T M^\nu O_\nu =
\tilde{M}_0^{\nu ~ {\rm diag}} + \tilde{M}_1^\nu \,.
\end{equation}
Noticing that the diagonalized matrix is still complex, we further
write that
\begin{equation}
\label{Mnu}
\begin{split}
&\tilde{M}_0^{\nu ~ {\rm diag}} +\tilde{M}_1^\nu \\[0.5cm]
=& -\frac{a^2}{M_{\tilde{Z}}}
\begin{pmatrix}
e^{i\frac{\delta_{\lambda_1}}{2}}&0&0\\
0&e^{i\frac{\beta_1}{2}}&0\\
0&0&e^{i\frac{\beta_2}{2}}
\end{pmatrix}
\begin{pmatrix}
m_{\nu_1} & 0 & 0 \\
0 & m_{\nu_2} & 0 \\
0 & 0 & m_{\nu_3}
\end{pmatrix}
\begin{pmatrix}
e^{i\frac{\delta_{\lambda_1}}{2}}&0&0\\
0&e^{i\frac{\beta_1}{2}}&0\\
0&0&e^{i\frac{\beta_2}{2}}
\end{pmatrix},
\end{split}
\end{equation}
the neutrino masses in our model are
\begin{equation}
\label{m3}
\begin{array}{lll}
m_{\nu_1} & = &
\displaystyle\frac{a^2}{M_{\tilde{Z}}}\lambda^\prime_1 \,,\\[3mm]
m_{\nu_2} & \simeq &\displaystyle\frac{a^2}{M_{\tilde{Z}}}
[\lambda_1^\prime+(v_{l_e}^2+v_{l_\mu}^2)
\displaystyle\frac{\Delta\lambda}{v_{l_\tau}^2+\Delta\lambda}
\cos(\delta_{\lambda_1}+\delta_Z)] \,,\\[3mm]
m_{\nu_3} & = & \displaystyle\frac{a^2}{M_{\tilde{Z}}}
\sqrt{{\lambda^\prime_2}^2+v_{l_\tau}^4
+2\lambda^\prime_2 v_{l_\tau}^2 \cos\delta_Z} \\,
\end{array}
\end{equation}
with the phases
\begin{equation}
\begin{split}
&\beta_1 \simeq \delta_{\lambda_1},\\
&\beta_2=\arctan\frac{v_{l_\tau}^2\sin{\delta_Z}}
{\lambda^\prime_2 + v_{l_\tau}^2\cos{\delta_Z}}\,.
\end{split}
\end{equation}
It is clear that $\nu_1$ and $\nu_2$ are almost degenerate with a mass
$\simeq\displaystyle\frac{a^2}{M_{\tilde{Z}}}\lambda_1^\prime$. Their
mass splitting is about
$\displaystyle\frac{a^2}{M_{\tilde{Z}}}(v_{l_e}^2+v_{l_\mu}^2)
\displaystyle\frac{\Delta\lambda}{v_{l_\tau}^2+\Delta\lambda}$.
$v_{l_\tau}^2+\Delta\lambda$ and $\Delta\lambda$ have the same order of
magnitude by definition, and we take
$(v^2_{l_e}+v^2_{l_\mu})\simeq\displaystyle\frac{\lambda^\prime_1}{10}$.
According to neutrino oscillation experiments \cite{pdg},
$\Delta m^2_{12} = 8.0 \times 10^{-5}$ eV$^2$ and
$|\Delta m^2_{23}|= 2.4 \times 10^{-3}$ eV$^2$,
this model typically gives that
\begin{equation}
m_{\nu_1}\simeq 2.0 \times 10^{-2} {\rm eV}\,, ~~~
m_{\nu_2}\simeq 2.2 \times 10^{-2} {\rm eV}\,, ~~~
m_{\nu_3}\simeq 5.4 \times 10^{-2} {\rm eV}\,.
\end{equation}
Naturally the phases in above formulae are $\mathcal{O}(1)$.
This makes us to take all the cosines to be $\mathcal{O}(1)$ for
simplicity in estimating the neutrino masses. And $m_{\nu_3}$ is
numerically fixed by choosing $\lambda_2^\prime$ and $v_{l_\tau}^2$.
Finally, we obtain the unitary matrix $U_\nu$ which diagonalizes
$M^\nu$,
\begin{equation}
U_\nu^T M^\nu U_\nu = -\frac{a^2}{M_{\tilde{Z}}}
\begin{pmatrix}
m_{\nu_1} & 0 & 0 \\
0 & m_{\nu_2} & 0 \\
0 & 0 & m_{\nu_3}
\end{pmatrix},
\end{equation}
\begin{equation}
U_\nu=O_\nu P^\dag
\end{equation}
with $P$ being the pure phase matrix appearing in Eq. (\ref{Mnu}).
\section{Charged lepton masses}
From Eq. (\ref{1}), the charged lepton mass matrix is obtained.
Considering the sneutrino and Higgs VEVs are complex, it is
\begin{equation}
M^l=
\begin{pmatrix}
0&\lambda_\mu v_{l_\mu}e^{i\delta_{l_\mu}}&\lambda_\tau v_{l_\mu}e^{i \delta_{l_\mu}}\\
0&\lambda_\mu v_{l_e}e^{i\delta_{l_e}}&\lambda_\tau v_{l_e}e^{i \delta_{l_e}}\\
0&0 & y_\tau v_de^{i\delta_{v_d}}
\end{pmatrix}.
\end{equation}
Here the electron mass is neglected. In this model, the electron mass
would be a loop contribution of SUSY breaking terms which also break the
flavor symmetry and the electroweak symmetry \cite{1,2}. $M^l$ in the
above equation basically fixes the mixing due to charged leptons with a
precision of $m_e/m_\mu$. It is standard to find the unitary matrix
$U_l$ which diagonalizes
$M^l {M^l}^\dag$,
\begin{equation}
U_l^\dag M^l {M^l}^\dag U_l=
\begin{pmatrix}
m_e^2 & 0 & 0 \\
0 & m_\mu^2 & 0 \\
0 & 0 & m_\tau^2
\end{pmatrix}.
\end{equation}
It can be expressed as
\begin{equation*}
U_l=P_lO_l \,,
\end{equation*}
where
\begin{equation}
P_l=\begin{pmatrix}
e^{i\delta_{l_\mu}}& 0 & 0 \\
0 & e^{i\delta_{l_e}} & 0 \\
0 & 0 & e^{i\delta_{v_d}}
\end{pmatrix},
\end{equation}
\begin{equation}
O_l\simeq\begin{pmatrix}
\displaystyle\frac{-v_{l_e}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}&
\displaystyle\frac{v_{l_\mu}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}
\displaystyle\frac{y_\tau v_d}{\sqrt{y_\tau^2 v_d^2+\lambda_\tau^2
(v_{l_e}^2+v_{l_\mu}^2)}}&
\displaystyle\frac{\lambda_\tau v_{l_\mu}}{\sqrt{y_\tau^2 v_d^2+
\lambda_\tau^2(v_{l_e}^2+v_{l_\mu}^2)}}\\
\displaystyle\frac{v_{l_\mu}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}&
\displaystyle\frac{v_{l_e}}{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}
\displaystyle\frac{y_\tau v_d}{\sqrt{y_\tau^2 v_d^2+\lambda_\tau^2
(v_{l_e}^2+v_{l_\mu}^2)}}&
\displaystyle\frac{\lambda_\tau v_{l_e}}{\sqrt{y_\tau^2 v_d^2+
\lambda_\tau^2(v_{l_e}^2+v_{l_\mu}^2)}}\\
0&\displaystyle\frac{-\lambda_\tau\sqrt{v_{l_e}^2+v_{l_\mu}^2}}
{\sqrt{y_\tau^2 v_d^2+\lambda_\tau^2(v_{l_e}^2+v_{l_\mu}^2)}}&
\displaystyle\frac{y_\tau v_d}{\sqrt{y_\tau^2 v_d^2+
\lambda_\tau^2(v_{l_e}^2+v_{l_\mu}^2)}}
\end{pmatrix}.
\end{equation}
\section{Lepton mixing matrix}
The lepton mixing matrix is $V=U_l^\dag U_\nu$. It is obtained
that $\nu_e-\nu_\mu$ mixing is
\begin{equation}
\label{mixing}
V_{e2}=\frac{v_{l_\mu}^2-v_{l_e}^2}{v_{l_e}^2+v_{l_\mu}^2}
\displaystyle e^{-i\frac{\beta_1}{2}} \,.
\end{equation}
The $\nu_\mu-\nu_\tau$ mixing is
\begin{small}
\begin{equation}
\label{mixing2}
\begin{split}
V_{\mu3}=& \displaystyle\frac{2v_{l_e}v_{l_\mu}v_{l_\tau}}
{\sqrt{v_{l_e}^2+v_{l_\mu}^2}(v_{l_\tau}^2+\Delta\lambda)}
\displaystyle\frac{y_\tau v_d}
{\sqrt{y_\tau^2 v_d^2+\lambda_\tau^2(v_{l_e}^2+v_{l_\mu}^2)}}
e^{-i\frac{\beta_2}{2}}\\
&\displaystyle -\frac{\lambda_\tau\sqrt{v_{l_e}^2+v_{l_\mu}^2}}
{\sqrt{y_\tau^2 v_d^2+\lambda_\tau^2(v_{l_e}^2+v_{l_\mu}^2)}}
e^{-i\delta_{v_d}-i\frac{\beta_2}{2}}\,.
\end{split}
\end{equation}
\end{small}
The $\nu_e-\nu_\tau$ mixing is
\begin{equation}
V_{e3}\simeq\displaystyle\frac{v_{l_\mu}^2-v_{l_e}^2}
{\sqrt{v_{l_e}^2+v_{l_\mu}^2}}
\frac{v_{l_\tau}}{v_{l_\tau}^2+\Delta\lambda}e^{-i\frac{\beta_2}{2}}\,.
\end{equation}
Experimental data for best values of these mixings are
$|V_{e2}|\simeq 0.54$, $|V_{\mu3}|\simeq 0.65$, and
$|V_{e3}|\simeq 0.15$ \cite{pdg}. Obviously, taking
$v_{l_\mu}\simeq 2v_{l_e}$, $|V_{e2}|$ is in agreement with data. The
value of $v_{l_\tau}$ is taken to be larger and still in the natural
range, $v_{l_\tau}\simeq 3v_{l_\mu}$. Choosing
$\Delta\lambda \simeq 0.3v_{l_\tau}^2$, it is easy to get
$|V_{e3}|\simeq 0.3|V_{e_2}|$.
For $|V_{\mu3}|$, there are two terms in Eq.(\ref{mixing2}), neglecting
the first term for simplicity, this mixing would be maximal if
$\lambda_\tau\sqrt{{v_{l_\mu}}^2+{v_{l_e}}^2}=y_\tau v_d$, namely
$\lambda_\tau\simeq 0.8$. Of course, a smaller $\lambda_\tau$ is more
natural. Therefore this model slightly favors the atmospheric neutrino
angle to be in the first octant.
The important CP violation in neutrino oscillations is given through
the invariant parameter $J$ \cite{j},
\begin{equation}
\label{jar}
\Im(V_{i\lambda}V_{j\rho}V_{i\rho}^*V_{j\lambda}^*)=
J\sum_{\kappa,\delta}\epsilon_{ijk}\epsilon_{\lambda\rho\delta},
\end{equation}
and
\begin{small}
\begin{equation}
\begin{split}
J \simeq
&\displaystyle\frac{2v_{l_e}v_{l_\mu}v_{l_\tau}(v_{l_e}^2-v_{l_\mu}^2)^2
\lambda_\tau y_\tau v_d}{(y_\tau^2 v_d^2+\lambda_\tau^2(v_{l_e}^2
+v_{l_\mu}^2))(v_{l_e}^2+v_{l_\mu}^2)^2(v_{l_\tau}^2+\Delta\lambda)}
\sin\delta \simeq 0.04 \sin\delta,\\[3mm]
\delta=&-\delta_{v_d} \,.
\end{split}
\end{equation}
\end{small}
$\delta_{v_d}$ is expected to be large, namely $|\sin\delta|\sim 0.1-1$.
This agrees with current preliminary experimental results
\cite{t2k-nova}.
\section{Majorana neutrino mass}
The effective Majorana mass in the neutrinoless double beta decay is
\begin{small}
\begin{equation}
\langle m\rangle_{ee}=|m_{\nu_1}{V_{e1}}^2+m_{\nu_2}{V_{e2}}^2+m_{\nu_3}{V_{e3}}^2|\,.
\end{equation}
\end{small}
In this work, it is
\begin{equation}
\langle m\rangle_{ee} =
\left|m_{\nu_1}|V_{e1}|^2 + m_{\nu_2}|V_{e2}|^2 +
\displaystyle m_{\nu_3}|V_{e3}|^2 e^{i(\delta_{\lambda_1}-\beta_2)}
\right|
\simeq 0.02 ~{\rm eV}\,.
\end{equation}
In the above formula, the $V_{e3}$ term has a Majorana phase
dependence, which is negligibly small anyway.
\section{Discussions}
Like gauge theories which are used to describe the elementary particle
interactions, SUSY is used for fermion masses. Our model is the minimal
SUSY SM with a vector-like triplet field extension, but SUSY breaks at a
high scale and the R-parity (lepton number) is not required. The
sneutrino VEVs result in a neutrino mass which is suppressed by the
Zino mass. This is a nice realization of the type-I seesaw mechanism
which, even does not need to introduce any right-handed neutrino. The
triplet field is originally for the realistic Higgs mass. However, it
also contributes to neutrino masses through a type-II seesaw mechanism.
The Zino related seesaw mechanism results in only one massive neutrino.
By including the triplet contribution, the neutrino masses can be
realistic. Compared to our previous studies \cite{2,3}, a more natural
pattern for neutrino masses is obtained.
To be numerically natural, let us return back to the original
superpotential in the beginning. The couplings are assumed to be taken
natural values. The field VEVs are mainly fixed by the soft parameters
in the Lagrangian, in addition to those in the superpotential. To fit
the lepton spectrum and mixing, we take $v_{l_e}\simeq 1$ GeV,
$v_{l_\mu}\simeq 2$ GeV, $v_{l_\tau}\simeq 6$ GeV, $v_d\simeq 10$ GeV,
and $v_u\simeq 228$ GeV. Note $v_{l_\tau}$ does not break the flavor
symmetry, it is natural that its value is more close to $v_d$. And the
large $v_u/v_d$ ratio is for explaining the top quark mass \cite{2}.
When $\lambda_1^\prime$, $\lambda_2^\prime$ and $v_{l_\tau}^2$ are in
the same order, the correct neutrino spectrum is obtained. In terms of
parameters in the superpotential, we have
$M_{\tilde{Z}}\simeq 3\times 10^{11}$ GeV.
$M_T \simeq (1-10) M_{\tilde{Z}}$, and $\lambda$'s $\simeq (0.01-0.1)$.
It is necessary to check the reliability of our approximation in
estimating the neutrino masses. That approximation about the phases
can be good when the quantities appear in the mass formulae are
hierarchical, say if $\lambda^\prime_1 \gg v_{l_e}^2+v_{l_\mu}^2$. As
it has been seen that this is indeed the case for $m_{\nu_2}$. In
$m_{\nu_3}$ (Eq.(\ref{m3})), $\lambda^\prime_2$, $\lambda^\prime_1$ and
$v_{l_\tau}^2$ are of the same order. This allows us to look at an
extreme case where the phase is $\pi$. In this case, there is a
possibility of inverted neutrino mass hierarchy, namely a very small
$m_{\nu_3}$. But this is achieved through a large cancellation between
$\lambda^\prime_2$ and $v_{l_\tau}^2$. Although this is possible, it
is unnatural.
The physics of neutrinos in this work is quite different from that in
Refs. \cite{2,3}. This is mainly due to the triplet. In Ref. \cite{2},
we introduced a singlet, the neutrino mass matrix $M_1^\nu$ was that
with only the $33$ matrix element nonvanishing. And in \cite{3}, the
triplet replaced the singlet for the Higgs mass in the beginning,
however, in the neutrino mass analysis, we took $\lambda_1^\nu$ to be
zero which essentially was the same as that for the singlet case.
Taking $\lambda_1^\nu$ to be zero was actually unreasonable because our
principle is to treat all the basic couplings close to $0.01-1$. As a
result, in Refs. \cite{2,3}, there was always one massless neutrino.
That led to that the Majorana mass $\langle m\rangle_{ee}$ is about
$10^{-3}$ eV. In addition, in \cite{3} it was wrong to say CP violation
is small in the lepton sector.
\section{Summary}
In summary, in the model of high scale SUSY for understanding the
fermion mass hierarchies, we have studied CP violation in the lepton
sector, and other aspects of neutrino physics in detail. In the
analysis, the phases of the Higgs and sneutrino VEVs, and contribution
of the $\lambda_1^\nu$ term in superpotential (\ref{2}), have been
included. This analysis is more complete than previous consideration.
The neutrino mass matrix, and the charged lepton one, are fixed by the
model. Its specific feature is the triplet contribution, the approximate
degeneracy of neutrinos $\nu_1$ and $\nu_2$ can be naturally explained.
This model could not predict exact values of the fermion masses because
of the flavor symmetry breaking as well as SUSY breaking. However, the
principle we follow is that all the coupling constants should be in the
natural parameter range which is about $(0.01-1)$. Taking triplet
contribution dominant, and inputting relevant
experimental data on leptons, we obtain that
(i) $m_{\nu_1} \simeq 0.020$ eV, $m_{\nu_2} \simeq 0.022$ eV,
$m_{\nu_3} \simeq 0.054$ eV. This normal ordering neutrino spectrum is
to be checked in JUNO experiment \cite{juno}.
(ii) CP violation in neutrino oscillation most probably is large. There
have been some experimental hint on this \cite{t2k-nova}. CP violation
in neutrino oscillations is a great study task experimentally
\cite{future-cp}.
(iii) The effective Majorana neutrino mass in the neutrinoless double
beta decay is about $0.02$ eV, it is within the detection ability of
future measurements \cite{majorana}.
(iv) $\theta_{23}$ is slightly favored being in the first octant.
(v) The electron neutrino mass to be measured in $\beta$ decays is
about $0.02$ eV. This is, however, still one order of magnitude lower
than the future limit of direct measurements \cite{katrin}.
(vi) The sum of three neutrino masses is close to
$\sum{m_\nu}\simeq 0.1$ eV. If the standard cosmology is correct,
astrophysics measurements on the cosmic microwave background has
constrained this sum to be $< 0.15$ eV \cite{planck}. It is interesting
to note that a recent analysis showed the sum is about $\sim 0.11$ eV
\cite{sum}.
Most of the above predictions are close to their experimental limits,
therefore, this model will soon be checked experimentally.
\begin{acknowledgments}
We would like to thank Gui-Jun Ding and Zhen-hua Zhao for very helpful
discussions. This work was supported in part by the National Natural
Science Foundation of China (No. 11375248 and 11875306).
\end{acknowledgments}
\newpage
|
\section{Introduction}
Gas-rich circumstellar disks around young stars provide a window to study the materials that are incorporated into forming planetary systems. Chemistry and dynamics in the disk can shift the balance of gas- versus ice-phase volatiles containing, e.g., carbon, nitrogen, and oxygen. In the core accretion paradigm \citep{pollack1996}, those materials that end up as rocks or ices are incorporated into the ``solid'' planetesimals, while the remaining gas can be accreted into natal planets' atmospheres. Correspondingly, it is essential to understand the form volatiles containing carbon, nitrogen, and oxygen take in the disk spatially and over time, to understand the initial elemental compositions of forming planets \citep[e.g.,][]{oberg2011co,piso2016,cridland2016,cridland2017}.
Many processes can alter the gas versus solid abundances in the disk, including snow lines \citep[e.g.,][]{oberg2011co}, chemistry \citep{bergin2014far,furuya2014,eistrup2016,schwarz2018}, mixing and/or diffusion of gas \citep[e.g.,][]{semenov2011,kama2016}, and redistribution of ices as dust grows and evolves \citep[e.g.,][]{hogerheijde2011,piso2016,krijt2016,oberg2016}. To date, observations of various carbon and oxygen carriers have suggested a substantial ``missing'' volatile mass within the disk molecular layer as traced by CO in the submillimeter \citep{favre2013,cleeves2015tw,zhang2017} and H$_2$O vapor in the far-infrared with {\em Herschel} \citep{bergin2010,hogerheijde2011,du2017}, though see also \citet{kamp2013} regarding model dependencies. For the few disks where estimates for both exist, more water in the disk surface is ``missing'' than CO when compared to interstellar abundances, and both have abundances lower than what simple desorption (thermal and non-thermal) models predict \citep[e.g., in TW Hya's disk;][]{hogerheijde2011,favre2013,schwarz2016,kama2016,cleeves2015tw}.
Absolute elemental abundances are often difficult to estimate due to uncertain hydrogen disk masses. In this context, relative elemental abundances, such as C/O and N/O are promising avenues toward robustly characterizing disk volatile compositions.
Recently, \citet{bergin2016} reported bright hydrocarbon rings of C$_2$H and $c-$C$_3$H$_2$ in the TW Hya \citep[see also][]{kastner2015} and DM Tau disks. Using chemical models, \citet{bergin2016} found the abundances of these small hydrocarbons were especially sensitive to the gas-phase C/O ratio of the disk. The observations required high C/O values, $>1$, to reproduce the observed line intensities \citep[see also][]{du2015}.
Such prospects for measuring C/O in disks are now especially exciting as we enter an era where the elemental compositions, including C/O, of exoplanets' atmospheres \citep[e.g.,][]{madhusudhan2011,kreidberg2014,macintosh2015,bonnefoy2016,lavie2017}.
In contrast to carbon and oxygen abundance estimates, there are few constraints on total nitrogen abundances or N/O ratios in disks owing in large part to the difficulty of observing the likely primary nitrogen carrier, N$_2$, \citep[e.g.,][and references therein]{schwarz2016}. Abundant nitrogen bearing species, such as N$_2$H$^+$ and HCN, are sensitive to other disk parameters than total N abundance, such as temperature structure, carbon abundance, and ionization rate. As such, interpreting these species in the context of bulk nitrogen abundance requires detailed knowledge of the source.
In this work, we constrain the carbon, nitrogen, and oxygen content of the warm molecular layer in the IM Lup protoplanetary disk using results from our previous study of CO and its isotopologues \citep{cleeves2016im} and new and archival ALMA observations of C$_2$H and HCN and H$^{13}$CN. The solar mass star IM Lup harbors a massive gas rich disk, $M_{\rm disk} \sim0.1-0.2$~M$_{\odot}$ based upon continuum (SED and resolved millimeter images) and CO multi-isotopologue multi-line data presented in \citep{cleeves2016im}. The source is relatively young at an age of $0.5-1$~Myr \citep{mawet2012}.
In \citet{cleeves2016im}, we found that IM Lup's CO is under-abundant by a factor of $\sim20$ compared to an interstellar CO abundance of $1.4\times10^{-4}$ per H based upon the dust-inferred disk mass. For comparison, this younger object appears to be ``missing'' less CO than the older TW Hya system, whose CO abundance is $\sim3-5\times$ less abundant than IM Lup \citep[e.g.,][]{favre2013}. However, based on this data alone it is difficult to tell whether the observed ``missing'' gas-phase CO is a result of missing carbon or missing oxygen or both; or, alternatively, missing total gas mass compared to dust. In the present paper, we explore what C/O and N/O abundance ratios are required in the disk's warm molecular layer to reproduce the observed C$_2$H, HCN, and HC$^{13}$N line intensities.
\section{Observations}\label{sec:obs}
\begin{figure*}[]
\begin{centering}
\includegraphics[width=1.0\textwidth]{imlup_c2h_avg_v4.pdf}
\caption{Channel maps of C$_2$H $N=3-2$ for the a) $J=\sfrac{5}{2} - \sfrac{3}{2}$ pair and b) $J=\sfrac{7}{2} - \sfrac{5}{2}$ pair, channel averaged by a factor of two for clarity. THe solid contour line indicates 3$\sigma$. The beam is in the lower left corner. The VLSR in km~s$^{-1}$ is indicated in the bottom right corner of each panel. The purple ellipse on the right panels indicates the scale of the millimeter thermal dust emission, $R=313$ AU \citep{cleeves2016im}. \label{fig:allchannel}}
\end{centering}
\end{figure*}
\subsection{ALMA Observations and Data Reduction}\label{sec:data}
\begin{figure}[]
\begin{centering}
\includegraphics[width=0.34\textwidth]{c2h_spectrum.pdf}
\caption{C$_2$H $N=3-2$ spectra extracted from a circular 4'' mask. The dotted line indicates the line centers for each pair in the rest frame of the $F=2-1$ (top) and $F=3-2$ (bottom) transitions. \label{fig:spec}}
\end{centering}
\end{figure}
IM Lup was observed as part of a survey of nitrogen isotope chemistry and deuterium chemistry presented in \citet{guzman2017} and \citet{huang2017}, respectively. Despite this source being one of the most gas-rich disks and bright in HCN $J=3-2$ \citep[see also][]{oberg2011discsii}, the H$^{13}$CN $J=3-2$ line was not detected at a per channel RMS of 3.6 mJy per beam in 0.5 km s$^{-1}$ channels \citep{huang2017}. The observations of
HCN $J=3-2$ and H$^{13}$CN $J=3-2$ are presented in \citet{huang2017} \citep[see also][]{guzman2017} and are not reproduced here.
The C$_2$H $N=3-2$ hyperfine complex was observed with ALMA as part of the Cycle 3 2015.1.00964.S program (PI: {\"O}berg) on 01 May 2016 with 41 antennae for 12~minutes on source.
The observations were calibrated by ALMA/NAASC staff using J1517-2422 for the bandpass, J1610-3958 for the phase and amplitude, and Titan for the flux calibration. We performed one additional round of phase self-calibration using the continuum within the spectral window containing the line. For the self-calibration, we use a solution interval of 30 seconds and average both polarizations. The continuum is estimated from line-free channels within the same spectral window, which is then subtracted in the $uv$-plane to obtain continuum-subtracted images. We detect two blended hyperfine pairs of C$_2$H $N=3-2$, the $J=\sfrac{5}{2} - \sfrac{3}{2}$ $F=2-1$ and $F=3-2$ pair and the $J=\sfrac{7}{2} - \sfrac{5}{2}$ $F=3-2$ and $F=4-3$ pair, see Table~\ref{tab:obs}. While the pairs are blended in velocity space, the inclined orientation of the Keplerian disk causes the emission from each of the pairs to be mostly spatially resolved on the sky. Images are generated using CASA 4.7 \citep{mcmullin2007} and the clean task. Figure~\ref{fig:allchannel} shows velocity-averaged channel maps and moment-0 maps tapered to 1'' to improve signal to noise. The moment-0 maps show the integrated combined emission from each pair after clipping the individual channels below 1$\sigma$. The disk-integrated spectrum of the observed C$_2$H transitions is shown in Figure~\ref{fig:spec} for each of the hyperfine pairs integrated within a 4'' radius circular aperture. Table~\ref{tab:obs} provides the C$_2$H $N=3-2$ disk integrated fluxes within this aperture, with errors estimated from a 4'' circular aperture in line-free portions of the spectrum combined with 10\% calibration uncertainty added in quadrature.
\begin{deluxetable}{ccc}
\tablecaption{C$_2$H $N=3-2$ Line Observations \label{tab:obs}}
\tabletypesize{\footnotesize}
\tablehead{{Transition } & Rest & Disk-Integrated \\
&Freq.& Flux Density (Blended) \\
& [GHz] & [Jy km s$^{-1}$] }
\startdata
$J=\sfrac{5}{2} - \sfrac{3}{2}$ $F=2-1$ & 262.067 & $0.56 \pm 0.08$ \\
$J=\sfrac{5}{2} - \sfrac{3}{2}$ $F=3-2 $& 262.065 & \\
$J=\sfrac{7}{2} - \sfrac{5}{2}$ $F=3-2$ & 262.006 & $0.76\pm 0.11$ \\
$J=\sfrac{7}{2} - \sfrac{5}{2}$ $F=4-3$ & 262.004 &
\enddata
\tablecomments{Uncertainties are quadrature combined RMS scatter and 10\% calibration uncertainty.}
\end{deluxetable}
\section{Methods}\label{sec:methods}
The following sections describe our methods to constrain carbon, oxygen, and nitrogen abundances in IM Lup's disk using chemical models to find what elemental abundances best reproduce the C$_2$H, HCN, and H$^{13}$CN observations, while also remaining consistent with the previous CO constraints \citep{cleeves2016im}.
\subsection{Disk Physical Model}\label{sec:physmod}
We use the underlying disk structure derived in \citet{cleeves2016im}, which we summarize here. The disk surface density follows the self-similarity solutions of \citet{lyndenbell1974}, with an inner power law and outer exponential taper in the disk surface density. The disk is vertically extended and flared, with a gas scale height of 12 AU at 100 AU. The millimeter grains form a thinner layer (25\% of the gas scale height) and contain 99\% of the total dust mass, $M_{\rm dust} = 0.0017$ M$_\odot$. We furthermore assume small grains follow the gas distribution and allow the millimeter grains to have a separate power law distribution with a cut off at the millimeter emission edge at $313$ AU. The small grains provide the greatest surface area per unit mass, and also fill most of the disk volume, and thus are the most important for the grain surface chemistry in the observable layers of the disk ($\gtrsim1$ scale height). In the model, we track the total surface area per unit volume throughout the disk as an input to the chemical calculations.
The gas and dust temperature structures are fixed, where the dust temperature is calculated assuming radiative equilibrium with a stellar $T_{\rm eff}=3900$~K and $R_*=2.5~R_\odot$ \citep{pinte2008,cleeves2016im}. We note \citet{alcala2017} provides updated stellar parameters for IM Lup; however, to within the uncertainties the new values are similar, and so we have chosen to keep these values fixed to more readily compare to previous work.
The gas temperature deviates from the dust temperature in the disk upper layers where there is both FUV heating from the central star and the external radiation field, where we fix the external radiation field to be $G_0=4$ Habing \citep{haworth2017,pinte2018,cleeves2016im}.
For the stellar high energy radiation field, we use the UV spectral template of TW Hya normalized to the observed {\it Swift} UVM2 flux density from IM Lup, and the ``quiescent'' X-ray template presented in \citet{cleeves2013a} normalized to the observed integrated X-ray luminosity provided in \citet{gunther2010} of $4.3\times10^{30}$~erg~s$^{-1}$. The wavelength dependent radiation transport is calculated for both UV and X-rays using the code of \citet{bethell2011u}.
\subsection{Chemical Modeling Procedure}\label{sec:chemmod}
We calculate the 2D chemical abundances of C$_2$H and HCN using a time-evolving gas-grain model first presented in \citet{fogel2011} and updated and expanded in \citet{cleeves2014par}. The chemical model takes into account the spatial changes in dust surface area per unit volume as described in Section~\ref{sec:modupd} and Appendix~\ref{app:grains}. The simulations are run over 0.5~Myr, corresponding to the lower age limit of IM Lup. However, we reach steady state in the warm molecular layer (the region probed by the observations) well before this time and the results are not affected by this choice.
Therefore, we find this assumption does not significantly impact our results. We use a non-deuterated chemical network with 5974 reactions and 638 species. We do not take into account carbon isotope chemistry. For the H$^{13}$CN $J=3-2$ line radiation transfer, we simply take the model HCN abundance and assume an isotope ratio of $\rm ^{12}C/^{13}C=70$ \citep{prantzos1996}.
The two main variables considered in the modeling are:
\begin{enumerate}
\item the initial C, N, and O abundances (Section~\ref{sec:modabun})
\item the cosmic ray ionization rate, $\zeta_{\rm CR}$.
\end{enumerate}
Based on theory \citep{cleeves2013a} and observations of the TW Hya disk \citep{cleeves2015tw}, the typical cosmic ray ionization in disks may be low, $\zeta_{\rm CR}\le2\times10^{-19}$ s$^{-1}$ per H$_2$. These low rates may be related to magnetized wind modulation or deflection by tangled magnetic fields within the disk \citep{cleeves2013a}. We consider models that have a cosmic ray ionization rate similar to TW Hya and values approximately one order of magnitude higher and one order of magnitude lower. These models correspond to the ``T Tauri minimum modulation'' (ttm; $\zeta_{\rm CR}=1.0\times10^{-20}$ s$^{-1}$ per H$_2$), ``Solar System maximum'' (ssx; $\zeta_{\rm CR}=2.0\times10^{-19}$ s$^{-1}$ per H$_2$), and the ``Solar System minimum'' (ssm; $\zeta_{\rm CR}=1.3\times10^{-18}$ s$^{-1}$ per H$_2$) from \citet{cleeves2013a}.
\subsubsection{Model updates}\label{sec:modupd}
For the present study, the chemical code has been updated in three ways. First, we have improved the grain surface chemistry to provide more ``realistic'' reaction rates. In the past, we have allowed all ice at a given time to participate in grain-surface chemistry, since our code does not treat the multi-layered nature of the ice like the models of, for example, \citet{hasegawa1993}, \citet{vasyunin2013}, \citet{garrod2013}, and \citet{furuya2016}. As such, without taking layered ices into account, we had implicitly enhanced the efficiency of grain surface chemistry, since more realistically we expect primarily the ice surface to be reactive. To approximate this behavior, we multiply the reaction rates by (number of monolayers)$^{-1}$,
except for atomic H, which we expect to exist mainly as part of the reactive surface. Incorporating this effect is an especially important addition given that most disks appear to have a large amount of settled dust mass, i.e., have a deficit of small grains in their surface, including IM Lup. Essentially, at a fixed volatile abundance, the ice-coating on a given grain can become ``thicker'' with decreasing total grain-surface area per volume, making less of the ice mobile and reactive.
The second chemical code update is to include a temperature-dependent sticking coefficient as described in \citet{he2016}, rather than assuming perfect sticking for all species. We adopt the generic fit from \citet{he2016} Table 1, except for gas-phase water, where we still assume perfect sticking due to its highly polar nature.
The third model update is the incorporation of N$_2$ self-shielding using the \citet{li2013} shielding functions and the vertically calculated H$_2$ and N$_2$ column densities. Without the latter, we over-predict the atomic N density and under-predict the N$_2$ density.
\subsubsection{Model abundances}\label{sec:modabun}
\begin{deluxetable}{llll}[bh!]
\tablecolumns{4}
\tablewidth{0pt}
\tablecaption{Fixed chemical abundances besides CO/C$^+$/H$_2$O ice relative to total H atoms. \label{tab:other}}
\tabletypesize{\footnotesize}
\tablehead{{Molecule} & Abundance & {Molecule} & Abundance }
\startdata
H$_2$ & $5.00\times10^{-1}$ & He & $1.40\times10^{-1}$ \\
N$_2$ & $3.75\times10^{-5}$ & CS & $4.00\times10^{-9}$ \\
SO & $5.00\times10^{-9}$ & HCO$^+$ & $9.00\times10^{-9}$ \\
H$_3^+$ & $1.00\times10^{-8}$ & C$_2$H & $8.00\times10^{-9}$ \\
Si$^+$ & $1.00\times10^{-9}$ & Mg$^+$ & $1.00\times10^{-9}$ \\
Fe$^+$ & $1.00\times10^{-9}$ &
\enddata
\tablecomments{Note: N$_2$ abundance varied in Section~\ref{sec:hcnfit}.}
\end{deluxetable}
The primary goal of this work is to constrain the total amount of volatile carbon, nitrogen, and oxygen in the upper layers of the disk, which may not have solar or interstellar bulk composition. Furthermore, we wish to remain agnostic to the specific process leading to deviations in the bulk abundances from interstellar values. To accomplish this, we take the simple approach of adjusting the initial chemical abundances in our models to explore a range of possible total C/H, O/H, and N/H abundances to jointly reproduce the ALMA observations of C$_2$H and HCN within our \citet{cleeves2016im} model framework.
We have examined the models to ensure that they reach a pseudo-steady state for the species of interest in the warm molecular layer between radii of 20 and 300 AU within a relatively short timescale ($\sim10^3$ years, or 0.2\% of the simulation time). Essentially, the chemistry in this layer quickly re-adjusts to the local conditions and is most sensitive to the bulk C, N, and O content rather than the details of the initial abundance profile. This feature of the warm molecular layer chemistry allows us to constrain the abundances without needing the full (and uncertain) chemical and physical history of the gas. To confirm this behavior, we have tested models where we move the carbon, oxygen, and nitrogen into different carriers, e.g., N$_2$ versus N versus NH$_3$ and achieve similar output abundances in the warm molecular layer ($z/r\gtrsim0.2$) to within a few percent.
The abundances of the species that are not varied are listed in Table~\ref{tab:other}, and are motivated by molecular cloud model abundances \citep{fogel2011,cleeves2016im}. Species whose abundances are altered in the initial conditions are in Table~\ref{tab:ini}. The baseline water abundance has been updated to reflect the high-end of interstellar water ice measurements, $\chi({\rm H_2O_{ice}}) = 8\times10^{-5}$ per H \citep{boogert2015}, and thus relative depletion factors in water reported here may be higher if the intrinsic interstellar water content is higher, e.g., hidden in larger interstellar grains \citep[i.e.][]{vandishoeck2014}.
The starting point of our depletion models is our 2016 paper which revealed CO to be under-abundant by a factor of 19 relative to an interstellar CO abundance of $1.3\times10^{-4}$ per H \citep{ripple2013}, updated from previously $1.4\times10^{-4}$ per H. For each abundance mixture, we require there to be sufficient elemental C and O to be able to produce a CO abundance of $7\times10^{-6}$, but not excess. For example, the carbon abundance can be $7\times10^{-6}$ per H, and the oxygen abundance can be this value or greater (carbon-limited models). Similarly, oxygen can have the same $7\times10^{-6}$ per H abundance, but with equal or excess carbon (oxygen-limited models). For the nitrogen depletion factors, our primary carrier is N$_2$, and to simulate nitrogen ``removal'' we directly reduce the initial N$_2$ abundance.
\begin{deluxetable}{cccc}[b!]
\tablecolumns{4}
\tablewidth{0pt}
\tablecaption{Abundances of key C and O species. \label{tab:ini}}
\tabletypesize{\footnotesize}
\tablehead{ C/O & C$^+$ & CO & H$_2$O(gr)}
\startdata
0.08& 0.0 & $7\times10^{-6}$ & $8\times10^{-5}$ \\
0.47& 0.0 & $7\times10^{-6}$ & $8\times10^{-6}$ \\
0.64& 0.0 & $7\times10^{-6}$ & $4\times10^{-6}$ \\
0.81& 0.0 & $7\times10^{-6}$ & $1.6\times10^{-6}$ \\
0.9& 0.0 & $7\times10^{-6}$ & $8\times10^{-7}$ \\
0.95& 0.0 & $7\times10^{-6}$ & $4\times10^{-7}$ \\
1.0& $3.5\times10^{-6}$ & $3.5\times10^{-6}$ & $3.5\times10^{-6}$ \\
1.86& $9.5\times10^{-6}$ & $3.5\times10^{-6}$ & $3.5\times10^{-6}$ \\
3.71& $2.25\times10^{-5}$ & $3.5\times10^{-6}$ & $3.5\times10^{-6}$
\enddata
\tablecomments{Models with C/O $\le 1$ are carbon limited, while models with C/O $\ge1$ are oxygen limited. In the latter case, the oxygen is split between CO and H$_2$O(gr).}
\end{deluxetable}
Even though the chemical reprocessing timescales are short in the upper layers ($z/r\ge0.25$) of the disk, we have nonetheless attempted to create ``realistic'' initial compositions rather than purely atomic. As seen in Table~\ref{tab:ini}, the oxygen and carbon are primarily in H$_2$O, CO and C$^+$ when needed. This choice does not affect the main goals of this study (reproducing the observables), but results in more realistic midplane chemical abundances.
\subsection{Model - Observation Comparison}\label{sec:linemod}
We simulate the C$_2$H $N=3-2$, HCN $J=3-2$, and H$^{13}$CN $J=3-2$ observations using the non-LTE radiative transfer code LIME v1.8\footnote{\url{https://github.com/lime-rt/lime}} \citep{brinch2010}. We use the collisional rates input files provided by the Leiden LAMDA database \citep{schoier2005}. All calculations take into account full non-LTE radiation transfer. The HCN and H$^{13}$CN collision rate data is assumed to be the same and sourced from \citet{green1974} and does not include hyperfine structure. The C$_2$H collision rate data is from \citet{spielfiedel2012}. The non-LTE analysis is especially important given that the C$_2$H is abundant in our models across a wide range of densities, down to $n_{\rm H}\sim10^5$ cm$^{-3}$. We simulate each of the C$_2$H $J=\sfrac{5}{2} - \sfrac{3}{2}$ and $J=\sfrac{7}{2} - \sfrac{5}{2}$ line pairs together since their emission covers the same frequency space, even if the emission is not blended spatially. While the CO gas extends out as far as $\sim1000$~AU, the HCN and C$_2$H and HCN extend to about $\sim500$ AU, and so we limit our emission models to this radius to only constrain the abundances where C$_2$H is well-detected, and discuss the implications of this in Section~\ref{sec:cono}.
The line and millimeter continuum data are modeled in LIME simultaneously as the millimeter continuum opacity is known to be high in this source, especially in the inner disk \citep{cleeves2016im}. The continuum at these wavelengths is is entirely midplane dominated, in a thin vertical layer. We have not attempted to adjust the inner disk opacity as in \citet{cleeves2016im}, and note that the main effect would be to reduce the observable line flux from the inner $R\lesssim40$~AU, where the C$_2$H indeed shows an inner depression (see Figure~\ref{fig:allchannel}).
The input gas velocities include Keplerian rotation around a solar mass star, thermal broadening, and a fixed turbulent velocity of 100~m~s$^{-1}$. The final spectral resolution of the C$_2$H, HCN, and H$^{13}$CN simulations is 0.28, 0.28, and 0.35 km~s$^{-1}$, but we simulate the cubes at $40\times$ higher spectral resolution than observed and average down to take into account blurring due to channel averaging. The simulations assume a fixed distance of 161~pc \citep[][DR1]{gaia}, and a fixed inclination of the disk midplane of $48^\circ$ \citep{cleeves2016im}. The updated DR2 distance is $158\pm3$ pc, which is sufficiently consistent with the DR1 value, and so we have kept the DR1 distance for ease of comparison.
The models are compared to the ALMA data in the visibility plane, where the \texttt{vis\_sample} package \citep{loomis2018}\footnote{\url{https://github.com/AstroChem/vis_sample}} is used to sample the LIME channel maps at the same spatial frequencies as were observed to create the model visibilities. To assess the goodness-of-fit for models, we compare the $\chi^2$ between the observed and simulated continuum subtracted visibilities. The total model grid size is nine values of C/O and three values of $\zeta_{\rm CR}$ for 27 models total.
\section{Results}\label{sec:results}
\subsection{Chemical Model Results}\label{chemresults}
Figure~\ref{fig:2dabun} presents the 2D chemical abundances for a selection of C/O ratios and the intermediate CR rate value.
It is clear that the C$_2$H is strongly sensitive to the C/O ratio in the gas, confirming early results of \citet{bergin2016}. The C$_2$H abundance is most sensitive for our models below C/O of 1.86, where there is a sharp column density transition straddling C/O of $\sim1$, clearly seen in the two orders of magnitude jump in C$_2$H column densities going from C/O of 1.86 to 0.95 in Figure~\ref{fig:column}. C$_2$H is essentially unaffected by the cosmic ray ionization rate for the three model values considered. This lack of dependence occurs because C$_2$H is abundant in a layer wrapping around the disk that is UV dominated rather than cosmic ray dominated \citep[see Figure~\ref{fig:2dabun} and][]{bergin2016}.
\begin{figure*}
\begin{centering}
\includegraphics[width=1\textwidth]{model2Dab_net_v5.pdf}
\caption{C$_2$H (top) and HCN (bottom) 2D abundances for different C/O ratios as labeled at the top of each column at an intermediate cosmic ray ionization rate of $\sim10^{-19}$~s$^{-1}$ per H$_2$. The dominant effect on the abundances of both C$_2$H and HCN is the C/O ratio rather than the cosmic ray ionization rate. \label{fig:2dabun}}
\end{centering}
\end{figure*}
\begin{figure*}
\begin{centering}
\includegraphics[width=1\textwidth]{model_col_v3.pdf}
\caption{C$_2$H (top) and HCN (bottom) chemical model column densities versus radius for different cosmic ray rates as indicated above each column, and at different C/O ratios as indicated by line color on the righthand side. Note the general monotonic trend of decreasing C$_2$H and HCN column density with decreasing C/O. \label{fig:column}}
\end{centering}
\end{figure*}
We can also see that the radial morphology of the column densities changes substantially with C/O (Figure~\ref{fig:column}). For high values $\ge1$, the C$_2$H column density forms a narrow ring or is centrally peaked. At low C/O ratios, the C$_2$H column density forms a wide ring that peaks outside of the millimeter dust disk ($R_{\rm mm}=313$~AU). This transition occurs once the layer of C$_2$H interior to $R\lesssim300$ AU becomes thin, with little column density compared to the outer disk.
Correspondingly, the radial morphology of C$_2$H can change from peaked to ringed even with a uniform C/O ratio, not necessarily requiring (but also not excluding) radial variations in C/O \citep{bergin2016}.
HCN is sensitive to both the C/O ratio in the gas and the cosmic ray ionization rate, especially for models with C/O $\ge 1$. Higher values of C/O and $\zeta_{\rm CR}$ generally increase the HCN abundance. At a given radius, similar column densities are achieved in models with high CR rates and low C/O, and in models with low CR rates and high C/O, though the overall effect on the radial column density profile is different (Figure~\ref{fig:column}).
The 2D abundance morphologies are similar between HCN and C$_2$H, though we find the HCN abundance is generally less than that of C$_2$H for a given C/O value. One morphological difference is that HCN extends vertically deeper, with a base of $z\sim60$ AU at $R=300$ AU compared to C$_2$H, which disappears below $z\sim80$~AU for C/O $< 1$.
\subsection{Fits to {\rm C$_2$H} Observations}\label{c2hfit}
Based upon the chemical modeling results in Section~\ref{chemresults}, we can use the C$_2$H observations to constrain the C/O ratio in the IM Lup disk's warm molecular layer using the grid of C/O models described in Table~\ref{tab:ini} and illustrated in Figures \ref{fig:2dabun} and \ref{fig:column}. The data and models are compared in the visibility plane with the procedure described in Section~\ref{sec:linemod}.
\begin{figure*}
\begin{centering}
\includegraphics[width=1\textwidth]{init_grid_nozoom_all2.pdf}
\caption{$\chi^2$ comparison between the observed visibilities and models for each of the lines or pairs (see top of each panel). Blended hyperfine components for the C$_2$H lines are modeled simultaneously. From left to right, $\chi^2_{\rm min}$ is $4.58\times10^{6}$, $4.39\times10^{6}$, $6.43\times10^{6}$, and $6.70\times10^{6}$, so in all cases the reduced $\chi^2$ is close to 1, with the minimum being the closet model-data fit. For H$^{13}$CN the models are consistent with non-detection in the image plane for C/O $<1$. Note: the C/O variations reflect the total elemental ratio in the simulations (gas and ice), see Section~\ref{sec:cono} for gas-phase C/O values. These values should be interpreted as surface ($z/r\gtrsim0.2$) constraints, i.e., the region of abundant C$_2$H (Section~\ref{chemresults}). \label{fig:chisq}}
\end{centering}
\end{figure*}
Figure~\ref{fig:chisq} shows the $(\chi^2$ - minimum $\chi^2 + 1)$ between the data visibilities and our model visibilities varying initial gas composition (C/O) and $\zeta_{\rm CR}$, where smaller values indicate better fits. Note, adding 1 allows us to plot the difference on log scale. For the C$_2$H $N=3-2$ $J=\sfrac{5}{2} - \sfrac{3}{2}$ pair of lines a C/O ratio of $\sim0.81$ is favored, while for the C$_2$H $N=3-2$ $J=\sfrac{7}{2} - \sfrac{5}{2}$ pair, C/O of $\sim0.7$ is the best match. Combining both line pairs favors the $\rm C/O=0.81$ model. We also find that the C$_2$H does not strongly distinguish between cosmic ray ionization rate values, as we expected from the models in Section~\ref{chemresults}. This C/O ratio corresponds to a water ice depletion factor of $50\times$ in the layer C$_2$H is present compared to an interstellar water ice abundance of $8\times10^{-5}$ per H \citep{boogert2015}.
In the same figure, we also show the HCN $J=3-2$ and H$^{13}$CN $J=3-2$ results for comparison. C/O values of $0.7 - 0.9$ also fit the HCN data reasonably well, while C/O of $\gtrsim1$ predicts {\em detectable} H$^{13}$CN $J=3-2$ image-plane emission, regardless of cosmic ray ionization rate, inconsistent with observations.
\subsection{Fits to {\rm HCN and H$^{13}$CN} Observations}\label{sec:hcnfit}
We now consider models with bulk nitrogen depletion to see whether we can arrive at a better fit for HCN $J=3-2$ at a fixed value for C/O. Taking bulk $\rm C/O=0.81$ from Section~\ref{c2hfit}, we reduce the nitrogen abundance from the fiducial value of $7.5\times10^{-5}$ N per H (i.e., $3.75\times10^{-5}$ N$_2$ per H, see Table~\ref{tab:other}). Figure~\ref{fig:chisqN} shows the $(\chi^2$ - $\chi_{\rm min}^2 + 1)$ values for this sub-grid of reduced nitrogen models.
The global best fit is the low $\zeta_{\rm CR}$ value, $1\times10^{-20}$ s$^{-1}$, and no nitrogen depletion. If instead we compare models within a fixed CR ionization rate, the intermediate CR ionization rate model favors no N-depletion, while the the high $\zeta_{\rm CR}=1.3\times10^{-18}$ s$^{-1}$ model, favors anywhere between no and $4\times$ reduction in bulk nitrogen. However, even this factor is small compared to the $\sim20\times$ depletion of CO, or that implied for water and oxygen not in CO based on the C$_2$H results, i.e., a $\sim50\times$ depletion.
\begin{figure}[ht!]
\begin{centering}
\includegraphics[width=0.33\textwidth]{init_grid_N_nozoom_log_v3.pdf}
\caption{$\chi^2$ comparison between the ALMA visibilities and models for HCN $J=3-2$ for differing amounts of initial nitrogen depletion. Colors are the same as in Figure~\ref{fig:chisq} and indicate CR ionization rate. $\chi^2_{\rm min}$ is $6.43\times10^{6}$, so again all models have $\Delta \chi^2$ is close to 1, with smaller values being better fits to the data. For intermediate to low CR rates, no nitrogen depletion is favored. For the higher rate, the models are insensitive to nitrogen depletion factors of $\le 4$ \label{fig:chisqN}}
\end{centering}
\end{figure}
\section{Discussion}\label{sec:discussion}
\subsection{Potential mechanisms behind elevated C/O and N/O ratios}
The relative differences in the abundances of the key volatile carriers are consistent with a picture of sequential loss of volatiles from the warm molecular layer. Water is the least volatile, freezing out at relatively high temperatures. Therefore it will be most impacted by the evolution of grains, through growth and settling, removing oxygen from the surface over time \citep[e.g.,][]{hogerheijde2011}. CO plays an active role in gas and grain surface chemistry, and as such it is gradually converted to species like CO$_2$ and CH$_3$OH, all of which can freeze out onto grains and then become depleted from the surface via dust settling \citep{bergin2014far,schwarz2018}. N$_2$, however, is not as chemically active in the disk surface. At cooler temperatures, below the region of CO freeze-out, N$_2$H$^+$ can form directly from N$_2$ and survive. In the surface, in the presence of CO, this formation pathway is hindered, and most of the surface chemistry requires N$_2$ be dissociated before forming other nitrogen-bearing species. As such, nitrogen will be less chemically ``processed'' than the key oxygen and carbon carriers and is expected to stay in its volatile N$_2$ form given its relatively low desorption temperature \citep{oberg2005}. Consequently, sequestration into ice and subsequent settling should theoretically be less effective for nitrogen-bearing species, consistent with our results from the ALMA observations.
\subsection{Gas-phase C/O and N/O constraints}\label{sec:cono}
The model grid presented in Section~\ref{sec:modabun} focuses on the total elemental abundances in both the gas and solid phases, as these are most relevant for the chemical modeling. The relevant values for comparing to observations of exoplanet atmospheres formed via core accretion are the C/O and N/O ratios specifically in the gas phase. Both CO and N$_2$ are primarily in the gas phase where our observations are most sensitive (i.e., the region where C$_2$H emits, $z/r\gtrsim0.2$), while H$_2$O is primarily ice with some gas-phase H$_2$O from UV photo-desorption (also producing OH).
Figure~\ref{fig:ratio} plots the 2D distributions of the {\em gas-phase} C/O and N/O ratios for our best fit model with an intermediate CR ionization rate ($\zeta_{\rm CR}\sim10^{-19}$ s$^{-1}$) and no nitrogen depletion. Whether we use an intermediate or one order of magnitude lower CR ionization rate does not significantly impact these results. At the surface, all oxygen that starts in H$_2$O ice is quickly dissociated to gas-phase atomic oxygen, such that the gas C/O ratio is equal to the bulk ratio of $\sim0.8$. Where water is both frozen-out and shielded from UV deeper in the disk, CO becomes the primary carbon and oxygen carrier in the gas phase, resulting in C/O $\sim1$. The layered C/O structure is visible here, where C/O in the gas is $0.8$ above normalized heights of $z/r>0.3$ and $\sim1$ below this layer until CO starts to freeze-out, at $z/r \lesssim 0.1$.
The N/O ratio on the other hand is $\gg 1$ throughout the disk atmosphere, and $\sim 10$ when most of the nitrogen is in the gas and CO is the primary oxygen carrier. In the layer where CO begins to freeze-out, N$_2$ is still in the gas due to its slightly lower binding energy \citep{fayolle2016}, and the gas-phase N/O ratio can be $>100$.
\begin{figure}[]
\begin{centering}
\includegraphics[width=0.43\textwidth]{bulk_ratio_v6.pdf}
\caption{ Gas-phase C/O and N/O ratios. Regions that do not have sufficient gas-phase C, O, or N to make a ratio are masked. Note the scale is saturated at the bottom edge of the N/O plot where CO is beginning to freeze out but N$_2$ remains as gas. \label{fig:ratio}}
\end{centering}
\end{figure}
At high velocities we find a discrepancy between the data and modeled C$_2$H, which can be seen in Figure~\ref{fig:channelmodel}. The high velocity emission of C$_2$H is weaker in the models than in the data. Given the signal to noise of the C$_2$H data, we did not attempt to radially or vertically vary the carbon, oxygen, and nitrogen abundances spatially in our fitting; however, from other sources with varied ring-like morphologies in C$_2$H \citep{bergin2016}, there is support for local variations in bulk C/O, N/O, etc. In this instance, the bright high velocity C$_2$H suggests that the inner disk has more excess carbon than the outer disk, which may point to slower sequestration of carbon-bearing volatiles in this region due to, e.g., reduced freeze-out in the warm inner disk.
\begin{figure*}
\begin{centering}
\includegraphics[width=1\textwidth]{channelexample_v6.pdf}
\caption{Sample model channel maps at the line center and at the line wings for HCN $J=3-2$ (left) and C$_2$H $N=3-2$ $J=\sfrac{7}{2}-\sfrac{5}{2}$ (right) lines, along with the data (top, respectively) for comparison. Rows show varying C/O ratios while columns show the low end and high end of the cosmic ray ionization rate considered.
\label{fig:channelmodel}}
\end{centering}
\end{figure*}
We also examined models where we allowed the disk to extend out to 1000 AU, beyond the $\sim500$~AU radial region where C$_2$H and HCN are observed. Since the model C$_2$H abundance continues to rise in the outer disk for the $\rm C/O=0.81$ case (see Figure~\ref{fig:2dabun}), we predict some emission beyond 500 AU, which was not seen. There may be a change in the gas surface density profile near this radius \citep{cleeves2016im,avenhaus2018}, or the disk may have a lower C/O ratio in this region, i.e., may have less efficient volatile sequestration at these very low disk densities. Higher signal to noise, resolved observations of C$_2$H toward this source will help disentangle these scenarios.
\section{Conclusions}\label{sec:conclusion}
Using detailed physical and chemical models constrained by CO and dust observations, along with C$_2$H and HCN data from ALMA, we constrain the C/O and N/O ratios in the molecular layer of the IM Lup disk. Our observations trace the properties of the disk where C$_2$H and HCN emit, primarily at normalized heights above $z/r\gtrsim0.2$. In these layers, the C$_2$H observations favor a super-solar elemental C/O ratio of $\sim0.8$. This high ratio is consistent with preferential loss of water ice from the surface, e.g., due to sequestration by an evolving and growing population of grains. We do not need to sequester any nitrogen for our best fit model, such that the N/O ratio is also super-solar, $\sim10$. The gas phase values of C/O and N/O vary spatially depending on the degree of UV shielding at a given location (Section~\ref{sec:cono}).
While grain sequestration of ices tends to remove volatiles from the surface, it implicitly carries them to the midplane and into the inner disk through settling and radial drift \citep[e.g.][]{oberg2016,piso2016,krijt2016}. If this interpretation is correct, this process would result in a large enhancement over interstellar values of water ice in the inner disk midplane, a mild enhancement of carbon-bearing ice, and relatively little nitrogen ice transport in the solid phase. Correspondingly, we expect the C/O and N/O ratios in midplane solids to both be lower than solar, with N/O much less than solar in the disk midplane.
Where settled ices eventually end up radially is still an open question. Radial drift of solids is thought to be quite efficient \citep[see review of][]{testi2014}, but this process may be slowed by the emergence of pressure variations in the disk that can effectively trap solids \citep[e.g.,][]{weidenschilling1980}. Now with ALMA, ringed radial structures are being observed, and may even be common \citep[e.g.,][]{alma2015,andrews2016,isella2016,loomis2017,huang2018,fedele2018}. In the absence of pressure traps, these grains should travel inward, thermally desorb, and enhance primarily oxygen, followed by carbon, and relatively little nitrogen. \citet{salyk2011} indeed found very low N/O ratios, $5\times10^{-4}$, in the inner disk with {\em Spitzer}; however, the nitrogen ``correction factors'' in the inner disk gas from HCN to total N are uncertain and require additional chemical modeling to constrain.
These results have interesting consequences for the debate regarding abundance measurements in disks \citep[see summary of][]{bergin2017}. At the low gas temperatures typical of disks, H$_2$ does not emit appreciably \citep{bergin2013hd,mcclure2016,bergin2017}. As a result, other uncertain mass tracers are typically used, such as the total dust mass multiplied by a conversion factor, or even optically thin CO emission itself. From our modeling, we find that we do not need to deplete nitrogen significantly, with a factor of $4-20\times$ difference between the CO depletion factor and that for nitrogen. These results would point to abundance variations between these species rather than an overall under-accounting of disk mass, which would impact all volatile abundances at similar if not equal levels. We of course cannot rule out with these data alone some missing gas mass, since for the higher CR models, a small amount of nitrogen depletion (a factor of a few) is allowed, even though these are not the global best fits. However, missing gas alone cannot explain the higher degree of depletion needed for CO and water ice. In the future, additional observations of nitrogen-bearing molecules like HCN may help to break this CO mass / gas mass degeneracy, where in larger disk surveys CO masses appear globally low relative to the dust masses scaled by the interstellar gas-to-dust ratio \citep{ansdell2016,miotello2016,long2017}.
Going forward, these results show that readily observable molecular tracers like CO, C$_2$H, and HCN, and their isotopologues combined with astrochemical models can be used to constrain the gas-phase C/O and N/O ratios within planet-forming disks. Such measurements may furthermore shed light on the inferred volatile composition of the ices, once we better understand the mechanism(s) of volatile loss from the warm molecular layer \citep[e.g.,][]{furuya2014,kama2016}. Future high sensitivity, resolved observations of many sources may further help shed light on ``typical'' gas-phase C/O and N/O ratios, how they spatially vary, and how these ratios may vary with time as planets are forming in the disk. Together, these can one day be compared with C/O measurements of exoplanet atmospheres, to eventually help unravel their formation locations and histories, where observations have already shown a wide range of exoplanet C/O values, even within a single planetary system \citep{bonnefoy2016,lavie2017}.
\acknowledgements{{\it Acknowledgements:} This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2015.1.00964.S and ADS/JAO.ALMA\#2013.1.00226.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan) and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. LIC acknowledges the support of NASA through Hubble Fellowship grant HST-HF2-51356.001-A awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555. J.H. acknowledges support from the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1144152. All simulations were carried out using the Smithsonian Institution High Performance Cluster (SI/HPC).
}
\providecommand{\od}{O'D} \providecommand{\accO}{\'O}
|
\section{Introduction}
Consider the family of matrices given by
\begin{equation}\label{eq:Jn_triang_fact}
J_n=
\setlength\arraycolsep{1.2pt}
\begin{pmatrix}
\mfrac{\alpha_1+1}{2} & \alpha_2 & \alpha_3 & \cdots & \alpha_n\\
0 & \mfrac{\alpha_2+1}{2} & \alpha_3 & \cdots & \vdots\\
0 & 0 & \mfrac{\alpha_3+1}{2} & \ddots & \vdots\\
\vdots & \ddots & \ddots & \ddots & \alpha_n\\
0 & \cdots & \cdots & 0 & \mfrac{\alpha_n+1}{2}
\end{pmatrix}
\setlength\arraycolsep{1.2pt}
\begin{pmatrix}
\mfrac{\alpha_1+1}{2} & 0 & \cdots & \cdots & 0\\
\alpha_1 & \mfrac{\alpha_2+1}{2} & \ddots & \ddots & \vdots\\
\vdots & \alpha_2 & \ddots & \ddots & \vdots\\
\vdots & \vdots & \ddots & \mfrac{\alpha_{n-1}+1}{2} & 0\\
\alpha_1 & \alpha_2 & \cdots & \alpha_{n-1} & \mfrac{\alpha_n+1}{2}
\end{pmatrix},
\end{equation}
where $\mathcal{A}=\{\alpha_1,\alpha_2,\ldots,\alpha_n\}\in\mathbb{R}$ are parameters. Such structured matrices arise when solving certain kinds of optimal control problems \cite{petersalg2009,petvar2017,MIDDLETON2004}, by means of the use of Nehari's Theorem \cite{chuichen12}. In particular, the eigenvalues and eigenvectors of $J_n$ are needed in order to construct an optimal controller and determine the performance achieved by it. Commonly, numerical methods are used to solve such problems, which frequently hide the behaviour of solutions in terms of the parameters defining the problem \cite{GaAp94,Tits}. In this work, we obtain a mathematical description of the eigenvalues of $J_n$ revealing several of their properties with respect to the parameter set $\mathcal{A}$.
For the aforementioned optimal control problem, it is highly desirable to possess mathematical descriptions that allow designers to study its solutions in a deeper way. They provide tools to understand or interpret physical properties of dynamical systems under the presence of feedback. Moreover, closed form solutions for such problems, in terms of the systems parameters, reveal the best achievable performance for certain configurations of control problems, pointing engineers in the correct direction when designing or implementing less sophisticated and, most of the time, more economical control schemes. A notable example of this is when reducing the high order solutions of optimal control problems \cite{antoulas2001survey}. In this context, we believe that the results presented in this work provide sensible tools for dealing with the related optimization problems, giving insight about the nature of the eigenvalues in terms of the defining parameters. We also think that these results might be the basis for finding new algorithms and methods for related problems.
As motivation, and to highlight that the structure of $J_n$ is non-trivial, we can see that for $n=1$ the eigenvalue of $J_1$ is given by $\lambda= 0.25(\alpha_1+1)^2$. For $J_2$ we have that the two eigenvalues are
\begin{align*}
\lambda=&\dfrac{(\alpha_1+1)^2+(\alpha_2+1)^2+4\alpha_1\alpha_2 \pm(\alpha_1 + \alpha_2)\sqrt{(\alpha_1+2)^2+(\alpha_2+2)^2+6\alpha_1\alpha_2-4}}{8}.
\end{align*}
It is of course possible to obtain algebraic expressions for $n=3$, however, it would take several lines of cumbersome expressions to write each of the three eigenvalues. In this paper we obtain results that allow to study the behaviour of the eigenvalues of $J_n$ for arbitrary values of $n$. For instance, if $\alpha_{i}\geq 0$, for all $i$,\footnote{ $\alpha_{i}\geq 0$, for all $i$ is the unique necessary condition in order to deal with optimal control problems as the aforementioned. In this paper, however, we are not restricted to that case and results are presented for $\alpha_{i} \in\mathbb{R}$, for all $i$} we can claim that the corresponding eigenvalues $\lambda_k$, $k=1,2,\dots n$, of $J_n$ are the solutions of
\begin{align}\label{eq:arctan}
\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right) +2\sum_{i=1}^{n}\arctan\left(\dfrac{\alpha_i}{\sqrt{4\lambda_k-1}}\right)=(2k-1)\dfrac{\pi}{2}
\end{align}
for $k=1,\ldots,n.$ Similar claims can be made whenever the eigenvalues of $J_n$ admit real solutions. Even though the properties of the $\arctan(\cdot)$ function clearly imply that \eqref{eq:arctan} should be equivalent to obtaining the roots of a polynomial, we believe that this new representation of the eigenvalue problem provides a lot of insight into the behaviour of the solutions in terms of the parameters $\alpha_i$, specially when considering the originating Control Theory problems. For example, our analysis shows that if all $\alpha_i\geq 0$, then $J_n$ has only real, positive and distinct eigenvalues. Moreover, if $\alpha_i=-\alpha_m=\alpha$ for some $i\neq m$, equation \eqref{eq:arctan} also holds, $J_n$ has an eigenvalue of multiplicity 2 at $(1-\alpha^2)/4$, while the remaining $n-2$ eigenvalues correspond to the eigenvalues of $J_{n-2}$. Also, from \eqref{eq:arctan}, it is possible to generate simple bounds for the values of the eigenvalues that do not seem obvious from the expression for $J_n$. Another benefit of the results proposed in this work correspond to the case of repeated values of the parameters. In fact, if $\alpha_i=\alpha$ for all $i$, \eqref{eq:arctan} collapses to
\begin{align}
\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right) +2n\arctan\left(\dfrac{\alpha}{\sqrt{4\lambda_k-1}}\right)=(2k-1)\dfrac{\pi}{2},
\end{align}
and this equation provides a faster and more accurate way to compute the eigenvalues for large $n$ when compared to standard computational methods (such as the ones used by Matlab).
Our study on the eigenvalues of $J_n$ was also motivated by the fact that non trivial eigenvalue problems with closed form solutions for structured matrices do exist.
For example, a tridiagonal Toeplitz matrix in $\mathbb{R}^{n\times n}$ has known eigenvalues in terms of the three parameters that define it \cite{yueh2005eigenvalues,noschese2013tridiagonal,BUCHHOLZER20121837}. Another well studied related problem is the inverse eigenvalue problem, which in simple words can be said equivalent to find, if possible, a matrix with fixed structure that possesses a set of eigenvalues and/or eigenvectors given a priori \cite{FRIEDLAND197715,chu1998inverse,datta2011solution,higgins2016inverse}. Our findings where sparked by a similar idea. We observed patterns when solving some eigenvalue problems which generated \eqref{eq:arctan} as a conjecture. Our goal was then to show whether it is always possible to connect \eqref{eq:arctan} to a characteristic polynomial, and eventually to our original matrices $J_n$. We notice that, in general, it is possible to write recursive relations for the characteristic polynomials of general tridiagonal matrices through the use of the \emph{Continuant} \cite{muir2003treatise,hearon1970roots}. We use this fact and other manipulations of $J_n$ in order to achieve our goal.
On the other hand, it is important to note that the family of matrices $J_n$ belongs to the class of $n\times n$ rank displacement matrices where only $\mathcal{O}(n)$ parameters are needed to define them (see \cite{kailath1995displacement,pan2012structured} and the references therein). In particular, given a fixed pair of matrices $\{A,B\}$, and a field $\mathbb{F}$, a \emph{Stein} displacement operator $L$ for square matrices is defined as $L:\mathbb{F}^{n\times n}\rightarrow \mathbb{F}^{n\times n}$, such that $L=\Delta_{A,B}$, where
\begin{align}
L(M)=\Delta_{A,B}(M)=M-AMB.
\end{align}
Then, the image $L(M)$ of the operator $L$ is called the \emph{displacement} of $M$ and the rank of $L(M)$ is called the \emph{displacement rank} of $M$. Such definition aims to exploit the structure of the matrix when the displacement rank of $M$ is low and $A,B$ are simple and sparse for matrix computations, inversions, etc \cite{PANstructured,PanWang2003}. In our current case, $J_n$ is similar to the product of two matrices which are the solutions of the Lyapunov equations
\begin{equation}
\Delta_{A,A^\top}(M)=C,\qquad \qquad
\Delta_{A^\top,A}(M)=K,
\end{equation}
with $C,K$ being of rank 1. It is clear that the solutions of the aforementioned equations are of displacement rank 1. However, it is well known that even if two matrices are structured, their product does not necessarily inherit the same explicit structure of the factors. Nevertheless it should be expected for it to also have a low displacement rank, whenever the factors have it \cite{kailath1995displacement}.
Given the previous observations, we believe that our results complement the field of displacement structured matrices. For example, in \cite{pan2012structured,PanWang2003} the inversion of displacement operators was suggested for the solution of Nenvalinna-Pick interpolation and Nehari problems, which are common generators of ``\emph{skew-Hankel-like}" matrices, and the main motivation of our present work. It is considered in \cite{PanWang2003} that the hidden structure of matrices similar to $J_n$ is hard to exploit, however numerical methods are readily available for certain computations. Fast algorithms for computations on similar matrices can be found in \cite{bostanetall2017} and the references therein. We consider that our results provide another connection between displacement methods and optimal control problems, providing an alternative analytical description for a type of problem that seldom allows one.
The presentation of our work is organized as follows. In section \ref{sec:polyJ} we obtain a recursive formula for the sequence of characteristic polynomials of $J_n$. In section \ref{sec:arctan} we derive a recursive formula for a sequence of polynomials that can be connected to \eqref{eq:arctan}. We show in Section \ref{sec:same} that both sequences have indeed the same roots. We present some applications and considerations of our findings in Section \ref{sec:numerics}. Final remarks are given in Section \ref{sec:conclusion}.
\textbf{Notation:} The imaginary unit is denoted by $j^2=-1$. For a number $Q\in\mathbb{C}$, $\Re\{Q\}$ and $\Im\{Q\}$ denote its real and imaginary parts respectively and $\angle\{Q\}$ denotes its argument. For any given square matrix $M$, $\sigma(M)$ denotes its spectrum and $\rho(M)$ its spectral radius.
\section{Characteristic polynomial of $J_n$}\label{sec:polyJ}
In this section we provide a recursive way to compute the characteristic polynomial of $J_n$ defined in \eqref{eq:Jn_triang_fact}. We first define the diagonal matrices
\begin{align}\label{eq:D}
D_\alpha=\begin{pmatrix}
\alpha_1 & & & \\
& \alpha_2 & & \\
& & \ddots & \\
& & & \alpha_n
\end{pmatrix},
\hspace{3mm}
D_\beta=\begin{pmatrix}
\beta_1 & & & \\
& \beta_2 & & \\
& & \ddots & \\
& & & \beta_n
\end{pmatrix}, \quad \text{with} \quad \beta_i \triangleq \dfrac{\alpha_i-1}{2\alpha_i}.
\end{align}
We note that the particular structure of $J_n$ in \eqref{eq:Jn_triang_fact} allow us to write
\begin{align}\label{eq:Jnfactored}
J_n=\left[ L_1-D_{\beta}\right] D_{\alpha}\left[ L_1^\top-D_{\beta}\right] D_{\alpha},
\end{align}
where $L_1$ is the upper triangular matrix with $1$s at the non-zero entries. The inverse of $L_1$ and $L_1^\top$ are straightforward to compute:
\begin{align}
L_1^{-1}=
\left(\begin{array}{ccccc}
1 & -1 & 0 & \cdots & 0\\
0 & 1 & -1 & \ddots &\vdots\\
\vdots & \ddots & 1 & \ddots & 0\\
\vdots & \ddots & \ddots & \ddots & -1 \\
0 & \cdots & \cdots & 0 & 1\end{array}\right),\quad (L_1^\top)^{-1}=(L_1^{-1})^\top.
\end{align}
We will obtain the characteristic polynomial of $J_n$ as the determinant of a tridiagonal matrix, obtaining a recursive relation for their computation.
\begin{proposition}\label{prop1}
The characteristic polynomial $P_n(x)$ of the matrix $J_n$ given by \eqref{eq:Jn_triang_fact}, satisfies the recursive equation
\begin{multline}\label{eq:recJn}
P_n(x)=\left[ \left(\dfrac{1}{\alpha_n}+\dfrac{1}{\alpha_{n-1}}\right)x -1-\dfrac{(\alpha_n-1)^2}{4\alpha_n}-\dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}\right] P_{n-1}(x) \nonumber\\
-\left[ \dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}-\dfrac{\alpha_{n-1}-1}{2}-\dfrac{1}{\alpha_{n-1}}x\right] ^2P_{n-2}(x),
\end{multline}
with initial conditions $P_0(x)=1$ and $P_1(x)=x/\alpha_1-(\alpha_1+1)^2/(4\alpha_1).$
\end{proposition}
\begin{proof}
The characteristic polynomial of $J_n$ can be computed as
\begin{align}
P_n(x)=\det\left(xI-J_n\right),
\end{align}
and from \eqref{eq:Jnfactored}, multiplying accordingly, we can write
\begin{multline}
\det(xI-J_n)=\\
\label{eq:det(xI-J_n)}
\det\left(xL_1^{-1}D_{\alpha}^{-1}(L_1^{-1})^\top-D_{\alpha}+L_1^{-1}D_{\alpha}D_{\beta}
+D_{\alpha}D_{\beta}(L_1^{-1})^\top-L_1^{-1}D_{\alpha}D_{\beta}^2(L_1^{-1})^\top\right).
\end{multline}
It can be checked that every term inside the determinant on the right hand side of \eqref{eq:det(xI-J_n)} is tridiagonal
\begin{align}
L_1^{-1}D_\alpha^{-1}(L_1^{-1})^\top=
\begin{pmatrix}
\dfrac{1}{\alpha_1}+\dfrac{1}{\alpha_2} & -\dfrac{1}{\alpha_2} & 0 & \cdots & 0 & 0 \\[3mm]
-\dfrac{1}{\alpha_2} & \dfrac{1}{\alpha_2}+\dfrac{1}{\alpha_3} & -\dfrac{1}{\alpha_3} & \cdots & 0 & 0\\
0 & -\dfrac{1}{\alpha_3} & \ddots & \cdots & 0 & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \cdots & \dfrac{1}{\alpha_{n-1}}+\dfrac{1}{\alpha_n} & -\dfrac{1}{\alpha_n}\\
0 & 0 & 0 & \cdots & -\dfrac{1}{\alpha_n} & \dfrac{1}{\alpha_n}
\end{pmatrix}
\end{align}
\begin{multline}
L_1^{-1}D_{\alpha}D_{\beta}^2(L_1^{-1})^\top=\\
\begin{pmatrix}
\alpha_1\beta_1^2+\alpha_2\beta_2^2 & -\alpha_2\beta_2^2 & 0 & \cdots & 0 & 0 \\
-\alpha_2\beta_2^2 & \alpha_2\beta_2^2+\alpha_3\beta_3^2 & -\alpha_3\beta_3^2 & \cdots & 0 & 0\\
0 & -\alpha_3\beta_3^2 & \ddots & \cdots & 0 & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \cdots & \alpha_{n-1}\beta_{n-1}^2+\alpha_n\beta_n^2 & -\alpha_n\beta_n^2\\
0 & 0 & 0 & \cdots & -\alpha_n\beta_n^2 & \alpha_n\beta_n^2
\end{pmatrix}
\end{multline}
\begin{equation}
D_{\alpha}-L_1^{-1}D_{\alpha}D_{\beta}-D_{\alpha}D_{\beta}(L_1^{-1})^\top=
\begin{pmatrix}
1 & \alpha_2 \beta_2 & 0 & \cdots & 0 & 0 \\
\alpha_2 \beta_2 & 1 & \alpha_3 \beta_3 & \cdots & 0 & 0\\
0 & \alpha_3 \beta_3 & \ddots & \cdots & 0 & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \cdots & 1 & \alpha_n \beta_n\\
0 & 0 & 0 & \cdots & \alpha_n \beta_n & 1
\end{pmatrix}.
\end{equation}
Now, we can relabel and reorder the elements inside of \eqref{eq:det(xI-J_n)} in order to compute the determinant of $\det(xI-J_n)$ as the determinant of a symmetric tridiagonal matrix
\begin{align}
\det(xI-J_n)=\det\begin{pmatrix}
\gamma_1(x) & \delta_1(x) & 0 & \cdots & 0 & 0 \\
\delta_1(x) & \gamma_2(x) & \delta_2(x) & \cdots & 0 & 0\\
0 & \delta_2(x) & \ddots & \cdots & 0 & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \cdots & \gamma_{n-1}(x) & \delta_{n-1}(x)\\
0 & 0 & 0 & \cdots & \delta_{n-1}(x) & \gamma_{n}(x)
\end{pmatrix}
\end{align}
where
\begin{align}
\gamma_1(x)=&\dfrac{x}{\alpha_1}-\dfrac{(\alpha_1+1)^2}{4\alpha_1} \\
\gamma_n(x)=&\left(\dfrac{1}{\alpha_{n-1}}+\dfrac{1}{\alpha_n}\right)x-1-\left(\dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}+\dfrac{(\alpha_{n}-1)^2}{4\alpha_{n}}\right),\quad n>1\\
\delta_n(x)=&\left(-x\dfrac{1}{\alpha_n}+\dfrac{(\alpha_{n}-1)^2}{4\alpha_{n}}-\dfrac{\alpha_n-1}{2}\right),\quad n\geq 1.
\end{align}
Using the \emph{Continuant} \cite{hearon1970roots} we can conclude that the characteristic polynomial of $J_n$ satisfies the recursive equation
\begin{multline}
P_n(x)=\left[ \left(\dfrac{1}{\alpha_n}+\dfrac{1}{\alpha_{n-1}}\right)x -1-\dfrac{(\alpha_n-1)^2}{4\alpha_n}-\dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}\right] P_{n-1}(x) \\
-\left[ \dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}-\dfrac{\alpha_{n-1}-1}{2}-\dfrac{1}{\alpha_{n-1}}x\right] ^2P_{n-2}(x)
\end{multline}
with initial conditions $P_0(x)=1$, $P_1(x)=x/\alpha_1-(\alpha_1+1)^2/(4\alpha_1)$.
\end{proof}
It is possible to verify that the recursion from Proposition \ref{prop1} generates the same polynomials as the ones obtained by directly computing $\det(xI-J_n)$ with $J_n$ defined as in \eqref{eq:Jn_triang_fact}. It can also be noted that the initial conditions $P_0(x)=\alpha_1$ and $P_1(x)=x-(\alpha_1+1)^2/(4\alpha_1)$ yield polynomials with the same roots.
\section{A sequence of polynomials whose roots are linked to \eqref{eq:arctan}}\label{sec:arctan}
As discussed in the introduction, for non-negative values of the parameters, we conjecture that the $n$ eigenvalues of $J_n$, $\lambda_k$, $k=1,\ldots,n$, satisfy the equations
\begin{align}
\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right) +2\sum_{i=1}^{n}\arctan\left(\dfrac{\alpha_i}{\sqrt{4\lambda_k-1}}\right)=(2k-1)\dfrac{\pi}{2}
\end{align}
for $k=1,\ldots,n$. Numerical examples suggest that these equations are indeed yielding the eigenvalues of $J_n$. We proceed to find a sequence of polynomials that have for roots the solutions of \eqref{eq:arctan}. In the first place, we note that these equations can be interpreted as the angle of a product of complex factors. A possibility is then to have a polynomial $R_n(x)$ with real coefficients, written as the sum of a polynomial with complex coefficients $Q_n(x)$ plus its conjugate $\overline{Q}_n(x)$. Therefore, some of the roots of the polynomials $R_n(x)$ (or $\Re\{Q_n(x)\}$) will be given by the values of $x$ where the complex number $Q_n(x)$ is pure imaginary, or, interchangeably, when $\angle\{Q_n(x)\}$, that is, the argument of the complex number $Q_n(x)$, is an odd multiple of $\pi/2$. We now have the following result.
\begin{proposition}\label{prop2}
Let $R_n(x)=Q_n(x)+\overline{Q}_n(x)$ be a polynomial of order $n$ with real coefficients and which has as roots the $n$ solutions of \eqref{eq:arctan} when $\alpha_i\geq 0$ for all $i$. Then, the real part of the complex polynomial $Q_n(x)$ satisfies the recurrence equation
\begin{align}\label{eq:recQn}
\Re\{Q_n(x)\}=&\left(\frac{(4x-1)-\alpha_n\alpha_{n-1}}{\alpha_{n-1}} \right) \left(\frac{\alpha_n +\alpha_{n-1}}{4} \right)\Re\{Q_{n-1}(x)\}\nonumber\\
& -\dfrac{\alpha_n}{\alpha_{n-1}} \left(x+ \frac{(\alpha_{n-1}^2 -1)}{4} \right)^2\Re\{Q_{n-2}(x)\},
\end{align}
with initial conditions $\Re\{Q_0(x)\}=1$ and $\Re\{Q_1(x)\}=x-(\alpha_1+1)^2/4.$
\end{proposition}
\begin{proof}
Firstly, we must note that there exists an infinite number of polynomials sharing roots with the solutions of \eqref{eq:arctan}, according to the earlier discussion. A general class of polynomials that shares roots with expressions similar to \eqref{eq:arctan} can be described as
\begin{align}\label{eq:atanRn(x)}
R_n(x)=(a+jb)\prod_{i=1}^{n}\left(v_i+jw_i\right)^2+(a-jb)\prod_{i=1}^{n}\left(v_i-jw_i\right)^2.
\end{align}
In this particular case we must have
\begin{align}
\dfrac{b}{a}=\dfrac{1}{\sqrt{4x-1}},\quad \dfrac{w_i}{v_i}=\dfrac{\alpha_i}{\sqrt{4x-1}}.
\end{align}
Now, we can claim that
\begin{align}\label{eq:Qn(x)}
R_n(x)=2\Re\{Q_n(x)\},\quad Q_n(x)=(a+jb)\prod_{i=1}^{n}\left(v_i+jw_i\right)^2.
\end{align}
Without loss of generality, and in order to work with monic polynomials, we consider
\begin{align}
v_i=\dfrac{\sqrt{4x-1}}{2},\quad w_i=\dfrac{\alpha_i}{2},\quad a=\sqrt{4x-1},\quad b=1.
\end{align}
It is possible to now write
\begin{align}
\left(\dfrac{\sqrt{4x-1}}{2}+j\dfrac{\alpha_i}{2} \right)^2=\underbrace{\dfrac{1}{4}\left((4x-1)-\alpha_i ^2\right)}_{F_i}+j\underbrace{\dfrac{\alpha_i}{2}\sqrt{4x-1}}_{G_i},
\end{align}
and $Q_n(x)$ can be now written as
\begin{align}
Q_n(x)=(a+jb)\prod_{i=1}^{n}\left(F_i+jG_i\right)
\end{align}
which is a complex polynomial satisfying the following recursions
\begin{align}\label{eq:ImQn1}
\Re\left\{Q_n(x)\right\}=&F_n\Re\left\{Q_{n-1}(x)\right\}-G_n\Im\{Q_{n-1}(x)\}\\\label{eq:ImQn2}
\Im\left\{Q_n(x)\right\}=&F_n\Im\left\{Q_{n-1}(x)\right\}+G_n\Re\{Q_{n-1}(x)\}.
\end{align}
Note how $a+jb$ disappears in these recursions. Solving for a recursion on $\Re\{Q_n(x)\}$ is not a hard task if we note that from \eqref{eq:ImQn1}
\begin{align}
\Im\left\{Q_{n-1}(x)\right\}=\dfrac{F_n\Re\{Q_{n-1}(x)\}-\Re\{Q_n(x)\}}{G_n},
\end{align}
and from \eqref{eq:ImQn2}
\begin{align}
\Im\left\{Q_{n-1}(x)\right\}=\dfrac{\Im\left\{Q_n(x)\right\}-G_n\Re\{Q_{n-1}(x)\}}{F_n}.
\end{align}
Solving for $\Im\left\{Q_n(x)\right\}$ we obtain
\begin{align}
\Im\left\{Q_n(x)\right\}=\left(\dfrac{F_n^2}{G_n}+G_n\right)\Re\{Q_{n-1}(x)\}-\dfrac{F_n}{G_n}\Re\{Q_n(x)\}.
\end{align}
If we shift this backwards by 1 and replace it in the right hand side of \eqref{eq:ImQn1} yields
\begin{align}\label{eq:recReQn}
\Re\{Q_n(x)\}=c_n(x) \Re\{Q_{n-1}(x)\}-d_n(x)\Re\{Q_{n-2}(x)\},
\end{align}
where
\begin{align}
c_n(x) \triangleq & F_n+F_{n-1}\dfrac{G_n}{G_{n-1}}=\dfrac{1}{4}\left((4x-1)-\alpha_n^2\right)+ \dfrac{1}{4}\left((4x-1)-\alpha_{n-1}^2\right)\frac{\alpha_n}{\alpha_{n-1}}\nonumber\\
& \hspace{2.6cm}=\left(\frac{(4x-1)-\alpha_n\alpha_{n-1}}{\alpha_{n-1}} \right) \left(\frac{\alpha_n +\alpha_{n-1}}{4} \right) \\
d_n(x) \triangleq & F_{n-1}^2\dfrac{G_n}{G_{n-1}}+G_{n}G_{n-1}=\dfrac{1}{16}\left((4x-1)-\alpha_{n-1}^2\right)^2\frac{\alpha_n}{\alpha_{n-1}}+
\dfrac{1}{4}\alpha_n\alpha_{n-1}(4x-1)\nonumber\\
& \hspace{3.4cm}=\dfrac{\alpha_n}{\alpha_{n-1}} \left(x+ \frac{(\alpha_{n-1}^2 -1)}{4} \right)^2.
\end{align}
Moreover, for $n=1$, we have that
\begin{align}
\dfrac{Q_1(x)}{a+jb}=\left(\dfrac{\sqrt{4x-1}}{2}+j\dfrac{\alpha_1}{2}\right)^2=\dfrac{1}{4}\left(4x-1-\alpha_1^2+j2\alpha_1\sqrt{4x-1}\right),
\end{align}
and $\Re\{Q_1(x)\}=0$ implies
\begin{align}
a\left(x-\dfrac{1+\alpha_1^2}{4}\right)-b\dfrac{\alpha_1}{2}\sqrt{4x-1}=a\left(\left(x-\dfrac{1+\alpha_1^2}{4}\right)-\dfrac{\alpha_1b}{2a}\sqrt{4x-1}\right)=0.
\end{align}
If $a\neq 0$ and since $b/a=1/\sqrt{4x-1}$, we have that
\begin{align}
x=\dfrac{1}{4}\left(1+\alpha_1\right)^2.
\end{align}
Therefore, with $\Re\{Q_0(x)\}=1$ and $\Re\{Q_1(x)\}=x-(\alpha_1 +1)^2/4$, the recursion \eqref{eq:recReQn} generates polynomials with roots at the solutions of \eqref{eq:arctan}.
\end{proof}
The previous result states that the solutions of \eqref{eq:arctan} can be associated to the roots of a sequence of polynomials. It can be checked numerically that for particular values of the parameters $\alpha_i\geq 0$, the polynomials generated by the recursion \eqref{eq:recReQn} share roots with the eigenvalues of $J_n$.
\begin{remark}
The particular case where $\alpha_i = 1$ for all $i$, corresponds to
\begin{align}
J_n=\left(\begin{array}{cccc}
1 & 1 & \cdots & 1\\
0 & 1 & \ddots & \vdots\\
\vdots & \ddots & \ddots & 1\\
0 & \cdots & 0 & 1 \end{array}\right)\left(\begin{array}{cccc}
1 & 0 & \cdots & 0\\
1 & 1 & \ddots & \vdots\\
\vdots & \ddots & \ddots & 0\\
1 & \cdots & 1 & 1 \end{array}\right),
\end{align}
namely, the product of the all-ones upper and lower triangular matrices. This matrix has well known eigenvalues for any $n$ given by \cite{elliott1953characteristic}
\begin{align}\label{eq:sigmacinfty}
\sigma(J_n)=\left\{\dfrac{1}{4}\sec^2\left(\dfrac{\pi}{2n+1}\right),\ldots,\dfrac{1}{4}\sec^2\left(\dfrac{n\pi}{2n+1}\right)\right\}.
\end{align}
Moreover, in this case
$$F_n=\left(x-\dfrac{1}{2}\right),\quad G_n=\dfrac{\sqrt{4x-1}}{2},$$
and therefore
$$\Re\{Q_n\}=(2x-1)\Re\{Q_{n-1}\}-x^2\Re\{Q_{n-2}\},$$
with $\Re\{Q_0\}=1$ and $\Re\{Q_1\}=x-1$. This recursion generates polynomials having the same roots as \eqref{eq:sigmacinfty}.
We can readily check that \eqref{eq:arctan} collapses to
\begin{align}
(2n+1)\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right)=(2n+1)\left(\dfrac{\pi}{2}-\arctan(\sqrt{4\lambda_k-1})\right)=\dfrac{(2k+1)\pi}{2},
\end{align}
from where we have that
\begin{align}
\lambda_k=\dfrac{1}{4}\left(1+\dfrac{1}{\tan^2\left(\dfrac{(2k+1)\pi}{2(2n+1)}\right)}\right)=\dfrac{1}{4}\sec^2\left(\dfrac{\pi(n-k)}{2n+1}\right),
\end{align}
for $k=1,\ldots,n,$ which again coincides with \eqref{eq:sigmacinfty}.
\end{remark}
\begin{remark}
The recursion \eqref{eq:recReQn} can be also associated to a tridiagonal matrix through the Generalized Continuant \cite{hearon1970roots}. Let $K_n$ be the matrix
\begin{align}
\displaystyle K_n = \begin{pmatrix} r_1 & s_2 & 0 & \ldots & 0 & 0 \\ t_2 & r_2 & s_3 & \ldots & 0 & 0 \\ 0 & t_3 & r_3 & \ldots & 0 & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\ 0 & 0 & 0 & \ldots & r_{n-1} & s_{n} \\ 0 & 0 & 0 & \ldots & t_{n} & r_n \end{pmatrix}
\end{align}
where $r_1=\Re\{Q_1\}=x-0.25(\alpha_{1}+1)^2$, $$r_i=\left(F_i+F_{i-1}\dfrac{G_i}{G_{i-1}}\right),$$
for $i>1$ and $s_i,t_i$ satisfy
$$s_it_i=\left(F_{i-1}^2\dfrac{G_i}{G_{i-1}}+G_{i}G_{i-1}\right),$$
for $i>1.$ Then, $\det(K_n)$ generates the same sequence of polynomials provided by \eqref{eq:recReQn}. \qed
\end{remark}
\begin{remark}
It is important to note that for Proposition \ref{prop2} to hold with respect to \eqref{eq:arctan}, every $\alpha_i$ must be non-negative. We would like to stress that this was made in order to highlight the inspiration we had for obtaining our results. The given sequence of polynomials $\Re\{Q_n(x)\}$ that satisfies \eqref{eq:recQn} is valid for any choice of the real parameters $\alpha_i$. Moreover, the eigenvalues of $J_n$ are indeed the roots of $\Re\{Q_n(x)\}$ as we will show in the following section.
\end{remark}
\section{Main results}\label{sec:same}
In this section we derive the main theorem of this work, which connects the recursive polynomials obtained in previous sections. We also derive and discuss some of its consequences.
\begin{theorem}\label{thm1}
Consider the $n\times n$ matrix $J_n$ defined in \eqref{eq:Jn_triang_fact} with $\alpha_{i}\in\mathbb{R}$ for $i=1,\ldots,n$. Its $n$ eigenvalues, $\lambda_k$, $k=1,\ldots,n$, are given by the $n$ roots of the polynomials $\Re\{Q_n(x)\}$.
\end{theorem}
\begin{proof}
The characteristic polynomial of $J_n$, $P_n(x)$, satisfies the recursion \eqref{eq:recJn}, namely
\begin{align}\label{eq:PnTheo}
P_n(x)=\gamma_n(x)P_{n-1}(x)-\delta_n(x)^2P_{n-2}(x)=\delta_{n}(x)^2\left(\dfrac{\gamma_n(x)}{\delta_{n}(x)^2}P_{n-1}(x)-P_{n-2}(x)\right)
\end{align}
where
\begin{align}
\label{a_modificado}
\gamma_n(x)&=\left(\dfrac{1}{\alpha_n}+\dfrac{1}{\alpha_{n-1}}\right)x -1-\dfrac{(\alpha_n-1)^2}{4\alpha_n}-\dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}\\\nonumber &=\dfrac{(\alpha_n+\alpha_{n-1})(4x-1-\alpha_n \alpha_{n-1})}{4\alpha_n \alpha_{n-1}},\\
\label{b_modificado}
\delta_n(x)&=\dfrac{(\alpha_{n-1}-1)^2}{4\alpha_{n-1}}-\dfrac{\alpha_{n-1}-1}{2}-\dfrac{1}{\alpha_{n-1}}x\\\nonumber
&=-\left( \dfrac{\alpha_{n-1}^2-1+4x}{4\alpha_{n-1}}\right) ,
\end{align}
with initial conditions $P_0(x)=1$, and $P_1(x)=\dfrac{1}{\alpha_1}x-1-\dfrac{(\alpha_1-1)^2}{4\alpha_1}.$
On the other hand, we have that the polynomials $R_n(x)$ defined in Proposition \ref{prop2} satisfy the relation
\begin{align}
R_n(x)=c_n(x)R_{n-1}(x)-d_n(x)R_{n-2}(x)=d_n(x)\left(\dfrac{c_n(x)}{d_n(x)}R_{n-1}(x)-R_{n-2}(x)\right),
\end{align}
where
\begin{align}
\label{c_modificado}
c_n(x)=&\frac{\left( 4x-1-\alpha_n\alpha_{n-1}\right) \left(\alpha_n +\alpha_{n-1} \right)}{4\alpha_{n-1}} \\
\label{d_modificado}
d_n(x)=&\dfrac{\alpha_n}{16\alpha_{n-1}} \left(4x-1+ \alpha_{n-1}^2 \right)^2,
\end{align}
with the initial conditions $R_0(x)=1$, and $R_1(x)=\left(x-\dfrac{(\alpha_1+1)^2}{4}\right)=\alpha_1 P_1(x).$
We claim that $R_n(x)=\theta_nP_n(x)$ for all $n$ where $\theta_n$ is a real constant for every $n$. This would make the roots of both polynomials equal for all $n$, proving the Theorem. In order to do this, we first note from \eqref{a_modificado}, \eqref{b_modificado}, \eqref{c_modificado} and \eqref{d_modificado} that, for all $n$
\begin{equation}
\dfrac{c_n(x)}{d_n(x)}=\dfrac{1}{\alpha_{n-1}}\dfrac{\gamma_n(x)}{\delta_n(x)^2},\qquad \;
d_n(x)=\alpha_n \alpha_{n-1} \delta_n(x)^2.
\end{equation}
In particular, for $n=2$, we have
\begin{align}
R_2(x)=&d_2(x)\left(\dfrac{c_2(x)}{d_2(x)}R_1(x)-R_0(x)\right).
\end{align}
Substituting the previous expressions and recalling that $R_0(x)=P_0(x)$ and $R_1(x)=\alpha_1 P_1(x)$ we have
\begin{align}
R_2(x)=\alpha_1 \alpha_2 \delta_2(x)^2\left(\dfrac{\gamma_2(x)}{\delta_2(x)^2}P_1(x)-P_0(x)\right)
\end{align}
which, according to \eqref{eq:PnTheo}, yields
\begin{align}
R_2(x)=\alpha_1 \alpha_2 P_2(x).
\end{align}
We then proceed by induction. Let us assume that for $k=1,\ldots,n-1$, the following equality holds:
\begin{align}\label{eq:induc}
R_{k}(x)=\left(\prod_{i=1}^{{k}}\alpha_i\right)P_{k}(x),
\end{align}
We know that
\begin{align}
R_n(x)=\alpha_{n} \alpha_{n-1} \delta_n(x)^2\left(\dfrac{1}{\alpha_{n-1}}\dfrac{\gamma_n(x)}{\delta_n(x)^2}R_{n-1}(x)-R_{n-2}(x)\right).
\end{align}
Substituting \eqref{eq:induc} we obtain
\begin{align}
R_n(x)=&\alpha_{n} \alpha_{n-1} \delta_n(x)^2\left(\dfrac{1}{\alpha_{n-1}}\dfrac{\gamma_n(x)}{\delta_n(x)^2}\left(\prod_{i=1}^{{n-1}}\alpha_i\right)P_{n-1}(x)-\left(\prod_{i=1}^{{n-2}}\alpha_i\right)P_{n-2}(x)\right).
\end{align}
Factorizing and reordering we finally obtain
\begin{align}
R_n(x)=&\alpha_{n} \alpha_{n-1} \left(\prod_{i=1}^{{n-2}}\alpha_i\right)\left(\gamma_n(x)P_{n-1}(x)-\delta_n(x)^2P_{n-2}(x)\right)=\left(\prod_{i=1}^{{n}}\alpha_i\right)P_n(x),
\end{align}
showing than indeed $P_n(x)$ and $R_n(x)$ share the same roots.
\end{proof}
Theorem \ref{thm1} states that the characteristic polynomial of $J_n$ can be expressed as the sum of a complex polynomial $Q_n(x)$ and its conjugate. This complex polynomial is a simple product of $n$ quadratic factors and a first order factor. The real parameters $\alpha_i$ appear each in one of the quadratic factors. There are a few obvious observations that can be made. For example, the parameters $\alpha_i$ can be arbitrarily permuted and the eigenvalues will not change. This particular property is to be expected when considering the optimal control problems that give rise to $J_n$. In said problems, it is entirely expected to have no numerical change on the result when permuting the problem parameters. We can also note that if $\alpha_i=0$ for some $i$, then there will be an eigenvalue of multiplicity $m$ at $x=0.25$, with $m$ equal to the number of parameters equal to $0$.
Some less obvious consequences are given in the following corollaries.
\begin{corollary}\label{cor:posit}
If the parameters $\alpha_i>0$ for all $i$, then the $n$ eigenvalues of $J_n$, say $\lambda_k$ for $k=1,\ldots,n$, satisfy \eqref{eq:arctan}. Also $\lambda_k$ are real positive, distinct, and monotonically decrease with $k$.
\end{corollary}
\begin{proof}
We first note that the first claim is true given the motivation behind Proposition \ref{prop2} and the results in Theorem \ref{thm1} that implies that the characteristic polynomial of $J_n$ share the same roots of the polynomial in Proposition \ref{prop2}.
For the second part, we also note that the right hand side of \eqref{eq:arctan} increases with $k$ and that the $n$ equations correspond to the $n$ intersections of the left hand side function with the constants on the right hand side. Moreover, since $\arctan(\cdot)$ is a continuous odd function and the arguments $\alpha_i(4x-1)^{-1/2}$ are always positive, the solutions must be positive if every $\alpha_i$ is positive. It can also be noted that
\begin{align}
\dfrac{d}{du}\left(\arctan\left(u\right) +2\sum_{i=1}^{n}\arctan\left(\alpha_iu\right)\right)>0,
\end{align}
for every $u>0$, whenever $\alpha_i>0$ for all $i$ and
\begin{align}
\dfrac{d}{dx}\left(\dfrac{\alpha_i}{\sqrt{4x-1}}\right)<0
\end{align}
for all $x>1/4$. This implies for $\alpha_i>0$ for all $i$, that the left hand side of \eqref{eq:arctan} is a strictly decreasing monotonic function that maps $x>1/4$ to the interval $(0,(2n+1)\pi/2)$ Hence, the $n$ eigenvalues $\lambda_k$ for $k=1,\ldots,n$ are positive, distinct and monotonically decrease with $k$ in this case.
\end{proof}
Corollary \eqref{cor:posit} is particularly useful to deal with the optimal control problem that motivates the study of the eigenvalues of $J_n$ since, in that case, each $\alpha_i$ is known to be positive. The properties of the nonlinear equation \eqref{eq:arctan} allow us to analyze the solution of the optimal control problem in a more intuitive way for a deeper understanding of the underlying problems.
\begin{corollary}\label{cor:repeat}
If $\alpha=\alpha_m=-\alpha_l$ for some $m\neq l$, then $J_n$ has an eigenvalue of multiplicity $2$ at $\lambda=(1-\alpha^2)/4$ and the remaining $n-2$ eigenvalues correspond to the solutions of \eqref{eq:arctan} for $k=0,\ldots,n-3$, removing the terms for $\alpha_l$ and $\alpha_m$.
\end{corollary}
\begin{proof}
The second part of the corollary follows directly from setting $\alpha_m=-\alpha_l$ in \eqref{eq:arctan} and from the fact that $\arctan(\cdot)$ is an odd function. This cancels out the two terms $2\arctan(\alpha_m/\sqrt{4x-1})$ and $2\arctan(\alpha_l/\sqrt{4x-1})$ and the solutions of these equations do not depend neither on $\alpha_m$ nor $\alpha_l$. We can alternatively see that the solutions of \eqref{eq:arctan} are equivalent to finding the roots of
\begin{align}
R_n(x)=(\sqrt{4x-1}+j)\prod_{i=1}^{n}\left(\sqrt{4x-1}+j\alpha_i\right)^2+(\sqrt{4x-1}-j)\prod_{i=1}^{n}\left(\sqrt{4x-1}-j\alpha_i\right)^2,
\end{align}
and setting $\alpha_m=\alpha=-\alpha_l$ yields
\begin{align}
R_n(x)=&(\sqrt{4x-1}+j)(4x-1+\alpha^2)^2\prod_{i=1,i\neq m,i\neq l}^{n}\left(\sqrt{4x-1}+j\alpha_i\right)^2+\\\nonumber
&(\sqrt{4x-1}-j)(4x-1+\alpha^2)^2\prod_{i=1,i\neq m,i\neq l}^{n}\left(\sqrt{4x-1}-j\alpha_i\right)^2,
\end{align}
and $R_n(x)=(4x-1+\alpha^2)^2\tilde{R}_{n-2}(x)$ where $\tilde{R}_{n-2}(x)$ corresponds to the polynomial $R_{n-2}(x)$ with the remaining $n-2$ parameters from the original problem relabelled.
\end{proof}
\begin{remark}
It is interesting to note that whenever two parameters satisfy $\alpha_m=-\alpha_j$, the remaining roots do not depend on them. This invariance property seems natural considering the problems that give rise to the study of these matrices. In particular, several optimal approximation problems in Hardy spaces involve the computation of the eigenvalues of related matrices \cite{chui2012discrete}. This invariance property could be key in studying the solutions to such problems or characterizing the equivalent eigenvalue problems. A similar thing can be said about the permutation property of the parameters $\alpha_{i}$.
We can also check that $J_n$ will have an eigenvalue at the origin if at least one $\alpha_i=-1$. Using Corollary \ref{cor:repeat}, $J_n$ will be nilpotent if and only if $\alpha_{i}=\pm 1$ for all $i$, and the number of parameters equal to $-1$ is equal to $\lceil n/2 \rceil$, where $\lceil \cdot \rceil$ denotes the ceiling function.
\end{remark}
\section{Examples and other miscellaneous derivations}\label{sec:numerics}
In this section we provide some examples that illustrate applications of the results and some of their practical features.
\subsection{Sensitivity of the eigenvalues to changes of the parameters}
Having \eqref{eq:arctan} allows us to easily compute the partial derivatives of the solutions with respect to the parameters $\alpha_i$. Since we have that the right hand side of \eqref{eq:arctan} is a constant, the expressions are valid for every eigenvalue. In particular, using the implicit function theorem, we have that any solution $\lambda_k$ of \eqref{eq:arctan}, as a function of the parameters $\alpha_i$, satisfies
\begin{align}\label{eq:partials}
\dfrac{\partial \lambda_k}{\partial \alpha_i}=2\dfrac{4\lambda_k-1}{4\lambda_k-1+\alpha_i^2}\left(\dfrac{1}{2\lambda_k}+4\sum_{j=1}^{n}\dfrac{\alpha_j}{4\lambda_k-1+\alpha_j^2}\right)^{-1}.
\end{align}
Note that the factor $(4x-1+\alpha_i^2)$ outside of the parenthesis cancels out with a term inside it. As stated in corollary \ref{cor:posit}, for positive parameters $\alpha_i$, we have positive and distinct eigenvalues. Even though it is to be expected from the expressions for $J_n$, in this case, \eqref{eq:partials} implies that the solutions $x$ become larger with the increase of any $\alpha_i$.
We can also claim, for a fixed set of $n+1$ parameters that, if $\alpha_i>0$ for all $i$, $\rho(J_n)=\overline{\lambda}_n<\overline{\lambda}_{n+1}=\rho(J_{n+1})$. To see this consider the left hand side of \eqref{eq:arctan} for $n$ parameters and evaluate it at $\overline{\lambda}_{n+1}$
\begin{align}
\arctan\left(\dfrac{1}{\sqrt{4\overline{\lambda}_{n+1}-1}}\right)+2\sum_{i=1}^n\arctan&\left(\dfrac{\alpha_{i}}{\sqrt{4\overline{\lambda}_{n+1}-1}}\right)= \\\nonumber
&\dfrac{\pi}{2}-2\arctan\left(\dfrac{\alpha_{n+1}}{\sqrt{4\overline{\lambda}_{n+1}-1}}\right).
\end{align}
The expression on the left above is strictly decreasing and equals $\pi/2$ only at $\overline{\lambda}_n$. Assuming that $\overline{\lambda}_n>\overline{\lambda}_{n+1}$ contradicts the latter.
\subsection{Numerical computations and simple bounds}
The alternative equation for computing the eigenvalues of $J_n$ when the parameters are non-negative is suitable for iterative methods that utilize derivatives, since these are readily available in analytical form. Moreover, Newton's method can be used to obtain easy bounds for the eigenvalues if we let
\begin{align}
f_{n,k}(x)=\arctan\left(\dfrac{1}{\sqrt{4x-1}}\right) +2\sum_{i=1}^{n}\arctan\left(\dfrac{\alpha_i}{\sqrt{4x-1}}\right)-(2k-1)\dfrac{\pi}{2}.
\end{align}
It is not hard to verify that for $n=2$ the largest eigenvalue of $J_2$ in this case satisfies
\begin{align}
\rho(J_2)\geq \dfrac{(\alpha_1+1)^2}{4}+\dfrac{(\alpha_2+1)^2}{4}.
\end{align}
Extrapolating, a simple, albeit conservative, lower bound for arbitrary $n$ is given by
\begin{align}
\rho(J_n)\geq x_0-\dfrac{f_{n,k}(x_0)}{f_{n,k}'(x_0)},
\end{align}
where
\begin{align}
x_0=\sum_{i=1}^{n}\dfrac{(\alpha_i+1)^2}{4}.
\end{align}
\subsection{Properties of the roots for negative parameters}
If any parameter $\alpha_{i}$ is negative, we cannot use \eqref{eq:arctan} in a straightforward fashion. However, $R_n(x)$ given by the recursion \eqref{eq:atanRn(x)} remains valid and all the eigenvalues of $J_n$ are roots of of $R_n(x).$ It should be clear to the reader that the roots of a polynomial are continuous functions of its parameters. So far we have shown some special cases where we can claim the exact location of some roots for particular parameter values. It is also a well known fact that in the real parameter case, roots with multiplicity come from complex conjugate pairs and split into real distinct roots as one parameter varies or they remain complex \cite{hinrichsen2005mathematical}. Hence, we have to consider the case of negative and even complex conjugate pairs as roots of $R_n(x)$ when at least one parameter is negative. We will illustrate this with the simplest case $n=1$. As we know the solution, we have that
\begin{align}
\lambda_1(\alpha)=\dfrac{(1+\alpha)^2}{4},
\end{align}
which evaluated at $Q_n(x)$ yields
\begin{align}
Q_n(\lambda_1(\alpha))=(\sqrt{(1+\alpha)^2-1}+j)\left(\sqrt{(1+\alpha)^2-1}+j\alpha\right)^2.
\end{align}
As we already know, when $\alpha\geq 0$, the solution is equivalent to solving \eqref{eq:arctan}. If $\alpha=-1$ the solution is $\lambda_1=0.$ Substituting in the previous expression we have that
\begin{align}
Q_n(\lambda_1(-1))=(\sqrt{-1}+j)\left(\sqrt{-1}-j\right)^2=0,
\end{align}
that is, both the real and imaginary part of $Q_n(x)$ are $0$ at $x=\lambda_1(-1).$ Clearly this case is not covered directly by \eqref{eq:arctan}. However, if $\alpha=-2$
\begin{align}
Q_n(\lambda_1(-2))=j\left(-j\right)^2=-j,
\end{align}
which has per argument $-\pi/2$, and is also not covered by \eqref{eq:arctan}.
We can distinguish a few extra cases for posing alternative equations. A real root that satisfies $4\lambda_k-1<0$ could be recovered by rewriting
\begin{align}
Q_n(x)&=(j\sqrt{1-4x}+j)\prod_{i=1}^n\left(j\sqrt{1-4x}+j\alpha_i\right)^2\\
\nonumber &=(-1)^nj(\sqrt{1-4x}+1)\prod_{i=1}^n\left(\sqrt{1-4x}+\alpha_i\right)^2.
\end{align}
Then we have (note the appearance of the superfluous solution $x=0.25$.)
\begin{align}\label{eq:Pu}
R_n(x)=(-1)^nj&\left((\sqrt{1-4x}+1)\prod_{i=1}^n\left(\sqrt{1-4x}+\alpha_i\right)^2+\right.\\\nonumber
&\left.(\sqrt{1-4x}-1)\prod_{i=1}^n\left(\sqrt{1-4x}-\alpha_i\right)^2\right),
\end{align}
and since we are looking for real roots we have that the left term inside the parenthesis is always positive. At the same time, the right term inside the parenthesis is only negative whenever $\sqrt{1-4x}<1$. We can also note that in this case, $R_n(x)$ is always pure imaginary. However, if there exists a value of $x$ that vanishes the expression between parenthesis, we will have that $1-4x<1$, or equivalently, $x>0.$ It is possible to claim that when only one parameter $\alpha_{m}$ is negative, a necessary and sufficient condition to have an eigenvalue in $(0,0.25)$ is $\alpha_{m}\in (-1,0).$
There is one exception to this. As we stated in Corollary \ref{cor:repeat}, one admitted solution for the case $\alpha_m=-\alpha_l=\alpha$ and located at $x=(1-\alpha^2)/4$, which is only negative for $|\alpha|> 1$. These solutions vanish both the real and imaginary part of $Q_n(x)$ and are the only case where real negative roots can be found as eigenvalues of $J_n$.
In general, if some $\alpha_i<0$, all the real positive eigenvalues of $J_n$ satisfying $\lambda_k>0.25$ are found by solving
\begin{align}\label{eq:atanex1}
\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right) +2\sum_{i=1}^{n}\arctan\left(\dfrac{\alpha_i}{\sqrt{4\lambda_k-1}}\right)=(2k-1)\dfrac{\pi}{2},
\end{align}
where $k\in\mathbb{Z},$ that is, the right hand side of the equation above can also be a negative odd multiple of $pi/2$. The number of feasible solutions could be studied by finding the global maximum and minimum of the sum of $\arctan(\cdot)$ functions, and counting the number of times that an odd multiple of $\pi/2$ falls between this interval. Since we are assuming real parameters, the remaining eigenvalues must come in complex conjugate pairs or belong to the interval $[0,0.25]$
For example, let us consider $n=5$ with $A=\{5,-0.1,3,-2,1.5\}$. In Figure \ref{fig:atanex1} we see all the solutions of \eqref{eq:atanex1} which are approximately $\lambda_1\approx 0.9821$ and $\lambda_2=22.5527$. It can be noted (note the zoom in of Figure \ref{fig:atanex1}) that since there are negative parameters, the slope of the sum of $\arctan(\cdot)$ functions is no longer negative for all $x$. It is possible, for some values of the parameters, for a single equation to have more than one solution.
\begin{figure}[!th]\begin{center}
\includegraphics[width=\textwidth]{atanex1.eps}\\
\caption{Graphical solutions of \eqref{eq:atanex1}} \label{fig:atanex1}
\end{center}
\end{figure}
We are still missing 3 eigenvalues. In Figure \ref{fig:Puex1} we have the plot of the imaginary part of $R_n(x)$ assuming $x<0.25$. There is a single cross with the $x$ axis at $\lambda_3\approx 0.2287<0.25$. The last two eigenvalues are a complex pair located at $\lambda_{4,5}=-0.7492\pm j0.03131$.
\begin{figure}[!th]\begin{center}
\includegraphics[width=\textwidth]{Puex1.eps}\\
\caption{Graphical solutions of \eqref{eq:Pu}} \label{fig:Puex1}
\end{center}
\end{figure}
\subsection{Root locus of a particular case}
To illustrate some of our previous comments we have plotted the root locus of $R_n(x)$ for the fixed parameters $A\setminus\{\alpha_{1}\}=\{3,-2,-5,1.5\}$ and $\alpha_{1}\in [-200,200]$. For these selections of the parameters, $J_n$ is given by
\begin{equation}\label{eq:Jnrlocus}
J_n=
\begin{pmatrix}
\left(\dfrac{\alpha_1}{2} + \dfrac{1}{2}\right)^2 - \dfrac{5\, \alpha_1}{2} & - \dfrac{21}{2} & 8 & \dfrac{5}{2} & \dfrac{15}{8} \\[3mm]
-\dfrac{7\, \mathrm{a1}}{2} & - \dfrac{25}{2} & 8 & \dfrac{5}{2} & \dfrac{15}{8} \\[3mm]
- 4\, \mathrm{a1} & -12 & \dfrac{29}{4} & \dfrac{5}{2} & \dfrac{15}{8}\\[3mm]
-\dfrac{\mathrm{a1}}{2} & - \dfrac{3}{2} & 1 & - \dfrac{7}{2} & \dfrac{15}{8}\\[3mm]
\dfrac{5\, \mathrm{a1}}{4} & \dfrac{15}{4} & - \dfrac{5}{2} & - \dfrac{25}{4} & \dfrac{25}{16}
\end{pmatrix}.
\end{equation}
We can observe that negative roots are only present when $\alpha_{1}$ equals any value in $\{-3,2,5,-1.5\}$ with locations $0.25(1-\alpha_i^2)$, that is $-6,-2,-0.75$ and $-0.3125$. There are no other real negative roots for any value of $\alpha_1$.
\begin{figure}[!h]\begin{center}
\includegraphics[width=\textwidth]{rootloc.eps}\\
\caption{Root locus for varying $\alpha_1$} \label{fig:rootloc}
\end{center}
\end{figure}
\subsection{Another closed form solution}
We can use our results to conclude some other simple facts. For example, if $\alpha_1=-1$, we have
\begin{align}
&R_n(x)=\\\nonumber
&=4x(\sqrt{4x-1}-j)\prod_{i=2}^{n}\left(\sqrt{4x-1}+j\alpha_i\right)^2+4x(\sqrt{4x-1}+j)\prod_{i=2}^{n}\left(\sqrt{4x-1}-j\alpha_i\right)^2 \\\nonumber
&=4x\left((\sqrt{4x-1}-j)\prod_{i=2}^{n}\left(\sqrt{4x-1}+j\alpha_i\right)^2+(\sqrt{4x-1}+j)\prod_{i=2}^{n}\left(\sqrt{4x-1}-j\alpha_i\right)^2\right),
\end{align}
and the positive roots of the expression between parenthesis are the solutions of
\begin{align}
-\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right)+2\sum_{i=2}^{n}\arctan\left(\dfrac{\alpha_{i}}{\sqrt{4\lambda_k-1}}\right)=(2k-1)\dfrac{\pi}{2},
\end{align}
with $k\in\mathbb{Z}.$ This is the same expression as the one obtained by setting $\alpha_1=-1$ in \eqref{eq:arctan}. However, if $\alpha_i=-1$ for all $i$ we have
\begin{align}
(1-2n)\arctan\left(\dfrac{1}{\sqrt{4\lambda_k-1}}\right)=(2k-1)\dfrac{\pi}{2},
\end{align}
which only has solutions when $k\leq 0$. In such case we have that the remaining $n-1$ eigenvalues in this case are
\begin{align}
\lambda_k=\dfrac{1}{4}\left(1+\cot^2\left(\dfrac{1+2k}{2n-1}\dfrac{\pi}{2}\right)\right)=\dfrac{1}{4}\csc^2\left(\dfrac{1+2k}{2n-1}\dfrac{\pi}{2}\right),\quad k=0,\ldots,n-2.
\end{align}
A simple numerical example corroborates these values. If $k=n-1$ one can note that the resulting value for the above expression is $0.25$. According to the discussion on the previous subsection, these correspond to the roots of
\begin{align}
R_n(x)=&4x\left((\sqrt{1-4x}-1)^{2n-1}+(\sqrt{1-4x}+1)^{2n-1}\right),
\end{align}
which are connected in a straightforward manner to the roots of the unity. Indeed, by setting $u=\sqrt{1-4x}$, we have that the roots in this case are given by the solutions of
\begin{align}
\left(\dfrac{u-1}{u+1}\right)^{2n-1}=-1.
\end{align}
\subsection{Limit cases}
We finally briefly review some limit cases. If a single parameter satisfies $|\alpha_i|\rightarrow \infty$, then one eigenvalue tends to $+\infty,$ and the remaining $n-1$ eigenvalues tend to the eigenvalues of $J_{n-1}$ with $\alpha_{i}$ removed from $A$. This can be easily obtained from any of the expressions for the characteristic polynomial of $J_n$. We can see in Figure \ref{fig:rootloc} that some branches of the root locus must have a loop, when a single parameter moves from $-\infty$ to $+\infty$.
\section{Conclusion}\label{sec:conclusion}
We have obtained the characteristic polynomial of a structured $n\times n$ matrix with $n$ real parameters as the solution of a recursive relation. The matrix in question is fairly non trivial and for some selections of the parameters it corresponds to well studied cases. The provided framework allows to obtain a surprising amount of insight on the behaviour of the $n$ eigenvalues, in terms of the respective parameters. For example, we have shown that when they are positive, all the eigenvalues are also real and positive. Moreover, they can be computed by means of solving $n$ equations (one for each eigenvalue) involving inverse trigonometric functions with interesting properties. While the alternative method does not always allow for an exact computation of the eigenvalues, it does allow for approximations and the study of other behaviours in terms of the parameters.
A full characterization of possible negative and complex eigenvalues, when we allow for the presence of negative parameters, is missing. However, we provided some initial discussion and numerical simulations to illustrate the case. This was mostly possible due to our main results.
We believe that the class of matrices with a characteristic polynomial given as the sum of two complex conjugate polynomials, plays an important role in the study of certain optimal control problems and model order reduction techniques.
\section*{Acknowledgement}
The authors were supported by the Advanced Center for Electrical
and Electronic Engineering, Basal Project FB0008, and by the Grants FONDECYT 3160738 and 1161241 CONICYT Chile.
\bibliographystyle{abbrv}
|
\section{Introduction}
\setcounter{equation}{0}
The phenomena of coagulation and fragmentation
are studied in various contexts of natural sciences.
Mathematically, they are considered to be `dual' to
each other at least in some naive sense
or to be simply the time-reversal of each other.
Hence, one naturally expects that
the coagulation-fragmentation dynamics
may lead to a nontrivial equilibrium
in the course of time provided the occurrence of
coagulation and fragmentation is prescribed
in an appropriate manner.
Many authors have examined
such situations through a nonlinear equation
called often the coagulation-fragmentation equation.
It takes the form
\begin{eqnarray}
\frac{\partial}{\partial t}c(t,x)
& = &
\frac{1}{2}\int_0^x\left[K(y,x-y)c(t,y)c(t,x-y)
-F(y,x-y)c(t,x)\right]dy \nonumber \\
& &
-\int_0^{\infty}\left[K(x,y)c(t,x)c(t,y)
-F(x,y)c(t,x+y)\right]dy, \label{1.1}
\end{eqnarray}
where $t, x>0$, and
the functions $K$ and $F$ are supposed to be given,
nonnegative, symmetric and depending on the mechanisms
of coagulation and fragmentation, respectively.
In the literature $c(t,x)$ represents
the `density' of clusters of size $x$
(or particles with mass $x$) at time $t$
and the equation (\ref{1.1}) is derived
heuristically by some physical arguments
or rigorously for some restricted cases.
(Among results of the latter kind
for both nonzero $K$ and $F$,
we refer \cite{EW00}.)
However, (\ref{1.1}) is not complete
for the full description of coagulation-fragmentation phenomena
since it usually emerges after
certain contraction procedure such as `propagation of chaos'
or under intuitive assumptions of asymptotic
independence among distributions of clusters.
In this paper we study a hierarchical system of equations,
for a special case of which
we establish a direct connection with
an infinite-dimensional stochastic dynamics
incorporating coagulation and fragmentation.
For each $k\in \N:=\{1,2,\ldots\}$,
the $k$th equation of the hierarchy reads
\begin{eqnarray}
\lefteqn{
\frac{\partial}{\partial t}c_k(t,z_1,\ldots,z_k)}
\label{1.2} \\
&=&
\frac{1}{2}\sum_{l=1}^k\int_0^{z_l}K(y,z_l-y)
c_{k+1}(t,z_1,\ldots,z_{l-1},y,z_l-y,z_{l+1},\ldots,z_k)dy
\nonumber \\
& &
-\frac{1}{2}\sum_{l=1}^k\int_0^{z_l}F(y,z_l-y)dy \
c_{k}(t,z_1,\ldots,z_k)
\nonumber \\
& &
-\sum_{l=1}^k\int_0^{\infty}K(z_l,y)
c_{k+1}(t,z_1,\ldots,z_l,y,z_{l+1},\ldots,z_k)dy
\nonumber \\
& &
+\sum_{l=1}^k\int_0^{\infty}F(z_l,y)
c_{k}(t,z_1,\ldots,z_{l-1},z_l+y,z_{l+1},\ldots,z_k)dy
\nonumber \\
& &
-\one_{\{k\ge 2\}}\sum_{l<m}^{k}K(z_l,z_m)
c_{k}(t,z_1,\ldots,z_k)
\nonumber \\
& &
+\one_{\{k\ge 2\}}\sum_{l<m}^{k}F(z_l,z_m)
c_{k-1}(t,z_1,\ldots,z_{l-1},z_l+z_m,z_{l+1},\ldots,
z_{m-1},z_{m+1},\ldots,z_k), \nonumber
\end{eqnarray}
where $\one_E$ stands in general for the indicator function of
a set $E$ and the sum $\sum_{l<m}^{k}$ is taken over
pairs $(l,m)$ of integers such that $1\le l<m \le k$.
If the last two terms on the right side of (\ref{1.2})
were absent, it is readily checked that
the system of equations is satisfied by the direct products
$c^{\otimes k}(t,z_1,\ldots,z_k):=c(t,z_1)\cdots c(t,z_k)$
of a solution to (\ref{1.1}). The equations
(\ref{1.2}) are considered to be much more informative
in the sense that interactions among an
arbitrary number of clusters are took into account.
In fact, a finite-system version of (\ref{1.2})
has been discussed for a pure coagulation model
(i.e. the case $F\equiv 0$) by Escobedo and Pezzotti \cite{EP}.
Their derivation of (\ref{1.2}) with $F\equiv 0$
starts from a finite set of evolution equations satisfied
by the so-called mass probability functions
associated with a stochastic coagulation model
known as the Marcus-Lushnikov process.
As pointed out in \cite{EP} the situation is similar
to the derivation of the BBGKY hierarchy in classic
kinetic theory although the underlying microscopic dynamics
for the BBGKY hierarchy is not stochastic but deterministic.
There is an extensive literature discussing
a stochastic dynamics which serves
as a basis of an infinite system called the Boltzmann hierarchy,
which is a thermodynamic limit of the BBGKY hierarchy.
(See the monograph of Petrina \cite{Petrina} and the references therein.)
It should be mentioned also that a number of articles
have discussed stochastic interacting systems
of finite particles to derive kinetic equations,
a special case of which is (\ref{1.1}),
in the limit as the number of particles tends to infinity.
(See e.g. a paper by Eibeck and Wagner \cite{EW}
and the references therein. Also,
for a systematic treatment in a general framework
related to such issues, see
Kolokoltsov's monograph \cite{Kol}.)
We intend to explore
the `coagulation-fragmentation hierarchy' (\ref{1.2})
by dealing with stochastic infinite systems directly
and derive (\ref{1.1}) as a macroscopic equation
for them through the limit under proper rescaling.
Such a limit theorem is regarded as
the law of large numbers for measure-valued processes
and related to the propagation of chaos.
(See {\it Remarks} at the end of \S 4.1 below.)
In the case where the mechanisms of coagulation and fragmentation
together enjoy the detailed balance condition, i.e.,
\be
K(x,y)M(x)M(y)=F(x,y)M(x+y) \label{1.3}
\ee
for some function $M$, equilibrium behaviors of
the solution $c(t,x)$ to (\ref{1.1}) with respect to
a stationary solution of the form
$x\mapsto M(x)e^{-bx}$ have been studied by
many authors. In particular, for the equation with
$K$ and $F$ being positive constants,
Aizenman and Bak \cite{AB} carried out
detailed analysis such as a uniform rate for
strong convergence to equilibrium.
(See also Stewart and Dubovski \cite{SD}.)
Lauren{\c c}ot and Mischler \cite{LM} studied
that convergence
under certain assumptions for $K, F$ and $M$
and suitable conditions on the initial state.
Such results include particularly an $H$-theorem,
namely the existence of a Lyapunov functional
of entropy type for the solution.
In what follows,
we shall restrict the discussion to the case where
\be
K(x,y)=xy \Hh(x,y), \quad
F(x,y)= (x+y) \Hc(x,y) \label{1.4}
\ee
for some homogeneous functions $\Hh$ and $\Hc$ of
common degree $\lambda\ge 0$, namely,
\be
\Hh(ax,ay)=a^{\lambda}\Hh(x,y), \quad
\Hc(ax,ay)=a^{\lambda}\Hc(x,y) \quad (a, x , y> 0). \label{1.5}
\ee
To avoid trivialities, we suppose also
that $\Hh$ and $\Hc$ are not identically zero. Therefore,
both $K$ and $F$ are necessarily unbounded but
of polynomial growth at most.
In case $\theta\Hh=\Hc$ for a constant
$\theta>0$, (\ref{1.3}) holds for $M(x)=\theta/x$.
The coagulation and fragmentation phenomena
have been discussed also in the probability literature.
See e.g. Bertoin's monograph \cite{B} for systematic accounts of
stochastic models and random operations
describing either phenomenon.
The choice (\ref{1.4}) is mainly motivated by
a coagulation-fragmentation process
studied by Mayer-Wolf et al \cite{MWZZ} and Pitman \cite{P}.
These papers concern the case where
$\Hh$ and $\Hc$ are constants.
Having the infinite-dimensional simplex
\[
\Omega_1
=\{\bx=(x_i)_{i=1}^{\infty}:~ x_1\ge x_2 \ge \ldots \ge 0,
\ \sum_{i}x_i=1 \}
\]
as its state space, the process keeps
the total sum 1 of cluster sizes fixed.
The special case $\Hh=\Hc \equiv \mbox{const.}$
corresponds to the Markov process
explored in \cite{T1} and \cite{T2},
which is associated with `the simplest split-merge operator'
originally introduced by A. Vershik in the context of
analysis of the infinite dimensional symmetric group.
This model was studied extensively in \cite{DMWZZ}
in a deep and explicit connection with
a discrete analogue generated by the random transposition,
for which one may refer to \cite{Sch} for instance.
For that discrete model, the
coagulation and fragmentation rates (\ref{1.4}) with
both $\Hh$ and $\Hc$ being constants,
naturally emerge as transition probabilities.
(See (2.2) in \cite{DMWZZ}.)
In these works it was shown that the celebrated
Poisson-Dirichlet distribution with parameter $\theta$
is a reversible distribution of the process, and
much efforts were made to prove
the uniqueness of a stationary distribution.
In particular, Diaconis et al \cite{DMWZZ}
succeeded in proving it for $\theta=1$ by giving
an effective coupling result with the discrete
coagulation-fragmentation processes.
(See also Theorem 1.2 in \cite{Sch} and
Theorem 7.1 in \cite{Gold}.)
We also refer the reader to \cite{Be}
for another result of interest
on a unique stationary distribution for
the model evolving with a different class of
coagulation-fragmentations.
The coagulation and fragmentation
we will be concerned with are only binary ones.
Cepeda \cite{C2} constructed stochastic models
incorporating coagulation and multi-fragmentation
on a larger state space than that of our models
(, i.e., $\Omega$ defined below). See Introduction
and the references in \cite{C2} for previous
works and development in the study of
related stochastic models.
By virtue of the homogeneity assumption on $\Hh$ and $\Hc$
we can consider the generalized process
associated with (\ref{1.4}) not only on $\Omega_1$
but also on the infinite-dimensional cone
\[
\Omega
=\{\bz=(z_i)_{i=1}^{\infty}:~ z_1\ge z_2 \ge \ldots \ge 0,
\ 0<\sum_{i}z_i<\infty \}.
\]
In fact, the major arguments in this paper
exploit some ingredients from theory of point processes.
The idea is that each $\bz=(z_i)\in\Omega$
can be identified with the locally finite point-configuration
\[
\xi=\sum_{i}\one_{\{z_i>0\}}\delta_{z_i}
\]
on the interval $(0,\infty)$, where
$\delta_{z_i}$ is the delta distribution concentrated at $z_i$.
Indeed, in one of our main results,
the notion of correlation measures
will make us possible to reveal
an exact connection between
hierarchical equations (\ref{1.2}) and
the coagulation-fragmentation process with rates (\ref{1.4}).
As another result based on the point process calculus
we will present a class of
coagulation-fragmentation processes having
the Poisson-Dirichlet distributions or
certain variants (including the laws of gamma
point processes) as their reversible distributions,
clarifying what mathematical structures
are responsible for this result.
That structure will be described in terms of
correlation functions
together with Palm distributions,
certain conditional laws for the point process.
(See (\ref{2.14}) and (\ref{2.15}) below.)
We mention also that
the reversibility will play some key roles
in discussing the existence of strong solutions to
(\ref{1.2}). As for the original equation (\ref{1.1}),
introducing rescaled models
which depend on the scaling parameter $N$,
one can discuss its derivation from
the associated measure-valued processes
in which each point is assigned mass $1/N$.
Such a result is formulated as a limit theorem
for Markov processes as $N\to\infty$
and one of the key steps is
to show the tightness of their laws,
which is far from trivial
because there is less restriction on
grows order of $K$ and $F$.
As will turn out later our setting of
the degrees of homogeneity plays
an essential role to overcome difficulties of this sort.
The organization of this paper is as follows.
In Section 2, we introduce
the coagulation-fragmentation process
associated with rates (\ref{1.4})
and give an equivalent description of the model
in terms of the corresponding point process.
Section 3 discusses
a weak version of (\ref{1.2}),
which will turn out to be satisfied by the
correlation measures of
our coagulation-fragmentation process.
After some preliminary arguments are made
for rescaled models in Section 4,
a solution to (\ref{1.1}) will be obtained
from properly rescaled empirical measures
in Section 5.
\section{The coagulation-fragmentation process
associated with split-merge transformations
\setcounter{equation}{0}
\subsection{Definition of the models}
As mentioned in Introduction
the rates $K$ and $F$ are supposed to
be of the form (\ref{1.4}) with $\Hh$ and $\Hc$ being
homogeneous functions of degree $\lambda\ge0$ throughout.
Notice that this is equivalent to the condition that
$K$ and $F$ are homogeneous functions
of degree $\lambda+2$ and $\lambda+1$,
respectively, for some $\lambda\ge 0$.
As far as coagulation rates are concerned,
the homogeneity, though mathematically a strong condition,
is satisfied typically by examples of kernels
used in the physical literature as seen in Table 1 in \cite{A}
(although our framework excludes any of such examples).
See also \cite{FL} and \cite{C},
which discuss the equation with homogeneous(-like) $K$.
In the rest, the following two conditions are
also imposed without mentioning: \\
(H1) $\Hh$ is a symmetric, nonnegative
measurable function on $(0,\infty)^2$ such that
\[
\Ch:=\sup\{\Hh(u,1-u) :~0<u<1 \}\in(0,\infty).
\]
(H2) $\Hc$ is a symmetric, nonnegative
measurable function on $(0,\infty)^2$ such that
\[
\Cc:= \int_0^1 \Hc(u,1-u)du \in(0,\infty).
\]
In general, a homogeneous function $H$
on $(0,\infty)^2$ is determined by
its degree $\lambda$ and the function
$h(u):=H(u,1-u)$ on $(0,1)$ through the relation
$H(x,y)=(x+y)^{\lambda}h(\frac{x}{x+y})$.
As for the fragmentation rate,
the homogeneity (\ref{1.5}) combined
with (\ref{1.4}) implies that the overall rate of
fragmentation of an $x$-sized cluster is
necessarily given by the power-law form:
\[
\frac{1}{2}\int_0^xF(y,x-y)dy
= \frac{1}{2}x^2\int_0^1\Hc(ux,(1-u)x)du
= \frac{\Cc}{2}x^{2+\lambda}.
\]
Such a situation is featured by the
coagulation-fragmentation equation
studied in \cite{BLL}, whose conditions
for coagulation rates are also well adapted to our setting.
(See {\it Remark} at the end of Section 5
for related discussions.)
Also, \cite{VZ} examined the interplay between
degrees of coagulation and fragmentation
in the context of stability analysis.
At the beginning of Section 4 we will
mention another role of such interplay
between them in the study of rescaled processes.
\medskip \\
\noindent
{\it Examples.}~
(i) Consider $H(x,y)=(xy)^{a}(x^b+y^b)^c$,
for which $\lambda=2a+bc$.
Without loss of generality, we can suppose that $b\ge 0$. Then
$H$ satisfies (H1) (resp. (H2)) if $a\ge 0$ (resp. $a>-1$). \\
(ii) Let $b>0$ and define ${H}(x,y)=(xy)^{a}|x^b-y^b|^c$,
for which $\lambda=2a+bc$.
${H}$ satisfies (H1) (resp. (H2))
if $a\ge 0$ and $c\ge 0$ (resp. $a>-1$ and $bc>-1$).
Indeed, it is readily observed that
${H}(u,1-u)/u^a \to 1 (u \downarrow 0)$
and
${H}(u,1-u)/|2u-1|^{bc} \to b2^{-(2a+bc)}
(u \to 1/2)$. \\
(iii) Given $\lambda\ge 0$, let
${H}(x,y)=(x^a+y^a)(x^{\lambda-a}+y^{\lambda-a})$.
It follows that (H1) (resp. (H2)) is satisfied
if $0\le a\le \lambda$ (resp. $-1<a<\lambda+1$). \\
(iv) An example of discontinuous homogeneous function
of degree $\lambda$ is
$H(x,y)=x^{\lambda}\one_{\{x\ge ay\}}
+y^{\lambda}\one_{\{y\ge ax\}}$, where $a>0$.
Another one is $H(x,y)=(xy)^{\lambda/2}
\one_{[a,1/a]}(x/y)$ with $0<a<1$.
For each example and any $\lambda\ge 0$,
both (H1) and (H2) hold.
\medskip
In order to define our coagulation-fragmentation process
as a continuous-time Markov process on $\Omega$,
suitable modifications to the formulation
in \cite{MWZZ} are made in the following manner.
Let $\Omega$ be equipped with the product topology
and $B(\Omega)$ (resp. $B(\Omega_1)$) be the Banach space of
bounded Borel functions on $\Omega$ (resp. $\Omega_1$)
with the sup norm $\Vert \cdot \Vert_{\infty}$.
For $\bz=(z_i)\in \Omega$ put $|\bz|=\sum z_i$.
A useful inequality is $\sum z_i^{1+a}\le |\bz|^{1+a}$
for any $a\ge 0$, which is implied by
$\sum (z_i/|\bz|)^{1+a}\le \sum (z_i/|\bz|) =1$.
Define a bounded linear operator $\wt{L}$ on $B(\Omega)$ by
\begin{eqnarray}
\wt{L}\Phi(\bz)
& = &
\frac{1}{2|\bz|^{2+\lambda}}\sum_{i\ne j}K(z_i,z_j)
\left(\Phi(M_{ij}\bz)-\Phi(\bz)\right) \nonumber \\
& & +\frac{1}{2|\bz|^{2+\lambda}}\sum_i
\int_0^{z_i}dy F(y,z_i-y)\left(\Phi(S_i^{(y)}\bz)-\Phi(\bz)\right), \label{2.1}
\end{eqnarray}
where $M_{ij}\bz\in\Omega$
(resp. $S_i^{(y)}\bz\in\Omega$)
is obtained from a sequence $\bz=(z_k)$ by merging
$z_i$ and $z_j$ into $z_i+z_j$
(resp. by splitting $z_i$ into $y$ and $z_i-y$)
and then by reordering. Noting that
$M_{ij}\bz=\bz$ (resp. $S_i^{(y)}\bz=\bz$)
whenever $z_iz_j=0$ (resp. $z_i=0$), we regard
the sum $\sum_{i\ne j}$ (resp. $\sum_i$) in (\ref{2.1})
as the sum taken over
$i\ne j$ (resp $i$) such that $z_iz_j>0$ (resp. $z_i>0$).
We adopt such convention
for the same kind of expression throughout.
The boundedness of $L$ is seen
easily from (H1) and (H2) in view of
alternative expression for (\ref{2.1})
\begin{eqnarray}
\wt{L}\Phi(\bz)
& = & \frac{1}{2}\sum_{i\ne j}
\frac{z_i}{|\bz|}\cdot\frac{z_j}{|\bz|}
\left(\frac{z_i+z_j}{|\bz|}\right)^{\lambda}
\Hh\left(\frac{z_i}{z_i+z_j},\frac{z_j}{z_i+z_j}\right)
\left(\Phi(M_{ij}\bz)-\Phi(\bz)\right) \nonumber \\
& & +\frac{1}{2}\sum_i
\left(\frac{z_i}{|\bz|}\right)^{2}
\left(\frac{z_i}{|\bz|}\right)^{\lambda}
\int_0^{1}du \Hc(u,1-u)\left(\Phi(S_i^{u}\bz)-\Phi(\bz)\right), \label{2.2}
\end{eqnarray}
where $S_i^{u}\bz:=S_i^{(uz_i)}\bz$.
Indeed, it follows that
$\Vert \wt{L}\Phi\Vert_{\infty}\le(\Ch\vee\Cc)
\Vert \Phi\Vert_{\infty}$.
Here and in what follows we use the notation
$a \vee b := \max\{a,b\}$ and $a \wedge b := \min\{a,b\}$.
The standard argument (e.g., \S 2 of Chapter 4 in \cite{EK})
shows that $ \wt{L}$ generates
a continuous-time Markov jump process
$\{\wt{Z}(t)=(\wt{Z}_i(t))_{i=1}^{\infty}:~t\ge 0\}$, say, on $\Omega$.
It is obvious that $|\wt{Z}(t)|=|\wt{Z}(0)|$
for all $t\ge 0$ a.s.
Similarly, $L_1$, the restriction of $\wt{L}$ on $B(\Omega_1)$, namely,
\begin{eqnarray}
L_1\Phi(\bx)
& = &
\frac{1}{2}\sum_{i\ne j}K(x_i,x_j)
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right) \nonumber \\
& & +\frac{1}{2}\sum_i
\int_0^{x_i}dy F(y,x_i-y)\left(\Phi(S_i^{(y)}\bx)-\Phi(\bx)\right) \nonumber \\
& = &
\frac{1}{2}\sum_{i\ne j} x_ix_j \Hh(x_i,x_j)
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right) \nonumber \\
& & +\frac{1}{2}\sum_i x_i^{2+\lambda}
\int_0^{1}du \Hc(u,1-u)\left(\Phi(S_i^{u}\bx)-\Phi(\bx)\right) \label{2.3}
\end{eqnarray}
generates a continuous-time Markov process
$\{X(t)=(X_i(t))_{i=1}^{\infty}:~t\ge 0\}$ on $\Omega_1$.
In the case where $\Hh\equiv 1$ and $\lambda=0$,
(\ref{2.3}) can be thought of as the generator of
continuous-time version of a Markov chain studied in
\cite{MWZZ}. Moreover, the operator
(\ref{2.3}) is a special case (more specifically,
the binary fragmentation case) of
the generator considered in \cite{C2}, although
in order for the model to be defined also on $\Omega$
we need homogeneity of $K$, whereas in \cite{C2}
certain continuity of $K$ is imposed.
For each $a>0$, define the dilation map
$D_a:\Omega \to \Omega $ by $D_a(\bz)=a\bz:=(az_i)$.
The relationship between $Z(t)$ and $X(t)$
described in the following lemma is fundamental.
\begin{lm
(i) Suppose that a process $\{\wt{Z}(t):~t\ge 0\}$
generated by $\wt{L}$ is given. Then
the $\Omega_1$-valued process
$\{X(t)=(X_i(t))_{i=1}^{\infty}:~t\ge 0\}$
defined by
$X_i(t)=\wt{Z}_i(t)/|\wt{Z}(t)|=\wt{Z}_i(t)/|\wt{Z}(0)|$
is a process generated by $L_1$. \\
(ii) Suppose that a $(0,\infty)$-valued random variable $V$
and a process $\{(X_i(t))_{i=1}^{\infty}:~t\ge 0\}$
generated by $L_1$ are mutually independent. Then
the $\Omega$-valued process
$\{\wt{Z}(t)=(\wt{Z}_i(t))_{i=1}^{\infty}:~t\ge 0\}$ defined
by $\wt{Z}_i(t)=VX_i(t)$ is a process generated by $\wt{L}$.
\end{lm
{\it Proof.}~
(i) Take $\Phi\in B(\Omega)$ arbitrarily
and denote by $E[\ \cdot \ |Z(0)=\bz]$
the expectation with respect to the process
generated by $L$ starting from $\bz\in\Omega$.
Since $v^{-1}(M_{ij}\bz)=M_{ij}(v^{-1}\bz)$ and
$|\bz|^{-1}(S_{i}^{(uz_i)}\bz)=S_{i}^u(|\bz|^{-1}\bz)$,
we see from (\ref{2.2})
\[
\wt{L}(\Phi\circ D_{1/v})(\bz)=(L_1\Phi)(v^{-1}\bz),
\quad \bz\in \Omega, v=|\bz|.
\]
Hence, for any $t>0$, by Fubini's theorem
\begin{eqnarray*}
\lefteqn{E\left[\Phi(X(t))\right]
-E\left[\Phi(X(0))\right]} \\
& = &
E\left[\Phi(|\wt{Z}(0)|^{-1}\wt{Z}(t))\right]
-E\left[\Phi(|\wt{Z}(0)|^{-1}\wt{Z}(0))\right] \\
& = &
\int_{\Omega}P(\wt{Z}(0)\in d\bz)
\left\{E\left[(\Phi\circ D_{1/|\bz|})
(\wt{Z}(t))|\wt{Z}(0)=\bz\right]
-(\Phi\circ D_{1/|\bz|})(\bz)\right\} \\
& = &
\int_{\Omega}P(\wt{Z}(0)\in d\bz)
\int_0^tdsE\left[\wt{L}(\Phi\circ D_{1/|\bz|})(\wt{Z}(s))|
\wt{Z}(0)=\bz\right] \\
& = &
\int_0^tds\int_{\Omega}P(\wt{Z}(0)\in d\bz)
E\left[L_1\Phi(|\bz|^{-1}\wt{Z}(s))|\wt{Z}(0)=\bz\right] \\
& = &
\int_0^tds E\left[L_1\Phi(|\wt{Z}(0)|^{-1}\wt{Z}(s))\right]
\ = \
\int_0^tds E\left[L_1\Phi(X(s))\right].
\end{eqnarray*}
This proves the first assertion. \\
(ii) Based on the relation
$L_1(\Phi\circ D_v)(\bx)=\wt{L}\Phi(v\bx)$
for $\bx\in \Omega_1$ and $v>0$, the proof of the second assertion is very similar to that for (i). So we omit it. \qed
\medskip
We call $V$ in Lemma 2.1 a lifting variable.
Roughly speaking, lifting a process on $\Omega_1$ generated by $L_1$
yields a process on $\Omega$ generated by $\wt{L}$.
We need to consider an unbounded operator
${L}\Phi(\bz)=|\bz|^{2+\lambda} \wt{L}\Phi(\bz)$ or
\begin{eqnarray}
{L}\Phi(\bz)
& = &
\frac{1}{2}\sum_{i\ne j}K(z_i,z_j)
\left(\Phi(M_{ij}\bz)-\Phi(\bz)\right)
\nonumber \\
& &
+\frac{1}{2}\sum_i\int_0^{z_i}dy F(y,z_i-y)
\left(\Phi(S_i^{(y)}\bz)-\Phi(\bz)\right). \label{2.4}
\end{eqnarray}
The corresponding process
$\{Z(t):~t\ge 0\}$ on $\Omega$ can be obtained from
a process $\{\wt{Z}(t):~t\ge 0\}$ generated by $\wt{L}$
with the same initial law by a random time-change
\[
{Z}(t):=\wt{Z}(|\wt{Z}(0)|^{2+\lambda}t).
\]
This can be shown by general theory of Markov processes,
e.g., Theorem 1.3 in Chapter 6 of \cite{EK}, or
more directly, by making the following observation:
for any $\Phi\in B(\Omega)$ such that
$L\Phi\in B(\Omega)$,
by the optional sampling theorem
\begin{eqnarray*}
\lefteqn{
\Phi(Z(t))-\Phi(Z(0))
-\int_0^tdu L\Phi({Z}(u)) } \\
& = &
\Phi(\wt{Z}(|\wt{Z}(0)|^{2+\lambda}t))-\Phi(\wt{Z}(0))
-\int_0^tdu|\wt{Z}(0)|^{2+\lambda}
\wt{L}\Phi(\wt{Z}(|\wt{Z}(0)|^{2+\lambda}u)) \\
& = &
\Phi(\wt{Z}(|\wt{Z}(0)|^{2+\lambda}t))-\Phi(\wt{Z}(0))
- \int_0^{|\wt{Z}(0)|^{2+\lambda}t}ds\wt{L}\Phi(\wt{Z}(s))
\end{eqnarray*}
is a martingale.
\subsection{Reformulation in terms of point processes}
We proceed to reformulate
the above-mentioned processes as Markov processes
taking values in a space of point-configurations on $(0,\infty)$.
To discuss in the setting of point processes,
we need the following notation.
Set $\Zp=\{0,1,2,\ldots\}$ and let $\cN$ be
the set of $\Zp$-valued Radon measures
on $(0,\infty)$. Each element $\eta$ of $\cN$
is regarded as a counting measure associated with
a locally finite point-configuration on $(0,\infty)$ with multiplicity.
We equip $\cN$ with the vague topology and
use the notation $|\eta|:=\int_{(0,\infty)}v\eta(dv)$
for $\eta\in\cN$ and
$\lg f,\nu\rg:=\int_{(0,\infty)}f(v)\nu(dv)$
for a measure $\nu$ and a Borel function
$f$ on $(0,\infty)$.
Denote by $B_+(S)$ the set of nonnegative
bounded Borel measurable functions
on a topological space $S$. For simplicity,
we set $B_+=B_+((0,\infty))$ and
use the notation $B_+^k$ instead of $B_+((0,\infty)^k)$
for $k=2,3,\ldots$.
As mentioned roughly in Introduction,
the subsequent argument is based on
the one-to-one correspondence between
$\bz=(z_i)\in \Omega$ and
\[
\Xi(\bz):=\sum_i\one_{\{z_i>0\}}\delta_{z_i}
\in \{\eta\in\cN:~|\eta|<\infty\}=:\cN_1.
\]
Clearly, if $\eta=\Xi(\bz)$, then $|\eta|=|\bz|$ and
$\eta([\epsilon,\infty))\le |\bz|/\epsilon$
for $\epsilon>0$. Note that the map
$\Xi:\Omega\to \cN_1$
is bi-measurable. It follows that
\be
\Xi(M_{ij}\bz)-\Xi(\bz)
=\delta_{z_i+z_j}-\delta_{z_i}-\delta_{z_j}
\quad \mbox{if} \quad z_i,z_j>0,
\label{2.5}
\ee
and
\be
\Xi(S_{i}^{(y)}\bz)-\Xi(\bz)
=\delta_{y}+\delta_{z_i-y}-\delta_{z_i}
\quad \mbox{if} \quad z_i>y>0.
\label{2.6}
\ee
Thus, employing $\Xi(\bz)$ rather than $\bz$ itself
enables us to avoid an unnecessary complication arising
from reordering of the sequence.
Besides, owing to the map $\Xi$,
the arguments below
make use of some effective tools
in theory of point processes,
such as correlation measures and
(reduced) Palm distributions.
(See e.g. \S 13.1 of \cite{DV2} for general accounts.)
We shall give their definitions
for a locally finite point process
$\xi=\sum\delta_{Z_i}$ on $(0,\infty)$.
In what follows, the domain of integration will be
suppresed as long as it is $(0,\infty)^k$ for some $k\in\N$,
which should be clear from context.
The first correlation measure $q_1$ is the mean measure of $\xi$,
and for $k=2,3,\ldots$ the $k$th correlation measure $q_k$
is the mean measure of the modified product counting measure
\be
\xi^{[k]}:=\sum_{i_1,\ldots,i_k(\ne)}
\delta_{(Z_{i_1},\ldots,Z_{i_k})} \label{2.7}
\ee
on $(0,\infty)^k$, where $\sum_{i_1,\ldots,i_k(\ne)}$ indicates that
the sum is taken over $k$-tuplets $(i_1,\ldots,i_k)$
such that $i_l\ne i_m$ whenever $l\ne m$.
The entire system $\{q_1,q_2,\ldots\}$ of
correlation measures determines uniquely the law of $\xi$.
(The identity (\ref{2.13}) below is
the structure underlying this fact.)
The density of $q_k$ is called
the $k$th correlation function of $\xi$ if it exists.
Furthermore, letting
$\{P_{z_1,\ldots,z_k}:~z_1,\ldots,z_k\in (0,\infty)\}$
be a family of Borel probability measures on $\cN$,
we call $P_{z_1,\ldots,z_k}$
the $k$th-order reduced Palm distribution of $\xi$
at $(z_1,\ldots,z_k)$ if
for any $G\in B_+((0,\infty)^k\times \cN)$
\begin{eqnarray}
\lefteqn{E\left[\sum_{i_1,\ldots,i_k(\ne)}
G(Z_{i_1},\ldots,Z_{i_k},\xi)\right]
= E\left[\int \xi^{[k]}(dz_1\cdots dz_k)
G(z_1,\ldots,z_k, \xi)\right]} \nonumber \\
& = &
\int q_k(dz_1\cdots dz_k) E_{z_1,\ldots,z_k}
\left[G(z_1,\ldots,z_k,\eta+\delta_{z_1}+\cdots+\delta_{z_k})\right].
\label{2.8}
\end{eqnarray}
Here and throughout, $E_{z_1,\ldots,z_k}$ denotes the expectation
with respect to $P_{z_1,\ldots,z_k}$, so that
\[
E_{z_1,\ldots,z_k}\left[\Psi(\eta)\right]
=
\int_{\cN}\Psi(\eta)P_{z_1,\ldots,z_k}(d\eta),
\quad \Psi\in B_+(\cN).
\]
(Rigorously speaking, certain measurability
of $P_{z_1,\ldots,z_k}$ in $(z_1,\ldots,z_k)$ is
required just as in the definition of
regular conditional distributions.
See Proposition 13.1.IV of \cite{DV2} for the case $k=1$.)
An intuitive interpretation for $P_{z_1,\ldots,z_k}$
is the conditional law of
$\xi-\delta_{z_1}-\cdots-\delta_{z_k}$
given $\xi(\{z_1\})\cdots\xi(\{z_k\})>0$.
We call an equality of the type (\ref{2.8})
the $k$th-order Plam formula for $\xi$.
To state the main result of this section,
let us recall some known results on
Poisson-Dirichlet point process,
which is by definition the point-configuration associated with
a random element of $\Omega_1$
distributed according to the Poisson-Dirichlet distribution.
For each $\theta>0$,
let ${\rm PD}(\theta)$ denote the Poisson-Dirichlet
distribution with parameter $\theta$.
(See e.g. \cite{King}, \cite{B}, \cite{Feng}
for the definition.) Suppose that $X=(X_i)$ is
a random element of $\Omega_1$
whose law is ${\rm PD}(\theta)$
and consider the associated point process
$\xi^{(\theta)}=\sum\delta_{X_i}$ on the interval $(0,1)$,
which we call simply the ${\rm PD}(\theta)$ process.
It was shown in \cite{W} that
the $k$th correlation function of $\xi^{(\theta)}$
takes the form
\be
(x_1,\ldots,x_{k}) \mapsto
\frac{\theta^k}{x_1\cdots x_k}
\left(1-
\sum_{j=1}^{k}x_{j}\right)^{\theta-1}
\one_{\Delta_k^{\circ}}(x_1,\ldots,x_{k}),
\label{2.9}
\ee
where
\[
\Delta_k^{\circ}=\left\{(x_1,\ldots,x_{k}):
x_1,\ldots,x_k>0, \
x_1+\cdots+x_k<1 \right\}.
\]
By a special case of Corollary 1 of \cite{Le}
the $k$th-order reduced Palm distribution
of $\xi^{(\theta)}$
at $(x_1,\ldots,x_{k})\in\Delta_k^{\circ}$ coincides with
the law of
$\xi^{(\theta)}\circ D_{1-x_1-\cdots-x_{k}}^{-1}$.
Moreover, such a self-similar property of Palm distributions
characterizes the family $\{{\rm PD}(\theta):~\theta>0\}$
as was shown in Theorem 2 (iii) of \cite{Le}.
\medskip
The next lemma shows how the aforementioned
notions are well adapted
for dealing with our generators. Denote by $B_c$ the totality of
bounded Borel functions on $(0,\infty)$ with
compact support and let $B_{+,c}=B_+\cap B_c$.
The support and the sup norm of a function $f$ are
denoted by ${\rm supp}(f)$ and $\Vert f\Vert_{\infty}$, respectively.
We now give a class of functions on $\Omega$
for which our generators act in a tractable manner. Set
\[
\widetilde{B}_{+}=\{\phi\in B_{+}:
~\phi-1\in B_c, \Vert \phi-1\Vert_{\infty}<1\}.
\]
For each $\phi\in \widetilde{B}_{+}$
we can define a function $\Pi_{\phi}$ on $\Omega$ by
$\Pi_{\phi}(\bz)=\prod_{i:z_i>0}\phi(z_i)$,
noting that the right side is a finite product.
By abuse of notation, we also write
$\Pi_{\phi}(\xi)$ for $\Pi_{\phi}(\bz)$
when $\xi=\Xi(\bz)$. Thus
\be
\Pi_{\phi}(\xi)
= \exp\left(\sum_i\one_{\{z_i>0\}}\log \phi(z_i) \right)
= e^{\lg \log \phi,\xi\rg}. \label{2.10}
\ee
An important remark is that the class
$\{\Pi_{\phi}:~\phi\in\widetilde{B}_{+}\}$ is
measure-determining on $\Omega$
because it includes all functions of the form
$\bz \mapsto \exp(-\lg f,\Xi(\bz)\rg)$ with $f\in B_{+,c}$.
Given a function on $\Omega$, we regard it also as
a function on $\Omega_{\le R}:=\{\bz\in\Omega:~|\bz|\le R\}$
for any $R>0$.
It is clear that $L$ restricted on $B(\Omega_{\le R})$ is bounded.
\begin{lm
Let $\phi\in \widetilde{B}_{+}$ and
set $f=\phi-1$. Then for any $\bz\in\Omega$
\be
0\le \Pi_{\phi}(\bz)
\le
(1-\Vert f\Vert_{\infty})^{-\xi({\rm supp}(f))}, \label{2.11}
\ee
where $\xi=\Xi(\bz)$. Moreover,
$\Pi_{\phi}\in B(\Omega_{\le R})$ for any $R>0$ and
\begin{eqnarray}
L\Pi_{\phi}(\bz)
& = &
\frac{1}{2}\sum_{i\ne j}K(z_i,z_j)
\left[\phi(z_i+z_j)-\phi(z_i)\phi(z_j)\right]
\prod_{k\ne i,j}\phi(z_k) \label{2.12} \\
& &
+\frac{1}{2}\sum_i \int_0^{z_i}dy F(y,z_i-y)
\left[\phi(y)\phi(z_i-y)-\phi(z_i)\right]
\prod_{k\ne i}\phi(z_k) \nonumber \\
& = &
\frac{1}{2}
\int\xi^{[2]}(dv_1dv_2)K(v_1,v_2)
\left[\phi(v_1+v_2)-\phi(v_1)\phi(v_2)\right]
\Pi_{\phi}(\xi-\delta_{v_1}-\delta_{v_2}) \nonumber \\
& &
+\frac{1}{2}\int\xi(dv)\int_0^vdy F(y,v-y)
\left[\phi(y)\phi(v-y)-\phi(v)\right]
\Pi_{\phi}(\xi-\delta_{v}). \nonumber
\end{eqnarray}
\end{lm
{\it Proof.}~
Let $\phi(0)=1$ and $f(0)=0$ by convention.
To prove (\ref{2.11}) observe that
\be
\prod_{i}\phi(z_i)
=
\prod_{i}(1+f(z_i))
=
1+\sum_{k=1}^{\infty}\frac{1}{k!}
\sum_{i_1,\ldots,i_k(\ne)}f(z_{i_1})\cdots f(z_{i_k}). \label{2.13}
\ee
So, using the notation
${\alpha \choose k}
=\alpha(\alpha-1)\cdots(\alpha-k+1)/k!$, we get
\begin{eqnarray*}
0 \le \Pi_{\phi}(\bz)
& \le &
1+\sum_{k=1}^{\infty}\frac{\Vert f\Vert_{\infty}^k}{k!}
\sum_{i_1,\ldots,i_k(\ne)}
\one_{\{z_{i_1},\ldots,z_{i_k}\in{\rm supp}(f)\}} \\
& = &
1+\sum_{k=1}^{\infty}\Vert f\Vert_{\infty}^k
{\xi({\rm supp}(f)) \choose k} \\
& = &
(1-\Vert f\Vert_{\infty})^{-\xi({\rm supp}(f))},
\end{eqnarray*}
and thus (\ref{2.11}) follows.
Putting
$\epsilon=\inf{\rm supp}(f)>0$, we get
\[
\xi({\rm supp}(f))\le \xi([\epsilon,\infty))
\le \epsilon^{-1}|\bz|.
\]
This combined with (\ref{2.11}) implies that $\Pi_{\phi}$
is bounded on $\Omega_{\le R}$. (\ref{2.12}) is verified by
direct calculations with the help of
(\ref{2.5}) and (\ref{2.6}). \qed
\medskip
\subsection{Reversible cases}
We demonstrate the power of
point process calculus by proving
an extension of the reversibility result
due to Mayer-Wolf et al \cite {MWZZ}.
(As for the stationarity result, a proof
based on the underlying Poisson process
can be found in \S 7.3 of \cite{Gold}.)
To this end, we recall that
the $k$th correlation function $q_{k}$ of the
${\rm PD}(\theta)$ process $\xi^{(\theta)}$
is given in (\ref{2.9})
and that the $k$th-order reduced Palm distribution
$P_{x_1,\ldots,x_k}$ of $\xi^{(\theta)}$ at
$(x_1,\ldots,x_k)\in \Delta_k^{\circ}$ is
the law of $\xi^{(\theta)}\circ D_{1-x_1-\cdots-x_k}^{-1}$.
In particular, for any $(x_1,x_2)\in\Delta_2^{\circ}$
\be
x_1x_2q_2(x_1,x_2)=\theta (x_1+x_2)q_1(x_1+x_2) \label{2.14}
\ee
and
\be
P_{x_1,x_2}=P_{x_1+x_2}. \label{2.15}
\ee
As will be shown in the next theorem,
these identities are responsible
for the reversibility of ${\rm PD}(\theta)$
with respect to processes associated with bounded operators
on $B(\Omega_1)$ of the form
\begin{eqnarray*}
L_1^{(Q,\theta)}\Phi(\bx)
& = &
\frac{1}{2}\sum_{i\ne j} x_ix_j Q(x_i,x_j)
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right) \nonumber \\
& & +\frac{\theta}{2}\sum_i x_i^{2}
\int_0^{1}du Q(ux_i,(1-u)x_i)
\left(\Phi(S_i^{u}\bx)-\Phi(\bx)\right),
\end{eqnarray*}
where $Q$ is any nonzero bounded,
symmetric nonnegative function on
$\{(x,y) |~x,y> 0, x+y \le 1\}$.
(It should be noted that Theorem 12 in \cite{MWZZ}
proved essentially the symmetry of $L_1^{(Q,\theta)}$ with
$Q\equiv \mbox{const.}$ with respect to
${\rm PD}(\theta)$.) More generally, we consider
\begin{eqnarray}
L_1^{\sharp}\Phi(\bx)
& = &
\frac{1}{2}\sum_{i\ne j}K_1(x_i,x_j)
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right) \nonumber \\
& & +\frac{1}{2}\sum_i x_i
\int_0^{1}du F_1(ux_i,(1-u)x_i)
\left(\Phi(S_i^{u}\bx)-\Phi(\bx)\right)
\label{2.16}
\end{eqnarray}
with $K_1$ and $F_1$ being symmetric nonnegative
functions on $\{(x,y) |~x,y>0, x+y \le 1\}$ such that
$(x_i)\mapsto
\sum_{i\ne j}K_1(x_i,x_j)\one_{\{x_ix_j>0\}}$ and
$(x_i)\mapsto \sum x_i \int_0^{1}du F_1(ux_i,(1-u)x_i)$
are bounded functions on $\Omega_1$. We may and do
suppose that $K_1(x,y)=0$ whenever $xy=0$.
\begin{th
(i) Let $X=(X_i)_{i=1}^{\infty}$ be a random element
of $\Omega_1$ and suppose that
the first and second correlation functions
$q_1$ and $q_2$ of $\xi:=\Xi(X)$ exist and satisfy
\be
K_1(x_1,x_2)q_2(x_1,x_2)P_{x_1,x_2}
=F_1(x_1,x_2)q_1(x_1+x_2) P_{x_1+x_2},
\label{2.17}
\ee
$\mbox{a.e.-}(x_1,x_2)\in \Delta_2^{o}$.
Here, $P_{x_1,x_2}$ and $P_{x_1+x_2}$ are
the reduced Palm distributions of $\xi$ and
the above equality is understood as the one
between two measures on $\cN$. Then
the process generated by $L_1^{\sharp}$ is
reversible with respect to the law of $X$.
\\
(ii) ${\rm PD}(\theta)$ is a reversible distribution
of the process generated by $L_1^{(Q,\theta)}$.
\end{th
{\it Proof.}~
(i) Since $L_1^{\sharp}$ is bounded,
we only have to check the symmetry
\[
E\left[\Phi(X)L_1^{\sharp}\Psi(X)\right]
=E\left[\Psi(X)L_1^{\sharp}\Phi(X)\right]
\]
for any $\Phi,\Psi\in B(\Omega_1)$.
We will prove stronger equalities
\begin{eqnarray*}
\lefteqn{\lg\lg\Phi,\Psi\rg\rg_{\rm coag}
\ := \
E\left[\Phi(X)\sum_{i\ne j}K_1(X_i,X_j)\Psi(M_{ij}X)\right]}
\\
& = &
E\left[\sum_{i}X_i\int_0^1duF_1(uX_i,(1-u)X_i)
\Phi(S_{i}^uX)\Psi(X)\right]
\ =: \
{}_{\rm frag}\lg\lg\Phi,\Psi\rg\rg,
\end{eqnarray*}
which are regarded as a sort of
coagulation-fragmentation duality.
Furthermore, it suffices to verify for functions
$\Phi=\Pi_{\phi}$ and $\Psi=\Pi_{\psi}$
with $\phi,\psi\in \widetilde{B}_+$. Thanks to the
first-order and second-order Palm formulae for $\xi=\sum\one_{\{X_i>0\}}\delta_{X_i}$,
similar calculations to (\ref{2.12}) show that
\begin{eqnarray*}
\lefteqn{
\lg\lg\Pi_{\phi},\Pi_{\psi}\rg\rg_{\rm coag}} \\
& = &
E\left[\sum_{i\ne j}K_1(X_i,X_j)\phi(X_i)\phi(X_j)
\psi(X_i+X_j)\prod_{k\ne i,j}(\phi(X_k)\psi(X_k))\right] \\
& = &
E\left[\int\xi^{[2]}(dx_1dx_2)K_1(x_1,x_2)\phi(x_1)\phi(x_2)
\psi(x_1+x_2)\Pi_{\phi\psi}(\xi-\delta_{x_1}-\delta_{x_2})\right] \\
& = &
\int_{\Delta_{2}^{\circ}}q_2(x_1,x_2)
K_1(x_1,x_2)\phi(x_1)\phi(x_2)\psi(x_1+x_2)
E_{x_1,x_2}\left[\Pi_{\phi\psi}(\eta)\right]dx_1dx_2
\end{eqnarray*}
and
\begin{eqnarray*}
\lefteqn{
{}_{\rm frag}\lg\lg\Pi_{\phi},\Pi_{\psi}\rg\rg} \\
& = &
E\left[\sum_{i}X_i\int_0^1du F_1(uX_i,(1-u)X_i)
\phi(uX_i)\phi((1-u)X_i)\psi(X_i)\prod_{k\ne i}(\phi(X_k)\psi(X_k))\right] \\
& = &
E\left[\int\xi(dv)v\int_0^1du F_1(uv,(1-u)v)
\phi(uv)\phi((1-u)v)\psi(v)
\Pi_{\phi\psi}(\xi-\delta_{v})\right] \\
& = &
\int_0^1q_1(v)v\int_0^1F_1(uv,(1-u)v)
\phi(uv)\phi((1-u)v)\psi(v)
E_{v}\left[\Pi_{\phi\psi}(\eta)\right]dudv.
\end{eqnarray*}
By virtue of (\ref{2.17}) we obtain the desired equality
$\lg\lg\Phi,\Psi\rg\rg_{\rm coag}
={}_{\rm frag}\lg\lg\Phi,\Psi\rg\rg$
after the change of variables $uv=:x_1,(1-u)v=:x_2$ \\
(ii) This assertion is immediate
by noting that (\ref{2.17}) with
\[
K_1(x_1,x_2)=x_1x_2Q(x_1,x_2) \quad \mbox{and}
\quad
F_1(x_1,x_2)=\theta(x_1+x_2)Q(x_1,x_2)
\]
is valid for the ${\rm PD}(\theta)$ process
because of (\ref{2.14}) and (\ref{2.15}).
The proof of Theorem 2.3 is complete.
\qed
\medskip
\noindent
{\it Remarks.}~
(i) It would be interesting to investigate the class
of (nonnegative unbounded) functions $Q$ for which
the operator $L_1^{(Q,\theta)}$ defines
a Markov process on $\Omega_1$.
For example, if taking $Q(x,y)=(xy)^{-1}$
is allowed in that sense, the operator
\begin{eqnarray*}
L_1^{(Q,\theta)}\Phi(\bx)
=
\frac{1}{2}\sum_{i\ne j}
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right)
+\frac{\theta}{2}\sum_i \int_0^{1}\frac{du}{u(1-u)}
\left(\Phi(S_i^{u}\bx)-\Phi(\bx)\right)
\end{eqnarray*}
could deserve further exploration
in connection with e.g. `asymptotic frequency'
of some exchangeable fragmentation-coalescence
process studied in \cite{Be}. It is pointed out that
at least the unbounded coagulation operator
in the above can be treated
within the fame work of \cite{C2}. \\
(ii) Alternative direction of generalization of
the processes reversible with respect to
${\rm PD}(\theta)$
is based on an obvious generalization of
(\ref{2.14}) and (\ref{2.15}), i.e.,
\[
x_1\cdots x_{k+1}q_{k+1}(x_1,\ldots,x_{k+1})
=\theta^k (x_1+\cdots +x_{k+1})
q_1(x_1+\cdots +x_{k+1})
\]
and
\[
P_{x_1,\ldots,x_{k+1}}=P_{x_1+\cdots +x_{k+1}},
\]
in which $k\in\N$ is arbitrary and
$(x_1,\ldots,x_{k+1})\in \Delta_{k+1}^{o}$.
The corresponding
process on $\Omega_1$
incorporates multiple-coagulation
and multiple-fragmentation. (cf.
the transition kernel (\ref{2.5}) in \cite{EW}
or Example 1.8 in \cite{Kol}.
See also \cite{C2} for more general scheme
for the multiple-fragmentation.)
One of the simplest examples of the generator
of such a process is
\begin{eqnarray*}
\lefteqn{
L_{1,k}^{(\theta)}\Pi_{\phi}(\bx)} \\
& := &
\sum_{i_1,\ldots,i_{k+1}(\ne)} x_{i_1}\cdots x_{i_{k+1}}
\left[\phi(x_{i_1}+\cdots+x_{i_{k+1}})
-\phi(x_{i_1})\cdots\phi(x_{i_{k+1}})\right]
\prod_{j\ne i_1,\ldots,i_{k+1}}\phi(x_j) \\
& &
+\theta^k\sum_i x_i^{k+1}\int_{\Delta_k}du_1\cdots du_k
\left[\phi(u_1x_i)\cdots\phi(u_{k}x_i)
\phi((1-|u|)x_i)-\phi(x_i)\right]
\prod_{j\ne i}\phi(x_j) \\
& = &
\int\xi^{[k+1]}(dv_1\cdots dv_{k+1})v_1\cdots v_{k+1} \\
& &
\hspace*{10mm} \times
\left[\phi(v_1+\cdots+v_{k+1})
-\phi(v_1)\cdots\phi(v_{k+1})\right]
\Pi_{\phi}(\xi-\delta_{v_1}-\cdots-\delta_{v_{k+1}}) \\
& &
+\theta^k\int\xi(dv)v^{k+1}
\int_{\Delta_k}du_1\cdots du_k \\
& &
\hspace*{10mm} \times
\left[\phi(u_1v)\cdots\phi(u_{k}v)
\phi((1-|u|)v)-\phi(v)\right]
\Pi_{\phi}(\xi-\delta_{v}),
\end{eqnarray*}
where $|u|=u_1+\cdots +u_{k}$ and $\xi=\Xi(\bx)$.
Clearly $L_{1,1}^{(\theta)}=2L_{1}^{(1,\theta)}$.
The calculations in the proof of Theorem 2.3
is easily modified to prove that
${\rm PD}(\theta)$ is still a reversible distribution
of the process generated by $L_{1,k}^{(\theta)}$.
The details are left to the reader. \\
(iii) The spectral gap of a suitable extension
$\overline{L_1^{(Q,\theta)}}$of $L_1^{(Q,\theta)}$ vanishes.
Indeed, letting $\Psi_{\delta}(\bz)=\sum_iz_i^{\delta}$ for
$\bz=(z_i)\in \Omega$ and $\delta>0$,
we see, with the help of Lemma 6.4 in \cite{H09},
that $\Psi_{\delta}$ is square integrable
with respect to ${\rm PD}(\theta)$ and that
its variance ${\rm var}(\Psi_{\delta})$ is given by
\[
{\rm var}(\Psi_{\delta})
\ = \
\frac{\Gamma(\theta+1)\Gamma(2\delta)}
{\Gamma(\theta+2\delta)}
+\Gamma(\delta)^2
\left(\frac{\theta\Gamma(\theta+1)}{\Gamma(\theta+2\delta)}
-\frac{\Gamma(\theta+1)^2}{\Gamma(\theta+\delta)^2}\right)
\ =: \
\chi_1(\delta)+\chi_2(\delta),
\]
where $\Gamma(\cdot)$ is Gamma function.
As $\delta \downarrow 0$,
${\rm var}(\Psi_{\delta})\to \infty$ since
$\chi_1(\delta)\to \infty$ and
\begin{eqnarray*}
\chi_2(\delta)& =&
\frac{\theta\Gamma(\theta+1)\Gamma(\delta+1)^2}
{\Gamma(\theta+2\delta)\Gamma(\theta+\delta)^2}
\cdot
\frac{\Gamma(\theta+\delta)^2
-\Gamma(\theta)\Gamma(\theta+2\delta)}{\delta^2} \\
& \to&
\frac{\theta^2}{\Gamma(\theta)^2}
\left(\Gamma'(\theta)^2-\Gamma(\theta)
\Gamma''(\theta)\right)
\quad (\mbox{by l'Hospital's rule}).
\end{eqnarray*}
As for Dirichlet form
$\cE(\Psi_{\delta}): =
E\left[\Psi_{\delta}(X)(-\overline{L_1^{(Q,\theta)}})
\Psi_{\delta}(X)\right]$
in which $X=(X_i)_{i=1}^{\infty}$ is
${\rm PD}(\theta)$-distributed,
by the dominated convergence theorem
\begin{eqnarray*}
\cE(\Psi_{\delta})
& = &
\frac{1}{2}E\left[\overline{L_1^{(Q,\theta)}}(\Psi_{\delta}^2)(X)
-2\Psi_{\delta}(X)\overline{L_1^{(Q,\theta)}}\Psi_{\delta}(X)\right] \\
& =&
\frac{1}{4} E\left[\sum_{i\ne j}X_iX_jQ(X_i,X_j)
\{\Psi_{\delta}(M_{ij}X)-\Psi_{\delta}(X)\}^2\right] \\
& &
+ \frac{\theta}{4}
E\left[\sum_{i}X_i^2\int_0^1du Q(uX_i,(1-u)X_i)
\{\Psi_{\delta}(S_{i}^uX)-\Psi_{\delta}(X)\}^2\right] \\
& =&
\frac{1}{4} E\left[\sum_{i\ne j}X_iX_jQ(X_i,X_j)
\left\{(X_i+X_j)^{\delta}-X_i^{\delta}-X_j^{\delta}\right\}^2\right] \\
& &
+ \frac{\theta}{4}
E\left[\sum_{i}X_i^2\int_0^1du Q(uX_i,(1-u)X_i)
\left\{(uX_i)^{\delta}+((1-u)X_i)^{\delta}
-X_i^{\delta}\right\}^2\right] \\
& \to &
\frac{1}{4} E\left[\sum_{i\ne j}X_iX_jQ(X_i,X_j)\right]
+\frac{\theta}{4} E\left[\sum_{i}X_i^2\int_0^1du
Q(uX_i,(1-u)X_i) \right]<\infty
\end{eqnarray*}
as $\delta \downarrow 0$.
(In fact, the second equality in the above needs
justification. This can be done by considering
bounded functions
$\Psi_{\delta, \epsilon}(\bx)
=\sum x_i^{\delta}\one_{\{x_i\ge \epsilon\}}$
on $\Omega_1$, taking the limit as
$\epsilon \downarrow 0$ and
applying Lebesgue's convergence theorem.)
Consequently,
$\cE(\Psi_{\delta})/{\rm var}(\Psi_{\delta})\to 0$,
and hence the exponential convergence to equilibrium
does not hold for the process generated by
$L_1^{(Q,\theta)}$. It seems, however,
that the exact speed of convergence is unknown
even in the case $Q\equiv \mbox{const}$.
\medskip
The equality (\ref{2.17}) is thought of as a
probabilistic counterpart of the detailed balance
condition (\ref{1.3}). It should be noted that
(\ref{2.17}) is equivalent to the validity of
two equalities
$K_1(x_1,x_2)q_2(x_1,x_2)=F_1(x_1+x_2)q_1(x_1+x_2)$
and $P_{x_1,x_2}=P_{x_1+x_2}$,
a.e.-$(x_1,x_2)\in \Delta_2^{o}$
because of the triviality that
$P_{x_1,x_2}(\cN)=P_{x_1+x_2}(\cN)=1$.
The reader may wonder
whether there is any distribution other
than Poisson-Dirichlet distributions
which enjoys the relation (\ref{2.17})
for some explicit $K_1$ and $F_1$.
The following examples are intended to
give answers to that question by discussing
certain deformations of ${\rm PD}(\theta)$.
\medskip
\noindent
{\it Examples.}~
Let $(X_i)_{i=1}^{\infty}$ be ${\rm PD}(\theta)$-distributed
and $\phi$ be a nonnegative measurable function
on $(0,1)$. Assume that $0<a:=E\left[\prod_{i}\phi(X_i)\right]<\infty$ and
define a probability measure $\wt{P}$ on $\Omega_1$ by
\[
\wt{P}(\bullet)=a^{-1}
E\left[\prod_{i}\phi(X_i):~(X_i)_{i=1}^{\infty}\in\bullet \right].
\]
It is not difficult to show that
the first and second correlation functions
$\wt{q}_1$ and $\wt{q}_2$ and
the first-order and second-order reduced
Palm distributions of $\sum \delta_{X_i}$
under $\wt{P}$ are given in terms of
those of the ${\rm PD}(\theta)$ process
(namely, $q_1, q_2, P_v$ and $P_{x_1,x_2}$) by
\[
\wt{q}_1(v)=a^{-1}\phi(v)q_1(v)
E_v\left[\prod_{i}\phi(X_i)\right],
\]
\[
\wt{q}_2(x_1.x_2)
=a^{-1}\phi(x_1)\phi(x_2)q_2(x_1,x_2)
E_{x_1,x_2}\left[\prod_{i}\phi(X_i)\right],
\]
\[
\wt{P}_v(\bullet)
=E_v\left[\prod_{i}\phi(X_i):~
\sum_i\delta_{X_i}\in \bullet \right]
\left(E_v\left[\prod_{i}\phi(X_i)\right]\right)^{-1}
\]
and
\[
\wt{P}_{x_1,x_2}(\bullet)
=E_{x_1,x_2}\left[\prod_{i}\phi(X_i):~
\sum_i\delta_{X_i}\in \bullet \right]
\left(E_{x_1,x_2}\left[\prod_{i}\phi(X_i)\right]\right)^{-1},
\]
respectively. Notice that the above formula for
$\wt{P}_v$ (resp. $\wt{P}_{x_1,x_2}$)
is valid in $\wt{q}_1(v)dv$-a.e.
(resp. $\wt{q}_1(x_1,x_2)dx_1dx_2$-a.e.) sense,
so that the denominator can be
considered to be positive.
Combining with (\ref{2.14}) and (\ref{2.15}), we obtain
\[
x_1x_2\phi(x_1+x_2)\wt{q}_2(x_1,x_2)
=\theta (x_1+x_2) \phi(x_1)\phi(x_2)\wt{q}_1(x_1+x_2)
\]
and $\wt{P}_{x_1,x_2}=\wt{P}_{x_1+x_2}$. Here are
two examples of $K_1$ and $F_1$. \\
(i) The above two identities show that
(\ref{2.17}) is satisfied by $\sum \delta_{X_i}$
under $\wt{P}$ when we choose
\[
K_1(x_1,x_2)=x_1x_2\phi(x_1+x_2) \quad \mbox{and} \quad
F_1(x_1,x_2)=\theta (x_1+x_2) \phi(x_1)\phi(x_2).
\]
In this case, (\ref{2.16}) reads
\begin{eqnarray*}
L_1^{\sharp}\Phi(\bx)
& = &
\frac{1}{2}\sum_{i\ne j}x_ix_j \phi(x_i+x_j)
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right) \nonumber \\
& & +\frac{\theta}{2}\sum_i x_i^2
\int_0^{1}du \phi(ux_i)\phi((1-u)x_i)
\left(\Phi(S_i^{u}\bx)-\Phi(\bx)\right),
\end{eqnarray*}
which defines a bounded operator
whenever $\phi$ is bounded. For example,
fixing $s\in (0,1)$ arbitrarily and
setting $\phi(u)=\one_{(0,s]}(u)$,
we see easily that $a=P(X_1\le s)$ and
the associated reversible distribution $\wt{P}$
is identified with the conditional law of
$(X_i)_{i=1}^{\infty}$ given that $X_1\le s$. \\
(ii) For another choice
\[
K_1(x_1,x_2)=x_1x_2 \quad \mbox{and} \quad
F_1(x_1,x_2)=\theta (x_1+x_2) \phi(x_1)\phi(x_2)/\phi(x_1+x_2)
\]
(\ref{2.16}) becomes
\begin{eqnarray*}
L_1^{\sharp}\Phi(\bx)
& = &
\frac{1}{2}\sum_{i\ne j}x_ix_j
\left(\Phi(M_{ij}\bx)-\Phi(\bx)\right) \nonumber \\
& & +\frac{\theta}{2}\sum_i \frac{x_i^2}{\phi(x_i)}
\int_0^{1}du \phi(ux_i)\phi((1-u)x_i)
\left(\Phi(S_i^{u}\bx)-\Phi(\bx)\right).
\end{eqnarray*}
(Compare with the transition kernel
studied in \cite{MWZZ}.)
This operator is bounded for any
uniformly positive, bounded function $\phi$ on $(0,1)$.
To check, we take $\phi(u)=\exp(bu)$ with
$b$ being an arbitrary real number
and then verify that $\wt{P}={\rm PD}(\theta)$
and $L_1^{\sharp}=L_1^{(1,\theta)}$.
The distribution $\wt{P}$ for $\phi(u)=\exp(bu^2)$
has been discussed as the equilibrium measure of
a certain model in population genetics.
(See \cite{H05} and the references therein.)
\medskip
To explore an analogue of Theorem 2.3
for processes on $\Omega$, we discuss
the process generated by $L$ in (\ref{2.4}).
The special case $\theta\Hh\equiv \Hc$ has
the generator
\begin{eqnarray}
L^{(H,\theta)}\Phi(\bz)
& := &
\frac{1}{2}\sum_{i\ne j} z_iz_j H(z_i,z_j)
\left(\Phi(M_{ij}\bz)-\Phi(\bz)\right) \nonumber \\
& & +\frac{\theta}{2}\sum_i z_i^{2}
\int_0^{1}du H(uz_i,(1-u)z_i)
\left(\Phi(S_i^{u}\bz)-\Phi(\bz)\right), \label{2.18}
\end{eqnarray}
where $H$ is a symmetric, nonnegative
homogeneous function $H$
of degree $\lambda\ge 0$ satisfying (H1) and $\theta>0$. By conditioning (or cut-off) argument, we get
\begin{th
(i) Let $Z=(Z_i)_{i=1}^{\infty}$ be a random element
of $\Omega$ and suppose that
the first and second correlation functions
$r_1$ and $r_2$ of $\xi:=\Xi(Z)$ exist and satisfy
\be
K(z_1,z_2)r_2(z_1,z_2)P_{z_1,z_2}
=F(z_1,z_2)r_1(z_1+z_2) P_{z_1+z_2},
\label{2.19}
\ee
$\mbox{a.e.-}(z_1,z_2)\in (0,\infty)^2$.
Here, $P_{z_1,z_2}$ and $P_{z_1+z_2}$ are
the reduced Palm distributions of $\xi$ and
the above equality is understood as the one
between two measures on $\cN$. Then
the process generated by $L$ is
reversible with respect to the law of $Z$.
\\
(ii) Let $L^{(H,\theta)}$ be as in (\ref{2.18}).
Suppose that
$(X_i)_{i=1}^{\infty}$ is $\rm{PD}(\theta)$-distributed.
Then, for any $(0,\infty)$-valued random variable $V$
independent of $X$,
the law of an $\Omega$-valued random element
$(VX_i)_{i=1}^{\infty}$ is a reversible distribution
of the process generated by $L^{(H,\theta)}$.
\end{th
{\it Proof.}~
(i) Let $R>0$ be such that $P(|Z|\le R)>0$.
First, consider the process generated by $L$
with initial distribution
$P^{(R)}(\bullet):=P(Z\in \bullet|~ |Z|\le R)$.
Then it is clear that such a process lies in
$\Omega_{\le R}$, and hence
its generator $L$ is essentially bounded.
So, just as in the proof of Theorem 2.3
the proof of the reversibility with respect to
$P^{(R)}$ can reduce to verifying that
the equalities corresponding to (\ref{2.19}) hold
for $\xi=\Xi(Z)$ under the conditional law $P^{(R)}$.
It is not difficult to show that,
under $P^{(R)}$, $\xi=\Xi(Z)$ has
the first and second correlation functions
\[
r_1^{(R)}(z):=r_1(z) \one_{\{z\le R\}}
\frac{P_{z}(|\eta|\le R-z)}{P(|Z|\le R)},
\]
\[
r_2^{(R)}(z_1,z_2):=r_2(z_1,z_2) \one_{\{z_1+z_2\le R\}}
\frac{P_{z_1,z_2}(|\eta|\le R-(z_1+z_2))}{P(|Z|\le R)}
\]
and the first order and second order
reduced Palm distributions
\[
P^{(R)}_z(\bullet)=P_z(\bullet~|~|\eta|\le R-z),
\quad P^{(R)}_{z_1,z_2}(\bullet)
=P_{z_1,z_2}(\bullet~|~|\eta|\le R-(z_1+z_2)).
\]
These formulas combined with (\ref{2.19}) yield
\[
K(z_1,z_2)r^{(R)}_2(z_1,z_2)P^{(R)}_{z_1,z_2}
=F(z_1,z_2)r^{(R)}_1(z_1+z_2) P^{(R)}_{z_1+z_2},
\]
which is sufficient to imply the reversibility of
the process $\{Z(t):~t\ge 0\}$
generated by $L$ with initial distribution $P^{(R)}$
for the aforementioned reason.
Before taking the limit as $R\to\infty$,
we interpret the reversibility obtained so far
in terms of conditional expectations
as follows: for any $n\in\N$,
$0<t_1<\cdots<t_n<T$ and
$\Phi_1,\ldots, \Phi_n\in B(\Omega)$
\begin{eqnarray*}
\lefteqn{
E\left[\Phi_1(Z(t_1))\cdots \Phi_n(Z(t_n))
|~|Z(0)|\le R\right]} \\
& = &
E\left[\Phi_1(Z(T-t_1))\cdots \Phi_n(Z(T-t_n))
|~|Z(0)|\le R\right].
\end{eqnarray*}
By letting $R\to \infty$ the required reversibility
has been proved. \\
(ii)
Consider the lifted process
$\{\wt{Z}(t)=(VX_i(t))_{i=1}^{\infty}:~t\ge 0\}$,
where $\{X(t)=(X_i(t))_{i=1}^{\infty}:~t\ge 0\}$
is generated by $L_1^{(H,\theta)}$,
independent of $V$ and
such that $X(0)=(X_i)_{i=1}^{\infty}$.
By Lemma 2.1 (ii) $\{\wt{Z}(t)\}$ is generated by
$\wt{L}$ with $\Hh=H$ and $\Hc=\theta H$. Accordingly
\[
Z(t):=\wt{Z}(|\wt{Z}(0)|^{2+\lambda}t)
=VX(V^{2+\lambda}t)
\]
is a process generated by $L^{(H,\theta)}$
and clearly $Z(0)=(VX_i)_{i=1}^{\infty}$.
Letting $\rho$ denote the law of $V$,
we see from the reversibility of $\{X(t)\}$
proved in Theorem 2.3 (ii) that for any $n\in\N$,
$0<t_1<\cdots<t_n<T$ and
$\Phi_1,\ldots, \Phi_n\in B(\Omega)$
\begin{eqnarray*}
\lefteqn{
E\left[\Phi_1(Z(t_1))\cdots \Phi_n(Z(t_n))\right]} \\
& = &
\int\rho(dv)E\left[\Phi_1(vX(v^{2+\lambda}t_1))
\cdots \Phi_n(vX(v^{2+\lambda}t_n))\right] \\
& =&
\int\rho(dv)E\left[\Phi_1(vX(v^{2+\lambda}T-v^{2+\lambda}t_1))
\cdots\Phi_n(vX(v^{2+\lambda}T-v^{2+\lambda}t_n))\right] \\
& =&
E\left[\Phi_1(Z(T-t_1))\cdots \Phi_n(Z(T-t_n))\right].
\end{eqnarray*}
This proves that the law of $Z(0)=(VX_i)_{i=1}^{\infty}$
is a reversible distribution of $\{Z(t):~t\ge 0\}$,
a process generated by $L^{(H,\theta)}$.
\qed
\medskip
In fact, alternative proof of Theorem 2.4 (ii)
exists and is based on (\ref{2.19})
together with the following static result
on the correlation measures
and the reduced Palm distributions of
`the lifted $\rm{PD}(\theta)$ process'
$\sum\delta_{VX_i}$.
\begin{lm
Let $(X_i)_{i=1}^{\infty}$ be
$\rm{PD}(\theta)$-distributed
and suppose that a $(0,\infty)$-valued
random variable $V$ independent of $(X_i)_{i=1}^{\infty}$ is given.
Then, for each $k\in\N$,
the $k$th correlation function $r_{k}$ on $(0,\infty)^k$ of $\sum\delta_{VX_i}$is given by
\be
r_{k}(z_1,\ldots,z_k)
=\frac{\theta^k}{z_1\cdots z_k}\int_{(|z|,\infty)}
\rho(dv)\left(1-\frac{|z|}{v}\right)^{\theta-1}, \label{2.20}
\ee
where $|z|=z_1+\cdots+z_k$ and
$\rho$ is the law of $V$.
Moreover, for any $z=(z_1,\ldots,z_k)\in(0,\infty)^k$
such that $P(V>|z|)>0$, the expectation $E_{z_1,\ldots,z_k}$
with resect to the $k$th-order reduced Palm distribution
of $\sum\delta_{VX_i}$ at $z=(z_1,\ldots,z_k)$
is characterized by the formula
\be
E_{z_1,\ldots,z_k}\left[\prod_i\phi(VX_i)\right]
=
\frac{\ds{\int_{(|z|,\infty)}\rho(dv)
\left(1-\frac{|z|}{v}\right)^{\theta-1}
E\left[\prod_i\phi((v-|z|)X_i)\right]}}
{\ds{\int_{(|z|,\infty)}\rho(dv)\left(1-\frac{|z|}{v}\right)^{\theta-1}}}
, \label{2.21}
\ee
in which $\phi\in\wt{B}_+$ is arbitrary. \par
In the special case where $V$ has the gamma density
\be
\rho_{\theta,b}(v):=\Gamma(\theta)^{-1}
b^{\theta}v^{\theta-1}e^{-bv}\one_{(0,\infty)}(v) \label{2.22}
\ee
with $b>0$, $\sum \delta_{VX_i}$
is a Poisson point process on $(0,\infty)$
with mean measure $\theta y^{-1}e^{-by}dy$.
\end{lm
{\it Proof.}~Let $f\in B_+^k$ be arbitrary.
By the assumed independence and
the Palm formula for the $\rm{PD}(\theta)$ process
$\sum \delta_{X_i}$
\begin{eqnarray*}
\lefteqn{
E\left[\sum_{i_1,\ldots,i_k(\ne)}f(VX_{i_1},\ldots,VX_{i_k})
\prod_{j\neq i_1,\ldots,i_k} \phi(VX_j) \right]} \\
& = &
\int\rho(dv)E\left[\sum_{i_1,\ldots,i_k(\ne)}
f(vX_{i_1},\ldots,vX_{i_k})
\prod_{j\neq i_1,\ldots,i_k} \phi(vX_j) \right] \\
& =&
\int\rho(dv)\int_{\Delta_k^{\circ}} f(vx_1,\ldots,vx_k)
\frac{\theta^k(1-|x|)^{\theta-1}}{x_1\cdots x_k}dx_1\cdots dx_k
E\left[\prod_{j} \phi(v(1-|x|)X_j) \right] \\
& =&
\int_{(0,\infty)^k}dz_1\cdots dz_k f(z_1,\ldots,z_k) \\
& &
\times \frac{\theta^k}{z_1\cdots z_k}
\int_{(|z|,\infty)}\rho(dv)
\left(1-\frac{|z|}{v}\right)^{\theta-1}
E\left[\prod_{j} \phi((v-|z|)X_j) \right].
\end{eqnarray*}
Taking $\phi\equiv 1$ yields (\ref{2.20}).
Therefore,
the above equalities suffice to imply (\ref{2.21}).
The last assertion follows from
\[
\frac{\theta^k}{z_1\cdots z_k}\int_{(|z|,\infty)}
\rho_{\theta,b}(v)
\left(1-\frac{|z|}{v}\right)^{\theta-1}dv
=\frac{\theta^k}{z_1\cdots z_k}e^{-b|z|}
=\prod_{i=1}^k\left(\frac{\theta}{z_i}e^{-bz_i}\right),
\]
which is nothing but the $k$th correlation function of
a Poisson point process on $(0,\infty)$ with
mean density $\theta y^{-1}e^{-by}$.
\qed
\medskip
\noindent
We call the above-mentioned Poisson process
the gamma point process with parameter $(\theta,b)$.
It is worth noting that in view of (\ref{2.20}) and
(\ref{2.21}) the equalities
\[
z_1z_2r_2(v_1,v_2)=\theta (z_1+z_2)r_1(z_1+z_2)
\quad \mbox{and} \quad
E_{z_1,z_2}=E_{z_1+z_2}
\]
which correspond to (\ref{2.14}) and (\ref{2.15}),
respectively, hold true for
any lifted $\rm{PD}(\theta)$ process,
and as a result it satisfies also (\ref{2.19}) with
\be
K(x,y)=xy H(x,y) \quad
\mbox{and} \quad
F(x,y)=\theta(x+y) H(x,y). \label{2.23}
\ee
In the forthcoming section it will be shown that
the time-dependent system of correlation measures
of the process generated by $L$ solves the
hierarchical equation (\ref{1.2}) weakly.
In connection with Theorem 2.4 (ii), we remark that
the system of the correlation functions $\{r_k\}$
given by (\ref{2.20}) is verified directly
to be a stationary solution to (\ref{1.2}) with (\ref{2.23}).
In these calculations, merely the following structure
is relevant:
\[
z_1\cdots z_k r_k(z_1,\cdots, z_k)
=\theta^k g(z_1+\cdots+z_k)
\]
for some function $g$ independent of $k$
(although (\ref{2.20}) shows us the exact form of $g$).
\section{Hierarchical equations for correlation measures
\setcounter{equation}{0}
The purpose of this section is to derive equations
(\ref{1.2}) as those describing the time evolution of
the correlation measures associated with
coagulation-fragmentation processes
introduced in the previous section.
As far as the reversible process generated by
$L_1^{(Q,\theta)}$ with $Q\equiv\mbox{const.}$ is concerned,
such an attempt is found essentially in \cite{MWZZ}
for the purpose of showing the uniqueness of stationary distributions.
In that paper, however, the stationary hierarchical equations
(called `the basic relations' on p.19) for the
correlation functions are incorrect,
overlooking a term coming from coagulation between
clusters with specific sizes given
(i.e., a term involving $p_{k-1}$).
It is not clear that developing such an approach could
make one possible to settle the uniqueness issue
in much more general setting.
More specifically, in the reversible case described
in Theorem 2.3 (resp. Theorem 2.4),
it seems reasonable to expect the existence of
a functional of distributions on
$\Omega_1$ (resp. $\Omega$) which decays
under the time evolution governed by
$L_1^{\sharp}$ (resp. $L$). (Notice that
the uniqueness of stationary distributions
cannot hold for the processes on $\Omega$.)
Intending only to derive an infinite system of equations
describing fully the time evolution of our model,
we begin by introducing a weak version of (\ref{1.2}),
namely, the equation obtained by operating on
a test function by its formal adjoint.
Since $K$ and $F$ are unbounded,
suitable integrability conditions must be required for solutions.
Let $B_{+,c}^k$ denote
the totality of functions in $B_{+}^k$ with compact support.
Also, the abbreviated notation $\bz_k=(z_1,\ldots,z_k)$ and
$d\bz_k=dz_1\cdots dz_k$ are used in the integral expressions. A family of measures $\{c_k(t,d\bz_k):~t\ge 0, k\in\N\}$
is said to be admissible if the following two conditions are fulfilled: \\
(A1) Each $c_k(t,\cdot)$ is a locally finite measure on $(0,\infty)^k$. \\
(A2) For any $f \in B_{+,c}^{k}$, $g \in B_{+,c}^{k-1}$
and $l\in\{1,\ldots,k\}$, the three functions below are
locally integrable on $[0,\infty)$:
\begin{eqnarray*}
{\rm (i)}~ t & \mapsto & \int f(\bz_k)c_k(t,d\bz_k), \\
{\rm (ii)}~ t & \mapsto &
\int z_l z_{l+1} f(z_1,\ldots,z_{l-1},z_{l}+z_{l+1},z_{l+2}\ldots,z_{k+1})
c_{k+1}(t,d\bz_{k+1}),
\\
{\rm (iii)}~ t & \mapsto &
\left\{
\begin{array}{ll}
\ds{\int z_{1}(z_{1}^{1+\lambda}\vee 1) c_{1}(t,dz_{1})} & (k=1), \\
\ds{\int z_{l}(z_{l}^{1+\lambda}\vee 1)
g(z_1,\ldots,z_{l-1},z_{l+1},\ldots,z_{k}) c_{k}(t,d\bz_{k})} & (k\ge 2).
\end{array}
\right.
\end{eqnarray*}
Given an admissible $\{c_k(t,d\bz_k):~t\ge 0, k\in\N\}$,
we call it a weak solution of the hierarchical
coagulation-fragmentation equation (\ref{1.2})
with kernels $K$ and $F$ given by (\ref{1.4})
if for any $t>0$, $k\in\N$ and $f\in B_{+,c}^k$
\begin{eqnarray}
\lefteqn{
\int f(\bz_k)c_k(t,d\bz_k)
-\int f(\bz_k)c_k(0,d\bz_k)}
\label{3.1} \\
&=&
\frac{1}{2}\sum_{l=1}^k\int_0^tds \int K(z_l,z_{l+1})
f({\rm Coag}_{l,l+1}\bz_{k+1})c_{k+1}(s,d\bz_{k+1})
\nonumber \\
& &
-\frac{1}{2}\sum_{l=1}^k\int_0^tds
\int\int_0^{z_l}dy F(y,z_l-y) f(\bz_k)c_{k}(s,d\bz_k)
\nonumber \\
& &
-\sum_{l=1}^k\int_0^tds \int K(z_l,z_{l+1})
f(z_1,\ldots,z_l,z_{l+2},\ldots,z_{k+1})
c_{k+1}(s,d\bz_{k+1})
\nonumber \\
& &
+\sum_{l=1}^k\int_0^tds \int \int_0^{z_l}dyF(y,z_l-y)
f(z_1,\ldots,z_{l-1},y,z_{l+1},\ldots,z_{k})c_{k}(s,d\bz_{k})
\nonumber \\
& &
-\one_{\{k\ge 2\}}\sum_{l<m}^{k}\int_0^tds \int K(z_l,z_m)
f(\bz_k)c_{k}(s,d\bz_{k})
\nonumber \\
& &
+\one_{\{k\ge 2\}}\sum_{l<m}^{k}\int_0^tds
\int\int_0^{z_l}dyF(y,z_l-y)
f({\rm Frag}_{l,m}^{(y)}\bz_{k-1})
c_{k-1}(s,d\bz_{k-1}), \nonumber
\end{eqnarray}
where
\[
{\rm Coag}_{l,l+1}\bz_{k+1}
=(z_1,\ldots,z_{l-1},z_l+z_{l+1},z_{l+2},\ldots,z_{k+1})
\]
and
\[
{\rm Frag}_{l,m}^{(y)}\bz_{k-1}
=(z_1,\ldots,z_{l-1},y,z_{l+1},\ldots,
z_{m-1},z_l-y,z_{m},\ldots,z_{k-1}).
\]
By (\ref{1.4}), the equation
(\ref{3.1}) actually takes a more specific form
\begin{eqnarray}
\lefteqn{
\int f(\bz_k)c_k(t,d\bz_k)
-\int f(\bz_k)c_k(0,d\bz_k)}
\label{3.2} \\
&=&
\frac{1}{2}\sum_{l=1}^k\int_0^tds \int z_l z_{l+1}
\Hh(z_l,z_{l+1})f({\rm Coag}_{l,l+1}\bz_{k+1})
c_{k+1}(s,d\bz_{k+1})
\nonumber \\
& &
-\frac{1}{2}\sum_{l=1}^k\int_0^tds \int z_l
\int_0^{z_l}dy\Hc(y,z_l-y)f(\bz_k)c_{k}(s,d\bz_k)
\nonumber \\
& &
-\sum_{l=1}^k\int_0^tds \int z_l z_{l+1}
\Hh(z_l,z_{l+1})f(z_1,\ldots,z_l,z_{l+2},\ldots,z_{k+1})
c_{k+1}(s,d\bz_{k+1})
\nonumber \\
& &
+\sum_{l=1}^k\int_0^tds \int z_l \int_0^{z_l}dy
\Hc(y,z_l-y)f(z_1,\ldots,z_{l-1},y,z_{l+1},\ldots,z_k)
c_{k}(s,d\bz_{k})
\nonumber \\
& &
-\one_{\{k\ge 2\}}\sum_{l<m}^{k}\int_0^tds \int z_l z_m
\Hh(z_l,z_m)f(\bz_k)c_{k}(s,d\bz_{k})
\nonumber \\
& &
+\one_{\{k\ge 2\}}\sum_{l<m}^{k}\int_0^t ds
\int z_l \int_0^{z_l} dy \Hc(y,z_l-y)
f({\rm Frag}_{l,m}^{(y)}\bz_{k-1}) c_{k-1}(s,d\bz_{k-1}) \nonumber \\
& =: &
I_1-I_2-I_3+I_4-I_5+I_6. \label{3.3}
\end{eqnarray}
Obviously, (A1) ensures that two integrals on
the left side of (\ref{3.1}) is finite.
Considering the terms on the right side,
we prepare the following bounds:
by homogeneity (\ref{1.5}), (H1) and (H2) together
\begin{eqnarray}
\Hh(x,y)
& = &
\Hh\left((x+y)\frac{x}{x+y},(x+y)\frac{x}{x+y}\right)
\le \Ch(x+y)^{\lambda} \label{3.4} \\
& \le & \Ch(1+x)^{\lambda} (y^{\lambda}\vee 1)
\ \le \ \Ch(1+x)^{\lambda} (y^{1+\lambda}\vee 1) \label{3.5}
\end{eqnarray}
and
\be
\int_0^{x}dy \Hc(y,x-y)
=x \int_0^1du \Hc(ux,(1-u)x) = \Cc x^{1+\lambda}. \label{3.6}
\ee
\begin{lm
Assume that $\{c_k(t,d\bz_k):~t\ge 0, k\in\N\}$ is admissible.
Then every term on the right side of (\ref{3.1})
(or equivalently of (\ref{3.2})) is finite.
\end{lm
{\it Proof.}~
We discuss $I_i$'s on (\ref{3.3})
instead of the terms on the right side of (\ref{3.1}).
It follows from (A2) (i)
that $I_2$ and $I_5$ are finite.
Also, $I_1$ is finite because of (\ref{3.4}) and (A2) (ii).
$I_3$ converges by (\ref{3.5}) together with (A2) (iii),
and similarly the finiteness of $I_4$ is
due to (\ref{3.6}) and (A2) (iii)
with
\[
g(z_{1},\ldots,z_{k-1}) =
\sup_{y>0} f(z_{1},\ldots,z_{{l-1}},y,z_{{l}},\ldots,z_{k-1}).
\]
Lastly, again by (A2) (i), $I_6$ is finite
since each function
\[
h_{l,m}(z_{1},\ldots,z_{k-1}):=
z_l \! \int_0^{z_l} \! dy \Hc(y,z_l-y)
f({\rm Frag}_{l,m}^{(y)}\bz_{k-1})
\]
is an element of $B_{+,c}^{k-1}$ for $k\ge 2$.
Indeed, taking $\epsilon>0$ so that
$f(z_{1},\ldots,z_{k})=0$ whenever $0<z_l<\epsilon$,
we see that $h_{l,m}(z_1,\ldots,z_{k-1})=0$
for any $z_l\in(0,\epsilon)$, and
analogously, taking $R>0$ so that
$f(z_{1},\ldots,z_{k})=0$ whenever
$z_l>R$ or $z_m>R$, we see that
$h_{l,m}(z_1,\ldots,z_{k-1})=0$ for any $z_l>2R$.
The proof of Lemma 3.1 is complete. \qed
\medskip
The main result of this section yields stochastic construction
of a solution to (\ref{1.2}) with kernels we are concerned with.
\begin{th
(i) Let $\{Z(t):~t\ge 0\}$ be the process
generated by $L$ and
suppose that $E[|Z(0)|^k]<\infty$ for all $k\in\N$.
For each $t\ge 0$ and $k\in \N$
denote by $r_k(t,d\bz_k)$
the $k$th correlation measure of $\Xi(Z(t))$.
Then
$\{r_k(t,d\bz_k):~t\ge 0, k\in\N\}$ is admissible and
solves weakly the hierarchical
coagulation-fragmentation equation (\ref{1.2}) with
kernels $K$ and $F$ given by (\ref{1.4}). \\
(ii) Let $\wh{Q}$ and $\check{Q}$ be
symmetric, nonnegative bounded functions on
$\{(x,y) |~x,y> 0, x+y \le 1\}$ and set
\[
K_1(x,y)=xy\wh{Q}(x,y) \quad \mbox{and} \quad
F_1(x,y)=(x+y)\check Q(x,y).
\]
Let $\{X(t):~t\ge 0\}$ be the process
generated by $L_1^{\sharp}$ in (\ref{2.16}).
For each $t\ge 0$ and $k\in \N$
denote by $q_k(t,d\bx_k)$
the $k$th correlation measure of $\Xi(X(t))$.
Then
$\{q_k(t,d\bz_k):~t\ge 0, k\in\N\}$ is admissible
and solves weakly the hierarchical
coagulation-fragmentation equation (\ref{1.2}) with
kernels $K_1$ and $F_1$.
\end{th
Recalling the definition of correlation measures
(cf. (\ref{2.7})),
the proof of this theorem is basically done by
calculating carefully $L\Phi(\bz)$ or
$L_1^{\sharp}\Phi(\bz)$ for
\be
\Phi(\bz)=\sum_{i_1,\ldots,i_k(\ne)}f(z_{i_1},\ldots,z_{i_k}),
\label{3.7}
\ee
where $f\in B_{+,c}^k$ is arbitrary.
Here, we understand that
$f(z_{i_1},\ldots,z_{i_k})=0$ when
$z_{i_1} \cdots z_{i_k}=0$, so that
\[
\sum_{i_1,\ldots,i_k(\ne)}f(z_{i_1},\ldots,z_{i_k})
=
\int f(y_1,\ldots,y_k)\xi^{[k]}(dy_1\cdots dy_k),
\]
where $\xi=\Xi(\bz)$.
Although such function's $\Phi$ on $\Omega$
may be unbounded, we can control its growth order
as will be seen in the next lemma. For each $a>0$
denote by $\cF_a$ the class of measurable functions $\Psi$
on $\Omega$ such that, for some constant $C<\infty$,
$|\Psi(\bz)|\le C|\bz|^a$ for all $\bz\in\Omega$.
\begin{lm
Let $f\in B_{+,c}^k$ and $\Phi$ be as in (\ref{3.7}).
Then $\Phi\in\cF_k$.
\end{lm
{\it Proof.}~
Define
$\epsilon =\inf\{\min\{z_1,\ldots,z_k\}:(z_1,\ldots,z_k)
\in {\rm supp}(f)\}$, which is strictly positive
because ${\rm supp}(f)$ is assumed to be
a compact subset of $(0,\infty)^k$. Therefore
\[
0 \le \Phi(z) \le \Vert f\Vert_{\infty}
\sum_{i_1,\ldots,i_k}
\frac{z_{i_1}}{\epsilon}\cdots \frac{z_{i_k}}{\epsilon}
\one_{\{(z_{i_1},\ldots,z_{i_k})\in {\rm supp}(f)\}}
\le \Vert f\Vert_{\infty}\frac{|\bz|^{k}}{\epsilon^k}.
\]
This proves Lemma 3.3. \qed
\medskip
\noindent
At the core of our proof of Theorem 3.2 is
\begin{lm
Let $f\in B_{+,c}^k$ and $\Phi$ be as in (\ref{3.7}).
Then for each $\bz=(z_i)_{i=1}^{\infty}\in \Omega$
\begin{eqnarray}
\lefteqn{L\Phi(\bz)}
\label{3.8} \\
& = &
\frac{1}{2}
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k+1}(\ne)}K(z_{i_l},z_{i_{l+1}})
f({\rm Coag}_{l,l+1}(z_{i_1},\ldots,z_{i_{k+1}}))
\nonumber \\
& &
-\frac{1}{2}\sum_{l=1}^k \sum_{i_1,\ldots,i_{k}(\ne)}
\int_0^{z_{i_l}}dyF(y,z_{i_l}-y)f(z_{i_1},\ldots,z_{i_k}). \nonumber \\
& &
-\sum_{l=1}^k \sum_{i_1,\ldots,i_{k+1}(\ne)}K(z_{i_l},z_{i_{l+1}})
f(z_{i_1},\ldots,z_{i_l},z_{i_{l+2}},\ldots,z_{i_{k+1}}) \nonumber \\
& &
+\sum_{l=1}^k \sum_{i_1,\ldots,i_{k}(\ne)}
\int_0^{z_{i_l}}dyF(y,z_{i_l}-y) f(z_{i_1},\ldots,z_{i_{l-1}},y,z_{i_{l+1}},\ldots,z_{i_{k}})
\nonumber \\
& &
-\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k}(\ne)}K(z_{i_l},z_{i_m})
f(z_{i_1},\ldots,z_{i_{k}}) \nonumber \\
& &
+\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k-1}(\ne)} \int_0^{z_{i_l}}dyF(y,z_{i_l}-y)
f({\rm Frag}_{l,m}^{(y)}(z_{i_1},\ldots,z_{i_{k-1}}))
\nonumber \\
& =: &
\frac{1}{2}\sum_{l=1}^k\Psi_l^{(1)}(\bz)
-\frac{1}{2}\sum_{l=1}^k\Psi_l^{(2)}(\bz)
-\sum_{l=1}^k \Psi_l^{(3)}(\bz) \nonumber \\
& &
+\sum_{l=1}^k \Psi_l^{(4)}(\bz)
-\one_{\{k\ge 2\}}\sum_{l<m}^{k} \Psi_{l,m}^{(5)}(\bz)
+\one_{\{k\ge 2\}}\sum_{l<m}^{k} \Psi_{l,m}^{(6)}(\bz). \nonumber
\end{eqnarray}
Moreover, $\Psi_l^{(1)}\in \cF_{k+1}$,
$\Psi_l^{(2)}\in \cF_{k}$, $\Psi_l^{(3)}\in \cF_{k+1+\lambda}$,
$\Psi_l^{(4)}\in \cF_{k+1+\lambda}$,
$\one_{\{k\ge 2\}}\Psi_{l,m}^{(5)}\in \cF_{k}$
and $\one_{\{k\ge 2\}}\Psi_{l,m}^{(6)}\in \cF_{k+2}$.
Thus $L\Phi$ belongs to
the linear span $\overline{\cF}$ of $\bigcup_{a>0}\cF_a$.
\end{lm
{\it Proof.}~
The main proof of (\ref{3.8}) consists of
almost algebraic calculations (which are completely
independent of other arguments)
and so it is postponed.
We here prove only the assertions for
$\Psi_l^{(i)} (i\in\{1,2,3,4\})$
and $\Psi_{l,m}^{(j)} (j\in\{5,6\})$.
The arguments below are based on similar observations to
those in the proof of Lemma 3.1.
From Lemma 3.3, it is evident that
$\Psi_l^{(2)}$ and $\Psi_{l,m}^{(5)}$ are elements of $\cF_k$.
Relying on the fact that the function $h_{l,m}$
in the proof of Lemma 3.1 belongs to $B_{+,c}^{k-1}$,
one can verify similarly that
$\Psi_{l,m}^{(6)}\in\cF_{k-1}$ for $k\ge 2$.
As for $\Psi_l^{(1)}$, taking $\epsilon>0$ and $R>\epsilon$
such that $[\epsilon, R]^k\supset {\rm supp}(f)$
and using (\ref{3.4}),
we observe as in the proof of Lemma 3.3 that
for $l\in\{1,\ldots,k\}$
\begin{eqnarray*}
|\Psi_{l}^{(1)}(\bz)|
&\le &
\Ch \sum_{i_1,\ldots,i_{k+1}(\ne)}
z_{i_l}z_{i_{l+1}}(z_{i_l}+z_{i_{l+1}})^{\lambda}
f({\rm Coag}_{l,l+1}(z_{i_1},\ldots,z_{i_{k+1}})) \\
& \le &
\Ch \Vert f\Vert_{\infty}
\sum_{i_1,\ldots,i_{k+1}} z_{i_l}z_{i_{l+1}} R^{\lambda}
\frac{z_{i_1}}{\epsilon}\cdots \frac{z_{i_{l-1}}}{\epsilon}
\frac{z_{i_{l+2}}}{\epsilon}\cdots \frac{z_{i_{k+1}}}{\epsilon} \\
& = &
\Ch R^{\lambda} \Vert f\Vert_{\infty}
\frac{|\bz|^{k+1}}{\epsilon^{k-1}},
\end{eqnarray*}
by which $\Psi_{l}^{(1)}\in\cF_{k+1}$. Analogously
\begin{eqnarray*}
|\Psi_{l}^{(3)}(\bz)|
&\le &
\Ch \sum_{i_1,\ldots,i_{k+1}(\ne)}
z_{i_l}z_{i_{l+1}}(z_{i_l}+z_{i_{l+1}})^{\lambda}
f(z_{i_1},\ldots,z_{i_l},z_{i_{l+2}},\ldots,z_{i_{k+1}}) \\
& \le &
\Ch \sum_{i_1,\ldots,i_{k+1}(\ne)} R z_{i_{l+1}} |z|^{\lambda}
f(z_{i_1},\ldots,z_{i_l},z_{i_{l+2}},\ldots,z_{i_{k+1}}) \\
& \le &
\Ch R |\bz|^{1+\lambda}
\sum_{i_1,\ldots,i_{k}(\ne)} f(z_{i_1},\ldots,z_{i_{k}}),
\end{eqnarray*}
which combined with Lemma 3.3 implies
that $\Psi_l^{(3)}\in\cF_{k+1+\lambda}$.
Lastly, we shall show that $\Psi_l^{(4)}\in\cF_{k+1+\lambda}$.
For $k=1$ we have by (\ref{3.6})
\[
|\Psi_1^{(4)}(\bz)|
\le \Cc \sum_i z_i^{2+\lambda} \Vert f\Vert_{\infty}
\le \Cc |\bz|^{2+\lambda} \Vert f\Vert_{\infty}
\]
and hence $\Psi_1^{(4)}\in\cF_{2+\lambda}$.
For $k\ge 2$, by noting that
\[
{\overline f}_l(z_1,\ldots,z_{k-1})
:= \sup_{y>0} f(z_{1},\ldots,z_{{l-1}},y,z_{{l}},\ldots,z_{{k-1}})
\]
belongs to $B_{+,c}^{k-1}$ and observing that
\begin{eqnarray*}
|\Psi_l^{(4)}(\bz)|
& \le &
\Cc\sum_{i_1,\ldots,i_{k}(\ne)} z_{i_l}^{2+\lambda}
{\overline f}_l(z_{i_1},\ldots,z_{i_{l-1}},z_{i_{l+1}},\ldots,z_{i_{k}}) \\
& \le &
\Cc |\bz|^{2+\lambda}\sum_{i_1,\ldots,i_{k-1}(\ne)}
{\overline f}_l(z_{i_1},\ldots,z_{i_{k-1}}),
\end{eqnarray*}
we deduce from Lemma 3.3
that $\Psi_l^{(4)}\in\cF_{k+1+\lambda}$.
Consequently, we have shown
that $L\Phi\in\overline{\cF}$.
\qed
\medskip
\noindent
{\it Proof of Theorem 3.2 (i).}~
Let $f\in B_{+,c}^k$ be arbitrary.
First, we must show the admissibility of $r_k(t,d\bz_k)$.
The conditions (A1) and (A2) (i) for $r_k(t,d\bz_k)$
are easily seen to hold by combining Lemma 3.3 with
\[
E[|Z(t)|^a]=E[|Z(0)|^a]=:m_a<\infty
\]
for all $a\ge 1$.
(A2) (ii) can be verified by a similar bound
to that for $|\Psi_l^{(1)}(\bz)|$ in the proof of Lemma 3.4.
Moreover, (A2) (iii) follows from observations that
\begin{eqnarray*}
\lefteqn{
\int z_{1} (z_{1}^{1+\lambda}\vee 1) r_{1}(t,dz_{1})
\le E\left[\sum_i(Z_{i}(t)+Z_{i}(t)^{2+\lambda})\right]} \\
& \le &
E\left[|Z(t)|+|Z(t)|^{2+\lambda}\right]
\ = \ m_{1}+m_{2+\lambda}
\end{eqnarray*}
and that for $k\ge 2$ and $l\in\{1,\ldots,k\}$
\begin{eqnarray*}
\lefteqn{\int z_{l} (z_{l}^{1+\lambda}\vee 1)
g(z_1,\ldots,z_{l-1},z_{l+1},\ldots,z_{k}) r_{k}(t,d\bz_{k})} \\
& \le &
C\int (z_{l}+z_{l}^{2+\lambda})
z_1\cdots z_{l-1} z_{l+1} \cdots z_{k} r_{k}(t,d\bz_{k}) \\
& \le &
C E\left[(|Z(t)|+|Z(t)|^{2+\lambda})|Z(t)|^{k-1}\right]
\ = \ C \left(m_k+m_{1+\lambda+k}\right),
\end{eqnarray*}
where $C$ is a finite constant.
Second, we claim that, for $\Phi$ given by (\ref{3.7}) and $t>0$
\be
\int f(\bz_k)r_k(t,d\bz_k)
-\int f(\bz_k)r_k(0,d\bz_k)
=\int_0^tds E[L\Phi(Z(s))].
\label{3.9}
\ee
Define, for each $R>0$,
$\Phi^{(R)}(\bz)=\one_{\{|\bz|\le R\}}\Phi(\bz)$.
Then $\Phi^{(R)}\in B(\Omega)$ by Lemma 3.3 and
clearly
$L\Phi^{(R)}(\bz)=\one_{\{|\bz|\le R\}}L\Phi(\bz)$.
Furthermore, by virtue of Lemma 3.4,
$\one_{\{|\bz|\le R\}}L\Phi(\bz)$ is bounded
and so is $L\Phi^{(R)}$.
These observations together imply that
\begin{eqnarray*}
E\left[\Phi^{(R)}(Z(t))\right]
-E\left[\Phi^{(R)}(Z(0))\right]
&=&
\int_0^tds E\left[L\Phi^{(R)}(Z(s))\right] \\
&=&
\int_0^tds E\left[\one_{\{|Z(s)|\le R\}}L\Phi(Z(s))\right].
\end{eqnarray*}
Noting that every moment of
$|Z(t)|=|Z(0)|$ is finite
by the assumption and that $L\Phi\in\overline{\cF}$,
we get (\ref{3.9}) by taking the limit as $R\to\infty$
with the help of Lebesgue's convergence theorem.
Integrating the right side of (\ref{3.8})
with respect to the law of $Z(s)$
and then plugging the resulting expression for the expectation
$E[L\Phi(Z(s))]$ into (\ref{3.9}) yield
(\ref{3.1}) with $r_k(s,d\bz_k)$ in place of $c_k(s,d\bz_k)$.
We thus obtained the required equations
for \{$r_k(t,d\bz_k):~t\ge 0, k\in \N\}$
and the proof of Theorem 3.2 (i) is complete,
provided that the identity (\ref{3.8}) is shown by
calculations which are self-contained.
\medskip
\noindent
{\it Proof of (\ref{3.8}).}~
Recalling the definition (\ref{2.4}) of $L$,
we now calculate $L\Phi(\bz)$ for $\Phi$
of the form (\ref{3.7}).
Observe that, for any $i\ne j$ such that $z_iz_j>0$,
the `coagulation difference' $\Phi(M_{ij}\bz)-\Phi(\bz)$
equals
\begin{eqnarray*}
\lefteqn{\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i,j\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_i+z_j,z_{i_l},\ldots,z_{i_{k-1}})} \\
& - &
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i,j\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_i,z_{i_l},\ldots,z_{i_{k-1}}) \nonumber \\
& - &
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i,j\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_j,z_{i_l},\ldots,z_{i_{k-1}})
\nonumber \\
& - &
\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k-2}(\ne)}^{\{i,j\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_i,z_{i_l},\ldots,z_{i_{m-2}},z_j,
z_{i_{m-1}},\ldots,z_{i_{k-2}})
\nonumber \\
& -&
\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k-2}(\ne)}^{\{i,j\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_j,z_{i_l},\ldots,z_{i_{m-2}},z_i,
z_{i_{m-1}},\ldots,z_{i_{k-2}}) \nonumber \\
=: \ \
\lefteqn{\Sigma_{ij}^{(1)}(\bz)
-\Sigma_{ij}^{(2)}(\bz)
-\Sigma_{ij}^{(3)}(\bz)
-\Sigma_{ij}^{(4)}(\bz)
-\Sigma_{ij}^{(5)}(\bz),} \nonumber
\end{eqnarray*}
where $\sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i,j\}^c}$ stands for the
sum taken over $(k-1)$-tuples $(i_1,\ldots,i_{k-1})$
of positive integers such that $i_1,\ldots,i_{k-1}\in\{i,j\}^c$
are mutually distinct.
Noting that $\Sigma_{ij}^{(2)}(\bz)=\Sigma_{ji}^{(3)}(\bz)$ and
$\Sigma_{ij}^{(4)}(\bz)=\Sigma_{ji}^{(5)}(\bz)$, we get
\begin{eqnarray}
\lefteqn{\frac{1}{2}
\sum_{i\ne j}K(z_i,z_j)(\Phi(M_{ij}\bz)-\Phi(\bz))} \label{3.10} \\
& = &
\frac{1}{2}
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k+1}(\ne)}K(z_{i_l},z_{i_{l+1}})
f(z_{i_1},\ldots,z_{i_{l-1}},z_{i_l}+z_{i_{l+1}},z_{i_{l+2}},\ldots,z_{i_{k+1}}) \nonumber \\
& &
-\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}K(z_{i_l},z_{i_{l+1}})
f(z_{i_1},\ldots,z_{i_{l-1}},z_{i_l},z_{i_{l+2}},\ldots,z_{i_{k+1}}) \nonumber \\
& &
-\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k}(\ne)}K(z_{i_l},z_{i_m})
f(z_{i_1},\ldots,z_{i_{k}}). \nonumber
\end{eqnarray}
Similarly, for each $i\in\N$ with $z_i>0$
and any $y\in(0,z_i)$,
the `fragmentation difference' $\Phi(S_{i}^{(y)}\bz)-\Phi(\bz)$
is expressed as
\begin{eqnarray*}
\lefteqn{\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},y,z_{i_l},\ldots,z_{i_{k-1}})} \\
& + &
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_i-y,z_{i_l},\ldots,z_{i_{k-1}})
\nonumber \\
& + &
\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k-2}(\ne)}^{\{i\}^c}
f({\rm Frag}_{l,m}^{(y)}(z_{i_1},\ldots,z_{i_{l-1}},z_i,z_{i_l},\ldots,z_{i_{k-2}}))
\nonumber \\
& + &
\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k-2}(\ne)}^{\{i\}^c}
f({\rm Frag}_{l,m}^{(z_i-y)}(z_{i_1},\ldots,z_{i_{l-1}},z_i,z_{i_l},\ldots,z_{i_{k-2}}))
\nonumber \\
& - &
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i\}^c}
f(z_{i_1},\ldots,z_{i_{l-1}},z_i,z_{i_l},\ldots,z_{i_{k-1}}) \nonumber \\
=: \ \
\lefteqn{\Sigma_{i}^{(6)}(y,\bz)
+\Sigma_{i}^{(7)}(y,\bz)
+\Sigma_{i}^{(8)}(y,\bz)
+\Sigma_{i}^{(9)}(y,\bz)
-\Sigma_{i}^{(10)}(\bz)} \nonumber
\end{eqnarray*}
because for $1\le l<m\le k$
\begin{eqnarray*}
\lefteqn{{\rm Frag}_{l,m}^{(y)}(z_{i_1},\ldots,z_{i_{l-1}},
z_i,z_{i_l},\ldots,z_{i_{k-2}})} \\
&= &
(z_{i_1},\ldots,z_{i_{l-1}},y,z_{i_l},\ldots,z_{i_{m-2}},
z_i-y,z_{i_{m-1}},\ldots,z_{i_{k-2}})
\end{eqnarray*}
and
\begin{eqnarray*}
\lefteqn{{\rm Frag}_{l,m}^{(z_i-y)}(z_{i_1},\ldots,z_{i_{l-1}},
z_i,z_{i_l},\ldots,z_{i_{k-2}})} \\
& = &
(z_{i_1},\ldots,z_{i_{l-1}},z_i-y,z_{i_l},\ldots,z_{i_{m-2}},
y,z_{i_{m-1}},\ldots,z_{i_{k-2}}).
\end{eqnarray*}
Here, $\sum_{i_1,\ldots,i_{k-1}(\ne)}^{\{i\}^c}$
indicates the sum taken over $(k-1)$-tuples $(i_1,\ldots,i_{k-1})$
of distinct positive integers which are different from $i$.
By noting two identities
$\Sigma_{i}^{(6)}(y,\bz)=\Sigma_{i}^{(7)}(z_i-y,\bz)$
and $\Sigma_{i}^{(8)}(y,\bz)=\Sigma_{i}^{(9)}(z_i-y,\bz)$
\begin{eqnarray}
\lefteqn{\frac{1}{2}
\sum_iz_i\int_0^{z_i}dy F(y,z_i-y)(\Phi(S_i^{(y)}\bz)-\Phi(\bz))} \label{3.11} \\
& = &
\sum_{l=1}^k \sum_{i_1,\ldots,i_{k}(\ne)}z_{i_l}
\int_0^{z_{i_l}}dy F(y, z_{i_l}-y)
f(z_{i_1},\ldots,z_{i_{l-1}},y,z_{i_{l+1}},\ldots,z_{i_{k}})
\nonumber \\
& &
+\one_{\{k\ge 2\}}\sum_{l<m}^{k}
\sum_{i_1,\ldots,i_{k-1}(\ne)}z_{i_l}
\int_0^{z_{i_l}}dy F(y, z_{i_l}-y)
f({\rm Frag}_{l,m}^{(y)}(z_{i_1},\ldots,z_{i_{k-1}}))
\nonumber \\
& &
-\frac{1}{2}\sum_{l=1}^k \sum_{i_1,\ldots,i_{k}(\ne)}
z_{i_l}\int_0^{z_{i_l}}dy F(y, z_{i_l}-y) f(z_{i_1},\ldots,z_{i_k}). \nonumber
\end{eqnarray}
Consequently, (\ref{3.8}) is deduced from (\ref{3.10}) and (\ref{3.11}).
\qed
\medskip
\noindent
{\it Proof of Theorem 3.2 (ii).}~
The proof is essentially the same as the proof of (i).
We just give some comments.
The admissibility is shown similarly by $|X(0)|=1$.
Concerning the analogue of Lemma 3.4,
(\ref{3.8}) with $K_1$ and $F_1$
in place of $K$ and $F$, respectively, holds true
since the proof of (\ref{3.8}) itself is almost algebraic.
Moreover, the assertions corresponding to
the second half of Lemma 3.4 (i.e., the assertions for
$\Psi_l^{(i)}$'s and $\Psi_{l,m}^{(i)}$'s)
are also valid for $\lambda=0$
as is easily seen from the proof.
We complete the proof of Theorem 3.2 \qed
\medskip
So far we have discussed only weak solutions.
The final result in this section shows,
under a certain condition on the initial distribution,
the existence of
a strong solution to (\ref{1.2}) with
$\Hc=\theta\Hh$ for some constant $\theta>0$,
which ensures reversibility of the underlying process
as was shown in Theorem 2.4 (ii).
We will require also for the initial distribution
to be regarded as a `perturbation'
from some reversible distribution,
namely the law of a lifted ${\rm PD}(\theta)$ process.
Recall that a lifted $\rm{PD}(\theta)$ process is
of the form $\sum\delta_{VX_i}$, where
a $(0,\infty)$-valued random variable $V$
and a $\rm{PD}(\theta)$-distributed random element
$(X_i)_{i=1}^{\infty}$ of $\Omega_1$ are mutually independent.
\begin{pr
Let $L^{(H,\theta)}$ be as in (\ref{2.18}) and
$\{Z(t)=(Z_i(t))_{i=1}^{\infty}: t\ge 0\}$
be a process generated by $L^{(H,\theta)}$.
Set $\xi(t)=\Xi(Z(t))$ and
denote by $\{\cT_t\}_{t\ge 0}$ the semigroup
associated with $\{\xi(t): t\ge 0\}$.
Suppose that the law of $\xi(0)$ is
absolutely continuous with respect to
the law of a lifted $\rm{PD}(\theta)$ process
$\sum \delta_{VX_i}$. \\
(i) For any $k\in\N$ and $t\ge 0$,
the $k$th correlation function of $\xi(t)$ is given by
\be
r_k(t,\bz_k):=
r_{k}(z_1,\ldots,z_k)
\int_{\cN}P_{z_1,\ldots, z_k}(d\eta)(\cT_t\Psi^*)
(\eta+\delta_{z_1}+\cdots+\delta_{z_k}), \label{3.12}
\ee
where $r_{k}(z_1,\ldots,z_k)$ is
the $k$th correlation function (\ref{2.20})
of $\sum \delta_{VX_i}$,
$P_{z_1,\ldots,z_k}$ is the
$k$th-order reduced Palm distribution of $\sum \delta_{VX_i}$
at $\bz_k=(z_1,\ldots,z_k)$ characterized by (\ref{2.21})
and $\Psi^*$ is the density of
the law of $\xi(0)$ with respect to
the law of $\sum \delta_{VX_i}$.
(It is understood that $r_k(t,\bz_k)=0$
whenever $P(V>{|\bz_k|})=0$.) \\
(ii) Suppose additionally that
$E[|Z(0)|^k]<\infty$ for all $k\in\N$.
Then the family of nonnegative measurable functions
$\{r_k(t,\bz_k):~t\ge 0, k\in\N\}$ given in (i) solves
the equation (\ref{1.2}) with
$K(x,y)=xy H(x,y)$ and $F(x,y)=\theta(x+y) H(x,y)$
in the following sense: for any $k\in N$ and $t\ge 0$
\be
r_k(t,\bz_k)-r_k(0,\bz_k)-
\int_0^t \cL_k(s,\bz_k)ds=0, \ \mbox{a.e.} \
\bz_k\in(0,\infty)^k, \label{3.13}
\ee
where $\cL_k(s,\bz_k)$ is the right side of (\ref{1.2})
with $c_{k+1}$, $c_{k}$, $c_{k-1}$ and $t$ replaced by
$r_{k+1}$, $r_{k}$, $r_{k-1}$ and $s$, respectively.
\end{pr
{\it Proof.}~(i) Let $f\in B_+^k$ be arbitrary
and $\Phi$ be as in (\ref{3.7}). Thus,
by abuse of notation as in the previous section
\[
\Phi(\xi)
=\sum_{i_1,\ldots,i_{k}(\ne)}f(z_{i_1},\ldots,z_{i_k})
\]
for $(z_i)_{i=1}^{\infty}\in\Omega$ and
$\xi=\Xi((z_i)_{i=1}^{\infty})=\sum\one_{\{z_i>0\}} \delta_{z_i}$.
By the assumption of absolute continuity
together with the reversibility implied by Theorem 2.4 (ii)
\begin{eqnarray*}
\lefteqn{
E\left[\sum_{i_1,\ldots,i_k(\ne)}
f(Z_{i_1}(t),\ldots,Z_{i_k}(t))\right]
\ = \
E\left[\Phi(\xi(t))\right]
\ = \
E\left[(\cT_t\Phi)(\xi(0))\right]} \\
& = &
\int_{\cN} (\cT_t\Phi)(\eta)P\left(\xi(0)\in d\eta\right)
\ = \
\int_{\cN} (\cT_t\Phi)(\eta) \Psi^*(\eta)
P\left(\sum_i \delta_{VX_i}\in d\eta\right) \\
& = &
\int_{\cN} \Phi(\eta) (\cT_t\Psi^*)(\eta)
P\left(\sum_i \delta_{VX_i}\in d\eta\right) \\
& = &
\int f(\bz_k)
r_{k}(\bz_k)d\bz_k
\int_{\cN} P_{z_1,\ldots,z_k}(d\eta)
(\cT_t\Psi^*)(\eta+\delta_{z_1}+\cdots+\delta_{z_k}),
\end{eqnarray*}
where the last equality is deduced from
the Palm formula (\ref{2.8}) combined with Lemma 2.5.
This proves (\ref{3.12}). \\
(ii) As an immediate consequence of Theorem 3.2 (i)
we have
\[
\int f(\bz_k)\left[r_k(t,\bz_k)-r_k(0,\bz_k)
-\int_0^t \cL_k(s,\bz_k)ds\right]d\bz_k=0
\]
for all $f\in B_{+,c}^k$.
Replace $f(\bz_k)$ by $f(\bz_k)z_1\cdots z_k$ to get
\be
\int f(\bz_k)\left[r_k(t,\bz_k)-r_k(0,\bz_k)
-\int_0^t \cL_k(s,\bz_k)ds\right]z_1\cdots z_kd\bz_k=0.
\label{3.14}
\ee
It is easily verified from the assumption on the moments
of $|Z(0)|$ that the signed measure
\[
\left[r_k(t,\bz_k)-r_k(0,\bz_k)
-\int_0^t \cL_k(s,\bz_k)ds\right]z_1\cdots z_kd\bz_k
\]
is expressed as a linear combination of (at most)
8 finite measures on $(0,\infty)^k$. Therefore,
(\ref{3.14}) implies that it must vanish
and accordingly (\ref{3.13}) holds. \qed
\medskip
\noindent
{\it Example.}~
In the case where the lifted process
$\sum \delta_{VX_i}$ is a gamma point process with parameter
$(\theta,b)$ (see at the end of Section 2),
the absolute continuity assumption in Proposition 3.5 is
satisfied e.g. when $\xi(0)$ is a Poisson process
on $(0,\infty)$ with mean density
of the form $e^{h(y)}\theta e^{-by}y^{-1}$ and
$\int|e^{h(y)}-1|e^{-by}y^{-1}dy<\infty$.
In that case, the density $\Psi^*$ mentioned
in Proposition 3.5 (i) is given by
\[
\Psi^*(\eta)=
\exp\left[\lg h,\eta\rg-\theta\int(e^{h(y)}-1)
e^{-by}y^{-1}dy\right].
\]
(See e.g. Lemma 2.4 of \cite{B}.) Also,
since the reduced Palm distributions
of any Poisson process are identical with its law,
(\ref{3.12}) becomes
\[
r_k(t,\bz_k)=
\frac{\theta^k}{z_1 \cdots z_k}
e^{-b(z_1+\cdots+z_k)}
E\left[ (\cT_t\Psi^*)
(\eta+\delta_{z_1}+\cdots+\delta_{z_k})\right],
\]
where $\eta$ is
a gamma point process with parameter $(\theta,b)$.
\section{Preliminary results for rescaled processes
\setcounter{equation}{0}
\subsection{Models with a scaling parameter}
Both this section and the subsequent section
are devoted to a derivation of the equation (\ref{1.1})
from properly rescaled coagulation-fragmentation processes.
In principle, the procedure is similar to that in \cite{EW00}
although that paper assumes
the conditions, among others, of the form
\[
K(x,y)= o(x)o(y) \quad \mbox{and} \quad
\int_0^xF(y,x-y)dy=o(x) \quad
\mbox{as} \quad x,y\to \infty
\]
for the rates.
In our situation, these conditions are never met
since by (\ref{1.4}) and (\ref{1.5})
\[
K(x,y)=xy(x+y)^{\lambda}
\Hh\left(\frac{x}{x+y},\frac{y}{x+y}\right)
\quad \mbox{and} \quad
\int_0^xF(y,x-y)dy=\Cc x^{2+\lambda}.
\]
However, it will turn out that our setting on the degrees of
homogeneity of $K$ and $F$ provides with us certain
effective ingredients to overcome difficulties due to such growth orders. More specifically, it will play
an essential role in the proof of Proposition 4.5 below.
In this connection we mention that
\cite{VZ} discussed the relation
between occurrence of `steady-state solutions'
for coagulation-fragmentation equations and
the degrees of homogeneity. According to
the authors' criterion based on analysis
of the moments in several basic examples
our setting on the degrees is in the region
corresponding to systems for which
steady states occur.
Let us specify the model we are concerned with
in the rest of the paper. Following \cite{EW00},
we introduce a scaling parameter $N=1,2,\ldots$
and modify the generator $L$ in (\ref{2.4})
by replacing $K$ by $K/N$.
To be more precise, define
\begin{eqnarray}
L^N\Phi(\bz)
& = &
\frac{1}{2N}\sum_{i\ne j} z_iz_j \Hh(z_i,z_j)
\left(\Phi(M_{ij}\bz)-\Phi(\bz)\right) \nonumber \\
& &
+\frac{1}{2}\sum_i z_i
\int_0^{z_i}dy \Hc(y,z_i-y)
\left(\Phi(S_i^{(y)}\bz)-\Phi(\bz)\right) \label{4.1}
\end{eqnarray}
and denote by
$\{Z^N(t)=(Z_i^N(t))_{i=1}^{\infty}:~t\ge 0\}$
the process generated by $L^N$.
The rescaled process we will study actually is
\[
\xi^N(t):=\frac{1}{N}\Xi(Z^N(t))
=\frac{1}{N}\sum_{i}\one_{(0,\infty)}(Z_i^N(t))
\delta_{Z_i^N(t)}.
\]
This process is regarded as a process taking values in
$\cM$, the totality of locally finite measures on $(0,\infty)$.
$\cM$ is equipped with the vague topology.
For $a\ge 0$, let $\psi_a$ denote the power function
$\psi_a(y)=y^a$, so that $\lg\psi_a, \zeta\rg$ stands for
the `$a$th moment' of $\zeta\in \cM$.
We consider $c_0\in\cM$ such that for some $\delta>1$
\be
0< \langle \psi_1, c_0 \rangle <\infty \quad
\mbox{and} \quad
\langle \psi_{2+\lambda+\delta}, c_0 \rangle <\infty. \label{4.2}
\ee
Note that (\ref{4.2}) implies that
$\lg\psi_a, c_0\rg<\infty$ for any $a\in[1,2+\lambda+\delta]$.
Concerning initial distributions of $\{Z^N(t):~t\ge 0\}$,
we suppose that
\be
E\left[\left|Z^N(0)\right|^{2+\lambda+\delta}\right]<\infty
\quad \mbox{for each $N=1,2,\ldots$}. \label{4.3}
\ee
It shall be required also
that $\xi^N(0)$ converges to $c_0$
in distribution as $N\to\infty$.
A typical case where such a convergence holds is
given in the following lemma, which is stated
in a general setting.
In the rest, ${\rm Po}(m)$ stands for the law of
a Poisson point process on $(0,\infty)$
with mean measure $m\in\cM$.
\begin{lm
Let $\zeta \in \cM$ be arbitrary.
Assume that $\sum \delta_{Y_i^N}$ is
${\rm Po}(N\zeta)$-distributed for each $N=1,2,\ldots$.
Then $\eta^N:=N^{-1}\sum \delta_{Y_i^N}$
converges to $\zeta$ in distribution as $N\to\infty$.
Moreover, if $\lg\psi_1, \zeta\rg<\infty$ and
$\lg\psi_a, \zeta\rg<\infty$ for some $a>1$, then
\be
E\left[\lg\psi_1, \eta^N\rg^a\right]
=E\left[\left(\frac{\sum Y_i^N}{N}\right)^a\right]
\to \lg\psi_1, \zeta\rg^a
\quad \mbox{as} \quad N\to\infty. \label{4.4}
\ee
In particular, $\sup_N
E\left[\left(N^{-1}\sum Y_i^N\right)^a\right] <\infty$.
\end{lm
Since the proof of Lemma 4.1 is rather lengthy and
not relevant to other parts of this paper,
the proof will be given in Appendix.
It is worth noting here that requiring
the law of $\sum \delta_{Y_i^N}$ to be Poisson
automatically implies that every correlation measure
of it is the direct product of the mean measure.
By looking at the limit points of
$\{\xi^N(t):~t\ge 0\}$ as $N\to\infty$
we intend to derive a weak solution to (\ref{1.1})
with $K$ and $F$ given by (\ref{1.4}), namely
\begin{eqnarray}
\lefteqn{
\frac{\partial}{\partial t}c(t,x)} \label{4.5} \\
& = &
\frac{1}{2}\int_0^x\left[y(x-y)\Hh(y,x-y)c(t,y)c(t,x-y)
-x \Hc(y,x-y)c(t,x)\right]dy \nonumber \\
& &
-\int_0^{\infty}\left[xy\Hh(x,y)c(t,x)c(t,y)
-(x+y)\Hc(x,y)c(t,x+y)\right]dy. \nonumber
\end{eqnarray}
One may realize that in the case
where both $\Hh$ and $\Hc$ are constants,
$a$ and $b$, say, respectively,
(\ref{4.5}) can be transformed by considering
the `size-biased version' $c^{\star}(t,x):=xc(t,x)$ into
\begin{eqnarray*}
x^{-1}\frac{\partial}{\partial t}c^{\star}(t,x)
& = &
\frac{1}{2}\int_0^x\left[a c^{\star}(t,y)c^{\star}(t,x-y)
-b c^{\star}(t,x)\right]dy \\
& &
-\int_0^{\infty}\left[a c^{\star}(t,x)c^{\star}(t,y)
-b c^{\star}(t,x+y)\right]dy.
\end{eqnarray*}
This equation is very similar to (\ref{1.1})
with $K$ and $F$ being constants, the solution of
which has been studied extensively in
\cite{AB} and \cite{SD}. However,
it is not clear whether or not there is any direct
connection between solutions of these two equations.
Turning to (\ref{4.5}), the weak form
with test functions $f\in B_{c}$
and initial measure $c_0$ reads
\begin{eqnarray*}
\lefteqn{
\int f(x)c(t,dx) -\int f(x)c_0(dx)} \\
&=&
\frac{1}{2}\int_0^tds
\int c(s,dx)c(s,dy)xy \Hh(x,y) (\Box f)(x,y) \\
& &
-\frac{1}{2}\int_0^tds
\int c(s,dx) x\int_0^{x}dy\Hc(y,x-y) (\Box f)(y,x-y),
\end{eqnarray*}
where $(\Box f)(x,y)=f(x+y)-f(x)-f(y)$.
A rough idea for its derivation can be described as follows.
For $f\in B_+\cup B_c$ and $\bz\in\Omega$, set
$\Phi_f(\bz)=\sum f(z_i)$,
adopting the convention that $f(0)=0$.
By calculating $L^N\Phi_f$ (cf. (\ref{3.8}) with $k=1$),
it is observed that for any $f\in B_{c}$
\[
M_f^N(t):=
\lg f, \xi^N(t) \rg -\lg f, \xi^N(0) \rg
-\frac{1}{N}\int_0^t L^N\Phi_f(Z^N(s))ds
\]
is a martingale and
\begin{eqnarray}
\lefteqn{
\lg f,\xi^N(t)\rg -\lg f,\xi^N(0)\rg-M_f^N(t)} \label{4.6} \\
&=&
\frac{1}{2}\int_0^tds
\int \xi^N(s)^{[2]}(dxdy)xy \Hh(x,y) (\Box f)(x,y) \nonumber \\
& &
-\frac{1}{2}\int_0^tds
\int \xi^N(s)(dx)x\int_0^{x}dy \Hc(y,x-y) (\Box f)(y,x-y), \nonumber
\end{eqnarray}
where
\[
\xi^N(s)^{[2]}
=\frac{1}{N^2}\sum_{i\ne j}
\one_{(0,\infty)}(Z_i^N(s)Z_j^N(s))
\delta_{(Z_i^N(s),Z_j^N(s))}.
\]
(By virtue of Lemma 3.4 with $k=1$, the
integrability of $M^N_f(t)$ is ensured by (\ref{4.3}).)
As far as the limit as $N\to\infty$ is concerned,
$\xi^N(s)^{[2]}$ in the right side of (\ref{4.6})
can be replaced by $\xi^N(s)^{\otimes 2}$
under a suitable assumption on the convergence of
$\xi^N(0)$. Indeed, letting $R>0$ be such that
${\rm supp}(f) \subset [0,R]$, we have by (\ref{3.4})
\begin{eqnarray*}
\lefteqn{
\left|\int \left(\xi^N(s)^{[2]}
-\xi^N(s)^{\otimes 2}\right)(dxdy)
xy \Hh(x,y) (\Box f)(x,y) \right|} \\
& = &
\left|\frac{1}{N^2}
\sum_{i} Z_i^N(s)^2 \Hh(Z_i^N(s),Z_i^N(s)) \left\{f(2Z_i^N(s))-2f(Z_i^N(s))\right\}\right| \\
& \le &
\frac{\Ch}{N^2}
\sum_{i} Z_i^N(s)^2 (2Z_i^N(s))^{\lambda}
\left|f(2Z_i^N(s))-2f(Z_i^N(s))\right| \\
& \le &
\frac{\Ch R(2R)^{\lambda}}{N^2}
\sum_{i} Z_i^N(s) \cdot 3\Vert f \Vert_{\infty} \\
& = & \frac{3\Ch2^{\lambda}R^{1+\lambda}}{N}
\Vert f \Vert_{\infty} \lg \psi_1,\xi^N(0)\rg.
\end{eqnarray*}
Thus, at least at formal level,
the derivation of a weak solution to (\ref{4.5})
from $\xi^N(t)$ would reduce to proving that
$M_f^N(t)$ vanishes
in a suitable sense as $N\to \infty$. \par
To this end, we shall calculate the
quadratic variation $\lg M_f^N\rg (t)$.
However, the square integrability of $M_f^N(t)$
is nontrivial under the condition (\ref{4.3}).
We will guarantee this in the next lemma
by the cutoff argument as in the proof of Theorem 3.2.
Also, the following inequality will be used to bound
the quadratic variation: for any $s,t\ge 0$ and $a\ge 0$
\be
(s+t)^a \le C_{1,a}(s^a+t^a), \label{4.7}
\ee
where $C_{1,a}=2^{a-1}\vee 1$.
This inequality for $a>1$ is deduced from
H\"older's inequality and the one for $0<a\le 1$ is
implied by the identity
\be
(s+t)^a-s^a-t^a
=a(a-1)\int_0^sdu\int_0^tdv(u+v)^{a-2}. \label{4.8}
\ee
\begin{lm
Suppose that
$E\left[\left|Z^N(0)\right|^{2+\lambda}\right]<\infty$.
For each $f\in B_{c}$,
$\{M_f^N (t):~t\ge 0\}$ is a square integrable
martingale with quadratic variation
\[
\lg M_f^N\rg (t)
=
\int_0^t \Gamma^N_f(Z^N(s))ds,
\]
where $\Gamma^N_f: \Omega\to \R_+$ is defined to be
\begin{eqnarray}
\Gamma^N_f(\bz)
& = &
\frac{1}{N^2}\left\{L^N((\Phi_f)^2)(\bz)
-2\Phi_f(\bz)L^N\Phi_f(\bz)\right\} \nonumber \\
& = &
\frac{1}{2N^3}\sum_{i\ne j} z_iz_j \Hh(z_i,z_j)
\left\{f(z_i+z_j)-f(z_i)-f(z_j)\right\}^2 \label{4.9} \\
& &
+\frac{1}{2N^2}\sum_i z_i^{2}
\int_0^{1}du \Hc(uz_i,(1-u)z_i)
\left\{f(uz_i)+f((1-u)z_i)-f(z_i)\right\}^2 \nonumber \\
& =: &
\frac{1}{2}\Sigma_K^N(\bz)
+\frac{1}{2}\Sigma_F^N(\bz). \nonumber
\end{eqnarray}
Moreover, the following estimates hold:
\be
\Sigma_K^N(\bz)
\le 18\Ch C_{1,\lambda}
\frac{\Vert f \Vert_{\infty}^2}{N}\cdot\frac{|\bz|}{N}
\sum_i\frac{z_i^{1+\lambda}}{N}
= 18\Ch C_{1,\lambda}\frac{\Vert f \Vert_{\infty}^2}{N}
\langle \psi_1,\xi^N \rangle
\langle \psi_{1+\lambda},\xi^N \rangle \label{4.10}
\ee
and
\be
\Sigma_F^N(\bz)
\le 9\Cc\frac{\Vert f \Vert_{\infty}^2}{N}
\sum_i\frac{z_i^{2+\lambda}}{N}
= 9\Cc\frac{\Vert f \Vert_{\infty}^2}{N}
\langle \psi_{2+\lambda},\xi^N \rangle, \label{4.11}
\ee
where $\xi^N=N^{-1}\Xi(\bz)
=N^{-1}\sum \one_{\{z_i>0\}}\delta_{z_i}$.
\end{lm
{\it Proof.}~
For any $R>0$, let
$\Phi_f^{(R)}(\bz)=\Phi_f (\bz)\one_{\{|\bz|\le R\}}$
and observe that
\begin{eqnarray*}
M_f^{N,R}(t)
& := &
\frac{1}{N}\Phi_f^{(R)}(Z^N(t))
-\frac{1}{N}\Phi_f^{(R)}(Z^N(0))
-\frac{1}{N}\int_0^t L^N\Phi_f^{(R)}(Z^N(s))ds \\
& = &
M_f^{N}(t) \one_{\{|Z^N(0)|\le R\}}
\end{eqnarray*}
is a bounded martingale with quadratic variation
\begin{eqnarray*}
\lg M_f^{N,R} \rg (t)
& = &
\frac{1}{N^2} \int_0^t
\left[L^N((\Phi_f^{(R)})^2)(Z^{N}(s))
-2\Phi_f^{(R)}(Z^{N}(s)) L^N\Phi_f^{(R)}(Z^{N}(s)) \right]ds \\
& = &
\int_0^t \Gamma_f^N(Z^N(s))ds
\one_{\{|Z^N(0)|\le R\}}.
\end{eqnarray*}
The expression (\ref{4.9}) for $\Gamma_f^N(\bz)$
is deduced from (\ref{4.1}). By (\ref{3.4}) and (\ref{4.7})
\begin{eqnarray*}
\Sigma_K^N(\bz)
& \le &
9\Ch C_{1,\lambda}\Vert f \Vert_{\infty}^2\sum_{i\ne j}
\frac{z_iz_j }{N^3} (z_i^{\lambda}+z_j^{\lambda}) \\
& = &
18\Ch C_{1,\lambda}\Vert f \Vert_{\infty}^2\sum_{i\ne j}
\frac{z_i^{1+\lambda}z_j }{N^3} \\
& \le &
18\Ch C_{1,\lambda}\frac{\Vert f \Vert_{\infty}^2}{N}
\cdot\frac{|\bz|}{N}\sum_i\frac{z_i^{1+\lambda}}{N}.
\end{eqnarray*}
This proves (\ref{4.10}), whereas
(\ref{4.11}) is immediate from (\ref{3.6}).
The assumption together with these two estimates
ensure the integrability of
$\int_0^t\Gamma_f^N(Z^N(s))ds$. Therefore,
by the monotone convergence theorem
\begin{eqnarray*}
E\left[(M_f^{N}(t))^2 \right]
& = &
\lim_{R\to\infty}E\left[(M_f^{N,R}(t))^2 \right] \\
& = &
\lim_{R\to\infty}
E\left[\int_0^t \Gamma_f^N(Z^N(s))ds
\one_{\{|z|\le R\}}\right] \\
& = &
E\left[\int_0^t \Gamma_f^N(Z^N(s))ds \right]<\infty.
\end{eqnarray*}
Once the square integrability of $M_f^{N}(t)$
is in hand,
one can show further by a similar argument that
$(M_f^{N}(t))^2-\int_0^t \Gamma_f^N(Z^N(s))ds$ is
a martingale. The proof of Lemma 4.2 is complete.
\qed
\medskip
\noindent
{\it Remarks.}~ (i) A heuristic derivation of (\ref{1.1})
based on the hierarchical structure discussed
in the previous section is available:
due to the rescaling $K \mapsto K/N$ and
$\xi \mapsto \xi^N :=\xi /N$, the equation
solved weakly by the family
$\{r_k^N(t,d\bz_k):~t\ge0, k\in\N\}$
of correlation measures of $\xi^N(t)$ becomes
\begin{eqnarray*}
\lefteqn{
\frac{\partial}{\partial t}r_k^N(t,z_1,\ldots,z_k)} \\
&=&
\frac{1}{2}\sum_{l=1}^k\int_0^{z_l}K(y,z_l-y)
r^N_{k+1}(t,z_1,\ldots,z_{l-1},y,z_l-y,z_{l+1},\ldots,z_k)dy \\
& &
-\frac{1}{2}\sum_{l=1}^k\int_0^{z_l}F(y,z_l-y)dy \
r^N_{k}(t,z_1,\ldots,z_k) \\
& &
-\sum_{l=1}^k\int_0^{\infty}K(z_l,y)
r^N_{k+1}(t,z_1,\ldots,z_l,y,z_{l+1},\ldots,z_k)dy \\
& &
+\sum_{l=1}^k\int_0^{\infty}F(z_l,y)
r^N_{k}(t,z_1,\ldots,z_{l-1},z_l+y,z_{l+1},\ldots,z_k)dy \\
& &
-\frac{\one_{\{k\ge 2\}}}{N}
\sum_{l<m}^{k}K(z_l,z_m)r^N_{k}(t,z_1,\ldots,z_k) \\
& &
+\frac{\one_{\{k\ge 2\}}}{N}\sum_{l<m}^{k}F(z_l,z_m)
r^N_{k-1}(t,z_1,\ldots,z_{l-1},z_l+z_m,z_{l+1},\ldots,
z_{m-1},z_{m+1},\ldots,z_k).
\end{eqnarray*}
So, the last two terms would be expected to
vanish in the limit as $N\to \infty$ and
the limits $r_k$ of $r_k^N$, if they exist,
would solve the same equations
as the ones satisfied by the direct products
$c^{\otimes k}(t,z_1,\ldots,z_k)
=c(t,z_1)\cdots c(t,z_k)$ of a solution to (\ref{1.1}).
This procedure has been accomplished in
\cite{EP} for a pure coagulation model. \\
(ii) We also give a remark
on the asymptotic equivalence
between the moment measures and
the correlation measures of the rescaled process
$\xi^N(t)$. (cf. Lemma 1.16 in \cite{Kol}.
The reader is cautioned that our terminology
`moment measure' is in conflict with that of \cite{Kol}.)
For each $k\in\{2,3,\ldots\}$,
under the assumption that
$\sup_N E[(|Z^N(0)|/N)^{k-1}]<\infty$,
it holds that for any $f\in B_{+,c}^k$
\[
\left|E\left[\frac{1}{N^k}\sum_{i_1,\ldots,i_k}
f(Z_{i_1}^N(t),\ldots,Z_{i_k}^N(t))\right]
-E\left[\frac{1}{N^k}\sum_{i_1,\ldots,i_k(\ne)}
f(Z_{i_1}^N(t),\ldots,Z_{i_k}^N(t))\right]\right|
\le \frac{C}{N}
\]
for some constant $C$ independent of $N$.
Indeed, it is sufficient to verify this by assuming that
$f(z_1,\ldots,z_k)=
z_1\cdots z_k\one_{[\epsilon,R]}(z_1)\cdots
\one_{[\epsilon,R]}(z_k)$ with $0<\epsilon<R$,
for which the above difference is dominated by
a finite sum of expectations of the form
\[
E\left[\frac{1}{N^k}\sum_{i_1,\ldots,i_j}
Z_{i_1}^N(t)^{n_1}\one_{[\epsilon,R]}(Z_{i_1}^N(t))
\cdots
Z_{i_j}^N(t)^{n_j}\one_{[\epsilon,R]}(Z_{i_j}^N(t))\right],
\]
where $j\in\{1,\ldots,k-1\}$, $n_1,\ldots,n_j\in \N$
is such that $n_1+\cdots+n_j=k$
and hence $n_{l}\ge 2$ for some $l\in\{1,\ldots,j\}$.
The desired bound follows by noting that
\[
Z_{i_l}^N(t)^{n_l}\one_{[\epsilon,R]}(Z_{i_l}^N(t))
\le
R Z_{i_l}^N(t)^{n_l-1}\one_{[\epsilon,R]}(Z_{i_l}^N(t))
\]
and observing that the above expectation is
less than or equal to
\[
E\left[\frac{R}{N^k}|Z^N(t)|^{k-1}\right]
=\frac{R}{N}
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{k-1}\right].
\]
\subsection{Key estimates for the martingale
Lemma 4.2 implies that for some constant
$C_2>0$ independent of $N, t$ and $f$
\begin{eqnarray}
\lefteqn{
C_2 \Vert f \Vert_{\infty}^{-2}
E\left[\lg M_f^N\rg (t)\right]} \nonumber \\
& \le &
\frac{1}{N}E\left[\frac{|Z^N(0)|}{N}\int_0^t ds
\sum_i\frac{Z^N_i(s)^{1+\lambda}}{N} \right]
+ \frac{1}{N}E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{2+\lambda}}{N} \right]. \label{4.12}
\end{eqnarray}
In fact, the two expectations on (\ref{4.12}) are
related to each other in such a way that
\begin{eqnarray}
\lefteqn{
E\left[\frac{|Z^N(0)|}{N}\int_0^t ds
\sum_i\frac{Z^N_i(s)^{1+\lambda}}{N} \right]} \label{4.13} \\
& \le &
\left(t
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{2+\lambda}\right]
\right)^{\frac{1}{1+\lambda}}
\left(E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{2+\lambda}}{N}\right]
\right)^{\frac{\lambda}{1+\lambda}}, \nonumber
\end{eqnarray}
which is a special case
($\alpha=1, \gamma=\lambda$ and $\epsilon=1$)
of the following lemma.
\begin{lm
For arbitrary $\alpha, \gamma \ge 0$ and $\epsilon>0$
\begin{eqnarray}
\lefteqn{
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{\alpha}\int_0^t ds
\sum_i\frac{Z^N_i(s)^{1+\gamma}}{N} \right]} \label{4.14} \\
& \le &
\left(t
E\left[\left(\frac{|Z^N(0)|}{N}
\right)^{1+\alpha+\frac{\alpha\gamma}{\epsilon}}\right]
\right)^{\frac{\epsilon}{\gamma+\epsilon}}
\left(E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{1+\gamma+\epsilon}}{N}\right]
\right)^{\frac{\gamma}{\gamma+\epsilon}}. \nonumber
\end{eqnarray}
\end{lm
{\it Proof.}~
Since (\ref{4.14}) for $\gamma=0$ is obvious,
we may assume that $\gamma>0$.
Then put $p=1+\gamma/\epsilon$
and $q=1+\epsilon/\gamma$, which are mutually conjugate.
By virtue of H\"older's inequality with respect to
the `weight' $E[N^{-1}\int_0^tds \sum_i \cdot ]$
\begin{eqnarray*}
\lefteqn{
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{\alpha}\int_0^t ds
\sum_i\frac{Z^N_i(s)^{1+\gamma}}{N} \right]} \\
& = &
E\left[\int_0^t \frac{ds}{N}\sum_i
\left(\frac{|Z^N(0)|^{\alpha}}{N^{\alpha}}
Z^N_i(s)^{\frac{1}{p}}\right)
Z^N_i(s)^{1+\gamma-\frac{1}{p}}\right] \\
& \le &
\left(E\left[\int_0^t \frac{ds}{N}\sum_i
\left(\frac{|Z^N(0)|}{N}\right)^{\alpha p}Z^N_i(s)\right]
\right)^{\frac{1}{p}}
\left(E\left[\int_0^t \frac{ds}{N} \sum_i
Z^N_i(s)^{(1+\gamma-\frac{1}{p})q}\right]
\right)^{\frac{1}{q}} \\
& = &
\left(t
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{\alpha p+1}\right]
\right)^{\frac{\epsilon}{\gamma+\epsilon}}
\left(E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{1+\gamma+\epsilon}}{N}\right]
\right)^{\frac{\gamma}{\gamma+\epsilon}}.
\end{eqnarray*}
Here, the final equality is due to
$(1+\gamma-\frac{1}{p})q=1+\gamma+\epsilon$.
(\ref{4.14}) has been obtained
since $\alpha p+1=1+\alpha+\alpha\gamma/\epsilon$.
\qed
\medskip
The expression in the right side of (\ref{4.13})
motivates us to define
\[
\overline{m}_{a}=
\sup_N E\left[\left(\frac{|Z^N(0)|}{N}\right)^{a}\right]
=\sup_N E\left[\lg \psi_1,\xi^N(0)\rg^a \right]
\]
for $a\ge 0$. We will discuss under the condition that
\be
\overline{m}_{2+\lambda+\delta}< \infty. \label{4.15}
\ee
This condition is stronger than (\ref{4.3}) and
valid in the case of Poisson processes
considered in Lemma 4.1. So, we shall focus
attention on estimation of the second term of (\ref{4.12}).
Before doing it, we prepare an elementary but
technically important lemma.
\begin{lm
Let $A,B>0$, a real number $C$ and $q>1$ be given.
If $A$ satisfies a `self-dominated inequality'
$A \le B A^{1/q}+ C$, then
\[
A
\le B^{\frac{q}{q-1}}+\frac{q}{q-1}(C \vee 0).
\]
\end{lm
{\it Proof.}~ Since the assertion is obvious
when $C\le 0$, we only have to consider the case $C>0$.
Denote by $\wt{A}$ the right side of the required inequality.
The equation $x=B x^{1/q}+C$ for $x>0$
has a unique root and
$x>B x^{1/q}+C$ for sufficiently large $x$.
Therefore, it is enough to show that
$\wt{A} > B \wt{A}^{1/q}+C.$
Observe that for $s,t>0$
\[
(s+t)^{\frac{1}{q}}-s^{\frac{1}{q}}
\ = \ \frac{1}{q} \int_0^tdu(s+u)^{\frac{1}{q}-1}
\ < \ \frac{1}{q} \int_0^tdu s^{\frac{1}{q}-1}
\ = \ \frac{1}{q}s^{\frac{1}{q}-1}t
\]
and thus $(s+t)^{1/q}<s^{1/q} + s^{(1/q)-1}t/q$.
Plugging $s=B^{\frac{q}{q-1}}$
and $t=\frac{q}{q-1}C$ gives
\begin{eqnarray*}
B\wt{A}^{1/q}+C
& < &
B\left(B^{\frac{1}{q-1}}
+B^{-1}\frac{q}{q-1}C\cdot\frac{1}{q}\right)+C \\
& = &
B^{\frac{q}{q-1}}+\frac{q}{q-1}C
\ = \
\wt{A}
\end{eqnarray*}
as desired.
\qed
\medskip
\noindent
The next proposition supplies the key to proceeding further.
\begin{pr
Assume that (\ref{4.2}) holds for some $\delta>1$.
Suppose that $\sum \delta_{Z_i^N(0)}$
has mean measure $Nc_0$ for each $N=1,2,\ldots$
and that (\ref{4.15}) holds.
Then, for each $t>0$, there exist constants
$\overline{C}(t)$, $\overline{C}_0(t)$ and $C^*(t)$
independent of $N$ such that
\be
E\left[\int_0^t ds
\langle \psi_{2+\lambda+\delta},\xi^N(s) \rangle \right]
=
E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{2+\lambda+\delta}}{N} \right]
\le \overline{C}(t) \label{4.16}
\ee
\be
E\left[\int_0^t ds
\langle \psi_{2+\lambda},\xi^N(s) \rangle \right]
=
E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{2+\lambda}}{N} \right]
\le \overline{C}_0(t) \label{4.17}
\ee
and
\be
E\left[\langle \psi_{\delta},\xi^N(t) \rangle \right]
=
E\left[\sum_i
\frac{Z^N_i(t)^{\delta}}{N} \right] \le C^*(t). \label{4.18}
\ee
\end{pr
{\it Proof.}~
Define bounded functions $\Psi_{\delta}^{(R)}(\bz)=
\sum_i z_i^{\delta}\one_{\{|\bz|\le R\}}$ on $\Omega$
for all $R>0$. By direct calculations
\begin{eqnarray*}
L^N\Psi_{\delta}^{(R)}(\bz)
& = &
\frac{1}{2N}\sum_{i\ne j}z_iz_j \Hh(z_i,z_j)
\left\{(z_i+z_j)^{\delta}-z_i^{\delta}-z_j^{\delta}\right\}
\one_{\{|\bz|\le R\}} \\
& &
+\frac{1}{2}\sum_i z_i^{2+\lambda+\delta}
\int_0^{1}du \Hc(u,1-u)
\left\{u^{\delta}+(1-u)^{\delta}-1\right\}
\one_{\{|\bz|\le R\}}.
\end{eqnarray*}
A crucial point here is that by $\delta>1$ and (\ref{4.8})
\[
\check{C}_{\delta}:=
\int_0^{1}du \Hc(u,1-u)\left\{1-u^{\delta}-(1-u)^{\delta}\right\}\in (0,\Cc).
\]
Taking expectation of the martingale
\[
\Psi_{\delta}^{(R)}(Z^N(t))
-\Psi_{\delta}^{(R)}(Z^N(0))
-\int_0^tds L^N\Psi_{\delta}^{(R)}(Z^N(s))
\]
and then letting $R\to \infty$ yield
\begin{eqnarray*}
\lefteqn{
\frac{\check{C}_{\delta}}{2}E\left[\int_0^t ds
\sum_i Z^N_i(s)^{2+\lambda+\delta} \right]
+E\left[\sum_i Z^N_i(t)^{\delta} \right]} \\
& = &
E\left[\sum_i Z^N_i(0)^{\delta} \right]
+ \frac{1}{2N}E\left[\int_0^t ds
\sum_{i \ne j} Z^N_i(s)Z^N_j(s)
\Hh(Z^N_i(s),Z^N_j(s)) \right. \\
& &
\hspace*{5cm}
\left. \cdot \left\{(Z^N_i(s)+Z^N_j(s))^{\delta}
-Z^N_i(s)^{\delta}-Z^N_j(s)^{\delta}\right\} \right] \\
& \le &
N\lg \psi_{\delta}, c_0\rg
+ \frac{\Ch}{2N}E\left[\int_0^t ds
\sum_{i \ne j} Z^N_i(s)Z^N_j(s)
(Z^N_i(s)+Z^N_j(s))^{\lambda+\delta} \right] \\
& \le &
N\lg \psi_{\delta}, c_0\rg
+ \frac{\Ch C_{1,\lambda+\delta}}{2N}
E\left[\int_0^t ds
\sum_{i \ne j} Z^N_i(s)Z^N_j(s)
\left\{Z^N_i(s)^{\lambda+\delta}
+Z^N_j(s)^{\lambda+\delta}\right\} \right] \\
& \le &
N\lg \psi_{\delta}, c_0\rg
+ \frac{\Ch C_{1,\lambda+\delta}}{N}
E\left[|Z^N(0)| \int_0^t ds
\sum_{i} Z^N_i(s)^{1+\lambda+\delta} \right].
\end{eqnarray*}
Note that all the expectations in the above are finite
by (\ref{4.15}). We have obtained
\begin{eqnarray}
\lefteqn{
\frac{\check{C}_{\delta}}{2}E\left[\int_0^t ds
\sum_i \frac{Z^N_i(s)^{2+\lambda+\delta}}{N} \right]
+E\left[\sum_i \frac{Z^N_i(t)^{\delta}}{N} \right]} \nonumber \\
& \le &
\Ch C_{1,\lambda+\delta}
E\left[\frac{|Z^N(0)|}{N} \int_0^t ds
\sum_{i} \frac{Z^N_i(s)^{1+\lambda+\delta}}{N} \right]
+\lg \psi_{\delta}, c_0\rg . \label{4.19}
\end{eqnarray}
In addition, Lemma 4.3 with
$\alpha=1, \gamma=\lambda+\delta$ and $\epsilon=1$ reads
\begin{eqnarray}
\lefteqn{
E\left[\frac{|Z^N(0)|}{N}\int_0^t ds
\sum_i\frac{Z^N_i(s)^{1+\lambda+\delta}}{N} \right]} \nonumber \\
& \le &
\left(t
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{2+\lambda+\delta}\right]
\right)^{\frac{1}{1+\lambda+\delta}}
\left(E\left[\int_0^t ds \sum_i
\frac{Z^N_i(s)^{2+\lambda+\delta}}{N}\right]
\right)^{\frac{\lambda+\delta}{1+\lambda+\delta}}. \label{4.20}
\end{eqnarray}
Set $p=1+\lambda+\delta$ and $q=1+1/(\lambda+\delta)$.
By combining the above two inequalities
\[
A_N\ := \ E\left[\int_0^t ds
\sum_i \frac{Z^N_i(s)^{2+\lambda+\delta}}{N} \right]
\le
B_N\left(A_N\right)^{1/q}
+\frac{2\lg \psi_{\delta}, c_0\rg }{\check{C}_{\delta}},
\]
where
\[
B_N
=\frac{2\Ch C_{1,\lambda+\delta}}{\check{C}_{\delta}}
\left(t
E\left[\left(\frac{|Z^N(0)|}{N}\right)^{2+\lambda+\delta}\right]
\right)^{1/p}.
\]
Applying Lemma 4.4 and (\ref{4.15}) and noting that
$\frac{q}{q-1}=p$, we obtain
\begin{eqnarray*}
A_N
& \le &
\left(B_N\right) ^{\frac{q}{q-1}}+ \frac{q}{q-1}\cdot
\frac{2\lg \psi_{\delta}, c_0\rg }{\check{C}_{\delta}} \\
& \le &
\left(\frac{2\Ch C_{1,\lambda+\delta}}{\check{C}_{\delta}}
\right)^{1+\lambda+\delta}
\overline{m}_{2+\lambda+\delta} t
+\frac{2(1+\lambda+\delta)\lg \psi_{\delta}, c_0\rg }
{\check{C}_{\delta}} \\
& =: &
\overline{C}(t).
\end{eqnarray*}
This proves (\ref{4.16}). The estimate (\ref{4.17})
can be deduced from (\ref{4.16}),
Lemma 4.3 ($\alpha=0, \gamma=1+\lambda$ and $\epsilon=\delta$) and (\ref{4.15}).
(\ref{4.18}) follows from (\ref{4.19}), (\ref{4.20}),
(\ref{4.15}) and (\ref{4.16}).
We complete the proof of Proposition 4.5. \qed
\begin{cor
Under the same assumptions as in Proposition 4.5,
for each $t>0$,
there exists a constant $C(t)$ independent of
$N\in\N$ and $f\in B_{c}$ such that
\be
E\left[\lg M_f^N\rg (t)\right]
\le
\frac{\Vert f \Vert_{\infty}^{2}}{N}C(t). \label{4.21}
\ee
\end{cor
{\it Proof.}~This is immediate from (\ref{4.12}),
(\ref{4.13}), (\ref{4.15}) and (\ref{4.17}).
\qed
\medskip
We end this section with a lemma
which will be used in the next section. Recall that
for $\gamma>0$ the function $\Psi_{\gamma}$
on $\Omega$ is defined by
$\Psi_{\gamma}(\bz)=\sum z_i^{\gamma}$.
\begin{lm
Let $\bz\in\Omega$ and put
$\xi^N=N^{-1}\Xi(\bz)
=N^{-1}\sum \one_{\{z_i>0\}}\delta_{z_i}$. \\
(i) For any $f\in B_c$
\begin{eqnarray}
\lefteqn{
\frac{1}{N}\left|L^N\Phi_f(\bz)\right|} \nonumber \\
& \le &
\Ch C_{1,\lambda} \Vert f \Vert_{\infty}
\lg \psi_1,\xi^N \rg
\lg \psi_{1+\lambda},\xi^N \rg
+\frac{\Cc}{2} \Vert f \Vert_{\infty}
\lg \psi_{2+\lambda},\xi^N \rg \label{4.22} \\
& \le &
\Ch C_{1,\lambda} \Vert f \Vert_{\infty}
\lg \psi_1,\xi^N \rg^{\frac{2+\lambda}{1+\lambda}}
\lg \psi_{2+\lambda},\xi^N \rg^{\frac{\lambda}{1+\lambda}}
+\frac{\Cc}{2} \Vert f \Vert_{\infty}
\lg \psi_{2+\lambda},\xi^N \rg \label{4.23}
\end{eqnarray}
provided that $\lg\psi_{2+\lambda},\xi^N\rg<\infty$. \\
(ii) Let $\gamma>1$ and assume that
$E\left[\left|Z^N(0)\right|
^{2+\lambda+\gamma}\right]<\infty$. Then
\be
\wt{M}_{\gamma}^N(t):=
\frac{1}{N}\Psi_{\gamma}(Z^N(t))
-\frac{1}{N}\Psi_{\gamma}(Z^N(0))
-\frac{1}{N}\int_0^tL^N\Psi_{\gamma}(Z^N(s))ds
\label{4.24}
\ee
is a martingale. Moreover,
\be
\frac{1}{N}\left|L^N\Psi_{\gamma}(\bz)\right|
\le
\Ch C_{1,\lambda+\gamma}
\lg \psi_1,\xi^N \rg
\lg \psi_{1+\lambda+\gamma},\xi^N \rg
+\frac{\check{C}_{\gamma}}{2}
\lg \psi_{2+\lambda+\gamma},\xi^N \rg \label{4.25}
\ee
provided that $\lg\psi_{2+\lambda+\gamma},\xi^N\rg<\infty$.
\end{lm
{\it Proof.}~(i) (\ref{4.22}) can be shown in just a similar
way to (\ref{4.10}) and (\ref{4.11}) in view of (\ref{4.9}) and
\begin{eqnarray*}
\frac{1}{N}L^N\Phi_f(\bz)
& = &
\frac{1}{2N^2}\sum_{i\ne j} z_iz_j \Hh(z_i,z_j)
\left\{f(z_i+z_j)-f(z_i)-f(z_j)\right\} \\
& &
+\frac{1}{2N}\sum_i z_i^{2}
\int_0^{1}du \Hc(uz_i,(1-u)z_i)
\left\{f(uz_i)+f((1-u)z_i)-f(z_i)\right\}.
\end{eqnarray*}
(\ref{4.23}) is a consequence of H\"older's inequality. \\
(ii) The proof is quite analogous to that of Lemma 4.2
in view of calculations at the beginning of
the proof of Proposition 4.5
with $\gamma$ in place of $\delta$.
So the details are left to the reader. \qed
\section{Derivation of the macroscopic equation
\setcounter{equation}{0}
\subsection{Tightness arguments
Before studying the limit of the rescaled processes discussed
in Section 4, relative compactness of their laws must be argued.
In fact, it will be convenient to consider,
rather than $\xi^N(t)$, the measure-valued process
\[
\mu^N(t):= \frac{1}{N}\sum_i Z^N_i(t)\delta_{Z^N_i(t)},
\]
which takes values in $\cM_f$,
the space of finite measures on $(0,\infty)$, almost surely
as long as $P(|Z^N(0)|<\infty)=1$.
Denote by $C_c$
the set of continuous functions on $(0,\infty)$
with compact support and set
$C_{+,c}=B_+\cap C_{c}$.
We begin the tightness argument by introducing
a metric on $\cM_f$ compatible with the weak topology by
\[
d_w(\nu,\nu')
=\sum_{k=0}^{\infty}2^{-k}
\left(|\lg h_k, \nu\rg -\lg h_k, \nu'\rg|\wedge 1\right),
\qquad \nu,\nu' \in\cM_f
\]
where $h_0\equiv 1$ and $\{h_1,h_2,\ldots\}\subset C_{+,c}$
is as in the proof of Proposition 3.17 of \cite{Re}.
(Alternatively, see A 7.7 of \cite{Ka}.)
In particular, denoting the vague convergence
by $\stackrel{v}{\to}$, we have,
for $\eta, \eta_1,\eta_2,\cdots \in\cM$,
$\eta_n \stackrel{v}{\to}\eta$ iff
$\lg h_k, \eta_n\rg \to \lg h_k,\eta\rg $
for all $k\in\N$.
$(\cM_f, d_w)$ is a complete, separable metric space.
Let $\stackrel{w}{\to}$ stand for the weak convergence.
Given $a,b, \gamma>0$ arbitrarily, let
\[
\cM_{a,b}^{\gamma}
=\{\nu\in\cM_f :~ \lg 1, \nu\rg\le a,
\lg \psi_{\gamma}, \nu\rg\le b\}.
\]
As will be seen in the next two lemmas,
it is a compact set in $\cM_f$
and plays an important role.
We introduce an auxiliary function $\varphi_R$
for each $R>0$ by
\[ \varphi_R(y)
=\left\{
\begin{array}{ll}
1 & (y\le R) \\
R+1-y & (R\le y \le R+1) \\
0 & (y\ge R+1),
\end{array} \right.
\]
which is bounded and continuous on $(0,\infty)$.
\begin{lm
Let $\cM_{a,b}^{\gamma}$ be as above and
$\nu_1,\nu_2,\ldots\in \cM_{a,b}^{\gamma}$.
Assume that $\nu_n\stackrel{w}{\to}\nu$
for some $\nu\in\cM_f$.
Then \\
(i) $\nu \in \cM_{a,b}^{\gamma}$. \\
(ii) For any $\alpha\in(0,\gamma)$,
$\lg \psi_{\alpha}, \nu_n\rg \to \lg \psi_{\alpha}, \nu\rg$
as $n\to \infty$.
\end{lm
{\it Proof }~
This is a special case of Lemma 4.1 in \cite{EW00}.
But we give a proof for completeness. \\
(i) It is clear that
$\lg 1, \nu\rg=\lim_{n\to\infty}\lg 1, \nu_n\rg\le a$.
Also,
\[
\lg \psi_{\gamma}\varphi_R, \nu\rg
=\lim_{n\to\infty}\lg \psi_{\gamma}\varphi_R, \nu_n\rg
\le \liminf_{n\to\infty}\lg \psi_{\gamma}, \nu_n\rg\le b.
\]
Letting $R\to\infty$, we get
$\lg \psi_{\gamma}, \nu\rg \le b$.
Hence $\nu\in \cM_{a,b}^{\gamma}$. \\
(ii) Fix $\alpha\in(0,\gamma)$ arbitrarily.
Observe that for each $R>0$
\begin{eqnarray*}
|\lg \psi_{\alpha},\nu_n \rg- \lg \psi_{\alpha},\nu \rg|
& \le &
|\lg \psi_{\alpha}\varphi_R,\nu_n \rg
- \lg \psi_{\alpha}\varphi_R,\nu \rg| \\
& &
+\lg \psi_{\alpha}(1-\varphi_R),\nu_n \rg
+\lg \psi_{\alpha}(1-\varphi_R),\nu \rg.
\end{eqnarray*}
The first term on the right side converges to 0
as $n\to \infty$. As for the second and
third terms, we have a uniform bound in $n$
\[
\lg \psi_{\alpha}(1-\varphi_R),\nu_n \rg
\le\lg \frac{\psi_{\alpha}}{\psi_{\gamma}} \cdot
\one_{[R,\infty)} \cdot \psi_{\gamma}, \nu_n \rg
\le R^{\alpha-\gamma}\lg \psi_{\gamma}, \nu_n\rg
\le bR^{\alpha-\gamma}
\]
and similarly
$\lg \psi_{\alpha}(1-\varphi_R),\nu \rg \le
bR^{\alpha-\gamma}$, which vanishes as
$R\to\infty$. Therefore,
$\lg \psi_{\alpha}, \nu_n\rg \to \lg \psi_{\alpha}, \nu\rg$
as $n\to\infty$. \qed
\medskip
\noindent
Since we know from Lemma 5.1 (i) that
$\cM_{a,b}^{\gamma}$ are closed subsets of
$(\cM_f,d_w)$, the next lemma is regarded as
a slight generalization of Lemma 4.2 in \cite{EW00},
which corresponding to the case $\gamma=1$.
Their proof will be arranged in an obvious manner.
\begin{lm
For any $a,b,\gamma>0$,
$\cM_{a,b}^{\gamma}$ is compact.
\end{lm
{\it Proof.}~
It is sufficient to prove that
any sequence $\{\nu_n\}\subset \cM_{a,b}^{\gamma}$
has a weakly convergent subsequence.
Since $\sup_n\lg 1,\nu_n\rg \le a$,
we can choose a subsequence $\{\nu_{n_k}\}$
for which
$a_0:=\lim_{k\to\infty}\lg 1,\nu_{n_k}\rg$ exists.
In case $a_0=0$, $\nu_{n_k}\stackrel{w}{\to}0$.
In case $a_0>0$, we may assume further that
$\lg 1,\nu_{n_k}\rg>a_0/2$ for all $k$ and consider
probability measures
$\tilde{\nu}_k:=\lg 1,\nu_{n_k}\rg^{-1}\nu_{n_k}$.
The family $\{\tilde{\nu}_k\}$ is tight because
\[
\tilde{\nu}_k([R,\infty))
= \frac{\lg \one_{[R,\infty)},\nu_{n_k}\rg}
{\lg 1,\nu_{n_k}\rg}
\le
R^{-\gamma}
\frac{\lg \psi_{\gamma}\one_{[R,\infty)},\nu_{n_k}\rg}
{\lg 1,\nu_{n_k}\rg}
< R^{-\gamma} \frac{2b}{a_0}.
\]
Taking its subsequence $\{\tilde{\nu}_{k_l}\}$
such that $\tilde{\nu}_{k_l}\stackrel{w}{\to}
\tilde{\nu}$ as $l\to \infty$ for some
$\tilde{\nu}\in\cM_f$, we see that
$\nu_{n_{k_l}}\stackrel{w}{\to}a_0\tilde{\nu}$
as desired. \qed
\medskip
We now prove the compact containment property of
the laws of $\{\mu^N(t):~ t\ge 0\}$.
Note the triviality
$\lg \psi_a,\mu^N(t)\rg=\lg \psi_{a+1},\xi^N(t)\rg$.
More generally, for any $f\in B_c$
$\lg f,\mu^N(t)\rg=\lg f^{\star},\xi^N(t)\rg$,
where $f^{\star}\in B_c$ is defined to be
$f^{\star}(y)=yf(y)$.
\begin{pr
Assume that (\ref{4.2}) holds for some $\delta>1$.
Suppose that $\sum\delta_{Z^N_i(0)}$ is
${\rm Po}(Nc_0)$-distributed for each $N=1,2,\ldots$.
Then for any $T>0$ and $\epsilon\in (0,1)$
there exist $a,b>0$ such that
\be
\inf_N P\left(\mu^N(t)\in \cM_{a,b}^{\delta-1}
\quad
\mbox{for all} \quad t\in [0,T] \right)
\ge 1-\epsilon. \label{5.1}
\ee
\end{pr
{\it Proof.}~ Lemma 4.1 with $a=2+\lambda+\delta$
ensures the validity of (\ref{4.15}), which enables us to apply
Proposition 4.5 and Lemma 4.7 with $\gamma=\delta$.
Since
\[
\lg 1,\mu^N(t)\rg =\lg \psi_1,\xi^N(t)\rg
=N^{-1}|Z^N(t)|=N^{-1}|Z^N(0)|=\lg \psi_1,\xi^N(0)\rg
\]
and
\begin{eqnarray*}
\lg \psi_{\delta-1},\mu^N(t)\rg
& = & \lg \psi_{\delta},\xi^N(t)\rg \\
& = &
\lg \psi_{\delta},\xi^N(0)\rg
+\frac{1}{N}\int_0^tL^N\Psi_{\delta}(Z^N(s))ds
+\wt{M}_{\delta}^N(t),
\end{eqnarray*}
Chebyshev's inequality and Doob's inequality for
submartingale (e.g. Corollary 2.17, Chapter 2 in \cite{EK})
together yield
\begin{eqnarray*}
\lefteqn{1-P\left(\mu^N(t)\in \cM_{a,b}^{\delta-1}
\quad
\mbox{for all} \quad t\in [0,T] \right)} \\
& \le &
P\left(\sup_{t\in [0,T]}\lg 1,\mu^N(t)\rg >a\right)
+P\left(
\sup_{t\in [0,T]}\lg \psi_{\delta-1},\mu^N(t)\rg >b \right) \\
& \le &
P\left(\lg \psi_1,\xi^N(0)\rg >a\right)
+P\left(\lg \psi_{\delta},\xi^N(0)\rg >\frac{b}{3}\right) \\
& &
+P\left(\frac{1}{N}\int_0^T\left|L^N
\Psi_{\delta}(Z^N(s))\right|ds >\frac{b}{3} \right)
+P\left(\sup_{t\in [0,T]}|\wt{M}_{\delta}^N(t)|>\frac{b}{3}\right)\\
& = &
a^{-1}E\left[\lg \psi_1,\xi^N(0)\rg \right]
+3b^{-1}E\left[\lg \psi_{\delta},\xi^N(0)\rg\right] \\
& &
+3b^{-1}
E\left[\frac{1}{N}\int_0^T\left|L^N
\Psi_{\delta}(Z^N(s))\right|ds\right]
+3b^{-1}E\left[|\wt{M}_{\delta}^N(T)|\right].
\end{eqnarray*}
Therefore, the proof of (\ref{5.1}) reduces to showing
that the four expectations in the above
are bounded in $N$. The first two ones are finite
and independent of $N$:
\[
E\left[\lg \psi_1,\xi^N(0)\rg \right]
=\lg \psi_1,c_0\rg
\quad \mbox{and} \quad
E\left[\lg \psi_{\delta},\xi^N(0)\rg \right]
=\lg \psi_{\delta},c_0\rg.
\]
For the third expectation we deduce
from (\ref{4.25}) and (\ref{4.20}) that
\begin{eqnarray*}
\lefteqn{ E_N \ := \
E\left[\frac{1}{N}\int_0^T\left|L^N\Psi_{\delta}(Z^N(s))
\right|ds\right]} \\
& \le &
\Ch C_{1,\lambda+\delta}
E\left[\int_0^T \lg \psi_1,\xi^N(0) \rg
\lg \psi_{1+\lambda+\delta},\xi^N(s) \rg ds \right]
+\frac{\check{C}_{\delta}}{2}
E\left[\int_0^T \lg \psi_{2+\lambda+\delta},\xi^N(s) \rg ds
\right] \\
& \le &
\Ch C_{1,\lambda+\delta}
\left(T \cdot
E\left[\lg \psi_1, \xi^N(0)\rg^{2+\lambda+\delta}\right]
\right)^{\frac{1}{1+\lambda+\delta}}
\left(E\left[\int_0^T
\lg \psi_{2+\lambda+\delta},\xi^N(s) \rg ds \right]
\right)^{\frac{\lambda+\delta}{1+\lambda+\delta}} \\
& &
+\frac{\check{C}_{\delta}}{2}
E\left[\int_0^T \lg \psi_{2+\lambda+\delta},\xi^N(s) \rg ds
\right].
\end{eqnarray*}
Thanking to
$\overline{m}_{2+\lambda+\delta}<\infty$ and (\ref{4.16}),
this is bounded in $N$ .
Lastly, in view of (\ref{4.24})
\begin{eqnarray*}
E\left[|\wt{M}_{\delta}^N(T)|\right]
& \le &
E\left[\langle \psi_{\delta},\xi^N(T) \rangle \right]
+E\left[\langle \psi_{\delta},\xi^N(0) \rangle \right]
+E_N \\
& \le &
C^*(T) + \lg \psi_{\delta},c_0\rg
+ E_N,
\end{eqnarray*}
where the last inequality follows from (\ref{4.18}).
We complete the proof of Proposition 5.3 since
we have already seen that $\sup_N E_N<\infty$. \qed
\medskip
Let $D([0,\infty),\cM_f)$ denote the space of
right continuous functions $\nu(\cdot)$
from $[0,\infty)$ into $\cM_f$ with left limits.
This space is equipped with a metric
which induces the Skorohod topology.
We now state the main result of this subsection,
which establishes tightness of the rescaled processes.
Roughly speaking,
the reason why a stronger assumption on $c_0$
than the one in Proposition 5.3 is made here
is that showing equicontinuity of
$t\mapsto \lg f,\xi^N(t)\rg$ requires to dominate
$N^{-1}E[|L^N\Phi_f(Z^N(t))|]$ uniformly in $t>0$.
\begin{th
Assume (\ref{4.2}) with $\delta=2+\lambda$,
i.e.,
\[
0< \langle \psi_1, c_0 \rangle <\infty \quad
\mbox{and} \quad
\langle \psi_{4+2\lambda}, c_0 \rangle <\infty.
\]
Suppose that $\sum \delta_{Z_i^N(0)}$ is
${\rm Po}(Nc_0)$-distributed for each $N=1,2,\ldots$.
Then the sequence $\{\cP^N\}_{N=1}^{\infty}$ of
the laws $\cP^N$ of $\{\mu^N(t):~t\ge 0\}$
on $D([0,\infty),\cM_f)$ is relatively compact.
\end{th
{\it Proof.}~ Lemma 5.2 and Proposition 5.3
together imply that for arbitrarily fixed $t\ge 0$
the family of the laws of
$\mu^N(t)$ $(N=1,2,\ldots)$ on $\cM_f$
is relatively compact.
Therefore, for the same reasoning as in
the proof of Theorem 2.1 in \cite{EW},
which exploits Corollary 7.4 in Chapter 3 of \cite{EK},
it is sufficient to prove that, for any $T>0$
and $\epsilon>0$, there exists $0<\Delta<1$ such that
\be
\limsup_{N\to\infty}
P\left(\max_{i:~t_i<T} \sup_{s\in[t_i,t_{i+1})}
d_w(\mu^N(s),\mu^N(t_i))>\frac{\epsilon}{2}\right)
\le \epsilon, \label{5.2}
\ee
where $t_i=i\Delta \ (i=0,1,2,\ldots)$.
By (\ref{5.1}) with $\delta=2+\lambda$
we can find $a,b>0$ such that
\[
\inf_NP\left(W^N_{a,b}(T)\right)
\ge 1-\frac{\epsilon}{2},
\]
where $W^N_{a,b}(T)$ is the event that
$\mu^N(t)\in \cM_{a,b}^{1+\lambda}$
for all $t\in [0,T+1]$. So, we only have to prove
\be
\limsup_{N\to\infty}
P\left(\left\{\max_{i:~t_i<T} \sup_{s\in[t_i,t_{i+1})}
d_w(\mu^N(s),\mu^N(t_i))>\frac{\epsilon}{2}\right\}
\cap W^N_{a,b}(T)\right)
\le \frac{\epsilon}{2}. \label{5.3}
\ee
It follows from (\ref{4.23}) that, on $W^N_{a,b}(T)$,
for any $t\in [0,T]$ and $s\in[t,t+\Delta)$
\begin{eqnarray*}
\lefteqn{|\lg h_k, \mu^N(s)\rg-\lg h_k, \mu^N(t)\rg|
\ = \
|\lg h_k^{\star},\xi^N(s)\rg-\lg h_k^{\star}, \xi^N(t)\rg|} \\
& \le &
\left|M^N_{h_k^{\star}}(s)-M^N_{h_k^{\star}}(t)\right|
+\frac{1}{N}\int_t^s
\left|L^N\Phi_{h_k^{\star}}(Z^N(u))\right|du \\
& \le &
2\sup_{u\in[0,T+1]}\left|M^N_{h_k^{\star}}(u)\right|
+
\Delta \left(\Ch C_{1,\lambda}
a^{\frac{2+\lambda}{1+\lambda}}
b^{\frac{\lambda}{1+\lambda}}
+\frac{\Cc}{2} b\right)\Vert h_k^{\star} \Vert_{\infty}.
\end{eqnarray*}
Hence, taking $k_0\in\N$ such that
$\sum_{k=k_0+1}^{\infty}2^{-k}<\epsilon/4$ and then
$\Delta$ sufficiently small so that
\[
\Delta \left(\Ch C_{1,\lambda}
a^{\frac{2+\lambda}{1+\lambda}}
b^{\frac{\lambda}{1+\lambda}}
+\frac{\Cc}{2} b\right)\Vert h_k^{\star} \Vert_{\infty}
<\frac{\epsilon}{8k_0}
\quad \mbox{for all $k\in\{1,\ldots,k_0\}$},
\]
we get
\begin{eqnarray*}
\lefteqn{P\left(\left\{\max_{i:~t_i<T} \sup_{s\in[t_i,t_{i+1})}
d_w(\mu^N(s),\mu^N(t_i))>\frac{\epsilon}{2}\right\}
\cap W^N_{a,b}(T)\right)} \\
& \le &
P\left(\sum_{k=1}^{k_0}2^{-k}
\left\{2\sup_{u\in[0,T+1]}
\left|M^N_{h_k^{\star}}(u)\right|
+\frac{\epsilon}{8k_0}\right\}
+\sum_{k=k_0+1}^{\infty}2^{-k}
>\frac{\epsilon}{2}\right) \\
& \le &
\sum_{k=1}^{k_0}P\left(
2\sup_{u\in[0,T+1]}\left|M^N_{h_k^{\star}}(u)\right|
+\frac{\epsilon}{8k_0}>\frac{\epsilon}{4k_0}\right) \\
& \le &
\frac{16k_0}{\epsilon}\sum_{k=1}^{k_0}E\left[
\left|M^N_{h_k^{\star}}(T+1)\right|\right],
\end{eqnarray*}
in which the last inequality is implied by
Doob's inequality. Since each expectation
in the above sum converges to 0 as $N\to\infty$
by Corollary 4.6, (\ref{5.3}) and hence (\ref{5.2})
are obtained.
The proof of Theorem 5.4 is complete. \qed
\subsection{Studying the limit laws
This section is devoted to the study of
the weak limit of an arbitrary convergent subsequence $\{\cP^{N_l}\}_{l=1}^{\infty}$, say, of
the laws $\cP^N$ of $\{\mu^N(t):~t\ge 0\}$
on $D([0,\infty),\cM_f)$.
Although our main concern will be
proving that under the limit law the weak form of
(\ref{4.5}) is satisfied almost surely,
some properties on the limit are shown in advance.
Let $C([0,\infty),\cM_f)$ be the space of
continuous functions from $[0,\infty)$ to $\cM_f$.
According to \S 10 of Chapter 3 of \cite{EK}
it is equipped with the metric
\[
d_U(\nu_1(\cdot),\nu_2(\cdot))
=\int_0^{\infty}e^{-u}\sup_{t\in[0,u]}
\{d_w(\nu_1(t),\nu_2(t)) \wedge 1\}du,
\]
which gives the topology of uniform convergence
on compact subsets of $[0,\infty)$.
\begin{lm
If $\{\mu^{N_l}(t):~t\ge 0\}$ converges to
a process $\{\mu(t):~t\ge 0\}$
in distribution on $D([0,\infty),\cM_f)$ as $l\to\infty$,
then $\mu(\cdot)\in C([0,\infty),\cM_f)$ a.s.
\end{lm
{\it Proof.}~As in the proof of Lemma 5.1 in \cite{EW},
we employ Theorem 10.2 in Chapter 3 of \cite{EK}.
To this end, observe that if
$\mu^N(t)\ne\mu^N(t-)=N^{-1}\sum z_i \delta_{z_i}$
for some $\bz\in\Omega$,
the signed measure $\mu^N(t)-\mu^N(t-)$ equals either
\[
\frac{1}{N}(z_i+z_j)\delta_{z_i+z_j}
-\frac{1}{N}\left(z_i\delta_{z_i}+z_j\delta_{z_j}\right)
\quad \mbox{for some $z_i,z_j>0$ with $i\ne j$}
\]
or
\[
\frac{1}{N}\left(y\delta_y+(z_i-y)\delta_{z_i-y}\right)
-\frac{1}{N}z_i\delta_{z_i} \quad
\mbox{for some $z_i>0$ and $y\in(0,z_i)$}.
\]
This implies that
\begin{eqnarray*}
\lefteqn{d_w(\mu^N(t),\mu^N(t-))} \\
& \le &
\sup_{i\ne j}\sum_{k=0}^{\infty}2^{-k}
\min\left\{1,
N^{-1}\left|h_k^{\star}(z_i+z_j)
-h_k^{\star}(z_i)-h_k^{\star}(z_j)\right|
\right\} \\
& &
\vee \ \sup_{i}\sup_{y\in(0,z_i)}\sum_{k=0}^{\infty}2^{-k}
\min\left\{1,
N^{-1}\left|h_k^{\star}(y)+h_k^{\star}(z_i-y)
-h_k^{\star}(z_i)\right|
\right\} \\
& \le &
\sum_{k=1}^{\infty}2^{-k}\min\{1,
3N^{-1}\Vert h_k^{\star}\Vert_{\infty}\}
\to 0 \quad \mbox{as} \quad N\to \infty,
\end{eqnarray*}
in which all the terms for $k=0$ in the sums vanish
because of $h_0^{\star}(y)=y$. So, the above mentioned
theorem proves the assertion. \qed
\medskip
\noindent
Given $\nu\in\cM$, we define two measures
$\nu^{\star}$ and $\nu_{\star}$ on $(0,\infty)$
by $\nu^{\star}(dy)=y\nu(dy)$
and $\nu_{\star}(dy)=y^{-1}\nu(dy)$, respectively.
For instance, $\mu^N(t)=\xi^N(t)^{\star}$
and conversely $\xi^N(t)=\mu^N(t)_{\star}$
\begin{lm
Suppose that the same assumptions
as in Proposition 5.3 hold.
If $\{\mu^{N_l}(t):~t\ge 0\}$ converges to
a process $\{\mu(t):~t\ge 0\}$
in distribution on $D([0,\infty),\cM_f)$ as $l\to\infty$,
then
\be
P\left(\mu(0)=c_0^{\star}, \ \lg 1, \mu(t)\rg
=\lg 1, \mu(0)\rg \ \mbox{for all}
\ t\ge 0 \right)=1 \label{5.4}
\ee
and for each $T>0$ there exist a constant $\wt{C}(T)$
such that
\be
E\left[\lg \psi_{\delta-1},\mu(T) \rg
+\int_0^T \left\{\lg 1,\mu(0) \rg
\lg \psi_{\lambda+\delta},\mu(s) \rg
+\lg \psi_{1+\lambda+\delta},\mu(s) \rg\right\} ds
\right] \le \wt{C}(T). \label{5.5}
\ee
\end{lm
{\it Proof.}~
By the assumption on initial distributions
and Lemma 4.1 together
\[
P(\mu(0)_{\star}=c_0)=1
\quad \mbox{or equivalently} \quad
P(\mu(0)=c_0^{\star})=1.
\]
For each $N$,
$P(\lg 1,\mu(t)^N\rg =\lg 1,\mu^N(0)\rg$ for all $t\ge 0)=1$
and as is seen easily
\[
\{\nu(\cdot)\in D([0,\infty),\cM_f):~
\ \lg 1, \nu(t)\rg =\lg 1, \nu(0)\rg
\ \mbox{for all} \ t\ge 0\}
\]
is a closed subset of $D([0,\infty),\cM_f)$.
So, by the assumed convergence in distribution
$P\left(\lg 1, \mu(t)\rg
=\lg 1, \mu(0)\rg \ \mbox{for all} \ t\ge 0 \right)=1$,
and summarizing, we have shown (\ref{5.4}).
We proceed to verification of (\ref{5.5}).
In estimating $E_N$ in the proof of Proposition 5.3
we have shown the existence of
a constant $\wt{C}_1(T)$ independent of $N$ such that
\[
E\left[\int_0^T\left\{\lg 1,\mu^N(0) \rg
\lg \psi_{\lambda+\delta},\mu^N(s) \rg
+\lg \psi_{1+\lambda+\delta},\mu^N(s) \rg\right\} ds
\right] \le \wt{C}_1(T).
\]
Combining this with (\ref{4.18}), we obtain
\begin{eqnarray}
\lefteqn{
E\left[\lg \psi_{\delta-1},\mu^N(T) \rg\right]}
\label{5.6} \\
& + &
E\left[\int_0^T\left\{\lg 1,\mu^N(0) \rg
\lg \psi_{\lambda+\delta},\mu^N(s) \rg
+\lg \psi_{1+\lambda+\delta},\mu^N(s) \rg\right\} ds
\right] \le \wt{C}(T), \nonumber
\end{eqnarray}
where $\wt{C}(T) =\wt{C}_1(T)+C^{*}(T)$.
Therefore, for any $a,R>0$ and $l\in\N$
\begin{eqnarray}
\lefteqn{
E\left[a \wedge \lg \psi_{\delta-1}
\varphi_{R},\mu^{N_l}(T) \rg \right]}
\label{5.7} \\
& + &
E\left[\int_0^T\min\left\{a, \lg 1,\mu^{N_l}(0) \rg
\lg \psi_{\lambda+\delta}\varphi_{R},\mu^{N_l}(s) \rg
+\lg \psi_{1+\lambda+\delta}\varphi_{R},\mu^{N_l}(s) \rg
\right\} ds \right] \le \wt{C}(T). \nonumber
\end{eqnarray}
It is not difficult to check that
the function on $D([0,\infty),\cM_f)$
\[
\nu(\cdot)\mapsto
\int_0^T\min\left\{a, \lg 1,\nu(0) \rg
\lg \psi_{\lambda+\delta}\varphi_{R},\nu(s) \rg
+\lg \psi_{1+\lambda+\delta}\varphi_{R},\nu(s) \rg
\right\} ds
\]
is bounded and continuous.
Hence, letting $l\to\infty$ in (\ref{5.7}) yields
(\ref{5.7}) with $\mu$ in placed of $\mu^{N_l}$.
Since $a$ and $R$ are arbitrary, (\ref{5.5}) holds true
by virtue of the monotone convergence theorem. \qed
\medskip
We are in a position to state the main result of
this section. It should be emphasized that
the continuity of $K$ and $F$ is required
only in the next theorem.
\begin{th
In addition to the assumptions of Theorem 5.4,
assume that $\Hh$ and $\Hc$ are continuous.
If $\{\mu^{N_l}(t):~t\ge 0\}$ converges
to a process $\{\mu(t):~t\ge 0\}$
in distribution on $D([0,\infty),\cM_f)$ as $l\to\infty$,
then $\{\xi^{N_l}(t):~t\ge 0\}$ converges in distribution
on $D([0,\infty),\cM)$ to $\{\xi(t):~t\ge 0\}$ defined by
$\xi(t)=\mu(t)_{\star}$. Moreover,
with probability 1, it holds that
for any $f\in B_c$ and $t\ge 0$
\begin{eqnarray}
\lg f,\xi(t)\rg -\lg f,c_0 \rg
& = &
\frac{1}{2}\int_0^tds
\int \xi(s)^{\otimes 2}(dxdy)xy \Hh(x,y) (\Box f)(x,y) \label{5.8} \\
& &
-\frac{1}{2}\int_0^tds
\int \xi(s)(dx) x\int_0^{x}dy\Hc(y,x-y) (\Box f)(y,x-y) \nonumber
\end{eqnarray}
with the integrals on the right side being
absolutely convergent.
\end{th
From the analytic view point,
this theorem particularly implies
the existence of a $\cM$-valued weak solution
to (\ref{4.5}) with symmetric continuous homogeneous
functions $\Hh$ and $\Hc$ of degree $\lambda\ge 0$ satisfying (H1) and (H2), respectively,
and with initial measure $c_0$ such that
$0< \langle \psi_1, c_0 \rangle <\infty$ and
$\langle \psi_{4+2\lambda}, c_0 \rangle <\infty$.
Unfortunately, the uniqueness of the solution
has not been proved and accordingly
the convergence of the laws of $\{\mu^N(t):~t\ge 0\}$
($N=1,2,\ldots)$ has not been obtained.
Concerning this point, some comments will be given
at the end of this section.
\medskip
\noindent
{\it Proof of Theorem 5.7.}~
In order to prove the first half
it is sufficient to verify continuity of the map
$\{\nu(t):~ t\ge 0\} \mapsto
\{\nu(t)_{\star}:~ t\ge 0\}$
from $D([0,\infty),\cM_f)$ to $D([0,\infty),\cM)$.
But, that continuity follows from an
equivalent condition to the convergence
with respect to the Skorohod topology
(e.g. Proposition 5.3 of Chapter 3 in \cite{EK})
in terms of the metric on the state space
$\cM_f$ or $\cM$ together with continuity of the map
$\nu\mapsto \nu_{\star}$ from $\cM_f$ to $\cM$.
(cf. Problem 13 of Chapter 3 in \cite{EK}.) \par
The proof of the last half is
divided into five steps. \par
{\it Step 1.}~
As was sketched roughly in \S 4.1,
the argument will be based on (\ref{4.6}).
So, for any $f\in C_c$ and $t>0$, set
for $\nu(\cdot)\in D([0,\infty),\cM_f)$
\begin{eqnarray*}
I_{t,f}(\nu(\cdot))
& = &
\frac{1}{2}\int_0^tds
\int \nu(s)^{\otimes 2}(dxdy)\Hh(x,y) (\Box f)(x,y) \\
& &
-\frac{1}{2}\int_0^tds
\int \nu(s)(dx) \int_0^{x}dy\Hc(y,x-y) (\Box f)(y,x-y)
\end{eqnarray*}
provided that both integrals converge absolutely.
Then, by the assumption and (\ref{4.23})
we have almost surely
\[
\lg f,\xi^N(t)\rg -\lg f,\xi^N(0)\rg-I_{t,f}(\mu^N(\cdot))
=
M_f^N(t)+D_f^N(t) \quad \mbox{for all $t\ge 0$},
\]
where
\begin{eqnarray*}
D_f^N(t)
& = &
\frac{1}{2}\int_0^t ds\int \left(\xi^N(s)^{[2]}
-\xi^N(s)^{\otimes 2}\right)(dxdy)xy \Hh(x,y)
(\Box f)(x,y) \\
& = &
\frac{1}{2N^2}\int_0^t ds\sum_{i} Z_i^N(s)^2
\Hh(Z_i^N(s),Z_i^N(s))\left\{f(2Z_i^N(s))-2f(Z_i^N(s))\right\}.
\end{eqnarray*}
By the same calculations as in \S 4.1
(see the observation after (\ref{4.6}))
it is readily shown that for some constant $C_f'$
independent of $N\in \N$ and $T>0$
\[
E\left[\sup_{t\in[0,T]} |D_f^N(t)|\right]
\le \frac{C_f'T}{N}.
\]
By combining this with (\ref{4.21})
and using Doob's inequality we get for any $\epsilon>0$
\begin{eqnarray}
\lefteqn{P\left(\sup_{t\in[0,T]}
\left|\lg f,\xi^N(t)\rg -\lg f,\xi^N(0)\rg
-I_{t,f}(\mu^N(\cdot))\right| > \epsilon \right)}
\nonumber \\
& \le &
P\left(\sup_{t\in[0,T]}
\left|M_f^N(t)\right| > \frac{\epsilon}{2} \right)
+ P\left(\sup_{t\in[0,T]}
\left|D_f^N(t)\right| > \frac{\epsilon}{2} \right)
\nonumber \\
& \le &
\left(\frac{2}{\epsilon}\right)^2
\frac{4\Vert f\Vert_{\infty}^2C(T)}{N}
+\frac{2}{\epsilon}\cdot\frac{C_f'T}{N}. \label{5.9}
\end{eqnarray}
Therefore, the main task in the rest of the proof is
to show, in a suitable sense, convergence of
$I_{t,f}(\mu^{N_l}(\cdot))$ to $I_{t,f}(\mu(\cdot))$
as $l \to \infty$. But $I_{t,f}$ cannot be defined
as a function on $D([0,\infty),\cM_f)$
and we need to handle by a cut-off argument. \par
{\it Step 2.}~
For each $R>0$ decompose $I_{t,f}$ in the form
\be
I_{t,f}(\nu(\cdot))
=
I_{t,f,R}^{(2)}(\nu(\cdot))
+\widetilde{I}_{t,f,R}^{(2)}(\nu(\cdot))
-I_{t,f,R}^{(1)}(\nu(\cdot))
-\widetilde{I}_{t,f,R}^{(1)}(\nu(\cdot)),
\label{5.10}
\ee
where
\[
I_{t,f,R}^{(2)}(\nu(\cdot))
= \frac{1}{2}\int_0^tds
\int \nu(s)^{\otimes 2}(dxdy)
\Hh(x,y) (\Box f)(x,y)\varphi_R(x+y),
\]
\[
\widetilde{I}_{t,f,R}^{(2)}(\nu(\cdot))
= \frac{1}{2}\int_0^tds
\int \nu(s)^{\otimes 2}(dxdy)
\Hh(x,y) (\Box f)(x,y)(1-\varphi_R(x+y)),
\]
\[
I_{t,f,R}^{(1)}(\nu(\cdot))
= \frac{1}{2}\int_0^tds \int \nu(s)(dx)
\int_0^{x}dy\Hc(y,x-y) (\Box f)(y,x-y)\varphi_R(x)
\]
and
\[
\widetilde{I}_{t,f,R}^{(1)}(\nu(\cdot))
=
\frac{1}{2}\int_0^tds \int \nu(s)(dx)
\int_0^{x}dy\Hc(y,x-y) (\Box f)(y,x-y)(1-\varphi_R(x)).
\]
Of course $I_{t,f,R}^{(2)}$ and $I_{t,f,R}^{(1)}$
should be dominant for $R$ large.
Putting $\delta=\lambda+2$, we actually claim that
\be
\sup_{t\in[0,T]}
\left|\widetilde{I}_{t,f,R}^{(2)}(\nu(\cdot))\right|
\le
\frac{\Ch C_{1,\lambda+\delta}}{R^{\delta}}
\Vert \Box f \Vert_{\infty} \int_0^T \lg 1,\nu(s) \rg \lg
\psi_{\lambda+\delta},\nu(s) \rg ds \label{5.11}
\ee
and
\be \sup_{t\in[0,T]}
\left|\widetilde{I}_{t,f,R}^{(1)}(\nu(\cdot))\right|
\le
\frac{\Cc}{2R^{\delta}}\Vert \Box f \Vert_{\infty} \int_0^T
\lg \psi_{1+\lambda+\delta},\nu(s) \rg ds \label{5.12}
\ee
whenever each integral on the right side is finite.
Indeed, by (\ref{3.4}) and (\ref{4.7})
\begin{eqnarray*}
\left|\widetilde{I}_{t,f,R}^{(2)}(\nu(\cdot))\right|
& \le &
\frac{\Ch}{2}
\int_0^tds \int \nu(s)^{\otimes 2}(dxdy)
(x+y)^{\lambda} |(\Box f)(x,y)|\one_{\{x+y>R\}} \\
& \le &
\frac{\Ch}{2}\Vert \Box f \Vert_{\infty}
\int_0^tds \int \nu(s)^{\otimes 2}(dxdy)
(x+y)^{\lambda+\delta}
\frac{\one_{\{x+y>R\}}}{R^{\delta}} \\
& \le &
\frac{\Ch C_{1,\lambda+\delta}}{2R^{\delta}}
\Vert \Box f \Vert_{\infty}
\int_0^tds \int \nu(s)^{\otimes 2}(dxdy)
(x^{\lambda+\delta}+y^{\lambda+\delta}),
\end{eqnarray*}
from which (\ref{5.11}) is immediate.
Similarly, by (\ref{3.6})
\begin{eqnarray*}
\left|\widetilde{I}_{t,f,R}^{(1)}(\nu(\cdot))\right|
& \le &
\frac{\Cc}{2}\Vert \Box f \Vert_{\infty}
\int_0^tds \int \nu(s)(dx)
x^{1+\lambda}\one_{\{x>R\}} \\
& \le &
\frac{\Ch}{2}\Vert \Box f \Vert_{\infty}
\int_0^tds \int \nu(s)(dx)x^{1+\lambda+\delta}
\frac{\one_{\{x>R\}}}{R^{\delta}} \\
& \le &
\frac{\Cc}{2R^{\delta}}\Vert \Box f \Vert_{\infty}
\int_0^tds \int \nu(s)(dx)x^{1+\lambda+\delta}
\end{eqnarray*}
and thus (\ref{5.12}) is valid.
Now consider (\ref{5.11}) and (\ref{5.12})
with $\nu(\cdot)$ being a random element
$\mu^N(\cdot)$ or $\mu(\cdot)$.
Thanking to Chebyshev's inequality,
taking expectations and then
using (\ref{5.6}), (\ref{5.4}) and (\ref{5.5}) lead to
\be
P\left(\sup_{t\in[0,T]}\left\{
\left|\widetilde{I}_{t,f,R}^{(2)}(\mu^N(\cdot))\right|
+\left|\widetilde{I}_{t,f,R}^{(1)}(\mu^N(\cdot))\right|\right\}
> \epsilon\right)
\le \frac{2\Ch C_{1,\lambda+\delta}+\Cc}{2\epsilon R^{\delta}}
\Vert \Box f \Vert_{\infty} \widetilde{C}(T) \label{5.13}
\ee
and
\be
P\left(\sup_{t\in[0,T]}\left\{
\left|\widetilde{I}_{t,f,R}^{(2)}(\mu(\cdot))\right|
+\left|\widetilde{I}_{t,f,R}^{(1)}(\mu(\cdot))\right|\right\}
> \epsilon\right)
\le \frac{2\Ch C_{1,\lambda+\delta}+\Cc}{2\epsilon R^{\delta}}
\Vert \Box f \Vert_{\infty} \widetilde{C}(T). \label{5.14}
\ee
{\it Step 3.}~ Let $f\in C_c$ be arbitrary.
Clearly proving that
\be
P\left(\lg f,\xi(t) \rg -\lg f,\xi(0) \rg
-I_{t,f}(\mu (\cdot))
=0 \quad \mbox{for all $t\ge 0$}\right)=1 \label{5.15}
\ee
is equivalent to showing that for any
$T>0$ and $\epsilon\in (0,1)$
\be
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi(t) \rg -\lg f,\xi(0) \rg
-I_{t,f}(\mu (\cdot))\right|>2\epsilon\right)=0. \label{5.16}
\ee
We claim here that the latter can be reduced to
establishing the inequality
\begin{eqnarray}
\lefteqn{
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi(t) \rg -\lg f,\xi(0) \rg
-I_{t,f,R}^{(2)}(\mu (\cdot))+I_{t,f,R}^{(1)}(\mu (\cdot))
\right|>\epsilon\right)} \label{5.17} \\
& \le &
\liminf_{l \to \infty}
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi^{N_l}(t) \rg
-\lg f,\xi^{N_l} (0) \rg
-I_{t,f,R}^{(2)}(\mu^{N_l} (\cdot))
+I_{t,f,R}^{(1)}(\mu^{N_l} (\cdot))
\right|>\epsilon\right) \nonumber
\end{eqnarray}
for each $R>0$. Indeed, by (\ref{5.10}) and (\ref{5.14})
\begin{eqnarray}
\lefteqn{
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi(t) \rg -\lg f,\xi(0) \rg
-I_{t,f}(\mu (\cdot))
\right|>2\epsilon\right)} \label{5.18} \\
& \le &
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi(t) \rg -\lg f,\xi(0) \rg
-I_{t,f,R}^{(2)}(\mu (\cdot))+I_{t,f,R}^{(1)}(\mu (\cdot))
\right|>\epsilon\right) \nonumber \\
& &
+\frac{2\Ch C_{1,\lambda+\delta}+\Cc}{2\epsilon R^{\delta}}
\Vert \Box f \Vert_{\infty} \widetilde{C}(T). \nonumber
\end{eqnarray}
On the other hand, by (\ref{5.9}) and (\ref{5.13})
\begin{eqnarray*}
\lefteqn{ \liminf_{l \to \infty}
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi^{N_l}(t) \rg
-\lg f,\xi^{N_l}(0) \rg
-I_{t,f,R}^{(2)}(\mu^{N_l} (\cdot))
+I_{t,f,R}^{(1)}(\mu^{N_l} (\cdot))
\right|>\epsilon\right) } \\
& \le &
\liminf_{l \to \infty}
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi^{N_l}(t) \rg
-\lg f,\xi^{N_l}(0) \rg-I_{t,f}(\mu^{N_l} (\cdot))
\right|>\frac{\epsilon}{2}\right)
\nonumber \\
& &
+\frac{2\Ch C_{1,\lambda+\delta}+\Cc}{2\epsilon R^{\delta}}
\Vert \Box f \Vert_{\infty} \widetilde{C}(T) \\
& \le &
\frac{2\Ch C_{1,\lambda+\delta}+\Cc}{2\epsilon R^{\delta}}
\Vert \Box f \Vert_{\infty} \widetilde{C}(T).
\end{eqnarray*}
Therefore,
this combined with (\ref{5.17}) and (\ref{5.18}) yields
\[
P\left(\sup_{t\in[0,T]}\left|\lg f,\xi(t) \rg
-\lg f,\xi(0) \rg -I_{t,f}(\mu (\cdot))
\right|>2\epsilon\right)
\le
\frac{2\Ch C_{1,\lambda+\delta}+\Cc}{\epsilon R^{\delta}}
\Vert \Box f \Vert_{\infty} \widetilde{C}(T)
\]
and the aforementioned claim
follows since $R>0$ is arbitrary. \par
{\it Step 4.}~
We shall prove (\ref{5.17}) for arbitrarily fixed
$T, R>0$, $\epsilon\in (0,1)$ and
$f\in \{h_1^{\star}, h_2^{\star}, \ldots\}(\subset C_c)$
at least.
But, because of the triviality that
$x > \epsilon$ if and only if $1 \wedge x > \epsilon $,
(\ref{5.17}) can be rewritten into
\be
P\left(1 \wedge \Upsilon (\mu(\cdot)) > \epsilon\right)
\le \liminf_{l\to\infty}
P\left(1 \wedge \Upsilon(\mu^{N_l}(\cdot))>\epsilon\right),
\label{5.19}
\ee
where $\Upsilon=\Upsilon_{T,f,R}$ is a Borel measurable
function on $D([0,\infty),\cM_f)$ defined by
\[
\Upsilon (\nu(\cdot))
=\sup_{t\in[0,T]}\left|\lg f_{\star},\nu(t) \rg
-\lg f_{\star},\nu(0) \rg
-I_{t,f,R}^{(2)}(\nu (\cdot))
+I_{t,f,R}^{(1)}(\nu (\cdot))\right|.
\]
Furthermore, (\ref{5.19}) can be reduced to
showing that for any $n\in\N$
\begin{eqnarray}
\lefteqn{
P\left(1 \wedge \Upsilon (\mu(\cdot)) > \epsilon \
|~\lg 1,\mu(0)\rg \le a_n\right)} \nonumber \\
& \le &
\liminf_{l\to\infty}
P\left(1 \wedge \Upsilon (\mu^{N_l}(\cdot))>\epsilon \
|~\lg 1,\mu^{N_l}(0)\rg \le a_n\right), \label{5.20}
\end{eqnarray}
where $a_1,a_2,\ldots$ are such that
$\lim_{n\to\infty}a_n=\infty$,
$a_n>\lg \psi_1,c_0\rg(=E[\lg 1,\mu^{N_l}(0)\rg])$
and $\lim_{l\to\infty}P(\lg 1,\mu^{N_l}(0)\rg \le a_n)
=P(\lg 1,\mu(0)\rg \le a_n)$ for each $n\in\N$.
Indeed, assuming (\ref{5.20}), we get
\begin{eqnarray*}
\lefteqn{
P\left(1 \wedge \Upsilon (\mu(\cdot)) >
\epsilon\right)} \\
& = &
P\left(1 \wedge \Upsilon (\mu(\cdot)) > \epsilon
\ |~\lg 1,\mu(0)\rg \le a_n\right) \
P\left(\lg 1,\mu(0)\rg \le a_n\right) \\
& \le &
\liminf_{l\to\infty}
P\left(1 \wedge \Upsilon (\mu^{N_l}(\cdot))>\epsilon \
|~\lg 1,\mu^{N_l}(0)\rg \le a_n\right) \
P\left(\lg 1,\mu(0)\rg \le a_n\right) \\
& \le &
\liminf_{l\to\infty}
\frac{P\left(1 \wedge \Upsilon (\mu^{N_l}(\cdot))>\epsilon\right)}
{P\left(\lg 1,\mu^{N_l}(0)\rg \le a_n\right)} \cdot
P\left(\lg 1,\mu(0)\rg \le a_n\right) \\
& \le &
\frac{
\liminf_{l\to\infty}
P\left(1 \wedge \Upsilon (\mu^{N_l}(\cdot))>\epsilon\right)}
{1-a_n^{-1}\lg \psi_1,c_0\rg} \cdot
P\left(\lg 1,\mu(0)\rg \le a_n\right),
\end{eqnarray*}
which tends to the right side of (\ref{5.19}) as $n\to\infty$.
For any $a>0$,
set $\cM_{\le a}=\{\nu\in\cM_f:~\lg 1,\nu\rg \le a\}$,
which is regarded as a closed subspace of $\cM_f$.
Accordingly $D([0,\infty),\cM_{\le a})$ is
a closed subspace of $D([0,\infty),\cM_{f})$.
Note that by (\ref{5.4})
\[
P\left(\mu(\cdot)\in D([0,\infty),\cM_{\le a}) \
|~\lg 1,\mu(0)\rg \le a\right)= 1
\]
and similarly
\[
P\left(\mu^{N_l}(\cdot)\in D([0,\infty),\cM_{\le a}) \
|~\lg 1,\mu^{N_l}(0)\rg \le a\right)= 1.
\]
By the assumption of
convergence of $\{\mu^{N_l}(t):~t\ge 0\}$
to $\{\mu(t):~t\ge 0\}$ in distribution as $l\to\infty$
together with
$\lim_{l\to\infty}P(\lg 1,\mu^{N_l}(0)\rg \le a_n)
=P(\lg 1,\mu(0)\rg \le a_n)$ $(n=1,2,\ldots)$
it is not difficult to verify that for each $n\in\N$
the sequence of the conditional laws
of $\{\mu^{N_l}(t):~t\ge 0\}$
on $D([0,\infty),\cM_{\le a_n})$
given $\lg 1,\mu^{N_l}(0)\rg \le a_n$
converges to the conditional law of $\{\mu(t):~t\ge 0\}$
on $D([0,\infty),\cM_{\le a_n})$
given $\lg 1,\mu(0)\rg \le a_n$. Therefore,
(\ref{5.20}) is naturally expected to follow
as a consequence of certain continuity of $\Upsilon$
restricted on $\cM_{\le a_n}$.
In fact, by virtue of
Theorem 10.2 in Chapter 3 of \cite{EK}
and Lemma 5.5 together, we can conclude (\ref{5.20})
as soon as the continuity of $\Upsilon$
restricted on $\cM_{\le a_n}$
with respect to the metric $d_U$
(defined at the beginning of this subsection)
is checked to hold. We show below more generally
that continuity of $\Upsilon$ on $\cM_{\le a}$
for any $a>0$.
\par
For this purpose, take an arbitrary sequence
$\{\nu_n(\cdot)\}_{n=1}^{\infty}$ of
$D([0,\infty),\cM_{\le a})$
and $\nu(\cdot)\in D([0,\infty),\cM_{\le a})$ such that
$d_{U}(\nu_n(\cdot), \nu(\cdot))\to 0$ as $n\to\infty$.
Then our task here is to show that
$\Upsilon (\nu_n(\cdot)) \to \Upsilon (\nu(\cdot))$
for any $f\in \{h_1^{\star}, h_2^{\star}, \ldots\}$.
From general inequalities of the form
\[
\left|\sup_t |\phi_1(t)| -\sup _t|\phi_2(t)|\right|
\le \sup_t|\phi_1(t)-\phi_2(t)|
\le \sup_t|\phi_1(t)|+\sup_t|\phi_2(t)|
\]
we deduce
\begin{eqnarray*}
\lefteqn{
\left|\Upsilon (\nu_n(\cdot))
-\Upsilon (\nu(\cdot))\right|} \\
& \le &
2\sup_{t\in[0,T]}
\left|\lg f_{\star},\nu_n(t) \rg
-\lg f_{\star},\nu(t) \rg \right| \\
& &
+ \sup_{t\in[0,T]}
\left|I_{t,f,R}^{(1)}(\nu_n (\cdot))
-I_{t,f,R}^{(1)}(\nu(\cdot))\right|
+\sup_{t\in[0,T]}
\left|I_{t,f,R}^{(2)}(\nu_n (\cdot))
-I_{t,f,R}^{(2)}(\nu(\cdot))\right| \\
&=: &
2s_n + s^{(1)}_n + s^{(2)}_n.
\end{eqnarray*}
Letting $i\in \N$ be such that $f=h_i^{\star}$
or $f_{\star}=h_i$, we have
\[
s_n= \sup_{t\in[0,T]}\left|\lg h_i,\nu_n(t) \rg
-\lg h_i,\nu(t) \rg \right|,
\]
which converges to 0 as $n\to \infty$
by $\sup_{t\in[0,T]} d_w(\nu_n(t), \nu(t))\to 0$.
As for $s_n^{(1)}$ observe that
\[
2s_n^{(1)}
\le
\int_0^Tdt\left| \lg g,\nu_n(t)\rg-\lg g,\nu(t)\rg \right|,
\]
where
$\ds{g(x)=\int_0^{x}dy\Hc(y,x-y)
(\Box h_i^{\star})(y,x-y)\varphi_R(x)}$.
Since the assumed continuity of $\Hc$ and (\ref{3.6})
together assure that
$g$ is a bounded continuous function on $(0,\infty)$,
$\lg g,\nu_n(t)\rg \to \lg g,\nu(t)\rg $
for each $t\ge0$. Furthermore,
\[
\sup_{t\in[0,T]}
\left| \lg g, \nu_n(t)\rg -\lg g, \nu(t)\rg \right|
\le
2a\Vert g \Vert_{\infty}.
\]
So, the dominated convergence theorem
proves that $s_n^{(1)}\to 0$ as $n\to \infty$.
Basically a similar strategy can be adopted
to $s_n^{(2)}$:
\[
2s_n^{(2)}
\le
\int_0^Tdt\left| \lg G,\nu_n(t)^{\otimes 2}\rg
-\lg G,\nu(t)^{\otimes 2}\rg \right|,
\]
where
$G(x,y):=\Hh(x,y) (\Box h_i^{\star})(x,y)\varphi_R(x+y)$
is verified to be bounded and continuous
on $(0,\infty)^2$ thanks to (\ref{3.4})
as well as the assumption that $\Hh$ is continuous.
Here, we claim that
$\nu_n(t)^{\otimes 2} \stackrel{w}{\to}
\nu(t)^{\otimes 2}$. In this respect,
we rely on a slight generalization of
Theorem 2.8 in \cite{Bi}, which implies
in particular that if a sequence $\{p_n\}$
of probability measures on $(0,\infty)$
converges weakly to $p$, then
$p_n^{\otimes 2} \stackrel{w}{\to} p^{\otimes 2}$.
Generalizing this assertion to finite measures
is easy by considering the normalized measures
and it follows that $\lg G,\nu_n(t)^{\otimes 2}\rg
\to \lg G,\nu(t)^{\otimes 2}\rg $
as $n\to \infty$ for each $t\ge0$.
In addition,
\[
\sup_{t\in[0,T]}
\left| \lg G, \nu_n(t)^{\otimes 2}\rg
-\lg G, \nu(t)^{\otimes 2}\rg \right|
\le
2a^2\Vert G \Vert_{\infty}.
\]
Hence again by the dominated convergence
theorem $s_n^{(2)}\to 0$.
Consequently we have proved
the continuity of $\Upsilon=\Upsilon_{T,f,R}$
on each $\cM_{\le a}$
with respect to $d_U$ for any $T,R>0$ and
$f\in \{h_1^{\star}, h_2^{\star}, \ldots\}$.
As was already discussed
this implies (\ref{5.15}) for those $f$'s. \par
{\it Step 5.}~
The remaining task is derivation of
the weak form (\ref{5.8}).
By combining (\ref{5.15}) for $f=h_i^{\star}$
$(i=1,2,\ldots)$ with
Lemma 5.6 (implying in particular (\ref{5.15})
for $f=h_0^{\star}$) and then
recalling the relation $\xi(t)^{\star}=\mu(t)$,
we have proved so far that, with probability 1,
for any $i\in\Zp$ and $t \ge 0$
\begin{eqnarray*}
\lg h_i,\mu(t)\rg -\lg h_i,c_0^{\star} \rg
& = &
\frac{1}{2}\int_0^tds \int \mu(s)^{\otimes 2}(dxdy)
\Hh(x,y)(\Box h_i^{\star})(x,y) \\
& &
-\frac{1}{2}\int_0^tds \int \mu(s)(dx)
\int_0^{x}dy\Hc(y,x-y) (\Box h_i^{\star})(y,x-y)
\end{eqnarray*}
with the integrals on the right side being
absolutely convergent. Since $\{h_0,h_1,\ldots\}$ is
measure-determining, the above equalities
are regarded as an equality among finite measures.
In other words, one can replace $h_i$ by
arbitrary bounded Borel functions $f$.
In particular, replacing $h_i$ by
$f_{\star}$ with $f\in B_c$ being arbitrary,
we obtain (\ref{5.8}).
The proof of Theorem 5.7 is complete. \qed
\medskip
\noindent
{\it Remark.}~
As mentioned earlier the uniqueness of weak
solutions of (\ref{4.5}) has not been proved.
There is quite an extensive literature concerning
the existence and/or uniqueness of solutions to
coagulation-fragmentation equations.
Well-posedness in the sense of
measure-valued solutions was studied in e.g. \cite{FL}
(for coagulation equations)
and \cite{C2} (for coagulation
multiple-fragmentation equations).
Below we take up a result in \cite{BLL}.
While it considers classical solutions,
the setting of that paper is well adapted to
a special case of our models.
Given $\lambda\ge 0$, let
$\Hc(x+y)=2 (x+y)^{\lambda}$
or $F(x+y)=2 (x+y)^{\lambda+1}$.
Then (\ref{1.1})
is rewritten into
\begin{eqnarray}
\frac{\partial}{\partial t}c(t,x)
& = &
-x^{2+\lambda} c(t,x)
+2\int_x^{\infty}y^{\lambda+1}c(t,y)dy \label{5.21} \\
& &
+\frac{1}{2}\int_0^x K(y,x-y)c(t,y)c(t,x-y)dy
-c(t,x)\int_0^{\infty}K(x,y)c(t,y)dy. \nonumber
\end{eqnarray}
This coincides with the equation (1.3) in \cite{BLL}
with $\alpha=\lambda+2$ and $\nu=0$.
One assumption made on $K$ in \cite{BLL} (cf. (1.5) there)
is the following: \\
for some $C>0$ and
$0\le \sigma \le \rho<\alpha(=\lambda+2)$
\[
0 \le K(x,y) \le C
\left[(1+x)^{\rho}(1+y)^{\sigma}+
(1+x)^{\sigma}(1+y)^{\rho}\right],
\quad x,y>0.
\]
Under our assumptions (\ref{1.4}) and (H1) on $K$,
this condition is fulfilled with $C=\Ch C_{1,\lambda}$,
$\rho=\lambda+1$ and $\sigma=1$. Therefore,
Theorem 2.2 (i) in \cite{BLL} implies, among other things,
the existence of a unique nonnegative classical solution
to (\ref{5.21}) which is local in time. As argued
in a closely related article \cite{BL}, showing
the existence of a global solution requires
a priori bound for the moments of solutions.
For such a purpose, an analogue to
our calculations in \S 4.2 could be useful,
although we will not pursue this point here.
\bigskip
\bigskip
\noindent
{\Large \bf Appendix}
\medskip
\medskip
\noindent
In this appendix we prove Lemma 4.1. \\
{\it Proof of Lemma 4.1.}~
By the assumption for any $f\in B_{+,c}$
\[
E\left[\exp(-\lg f,\eta^N\rg )\right]
=E\left[\exp\left(-\frac{1}{N}\sum_if(Y_i^N)\right)\right]
=\exp\left(-N\lg 1-e^{-f/N}, \zeta\rg \right).
\]
As $N\to \infty$
the most right side converges to
$\exp(-\lg f, \zeta\rg)$ by Lebesgue's convergence theorem.
This proves the first assertion. To prove
the convergence (\ref{4.4}) in case $a(=:n)\in \N$
it suffices to give the moment formula of the form
\[
E\left[\left(\sum_iY_i^N\right)^n\right]
=
n!\sum_{k=1}^n\frac{N^k}{k!}\sum
\frac{\lg\psi_{n_1}, \zeta\rg\cdots \lg\psi_{n_k}, \zeta\rg}
{n_1!\cdots n_k!},
\]
where the inner summation is taken over
$k$-tuples $(n_1,\ldots,n_k)$ of positive integers
such that $n_1+\cdots+n_k=n$.
But, the proof of the above formula clearly reduces to
showing that for each $R>0$
\[
E\left[\left(\sum_iY_i^N\one_{(0,R]}(Y_i^N)
\right)^n\right]
=n!\sum_{k=1}^n\frac{N^k}{k!}\sum
\frac{\lg\psi_{n_1}^{(R)}, \zeta\rg
\cdots \lg\psi_{n_k}^{(R)}, \zeta\rg}{n_1!\cdots n_k!},
\]
where $\psi_{l}^{(R)}(y)=y^l\one_{(0,R]}(y)$, and
this version is derived by comparing the coefficients
of $t^n$ after expanding in $t$ each side of
\[
E\left[\exp\left(t\sum_iY_i^N\one_{(0,R]}(Y_i^N)
\right)\right]
=\exp\left(N\int_{(0,R]} \zeta(dy)(e^{ty}-1) \right).
\]
It remains to prove (\ref{4.4}) in case
$a\in (1,\infty)\setminus \N$.
For such an $a$, put $n=[a]$ and $\alpha=a-[a]\in (0,1)$.
Let $C_{3,\alpha}=\alpha/\Gamma(1-\alpha)$.
Combining the aforementioned moment formula
with the identity
$y^{\alpha}=C_{3,\alpha}
\int ds s^{-(1+\alpha)}(1-e^{-sy})$
for $y>0$ we deduce
\begin{eqnarray*}
\lefteqn{
E\left[\left(\sum_iY_i^N\right)^a\right]} \\
&=&
C_{3,\alpha}\int \frac{ds}{s^{1+\alpha}}
\left\{E\left[\left(\sum_iY_i^N\right)^n\right]
-E\left[\left(\sum_iY_i^N\right)^n
\exp\left(-s\sum_iY_i^N\right)\right] \right\} \\
&=&
C_{3,\alpha}n!\int \frac{ds}{s^{1+\alpha}}
\left\{\sum_{k=1}^n\frac{N^k}{k!} \sum
\frac{\int y^{n_1}\zeta(dy)\cdots \int y^{n_k}\zeta(dy)}
{n_1!\cdots n_k!}\right. \\
& &
-\left.\sum_{k=1}^n\frac{N^k}{k!}\sum
\frac{\int y^{n_1}e^{-sy}\zeta(dy)\cdots \int y^{n_k}e^{-sy}\zeta(dy)}
{n_1!\cdots n_k!}
E\left[\exp\left(-s\sum_iY_i^N\right)\right] \right\}.
\end{eqnarray*}
Here, the second equality follows from the fact that
under the transformed measure
\[
\tilde{P}_s\left(\sum \delta_{Y^N_i}\in\bullet \right)
:= E\left[\exp\left(-s\sum_iY_i^N\right);
\ \sum \delta_{Y^N_i}\in\bullet
\right]
E\left[\exp\left(-s\sum_iY_i^N\right)\right]^{-1}
\]
$\sum \delta_{Y^N_i}$ is
${\rm Po}(e^{-sy}\zeta(dy))$-distributed.
(See e.g. \cite{B}, p.80, Lemma 2.4.)
For $s>0$ set
$\Lambda(s) =\int(1-e^{-sy})\zeta(dy)$ so that
$\Lambda(s)\le s\int y\zeta(dy)$ and
\[
E\left[\exp\left(-s\sum_iY_i^N\right)\right]
= \exp\left(-N\Lambda(s)\right).
\]
By the change of variable $u:=Ns$ in the integral
with respect to $ds$
\begin{eqnarray*}
\lefteqn{
E\left[\left(\frac{\sum_iY_i^N}{N}\right)^a\right]} \\
&=&
C_{3,\alpha}n!\int \frac{du}{u^{1+\alpha}}
\left\{\sum_{k=1}^n\frac{N^{k-n}}{k!}\sum
\frac{\int y^{n_1}\zeta(dy)\cdots \int y^{n_k}\zeta(dy)}
{n_1!\cdots n_k!}\right. \\
& &
-\left.\sum_{k=1}^n\frac{N^{k-n}}{k!}\sum
\frac{\int y^{n_1}e^{-\frac{uy}{N}}\zeta(dy)\cdots
\int y^{n_k}e^{-\frac{uy}{N}}\zeta(dy)}{n_1!\cdots n_k!}
\exp\left(-N\Lambda\left(\frac{u}{N}\right)\right)\right\} \\
&=&
C_{3,\alpha}n!\sum_{k=1}^n\frac{N^{k-n}}{k!}\sum
\int \frac{du}{u^{1+\alpha}}
\left\{\frac{\int y^{n_1}\zeta(dy)\cdots \int y^{n_k}\zeta(dy)}
{n_1!\cdots n_k!}\right. \\
& &
\hspace*{2cm}\left. -\frac{\int y^{n_1}e^{-\frac{uy}{N}}\zeta(dy)
\cdots \int y^{n_k}e^{-\frac{uy}{N}}\zeta(dy)}{n_1!\cdots n_k!}
\exp\left(-N\Lambda\left(\frac{u}{N}\right)\right)\right\}.
\end{eqnarray*}
Here, we claim that each integral with respect to $du$
in the last expression is not only finite
but also bounded in $N$.
Indeed, for such an integral corresponding to
$k=1$ (or $n_1=n$), we can observe by Fubini's theorem
\begin{eqnarray*}
\lefteqn{
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left|\int y^{n}\zeta(dy)
- \int y^{n}e^{-\frac{uy}{N}}\zeta(dy)
\exp\left(-N\Lambda\left(\frac{u}{N}\right)\right) \right|} \\
& = &
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left(\int y^{n}\zeta(dy)- \int y^{n}e^{-\frac{uy}{N}}\zeta(dy)\right) \\
& &
+C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\int y^{n}e^{-\frac{uy}{N}}\zeta(dy)
\left\{1-\exp\left(-N\Lambda\left(\frac{u}{N}\right)\right)\right\} \\
& \le &
\int y^n\left(C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
(1-e^{-\frac{uy}{N}})\right)\zeta(dy) \\
& &
+\int y^{n}\zeta(dy)C_{3,\alpha}
\int \frac{du}{u^{1+\alpha}}
\left\{1-\exp\left(-u\int y\zeta(dy)\right)\right\} \\
& = &
\int y^n \left(\frac{y}{N}\right)^{\alpha}\zeta(dy)
+\int y^{n}\zeta(dy)\left(\int y \zeta(dy)\right)^{\alpha}.
\end{eqnarray*}
It is not difficult to obtain
analogous estimates for any integral
corresponding to $k\in\{2,\ldots,n-1\}$. Thus,
we have identified the main term as $N\to\infty$:
\begin{eqnarray*}
\lefteqn{
E\left[\left(\frac{\sum_iY_i^N}{N}\right)^a\right]+o(1)} \\
& = &
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left\{\left(\int y\zeta(dy)\right)^n
-\left(\int ye^{-\frac{uy}{N}}\zeta(dy)\right)^n
\exp\left(-N\Lambda\left(\frac{u}{N}\right)\right)\right\}.
\end{eqnarray*}
Lastly, this integral converges to
$\lg \psi_1,\zeta\rg^{a}$ since as before
\begin{eqnarray*}
\lefteqn{
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left|\left(\int y\zeta(dy)\right)^n
-\left(\int ye^{-\frac{uy}{N}}\zeta(dy)\right)^n\right|} \\
& \le &
\int y \left(
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left(1-e^{-\frac{uy}{N}}\right)\right)\zeta(dy)
\cdot n\left(\int y\zeta(dy)\right)^{n-1} \\
& = &
\int y\left(\frac{y}{N}\right)^{\alpha}\zeta(dy)
\cdot n\left(\int y\zeta(dy)\right)^{n-1}
\ \to \ 0
\end{eqnarray*}
and by Lebesgue's convergence theorem
\begin{eqnarray*}
\lefteqn{
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left(\int ye^{-\frac{uy}{N}}\zeta(dy)\right)^n
\left\{1-\exp\left(-N\Lambda\left(\frac{u}{N}\right)\right)\right\}} \\
&\to &
C_{3,\alpha}\int \frac{du}{u^{1+\alpha}}
\left(\int y\zeta(dy)\right)^n
\left\{1-\exp\left(-u\int y \zeta(dy)\right)\right\} \\
&=&
\left(\int y\zeta(dy)\right)^{n+\alpha}
\ = \
\lg \psi_1,\zeta\rg^{a}.
\end{eqnarray*}
Consequently (\ref{4.4}) holds
and the proof of Lemma 4.1 is complete.
\qed
|
\section{Introduction}
Piecewise-smooth stochastic systems are used as models of physical and biological systems \cite{Reimann2002,Gennes2005dryfriction,SanoKanazawa2016}. The interrelation between noise and discontinuities in such systems has attracted considerable attention recently
\cite{KawaradaHayakawa2004Non-Gaussian,Hayakawa2005Langevin,BauleCohenTouchette2010path,MenzelGoldenfeld2011,BauleTouchetteCohen2011path,
BauleSollich2012,BauleSollich2013,ChenJust2013,ChenJust2014,SanoHayakawa2014,GeffertJust2017}.
Some of them can be modeled by stochastic differential equations (SDEs) with piecewise-smooth drifts. Particularly, we consider in this paper the problems that can be modeled by the Langevin equation
\begin{equation}
\dot{v}(t)=\Phi(v)+\sqrt{2D}\xi(t),
\label{aa}
\end{equation}
where the overdot denotes the time derivative, the drift $\Phi(v)$ is discontinuous at some points, and $D>0$ represents the strength of the Gaussian white noise $\xi(t)$ that is characterized by the zero mean $\langle \xi(t) \rangle=0$ and the correlation $\langle \xi(t)\xi(t') \rangle=\delta(t-t')$. Here, the notation $\langle \cdots \rangle$ stands for the average over all possible realizations of the noise, and $\delta$ denotes the Dirac delta function. The initial condition for Eq.~(\ref{aa}) is set to be $v(0)=v_0$.
The theory of piecewise-smooth SDEs is only in its infancy
compared to its noiseless counterpart \cite{BernardoBudd2008}. For a few simple piecewise-smooth drifts, the propagators of Eq.~(\ref{aa}) are known analytically. For instance, when the drift is pure dry friction \cite{Gennes2005dryfriction} (also called solid friction or Coulomb friction), the
propagator is available in closed analytic form \cite{CaugheyDienes1961,Karatzas1984,TouchetteStraeten2010Brownian}. More generally, when the
drift is piecewise constant with a discontinuity (called the Brownian motion with a two-valued drift), the propagator can be expressed in terms of convolution integrals \cite{Karatzas1984,SimpsonKuske2014TwoValued}. Moreover, the distribution of the occupation time can also be obtained analytically \cite{Simpson2014OPT}. When the drift contains both dry friction and viscous friction, the propagator can be expressed as a sum of series \cite{TouchetteStraeten2010Brownian} or in connection with a Laplace transform \cite{TouchetteThomas2012Brownian}. For Eq.~(\ref{aa}) with dry friction the first two moments of the displacement and other integral functionals have also been obtained by solving backward Komogorov equations \cite{ChenJust2014II} or using the method based on the Pugachev-Sveshnikov equation \cite{Berezin2018}. However, there are vast cases that cannot be solved analytically by using existing theoretical methods. In those cases, we should resort to some effective numerical methods if we want to know the dynamics of Eq.~(\ref{aa}).
For instance, one can employ some numerical schemes to solve the SDE (\ref{aa}) directly. The Euler-Maruyama scheme is one of the simplest schemes that can be applied to obtain approximate results \cite{Leobacher2016}. However, there are errors arising from the approximations to discontinuities and the derivative. To address this issue, the so-called exact simulation was developed for solving Brownian motions with drift admitting a unique jump \cite{Etore2014, papaspiliopoulos2016}. The exact simulation involves only computer representation errors, enabling one to get exact samplings for the considered SDEs. In addition, the algorithm can be generalized to solve Brownian motions with drift admitting several jumps \cite{Dereudre2017}. Nevertheless, it requires heavy calculations to realize the exact simulation.
In this paper, we intend to solve the following Fokker-Planck equation directly, which governs the propagator of the model (\ref{aa}) with the Gaussian white noise:
\begin{equation}
\partial_t\, p=-\partial_v[\Phi(v)p]+D\partial^2_v p,
\label{ac}
\end{equation}
where $p=p(v,t|v_0,0)$ denotes the propagator with the initial condition $p(v,0|v_0,0)=\delta(v-v_0)$.
To solve Eq.~(\ref{ac}) with drift-admitting jumps, we need to apply two matching conditions at each jump of the drift, i.e., the continuity of the propagator and the continuity of the probability current (or flux)
\begin{equation}
f(v,t|v_0,0)=-\Phi(v)p+D\partial_v\, p. \label{ad}
\end{equation}
When the drift is continuous, there are many numerical methods that can be used to solve Eq.~(\ref{ac}); see for instance \cite{ChangCooper1970,LarsenLevermore1985,Langtangen1991,DrozdovMorillo1996,ZhangWei1997,Wei2000,LiuYu2014,PareschiZanella2018}. However, to the best of our knowledge, there are only few numerical results in the literature considering the cases with drift-admitting jumps. In \cite{MenzelGoldenfeld2011}, the authors transformed the Fokker-Planck equation with pure dry friction to a Schr\"{o}dinger equation with a delta potential, and then investigated the displacement statistics by solving a corresponding Brinkman hierarchy numerically. By treating the discontinuous drift carefully using a finite volume method \cite{ZhangChen2018} or an immerse interface method \cite{ZhangChen2017}, second-order schemes were developed for solving Eq.~(\ref{ac}). In this paper, we attempt to derive a finite difference scheme based on a grid staggered by flux points and solution points (see e.g. Fig.~\ref{fig_grid}). It will be seen later that the aforementioned matching conditions at jumps can be easily satisfied by using this grid, resulting in a simple way to treat the cases with drift-admitting jumps.
The rest of this paper is arranged as follows. In Sec.~\ref{sec_2} we take as an example the case with drift admitting two jumps to describe the procedure of the main algorithm for the \blue{spatial} discretization. The corresponding staggered grid is also introduced. Then we present the finite difference scheme in Sec.~\ref{sec_3}. Some benchmark problems are solved numerically in Sec.~\ref{sec_4} to show the validity of the scheme. In Sec.~\ref{sec_5}, we extend the algorithm to study the displacement of the Brownian motion with pure dry friction. Finally, conclusions are drawn in Sec.~\ref{sec_6}.
\section{Staggered grid}
\label{sec_2}
We describe the algorithm by assuming that the drift in Eq.~(\ref{aa}) admits two jumps at $v=v_{d_1}, v_{d_2}$ ($v_{d_1}<v_{d_2}$)\red{, respectively}. For other cases, the algorithm can be generalized straightforwardly
according to the number of jumps.
For Eq.~(\ref{ac}) defined for $v\in (-\infty,\infty)$, we first truncate the domain into a finite interval, denoted by $[v_{_L},v_{_R}]$, containing the two discontinuous points. Then by using these two points we partition the interval into three subdomains, i.e., $\Omega_1=[v_{_L}, v_{d_1}]$, $\Omega_2=[v_{d_1}, v_{d_2}]$ and $\Omega_3=[v_{d_2},v_{_R}]$. As illustrated in Fig.~\ref{fig_grid}, a grid staggered by flux points and solution points is used for the partitioned subdomains. In particular, the discontinuous points $v_{d_1}$ and $v_{d_2}$ are both set to be solution points such that the continuity condition\red{s} of the propagator are satisfied automatically for the discrete method.
In each subdomain $\Omega_i$, the grid points are set to be uniformly distributed with the solution points defined by
\begin{align}
\begin{cases}
v_{1,j}=v_{_L}+(j-1/2)h_1, & 1\leqslant j \leqslant N_1,\\
v_{2,j}=v_{d_1}+(j-1)h_2, & 1\leqslant j \leqslant N_2,\\
v_{3,j}=v_{d_2}+(j-1)h_3, & 1\leqslant j \leqslant N_3,\\
\end{cases}
\end{align}
where $N_i$ are the numbers of solution points and $h_i$ the \blue{spatial} steps for the subdomains,
\begin{align}
\begin{cases}
h_1=(v_{d_1}-v_{_L})/(N_1-1/2),\\
h_2=(v_{d_2}-v_{d_1})/(N_2-1),\\
h_3=(v_{_R}-v_{d_2})/(N_3-1/2).
\end{cases}
\end{align}
Especially, we have $
v_{1,N_1}=v_{2,1}=v_{d_1}
$ and
$
v_{2,N_2}=v_{3,1}=v_{d_2}
$.
The flux points $v_{i,j+1/2}$ are defined by
\begin{equation}
\begin{cases}
v_{1,j+1/2}=v_{_L}+jh_1, & 0\leqslant j \leqslant N_1-1,\\
v_{2,j+1/2}=v_{d_1}+(j-1/2)h_2, & 1\leqslant j \leqslant N_2-1,\\
v_{3,j+1/2}=v_{d_2}+(j-1/2)h_3, & 1\leqslant j \leqslant N_3.\\
\end{cases}
\end{equation}
Particularly, we have $ v_{1,1/2}=v_{_L} $ and $ v_{3,N_3+1/2} = v_{_R} $, i.e., the end points
of the interval $ [v_{_L}, v_{_R}] $ are both flux points, which are designed to impose boundary conditions.
\begin{figure*}
\begin{center}
\includegraphics[width=0.9\linewidth]{grid1}
\end{center}
\caption{Illustration of the grid \blue{staggered by flux points and solution points for the case with two jumps at $v=v_{d_1}$ and $v=v_{d_2}$, respectively. The two jumps are both set to be
solution points and used to divide the computational domain $[v_{_L},v_{_R}]$ into three subdomains $\Omega_i$ ($i=1,2,3$).}}
\label{fig_grid}
\end{figure*}
Given initial values at the solution points of the staggered grid, a finite difference scheme \red{for Eq.~(\ref{ac})} can be constructed by the following procedure:
\begin{enumerate}
\item[(i)] Within each subdomain $\Omega_i$, obtain the solutions at the flux points by using interpolation schemes. For the purpose of stability, \emph{upwind interpolations} are used here according to the sign of the drift $\Phi(v)$. If $\Phi(v)$ changes its sign within $\Omega_i$, \mage{we need to
split the drift into an appropriate form to apply upwind interpolations. Here we split the drift into two parts: $\Phi(v)=\min\{\Phi(v),0\}+\max\{\Phi(v),0 \}$, ensuring that each part does not change its sign. Then using this split form we can approximate the term $\Phi(v)p$ appearing in Eq.~(\ref{ac}) at the flux point $v_{i,j+1/2}$ by}
\begin{align}
\min&\{\Phi(v_{i,j+1/2}),0\}p^+_{i,j+1/2}\nonumber\\
&+\max\{\Phi(v_{i,j+1/2}),0 \}p^-_{i,j+1/2},
\label{eq_drift_spliting}
\end{align}
where $p^+_{i,j+1/2}$ and $p^-_{i,j+1/2}$ are the approximate values of $p$ at $v_{i,j+1/2}$, respectively, obtained by using interpolations with
stencils as illustrated in Fig.~\ref{fig_grid_ill};
\item[(ii)] Evaluate the first derivative of $p$ at flux points by using
\emph{difference schemes} in each subdomain $\Omega_i$;
\item[(iii)] Obtain the values of the current (\ref{ad}) at fluxes points by using the above two steps. Then approximate the values of the derivative of the current at solution points by using a difference scheme, which is designed for the domain $[v_{_L}, v_{_R}]$ directly since the current (\ref{ad}) is theoretically continuous everywhere.
\end{enumerate}
\begin{figure}
\begin{center}
\includegraphics[width=0.75\linewidth]{grid_upwind}
\end{center}
\caption{Illustration of the stencils \blue{ used to reconstruct the values $p_{i,j+1/2}^\pm$ required by the approximation (\ref{eq_drift_spliting}). Here only the stencils of the fifth-order interior interpolations are presented. Near the boundaries the stencils should be adjusted accordingly. } }
\label{fig_grid_ill}
\end{figure}
\section{Scheme}
\label{sec_3}
To discretize the right side of Eq.~(\ref{ac}), we follow the aforementioned procedure: first calculate the values of the probability current at flux points and then derive a difference scheme to evaluate the derivative of the current. Here the \blue{spatial} scheme is designed to be fifth-order for the cases with smooth drifts. Finally, a third-order Runge-Kutta scheme is employed to solve the resulting ordinary differential system.
\subsection{Evaluation of the \red{p}robability current}
\label{sec_3a}
There are two terms appearing in the current (\ref{ad}). For the first term $\Phi(v)p$, we use interpolation schemes to reconstruct the required values in the \red{approximation} (\ref{eq_drift_spliting}). For the second term $D\partial_v p$, we derive difference schemes to approximate it.
\blue{In the following, we will present the schemes in matrix forms, where the entries of the matrices are all easily obtained by using Lagrangian interpolations according to specified stencils.
For example, if we consider the stencil $S=\{v_{1,1},v_{1,2},\dots,v_{1,5}\}$,
then at any point $v$, the values of $p$ and $\partial_v p$ are approximated respectively by
\begin{align}
p_{_I}(v)=&\sum_{k=1}^5 l_k(v) p_{1,k}, \\
\frac{d}{dv} p_{_I}(v)=&\sum_{k=1}^5 \frac{d}{dv} l_k(v) p_{1,k},
\end{align}
where
\begin{equation}
l_{k}(v)=\prod_{s=1,s\neq k}^5 \frac{v-v_{1,s}}{v_{1,k}-v_{1,s}}.
\end{equation}
}
Since the grid points are slightly different in different subdomains (see Fig.~\ref{fig_grid}),
we describe the schemes separately for each subdomain $\Omega_i$.
\blue{For convenience of notations, we will first consider the subdomain $\Omega_2$, and then
$\Omega_1$ and $\Omega_3$.}
\subsubsection{Subdomain $\Omega_2$}
To describe the scheme compactly, let us introduce the vectors
$\mathbf{p}^\pm_2=[ p^\pm_{2,3/2},p^\pm_{2,5/2},\dots,p^\pm_{2,N_2-1/2} ]^T$ and
the vector $\mathbf{p}_2=[ p_{2,1},p_{2,2},\dots,p_{2,N_2} ]^T$,
where the ``\blue{$T$}'' denotes the transpose operation. Then according to the grid point distribution in $\Omega_2$ we can compute the values at flux points by the fifth-order interpolation schemes $ \mathbf{p}^\pm_2=I_2^\pm \mathbf{p}_2 $, \blue{where $I_2^\pm$ are both $(N_2-1) \times N_2$
matrices. Here}
\begin{equation}
I_2^+=\begin{bmatrix}
\frac{35}{128} & \frac{35}{32} & -\frac{35}{64} & \frac{7}{32} &
-\frac{5}{128} & & \\[3pt]
-\frac{5}{128} & \frac{15}{32} & \frac{45}{64} & -\frac{5}{32} &
\frac{3}{128} & & \\[3pt]
& \ddots & \ddots & \ddots & \ddots &
\ddots & \\[3pt]
& & -\frac{5}{128} & \frac{15}{32} & \frac{45}{64} & -\frac{5}{32} &
\frac{3}{128} \\[3pt]
& & \frac{3}{128} & -\frac{5}{32} & \frac{45}{64} & \frac{15}{32} &
-\frac{5}{128} \\[3pt]
& & -\frac{5}{128} & \frac{7}{32} & -\frac{35}{64} & \frac{35}{32} &
\frac{35}{128}
\end{bmatrix},
\label{eq_matrix_I2p}
\end{equation}
and $I_2^-$ is defined by letting its entries satisfy that $[I_2^-]_{j,k}=[I_2^+]_{N_2-j,N_2+1-k}$, $1\leqslant j \leqslant N_2-1$, $1\leqslant k \leqslant N_2$.
Denote the approximation to $\partial_v p$ at the flux point $v_{2,j+1/2}$ as $(\partial_v p)_{2,j+1/2}$ and introduce the vector
\[\mathbf{p}_{2,v}=[ (\partial_v p)_{2,3/2},(\partial_v p)_{2,5/2},\dots,(\partial_v p)_{2,N_2-1/2}]^T.\]
We can write the difference scheme as
$
\mathbf{p}_{2,v}=A_2\mathbf{p}_2/h_2,
$
where the $(N_2-1)\times N_2$ matrix $A_2$ is
\begin{equation}
A_2=\begin{bmatrix}
-\frac{11}{12} & \frac{17}{24} & \frac{3}{8} & -\frac{5}{24} & \frac{1}{24} &
& & \\[3pt]
\frac{1}{24} & -\frac{9}{8} & \frac{9}{8} & -\frac{1}{24} & & & & \\[3pt]
-\frac{3}{640} & \frac{25}{384} & -\frac{75}{64} & \frac{75}{64} &
-\frac{25}{384} & \frac{3}{640} & & \\[3pt]
& \ddots & \ddots & \ddots & \ddots &
\ddots & \ddots & \\[3pt]
& & -\frac{3}{640} & \frac{25}{384} & -\frac{75}{64} & \frac{75}{64} &
-\frac{25}{384} & \frac{3}{640} \\[3pt]
& & & & \frac{1}{24} & -\frac{9}{8} & \frac{9}{8} & -\frac{1}{24} \\[3pt]
& & & -\frac{1}{24} & \frac{5}{24} & -\frac{3}{8} & -\frac{17}{24} &
\frac{11}{12}
\end{bmatrix},
\label{eq_matrix_a2}
\end{equation}
such that the difference scheme is sixth-order at $v_{2,j+1/2}$ with $3\leqslant j \leqslant N_2-3$ and fourth-order at the other flux points.
\subsubsection{Subdomain $\Omega_1$}
Introduce $\mathbf{p}^\pm_1=[ p^\pm_{1,1/2},p^\pm_{1,3/2},\dots,p^\pm_{1,N_1-1/2} ]^T$ and $\mathbf{p}_1=[ p_{1,1},p_{1,2},\dots,p_{1,N_1} ]^T$.
We first compute the right vector $\mathbf{p}^+_1$ by using the fifth-order interpolation scheme $ \mathbf{p}^+_1=I_1^+\mathbf{p}_1 $ with the $N_1\times N_1$ matrix $I_1^+$ written as
\blue{
\begin{equation*}
I_1^+=\begin{bmatrix}
\mathbf{a} \\
(I_2^+)_{(N_1-1)\times N_1}
\end{bmatrix}.
\end{equation*}
Here
\begin{equation}
\mathbf{a}=
\begin{bmatrix}
\frac{315}{128} & -\frac{105}{32} & \frac{189}{64} & -\frac{45}{32} &
\frac{35}{128} & 0 & \dots & 0
\end{bmatrix}
\label{eq_def_b}
\end{equation}
is a $1\times N_1$ vector and the matrix $(I_2^+)_{(N_1-1)\times N_1}$ is defined in Eq.~(\ref{eq_matrix_I2p}) by replacing $N_2$ with $N_1$ (The same notation method will be used throughout this paper).
}For the left values, first let $p^-_{1,1/2}=p^+_{1,1/2}$. Then we derive the interpolation scheme $ \mathbf{p}^-_1=I_1^-[p^-_{1,1/2},\mathbf{p}_1^T]^T $ with the $N_1\times (N_1+1)$
matrix $I_1^-$ reading as
\begin{equation}
I_1^-= \begin{bmatrix}
1 & & & & & & & \\[3pt]
-\frac{1}{7} & \frac{5}{8} & \frac{5}{8} & -\frac{1}{8} & \frac{1}{56} & &
& \\[3pt]
\frac{3}{35} & -\frac{1}{4} & \frac{3}{4} & \frac{9}{20} & -\frac{1}{28} &
& & \\[3pt]
& \frac{3}{128} & -\frac{5}{32} & \frac{45}{64} & \frac{15}{32} &
-\frac{5}{128} & & \\[3pt]
& & \ddots & \ddots & \ddots & \ddots &
\ddots & \\[3pt]
& & & \frac{3}{128} & -\frac{5}{32} & \frac{45}{64} & \frac{15}{32} &
-\frac{5}{128} \\[3pt]
& & & -\frac{5}{128} & \frac{7}{32} & -\frac{35}{64} & \frac{35}{32} &
\frac{35}{128}
\end{bmatrix}.
\label{eq_matrix_I1m}
\end{equation}
Introducing the vector
\[\mathbf{p}_{1,v}=[ (\partial_v p)_{1,1/2},(\partial_v p)_{1,3/2},\dots,(\partial_v p)_{1,N_1-1/2} ]^T,\]
we can derive the difference scheme $ \mathbf{p}_{1,v}=A_1 \mathbf{p}_1/h_1 $ with the $N_1\times N_1$ matrix $A_1$ written as
\blue{
\begin{equation*}
A_1=\begin{bmatrix}
\mathbf{b}\\
(A_2)_{(N_1-1)\times N_1}
\end{bmatrix},
\end{equation*}
where
\begin{equation}
\mathbf{b}=\begin{bmatrix}
-\frac{31}{8} & \frac{229}{24} & -\frac{75}{8} & \frac{37}{8} &
-\frac{11}{12} & 0 & \dots & 0 \\
\end{bmatrix}
\end{equation}
is a $1\times N_1$ vector and $A_2$ is defined in Eq.~(\ref{eq_matrix_a2}),
such that } the difference scheme is sixth-order at $v_{1,j+1/2}$ with $ 3\leqslant j\leqslant N_1-3$ and fourth-order at the other flux points.
\subsubsection{Subdomain $\Omega_3$}
\blue{The schemes for this subdomain are basically the same as
those of the subdomain $\Omega_2$ if we swap the direction.} Introduce
$\mathbf{p}^\pm_3=[ p^\pm_{3,3/2},p^\pm_{3,5/2},\dots,p^\pm_{3,N_3+1/2} ]^T$ and $\mathbf{p}_3=[ p_{3,1},p_{3,2},\dots,p_{3,N_3} ]^T$.
We first reconstruct the vector $\mathbf{p}^-_3$ by using the fifth-order interpolation scheme $\mathbf{p}^-_3=I_3^-\mathbf{p}_3$ with \blue{the entries of} the $N_3\times N_3$ matrix $I_3^-$ defined by
\blue{
$
[I_3^-]_{j,k}=[(I_1^+)_{N_3\times N_3}]_{N_3+1-j,N_3+1-k}
$.
}Then the right values are calculated by the fifth-order interpolation scheme $\mathbf{p}^+_3=I_3^+[\mathbf{p}_3^T,p^+_{3,N_3-1/2}]^T$, where the
entries of the $N_3\times (N_3+1)$ matrix $I_3^+$ are derived to be
\blue{
$
[I_3^+]_{j,k}=[(I_1^-)_{N_3\times (N_3+1)}]_{N_3+1-j,N_3+2-k}
$}. Note that we have assumed $p^+_{3,N_3-1/2}=p^-_{3,N_3-1/2}$ here.
Similarly, by introducing the vector
\[
\mathbf{p}_{3,v}=[ (\partial_v p)_{3,3/2},(\partial_v p)_{3,5/2},\dots,(\partial_v p)_{3,N_3+1/2} ]^T,
\]
the derivative values at flux points are approximated by
$ \mathbf{p}_{3,v}=A_3 \mathbf{p}_{3} /h_3 $ with \blue{the entries} of the $N_3\times N_3$ matrix $A_3$ defined by $\blue{ [A_3]_{j,k}=-[(A_1)_{N_3\times N_3}]_{N_3+1-j,N_3+1-k}}$, such that the difference scheme is sixth-order at $v_{3,j+1/2}$ with $3\leqslant j \leqslant N_3-3$ and fourth-order at the other flux points.
\subsubsection{Imposing boundary conditions}
Using the above schemes derived for subdomains, we can obtain the values of the probability current at all flux points. However, it is noted that we have not used any boundary conditions
so far. As we will see later in Sec.~\ref{sec_4}, depending on the signs of the drift at the domain boundaries we may need to set the computational domain to be large enough and impose reflecting boundary conditions \cite{Veestraeten2004} appropriately. In that cases, we just \red{reset} the current values at the boundaries to be zero.
\subsection{Derivative of the current}
Now we are ready to derive a difference scheme to compute the derivative of the current using
the values at flux points obtained in Sec.~\ref{sec_3a}. To get a correct solution, it is no doubt that we have to consider information transmission between different subdomains $\Omega_i$. As mentioned before, although the derivative of the propagator is not continuous at jumps, the current is continuous everywhere. Therefore, we can derive a difference scheme for the whole domain directly to approximate the derivative of the current. However, as it allows different \blue{spatial} steps in different subdomains, we have to pay attention to the solution points near the jumps.
For convenience of notations, let us introduce the flux vector
$
\mathbf{f}=[ \mathbf{f}_1^T,\mathbf{f}_2^T,\mathbf{f}_3^T ]^T
$, where
\begin{align}
\mathbf{f}_1=[ f_{1,1/2},f_{1,3/2},\dots, f_{1,N_1-1/2} ]^T,\\
\mathbf{f}_2=[ f_{2,3/2},f_{1,5/2},\dots,f_{2,N_2-1/2} ]^T,\\
\mathbf{f}_3=[ f_{3,3/2},f_{1,5/2},\dots,f_{3,N_3+1/2} ]^T.
\end{align}
The values of the derivative $\partial_v f$ at solution points are denoted by
$
\mathbf{f}_v=[ \mathbf{f}_{1,v}^T,\mathbf{f}_{2,v}^T,\mathbf{f}_{3,v}^{T} ]^T
$ with
\begin{align}
\mathbf{f}_{1,v}=[ (\partial_v f)_{1,1},(\partial_v f)_{1,2},\dots,(\partial_v f)_{1,N_1} ]^T, \\
\mathbf{f}_{2,v}=[ (\partial_v f)_{2,2},(\partial_v f)_{2,3},\dots,(\partial_v f)_{2,N_2} ]^T, \\
\mathbf{f}_{3,v}=[ (\partial_v f)_{3,2},(\partial_v f)_{3,3},\dots,(\partial_v f)_{3,N_3} ]^T.
\end{align}
Then we attempt to derive a \red{derivative} matrix $A$ such that
$\mathbf{f}_v=A \mathbf{f}$. Here the size of $A$ is $N_v \times (N_v+1)$ with $N_v=N_1+N_2+N_3-2$.
By observing the distribution of the grid points, we design the difference scheme by
using the stencils
\begin{align}
[\mathbf{f}_v]_j =
\begin{cases}
\sum_{k=1}^5 a_{j,k}[\mathbf{f}]_k, & 1\leqslant j\leqslant 2, \\[2pt]
\sum_{k=1}^6 a_{j,k}[\mathbf{f}]_{j+k-3}, & 3\leqslant j \leqslant N_v-2,\\[2pt]
\sum_{k=1}^5 a_{j,k}[\mathbf{f}]_{N_v+k-4}, & N_v-1\leqslant j \leqslant N_v,\\
\end{cases}
\label{eq_diff}
\end{align}
where $[\mathbf{f}]_k$ denotes the $k$-th entry of the vector $\mathbf{f}$ and
the coefficients $a_{j,k}$ can be determined directly by using Lagrangian interpolations.
Hence the difference matrix $A$ can be easily written down following the above stencils.
We first present the coefficients of the cases with stencils in a single subdomain.
The results are as follows:
\begin{align}
[a_{j,k}]_{1\leqslant k \leqslant 5}&=
\begin{cases}
\frac{1}{h_1}
[
-\frac{11}{12} , \frac{17}{24} , \frac{3}{8} , -\frac{5}{24} , \frac{1}{24}
], & j=1, \\[2pt]
\frac{1}{h_1}
[
\frac{1}{24} , -\frac{9}{8} , \frac{9}{8} , -\frac{1}{24} , 0
], & j=2,
\end{cases}
\\
[a_{j,k}]_{1\leqslant k \leqslant 6}&=
\begin{cases}
\frac{1}{h_1} \mathbf{c}, & 3\leqslant j\leqslant N_1-3,\\[2pt]
\frac{1}{h_2} \mathbf{c}, & N_1+3\leqslant j \leqslant N_1+N_2-4,\\[2pt]
\frac{1}{h_3} \mathbf{c}, & N_1+N_2+2\leqslant j \leqslant N_v-2,
\end{cases}
\\
[a_{j,k}]_{1\leqslant k \leqslant 5}&=
\begin{cases}
\frac{1}{h_3}
[
0 , \frac{1}{24} , -\frac{9}{8} , \frac{9}{8} , -\frac{1}{24}
], & j=N_v-1, \\[2pt]
\frac{1}{h_3}[
-\frac{1}{24}, \frac{5}{24}, -\frac{3}{8}, -\frac{17}{24}, \frac{11}{12}], & j=N_v,
\end{cases}
\end{align}
where the vector
\begin{equation}
\mathbf{c} =
\begin{bmatrix}
-\tfrac{3}{640} & \tfrac{25}{384} & -\tfrac{75}{64} &
\tfrac{75}{64} & -\tfrac{25}{384} & \tfrac{3}{640}
\end{bmatrix}.
\end{equation}
For the other cases, we have to pay attention to the fact that the stencils (\ref{eq_diff}) are across the jumps, as illustrated in Fig.~\ref{fig_discontinuities}.
When $N_1-2\leqslant j \leqslant N_1+2$, \red{by using Lagrangian interpolations}
we can determine the coefficients to be
\begin{equation}
a_{j,k}=d_{j-N_1+3,k}(h_1,h_2), \quad 1\leqslant k \leqslant 6,
\end{equation}
where $d_{s,k}(x,y)$ are functions of $x$ and $y$, as shown in Tab.~\ref{diff_coefficients_cross}.
Similarly, when $N_1+N_2-3\leqslant j \leqslant N_1+N_2+1$, we have
\begin{equation}
a_{j,k}=d_{j-N_1-N_2+4,k}(h_2,h_3), \quad 1\leqslant k \leqslant 6.
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{jump_grid}
\end{center}
\caption{Illustration of the grid points near a jump. \blue{ The spatial steps on the left and on the right of the jump are $h_1$ and $h_2$, respectively.} }
\label{fig_discontinuities}
\end{figure}
\begin{table*}
\begin{center}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}cccc}
\hline
& $ s=1 $ & $ s=2 $ & $ s=3 $ \\
\cline{2-4}
\vspace{-0.8em}\\
$d_{s,1}(x,y)$
& $ -\frac{3}{64 (9 x+y)} $
& $ \frac{-8 x^2-4 x y+3 y^2}{24 x (7 x+y) (7 x+3 y)} $
& $ \frac{60 y^3-69 x y^2}{20 x(x+y) (5 x+y) (5 x+3 y)} $
\\[4pt]
$d_{s,2}(x,y)$
& $ \frac{23 x+2 y}{336 x^2+48 x y}$
& $ \frac{16 x^2+12 x y-3 y^2}{8 x (5 x+y) (5 x+3 y)} $
& $ \frac{5 (23 x-18 y) y^2}{6 x (x+y) (3 x+y) (3 x+5 y)} $
\\[4pt]
$d_{s,3}(x,y)$
& $ -\frac{9 (21 x+4 y)}{32 x (5 x+y)}$
& $ -\frac{88 x^2+116 x y+21 y^2}{24 x (x+y) (3 x+y)} $
& $ \frac{15 y^2 (8 y-23 x)}{4 x (x+y) (x+3 y) (x+5 y)} $
\\[4pt]
$d_{s,4}(x,y)$
& $ \frac{57 x+18 y}{48 x^2+16 x y} $
& $ \frac{32 x^2+124 x y+69 y^2}{24 x (x+y) (x+3 y)} $
& $ \frac{15 x^2 (23 y-8 x)}{4 y (x+y) (3 x+y) (5 x+y)} $
\\[4pt]
$d_{s,5}(x,y)$
& $ -\frac{17 x+8 y}{192 x (x+y)} $
& $ -\frac{x^3 (x+24 y)}{y (x+y) (3 x+y) (5 x+y) (7 x+y)} $
& $ \frac{5 x^2 (18 x-23 y)}{6 y (x+y) (x+3 y) (5 x+3 y)} $
\\[4pt]
$d_{s,6}(x,y)$
& $ \frac{18 x^4}{(x+y) (3 x+y) (5 x+y) (7 x+y) (9 x+y)} $
& $ \frac{x^3 (x+8 y)}{3 y (x+y) (x+3 y) (5 x+3 y) (7 x+3 y)} $
& $ \frac{3 x^2 (23 y-20 x)}{20 y (x+y) (x+5 y) (3 x+5 y)} $ \\[4pt]
\hline
\end{tabular*}
\end{center}
\caption{Coefficients that determine the difference scheme (\ref{eq_diff}) for the cases $N_1-2\leqslant j \leqslant N_1+2$ and $N_1+N_2-3\leqslant j \leqslant N_1+N_2+1$. \blue{When
$s=4$ and $5$, the coefficients are determined by the relation $d_{s,j}(x,y)=-d_{6-s,7-j}(y,x)$ for
$1\leqslant j \leqslant 6$.}}
\label{diff_coefficients_cross}
\end{table*}
\subsection{Time-marching scheme}
Approximating the right side of Eq.~(\ref{ac})
by using the above finite difference scheme, we obtain
a semi-discretized system, denoted by
\begin{equation}
\frac{d \mathbf{p}}{dt}= R(\mathbf{p},t),
\end{equation}
where $\mathbf{p}$ stands for the vector of the unknowns at solution points, and $R(\mathbf{p},t)$
represents the right hand side term. Then many time-marching schemes can be used to solve this
system. In this paper we employ a traditional third-order Runge-Kutta scheme, written as
\begin{align}
\mathbf{p}^{n+1}=\mathbf{p}^n+\frac{1}{9}\tau (2K_1+3K_2+4K_3),\label{rk_1}\\
\begin{cases}
K_1=R( \mathbf{p}^n, t_n ),\\
K_2=R( \mathbf{p}^n+\frac{1}{2}\tau K_1, t_n+\frac{1}{2}\tau ),\\
K_3=R( \mathbf{p}^n+\frac{3}{4}\tau K_2, t_n+\frac{3}{4}\tau ),
\end{cases}
\label{rk_2}
\end{align}
where $\mathbf{p}^n$ denotes the value of $\mathbf{p}$ at time $t_n$ and $\tau$ is the time step.
\section{Numerical examples}
\label{sec_4}
In this section, some benchmark problems are solved numerically to show
the validity of the algorithm presented above. The discrete $L^2$-norm error for the case with two jumps is defined by
\begin{align}
L^2 \mbox{ error}=&\Bigg[\sum_{j=1}^{N_1-1} e_{1,j}^2h_1 +\frac{1}{2}e_{1,N_1}^2(h_1+h_2)+\sum_{j=2}^{N_2-1}e_{2,j}^2h_2\nonumber\\
&+\frac{1}{2} e_{2,N_2}^2(h_2+h_3)+\sum_{j=2}^{N_3}e_{3,j}^2h_3\Bigg]^{1/2},
\label{eq_l2error}
\end{align}
where $e_{i,j}=p_{i,j}-p(v_{i,j})$ are the errors between numerical results and exact solutions.
\red{In addition,} the $L^\infty$-norm error is defined by
\begin{equation}
L^\infty \mbox{ error}=\max_{i,j}\{|e_{i,j}|\}.
\label{eq_linfty}
\end{equation}
For other cases, the errors are defined similarly.
The numerical convergence rate is defined by
\begin{equation}
\mbox{rate}=-\ln( E_{M}/E_{N} )/\ln( M/N ),
\label{eq_rate}
\end{equation}
where $E_{M}$ and $E_{N}$ are the errors corresponding to the cases with $M$ and $N$ solution points, respectively.
In numerical computations, the initial condition of Eq.~(\ref{ac}) given by a delta function cannot be used directly. Instead, if an exact solution to Eq.~(\ref{ac}) is available we choose the initial condition to be $p(v,\tau_0|v_0,0)$ and start the computation from $t=\tau_0$. Here
$\tau_0$ is a constant that can be chosen appropriately for the considered problems. Otherwise,
the initial condition is set to be Gaussian,
\begin{equation}
p(v,\tau_0|v_0,0)=\frac{1}{ \sqrt{4\pi D \tau_0} }e^{ -[v-v_0-\Phi(v_0)\tau_0]^2/(4D \tau_0) },
\label{eq_v}
\end{equation}
which mimics the delta function when $\tau_0$ is small. For convenience, \red{$D=0.5$ and $\tau_0=0.01$} are chosen for all test cases in this section.
It should be noted that the proposed finite difference scheme can also be applied to solve problems with continuous drifts. In the following, we first show that the scheme is actually fifth-order for smooth cases. Then we pay attention to the cases with drift-admitting jumps, where a second-order convergence rate is observed.
\subsection{Smooth drifts}
The following \red{two} examples are used to confirm \red{that
the scheme described in Sec.~\ref{sec_3} is fifth-order for the cases with smooth drifts.}
Here $v_{d_1}=0$ and $v_{d_2}=1$ are used to divide all the computational domains (see Fig.~\ref{fig_grid}) and the time step is set to be $\tau=0.01\min\{h_1^2,h_2^2,h_3^2\}$.
\subsubsection{Constant drift} \label{sec_4aa}
When $\Phi(v)=\mu$ with $\mu$ being a constant, Eq.~(\ref{aa}) corresponds to the Brownian motion with a constant drift, whose propagator is simply Gaussian,
\begin{equation}
p(v,t|v_0,0)=\frac{1}{ \sqrt{4\pi Dt} }e^{ -(v-v_0-\mu t)^2/(4Dt) }.
\label{eq_constant}
\end{equation}
Here $\mu=1$ and $v_0=0$ are chosen and the solution domain is truncated to be
$[-5,10]$. While a zero current boundary condition is set for the left boundary, i.e., $f(-5,t|v_0,0)=0$, no boundary condition is needed for the right due to the positiveness of the chosen $\mu$. The numerical results obtained in Tab.~\ref{tab_constant} show that for this smooth case the algorithm proposed in this paper achieves a fifth-order convergence rate approximately, \blue{ while only approximately second-order for the Chang-Cooper scheme \cite{ChangCooper1970} (see Appendix \ref{app_sec1}), which is one of the most popular schemes for solving Fokker-Planck equations.} As we can see from Fig.~\ref{fig_constant}, when $t=8$ the current is much large than zero at the right boundary\red{. But} the proposed scheme still produces a solution that matches with the exact solution.
This means that the computational domain is not necessary to be large to avoid boundary refection here. It is no doubt that this property is very desirable in numerical simulations.
\begin{table}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}cccccccc}
\hline
\multicolumn{8}{c}{The current method} \\
\hline
$N_1$ &$N_2$ &$N_3$ & $N_v$ & $L^2$ error & rate & $L^\infty$ error & rate \\
\hline
40 & 10 & 20 & 68 & 1.39E-03 & -- & 1.94E-03\\
80 & 20 & 40 & 138 & 5.16E-05 &4.66& 5.21E-05& 5.11\\
160 & 40 & 80 & 278& 1.77E-06& 4.82 & 1.46E-06 &5.11\\
320 & 80 & 160 & 558 & 5.09E-08 & 5.09 &4.04E-08 & 5.15
\\\hline
\multicolumn{8}{c}{ \blue{The Chang-Cooper scheme} } \\
\hline
-- &-- &-- & \blue{ $N_v$ } & \blue{$L^2$ error} & \blue{rate} & \blue{$L^\infty$ error } &
\blue{ rate } \\
\hline
--& -- & -- & 68 & 6.73E-01 & -- & 9.35E-01\\
--& -- & -- & 138 & 2.50E-01 &1.40& 3.25E-01 &1.49\\
--& -- & -- & 278 & 3.80E-02& 2.69 & 5.80E-02 &2.46\\
--& -- & -- & 558 & 1.02E-02 & 1.89 &1.58E-02 & 1.87
\\\hline
\end{tabular*}
\caption{Accuracy test for Eq.~(\ref{ac}) with $\Phi(v)=1$ at time $t=1$.
The total number of solution points is $N_v=N_1+N_2+N_3-2$. \blue{The Chang-Cooper scheme for Eq.~(\ref{ac}) with constant drift $\Phi(v)=\mu$ is presented in Appendix \ref{app_sec1}.
Here, the errors are computed according to Eqs.~(\ref{eq_l2error}) and (\ref{eq_linfty}), and the rates are defined by Eq.~(\ref{eq_rate}).}}
\label{tab_constant}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{constant_t8}
\end{center}
\caption{Numerical result of Eq.~(\ref{ac}) with $\Phi(v)=1$ at time $t=8$. Here, $N_1=40$, $N_2=10$ and $N_3=20$ \blue{ are used to compute the numerical result, which matches well with the exact solution (\ref{eq_constant}). } }
\label{fig_constant}
\end{figure}
\subsubsection{Ornstein-Uhlenbeck process}
When $\Phi(v)=-\gamma v$ with $\gamma$ being a constant, Eq.~(\ref{aa}) corresponds to the Ornstein-Uhlenbeck process.
In this case, it is well known that Eq.~(\ref{ac}) admits the solution
\begin{equation}
p(v,t|v_0,0)=\sqrt{ \frac{\gamma}{2\pi D (1-e^{-2\gamma t}) } }\exp\left( -\frac{\gamma ( v-e^{ -\gamma t } v_0 )^2 }{ 2D( 1-e^{-2\gamma t}) } \right).
\label{eq_ou}
\end{equation}
For $\gamma>0$ and $t\rightarrow \infty$, the solution tends to the stationary solution
\begin{equation}
p_{_{\mathrm{OU}}}(v)=\sqrt{ \frac{\gamma}{2\pi D } }e^{-\gamma v^2/(2D) }.
\end{equation}
For $\gamma\leqslant 0$, no stationary solution exists.
Here we consider computations for the two cases $\gamma=1$ and $-1$. The computational domain $[-5,5]$ is chosen for both the cases. While zero current boundary conditions are set for the first case, no particular boundary condition is needed for the second. As we can see in Tab.~\ref{tab_ou} that the proposed scheme achieves fifth-order accuracy approximately for the two cases. Numerical results for larger time as shown in Fig.~\ref{fig_ou} confirm that the scheme is also \red{valid} for long time simulations, even for negative $\gamma$ without a large computational domain.
\begin{table}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}cccccccccccccc}
\hline
&&&& \multicolumn{4}{c}{$\gamma=1$}\\
\cline{5-8}
$N_1$ &$N_2$ &$N_3$ & $N_v$ & $L^2$ error & rate & $L^\infty$ error & rate \\
\hline
40 & 10 & 20 & 68 & 4.11E-04 & -- & 3.93E-04 &--\\
80 & 20 & 40 & 138 & 5.18E-05 & 2.93 & 5.63E-05 & 2.74
\\
160 & 40 & 80 & 278& 2.06E-06 &4.60 &2.03E-06 &4.75
\\
320 & 80 & 160 & 558 & 4.10E-08& 5.62 &3.86E-08 &5.69
\\\hline
&&&& \multicolumn{4}{c}{$\gamma=-1$}\\
\cline{5-8}
$N_1$ &$N_2$ &$N_3$ & $N_v$ & $L^2$ error & rate & $L^\infty$ error & rate \\
\hline
40 & 10 & 20 & 68 & 3.29E-04 & -- & 2.87E-04 &--
\\
80 & 20 & 40 & 138 & 5.59E-05 & 2.50 & 4.53E-05& 2.61
\\
160 & 40 & 80 & 278& 2.34E-06 & 4.53 & 1.73E-06& 4.67
\\
320 & 80 & 160 & 558 & 4.75E-08 & 5.60 & 3.36E-08 & 5.65
\\\hline
\end{tabular*}
\caption{Accuracy test for Eq.~(\ref{ac}) with $\Phi(v)=-\gamma v$ at time $t=0.5$ \blue{for two different values of $\gamma$}. Here $N_v=N_1+N_2+N_3-2$. \blue{The errors are computed according to Eqs.~(\ref{eq_l2error}) and (\ref{eq_linfty}), and the rates are defined by Eq.~(\ref{eq_rate}).} }
\label{tab_ou}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{ou_gamma1}\\\vspace{0.5em}
\includegraphics[width=0.9\linewidth]{ou_gamma_m1}
\end{center}
\caption{Numerical results of Eq.~(\ref{ac}) with $\Phi(v)=-\gamma v$ for $\gamma=1$ and $\gamma=-1$ \blue{at different time}. Here, $N_1=40$, $N_2=10$ and $N_3=20$ \blue{ are used to compute the numerical results, which match well with the exact solution (\ref{eq_ou})}. }
\label{fig_ou}
\end{figure}
\subsection{Drifts admitting one jump}
Since drifts admitting only one jump are considered, the computational domain is only divided into two subdomains by the jump \red{here}. Then we can modify the proposed finite difference scheme just by removing the subdomain $\Omega_2$ as shown in Fig.~\ref{fig_grid}. Here the time step is chosen appropriately to be $\tau=0.01 \min\{h_1^2,h_2^2\}$.
\subsubsection{Pure dry friction}\label{sec_4bc}
When $\Phi(v)=-\mu\, \text{sgn}(v)$ \red{with $\mu$ being positive constant and ``sgn" denoting the sign function}, Eq.~(\ref{aa}) corresponds to the Brownian motion with pure dry friction \cite{Gennes2005dryfriction}, whose propagator is known in closed analytic form \cite{CaugheyDienes1961,Karatzas1984,TouchetteStraeten2010Brownian}
\begin{align}
p(v,t|v_0,0)=\tfrac{\mu}{D}\hat{p}\left( \tfrac{\mu}{D}v,\tfrac{\mu^2}{D}t\big|\tfrac{\mu}{D}v_0,0 \right),
\label{eq_dry_sol}
\end{align}
where
\begin{align*}
\hat{p}(x,\tau|x_0,0)=& \frac{e^{-\tau/4}}{2\sqrt{\pi \tau}}e^{-(|x|-|x_0|)/2}e^{-(x-x_0)^2/(4\tau)}\nonumber\\
& + \frac{e^{-|x|}}{4}\left[
1+\mathrm{erf}\left( \tfrac{\tau-(|x|+|x_0|)}{2\sqrt{\tau}} \right)
\right]
\end{align*}
with $
\mathrm{erf}(x)=2\int_0^x \exp(-z^2)dz/\sqrt{\pi}
$ denoting the error function. The exact solution at time $t=\tau_0$ is set to be the initial condition with $v_0=2$, the computational domain is chosen to be $[-4,8]$ and zero current conditions are set at the boundaries. Figure \ref{fig_dry} shows that the numerical results are consistent to the exact solutions. Since the solution admits a cusp at $v=0$ like peakons \cite{KalischRaynaud2006}, second-order accuracy is observed in Tab.~\ref{tab_dry}, as expected. \red{In addition, the results for a more general case with a two-valued drift are presented in Appendix \ref{app_sec2}}.
\begin{figure}
\includegraphics[width=0.9\linewidth]{dry_t1}\\\vspace{0.5em}
\includegraphics[width=0.9\linewidth]{dry_t5}
\caption{Propagators of \blue{Eq.~(\ref{aa}) with the drift $\Phi(v)= -\mathrm{sgn}(v)$} at $t=1$ and $t=5$. Here $v_0=2$ and $N_1=N_2=50$ \blue{are chosen to compute the numerical results, which
matches well with the exact solution (\ref{eq_dry_sol}).} }
\label{fig_dry}
\end{figure}
\begin{table}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}ccccccc}
\hline
$N_1$ &$N_2$ & $N_v$ & $L^2$ error & rate & $L^\infty$ error & rate \\
\hline
50 & 50 & 99 & 4.73E-03 &--& 2.80E-03&--
\\
100 & 100 & 199 & 1.21E-03 &1.96 &6.94E-04 &2.00
\\
200 & 200 & 399 & 3.01E-04& 2.00& 1.72E-04& 2.00
\\
400 & 400 & 799 & 7.57E-05 &1.99 &4.32E-05& 1.99
\\\hline
\end{tabular*}
\caption{Accuracy test for \blue{Eq.~(\ref{aa}) with the drift $\Phi(v)=-\mathrm{sgn}(v)$} at $t=1$. Here, $v_0=2$ and $N_v=N_1+N_2-1$. \blue{The errors are computed according to Eqs.~(\ref{eq_l2error}) and (\ref{eq_linfty}), and the rates are defined by Eq.~(\ref{eq_rate}).}}
\label{tab_dry}
\end{table}
\subsubsection{Other drifts admitting one jump}
Additional to the pure dry friction case, we consider here
another two drifts admitting one jump, which are studied by using exact simulations
of Eq.~(\ref{aa}) in \cite{Etore2014}
and \cite{papaspiliopoulos2016}, respectively.
The drift \red{studied in \cite{Etore2014}} is
\begin{equation}
\Phi(v)=
\begin{cases}
\frac{3\pi}{2}-\frac{\pi}{2}\cos\left(\frac{\pi}{5}v\right), & v<0, \\
-\frac{\pi}{2}\cos\left(\frac{\pi}{5}v\right), & v>0.
\end{cases}
\label{eq_drift_one_r1}
\end{equation}
We choose the computational domain to be $[-2,3]$ and impose
zero current conditions at the domain boundaries. The initial condition is set to be Eq.~(\ref{eq_v}) with $v_0=0$. However, \blue{in Eq.~(\ref{eq_drift_one_r1}) we did not define the value of $\Phi(0)$ since the proposed finite difference scheme does not involve the values of the drift at discontinuous points. Therefore, we have to define the value of $\Phi(0)$ involved in the initial condition (\ref{eq_v}). In this case, we define $\Phi(0)=[\Phi(0-)+\Phi(0+)]/2$. The same definition will be used throughout this paper.} The profile of the numerical propagator
as shown in Fig.~\ref{fig_one_jump_r1}(a) agrees with the result obtained in \cite{Etore2014} (see Fig.~4 therein). Moreover, the result obtained for a coarse grid matches with the fine-grid solution.
\begin{figure}
\includegraphics[width=0.9\linewidth]{one_jump_1}\vspace{0.5em}
\includegraphics[width=0.9\linewidth]{one_jump_2}
\caption{Propagators of Eq.~(\ref{aa}) with the drifts (\ref{eq_drift_one_r1}) \blue{ [(a)]
and (\ref{eq_drift_one_r2}) [(b)]} at $t=1$; $v_0=0$. Here the coarse-grid solutions obtained with $N_1=N_2=50$ match well with the fine-grid solutions obtained with $N_1=N_2=400$.}
\label{fig_one_jump_r1}
\end{figure}
The drift \red{investigated in \cite{papaspiliopoulos2016}} is
\begin{equation}
\Phi(v)=
\begin{cases}
\sin\left(v-\tfrac{\pi}{4}\right), & v<0, \\
\sin\left(v-\tfrac{7\pi}{6}\right), & v>0.
\end{cases}
\label{eq_drift_one_r2}
\end{equation}
The solution domain is truncated to be $[-5,4]$ and zero current conditions are imposed at the domain boundaries. As we can see from Fig.~\ref{fig_one_jump_r1}(b), the results agree with that obtained in \cite{papaspiliopoulos2016} (see Fig.~1 therein).
Moreover, the results obtained by using a coarse grid and a fine grid match with each other.
\subsection{Drifts admitting two jumps}
Here, the two examples presented in \cite{Dereudre2017} are considered. In both cases, the time step is chosen to be $\tau=0.01\min\{h_1^2,h_2^2,h_3^2\}$.
The first drift admitting two jumps reads as follows,
\begin{align}
\Phi(v)=\begin{cases}
0, & v<0,\\
1, & 0<v<1,\\
0, & v>1,
\end{cases} \label{eq_drift_two_1}
\end{align}
which is piecewise-constant. In this case, the computational domain is chosen to be $[-4,6]$ and no numerical boundary condition is needed for the proposed scheme. The numerical solutions as shown
in Fig.~\ref{fig_two_jumps_1}\blue{(a)} agree with the result presented in \cite{Dereudre2017} (see Fig.~6(a) therein). In addition, the results obtained by using a coarse grid and a fine grid are consistent.
\begin{figure}
\includegraphics[width=0.9\linewidth]{two_jumps_1}\\\vspace{0.5em}
\includegraphics[width=0.9\linewidth]{two_jumps_2}
\caption{Propagators of Eq.~(\ref{aa}) with the drift (\ref{eq_drift_two_1}) at time $t=1$ \blue{[(a)] and the drift (\ref{eq_drift_two_2}) at time $t=0.6$ [(b)]}. Here, $v_0=0.5$ is chosen for both cases. The coarse-grid solutions are obtained with $N_1=60$, $N_2=10$ and $N_3=50$, and the fine-grid solutions are obtained with $N_1=480$, $N_2=80$ and $N_3=400$. \blue{The coarse-grid solutions match well with the fine-grid solutions.} }
\label{fig_two_jumps_1}
\end{figure}
The second drift is
\begin{align}
\Phi(v)=\begin{cases}
-2\cos(v), & v<0,\\
\sin(v), & 0<v<1,\\
\cos(v-1)+\sin(1), & v>1.
\end{cases}\label{eq_drift_two_2}
\end{align}
The computational domain is chosen to be $[-4,5]$, which is large enough for us to impose
zero current conditions at the domain boundaries for time $t=0.6$ with $v_0=0.5$. As shown in Fig.~\blue{\ref{fig_two_jumps_1}(b)}, the profile of the propagator agrees with that presented in \cite{Dereudre2017} (see Fig.~6(b) therein). Again, it is observed that the results obtained by using a coarse grid and a fine grid are consistent.
\section{Extension to functionals}
\label{sec_5}
Functionals of a stochastic process have been investigated intensively in the past and have found numerous applications in physics. Here we consider
the functional
\begin{equation}
u(t)=\int_0^t K(v(s))ds
\end{equation}
with an integrable kernel $K(v)$, where the stochastic process $v(t)$ obeys
the Langevin equation (\ref{aa}). In particular, we have $u(0)=0$.
The joint propagator of $u$ and $v$, denoted by $p=p(u,v,t|v_0,0)$, is governed by the following Fokker-Planck equation
\begin{align}
\partial_t\,p
=-K(v)\partial_u\, p
-\partial_v[ \Phi(v) p ]
+\partial_v^2\, p
\label{eq_two}
\end{align}
with the initial condition
\begin{equation}
p(u,v,0|v_0,0)=\delta(u)\delta(v-v_0).
\end{equation}
\subsection{Scheme}
To solve Eq.~(\ref{eq_two}), we have to use the same matching conditions at the discontinuities of the drift $\Phi(v)$ (in the $v$ direction) as Eq.~(\ref{ac}), while in the $u$ direction we just need to use the continuous condition as usual.
Therefore, the scheme derived for Eq.~(\ref{ac}) can be applied directly for the $v$ direction.
In the $u$ direction, we choose the computational domain to be $[u_{_L},u_{_R}]$. Then a uniform staggered grid as shown in Fig.~\ref{fig_grid_x} is used. Here the flux points and solution points are
\begin{align}
&u_{k+1/2}= u_{_L}+k h_u, \quad 0\leqslant k \leqslant N_u, \label{eq_flux_u}\\
&u_{k}=u_{_L}+(k-1/2)h_u, \quad 1\leqslant k \leqslant N_u, \label{eq_sol_u}
\end{align}
where $N_u$ denotes the number of solution points in the $u$ direction and
the step $h_u=(u_{_R}-u_{_L})/N_u$.
\begin{figure}
\begin{center}
\includegraphics[width=0.98\linewidth]{grid_x}
\end{center}
\caption{ Illustration of the uniform grid \blue{staggered by flux points (\ref{eq_flux_u}) and solution points (\ref{eq_sol_u})} for the $u$ direction. }\label{fig_grid_x}
\end{figure}
Similar to
the approximation to the term $\Phi(v)p$ appearing in the current (\ref{ad}), we
approximate the term $K(v)\partial_u p$ at the point $(u_k,v_j)$ by
\[
\min\{ K(v_j),0 \}(\partial_u p)_{k,j}^+ +\max\{ K(v_j),0 \}(\partial_u p)_{k,j}^-,
\]
where $(\partial_u p)_{k,j}^+$ and $(\partial_u p)_{k,j}^-$ are obtained by the following procedure. For a fixed $v_j$, we first reconstruct the values of the propagator at $u_{k+1/2}$ from
the values at solution points by using fifth-order interpolations. Introducing the following two $(N_u+1)\times N_u$ interpolation matrices,
\begin{equation*}
I_u^-=
\begin{bmatrix}
(\mathbf{a})_{1\times N_u} \\[3pt]
(I_3^-)_{N_u\times N_u}
\end{bmatrix},\quad
I_u^+=
\begin{bmatrix}
(I_1^+)_{N_u\times N_u} \\[3pt]
\widetilde{\mathbf{a}}
\end{bmatrix},
\end{equation*}
where $\mathbf{a}$ is defined in Eq.~(\ref{eq_def_b}) and
$\widetilde{\mathbf{a}}$ is defined by letting its entries satisfy that
$
[\widetilde{\mathbf{a}}]_{1,j}=[\mathbf{a}_{1\times N_u}]_{1,N_u+1-j}
$, we can express the reconstructions as
$
\mathbf{p}_{\bullet j}^{\pm}=I_u^\pm \mathbf{p}_{\bullet j},
$
where \red{the vectors} $\mathbf{p}_{\bullet j}^\pm = [ p_{1/2,j}^\pm,p_{3/2,j}^\pm,\dots,p_{N_u+1/2,j}^\pm]^{T} $ and $\mathbf{p}_{\bullet j} = [ p_{1,j},p_{2,j},\dots,p_{N_u,j}]^{T} $.
Then we approximate the derivative $\partial_u p$ at solution points as
$
(\partial_u p)_{\bullet j}^\pm=A_u\mathbf{p}_{\bullet j}^{\pm}/h_u,
$
where the vectors $(\partial_u p)_{\bullet j}^\pm = [ (\partial_u p)_{1,j}^\pm,(\partial_u p)_{2,j}^\pm,\dots,(\partial_u p)_{N_u,j}^\pm ]^{T} $ and the $N_u\times (N_u+1)$ difference matrix $A_u$ is defined by
$
\blue{
A_u=(A_2)_{(N_u-1)\times N_u}
}
$
such that the difference scheme is sixth-order at $u_k$ with $3\leqslant k \leqslant N_u-2$ and fourth-order at the other solution points.
Here, the third-order Runge-Kutta scheme (\ref{rk_1}) is still used to solve the resulting ordinary differential system.
\subsection{Displacement}
Particularly, in this work we focus on the displacement $u(t)=\int_0^t v(s)ds$ associated with the Brownian motion with pure dry friction, i.e.,
$\Phi(v)=-\text{sgn}(v)$ and $K(v)=v$ are chosen for Eq.~(\ref{eq_two}). Here the
the computational domain in the $v$ direction is divided into two subdomains by $v=0$.
We set the initial condition to be
\begin{align}
p(u,v,\tau_0|v_0,0)=&\frac{1}{ 4\pi D \tau_0 }e^{ -(u-v_0 \tau_0)^2/(4D \tau_0) }\nonumber\\
&\times e^{ -(v-v_0-\Phi(v_0)\tau_0)^2/(4D \tau_0)}
\label{eq_uv}
\end{align}
and start the computations at $t=\tau_0$ \red{with} $\tau_0$ chosen to be $0.01$.
In addition, the time step is chosen to be $\tau=0.01\min\{h_u,h_1,h_2\}$.
As mentioned in the Introduction, for the Brownian motion with pure dry friction, analytic expressions of the first two moments of the displacement are available by solving a backward Komogorov equation \cite{ChenJust2014II} or using the method based on the Pugachev-Sveshnikov equation \cite{Berezin2018}. For instance,
when $v_0=0$ and $D=1$ we can inverse the expressions (70) and (73) in \cite{ChenJust2014II} to obtain the first two moments as
\begin{align}
M_1(t)=&0, \label{eq_mom1}\\
M_2(t)=& \left[\tfrac{1}{8} t^4+\tfrac{5}{6} t^3-2 t^2 +6 t-10\right]\text{erfc}\left(\tfrac{\sqrt{t}}{2}\right)\nonumber\\
& +\Big[64-\sqrt{\tfrac{t}{\pi}} \left(\tfrac{1}{4} t^3+\tfrac{7}{6} t^2- \tfrac{13}{3} t+10\right)\Big] e^{-t/4} \nonumber\\
& + 10t-54, \label{eq_mom2}
\end{align}
where $\text{erfc}(x)$ is the complementary error function.
For different time $t$, computational domain\red{s} can be chosen differently. For simplicity, the domain in the $v$ direction is fixed to be $[-6,6]$, while the domain for the $u$ direction is set to be dependent on time. To compute the results shown in Fig.~\ref{fig_diss}, the computational domain $u\in [-10,10]$ is chosen for $t=0.1$ and $t=1$, $u\in [-15,15]$ for $t=2.5$, and $u\in [-30,30]$ for $t=5$. While zero current conditions are set at the domain boundaries in the $v$ direction, no particular boundary condition is needed in the $u$ direction. \red{N}umerical evolution of the joined propagator is shown in Fig.~\ref{fig_diss}. In addition, the corresponding propagators of the displacement obtained by
\begin{align}
p_{\text{dis}}( u_k,t|0 )=& \sum_{j=1}^{N_1-1}p_{j,k} h_1 +\frac{1}{2}p_{N_1,k}(h_1+h_2)\nonumber\\
& +\sum_{j=N_1+1}^{N_1+N_2-1}p_{j,k} h_2
\label{eq_dis}
\end{align}
are illustrated in Fig.~\ref{fig_dis_x}. To confirm the
correctness of the results, the first two moments of the displacement are computed numerically by
\begin{align}
\widetilde{M}_s(t)=h_u\sum_{k=1}^{N_u} (u_k)^s p_{\text{dis}}( u_k,t|0 ),\quad s=1,2.
\label{eq_moments}
\end{align}
As shown in Fig.~\ref{fig_diss_mom}, the
results agree with the analytical expressions (\ref{eq_mom1}) and (\ref{eq_mom2}), indicating the validity of the numerical method.
\begin{figure*}
\begin{center}
\includegraphics[width=0.4\linewidth]{dis_t01}\quad
\includegraphics[width=0.4\linewidth]{dis_t1} \\\vspace{0.5em}
\includegraphics[width=0.4\linewidth]{dis_t2p5}\quad
\includegraphics[width=0.4\linewidth]{dis_t5}
\end{center}
\caption{Contour plots of the joint propagator $p(u,v,t|v_0,0)$ of Eq.~(\ref{eq_two}) at different time with $\Phi(v)=-\text{sgn}(v)$, $K(v)=v$, $D=1$ and $v_0=0$. Here $N_1=N_2=100$ and $N_u=200$ \blue{are chosen to compute the numerical results}.}
\label{fig_diss}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.4\linewidth]{dis_x_t01}\quad
\includegraphics[width=0.4\linewidth]{dis_x_t1}\\\vspace{0.5em}
\includegraphics[width=0.4\linewidth]{dis_x_t2p5}\quad
\includegraphics[width=0.4\linewidth]{dis_x_t5}
\end{center}
\caption{Propagators of the displacement (\ref{eq_dis}) correspond to the numerical joint propagator
$p(u,v,t|v_0,0)$ at different time, as shown in Fig.~\ref{fig_diss}.}
\label{fig_dis_x}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.4\linewidth]{mom_1}\quad
\includegraphics[width=0.4\linewidth]{mom_2}
\end{center}
\caption{The first two moments of the displacement. Lines correspond to the exact solutions (\ref{eq_mom1}) and (\ref{eq_mom2}), and points to the numerical solutions obtained by using Eq.~(\ref{eq_moments}) for the displacement distributions as shown in Fig.~\ref{fig_dis_x}.}
\label{fig_diss_mom}
\end{figure*}
\section{Conclusions}
\label{sec_6}
We have derived in this paper a finite difference scheme to solve Fokker-Planck equations with drift-admitting jumps. The scheme is based on a grid staggered by flux points and solution points. In particular, the positions of the jumps are set to be solution points and used to split the solution domain into subdomains, such that we do not have to do much work to deal with the matching conditions of the propagator and the probability current at the jumps. Some benchmark problems have been computed numerically to show the validity of the scheme. The results showed that the scheme is fifth-order for the cases with smooth drifts and second-order for the cases with discontinuous drifts.
One of the desirable properties of the scheme is that, depending on the signs of the drift $\Phi(v)$ at the domain boundaries, we may not need to specify boundary conditions for the proposed scheme and could use a small computational domain to get a correct solution. This property is in particular useful when we extend the scheme to study functionals of a process, where no boundary condition is needed at the domain boundaries of the functionals. The displacement statistics of the Brownian motion with pure dry friction have been computed to show the effectiveness of the extended scheme.
The proposed numerical approach may be generalized to solve other problems involving
discontinuous drifts, e.g., problems with both discontinuous drifts and some colored noises \cite{GeffertJust2017}, and high-dimensional problems with drift-admitting jumps \cite{DasPuri2017}.
\begin{acknowledgments}
This work was supported by the National Natural Science Foundation of China (Grant No. 11601517) and the Basic Research Foundation of National University of Defense
Technology (No. ZDYYJ-CYJ20140101).
\end{acknowledgments}
|
\chapter{Introduction}\label{cha:introduction}
\pagenumbering{arabic}
The goal of this dissertation is to present synthetic homotopy theory in the setting
of \emph{homotopy type theory}. We will present various results in this
framework, most notably the construction of the Atiyah-Hirzebruch and Serre
spectral sequences for cohomology, which have been fully formalized in the Lean
proof assistant.
Homotopy type theory, often abbreviated HoTT, is a version of type theory. Type
theory is a language for formal mathematics, in which every object has a
computational interpretation, so that it can also function as a programming
language. It can be used as a foundation of mathematics as an alternative to set
theory.
A key feature of HoTT is that the equality in a space corresponds to the path
spaces; a path between two points $a$ and $b$ is a proof that $a=b$. Two paths
that are not homotopic give different (unequal) proofs of this equality. The
fact that we identify proofs of an equality with a path means that every
construction in HoTT respects paths.
Many different researchers contributed to the homotopical interpretation of type theory.
Steve Awodey and Michael Warren gave a model of type theory in abstract homotopy theory~\cite{awodey2009homotopy}.
Benno van den Berg and Richard Garner published a paper addressing the coherence issue~\cite{berg2010models}.
Independently, Vladimir Voevodsky gave a model of type theory without identity types in simplicial sets
and formulated the \emph{univalence axiom}, which he proved consistent~\cite{voevodsky2006,voevodsky2009typesystems}.
The univalence axiom states that
homotopy equivalences between two types (spaces) corresponds to equality between
them~\cite{voevodsky2014univalence}. This means that every construction done in
HoTT automatically respects homotopy equivalence, which is a very convenient
property. Also, Voevodsky proved that a consequence of the univalence axiom is
\emph{function extensionality}. This states that two functions are equal when
they are homotopic.
The fact that all constructions are homotopy invariant also leads to some
challenges. It is not always clear whether we can define a concept of homotopy theory in
homotopy type theory. For example, \emph{singular homology} is a homotopy
invariant notion, but in the construction we use the set of all simplices in a
space, which is not a homotopy invariant notion. In this case, we can define
homology in a different way (see \autoref{sec:spectral-sequence-homology}).
However, for other definitions, such as the Grassmannian manifolds, it is an
open problem whether they can be constructed in homotopy type theory.
A new concept in homotopy type theory is the concept of \emph{higher inductive
types}. These are types that generalize both cell complexes in homotopy theory,
and inductively generated types (like $\textbf{N}$) in type theory. Higher inductive
types can be used to construct many spaces and operations on spaces often
encountered in homotopy theory.
Type theory is a convenient language for computer proof assistants. These are
programs that allow you to write formal proofs in a specified language, and
then the computer checks whether the proof is correct and complete. There are
many major results formalized in proof assistants, such as the four colour
theorem~\cite{gonthier2005fourcolour}, Feit-Thompson theorem~\cite{gonthier2013oddorder}
and the Kepler conjecture (Hales' Theorem)~\cite{hales2017kepler}.
HoTT is a type theory, and it has been implemented in various proof assistants,
such as Coq~\cite{bauer2016coqhott}, Agda~\cite{hottagda}, cubicaltt~\cite{cubicaltt}, Lean~\cite{vandoorn2017leanhott}
and various experimental proof assistants. One disadvantage of formally verifying
proofs in a proof assistant is that it takes a lot of work spelling out all
details. For example, doing very basic homotopy theory (not using homotopy type
theory) already takes a lot of effort~\cite{zhan2017auto2}. In HoTT this effect
is mitigated, because many homotopical concepts are close to the foundations of
the type theory, making formal proofs only a little more work than a paper proof.
Various results have been proven and formalized in HoTT, such as the the
Seifert--van Kampen theorem~\cite{favonia2016seifert}, the Blakers--Massey
theorem~\cite{favonia2016blakersmassey} and a development of cellular
cohomology~\cite{buchholtz2018cellular}. Another main result (which has not been
formalized) is the computation of $\pi_4(\S^3)$~\cite{brunerie2016spheres}, which
relies on conjectured properties of the smash product, which we will discuss in
\autoref{sec:smash-product}.
HoTT gives novel proof methods and new insights to homotopy theory. A basic
property of HoTT is \emph{path induction}, which states that when proving
something for a path with one free endpoint, one may assume that the path is the
constant path. This corresponds to the fact that the path space with one fixed
endpoint is contractible. Another technique is the encode-decode method, for
calculating the path space of certain spaces~\cite{licatashulman2013}. Moreover,
the proof of the Blakers--Massey theorem has been translated back to homotopy
theory, resulting in a new proof with novel ideas~\cite{rezk2014blakersmassey}.
Homotopy type theory has models in most model categories~\cite{awodey2009homotopy,berg2010models},
which are categorical models for homotopy theory. These models were
inspired by the groupoid model~\cite{hofmann1998groupoid}. Other models for HoTT
include the simplicial set model~\cite{voevodsky2009typesystems,kapulkin2012simplicialnew,streicher2014simplicial} and the cubical
set model~\cite{bezem2014cubicalsets,cohen2016cubical}. More generally, all Grothendieck
$(\infty,1)$-toposes model HoTT~\cite{cisinski2014models}.\footnote{General Grothendieck $(\infty,1)$-toposes model HoTT with universes \'a la Tarski.
This notion is weaker than universes \'a la Russell, which are usually considered in HoTT. We explain Russell universes in \autoref{sec:universes}.} Moreover, it is
conjectured that all elementary $(\infty,1)$-toposes form models of HoTT~\cite{shulman2017topos}.
\subsubsection*{Type Theory}
Homotopy type theory is based on Martin-L\"of type theory (also called intuitionistic type theory
or constructive type theory)~\cite{martinlof1975typetheory,martinlof1984typetheory}. In this type theory there are types, like the
integers $\ensuremath{\mathbb{Z}}\xspace$, vectors $\R^n$; and complex functions $\mathbb{C}\to\mathbb{C}$. There are also
terms, which have a unique type.\footnote{To be more precise: in many type theories there are terms with multiple types, for example due to universe cumulativity, but we will ignore these issues. Moreover, the type theory of Lean has unique typing~\cite{carneiro2018leantheory}.}
For example the number $-2$ has type $\ensuremath{\mathbb{Z}}\xspace$
(written as $-2:\ensuremath{\mathbb{Z}}\xspace$), the vector $(1,2,3,\ldots,n)$ has type $\R^n$ and we have the
exponential function $\exp:\mathbb{C}\to\mathbb{C}$. One can think of types as sets of objects
(and indeed, there is a model of type theory where the types are exactly sets),
but there are different interpretations, such as the types-as-spaces
interpretation that homotopy type theory provides. The fact that terms have a unique type means that the
$2:\ensuremath{\mathbb{Z}}\xspace$ and the $2:\R$ are different objects. It might be helpful to think of data types in a programming language, in which the \texttt{int} $2$ is stored differently in memory than the \texttt{float} $2$. Of course, the canonical inclusion
$i:\ensuremath{\mathbb{Z}}\xspace\hookrightarrow\R$ does satisfy $i(2)=2$. Type theory has a primitive
notion of computation, so that for example $2+3$ computes to $5$. Every function
that is explicitly defined in type theory therefore describes an algorithm
that can be executed. This means that type theory can be used as a programming
language, and many programming languages make use of a type system.
The congruence closure of this notion of computation is called
\emph{definitional equality} or \emph{judgmental equality},
and if two terms are judgmentally equal, one can replace one for the other in any term.
There are several methods to construct new types out of existing ones. For
example we can form the function type $A\to B$ for types $A$ and $B$, the
cartesian product type $A\times B$ and the coproduct or sum $A+B$. Propositions
can also be interpreted as types by the \emph{Curry-Howard isomorphism}~\cite{curry1958combinatorylogic,howard1980formulae}, and
under this interpretation $A\times B$ is the conjunction of $A$ and $B$, the sum
$A+B$ is the disjunction and $A\to B$ is the implication. Furthermore, there are
dependent function types $\prd{x:A}P(x)$ and dependent sum types $\sm{x:A}P(x)$,
which correspond to the universal quantification $\forall(x:A), P(x)$ and
existential quantification $\exists(x:A), P(x)$, respectively. So for example
the transitivity of $\le$ on $\textbf{N}$ can be expressed as $\prd{k,m,n:\textbf{N}}k \le m \to
m \le n \to k \le n$, and a term of this type is a proof that $\le$ is
transitive. The $P$ in $\prd{x:A}P(x)$ and $\sm{x:A}P(x)$ is called a
\emph{dependent type}, since it is a type depending on a term $x:A$. It has type
$P:A\to\ensuremath{\mathsf{Type}}\xspace$, where $\ensuremath{\mathsf{Type}}\xspace$ is the universe of (small) types. The dependent
function type $\prd{x:A}P(x)$ consists of functions $f$ that send terms $a:A$
to a term $f(a):P(a)$. Note that the type of $f(a)$ depends on the input $a$.
The dependent sum type $\sm{x:A}P(x)$ consists of dependent pairs $(a,x)$ with
$a:A$ and $x:P(a)$, where the type of $x$ depends on $a$.
Given two terms $a, b : A$, we can form the \emph{identity type} which we write as $a =_A b$ or $a=b$. As a proposition we view $a=_A b$ as the statement that $a$ and $b$ are
equal. In homotopy type theory these identity types correspond to the path space
of the type $A$.
\subsubsection*{Homotopy Type Theory}
There are various versions of dependent type theory with different rules for the
identity type. Some type theories have a \emph{reflection rule}, which states that
if we have a proof $p:a=b$, then $a$ and $b$ are judgmentally equal. Type theories with this rule are often called \emph{extensional}.
This is a convenient rule, but these type theories have meta-theoretic properties that are often seen as undesirable.
For example, checking whether a term $t$ has type $A$ is not decidable anymore.
Since this operation can be viewed as ``checking the correctness of a proof,'' one often wants to work in a type theory with decidable type-checking.
In \emph{intensional} type theory, without the reflection rule, multiple approaches can be taken for the identity type.
In some versions, there is a rule that any two proofs of the same
equality are themselves equal. This rule, often called \emph{uniqueness of
identity proofs} or \emph{axiom K} states that if $p,q : a = b$, then there is a
proof of $p = q$. In homotopy type theory, this rule is rejected. In the
types-as-spaces interpretation of homotopy type theory, terms of the identity
type $a =_A b$ are interpreted as paths in $A$ from $a$ to $b$.
We have familiar operations on paths: given two paths $p:a=_Ab$ and $q:b=_Ac$,
we write $p\cdot q:a=c$ for the concatenation of $p$ and $q$.
Furthermore, we have the inverse path $p^{-1}:b=a$ and the constant path $\ensuremath{\operatorname{\mathsf{refl}}}_a:a=a$.
We also have higher paths, the identity type $p =_{a=_Ab} q$ consists of homotopies from path $p$ to $q$. We can form
higher path types between two homotopies, and there are also operations on these higher paths.
In this way every type comes equipped with the structure of a higher groupoid.
In 2011, higher inductive types were introduced in homotopy type
theory~\cite{shulman2011intervalimpliesfunext,lumsdaine2011hits,shulman2011HoTThits,shulman2011pi1S1}. With
ordinary inductive types we specify constructors that generate the type, for
example the natural numbers are generated by zero $0:\textbf{N}$ and the successor
function $\mathsf{succ}:\textbf{N}\to\textbf{N}$. Higher inductive types are generated not only
by these ``point constructors'' but also by ``path constructors,'' which specify
the inhabitants of paths or higher paths in the type. For example, the circle
$\S^1$ is generated by a point $\star:\S^1$ and a loop $\ell:\star=\star$. The
rest of the structure of $\S^1$ is built from these constructors. Using higher
inductive types we can construct many other spaces in homotopy theory,
such as Eilenberg-MacLane spaces and homotopy pushouts.
As mentioned before in this introduction, we can use HoTT to do homotopy theory.
We think of types as spaces and we think of maps between types as continuous
maps between those spaces. Then we can define usual notions in homotopy theory,
as long as they are homotopy invariant: homotopy equivalences, suspensions,
spheres, etcetera. This is a \emph{synthetic} way to do homotopy theory: many
concepts, such as spaces and paths are uninterpreted constants of the type
theory. This is opposed to \emph{analytic} homotopy theory, where one studies
topological spaces up to homotopy equivalence. This distinction is similar to
the distinction for elementary geometry, which we can do synthetically (points
and lines are undefined concepts) or analytically (we are working in $\R^2$).
Synthetic geometry limits the things one can state or prove, but these proofs
are applicable in every model of the axioms. The same is true for synthetic
homotopy theory: the proofs performed synthetically are true in all models of
HoTT.
In this dissertation I will not be very precise about the exact rules of the type theory we are using. We will
present the constructions and proofs in such a way that they can be performed in the ``HoTT
book''~\cite{hottbook}. Most of the results in this dissertation have been
formalized in the Lean proof assistant~\cite{moura2015lean}. The HoTT mode we
used in Lean has very similar rules to the HoTT book, and the differences are
not relevant for the constructions in this dissertation. A concept closely related to
homotopy type theory is \emph{univalent mathematics}, a term coined by Vladimir
Voevodsky for the development of mathematics where one takes homotopy types as
primitive objects, and reasons about them using type-theoretic reasoning and the
univalence axiom. This is pursued in the proof assistant UniMath~\cite{unimath}. There are also
radically different type theories which are studied in homotopy type
theory. These are called ``cubical type theories'' because they all have a
primitive notion of cubes. Examples include the cubical type theory described in~\cite{cohen2016cubical},
which was implemented in the proof assistant cubicaltt~\cite{cubicaltt},
and computational higher-dimensional type theory~\cite{angiuli2017computational},
on which the proof assistant RedPRL is based~\cite{redprl}.
These type theories are extensions of the type theory presented in
the HoTT book, which we will call book-HoTT. In book-HoTT the univalence axiom
is an axiom: an uninterpreted constant of a certain type. This breaks the
computational behavior of the type theory. For example not every closed term of
type $\textbf{N}$ computes to either $0$ or the successor of another number. These
cubical type theories add primitive concepts to the theory to make the
univalence axiom provable, and therefore all terms in these system do compute.
We will often want to compare homotopy type theory with ordinary homotopy
theory. We will use the adverb ``classically'' to refer to the concepts and
theorems in homotopy theory that do not involve HoTT.\footnote{This use of classically has nothing to do with the word classical in ``classical logic,'' involving the law of excluded middle or the axiom of choice. In homotopy type theory one can consistently assume the law of excluded middle or the axiom of choice, formulated in a precise way so that it corresponds to what it usually means. However, doing so removes the computational content of all notions defined using it.}
Conversely, we will say that something is provable in HoTT if we can prove it in book-HoTT.
\subsubsection*{Contents}
In \autoref{cha:preliminaries} we review the basic concepts in homotopy type
theory. For a more detailed and thorough exposition, we refer to~\cite{hottbook}.
Alternative introductions can be found in~\cite{favonia2017thesis} and~\cite{brunerie2016spheres}. In
\autoref{sec:martin-lof-type} we introduce the basic concepts of type theory:
functions, pairs, universes, and inductive types such as the identity type. In
\autoref{sec:homotopy-type-theory} we will introduce the basics of homotopy type
theory. In particular we will formally state the univalence axiom and present
higher inductive types. In \autoref{sec:lean} we will discuss the Lean proof
assistant in more detail.
In \autoref{cha:high-induct-types} we will study higher inductive types
internally in HoTT. The main problem we will focus on is the interdefinability
of higher inductive types. In particular, we try to construct various higher
inductive types from the homotopy pushout. We will define the propositional
truncation in \autoref{sec:prop-trunc}, nonrecursive higher inductive types with
2-path constructors in \autoref{sec:non-recursive-2} and work towards defining
certain \emph{localizations} in \autoref{sec:colimits}.
In \autoref{cha:homotopy-theory} we present some synthetic homotopy theory in
HoTT. In \autoref{sec:computing-pi3s2} we will describe the formalization of the
long exact sequence of homotopy groups and its application to compute
$\pi_3(\S^2)$. Although this construction has been described before in HoTT
in~\cite[Section 8.4]{hottbook} and~\cite[Section 2.5.1]{brunerie2016spheres},
no formally verified proof has been given before. In
\autoref{sec:eilenb-macl-spac} we will study Eilenberg-MacLane spaces, which are
spaces with only one nontrivial homotopy group. Eilenberg-MacLane spaces have
been defined in HoTT before~\cite{licata2014em}. Here we prove the (classically
known) results that Eilenberg-MacLane spaces are unique, and give an equivalence
of categories between the category of (abelian) groups and an appropriate class of pointed
types. In \autoref{sec:smash-product} we will discuss the smash product. The
ultimate goal is to prove that the smash product forms a 1-coherent symmetric monoidal
product on pointed types, and we will give one approach towards proving this
using a Yoneda-style argument.
In \autoref{cha:serre-spectr-sequ} we develop the theory of spectral sequences
in HoTT. We give the construction of a spectral sequence from an exact couple
(in \autoref{sec:exact-couples}) and show how to construct an exact couple from
a tower of spectra (in \autoref{sec:spectra}). We construct the
classically-known Atiyah-Hirzebruch and Serre spectral sequences for cohomology
(in \autoref{sec:spectral-sequence-cohomology}), and give some ideas towards
doing the same for their counterparts in homology (in
\autoref{sec:spectral-sequence-homology}).
\chapter{Preliminaries}\label{cha:preliminaries}
In this chapter we will give a brief overview of type theory and homotopy type
theory. We cannot cover all the subtleties, so readers new to (homotopy) type
theory should consult the homotopy type theory book~\cite{hottbook}.
In \autoref{sec:lean} we will discuss the proof assistant \emph{Lean}. All main
results in this dissertation have been formalized in Lean.
\section{Martin-L\"of Type Theory}\label{sec:martin-lof-type}
As mentioned in the introduction, homotopy type theory is based on a system called \emph{Martin-L\"of type theory} or \emph{intuitionistic type theory}.
There are types and there are terms, which have a unique type. There is a notion of computation. Two terms $t$ and $s$ are considered \emph{judgmentally equal} or \emph{definitionally equal},
denoted $t \equiv s$ if $t$ and $s$ compute to the same term.
We are working in dependent type theory, which means that types can depend on terms.
For example, there is a type of vectors of length $n:\textbf{N}$ in type $A$, denoted $\ensuremath{\operatorname{\mathsf{vector}}}_A(n)$.
In this case $\ensuremath{\operatorname{\mathsf{vector}}}_A$ is a dependent type over $\textbf{N}$.
An example term in this type family is $(5,6,7,8) : \ensuremath{\operatorname{\mathsf{vector}}}_\textbf{N}(4)$.
When we say that a term has a unique type, we mean that it has a unique type up to definitional equality.
In our example, we also have that $(5,6,7,8) : \ensuremath{\operatorname{\mathsf{vector}}}_\textbf{N}(2+2)$, because $2+2\equiv 4$.
More generally, if we have two definitionally equal types $A\equiv B$ and if $t : A$, then $t : B$.
Logically (under the types-as-propositions interpretation) dependent types are predicates.
We will explain the topological interpretation of dependent types at the end of \autoref{sec:pair-types}.
In the remainder of this section we will discuss the type formers of Martin-L\"of type theory more closely.
\subsection{Function Types}\label{sec:function-types}
Given a type $A$ and a family of types $B$ depending on $A$, we can form the \emph{dependent function type} (also called \emph{product type} or \emph{pi type}) $$(x : A) \to B(x)\qquad\text{or}\qquad\prd{x:A}B(x).$$
We will use the former notation in this document. A term $f:(x : A)\to B(x)$ is a function that sends each element $a : A$ to an element\footnote{Formally, $B(a)$ is the term $B(x)$ where we substitute $a$ for $x$. In \autoref{sec:universes} we will see that we can treat $B$ as a function into a universe, and that alternatively we can view $B(x)$ and $B(a)$ as function applications.} $f(a):B(a)$. We also use the notation $fa$ or $f\ a$ for $f(a)$. Note that the type of $f(x)$ depends on $x$. We can form functions using \emph{lambda-abstraction}. Given a term $t(x) : B(x)$, we can form the term $\lam{x}t(x): (x : A) \to B(x)$, which is the function $x\mapsto t(x)$, i.e. the function that sends $x$ to $t(x)$. We get the computation rule
$$(\lam{x}t(x))a\equiv t(a)$$
for $a : A$, which is called the \emph{beta-rule} or \emph{beta-reduction}. We also have an \emph{eta-rule}, which states that every function is a lambda abstraction. This means that for $f : (x : A) \to B(x)$ we have
$$f\equiv \lam{x}f(x).$$
We will often define functions by writing $f(x)\ensuremath{\vcentcolon\equiv} t$ (where $x$ may occur in $t$), which formally means that we define $f$ as $\lam{x}t$.
An important special case occurs when $B$ does not depend on $A$. In this case the dependent function type $(x : A) \to B$ is written as $A \to B$, which is the type of functions from type $A$ to type $B$.
Logically, the type $A \to B$ is interpreted as the implication $A \Rightarrow B$ and the type $(x : A) \to B(x)$ is interpreted as the universal quantification $\forall(x : A), B(x)$. Topologically, a function $f : A \to B$ corresponds to a continuous map from $A$ to $B$. The type $A \to B$ is the mapping space from $A$ to $B$. We will explain the topological interpretation of $(x : A) \to B(x)$ at the end of \autoref{sec:pair-types}.
We can define the identity function
$$\idfunc\equiv\idfunc[A]\ensuremath{\vcentcolon\equiv} \lam{x:A}x:A\to A$$
and the composition of functions: if $f : A \to B$ and $g : B \to C$, then $g \o f\ensuremath{\vcentcolon\equiv} \lam{x}g(f(x)):A\to C$. Given $b:B$, we also have a constant function $\ensuremath{\mathbf{0}}_b\ensuremath{\vcentcolon\equiv} \lam{x}b:A \to B$.
We will often write some arguments of a function implicitly. Such arguments are written with curly braces in the type. For example, given a dependent type $C$ over $\textbf{N}$, we write
$$g : \{n : \textbf{N}\} \to C(n) \to C(n+1)$$
to emphasize that the first argument of $g$ is implicit. In this case, for $c : C(n)$ we will write $g(c)$ for $g$ applied (implicitly) to $n$ and applied to $c$. The curly braces are only to indicate how we write function application for functions with this type, for all other purposes the types $\{x : A\} \to B(x)$ and $(x : A) \to B(x)$ are the same.
\subsection{Pair Types}\label{sec:pair-types}
Given a type family $B$ depending on a type $A$, we can form the \emph{dependent pair type} (also called \emph{dependent sum type} or \emph{sigma type})
$$(x : A) \times B(x)\qquad\text{or}\qquad\sm{x:A}B(x).$$
We will use the former notation in this document. A term of type $(x : A)\times B(x)$ is a pair consisting of an element $a : A$ and an element $b : B(a)$. Given $a : A$ and $b : B(a)$, we can form the term $(a,b):(x : A)\times B(x)$, and we have projections
$$p_1:(x : A)\times B(x)\to A\quad \text{ and }\quad p_2:(z : (x : A)\times B(x))\to B(p_1(z)).$$
We will sometimes write $x.i$ for $p_i(x)$. There are beta rules $p_1(a,b)\equiv a$ and $p_2(a,b)\equiv b$ and an eta rule stating that for any $z:(x : A)\times B(x)$ we have $z\equiv (p_1z, p_2z)$. In Lean, there is no eta rule for dependent pair types, but instead there is an induction principle, similar to those of inductive types (see \autoref{sec:inductive-types}).
If $B(x)$ does not depend on $x$, we write $(x : A) \times B$ simply as $A\times B$. In this case we retrieve the usual cartesian product of $A$ and $B$.
Logically we can think of $A\times B$ as the conjunction of $A$ and $B$, as
described above. Furthermore, we can think of $(x : A)\times B(x)$ as a
proof-relevant version of the existential quantifier $\exists(x : A). B(x)$. It
is proof-relevant in the sense that from a proof of $(x : A)\times B(x)$ we can
extract a witness $a : A$ such that $B(a)$ holds. In \autoref{sec:truncatedness}
we will define an existential quantifier from which the witness cannot be
extracted.
Topologically, we think of $A\times B$ as the product space of $A$ and $B$.
The map $p_1 : (x : A) \times B(x) \to A$ corresponds to a \emph{fibration}. A fibration is a map that has the homotopy lifting property with respect to any space, which is given by transport, to be defined in \autoref{sec:paths}. Under this interpretation, $(x : A) \times B(x)$ is the total space of the fibration $p_1$, and $B(a)$ is the \emph{fiber} of $p_1$ at point $a$. The type $(x : A)\to B(x)$ is the type of sections of $p_1$. These observations are usually summarized as ``dependent types correspond to fibrations.'' We will often call dependent functions \emph{sections}.
\subsection{Universes}\label{sec:universes}
In our discussions below we need one or more \emph{universes} in our type theory. There are different styles of universes in type theory~\cite{martinlof1984typetheory}, we will describe the universes \'a la Russell. A universe $\ensuremath{\mathsf{Type}}\xspace$ is a type that has types as its terms. That is to say, if $A:\ensuremath{\mathsf{Type}}\xspace$, then $A$ is a type. It is closed under all type-forming operations. For example, for pi-types this means that if $A : \ensuremath{\mathsf{Type}}\xspace$ and for $a:A$ we have $B(a):\ensuremath{\mathsf{Type}}\xspace$, then
$$(a : A) \to B(a):\ensuremath{\mathsf{Type}}\xspace.$$
We can now interpret dependent types in $\ensuremath{\mathsf{Type}}\xspace$, such as $B$ above, as functions $B:A\to\ensuremath{\mathsf{Type}}\xspace$.
In the proofs in this document we can often get away with assuming only a single universe. However, it is useful to have the property that all types have a type themselves, and we cannot do that with a single universe $\ensuremath{\mathsf{Type}}\xspace$, because positing $\ensuremath{\mathsf{Type}}\xspace:\ensuremath{\mathsf{Type}}\xspace$ is inconsistent~\cite{girard1972paradox}. Instead, we will assume that we have a tower of universes $$\ensuremath{\mathsf{Type}}\xspace_0:\ensuremath{\mathsf{Type}}\xspace_1:\ensuremath{\mathsf{Type}}\xspace_2:\cdots$$
such that for every type $A$ there is an $i$ such that $A:\ensuremath{\mathsf{Type}}\xspace_i$. In this case every dependent type can be interpreted as a function $A\to\ensuremath{\mathsf{Type}}\xspace_i$ for some $i$. As is customary, we usually omit writing universe levels explicitly, and we will perform constructions polymorphic over all universes. For example, if we write $$\idfunc : \{A : \ensuremath{\mathsf{Type}}\xspace\} \to A \to A,$$ we really mean that for any universe level $i$ we have $$\idfunc[i] : \{A : \ensuremath{\mathsf{Type}}\xspace_i\} \to A \to A.$$
One rule that is sometimes assumed is \emph{universe cumulativity}, which states that if $A : \ensuremath{\mathsf{Type}}\xspace_i$, then $A : \ensuremath{\mathsf{Type}}\xspace_j$ for $j\geq i$.
This can be problematic, and lead to violation of nice properties of the type theory, such as subject reduction or canonicity~\cite{luo2012universes}.
In this document (and in Lean), we do not assume universe cumulativity.
Instead, using inductive types (see \autoref{sec:inductive-types}) we can construct for $A : \ensuremath{\mathsf{Type}}\xspace_i$ a new type $\ensuremath{\operatorname{\mathsf{lift}}} A:\ensuremath{\mathsf{Type}}\xspace_j$ for $j\geq i$ such that $A\simeq \ensuremath{\operatorname{\mathsf{lift}}} A$.
\subsection{Inductive Types}\label{sec:inductive-types}
\emph{Inductive types} are types that are inductively generated by some
\emph{constructors}. A simple example is $\textbf{N}$, which is inductively generated by
$0$ and the successor function $S\ensuremath{\vcentcolon\equiv}\lam{x}x+1$. In this section we will
discuss some inductive types that we will need in this dissertation. We will
talk about the empty type, the unit type, the booleans, the natural numbers and
the sum type. The dependent pair type (\autoref{sec:pair-types}) is also an inductive type.
\subsubsection*{The empty type}
The empty type $\ensuremath{\mathbf{0}}\xspace:\ensuremath{\mathsf{Type}}\xspace_0$ is a type without inhabitants. There are no constructors, and we have as induction principle that if $P:\ensuremath{\mathbf{0}}\xspace\to\U_i$, then $$\ind{\ensuremath{\mathbf{0}}\xspace}:(x : \ensuremath{\mathbf{0}}\xspace) \to P(x).$$ This conveys that $\ensuremath{\mathbf{0}}\xspace$ indeed has no inhabitants, because if we view $P$ as a predicate, we can prove anything about all inhabitants of $\ensuremath{\mathbf{0}}\xspace$. We can define negation $\neg A\ensuremath{\vcentcolon\equiv} A \to \ensuremath{\mathbf{0}}\xspace$.
\subsubsection*{The unit type}
The unit type $\ensuremath{\mathbf{1}}\xspace:\ensuremath{\mathsf{Type}}\xspace_0$ is a type with exactly one inhabitant $\star:\ensuremath{\mathbf{1}}\xspace$. The induction principle states that if $P:\ensuremath{\mathbf{1}}\xspace\to\U_i$, then $$\ind{\ensuremath{\mathbf{1}}\xspace}:P(\star) \to (x : \ensuremath{\mathbf{1}}\xspace) \to P(x).$$ This states that $\star$ is the only inhabitant of $\ensuremath{\mathbf{1}}\xspace$, because if we can prove something for $\star$, then it holds for all inhabitants of $\ensuremath{\mathbf{1}}\xspace$. There is a computation rule $$\ind{\ensuremath{\mathbf{1}}\xspace}(p,\star)\equiv p.$$
\subsubsection*{The booleans}
The type of booleans $\ensuremath{\mathbf{2}}\xspace:\ensuremath{\mathsf{Type}}\xspace_0$ has exactly two inhabitants ${1_{\bool}},{0_{\bool}}:\ensuremath{\mathbf{2}}\xspace$. Its induction principle states that if $P:\ensuremath{\mathbf{2}}\xspace\to\U_i$, then $$\ind{\ensuremath{\mathbf{2}}\xspace}:P({1_{\bool}}) \to P({0_{\bool}}) \to (x : \ensuremath{\mathbf{2}}\xspace) \to P(x).$$
The computation rules are
$$\ind{\ensuremath{\mathbf{2}}\xspace}(p_{1_{\bool}},p_{0_{\bool}},{1_{\bool}})\equiv p_{1_{\bool}}\qquad\text{and}\qquad
\ind{\ensuremath{\mathbf{2}}\xspace}(p_{1_{\bool}},p_{0_{\bool}},{0_{\bool}})\equiv p_{0_{\bool}}.$$
\subsubsection*{The natural numbers}
A more interesting type is the type of natural numbers $\textbf{N}:\ensuremath{\mathsf{Type}}\xspace_0$. It has a constructor $0:\textbf{N}$ and a unary constructor $S:\textbf{N} \to \textbf{N}$, and it is freely generated by these constructors. This means that if $P:\textbf{N}\to\U_i$ and if we have $p_0:P(0)$ and $p_S : (k : \textbf{N}) \to P(k) \to P(S\ k)$, then $$\ind{\textbf{N}}(p_0,p_S):(n : \textbf{N}) \to P(n).$$
If we view $P$ as a predicate, this is the usual induction principle for $\textbf{N}$: to prove something for all numbers we need to prove it for $0$ and we need to prove it for $k+1$ assuming it holds for $k$, for an arbitrary $k$. However, this induction principle also allows us to define (dependent) functions from $\textbf{N}$. These functions satisfy the computation rules
$$\ind{\nat}(p_0,p_S,0)\equiv p_0\qquad\text{and}\qquad
\ind{\nat}(p_0,p_S,S\ n)\equiv p_S(n,\ind{\nat}(p_0,p_S,n)).$$
Often, we will want to give a name $f$ to $\ind{\nat}(p_0,p_S)$, and we will instead denote the recursive definition of $f$ using pattern matching notation:
$$f(0)\ensuremath{\vcentcolon\equiv} p_0\qquad\text{and}\qquad f(S\ n)\ensuremath{\vcentcolon\equiv} p_S(n,f(n)).$$
For example, we can define addition and multiplication $+,\cdot:\textbf{N}\to\textbf{N}\to\textbf{N}$ recursively (in the second argument) as
\begin{align*}
n+0&\ensuremath{\vcentcolon\equiv} n & n\cdot 0&\ensuremath{\vcentcolon\equiv} 0\\
n+(S\ m)&\ensuremath{\vcentcolon\equiv} S(n + m) & n \cdot (S\ m)&\ensuremath{\vcentcolon\equiv} n \cdot m + n.
\end{align*}
Note that $n+1\equiv S\ n$, and we will often write $n+1$ instead of $S\ n$ from now on.
\subsubsection*{The sum type}
Given two types $A$ and $B$, we can form the \emph{sum type} or \emph{coproduct} $A+B$ with constructors $\ensuremath\inlsym\xspace:A\to A+B$ and $\ensuremath\inrsym\xspace : B \to A + B$.
The induction principle states that for $P:A+B\to\ensuremath{\mathsf{Type}}\xspace$ with maps $p_{\ensuremath\inlsym\xspace} : (a : A) \to P(\ensuremath\inlsym\xspace a)$ and $p_{\ensuremath\inrsym\xspace} : (b : B) \to P(\ensuremath\inrsym\xspace b)$ we get a section
$$\ind{{+}}(p_{\ensuremath\inlsym\xspace},p_{\ensuremath\inrsym\xspace}):(x : A + B) \to P(x)$$
with computation rules
$$\ind{{+}}(p_{\ensuremath\inlsym\xspace},p_{\ensuremath\inrsym\xspace},\ensuremath\inlsym\xspace(a))\equiv p_{\ensuremath\inlsym\xspace}(a)\qquad\text{and}\qquad
\ind{{+}}(p_{\ensuremath\inlsym\xspace},p_{\ensuremath\inrsym\xspace},\ensuremath\inrsym\xspace(b))\equiv p_{\ensuremath\inrsym\xspace}(b)$$
Logically, the type $A+B$ is the proof-relevant disjunction of $A$ and $B$. It is
proof-relevant in the sense that a proof of $A + B$ is of the form
$\ensuremath\inlsym\xspace a$ or $\ensuremath\inrsym\xspace b$. Therefore, a proof comes with a proof of either $A$ or $B$. In
\autoref{sec:truncatedness} we will see a disjunction that does not have this
property.
\subsubsection*{General Inductive Types}
In \autoref{sec:inductive-types} we saw various instances of inductive types.
Also the sigma-types from \autoref{sec:pair-types} (without eta rule) are an instance of an
inductive type. We will now explain inductive types and families of inductive
types in general. For a more detailed description, see~\cite[Section ``Inductive Types'']{carneiro2018leantheory}.
When defining an inductive type, we have to list its constructors. For example, we could define the sum type as follows. Given $A\ B : \ensuremath{\mathsf{Type}}\xspace$, we define
\begin{inductive}
\texttt{inductive} $A + B : \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \ensuremath\inlsym\xspace : A \to A + B;$\\
$\bullet\ \ensuremath\inrsym\xspace : B \to A + B.$
\end{inductive}
This defines the type $A+B$ with constructors $\ensuremath\inlsym\xspace$ and $\ensuremath\inrsym\xspace$ of the specified
type. Each constructor must have as target the inductive type currently being
defined (in this case $A+B$).\footnote{For \emph{higher inductive types}
(\autoref{sec:high-induct-types}) the conclusion can also be a (higher) path in
the type currently being defined.} Constructors can be \emph{recursive},
meaning that the type being defined can occur in the domain of a constructor.
For example, here is the type of $\omega$-branching trees with leaves labeled by
a type $C$.
\begin{inductive}
\texttt{inductive} $\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C : \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \ensuremath{\operatorname{\mathsf{leaf}}} : C\to\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C;$\\
$\bullet\ \ensuremath{\operatorname{\mathsf{node}}} : (\textbf{N} \to \ensuremath{\operatorname{\mathsf{\omega-tree}}}_C)\to\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C.$
\end{inductive}
A restriction on recursive constructors is that the inductive type being defined can only occur in strictly positive positions, that is as the target of one of the arguments of the constructor.
Every inductive type has an induction principle. We can algorithmically find the type of the induction principle from the constructors. The first argument of the induction principle (often left implicit) is the \emph{motive}, which is an arbitrary type family over the inductive type being defined, for $\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C$ this has type $P:\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C\to\ensuremath{\mathsf{Type}}\xspace$. Then for every constructor $c$ there is an argument that mimics the type of $c$ and has as target $P(c(\cdots))$.
For $\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C$ these arguments have type $p_{\ensuremath{\operatorname{\mathsf{leaf}}}}:(c:C)\to P(\ensuremath{\operatorname{\mathsf{leaf}}} c)$ and
$$p_{\ensuremath{\operatorname{\mathsf{node}}}}:(f:\textbf{N}\to\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C)\to ((n : \textbf{N}) \to P(f\ n))\to P(\ensuremath{\operatorname{\mathsf{node}}} f).$$
Note that for each recursive argument $f$ of the constructor we assume an \emph{induction hypothesis} of type $P(f(\cdots))$. The induction principle then gives a section of $P$. So for example we get
$$\ind{\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C}(p_{\ensuremath{\operatorname{\mathsf{leaf}}}},p_{\ensuremath{\operatorname{\mathsf{node}}}}):(x : \ensuremath{\operatorname{\mathsf{\omega-tree}}}_C) \to P(x).$$
Finally, the computation rules states that if the induction principle acts on a constructor, then it will reduce to the argument corresponding to that constructor. For $\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C$ this means (abbreviating $s\ensuremath{\vcentcolon\equiv} \ind{\ensuremath{\operatorname{\mathsf{\omega-tree}}}_C}(p_{\ensuremath{\operatorname{\mathsf{leaf}}}},p_{\ensuremath{\operatorname{\mathsf{node}}}})$)
$$s(\ensuremath{\operatorname{\mathsf{leaf}}}(c))\equiv p_{\ensuremath{\operatorname{\mathsf{leaf}}}}(c)\qquad\text{and}\qquad
s(\ensuremath{\operatorname{\mathsf{node}}}(f))\equiv p_{\ensuremath{\operatorname{\mathsf{node}}}}(f,\lam{n}s(f\ n)),$$
where applying $s$ to the recursive constructor leads to a recursive call of $s$.
One important generalization of inductive types are families of inductive types. In this case, a family of types $P$ is being defined simultaneously indexed over some type $I$. In this case, constructors must have as target $P(t)$ where $t$ is a term of type $I$ formed by the (nonrecursive) arguments of the constructor. An example of an inductive family of types is the type of vectors in $A$ of some length $n:\textbf{N}$.
\begin{inductive}
\texttt{inductive} $\ensuremath{\operatorname{\mathsf{vector}}}_A : \textbf{N} \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \mathsf{nil} : \ensuremath{\operatorname{\mathsf{vector}}}_A(0);$\\
$\bullet\ \mathsf{cons} : \{n : \textbf{N}\} \to A \to \ensuremath{\operatorname{\mathsf{vector}}}_A(n) \to \ensuremath{\operatorname{\mathsf{vector}}}_A(n+1).$
\end{inductive}
Note that the parameter $A$ remains fixed in the definition of $\ensuremath{\operatorname{\mathsf{vector}}}_A(n)$, while the index $n:\textbf{N}$ is not: the constructor $\mathsf{cons}$ constructs a vector of length $n+1$ from a vector of length $n$.
The induction principle can again be extracted algorithmically. It is important that the motive also quantifies over all indices of the inductive family. For vectors it states that given a motive
$$P:\{n : \textbf{N}\} \to \ensuremath{\operatorname{\mathsf{vector}}}_A(n)\to \ensuremath{\mathsf{Type}}\xspace$$
and induction steps
\begin{align*}
p_{\mathsf{nil}}&:P(\mathsf{nil})\\
p_{\mathsf{cons}}&:(n : \textbf{N}) \to (a : A) \to (x : \ensuremath{\operatorname{\mathsf{vector}}}_A(n))\to P(x) \to P(\mathsf{cons}(a,x)),
\end{align*}
we get a section
$$\ind{\ensuremath{\operatorname{\mathsf{vector}}}}(p_{\mathsf{nil}},p_{\mathsf{cons}}):\{n : \textbf{N}\} \to (x : \ensuremath{\operatorname{\mathsf{vector}}}_A(n)) \to P(x)$$
with the expected computation rules.
A very important inductive family of types is the \emph{identity type}.\footnote{also called \emph{path type}, \emph{identification type} or \emph{equality type}.} This is a family of types with parameters $A:\ensuremath{\mathsf{Type}}\xspace$ and $a:A$ and is defined as
\begin{inductive}
\texttt{inductive} $\ensuremath{\operatorname{\mathsf{Id}}}_A(a,{-}) : A \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \ensuremath{\operatorname{\mathsf{refl}}}_a : \ensuremath{\operatorname{\mathsf{Id}}}_A(a,a).$
\end{inductive}
We also denote the type $\ensuremath{\operatorname{\mathsf{Id}}}_A(a_1,a_2)$ by $a_1=_A a_2$ or $a_1=a_2$ and $\ensuremath{\operatorname{\mathsf{refl}}}_a$ by $\ensuremath{\operatorname{\mathsf{refl}}}$, $1_a$ or $1$. Its induction principle states that for a family $P:(a' : A) \to a = a' \to \ensuremath{\mathsf{Type}}\xspace$ and a term $p_{\ensuremath{\operatorname{\mathsf{refl}}}}: P(a,1_a)$ we find a section
$$\ind{=}(p_{\ensuremath{\operatorname{\mathsf{refl}}}}):(a' : A) \to (p : a = a') \to P(a',p).$$
In words: we may assume that a path with free right endpoint (that is, the right hand side of the equality is a variable) is reflexivity.
Logically, the identity type corresponds to equality. Under this interpretation, a term of type $a_1=a_2$ is a proof that $a_1$ and $a_2$ are equal. Homotopically, the identity type corresponds to the path space of $A$, and we will explore this interpretation more in \autoref{sec:paths}.
\section{Homotopy Type Theory}\label{sec:homotopy-type-theory}
We will now discuss in more detail the homotopical interpretation of types, and the basic concepts of homotopy type theory.
\subsection{Paths}\label{sec:paths}
Elements of an identity type form paths in the space. We can define the usual operations on paths.
Given a path $p:a=_Ab$, we can define the \emph{inverse} $p^{-1}:b=_Aa$. We can do this by path induction. Define the family $$P\ensuremath{\vcentcolon\equiv} \lam{x:A}{q:a=_Ax}x=_Aa:(x:A)\to a=_Ax \to \ensuremath{\mathsf{Type}}\xspace.$$
We now have $\ensuremath{\operatorname{\mathsf{refl}}}_a:P(a,p)\equiv a =_A a$, and therefore we get
$$p^{-1}\ensuremath{\vcentcolon\equiv}\ind{=}(\ensuremath{\operatorname{\mathsf{refl}}}_a,b,p):b=_Aa.$$
The computation rule gives that $\ensuremath{\operatorname{\mathsf{refl}}}_a^{-1}\equiv \ensuremath{\operatorname{\mathsf{refl}}}_a$.
We can explain the proof in words more intuitively. Path induction states that we may assume that a path with a free endpoint is reflexivity. Since $p$ has a free endpoint ($b$ is a variable), we may assume that $b\equiv a$ and $p\equiv\ensuremath{\operatorname{\mathsf{refl}}}_a$. In this case, we can define $$p^{-1}\equiv \ensuremath{\operatorname{\mathsf{refl}}}_a^{-1}\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{refl}}}_a:a=a.$$
The map path inversion we have defined this way has type
$$\{a\ b : A\} \to a = b \to b = a.$$
We can also define path \emph{concatenation}. Given $p:a=_Ab$ and $q:b=_Ac$, we define $p\cdot q:a=_Ac$ again by path induction. We will only give the intuitive argument and leave the formal proof to the reader. Since $q$ has free endpoint $c$, we may assume that $c\equiv b$ and $q\equiv\ensuremath{\operatorname{\mathsf{refl}}}_b$. In this case, we define $p\cdot \ensuremath{\operatorname{\mathsf{refl}}}_b\ensuremath{\vcentcolon\equiv} p:a=b$.
We can also define higher paths. For example, given $p:a=b$ and $q:b=c$ and $r:c=d$, we have a path
$$p\cdot(q \cdot r)=(p\cdot q)\cdot r,$$
which is the \emph{associativity} of path concatenation. We can prove this by path induction on $r$: if $r$ is reflexivity, then both sides reduce to $p\cdot q$.
By using path induction, we can also prove the following equalities:
\begin{align*}
p\cdot 1 &= p & p\cdot p^{-1} &= 1\\
1\cdot p &= p & p^{-1}\cdot p &= 1.
\end{align*}
It is trickier to prove the Eckmann-Hilton property of equality, which states that given $a:A$ and $p,q:\ensuremath{\operatorname{\mathsf{refl}}}_a=\ensuremath{\operatorname{\mathsf{refl}}}_a$, we have $p\cdot q=q\cdot p$.
The problem is that cannot apply path induction to $p$ or $q$ directly. We omit the proof here and refer to~\cite[Theorem 2.1.6]{hottbook}.
Given a map $f:A\to B$, we can prove that $f$ respects paths. Given a path
$p:a=_Aa'$, we define $\apfunc{f}(p):f(a)=_Bf(b)$ by path induction: for
reflexivity we define $\apfunc{f}(\ensuremath{\operatorname{\mathsf{refl}}}_a)\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{refl}}}_{f(a)}$. We will sometimes
abuse notation and write $f(p)$ for $\apfunc{f}(p)$. From a logical perspective this just states that functions respect equality,
but from a homotopical perspective, this states that functions respect paths, which is in line with our intuition that all functions are continuous in HoTT.
We can compute what $\apfunc{}$ does when our map is the identity map, a constant map or a composition of maps:
\begin{align*}
\apfunc{\ensuremath{\operatorname{\mathsf{id}}}_A}(p)&=p&\apfunc{\ensuremath{\mathbf{0}}_b}(p)&=\ensuremath{\operatorname{\mathsf{refl}}}_b&\apfunc{g\o f}(p)=\apfunc{g}(\apfunc{f}(p)).
\end{align*}
All three of these properties are easily proven by path induction. Also, we can compute $\apfunc{}$ when we apply it to inverses or concatenations of paths:
\begin{align*}
\apfunc{f}(p\cdot q)&=\apfunc{f}(p)\cdot\apfunc{f}(q)&\apfunc{f}(p^{-1})&=(\apfunc{f}(p))^{-1}.
\end{align*}
Given a dependent type $P:A\to\ensuremath{\mathsf{Type}}\xspace$ and a path $p:a=_Aa'$, we can define the
\emph{transport} function $\ensuremath{\operatorname{\mathsf{transport}}}^P(p): P(a)\to P(a')$. We define it by path induction; for reflexivity we define $\ensuremath{\operatorname{\mathsf{transport}}}^P(\ensuremath{\operatorname{\mathsf{refl}}}_a)\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{id}}}_{P(a)}$. When $P$ is known from context we will write $p_*(b)$ for $\ensuremath{\operatorname{\mathsf{transport}}}^P(p,b)$.
By path induction we can prove basic equalities about transports. We have
\begin{align*}
\ensuremath{\operatorname{\mathsf{transport}}}^P(p \cdot q,x) &= \ensuremath{\operatorname{\mathsf{transport}}}^P(q,\ensuremath{\operatorname{\mathsf{transport}}}^P(p,x))\\
\ensuremath{\operatorname{\mathsf{transport}}}^{\lam{a}B}(p,x) &= x\\
\ensuremath{\operatorname{\mathsf{transport}}}^{P\o f}(p,x) &= \ensuremath{\operatorname{\mathsf{transport}}}^P(\apfunc{f}(p),x)\\
f_{a'}(\ensuremath{\operatorname{\mathsf{transport}}}^P(p,x)) &= \ensuremath{\operatorname{\mathsf{transport}}}^Q(p,f_a(x))&&\text{for $f:(a:A)\to P(a)\to Q(a)$.}
\end{align*}
\subsection{Equivalences}\label{sec:equivalences}
In this section we talk about maps between types that have an inverse in a suitable way. Before we can give the definition, we need to define homotopy.
Given two dependent maps $f,g:(a:A)\to B(a)$, a \emph{homotopy} $h:f\sim g$ is a proof that $f$ and $g$ are pointwise equal:
$$(f\sim g)\ensuremath{\vcentcolon\equiv} (a:A)\to f(a)=_{B(a)}g(a).$$
Recall that all maps are considered continuous, so this actually gives a continuous deformation of $f$ to $g$, which is exactly what a homotopy is in topology.
\begin{defn}Suppose given a function $f:A\to B$.
\begin{itemize}
\item A \emph{left-inverse} of $f$ is an inhabitant of $(g:B\to A)\times g\o f\sim\ensuremath{\operatorname{\mathsf{id}}}_A$.
\item Similarly, a \emph{right-inverse} of $f$ is an inhabitant of $(h:B\to A)\times f\o h\sim\ensuremath{\operatorname{\mathsf{id}}}_B$.
\item We say that $f$ is an \emph{equivalence} or $\ensuremath{\mathsf{isequiv}}(f)$ if $f$ has both a left and a right inverse. We will denote its left-inverse by $f^{-1}$. We can then show that $f^{-1}$ is also a right inverse of $f$.
\item The type of equivalences between $A$ and $B$ is $(A\simeq B)\ensuremath{\vcentcolon\equiv}(f:A\to B)\to \ensuremath{\mathsf{isequiv}}(f)$. Given an element $f:A\simeq B$, we will also use $f$ to denote the underlying map $A\to B$.
\end{itemize}
\end{defn}
It is easy to show that the identity map $\ensuremath{\operatorname{\mathsf{id}}}_A:A\simeq A$ is an equivalence. Moreover, if $g:B\to C$ and $f:A\to B$ are both equivalences, then $g\o f$ and $f^{-1}$ are also equivalences. This shows that equivalences are reflexive, symmetric and transitive.
A very important property is that any two inhabitants of $\ensuremath{\mathsf{isequiv}}(f)$ are equal: if $p,q:\ensuremath{\mathsf{isequiv}}(f)$, then $p=q$. We will not prove this here, but it is shown in~\cite[Theorem 4.3.2]{hottbook}. This property is the reason that we define the notion of equivalences this way.
If we would define $\ensuremath{\mathsf{isequiv}}(f)$ by requiring a map that is both a left \emph{and} a right inverse of $f$, then this property would not hold.
Given two equivalences $f,f':A\simeq B$, it does not matter whether we compare them as functions or equivalences:
$$(f=_{A\simeq B}f')\simeq (f=_{A\to B}f') \simeq (f\sim f').$$
By path induction we also get a map $(A=_{\ensuremath{\mathsf{Type}}\xspace}B)\to(A\simeq B)$, because if the path $p:A=B$ is reflexivity, we can just take $\ensuremath{\operatorname{\mathsf{id}}}_A:A\simeq A$ as our equivalence. In plain Martin L\"of type theory one cannot characterize what the type $A=B$ is. This is where the univalence axiom comes in. The \emph{univalence axiom} states that the map
$$(A=B)\to (A\simeq B)$$
is an equivalence. In particular this means that we get a map in the other direction: given an equivalence $e:A\simeq B$, we get an equality $\ua(e):A=B$.
\subsection{More on paths}\label{sec:more-paths}
In this section we will discuss dependent paths, or pathovers; higher paths, such as squares and cubes; and paths in type formers.
\subsubsection*{Pathovers}
We will often need to relate elements in two different fibers of a dependent type. Suppose we have a family $P:A\to\ensuremath{\mathsf{Type}}\xspace$ with $x:P(a)$ and $x':P(a')$. If we have a path $p:a=a'$, we can form the type $x =_p^P x'$ of \emph{dependent paths} or \emph{pathovers} over $p$. There are four equivalent ways to define this:
\begin{enumerate}
\item \label{item:pathover-tr} We can define $(x =_p^P x')\ensuremath{\vcentcolon\equiv}(\ensuremath{\operatorname{\mathsf{transport}}}^P(p,x)=x')$
\item We can define $(x =_p^P x')\ensuremath{\vcentcolon\equiv}(x=\ensuremath{\operatorname{\mathsf{transport}}}^P(p^{-1},x'))$
\item We can define $(x =_p^P x')$ by path induction on $p$. If $p\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{refl}}}_a$, we define
$(x =_p^P x')\equiv(x =_{\ensuremath{\operatorname{\mathsf{refl}}}_a}^P x')\ensuremath{\vcentcolon\equiv}(x=_{P(a)}x')$.
\item \label{item:pathover-ind} We can define $(x =_p^P x')$ by a family of inductive types. For fixed $A:\ensuremath{\mathsf{Type}}\xspace$ and $P:A\to\ensuremath{\mathsf{Type}}\xspace$ and $a:A$ and $x:P(a)$ we have the following family:
\begin{inductive}
\texttt{inductive} $x =_{(-)}^P {(-)} : \{a' : A\} \to a = a' \to P(a') \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \ensuremath{\operatorname{\mathsf{refl}}} : x =_{\ensuremath{\operatorname{\mathsf{refl}}}_a}^P x.$
\end{inductive}
\end{enumerate}
It does not matter which of these definitions we pick, because we can prove that all of them are equivalent.\footnote{In Lean, we chose option \ref{item:pathover-ind}. Option \ref{item:pathover-tr} would probably be slightly more convenient to work with, because then this characterization becomes a definitional equality. In practice it will not matter much, though.}
We have the following equivalences between pathovers
\begin{align*}
(x=^{\lam{a}B}_px') &\simeq (x =_B x') & (x =^{P \o f}_p x') \simeq (x =^P_{\apfunc{f}(p)}x').
\end{align*}
We can do operations on pathovers, similar to the operations on paths. We have concatenation and inversion, and we will abuse notation and denote them with the same notation.
\begin{align*}
({-})\cdot ({-})&: x_1 =^P_{p} x_2 \to x_2 =^P_{q} x_3 \to x_1 =^P_{p\cdot q} x_3\\
({-})^{-1}&: x_1 =^P_{p} x_2 \to x_2 =^P_{p^{-1}} x_1.
\end{align*}
We have a dependent version of $\apfunc{}$. Given a dependent map $f:(a:A)\to P(a)$, we get
$$\apd_f: (p : a = a') \to f(a) =^P_p f(a').$$
A variant to $\apd$ is the following. Given $f:A\to B$, a family $P:B\to\ensuremath{\mathsf{Type}}\xspace$ and a section $g:(a:A)\to P(f(a))$, we define
\begin{align}\widetilde{\mathsf{apd}}_g: (p : a = a') \to g(a) =^P_{\apfunc{f}(p)} g(a').\label{eq:apdtilde}\end{align}
The difference between $\apd$ and $\widetilde{\mathsf{apd}}$ is over which path they lie.
Furthermore, if we have a map $f:A\to B$ and two families $P:A\to\ensuremath{\mathsf{Type}}\xspace$ and $Q:B\to\ensuremath{\mathsf{Type}}\xspace$ and a fiberwise map $g:(a:A)\to P(a)\to Q(f(a))$, then we get a fiberwise version of $\apfunc{}$:
\begin{equation}\ensuremath{\mathsf{apo}}_g: x =^P_p x'\to g_a(x) =^Q_{\apfunc{f}(p)} g_{a'}(x').\label{eq:apo}\end{equation}
\subsubsection*{Squares}
For higher paths, it is convenient to define a separate notion of a square in a type:
\begin{center}
\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$a_{00}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$a_{20}$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$a_{02}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$a_{22}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node [above] {$p_{10}$} (tr)
edge[double equal sign distance] node [right] {$p_{01}$} (bl)
(bl) edge[double equal sign distance] node [above] {$p_{12}$} (br)
(tr) edge[double equal sign distance] node [right] {$p_{21}$} (br);
\end{tikzpicture}
\end{center}
Suppose given four paths as in the diagram above, that is
\begin{align*}
p_{10}&:a_{00}=a_{20}&p_{01}&:a_{00}=a_{02}\\
p_{12}&:a_{02}=a_{22}&p_{21}&:a_{20}=a_{22}.
\end{align*}
We have a type of squares $\ensuremath{\mathsf{square}}(p_{10},p_{12},p_{01},p_{21})$, which we can define in either of the two following equivalent ways
\begin{enumerate}
\item We can define $\ensuremath{\mathsf{square}}(p_{10},p_{12},p_{01},p_{21})\ensuremath{\vcentcolon\equiv} (p_{10}\cdot p_{21} = p_{01} \cdot p_{12})$.
\item $\ensuremath{\mathsf{square}}(p_{10},p_{12},p_{01},p_{21})$ is defined as an inductive family of types. For a fixed $a_{00}:A$ we define the family
\begin{inductive}
\texttt{inductive} $\ensuremath{\mathsf{square}}({-},{-},{-},{-}) : \{a_{20}\ a_{02}\ a_{22} : A\} \to a_{00} = a_{20} \to a_{02} = a_{22} \to a_{00} = a_{02} \to a_{20} = a_{22} \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \ensuremath{\operatorname{\mathsf{refl}}} : \ensuremath{\mathsf{square}}(\ensuremath{\operatorname{\mathsf{refl}}}_{a_{00}},\ensuremath{\operatorname{\mathsf{refl}}}_{a_{00}},\ensuremath{\operatorname{\mathsf{refl}}}_{a_{00}},\ensuremath{\operatorname{\mathsf{refl}}}_{a_{00}}).$
\end{inductive}
\end{enumerate}
We will usually write squares using diagrams as above. There are various operations on squares. For example, we can horizontally concatenate them. If we can fill each of the individual squares below, we can fill the outer rectangle (which has as top $p_{10}\cdot p_{30}$ and as bottom $p_{12}\cdot p_{32}$).
\begin{center}
\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$a_{00}$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$a_{02}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$a_{20}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$a_{22}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tr] (tr2) {$a_{40}$};
\ensuremath{\operatorname{\mathsf{node}}} (br2) at (tr2 |- bl) {$a_{42}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node [above] {$p_{10}$} (tr)
edge[double equal sign distance] node [right] {$p_{01}$} (bl)
(bl) edge[double equal sign distance] node [above] {$p_{12}$} (br)
(tr) edge[double equal sign distance] node [above] {$p_{30}$} (tr2)
edge[double equal sign distance] node [right] {$p_{21}$} (br)
(br) edge[double equal sign distance] node [above] {$p_{32}$} (br2)
(tr2) edge[double equal sign distance] node [right] {$p_{41}$} (br2);
\end{tikzpicture}
\end{center}
We can also vertically concatenate squares, and horizontally or vertically invert squares.
Given a homotopy $h:f\sim g$ between nondependent functions $f,g:A\to B$ and a path $p:a=_Aa'$, we get the following \emph{naturality square}.
\begin{center}
\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$f(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$g(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$f(a')$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$g(a')$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node [above] {$h(a)$} (tr)
edge[double equal sign distance] node [left] {$\apfunc{f}(p)$} (bl)
(bl) edge[double equal sign distance] node [above] {$h(a')$} (br)
(tr) edge[double equal sign distance] node [right] {$\apfunc{g}(p)$} (br);
\end{tikzpicture}
\end{center}
\subsubsection*{Squareovers and cubes}
Going up further, we have the type of \emph{squareovers}. A squareover is a square in a dependent type over a square. Suppose that we have a dependent type $P:A\to\ensuremath{\mathsf{Type}}\xspace$, a square $s$ in $A$ and a dependent path over each of the sides of the square, as in the following diagram.
\begin{center}
\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$x_{00}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$x_{20}$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$x_{02}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$x_{22}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node [above] {$q_{10}$} (tr)
edge[double equal sign distance] node [right] {$q_{01}$} (bl)
(bl) edge[double equal sign distance] node (t) [above] {$q_{12}$} (br)
(tr) edge[double equal sign distance] node [right] {$q_{21}$} (br);
\ensuremath{\operatorname{\mathsf{node}}}[below = of bl] (tl) {$a_{00}$};
\ensuremath{\operatorname{\mathsf{node}}} (tr) at (br |- tl) {$a_{20}$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$a_{02}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$a_{22}$};
\ensuremath{\operatorname{\mathsf{node}}} (m) at ($(tl)!0.5!(br)$) {$s$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node (b) [above] {$p_{10}$} (tr)
edge[double equal sign distance] node [left] {$p_{01}$} (bl)
(bl) edge[double equal sign distance] node [above] {$p_{12}$} (br)
(tr) edge[double equal sign distance] node [right] {$p_{21}$} (br)
(t) edge[->, shorten <= 3mm] (b);
\end{tikzpicture}
\end{center}
We have the type of \emph{squareovers} or \emph{dependent squares}, which fill the top square and lie over the bottom square. We can again define this using multiple methods, but the most convenient method here is to define it as an inductive family. We take as parameters the type $A$, the family $P$ and the points $a_{00}$ and $x_{00}$ and let all the other arguments be indices. We have a ``reflexivity squareover'' when the square $s$ is the reflexivity square and each of the four pathovers are reflexivity pathovers.
We can also define a type of cubes. Given six squares in a type with twelve paths as sides, fitting together in a cube, we can define the type of fillers of the cube. This is again done using a family of inductive types, where we give a cube filler when all the six sides are reflexivity squares. Of course, we could continue by defining cubeovers and 4-cubes, but we will not need them in this dissertation.
\subsection*{Paths in type formers}
In each of the type formers of \autoref{sec:martin-lof-type} we can compute what the paths in that type are, and what the operations of paths are in that type.
As a simple example, consider the cartesian product type $A\times B$. A path in the cartesian product is just a pair of paths.
$$(x=_{A\times B}y)\simeq(p_1x=_Ap_1y)\times (p_2x=_Bp_2y)$$
In particular, given paths $r:p_1x=p_1y$ and $s:p_2x=p_2y$, we get a path
$x=y$, which we will denote $(r,s)$. Given maps $f:A\to A'$ and $g:B\to B'$, we get the map $f\times g:A\times B\to A'\times B'$ and we can compute
$$\apfunc{f\times g}(r,s)=(\apfunc{f}(r),\apfunc{g}(s))$$.
Given families $P,Q:A\to\ensuremath{\mathsf{Type}}\xspace$, we can compute transport:
$$\ensuremath{\operatorname{\mathsf{transport}}}^{\lam{a}P(a)\times Q(a)}(p,(x,y))=(\ensuremath{\operatorname{\mathsf{transport}}}^P(p,x),\ensuremath{\operatorname{\mathsf{transport}}}^Q(p,y))$$
Pathovers in a family of cartesian products are also pairs of pathovers:
$$(x,y)=^{\lam{a}P(a)\times Q(a)}_p(x',y')\simeq(x=^P_px')\times (y=^Q_p y').$$
In sigma-types the relations are a bit more difficult, since the second component depends on the first. In the type $(a:A)\times B(a)$ paths are pairs of a path and a path over that path:
$$(x=_{(a:A) \times B(a)}y)\simeq(r:p_1x=_Ap_1y)\times (p_2x=^B_rp_2y)$$
We will also denote in this case the map from right to left by $({-},{-})$. Given a map $f:A\to A'$ and a fiberwise map $g:(a:A)\to B(a)\to B'(f(a))$, we get a functorial action of the sigma type: $f\times g: ((a:A)\times B(a))\to ((a':A')\times B(a'))$. In this case, we can compute
$$\apfunc{f\times g}(r,s)=(\apfunc{f}(r),\ensuremath{\mathsf{apo}}_g(s)),$$
where $\ensuremath{\mathsf{apo}}$ is defined in \eqref{eq:apo}.
We leave the rule for transports as an exercise to the reader, but the rule for pathovers in a family of sigma-types is the following. For
$B:A\to\ensuremath{\mathsf{Type}}\xspace$ and $C:(a:A)\to B(a)\to\ensuremath{\mathsf{Type}}\xspace$ we get:\footnote{We could define a new notion ``path over a pathover,'' but the rule given here suffices for all the cases we considered.}
$$((a,b)=^{\lam{a}(b:B(a))\times C(a,b)}_p(a',b'))\simeq(q:a=^P_pa')\times (y=^{\lam{x:(a{:}A)\times B(a)}Q(p_1x,p_2x)}_{(p,q)} y').$$
For dependent function types the situation is a bit more complicated. Given $f,g:(a:A)\to
B(a)$, by path induction we get a map
$$\mathsf{happly}:(f=g)\to f\sim g.$$
However, we cannot show in plain Martin-L\"of type theory that this map gives
rise to an equivalence. In homotopy type theory we can use the univalence axiom
(see \autoref{sec:equivalences}) to show that $\mathsf{happly}$ is an equivalence. We
skip the proof here, but refer the reader to~\cite[Section 4.9]{hottbook}. Using univalence we can also prove the other properties. The general rule for pathovers in a dependent function type is complicated, but two important special cases are the following. In the first case, the domain does not depend on the path. We have types $A$ and $B$ and a family $C:A\to B\to \ensuremath{\mathsf{Type}}\xspace$ and then we can prove:
$$(f=^{\lam{a}(b:B)\to C(a,b)}_pg)\simeq(b:B)\to f(b)=^{C({-},b)}_{p} g(b).$$
The second case is for nondependent functions. Given a type $A$ and two families $B,C:A\to\ensuremath{\mathsf{Type}}\xspace$, we have
$$(f=^{\lam{a}B(a)\to C(a)}_pg)\simeq(b:B(a))\to f(b)=^{C}_{p} g(p_*(b)).$$
We characterized paths in the universe in \autoref{sec:equivalences} using the univalence axiom. We will not need to do much path algebra in inductive types, except for the identity type, pathover type and square type.
A pathover in a family of identity types is a square. Suppose given types $A$ and $B$ and functions $f, g : A \to B$, a path $p:a=_Aa'$ and paths $q:f(a)=g(a)$ and $r:f(a')=g(a')$. Then the pathover type becomes equivalent to the square type shown below.
$$(q=^{\lam{a}f(a)=g(a)}_{p}r)\simeq
\begin{tikzpicture}[node distance=10mm,baseline=(l.base)]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$f(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$g(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$f(a')$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$g(a')$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node [above] {$q$} (tr) edge[double
equal sign distance] node [right] (l) {$\apfunc{f}(p)$} (bl) (bl)
edge[double equal sign distance] node [above] {$r$} (br) (tr) edge[double
equal sign distance] node [right] {$\apfunc{g}(p)$} (br);
\end{tikzpicture}$$
We also sometimes encounter a pathover in a dependent family of
pathovers. In that case we get a squareover. Suppose we are given
functions $f,g:A\to B$, and a homotopy $h:f\sim g$, a dependent family
$C:B\to\ensuremath{\mathsf{Type}}\xspace$ and sections $c:(a:A)\to C(f(a))$ and $c':(a:A)\to
C(g(a))$. We want to characterize a pathover in the family
$P\ensuremath{\vcentcolon\equiv}\lam{a}c(a)=^C_{h(a)}c'(a):A\to\ensuremath{\mathsf{Type}}\xspace$. If we are also given a
path $p:a=_Aa'$ and two pathovers $q:c(a)=^C_{h(a)}c'(a)$ and
$q':c(a')=^C_{h(a')}c'(a')$, then the pathover $q=^P_pq'$ is equivalent to the
following squareover, where $\widetilde{\mathsf{apd}}$ is defined in \eqref{eq:apdtilde}, and the bottom square is a naturality square.
\begin{center}
\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$c(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$c'(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$c(a')$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$c'(a')$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node [above] {$q$} (tr)
edge[double equal sign distance] node [left] {$\widetilde{\mathsf{apd}}_b(p)$} (bl)
(bl) edge[double equal sign distance] node (t) [above] {$q'$} (br)
(tr) edge[double equal sign distance] node [right] {$\widetilde{\mathsf{apd}}_{b'}(p)$} (br);
\ensuremath{\operatorname{\mathsf{node}}}[below = of bl] (tl) {$f(a)$};
\ensuremath{\operatorname{\mathsf{node}}} (tr) at (br |- tl) {$g(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$f(a')$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$g(a')$};
\ensuremath{\operatorname{\mathsf{node}}} (m) at ($(tl)!0.5!(br)$) {nat.};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node (b) [above] {$h(a)$} (tr)
edge[double equal sign distance] node [left] {$\apfunc{f}(p)$} (bl)
(bl) edge[double equal sign distance] node [below] {$h(a')$} (br)
(tr) edge[double equal sign distance] node [right] {$\apfunc{g}(p)$} (br)
(t) edge[->, shorten <= 3mm] (b);
\end{tikzpicture}
\end{center}
Lastly, we will mention that a pathover in a family of squares is a cube, but we will not explain the details here.
\subsection{Truncated Types}\label{sec:truncatedness}
In HoTT we can define iterated path spaces in any type. In certain types, if we iterate path spaces enough times, these path spaces do not contain any information. These types are called \emph{truncated}. The notion of an $n$-truncated type, was introduced in 2009 by Vladimir Voevodsky under the name ``a type of h-level $n+2$.''
We define the notion that $A$ is $n$-truncated, or that $A$ is an \emph{$n$-type} or $\istrunc{n}(A)$ recursively for $n\geq-2$. We say that a type $A$ is $(-2)$-truncated or \emph{contractible} if it has exactly one inhabitant, i.e. if we can prove
$$(a_0:A)\times (a:A)\to a=a_0.$$
A type $A$ is $(n+1)$-truncated if for all $a\ a':A$ the type $a=_Aa'$ is $n$-truncated.
We can show that $\ensuremath{\mathbf{1}}\xspace$ is contractible and that every contractible type is equivalent to $\ensuremath{\mathbf{1}}\xspace$.
The $(-1)$-truncated types are called \emph{mere propositions} or \emph{propositions} for short. A type $A$ is a proposition precisely when any two of its inhabitants are equal, i.e. if we can prove
$$(a\ a':A)\to a=a'.$$
We call these types propositions because these types correspond to truth values, and do not contain any further information. In particular, if a proposition is inhabited, then it is contractible.
It is easy to see that $\ensuremath{\mathbf{0}}\xspace$ and $\ensuremath{\mathbf{1}}\xspace$ are mere propositions, and in \autoref{sec:equivalences} we saw that the statement $\ensuremath{\mathsf{isequiv}}(f)$ is a mere proposition.
One level up, the $0$-types are called \emph{sets}. These are the types for which uniqueness of identity proofs holds. Examples of sets are $\textbf{N}$ and $\ensuremath{\mathbf{2}}\xspace$.
On the next level we have the $1$-types or \emph{groupoids}. Below we list some properties of truncated types, see~\cite[Section 7.1]{hottbook} for their proofs.
\begin{lem}\mbox{}
\begin{itemize}
\item If $A$ is $n$-truncated, then $A$ is $m$-truncated for all $m\geq n$.
\item If $A$ is $n$-truncated and $A\simeq B$, then $B$ is $n$-truncated.
\item If $A$ and $B$ are $n$-truncated types, then $A\times B$ and $A\simeq B$ are $n$-truncated.
If $n\geq0$, then $A+B$ is also $n$-truncated.
\item If $B:A\to\ensuremath{\mathsf{Type}}\xspace$ is a family of $n$-truncated types (i.e. $(a:A)\to\istrunc{n}(B(a))$), then $(a:A)\to B(a)$ is $n$-truncated. If moreover $A$ is also $n$-truncated, then $(a:A)\times B(a)$ is also $n$-truncated.
\item Given $a_0:A$, the type $(a:A)\times (a_0=a)$ is contractible.
\item The type $\istrunc{n}A$ is a mere proposition.
\end{itemize}
\end{lem}
We define the \emph{subuniverse of $n$-types} as $\ensuremath{\mathsf{Type}}\xspace_{\leq n}\ensuremath{\vcentcolon\equiv} (X:\ensuremath{\mathsf{Type}}\xspace)\times\istrunc{n}(X)$. For $X:\ensuremath{\mathsf{Type}}\xspace_{\leq n}$ we will also write $X$ for the underlying type of $X$. We write $\ensuremath{\mathsf{Prop}}\xspace\ensuremath{\vcentcolon\equiv} \ensuremath{\mathsf{Type}}\xspace_{\leq -1}$ and $\ensuremath{\mathsf{Set}}\xspace\ensuremath{\vcentcolon\equiv}\ensuremath{\mathsf{Type}}\xspace_{\leq0}$.
We can do set-level mathematics in the subuniverse of sets. For example, we can define a \emph{group} to be a set with operations satisfying the following axiomatization:\footnote{From these equalities the fact that $e$ is a left-identity and $i$ is a left-inverse can be derived.}
\begin{align*}
\ensuremath{\mathsf{Group}}&\ensuremath{\vcentcolon\equiv}(G:\ensuremath{\mathsf{Set}}\xspace)\times (m:G\to G \to G)\times (i:G\to G)\times (e:G)\times ((x\ y\ z: G) \to {}\\
&\mathrel{\hphantom{\ensuremath{\vcentcolon\equiv}}} m(x,m(y,z))=m(m(x,y),z)\times m(x,e)=x\times m(x,i(x))=e).
\end{align*}
A group $G$ is \emph{abelian} if it moreover satisfies $m(x,y)=m(y,x)$ for all $x,y:G$. This gives the usual notion of groups, and we can perform all basic group theory in this setting.
\subsubsection*{Truncations}
We can turn every type $A$ into an $n$-type $\|A\|_n$ in a universal way, which is called the \emph{$n$-truncation} of $A$. It comes with a map $|{-}|_n:A\to\|A\|_n$ and has the following induction principle. Suppose given $P:\|A\|_n\to\ensuremath{\mathsf{Type}}\xspace$ such that $P(x)$ is $n$-truncated for all $x:\|A\|_n$. If we are given a dependent map $f:(a:A)\to P(|a|_n)$, we get a section
$$\ind{\|{-}\|}(f):(x:\|A\|_n)\to P(x)$$
such that $\ind{\|{-}\|}(f,|a|_n)\equiv f(a)$.
We will now state some properties of the $n$-truncation, for the proofs we refer to~\cite[Section 7.3]{hottbook}.
\begin{lem}\mbox{}
\begin{itemize}
\item The truncation is functorial. Given $f:A\to B$, we get a map $\|f\|_n : \|A\|_n \to \|B\|_n$. This map respects composition and identities: $\|g\o f\|_n\sim \|g\|_n \o \|f\|_n$ and $\|\ensuremath{\operatorname{\mathsf{id}}}_A\|_n\sim \ensuremath{\operatorname{\mathsf{id}}}_{\|A\|_n}$.
\item $A$ is an $n$-type iff $|{-}|_n : A \to \|A\|_n$ is an equivalence.
\item The equality type in the truncation is truncated equality, but shifted: $$(|a|_{n+1} =_{\|A\|_{n+1}} = |a'|_{n+1})\simeq \|a =_A a'\|_n.$$
\item Truncating twice is the same as truncating once: $$\|\|A\|_n\|_k\simeq \|A\|_{\min(n,k)}.$$
\end{itemize}
\end{lem}
In particular the \emph{propositional truncation} $\|A\|\ensuremath{\vcentcolon\equiv}\|A\|_{-1}$ of $A$ is a proposition stating that $A$ is \emph{merely inhabited}~\cite{awodey2004Propositions}.
We can use it to define \emph{proof irrelevant} versions of the disjunction or existential quantifier. We have the \emph{mere disjunction}
\begin{align*}(P\vee Q)&\ensuremath{\vcentcolon\equiv} \|P+Q\|
\intertext{and the \emph{mere existential}}
(\exists(x:A).P(x))&\ensuremath{\vcentcolon\equiv} \|(x:A)\times P(x)\|.
\end{align*}
We say that there \emph{merely exists} $x:A$ such that $P(x)$ holds if $\exists(x:A).P(x)$ is inhabited, to contrast with constructing an element in the untrucated dependent pair type. If we construct an element of $(x:A)\times P(x)$, we will sometimes say that there \emph{purely exists} an $x$ such that $P(x)$ holds, but often we will drop the adverb \emph{purely}.
\subsubsection*{Connected types}
A type is truncated if the type contains no interesting information in a high enough dimension. Dually, a type is \emph{connected} if it contains no interesting information in a low enough dimension.
We say that a type $A$ is \emph{$n$-connected} for $n\geq-2$ if $\|A\|_n$ is contractible. From the definition we see that every type is $(-2)$-connected. A type is $(-1)$-connected precisely when it is merely inhabited. A type is called 0-connected
when $A$ has exactly one \emph{connected component}. A 1-connected type is called \emph{simply connected}.
\subsubsection*{Fibers}
We can extend the notion of truncated types and connected types to functions. Given a function $f:A\to B$ and a point $b:B$, we define the \emph{fiber} of $f$ at $b$ to be
$$\ensuremath{\operatorname{\mathsf{fib}}}_f(b)\ensuremath{\vcentcolon\equiv}(a:A)\times f(a)=b.$$
The fiber of the projection $p_1:((a:A)\times B(a))\to A$ at $a:A$ is equivalent to $B(a)$, which explains the terminology that $B(a)$ is the fiber of $B$ over $a$.
We say that a function $f:A\to B$ is $n$-truncated ($n$-connected) when for all $b:B$ the type $\ensuremath{\operatorname{\mathsf{fib}}}_f(b)$ is $n$-truncated ($n$-connected). The function $f$ is $(-2)$-truncated precisely when it is an equivalence. The function $f$ is $(-1)$-truncated, or an \emph{embedding}, if for all $a\ a':A$ the map $\apfunc{f}:a=_Aa'\to f(a)=_Bf(a')$ is an equivalence. A map $f:A\to B$ between sets is an embedding iff it is \emph{injective}, i.e. if we have a map $f(a)=f(a')\to a=a'$ for all $a\ a':A$. On the other hand, a $(-1)$-connected map is called a \emph{surjection}, which means that for every $b:B$ there merely exists an $a:A$ such that $f(a)=b$.
Every map can be factorized as an $n$-connected map followed by an $n$-truncated map in a unique way, which means that these classes form an \emph{orthogonal factorization system}~\cite{rijke2017modalities}.
Similar to the universe of $n$-truncated types, we have a universe of $n$-connected types: $$\ensuremath{\mathsf{Type}}\xspace_{>n}\ensuremath{\vcentcolon\equiv} (A:\ensuremath{\mathsf{Type}}\xspace)\times\isconn{n}(A).$$
\subsection{Pointed Types}\label{sec:pointed}
A lot of homotopy theory is done in the $(\infty,1)$-category of pointed types
where the morphisms are maps that preserve the basepoints of the types. Below are the
basic definitions for pointed types.
\begin{defn}\label{def:pointed-types-basic}\mbox{}
\begin{enumerate}
\item A type $A$ is \emph{pointed} if $A$ has a distinguished basepoint
$a_0:A$. For example, $\ensuremath{\mathbf{1}}\xspace$ is pointed by $\star$ and $\ensuremath{\mathbf{2}}\xspace$ is pointed
with ${0_{\bool}}$. We will also write $\ensuremath{\S^0}$ for the pointed type $\ensuremath{\mathbf{2}}\xspace$. $A\times B$ is pointed if both $A$ and $B$ are
pointed,\footnote{More formally, we have to specify the basepoint of $A\times
B$, because being pointed is structure on a type, not a property of the
type, but there is only one choice of basepoint in this example and other
examples where we leave the basepoint implicit.} $(a:A)\to B(a)$ is
pointed if $B$ is a family of pointed types, and $(a:A)\times B(a)$ is
pointed if $A$ is pointed and $B(a_0)$ is pointed.
\item The type of \emph{pointed types} is $\ensuremath{\mathsf{Type}}\xspace^*\ensuremath{\vcentcolon\equiv}(A:\ensuremath{\mathsf{Type}}\xspace)\times A$.
Given a pointed type $A:\ensuremath{\mathsf{Type}}\xspace^*$, we will also write $A$ for its underlying
type.
\item Given two pointed types $A,B:\ensuremath{\mathsf{Type}}\xspace^*$, a \emph{pointed map} $f:A\to^*B$
is a pair consisting of a map $f:A\to B$ and a path $f_0$ stating that $f$
preserves the basepoint, that is $f_0:f(a_0)=b_0$. The type $A\to^*B$ is
pointed with basepoint
$\ensuremath{\mathbf{0}}\equiv\ensuremath{\mathbf{0}}_{A,B}\ensuremath{\vcentcolon\equiv}(\lam{a}b_0,\ensuremath{\operatorname{\mathsf{refl}}}_{b_0})$.
\item We have an identity pointed map $\ensuremath{\operatorname{\mathsf{id}}}\equiv\ensuremath{\operatorname{\mathsf{id}}}_A:A\to^*A$ defined as
$(\lam{x}x,\ensuremath{\operatorname{\mathsf{refl}}}_{a_0})$ and if $g:B\to^*C$ and $f:A\to^*B$ we have a
composite $g\o f:A\to^* C$ defined as $(\lam{x}g(f(x)),\mapfunc{g}(f_0)\cdot
g_0)$.
\item More generally, Given a pointed type $A:\ensuremath{\mathsf{Type}}\xspace^*$ and a family of types
$B:A\to\ensuremath{\mathsf{Type}}\xspace$ with a basepoint $b_0:B(a_0)$, a \emph{pointed dependent map}
$f:(a:A)\to^*B(a)$ is a pair consisting of a dependent map $f:(a:A)\to B(a)$
and a path $f_0:f(a_0)=b_0$. If we require that $B$ is a family of pointed
types, i.e. $B:A\to\ensuremath{\mathsf{Type}}\xspace^*$, then $(a:A)\to^* B(a)$ is pointed with
basepoint $(\lam{a}b_0(a),\ensuremath{\operatorname{\mathsf{refl}}}_{b_0(a_0)}).$
\item Given two pointed dependent maps $f,g:(a:A)\to^* B(a)$, a \emph{pointed
homotopy} $h:f\sim^* g$ is a pointed dependent map $(a:A)\to^* f(a)=g(a)$.
This is well-defined, since the type $f(a_0)=g(a_0)$ is pointed by $f_0\cdot g_0^{-1}$.
Expanding the definition, this means that $h$ is a pair of a homotopy
$h:f\sim g$ and a 2-path stating that $h$ relates the basepoint-preserving
paths of $f$ and $g$. This means that we have $h_0:h(a_0)=f_0\cdot g_0^{-1}$, or equivalently,
$h_0:h(a_0)\cdot g_0=f_0$. We say that a diagram of pointed types commutes
if there are pointed homotopies between the corresponding composites of
pointed maps.
\item A pointed map $e:A\to^* B$ is a \emph{pointed equivalence} if it has a
left-inverse and a right-inverse. That is, there is $\ell:B\to^* A$ such that
$\ell\o e\sim^* \ensuremath{\operatorname{\mathsf{id}}}_A$ and $r:B\to^* A$ such that $e\o r\sim^* \ensuremath{\operatorname{\mathsf{id}}}_B$. The
type of pointed equivalences between $A$ and $B$ is denoted $A\simeq^* B$.
The identity map is a pointed equivalence and pointed equivalences are
closed under composition.
\item Given $A:\ensuremath{\mathsf{Type}}\xspace^*$, we define its \emph{loop space} $\Omega A:\ensuremath{\mathsf{Type}}\xspace^*\ensuremath{\vcentcolon\equiv}
(a_0=a_0,\ensuremath{\operatorname{\mathsf{refl}}}_{a_0})$. We define the \emph{iterated loop space} $\Omega^nA$ by iteration as $\Omega^0A\ensuremath{\vcentcolon\equiv} A$ and $\Omega^{n+1}A\ensuremath{\vcentcolon\equiv} \Omega(\Omega^nA)$.
\item We define the $n$-th homotopy group of $A$ as the set-truncation of the iterated loop space, i.e. $\pi_n(A)\ensuremath{\vcentcolon\equiv}\|\Omega^nA\|_0$. This is a group for $n\geq1$ that is abelian for $n\geq 2$.
\item Given a pointed map $f:A\to^* B$, we define the \emph{pointed fiber of
$f$} $\ensuremath{\operatorname{\mathsf{fib}}}_f:\ensuremath{\mathsf{Type}}\xspace^*$ as $\ensuremath{\operatorname{\mathsf{fib}}}_f(b_0)\equiv(x:A)\times f(a)=b_0$ with basepoint $(a_0,f_0)$.
There is a pointed map $p_1:\ensuremath{\operatorname{\mathsf{fib}}}_f\to^*A$ defined as
$(\lam{x}p_1(x),\ensuremath{\operatorname{\mathsf{refl}}}_{a_0})$.
\end{enumerate}
\end{defn}
Here are some basic properties of pointed types. We omit the proofs.
\begin{lem}\label{lem:pointed-types-basic}\mbox{}
\begin{enumerate}
\item Suppose given a pointed map $f:A\to^* B$. The type of proofs that $f$ is an
equivalence is equivalent to the type that $f$ is a pointed equivalence. In
particular, being a pointed equivalence is a property. Also, we can define a
pointed equivalence $X\simeq^* Y$ by giving a map $e:X\to Y$ that is both
an equivalence and pointed.
\item Suppose given $A,B:\ensuremath{\mathsf{Type}}\xspace^*$. Univalence implies \emph{univalence for pointed
types}: the canonical map $(A = B)\to (A\simeq^* B)$ is an equivalence.
\item Suppose given pointed maps $f,g:(a:A)\to^* B(a)$. Function extensionality
implies \emph{function extensionality for pointed maps:} the canonical
map $(f = g) \to (f \sim^* g)$ is an equivalence.
\item We have the usual categorical laws:
\begin{align*}
f\o\ensuremath{\operatorname{\mathsf{id}}} &\sim^* f&\ensuremath{\operatorname{\mathsf{id}}}\o f&\sim^* f&(h\o g)\o f&\sim^* h\o (g\o f)\\
f\o\ensuremath{\mathbf{0}} &\sim^* \ensuremath{\mathbf{0}}&\ensuremath{\mathbf{0}}\o f&\sim^*\ensuremath{\mathbf{0}}
\end{align*}
The two homotopies showing $\ensuremath{\mathbf{0}}\o\ensuremath{\operatorname{\mathsf{id}}}\sim^*\ensuremath{\mathbf{0}}$ are equal. This is also
true for the two homotopies of $\ensuremath{\operatorname{\mathsf{id}}}\circ\ensuremath{\mathbf{0}}\sim^*\ensuremath{\mathbf{0}}$ and of $\ensuremath{\operatorname{\mathsf{id}}}\o\ensuremath{\operatorname{\mathsf{id}}}\sim^*\ensuremath{\operatorname{\mathsf{id}}}$
and of $\ensuremath{\mathbf{0}}\o \ensuremath{\mathbf{0}}\sim^*\ensuremath{\mathbf{0}}$.
\item
We can form iterated pointed maps $(A\to^* B \to^* C)\ensuremath{\vcentcolon\equiv} (A\to^*
(B\to^* C))$. To show that such a map preserves the basepoint, we need to
give an equality between pointed maps, or equivalently, we can give a
pointed homotopy between pointed maps. For example, the above homotopies
involving $\ensuremath{\mathbf{0}}$ imply that we have precomposition and postcomposition maps.
For $f:A\to^* B$ we have a pointed map $({-})\o f:(B\to^* C)\to^* A \to^* C$
and for $g:B\to^* C$ we have a pointed map $g\o({-}):(A\to^* B)\to^* A \to^*
C$. We will also write $f\to C$ resp. $A \to g$ for these maps.
Precomposition and postcomposition commute, which means that the
following square commutes.
\begin{center}
\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$(A\to^* B)$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$(A\to^* B)$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$(A'\to^* B)$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$(A'\to^* B')$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$g\o({-})$} (tr)
edge[->] node [right] {$({-})\o f$} (bl)
(bl) edge[->] node [above] {$g\o({-})$} (br)
(tr) edge[->] node [right] {$({-})\o f$} (br);
\end{tikzpicture}
\end{center}
Moreover, if $f$ or $g$ are constant, then these maps are pointed homotopic
to constant maps, which gives a pointed map
$$({-})\o({-}):(B\to^* C)\to^*(A\to^* B)\to^* A\to^* C.$$
\item\label{item:fiber-composition}
There are also dependent versions of these composition maps. In particular,
if $g:(a:A)\to B(a)\to^* C(a)$, then we have a map
$$g\o({-}):((a:A)\to^* B(a))\to^* (a:A)\to^* C(a).$$
We have an equivalence $$\ensuremath{\operatorname{\mathsf{fib}}}_{g\o({-})}\simeq^* ((a:A)\to^* \ensuremath{\operatorname{\mathsf{fib}}}_{ga}).$$
\item $\Omega$ and $\Omega^n$ are pointed functors. For $\Omega$ this means
that given a pointed map $f:A\to^* B$, we can define $\Omega f:\Omega A \to^*
\Omega B$, with pointed homotopies $\Omega(g\o f)\sim^* \Omega g \o \Omega
f$ and $\Omega\ensuremath{\operatorname{\mathsf{id}}}\sim^* \ensuremath{\operatorname{\mathsf{id}}}$ and $\Omega\ensuremath{\mathbf{1}}\xspace\simeq^*\ensuremath{\mathbf{1}}\xspace$. This also
implies that $\Omega\ensuremath{\mathbf{0}}\sim^*\ensuremath{\mathbf{0}}$ and that if $e:A\simeq^* B$ then
$\Omega e:\Omega A\simeq^* \Omega B$.
\item \label{item:pointed-function-extensionality}
There is a pointed version of function extensionality for pointed types.
If $B$ is a family of pointed types, we have a pointed equivalence
$$e_B:\Omega((a:A)\to^* B(a))\simeq^* ((a:A)\to^* \Omega B(a)).$$
This equivalence is natural in $B$. This means that given a fiberwise
pointed map $f:(a:A)\to B(a)\to^* C(a)$, the following square commutes.
\begin{center}
\begin{tikzpicture}
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\Omega((a:A)\to^* B(a))$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$((a:A)\to^* \Omega B(a))$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$\Omega((a:A)\to^* C(a))$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$((a:A)\to^* \Omega C(a))$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$e_B$} (tr)
edge[->] node [right] {$\Omega(f\o({-}))$} (bl)
(bl) edge[->] node [above] {$e_C$} (br)
(tr) edge[->] node [right] {$\Omega f \o ({-})$} (br);
\end{tikzpicture}
\end{center}
\item The fiber of a pointed map is functorial. This means that given a commuting square, we get a pointed map from the fiber of the top map to the fiber of the bottom map.
\begin{center}
\begin{tikzpicture}
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\ensuremath{\operatorname{\mathsf{fib}}}_f$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (t) {$A$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of t] (tr) {$B$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$\ensuremath{\operatorname{\mathsf{fib}}}_{f'}$};
\ensuremath{\operatorname{\mathsf{node}}} (b) at (t |- bl) {$A'$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$B'$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$p_1$} (t)
edge[->, dashed] (bl)
(t) edge[->] node [above] {$f$} (tr)
edge[->] node [right] {$g$} (b)
(bl) edge[->] node [above] {$p_1$} (b)
(b) edge[->] node [above] {$f'$} (br)
(tr) edge[->] node [right] {$h$} (br);
\end{tikzpicture}
\end{center}
Moreover, if the left and the right sides of the squares are equivalences, then the functorial action is an equivalence. Lastly, $p_1$ is natural, which means that the left square commutes.
\item Given a pointed map $f:A\to^*B$, we have a equivalence $\Omega\ensuremath{\operatorname{\mathsf{fib}}}_f\simeq^*\ensuremath{\operatorname{\mathsf{fib}}}_{\Omega f}$ that is natural in $f$. This means that if we have a commuting square with top $f$ and bottom $f'$, then the following square commutes (the left and the right side come from the functorial action of $\ensuremath{\operatorname{\mathsf{fib}}}$).
\begin{center}
\begin{tikzpicture}
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\Omega\ensuremath{\operatorname{\mathsf{fib}}}_f$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$\ensuremath{\operatorname{\mathsf{fib}}}_{\Omega f}$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$\Omega\ensuremath{\operatorname{\mathsf{fib}}}_{f'}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$\ensuremath{\operatorname{\mathsf{fib}}}_{\Omega f'}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$\sim$} (tr)
edge[->] (bl)
(bl) edge[->] node [above] {$\sim$} (br)
(tr) edge[->] (br);
\end{tikzpicture}
\end{center}
\item We have a pointed equivalence $(\ensuremath{\S^0}\to^* X)\simeq^* X$ natural in $X$.
\item A pointed type $A$ is $n$-connected iff $\pi_k(A)$ is trivial (contractible) for all $k\leq n$. If a type $A$ is $n$-truncated, then $\pi_k(A)$ is trivial for $k>n$ (however, the converse is not true in general).
\end{enumerate}
\end{lem}
\subsection{Higher Inductive Types} \label{sec:high-induct-types}
Higher inductive types are a generalization of inductive types where we specify not only the generating points in the type by constructors, but also the generating paths and higher paths.
The idea is that the type together with its (higher) path spaces are freely generated by these constructors.
A simple example is the \emph{interval}. The interval $I$ is generated by two points $0,1:I$ and a path $\ensuremath{\mathsf{seg}}\xspace:0=1$. Using a syntax similar to that of inductive types, we could write
\begin{inductive}
\texttt{HIT} $I:\ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ 0, 1 : I$; \\
$\bullet\ \ensuremath{\mathsf{seg}}\xspace : 0=_I1$.
\end{inductive}
Note that this is not an inductive type, since the last constructor does not specify an element in $I$, but an element in the path space of $I$. We get an induction principle for higher inductive types, similar to the induction principle for inductive types. We first give a special case, the nondependent induction principle, also called the \emph{recursion principle}. For the interval this states the following. Given a type $X$, if we have points $x_0\ x_1:X$ and a path $p:x_0=_Xx_1$, then we get a map $\rec{I}(x_0,x_1,p):I\to X$. On the points this has the expected computation rules:
$$\rec{I}(x_0,x_1,p,0)\equiv x_0\qquad\text{and}\qquad\rec{I}(x_0,x_1,p,1)\equiv x_1.$$
We want a similar computation rule on paths. We can apply the induction principle to $\ensuremath{\mathsf{seg}}\xspace$ using $\apfunc{}$. The resulting computation rule is
$$\apfunc{\rec{I}(x_0,x_1,p)}(\ensuremath{\mathsf{seg}}\xspace)=p.$$
Note that for this case we postulate a member of the identity type instead of making this a definitional equality. There are various reasons for this. Firstly, in this type theory, there is no justification for this equality to be definitional. There are various ways to define $\apfunc{}$, and there is no good reason for the computation rules to favor this definition. Secondly, in the early proof assistants for HoTT there was no support for definitional computation rules on path constructors, but there was a trick to get it for the point constructors~\cite{licata2011trick}. In fact, calling this rule a ``computation rule'' is not quite accurate, since there is no computation going on. We will still keep using this terminology, so that we have the same terminology as for inductive types. In the cubical type theories mentioned in the introduction we can make these terms reduce judgmentally, making them convenient for working with higher inductive types.
The induction principle for the interval is the following. Suppose given a family $P:I\to\ensuremath{\mathsf{Type}}\xspace$ with elements $x_0:P(0)$ and $x_1:P(1)$. We need to relate $x_0$ and $x_1$ in some way, but we cannot ask that they are equal, since they live in different types. Instead, we require a pathover $p:x_0=^P_{\ensuremath{\mathsf{seg}}\xspace}x_1$. In this case we get a dependent map
$\ind{I}(x_0,x_1,p):(i:I)\to P(i)$ with computation rules on points
$$\ind{I}(x_0,x_1,p,0)\equiv x_0\qquad\text{and}\qquad\ind{I}(x_0,x_1,p,1)\equiv x_1.$$
For the computation rule on paths, we need to use $\apd$ to apply the induction principle to $\ensuremath{\mathsf{seg}}\xspace$, and we get
$$\apd_{\ind{I}(x_0,x_1,p)}(\ensuremath{\mathsf{seg}}\xspace)=p.$$
A more interesting example of a higher inductive type is the \emph{(graph) quotient} which we will call a \emph{quotient} in this dissertation. Given $A : \U$ and $R : A \to A \to \U$, the quotient is the following higher inductive type.
\begin{inductive}
\texttt{HIT} $\ensuremath{\operatorname{\mathsf{quotient}}}_A(R) \ensuremath{\vcentcolon=}$ \\
$\bullet\ i : A \to \ensuremath{\operatorname{\mathsf{quotient}}}_A(R)$; \\
$\bullet\ \mathsf{glue} : (a\ a' : A) \to R(a,a')\to i(a) = i(a')$.
\end{inductive}
We will sometimes use the notation $[{-}]_0$ for $i$ and $[{-}]_1$ for $\mathsf{glue}$.
A very similar higher inductive type is the \emph{homotopy pushout}, or \emph{pushout} for short. Given two maps $f:A\to B$ and $g:A\to C$, their pushout is the following HIT.
\begin{inductive}
\texttt{HIT} $\ensuremath{\operatorname{\mathsf{pushout}}}(f,g) \ensuremath{\vcentcolon=}$ \\
$\bullet\ \ensuremath\inlsym\xspace : B \to \ensuremath{\operatorname{\mathsf{pushout}}}(f,g)$ \\
$\bullet\ \ensuremath\inrsym\xspace : C \to \ensuremath{\operatorname{\mathsf{pushout}}}(f,g)$ \\
$\bullet\ \mathsf{glue} : (a : A) \to \ensuremath\inlsym\xspace(f(a)) = \ensuremath\inrsym\xspace(g(a))$
\end{inductive}
We denote $\ensuremath{\operatorname{\mathsf{pushout}}}(f,g)$ by $B+_AC$ if $f$ and $g$ are clear from the context. In this section, we will define other higher inductive types in terms of the pushout. However, we could also start with the quotient, by the following lemma.
\begin{lem}
The pushout and quotient are interdefinable in MLTT.
\end{lem}
\begin{proof}
We will only give the definitions of the pushout and the quotient in terms of the other.
Showing that these definitions are correct is easy, and we omit it here.
If we have quotients, we can define the pushout of $f: A \to B$ and $g:B\to C$ as the quotient of
$B+C$ under the relation $R:B+C\to B+C\to\U$, which is inductively generated by
$\ensuremath{\operatorname{\mathsf{mk}}} : (a : A) \to R (f(a),g(a))$.
On the other hand, if we have pushouts, we can define the quotient of $A$ under $R$ as
follows. Let $T\ensuremath{\vcentcolon\equiv}(a\ a' : A)\times R(a,a')$ be the total space of $R$. Then the quotient of
$A$ under $R$ is the pushout of $f\ensuremath{\vcentcolon\equiv}\ensuremath{\mathsf{pair}}\xspace{\pi_1}{\pi_2} : T+T \to A$ and
$g\ensuremath{\vcentcolon\equiv}\ensuremath{\mathsf{pair}}\xspace\ensuremath{\operatorname{\mathsf{id}}}\id:T+T\to T$.
\end{proof}
Many higher inductive types can be defined in terms of the homotopy pushout (or equivalently, the quotient):
\begin{itemize}
\item The \emph{cofiber} of a map $f : A \to B$ is defined as $C_f\ensuremath{\vcentcolon\equiv} B+_A\ensuremath{\mathbf{1}}\xspace$.
The maps are $f$ and $!$.
\item The \emph{suspension} $\Sigma A$ of type $A$ is defined as $\Sigma A \ensuremath{\vcentcolon\equiv}
\ensuremath{\mathbf{1}}\xspace+_A\ensuremath{\mathbf{1}}\xspace$, i.e. as the cofiber of the map $A \to \ensuremath{\mathbf{1}}\xspace$. The points are
called $\mathsf{N}$ and $\mathsf{S}$ and $\mathsf{glue}$ is called $\mathsf{merid}$.
\item The \emph{wedge sum} of a family of pointed types $A : I \to \U^*$ is defined as the
cofiber of the map $I \to (i : I) \times A(i)$, which sends $i$ to the pair
$(i,\ensuremath{\operatorname{\mathsf{pt}}}_{A(i)})$. The binary wedge $A\vee B$ of two pointed types $A\ B:\U^*$ can
equivalently be described as the pushout of $A+_\ensuremath{\mathbf{1}}\xspace B$ where the maps come
from the basepoints of $A$ and $B$.
\item The \emph{smash product} $A\wedge B$ of $A$ and $B$ can be defined as the cofiber
of the map $A\vee B\to A\times B$, which sends $\ensuremath\inlsym\xspace(a)$ to $(a,b_0)$ and $\ensuremath\inrsym\xspace(b)$
to $(a_0,b)$ and $\mathsf{glue}(\star)$ to $\ensuremath{\operatorname{\mathsf{refl}}}_{(a_0,b_0)}$. We will discuss the smash product in~\cref{sec:smash-product}.
\item The \emph{$n$-sphere} $\S^n$ is defined inductively for $n\geq0$: $\S^0\ensuremath{\vcentcolon\equiv}\ensuremath{\mathbf{2}}\xspace$ and
$\S^{n+1}\ensuremath{\vcentcolon\equiv}\Sigma\S^n$. The $n$-sphere is pointed with point $\mathsf{N}$ for $n\geq1$ and with ${0_{\bool}}$ for $n=0$.
We could also start counting at $n=-1$, defining $\S^{-1}=\ensuremath{\mathbf{0}}\xspace$, but we often only want to consider the pointed spheres.
\end{itemize}
Another higher inductive type that we will study is the \emph{sequential colimit} or \emph{colimit} for short. This is the following HIT for $A:\textbf{N}\to\ensuremath{\mathsf{Type}}\xspace$ and $f:(n:\textbf{N})\to A(n)\to A(n+1)$:
\begin{inductive}
\texttt{HIT} $\ensuremath{\operatorname{\mathsf{colim}}}(A,f) \ensuremath{\vcentcolon=}$ \\
$\bullet\ \iota : (n : \textbf{N}) \to A(n) \to \ensuremath{\operatorname{\mathsf{colim}}}(A,f)$; \\
$\bullet\ \kappa : (n : \textbf{N}) \to (a : A(n)) \to i_{n+1}(f_n(a))=i_n(a)$.
\end{inductive}
We can define $\ensuremath{\operatorname{\mathsf{colim}}}(A,f)$ using quotients, namely as $\ensuremath{\operatorname{\mathsf{quotient}}}(B,R)$ where
$B=(n : \textbf{N})\times A(n)$ is the total space of $A$ and $R:B\to B\to\U$ is inductively generated by
$\ensuremath{\operatorname{\mathsf{mk}}}:(n : \textbf{N})\to(a : A(n))\to R(f_n(a),a)$. We will discuss the colimit more in \autoref{sec:colimits}
We will use the following properties of these higher inductive types. For the proof we refer to~\cite[Chapter 8]{hottbook}
\begin{lem}\mbox{}
\begin{itemize}
\item If $A$ is $n$-connected, then $\Sigma A$ is $(n+1)$-connected.
\item The suspension is left-adjoint to the loop space: $\Sigma\dashv\Omega$. That means that for any two pointed types $A$ and $B$ there is a pointed equivalence
$$(\Sigma A\to^* B)\simeq^*(A\to^*\Omega B)$$
that is natural in $A$ and $B$.
\item We have the following equivalence: $\Omega \S^1\simeq \ensuremath{\mathbb{Z}}\xspace$. Therefore $\S^1$ is1-truncated and $\pi_1(\S^1)\simeq \ensuremath{\mathbb{Z}}\xspace$.
\end{itemize}
\end{lem}
In particular, by the above lemma we know that $\S^n$ is $(n-1)$-connected, and hence that $\pi_k(\S^n)$ is trivial for $k<n$.
Another higher inductive type is the \emph{torus}, which is the following higher inductive type
\begin{inductive}
\texttt{HIT} $T^2 \ensuremath{\vcentcolon=}$ \\
$\bullet\ \star : T^2$; \\
$\bullet\ \ell_1\ \ell_2 : \star = \star$;\\
$\bullet\ \ell_1 \cdot \ell_2 = \ell_2 \cdot \ell_1$.
\end{inductive}
The last constructor of the torus is a 2-path constructor. In general, HITs can have as constructor any higher path. We say that a HIT is an \emph{$n$-HIT} if its highest path constructor has dimension $n$. So the torus is a 2-HIT and all the other HITs we have seen are 1-HITs.
Higher inductive types can also have \emph{recursive} constructors. If a higher inductive type has at least one recursive constructor, we will call it a recursive HIT. For example, we can encode the propositional truncation as a HIT with a recursive path constructor:
\begin{inductive}
\texttt{HIT} $\|A\| \ensuremath{\vcentcolon=}$ \\
$\bullet\ |{-}|:A\to\|A\|$; \\
$\bullet\ (x_1\ x_2 : \|A\|) \to x_1 = x_2$.
\end{inductive}
Higher truncations can also be encoded using HITs~\cite[Section 7.3]{hottbook}.
\section{Lean}\label{sec:lean}
Lean~\cite{moura2015lean} is an interactive theorem prover that is mainly developed at Microsoft
Research and Carnegie Mellon University.\footnote{The contents of this section are based on~\cite{vandoorn2017leanhott}, which was written with Jakob von Raumer and Ulrik Buchholtz.}
The project was started in 2013 by Leonardo de Moura to
bridge the gap between interactive theorem proving and automated theorem proving. Lean is an
open-source program released under the Apache License 2.0.
In its short history, Lean has undergone several major changes. The second version (Lean~2) supports
two kernel modes. The standard mode is for proof irrelevant reasoning, in which "Prop", the bottom
universe, contains types whose objects are considered to be judgmentally equal. This is incompatible
with homotopy type theory, so there is a second HoTT mode without "Prop". In 2016, the third major
version of Lean (Lean~3) was released~\cite{ebner2017metaprogramming}.
In this version, many components of Lean have been rewritten. Of note, the unification procedure
has been restricted, since the full higher-order unification that is available in Lean~2 can lead
to timeouts and error messages that are unrelated to the actual mistakes. Due to certain design
decisions, such as proof erasure in the virtual machine and a function definition package that
requires axiom K~\cite{goguen2006eliminating}, the homotopy type theory mode is currently not natively
supported in Lean~3. However, a trick found by Gabriel Ebner allows us to build a homotopy type theory library in Lean~3.
In this library, we do not use singleton elimination, which is the feature of "Prop" that is inconsistent with univalence.
Singleton elimination is the property that some "Prop"-valued inductive types can eliminate to all universe levels.
Gabriel Ebner also wrote a piece of code that no definition in this library uses singleton elimination in its definition.
Porting the HoTT library from Lean~2 to Lean~3 is a lot of work, because of the changes in the elaborator and in the syntax.
All major results in this dissertation are only formalized in Lean~2 and not yet in Lean~3.
The HoTT~3 library can be found at \url{https://github.com/gebner/hott3}.
The HoTT kernel of Lean~2 provides the following primitive notions:
\begin{itemize}
\item \emph{Type universes} "Type.{u} : Type.{u + 1}" for each universe level $u \in \mathbb{N}$.
In Lean, this chain of universes is non-cumulative, and all universes are predicative.
\item \emph{Function types} "A → B : Type.{max u v}" for types "A : Type.{u}" and "B : Type.{v}" as well as
\emph{dependent function types} "Πa, B a : Type.{max u v}" for each type "A : Type.{u}" and type family
"B : A → Type.{v}". These come with the usual "β" and "η" rules.
\item \emph{inductive types} and \emph{inductive type families}, as proposed by
Peter Dybjer~\cite{dybjer1994inductive}.
Every inductive definition adds its constructors and dependent recursors to the environment.
Pattern matching is \emph{not} part of the kernel
\item two kinds of \emph{higher inductive types}:
"n"-truncation and (typal) quotients.
\end{itemize}
Outside the kernel, Lean's elaborator uses backtracking search to infer implicit information. It
does the following simultaneously.
\begin{itemize}
\item The elaborator fills in \emph{implicit arguments} that can be inferred from the context,
such as the type of the term to be constructed and the given explicit arguments. Users mark
implicit arguments with curly braces. For example, the type of equality is
"eq : Π{A : Type}, A → A → Type", which allows the user to write "eq a₁ a₂" or "a₁ = a₂" instead
of "@eq A a₁ a₂". The symbol "@" allows the user to fill in implicit arguments explicitly. The
elaborator supports both first-order unification and higher-order unification.
\item We can mark functions as \emph{coercions}, which are then ``silently'' applied when needed.
For example, we have the type of equivalences "A ≃ B", which is a structure consisting of a function
"A → B" with a proof that the function is an equivalence. The map "(A ≃ B) → (A → B)" is marked as
a coercion. This means that we can write "f a" for "f : A ≃ B" and "a : A", and the coercion is
inserted automatically.
\item Lean was designed with \emph{type classes} in mind, which can provide canonical inhabitants of
certain types. This is especially useful for algebraic structures and for type properties like
truncatedness and connectedness. Type class instances can refer to other type classes, so that we
can chain them together. This makes it possible for Lean to automatically infer why types are
"n"-truncated if our reasoning requires this, for example when we are eliminating out of a
truncated type. For example we show that the type of functors between categories "C" and "D" is
equivalent to an iterated sigma type.
\begin{lstlisting}[gobble=2]
(Σ (F₀ : C → D) (F₁ : Π {a b}, hom a b → hom (F₀ a) (F₀ b)),
(Π (a), F₁ (ID a) = ID (F₀ a)) ×
(Π {a b c} (g : hom b c) (f : hom a b),
F₁ (g ∘ f) = F₁ g ∘ F₁ f)) ≃ functor C D
\end{lstlisting}
Note the use of coercions here: "F₀ : C → D" really means a function from the objects of "C" to
the objects of "D". From this equivalence, Lean's type class inference can automatically infer
that "functor C D" is a set if the objects of "D" form a set. Type class inference will repeatedly
apply the rules when sigma-types and pi-types are sets, and use the facts that hom-sets are sets
and that equalities in sets are sets (in total 20 rules are applied for this example).
\item Instead of giving constructions by explicit terms, we can also make use of
Lean's \emph{tactics}, which give us an alternative way to construct terms step by
step. This is especially useful if the proof term is large, or if the elaboration relies heavily on
higher-order unification.
\item We can define custom syntax, including syntax with binding.
In the following example we declare two custom notations.
\begin{lstlisting}
infix ⬝ := concat
notation `Σ` binders `, ` r:(scoped P, sigma P) := r
\end{lstlisting}
The first line allows us to write "p ⬝ q" for path concatenation "concat p q". The second line
allows us to write "Σ x, P x" instead of "sigma P". This notation can also be chained:
"Σ (A : Type) (a : A), a = a" means "sigma (λ(A : Type), sigma (λ(a : A), a = a))".
\end{itemize}
All main results in this dissertation have been formalized in Lean. Some corollaries or examples have not been formalized, in which case we will explicitly mention this.
The formalizations are separated in two Github repositories: the Lean-HoTT library\footnote{\url{https://github.com/leanprover/lean2/blob/master/hott/hott.md}}
and the ``spectral'' repository, which was originally a repository to formalize spectral sequences, but now also contain many other results in synthetic homotopy theory.\footnote{\url{https://github.com/cmu-phil/Spectral/}}
Below is a table with the locations of the formal results in the libraries.
\begin{xcenter}
{\footnotesize
\begin{tabular}{|c|c|c|} \hline
\textbf{Theorem} & \textbf{File} & \textbf{Name} \\ \hline
\autoref{thm:main} & \texttt{hott/hit/prop\_trunc.hlean} & \texttt{ptrunc\_equiv\_trunc} \\
\autoref{thm:simple-two-quotient-def} & \texttt{hott/hit/two\_quotient.hlean} & \texttt{simple\_two\_quotient.rec} \\
\autoref{thm:colim_sm} & \texttt{Spectral/colimit/seq\_colim.hlean} & \texttt{sigma\_seq\_colim\_over\_equiv} \\
\autoref{cor:eq_colim} & \texttt{Spectral/colimit/seq\_colim.hlean} & \texttt{seq\_colim\_eq\_equiv} \\
\autoref{thm:les-homotopy} & \texttt{hott/homotopy/LES\_of\_homotopy\_groups.hlean} & \texttt{is\_exact\_LES\_of\_homotopy\_groups} \\
\autoref{cor:homotopy-group-spheres-1} & \texttt{hott/homotopy/sphere2.hlean} & \texttt{$\pi$2S2} and \texttt{$\pi$nS3\_eq\_$\pi$nS2} \\
\autoref{cor:homotopy-group-spheres-2} & \texttt{hott/homotopy/sphere2.hlean} & \texttt{$\pi$nSn} and \texttt{$\pi$3S2} \\
\autoref{thm:EM-equiv-categories} & \texttt{Spectral/homotopy/EM.hlean} & \texttt{AbGrp\_equivalence\_cptruncconntype${}^\prime$} \\
\autoref{thm:smash-adjoint} & \texttt{Spectral/homotopy/smash\_adjoint.hlean} & \texttt{smash\_adjoint\_pmap} \\
\autoref{thm:exact-couple-convergence} & \texttt{Spectral/algebra/spectral\_sequence.hlean} & \texttt{is\_built\_from\_infpage} \\
\autoref{thm:spectral-sequence-spectrum} & \texttt{Spectral/algebra/spectral\_sequence.hlean} & \texttt{converges\_to\_sequence} \\
\autoref{thm:atiyah-hirzebruch-reduced} & \texttt{Spectral/cohomology/serre.hlean} & \texttt{atiyah\_hirzebruch\_convergence} \\
\autoref{thm:serre-spectral-sequence} & \texttt{Spectral/cohomology/serre.hlean} & \texttt{serre\_convergence} \\
\hline
\end{tabular}
}
\end{xcenter}
\chapter{Higher Inductive Types}\label{cha:high-induct-types}
In this chapter we will study properties of Higher Inductive Types (HITs), which
we introduced in \Cref{sec:high-induct-types}. There is no uniformly accepted
scheme of which HITs are allowed, and the semantics of HITs is a topic of
current research. There are semantic interpretations of a large class of Higher
Inductive Types~\cite{lumsdaine2017HITsemantics}, but there are still open
questions. Firstly, a general scheme for higher inductive types is unknown,
although~\cite{altenkirch2016qiits} is a step in the right direction. Secondly,
it is unknown whether universes can be closed under higher inductive types.
This is unknown even in the case for homotopy pushouts. In this chapter, we do
not study the semantics of higher inductive types. Instead, we will work
internally in a type theory that has some specific HITs, and construct other
HITs from the ones we started with.
In particular, we are interested in the case where we start with the quotient, or equivalently, the homotopy pushout.
One HIT from \Cref{sec:high-induct-types} that we have not yet defined using
quotients is the $n$-truncation. In \Cref{sec:prop-trunc} we will define the
propositional truncation using quotients. A construction of the $n$-truncations is given by the join
construction~\cite{rijke2017join}. This shows that we can define certain
recursive HITs using quotients. We will make a start on defining a bigger class of recursive HITs using quotients in \cref{sec:colimits}.
Another class of HITs we want to construct is HITs with higher path constructors. We construct
nonrecursive 2-HITs in \Cref{sec:non-recursive-2}, using a method very similar to the hubs and
spokes method~\cite[Section 6.7]{hottbook}.
One might wonder after these examples whether all HITs can be reduced to
quotients. This turns out to be false. In~\cite[Section
9]{lumsdaine2017HITsemantics} the authors describe a specific recursive 1-HITs
that cannot be reduced to quotients. Still, it is worthwhile to see which
higher inductive types can be constructed from quotients, for example if one is
interested in a model of HoTT with homotopy pushouts, but without the extra
structure to model all HITs.
\section{Propositional Truncation}\label{sec:prop-trunc}
In this section we will construct the propositional truncation from quotients.\footnote{The contents of this section have been published in~\cite{vandoorn2016proptrunc}. However, \autoref{cor:prop-trunc-univ-set} is new.}
Given a type $A$, define $\{A\}$ as the quotient of $A$ by the indiscrete relation
$R\ensuremath{\vcentcolon\equiv}\l(a\ a':A),\ensuremath{\mathbf{1}}\xspace$. We will call the type $\{A\}$ the \emph{one-step truncation}, since
repeating it will give the propositional truncation. We will denote its point constructor by $f : A \to \{A\}$ and its path constructor by $e : (x\ y : A) \to f(x) = f(y)$. We call a function $g : A \to B$ \emph{weakly constant} if $(x\ y : A) \to g(x) = g(y)$ is inhabited. Note that maps $\{A\}\to B$ correspond exactly to weakly constant maps $A\to B$.
Given a type $A$, we define a sequence
$\{A\}_{-}:\textbf{N}\to\U$ by
\begin{align}
\begin{aligned}
\{A\}_0&:\equiv A\\ \{A\}_{n+1}&:\equiv \{\{A\}_n\}
\end{aligned}
\label{eq:An}
\end{align}
We have map $f_n:\equiv f : \{A\}_n\to \{A\}_{n+1}$, which is the constructor of the one-step
truncation. This gives the sequence
\begin{equation}
A\xrightarrow{f}\{A\}\xrightarrow{f}\{\{A\}\}\xrightarrow{f}\cdots \label{eq:prop-sequence}
\end{equation}
We define $\{A\}_\infty=\ensuremath{\operatorname{\mathsf{colim}}}(\{A\}_{-},f_{-})$. We will prove that $\{A\}_\infty$ is the
propositional truncation of $A$, in the sense that the construction $A\mapsto \{A\}_\infty$ has the
same formation, introduction, elimination and computation rules for the propositional truncation.
We have already shown the formation rule of the propositional truncation (note that $\{A\}_\infty$
lives in the same universe as $A$).
We also easily get the point constructor of the propositional truncation, because that is just the
map $i_0:A\to \{A\}_\infty$. The path constructor $(x, y : \{A\}_\infty) \to x = y$, i.e. the
statement that $\{A\}_\infty$ is a mere proposition, is harder to define. We will postpone this
until after we have defined the elimination and computation rules.
The elimination principle --- or induction principle --- for the propositional truncation is the
following statement. Suppose we are given a family of propositions $P : \{A\}_\infty \to \ensuremath{\mathsf{Prop}}\xspace$ with
a section $h : (a : A) \to P(i_0(a))$. We then have to construct a map
$k : (x : \{A\}_\infty) \to P(x)$. To construct $k$, take an $x : \{A\}_\infty$. Since $x$ is in a
colimit, we can apply induction on $x$. Notice that we construct an element in $P(x)$, which is a
mere proposition, so we only have to define $k$ on the point constructors. This means that we can
assume that $x\equiv i_n(a)$ for some $n : \textbf{N}$ and $a:\{A\}_n$. Now we apply induction on $n$.
If $n\equiv0$, then we can choose $k(i_0(a)):\equiv h(a):P(i_0(a))$.
If $n\equiv \ell+1$ for some $\ell:\textbf{N}$, we know that $a:\{\{A\}_\ell\}$, so we can induct on
$a$. The path constructor of this induction is again automatic. For the point constructor, we can
assume that $a\equiv f(b)$. In this case we need to define $k(i_{\ell+1}(f(b))) :
P(i_{\ell+1}(f(b)))$. By induction hypothesis, we have an element $y : P(i_\ell(b))$. Now we can
transport $x$ along the equality $(g_\ell(b))^{-1} : i_\ell(b)=i_{\ell+1}(f(b))$. This gives the
desired element in $P(i_{\ell+1}(f(b)))$.
We can write the proof in pattern matching notation:
\begin{itemize}
\item $k(i_0(a)):\equiv h(a)$
\item $k(i_{n+1}(f_n(a))):\equiv(g_n(b))_*^{-1}(k(i_n(b)))$
\end{itemize}
The definition $k\ (i_{0}\ a) :\equiv h\ a$ is also the judgmental computation rule for the point constructors of the propositional truncation.
For the remainder of this section we will prove that $\{A\}_\infty$ is a mere proposition. We will
need the following two lemmas.
\begin{lem}\label{lem:pieq}
Let $X$ be a type with $x : X$. Then the type $(y:X) \to x=y$ is a mere proposition.
\end{lem}
\begin{proof}
To prove that $(y:X) \to x=y$ is a mere proposition, we assume that it is inhabited and show that
it is contractible. Let $f : (y:X) \to x=y$. From this, we conclude that $X$ is contractible
with center $x$. Now given any $g : (y:X) \to x=y$, we know that $f$ and $g$ are pointwise
equal, because their codomain is contractible. By function extensionality we conclude that $f=g$,
finishing the proof.
\end{proof}
\begin{lem}\label{lem:apconstant}
If $g:X\to Y$ is weakly constant, then for every $x, x' : X$, the function $\text{ap}_g:x=x'\to
g(x)=g(x')$ is weakly constant. That is, $\ap gp=\ap gq$ for all $p,q:x=x'$.
\end{lem}
\begin{proof}
Let $q : (x, y : X) \to g(x)=g(y)$ be the proof that $g$ is weakly constant, and fix $x : X$. We
first prove that for all $y : X$ and $p : x = y$ we have
\begin{equation}\label{eq:apconstanteq}
\ap{g}{p} = q(x,x)^{-1}\ \cdot\ q(x,y).
\end{equation}
This follows from path induction, because if $p$ is reflexivity, then $\ap{g}{\ensuremath{\operatorname{\mathsf{refl}}}_x} \equiv
\ensuremath{\operatorname{\mathsf{refl}}}_{g(x)} = q(x,x)^{-1}\ \cdot\ q(x,x)$. The right hand side of \eqref{eq:apconstanteq} does not
depend on $p$, hence $\text{ap}_g$ is weakly constant.
\end{proof}
To prove that $\{A\}_\infty$ is a mere proposition, we need to show
$(x, y : \{A\}_\infty) \to x=y$. Since $(y : \{A\}_\infty) \to x=y$ is a mere proposition, we can use the
induction principle for the propositional truncation on $x$, which we have just proven for
$\{A\}_\infty$. This means we only have to show that for all $a : A$ we have
$(y : \{A\}_\infty) \to i_0(a)=y$. We do not know that $i_0(a)=y$ is a mere proposition,\footnote{Of
course, we do know that it is a mere proposition after we have finished the proof that
$\{A\}_\infty$ is a mere proposition.} so we will just use the regular induction principle for
colimits on $y$. We then have to construct two inhabitants of the following two types:
\begin{enumerate}
\item For the point constructor we need $p(a,b) : i_0(a) = i_n(b)$ for all $a : A$ and $b :
\{A\}_n$.
\item We have to show that $p$ respects path constructors:
\begin{equation}
p(a,f(b))\ \cdot\ g(b) = p(a,b).\label{eq:coh}
\end{equation}
\end{enumerate}
We have a map $f^n:A \to \{A\}_n$ defined by induction on $n$, which repeatedly applies $f$. We also
have a path $g^n(a): i_n(f^n(a)) = i_0(a)$, which is a concatenation of instances of $g$.
\begin{figure}\begin{center}
\begin{tikzpicture}[node distance=3cm,font=\footnotesize,
thick,main node/.style={font=\large}]
\ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$a$] (a) at (0,0) {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$f^n(a)$]
(fa) [above of=a] {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$f^{n+1}(a)$] (ffa) [above of=fa]
{$\bullet$}; \tikzset{node distance=2.5cm}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$b$] (b) [right
of=fa] {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$f(b)$] (fb) [right of=ffa] {$\bullet$};
\tikzset{node distance=2cm}; \ensuremath{\operatorname{\mathsf{node}}} (Am1) [right of=fb] {$\{A\}_{n + 1}$}; \ensuremath{\operatorname{\mathsf{node}}} (Am) [right
of=b] {$\{A\}_n$}; \ensuremath{\operatorname{\mathsf{node}}} (An) [right of=a] {$A$}; \draw ($(ffa)!0.4!(fb)$) ellipse (2.7cm
and 1cm); \draw ($(fa)!0.4!(b)$) ellipse (2.4cm and 1cm); \draw (a) circle (1cm); \path (fa) edge node [left] {$g^n$} (a) edge node [right]
{$g$} (ffa) (a) edge [bend left=60] node [above left] {$g^{n+1}$} (ffa) (fb) edge [bend
right=20] node [above] {$e$} (ffa) edge node [left] {$g$} (b);
\end{tikzpicture}
\caption{The definition of $p$. The applications of $i$ and the arguments of the paths are
implicit.}
\label{fig:defp}
\end{center}\end{figure}
We can now define $p(a,b)$ as displayed in \Cref{fig:defp}, which is the concatenation
\begin{align*}
i_0(a) &= i_{n+1}(f^{n+1}(a)) &&\text{(using $g^{n+1}$)}\\ &\equiv i_{n+1}(f(f^n(a)))\\ &=
i_{n+1}(f(b)) &&\text{(using $e$)}\\ &= i_n(b) &&\text{(using $g$)}
\end{align*}
Note that by definition $g^{n+1}(a) \equiv g(f^n(a))\ \cdot\ g^n(a)$, so the triangle on the left of
\Cref{fig:defp} is a definitional equality.
\begin{figure}\begin{center}
\begin{tikzpicture}[node distance=3cm,font=\footnotesize,
thick,main node/.style={font=\large}]
\ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$a$] (a) at (0,0) {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$f^n(a)$]
(fa) [above of=a] {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$f^{n+1}(a)$] (ffa) [above of=fa]
{$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$f^{n+2}(a)$] (fffa) [above of=ffa]{$\bullet$};
\tikzset{node distance=2.5cm}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$b$] (b) [right of=fa]
{$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$f(b)$] (fb) [right of=ffa] {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$f(f(b))$] (ffb) [right of=fffa]{$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}} (Am2)
[right of=ffb]{$\{A\}_{n + 2}$}; \ensuremath{\operatorname{\mathsf{node}}} (Am1) [right of=fb] {$\{A\}_{n + 1}$};
\ensuremath{\operatorname{\mathsf{node}}} (Am) [right of=b] {$\{A\}_n$}; \ensuremath{\operatorname{\mathsf{node}}} (An) [right of=a] {$A$}; \draw
($(fffa)!0.45!(ffb)$) ellipse (2.9cm and 1cm); \draw ($(ffa)!0.4!(fb)$) ellipse
(2.7cm and 1cm); \draw ($(fa)!0.4!(b)$) ellipse (2.4cm and 1cm); \draw (a)
circle (1cm); \path (fa) edge node
[left] {$g^n$} (a) edge node [right] {$g$} (ffa) (fb) edge [bend left=20] node
[below] {$e$} (ffa) edge node [left] {$g$} (b) edge node [left] {$g$} (ffb)
(fffa) edge node [right] {$g$} (ffa) edge [bend left=20] node [above] {$e$}
(ffb);
\end{tikzpicture}
\caption{The coherence condition for $p$. The applications of $i$ and the arguments of the paths are
implicit.}
\label{fig:cohp}
\end{center}\end{figure}
Now we have to show that this definition of $p$ respects the path constructor of the colimit, which
means that we need to show~\eqref{eq:coh}. This is displayed in \Cref{fig:cohp}. We only need to
fill the square in \Cref{fig:cohp}. To do this, we first need to generalize the statement,
because we want to apply path induction. Note that if we give the applications of $i$ explicitly,
the bottom and the top of this square are $$\ap i{e(f^{n+1}(a),f(b))}$$ and $$\ap
i{e(f^{n+2}(a),f(f(b)))},$$ respectively. This means we can apply the following lemma to prove this
equality.
\begin{figure}\begin{center}
\begin{tikzpicture}[node distance=3cm,font=\footnotesize,
thick,main node/.style={font=\large,
}]
\ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$i(x)$] (x) at (0,0) {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=left:$i(f(x))$]
(fx) [above of=x] {$\bullet$}; \tikzset{node distance=2.5cm}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$i(y)$]
(y) [right of=x] {$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}}[main node,label=right:$i(f(y))$] (fy) [right of=fx]
{$\bullet$}; \ensuremath{\operatorname{\mathsf{node}}} (Ak1) [right of=fy] {$\{A\}_{n + 1}$}; \ensuremath{\operatorname{\mathsf{node}}} (Ak) [right of=y]
{$\{A\}_n$}; \draw ($(fx)!0.5!(fy)$) ellipse (2.8cm and 1cm); \draw ($(x)!0.5!(y)$) ellipse
(2.6cm and 1cm); \path (x) edge [bend right=20]
node [below] {$\ap ip$} (y) edge node [right] {$g$} (fx) (fy) edge [bend right=20] node
[above] {$\ap i{p'}$} (fx) edge [bend left=20] node [below] {$\ap i{\ap fp}$} (fx) edge node
[left] {$g$} (y);
\end{tikzpicture}
\caption{The situation in \Cref{lem:cohplemma}.}
\label{fig:cohplemma}
\end{center}\end{figure}
\begin{lem}\label{lem:cohplemma}
Suppose we are given $x\ y : \{A\}_n$, $p:x=y$ and $p' : f(x) = f(y)$. Then we can fill the outer
square in \Cref{fig:cohplemma}, i.e.
$$g(x)\ \cdot\ \ap ip = \ap i{p'}\ \cdot\ g(y).$$
\end{lem}
\begin{proof}
We can fill the inner square of the diagram by induction on $p$, because if $p$ is reflexivity, then
the inner square reduces to
$$g(x)\ \cdot\ \ensuremath{\operatorname{\mathsf{refl}}}_{i(x)}=\ensuremath{\operatorname{\mathsf{refl}}}_{i(f(x))}\ \cdot\ g(x).$$ To show that the two paths in the top
are equal, first note that $i_{k} : \{A\}_{k} \to \{A\}_\infty$ is weakly constant. To see this,
look at \Cref{fig:defp}. The path from $f^n(a)$ to $b$ in that figure gives a proof of
$i_n(f^n(a))=i_n(b)$ that does not use the form of $f^n(a)$, so we also have $i_k(u)=i_k(v)$ for
$u,v :\{A\}_k$. Since $i_{n+1}$ is weakly constant, by \Cref{lem:apconstant} the
function $$\text{ap}_{i_{n+1}} : f(x)=f(y)\to i_{n+1}(f(x))=i_{n+1}(f(y))$$ is also weakly
constant. This means that the two paths in the top are equal, proving the Lemma.
\end{proof}
We have now given the proof of the following theorem:
\begin{thm}\label{thm:main}
The map $A\mapsto \{A\}_\infty$ satisfies all the properties of the propositional truncation $\|{-}\|$, including the universe level and judgmental computation rule on point constructors.
\end{thm}
We will mention two corollaries of this result. An alternate proof of the first one is given in~\cite{kraus2014anonymousexistence}.
\begin{cor}\label{c:hstable}
Given a weakly constant function $h : A \to A$, there is a function $\|A\|\to A$.
\end{cor}
\begin{proof}
The weakly constant function $h$ gives a function $\tilde h : \{A\}\to A$. The HIT $\{{-}\}$ is
functorial (just like all other HITs), so by its functorial action we get a map $\{\tilde
h\}:\{\{A\}\}\to\{A\}$, which we can compose with $\tilde h$ to get a map $\{\{A\}\}\to A$. By
induction on $n$ we get a map $k_n : \{A\}_n \to A$. Formally, we define
\begin{align*}
k_0(a)&:\equiv a\\ k_{n+1}(x)&:\equiv \tilde h(\{k_n\}(x))
\end{align*}
However, this sequence of maps does not form a cocone, because the triangles do not commute. (For
example for the first triangle we have to show $h(a)=a$ for all $a$.) But we can easily modify the
definition by postcomposing with $h$. Define $h_n:\equiv h\circ k_n : \{A\}_n\to A$. Now we get a
cocone; all triangles commute because $h$ is weakly constant. By the universal property of the sequential colimit we get
a map $\|A\|\to A$.
\end{proof}
We can also construct maps out of the propositional truncation into a set by giving a weakly constant function.
An alternate proof was given in~\cite{kraus2014universalproperty}.
\begin{cor}\label{cor:prop-trunc-univ-set}
Suppose given a weakly constant function $g:A\to B$ where $B$ is a set. Then there is a map $\tilde g:\|A\|\to B$ such that $\tilde g(|a|)=g(a)$.
\end{cor}
\begin{proof}
First note that given any map $h:\{X\}\to B$, we get a map $h':\{\{X\}\}\to B$ such that $h'\o f\sim h$.
Namely, on point constructors we define $h'(f(x))\ensuremath{\vcentcolon\equiv} h(x)$ for $x:\{X\}$.
Now given $x',y':\{X\}$, we want to define $h'$ on $e(x',y')$. We perform induction on both $x$ and $y$.
In the case that both $x$ and $y$ are point constructors, $x'\equiv f(x)$ and $y'\equiv f(y)$ we can define
$$\apfunc{h'}(e(x',y'))\ensuremath{\vcentcolon=}\apfunc{h}(e(x,y)):h'(f(x'))\equiv h(f(x))=h(f(y))\equiv h'(f(y')).$$
In the other three cases, we are constructing a 2-path (or 3-path) in $B$, which is automatically filled because $B$ is a set.
This finishes the construction of $h'$, which satisfies $h'\o f\sim h$ by definition.
Now we can define a cocone $g_n:\{A\}_n\to B$ as follows. $g_0$ and $g_1$ are given by $g$.
We now define $g_{n+2}\ensuremath{\vcentcolon\equiv} g_{n+1}'$. These $g$'s form a cocone because $g_{n+2}'\o f\sim g_{n+1}$.
This gives a map $\tilde g:\|A\|\to B$ such that $\tilde g(|a|)=g(a)$.
\end{proof}
An alternative construction of the propositional truncation using non-recursive HITs has been given in~\cite{kraus2016hits}. All results in this section have been fully formalized.
\section{Non-recursive 2-HITs}\label{sec:non-recursive-2}
We can also define nonrecursive 2-HITs using quotients.\footnote{A summary of this section also appeared in~\cite{vandoorn2017leanhott}.} There are various 2-HITs we would like to construct, such as the torus (as formulated in \autoref{sec:high-induct-types}), groupoid quotients, and Eilenberg-MacLane spaces $K(G,1)$. The construction of 2-HITS uses a method similar to the hubs-and-spokes method described in~\cite[Sect.~6.7]{hottbook}.
The idea behind the hubs-and-spokes method is that for any path $p : x =_A x$ we can define a map $f : \S^1 \to A$
with $\apfunc{f}(\ensuremath{\operatorname{\mathsf{loop}}}) = p$ by circle induction. Then we can prove the equivalence
$$(p = 1) \simeq (x_0 : A) \times (z : \S^1) \to f(z) = x_0.$$
This equivalence informally
states that filling in a loop is the same as adding a new point $x_0$, the \emph{hub}, and
\emph{spokes} $f(z) = x_0$ for every $z : \S^1$, similar to the spokes in a wheel.
This means that in a higher inductive type, we can replace a 2-path constructor $p = 1_x$ by
a new point constructor $x_0 : A$ and a family of 1-path constructors $(z : \S^1) \to f(z) = x_0$. 2-path constructors of the form $p = q$ can be replaced by the equivalent path constructor $p \cdot q^{-1} = 1$.
This construction reduces certain 2-HITs to 1-HITs. However, this reduction is not a quotient, since this family of path constructors refers to other path constructors (in the definition of $f$), which is not allowed in quotients. If we use quotients, we need to take the quotient twice. We first define a quotient with only the 1-paths (and the hubs), and then use another quotient to add the spokes. In this section we will describe this construction of 2-HITs from quotients.
To be more formal, let us first prove a slightly more general version of the above equivalence.
\begin{lem}\label{lem:twopathlemma}
Given a path $p : a =_A a$ and $f : A \to B$, we have an equivalence
$$e:((b_0 : B) \times (z : \S^1) \to f(\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}(a,p,z)) = b_0) \simeq (\apfunc{f}(p) = 1),$$
where $\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}} : (y : P)\to y = y \to \S^1 \to P$ is the nondependent eliminator of the circle $\S^1$.
\end{lem}
\begin{proof}
This follows from the following chain of equivalences.
\begin{align*}
&\mathrel{\hphantom{\simeq}}(b_0 : B) \times (z : \S^1) \to f(\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}(a,p,z)) = b_0 \\
&\simeq (b_0 : B) \times (q : f(a) = b_0) \times q =_{\ensuremath{\operatorname{\mathsf{loop}}}}^{f(\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}(a,p,{-}))=b_0} q\\
&\simeq 1_{f(a)} =_{\ensuremath{\operatorname{\mathsf{loop}}}}^{f(\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}(a,p,{-}))=f(a)} 1_{f(a)}\\
&\simeq \apfunc{f\circ\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}(a,p)}(\ensuremath{\operatorname{\mathsf{loop}}}) = \apfunc{\ensuremath{\mathbf{0}}_{f(a)}}(\ensuremath{\operatorname{\mathsf{loop}}})\\
&\simeq \apfunc{f}(p) = 1.\qedhere
\end{align*}
\end{proof}
More formally, for $A:\ensuremath{\mathsf{Type}}\xspace$ and $R:A\to A\to\ensuremath{\mathsf{Type}}\xspace$ we will define words in $R$ to be the following inductive family of types:
\begin{inductive}
\texttt{inductive} $\ensuremath{\operatorname{\mathsf{words}}}_R : A \to A \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ [{-}] : \{a\ a' : A\} \to R(a,a') \to \ensuremath{\operatorname{\mathsf{words}}}_R(a,a');$\\
$\bullet\ \langle{-}\rangle : \{a\ a' : A\} \to a = a' \to \ensuremath{\operatorname{\mathsf{words}}}_R(a,a');$ \\
$\bullet\ {-}^{-1}:\{a\ a' : A\} \to \ensuremath{\operatorname{\mathsf{words}}}_R(a,a') \to \ensuremath{\operatorname{\mathsf{words}}}_R(a',a);$; \\
$\bullet\ {-} \cdot {-}: \{a_1\ a_2\ a_3 : A\} \to \ensuremath{\operatorname{\mathsf{words}}}_R(a_1,a_2) \to \ensuremath{\operatorname{\mathsf{words}}}_R(a_2,a_3) \to \ensuremath{\operatorname{\mathsf{words}}}_R(a_1,a_3).$
\end{inductive}
A \emph{specification for a \emph(nonrecursive\emph) 2-HIT} consists of a type $A$ and two families
$R : A \to A \to \ensuremath{\mathsf{Type}}\xspace$ and $S : \{a\ a' : A\} \to \ensuremath{\operatorname{\mathsf{words}}}_R(a,a') \to \ensuremath{\operatorname{\mathsf{words}}}_R(a,a') \to \ensuremath{\mathsf{Type}}\xspace$.
Using this, we define the 2-HIT $\ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S)$ with constructors
\begin{inductive}
\texttt{HIT} $\ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S) \ensuremath{\vcentcolon=}$ \\
$\bullet\ [{-}]_0 : A \to \ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S);$\\
$\bullet\ [{-}]_1 : \{a\ a' : A\} \to R(a,a') \to [a]_0=[a']_0;$ \\
$\bullet\ [{-}]_2:\{a\ a' : A\} \to \{t\ t' :\ensuremath{\operatorname{\mathsf{words}}}_R(a,a')\} \to S(t,t') \to
\overline{[t]_1}=\overline{[t']_1}$.
\end{inductive}
where $\overline{[t]_1}$ is the action of $[{-}]_1$ on words in $R$. So if $t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a')$, then $\overline{[t]_1}:[a]_0=[a']_0$ is defined by recursion over $t$. For example, the recursive steps for concatenation of words is
$$\overline{[t_1 \cdot t_2]_1} \ensuremath{\vcentcolon\equiv} \overline{[t_1]_1} \cdot \overline{[t_2]_1}.$$
Before we define $\ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S)$, we first define a special case with only reflexivities on the right hand side of 2-path constructors. This is the following HIT, where $Q$ has type
$\{a : A\} \to \ensuremath{\operatorname{\mathsf{words}}}_R(a,a) \to \ensuremath{\mathsf{Type}}\xspace$.
\begin{inductive}
\texttt{HIT} $\ensuremath{\operatorname{\mathsf{simple-two-quotient}}}(A,R,Q) \ensuremath{\vcentcolon=}$ \\
$\bullet\ [{-}]_0 : A \to \ensuremath{\operatorname{\mathsf{simple-two-quotient}}}(A,R,Q);$\\
$\bullet\ [{-}]_1 : \{a\ a' : A\} \to R(a,a') \to [a]_0=[a']_0;$ \\
$\bullet\ [{-}]_2:\{a : A\} \to \{t :\ensuremath{\operatorname{\mathsf{words}}}_R(a,a)\} \to Q(t) \to
\overline{[t]_1}=1$.
\end{inductive}
To define this, we first define a new type where we add a \emph{hub} to $A$ for every path specified by $Q$.
$$B \ensuremath{\vcentcolon\equiv} A + (a : A) \times (t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a)) \times Q t.$$
Then we quotient this as specified by $R$, to obtain the 1-paths.
$$C \ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{quotient}}}_B(R_B),$$
where the inductive family of types $R_B$ is defined as follows.
\begin{inductive}
\texttt{inductive} $R_B : B \to B \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \{a\ a' : A\} \to R(a, a') \to R_B(\ensuremath\inlsym\xspace a,\ensuremath\inlsym\xspace a').$
\end{inductive}
We now define $D\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{simple-two-quotient}}}(A,R,Q') \ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{quotient}}}_C(R_C)$ where $R_C$ is defined as the following inductive family of types (we write $u_t\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}([\ensuremath\inlsym\xspace a]_0,\overline{[t]_1}):\S^1\to C$)
\begin{inductive}
\texttt{inductive} $R_C : C \to C \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ \{a : A\} \to \{t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a)\} \to (q : Q(t)) \to (x : \S^1) \to R_C(u_t(x),[\ensuremath\inrsym\xspace (a,t,q)]_0).$
\end{inductive}
We will now define the expected constructors, eliminators, and computation rules for this two-quotient.
\begin{thm}\label{thm:simple-two-quotient-def}
The type $D\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{simple-two-quotient}}}(A,R,Q)$ is the HIT as specified above. This means that
\begin{itemize}
\item There is a 0-path constructor $\br{{-}}_0 : A \to D$;
\item There is a 1-path constructor $\br{{-}}_1 : \{a\ a' : A\} \to R(a,a') \to \br{a}_0=\br{a'}_0$;
\item There is a 2-path constructor $\br{{-}}_2 : \{a : A\} \to \{t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a)\} \to Q(t) \to \overline{\br{t}_1}=1$;
\item There is an induction principle that states the following: given a family $P: D \to \ensuremath{\mathsf{Type}}\xspace$ with $s_0 : (a : A) \to P\br{a}_0$ and
$$s_1 : \{a\ a' : A\} \to (r : R(a,a')) \to s_0(a) =_{\br{r}_1}^P s_0(a')$$
$$s_2 : \{a : A\} \to \{t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a)\} \to (q : Q(t)) \to
\overline{s_1}(t)=_{\br{q}_2}1,$$
then $P$ has a section $f : (d : D) \to P(d)$ that computes on the point and 1-path constructors: $f\br{a}_0\equiv s_0(a)$ and $\apd_f\br{r}_1=s_1(r)$.
\item There is a recursion principle that states the following: given $P : \ensuremath{\mathsf{Type}}\xspace$ with $p_0 : A \to P$ and
$$p_1 : \{a\ a' : A\} \to R(a,a') \to p_0(a) = p_0(a')$$
$$p_2 : \{a : A\} \to \{t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a)\} \to Q(t) \to \overline{p_1}(t)=1,$$
then there is a map $g : D \to P$ that computes on the point 1-path and 2-path constructors. This means that $g\br{a}_0\equiv p_0(a)$ for $a : A$, and that there is a path
$\iota_1:\apfunc{g}\br{r}_1=p_1(r)$ for $r : R(a,a')$ and a filler of the following square for $q : Q(t)$.
\begin{center}
\begin{tikzpicture}[thick, node distance=3cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\apfunc{g}\overline{\br{t}_1}$};
\ensuremath{\operatorname{\mathsf{node}}}[right of = tl] (tr) {$1$};
\ensuremath{\operatorname{\mathsf{node}}} (bl) at (0,-2) {$\overline{p_1}(t)$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$1$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node {$\apfunc{\apfunc{g}}\br{q}_2$} (tr)
edge[double equal sign distance] node[left] {$\overline{\iota_1}$} (bl)
(tr) edge[double equal sign distance] node {$1$} (br)
(bl) edge[double equal sign distance] node {$p_2(q)$} (br);
\end{tikzpicture}
\end{center}
\end{itemize}
\end{thm}
\begin{rmk}\mbox{}
\begin{itemize}
\item We do not prove a computation rule for the induction principle on 2-paths. Although we strongly expect this to be true, it will involve an elaborate computation. This computation rule is not necessary to define $\ensuremath{\operatorname{\mathsf{simple-two-quotient}}}(A,R,Q)$ up to equivalence. If we had another type with these exact constructors, eliminators and computation rules, we can prove that it is equivalent to this one. Furthermore, in many examples of two-quotients we will 1-truncate the result, such as for Eilenberg-MacLane spaces and groupoid quotients (see \autoref{sec:eilenb-macl-spac}). After the 1-truncation, the computation rules on the 2-paths are automatic, since these 3-paths can be constructed just from the assumption that the type family is truncated.
\item We do not define the recursion principle as a special case of the induction principle. We can define it is a much simpler way, so that we can compute its action on 2-paths more easily.
\item We use overlines to denote elimination out of the inductive type $\ensuremath{\operatorname{\mathsf{words}}}_R$. The exact type and definition of the overline depends on the type of the object we overline.
For example
$$\overline{\br{{-}}_1}: \ensuremath{\operatorname{\mathsf{words}}}_R(a,a')\to\br{a}_0=\br{a'}_0$$
is defined recursively by path concatenation and path inversion. In contrast
$$\overline{s_1}: (t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a'))\to s_0(a)=_{\overline{\br{t}_1}}^P s_0(a')$$
is defined recursively by pathover concatenation and pathover inversion and
$$\overline{\iota_1}: (t : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a'))\to \apfunc{g}\overline{\br{t}_1}=\overline{p_1}(t)$$ is defined recursively by horizontal concatenation and horizontal inversion and by using the rules
$\apfunc{g}(p \cdot q)=\apfunc{g}(p)\cdot\apfunc{g}(q)$ and $\apfunc{g}(p^{-1})=(\apfunc{g}(p))^{-1}$.
\end{itemize}
\end{rmk}
\begin{proof}
\textbf{Constructors.}\\
We define for $a : A$ the point constructor
$$\br{a}_0\ensuremath{\vcentcolon\equiv} [[\ensuremath\inlsym\xspace a]_0]_0:D$$
and for $r : R(a,a')$ the 1-path constructor
$$\br{r}_1\ensuremath{\vcentcolon\equiv}\apfunc{[{-}]_0}[r]_1:\br{a}_0=\br{a'}_0$$
from the path constructors of $C$.
The 2-path constructor $\br{q}_2:\overline{\br{t}_1}=1$ for $q : Q(t)$ is
defined as the concatenation
$\overline{\br{t}_1}=\apfunc{[{-}]_0}\overline{[t]_1}=1.$ Here the first
equality is by a general lemma about $\ensuremath{\operatorname{\mathsf{words}}}_R$ that states that
$\overline{\apfunc{f}(h(r))}=\apfunc{f}(\overline{h(r)})$. The second path
uses \autoref{lem:twopathlemma} and is defined as
$e([[\ensuremath\inrsym\xspace(a,t,q)]_0]_0,[q,{-}])$, where
$$[q,{-}]:[\ensuremath{\operatorname{\S^1\hspace{-1mm}{.}\mathsf{rec}}}([\ensuremath\inlsym\xspace a]_0,\overline{[t]_1},x)]_0\equiv [u_t(x)]_0=[[\ensuremath\inrsym\xspace(a,t,q)]_0]_0$$
is the path constructor $[q,x]_1$ of $D$.\\ \mbox{} \\
\textbf{Induction Principle.}\\
For the induction principle, suppose given $P$, $s_0$, $s_1$ and $s_2$ as in the theorem statement. We first define $f_0 : (c : C) \to P[c]_0$ by induction on $c : C$.
We define $$f_0[\ensuremath\inlsym\xspace a]\ensuremath{\vcentcolon\equiv} s_0(a)$$ and (denoting $b\ensuremath{\vcentcolon\equiv}\ensuremath{\mathsf{base}}\xspace:\S^1$)
$$f_0[\ensuremath\inrsym\xspace(a,t,q)]\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{transport}}}^P([q,b]_1,s_0(a)).$$
For the path constructor, we need to construct for $r : R(a,a')$ the pathover
$$\apd_{f_0}[r]_1:s_0(a)=_{[r]_1}^{P[{-}]_0}s_0(a').$$
Here we can use $s_1$, and then apply the equivalence
$$y=_{\apfunc{f}(p)}^P y' \simeq y =_p^{P \circ f} y'.$$
Note that this equivalence holds by reflexivity in a cubical type theory. In the remainder of this proof we will denote any occurrence of this and similar by a tilde for readability. So we define
$$\apd_{f_0}[r]_1\ensuremath{\vcentcolon=}\widetilde{s_1}(r).$$
This defines $f_0$, which is $f$ applied to the point constructor of $D$, that is, $f[c]_0\ensuremath{\vcentcolon\equiv} f_0(c)$. Now we need to define for $x:\S^1$ the pathover
$$\apfunc{f}[q,x]_1:f_0(u_t(x))=_{[q,x]_1}^P
([q,b]_1)_*(s_0a).$$
We will fill this pathover by induction to $x$. For $x\equiv b$ we can constructor the resulting pathover by
$$1_{[q,b]_1}:\apfunc{f}[q,b]_1:s_0a=_{[q,b]_1}^P ([q,b]_1)_*(s_0a),$$
where in general $1_p:y=_p^P p_*(y)$ can be easily defined by induction on $p$. When $x$ varies along loop, we need to construct a pathover between two pathovers and this corresponds to the following squareover. The bottom square is a square in $D$, namely the naturality square of $$[(q,{-})]_1:(x : \S^1) \to [u_t(x)]_0=[[\ensuremath\inrsym\xspace(a,t,q)]_0]_0$$ applied to the path $\ensuremath{\operatorname{\mathsf{loop}}}$, and the top square is the squareover we need to fill.
\begin{center}
\begin{tikzpicture}[thick, node distance=3cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$s_0(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[right of = tl] (tr) {$[q,b]_{1*}(s_0(a))$};
\ensuremath{\operatorname{\mathsf{node}}} (bl) at (0,-2) {$s_0(a)$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$[q,b]_{1*}(s_0(a))$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node {$1_{[q,b]_1}$} (tr)
edge[double equal sign distance] node[left]{$\widetilde{\mathsf{apd}}_{f_0\circ u_t}(\ensuremath{\operatorname{\mathsf{loop}}})$} (bl)
(tr) edge[double equal sign distance] node {$\widetilde{\mathsf{apd}}_{\ensuremath{\mathbf{0}}_{[q,b]_{1*}(s_0(a))}}(\ensuremath{\operatorname{\mathsf{loop}}})$} (br)
(bl) edge[double equal sign distance] node (b) {$1_{[q,b]_1}$} (br);
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,-4.5) {$\br{a}_0$};
\ensuremath{\operatorname{\mathsf{node}}}[right of = tl] (tr) {$[[\ensuremath\inrsym\xspace(a,t,q)]_0]_0$};
\ensuremath{\operatorname{\mathsf{node}}} (bl) at (0,-6.5) {$\br{a}_0$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$[[\ensuremath\inrsym\xspace(a,t,q)]_0]_0$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node (t) {$[q,b]_1$} (tr)
edge[double equal sign distance] node[left]{$\apfunc{[u_t({-})]_0}(\ensuremath{\operatorname{\mathsf{loop}}})$} (bl)
(tr) edge[double equal sign distance] node {$\apfunc{\ensuremath{\mathbf{0}}_{[[\ensuremath\inrsym\xspace(q)]_0]_0}}(\ensuremath{\operatorname{\mathsf{loop}}})$} (br)
(bl) edge[double equal sign distance] node {$[q,b]_1$} (br)
(1.3,-2.5) edge[->] (1.3,-3.5);
\end{tikzpicture}
\end{center}
We will first focus on the left side of the squareover. We compute
\begin{align*}
\widetilde{\mathsf{apd}}_{f_0\circ u_t}(\ensuremath{\operatorname{\mathsf{loop}}})
&= \widetilde{\mathsf{apd}}_{f_0}(\apfunc{u_t}(\ensuremath{\operatorname{\mathsf{loop}}}))\\
&= \widetilde{\mathsf{apd}}_{f_0}(\overline{[t]_1})\\
&= \widetilde{\overline{\widetilde{s_1}}(t)}\\
&= \widetilde{\overline{s_1}(t)}\\
&= \widetilde{(\br{q}_2^{-1})_*1} && \text{(using $s_2$)} \\
&\equiv\vcentcolon \widetilde{1}.
\end{align*}
Here with $\widetilde{1}$ we mean the pathover $1:s_0(a) =_{1_{\br{a}_0}}^{P}s_0(a)$ but transported along the path
$$1_{\br{a}_0} \stackrel{\br{q}_2}= \overline{\br{t}_1} = \apfunc{[{-}]_0}\overline{[t]_1} =
\apfunc{[{-}]_0}(\apfunc{u_t}(\ensuremath{\operatorname{\mathsf{loop}}}))=\apfunc{[u_t({-})]_0}(\ensuremath{\operatorname{\mathsf{loop}}}).$$
By unfolding the definition of $\br{q}_2$ this can be simplified to the following concatenation:
$$1_{\br{a}_0} \stackrel{e}= \apfunc{[{-}]_0}\overline{[t]_1} =
\apfunc{[{-}]_0}(\apfunc{u_t}(\ensuremath{\operatorname{\mathsf{loop}}}))=\apfunc{[u_t({-})]_0}(\ensuremath{\operatorname{\mathsf{loop}}}).$$
The right side of the squareover is easier to manipulate:
$$\widetilde{\mathsf{apd}}_{\ensuremath{\mathbf{0}}_{[q,b]_{1*}(s_0(a))}}(\ensuremath{\operatorname{\mathsf{loop}}})=\widetilde{1},$$
where in this case we mean the pathover $1:s_0(a) =_{1_{\br{a}_0}}^{P}s_0(a)$ transported along the path
$$1_{\br{a}_0} = \apfunc{\ensuremath{\mathbf{0}}_{[[\ensuremath\inrsym\xspace(q)]_0]_0}}(\ensuremath{\operatorname{\mathsf{loop}}}).$$
Now in both the left and the right side these transports only act on the path they lie over. This means that we can ``push them down'' to the base square.
After we do that, we have a vertically degenerate squareover, and we only have to show that the square over which it lies is also vertically degenerate, which is a straightforward calculation.
This finishes the definition of $f$. The computation rule $f\br{a}_0\equiv s_0(a)$ follows directly from the computation rule for the quotient. Furthermore, we have
$$\apd_f\br{r}_1\equiv\apd_f\mapfunc{[{-}]_0}[r]_1=\widetilde{\mathsf{apd}}_{f\circ[{-}]_0}[r]_1\equiv\widetilde{\mathsf{apd}}_{f_0}[r]_1=s_1(r).$$
\textbf{Recursion Principle.}\\
For the recursion principle, suppose given $P, p_0, p_1, p_2$ as in the theorem statement. We first define $g_0 : C \to P$ by
\begin{align*}
g_0[\ensuremath\inlsym\xspace a]_0&\ensuremath{\vcentcolon\equiv} p_0(a)\\
g_0[\ensuremath\inrsym\xspace (a,t,q)]_0&\ensuremath{\vcentcolon\equiv} p_0(a)\\
\apfunc{g_0}[r]_1&\ensuremath{\vcentcolon=} p_1(r).
\end{align*}
We define $g:D\to P$ by $g[c]_0\ensuremath{\vcentcolon\equiv} g_0(c)$ and then we need to define
$\apfunc{g}[q,x]_1 : g_0(u_t(x)) = p_0(a)$, which we do by induction to $x$. For $x\equiv b$, this can be done by reflexivity, so $\apfunc{g}[q,b]_1\ensuremath{\vcentcolon=} 1_{p_0(a)}$. When $x$ varies over $\ensuremath{\operatorname{\mathsf{loop}}}$, we need to fill the following square.
\begin{center}
\begin{tikzpicture}[thick, node distance=3cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$p_0(a)$};
\ensuremath{\operatorname{\mathsf{node}}}[right of = tl] (tr) {$p_0(a)$};
\ensuremath{\operatorname{\mathsf{node}}} (bl) at (0,-2) {$p_0(a)$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$p_0(a)$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[double equal sign distance] node {$1$} (tr)
edge[double equal sign distance] node[left] {$\apfunc{g_0\circ u_t}(\ensuremath{\operatorname{\mathsf{loop}}})$} (bl)
(tr) edge[double equal sign distance] node {$\apfunc{\ensuremath{\mathbf{0}}_{p_0(a)}}(\ensuremath{\operatorname{\mathsf{loop}}})$} (br)
(bl) edge[double equal sign distance] node {$1$} (br);
\end{tikzpicture}
\end{center}
This can be done by the following calculation.
$$\apfunc{g_0\circ u_t}(\ensuremath{\operatorname{\mathsf{loop}}})=\apfunc{g_0}\apfunc{u_t}(\ensuremath{\operatorname{\mathsf{loop}}})=\apfunc{g_0}\overline{[t]_1}=
\overline{p_1}(t)\stackrel{p_2}=1=\apfunc{\ensuremath{\mathbf{0}}_{p_0(a)}}(\ensuremath{\operatorname{\mathsf{loop}}}).$$
This completes the definition of $g$. The computation rule $g\br{a}_0\equiv p_0(a)$ follows from the computation rule for quotients on points. We can define the computation rule on paths as the composite
$$\iota_1 : \apfunc{g}\br{r}_1 \equiv \apfunc{g}\apfunc{[{-}]_0}[r]_1 = \apfunc{g_0}[r]_1 = p_1(r).$$
The fact that $g$ has the correct computation rule for 2-paths requires some complicated path algebra, which we will omit here.
\end{proof}
We can now define the general version of the 2-quotient, $\ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S)$, to be equal to
$\ensuremath{\operatorname{\mathsf{simple-two-quotient}}}(A,R,Q)$ where $Q$ is the inductive family
\begin{inductive}
\texttt{inductive} $Q : \{a : A\} \to \ensuremath{\operatorname{\mathsf{words}}}_R(a, a) \to \ensuremath{\mathsf{Type}}\xspace \ensuremath{\vcentcolon=}$ \\
$\bullet\ (a\ a' : A) \to (t\ t' : \ensuremath{\operatorname{\mathsf{words}}}_R(a,a')) \to (s : S(t,t')) \to Q(t \cdot t^{-1}).$
\end{inductive}
We then show that $\ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S)$ and $\|\ensuremath{\operatorname{\mathsf{two-quotient}}}(A,R,S)\|_n$ have the right elimination
principles and computation rules (it requires some work to show that the eliminator of the truncated 2-quotient has the right computation rules on 2-paths).
This allows us to define all nonrecursive HITs with point, 1-path and 2-path constructors.
For example, we define the torus $T^2 := \ensuremath{\operatorname{\mathsf{two-quotient}}}(\ensuremath{\mathbf{1}}\xspace,R,S)$ where
$R(⋆,⋆) = \ensuremath{\mathbf{2}}\xspace$ (giving two path constructors $p$ and $q$ from the basepoint to itself) and
$Q$ is generated by the constructor
$s_0 : S ({0_{\bool}} \cdot {1_{\bool}}) ({1_{\bool}} \cdot {0_{\bool}})$, which determines a path $p \cdot q = q \cdot p$.
We also define the \emph{groupoid quotient}: For a groupoid $G$ we define its quotient as
$\|\ensuremath{\operatorname{\mathsf{two-quotient}}}(G, \ensuremath{\operatorname{\mathsf{hom}}}_G, S)\|_1$ where:
\begin{inductive}
\texttt{inductive} $S \ensuremath{\vcentcolon=}$ \\
$\bullet\ (a\ b\ c : G) \to (g : \ensuremath{\operatorname{\mathsf{hom}}}(b,c)) \to (f : \ensuremath{\operatorname{\mathsf{hom}}}(a,b)) \to S (g \circ f) (f \cdot g)$
\end{inductive}
If $G$ is just a group (considered as a groupoid with a single object), then the groupoid quotient
of $G$ is exactly the Eilenberg-MacLane space $K(G,1)$. For more information, see \autoref{sec:eilenb-macl-spac}.
\section{Colimits}\label{sec:colimits}
\lstDeleteShortInline"
We can ask whether we can use the construction of \Cref{sec:prop-trunc} can be generalized to
construct other higher inductive types.\footnote{The work in this section is joint work with Egbert Rijke and Kristina Sojakova.}
The general idea is that we can construct a recursive higher
inductive type as a sequential colimit of repeatedly applying a nonrecursive version of the
HIT. This does not work in general: if a constructor is infinitary, there is no reason why the type after $\omega$ many steps is the desired type. However, this does work for a general class of higher inductive types, the \emph{$\omega$-compact localizations}. In this section we will show various properties of colimits that are used in the proof of this fact. The full proof will appear in an upcoming preprint.
\begin{defn}
Suppose given a type $A$, families $P, Q : A \to \ensuremath{\mathsf{Type}}\xspace$ and $F : \{a : A\} \to P(a) \to Q(a)$.
A type $X$ is \emph{$F$-local} if for all $a : A$ the map
$$\psi_X(a)\ensuremath{\vcentcolon\equiv} \lam{f}f \o F(a) : (Q(a) \to X) \to (P(a) \to X)$$
is an equivalence.
The \emph{$F$-localization} $L_FX$ or $LX$ of $X$ turns $X$ into a $F$-local type in a universal
way. This means there is a map $\ell_X : X \to LX$ such that for any $F$-local type $Y$ there is an
equivalence of maps $(LX \to Y) \to (X \to Y)$ given by precomposition with $\ell_X$. $L_FX$ can
be given as a higher inductive type with the following constructors:
\begin{lstlisting}[gobble=4]
HIT L F X : Type :=
| incl : X → L X
| rinv : Π{a} (f : P a → L X), Q a → L X
| isri : Π{a} (f : P a → L X) (x : P a), rinv f (F x) = f x
| linv : Π{a} (f : P a → L X), Q a → L X
| isli : Π{a} (f : Q a → L X) (x : Q a), linv (f ∘ F) x = f x.
\end{lstlisting}
\end{defn}
For a sequence $(A_n,f_n)_n$ we denote the colimit by $\ensuremath{\operatorname{\mathsf{colim}}}(A)$ or $A_\infty$. Also, for any type $X$, we can define a new sequence $(X\to A_n, f_n \o ({-}))_n$. Note that there is a canonical map
$$\xi_X : \ensuremath{\operatorname{\mathsf{colim}}}(X\to A_n)\to (X\to A_\infty).$$
It is defined by $\xi_X(i_n(f))\ensuremath{\vcentcolon\equiv} i_n\o f$ and $\xi_X(\kappa(f))\ensuremath{\vcentcolon\equiv} \kappa_n\o f$, where $\kappa$ is the path constructor of the colimit.
\begin{defn}
A type $X$ is said to be \emph{$\omega$-compact} if the map $\xi_X$ is an equivalence for all sequences $(A_n,f_n)_n$.
\end{defn}
Examples of $\omega$-compact types are the finite types. Moreover, the $\omega$-compact types are closed under dependent pair types and pushouts. A non-example of an $\omega$-compact type is $\textbf{N}$. We will omit the details here.
\begin{thm}\label{thm:localization}
Assume that for all $a : A$ the types $P(a)$ and $Q(a)$ are $\omega$-compact. Then we can
construct the $F$-localization in MLTT$+$quotients.
\end{thm}
We will not prove this theorem here, but defer it to an upcoming preprint. However, we will develop machinery here that is crucial to prove this theorem. In particular we prove that sigma-types commute with sequential colimits.
\subsubsection*{Type Sequences}
\begin{defn}
A \emph{type sequence} $\sequence{A}{f}$ consists of a diagram of the form
\begin{equation*}
\begin{tikzcd}
A_0 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_0"] & A_1 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_1"] & A_2 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_2"] & \cdots
\end{tikzcd}
\end{equation*}
Thus, the type of all sequences of types is
\begin{equation*}
\mathrm{Seq} \ensuremath{\vcentcolon\equiv} (A:\nat\to\ensuremath{\mathsf{Type}}\xspace)\times(n:\nat) \to A_n\to A_{n+1}
\end{equation*}
\end{defn}
Recall that the relation $\leq$ on the natural numbers is defined as an inductive family of types $\leq\mathop{:}\textbf{N}\to\textbf{N}\to\ensuremath{\mathcal{U}}\xspace$ with
\begin{align*}
r & : (n:\textbf{N}) \to n\leq n \\
s & : (n,m:\textbf{N})\to n\leq m \to n\leq m+1.
\end{align*}
It follows that $n\leq m$ is a a mere proposition for each $n,m:\textbf{N}$.
\begin{defn}
Let $\sequence{A}{f}$ be a type sequence. For any $n,m:\nat$, we define
\begin{equation*}
f^{n\leq m} : A_n\to A_m.
\end{equation*}
where we leave the proof that $n\leq m$ implicit.
\end{defn}
\begin{proof}[Construction]
We define $f^{n\leq m}$ by induction on the proof that $n\leq m$ by taking
\begin{align*}
f^{n\leq n} & \ensuremath{\vcentcolon\equiv} \idfunc[A_n] \\
f^{n\leq m+1} & \ensuremath{\vcentcolon\equiv} f_m\circ f^{n\leq m} \qedhere
\end{align*}
\end{proof}
\begin{defn}
Let $\sequence{A}{f}$ be a type sequence. For any $n,k:\nat$, we define
$f_n^k:A_n\to A_{n+k}$ to be $f^{n\leq n+k}(p)$, where $p$ is the canonical proof that $n\leq n+k$.
\end{defn}
\begin{defn}
A \emph{sequence $\sequence{B}{g}$ of types over $\sequence{A}{f}$} consists of a diagram of the form
\begin{equation*}
\begin{tikzcd}
B_{0} \ensuremath{\operatorname{\mathsf{arrow}}}[r,"g_0"] \ensuremath{\operatorname{\mathsf{arrow}}}[d,->>] & B_{1} \ensuremath{\operatorname{\mathsf{arrow}}}[r,"g_1"] \ensuremath{\operatorname{\mathsf{arrow}}}[d,->>] & B_{2} \ensuremath{\operatorname{\mathsf{arrow}}}[r,"g_2"] \ensuremath{\operatorname{\mathsf{arrow}}}[d,->>] & \cdots \\
A_0 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_0"] & A_1 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_1"] & A_2 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_2"] & \cdots
\end{tikzcd}
\end{equation*}
where each $g_n$ has type $(a:A_n)\to B_n(x)\to B_{n+1}(f_n(a))$, implicitly rendering the
squares commutative.
We say that a sequence $\sequence{B}{g}$ over $\sequence{A}{f}$ is \emph{equifibered} if each $g_n$ is a family of equivalences.
\end{defn}
\begin{defn}
Let $\sequence{A}{f}$ and $\sequence{A'}{f'}$ be type sequences.
A \emph{natural transformation} $\sequence{A}{f}\to\sequence{A'}{f'}$
is a pair $\sequence{\tau}{H}$ consisting of a family of maps
\begin{equation*}
\tau : (n:\textbf{N}) \to A_n \to A'_n
\end{equation*}
and a family $H_n$ of homotopies witnessing that the diagram
\begin{equation*}
\begin{tikzcd}
A_{0} \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_0"] \ensuremath{\operatorname{\mathsf{arrow}}}[d,"\tau_0"] & A_{1} \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_1"] \ensuremath{\operatorname{\mathsf{arrow}}}[d,"\tau_1"] & A_{2} \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_2"] \ensuremath{\operatorname{\mathsf{arrow}}}[d,"\tau_2"] & \cdots \\
A'_0 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"{f'_0}"] & A'_1 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"{f'_1}"] & A'_2 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"{f'_2}"] & \cdots
\end{tikzcd}
\end{equation*}
commutes.
\end{defn}
\begin{defn}
A \emph{natural equivalence} is a natural
transformation $(\tau,H)$ such that each $\tau_n$ is an equivalence.
The type of natural equivalences from $\sequence{A}{f}$ to $\sequence{A'}{f'}$
is called $\mathsf{NatEq}(\sequence{A}{f},\sequence{A'}{f'})$.
\end{defn}
\begin{lem}
The canonical dependent function $\mathsf{idtonateq}$
\begin{equation*}
(\sequence{A}{f}=\sequence{A'}{f'})\to \mathsf{NatEq}(\sequence{A}{f},\sequence{A'}{f'})
\end{equation*}
that sends $\ensuremath{\operatorname{\mathsf{refl}}}_{\sequence{A}{f}}$ to the identity natural transformation, is
an equivalence.
\end{lem}
\begin{proof}
Straightforward application of univalence.
\end{proof}
Every type sequence $\sequence{B}{g}$ over $\sequence{A}{f}$ gives rise to a natural transformation, by the following definition.
\begin{defn}
Let $\sequence{B}{g}$ be a sequence over $\sequence{A}{f}$. Then we define the
sequence $\msm{\sequence{A}{f}}{\sequence{B}{g}}$ to consist of the diagram
\begin{equation*}
\begin{tikzcd}[column sep=large]
(a:A_0)\times B_0(a) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"\pairr{f_0,g_0}"] & (a:A_1)\times B_1(a) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"\pairr{f_1,g_1}"]
& (a:A_2)\times B_2(a) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"\pairr{f_2,g_2}"] & \cdots
\end{tikzcd}
\end{equation*}
where we take the usual definition
\begin{equation*}
\pairr{f_n,g_n} \ensuremath{\vcentcolon\equiv} \lam{\pairr{a,b}}\pairr{f_n(a),g_n(a,b)}.
\end{equation*}
Furthermore, we define a natural transformation
\begin{equation*}
\sequence{\pi}{\theta}:\msm{\sequence{A}{f}}{\sequence{B}{g}}\to \sequence{A}{f}
\end{equation*}
by taking
\begin{align*}
\pi_n & \ensuremath{\vcentcolon\equiv} \proj1 & & : ((a:A_n)\times B_n(a))\to A_n \\
\theta_n(a,b) & \ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{refl}}}_{f_n(a)} & & : f_n(\proj 1(a,b))= \proj 1(f_n(a),g_n(b)).
\end{align*}
\end{defn}
We will now look at the shift operation on type sequences, in particular to bring up subtleties that come up in the formalization of mathematics in homotopy type theory. The issue we face is that equality in the natural numbers is not always strict. For instance, when addition is defined by induction on the second argument, then $n+0$ is judgmentally equal to $n$, while $0+n$ is not. This implies that sometimes we might have to \emph{transport} along the equalities in the natural numbers (such as $n=0+n$), and this complicates the formalization process.
We define the shift operation.
\begin{defn}
For any type sequence $\sequence{A}{f}$ we define a new type sequence $(S(A),S(f))$ by taking
\begin{align*}
S(A)_n & \ensuremath{\vcentcolon\equiv} A_{n+1} \\
S(f)_n & \ensuremath{\vcentcolon\equiv} f_{n+1}.
\end{align*}
\end{defn}
Of course we can iterated the shift operation, defining a type sequence $(S^k(A),S^k(f))$ for every $k:\textbf{N}$. However, while the type $S^k(A)_n$ is $A_{n+k}$, the function $S^k(f)_n$ is some function $A_{n+k}\to A_{(n+1)+k}$ that is not judgmentally equal to a function of the form $f_m$ for some $m:\textbf{N}$. Therefore, we make an alternative definition of the $k$-shift that is different from $S^k$, the type sequence obtained from iterating the shift $S$.
\begin{defn}
Given a type sequence $(A,f)$, we define $S_k(A,f)\equiv(S_k(A),S_k(f))$ to be the type sequence given by
\begin{align*}
S_k(A)_n & \ensuremath{\vcentcolon\equiv} A_{k+n} \\
S_k(f)_n & \ensuremath{\vcentcolon\equiv} f_{k+n}.
\end{align*}
Given a dependent sequence $(B,g)$ over $(A,f)$, we also define $S_k(B,g)\equiv (S_k(B),S_k(g))$ by
\begin{align*}
S_k(B)_n & \ensuremath{\vcentcolon\equiv} B_{k+n} \\
S_k(g)_n & \ensuremath{\vcentcolon\equiv} g_{k+n}.
\end{align*}
\end{defn}
Note that the sequence $(S_{k+1}(A),S_{k+1}(f))$ is not judgmentally equal to the sequence $S(S_k(A),S_k(f))$, since in general we do not have $(k+1)+n\equiv (k+n)+1$. Therefore we have the following lemma.
\begin{lem}\label{lem:iterate_succ}
For any $k,n:\nat$ and $a : A_k$, one has $q_{k,n}(a):\dpath{A}{p(k,n)}{f_k^{n+1}(a)}{f_{k+1}^n(f_k(a))}$ where $p(k,n):(k+n)+1=(k+1)+n$ is the canonical path in $\nat$.
\end{lem}
\begin{proof}
By induction on $n:\textbf{N}$.
\end{proof}
\begin{cor}
For any type sequence $\sequence{A}{f}$, the type sequence $(S_{k+1}(A),S_{k+1}(f))$ is naturally equivalent to the type sequence $(S(S_k(A)),S(S_k(f)))$.
\end{cor}
\subsubsection*{Sequential Colimits}
\begin{rmk}
The induction principle for sequential colimits tells us how to construct a dependent function $f:(a:A_\infty)\to P(a)$ for a type family $P:A_\infty\to\ensuremath{\mathsf{Type}}\xspace$.
Given $s:(a:A_\infty)\to P(a)$, we get
\begin{align*}
\lam{n}{a} s(\iota_n(a)) & : (n:\textbf{N})(a:A_n)\to P(\iota_n(a)) \\
\lam{n}{a} \apd{s}{\kappa_n(a)} & : (n:\textbf{N})(a:A_n)\to s(\iota_n(a)) =_{\kappa_n(a)}^P s(\iota_{n+1}(f_n(a)))
\end{align*}
In other words, we have a canonical map
\begin{align*}
&\Big((a:A_\infty)\to P(a)\Big)\to\\
&\Big((h:(n:\textbf{N})(a:A_n)\to P(\iota_n(a)))\times(n:\textbf{N})(a:A_n)\to h_n(a) =_{\kappa_n(a)}^P h_{n+1}(f_n(a))\Big)
\end{align*}
Now we can state the induction principle and computation rule concisely: the canonical map described above comes equipped with a section. We assume that that the computation rule is strict on the point constructors.
\end{rmk}
The universal property of sequential colimits is a straightforward consequence of the induction principle.
\begin{thm}
Let $\sequence{A}{f}$ be a type sequence, and let $X$ be a type. Then the canonical map
\begin{equation*}
(A_\infty\to X)\to (h:(n:\textbf{N}) \to A_n\to X)\times(n:\textbf{N})\to h_n\sim h_{n+1}\circ f_n
\end{equation*}
is an equivalence.
\end{thm}
The following theorem is a descent theorem for sequential colimits.
\begin{thm}\label{thm:descent}
Consider a sequence $\sequence{A}{f}$. The type $A_\infty\to\ensuremath{\mathcal{U}}\xspace$ is equivalent to the type of equifibered type sequences over $\sequence{A}{f}$.
\end{thm}
\begin{proof}
By the universal property of $A_\infty$ and by univalence we have
\begin{align*}
(A_\infty\to \ensuremath{\mathcal{U}}\xspace) & \eqvsym (B:(n:\textbf{N})\to A_n\to \ensuremath{\mathcal{U}}\xspace)\times (n:\textbf{N}) \to B_n\sim B_{n+1}\circ f_n \\
& \eqvsym (B:(n:\textbf{N}) \to A_n\to \ensuremath{\mathcal{U}}\xspace)\times(n:\textbf{N})(x:A_n)\to B_n(x)\eqvsym B_{n+1}(f_n(x))\qedhere
\end{align*}
\end{proof}
\begin{lem}\label{lem:seq_colim_functor}
Suppose given a natural transformation $(\tau,H):\sequence{A}{f}\to\sequence{A'}{f'}$.
\begin{enumerate}
\item\label{part:seq_colim_functor} We get a function $\mathsf{colim}(\tau,H)$ or $\tau_\infty:A_\infty\to A'_\infty$.
\item\label{part:1_functoriality} The sequential colimit is 1-functorial. This means the following three things. If $(\sigma,K):\sequence{A'}{f'}\to\sequence{A''}{f''}$, then $(\tau\circ\sigma)_\infty\sim\tau_\infty\circ\sigma_\infty$. Moreover, $1_\infty\sim \idfunc$, where $1$ is the identity natural transformation. Lastly, if $(\tau',H'):\sequence{A}{f}\to\sequence{A'}{f'}$ and $q:(n:\textbf{N})\to \tau_n\sim\tau'_n$ and we can fill the following square for all $a:A_n$
\begin{center}\begin{tikzcd}[column sep=25mm]
\tau_{n+1}(f_na)
\ar[r,equal,"{q_{n+1}(f_na)}"]
\ar[d,equal,"{H_n(a)}"] &
\tau'_{n+1}(f_na)
\ar[d,equal,"{H'_n(a)}"]\\
f'_n(\tau_n(a))
\ar[r,equal,"{\apfunc{f'_n}(q_n(a))}"] &
f'_n(\tau'_n(a))
\end{tikzcd}\end{center}
then $\tau_\infty\sim\tau'_\infty$.
\item\label{part:functor_equivalence} If $\tau$ is a natural equivalence, then $\tau_\infty$ is an equivalence.
\end{enumerate}
\end{lem}
\begin{proof}\mbox{}
\begin{enumerate}
\item
We define $\tau_\infty(\iota_n(a))\ensuremath{\vcentcolon\equiv}\iota_n(\tau_n(a))$ and
$$\apfunc{\tau_\infty}(\kappa_n(a))\vcentcolon=\apfunc{\iota_{n+1}}(H(a))\cdot\kappa_n(\tau_n(a)):
\iota_{n+1}(\tau_{n+1}(f_na))=\iota_n(\tau_n(a)).$$
\item All three parts are by induction on the element of $A_\infty$, and all parts are straightforward.
\item We define $(\tau_\infty)^{-1}\ensuremath{\vcentcolon\equiv}(\tau^{-1})_\infty$ where $\tau^{-1}$ is the natural transformation by inverting $\tau_n$ for each $n$. Now we can check that this is really the inverse by using all three parts of the 1-functoriality.
$$\tau_\infty^{-1}\circ\tau_\infty\sim (\tau^{-1}\circ\tau)_\infty\sim 1_\infty\sim\idfunc[A_\infty].$$
For the second homotopy we need to show that we can fill a certain square, which is straightforward.
The other composite is homotopic to the identity by a similar argument.\qedhere
\end{enumerate}
\end{proof}
The following lemma states that $\iota_0$ is an equivalence if all maps in the sequence are an equivalence. We will have a more general result in \autoref{cor:trunc_colim}\ref{part:iota_is_trunc_conn}, but in that proof we will use some special cases of this lemma.
\begin{lem}\label{lem:equiv_equiseq}
Suppose given a sequence $\sequence{A}{f}$ where $f_n$ is an equivalence for all $n$.
Then $\iota_0:A_0\to A_\infty$ is an equivalence.
\end{lem}
\begin{proof}
First note that the map $f^{0\le n}:A_0\to A_n$ is an equivalence, which is an easy induction on the proof that $0\le n$, because $f^{0\le0}\equiv\idfunc$ is an equivalence and $f^{0\le n+1}\equiv f_n\circ f^{0\le n}$ is a composition of two equivalences.
Also note that we have paths $\kappa^{n\le m}(a):\iota_m(f^{n\le m}(a))=\iota_n(a)$ for $a:A_n$.
Now we define $\iota_0^{-1}:A_\infty\to A_0$ as
$$\iota_0^{-1}(\iota_n(a))\ensuremath{\vcentcolon\equiv} (f^{0\le n})^{-1}(a)$$
and we define
$$\apfunc{\iota_0^{-1}}(\kappa_n(a)):(f^{0\le n})^{-1}(f_n^{-1}(f_n(a)))=(f^{0\le n})^{-1}(a)$$
as $\apfunc{(f^{0\le n})^{-1}}(\ell_n(a))$, where $\ell_n(a):f_n^{-1}(f_n(a))$ is the canonical path.
Now $\iota_0^{-1}\circ\iota_0\sim\idfunc$ is true by definition. To show that for $x:A_\infty$ we have
$p(x):\iota_0(\iota_0^{-1}(x))=x$, we use induction on $x$. If $x\equiv\iota_n(a)$, we have
\begin{align*}
\iota_0(\iota_0^{-1}(\iota_n(a)))
&\equiv \iota_0((f^{0\le n})^{-1}(a))\\
&=\iota_n(f^{0\le n}((f^{0\le n})^{-1}(a)))\\
&=\iota_n(a).
\end{align*}
If we write $r^{0\le n}: f^{0\le n} \circ (f^{0\le n})^{-1}\sim\idfunc$ for the canonical homotopy, then we explicitly define $p(\iota_n(a))$ as
$$p(\iota_n(a))\ensuremath{\vcentcolon\equiv}(\kappa^{0\le n}((f^{0\le n})^{-1}(a)))^{-1}\cdot \apfunc{\iota_n}(r^{0\le n}(a)).$$
If $x$ varies over $\kappa_n(a)$, then we need to fill the following square.
\begin{center}\begin{tikzcd}[column sep=25mm]
\iota_0(\iota_0^{-1}(\iota_{n+1}(f_na)))
\ar[r,equal,"{p(\iota_{n+1}(f_na))}"]
\ar[d,equal,"{\mapfunc{\iota_0\circ\iota_0^{-1}}(\kappa_n(a))}"] &
\iota_{n+1}(f_na)
\ar[d,equal,"{\kappa_n(a)}"]\\
\iota_0(\iota_0^{-1}(\iota_n(a)))
\ar[r,equal,"{p(\iota_n(a))}"] &
\iota_n(a)
\end{tikzcd}\end{center}
If we unfold the definitions of $\iota_0^{-1}$ and $p$, we can fill this as the horizontal concatenation of the following two squares (where we have left out some arguments to the paths)
\begin{center}\begin{tikzcd}[column sep=15mm]
\iota_0((f^{0\le n})^{-1}(f_n^{-1}(f_na)))
\ar[r,equal,"{(\kappa^{0\le n})^{-1}}"]
\ar[d,equal,"{\mapfunc{\iota_0\circ(f^{0\le n})^{-1}}(\ell)}"] &
\iota_n(f^{0\le n}((f^{0\le n})^{-1}(f_n^{-1}(f_na))))
\ar[d,equal,"{\mapfunc{\iota_n\circ f^{0\le n}\circ (f^{0\le n})^{-1}}(\ell)}"] \\
\iota_0((f^{0\le n})^{-1}(a))
\ar[r,equal,"{(\kappa^{0\le n})^{-1}}"] &
\iota_n(f^{0\le n}((f^{0\le n})^{-1}(a)))
\end{tikzcd}\end{center}
\begin{center}\begin{tikzcd}[column sep=15mm]
\iota_n(f^{0\le n}((f^{0\le n})^{-1}(f_n^{-1}(f_na))))
\ar[rr,equal,"{\kappa^{-1}\cdot\apfunc{\iota_{n+1}}(\apfunc{f}(r^{0\le n})\cdot r)}"]
\ar[dd,equal,"{\mapfunc{\iota_n\circ f^{0\le n}\circ (f^{0\le n})^{-1}}(\ell)}"]
\ar[rd,equal,"{\mapfunc{\iota_n}(r^{0\le n})}"] & &
\iota_{n+1}(f_na)
\ar[dd,equal,"{\kappa}"]\\
& \iota_0(f_n^{-1}(f_na))
\ar[rd,equal,"{\mapfunc{\iota_n}(\ell)}"] & \\
\iota_n(f^{0\le n}((f^{0\le n})^{-1}(a)))
\ar[rr,equal,"{\apfunc{\iota_n}(r^{0\le n})}"] & &
\iota_n(a)
\end{tikzcd}\end{center}
The first square is a naturality square, as is the bottom-left part of the second square.
We can use the triangle equalities of $f$ to rewrite the $r$ in the top part to $\apfunc{f}(\ell)$. After doing that, the top-right square becomes the following naturality square.
\begin{center}\begin{tikzcd}[column sep=20mm]
\iota_n(f_n(f^{0\le n}((f^{0\le n})^{-1}(f_n^{-1}(f_na)))))
\ar[r,equal,"{\apfunc{\iota_{n+1}\circ f}(r^{0\le n}\cdot \ell)}"]
\ar[d,equal,"{\kappa}"] &
\iota_{n+1}(f_na)
\ar[d,equal,"{\kappa}"]\\
\iota_0(f_n^{-1}(f_na))
\ar[r,equal,"{\mapfunc{\iota_n}(r^{0\le n}\cdot\ell)}"] &
\iota_n(a)
\end{tikzcd}\end{center}
\end{proof}
\begin{lem}\label{lem:colim_shift_one}
For any type sequence $(A,f)$, the colimits of $(A,f)$ and $S(A,f)$ are equivalent.
\end{lem}
\begin{proof}
We construct a map $\varphi:A_\infty \to S(A)_\infty$ by induction on $A_\infty$, by taking
\begin{align*}
(x:A_n) & \mapsto \iota^{S(A),S(f)}_n(f_n(x)) \\
(x:A_n) & \mapsto \kappa^{S(A),S(f)}_n(f_n(x)).
\end{align*}
Next, we construct a map $\psi:S(A)_\infty\to A_\infty$ by induction on $S(A)_\infty$, by taking
\begin{align*}
(x:S(A)_n) & \mapsto \iota^{A,f}_{n+1}(x) \\
(x:S(A)_n) & \mapsto \kappa^{A,f}_{n+1}(x).
\end{align*}
Then we prove that $\psi\circ \varphi\sim \idfunc$ by induction on $A_\infty$, by taking
\begin{align*}
(x:A_n) & \mapsto \kappa^{A,f}_n(x)
\end{align*}
Now we compute
\begin{align*}
\mathsf{ap}_{\psi\circ\varphi}(\kappa^{A,f}_n(x))) & = \mathsf{ap}_\psi(\mathsf{ap}_\varphi(\kappa^{A,f}_n(x))) \\
& = \mathsf{ap}_\psi(\kappa^{S(A),S(f)}_n(f_n(x))) \\
& = \kappa^{A,f}_{n+1}(f_n(x))
\end{align*}
from the computation rules of $A_\infty$ and $S(A)_\infty$.
We construct the homotopy $\varphi\circ\psi\sim\idfunc$ by induction on $A_\infty$, by taking
\begin{align*}
(x:S(A)_n) & \mapsto \kappa_{n+1}(S(f)_n(x))
\end{align*}
Now we compute
\begin{align*}
\mathsf{ap}_{\varphi\circ\psi}(\kappa^{S(A),S(f)}_n(x)) & = \mathsf{ap}_\varphi(\mathsf{ap}_\psi(\kappa^{S(A),S(f)}_n(x))) \\
& = \mathsf{ap}_\varphi(\kappa^{A,f}_{n+1}(x)) \\
& = \kappa^{S(A),S(f)}_{n+1}(f_{n+1}(x)).\qedhere
\end{align*}
\end{proof}
\begin{lem}\label{lem:colim_shift_k}
For any type sequence $(A,f)$, we have an equivalence
\begin{equation*}
\mathsf{kshift\underline{~}equiv}_{k} : \mathsf{colim}(A,f)\eqvsym\mathsf{colim} (S_k(A,f)).
\end{equation*}
\end{lem}
The shift operations and the corresponding equivalences on the sequential colimits can be used to turn an arbitrary sequence $\sequence{B}{g}$ over $\sequence{A}{f}$ into an equifibered sequence over $\sequence{A}{f}$.
\begin{defn}
Given a dependent sequence $(B,g)$ over $(A,f)$ and $x:A_0$, we define a type sequence $(B[x],g[x])$ by
\begin{align*}
B[x]_n & \ensuremath{\vcentcolon\equiv} B_n(f^n(x)) \\
g[x]_n & \ensuremath{\vcentcolon\equiv} g_n(f^n(x),\mathord{\hspace{1pt}\text{--}\hspace{1pt}}).
\end{align*}
\end{defn}
\begin{defn}
Given any sequence $\sequence{B}{g}$ over $\sequence{A}{f}$, we define an equifibered sequence
$\sequence{\square B}{\square g}$ over the sequence $\sequence{A}{f}$.
\end{defn}
\begin{constr}
For $x:A_n$ we define
\begin{equation*}
(\square B)_n(x) \ensuremath{\vcentcolon\equiv} S_n(B)[x]_{\infty} \equiv \mathsf{colim}_m(B_{n+m}(f^m(x))).
\end{equation*}
Now note that
\begin{align*}
(\square B)_{n+1}(f(x)) & \equiv \mathsf{colim}_m(B_{(n+1)+m}(f^m(f(x))) \\
& \simeq \mathsf{colim}_m(B_{n+(m+1)}(f^{m+1}(x))\\
& \simeq \mathsf{colim}_m(B_{n+m}(f^m(x))\\
& \equiv (\square B)_n(x)
\end{align*}
The first equivalence $u_{n,m}$ is given by transporting along the dependent path in \autoref{lem:iterate_succ} in the family $B$. This forms a natural equivalence, because $\mathsf{transport}$ is natural. The second equivalence is given by applying \autoref{lem:colim_shift_one}. We call the composite equivalence $F$, which shows that $\square B$ is an equifibered sequence.
\end{constr}
\begin{defn}
Let $\sequence{B}{g}$ be a sequence over $\sequence{A}{f}$. Then we define
\begin{equation*}
B_\infty : A_\infty\to\ensuremath{\mathcal{U}}\xspace
\end{equation*}
to be the family over $A_\infty$ associated to the equifibered sequence $(\square B,\square g)$ via the equivalence of \autoref{thm:descent}.
\end{defn}
By construction of $B_\infty$ we get the equality
$$r(y):\transfib{B_\infty}{\kappa_n(x)}{y}=F(y)$$
for $y:B_\infty(\iota_{n+1}(f_n(x)))$ witnessing that $B_\infty$ is defined by the equivalence $F$ on the path constructor.
We now state our main result, which could be seen as a flattening lemma for sequential colimits,
with the added generality that the sequence $\sequence{B}{g}$ over
$\sequence{A}{f}$ is not required to be equifibered.
\begin{thm}\label{thm:colim_sm}
Let $P\ensuremath{\vcentcolon\equiv}\sequence{P}{f}$ be a sequence over $A\ensuremath{\vcentcolon\equiv}\sequence{A}{a}$. Then we have a
commuting triangle
\begin{equation*}
\begin{tikzcd}[column sep=huge]
\mathsf{colim}(\msm{A}{P}) \ensuremath{\operatorname{\mathsf{arrow}}}[rr,"{\alpha}"] \ensuremath{\operatorname{\mathsf{arrow}}}[dr,swap,"{p\ensuremath{\vcentcolon\equiv}\mathsf{rec}(\iota_n\circ \proj 1,\mathord{\hspace{1pt}\text{--}\hspace{1pt}})}"]
& & (x:A_\infty)\times P_\infty(x) \ensuremath{\operatorname{\mathsf{arrow}}}[dl,"{\proj 1}"] \\
& A_\infty
\end{tikzcd}
\end{equation*}
in which $\alpha$ is an equivalence.
\end{thm}
The strategy of the proof is to first show that $(x:A_\infty)\times P_\infty(x)$ has the induction principle of $\ensuremath{\operatorname{\mathsf{colim}}}((x:A_n)\times P_n(x),(a_n,f_n))_n$. This simplifies giving the equivalence, because
$(x:A_\infty)\times P_\infty(x)$ is a 2-HIT, being a sigma-type of two 1-HITs, while $\ensuremath{\operatorname{\mathsf{colim}}}((x:A_n)\times P_n(x),(a_n,f_n))_n$ is a 1-HIT. Before we continue, we first define $\alpha$.
The map $\alpha$ is defined by induction on $\mathsf{colim}(\msm{A}{P})$. On the point constructors we define
\begin{equation*}
\alpha(\iota_n(x,y))\ensuremath{\vcentcolon\equiv} \pairr{\iota_n(x),\iota_0(y)}.
\end{equation*}
For the path constructor we need to define
$$\kappa'_n(x,y):(\iota_{n+1}(a_nx),\iota_0(f_n(x,y)))=(\iota_n(x),\iota_0(y))$$
The first components are equal by $\kappa_n(x)$. By the definition of $P_\infty$, transporting along $\kappa_n(x)$ takes $\iota_0(f_n(x,y))$ to $\iota_1(f_n(x,y))$, which is equal to $\iota_0(y)$ by $\kappa_0(y)$. Explicitly, we define
$$\apfunc{\alpha}(\kappa_n(x,y))\vcentcolon=\kappa'_n(x,y)\ensuremath{\vcentcolon\equiv} (\kappa_n(x),r(\iota_0(f_n(x,y)))\cdot\kappa_0(y)).$$
\begin{thm}\label{thm:sigma-colim-induction}
Let $E: (x:A_{\infty})\to P_{\infty}(x) \to \ensuremath{\mathcal{U}}\xspace$ such that
\begin{enumerate}
\item For each $n:\textbf{N}$, $x:A_n$, $y:P_n(x)$, a term $e_n(x,y):E(\iota_n(x),\iota_0(y))$.
\item For each $n:\textbf{N}$, $x:A_n$, $y:P_n(x)$, a path
\begin{equation*}
w_{n}(x,y):e_{n+1}(a_nx,f_n(x,y))=_{\kappa'_n(x,y)}^Ee_n(x,y).
\end{equation*}
\end{enumerate}
Then there exists a function $s:(x:A_\infty)(y:P_\infty(x))\to E(y)$.
\end{thm}
\begin{proof}
We define the function $s$ by induction on both $x$ and $y$. We need to consider four cases, since both $x$ and $y$ can be a point constructor or vary over a path constructor.
\emph{(point-point)}
Fix $x:A_n$, we first define $g(n,x):(p:P_{\infty}(\iota_n(x))) \to E(\iota_n(x),p)$. To obtain $g(n,x)$, we do induction on $p:P_\infty(\iota_n(x))$. Fix $y:P_{n+k}(a_n^k(x))$, we need to construct a term of type $g_\ast(k,n,x,y) : E(\iota_n(x),\iota_k(y))$. Proceed by induction on $k$. We can define
$$g_\ast(0,n,x,y)\ensuremath{\vcentcolon\equiv} e_n(x,y) : E(\iota_n(x),\iota_0(y)).$$
Assume that $g_\ast(k)$ is defined.
We need to define $g_\ast(k+1,n,x,y):E(\iota_n(x),\iota_{k+1}(y))$, where $y:P(n+(k+1),a_n^{k+1}(x))$. However, the type of $y$ is equivalent to the type $P((n+1)+k,a_{n+1}^k(a_n(x)))$ via the equivalence $u_{n,k}$. Therefore, it suffices to define for $z:P_{(n+1)+k}(a_{n+1}^k(a_n(x)))$
\begin{equation*}
g_\ast(k+1,n,x,u_{n,k}(z)) : E(\iota_n(x),\iota_{k+1}(u_{n,k}(z))).
\end{equation*}
By induction hypothesis we have $g_\ast(k,n+1,a_n(x),z):E(\iota_{n+1}(a_n(x)),\iota_k(z))$, so it suffices to show that
$$\kappa^\ast_{n,k}(x,z):(\iota_{n+1}(a_n(x)),\iota_k(z))=(\iota_n(x),\iota_{k+1}(u_{n,k}(z))).$$
This construction is similar to that of $\kappa'_n(x,y)$.
The first components are equal by $\kappa_n(x)$, and for the second components we need to show that $\transfib{P_\infty}{\kappa_n(x)}{\iota_k(z)}=\iota_{k+1}(u_{n,k}(z))$. This follows from the computation rule of $P_\infty$ on paths, since the equivalence used to define $P_\infty$ sends $\iota_k(z)$ to $\iota_{k+1}(u_{n,k}(z))$. Specifically,
$$\kappa^\ast_{n,k}(x,z)\ensuremath{\vcentcolon\equiv}(\kappa_n(x),r(\iota_k(y))).$$
This finishes the construction of $g_\ast$, hence also of $g$ on points. By construction, we get the following equation:
$$\mu_{n,k}(x,z):\dpath{E}{\kappa^\ast_{n,k}(x,z)}{g_\ast(k,n+1,a_n(x),z)}{g_\ast(k+1,n,x,u_{n,k}(z))}$$
\emph{(point-path)} To define $g$ on paths $\kappa_k(y):\iota_{k+1}(f(y))=\iota_k(y)$, we need to give a dependent path
$$\nu(k,n,x,y):g_\ast(k+1,n,x,f(y))=^{E(\iota_n(x))}_{\kappa_k(y)}g_\ast(k,n,x,y).$$
We do this by induction on $k$. For $k=0$ note that $u_{n,0}$ is the identity function, and the goal definitionally reduces to
$$\nu(k,n,x,y):\transfib{E}{\kappa^\ast_{n,0}(x,f_n(x,y)}{e_{n+1}(a_n(x),f_n(x,y)}=^{E(\iota_n(x))}_{\kappa_0(y)}e_n(x,y).$$
Note that $\kappa'_n(x,y)=\kappa^\ast_{n,0}(f_n(x,y)\cdot(1,\kappa_0(y))$, which means we get this from $w_n(x,y)$. Now suppose that $\nu(k)$ is defined. We need to define for $y:P(n+(k+1),a_n^{k+1}(x))$
$$\nu(k+1,n,x,y):g_\ast(k+2,n,x,f(y))=^{E(\iota_n(x))}_{\kappa_k(y)}g_\ast(k+1,n,x,y).$$
Now we again write $y=u_{n,k}(z)$ for $z:P((n+1)+k,a_{n+1}^k(a_n(x)))$ and we equivalently need to give
$$\nu(k+1,n,x,u_{n,k}(z)):g_\ast(k+2,n,x,f(u_{n,k}(z)))=^{E}_{(1,\kappa_k(y))}g_\ast(k+1,n,x,u_{n,k}z).$$
We will define this as the composition of a square that we will give later in the proof.
\emph{(path-point)} We have defined $s$ on points constructors of $A_\infty$. To define it on the path $\kappa_n(x):\iota_{n+1}(a_n(x))=\iota_n(x)$ we need a path $g(n+1,a_n(x))=g(n,x)$ over $\kappa_n(x)$. By function extensionality, we can characterize dependent paths in a function type, which means we need to show:
\begin{equation*}
(p:P_\infty(\iota_{n+1}(a_n(x))))\to g(n+1,a_n(x),p) =^E_{(\kappa_n(x),1)} g(n,x,\transfib{P_\infty}{\kappa_n(x)}{p}).
\end{equation*}
Now for $p:P_\infty(\iota_{n+1}(a_n(x)))$, we can apply the path $r(p)$, which means we need to construct the following path (note that $r(p)$ is added to the path, since $g$ is a dependent function):
\begin{equation*}
g(n+1,a_n(x),p) =^E_{(\kappa_n(x),r(p))} g(n,x,F(p)).
\end{equation*}
We proceed by induction on $p$. If $p\equiv\iota_k(y)$ for $k:\textbf{N}$, $y:P_{(n+1)+k}(a_{n+1}^k(a_n(x)))$, then $F(p)\equiv \iota_{k+1}(u_{n,k}(y))$ and we need a path
\begin{equation*}
g_\ast(k,n+1,a_n(x),y) =^E_{(\kappa_n(x),r(\iota_k(y))))} g_\ast(k+1,n,x,u_{n,k}(y)).
\end{equation*}
Now the path $(\kappa_n(x),r(\iota_k(y)))\equiv\kappa^\ast_{n,k}(x,y)$, hence this dependent path is given by $\mu_{n,k}(x,y)$.
\emph{(path-path)} If $p$ varies over $\kappa_k(y)$, we need to give a dependent path in a family of dependent paths.
This is equivalent to filling the following dependent square in the family $E$,
which lies over the naturality square form by applying $\lam{p}(\kappa_n(x),r(p))$ to the path $\kappa_k(y)$.\footnote{The
left and right sides of the square are not quite correct,
the dependent function applied to $\kappa_k(y)$ are pathovers lying over $\kappa_k(y)$, and not $(1,\kappa_k(y))$.
However, pathovers lying over $\kappa_k(y)$ in the family $E(\iota_n(x))$ are equivalent to pathovers lying over $(1,\kappa_k(y))$ in the family $E$,
and this equivalence commutes with all operations we perform, therefore we omit them in this proof.
The following calculations are only type correct when these equivalences are inserted back.
Furthermore, we omit some other details. For example, if $p = q$,
then $\mapdep{g}{p}$ and $\mapdep{g}{q}$ have different types:
the former is a dependent path over $p$ and the latter one over $q$.
However, if you modify the path over which they lie, they become equal.
These ``modifications'' can be pushed down to the square in $(x:A_\infty)\times P_\infty(x)$,
and the proof still goes through. For the full details, consult the formal proof.}
\begin{xcenter}\begin{tikzcd}[column sep=25mm]
g_\ast(k+1,n+1,a_n(x),f_{(n+1)+k)}(y))
\ar[r,equal,"{\mu_{n,k+1}(x,f_{(n+1)+k}(y))}","{\kappa^\ast_{n,k+1}(x,f_{n+k}(y))}" swap]
\ar[d,equal,"{\mapdep{g(n+1,a_n(x))}{\kappa_k(y))})}","{(1,\kappa_k(y)}" swap] &
g_\ast(k+2,n,x,u_{n,(k+1)}(f_{(n+1)+k}(y)))
\ar[d,equal,"{\mapdep{g(n,x)\circ F}{\kappa_k(y))})}","{(1,\kappa_k(y)}" swap]\\
g_\ast(k,n+1,a_n(x),y)
\ar[r,equal,"{\mu_{n,k}(x,y)}","{\kappa^\ast_{n,k}(x,y)}" swap] &
g_\ast(k+1,n,x,u_{n,k}(y))
\end{tikzcd}\end{xcenter}
Below and to the left of each equal sign we give the path in $(x:A_\infty)\times P_\infty(x)$ over which the pathover lie. Above and to the right of each equal sign we give the value of the dependent path.
Now $\mapdep{g(n+1,a_n(x))}{\kappa_k(y)}$ (occurring in the left pathover) is equal to $\nu(k,n+1,a_n(x),y)$ by definition of $g$. On the right, we have ($\delta$ is the naturality of $u_{n,k}$)
\begin{align*}
\mapdep{g(n,x)\circ F}{\kappa_k(y)} &=
\mapdep{g(n,x)}{\map{F}{\kappa_k(y)}}\\
&= \mapdep{g(n,x)}{\map{\iota_{k+2}}{\delta(y)}\cdot\kappa_{k+1}(u_{n,k}(y))}\\
&= \mapdep{g_\ast(k+2,n,x)}{\delta(y)} \cdot \mapdep{g(n,x)}{\kappa_{k+1}(u_{n,k}(y))}\\
&= \mapdep{g_\ast(k+2,n,x)}{\delta(y)} \cdot \nu(k+1,n,x,u_{n,k}(y))\\
\end{align*}
Now we can move the first part of the expression to the top of the square, which means we need to fill the following squareover (where we made some arguments implicit).
\begin{center}\begin{tikzcd}[column sep=25mm]
g_\ast(f(y))
\ar[r,equal,"{\mu(f(y))}","{\kappa^\ast(f(y))}" swap]
\ar[d,equal,"{\nu(k,n+1,a_n(x),y)}","{(1,\kappa(y))}" swap] &
g_\ast(u(f(y)))
\ar[r,equal,"{\mapdep{g_\ast(k+2,n,x)}{\delta(y)}}","{(1,\delta(y))}" swap] &
g_\ast(f(u(y)))
\ar[d,equal,"{\nu(k+1,n,x,u_{n,k}(y))}","{(1,\kappa(u_{n,k}(y)))}" swap]\\
g_\ast(y)
\ar[rr,equal,"{\mu(y)}","{\kappa^\ast(y)}" swap] & &
g_\ast(u_{n,k}(y))
\end{tikzcd}\end{center}
Note that in this squareover $g$ does not occur, and $\nu$ only occurs on the left side (applied to $k$) and on the right side (applied to $k+1$). Therefore, the top, bottom and left side form a valid open box, and we define $\nu(k+1,n,x,u_{n,k}(y))$ to be the composition of \emph{this} open box. This inductively defines $\nu$, and makes the filler for this square automatic. This finishes the proof.
\end{proof}
\begin{proof}[Proof (of Theorem \ref{thm:colim_sm})]
We first define a map $$\beta: \big((x:A_\infty)\times P_\infty(x)\big) \to \mathsf{colim}(\msm{A}{P}).$$
We do this by induction on $x:A_\infty$ and $p:P_\infty(x)$ individually, so we get four cases again
(we do not use our newly defined induction principle, because we have not proven a computation rule for it).
\emph{(point-point)} Suppose $x:A_n$ and $y:P_{n+k}(a_n^k(x))$. We define
$$\beta(\iota_n(x),\iota_k(y))\ensuremath{\vcentcolon\equiv}\iota_{n+k}(a_n^k(x),y).$$
\emph{(point-path)} To show that the second argument respects $\kappa_k(y)$, we define
$$\mapfunc{\beta(\iota_n(x))}(\kappa_k(y))\vcentcolon=\kappa_{n+k}(a_n^k(x),y):\iota_{n+(k+1)}(a_n^{k+1}(x),f(y))=\iota_{n+k}(a_n^k(x),y).$$
\emph{(path-point)} To show that the first argument respects $\kappa_n(x)$, we need to give a dependent path
$$\beta(\iota_{n+1}(a_n(x)))=^{P_\infty({-})\to\mathsf{colim}(\msm{A}{P})}_{\kappa_n(x)}\beta(\iota_n(x)).$$
By function extensionality, this is equivalent to showing for $p:P_\infty(\iota_{n+1}(a_n(x)))$ that
$$\beta(\iota_{n+1}(a_n(x)),p)=\beta(\iota_n(x),\transfib{P_\infty}{\kappa_n(x)}{p}).$$
We apply $\apfunc{\beta(\iota_n(x))}^{-1}{(r(p))}$ on the right, so that we have to show
$$\beta(\iota_{n+1}(a_n(x)),p)=\beta(\iota_n(x),F(p)).$$
Now we apply induction on $p$. If $p\equiv\iota_k(y)$, then $F(p)\equiv \iota_{k+1}(u_{n,k}(y))$ and we need to show
$$\mu_{n,k}(x,y):\iota_{(n+1)+k}(a_{n+1}^k(a_n(x)),y)=\iota_{n+(k+1)}(a_n^{k+1},u_{n,k}(y)).$$
But the triples $((n+1)+k,a_{n+1}^k(a_n(x)),y)$ and $(n+(k+1),a_n^{k+1},u_{n,k}(y))$ are equal: the first two components by \autoref{lem:iterate_succ} and the last component because $u_{n,k}$ was defined by transporting along the equality of the first components. Let us call this equality $s$. So we define $\mu_{n,k}(x,y)$ by applying $\iota$ to $s$.
\emph{(path-path)} Suppose $p$ varies along $\kappa_k(y)$, we need to construct a proof of a pathover in an equality type. This is equivalent to filling the following square.
\begin{center}\begin{tikzcd}[column sep=25mm]
\beta(\iota_{n+1}(a_n(x)),\iota_{k+1}(f(y)))
\ar[r,equal,"{\mu_{n,k+1}(x,f(y))}"]
\ar[d,equal,"{\mapfunc{\beta(\iota_{n+1}(a_nx))}(\kappa_k(y)})"] &
\beta(\iota_n(x),\iota_{k+2}(u_{n,k+1}(f(y))))
\ar[d,equal,"{\mapfunc{\beta(\iota_n(x))\circ F}(\kappa_k(y))}"]\\
\beta(\iota_{n+1}(a_n(x)),\iota_k(y))
\ar[r,equal,"{\mu_{n,k}(x,y)}"] &
\beta(\iota_n(x),\iota_{k+1}(u_{n,k}(y)))
\end{tikzcd}\end{center}
By simplifying the left and right path, this reduces to
\begin{center}\begin{tikzcd}[column sep=10mm]
\beta(\iota(a(x)),\iota(f(y)))
\ar[r,equal,"{\mu(f(y))}"]
\ar[d,equal,"{\kappa(a^k(a(x)),y)}"] &
\beta(\iota(x),\iota(u(f(y))))
\ar[r,equal,"{\mapfunc{\beta(\iota(a(x)),\iota({-}))}(\kappa(y))}"] &[15mm]
\beta(\iota(x),\iota(f(u(y))))
\ar[d,equal,"{\kappa(a^k(x),u(y))}"]\\
\beta(\iota(a(x)),\iota_k(y))
\ar[rr,equal,"{\mu(y)}"] & &
\beta(\iota(x),\iota(u(y)))
\end{tikzcd}\end{center}
Now the concatenation of the two paths on the top reduces to the function $i\ensuremath{\vcentcolon\equiv}\lam{n}{x}{y}\iota_{n+1}(a(x),f(y))$ applied to $s$. Then the square is exactly the naturality square of the homotopy $\kappa:i\sim\iota$ applied to the path $s$. This finishes the definition of $\beta$.
Now we need to show that $\beta\circ\alpha\sim\idfunc$. Take $p:\mathsf{colim}(\msm{A}{P})$, we apply induction to $p$. If $p\equiv\iota_n(x,y)$, then the equality holds by reflexivity:
$$\beta(\alpha(p))\equiv\beta(\iota_n(x),\iota_0(y))\equiv\iota_n(x,y)\equiv p.$$
If $p$ varies over $\kappa_n(x,y)$, we need to fill a square with two degenerate sides, so we need to prove that
$\apfunc{\beta\circ\alpha}(\kappa_n(x,y))=\kappa_n(x,y).$
We can show this as follows.
\begin{align*}
&\mathrel{\hphantom{=}}\apfunc{\beta\circ\alpha}(\kappa_n(x,y))\\
&=\apfunc{\beta}(\kappa_n(x),r(\iota_0(f(y)))\cdot\kappa_0(y))\\
&=\apfunc{\beta}(\kappa'_n(x,y))\\
&=\mu_{n,1}(x,f(y))\cdot\apfunc{\beta(\iota_n(x))}^{-1}{(r(\iota_0(f(y))))}\cdot\apfunc{\beta(\iota_n(x))}(r(\iota_0(f(y)))\cdot\kappa_0(y))\\
&=\mu_{n,1}(x,f(y))\cdot\apfunc{\beta(\iota_n(x))}(\kappa_0(y))\\
&=\apfunc{\beta(\iota_n(x))}(\kappa_0(y))\\
&=\kappa_n(x,y)
\end{align*}
In the third step we use that $\apfunc{\beta}(p,q)=\mapdep{\beta}{p}(f(y))\cdot\mapfunc{\beta{\iota_n(x)}(q)}$ and in the fifth step that $\mu_{n,k}(x,y)=1$ for any \emph{numeral} $k$.
Finally we need to show that $\alpha\circ\beta\sim\idfunc$. Take $p:(x:A_\infty)\times P_\infty(x)$. We apply the induction principle proven in \autoref{thm:sigma-colim-induction} to $p$. Suppose that $p\equiv(\iota_n(x),\iota_0(y))$. Then the equality holds by reflexivity:
$$\alpha(\beta(p))\equiv\alpha(\iota_n(x,y))\equiv(\iota_n(x),\iota_0(y))\equiv p.$$
If $p$ varies over $\kappa'_n(x,y)$, then we have to show (similar to the proof $\beta\circ\alpha\sim\idfunc$) that
$$\apfunc{\alpha\circ\beta}(\kappa'_n(x,y))\kappa'_n(x,y).$$
But by the previous computation, $\apfunc{\beta}(\kappa'_n(x,y))=\kappa_n(x,y)$, so we have
$$\apfunc{\alpha\circ\beta}(\kappa'_n(x,y))=\apfunc{\alpha}(\kappa_n(x,y))=\kappa'_n(x,y).$$
This finishes the proof.
\end{proof}
\begin{cor}\label{cor:eq_colim}
Consider a sequence $(A_n,f_n)_n$. Then for any $a,a':A_n$ there is an equivalence
\begin{equation*}
\eqv{(\iota_n(a)=_{A_\infty}\iota_n(a'))}{\mathsf{colim}(f^k(a)=_{A_{n+k}}f^k(a'))}.
\end{equation*}
\end{cor}
\begin{proof
We first prove this for $n \equiv 0$.
Note that for any $a:A_0$, we have the diagram
\begin{equation*}
\begin{tikzcd}
\lam{a':A_0}a=a' \ensuremath{\operatorname{\mathsf{arrow}}}[r] \ensuremath{\operatorname{\mathsf{arrow}}}[d,->>] & \lam{a':A_1} f(a)=a' \ensuremath{\operatorname{\mathsf{arrow}}}[r] \ensuremath{\operatorname{\mathsf{arrow}}}[d,->>] & \lam{a':A_2} f^2(a)=a' \ensuremath{\operatorname{\mathsf{arrow}}}[r] \ensuremath{\operatorname{\mathsf{arrow}}}[d,->>] & \cdots \\
A_0 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_0"] & A_1 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_1"] & A_2 \ensuremath{\operatorname{\mathsf{arrow}}}[r,"f_2"] & \cdots
\end{tikzcd}
\end{equation*}
This defines a type family $P:A_\infty\to\ensuremath{\mathsf{Type}}\xspace$ with
$$P(\iota_n(a'))\ensuremath{\vcentcolon\equiv}\mathsf{colim}_k(f^{0\le n+k}(a)=_{A_{n+k}}f^k(a')).$$
Now we use \autoref{thm:colim_sm} to see that the total space of $P$ is contractible.
\begin{align*}
(a':A_\infty)\times P(a')
& \eqvsym \mathsf{colim}_n((a':A_n)\times f^{n}(a)=a') \\
& \eqvsym \mathsf{colim}_n(\ensuremath{\mathbf{1}}\xspace) \\
& \eqvsym \ensuremath{\mathbf{1}}\xspace.
\end{align*}
Since $\iota_0(\ensuremath{\operatorname{\mathsf{refl}}}{a}):P(\iota_0(a))$ and noting that $f^{0\le 0+k}(a)\equiv f^k(a)$ we can now conclude by the total space method to characterize the identity type that
$$(\iota_0(a)=_{A_\infty}\iota_0(a'))\simeq P(\iota_0(a'))\equiv\mathsf{colim}(f^k(a)=_{A_{0+k}}f^k(a')).$$
For general $n$, we use \autoref{lem:colim_shift_k}, which gives us an equivalence
$\mathsf{kshift\underline{~}equiv}_n: A_\infty\simeq \mathsf{colim}(S_n(A,f))$. For $a,a':A_n$ we can now compute:
\begin{align*}
(\iota_n(a)=_{A_\infty}\iota_n(a'))
& \eqvsym (\mathsf{kshift\underline{~}equiv}_n(\iota_n(a))=_{\mathsf{colim}(S_n(A,f))}\mathsf{kshift\underline{~}equiv}_n(\iota_n(a'))) \\
& \eqvsym (\iota_0(a)=_{\mathsf{colim}(S_n(A,f))}\iota_0(a')) \\
& \eqvsym \mathsf{colim}(S_n(f)^k(a)=_{S_n(A)_{0+k}}S_n(f)^k(a')) \\
& \eqvsym \mathsf{colim}(f^k(a)=_{A_{n+k}}f^k(a'))..
\end{align*}
The last equivalence comes from a natural equivalences of the sequences, because there is a dependent path between
$S_n(f)^k(a)$ and $f^k(a)$ over the canonical path that $n+(0+k)=n+k$.
\end{proof}
\begin{cor}\label{cor:fiber_functor}
Suppose given a natural transformation $\tau : \sequence{A'}{f'}\to\sequence{A}{f}$ and a point $a:A_n$. Then
$$\hfib{\tau_\infty}{\iota_n(a)}\simeq\ensuremath{\operatorname{\mathsf{colim}}}(\square\mathsf{fib}_\tau[a])\equiv\ensuremath{\operatorname{\mathsf{colim}}}_k(\hfib{\tau_{n+k}}{f^k(a)}).$$
\end{cor}
\begin{proof}
Consider the following diagram, where the equivalences on the top are given by \autoref{thm:colim_sm} and the fact that the total space of the fiber of a function is the domain of that function.
\begin{center}\begin{tikzcd}[column sep=10mm]
(x:A_\infty)\times (\mathsf{fib}_\tau)_\infty(x) \ar[r,"{\sim}"] \ar[dr,"{\pi_1}"] &
\ensuremath{\operatorname{\mathsf{colim}}}_k((x:A_n))\times \hfib{\tau_n}{x}) \ar[r,"{\sim}"]\ar[d,"p"] &
A'_\infty \ar[dl,"{\tau_\infty}"] \\
& A_\infty &
\end{tikzcd}\end{center}
This diagram commutes: the left triangle commutes by \autoref{thm:colim_sm} and the right triangle commutes by the 1-functoriality of the colimit, \autoref{lem:seq_colim_functor}. Therefore,
$$\hfib{\tau_\infty}{\iota_n(a)}\simeq\hfib{\pi_1}{\iota_n(a)}\simeq (\mathsf{fib}_\tau)_\infty(\iota_n(a))\equiv \ensuremath{\operatorname{\mathsf{colim}}}(\square\mathsf{fib}_\tau[a]).$$
\end{proof}
\begin{cor}\label{cor:trunc_colim}
Consider a sequence $\sequence{A}{f}$ and some $k\geq-2$.
\begin{enumerate}
\item\label{part:colim_is_trunc} If $A_n$ is $k$-truncated for all $n:\nat$, then $A_\infty$ is $k$-truncated.
\item\label{part:colim_trunc} We have an equivalence $$\trunc{k}{A_\infty}\simeq\ensuremath{\operatorname{\mathsf{colim}}}(\trunc{k}{A_n},\trunc{k}{f_n})_n.$$
\item\label{part:colim_is_connected} If $A_n$ is $k$-connected for all $n:\nat$, then $A_\infty$ is $k$-connected.
\item\label{part:functor_is_trunc_conn} Given a natural transformation $(\tau,H):\sequence Af \to \sequence{A'}{f'}$ such that $\tau_n$ is $k$-truncated ($k$-connected) for all $n$, then $\tau_\infty$ is $k$-truncated ($k$-connected).
\item\label{part:iota_is_trunc_conn} If $f_n$ is $k$-truncated ($k$-connected) for all $n$, then $\iota_0$ is $k$-truncated ($k$-connected).
\end{enumerate}
\end{cor}
\begin{rmk}
By \autoref{lem:colim_shift_k} we can generalize the quantification ``for all $n:\nat$'' in this Corollary to the weaker ``there exists an $m:\nat$ such that for all $n\ge m$''. In part \ref{part:iota_is_trunc_conn} the conclusion then becomes that $\iota_m$ is $k$-truncated ($k$-connected).
\end{rmk}
\begin{proof}\mbox{}
\begin{enumerate}
\item We prove this by induction on $k$. Suppose $k=-2$, then $f_n$ is an equivalence for all $n$. Therefore $A_\infty\simeq A_0$ by \autoref{lem:equiv_equiseq}, hence $A_\infty$ is contractible.
Now suppose $k\equiv k'+1$. Take $x,x' : A_\infty$, we need to show that $x=x'$ is $k'$-truncated. Since being truncated is a mere proposition, by induction on $x$ and $x'$ we may assume that $x\equiv\iota_n(a)$ and $x'\equiv\iota_m(a')$. Now
$\iota_n(a)=\iota_{\max(n,m)}(f^{n\le \max(n,m)}(a))$ and $\iota_m(a')=\iota_{\max(n,m)}(f^{m\le \max(n,m)}(a'))$, therefore the type $\iota_n(a)=\iota_m(a')$ is equivalent to
$$\iota_{\max(n,m)}(f^{n\le \max(n,m)}(a))=\iota_{\max(n,m)}(f^{m\le \max(n,m)}(a')).$$ Therefore it suffices to show that the latter equality type is $k'$-truncated. By \autoref{cor:eq_colim} we need to show that
$$\mathsf{colim}(f^\ell(f^{n\le \max(n,m)}(a))=f^\ell(f^{m\le \max(n,m)}(a')))_\ell$$
is $k'$-truncated, which follows from the induction principle and the fact that $A_{\max(n,m)+\ell}$ is $(k'+1)$-truncated.
\item From the functoriality of the sequential colimit, we get a function
$$A_\infty\to\ensuremath{\operatorname{\mathsf{colim}}}_n(\trunc{k}{A_n},\trunc{k}{f_n}).$$
Because the right hand side is $k$-truncated, this induces a map
$$g:\trunc{k}{A_\infty}\to\ensuremath{\operatorname{\mathsf{colim}}}(\trunc{k}{A_n},\trunc{k}{f_n})_n.$$
For the other direction, we define the function
$$h:\ensuremath{\operatorname{\mathsf{colim}}}(\trunc{k}{A_n},\trunc{k}{f_n})_n\to\trunc{k}{A_\infty}$$
by
$$h(\iota_n(\tproj{k}{a}))\ensuremath{\vcentcolon\equiv}\tproj{k}{\iota_n(a)}$$
and
$$\mapfunc{h}(\kappa_n(\tproj{k}{a}))\vcentcolon=\mapfunc{\tprojf{k}}(\kappa_n(a)).$$ It is straightforward to show that both $h\circ g$ and $g\circ h$ are homotopic to the identity.
\item Since $A_n$ is $k$-connected, $\trunc{k}{A_n}$ is contractible, and therefore $\ensuremath{\operatorname{\mathsf{colim}}}_n(\trunc{k}{A_n})\simeq \trunc{k}{A_\infty}$ is contractible.
\item A function is $k$-truncated ($k$-connected) whenever its fibers are $k$-truncated ($k$-connected).
Let $x:A_\infty$. We need to show a proposition, so we may assume that $x\equiv \iota_n(a)$ for some $a:A_n$. Now
$\hfib{\tau_\infty}{\iota_n(a)}\simeq\mathsf{colim}(\square\mathsf{fib}_\tau[a])$ by \autoref{cor:fiber_functor}. Since $\hfib{\tau_n}{x}$ is $k$-truncated ($k$-connected) for all $n$, we know that $\mathsf{colim}(\square\mathsf{fib}_\tau[a])$ is $k$-truncated ($k$-connected) for all $n$, by part \ref{part:colim_is_trunc} or \ref{part:colim_is_connected}.
\item Consider the natural transformation
\begin{center}\begin{tikzcd}[column sep=10mm]
A_0 \ar[r,equal] \ar[d,equal] &
A_0 \ar[r,equal] \ar[d,equal,"f"] &
A_0 \ar[r,equal] \ar[d,equal,"{f^{0\le2}}"] &
A_0 \ar[r,equal] \ar[d,equal,"{f^{0\le3}}"] &
\cdots \ar[r,equal] &
\ensuremath{\operatorname{\mathsf{colim}}}(A_0)_n \ar[d] \\
A_0 \ar[r] &
A_1 \ar[r] &
A_2 \ar[r] &
A_3 \ar[r] &
\cdots \ar[r] &
A_\infty
\end{tikzcd}\end{center}
The maps $f^{0\le n}:A_0\to A_n$ are $k$-truncated ($k$-connected) and form a natural transformation. Therefore, by part \ref{part:functor_is_trunc_conn} the map $f^{0\le\infty}:\ensuremath{\operatorname{\mathsf{colim}}}_n(A_0)\to A_\infty$ is $k$-truncated ($k$-connected). The fiber of $\iota_0$ over $x:A_\infty$ is the same as the fiber of $f^{0\le\infty}$ over $x$, and therefore $\iota_0$ is $k$-truncated ($k$-connected).
\end{enumerate}
\end{proof}
We can use this machinery, in particular \autoref{thm:colim_sm}, to define the localization for maps between $\omega$-compact types. We will omit the construction here, but this will be published in an upcoming preprint.
\chapter{Homotopy Theory}\label{cha:homotopy-theory}
As discussed in the introduction, one very useful application of HoTT is synthetic homotopy
theory. Many results in homotopy theory have been stated and proven in HoTT in a synthetic way. Most
of these results have also been formalized in a proof assistant. This is important,
because one of the advantages of HoTT is to make verification of proofs by a proof assistant
practically possible. Formalizing results that have been proved internally in HoTT provides more evidence for this.
In this chapter we will look at various topics in homotopy theory and give proofs for them in HoTT
that are fully checked by the Lean proof assistant. In \cref{sec:computing-pi3s2} we will describe
a formalization of the proof that $\pi_3(\S^2)=\ensuremath{\mathbb{Z}}\xspace$. This was already known to be provable in HoTT, but no fully
formalized proof has been given before. We will discuss some new properties proven about
Eilenberg-MacLane spaces in HoTT in \cref{sec:eilenb-macl-spac}, namely that the Eilenberg-MacLane
space functor induces an equivalence of categories. In \cref{sec:smash-product} we prove the
adjunction of the smash product and pointed maps, from which we can conclude that the smash product
is associative.
None of these results have been formalized before, even including formalization in foundations other than HoTT.
In fact, not much homotopy theory has been formalized in other foundations.
The most notable examples of formalizations are the formalization of basic properties of the fundamental group~\cite{zhan2017auto2}
and the formalization of singular homology theory~\cite[\texttt{Multivariate/homology.ml}]{hollight}.
\section{Computing \texorpdfstring{$\pi_3(\S^2)$}{pi\textunderscore3(S\textasciicircum2)}}%
\label{sec:computing-pi3s2}
Computing that $\pi_3(\S^2)=\ensuremath{\mathbb{Z}}\xspace$ has been done before in Homotopy Type Theory, but it has not been
formalized in a proof assistant before. In this section we will discuss some considerations of
formalizing the proof that $\pi_3(\S^2)=\ensuremath{\mathbb{Z}}\xspace$. The Hopf fibration was formalized in Lean by Ulrik
Buchholtz and was formalized before in Agda by Guillaume Brunerie. The remaining results are formalized by the author.
\subsection{The long exact sequence of homotopy groups}\label{sec:les-homotopy}
We start with an important result in homotopy theory, the long exact sequence of homotopy
groups.
This has been proven before in HoTT. Two different proofs are given in~\cite[Section 8.4]{hottbook}
and~\cite[Section 2.5.1]{brunerie2016spheres},
although these proofs have not been formalized. There have been previous
formalizations of parts of this
result~\cite{avigad2015limits,voevodsky2015lecture,unimath}; however none of
these formalizations are complete in the sense that they can be used to deduce
the results in this section.
The statement is as follows.
\begin{thm}[Long exact sequence of homotopy groups]\label{thm:les-homotopy}
Suppose $f : X \to Y$ is a pointed map. Then the following is an exact sequence
\begin{center}\begin{tikzpicture}[node distance=3cm, thick, node
distance=12mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$\pi_0(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$\pi_0(X)$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$\pi_0(F)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of Y] (OY) {$\pi_1(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of X] (OX) {$\pi_1(X)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of F] (OF) {$\pi_1(F)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OY] (O2Y) {$\pi_2(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OX] (O2X) {$\pi_2(X)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OF] (O2F) {$\pi_2(F)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 5mm of O2X] (dots) {$\vdots$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] node [above] (f){$\pi_0(f)$} (Y)
(F) edge[->] node [below] (f){$\pi_0(p_1)$} (X)
(OY) edge[->] node [left] (f){$\pi_0(\delta)\quad\mbox{}$} (F)
(OX) edge[->] node [above] (f){$\pi_1(f)$} (OY)
(OF) edge[->] node [below] (f){$\pi_1(p_1)$} (OX)
(O2Y) edge[->] node [left] (f){$\pi_1(\delta)\quad\mbox{}$} (OF)
(O2X) edge[->] node [above] (f){$\pi_2(f)$} (O2Y)
(O2F) edge[->] node [below] (f){$\pi_2(p_1)$} (O2X);
\end{tikzpicture}
\end{center}
Here $F\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{fib}}}_f$ is the fiber of $f$, $p_1:F\to X$ is the first projection, and
$\delta:\Omega Y \to F$ is defined in the proof.
\end{thm}
First of all, we have to carefully formulate the statement of this theorem in type theory. The naive
thing to do is to say that there is a sequence $A : \textbf{N} \to \ensuremath{\mathsf{Set}}\xspace^*$ and maps $f : (n : \textbf{N}) \to
A_{n+1} \to A_n$ such that
$$A_0\ensuremath{\vcentcolon\equiv}\pi_0(Y),\quad A_1\ensuremath{\vcentcolon\equiv}\pi_0(X),\quad A_2\ensuremath{\vcentcolon\equiv}\pi_0(F),$$
and so forth. Continuing, this means that
$$A_{3n}=\pi_n(Y),\quad A_{3n+1}=\pi_n(X),\quad A_{3n+2}=\pi_n(F).$$
However, there is no way to make these equalities definitional, the elimination principle for the
natural numbers does not allow for computation rules like that. This means that the map
$f_{3n}:A_{3n+1}\to A_{3n}$ cannot be compared directly to $\pi_n(f)$ since the domain and codomain
are note definitionally equal. Setting things up this way is possible, but makes reasoning about it
unnecessarily complicated. Instead, we change the indexing set, using $\textbf{N}\times\ensuremath{\operatorname{\mathsf{fin}}}_3$ instead of $\textbf{N}$. We will work with a general
notion of sequences with a flexible choice of indexing set.
\begin{defn}\label{def:chain-complex}
A \emph{successor structure} is a type $I$ with endomap $S : I \to I$ called the
\emph{successor}. We will write $i+n$ for $i:I$ and $n:\textbf{N}$ to mean iterated application of the
successor function, $i+n\ensuremath{\vcentcolon\equiv} S^n(i)$.
A \emph{chain complex} indexed by a successor structure $I$ is a family of pointed sets
$A : I \to \ensuremath{\mathsf{Set}}\xspace^*$ and maps $f : (i : I) \to A_{i+1} \to A_i$ with the property that
$(i : I) \to (a : A_{i+2}) \to f_i(f_{i+1}(a))=a_0^i$ where $a_0^i$ is the basepoint of $A_i$. We call a chain complex \emph{exact} or a \emph{long exact sequence} if
$$(i : I) \to (a : A_{i+1}) \to f_i(a) = a_0^i \to \|(a' : A_{i+2}) \times f_{i+1}(a')=a\|.$$
A \emph{type-valued chain complex} is the same, except that $A_i$ is only required to be a pointed type (not a pointed set). A type-valued chain complex is \emph{exact} or a \emph{type-valued exact sequence} if the above property holds without any propositional truncation, i.e. if
$$(i : I) \to (a : A_{i+1}) \to f_i(a) = a_0^i \to (a' : A_{i+2}) \times f_{i+1}(a')=a.$$
\end{defn}
\begin{rmk}
Note that a type-valued exact sequence gives part of the structure of a \emph{fiber sequence}. A \emph{fiber sequence} is a sequence where $A_{i+2}$ ``is'' the fiber of $f_i$. This means that $(A_{i+2},f_{i+1})=(\ensuremath{\operatorname{\mathsf{fib}}}_{f_i},p_1)$ for all $i$. Using univalence this can be unpacked in an equivalence and a commuting triangle.
In a type-valued exact sequence we just require two maps back and forth $A_{i+2}\leftrightarrow \ensuremath{\operatorname{\mathsf{fib}}}_{f_i}$ such that the corresponding triangles commute, but we do not require that these maps are mutual inverses. In the text below we will have sequences that are not fiber sequences, so we require this additional generality.
\end{rmk}
\begin{ex}\label{ex:succ_structure}
Some useful examples of successor structures are $(\textbf{N},\lam{n}n+1)$ and $(\ensuremath{\mathbb{Z}}\xspace,\lam{n}n+1)$. Sequences over these successor structures correspond to one-sided and two-sided infinite sequences. We can also mimic one-sided infinite sequences in the other direction using the successor structure $(\textbf{N},\lam{n}n-1)$ (with the convention that $0-1=0$). This has the disadvantage that there is one extra map $A_0\to A_0$. Whenever we use $\textbf{N}$ as successor structure in this section, we use $\lam{n}n+1$ as its successor.
Furthermore, if $N$ is a successor structure and $k : \textbf{N}$, then we define a successor structure on
$N \times \ensuremath{\operatorname{\mathsf{fin}}}_{k+1}$ by defining
$$S(n,i)\ensuremath{\vcentcolon\equiv}\begin{cases}(n+1,0) & \text{if $i=k$}\\
(n,i+1) & \text{otherwise}\end{cases}$$
Note that $n+1$ is addition in the successor structure $N$.
\end{ex}
We now build the long exact sequence of homotopy groups in five steps. The order of these steps is somewhat arbitrary and can be altered.
We perform the 0-truncation of the sequence as the last step, so that the intermediate sequences contain as much information as possible.
\begin{enumerate}[(1)]
\item First we define the fiber sequence of $f$.
\item Then we show that this sequence is equivalent to a sequence involving iterated loop spaces.
\item We fix some negation signs in the exact sequence.
\item We index the sequence over $\textbf{N}\times\ensuremath{\operatorname{\mathsf{fin}}}_3$.
\item We 0-truncate the sequence to obtain the sequence in \cref{thm:les-homotopy}.
\end{enumerate}
We first need some lemmas about fibers.
\begin{lem}\label{lem:fibers}
Suppose given a pointed map $f : A \to^* B$. Let $p_1 : \ensuremath{\operatorname{\mathsf{fib}}}_f \to^* A$ be the first projection. Then there is a pointed natural equivalence $e_f:\ensuremath{\operatorname{\mathsf{fib}}}_{p_1} \simeq^* \Omega B$.
Furthermore, if $q_1:\ensuremath{\operatorname{\mathsf{fib}}}_{p_1}\to\ensuremath{\operatorname{\mathsf{fib}}}_f$ is the first projection, we get a commuting square
\begin{center}\begin{tikzpicture}[node distance=2cm, thick, node
distance=12mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\Omega A$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$\Omega B$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$\ensuremath{\operatorname{\mathsf{fib}}}_{q_1}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$\ensuremath{\operatorname{\mathsf{fib}}}_{p_1}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node {$-\Omega f$} (tr)
edge[<-] node {$e_{p_1}$} (bl)
(tr) edge[<-] node {$e_f$} (br)
(bl) edge[->] node {$r_1$} (br);
\end{tikzpicture}
\end{center}
where $r_1$ is (also) the first projection. We write $-\Omega f$ for the map $\Omega f \o ({-})^{-1}$.
\end{lem}
\begin{proof}
The underlying equivalence is the following composite
\begin{align*}
\ensuremath{\operatorname{\mathsf{fib}}}_{p_1}&\simeq((a,p):\ensuremath{\operatorname{\mathsf{fib}}}_f)\times a = a_0\\
&\simeq (a:A)\times a = a_0 \times f(a)=b_0\\
&\simeq f(a_0)=b_0\\
&\simeq b_0=b_0 \equiv \Omega B
\end{align*}
This equivalence sends $((a, p), q):\ensuremath{\operatorname{\mathsf{fib}}}_{p_1}$ (with $p : fa=b_0$ and $q:a=a_0$) to
$f_0^{-1} \cdot \ap fq \cdot p$. So there is a path
$$r(a,p,q): e_f((a, p), q)=f_0^{-1} \cdot f(q^{-1}) \cdot p.$$
This path satisfies $r(a_0,f_0,1)=1$ (equality is type correct since $e_f((a_0, f_0), q)\equiv f_0^{-1} \cdot q$). We also have $e_f^{-1}(p)=((a_0,f_0\cdot p),1)$ for $p:\Omega B$.
Now $e$ respects the basepoint, because $$e(a_0,f_0,1)=f_0^{-1}\cdot f_0=1.$$
We will not prove naturality here, since it is not required for the results in this section.
For the commuting square, we will prove that
$$h:e_f \o r_1 \o e_{p_1}^{-1}\sim^*-\Omega f$$
For the underlying homotopy, we compute for $p:\Omega A$
\begin{align*}
e_f(r_1(e_{p_1}^{-1} p))&=e_f(r_1(((a_0,f_0),1 \cdot p),1))\\
&=e_f((a_0,f_0),p)\\
&=f_0^{-1} \cdot f(p^{-1}) \cdot f_0 \equiv -\Omega f(p).
\end{align*}
To show that $h$ respects the basepoint, suppose that $p\equiv1$. In that case, the first two steps of the above equation becomes definitional equalities. Since we know that $r(a_0,f_0,1)=1$, the last equality is also reflexivity. Since the maps $e_f\o r\o e_{p_1}^{-1}$ and $-\Omega f$ respect the basepoints using the same path, this shows that $h$ is a pointed homotopy, which finishes the proof.
\end{proof}
\subsubsection{Step 1}
Denote $\ensuremath{\operatorname{\mathsf{arrow}}}^*\ensuremath{\vcentcolon\equiv}(X\ Y : \U^*_i)\times (X \to^* Y)$. We define $F:\ensuremath{\operatorname{\mathsf{arrow}}}^*\to\ensuremath{\operatorname{\mathsf{arrow}}}^*$ by
$F(X,Y,f)\ensuremath{\vcentcolon\equiv}(\ensuremath{\operatorname{\mathsf{fib}}}_f,X,p_1)$.
Given a pointed map $f:X\to^* Y$, we define its fiber sequence $A:\textbf{N}\to\U$ by $A_n\ensuremath{\vcentcolon\equiv} p_2(F^n(X,Y,f))$, and we define $f_n:A_{n+1}\to A_n$ by $p_3(F^n(X,Y,f))$ (which is well-typed, since $A_{n+1}\equiv p_1(F^n(X,Y,f))$ by unfolding the definition of $F$). It is easy to show that $(A_n,f_n)_n$ is a type-valued exact sequence, since $A_{n+2}$ is (definitionally) the fiber of $f_n$.
Note that by \autoref{lem:fibers} there is a pointed equivalence $e_f:A_3\simeq^* \Omega Y$. We define the diagonal map $\delta\ensuremath{\vcentcolon\equiv} p_1\o e_f^{-1} : \Omega Y \to \ensuremath{\operatorname{\mathsf{fib}}}_f$.
\subsubsection{Step 2}
Define the sequence $B:\textbf{N}\to\U$ and $g_n:B_{n+1}\to B_n$ by
\begin{align*}
B_0&\ensuremath{\vcentcolon\equiv} Y&&\\
B_1&\ensuremath{\vcentcolon\equiv} X & g_0&\ensuremath{\vcentcolon\equiv} f\\
B_2&\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{fib}}}_f & g_1&\ensuremath{\vcentcolon\equiv} p_1\\
B_{n+3}&\ensuremath{\vcentcolon\equiv} \Omega B_n & g_2&\ensuremath{\vcentcolon\equiv} \delta\\
&& g_{n+3}&\ensuremath{\vcentcolon\equiv} -\Omega g_n
\end{align*}
Note that $g_2$ has the correct type, since $A_3\equiv B_3$.
Now we can show that $(B,g)$ is a type-valued exact sequence by showing that it is equivalent to $(A,f)$.
\begin{lem}
There is a natural equivalence $(A_n, f_n)_n\simeq(B_n,g_n)_n$. This means that there are pointed equivalences $\eta_n:A_n\simeq^* B_n$ such that for all $n:\textbf{N}$ we have $$\eta_n\o f_n\sim^*g_n\o \eta_{n+1}.$$
\end{lem}
\begin{proof}
We define the equivalence $\eta_n$ by induction on $n$. Note that $A_k\equiv B_k$ for $k=0,1,2$. Now suppose we have an equivalence $\eta_k: A_k\simeq B_k$. Then by \autoref{lem:fibers} we have
$$A_{k+3}\equiv\ensuremath{\operatorname{\mathsf{fib}}}_{f_{k+1}}\stackrel{e_{f_k}}\simeq \Omega A_k\stackrel{\Omega\eta_k}\simeq \Omega B_k\equiv B_{k+3}.$$
We also show the naturality by induction on $n$.\\
For $n\equiv0$ we have $\idfunc[Y]\o f\sim^* f\o\idfunc[X].$\\
For $n\equiv1$ we have $\idfunc[X]\o p_1\sim^* p_1\o\idfunc[\ensuremath{\operatorname{\mathsf{fib}}}_f].$\\
For $n\equiv2$ we have
$$\idfunc[\ensuremath{\operatorname{\mathsf{fib}}}_f]\o p_1\equiv p_1\sim^* (p_1 \o e_f^{-1}) \o e_f \sim^* \delta\o (\Omega\idfunc[Y] \o e_f).$$
Now suppose the naturality holds for $k$, then we get the following diagram.
\begin{center}\begin{tikzpicture}[node distance=2cm, thick, node
distance=12mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\Omega B_{k+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$\Omega B_k$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (l) {$\Omega A_{k+1}$};
\ensuremath{\operatorname{\mathsf{node}}} (r) at (tr |- l) {$\Omega A_k$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of l] (bl) {$A_{k+4}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (r |- bl) {$A_{k+3}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node {$-\Omega g_k$} (tr)
edge[<-] node {$\Omega\eta_{k+1}$} (l)
(tr) edge[<-] node {$\Omega\eta_k$} (r)
(l) edge[->] node {$-\Omega f_k$} (r)
edge[<-] node {$e_{f_{k+1}}$} (bl)
(r) edge[<-] node {$e_{f_k}$} (br)
(bl) edge[->] node {$f_{k+3}$} (br);
\end{tikzpicture}\end{center}
The bottom square can be filled by the second part of \autoref{lem:fibers}. The top square can be filled by applying the functor $\Omega$ to the naturality for $k$ and then noticing that $({-})^{-1} \o \Omega\eta_k\sim^* \Omega\eta_k \o ({-})^{-1},$ which is easily proven for an arbitrary pointed map.
\end{proof}
\subsubsection{Step 3}
We now remove the inverses in our sequence. More precisely, we define a second sequence $h_n:B_{n+1}\to B_n$ by
$$h_0\ensuremath{\vcentcolon\equiv} f\qquad h_1\ensuremath{\vcentcolon\equiv} p_1\qquad h_2\ensuremath{\vcentcolon\equiv} \delta\qquad h_{n+3}\ensuremath{\vcentcolon\equiv} \Omega h_n.$$
To show that $(B,h)$ is a type-valued exact sequence we use the following lemma.
\begin{lem}\label{lem:LESstep3}
Suppose $N$ is a successor structure and $(B,g)$ is a type-valued exact sequence over $N$. Suppose $h_n:B_{n+1}\to^* B_n$ is another sequence of maps, and suppose that there are pointed maps $e_n, \ell_n, r_n : B_n \to^* B_n$ such that $e_n$ is an equivalence and the following diagrams commute as homotopies (not necessarily pointed):
\begin{center}\begin{tikzpicture}[node distance=2cm, thick, node distance=12mm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$B_{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$B_{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of bl] (br) {$B_n$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node {$h_n$} (br)
(bl) edge[->] node {$e_{n+1}$} (tl)
edge[->] node {$g_n$} (br);
\ensuremath{\operatorname{\mathsf{node}}}[right = of tr] (tl) {$B_{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$B_n$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$B_{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$B_n$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node {$h_n$} (tr)
edge[->] node {$\ell_{n+1}$} (bl)
(tr) edge[->] node {$e_n$} (br)
(bl) edge[->] node {$h_n$} (br);
\ensuremath{\operatorname{\mathsf{node}}}[right = of tr] (tl) {$B_{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$B_n$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$B_{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$B_n$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node {$h_n$} (tr)
edge[->] node {$e_{n+1}$} (bl)
(tr) edge[<-] node {$r_n$} (br)
(bl) edge[->] node {$h_n$} (br);
\end{tikzpicture}\end{center}
Then $(B,h)$ is a type-valued exact sequence over $N$.
\end{lem}
\begin{proof}
First we need to show that for $x:B_{n+2}$ we have $h_n(h_{n+1}(x))=b_0^n$. We compute
\begin{align*}
h_n(h_{n+1}(x))&=r_n(h_n(e_{n+1}(h_{n+1}(x))))\\
&=r_n(g_n(h_{n+1}(x)))\\
&=r_n(g_n(g_{n+1}(e_{n+2}^{-1}(x))))\\
&=r_n(b_0^n)\\
&=b_0^n.
\end{align*}
For exactness, suppose that $y:B_{n+1}$ such that $h_n(y)=b_0^n$. Then
$g_n(e_{n+1}^{-1}(y))=h_n(y)=b_0^n$, therefore, by exactness of $g$ there (purely) exists
an $x:B_{n+2}$ such that $g_{n+1}(x)=e_{n+1}^{-1}(y)$. Now we compute
\begin{align*}
h_{n+1}(\ell_{n+2}(e_{n+2}(x)))&=e_{n+1}(h_{n+1}(e_{n+2}(x)))\\
&=e_{n+1}(g_{n+1}(x))\\
&=e_{n+1}(e_{n+1}^{-1}(y))\\
&=y.
\end{align*}
This finishes the proof.
\end{proof}
\begin{lem}
The sequence $(B,h)$ is a type-valued exact sequence.
\end{lem}
\begin{proof}
We first define for $k\ge2$ we the pointed equivalence
$\epsilon_n^k:B_n\simeq^* B_n$ by induction on $n$. For $n\le k$
$\epsilon_n^k\ensuremath{\vcentcolon\equiv}\idfunc:B_n\simeq^* B_n$ we define
$\epsilon_{n+3}^k\ensuremath{\vcentcolon\equiv} -\Omega\epsilon_n^k:B_{n+3}\simeq^* B_{n+3}$ for $n+3>k$.
Now define $e_n\ensuremath{\vcentcolon\equiv}\epsilon_n^3$ and $\ell_n\ensuremath{\vcentcolon\equiv}\epsilon_n^4$ and $r_n\ensuremath{\vcentcolon\equiv}\epsilon_n^2$.
We apply \autoref{lem:LESstep3} using these equivalences to obtain the desired result. To do this we need to check three commuting triangles. We will check $h_n\o e_{n+1}\sim g_n$, the other two proofs are similar. Apply induction on $n$. For $n=0,1,2$ it is trivial, reducing to $g_n\o\ensuremath{\operatorname{\mathsf{id}}} \sim g_n$. Suppose the homotopy is true for $n=k$. Then
$$h_{k+3}\o e_{k+4}\equiv \Omega h_k \o -\Omega e_{k+1} \sim -\Omega (h_k \o e_{k+1}) \sim -\Omega g_k\equiv g_{k+3}.$$
\end{proof}
\subsubsection{Step 4}
We now define a type-valued chain complex over $\textbf{N}\times\ensuremath{\operatorname{\mathsf{fin}}}_3$, which has a successor structure by \autoref{ex:succ_structure}. Let $\rho_X$ be the equivalence $\Omega^{n+1}X \simeq^* \Omega^n(\Omega X)$. We now define the sequence $C:\textbf{N}\times\ensuremath{\operatorname{\mathsf{fin}}}_3$ and $k_n:C_{n+1}\to C_n$ by
\begin{align*}
C_{(n,0)}&\ensuremath{\vcentcolon\equiv} \Omega^nY& k_{(n,0)}&\ensuremath{\vcentcolon\equiv} \Omega^nf\\
C_{(n,1)}&\ensuremath{\vcentcolon\equiv} \Omega^nX& k_{(n,1)}&\ensuremath{\vcentcolon\equiv} \Omega^np_1\\
C_{(n,2)}&\ensuremath{\vcentcolon\equiv} \Omega^n\ensuremath{\operatorname{\mathsf{fib}}}_f & k_{(n,2)}&\ensuremath{\vcentcolon\equiv} \Omega^n\delta \o \rho_X
\end{align*}
In a diagram, $(C,k)$ looks like the following.
\begin{center}\begin{tikzpicture}[node distance=3cm, thick, node
distance=12mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$Y$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$X$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$F$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of Y] (OY) {$\Omega Y$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of X] (OX) {$\Omega X$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of F] (OF) {$\Omega F$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OY] (O2Y) {$\Omega^2Y$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OX] (O2X) {$\Omega^2X$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OF] (O2F) {$\Omega^2F$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 5mm of O2X] (dots) {$\vdots$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] node [above] (f){$f$} (Y)
(F) edge[->] node [below] (f){$p_1$} (X)
(OY) edge[->] node [left] (f){$\delta\quad\mbox{}$} (F)
(OX) edge[->] node [above] (f){$\Omega f$} (OY)
(OF) edge[->] node [below] (f){$\Omega p_1$} (OX)
(O2Y) edge[->] node [left] (f){$\Omega\delta\quad\mbox{}$} (OF)
(O2X) edge[->] node [above] (f){$\Omega^2f$} (O2Y)
(O2F) edge[->] node [below] (f){$\Omega^2p_1$} (O2X);
\end{tikzpicture}
\end{center}
There is an equivalence $e:\textbf{N} \simeq \textbf{N}\times\ensuremath{\operatorname{\mathsf{fin}}}_3$ that sends $n$ to its quotient and remainder when dividing $n$ by 3. The proof of the following lemma is straightforward and omitted.
\begin{lem}
The sequence $(B,h)$ is naturally equivalent to $(C,k)$ over the equivalence $e$.
Therefore, $(C,k)$ is a type-valued exact sequence.
\end{lem}
\subsubsection{Step 5}
If we 0-truncate the sequences at step 4, we get the sequence $(D,\ell)\ensuremath{\vcentcolon\equiv}(\|C\|_0,\|k\|_0)$. This is exactly the sequence in \autoref{thm:les-homotopy}. It is now easy to show that this is a long exact sequence.
\begin{proof}[Proof of \autoref{thm:les-homotopy}]
First note that it is a chain complex by the following computation:
$$\ell_n \o \ell_{n+1}\sim \|k_n \o k_{n+1}\|_0\sim \|0\|_0 \sim 0.$$
To show that it is exact, suppose given $x : D_{n+1}$ and $p:\ell_n(x)=d_0^n$. We need to construct an element in a proposition, so we may assume by induction that $x\equiv |y|_0$. Now the type of $p$ reduces to $|k_n(y)|_0=|c_0^n|_0$, which is equivalent to $\|k_n(y)=c_0^n\|_{-1}$ by the characterization of the identity type in truncations. Therefore, the latter type is inhabited, and by induction, we may assume that we have a path $k_n(y)=c_0^n$. By exactness of $(C,k)$ we get an element $z:C_{n+2}$ such that $q:k_{n+1}(z)=y$. Now we can find $|z|_0:D_{n+2}$ and the path $\mapfunc{|{-}|_0}(q):\ell_{n+1}(|z|_0)=x$, showing exactness.
\end{proof}
\subsection{Computation of homotopy groups}
An important application of the long exact sequence of homotopy groups comes in combination with the Hopf fibration. Combining these tools, we can compute more homotopy groups of spheres. The Hopf fibration was constructed in~\cite[Theorem 8.5.1]{hottbook} and has been formalized by Ulrik Buchholtz. We will not give the construction here.
\begin{thm}[Hopf Fibration] \label{thm:hopf}
There is a pointed map $\S^3\to\S^2$ with fiber $\S^1$.
\end{thm}
The quaternionic Hopf fibration has also been constructed in HoTT and formalized in Lean~\cite{buchholtz2016cayleydickson}. This gives a fibration $\S^7\to \S^4$ with fiber $\S^3$.
\begin{cor}\label{cor:homotopy-group-spheres-1}
$\pi_2(\S^2)=\ensuremath{\mathbb{Z}}\xspace$ and $\pi_n(\S^3)=\pi_n(\S^2)$ for $n\geq 3$.
\end{cor}
\begin{proof}
We know by the connectedness of spheres that $\pi_1(\S^3)$ and $\pi_2(\S^3)$ are trivial, and by the truncatedness of the circle that $\pi_k(\S^1)$ is trivial for $k>1$ and $\ensuremath{\mathbb{Z}}\xspace$ for $k=1$. We now get the following long exact sequence, from which the result immediately follows.
\begin{center}\begin{tikzpicture}[node distance=3cm, thick, node
distance=12mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$\ensuremath{\mathbb{Z}}\xspace$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 8mm of Y] (OY) {$\pi_2(\S^2)$};
\ensuremath{\operatorname{\mathsf{node}}} (OX) at (X |- OY) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (OF) at (F |- OY) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 8mm of OY] (O2Y) {$\pi_3(\S^2)$};
\ensuremath{\operatorname{\mathsf{node}}} (O2X) at (X |- O2Y) {$\pi_3(\S^3)$};
\ensuremath{\operatorname{\mathsf{node}}} (O2F) at (F |- O2Y) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 8mm of O2Y] (O3Y) {$\pi_4(\S^2)$};
\ensuremath{\operatorname{\mathsf{node}}} (O3X) at (X |- O3Y) {$\pi_4(\S^3)$};
\ensuremath{\operatorname{\mathsf{node}}} (O3F) at (F |- O3Y) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 5mm of O3X] (dots) {$\vdots$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] (Y)
(F) edge[->] (X)
(OY) edge[->] (F)
(OX) edge[->] (OY)
(OF) edge[->] (OX)
(O2Y) edge[->] (OF)
(O2X) edge[->] (O2Y)
(O2F) edge[->] (O2X)
(O3Y) edge[->] (O2F)
(O3X) edge[->] (O3Y)
(O3F) edge[->] (O3X);
\end{tikzpicture}
\end{center}
\end{proof}
The last ingredient we need is the Freudenthal Suspension
Theorem. This has been formalized before by Dan Licata in Agda, and our
formalization is a direct port of that proof to Lean. For the proof we refer to~\cite[Section 8.6]{hottbook}.
\begin{thm}[Freudenthal Suspension Theorem] \label{thm:freudenthal}
Suppose that $X$ is $n$-connected. Then $\|X\|_{2n}\simeq \|\Omega\Sigma X\|_{2n}$.
\end{thm}
We can combine these results to compute the following homotopy groups.
\begin{cor}\label{cor:homotopy-group-spheres-2}
$\pi_n(\S^n)=\ensuremath{\mathbb{Z}}\xspace$ and $\pi_3(\S^2)=\ensuremath{\mathbb{Z}}\xspace$
\end{cor}
\begin{proof}
Note that $\S^n$ is $(n-1)$-connected. Therefore, by the Freudenthal suspension theorem we have
$$\|\S^n\|_{2(n-1)}\simeq \|\Omega\S^{n+1}\|_{2(n-1)}.$$
For $n\geq 2$ we have $2(n-1)\geq n$, and therefore we also have
$$\|\S^n\|_{n}\simeq \|\Omega\S^{n+1}\|_{n}.$$
Taking the $n$-th homotopy group, we get
$$\pi_n(\S^n)\simeq \pi_{n+1}(\S^{n+1}).$$
Combining this with \autoref{cor:homotopy-group-spheres-1}, we also get $\pi_3(\S^2)\simeq\ensuremath{\mathbb{Z}}\xspace$, as desired.
\end{proof}
\section{Eilenberg-MacLane Spaces}\label{sec:eilenb-macl-spac}
In this section we give an important equivalence between groups and Eilenberg-MacLane spaces~\cite{eilenberg1945spaces}.\footnote{Some of the contents of this section have been published in~\cite{buchholtz2018groups}.
The work in this section is joint work with Ulrik Buchholtz and Egbert Rijke.} Eilenberg-MacLane space are play an important role in homotopy theory, since they are spaces with simple homotopy groups. Therefore, they can be used to build up more complicated spaces with complicated homotopy groups. Also, they can be used to define homology and cohomology in HoTT, see Sections \ref{sec:spectral-sequence-cohomology} and \ref{sec:spectral-sequence-homology}.
We prove in this section that the category of $n$-connected $(n+1)$-truncated pointed types is equivalent to the
category of groups for $n = 0$ and the category of abelian groups for $n \geq 1$.
If $G$ is a (pre-)groupoid, the groupoid quotient is a higher inductive type with constructors
\begin{inductive}
\texttt{HIT} $\ensuremath{\operatorname{\mathsf{groupoid-quotient}}}(G) :=$ \\
$\bullet\ i : G_0 \to \ensuremath{\operatorname{\mathsf{groupoid-quotient}}}(G)$; \\
$\bullet\ p : (x\ y : G_0) \to \ensuremath{\operatorname{\mathsf{hom}}}(x,y)\to x=y$; \\
$\bullet\ q : (x\ y\ z : G_0) \to (g : \ensuremath{\operatorname{\mathsf{hom}}}(y,z)) \to (f : \ensuremath{\operatorname{\mathsf{hom}}}(x,y)) \to
p(g \circ f) = p(f) \cdot p(g)$; \\
$\bullet\ \ensuremath{\varepsilon} : \istrunc{1}(\ensuremath{\operatorname{\mathsf{groupoid-quotient}}}(G))$.
\end{inductive}
The groupoid quotient can be constructed purely from homotopy pushouts.
The untruncated version was constructed in \autoref{sec:non-recursive-2}.
Then we can apply the 1-truncated afterwards, and we can also construct truncations from homotopy pushouts~\cite{rijke2017join}.
In~\cite{licata2014em} the authors define Eilenberg-MacLane spaces. We use the same approach as in
that paper. We first quickly review the results in that paper.
\subsection{Construction of Eilenberg-MacLane spaces}
If $G$ is any group, the 1-dimensional Eilenberg-MacLane space $K(G,1)$ can be defined by viewing
$G$ as a groupoid, and taking the groupoid quotient of $G$. It is not hard to see that $K(G,1)$ is
0-connected and 1-truncated. Using an encode-decode proof, we can show that $\Omega K(G,1) \simeq G$
and that this equivalence sends concatenation to multiplication. Hence the composite
$\pi_1K(G,1)\simeq \|G\|_0\simeq G$ is a group isomorphism.
If $G$ is abelian, the higher Eilenberg-MacLane spaces can be defined recursively
as $$K(G,n+1):\equiv \|\Sigma K(G,n)\|_{n+1}$$ for $n\geq 1$. This definition is slightly different
than the one given in~\cite{licata2014em}, where $K(G,n+1)$ was defined using the iterated
suspension as $\|\Sigma^n K(G,1)\|_{n+1}$. We chose to modify the definition, since a lot of
properties of Eilenberg-MacLane spaces are proven by induction on $n$, so it is more convenient to
have $K(G,n+1)$ defined directly in terms of $K(G,n)$.
It is easy to show that $K(G,n)$ is $(n-1)$-connected and $n$-truncated. It is trickier to show that
$\Omega K(G,n+1)\simeq K(G,n)$. This is done separately for $n=1$ and for $n\geq 2$.
For $n=1$ we need the result that for every type $X$ with a coherent h-structure, the type $\|\Sigma
X\|_2$ is a \emph{delooping} of $X$, which means that $\Omega \|\Sigma X\|_2 \simeq X$. If $G$ is abelian, then $K(G,1)$ can be equipped with a coherent h-structure,
showing that $\Omega K(G,2)\simeq K(G,1)$.
For $n\geq 2$, this can be done using the Freudenthal suspension theorem,
\cref{thm:freudenthal}. Then the equivalence follows from the following chain of equivalences:
$$\Omega K(G,n+1)\equiv \Omega\|\Sigma K(G,n)\|_{n+1}\simeq \|\Omega\Sigma K(G,n)\|_n\simeq
\|K(G,n)\|_n\simeq K(G,n).$$ The Freudenthal Suspension Theorem is applied in the third step, which
is allowed since $K(G,n)$ is $(n-1)$-connected and $n \leq 2(n-1)$ for $n\geq2$.
This finishes the proof sketch that $\operatorname{K-\mathsf{loop}}(G,n):\Omega K(G,n+1)\simeq K(G,n)$. By induction,
$\Omega^n K(G,n+1)\simeq K(G,1)$, hence we get the following group isomorphism
$\pi_{n+1}(K(G,n+1))\simeq \pi_1(K(G,1))\simeq G$.
\subsection{Uniqueness}
In this section we prove that Eilenberg-MacLane spaces are unique, which means that if $X$ and $Y$
are both $(n-1)$-connected, $n$-truncated pointed types such that $\pi_n(X)\simeq \pi_n(Y)$, then
$X\simeq Y$. Note that from these assumptions one can show that $\pi_k(X)\simeq 1\simeq \pi_k(Y)$
for $k < n$ since $X$ and $Y$ are $(n-1)$-connected, but also for $k > n$ since $X$ and $Y$ are
$n$-truncated. Hence from the assumptions we actually have that $\pi_k(X)\simeq\pi_k(Y)$ for all
natural numbers $k$.
This is similar to Whitehead's Theorem, which states that if $f : X \to Y$ is a pointed map
that induces an equivalence on all homotopy groups, then $f$ is an equivalence. Whitehead's Theorem
is not true in general, but it is true under the assumption that both $X$ and $Y$ are $n$-truncated
for some $n$. For the special case that $X$ and $Y$ are both $(n-1)$-connected and $n$-truncated one
does not need to find a map between $X$ and $Y$ to show that they are equivalent, as long as they
have isomorphic homotopy groups.
We first give an elimination principle for $K(G,n)$.
\begin{defn}\label{def:Kelim}
Suppose that $X$ is an $n$-truncated pointed type, and suppose that for some group
$G$ there is an map $\phi : G \to \Omega^n X$ that sends multiplication to
concatenation. Then there is a pointed map $\operatorname{K-\mathsf{elim}}(\phi, n) : K(G,n)\to X$.
\end{defn}
\begin{proof}[Construction]
We construct this by induction on $n$.
For $n=1$ this follows directly from the induction principle of $K(G,1)$. For $n=k+1>1$ we can define the group homomorphism $\widetilde\phi$ as the composite $G \xrightarrow{\phi} \Omega^{k+1} X \simeq \Omega^k(\Omega X)$, and apply the induction hypothesis to get a map
$\operatorname{K-\mathsf{elim}}(\widetilde\phi, k):K(G,k)\to^* \Omega X$. By the adjunction $\Sigma\dashv\Omega$ we get a pointed map $\Sigma K(G,k)\to^* X$, and by the elimination principle of the truncation we get a map
$K(G,k+1)\equiv\|\Sigma K(G,k)\|_{k+1}\to^* X$.
\end{proof}
\begin{lem}\label{lem:Kelim1}
There is a pointed homotopy making the following diagram commute.
\begin{center}
\begin{tikzpicture}[node distance=3cm,
thick,main node/.style={font=\sffamily\bfseries}]
\ensuremath{\operatorname{\mathsf{node}}}[main node] (Kn) at (0,0) {$K(G,n)$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (OK) at (4,0) {$\Omega K(G,n+1)$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (OX) at (2,-2) {$\Omega X$};
\path[every node/.style={font=\sffamily\small}]
(Kn) edge [->] node {$\sim$} (OK)
edge [->] node [below left] {$\operatorname{K-\mathsf{elim}}(\widetilde\phi,n)$} (OX)
(OK) edge [->] node [below right] {$\Omega(\operatorname{K-\mathsf{elim}}(\phi,n+1))$} (OX);
\end{tikzpicture}
\end{center}
\end{lem}
\begin{proof}
This follows by unwinding the definition of the function $\operatorname{K-\mathsf{elim}}(\phi,n+1)$ in terms of $\operatorname{K-\mathsf{elim}}(\phi,n)$.
\end{proof}
\begin{lem}\label{lem:Kelim2}
The following diagram commutes.
\begin{center}
\begin{tikzpicture}[node distance=3cm,
thick,main node/.style={font=\sffamily\bfseries}]
\ensuremath{\operatorname{\mathsf{node}}}[main node] (OK) at (0,0) {$\Omega^nK(G,n)$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (G) at (4,0) {$G$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (OX) at (2,-2) {$\Omega^n X$};
\path[every node/.style={font=\sffamily\small}]
(OK) edge [->] node {$\sim$} (G)
edge [->] node [below left] {$\Omega^n(\operatorname{K-\mathsf{elim}}(\phi,n))$} (OX)
(OX) edge [<-] node [below right] {$\phi$} (G);
\end{tikzpicture}
\end{center}
\end{lem}
\begin{proof}
This follows by repeatedly applying \autoref{lem:Kelim1}.
\end{proof}
\begin{thm}\label{thm:em-unique}
Suppose that $X$ is an $(n-1)$-connected $n$-truncated pointed type, and suppose that for some group
$G$ there is an equivalence $\phi : G \simeq \Omega^n X$ that sends multiplication to
concatenation. Then the map $\operatorname{K-\mathsf{elim}}(\phi, n) : K(G,n)\to X$ is an equivalence.
In particular this means that if $X$ is an $(n-1)$-connected $n$-truncated pointed type, and there is a group isomorphism $e : \pi_n(X) \simeq G$, then $X\simeq^* K(G,n)$.
\end{thm}
\begin{proof}
We apply Whitehead's principle for truncated types. This states that a \emph{weak equivalence} (a map inducing an isomorphism on all homotopy groups) between truncated types is an equivalence. The proof can be found in~\cite[Theorem 8.8.3]{hottbook}. Since both $K(G,n)$ and $X$ are $(n-1)$-connected and $n$-truncated, the map $\operatorname{K-\mathsf{elim}}(\phi,n)$ trivially induces an isomorphism on all homotopy groups for all levels other than $n$. It also induces an isomorphism on level $n$ by \autoref{lem:Kelim2}. This finishes the proof.
\end{proof}
\begin{cor}
The type of $(n-1)$-connected, $n$-truncated pointed types is equivalent to the type of groups for
$n=1$ and equivalent to the type of abelian groups for $n \geq 2$.
\end{cor}
\begin{proof}
The maps back and forth are $K({-},n)$ and $\pi_n$. The composites are homotopic to the identity map, since $\pi_n(K(G,n))\simeq G$ and $K(\pi_n(X),n)\simeq^* X$ (the last equivalence comes from \autoref{thm:em-unique}).
\end{proof}
\subsection{Equivalence of categories}
\begin{defn}
If $\phi : G \to H$ is a homomorphism between groups, then there is a pointed map $K(\phi, n) :
K(G, n) \to K(H, n)$. This action is functorial, i.e. it respects composition and identity maps.
\end{defn}
\begin{proof}[Construction]
The functorial action comes from \autoref{def:Kelim}. We omit the proof of the other properties.
\end{proof}
To show that we get the desired equivalence of categories, we need to fill the following naturality squares. We will omit the proofs here.
\begin{center}
\begin{tikzpicture}[node distance=3cm,
thick,main node/.style={font=\sffamily\bfseries}]
\ensuremath{\operatorname{\mathsf{node}}}[main node] (KG) at (0,0) {$\pi_n(K(G,n))$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (KH) at (4,0) {$\pi_n(K(H,n))$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (G) at (0,-4) {$G$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (H) at (4,-4) {$H$};
\path[every node/.style={font=\sffamily\small}]
(KG) edge [->] node {$\pi_n(K(\phi,n))$} (KH)
(G) edge [->] node [left] {$\sim$} (KG)
edge [->] node {$\phi$} (H)
(H) edge [->] node [right] {$\sim$} (KH);
\end{tikzpicture}
\qquad
\begin{tikzpicture}[node distance=3cm,
thick,main node/.style={font=\sffamily\bfseries}]
\ensuremath{\operatorname{\mathsf{node}}}[main node] (X) at (0,0) {$X$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (Y) at (4,0) {$Y$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (KX) at (0,-4) {$K(\pi_n(X),n)$};
\ensuremath{\operatorname{\mathsf{node}}}[main node] (KY) at (4,-4) {$K(\pi_n(Y),n)$};
\path[every node/.style={font=\sffamily\small}]
(X) edge [->] node {$f$} (Y)
(KX) edge [->] node [left] {$\sim$} (X)
edge [->] node {$K(\pi_n(f),n)$} (KY)
(KY) edge [->] node [right] {$\sim$} (Y);
\end{tikzpicture}
\end{center}
These diagrams show the following result.
\begin{thm}\label{thm:EM-equiv-categories}
$K({-},n)$ is an equivalence from the category of $(n-1)$-connected $n$-truncated pointed types to
the category of groups (for $n=1$) or abelian groups (for $n\geq2$).
\end{thm}
\begin{rmk}
In particular this shows that the type of pointed maps between two $(n-1)$-connected $n$-truncated types is a set.
This is a special case of the more general fact that the type of pointed maps from an $n$-connected type to a $(n+k+1)$-truncated type is $k$-truncated (for $n\geq-1$).
\end{rmk}
\begin{rmk}
It would be interesting, but a lot more work, to do this one level up. In that case, it should be possible to show that crossed modules or 2-groups correspond to pointed connected 2-types. Furthermore, pointed $(n-2)$ connected $n$-types should correspond to braided 2-groups for $n=3$ and to symmetric 2-groups for $n\geq 4$. A start of this project was given in~\cite{raumer2016doublegroupoids}.
\end{rmk}
\section{The Smash Product}\label{sec:smash-product}
In this section we will discuss the smash product and its
properties.\footnote{The work in this section is joint work with Stefano
Piceghello. Parts of this section are based on ideas from Robin Adams, Marc
Bezem, Ulrik Buchholtz and Egbert Rijke.} The smash product has many uses in
homotopy theory. It can be used to define generalized homology theory (see
\autoref{sec:spectral-sequence-homology}) and it is used to define the cup
product for cohomology~\cite[Section 5.1]{brunerie2016spheres}.
The goal is to prove that the smash product defines a \emph{1-coherent symmetric
monoidal product on pointed types}~\cite[Definition 4.1.1]{brunerie2016spheres},
which we repeat in \autoref{def:symmonprod}. Our proof strategy is to show that
the smash product is left adjoint to pointed maps and then use a Yoneda-style
argument to show that we get a 1-coherent symmetric monoidal product.
This proof is known in 1-category theory~\cite[Chapter 2, Theorem
5.3]{eilenberg1966closedcategories}. Suppose given a closed category\footnote{A
\emph{closed category} is a category with internal hom-objects. We can view pointed
types as a higher closed category, where the internal hom-object is the type of pointed
maps, pointed by the constant map.} $\mathcal C$ with internal hom $[{-},{-}]:\mathcal C^{\mathrm{op}}\times \mathcal C\to \mathcal C$.
Moreover suppose that for every $A,B:\mathcal C$ the functor $[A,[B,{-}]]:\mathcal C\to \mathcal C$ is
representable as a $\mathcal C$-enriched functor. This means that there is an object
$A\otimes B:\mathcal C$ and a $\mathcal C$-enriched natural transformation $[A\otimes
B,C]\cong [A,[B,C]]$. Then $\mathcal C$ is a monoidal closed category. We will spell
out the precise formulation for pointed types in \autoref{def:naturality}, where we will call
$\ensuremath{\mathsf{Type}}\xspace^*$-enriched functors \emph{pointed functors} and $\ensuremath{\mathsf{Type}}\xspace^*$-enriched
natural transformations \emph{pointed natural transformations}.
In this section we will prove two main claims.
\begin{itemize}
\item We prove that $A\wedge B$ represents the functor $A\to^* B \to^* ({-})$ on
pointed types. In other words, that we have a natural equivalence
$$(A\wedge B \to^* C)\simeq^* (A\to^* B \to^* C).$$
\item We prove that if we have a \emph{pointed} natural equivalence
$$(A\wedge B \to^* C)\simeq^* (A\to^* B \to^* C),$$
then the smash product forms a 1-coherent symmetric
monoidal product on pointed types.
\end{itemize}
There is still a gap in this argument: we still need to show that the natural
equivalence above is a pointed natural equivalence. We did not manage to do
this, because of the high level of the path algebra involved, but we do not
expect theoretical difficulties.
In this section, all types, maps, homotopies and equivalences are pointed,
unless mentioned otherwise. We will denote pointed homotopies using equalities
in diagrams. We will start with defining some categorical properties of pointed
types. We will use the notation established in \autoref{sec:pointed}.
\subsection{The Category of Pointed Types}\label{sec:pointed-category}
\begin{defn}\label{def:naturality}
Suppose we are given $F:\ensuremath{\mathsf{Type}}\xspace^*\to\ensuremath{\mathsf{Type}}\xspace^*$. We say that $F$ is a \emph{1-coherent functor} if
\begin{itemize}
\item $F$ acts on pointed maps: given $f : A \to A'$, there is a pointed map $Ff : F(A) \to F(A');$
\item it respects identities: $F(\ensuremath{\operatorname{\mathsf{id}}}_A)\sim \ensuremath{\operatorname{\mathsf{id}}}_{FA};$
\item it respects composition: $F(f' \o f) \sim Ff' \o Ff.$
\end{itemize}
We will call a 1-coherent functor a \emph{functor} for short.\footnote{While
this is an abuse of terminology, it will not cause confusion in practice. Note that internally in the language of HoTT
it is an open problem whether we can even formulate the type of fully coherent functors.}
We say that a functor $F$ is a \emph{pointed functor} if moreover
$F\ensuremath{\mathbf{1}}\xspace=\ensuremath{\mathbf{1}}\xspace$, where $\ensuremath{\mathbf{1}}\xspace$ is the unit type (which is the zero object in
pointed types). In this case we can show that
$F(\ensuremath{\mathbf{0}}_{A,B})=\ensuremath{\mathbf{0}}_{FA,FB}$, where $\ensuremath{\mathbf{0}}_{A,B}$ is the constant map.
Let $F$, $G$ be functors of pointed types and suppose that $\theta$ is a
family of pointed maps $(X : \ensuremath{\mathsf{Type}}\xspace^*) \to F(X) \to G(X)$. We say that $\theta$
is a (1-coherent) \emph{natural transformation} or \emph{natural} if for every $f : A \to
B$ there is a diagram:
\begin{center}
\begin{tikzcd}
F(A)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "F(f)"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\theta_A"]
& F(B)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "\theta_B"]
\\
G(A)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "G(f)"]
& G(B)
\end{tikzcd}
\end{center}
That is, a pointed homotopy
\[p_\theta(f) : \theta_B \o F(f) \sim G(f) \o \theta_A.\]
We say that $\theta$ is \emph{pointed natural} if $\theta$ is natural and $p_\theta(\ensuremath{\mathbf{0}}) = (p_\theta)_0$, where
\[(p_\theta)_0 : G(\ensuremath{\mathbf{0}}) \o \theta_A \sim \ensuremath{\mathbf{0}} \o \theta_A \sim \ensuremath{\mathbf{0}} \sim \theta_B \o \ensuremath{\mathbf{0}} \sim \theta_B \o F(\ensuremath{\mathbf{0}})\]
is the canonical proof of the pointed homotopy $G(\ensuremath{\mathbf{0}}) \o \theta_A \sim \theta_B \o F(\ensuremath{\mathbf{0}})$.
For $n$-ary functions $F:\ensuremath{\mathsf{Type}}\xspace^*\to\cdots\to\ensuremath{\mathsf{Type}}\xspace^*$ we define functoriality
similarly. We say that transformations between $n$-ary functors are natural if
they are natural in all arguments.
\end{defn}
\begin{rmk}
We could define a notion of \emph{weak naturality}, which is like naturality,
but where the homotopy is not required to be pointed. However, this is
generally ill-behaved. For example, if $\theta$ is weakly natural, neither
$X\to\theta$ nor $\theta\to X$ needs to be weakly natural.
\end{rmk}
\begin{defn}\label{def:symmonprod}
A 1-coherent symmetric monoidal product for pointed types is a binary operation $\otimes:\ensuremath{\mathsf{Type}}\xspace^*\to\ensuremath{\mathsf{Type}}\xspace^*\to\ensuremath{\mathsf{Type}}\xspace^*$ that is functorial. Explicitly, this means that
\begin{itemize}
\item Given $f : A \to A'$ and $g : B \to B'$, there is a map $f\otimes g : A \otimes B \to A' \otimes B.$
\item It respects identities: $\ensuremath{\operatorname{\mathsf{id}}}_A\otimes\ensuremath{\operatorname{\mathsf{id}}}_B\sim \ensuremath{\operatorname{\mathsf{id}}}_{A\otimes B}.$
\item It respects composition: $(f' \o f) \otimes (g' \o g) \sim (f' \otimes g') \o (f \otimes g).$
\end{itemize}
Furthermore, there is a pointed type $I$ and natural equivalences
\begin{itemize}
\item $\alpha : (A \otimes B) \otimes C \simeq A \otimes (B \otimes C)$ (associativity of the smash product);
\item $\lambda : I \otimes B \simeq B$ (left unitor for the smash product);
\item $\gamma : A \otimes B \simeq B \otimes A$ (braiding for the smash product).
\end{itemize}
With pointed homotopies filling the following three diagrams.
\begin{center}
\begin{tikzcd}
&((A \otimes B) \otimes (C \otimes D))
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, "\alpha"]
\\
(((A \otimes B) \otimes C) \otimes D)
\ensuremath{\operatorname{\mathsf{arrow}}}[ru, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\alpha \otimes D"]
&& (A \otimes (B \otimes (C \otimes D)))
\\
((A \otimes (B \otimes C)) \otimes D)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, "\alpha"]
&& (A \otimes ((B \otimes C) \otimes D))
\ensuremath{\operatorname{\mathsf{arrow}}}[u, swap, "A \otimes \alpha"]
\end{tikzcd}
\end{center}
\begin{center}
\begin{tikzcd}
((I \otimes A) \otimes B)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, swap, "\lambda \otimes B"]
&& (I \otimes (A \otimes B))
\ensuremath{\operatorname{\mathsf{arrow}}}[dl, "\lambda"]
\\
& (A \otimes B)
\end{tikzcd}
\end{center}
\begin{center}
\begin{tikzcd}
((A \otimes B) \otimes C)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\gamma \otimes C"]
&(A \otimes (B \otimes C))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "\gamma"]
& ((B \otimes C) \otimes A)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "\alpha"]
\\
((B \otimes A) \otimes C))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "\alpha"]
& (B \otimes (A \otimes C))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "B \otimes \gamma"]
& (B \otimes (C \otimes A))
\end{tikzcd}
\end{center}
\end{defn}
We have a version of the Yoneda Lemma for pointed types.
\begin{lem}[Yoneda]\label{lem:yoneda}
Let $A$, $B$ be pointed types, and assume, for all pointed types $X$, a pointed equivalence $\phi_X : (B \to X) \simeq (A \to X)$, natural in $X$, i.e. for all $f : X \to X'$ there is a homotopy \[ p_\phi(f) : (A \to f) \o \phi_X \sim \phi_X' \o (B \to f) \]
Then there exists a pointed equivalence $\psi_\phi : A \simeq B$.
\end{lem}
\begin{proof}
We define $\psi_\phi \ensuremath{\vcentcolon\equiv} \phi_B(\idfunc[B]) : A \to B$ and $\psi_\phi^{-1} \ensuremath{\vcentcolon\equiv} \phi_A^{-1}(\idfunc[A])$. The given naturality square for $X \ensuremath{\vcentcolon\equiv} B$ and $g \ensuremath{\vcentcolon\equiv} \psi_\phi^{-1}$ yields $\psi_\phi^{-1} \o \phi_B (\idfunc[B]) \judgeq \psi_\phi^{-1} \o \psi_\phi \sim \phi_A (\psi_\phi^{-1} \o \idfunc[B]) \judgeq \phi_A (\phi_A^{-1} (\idfunc[A])) \sim \idfunc[A]$, and similarly for the inverse composition.
\end{proof}
\begin{lem}\label{lem:yoneda-pointed}
Assume $A$, $B$, $\phi_X$ and $p$ as in \autoref{lem:yoneda}, and assume moreover that $\phi_X$ is pointed natural. Then there is a pointed homotopy $(\psi_\phi \to X) \sim \phi_X$.
\end{lem}
\begin{proof}
Let $f : B \to X$. The underlying homotopy is obtained by:
\begin{align*}
(\psi_\phi \to X)(f) &\judgeq f \o \psi_\phi\\
&\sim \phi_X (f \o \idfunc) &&\text{(by $p_\phi(f)(\idfunc)$)}\\
&\sim \phi_X (f) &&\text{(by $\mapfunc{\phi_X}(\mathsf{u}_f)$)}
\end{align*}
To show that this is a pointed homotopy, we need to prove that the following diagram commutes:
\begin{center}
\begin{tikzcd}[column sep=4em]
(\psi_\phi \to X)(\ensuremath{\mathbf{0}})
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals, "p_\phi(\ensuremath{\mathbf{0}})(\idfunc)\cdot\mapfunc{\phi_X}(\mathsf{u}_\ensuremath{\mathbf{0}})"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, equals, swap, "\mathsf{z}_{\psi_\phi}"]
&&\phi_X(\ensuremath{\mathbf{0}})
\ensuremath{\operatorname{\mathsf{arrow}}}[dl, equals, "(\phi_X)_0"]
\\
&\ensuremath{\mathbf{0}}
\end{tikzcd}
\end{center}
where the top-left expression is definitionally equal to $\ensuremath{\mathbf{0}} \o \phi_X(\idfunc)$, the horizontal path comes from the underlying homotopy and $(\phi_X)_0$ is the canonical path from $\phi_X(\ensuremath{\mathbf{0}})$ to $\ensuremath{\mathbf{0}}$. Since $\phi_X$ is pointed natural, we have that
$p_{\phi_X}(\ensuremath{\mathbf{0}})(\idfunc) = (p_{\phi_X})_0(\idfunc)$, which is the concatenation:
\begin{align*}
\ensuremath{\mathbf{0}}\o \phi_X(\idfunc)
&= \ensuremath{\mathbf{0}} &&\text{(by $\mathsf{z}_{q_X(\idfunc)}$)}\\
&= \phi_X(\ensuremath{\mathbf{0}}) &&\text{(by $(\phi_X)_0^{-1}$)}\\
&= \phi_X(\ensuremath{\mathbf{0}}\o 1) &&\text{(by $(\mapfunc{\phi_X}(\mathsf{z}_{\idfunc}))^{-1}$)}
\end{align*}
The diagram then commutes by cancellation of inverses and using that $\mathsf{z}_{\idfunc} = \mathsf{u}_\ensuremath{\mathbf{0}}$.
\end{proof}
\subsection{Basic Properties of the Smash Product}\label{sec:smash-basic}
\begin{defn}
The smash of $A$ and $B$ is the HIT generated by the point constructor $(a,b)$ for $a:A$ and $b:B$
and two auxiliary points $\ensuremath{\operatorname{\mathsf{auxl}}},\ensuremath{\operatorname{\mathsf{auxr}}}:A\wedge B$ and path constructors $\ensuremath{\operatorname{\mathsf{gluel}}}_a:(a,b_0)=\ensuremath{\operatorname{\mathsf{auxl}}}$
and $\ensuremath{\operatorname{\mathsf{gluer}}}_b:(a_0,b)=\ensuremath{\operatorname{\mathsf{auxr}}}$ (for $a:A$ and $b:B$). $A\wedge B$ is pointed with point $(a_0,b_0)$.
\end{defn}
\begin{rmk}
This definition of $A\wedge B$ is basically the pushout of
$\ensuremath{\mathbf{2}}\xspace\leftarrow A+B\to A \times B$. A more traditional definition of $A\wedge B$ is the pushout
$\ensuremath{\mathbf{1}}\xspace\leftarrow A\vee B\to A \times B$; here $\vee$ denotes the wedge product, which can be
equivalently described as either the pushout $A\leftarrow \ensuremath{\mathbf{1}}\xspace\to B$ or
$\ensuremath{\mathbf{1}}\xspace\leftarrow \ensuremath{\mathbf{2}}\xspace\to A + B$. These two definitions of $A\wedge B$ are equivalent, because in
the following diagram the top-left square and the top rectangle are pushout squares, hence the
top-right square is a pushout square by applying the pushout lemma. Another application of the
pushout lemma then states that the two definitions of $A\wedge B$ are equivalent.
\begin{center}
\begin{tikzcd}
\ensuremath{\mathbf{2}}\xspace \ensuremath{\operatorname{\mathsf{arrow}}}[r]\ensuremath{\operatorname{\mathsf{arrow}}}[d] & A+B \ensuremath{\operatorname{\mathsf{arrow}}}[r]\ensuremath{\operatorname{\mathsf{arrow}}}[d] & \ensuremath{\mathbf{2}}\xspace \ensuremath{\operatorname{\mathsf{arrow}}}[d] \\
\ensuremath{\mathbf{1}}\xspace \ensuremath{\operatorname{\mathsf{arrow}}}[r] & A\vee B \ensuremath{\operatorname{\mathsf{arrow}}}[r]\ensuremath{\operatorname{\mathsf{arrow}}}[d] & \ensuremath{\mathbf{1}}\xspace \ensuremath{\operatorname{\mathsf{arrow}}}[d] \\
& A\times B \ensuremath{\operatorname{\mathsf{arrow}}}[r] & A\wedge B
\end{tikzcd}
\end{center}
\end{rmk}
\begin{lem}\label{lem:smash-general}
The smash product is functorial: if $f:A\to A'$ and $g:B\to B'$, then
$f\wedge g:A\wedge B\to A'\wedge B'$. We write $A\wedge g$ or $f\wedge B$ if one of the
functions is the identity function. Moreover, if $p:f\sim f'$ and $q:g\sim g'$, then $p\wedge q:f\wedge g\sim f'\wedge g'$; this operation preserves reflexivities, symmetries and transitivies. We will write $p \wedge g$ or $f \wedge q$ if one of the homotopies is reflexivity.
\end{lem}
\begin{lem}\label{lem:interchange}
The smash product preserves composition, which gives rise to the interchange law:
\[i:(f_2 \o f_1)\wedge (g_2 \o g_1) \sim f_2 \wedge g_2 \o f_1 \wedge g_1\]
for maps $A_1\xrightarrow{f_1}A_2\xrightarrow{f_2}A_3$ and $B_1\xrightarrow{g_1}B_2\xrightarrow{g_2}B_3$.
\end{lem}
\begin{proof}
Let us denote the basepoints of $A_i$ and $B_i$ with $a_i$ and $b_i$ respectively. We first apply induction on the paths that all the maps in the statement respect the basepoint. We verify the underlying homotopy of $i$ by induction on terms $x$ of the domain $A_1 \wedge B_1$ of the two maps; this can be defined on point constructors $(a,b)$, $\ensuremath{\operatorname{\mathsf{auxl}}}$ and $\ensuremath{\operatorname{\mathsf{auxr}}}$ to be the identity path. If $x$ varies over $\ensuremath{\operatorname{\mathsf{gluel}}}_a$, we need to fill the following square:
\begin{equation}\label{eq:i-gluel}
\begin{tikzcd}
(f_2(f_1(a)), b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[r,equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals,"\mapfunc{(f_2 \o f_1)\wedge (g_2 \o g_1)}(\ensuremath{\operatorname{\mathsf{gluel}}}_a)"]
& (f_2(f_1(a)), b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"\mapfunc{f_2 \wedge g_2 \o f_1 \wedge g_1}(\ensuremath{\operatorname{\mathsf{gluel}}}_a)"]
\\
\ensuremath{\operatorname{\mathsf{auxl}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[r,swap,equals,"1"]
&\ensuremath{\operatorname{\mathsf{auxl}}}
\end{tikzcd}
\end{equation}
This reduces to proving that
\[\mapfunc{(f_2(f_1(a)),-)}(g_2\o g_1)_0 \cdot \ensuremath{\operatorname{\mathsf{gluel}}}_{f_2(f_1(a))} = \mapfunc{(f_2(f_1(a)),-)}(\mapfunc{g_2}{(g_1)}_0 \cdot {(g_2)}_0) \cdot \ensuremath{\operatorname{\mathsf{gluel}}}_{f_2(f_1(a))}\]
Since we assumed that ${(g_1)}_0$ and ${(g_2)}_0$ are the identity path, the claim is easily verified. The case for $x$ varying over $\ensuremath{\operatorname{\mathsf{gluer}}}_b$ is entirely analogous, giving the square:
\begin{equation}\label{eq:i-gluer}
\begin{tikzcd}
(a_3, g_2(g_1(b))
\ensuremath{\operatorname{\mathsf{arrow}}}[r,equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals,"\mapfunc{(f_2 \o f_1)\wedge (g_2 \o g_1)}(\ensuremath{\operatorname{\mathsf{gluer}}}_b)"]
& (a_3, g_2(g_1(b))
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"\mapfunc{f_2 \wedge g_2 \o f_1 \wedge g_1}(\ensuremath{\operatorname{\mathsf{gluer}}}_b)"]
\\
\ensuremath{\operatorname{\mathsf{auxr}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[r,swap,equals,"1"]
&\ensuremath{\operatorname{\mathsf{auxr}}}
\end{tikzcd}
\end{equation}
The resulting homotopy is pointed, as $i(a_1,b_1) \judgeq 1$ and the proofs that the two maps respect the basepoint are assumed to be the identity path.
\end{proof}
\begin{lem}\label{lem:smash-zero}
There are homotopies
\begin{align*}
t_g : \ensuremath{\mathbf{0}}\wedge g\sim \ensuremath{\mathbf{0}} && t'_f : f\wedge\ensuremath{\mathbf{0}}\sim\ensuremath{\mathbf{0}}
\end{align*}
such that the following diagrams
commute for given homotopies $p : g\sim g'$ and $q : f\sim f'$.
\begin{equation}\label{eq:t-triangles}
\begin{tikzcd}
\ensuremath{\mathbf{0}}\wedge g
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals,"1\wedge p"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr,equals,swap,"t_g"]
&& \ensuremath{\mathbf{0}}\wedge g'\ensuremath{\operatorname{\mathsf{arrow}}}[dl,equals,"t_{g'}"]
&f\wedge \ensuremath{\mathbf{0}}
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals,"q\wedge 1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr,equals,swap, "t'_f"]
&& f'\wedge \ensuremath{\mathbf{0}}\ensuremath{\operatorname{\mathsf{arrow}}}[dl,equals,"t'_{f'}"]
\\
& \ensuremath{\mathbf{0}}
&&& \ensuremath{\mathbf{0}}
\end{tikzcd}
\end{equation}
\end{lem}
\begin{proof}
We will define the homotopy $t_g : \ensuremath{\mathbf{0}} \wedge g$, with $\ensuremath{\mathbf{0}} : A_1 \to A_2$ and $g : B_1 \to B_2$ (with the notational convention for the basepoints as in \autoref{lem:interchange}); the definition for $t'_f$ is analogous. First, we apply induction on the path that $g$ respects the basepoint. The underlying homotopy of $t_g$ is given by induction on terms $x : A_1 \wedge B_1$. On point constructors, we define:
\begin{align*}
t_g (a,b) &\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{gluer}}}_{g(b)} \cdot \ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1} && : (a_2, g(b)) = (a_2, b_2)\\
t_g (\ensuremath{\operatorname{\mathsf{auxl}}}) &\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{gluel}}}_{a_2}^{-1} && : \ensuremath{\operatorname{\mathsf{auxl}}} = (a_2, b_2)\\
t_g (\ensuremath{\operatorname{\mathsf{auxr}}}) &\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1} && : \ensuremath{\operatorname{\mathsf{auxr}}} = (a_2, b_2)
\end{align*}
If $x$ varies over $\ensuremath{\operatorname{\mathsf{gluel}}}_a$, after some reductions, we need to fill the following square:
\begin{equation}\label{eq:t-gluel}
\begin{tikzcd}[column sep=7em]
(a_2, g(b_1))
\ensuremath{\operatorname{\mathsf{arrow}}}[r,equals,"\ensuremath{\operatorname{\mathsf{gluer}}}_{b_2} \cdot \ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals, "\ensuremath{\operatorname{\mathsf{gluel}}}_{a_2}"]
& (a_2, b_2)
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"1"]
\\
\ensuremath{\operatorname{\mathsf{auxl}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[r,swap,equals, "\ensuremath{\operatorname{\mathsf{gluel}}}_{a_2}^{-1}"]
& (a_2, b_2)
\end{tikzcd}
\end{equation}
Similarly, if $x$ varies over $\ensuremath{\operatorname{\mathsf{gluer}}}_b$, we need to fill the following square:
\begin{equation}\label{eq:t-gluer}
\begin{tikzcd}[column sep=7em]
(a_2, g(b))
\ensuremath{\operatorname{\mathsf{arrow}}}[r,equals,"\ensuremath{\operatorname{\mathsf{gluer}}}_{g(b)} \cdot \ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals, "\ensuremath{\operatorname{\mathsf{gluer}}}_{g(b)}"]
& (a_2, b_2)
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"1"]
\\
\ensuremath{\operatorname{\mathsf{auxr}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[r,swap,equals, "\ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1}"]
& (a_2, b_2)
\end{tikzcd}
\end{equation}
The squares in (\ref{eq:t-gluel}) and (\ref{eq:t-gluer}) can both be filled by simple path algebra. The resulting homotopy is pointed, as $t_g(a_1,b_1)$ is equal to the identity path and the proof that $g$ respects the basepoint is also assumed to be the identity path. Finally, for $p : g \sim g'$, the diagram on the left in (\ref{eq:t-triangles}) commutes by induction on $p$.
\end{proof}
\begin{lem}\label{lem:smash-coh}
Suppose that we have maps $A_1\xrightarrow{f_1}A_2\xrightarrow{f_2}A_3$ and $B_1\xrightarrow{g_1}B_2\xrightarrow{g_2}B_3$
and suppose that either $f_1$ or $f_2$ is constant. Then there are two homotopies
$(f_2 \o f_1)\wedge (g_2 \o g_1)\sim \ensuremath{\mathbf{0}}$, one of which uses the interchange law and one that does not. These two homotopies are equal. Specifically, the following two diagrams commute:
\begin{center}
\begin{tikzcd}
(f_2 \o \ensuremath{\mathbf{0}})\wedge (g_2 \o g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, equals, "i"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, "\mathsf{z}' \wedge (g_2 \o g_1)"]
&(f_2 \wedge g_2)\o (\ensuremath{\mathbf{0}} \wedge g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, equals, "(f_2 \wedge g_2) \o t_{g_1}"]
\\
& (f_2 \wedge g_2)\o \ensuremath{\mathbf{0}}
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals, "\mathsf{z}'"]
\\
\ensuremath{\mathbf{0}}\wedge (g_2 \o g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[r,equals, swap, "t_{g_2 \o g_1}"]
& \ensuremath{\mathbf{0}}
\end{tikzcd}
\qquad
\begin{tikzcd}
(\ensuremath{\mathbf{0}} \o f_1)\wedge (g_2 \o g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, equals, "i"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, "\mathsf{z} \wedge (g_2 \o g_1)"]
& (\ensuremath{\mathbf{0}} \wedge g_2)\o (f_1 \wedge g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals, "t_{g_2} \o (f_1 \wedge g_1)"]
\\
& \ensuremath{\mathbf{0}}\o (f_1 \wedge g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals, "\mathsf{z}"]
\\
\ensuremath{\mathbf{0}}\wedge (g_2 \o g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[r,swap, equals, "t_{g_2 \o g_1}"]
& \ensuremath{\mathbf{0}}
\end{tikzcd}
\end{center}
\end{lem}
\begin{proof}
We start by filling the diagram on the left. First apply induction on the paths that $f_2$, $g_1$ and $g_2$
respect the basepoint. In this case $f_2\o\ensuremath{\mathbf{0}}$ is definitionally equal to $\ensuremath{\mathbf{0}}$, and the canonical
proof that $f_2\o \ensuremath{\mathbf{0}}\sim\ensuremath{\mathbf{0}}$ is (definitionally) equal to reflexivity. This means that the homotopy
$(f_2 \o \ensuremath{\mathbf{0}})\wedge (g_2 \o g_1)\sim\ensuremath{\mathbf{0}}\wedge (g_2 \o g_1)$ is also equal to reflexivity, and also the
path that $f_2 \wedge g_2$ respects the basepoint is reflexivity, hence the homotopy
$(f_2 \wedge g_2)\o \ensuremath{\mathbf{0}}\sim\ensuremath{\mathbf{0}}$ is also reflexivity. This means we need to fill the following square:
\begin{center}
\begin{tikzcd}
(f_2 \o \ensuremath{\mathbf{0}})\wedge (g_2 \o g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, equals,"i"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, equals,"1"]
& (f_2 \wedge g_2)\o (\ensuremath{\mathbf{0}} \wedge g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"(f_2\wedge g_2)\o t_{g_1}"]
\\
\ensuremath{\mathbf{0}} \wedge (g_2 \o g_1)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, equals,"t_{g_1 \o g_2}"]
& \ensuremath{\mathbf{0}}
\end{tikzcd}
\end{center}
For the underlying homotopy, take $x : A_1\wedge B_1$ and apply induction on $x$. Suppose
$x\equiv(a,b)$ for $a:A_1$ and $b:B_1$. With the notational convention for basepoints as in \autoref{lem:interchange}, we have to fill the square (we use that the paths that the maps respect the basepoints are reflexivity):
\begin{equation}\label{eq:pent-left-ab}
\begin{tikzcd}[column sep=5em]
(a_3,g_2(g_1(b)))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals,"1"]
& (a_3,g_2(g_1(b)))
\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluer}}}_{g_1(b)}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1})"]
\\
(a_3,g_2(g_1(b)))
\ensuremath{\operatorname{\mathsf{arrow}}}[r,swap,equals,"\ensuremath{\operatorname{\mathsf{gluer}}}_{g_2(g_1(b))}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_3}^{-1}"]
& (a_3,b_3)
\end{tikzcd}
\end{equation}
Now $\mapfunc{h\wedge k}(\ensuremath{\operatorname{\mathsf{gluer}}}_z)=\ensuremath{\operatorname{\mathsf{gluer}}}_{k(z)}$, so by general groupoid laws we see that the path on the bottom is equal to the path on the right, which means we can fill the square. For the other point constructors, the squares to fill are similar. If $x \judgeq \ensuremath{\operatorname{\mathsf{auxl}}}$, we have:
\begin{equation}\label{eq:pent-left-auxl}
\begin{tikzcd}[column sep=5em]
\ensuremath{\operatorname{\mathsf{auxl}}} \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals,"1"] &
\ensuremath{\operatorname{\mathsf{auxl}}} \ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluel}}}_{a_2}^{-1})"] \\
\ensuremath{\operatorname{\mathsf{auxl}}} \ensuremath{\operatorname{\mathsf{arrow}}}[r,swap, equals,"\ensuremath{\operatorname{\mathsf{gluel}}}_{a_3}^{-1}"] &
(a_3,b_3)
\end{tikzcd}
\end{equation}
We can fill this square, as the path on the bottom is definitionally equal to $\ensuremath{\operatorname{\mathsf{gluel}}}_{a_3}^{-1}$ (as we applied path induction on the path that $f_2$ respects the basepoint) and the path on the right also reduces to $\ensuremath{\operatorname{\mathsf{gluel}}}_{a_3}^{-1}$ using that $\mapfunc{h\wedge k}(\ensuremath{\operatorname{\mathsf{gluel}}}_z)=\ensuremath{\operatorname{\mathsf{gluel}}}_{h(z)}$. Similarly, we can fill the square for $x \judgeq \ensuremath{\operatorname{\mathsf{auxr}}}$, which is:
\begin{equation}\label{eq:pent-left-auxr}
\begin{tikzcd}[column sep=5em]
\ensuremath{\operatorname{\mathsf{auxr}}} \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d,swap,equals,"1"] &
\ensuremath{\operatorname{\mathsf{auxr}}} \ensuremath{\operatorname{\mathsf{arrow}}}[d,equals,"\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1})"] \\
\ensuremath{\operatorname{\mathsf{auxr}}} \ensuremath{\operatorname{\mathsf{arrow}}}[r,swap, equals,"\ensuremath{\operatorname{\mathsf{gluer}}}_{b_3}^{-1}"] &
(a_3,b_3)
\end{tikzcd}
\end{equation}
If $x$ varies over $\ensuremath{\operatorname{\mathsf{gluel}}}_a$, after some reductions, we need to fill the following cube, where the front and the back are the squares in (\ref{eq:pent-left-ab}) for $(a,b_1)$ and (\ref{eq:pent-left-auxl}) respectively; the left square is degenerate; the other three sides are the squares in the definition of $i$ and $t$ to show that they respect $\ensuremath{\operatorname{\mathsf{gluel}}}_a$ (given in (\ref{eq:i-gluel}) and (\ref{eq:t-gluel}) respectively), where we also apply $f_2 \wedge g_2$ to the square on the right. We suppress in the diagram the arguments of $\ensuremath{\operatorname{\mathsf{gluer}}}$ in $\ensuremath{\operatorname{\mathsf{gluer}}}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}^{-1}$ (which match, so the concatenation results equal to the identity path).
\begin{equation}\label{eq:pent-left-gluel}
\begin{tikzcd}[column sep=5em]
& \ensuremath{\operatorname{\mathsf{auxl}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, near end, "1"]
&& \ensuremath{\operatorname{\mathsf{auxl}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, equals, "\mapfunc{f_2\wedge g_2} (\ensuremath{\operatorname{\mathsf{gluel}}}_{a_2}^{-1})"]
\\
(a_3,b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "\ensuremath{\operatorname{\mathsf{gluel}}}_{a_3}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals, crossing over, near end, "1"]
&& (a_3,b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, near start, "\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluel}}}_{a_2})"]
\\
& \ensuremath{\operatorname{\mathsf{auxl}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals, near start, "\ensuremath{\operatorname{\mathsf{gluel}}}_{a_3}^{-1}"]
&& (a_3,b_3)
\\
(a_3,b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "\ensuremath{\operatorname{\mathsf{gluel}}}_{a_3}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals, "\ensuremath{\operatorname{\mathsf{gluer}}}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}^{-1}"]
&& (a_3,b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, swap, equals, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[from=uu, equals, crossing over, very near start, "\mapfunc{f_2 \wedge g_2}(\ensuremath{\operatorname{\mathsf{gluer}}}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}^{-1})"]
\end{tikzcd}
\end{equation}
Similarly, if $x$ varies over $\ensuremath{\operatorname{\mathsf{gluer}}}_b$, we need to fill the cube below: the front and the back are the squares in (\ref{eq:pent-left-ab}) for $(a_1,b)$ and (\ref{eq:pent-left-auxr}) respectively; the left square is again degenerate; the other three sides come from the fact that $i$ and $t$ respect $\ensuremath{\operatorname{\mathsf{gluer}}}_b$ (given in (\ref{eq:i-gluer}) and (\ref{eq:t-gluer}) respectively). Again, we omit the arguments of $\ensuremath{\operatorname{\mathsf{gluer}}}$ in $\ensuremath{\operatorname{\mathsf{gluer}}}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}^{-1}$ (in this case, not a priori judgmentally equal).
\begin{equation}\label{eq:pent-left-gluer}
\begin{tikzcd}[column sep=4em]
& \ensuremath{\operatorname{\mathsf{auxr}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, near end,"1"]
&& \ensuremath{\operatorname{\mathsf{auxr}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[dd,equals,"\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluer}}}_{b_2}^{-1})"]
\\
(a_3,g_2(g_1(b)))
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals, near end, crossing over, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "\ensuremath{\operatorname{\mathsf{gluer}}}_{g_2(g_1(b))}"]
&& (a_3,g_2(g_1(b)))
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, near start, "\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluer}}}_{g_1(b)})"]
\\
& \ensuremath{\operatorname{\mathsf{auxr}}}
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals, near start, "\ensuremath{\operatorname{\mathsf{gluer}}}_{b_3}^{-1}"]
&& (a_3,b_3)
\\
(a_3,g_2(g_1(b)))
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals,"\ensuremath{\operatorname{\mathsf{gluer}}}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, near end, "\ensuremath{\operatorname{\mathsf{gluer}}}_{g_2(g_1(b))}"]
&& (a_3,b_3)
\ensuremath{\operatorname{\mathsf{arrow}}}[from=uu, equals, crossing over, very near start, "\mapfunc{f_2\wedge g_2}(\ensuremath{\operatorname{\mathsf{gluer}}}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}^{-1})"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, swap, equals, "1"]
\end{tikzcd}
\end{equation}
In order to fill the cubes in (\ref{eq:pent-left-gluel}) and (\ref{eq:pent-left-gluer}), we generalize the paths and fill the cubes by path induction. The cube in (\ref{eq:pent-left-gluel}) can be generalized to a cube:
\begin{center}
\begin{tikzcd}[column sep=3em]
& h(y)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, near end,"1"]
&& h(y)
\ensuremath{\operatorname{\mathsf{arrow}}}[dd,equals,"\mapfunc{h}(p_l^{-1})"]
\\
h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals, near end, crossing over, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "q_l"]
&& h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, near start, "\mapfunc{h}(p_l)"]
\\
& h(y)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals, near start, "q_l^{-1}"]
&& h(x)
\\
h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals,"q_r\cdot q_r^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "q_l"]
&& h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[from=uu, equals, crossing over, near start, "\mapfunc{h}(p_r\cdot p_r^{-1})"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, swap, equals, "1"]
\end{tikzcd}
\end{center}
for $X$ and $X'$ pointed types; a map $h : X \to X'$; terms $x$, $y$ $z : X$; paths $p_l : x = y$, $p_r : x = z$, $q_l : h(x) = h(y)$, $q_r : h(x) = h(z)$; and 2-paths $s_l : \mapfunc{h}(p_l) = q_l$ (for the back and the top) and $s_r : \mapfunc{h}(p_r) = q_r$ (for the right side). This cube is filled by path induction on $s_l$, $s_r$, $p_l$ and $p_r$. The cube in (\ref{eq:pent-left-gluer}) can be generalized to a similar cube:
\begin{center}
\begin{tikzcd}[column sep=3em]
& h(y)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals,"1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, near end,"1"]
&& h(y)
\ensuremath{\operatorname{\mathsf{arrow}}}[dd,equals,"\mapfunc{h}(p_b)"]
\\
h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, equals, near end, crossing over, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, equals, "1"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "q_l"]
&& h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, near start, "\mapfunc{h}(p_l)"]
\\
& h(y)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals, near start, "q_b"]
&& h(z)
\\
h(x)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, equals,"q_l\cdot q_b"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, equals, "q_l"]
&& h(z)
\ensuremath{\operatorname{\mathsf{arrow}}}[from=uu, equals, crossing over, near start, "\mapfunc{h}(p_l\cdot p_b)"]
\ensuremath{\operatorname{\mathsf{arrow}}}[ur, swap, equals, "1"]
\end{tikzcd}
\end{center}
for paths $p_l : x = y$, $p_b : y = z$, $q_l : h(x) = h(y)$, $q_b : h(y) = h(z)$ and for 2-paths $s_l : \mapfunc{h}(p_l) = q_l$ (for the top) and $s_b : \mapfunc{h}(p_b) = q_b$ (for the back).
The diagram on the right is similar to the previous case. It is not hard to show that these homotopies are pointed.
\end{proof}
\begin{thm}\label{thm:smash-functor-right}
Given pointed types $A$, $B$ and $C$, the functorial action of the smash product induces a map
$$({-})\wedge C:(A\to B)\to(A\wedge C\to B\wedge C)$$
that is natural in $A$ and $B$ and dinatural in $C$.
\end{thm}
The naturality and dinaturality means that the following squares commute for $f : A' \to A$ $g:B\to B'$ and $h:C\to C'$.
\begin{center}
\begin{tikzcd}[column sep=5em]
(A\to B) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge C"]\ensuremath{\operatorname{\mathsf{arrow}}}[d,"f\to B"] &
(A\wedge C\to B\wedge C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,"f\wedge C\to B\wedge C"] \\
(A'\to B) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge C"] &
(A'\wedge C\to B\wedge C)
\end{tikzcd}
\qquad
\begin{tikzcd}[column sep=5em]
(A\to B) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge C"]\ensuremath{\operatorname{\mathsf{arrow}}}[d,"A\to g"] &
(A\wedge C\to B\wedge C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,"A\wedge C\to g\wedge C"] \\
(A\to B') \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge C"] &
(A\wedge C\to B'\wedge C)
\end{tikzcd}
\begin{tikzcd}[column sep=5em]
(A\to B) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge C"]\ensuremath{\operatorname{\mathsf{arrow}}}[d,"({-})\wedge C'"] &
(A\wedge C\to B\wedge C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,"A\wedge C\to B\wedge h"] \\
(A\wedge C'\to B\wedge C') \ensuremath{\operatorname{\mathsf{arrow}}}[r,"A\wedge h\to B\wedge C'"] &
(A\wedge C\to B\wedge C')
\end{tikzcd}
\end{center}
\begin{proof}
First note that $\lam{f}f\wedge C$ preserves the basepoint so that the map is indeed pointed.
Let $k:A\to B$. Then as homotopy the naturality in $A$ becomes
$(k\o f)\wedge C=k\wedge C\o f\wedge C$. To prove an equality between pointed maps, we need to give
a pointed homotopy, which is given by interchange. To show that this homotopy is pointed, we need to
fill the following square (after reducing out the applications of function extensionality), which follows from \autoref{lem:smash-coh}.
\begin{center}
\begin{tikzcd}
(\ensuremath{\mathbf{0}} \o f)\wedge C \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals]\ensuremath{\operatorname{\mathsf{arrow}}}[dd,equals] &
(\ensuremath{\mathbf{0}} \wedge C)\o (f \wedge C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] \\
& \ensuremath{\mathbf{0}} \o (f \wedge C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] \\
\ensuremath{\mathbf{0}}\wedge C \ensuremath{\operatorname{\mathsf{arrow}}}[r,equals] &
\ensuremath{\mathbf{0}}
\end{tikzcd}
\end{center}
The naturality in $B$ is almost the same: for the underlying homotopy we need to show
$i:(g \o k)\wedge C = g\wedge C \o k\wedge C$. For the pointedness we need to fill the following
square, which follows from the left pentagon in \autoref{lem:smash-coh}.
\begin{center}
\begin{tikzcd}
(g \o \ensuremath{\mathbf{0}})\wedge C \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals]\ensuremath{\operatorname{\mathsf{arrow}}}[dd,equals] &
(g \wedge C)\o (\ensuremath{\mathbf{0}} \wedge C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] \\
& (g\wedge C) \o \ensuremath{\mathbf{0}}\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] \\
\ensuremath{\mathbf{0}}\wedge C \ensuremath{\operatorname{\mathsf{arrow}}}[r,equals] &
\ensuremath{\mathbf{0}}
\end{tikzcd}
\end{center}
The dinaturality in $C$ is a bit harder. For the underlying homotopy we need to show
$B\wedge h\o k\wedge C=k\wedge C'\o A\wedge h$. This follows from applying interchange twice:
$$B\wedge h\o k\wedge C\sim(\idfunc[B]\o k)\wedge(h\o\idfunc[C])\sim(k\o\idfunc[A])\wedge(\idfunc[C']\o h)\sim k\wedge C'\o A\wedge h.$$
To show that this homotopy is pointed, we need to fill the following square:
\begin{xcenter}
\begin{tikzcd}
B\wedge h\o \ensuremath{\mathbf{0}}\wedge C \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals]\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] &
(\idfunc[B]\o \ensuremath{\mathbf{0}})\wedge(h\o\idfunc[C]) \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals]\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] &
(\ensuremath{\mathbf{0}}\o\idfunc[A])\wedge(\idfunc[C']\o h)\ensuremath{\operatorname{\mathsf{arrow}}}[r, equals]\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] &
\ensuremath{\mathbf{0}}\wedge C'\o A\wedge h\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] \\
B\wedge h\o \ensuremath{\mathbf{0}} \ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] &
\ensuremath{\mathbf{0}}\wedge(h\o\idfunc[C]) \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals]\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] &
\ensuremath{\mathbf{0}}\wedge(\idfunc[C']\o h) \ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] &
\ensuremath{\mathbf{0}}\o A\wedge h\ensuremath{\operatorname{\mathsf{arrow}}}[d,equals] \\
B\wedge h\o \ensuremath{\mathbf{0}} \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals] &
\ensuremath{\mathbf{0}} \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals] &
\ensuremath{\mathbf{0}} \ensuremath{\operatorname{\mathsf{arrow}}}[r, equals] &
\ensuremath{\mathbf{0}}
\end{tikzcd}
\end{xcenter}
The left and the right squares are filled by \autoref{lem:smash-coh}. The squares in the middle
are filled by (corollaries of) \autoref{lem:smash-general}.
\end{proof}
\subsection{Adjunction}\label{sec:smash-adjunction}
\begin{lem}\label{lem:unit-counit}
There is a unit $\eta_{A,B}\equiv\eta:A\to B\to A\wedge B$ natural in $A$ and counit
$\epsilon_{B,C}\equiv\epsilon : (B\to C)\wedge B \to C$ dinatural in $B$ and natural in $C$.
These maps satisfy the unit-counit laws:
$$(A\to\epsilon_{A,B})\o \eta_{A\to B,A}\sim \idfunc[A\to B]\qquad
\epsilon_{B,B\wedge C}\o \eta_{A,B}\wedge B\sim\idfunc[A\wedge B].$$
\end{lem}
Note: $\eta$ is also dinatural in $B$, but we do not need this.
\begin{proof}
We define $\eta ab=(a,b)$. We define the path that $\eta a$ respects the basepoint as
$$(\eta a)_0\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{gluel}}}_a\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}^{-1}:(a,b_0)=(a_0,b_0).$$ Also, $\eta$ itself respects the basepoint. To show this, we need to give $\eta_0:\eta (a_0)\sim \ensuremath{\mathbf{0}}$. The underlying maps are homotopic, by $$\eta_0b\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{gluer}}}_b\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}^{-1}:(a_0,b)=(a_0,b_0).$$ To show that
this homotopy is pointed, we need to show that the two given proofs of $(a_0,b_0)=(a_0,b_0)$ are
equal, but they are both equal to reflexivity:
$$\eta_{00}:\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}^{-1}=1=\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}^{-1}.$$
This defines the unit. To show that it is natural in $A$, we need to give the following pointed homotopy $p_\eta(f)$ for $f:A\to A'$.
\begin{center}
\begin{tikzcd}
A \ensuremath{\operatorname{\mathsf{arrow}}}[r,"\eta"]\ensuremath{\operatorname{\mathsf{arrow}}}[d,"f"] &
(B\to A \wedge B)\ensuremath{\operatorname{\mathsf{arrow}}}[d,"B\to f\wedge B"] \\
A' \ensuremath{\operatorname{\mathsf{arrow}}}[r,"\eta"] &
(B\to A'\wedge B)
\end{tikzcd}
\end{center}
We may assume that $f_0$ is reflexivity. For the underlying homotopy we need to define for $a:A$ that $p_\eta(f,a):\eta(fa)\sim f\wedge B \circ \eta a$, which is another pointed homotopy. For $b:B$ we have $\eta(fa,b)\equiv(fa,b)\equiv(f\wedge B)(\eta ab).$
The homotopy $p_\eta(f,a)$ is pointed, since $$(f\wedge B \circ \eta a)_0=\apfunc{f\wedge B}(\ensuremath{\operatorname{\mathsf{gluel}}}_a\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}^{-1})=\ensuremath{\operatorname{\mathsf{gluel}}}_{fa}\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0'}^{-1}=(\eta(fa))_0.$$
Now we need to show that $p_\eta(f)$ is pointed, for which we need to fill the following diagram.
\begin{center}
\begin{tikzcd}
\eta(fa_0) \ensuremath{\operatorname{\mathsf{arrow}}}[equals,rr,"{p_\eta(f,a_0)}"]\ensuremath{\operatorname{\mathsf{arrow}}}[dr,equals,"{\eta_0}"] & &
f\wedge B \circ \eta a_0\ensuremath{\operatorname{\mathsf{arrow}}}[dl,equals,"{f\wedge B\circ\eta_0}"] \\
& \ensuremath{\mathbf{0}}_{B,A'\wedge B} &
\end{tikzcd}
\end{center}
These pointed homotopies have equal underlying homotopies, since for $b:B$ we have
$$p_\eta(f,a_0,b)\cdot\apfunc{f\wedge B}(\eta_0 b)=1\cdot\apfunc{f\wedge B}(\ensuremath{\operatorname{\mathsf{gluer}}}_b\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}^{-1})=\ensuremath{\operatorname{\mathsf{gluer}}}_{b}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}^{-1}=\eta_0b.$$
We will skip the proof that these homotopies respect the point in the same way.
To define the counit, given $x:(B\to C)\wedge B$, we construct
$\epsilon (x):C$ by induction on $x$. If $x\equiv(f,b)$, we set $\epsilon(f,b)\ensuremath{\vcentcolon\equiv} f(b)$. If $x$
is either $\ensuremath{\operatorname{\mathsf{auxl}}}$ or $\ensuremath{\operatorname{\mathsf{auxr}}}$, then we set $\epsilon (x)\ensuremath{\vcentcolon\equiv} c_0:C$. If $x$ varies over $\ensuremath{\operatorname{\mathsf{gluel}}}_f$,
then we need to show that $f(b_0)=c_0$, which is true by $f_0$. If $x$ varies over $\ensuremath{\operatorname{\mathsf{gluer}}}_b$, we
need to show that $\ensuremath{\mathbf{0}}(b)=c_0$ which is true by reflexivity. Now $\epsilon_0\ensuremath{\vcentcolon\equiv} 1:\epsilon(\ensuremath{\mathbf{0}}_{B,C},b_0)=c_0$ shows that $\epsilon$ is pointed.
We will skip the proof that the counit is dinatural in $B$ and natural in $C$.
Finally, we need to show the unit-counit laws. For the underlying homotopy of the first one, let
$f:A\to B$. We need to show that $p_f:\epsilon\o\eta f\sim f$. We define $p_f(a)=1:\epsilon(f,a)=f(a)$. To show that $p_f$ is a pointed homotopy, we need to show that
$p_f(a_0)\cdot f_0=\mapfunc{\epsilon}(\eta f)_0\cdot \epsilon_0$, which reduces to
$f_0=\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluel}}}_f\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_0^{-1})$, but we can reduce the right hand side: (note:
$\ensuremath{\mathbf{0}}_0$ denotes the proof that $\ensuremath{\mathbf{0}}(a_0)=b_0$, which is reflexivity)
$$\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluel}}}_f\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_0^{-1})=\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluel}}}_f)\cdot(\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluel}}}_0))^{-1}=f_0\cdot \ensuremath{\mathbf{0}}_0^{-1}=f_0.$$
Now we need to show that $p$ itself respects the basepoint of $A\to B$, i.e. that the composite
$\epsilon\o\eta(\ensuremath{\mathbf{0}})\sim\epsilon\o\ensuremath{\mathbf{0}}\sim\ensuremath{\mathbf{0}}$ is equal to $p_{\ensuremath{\mathbf{0}}_{A,B}}$. The underlying
homotopies are the same for $a : A$; on the one side we have
$\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluer}}}_{a}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{a_0}^{-1})$ and on the other side we have reflexivity
(note: this type checks since $\ensuremath{\mathbf{0}}_{A,B}a\equiv\ensuremath{\mathbf{0}}_{A,B}a_0$). These paths are equal, since
$$\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluer}}}_{a}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{a_0}^{-1})=\mapfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluer}}}_{a})\cdot(\mapfunc\epsilon(\ensuremath{\operatorname{\mathsf{gluer}}}_{a_0}))^{-1}=1\cdot1^{-1}\equiv1.$$
Both pointed homotopies are pointed in the same way, which requires some path-algebra, and we skip the proof here.
For the underlying homotopy of the second unit-counit law, we need to show for $x:A\wedge B$ that
$q(x):\epsilon((\eta\wedge B)x)=x$, which we prove by induction to $x$. If $x\equiv(a,b)$, then we can define $q(a,b)\defeq1_{(a,b)}$.
If $x$ is $\ensuremath{\operatorname{\mathsf{auxl}}}$ or $\ensuremath{\operatorname{\mathsf{auxr}}}$, then the left-hand side reduces to $(a_0,b_0)$,
so we can define $q(\ensuremath{\operatorname{\mathsf{auxl}}})\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}$ and $q(\ensuremath{\operatorname{\mathsf{auxr}}})\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}$. The following computation shows that $q$ respects $\ensuremath{\operatorname{\mathsf{gluel}}}_a$:
\begin{align*}
\apfunc{\epsilon\circ\eta\wedge B}(\ensuremath{\operatorname{\mathsf{gluel}}}_a)\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}&= \apfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluel}}}_{\eta a})\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}=(\eta a)_0\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}=\ensuremath{\operatorname{\mathsf{gluel}}}_a\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}^{-1}\cdot\ensuremath{\operatorname{\mathsf{gluel}}}_{a_0}\\
&=\ensuremath{\operatorname{\mathsf{gluel}}}_a.
\end{align*}
To show that it respects $\ensuremath{\operatorname{\mathsf{gluer}}}_b$ we compute
\begin{align*}
\apfunc{\epsilon\circ\eta\wedge B}(\ensuremath{\operatorname{\mathsf{gluer}}}_b)\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}&=
\apfunc{\epsilon({-},b)}(\eta_0)\cdot\apfunc{\epsilon}(\ensuremath{\operatorname{\mathsf{gluer}}}_b)\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}=
\apfunc{\lam{f}fb}(\eta_0)\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}\\
&=\eta_0b\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}=
\ensuremath{\operatorname{\mathsf{gluer}}}_b.
\end{align*}
To show that $q$ is a pointed homotopy, we need to show that $(\epsilon\circ\eta\wedge B)_0=1$, For this we compute $$(\epsilon\circ\eta\wedge B)_0=\apfunc{\epsilon({-},b_0)}(\eta_0)=\eta_0b_0=\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}\cdot\ensuremath{\operatorname{\mathsf{gluer}}}_{b_0}^{-1}=1.$$
\end{proof}
\begin{defn}
The function $e\equiv e_{A,B,C}:(A\to B\to C)\to(A\wedge B\to C)$ is defined as the composite
$$(A\to B\to C)\xrightarrow{({-})\wedge B}(A\wedge B\to (B\to C)\wedge B)\xrightarrow{A\wedge B \to\epsilon}(A\wedge B\to C).$$
\end{defn}
\begin{lem}
The function $e$ is invertible, hence gives a pointed equivalence $$(A\to B\to C)\simeq(A\wedge B\to C).$$
\end{lem}
\begin{proof}
Define
$$e^{-1}_{A,B,C}:(A\wedge B\to C)\xrightarrow{B\to({-})}((B\to A\wedge B)\to (B\to
C))\xrightarrow{\eta\to(B\to C)}(A\to B\to C).$$ It is easy to show that $e$ and $e^{-1}$ are
inverses as unpointed maps from the unit-counit laws (\autoref{lem:unit-counit}) and naturality of $\eta$ and $\epsilon$.
\end{proof}
\begin{lem}\label{lem:e-natural}
The function $e$ is natural in $A$, $B$ and $C$.
\end{lem}
\begin{proof}
\textbf{Naturality of $e$ in $A$}. Suppose that $f:A'\to A$. Then the following diagram commutes. The left square commutes by naturality of $({-})\wedge B$ in the first argument and the right square commutes because composition on the left commutes with composition on the right.
\begin{center}
\begin{tikzcd}
(A\to B\to C) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge B"]\ensuremath{\operatorname{\mathsf{arrow}}}[d,"f\to B\to C"] &
(A\wedge B\to (B\to C)\wedge B) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"A\wedge B\to\epsilon"]\ensuremath{\operatorname{\mathsf{arrow}}}[d,"f\wedge B\to\cdots"] &
(A\wedge B\to C)\ensuremath{\operatorname{\mathsf{arrow}}}[d,"f\wedge B\to C"] \\
(A'\to B\to C) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"({-})\wedge B"] &
(A'\wedge B\to (B\to C)\wedge B) \ensuremath{\operatorname{\mathsf{arrow}}}[r,"A\wedge B\to\epsilon"] &
(A'\wedge B\to C)
\end{tikzcd}
\end{center}
\textbf{Naturality of $e$ in $C$}. Suppose that $f:C\to C'$. Then in the following diagram the left square commutes by naturality of $({-})\wedge B$ in the second argument (applied to $B\to f$) and the right square commutes by applying the functor $A\wedge B \to({-})$ to the naturality of $\epsilon$ in the second argument.
\begin{center}
\begin{tikzcd}
(A\to B\to C) \ensuremath{\operatorname{\mathsf{arrow}}}[r]\ensuremath{\operatorname{\mathsf{arrow}}}[d] &
(A\wedge B\to (B\to C)\wedge B) \ensuremath{\operatorname{\mathsf{arrow}}}[r]\ensuremath{\operatorname{\mathsf{arrow}}}[d] &
(A\wedge B\to C)\ensuremath{\operatorname{\mathsf{arrow}}}[d] \\
(A\to B\to C') \ensuremath{\operatorname{\mathsf{arrow}}}[r] &
(A\wedge B\to (B\to C')\wedge B) \ensuremath{\operatorname{\mathsf{arrow}}}[r] &
(A\wedge B\to C')
\end{tikzcd}
\end{center}
\textbf{Naturality of $e$ in $B$}. Suppose that $f:B'\to B$. Here the diagram is a bit more
complicated, since $({-})\wedge B$ is dinatural (instead of natural) in $B$. Then we get the
following diagram. The front square commutes by naturality of $({-})\wedge B$ in the second argument
(applied to $f\to C$). The top square commutes by naturality of $({-})\wedge B$ in the third
argument, the back square commutes because composition on the left commutes with composition on the
right, and finally the right square commutes by applying the functor $A\wedge B' \to({-})$ to the
naturality of $\epsilon$ in the first argument.
\begin{center}
\begin{tikzcd}[row sep=scriptsize, column sep=-4em]
& (A\wedge B\to (B\to C)\wedge B) \ensuremath{\operatorname{\mathsf{arrow}}}[rr] \ensuremath{\operatorname{\mathsf{arrow}}}[dd] & & (A\wedge B'\to (B\to C)\wedge B)\ensuremath{\operatorname{\mathsf{arrow}}}[dd] \\
(A\to B\to C) \ensuremath{\operatorname{\mathsf{arrow}}}[ur] \ensuremath{\operatorname{\mathsf{arrow}}}[rr, crossing over] \ensuremath{\operatorname{\mathsf{arrow}}}[dd] & & (A\wedge B'\to (B\to C)\wedge B') \ensuremath{\operatorname{\mathsf{arrow}}}[ur] \\
& (A\wedge B\to C)\ensuremath{\operatorname{\mathsf{arrow}}}[rr] & & (A\wedge B'\to C) \\
(A\to B'\to C) \ensuremath{\operatorname{\mathsf{arrow}}}[rr] & & (A\wedge B'\to (B'\to C)\wedge B') \ensuremath{\operatorname{\mathsf{arrow}}}[ur] \ensuremath{\operatorname{\mathsf{arrow}}}[from=uu, crossing over]
\end{tikzcd}
\end{center}
\end{proof}
\begin{rmk}
Instead of showing that $e$ is natural, we could show that $e^{-1}$ is natural. In
that case we need to show that the map $A\to({-}):(B\to C)\to(A\to B)\to(A\to C)$ is natural in
$A$, $B$ and $C$. This might actually be easier, since we do not need to work with any higher
inductive type to prove that.
\end{rmk}
We have now obtained the following theorem
\begin{thm}\label{thm:smash-adjoint}
There is an equivalence
$$(A\to B\to C)\simeq(A\wedge B\to C)$$
natural in $A$, $B$ and $C$.
\end{thm}
\begin{rmk}
We can state \autoref{thm:smash-adjoint} as an adjunction $({-})\wedge B\dashv B\to({-})$ or by saying that $A\wedge B$
represents the functor $A\to B\to ({-})$.
In \autoref{sec:smash-monoidal} we show that the smash product forms a
1-coherent symmetric monoidal product from the assumption that this adjunction
is pointed in $C$. Explicitly, this means that the naturality of $e$ in $C$
applied to the map $\ensuremath{\mathbf{0}}_{C,C'}:C\to C'$ is equal to the composite
$$(A\wedge B \to \ensuremath{\mathbf{0}}_{C,C'})\o e_{A,B,C}\sim \ensuremath{\mathbf{0}}\o e_{A,B,C}\sim \ensuremath{\mathbf{0}}
\sim e_{A,B,C'}\o \ensuremath{\mathbf{0}} \sim e_{A,B,C'}\o (A\to B\to \ensuremath{\mathbf{0}}_{C,C'}).$$
To prove this, we need that the counit $\ensuremath{\varepsilon}$ is pointed natural in $C$. To
prove that, we need to show that the map $({-})\wedge C$, defined in
\autoref{thm:smash-functor-right}, is pointed natural in $B$. In order to
prove that, we need to show that in the situation of \autoref{lem:smash-coh},
if both $f_1$ and $f_2$ are (judgmentally) the constant map, then the two
pentagons stated in that lemma are equal (transported appropriately in order
to make this equality type check). This can be formulated as a 3-path in a
type of pointed maps, which is hard to fill.
\end{rmk}
\subsection{Symmetric monoidal product}\label{sec:smash-monoidal}
In this section we will prove that the smash product is a 1-coherent symmetric
monoidal product~\autoref{def:symmonprod}, from the assumption that the
adjunction from \autoref{sec:smash-adjunction} is pointed natural in $C$. We
will need to following pointed equivalences. Without the proof that $e$ is
pointed natural, parts of this section are still true. In particular, the
natural equivalences defined in \autoref{def:smash-alrg} do not require pointed
naturality of $e$.
\begin{defn}\label{def:b-and-tw}
We define the pointed equivalences:
\[\mathsf{b} : (\ensuremath{\S^0} \to X) \simeq X\] where $\ensuremath{\S^0}$ is the type of booleans (pointed in ${0_{\bool}}$) with underlying map defined with $\mathsf{b}(f) \ensuremath{\vcentcolon\equiv} f({1_{\bool}})$, and
\[\mathsf{tw} : (A \to B \to X) \simeq (B \to A \to X)\]
with underlying map defined with $\mathsf{tw}(f) \ensuremath{\vcentcolon\equiv} \lam{b}\lam{a}f(a)(b)$.
\end{defn}
Using \autoref{lem:yoneda} (Yoneda) we can prove associativity, left and right
unitality and braiding equivalences for the smash product, in the following way.
\begin{defn}\label{def:equiv-precursors}
The following pointed equivalences are defined for $A$, $B$, $C$ and $X$ pointed types:
\begin{itemize}
\item $\overline{\alpha}_X : (A \wedge (B \wedge C) \to X) \simeq ((A \wedge B) \wedge C \to X)$ as the composition of the equivalences:
\begin{align*}
A \wedge (B \wedge C)\to X&\simeq A \to B\wedge C\to X && (e^{-1})\\
&\simeq A \to B\to C\to X && (A \to e^{-1})\\
&\simeq A \wedge B\to C\to X && (e)\\
&\simeq (A \wedge B)\wedge C\to X. && (e)
\end{align*}
\item $\overline{\lambda}_X : (B \to X) \simeq (\ensuremath{\S^0} \wedge B \to X)$ as the composition of the equivalences:
\begin{align*}
B \to X &\simeq \ensuremath{\S^0} \to B \to X && (\mathsf{b}^{-1})\\
&\simeq \ensuremath{\S^0} \wedge B \to X && (e)
\end{align*}
\item $\overline{\rho}_X : (A \to X) \simeq (A \wedge \ensuremath{\S^0} \to X)$ as the composition of the equivalences:
\begin{align*}
A \to X &\simeq A \to \ensuremath{\S^0} \to X && (A \to \mathsf{b}^{-1})\\
&\simeq A \wedge \ensuremath{\S^0} \to X && (e)
\end{align*}
\item $\overline{\gamma}_X : (B \wedge A \to X) \simeq (A \wedge B \to X)$ as the composition of the equivalences:
\begin{align*}
B \wedge A \to X &\simeq B \to A \to X && (e^{-1})\\
&\simeq A \to B \to X && (\mathsf{tw})\\
&\simeq A \wedge B \to X && (e)
\end{align*}
\end{itemize}
\end{defn}
\begin{rmk}\label{rmk:alrg-pointed-natural}
The equivalences in \autoref{def:equiv-precursors} are natural in all their
arguments and from the assumption that $e$ is pointed natural in $C$ we can
show that these maps are all pointed natural in $X$.
\end{rmk}
\begin{defn}\label{def:smash-alrg}
We define the following equivalences, natural in all their arguments, with inverses provided as in \autoref{lem:yoneda}:
\begin{itemize}
\item $\alpha\ensuremath{\vcentcolon\equiv}\overline{\alpha}_{A \wedge (B \wedge C)}(\idfunc) : (A \wedge B) \wedge C \simeq A \wedge (B \wedge C)$ (associativity of the smash product), with inverse $\alpha^{-1}\ensuremath{\vcentcolon\equiv}\overline{\alpha}^{-1}_{(A \wedge B) \wedge C}(\idfunc)$;
\item $\lambda \ensuremath{\vcentcolon\equiv} \overline{\lambda}_B(\idfunc) : \ensuremath{\S^0} \wedge B \simeq B$ and $\rho \ensuremath{\vcentcolon\equiv} \overline{\rho}_A(\idfunc) : A \wedge \ensuremath{\S^0} \simeq A$ (left- and right unitors for the smash product), with inverses $\lambda^{-1}\ensuremath{\vcentcolon\equiv} \overline{\lambda}_{\ensuremath{\S^0}\wedge B}^{-1}(\idfunc)$ and $\rho^{-1}\ensuremath{\vcentcolon\equiv} \overline{\rho}_{A\wedge \ensuremath{\S^0}}^{-1}(\idfunc)$, respectively;
\item $\gamma \ensuremath{\vcentcolon\equiv} \overline{\gamma}_{B\wedge A} (\idfunc) : A \wedge B \simeq B \wedge A$ (braiding for the smash product), with inverse $\gamma^{-1} \ensuremath{\vcentcolon\equiv} \overline{\gamma}_{A \wedge B}^{-1} (\idfunc)$.
\end{itemize}
$\alpha$, $\lambda$, $\rho$ and $\gamma$ are natural in all their arguments,
as $\overline{\alpha}$, $\overline{\lambda}$, $\overline{\rho}$ and $\overline{\gamma}$ are. Note that these
definitions do \emph{not} require pointed naturality of $e$.
\end{defn}
\begin{lem}\label{lem:bar-homotopy}
There are pointed homotopies
\begin{align*}
\overline{\alpha}_X &\sim \alpha \to X
& \overline{\lambda}_X &\sim \lambda \to X
\\
\overline{\rho}_X &\sim \rho \to X
& \overline{\gamma}_X &\sim \gamma \to X
\end{align*}
\end{lem}
\begin{proof}
This follows directly from \autoref{lem:yoneda-pointed} and \autoref{rmk:alrg-pointed-natural} (this does require pointed naturality of $e$).
\end{proof}
\begin{thm}[Associativity pentagon]\label{thm:smash-associativity-pentagon}
For $A$, $B$, $C$ and $D$ pointed types, there is a homotopy
\[\alpha \o \alpha \sim (A \wedge \alpha) \o \alpha \o (\alpha \wedge D)\]
corresponding to the commutativity of the following diagram:
\begin{center}
\begin{tikzcd}
&((A \wedge B) \wedge (C \wedge D))
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, "\alpha"]
\\
(((A \wedge B) \wedge C) \wedge D)
\ensuremath{\operatorname{\mathsf{arrow}}}[ru, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\alpha \wedge D"]
&& (A \wedge (B \wedge (C \wedge D)))
\\
((A \wedge (B \wedge C)) \wedge D)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, "\alpha"]
&& (A \wedge ((B \wedge C) \wedge D))
\ensuremath{\operatorname{\mathsf{arrow}}}[u, swap, "A \wedge \alpha"]
\end{tikzcd}
\end{center}
\end{thm}
\begin{proof}
We articulate the proof in several steps. A map homotopic to both sides of the sought homotopy will be constructed via the equivalence
\begin{align*}
\overline{\alpha}^4 : (A \wedge (B \wedge (C \wedge D)) \to X) &\simeq (((A \wedge B) \wedge C) \wedge D \to X)
\intertext{(natural in all its arguments), defined as the composite:}
A \wedge (B \wedge (C \wedge D)) \to X
&\simeq A \to B \wedge (C \wedge D) \to X && \text{($e^{-1}$)}\\
&\simeq A \to B \to C \wedge D \to X &&\text{($A \to e^{-1}$)}\\
&\simeq A \to B \to C \to D \to X &&\text{($A \to B \to e^{-1}$)}\\
&\simeq A \wedge B \to C \to D \to X &&\text{($e$)}\\
&\simeq (A \wedge B) \wedge C \to D \to X &&\text{($e$)}\\
&\simeq ((A \wedge B) \wedge C) \wedge D \to X && \text{($e$)}
\intertext{giving $\overline{\alpha}^4(\idfunc) : ((A \wedge B) \wedge C) \wedge D) \simeq A \wedge (B \wedge (C \wedge D))$. Moreover, in order to simplify the expressions of $\alpha \wedge D$ and $A \wedge \alpha$, we also define:}
\overline{\alpha}^R : ((A \wedge (B \wedge C)) \wedge D \to X) &\simeq (((A \wedge B) \wedge C) \wedge D \to X)
\intertext{as the composite:}
(A \wedge (B \wedge C)) \wedge D \to X
&\simeq A \wedge (B \wedge C) \to D \to X &&\text{($e^{-1}$)}\\
&\simeq (A \wedge B) \wedge C \to D \to X &&\text{($\overline{\alpha}$)}\\
&\simeq ((A \wedge B) \wedge C) \wedge D \to X &&\text{($e$)}
\intertext{and}
\overline{\alpha}^L : (A \wedge (B \wedge (C \wedge D)) \to X) &\simeq (A \wedge ((B \wedge C) \wedge D) \to X)
\intertext{as the composite:}
A \wedge (B \wedge (C \wedge D)) \to X
&\simeq A \to B \wedge (C \wedge D) \to X &&\text{($e^{-1}$)}\\
&\simeq A \to (B \wedge C) \wedge D \to X &&\text{($A \to \overline{\alpha}$)}\\
&\simeq A \wedge ((B \wedge C) \wedge D) \to X &&\text{($e$)}
\end{align*}
also natural in their arguments. Evaluating these equivalences to the identity function, we get new arrows that fit in the original diagram:
\begin{center}
\begin{tikzcd}
&((A \wedge B) \wedge (C \wedge D))
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, "\alpha"]
\\
(((A \wedge B) \wedge C) \wedge D)
\ensuremath{\operatorname{\mathsf{arrow}}}[ru, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\alpha \wedge D"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, bend left=40, "\overline{\alpha}^R(\idfunc)"]
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, "\overline{\alpha}^4(\idfunc)"]
&& (A \wedge (B \wedge (C \wedge D)))
\\
((A \wedge (B \wedge C)) \wedge D)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, "\alpha"]
&& (A \wedge ((B \wedge C) \wedge D))
\ensuremath{\operatorname{\mathsf{arrow}}}[u, swap, "A \wedge \alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[u, bend left=40, "\overline{\alpha}^L(\idfunc)"]
\end{tikzcd}
\end{center}
The theorem is then proved once we show the chain of homotopies:
\begin{equation}\label{eq:alphafour}
\alpha \o \alpha
\sim \overline{\alpha}^4(\idfunc)
\sim \overline{\alpha}^L(\idfunc) \o \alpha \o \overline{\alpha}^R(\idfunc)
\sim (A \wedge \alpha) \o \alpha \o (\alpha \wedge D)
\end{equation}
To verify the first homotopy in (\ref{eq:alphafour}), we see that:
\begin{align*}
\alpha \o \alpha
&\judgeq \overline{\alpha}(\idfunc) \o \overline{\alpha}(\idfunc)\\
&\sim (\overline{\alpha} \o \overline{\alpha}) (\idfunc) &&\text{(naturality of $\overline{\alpha}$)}\\
&\judgeq (e \o e \o (A \to e^{-1}) \o e^{-1} \o e \o e \o (A \to e^{-1}) \o e^{-1})(\idfunc)\\
&\sim (e \o e \o (A \to e^{-1}) \o e \o (A \to e^{-1}) \o e^{-1})(\idfunc) &&\text{(cancelling)}\\
&\sim (e \o e \o e \o (B \to A \to e^{-1}) \o (A \to e^{-1}) \o e^{-1})(\idfunc) &&\text{(naturality of $e$)}\\
&\judgeq \overline{\alpha}^4(\idfunc)
\end{align*}
The second homotopy in (\ref{eq:alphafour}) is verified by (right-to-left):
\begin{align*}
\overline{\alpha}^L(\idfunc) \o \alpha \o \overline{\alpha}^R(\idfunc)
&\judgeq \overline{\alpha}^L(\idfunc) \o \overline{\alpha}(\idfunc) \o \overline{\alpha}^R(\idfunc)\\
&\sim (\overline{\alpha}^R \o \overline{\alpha} \o \overline{\alpha}^L)(\idfunc) \\ &\mbox{}\qquad\text{(naturality of $\overline{\alpha}$ and $\overline{\alpha}^R$)}\\
&\judgeq (e \o \overline{\alpha} \o e^{-1} \o e \o e \o (A \to e^{-1}) \o e^{-1} \o e \o (A \to \overline{\alpha}) \o e^{-1})(\idfunc)\\
&\sim (e \o \overline{\alpha} \o e \o (A \to e^{-1}) \o (A \to \overline{\alpha}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(cancelling)}\\
&\sim (e \o \overline{\alpha} \o e \o (A \to (e^{-1} \o \overline{\alpha})) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(functoriality of $A\to -$)}\\
&\judgeq (e \o e \o e \o (A \to e^{-1}) \o e^{-1} \o e \\
&\hspace{3em}\o (A \to (e^{-1} \o e \o e \o (B \to e^{-1}) \o e^{-1})) \o e^{-1})(\idfunc)\\
&\sim (e \o e \o e \o (A \to ((B \to e^{-1}) \o e^{-1})) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(cancelling)}\\
&\sim (e \o e \o e \o (B \to A \to e^{-1}) \o (A \to e^{-1}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(funct. of $A \to -$)}\\
&\judgeq \overline{\alpha}^4(\idfunc)
\end{align*}
In order to prove the last homotopy in (\ref{eq:alphafour}), it is sufficient to show that $\overline{\alpha}^R(\idfunc) \sim \alpha \wedge D$ and that $\overline{\alpha}^L(\idfunc) \sim A \wedge \alpha$. We have:
\begin{align*}
\overline{\alpha}^R(\idfunc)
&\judgeq e(\overline{\alpha} (e^{-1}(\idfunc)))\\
&\sim e (\overline{\alpha} (\eta))\\
&\sim e (\eta \o \overline{\alpha}(\idfunc)) &&\text{(naturality of $\overline{\alpha}$)}\\
&\judgeq \epsilon \o (\eta \o \alpha) \wedge D\\
&\sim \epsilon \o (\eta \wedge D) \o (\alpha \wedge D) &&\text{(distrib. of $\wedge$)}\\
&\sim \alpha \wedge D &&\text{(\autoref{lem:unit-counit})}
\end{align*}
and, lastly,
\begin{align*}
\overline{\alpha}^L(\idfunc)
&\judgeq e(\overline{\alpha} \o e^{-1}(\idfunc))\\
&\sim e(\overline{\alpha} \o \eta)\\
&\sim e((\alpha \to A \wedge (B \wedge (C \wedge D))) \o \eta) &&\text{(\autoref{lem:bar-homotopy})}\\
&\sim e((B \wedge (C \wedge D) \to A \wedge \alpha) \o \eta) &&\text{(dinaturality of $\eta$)}\\
&\sim (A \wedge \alpha) \o e(\eta) &&\text{(naturality of $e$)}\\
&\sim A \wedge \alpha &&\text{(\autoref{lem:unit-counit})}
\end{align*}
thus proving the desired homotopy.
\end{proof}
\begin{thm}[Unitors triangle]\label{thm:smash-unitors-triangle}
For $A$ and $B$ pointed types, there is a homotopy
\[(A \wedge \lambda) \o \alpha \sim (\rho \wedge B)\]
corresponding to the commutativity of the following diagram:
\begin{center}
\begin{tikzcd}
((A \wedge \ensuremath{\S^0}) \wedge B)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, swap, "\rho \wedge B"]
&& (A \wedge (\ensuremath{\S^0} \wedge B))
\ensuremath{\operatorname{\mathsf{arrow}}}[dl, "A \wedge \lambda"]
\\
& (A \wedge B)
\end{tikzcd}
\end{center}
\end{thm}
\begin{proof}
By an argument similar to the one for $\overline{\alpha}^L$ and $\overline{\alpha}^R$ in \autoref{thm:smash-associativity-pentagon}, one can verify the homotopies $A \wedge \lambda \sim (e \o (A \to \overline{\lambda}) \o e^{-1})(\idfunc)$ and $\rho \wedge B \sim (e \o \overline{\rho} \o e)(\idfunc)$, simplifying the expressions in the sought homotopy. Then:
\begin{align*}
(A \wedge \lambda) \o \alpha
&\sim e(\overline{\lambda} \o e^{-1}(\idfunc)) \o \overline{\alpha}(\idfunc) &&\text{(simplification)}\\
&\sim \overline{\alpha}(e(\overline{\lambda} \o e^{-1}(\idfunc)) &&\text{(naturality of $\overline{\alpha}$)}\\
&\judgeq e(e(e^{-1} \o e^{-1} (e(\overline{\lambda} \o e^{-1}(\idfunc)))))\\
&\sim e(e(e^{-1} \o \overline{\lambda} \o e^{-1}(\idfunc))) &&\text{(cancelling)}\\
&\judgeq e(e(e^{-1} \o e \o \mathsf{b}^{-1} \o e^{-1}(\idfunc)))\\
&\sim e(e(\mathsf{b}^{-1} \o e^{-1}(\idfunc))) &&\text{(cancelling)}\\
&\judgeq (e \o \overline{\rho} \o e^{-1})(\idfunc)\\
&\sim \rho \wedge B &&\text{(simplification)}
\end{align*}
gives the desired homotopy.
\end{proof}
\begin{thm}[Braiding-unitors triangle]\label{thm:smash-braiding-unitors}
For a pointed type $A$, there is a homotopy
\[\lambda \o \gamma \sim \rho\]
corresponding to the commutativity of the following diagram:
\begin{center}
\begin{tikzcd}
(A \wedge \ensuremath{\S^0})
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, "\gamma"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, swap, "\rho"]
&& (\ensuremath{\S^0} \wedge A)
\ensuremath{\operatorname{\mathsf{arrow}}}[dl, "\lambda"]
\\
& A
\end{tikzcd}
\end{center}
\end{thm}
\begin{proof}
We have:
\begin{align*}
\lambda \o \gamma
&\judgeq \overline{\lambda}(\idfunc) \o \overline{\gamma}(\idfunc)\\
&\sim (\overline{\gamma} \o \overline{\lambda})(\idfunc) &&\text{(naturality of $\overline{\gamma}$)}\\
&\judgeq (e \o \mathsf{tw} \o e^{-1} \o e \o \mathsf{b}^{-1})(\idfunc)\\
&\sim (e \o \mathsf{tw} \o \mathsf{b}^{-1})(\idfunc) &&\text{(cancelling)}\\
&\sim (e \o (A \to \mathsf{b}^{-1}))(\idfunc)\\
&\judgeq \overline{\rho}(\idfunc) \judgeq \rho
\end{align*}
where the last homotopy is given by $(A \to c) \o \mathsf{b} \sim \mathsf{tw} : (\ensuremath{\S^0} \to A \to X) \to (A \to X)$.
\end{proof}
\begin{lem}\label{lem:pentagon-c}
The following diagram commutes, for $A$, $B$, $C$ and $X$ pointed types:
\begin{center}
\begin{tikzcd}[column sep=7em]
(B \wedge C \to A \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "e^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, "\mathsf{tw}"]
& (B \to C \to A \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "B \to \mathsf{tw}"]
\\
& (B \to A \to C \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "\mathsf{tw}"]
\\
(A \to B \wedge C \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "A \to e^{-1}"]
& (A \to B \to C \to X)
\end{tikzcd}
\end{center}
\end{lem}
\begin{proof}
Unfolding the definition of $e^{-1}$, we get the diagram:
\begin{xcenter}
\begin{tikzcd}[column sep=6em,every node/.style={font=\sffamily\small}]
(B \wedge C \to A \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, bend left=10, "e^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "C\to -"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dd, swap, "\mathsf{tw}"]
& ((C \to B \wedge C) \to C \to A \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "\eta \to C \to A \to X"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "(C \to B \wedge C) \to \mathsf{tw}"]
& (B \to C \to A \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "B \to \mathsf{tw}"]
\\
& ((C \to B \wedge C) \to A \to C \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "\eta \to A \to C \to X"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\mathsf{tw}"]
& (B \to A \to C \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "\mathsf{tw}"]
\\
(A \to B \wedge C \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[rr, swap, bend right=10, "A \to e^{-1}"]
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "A \to (C \to -)"]
& (A \to (C \to B \wedge C) \to C \to X)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "A \to (\eta \to C \to X)"]
& (A \to B \to C \to X)
\end{tikzcd}
\end{xcenter}
where the squares on the right are instances of naturality of $\mathsf{tw}$, while the commutativity of the pentagon on the left follows easily from the definition of $\mathsf{tw}$.
\end{proof}
\begin{thm}[Associativity-braiding hexagon]\label{thm:smash-associativity-braiding}
For pointed types $A$, $B$ and $C$, there is a homotopy
\[\alpha \o \gamma \o \alpha \sim (B \wedge \gamma) \o \alpha \o (\gamma \wedge C)\]
corresponding to the commutativity of the following diagram:
\begin{center}
\begin{tikzcd}
((A \wedge B) \wedge C)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "\alpha"]
\ensuremath{\operatorname{\mathsf{arrow}}}[d, swap, "\gamma \wedge C"]
&(A \wedge (B \wedge C))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "\gamma"]
& ((B \wedge C) \wedge A)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "\alpha"]
\\
((B \wedge A) \wedge C))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "\alpha"]
& (B \wedge (A \wedge C))
\ensuremath{\operatorname{\mathsf{arrow}}}[r, swap, "B \wedge \gamma"]
& (B \wedge (C \wedge A))
\end{tikzcd}
\end{center}
\end{thm}
\begin{proof}
The proof is structured similarly to the one for \autoref{thm:smash-associativity-pentagon}: the homotopies
\begin{align*}
B \wedge \gamma &\sim \overline{\gamma}^L(\idfunc) &\text{with\ \ } \overline{\gamma}^L &\ensuremath{\vcentcolon\equiv} e \o (B \to \overline{\gamma}) \o e^{-1}\\
\gamma \wedge C &\sim \overline{\gamma}^R(\idfunc) &\text{with\ \ } \overline{\gamma}^R &\ensuremath{\vcentcolon\equiv} e \o \overline{\gamma} \o e^{-1}
\end{align*}
can be proven in exactly the same way and, using these simplifications, we will show that both sides of the sought homotopy are homotopic to the same equivalence. Indeed we have:
\begin{align*}
\alpha \o \gamma \o \alpha
&\judgeq \overline{\alpha}(\idfunc) \o \overline{\gamma}(\idfunc) \o \overline{\alpha}(\idfunc)\\
&\sim (\overline{\alpha} \o \overline{\gamma} \o \overline{\alpha})(\idfunc) \\ &\mbox{}\qquad\text{(naturality of $\overline{\gamma}$ and $\overline{\alpha}$)}\\
&\judgeq (e \o e \o (A \to e^{-1}) \o e^{-1} \o e \o \mathsf{tw} \o e^{-1} \o e \o e \o (B \to e^{-1}) \o e^{-1})(\idfunc)\\
&\sim (e \o e \o (A \to e^{-1}) \o \mathsf{tw} \o e \o (B \to e^{-1}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(cancelling)}\\
&\sim (e \o e \o \mathsf{tw} \o (B \to \mathsf{tw}) \o e^{-1} \o e \o (B \to e^{-1}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(\autoref{lem:pentagon-c})}\\
&\sim (e \o e \o \mathsf{tw} \o (B \to \mathsf{tw}) \o (B \to e^{-1}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(cancelling)}
\end{align*}
and
\begin{align*}
(B \wedge \gamma) \o \alpha \o (\gamma \wedge C)
&\sim \overline{\gamma}^L(\idfunc) \o \overline{\alpha} \o \overline{\gamma}^R(\idfunc) \\ &\mbox{}\qquad\text{(simplification)}\\
&\sim (\overline{\gamma}^R \o \overline{\alpha} \o \overline{\gamma}^L)(\idfunc) \\ &\mbox{}\qquad\text{(naturality of $\overline{\alpha}$ and $\overline{\gamma}^R$)}\\
&\judgeq (e \o \overline{\gamma} \o e^{-1} \o e \o e \o (B \to e^{-1}) \o e^{-1} \o e \o (B \to \overline{\gamma}) \o e^{-1})(\idfunc)\\
&\sim (e \o \overline{\gamma} \o e \o (B \to e^{-1}) \o (B \to \overline{\gamma}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(cancelling)}\\
&\sim (e \o \overline{\gamma} \o e \o (B \to (e^{-1} \o \overline{\gamma})) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(functoriality of $B \to -$)}\\
&\judgeq (e \o e \o \mathsf{tw} \o e^{-1} \o e \o (B \to (e^{-1} \o e \o \mathsf{tw} \o e^{-1})) \o e^{-1})(\idfunc)\\
&\sim (e \o e \o \mathsf{tw} \o (B \to \mathsf{tw}) \o (B \to e^{-1}) \o e^{-1})(\idfunc) \\ &\mbox{}\qquad\text{(cancelling)}
\end{align*}
proving the commutativity of the diagram.
\end{proof}
\begin{thm}[Double braiding]\label{thm:smash-double-braiding}
For $A$ and $B$ pointed types, there is a homotopy
\[\gamma \o \gamma \sim \idfunc\]
corresponding to the commutativity of the following diagram:
\begin{center}
\begin{tikzcd}
(A \wedge B)
\ensuremath{\operatorname{\mathsf{arrow}}}[r, "\gamma"]
\ensuremath{\operatorname{\mathsf{arrow}}}[dr, equals]
& (B \wedge A)
\ensuremath{\operatorname{\mathsf{arrow}}}[d, "\gamma"]
\\
& (A \wedge B)
\end{tikzcd}
\end{center}
\end{thm}
\begin{proof}
Using that $\mathsf{tw} \o \mathsf{tw} \sim \idfunc$, we get:
\begin{align*}
\gamma \o \gamma
&\judgeq \overline{\gamma}(\idfunc) \o \overline{\gamma}(\idfunc)\\
&\sim (\overline{\gamma} \o \overline{\gamma})(\idfunc) &&\text{(naturality of $\overline{\gamma}$)}\\
&\judgeq (e \o \mathsf{tw} \o e^{-1} \o e \o \mathsf{tw} \o e^{-1})(\idfunc)\\
&\sim \idfunc &&\text{(cancelling)}
\end{align*}
as desired.
\end{proof}
Finally we get the result of this section.
\begin{thm}
$\wedge$ is a 1-coherent symmetric monoidal product, assuming that $e$ is pointed natural in $C$.
\end{thm}
\begin{proof}
This follows immediately from the theorems in this section.
\end{proof}
\chapter{The Serre Spectral Sequence}\label{cha:serre-spectr-sequ}
Spectral sequences are important tools in algebraic topology.\footnote{The work in this chapter is joint work with Jeremy Avigad, Steve Awodey, Ulrik Buchholtz, Egbert Rijke and Mike Shulman.} They give a
relationship between certain homotopy, homology and cohomology groups, in a way
that generalizes long exact sequences. This generalization comes at a cost of
being a lot more complicated than a long exact sequence.
In this chapter we will start the study of spectral sequences in homotopy type
theory. We will introduce the notion of spectral sequences, and then construct
the Atiyah-Hirzebruch and Serre spectral sequences for cohomology. We follow the
construction due to Michael Shulman given in~\cite{shulman2013spectral}. We will also give a sketch on
how to construct the analogues for homology, and look at some of the
applications of these spectral sequences.
There are a couple of notable differences between spectral sequences in homotopy
type theory compared to classical homotopy theory.
\begin{itemize}
\item As always, in HoTT all constructions have to be homotopy invariant, so
we cannot use classical constructions that are not homotopy invariant. For
example, the construction of the Serre spectral sequence for homology in~\cite{hatcher2004spectral}
uses CW-approximation of a space and the skeleton
of the obtained CW-complex to construct the spectral sequence. These
operations are not homotopy invariant, and therefore cannot be performed in
HoTT.
\item Another difference is that homology and cohomology are defined
differently in HoTT than in classical homotopy theory. In classical homotopy
theory (co)homology is defined using singular (co)homology. Since the intermediate
steps in the construction of singular (co)homology is not homotopy invariant, we use a different definition of
(co)homology (see \autoref{def:cohomology}), which impacts the definition of spectral sequences involving (co)homology.
\item
The first page of a spectral sequence is often not homotopy invariant, and
therefore cannot be constructed in HoTT. For this reason, we start counting
the pages of spectral sequences at 2.
\item HoTT offers a convenient language for formalizing proofs. Therefore, we
have formalized all constructed spectral sequences in this chapter.
\end{itemize}
The spectral sequences we construct are not the most general version of these
spectral sequences. The spectral sequences we construct are still more general
than the formulation of the Serre spectral sequence in many textbooks (we give a
version of generalized and parametrized cohomology), but there exist more
general versions. There are two places where we compromised on generality for
the sake of making the formalization easier. The first compromise is that we
only formalized exact couples for graded $R$-modules for a ring $R$ (which is
not graded). More generally we could do this for any abelian category, which
would require building up the theory of abelian categories (this is done in
UniMath~\cite{unimath}). Furthermore, we did not look at convergence of spectral
sequences in the most general sense, since that can get quite complicated and
subtle. Instead, we only look at spectral sequences that are eventually constant
pointwise, so the $\infty$-page is just the eventual value. This restriction
adds the condition to the spectral sequences we construct that the coefficients
are only in truncated spectra.
\section{Spectral Sequences}\label{sec:spectral-sequences}
A spectral sequence consists of a sequence of pages, each of them containing a two-dimensional grid of abelian groups. There are maps between these groups, called \emph{differentials}. These differentials form (co)chain complexes, and the (co)homology of these complexes determine the groups on the next page. In \ref{fig:spectral-sequence-pages} we show an example of two pages of a spectral sequence, where each dot represents an abelian group. In this figure only the two first quadrants are shown, because in simple applications all other groups are trivial, though that need not be the case in general.
\begin{figure}[ht]
\centering
\begin{subfigure}{.5\textwidth}
\centering
\begin{tikzpicture}
\draw [thick] (0.5,0.5) -- (0.5,5.5);
\draw [thick] (0.5,0.5) -- (5.5,0.5);
\ensuremath{\operatorname{\mathsf{node}}} at (5.3,0.3) {$p$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.3,5.3) {$q$};
\ensuremath{\operatorname{\mathsf{node}}} (x11) at (1,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x12) at (1,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x13) at (1,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x14) at (1,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x15) at (1,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x21) at (2,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x22) at (2,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x23) at (2,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x24) at (2,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x25) at (2,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x31) at (3,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x32) at (3,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x33) at (3,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x34) at (3,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x35) at (3,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x41) at (4,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x42) at (4,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x43) at (4,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x44) at (4,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x45) at (4,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x51) at (5,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x52) at (5,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x53) at (5,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x54) at (5,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x55) at (5,5) {$\bullet$};
\path
(x12) edge[->] (x31)
(x13) edge[->] (x32)
(x14) edge[->] (x33)
(x15) edge[->] (x34)
(x22) edge[->] (x41)
(x23) edge[->] (x42)
(x24) edge[->] (x43)
(x25) edge[->] (x44)
(x32) edge[->] (x51)
(x33) edge[->] (x52)
(x34) edge[->] (x53)
(x35) edge[->] (x54)
(x42) edge (5.5,1.25)
(x43) edge (5.5,2.25)
(x44) edge (5.5,3.25)
(x45) edge (5.5,4.25)
(x52) edge (5.5,1.75)
(x53) edge (5.5,2.75)
(x54) edge (5.5,3.75)
(x55) edge (5.5,4.75)
(2,5.5) edge[->] (x35)
(3,5.5) edge[->] (x45)
(4,5.5) edge[->] (x55)
(5,5.5) edge (5.5,5.25)
;
\end{tikzpicture}
\caption{The page $E_2^{p,q}$}
\label{fig:E2-page}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\begin{tikzpicture}
\draw [thick] (0.5,0.5) -- (0.5,5.5);
\draw [thick] (0.5,0.5) -- (5.5,0.5);
\ensuremath{\operatorname{\mathsf{node}}} at (5.3,0.3) {$p$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.3,5.3) {$q$};
\ensuremath{\operatorname{\mathsf{node}}} (x11) at (1,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x12) at (1,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x13) at (1,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x14) at (1,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x15) at (1,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x21) at (2,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x22) at (2,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x23) at (2,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x24) at (2,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x25) at (2,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x31) at (3,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x32) at (3,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x33) at (3,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x34) at (3,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x35) at (3,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x41) at (4,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x42) at (4,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x43) at (4,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x44) at (4,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x45) at (4,5) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x51) at (5,1) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x52) at (5,2) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x53) at (5,3) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x54) at (5,4) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} (x55) at (5,5) {$\bullet$};
\path
(x13) edge[->] (x41)
(x14) edge[->] (x42)
(x15) edge[->] (x43)
(x23) edge[->] (x51)
(x24) edge[->] (x52)
(x25) edge[->] (x53)
(x33) edge (5.5,1.3333)
(x34) edge (5.5,2.3333)
(x35) edge (5.5,3.3333)
(x43) edge (5.5,2)
(x44) edge (5.5,3)
(x45) edge (5.5,4)
(x53) edge (5.5,2.6667)
(x54) edge (5.5,3.6667)
(x55) edge (5.5,4.6667)
(1.75,5.5) edge[->] (x44)
(2.75,5.5) edge[->] (x54)
(3.25,5.5) edge[->] (x45)
(4.25,5.5) edge[->] (x55)
(3.75,5.5) edge (5.5,4.3333)
(4.75,5.5) edge (5.5,5)
(5.25,5.5) edge (5.5,5.3333)
;
\end{tikzpicture}
\caption{$E_3^{p,q}$}
\label{fig:E3-page}
\end{subfigure}
\caption{Two pages of a spectral sequence.}
\label{fig:spectral-sequence-pages}
\end{figure}
Before we start, we define the notion of a graded abelian group. We will give a
nonstandard definition that is equivalent to the standard one.
\begin{defn}\label{def:graded}
For an abelian group $G$, an \emph{$G$-graded abelian group} is
a family of abelian groups indexed over $G$. If $M$ and $M'$ are $G$-graded
abelian groups, the type of \emph{graded abelian group homomorphism from $M$ to
$M'$} is a triple consisting of a \emph{degree} $e : G \simeq G$ (this is an equivalence of types, not a group isomorphism), a proof of
$(g : G) \to e(g)=g+e(0)$ and a term of type
$$\{x\ y : I\} \to (p : e(x)=y) \to M_x \to M'_y.$$
We will denote the type of homomorphisms as $M\to M'$.
For $\phi:M\to M'$ we write $\deg_\phi$ for the first projection. We will
often call $\deg_\phi(0)$ the degree of $\phi$. For $x : I$ we will write
$$\phi_x\ensuremath{\vcentcolon\equiv}\phi_{\ensuremath{\operatorname{\mathsf{refl}}}_x}:M_x\to M'_{\deg_\phi(x)}$$
and
$$\phi_{[x]}\ensuremath{\vcentcolon\equiv}\phi_{p_x}:M_{\deg_\phi^{-1}(x)}\to M'_x$$
where $p_x:\deg_\phi(\deg_\phi^{-1}(x))=x$ is the proof obtained from the equivalence $\deg_\phi$.
\end{defn}
\begin{rmk}
This definition looks a bit cumbersome, since the condition on $e$ forces
$e$ to be homotopic to the function $\lam{g}g+h$ for some group element $h$.
Furthermore, the type of $\phi$ is equivalently $(x : I) \to M_x \to
M'_{x+h}$. We will now discuss why we made these choices.
To see why this is more convenient, we consider the composition of two graded homomorphisms. Suppose we
have two graded homomorphisms $\phi:M\to M'$ and $\psi:M'\to M''$ of degrees $h:G$ and $k:G$,
respectively. Then the pointwise composition $\lam{g:G}{m:M_g}\psi_{g+h}(\phi_g(m))$ has type
$(g : G) \to M_g \to M''_{(g+h)+k}$. So to get a graded homomorphism of degree $h+k$, with the more straightforward representation, we would need to
transport along the equality $(g+h)+k=g+(h+k)$. Since compositions are ubiquitous, this would happen all over the place.
However, in our setting, the composite of two graded homomorphisms of degree $e$ and
$e'$ will have degree $e' \o e$, without using any transports.
We eliminated a transport to define composition, but there are other places where we cannot get rid
of them so easily. For example, given morphisms $\phi:M\to M'$ and $\psi:M'\to M''$ with
$\psi\o\phi=0$ (the graded map that is constantly 0), we are interested in the
homology of $\phi$ and $\psi$. This is the kernel of $\psi$ quotiented by the image of $\phi$ in
$M'_x$. However, if $\phi_x$ has type $M_x \to M'_{\deg_\phi(x)}$, there is no map that (without
transports) lands in $M'_x$. We would need to transport along the path
$p_x:\deg_\phi^{-1}(\deg_\phi(x))=x$ and take the image of this composite:
$$M_{\deg_\phi^{-1}(x)} \xrightarrow{\phi_{\deg_\phi^{-1}(x)}} M'_{\deg_\phi(\deg_\phi^{-1}(x))} \xrightarrow{\sim} M'_x.$$
For this reason, we allow graded homomorphisms to be applied to paths, so that we have a ``built-in''
transport. Then we can define the homology as $H_x\ensuremath{\vcentcolon\equiv} \ker(\psi_x)/\im(\phi_{[x]})$, or diagramatically
$$M_{\deg_\phi^{-1}(x)}\xrightarrow{\phi_{[x]}}M'_x\xrightarrow{\psi_x}M''_{\deg_\psi(x)}.$$
For the construction of spectral sequences, we do not actually need the
second component of a graded homomorphism: all constructions also work if
the degrees are arbitrary equivalences of type $I\simeq I$, where $I$ is an
arbitrary set. This is the definition used in the formalization. In this
document we add this condition, so that our definition is equivalent to the
usual definition of graded morphism.
\end{rmk}
\begin{defn}
A \emph{spectral sequence} consists of the following data.
\begin{itemize}
\item A sequence $E_r$ of abelian groups graded over $\ensuremath{\mathbb{Z}}\xspace\times\ensuremath{\mathbb{Z}}\xspace$ for $r\geq2$. $E_r$ is called the \emph{$r$-page} of the spectral sequence;
\item \emph{differentials}, which are graded morphisms $d_r:E_r\to E_r$ such that $d_r\circ d_r=0$;
\item isomorphisms $\alpha_r^{p,q}:H^{p,q}(E_r)\simeq E_{r+1}^{p,q}$ where $H^{p,q}(E_r)=\ker(d_r^{p,q})/\im(d_r^{[p,q]})$ is the cohomology of the cochain complex determined by $d_r$.
\end{itemize}
\end{defn}
We use the notation for cohomologically indexed spectral sequences, since we
will construct spectral sequences in cohomology in this chapter. For the
spectral sequences in this chapter, the degree of $d_r$ will be $(r,1-r)$, which
signifies a cohomologically indexed spectral sequence.
As mentioned before, we start counting the pages at 2, since the first page of
the spectral sequences we construct will not be homotopy invariant. In the
formalization we start counting at 0 for convenience. Also, in the
formalization, we assume that the grading of $E_r$ is over some set $I$
instead of fixing it to $\ensuremath{\mathbb{Z}}\xspace\times \ensuremath{\mathbb{Z}}\xspace$. It is not clear whether this extra
generality is useful. Instead of abelian groups, we could take objects of an
arbitrary abelian category, but for concreteness and to simplify things, we
choose to develop the theory only for abelian groups. In the formalization we
developed the theory for graded $R$-module for a (non-graded) ring $R$, but we
have only applied it to abelian groups so far.
Note that $(E_r,d_r)$ determines $E_{r+1}$ but \emph{not} $d_{r+1}$.
Furthermore, $E_r$ is a subquotient (subgroup of a quotient) of $E_2$, so if
$E_2^{p,q}$ is trivial, then $E_r^{p,q}$ is trivial for all $r$.
In many cases, the spectral sequence will \emph{converge}. That means that for
a fixed $(p,q):\ensuremath{\mathbb{Z}}\xspace\times \ensuremath{\mathbb{Z}}\xspace$ the sequence $E_r^{p,q}$ will be constant for $r$
large enough. For example, suppose that the degree of $d_r$ is $(r,r-1)$, and
$E_2$ is limited to the first quadrant. Now for any $(p,q)$ all differentials
in or out of $E_r^{p,q}$ will go out the first quadrant for sufficiently large
$r$. This means that the image of $d_r^{[p,q]}$ is trivial, and the kernel of
$d_r^{p,q}$ is the full group. This implies that $E_{r+1}^{p,q}\simeq
E_r^{p,q}$, so the spectral sequence converges.
Whenever a spectral sequence converges, we write $E_\infty^{p,q}$ for the
eventual value of $E_r^{p,q}$ for $r$ large enough. Now the power of spectral
sequences is that there is often a relation between $E_2^{p,q}$ and
$E_\infty^{p,q}$. This relation does not specify $E_\infty^{p,q}$ exactly, but
specifies that $E_\infty^{p,q}$ build up some group $D^n$ for the diagonals
where $p+q=n$.
\begin{defn}
Suppose given an abelian group $D$ and a finite sequence of abelian groups
$(E^n)_n$. We say that \emph{$D$ is built from $(E^n)_n$} if there is a
sequence of abelian groups $(D^n)_n$ and short exact sequences
\begin{align*}
E^0\to &D \to D^1\\
&\vdots\\
E^k\to &D^k \to D^{k+1}\\
E^{k+1}\to &D^{k+1} \to D^{k+2}\\
&\vdots\\
E^m\to &D^m \to 0\\
\end{align*}
The sequence $(D^n)_n$ is called a \emph{cofiltration} of $D$, they are successive quotients of $D$.
\end{defn}
\begin{defn}
Given a graded abelian group $D^n$ and a bigraded abelian group $C^{p,q}$, we write
$$E_2^{p,q}=C^{p,q}\Rightarrow D^{p+q}$$
if there is a spectral sequence $E$ such that
\begin{itemize}
\item $E_2^{p,q}=C^{p,q}$;
\item $E$ converges to $E_\infty$;
\item $D^n$ is built from $E_\infty^{p,q}$ where $p+q=n$.
\end{itemize}
\end{defn}
\begin{rmk}
This definition implicitly requires that for $p+q=n$ only finitely many $E_\infty^{p,q}$ are nontrivial.
This is sufficient for the spectral sequences we consider in this chapter, but this condition can be relaxed in more general constructions of spectral sequences.
\end{rmk}
\section{Exact Couples}\label{sec:exact-couples}
As we said before, the pair $(E_r,d_r)$ in a spectral sequence specifies
$E_{r+1}$, but not $d_{r+1}$. If we have some more information about page $r$,
then we can construct page $r+1$ and the extra information for page $r+1$. Now
we can iterate this construction and obtain a spectral sequence by forgetting
about the extra information.
An \emph{exact couple} exactly gives this extra information~\cite{massey1952exactcouple}. From it, we can compute the \emph{derived exact
couple}, which gives us the information next page of the spectral sequence.
\begin{defn}
An \emph{exact couple} is a pair $(D,E)$ of $\ensuremath{\mathbb{Z}}\xspace\times\ensuremath{\mathbb{Z}}\xspace$-graded abelian groups with graded homomorphisms
\begin{center}
\begin{tikzpicture}[thick,node distance=1cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$D$};
\ensuremath{\operatorname{\mathsf{node}}}[below right = of tl] (b) {$E$};
\ensuremath{\operatorname{\mathsf{node}}}[above right = of b] (tr) {$D$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$i$} (tr)
(tr) edge[->] node [below right] {$j$} (b)
(b) edge[->] node [below left] {$k$} (tl);
\end{tikzpicture}
\end{center}
that is exact in all three vertices. This means that for all $p:\deg_j(x)=_Iy$ and
$q:\deg_k(y)=z$ that $\ker(k_q)=\im(j_p)$, and similarly for the other two pairs of maps.
For an exact couple we will write $\iota\ensuremath{\vcentcolon\equiv}\deg_i$ and $\eta\ensuremath{\vcentcolon\equiv}\deg_j$ and $\kappa\ensuremath{\vcentcolon\equiv}\deg_k$ for the degrees.
\end{defn}
\begin{lem}
Given an exact couple $(D,E,i,j,k)$, we can define a \emph{derived exact couple} $(D',E',i',j',k')$
where $E'$ is the homology of $d\ensuremath{\vcentcolon\equiv} j\o k:E\to E$. The degrees of the derived maps are
$\deg_{i'}\equiv\iota$, $\deg_{k'}\equiv\kappa$ and $\deg_{j'}\equiv\eta\o\iota^{-1}$.
\begin{center}
\begin{tikzpicture}[thick,node distance=1cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$D'$};
\ensuremath{\operatorname{\mathsf{node}}}[below right = of tl] (b) {$E'$};
\ensuremath{\operatorname{\mathsf{node}}}[above right = of b] (tr) {$D'$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$i'$} (tr)
(tr) edge[->] node [below right] {$j'$} (b)
(b) edge[->] node [below left] {$k'$} (tl);
\end{tikzpicture}
\end{center}
\end{lem}
\begin{proof}
In this proof we will be explicit about the grading of $D$ and $E$, which is a lot trickier (at least in intensional type theory) than a proof without the grading. For a proof that does not take the grading into account, see for example~\cite[Lemma 1.1]{hatcher2004spectral}.
We define for $x:\ensuremath{\mathbb{Z}}\xspace\times \ensuremath{\mathbb{Z}}\xspace$ the graded abelian groups $D'$ and $E'$ by $D_x'=\im i_{[x]}$ and $E_x'=\ker d_x/\im d_{[x]}$. Now $i_x':D_x'\to D_{\iota x}'$ is defined as the composite
$$D_x' \hookrightarrow D_x \xrightarrow{i} D_{\iota x}'.$$
This is sufficient to define $i'$ on all paths $\iota x=y$ as a function $D_x \to D_y$.\\
We first define $j_{px}':D_{\iota x}'\to E_{\eta x}'$ for the canonical path $px:\eta(\iota^{-1}(\iota x))=\eta x$, which is sufficient to define $j'$ in general.
Note that $D_{\iota x}'\equiv \im i_{[\iota x]}\simeq\im i_x$, so to define $j_{px}'$ it is sufficient to define $\tilde\jmath : D_x \to E_{\eta x}'$ such that $(a : D_x) \to i_x(a)=0 \to \tilde\jmath(a)=0$. We define $\tilde\jmath(a)\ensuremath{\vcentcolon\equiv} [j_xa]$. This is well-defined, since $j_xa\in\ker d_{\eta x}$,\footnote{We use the set-theoretical notation $g\in H$ to say that a group element $g:G$ is in subgroup $H$. Formally, a subgroup $H$ is an element of $G\to\ensuremath{\mathsf{Prop}}\xspace$ (containing 0, and closed under addition and negation) and $g\in H$ is defined as $H(g)$. Note that $(g:G)\times H(g)$ can be endowed with a group structure, which is $H$ viewed as a group.} because
$$d_{\eta x}(j_xa)=j_{\kappa(\eta x)}(k_{\eta x}(j_xa))=j_{\kappa(\eta x)}\ensuremath{\mathbf{0}}=\ensuremath{\mathbf{0}}.$$
Now suppose that $i_x(a)=0$. Without loss of generality we may assume that $x\equiv \kappa y$. By exactness, this means that $a\in\im k_y$, so there is $b:E_y$ such that $k_y(b)=a$.
Now $j_xa=j_x(k_yb)\equiv d_yb$, so
$$j_xa\in\im d_y\simeq \im d_{[\deg_d y]}\equiv \im d_{[\eta x]}.$$
This shows that $\tilde\jmath(a)=0$, completing the definition of $j'$. Note that
$j_{px}'(i_xa)=[j_xa]$.
To define $k_x':E_x'\to D_{\kappa x}'$, first note that if $a\in\ker d_x$, then $k_xa\in\ker j_{\kappa x}=\im i_{[\kappa x]}$ by exactness. Now we need to show that if $a\in\im d_{[x]}$, then $k_xa=0$. By assumption, we have $b:E_{\eta^{-1}\kappa^{-1} x}$ such that $d_{[x]}(b)=a$. Now we compute (using $k_xj_{[x]}=\ensuremath{\mathbf{0}}$)
$$k_xa=k_x(d_{[x]}b)=k_x(j_{[x]}(k_{[\eta^{-1} x]}b))=0.$$
This defines $k$.
Showing exactness of the derived couple involves some diagram chasing.
To show that $j'i'=\ensuremath{\mathbf{0}}$ it is sufficient to show that for all $a:D_{\iota x}'$ we have
$j'_{p(\iota x)}(i'_{\iota x}a)=0$. Since $a\in\im i_{[\iota x]}\simeq i_x$ we know that $a=i_x b$ for some $b: D_x$. We compute
$$j'_{p(\iota x)}(i'_{\iota x}a)=j'_{p(\iota x)}(i_{\iota x}a)=[j_{\iota x}a]=[j_{\iota x}(i_x b)]=[0]=0.$$
To show that $\ker j'\subseteq \im i'$, it is sufficient to show that
$\ker j'_{p(\kappa x)}\subseteq \im i'_{\kappa x}$. Suppose $a : D'_{\iota(\kappa x)}$ such that $j'_{p(\kappa x)}(a)=0$, we know that $a=i_{\kappa x}(b)$ for some $b$. Now
$$0=j'_{p(\kappa x)}(a)=j'_{p(\kappa x)}(i_{\kappa x}(b))=[j_{\kappa x}(b)],$$
which means that $j_{\kappa x}(b)\in\im d_{[{\kappa (\eta x)}]}\simeq \im d_x$. This means that for some $c: E_x$ we have $j_{\kappa x}(b)=d_x(c)=j_{\kappa x}(k_x c)$. This means that
$j_{\kappa x}(b-k_x c)=0$, hence $b-k_xc\in \ker j_{\kappa x}=\im i_{[\kappa x]}$. This means that we can define $b-k_xc: D'_{\kappa x}$. Now we compute
$$i'_{\kappa x}(b-k_xc)=i_{\kappa x}b-i_{\kappa x}(k_xc)=a-0=a,$$
which means $a\in\im i'_{\kappa x}$, as desired.
We will omit the other cases, which are similar but easier.
\end{proof}
Repeating the process of deriving exact couples, we get a sequence of exact couples $(D_r,E_r,i_r,j_r,k_r)$.\footnote{We will now put the grading of $D$, $E$ and the maps as superscript, so that we can put the page as subscript.} We get a spectral sequence $(E_r,d_r)$ where $d_r\ensuremath{\vcentcolon\equiv} j_r\o k_r$. Note that
$$\deg_{d_r}=\deg_{j_r}\o\deg_{k_r}=\eta\o\iota^r\o\kapp
$$
Given some extra conditions on the exact couple, we can show that this spectral sequence converges.
\begin{defn}\label{def:bounded}
We call an exact couple \emph{bounded} if for every $x:\ensuremath{\mathbb{Z}}\xspace\times\ensuremath{\mathbb{Z}}\xspace$ there is are bounds $B_x : \textbf{N}$ such that for all $s \geq B_x$ we have
$$E^{\iota^{-s}(x)}=0 \qquad\text{and}\qquad D^{\iota^s(x)}=0$$
\end{defn}
\begin{rmk}\label{rmk:bounded}
The condition on $D$ also shows that if you go sufficiently far in the $\iota$-direction, then $E$ is trivial, since $D \xrightarrow{j} E \xrightarrow{k} D$ is exact and the occurrences of $D$ will be trivial. Converse, the condition on $E$ shows that if you go sufficiently far in the
$\iota^{-1}$ direction, $i$ will be an equivalence, by the following exact sequence.
$$E \xrightarrow{k} D \xrightarrow{i} D \xrightarrow{j} E$$
We call $x:\ensuremath{\mathbb{Z}}\xspace\times\ensuremath{\mathbb{Z}}\xspace$ a \emph{stable index} whenever $i^{[\iota^{-s}x]}$ is surjective for \emph{all} $s\geq0$.
\end{rmk}
Given a bounded exact couple, the pages stabilize pointwise, which is the content of the next lemma.
\begin{lem}\label{lem:exact-couple-stabilize}
For a bounded exact couple $(D,E,i,j,k)$ we have for all sufficiently large $r$ that $D_{r+1}^x=D_r^x$ and $E_{r+1}^x=E_r^x$.
\end{lem}
\begin{proof}
Note that $E_{r+1}^x=\ker d_r^{x}/\im d_r^{[x]}$. Since $d_r$ has degree
$\eta\o\iota^r\o\kappa$, and because $\ensuremath{\mathbb{Z}}\xspace\times \ensuremath{\mathbb{Z}}\xspace$ is an abelian group, the
degrees commute.\footnote{In the formalization, we do not assume that the
degrees are shifts by a group element, and we explicitly assume that
$\kappa\iota=\iota\kappa$ and $\iota\eta=\eta\iota$.} The codomain of $d_r^x$
is $E_r^{\eta(\iota^r(\kappa x))}=E_r^{\iota^r(\eta(\kappa x))}$, which is
trivial for sufficiently large $r$ by \autoref{rmk:bounded}. Also, the domain
of $d_r^{[x]}$ is $E_r^{\kappa^{-1}(\iota^{-r}(\eta^{-1}
x))}=E_r^{\iota^{-r}(\kappa^{-1}(\eta^{-1} x))}$, which is trivial for sufficiently large $r$ by the definition of boundedness.
To show that $D$ stabilizes, first note that if $i_r^{[\iota^{-1} x]}$ is surjective, then $i_{r+1}^{[x]}$ is surjective. The reason is that
$$i_{r+1}: D_{r+1}^{\iota^{-1} x}\xrightarrow{\sim} D_r^{\iota^{-1} x} \to D_{r+1}^{x}$$
is now a composite of two surjective maps. This means that if the maps $i_{r_0}^{[\iota^{-s}x]}$ are surjections for all $s\geq B+1$, then the maps $i_{r_0+1}^{[\iota^{-s}x]}$ will be surjections for all $s\geq B$. In this case, for $r\geq r_0+B$ we have that $i_r^{[x]}$ is a surjection, hence that $D_{r+1}^x=D_r^x$. Since $i_0^{[\iota^{-s}x]}$ are surjections for sufficiently large $s$ by \autoref{rmk:bounded}, we finish the proof.
\end{proof}
By the proof of \autoref{lem:exact-couple-stabilize} we get explicit bounds $B_x^D$ and $B_x^E$ such that $D_r^x=D_{B_x^D}^x$ and $E_{r'}^x=E_{B_x^E}^x$ for all $r\geq B_x^D$ and $r'\geq B_x^E$. We define $D_\infty^x\ensuremath{\vcentcolon\equiv} D_{B_x^D}^x$ and $E_\infty^x\ensuremath{\vcentcolon\equiv} E_{B_x^E}^x$. Both $B_x^D$ and $B_x^E$ will be the maximum of $B_y$ for some sequence of indices $y$.
\begin{thm}[Convergence Theorem]\label{thm:exact-couple-convergence}
Let $(D,E,i,j,k)$ be a bounded exact couple and let $x$ be a stable index.
Then $D^{\kappa x}$ is built from $(E_\infty^{\iota^n(x)})_{0\le n < B_{\kappa x}}$.
\end{thm}
\begin{proof}
Define
$C^n = D_\infty^{\kappa(\iota^nx)}$. Let $n:\textbf{N}$ be arbitrary, then for sufficiently large $r$ the following is a short exact sequence
$$0 \xrightarrow{j_r} E_r^{\iota^nx} \xrightarrow{k_r} D_r^{\kappa(\iota^nx)} \xrightarrow{i_r}
D_r^{\iota(\kappa(\iota^nx))} \xrightarrow{j_r} 0.$$
This is the case, because for sufficiently large $r$ the domain of $j_r^{[\iota^nx]}$ and the codomain of $j_r{\iota(\kappa(\iota^nx))}$ are contractible. Now (possibly by increasing $r$) these groups are in the stable range, so we get a short exact sequence
$$0 \to E_\infty^{\iota^nx} \to C^n \to C^{n+1} \to 0.$$ Moreover, we have
$C^0\equiv D_\infty^{\kappa x}=D^{\kappa x}$ because $x$ is a stable index.
Lastly, for $s\geq B_{\kappa x}$ we know that $C^s$ is trivial, because
$D^{\kappa(\iota^n)}$ is trivial by the condition of being bounded. This shows
that $D^{\kappa x}$ is built from $(E_\infty^{\iota^n(x)})_{0\le n < B_{\kappa
x}}$.
\end{proof}
\section{Spectra}\label{sec:spectra}
We have not yet discussed how to get an exact couple in the first place. Recall that from a pointed
map we get a long exact sequence of homotopy groups. For a \emph{sequence} of pointed maps we get a sequence of long exact sequences. However, we do not want to do this for pointed maps, but for maps between \emph{spectra}.
You can think of a spectrum as a generalized space with negative dimensions. Suppose we are given a pointed type $X$ and a chosen delooping $Y$ of $X$. That is, $Y$ is a pointed type such that $\Omega Y\simeq^* X$. Now the $(n+1)$-th homotopy group of $Y$ is equal to the $n$-th homotopy group of $X$. The $0$-th homotopy group of $Y$ is new information, and we can think of it as the $(-1)$-th homotopy group of $X$. Spectra go further on this idea: it is a pointed type with infinitely many deloopings.
\begin{defn}
A \emph{prespectrum} is a pair consisting of a sequence of pointed types $Y:\ensuremath{\mathbb{Z}}\xspace\to\ensuremath{\mathsf{Type}}\xspace^*$ and a sequence of pointed maps $e:(n:\ensuremath{\mathbb{Z}}\xspace) \to Y_n \to^* \Omega Y_{n+1}$. An \emph{$\Omega$-spectrum} or \emph{spectrum} is a prespectrum $(Y,e)$ where $e_n$ is a pointed equivalence for all $n$. We will often just write $Y$ for the pair $(Y,e)$, and we denote the type of (pre)spectra by $\ensuremath{\mathsf{Prespectrum}}$ and $\ensuremath{\mathsf{Spectrum}}$.\\
A map between (pre)spectra $(Y,e)\to (Y',e')$ is a pair consisting of $f:(n : \ensuremath{\mathbb{Z}}\xspace) \to Y_n \to Y_n'$ and $p : (n : \textbf{N}) \to e_n' \o f_n \sim^* \Omega f_{n+1} \o e_n$.
\end{defn}
\begin{rmk}
Usually a (pre)spectrum is indexed over $\textbf{N}$ and not over $\ensuremath{\mathbb{Z}}\xspace$. We index it over $\ensuremath{\mathbb{Z}}\xspace$ so that we do not have to do a case split in --- for example --- the definition of homotopy group of a spectrum, see \autoref{def:spectrum-homotopy-group}.
\end{rmk}
\begin{ex}\mbox{}
\begin{itemize}
\item If $A$ is an abelian group, we have $HA:\ensuremath{\mathsf{Spectrum}}$ where $(HA)_n=K(A,n)$ for $n\geq0$ and $(HA)_n=\ensuremath{\mathbf{1}}\xspace$ for $n<0$.
\item Given $Y:\ensuremath{\mathsf{Spectrum}}$ and $k:\ensuremath{\mathbb{Z}}\xspace$, we can define two new spectra $\Omega^kY$ and $\Sigma^k Y:\ensuremath{\mathsf{Spectrum}}$ with
\begin{align*}
(\Omega^kY)_n&\ensuremath{\vcentcolon\equiv} Y_{n-k}&(\Sigma^kY)_n&\ensuremath{\vcentcolon\equiv} Y_{n+k}
\end{align*}
\item Given a spectrum map $f: X \to Y$, we have a spectrum $\ensuremath{\operatorname{\mathsf{fib}}}_f:\ensuremath{\mathsf{Spectrum}}$ with $(\ensuremath{\operatorname{\mathsf{fib}}}_f)_n\ensuremath{\vcentcolon\equiv} \ensuremath{\operatorname{\mathsf{fib}}}_{f_n}$. Furthermore we have a spectrum map $p_1:\ensuremath{\operatorname{\mathsf{fib}}}_f\to X$. This follows from the following two facts about fibers (which we will not prove here).
\begin{enumerate}
\item Given a pointed map $g : A \to B$, there is a pointed equivalence $e_1:\Omega\ensuremath{\operatorname{\mathsf{fib}}}_g\simeq^*$ with a pointed homotopy
\begin{center}
\begin{tikzpicture}[thick,node distance=1cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\Omega\ensuremath{\operatorname{\mathsf{fib}}}_g$};
\ensuremath{\operatorname{\mathsf{node}}}[below right = of tl] (r) {$A$};
\ensuremath{\operatorname{\mathsf{node}}}[below left = of r] (bl) {$\ensuremath{\operatorname{\mathsf{fib}}}_{\Omega g}$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above right] {$\Omega p_1$} (r)
(bl) edge[->] node [below right] {$p_1$} (r)
(tl) edge[->] node [left] {$e_1$} (bl);
\end{tikzpicture}
\end{center}
\item $\ensuremath{\operatorname{\mathsf{fib}}}$ is a functor from pointed maps to pointed types and $p_1$ is a natural transformation. This means the following. Suppose we are given a square of pointed maps and a homotopy filling the following square.
\begin{center}
\begin{tikzpicture}[thick,node distance=1cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$A$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$A'$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$B$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$B'$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$f$} (tr)
edge[->] node [right] {$h$} (bl)
(bl) edge[->] node [above] {$g$} (br)
(tr) edge[->] node [right] {$h'$} (br);
\end{tikzpicture}
\end{center}
Then there is a pointed map $e_2:\ensuremath{\operatorname{\mathsf{fib}}}_f\to\ensuremath{\operatorname{\mathsf{fib}}}_g$, functorial in $(h,h')$. In particular this means that if $h$ and $h'$ are equivalences, then $e_2$ is. The naturality of $p_1$ means that we have the following pointed homotopy.
\begin{center}
\begin{tikzpicture}[thick,node distance=1cm]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$\ensuremath{\operatorname{\mathsf{fib}}}_f$};
\ensuremath{\operatorname{\mathsf{node}}}[right = of tl] (tr) {$A$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$\ensuremath{\operatorname{\mathsf{fib}}}_g$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$B$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node [above] {$p_1$} (tr)
edge[->] node [right] {$e_2$} (bl)
(bl) edge[->] node [above] {$p_1$} (br)
(tr) edge[->] node [right] {$h$} (br);
\end{tikzpicture}
\end{center}
\end{enumerate}
\end{itemize}
\end{ex}
Given an $\Omega$-spectrum $Y$ and $n : \ensuremath{\mathbb{Z}}\xspace$, we define can the $n$-th homotopy group of $Y$ to be
$$\pi_n(Y)\ensuremath{\vcentcolon=}\pi_{n+k}(Y_k):\ensuremath{\mathsf{AbGroup}}$$
for any $k$ such that $n+k\geq 0$. This is independent of $k$, because
$$\pi_{n+(k+1)}(Y_{k+1})\simeq \pi_{n+k}(\Omega Y_{k+1})\simeq \pi_{n+k}(Y_k).$$
For concreteness, in the following definition we pick $k=2-n$. We make this choice so that $\pi_n(Y)$ directly carries the structure of an abelian group.
The homotopy group of a prespectrum $Y$ is a bit different, since $\pi_{n+k}(Y_k)$ is not independent of $k$. In this case, it is the colimit as $k\to\infty$. We make the substitution
$\ell=n+k-2$ to make the index of the homotopy group always positive.
\begin{defn}\label{def:spectrum-homotopy-group}
Given an $\Omega$-spectrum $Y$ and $n : \ensuremath{\mathbb{Z}}\xspace$, we define the \emph{$n$-th homotopy group of $Y$} as
$$\pi_n(Y)\ensuremath{\vcentcolon\equiv} \pi_2(Y_{2-n}).$$
For a prespectrum $Y$ we define
$$\pi_n(Y)\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{colim}}}_{\ell\to\infty}(\pi_{\ell+2}(Y_{\ell+2-n})).$$
\end{defn}
Note that the homotopy group of a prespectrum is a set by \autoref{cor:trunc_colim}\ref{part:colim_is_trunc},
and the colimit can be equipped with a group structure, making $\pi_n(Y)$ an abelian group for a prespectrum $Y$.
The long exact sequence of homotopy groups for pointed types, constructed in \autoref{sec:les-homotopy}, induces one on spectra.
\begin{thm}\label{thm:spectrum-LES}
Given a spectrum map $f:X\to Y$ with fiber $F$, we get the following long exact sequence of homotopy groups indexed over $\ensuremath{\mathbb{Z}}\xspace\times \ensuremath{\operatorname{\mathsf{fin}}}_3$.
\begin{center}\begin{tikzpicture}[thick, node distance=18mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$\pi_k(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$\pi_k(X)$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$\pi_k(F)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of Y] (OY) {$\pi_{k+1}(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of X] (OX) {$\pi_{k+1}(X)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of F] (OF) {$\pi_{k+1}(F)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OY] (O2Y) {$\pi_{k+2}(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OX] (O2X) {$\pi_{k+2}(X)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OF] (O2F) {$\pi_{k+2}(F)$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 5mm of O2X] {$\vdots$};
\ensuremath{\operatorname{\mathsf{node}}}[below = 5mm of X] {$\vdots$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] node [above] (f){$\pi_k(f)$} (Y)
(F) edge[->] node [below] (f){$\pi_k(p_1)$} (X)
(OY) edge[->] (F)
(OX) edge[->] node [above] (f){$\pi_{k+1}(f)$} (OY)
(OF) edge[->] node [below] (f){$\pi_{k+1}(p_1)$} (OX)
(O2Y) edge[->] (OF)
(O2X) edge[->] node [above] (f){$\pi_{k+2}(f)$} (O2Y)
(O2F) edge[->] node [below] (f){$\pi_{k+2}(p_1)$} (O2X);
\end{tikzpicture}
\end{center}
\end{thm}
We will use the following lemma. Recall the definition of successor structure from \autoref{def:chain-complex}.
\begin{lem}\label{lem:splice}
Suppose given two successor structures $N$ and $M$, and for each $n:N$ let $G^n$ be a long exact sequence index by $M$. Let $m:M$ and $k\geq 2$. Suppose that
\begin{itemize}
\item for all $n:N$, $G_m^{n+1} \simeq G_{m+k}^n$ and $G_{m+1}^{n+1} \simeq G_{m+k+1}^n$
\item for all $n:N$ the following diagram commutes.
\begin{center}\begin{tikzpicture}[thick, node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$G_{m+k}^n$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$G_{m+k+1}^n$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of Y] (OY) {$G_m^{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of X] (OX) {$G_{m+1}^{n+1}$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] (Y)
(OY) edge[->] node [sloped, above] {$\sim$} (Y)
(OX) edge[->] node [sloped, above] {$\sim$} (X)
(OX) edge[->] (OY);
\end{tikzpicture}\end{center}
\end{itemize}
Then there is a long exact sequence $H:N\times\ensuremath{\operatorname{\mathsf{fin}}}_{k-1}\to\ensuremath{\mathsf{Set}}\xspace^*$ with $H_{(n,\ell)}\ensuremath{\vcentcolon\equiv} G_{m+\ell}^n$
\end{lem}
For $k=3$ the hypotheses can be represented in the diagram below.
\begin{center}\begin{tikzpicture}[thick, node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$G_m^n$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$G_{m+1}^n$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$G_{m+2}^n$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of F] (G) {$G_{m+3}^n$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of G] (H) {$G_{m+4}^n$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of H] (I) {$\cdots$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of Y] (OY) {$G_m^{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of X] (OX) {$G_{m+1}^{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of F] (OF) {$G_{m+2}^{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of G] (OG) {$G_{m+3}^{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of H] (OH) {$G_{m+4}^{n+1}$};
\ensuremath{\operatorname{\mathsf{node}}} (OI) at (I |- OH) {$\cdots$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of OY] (O2Y) {$G_{m}^{n+2}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of OX] (O2X) {$G_{m+1}^{n+2}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of OF] (O2F) {$G_{m+2}^{n+2}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of OG] (O2G) {$G_{m+3}^{n+2}$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 12mm of OH] (O2H) {$G_{m+4}^{n+2}$};
\ensuremath{\operatorname{\mathsf{node}}} (O2I) at (OI |- O2H) {$\cdots$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 5mm of O2F] {$\vdots$};
\ensuremath{\operatorname{\mathsf{node}}}[below = 5mm of F] {$\vdots$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] (Y)
(F) edge[->] (X)
(G) edge[->] (F)
(H) edge[->] (G)
(I) edge[->] (H)
(OY) edge[->] node [sloped, below, pos = 0.4] {$\sim$} (G)
(OX) edge[->] node [sloped, above, pos = 0.6] {$\sim$} (H)
(OX) edge[->] (OY)
(OF) edge[->] (OX)
(OG) edge[->] (OF)
(OH) edge[->] (OG)
(OI) edge[->] (OH)
(O2Y) edge[->] node [sloped, below, pos = 0.4] {$\sim$} (OG)
(O2X) edge[->] node [sloped, above, pos = 0.6] {$\sim$} (OH)
(O2X) edge[->] (O2Y)
(O2F) edge[->] (O2X)
(O2G) edge[->] (O2F)
(O2H) edge[->] (O2G)
(O2I) edge[->] (O2H);
\end{tikzpicture}\end{center}
\begin{proof}[Proof (\autoref{lem:splice})]
The map $H_{(n,\ell+1)}\to H_{(n,\ell)}$ is defined to be the given map $G_{m+\ell+1}^n\to G_{m+\ell}^n$. The map $H_{(n+1,0)}\to H_{(n,k-1)}$ is defined to be the composite
$$G_m^{n+1}\xrightarrow{\sim}G_{m+k}^n\to G_{m+k-1}^n.$$
It is easy to check that this is a long exact sequence from the conditions.
\end{proof}
\begin{proof}[Proof (\autoref{thm:spectrum-LES})]
For each $n:\ensuremath{\mathbb{Z}}\xspace$ we get a long exact sequence of homotopy groups for $f_{2-n}$ by
\autoref{thm:les-homotopy}. We splice them together using \autoref{lem:splice} with $N=(\ensuremath{\mathbb{Z}}\xspace,\lam{n}n+1)$ and $M=(\textbf{N},\lam{n}n+1)$ and with $k=3$ and $m=(2,0)$. This means that the resulting sequence is
\begin{center}\begin{tikzpicture}[thick, node distance=18mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$\pi_2(Y_{2-n})$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$\pi_2(X_{2-n})$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$\pi_2(F_{2-n})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of Y] (OY) {$\pi_2(Y_{2-(n+1)})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of X] (OX) {$\pi_2(X_{2-(n+1)})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of F] (OF) {$\pi_2(F_{2-(n+1)})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OY] (O2Y) {$\pi_2(Y_{2-(n+2)})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OX] (O2X) {$\pi_2(X_{2-(n+2)})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of OF] (O2F) {$\pi_2(F_{2-(n+2)})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = 5mm of O2X] {$\vdots$};
\ensuremath{\operatorname{\mathsf{node}}}[below = 5mm of X] {$\vdots$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] node [above] (f){$\pi_2(f)$} (Y)
(F) edge[->] node [below] (f){$\pi_2(p_1)$} (X)
(OY) edge[->] (F)
(OX) edge[->] node [above] (f){$\pi_2(f)$} (OY)
(OF) edge[->] node [below] (f){$\pi_2(p_1)$} (OX)
(O2Y) edge[->] node [left] (f){$\pi_2(\delta)\quad\mbox{}$} (OF)
(O2X) edge[->] node [above] (f){$\pi_2(f)$} (O2Y)
(O2F) edge[->] node [below] (f){$\pi_2(p_1)$} (O2X);
\end{tikzpicture}
\end{center}
We still need to check the conditions for the Lemma. The first isomorphism is given by the following composition
$$\pi_2(Y_{2-(n+1)})\simeq\pi_2(\Omega Y_{2-(n+1)+1})\simeq\pi_2(\Omega Y_{2-n})\equiv \pi_3(Y_{2-n}),$$
The second isomorphism is the same, replacing $Y$ by $X$. The square commutes because the two isomorphisms are both natural in $Y$.
\end{proof}
Suppose given a sequence of spectra $A$ and a sequence of spectrum maps
$$\cdots \to A_{s} \xrightarrow{f_{s}} A_{s-1} \xrightarrow{f_{s-1}} A_{s-2} \to \cdots $$
Let $B_s\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{fib}}}_{f_s}$. Then $D^{n,s}\ensuremath{\vcentcolon\equiv}\pi_n(A_s)$ and $E^{n,s}\ensuremath{\vcentcolon\equiv}\pi_n(B_s)$ are
graded abelian groups and the maps of the long exact sequences become graded
homomorphisms. This gives exactly the data of an exact couple.
For cohomology, it is customary to reindex the pages of the spectral sequence
with the base change $(p,q)=(s-n,-s)$, or equivalently $(n,s)=(-(p+q),-q)$.
\begin{thm}\label{thm:spectral-sequence-spectrum}
Given a sequence of spectra
$$\cdots \to A_{s} \xrightarrow{f_{s}} A_{s-1} \xrightarrow{f_{s-1}} A_{s-2}
\to \cdots $$ with fibers $B_s\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{fib}}}_{f_s}$, suppose for all $n$ there is
a $\beta_n$ such that for all $s\le \beta_n$ we have $\pi_n(A_s)=0$ and suppose that
for all $n$ there is a $\gamma_n$ such that for all $s>\gamma_n$ the map $\pi_n(f_s)$ is
an isomorphism. Then the exact couple constructed from this sequence is
bounded. This spectral sequence gives
$$E_2^{p,q}=\pi_{-(p+q)}(B_{-q})\Rightarrow \pi_{-(p+q)}(A_{\gamma_{-(p+q)}}).$$
\end{thm}
\begin{proof}
Note that for this spectral sequence we have
$\iota(n,s)\equiv\deg_i(n,s)\equiv(n,s-1)$ and $\kappa=\ensuremath{\operatorname{\mathsf{id}}}$. This means that
we need to show that for all $(n,s):\ensuremath{\mathbb{Z}}\xspace\times\ensuremath{\mathbb{Z}}\xspace$ there is a bound $\beta'_{n,s}$
such that for all $t\geq \beta'_{n,s}$ we have
$$E^{n,s+t}\equiv \pi_n(B_{s+t})=0\quad\text{ and }\quad D^{n,s-t}\equiv\pi_n(A_{s-t})=0.$$
Note that the right equation holds if $s-t\le \beta_n$, i.e. if $t\ge s-\beta_n$. By
the long exact sequence of homotopy groups we know that if $f_s:A_s\to
A_{s-1}$ induces an equivalence on both $\pi_n$ and $\pi_{n+1}$, then
$\pi_n(B_{s+t})=0$. So if we define
$$\beta'_{n,s}\ensuremath{\vcentcolon\equiv}\max(s-\beta_n,\gamma_n-s,\gamma_{n+1}-s),$$
we know that the exact couple is bounded with bound $\beta'$.
Now note that $x=(n,\gamma_n)$ is a stable index, because $\pi_n(f_{\gamma_n+t})$ is
surjective for all $t\geq0$. Therefore, by
\autoref{thm:exact-couple-convergence} we know that $D^{n,\gamma_n}$ is built from
$(E_\infty^{n,\gamma_n-s})_{0\le s\le \beta'_{n,\gamma_n}}$. If we apply the reindexing
$(p,q)=(s-n,-s)$, we get the desired relation
$$E_2^{p,q}=\pi_{-(p+q)}(B_{-q})\Rightarrow \pi_{-(p+q)}(A_{\gamma_{-(p+q)}}).$$
\end{proof}
\section{Spectral Sequences for Cohomology}\label{sec:spectral-sequence-cohomology}
Cohomology groups are algebraic invariants of types. They are often easier to
compute than homotopy groups, but they can also be used to compute certain
homotopy groups, often via the universal coefficient theorem and the Hurewicz
theorem (neither of which have been proven in HoTT yet).
The intermediate steps of most classical constructions of the singular cohomology are not homotopy invariant.
\emph{Cellular cohomology} is only defined for cell complexes and not for arbitrary
spaces, but it can be defined in HoTT~\cite{buchholtz2018cellular}. \emph{Singular
cohomology} is defined as a quotient of a large abelian group that is not
homotopy invariant, which makes this definition impossible in HoTT. However,
classically, Eilenberg-MacLane spaces represent cohomology, and we can use this
fact as the \emph{definition} of cohomology in HoTT~\cite{cavallo2015cohomology}.
Normally cohomology groups have coefficients in an abelian group, but more
generally they can have coefficients in a spectrum, or even a family of spectra.
In this section we will define cohomology groups and construct the
Atiyah-Hirzebruch spectral sequence for cohomology. This is a generalization of
the spectral sequence defined in~\cite{atiyah1961spectral} in the special case
of topological K-theory. From the Atiyah-Hirzebruch spectral sequence we can
construct the Serre spectral sequence, sometimes also called the Leray-Serre spectral sequence.
\begin{defn}
Suppose given $X:\ensuremath{\mathsf{Type}}\xspace^*$ and $Y:X\to\ensuremath{\mathsf{Spectrum}}$. We define $(x:X)\to^* Yx:\ensuremath{\mathsf{Spectrum}}$ such that
$((x:X)\to^* Yx)_n\ensuremath{\vcentcolon\equiv} (x:X)\to^*(Yx)_n$. If $Y$ does not depend on $X$, we write $X\to^* Y$.
For an unpointed type $X:\ensuremath{\mathsf{Type}}\xspace$ and $Y:X\to\ensuremath{\mathsf{Spectrum}}$, we similarly define $(x:X)\to Yx:\ensuremath{\mathsf{Spectrum}}$ such that
$((x:X)\to Yx)_n\ensuremath{\vcentcolon\equiv} (x:X)\to(Yx)_n$ (this has as basepoint the constant map into the basepoint of $(Yx)_n$), and abbreviate this to $X\to Y$ if $Y$ does not depend on $X$.
\end{defn}
These spectra are well-defined, since we have
$$\Omega((a:A)\to^*(Ba))\simeq (a:A)\to^*\Omega(Ba)$$
and
$$\Omega((a:A)\to(Ba))\simeq (a:A)\to\Omega(Ba).$$
Moreover, they satisfy the expected properties of dependent product. In particular, if $X:\ensuremath{\mathsf{Type}}\xspace^*$ and $Y,Z:X\to\ensuremath{\mathsf{Spectrum}}$ and moreover if we have a fiberwise spectrum map $f:(x:X)\to Yx \to Zx$, this induces a map on the dependent products
$$\Pi_f:((x:X)\to Yx)\to ((x:X)\to Zx).$$
\begin{defn}\label{def:cohomology}
Suppose given $X:\ensuremath{\mathsf{Type}}\xspace^*$, $Y:X\to\ensuremath{\mathsf{Spectrum}}$ and $n:\ensuremath{\mathbb{Z}}\xspace$. We define the \emph{geneneralized, parametrized, reduced cohomology} of $X$ with coefficients in $Y$ as\footnote{We will write $\lam{x}Yx$ in $\eta$-expanded form to remember that this is parametrized cohomology.}
$$\tilde H^n(X;\lam{x}Yx)\ensuremath{\vcentcolon\equiv} \pi_{-n}((x:X)\to^* Yx)\simeq \|(x:X)\to^* (Yx)_n\|_0.$$
If $Y$ does not depend on $X$, we have the \emph{unparametrized cohomology} as
$$\tilde H^n(X;Y)\ensuremath{\vcentcolon\equiv} \pi_{-n}(X\to^* Yx)\simeq \|X\to^* Y_n\|_0.$$
If $X:\ensuremath{\mathsf{Type}}\xspace$ is an arbitrary type, we define the \emph{unreduced cohomology} as
$$H^n(X;\lam{x}Yx)\ensuremath{\vcentcolon\equiv} \pi_{-n}((x:X)\to Yx)\simeq \|(x:X)\to (Yx)_n\|_0\simeq \tilde H^n(X_+;\lam{x}Y_+x).$$
Here $X_+\ensuremath{\vcentcolon\equiv} X+1:\ensuremath{\mathsf{Type}}\xspace^*$ and $Y_+:X_+\to\ensuremath{\mathsf{Spectrum}}$ is defined as $Y_+(\ensuremath\inlsym\xspace(x))\ensuremath{\vcentcolon\equiv} Yx$ and $Y_+(\ensuremath\inrsym\xspace(\star))\ensuremath{\vcentcolon\equiv} 1$.
If $X:\ensuremath{\mathsf{Type}}\xspace^*$ and $A:X\to\ensuremath{\mathsf{AbGroup}}$, we define the \emph{ordinary cohomology} as
$$\tilde H^n(X;\lam{x}Ax)\ensuremath{\vcentcolon\equiv} \tilde H^n(X;\lam{x}H(Ax).$$
We can combine the attributes ordinary/generalized, parametrized/unparametrized and reduced/unreduced for cohomology however we want, leading to eight different notions.
We define
$$\tilde H^n(X)\ensuremath{\vcentcolon\equiv} \tilde H^n(X;\ensuremath{\mathbb{Z}}\xspace)$$
and similarly for unreduced cohomology.
\end{defn}
Unparametrized cohomology satisfies the Eilenberg-Steenrod axioms for cohomology. Although we will not use this fact in this chapter, for completeness we will state it here.
To give the definition we need to introduce one more concept.
\begin{defn}\label{def:choice}
A type $X$ has $n$-choice for $n\geq-2$ if for all $P:X\to\ensuremath{\mathsf{Type}}\xspace$ the canonical map
$$\|(x:X)\to Px\|_n\to ((x:X) \to \|Px\|_n)$$
is an equivalence.
\end{defn}
Note that in particular $\ensuremath{\operatorname{\mathsf{fin}}}_k$ has $n$-choice for all $k,n$.
\begin{defn}
A \emph{unparametrized reduced cohomology theory} is a contravariant functor
$\tilde E^n:\ensuremath{\mathsf{Type}}\xspace^*\to\ensuremath{\mathsf{AbGroup}}$ for every $n:\ensuremath{\mathbb{Z}}\xspace$ satisfying the
\emph{Eilenberg-Steenrod axioms}. Functoriality means that for a pointed map
$f:X\to^* Y$ there is a map $\tilde E^n(f):\tilde E^n(Y)\to \tilde E^n(X)$ such that $\tilde E(\ensuremath{\operatorname{\mathsf{id}}})\sim^*\ensuremath{\operatorname{\mathsf{id}}}$ and $\tilde E(g\o
f)\sim^*\tilde E(f)\o\tilde E g$. The Eilenberg-Steenrod axioms are
\begin{itemize}
\item \emph{(Suspension axiom)} There is a natural transformation $\tilde E^{n+1}(\Sigma X)\simeq \tilde E^n(X)$.
\item \emph{(Exactness)} Given a cofiber sequence $X\xrightarrow{f}Y\xrightarrow{g}Z$, the sequence
$$\tilde E^n(Z)\xrightarrow{\tilde E^n(g)}\tilde E^nY\xrightarrow{\tilde E^n(f)}\tilde E^n(X)$$
is exact at $\tilde E^n(Y)$.
\item \emph{(Additivity)} Suppose given a type $I$ satisfying 0-choice and $X:I\to\ensuremath{\mathsf{Type}}\xspace^*$. Then the canonical homomorphism
$$\tilde E^n\big(\bigvee_i Xi\big)\to ((i:I)\to \tilde E^n(Xi))$$
is an isomorphism.
\end{itemize}
A cohomology theory is called \emph{ordinary} if it also satisfies the following axiom.
\begin{itemize}
\item \emph{(Dimension)} If $n\neq 0$, then $\tilde E^n(\ensuremath{\S^0})$ is trivial.
\end{itemize}
\end{defn}
The following theorem has been proven in~\cite{cavallo2015cohomology}. We will not repeat the proof here.
\begin{thm}\label{thm:cohomology-theory}
Unparametrized generalized reduced cohomology is a cohomology theory. Ordinary cohomology also satisfies the dimension axiom.
\end{thm}
We will not use \autoref{thm:cohomology-theory} in the remainder of this chapter.
To construct the Atiyah-Hirzebruch spectral sequence, we need the Postnikov tower of a spectrum.
\begin{defn}
We say that for $k:\ensuremath{\mathbb{Z}}\xspace$ a spectrum $Y$ is \emph{$k$-truncated} if $Y_n$ is $(k+n)$-truncated for all $n:\ensuremath{\mathbb{Z}}\xspace$ (using the convention that any type is $\ell$-truncated for $\ell\leq-2$).
The \emph{$k$-truncation} of a spectrum $Y$, written $\|Y\|_k$, is defined as $(\|Y\|_k)_n\ensuremath{\vcentcolon\equiv} \|Y_n\|_{k+n}$ where we define $\|A\|_\ell=\ensuremath{\mathbf{1}}\xspace$ for $\ell\leq-2$.
\end{defn}
\begin{lem}\label{lem:spectrum-trunc-properties}
The usual properties of truncations also hold for spectra. In particular we will use that there is a spectrum map $|{-}|_k:Y\to\|Y\|_k$ and that if $Z$ is $k$-truncated, then a spectrum map $f:Y\to Z$ induces a spectrum map $\|Y\|_k\to Z$.
\end{lem}
\begin{proof}
The underlying maps are the corresponding facts for pointed maps. The fact that these maps are spectrum maps comes from the fact that these operations commute with taking loop spaces. We omit the details here.
\end{proof}
\begin{lem}[Postnikov Tower for spectra]\label{lem:postnikov-tower-spectra}
For $s:\ensuremath{\mathbb{Z}}\xspace$ and $Y:\ensuremath{\mathsf{Spectrum}}$ there is a spectrum map $f^s:\|Y\|_s\to\|Y\|_{s-1}$ that levelwise has fiber $\Sigma^n(H\pi_s(Y))$. That is,
$$(\ensuremath{\operatorname{\mathsf{fib}}}_{f^s})_k\simeq^*(\Sigma^s(H\pi_s(Y)))_k.$$
\end{lem}
We should be able to extend this equivalence to a spectrum equivalence, but we do not need this strengthening for the remainder of the proof.
\begin{proof}
Note that $\|Y\|_{s-1}$ is $(s-1)$-truncated, and therefore $s$-truncated. By
the elimination of spectrum truncation in
\autoref{lem:spectrum-trunc-properties} we get a spectrum map
$f^s:\|Y\|_s\to\|Y\|_{s-1}$. For the levelwise pointed equivalence, we need to
show that
$$\ensuremath{\operatorname{\mathsf{fib}}}_{f^s_k}\simeq^*K(\pi_s(Y),s+k)$$
To show this, by \autoref{thm:em-unique} we need to show that $\ensuremath{\operatorname{\mathsf{fib}}}_{f^s_k}$
is $(s+k)$-truncated, $(s+k-1)$-connected and $\pi_{s+k}(\ensuremath{\operatorname{\mathsf{fib}}}_{f^s_k})\simeq
\pi_s(Y)$.
Note that $f^s_k:\|Y_k\|_{s+k}\to\|Y_k\|_{s+k-1}$, so the truncatedness
follows because the domain and codomain of $f^s_k$ are both $(s+k)$-truncated.
For the connectedness, we know that $|{-}|_{s+k-1}:Y_k\to \|Y_k\|_{s+k-1}$ is
$(s+k-1)$-connected, and the elimination principle for truncations preserve
connectedness, therefore $f^s_k$ is $(s+k-1)$-connected. To compute the
homotopy group, we look at a piece of the long exact sequence for homotopy
groups for $f^s_k$ at level $s+k$ and $s+k+1$.
\begin{center}\begin{tikzpicture}[node distance=10mm]
\ensuremath{\operatorname{\mathsf{node}}} (Y) at (0,0) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of Y] (X) {$\pi_{s+k}(Y)$};
\ensuremath{\operatorname{\mathsf{node}}}[left = of X] (F) {$\pi_{s+k}(\ensuremath{\operatorname{\mathsf{fib}}}_{f^s_k})$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of Y] (OY) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of X] (OX) {$0$};
\ensuremath{\operatorname{\mathsf{node}}}[above = of F] (OF) {$\bullet$};
\path[every node/.style={font=\sffamily\small}]
(X) edge[->] (Y)
(F) edge[->] (X)
(OY) edge[->] (F)
(OX) edge[->] (OY)
(OF) edge[->] (OX);
\end{tikzpicture}
\end{center}
Since we have the exact sequence
$0\to\pi_{s+k}(\ensuremath{\operatorname{\mathsf{fib}}}_{f^s_k})\to\pi_{s+k}(Y)\to0$, the middle map must be an
equivalence, which finishes the proof.
\end{proof}
For a spectrum $Y$, we get the Postnikov tower
$$\cdot\to\|Y\|_s\to\|Y\|_{s-1}\to\|Y\|_{s-2}\to\cdots$$
This satisfies the conditions of \autoref{thm:spectral-sequence-spectrum}, but
unfortunately the spectral sequence constructed from this is trivial. We need another ingredient to get an interesting spectral sequence.
\begin{lem}\label{lem:spi-functor}
Suppose given $X:\ensuremath{\mathsf{Type}}\xspace^*$ and two family of spectra $Y,Z:X\to\ensuremath{\mathsf{Spectrum}}$. A family of spectrum maps
$$f:(x:X)\to Yx \to Zx$$
induces a spectrum map between the spectra of sections for $Y$ and $Z$:
$$f\o({-}):((x:X)\to Yx)\to((x:X)\to Zx).$$
Moreover, the fiber of this spectrum map is levelwise $(x:X)\to \ensuremath{\operatorname{\mathsf{fib}}}_{fx}$, that is
$$(\ensuremath{\operatorname{\mathsf{fib}}}_{f\o({-})})_n \simeq^* ((x:X)\to \ensuremath{\operatorname{\mathsf{fib}}}_{fx})_n.$$
\end{lem}
The levelwise equivalence should be extendable to a spectrum equivalence, but we do not need that in this chapter.
\begin{proof}
We define (see \autoref{lem:pointed-types-basic}.\ref{item:fiber-composition})
$$(f\o({-}))_n\ensuremath{\vcentcolon\equiv} f_n \o ({-}):((x:X)\to^* (Yx)_n)\to^*((x:X)\to^* (Zx)_n).$$
This is a spectrum map because of the pointed function extensionality mentioned in \autoref{lem:pointed-types-basic}.\ref{item:pointed-function-extensionality}.
By \autoref{lem:pointed-types-basic}.\ref{item:fiber-composition} the fiber of this map is levelwise $(x:X)\to \ensuremath{\operatorname{\mathsf{fib}}}_{fx}$.
\end{proof}
We now have all the ingredients of the Atiyah-Hirzebruch spectral sequence.
\begin{thm}[Atiyah-Hirzebruch spectral sequence for reduced cohomology]\label{thm:atiyah-hirzebruch-reduced}
If $X:\ensuremath{\mathsf{Type}}\xspace^*$ is a pointed type and $Y:X\to k\operatorname{-\ensuremath{\mathsf{Spectrum}}}$ is a family of
$k$-truncated spectra over $X$, then we get a spectral sequence with
$$E_2^{p,q}=\tilde H^p(X;\lam{x}\pi_{-q}(Yx))\Rightarrow \tilde H^{p+q}(X;\lam{x}Yx).$$
\end{thm}
\begin{proof}
Define $A_s\ensuremath{\vcentcolon\equiv}((x:X)\to^* \|Yx\|_s)$ and consider the sequence of spectra
$$\cdots \to A_s\xrightarrow{f_s} A_{s-1} \xrightarrow{f_{s-1}} A_{s-2} \to \cdots$$
where $f_s$ is the map induced by the Postnikov tower. By
\autoref{lem:spi-functor} and \autoref{lem:postnikov-tower-spectra} $f_s$
levelwise has fiber $B_s\ensuremath{\vcentcolon\equiv}(x:X)\to^*\Sigma^sH\pi_s(Yx)$. We want to apply
\autoref{thm:spectral-sequence-spectrum}, so we need to check the conditions
of that theorem. For $n:\ensuremath{\mathbb{Z}}\xspace$ we define $\beta_n\ensuremath{\vcentcolon\equiv} n-1$. Notice that $A_s$
is $s$-truncated, and thus for $s\le B_n$ we have
$$\pi_n(A_s)\ensuremath{\vcentcolon\equiv} \pi_n((x:X)\to^* \|Yx\|_s)=0.$$
For $n:\ensuremath{\mathbb{Z}}\xspace$ define $\gamma_n\ensuremath{\vcentcolon\equiv} k$. Then for $s\geq \gamma_n$ the spectrum
$A_s$ is levelwise equivalent to $(x:X)\to^* Yx$, so for $s>\gamma_n$ the map
$A_s\to A_{s-1}$ becomes levelwise the identity map under that equivalence.
This means that $f_s$ is an equivalence, so in particular $\pi_n(f_s)$ is an
isomorphism. By \autoref{thm:spectral-sequence-spectrum} we now get the
spectral sequence
$$E_2^{p,q}=\pi_{-(p+q)}(B_{-q})\Rightarrow \pi_{-(p+q)}(A_k).$$
We now compute
\begin{align*}
\pi_{-(p+q)}(B_{-q})&\simeq\pi_{-(p+q)}((x:X) \to^* \Sigma^{-q}H\pi_{-q}(Yx))\\
&\simeq\tilde H^{p+q}(X;\lam{x}\Sigma^{-q}H\pi_{-q}(Yx))\\
&\simeq\tilde H^{p}(X;\lam{x}\pi_{-q}(Yx))
\intertext{and}
\pi_{-(p+q)}(A_k)&\simeq \pi_{-(p+q)}((x:X)\to^*\|Yx\|_k)\\
&\simeq\pi_{-(p+q)}((x:X)\to^* Yx)\\
&\simeq\tilde H^{p+q}(X;\lam{x}Yx).
\end{align*}
\end{proof}
We also have the corresponding spectral sequence for unreduced cohomology.
\begin{cor}[Atiyah-Hirzebruch spectral sequence for unreduced cohomology]\label{cor:atiyah-hirzebruch-unreduced}
If $X:\ensuremath{\mathsf{Type}}\xspace$ is any type and $Y:X\to k\operatorname{-\ensuremath{\mathsf{Spectrum}}}$ is a family
of $k$-truncated spectra over $X$, then
$$E_2^{p,q}=H^p(X;\lam{x}\pi_{-q}(Yx))\Rightarrow H^{p+q}(X;\lam{x}Yx).$$
\end{cor}
\begin{proof}
Apply \autoref{thm:atiyah-hirzebruch-reduced} to $X_+$ and $Y_+$ (defined in \autoref{def:cohomology}).
\end{proof}
From the Atiyah-Hirzebruch spectral sequence we can construct the Serre spectral sequence.
\begin{thm}[Serre spectral sequence for cohomology]\label{thm:serre-spectral-sequence}
Suppose given $B:\ensuremath{\mathsf{Type}}\xspace$, a family of types $F:B\to\ensuremath{\mathsf{Type}}\xspace$ and a spectrum
$Y:\ensuremath{\mathsf{Spectrum}}$ that is $k$-truncated. Then
$$E_2^{p,q}=H^p(B;\lam{b}H^q(Fb;Y))\Rightarrow H^{p+q}((b:B)\times Fb;Y).$$
\end{thm}
\begin{proof}
Apply \autoref{cor:atiyah-hirzebruch-unreduced} to the type $B$ and and the family of spectra $\lam{b}Fb\to Y$, which is $k$-truncated. Then we get
$$E_2^{p,q}=H^p(B;\lam{b}\pi_{-q}(Fb\to Y))\Rightarrow H^{p+q}(B;\lam{b}Fb\to Y).$$
Note that $\pi_{-q}(Fb\to Y)\simeq H^q(Fb;Y)$, so the second page is the desired group, and for the $\infty$-page we compute
\begin{align*}
H^{p+q}(B;\lam{b}Fb\to Y) &= \pi_{-(p+q)}((b:B) \to Fb \to Y)\\
&= \pi_{-(p+q)}(((b:B) \times Fb)\to Y)\\
&= H^{p+q}((b:B)\times Fb;Y).
\end{align*}
\end{proof}
Equivalent to the data given in \autoref{thm:serre-spectral-sequence} is a map
$X\to B$ and a $k$-truncated spectrum $Y$. In that case we get the spectral
sequence
$$E_2^{p,q}=H^p(B;\lam{b}H^q(\ensuremath{\operatorname{\mathsf{fib}}}_f(b);Y))\Rightarrow H^{p+q}(X;Y).$$ Analogous
to the proof of \autoref{thm:serre-spectral-sequence} we also have a version
when $Y$ is parametrized over $B$. In that case we get
$$E_2^{p,q}=H^p(B;\lam{b}H^q(\ensuremath{\operatorname{\mathsf{fib}}}_f(b);Y))\Rightarrow H^{p+q}(X;\lam{x}Y(fx)).$$
We get a useful special case of the Serre spectral sequence when the family
$\lam{b}H^q(Fb;Y)$ is constant. This happens in particular when $B$ is simply
connected.
\begin{cor}\label{cor:serre-spectral-sequence-conn}
Suppose given a simply connected pointed type $B:\ensuremath{\mathsf{Type}}\xspace^*$, a family of types $F:B\to\ensuremath{\mathsf{Type}}\xspace$ and a spectrum
$Y:\ensuremath{\mathsf{Spectrum}}$ that is $k$-truncated. Then
$$E_2^{p,q}=H^p(B;H^q(Fb_0;Y))\Rightarrow H^{p+q}((b:B)\times Fb;Y).$$
\end{cor}
\begin{proof}
Apply \autoref{thm:serre-spectral-sequence}. The family
$\lam{b}H^q(Fb;Y):B\to\ensuremath{\mathsf{AbGroup}}$ is a family of sets. Since $B$ is simply
connected, every such family is constant, so all fibers are equal to $H^q(Fb_0;Y)$.
\end{proof}
In the spectral sequences constructed we assumed that the spectra were truncated.
The reason we need this assumption is that the notion of convergence we used for spectral sequence is the eventual value of the sequence.
If we had a stronger notion of convergence, we might be able to relax the truncatedness condition.
However, there is another reason why the spectral sequence can become (pointwise) eventually constant.
Instead of assuming that the spectra are truncated, we can pose a restriction on the base space.
\begin{defn}\label{def:weak-pointed-choice}
We say that $X:\ensuremath{\mathsf{Type}}\xspace^*$ satisfies \emph{weak pointed choice} if there is a natural number $n$ such that for all families $Y:X\to\ensuremath{\mathsf{Type}}\xspace^*$ of $n$-connected types the type of dependent pointed maps $(x:X)\to^* Y(x)$ is 0-connected.
\end{defn}
\begin{ex}\mbox{}
\begin{itemize}
\item The spheres $\S^n$ satisfy weak pointed choice. The proof is easy for $n=0$, which we will skip.
For $\S^{n+1}$, note that
$$((x:S^{n+1})\to^* Y(x))\simeq \star =_{\ensuremath{\mathsf{surf}}\xspace}^{\Omega^nY({-})} \star$$
where $\ensuremath{\mathsf{surf}}\xspace:\Omega^{n+1}\S^{n+1}$ is the surface of $\S^{n+1}$ and
$\star:\Omega^nY(\ensuremath{\mathsf{base}}\xspace)$ is the basepoint. Now if $Y$ is a family of $(n+1)$-connected types, then $\Omega^nY({-})$ is a family of 1-connected types, and a pathover in that family is 0-connected, as desired.
\item Suppose $I$ is a type that satisfies 0-choice (see \autoref{def:choice}). Then the collection of types that satisfy weak pointed choice are closed under $I$-indexed wedges. This follows from the dependent universal property of the wedge.
$$\big(\bigvee_{i:I}X_i\to^* P(x)\big)\simeq (i:I) \to (x : X_i) \to^* P(\ensuremath{\operatorname{\mathsf{in}}}_i(x)).$$
\end{itemize}
\end{ex}
\begin{thm}
If $X:\ensuremath{\mathsf{Type}}\xspace^*$ satisfies weak pointed choice and $Y:X\to\ensuremath{\mathsf{Spectrum}}$ is any family of spectra, we get the spectral sequence in \autoref{thm:atiyah-hirzebruch-reduced}:
$$E_2^{p,q}=\tilde H^p(X;\lam{x}\pi_{-q}(Yx))\Rightarrow \tilde H^{p+q}(X;\lam{x}Yx).$$
\end{thm}
\begin{proof}
The proof is the mostly the same as for \autoref{thm:atiyah-hirzebruch-reduced}. The only difference is in showing that the sequence stabilizes on homotopy groups when $s$ is large. Suppose $X$ satisfies choice with respect to $k$-connected families.
For $n:\ensuremath{\mathbb{Z}}\xspace$ define $\gamma_n\ensuremath{\vcentcolon\equiv} n+k$. Then for $s>\gamma_n$ we know that the fiber of $f_s$ has as $k$-th homotopy group
$$\pi_\ell(B_s)=\pi_0(\Omega^\ell B_s)=\|(x:X)\to^* K(\pi_s(Yx),s-\ell)\|_0.$$
This is a product into a family of $(s-\ell-1)$-connected types, which for $\ell=n$ and $\ell=n-1$ is a family of at least $k$-connected types. By the weak choice principle on $X$ this type is 0-connected, so these homotopy groups are trivial. Now by the long exact sequence of homotopy groups for $f_s$ the map $\pi_n(f_s)$ is an isomorphism, as required.
\end{proof}
\section{Spectral Sequences for Homology}\label{sec:spectral-sequence-homology}
Homology theory has not been developed as much as cohomology theory in HoTT. It is known that the homology given by a prespectrum forms a homology theory~\cite{graham2017homology}. Lemma 18 in that paper was not proven carefully, but it follows from the results in \autoref{sec:smash-product}.
In this section, we sketch the construction of the Atiyah-Hirzebruch and Serre spectral sequences for homology~\cite{serre1951homology}. The results in the section are not proven in HoTT, and are therefore stated as remarks without proof.
If $X:\ensuremath{\mathsf{Type}}\xspace^*$ and $Y:\ensuremath{\mathsf{Prespectrum}}$, we can define $X\wedge Y:\ensuremath{\mathsf{Prespectrum}}$ with
$$(X\wedge Y)_n\ensuremath{\vcentcolon\equiv} X\wedge Y_n.$$
To show that it is a prespectrum, recall the adjunction between the suspension and the loop. For pointed types $X$ and $Y$ we have a natural equivalence
$$(\Sigma X \to^* Y)\simeq (X \to^* \Sigma Y).$$
Therefore, to characterize a prespectrum, it is sufficient to give a map $f_n:\Sigma Y_n \to Y_{n+1}$. This is given for the smash prespectrum as the composite
$$\Sigma(X\wedge Y_n)\xrightarrow{\sim} X \wedge \Sigma Y_n \xrightarrow{X\wedge f_n} X \wedge Y_{n+1}.$$
We can define reduced homology
$$\tilde H_n(X;Y)\ensuremath{\vcentcolon\equiv} \pi_n(X\wedge Y).$$
For the construction of parametrized homology we need to generalize the smash product.
\begin{defn}
Given $A : \ensuremath{\mathsf{Type}}\xspace^*$ and $B : A \to \ensuremath{\mathsf{Type}}\xspace^*$, we define the parametrized smash
$$(x : A) \wedge B(x)$$
to be the pushout
\begin{center}
\begin{tikzpicture}[node distance=10mm,baseline={([yshift={-\ht\strutbox}]current bounding box.north)}]
\ensuremath{\operatorname{\mathsf{node}}} (tl) at (0,0) {$A+B$};
\ensuremath{\operatorname{\mathsf{node}}}[right = 25mm of tl] (tr) {$2$};
\ensuremath{\operatorname{\mathsf{node}}}[below = of tl] (bl) {$(x:A)\times B(x)$};
\ensuremath{\operatorname{\mathsf{node}}} (br) at (tr |- bl) {$(x : A) \wedge B(x)$};
\path[every node/.style={font=\sffamily\small}]
(tl) edge[->] node {} (tr)
edge[->] node {} (bl)
(tr) edge[->] node {} (br)
(bl) edge[->] node {} (br);
\end{tikzpicture}\quad
\begin{tikzpicture}[baseline={([yshift={-\ht\strutbox}]current bounding box.north)}]
\draw[fill=black!5, thin] (0,0) ellipse (20mm and 8mm);
\ensuremath{\operatorname{\mathsf{node}}}[label=below:{$a_0$}] at (0,-1.98) {$\bullet$};
\ensuremath{\operatorname{\mathsf{node}}} at (2.3,-2) {$A$};
\ensuremath{\operatorname{\mathsf{node}}} at (2.3,0) {$B$};
\path[every node/.style={font=\sffamily\scriptsize}]
(-2,-2) edge[thick] (2,-2)
(0,-1.1) edge[->] (0,-1.8)
(0,0) edge[bend left = 10, thick] (-2,0)
edge[bend left = 10, thick] node[below] {$b_0$} (2,0)
edge[bend left = 5, thick] node[left] {$B(a_0)$} (0,0.8)
edge[bend left = 5, thick] (0,-0.8);
\end{tikzpicture}
\end{center}
\end{defn}
\begin{rmk}The strategy for constructing the spectral sequences for homology is as follows.
\begin{itemize}
\item The parametrized smash is (should be) left adjoint to pointed dependent maps.
That means that there is a natural equivalence
$$(((x:A) \wedge Bx) \to^* C) \simeq (x:A) \to^* Bx \to^* C.$$
\item From this we get (natural) equivalences
$$\Sigma((x:A)\wedge Bx) \simeq ((x:A) \wedge \Sigma(Bx));$$
$$((x:A)\wedge Bx) \wedge C \simeq (x:A) \wedge (Bx \wedge C);$$
$$(x:A_+)\wedge B_+x) \wedge C \simeq (x:A) \times Bx.$$
The proofs of these properties should be similar to the proofs in \autoref{sec:smash-monoidal}.
\item Therefore, for $X:\ensuremath{\mathsf{Type}}\xspace^*$ and $Y:X\to\ensuremath{\mathsf{Prespectrum}}$ we have a prespectrum $(x : X)\wedge Yx$. The maps are given by the above equivalence.
\item We can now define parametrized (reduced, generalized) homology as
$$H_n(X;\lam{x}Yx) \ensuremath{\vcentcolon\equiv} \pi_n((x : X) \wedge Yx).$$
We can define unreduced homology by adding a point to $X$, in the same way as for cohomology.
\item As before, given $X:\ensuremath{\mathsf{Type}}\xspace^*$ and $Y:X\to\ensuremath{\mathsf{Spectrum}}$, we can again form the Postnikov tower of $Yx$ for any $x:X$. We now want to take the parametrized smash over $X$, but there is no hope to compute the fiber of this spectrum.
\item However, we should be able to do it when we work in spectra. The forgetful
functor $\ensuremath{\mathsf{Spectrum}} \to \ensuremath{\mathsf{Prespectrum}}$ has a left adjoint, called
\emph{spectrification}. The spectrification $LY$ of a prespectrum $Y$ can be
constructed either as a higher inductive family of types~\cite{shulman2011spectrification} or as the colimit
$$(LY)_n\ensuremath{\vcentcolon\equiv}\ensuremath{\operatorname{\mathsf{colim}}}_{k\to\infty}\Omega^k Y_{n+k}.$$
For neither definition a careful proof of the adjunction has been given.
\item We can now define the parametrized smash of spectra as the spectrification of the parametrized smash for prespectra. This should preserve cofiber sequences of spectra, in the sense that if
$$Ax \to Bx \to Cx$$
is a family of cofiber sequences of spectra indexed by $x:\ensuremath{\mathsf{Type}}\xspace^*$, the following sequence is also a cofiber sequence of spectra
$$((x:X)\wedge Ax)\to ((x:X)\wedge Bx)\to ((x:X)\wedge Cx)$$
\item A sequence of spectra should be a fiber sequence of spectra if and only if it is a cofiber sequence of spectra. This is true classically, and should also hold in HoTT.
\item Assuming that all the above properties have been proven, we can get the Atiyah-Hirzebruch spectral sequence for reduced homology. Suppose given a pointed type $X$ and $Y:X\to\ensuremath{\mathsf{Spectrum}}$ a family of spectra. We can apply
\autoref{thm:spectral-sequence-spectrum} to the iterated fiber sequence
$$((x:X)\wedge \Sigma^nH)\to((x:X)\wedge \|Yx\|_s)\to ((x:X)\wedge
\|Yx\|_{s-1}).$$ To satisfy the conditions for that theorem we need to assume
some conditions on $X$ and/or $Y$. In particular it is sufficient if $Y$ is a family of
truncated and connected spectra, but weaker conditions might also suffice.
Using homological indexing (where $p$ and $q$ have their sign reversed) we get
$$E^2_{p,q}=\pi_{p+q}(B_q)\Rightarrow \pi_{p+q}(A_{\gamma_{p+q}}).$$
Now we compute
\begin{align*}
\pi_{p+q}(B_q)&\simeq\pi_{p+q}((x:X) \wedge \Sigma^qH\pi_q(Yx))\\
&\simeq\tilde H_{p+q}(X;\lam{x}\Sigma^qH\pi_q(Yx))\\
&\simeq\tilde H_p(X;\lam{x}\pi_q(Yx))
\intertext{and}
\pi_{p+q}(A_k)&\simeq \pi_{p+q}((x:X)\wedge\|Yx\|_k)\\
&\simeq\pi_{p+q}((x:X)\to Yx)\\
&\simeq\tilde H_{p+q}(X;\lam{x}Yx).
\end{align*}
This gives the desired spectral sequence:
$$E^2_{p,q}=\tilde H_p(X;\lam{x}\pi_q(Yx))\Rightarrow \tilde H_{p+q}(X;\lam{x}Yx).$$
\item We get the Atiyah-Hirzebruch spectral sequence for \emph{unreduced} homology in the same way as for cohomology, by applying the version for reduced homology to $X_+$ and $Y_+$.
\item We get the Serre spectral sequence for homology also in the same way. Suppose given
$B:\ensuremath{\mathsf{Type}}\xspace$ and $F:B\to\ensuremath{\mathsf{Type}}\xspace$ and a truncated spectrum $Y$. Applying the Atiyah-Hirzebruch spectral sequence for unreduced homology to the type $B$ and the spectrum $\lam{b}Fb\wedge Y$ we get
$$E^2_{p,q}=H_p(B;\lam{b}\pi_q(Fb\wedge Y))\Rightarrow H_{p+q}(B;\lam{b}Fb\wedge Y).$$
The second page is what we want. For the $\infty$-page we compute
\begin{align*}
H_{p+q}(B;\lam{b}Fb\to Y) &= \pi_{p+q}((b:B_+) \wedge (F_+b \wedge Y))\\
&= \pi_{p+q}(((b:B_+) \wedge F_+b)\wedge Y)\\
&= \pi_{p+q}(((b:B) \times Fb)\wedge Y)\\
&= H_{p+q}((b:B)\times Fb;Y).
\end{align*}
This gives the Serre spectral sequence for homology:
$$E^2_{p,q}=H_p(B;\lam{b}H_q(Fb;Y))\Rightarrow H_{p+q}((b:B)\times Fb;Y).$$
\end{itemize}
\end{rmk}
\begin{rmk}
We can also use the parametrized smash to get a spectral sequence for reduced
homology and reduced cohomology. Suppose given $B:\ensuremath{\mathsf{Type}}\xspace^*$ and a family of types $F:B\to\ensuremath{\mathsf{Type}}\xspace^*$ and a spectrum $Y:\ensuremath{\mathsf{Spectrum}}$ that is $k$-truncated. Then we get the following two spectral sequences
$$E_2^{p,q}=\tilde H^p(B;\lam{b}\tilde H^q(Fb;Y))\Rightarrow \tilde H^{p+q}((b:B)\wedge Fb;Y);$$
$$E^2_{p,q}=\tilde H_p(B;\lam{b}\tilde H_q(Fb;Y))\Rightarrow \tilde H_{p+q}((b:B)\wedge Fb;Y).$$
For homology, the proof is the same as above. For cohomology, we apply the Atiyah-Hirzebruch spectral sequence for reduced cohomology to the pointed type $B$ and the family of spectra $\lam{b}Fb\to^* Y$. We get the desired spectral sequence by the adjunction between parametrized smash and dependent pointed maps.
These spectral sequences generalize \autoref{thm:serre-spectral-sequence} and the corresponding version for homology: we get those versions back when we add a point to $B$ and $F$. Whether this extra generality is useful is unknown.
\end{rmk}
\section{Applications of Spectral Sequences}\label{sec:applications-spectral-sequences}
Classically, there are many applications of the Serre and Atiyah-Hirzebruch spectral sequences.
Here we will list some of these applications, and give thoughts on how to translate these results in HoTT.
The results in this section have not been formalized. Before we start, we compute the cohomology of spheres.
\begin{lem}\label{lem:cohomology-spheres}
If $n\geq 1$, then
\begin{equation}H^k(\S^n;A)=\begin{cases}A&\text{if $k\in\{0,n\}$}\\0&\text{otherwise.}\end{cases}\label{eq:cohomology-spheres}\end{equation}
\end{lem}
This is a special case of the universal coefficient theorem, which we do not have yet in HoTT.
However, we can prove these equalities directly from the definition of cohomology.
\begin{proof}
For $k=0$ we have
$$H^0(\S^n;A)=\|\S^n\to A\|_0=(\S^n\to A) = A,$$
where we use that $\S^n$ is 0-connected. For $k\neq 0$ we have
$$H^k(\S^n;A)=\tilde H^k(\S^n+1;A)=\tilde H^k(\S^n;A)=\|\S^n\to^* K(A,k)\|_0=\|\Omega^n K(A,k)\|_0.$$
Now for $n<k$ the type $\Omega^n K(A,k)$ is 0-connected, hence the result is contractible. For $n=k$ the result is $A$, and for $n>k$ the type $\Omega^n K(A,k)$ itself is contractible.
\end{proof}
The first application is the path fibration. Suppose given a simply connected pointed type $B$ we have a map $\ensuremath{\mathbf{1}}\xspace\to B$ that has fiber $\Omega B$.\footnote{It is called the \emph{path fibration} because classically to get a Serre fibration we need to take the path space $PB$ instead of $\ensuremath{\mathbf{1}}\xspace$.} In other words, we have the fiber sequence
$$\Omega B \to \ensuremath{\mathbf{1}}\xspace \to B.$$
Now the Serre spectral sequence for cohomology gives (say, with integer coefficients)
$$E_2^{p,q}=H^p(B;H^q(\Omega B))\Rightarrow H^{p+q}(\ensuremath{\mathbf{1}}\xspace).$$
Note that the $\infty$-page vanishes, except when $p+q=0$, when the coefficient is $\ensuremath{\mathbb{Z}}\xspace$.
For ordinary cohomology $H^n$ is trivial for $n<0$, which means that the second page is only nontrivial in the first quadrant of the plane,
hence this is true for all pages, including the $\infty$-page.
Therefore, the $\infty$-page has one group $\ensuremath{\mathbb{Z}}\xspace$ at the origin, and trivial groups everywhere else, as shown in \autoref{fig:infty-path-fibration}.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}
\draw [thick] (0.5,0.5) -- (0.5,3.5);
\draw [thick] (0.5,0.5) -- (5.5,0.5);
\ensuremath{\operatorname{\mathsf{node}}}[font=\normalsize] at (5.4,0.3) {$p$};
\ensuremath{\operatorname{\mathsf{node}}}[font=\normalsize] at (0.3,3.4) {$q$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,2) {$1$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,3) {$2$};
\ensuremath{\operatorname{\mathsf{node}}} at (1,0.1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} at (2,0.1) {$1$};
\ensuremath{\operatorname{\mathsf{node}}} at (3,0.1) {$2$};
\ensuremath{\operatorname{\mathsf{node}}} at (4,0.1) {$3$};
\ensuremath{\operatorname{\mathsf{node}}} at (5,0.1) {$4$};
\ensuremath{\operatorname{\mathsf{node}}} (x11) at (1,1) {$\ensuremath{\mathbb{Z}}\xspace$};
\ensuremath{\operatorname{\mathsf{node}}} (x12) at (1,2) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x13) at (1,3) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x21) at (2,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x22) at (2,2) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x23) at (2,3) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x31) at (3,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x32) at (3,2) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x33) at (3,3) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x41) at (4,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x42) at (4,2) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x43) at (4,3) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x51) at (5,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x52) at (5,2) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} (x53) at (5,3) {$0$};
\end{tikzpicture}
\end{center}
\caption{$E_\infty^{p,q}$ for the path fibration.}
\label{fig:infty-path-fibration}
\end{figure}
This gives a relation between the cohomology of $B$ and the cohomology of $\Omega B$.
If we know the cohomology for one of the spaces one of them,
then we can sometimes compute the cohomology from the other using this.
Using the Serre spectral sequence for homology, we have the same relationship between the homology of $B$ and the homology of $\Omega B$.
The computations in the next example will work exactly the same for homology.
\begin{ex}\label{ex:cohomology-KZ2}
As an example, we can compute the cohomology groups of $B = K(\ensuremath{\mathbb{Z}}\xspace,2)$
(which is the complex projective space $\mathbf{CP}^\infty$).
Its loop space is $\Omega K(\ensuremath{\mathbb{Z}}\xspace,2) = K(\ensuremath{\mathbb{Z}}\xspace,1) = \S^1$, and by \autoref{lem:cohomology-spheres} we have
$$H^n(\S^1)=\begin{cases}\ensuremath{\mathbb{Z}}\xspace&\text{if $n=0,1$}\\ 0&\text{otherwise.}\end{cases}$$
The resulting second page of the spectral sequence is shown in \autoref{fig:two-page-KZ2-path-fibration}, all other groups on the second page are trivial.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[every node/.style={font=\small},x=15mm,y=15mm]
\draw [thick] (0.5,0.5) -- (0.5,2.5);
\draw [thick] (0.5,0.5) -- (5.5,0.5);
\ensuremath{\operatorname{\mathsf{node}}}[font=\normalsize] at (5.4,0.3) {$p$};
\ensuremath{\operatorname{\mathsf{node}}}[font=\normalsize] at (0.3,2.4) {$q$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,2) {$1$};
\ensuremath{\operatorname{\mathsf{node}}} at (1,0.1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} at (2,0.1) {$1$};
\ensuremath{\operatorname{\mathsf{node}}} at (3,0.1) {$2$};
\ensuremath{\operatorname{\mathsf{node}}} at (4,0.1) {$3$};
\ensuremath{\operatorname{\mathsf{node}}} at (5,0.1) {$4$};
\ensuremath{\operatorname{\mathsf{node}}} (x11) at (1,1) {$H^0\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x12) at (1,2) {$H^0\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x21) at (2,1) {$H^1\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x22) at (2,2) {$H^1\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x31) at (3,1) {$H^2\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x32) at (3,2) {$H^2\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x41) at (4,1) {$H^3\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x42) at (4,2) {$H^3\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x51) at (5,1) {$H^4\!(B)$};
\ensuremath{\operatorname{\mathsf{node}}} (x52) at (5,2) {$H^4\!(B)$};
\path
(x12) edge[->] (x31)
(x22) edge[->] (x41)
(x32) edge[->] (x51)
(x42) edge (5.5,1.25)
(x52) edge (5.5,1.75);
\end{tikzpicture}
\end{center}
\caption{$E_2^{p,q}$ for the path fibration of $K(\ensuremath{\mathbb{Z}}\xspace,2)$.}
\label{fig:two-page-KZ2-path-fibration}
\end{figure}
Note that the shown differentials are the only nontrivial differentials on the second page, and all differentials on all later pages are also trivial. This means that $E_3=E_\infty$, depicted in \autoref{fig:infty-path-fibration}. Note that there are no nontrivial differentials going in or out of the $H^0(B)$ and $H^1(B)$ in the bottom line. This means that
$$H^0(B)=E_\infty^{0,0}=\ensuremath{\mathbb{Z}}\xspace$$
and
$$H^1(B)=E_\infty^{1,0}.$$
All other groups displayed on the second page vanish on the $\infty$-page. Therefore, all shown differentials must be isomorphisms.
This means that $H^{n+2}(B)=H^n(B)$, which shows that $H^n(B)$ is $\ensuremath{\mathbb{Z}}\xspace$ for even $n$ and $0$ for odd $n$.
\end{ex}
Another simple application of the Serre spectral sequence is to compute the homology and cohomology groups of $\Omega \S^n$, given in~\cite[Example 1.5]{hatcher2004spectral}.
In this case, we know the (co)homology of the base space $\S^n$, and from it we can deduce the (co)homology of the loop space $\Omega \S^n$. We will do the computation here for cohomology.
\begin{ex}\label{ex:cohomology-OSn}
If we take the Serre spectral sequence for the path fibration of $B=S^n$ for $n\geq 2$, then the second page has entries
$$E_2^{p,q}=H^p(S^n;H^q(\Omega S^n))=\begin{cases}H^q(\Omega S^n)&\text{if $p=0,n$}\\ 0&\text{otherwise.}\end{cases}$$
using \autoref{lem:cohomology-spheres}. Therefore, the only nontrivial groups are in the columns $p=0$ and $p=n$.
This means that by looking at the degree of the differentials, the only nonzero differentials can occur in page $n$, as shown in \autoref{fig:n-page-path-fibration-sphere}.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[every node/.style={font=\small},x=25mm,y=12mm]
\draw [thick] (0.5,0.5) -- (2.5,0.5);
\draw [thick] (0.5,0.5) -- (0.5,4.5);
\ensuremath{\operatorname{\mathsf{node}}}[font=\normalsize] at (2.4,0.3) {$p$};
\ensuremath{\operatorname{\mathsf{node}}}[font=\normalsize] at (0.4,4.4) {$q$};
\ensuremath{\operatorname{\mathsf{node}}} at (1,0.1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} at (2,0.1) {$n$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,1) {$0$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,2) {$n-1$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,3) {$2(n-1)$};
\ensuremath{\operatorname{\mathsf{node}}} at (0.1,4) {$3(n-1)$};
\ensuremath{\operatorname{\mathsf{node}}} (x11) at (1,1) {$H_0(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x12) at (1,2) {$H_{n-1}(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x13) at (1,3) {$H_{2(n-1)}(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x14) at (1,4) {$H_{3(n-1)}(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x21) at (2,1) {$H_0(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x22) at (2,2) {$H_{n-1}(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x23) at (2,3) {$H_{2(n-1)}(\Omega S^n)$};
\ensuremath{\operatorname{\mathsf{node}}} (x24) at (2,4) {$H_{3(n-1)}(\Omega S^n)$};
\path
(x12) edge[->] (x21)
(x13) edge[->] (x22)
(x14) edge[->] (x23)
(1.5,4.5) edge[->] (x24);
\end{tikzpicture}
\end{center}
\caption{$E_n^{p,q}$ for the path fibration of $S^n$.}
\label{fig:n-page-path-fibration-sphere}
\end{figure}
Because all later differentials are trivial, $E_{n+1}=E_\infty$, which is depicted in \autoref{fig:infty-path-fibration}. This means that all differentials on page $n$ from the $p=0$ column to the $p=n$ column must be isomorphisms, except for the differential from $(0,0)$ to $(n,-(n-1))$. Hence we can conclude by induction that
\begin{equation}H^k(\Omega\S^n)\begin{cases}\ensuremath{\mathbb{Z}}\xspace&\text{if $n-1\mid k$}\\0&\text{otherwise.}\end{cases}\label{eq:cohomology-loop-spheres}\end{equation}
\end{ex}
As a generalization of \autoref{ex:cohomology-KZ2}, we can construct the \emph{Gysin sequence} from the Serre spectral sequence~\cite[Theorem 3.3.3]{holmbergperoux2014models}.
The Gysin sequence for homology states that if $f:E\to B$ is a pointed map with fiber $\S^{n-1}$ for $n\geq 2$ and if $B$ is simply connected, then there exists a long exact sequence
$$\cdots \to H_i(E)\to H_i(B)\to H_{i-n}(B)\to H_{i-1}(E) \to \cdots.$$
There is also an analogue for cohomology, which states that under the same assumptions there exists a long exact sequence of cohomology groups
$$\cdots \to H^{i-1}(E)\to H^{i-n}(B)\to H^i(B)\to H^i(E) \to \cdots.$$
The proof given in~\cite[Theorem 3.3.3]{holmbergperoux2014models} works the same in HoTT.
An alternative construction of the Gysin sequence in HoTT is given in~\cite[Section 6.1]{brunerie2016spheres}, which was used as a main ingredient to compute $\pi_4(\S^3)$.
We can also generalize \autoref{ex:cohomology-OSn} to get the \emph{Wang sequence}.
For homology this states that if $E\to \S^n$ is a pointed map for $n\geq 2$ with fiber $F$, then there exists a long exact sequence
$$\cdots \to H_i(F)\to H_i(E)\to H_{i-n}(F)\to H_{i-1}(F) \to \cdots.$$
Again, a similar long exact sequence holds for cohomology, and the proof given in~\cite[Theorem 3.3.6]{holmbergperoux2014models} works the same in HoTT.
As another application, we can prove the Hurewicz theorem from the Serre spectral sequence~\cite{holmbergperoux2014models}.
The Hurewicz theorem only holds for homology, and requires the Serre spectral sequence for homology.
The theorem states that if $X$ is a simply connected pointed type, $n\geq 2$ and $\pi_q(X)$ is trivial for $q<n$, then $H_q(X)=0$ for $q<n$ and $H_n(X)=\pi_n(X)$.
For $n=1$ the Hurewicz theorem states that for a 0-connected pointed type $X$, the first homology group $H_1(X)$ is the abelianization of $\pi_1(X)$.
In the proof given in the aforementioned reference, the case for $n=1$ needs to be proven separately,
but then the case for $n\geq2$ follows from that using the Serre spectral sequence.
Since the case for $n=1$ seems easier than the general case, this should be very helpful to prove the Hurewicz theorem in HoTT.
An application for the Atiyah-Hirzebruch spectral sequence would be to compute cohomology groups of generalized cohomology theories.
One such generalized cohomology theory is K-theory. Although K-theory has not been precisely defined yet in HoTT,
one possible idea by Ulrik Buchholtz is to define it using Snaith's theorem~\cite{snaith1981ktheory}.
If we have defined K-theory, we could try to compute its cohomology groups using the Atiyah-Hirzebruch spectral sequence.
With the current machinery, we can compute the cohomology groups of all types that satisfy weak pointed choice (cf. \autoref{def:weak-pointed-choice}),
which probably includes all finite CW-complexes.
An application that is probably trickier in HoTT is the Serre class theorem. A \emph{Serre class} is a class $\mathcal C$ of abelian groups such that for every short exact sequence
$0\to A \to B \to C \to 0$ of abelian groups we have $B\in \mathcal C$ iff $A,C\in \mathcal C$. In particular, any Serre class is closed under taking subgroups and quotient groups. Classical examples of Serre classes include
\begin{itemize}
\item finite abelian groups;
\item finitely generated abelian groups;
\item torsion abelian groups.
\end{itemize}
However, constructively, the first two classes are not closed under either taking subgroups or quotient groups (torsion abelian groups do form a Serre class constructively).
The Serre class theorem is a theorem about certain Serre classes that satisfy some extra properties.
This include the three examples mentioned above.
If $\mathcal C$ is such a Serre class and if $X$ is a simply connected
type,\footnote{or path-connected and \emph{abelian}. A space $X$ is abelian if the action of $\pi_1(X)$ on $\pi_n(X)$ is trivial for all $n\geq 1$.}
then the theorem states that $\pi_n(X)\in\mathcal C$ for all $n$ iff $H_n(X)\in\mathcal C$ for all $n$.
More general is the Hurewicz theorem modulo a Serre class, which states that if $\pi_i(X)\in\mathcal C$ for all $i<n$,
then the kernel and the cokernel of the Hurewicz homomorphism $h:\pi_n(X)\to H_n(X)$ belong to $\mathcal C$.
As a corollary of the Serre class theorem, we know that the homotopy groups of the spheres are finitely generated, since their homology groups are finitely generated.
Moreover, the homotopy groups of simply connected finite CW-complexes are also finitely generated, using cellular cohomology~\cite{buchholtz2018cellular}.
A classical proof of the Serre class theorem can be found in~\cite[Section 1.1]{hatcher2004spectral}.
It is not straightforward to adapt the proof of the Serre class theorem to a proof in HoTT. One difficulty is that the classical proof uses the \emph{universal coefficient theorem} for homology.
This theorem is not yet proven in HoTT.
The universal coefficient theorem relates the homology group $H_n(X;A)$ with coefficients in any abelian group $A$ to the the homology group $H_n(X)$ with integer coefficients.
There is also a dual universal coefficient theorem for cohomology that relates the cohomology group $H^n(X;A)$ with the homology group $H_n(X)$.
It is not clear how to prove or even formulate the universal coefficient theorem in HoTT.
The universal coefficient theorem for homology uses the $\mathsf{Tor}$ functor, whose definition requires projective resolutions.
Similarly, the universal coefficient theorem for cohomology uses the $\mathsf{Ext}$ functor, whose definition requires injective resolutions.
Basic properties of projective and injective resolutions are classically proven with the axiom of choice~\cite{blass1979injectivity},
so it is not clear whether we can prove the universal coefficient theorem without the axiom of choice.
Another problem with proving the universal coefficient theorem is that classically it is proven algebraically for chain complexes.
Since homology and cohomology groups of spaces are defined as the (co)homology of a chain complex, the universal coefficient theorem then applies to spaces.
Even if we could solve the issues with the axiom of choice, and we could prove the universal coefficient theorem for chain complexes in HoTT,
it does not directly follow that it is true for spaces, since these groups are \emph{not} defined as the (co)homology of chain
complexes.\footnote{Cellular (co)homology~\cite{buchholtz2018cellular} is defined as the (co)homology of a chain complex,
and therefore we could prove it for finite CW complexes, but it would not follow for arbitrary types.}
Therefore, in HoTT, it seems fruitful to prove the universal coefficient theorem directly for spaces,
using the definition in terms of Eilenberg-MacLane spaces, but this is an open problem as of now.
A good first step might be to try to prove a special case of the universal coefficient theorem where the $\mathsf{Tor}$ and $\mathsf{Ext}$ functors vanish,
although that will not be sufficient to prove the Serre class theorem.
Proving the Serre class theorem in HoTT will be tricky, and it might be necessary to reformulate or weaken some notions to get a usable result in HoTT.
If we manage to prove these results in HoTT, we can get a lot of information about the homotopy groups of spheres.
One additional ingredient that is required is the fact that the cup product structure of the cohomology groups respect the Serre spectral sequence.
From these ingredients we can classically show the following:
\begin{itemize}
\item The groups $\pi_i(\S^n)$ are finite for $i>n$, except for $\pi_{4k-1}(\S^{2k})$, which are the direct sum of $\ensuremath{\mathbb{Z}}\xspace$ and a finite group~\cite[Theorem 1.21]{hatcher2004spectral}.
\item For a prime $p$ the $p$-torsion subgroup of $\pi_i(\S^3)$ is 0 for $i<2p$ and $\ensuremath{\mathbb{Z}}\xspace_p$ for $i=2p$~\cite[Example 1.18]{hatcher2004spectral}.
\item From the two above results we can immediately conclude that $\pi_4(\S^3)=\ensuremath{\mathbb{Z}}\xspace_2$.
\item Using additionally the localization of a space at a prime, we can show that for $p$ a prime, the $p$-torsion subgroup of $\pi_i(\S^{n+3})$ is 0 for $i<n+2p$ and $\ensuremath{\mathbb{Z}}\xspace_p$ for $i=n+2p$~\cite[Theorem 1.28]{hatcher2004spectral}.
\item We can compute more homotopy groups of spheres using significantly more machinery.
For this we need the EHP sequence, Steenrod squares and Serre's theorem, which computes the cohomology rings of $K(\ensuremath{\mathbb{Z}}\xspace_2,n)$ and $K(\ensuremath{\mathbb{Z}}\xspace,n)$ and $K(\ensuremath{\mathbb{Z}}\xspace_{2^k},n)$ with coefficients in $\ensuremath{\mathbb{Z}}\xspace_2$.
If we have all these results, we can compute $\pi_{n+i}(\S^n)$ for all $n$ and $i\leq 3$~\cite[Theorem 1.40]{hatcher2004spectral}.
\end{itemize}
\chapter*{Conclusion}
\addcontentsline{toc}{chapter}{Conclusion}
\lstMakeShortInline"
In this dissertation I have shown that homotopy type theory is a practical language to prove involved theorems in homotopy theory,
most notably the construction of two important spectral sequences: the Atiyah-Hirzebruch and the Serre spectral sequences for cohomology.
The discovery of these spectral sequences in classical homotopy theory was an important milestone,
and we expect that the corresponding proof in HoTT will lead to many useful corollaries in synthetic homotopy theory.
That said, many applications of these spectral sequences require more machinery, such as the universal coefficient theorem, Serre classes and the Hurewicz theorem.
The first two of these three results might be problematic to prove in HoTT, because of their dependence on the axiom of choice.
I hope that an adapted or weaker version of these theorems can be found, which avoids the use of the axiom of choice, or alternatively,
that when looking at their applications, we can avoid the use of choice. For example, we cannot prove constructively that finitely generated abelian groups form a Serre class,
but it is conceivable that we can still prove that all homotopy groups of spheres are finitely generated without resorting to the axiom of choice.
That said, we could also assume the axiom of choice and continue proving results in synthetic homotopy theory using it.
However, then the resulting theorem would not hold anymore in all models of HoTT.
It would be interesting to see other spectral sequences proven in HoTT, such as the Adams spectral sequence and the Eilenberg-Moore spectral sequences~\cite{hatcher2004spectral}.
\subsection*{Thoughts on formalization}
This dissertation also shows that homotopy type theory provides a good language for the computer formalization of results in homotopy theory.
Through the formal methods community there is a strong desire that formal methods will be adopted in a large scale by general mathematicians.
The main bottlenecks for this adoption are
\begin{enumerate}
\item the necessary expertise of formalization in the proof assistant of choice;
\item the vast number of proof assistant in existence;
\item the amount of work it takes to formalize mathematics compared to writing it on paper.
\end{enumerate}
It definitely takes time to learn a proof assistant, familiarize oneself with the library and get enough practice to use a proof assistant efficiently.
Moreover, in my experience, learning to use a proof assistant takes longer than learning to use other programs, like LaTeX or Mathematica.
Still, I do not think this is the main bottleneck to the adoption of proof assistants.
Various courses that integrate the use of proof assistants have been taught, and students taking those courses will get a level of proficiency of using that proof assistant.
The second concern is the number of proof assistants in existence, each with a separate library and the near-impossibility to translate theorems and proofs between two proof assistants.
There are translation procedures between some proof assistants, such as~\cite{mclaughlin2006interpretation},
but such translations are often incomplete, and only specific to two proof assistants.
However, I think the main bottleneck is the amount of extra time it takes to formalize mathematics compared to writing a paper proof.
Rough estimates for the formalization time is about one week to formalize a page of a mathematical paper or textbook~\cite{asperti2010itp}.
My experience with formalizing synthetic homotopy theory specifically is a little different.
Paper proofs given in synthetic homotopy theory are often quite detailed, and the techniques used are often very close to the underlying type theory.
I would argue that this is necessary; we do not have much experience with proving theorems in synthetic homotopy theory yet, and it is not always clear which results are hard to prove.
Some results turn out more difficult to prove than initially thought.
For example, the ``basic property'' of the smash product that it forms a 1-coherent symmetric monoidal product (see \autoref{sec:smash-product})
was assumed with a vague proof sketch in~\cite{brunerie2016spheres} to prove $\pi_4(\S^3)=\ensuremath{\mathbb{Z}}\xspace_2$, but this result is still open as of now.
Another example is \autoref{thm:colim_sm}, which was originally thought to be a basic result about colimits by Egbert Rijke and me, but the proof turned out to be much harder than expected.
Because paper proofs in synthetic homotopy theory are often proven with many details, in my experience, giving a fully formal proof is not much more work.
In cases where the formal proof is a lot more work, the paper proof sometimes omitted showing the case of the path constructor when inducting over a HIT,
which is the hardest --- but least enlightening --- part of the proof.
In the paper proofs of this dissertation, I have also sometimes omitted these steps, because they are tedious to work through and not very enlightening.
However, the formal proofs (of course) contain all the details.
For some theorems the formalization did take substantially more work.
For the formalization of spectral sequences, a substantial algebra library had to be developed,
consisting of basic group theory, ring theory, modules over a ring and graded modules.
This took many man-hours of work, which would have no counterpart in a paper proof.
Another reason why formalization is more work, is that necessarily such proofs have to be encoded in the corresponding logic, intensional type theory.
Most of the time, this is straightforward, but in some cases it takes a bit more work.
Especially when dealing with sequences of types, in intensional type theory one has to work explicitly with transports (or its relatives, like pathovers or heterogenous equality),
which is especially laborious in the proof-relevant setting of HoTT.
Sometimes an ``encoding trick'' is useful when dealing with these dependent types. In this dissertation some of these tricks have been given.
In \autoref{sec:les-homotopy} we defined a chain complex over an arbitrary successor structure,
because we wanted to not only index chain complexes over $\textbf{N}$ or $\ensuremath{\mathbb{Z}}\xspace$, but also over $\textbf{N}\times\ensuremath{\operatorname{\mathsf{fin}}}_3$ or similar types, to get a more convenient computational content.
These successor structures also turned out to be useful for spectra in \autoref{sec:spectra}, so that we can apply the same notion to spectra indexed over $\textbf{N}$ and spectra indexed over $\ensuremath{\mathbb{Z}}\xspace$.
The reason that spectra indexed over $\ensuremath{\mathbb{Z}}\xspace$ are useful (traditionally they are only indexed over $\textbf{N}$) is that for certain definitions,
such as the homotopy group of a spectrum, no case-splits are required when they are indexed over $\ensuremath{\mathbb{Z}}\xspace$.
Another encoding trick was given in the definition of graded morphisms. In order to define the composition of graded morphisms more easily,
we defined the degree of a graded morphism to be an automorphism of the indexing set.
In order to avoid dealing with transports everywhere, we defined a graded morphism to act on a path in the indexing set, see \autoref{sec:spectral-sequences}.
Formalizing in Lean is a fun activity, and Lean is a good language for formalization.
In Lean 2 one of the main annoyances when formalizing was the unpredictability of the elaborator, which was greatly improved in Lean 3.
Another issue was the ability to simplify expressions. There was a tactic "esimp" that simplified by evaluation,
but it was quite slow, and would sometimes use $99\%$ of the elaboration time of a proof.
I do not have enough experience with "dsimp" in Lean 3 to see whether it has similar issues.
In 2016 Leonardo de Moura decided that he would stop supporting homotopy type theory in Lean~3. It was quite devastating to hear this.
I am glad that Gabriel Ebner has found a method to do homotopy type theory in Lean~3 safely, by avoiding the use of "Prop".
Since then, I have been slowly working on porting the Lean 2 HoTT library to Lean~3, although the progress has been slow.
The main reasons for this are:
\begin{itemize}
\item The elaborator in Lean~3 is weaker to make it more robust, which causes many proofs to break.
\item Some tactics do not work without using "Prop" or the "Prop"-valued equality.
Gabriel Ebner has modified the "simp" and "rewrite" tactic to work in HoTT.
I have written an "induction"-tactic, since the default induction tactic does not allow custom induction principles to eliminate to only non-"Prop" sorts.
\item The notation "!" has been removed in Lean~3. This was used in Lean~2 to turn (some) explicit arguments into implicit ones.
\item There are many small differences in Lean~2 and Lean~3 in syntax for tactics, proof styles, attributes, universe levels and declarations.
None of these issues take much time to fix, but the sheer number of them add up.
\end{itemize}
Despite this, a significant part of the library has been ported, and I am planning to continue this so that Lean~3 (and later Lean~4) can be used to formalize results in homotopy type theory.
The current implementation in Lean is probably not the ultimate proof assistant for HoTT in the long-term.
Many cubical type theories have been developed over the last few years,
and a few proof assistants have been developed using a cubical type theory as their underlying logic.
Cubical type theory offers many advantages when reasoning about higher inductive types and when doing higher path algebra, since more relations hold strictly.
For example, the computation rule of the induction principle for a higher inductive type holds judgmentally in cubical type theory.
This is very convenient when working with HITs, especially HITs with higher path constructors.
It is conceivable that a cubical type theory can be implemented in Lean, although it will require some hacking in the "C++" code,
and many features of Lean will need to be modified to work well with the cubical structure.
This will be a big project, and it is probably smart not do this project until the different variants of cubical type theory have been studied more.
In particular, current versions of cubical type theory do not satisfy regularity, which states that the induction principle for paths has judgmental computation rules.\footnote{It
is possible to have two notions of paths: a path type with all the cubical structure, and an identity type with an induction principle and a judgmental computation rule.
However, in current cubical type theories these cannot be the same type.}
Some constructions in HoTT are done by doing a long string of path inductions, and such proofs will be harder to reason with in cubical type theory. That said, it would be interesting to perform some constructions of this dissertation in one of the cubical type theories to see whether the proof would significantly simplify. In particular the proofs in \autoref{sec:non-recursive-2} would simplify when the induction principle of higher inductive types reduces definitionally when applied to path constructors.
\chapter*{Acknowledgements}
\addcontentsline{toc}{chapter}{Acknowledgements}
First and foremost I would like to thank my advisor Jeremy Avigad, who was always ready to give useful feedback, proofread drafts of all my written work and provide support.
Futhermore, I would like to thank Steve Awodey for always being ready to answer any questions I have about HoTT or category theory.
I would like to thank Mike Shulman for many helpful remarks and insights whenever I show my work.
I also want to thank Ulrik Buchholtz, Egbert Rijke, Jakob von Raumer, Stefano Piceghello and Kristina Sojakova for the collaborations and discussions.
I am grateful towards Leonardo de Moura for all his help with getting me up to speed with Lean, and answering all my stupid questions and ideas I brought up early in the development of Lean.
I would like to thank Marc Bezem and Dan Christensen to invite me for academic visits.
More generally, I would like to thank everyone in the HoTT community for maintaining such a good research community.
It is very nice to be part of such a friendly and collaborative research community,
where it is normal to have unfinished projects on Github or discuss half-baked ideas on a mailing list.
For moral support, I would like to thank my parents, Peter van Doorn and Judith van Wakeren, for supporting me during times when I was struggling. Dank jullie wel!
Lastly I would like to thank Cecilia Hornberger for the moral support over the last months.
I gratefully acknowledge the support of the Air Force Office of Scientific Research through MURI
grant FA9550-15-1-0053. Any opinions, findings and conclusions or recommendations expressed in this
material are those of the authors and do not necessarily reflect the views of the AFOSR.
The author would like to thank the Isaac Newton Institute for Mathematical Sciences, Cambridge,
for support and hospitality during the programme Big Proof where work on this paper
was undertaken. This work was supported by EPSRC grant no EP/K032208/1.
This material is based upon work supported by the National Science Foundation under Grant Number DMS 1641020.
\clearpage
\phantomsection
\addcontentsline{toc}{chapter}{Bibliography}
\Urlmuskip=0mu plus 1mu
\bibliographystyle{amsalphaurl}
\section*{Notes}\addcontentsline{toc}{section}{Notes}\markright{\textsc{\@chapapp{} \thechapter{} Notes}}}
\newcommand{\sectionExercises}[1]{\phantomsection\section*{Exercises}\addcontentsline{toc}{section}{Exercises}\markright{\textsc{\@chapapp{} \thechapter{} Exercises}}}
\makeatother
\newcommand{\equiv}{\equiv}
\let\judgeq\equiv
\newcommand{\ensuremath{\vcentcolon\equiv}}{\vcentcolon\equiv}
\newcommand{\define}[1]{\textbf{#1}}
\newcommand{\ensuremath{\mathsf{Vec}}}{\ensuremath{\mathsf{Vec}}}
\newcommand{\ensuremath{\mathsf{Fin}}}{\ensuremath{\mathsf{Fin}}}
\newcommand{\ensuremath{\mathsf{fmax}}}{\ensuremath{\mathsf{fmax}}}
\newcommand{\seq}[1]{\langle #1\rangle}
\def\textstyle\prod{\textstyle\prod}
\makeatletter
\def\prd#1{\@ifnextchar\bgroup{\prd@parens{#1}}{\@ifnextchar\sm{\prd@parens{#1}\@eatsm}{\prd@noparens{#1}}}}
\def\prd@parens#1{\@ifnextchar\bgroup%
{\mathchoice{\@dprd{#1}}{\@tprd{#1}}{\@tprd{#1}}{\@tprd{#1}}\prd@parens}%
{\@ifnextchar\sm%
{\mathchoice{\@dprd{#1}}{\@tprd{#1}}{\@tprd{#1}}{\@tprd{#1}}\@eatsm}%
{\mathchoice{\@dprd{#1}}{\@tprd{#1}}{\@tprd{#1}}{\@tprd{#1}}}}}
\def\@eatsm\sm{\sm@parens}
\def\prd@noparens#1{\mathchoice{\@dprd@noparens{#1}}{\@tprd{#1}}{\@tprd{#1}}{\@tprd{#1}}}
\def\lprd#1{\@ifnextchar\bgroup{\@lprd{#1}\lprd}{\@@lprd{#1}}}
\def\@lprd#1{\mathchoice{{\textstyle\prod}}{\prod}{\prod}{\prod}({\textstyle #1})\;}
\def\@@lprd#1{\mathchoice{{\textstyle\prod}}{\prod}{\prod}{\prod}({\textstyle #1}),\ }
\def\tprd#1{\@tprd{#1}\@ifnextchar\bgroup{\tprd}{}}
\def\@tprd#1{\mathchoice{{\textstyle\prod_{(#1)}}}{\prod_{(#1)}}{\prod_{(#1)}}{\prod_{(#1)}}}
\def\dprd#1{\@dprd{#1}\@ifnextchar\bgroup{\dprd}{}}
\def\@dprd#1{\prod_{(#1)}\,}
\def\@dprd@noparens#1{\prod_{#1}\,}
\def\lam#1{{\lambda}\@lamarg#1:\@endlamarg\@ifnextchar\bgroup{.\,\lam}{.\,}}
\def\@lamarg#1:#2\@endlamarg{\if\relax\detokenize{#2}\relax #1\else\@lamvar{\@lameatcolon#2},#1\@endlamvar\fi}
\def\@lamvar#1,#2\@endlamvar{(#2\,{:}\,#1)}
\def\@lameatcolon#1:{#1}
\let\lamt\lam
\def\lamu#1{{\lambda}\@lamuarg#1:\@endlamuarg\@ifnextchar\bgroup{.\,\lamu}{.\,}}
\def\@lamuarg#1:#2\@endlamuarg{#1}
\def\fall#1{\forall (#1)\@ifnextchar\bgroup{.\,\fall}{.\,}}
\def\exis#1{\exists (#1)\@ifnextchar\bgroup{.\,\exis}{.\,}}
\def\textstyle\sum{\textstyle\sum}
\def\sm#1{\@ifnextchar\bgroup{\sm@parens{#1}}{\@ifnextchar\prd{\sm@parens{#1}\@eatprd}{\sm@noparens{#1}}}}
\def\sm@parens#1{\@ifnextchar\bgroup%
{\mathchoice{\@dsm{#1}}{\@tsm{#1}}{\@tsm{#1}}{\@tsm{#1}}\sm@parens}%
{\@ifnextchar\prd%
{\mathchoice{\@dsm{#1}}{\@tsm{#1}}{\@tsm{#1}}{\@tsm{#1}}\@eatprd}%
{\mathchoice{\@dsm{#1}}{\@tsm{#1}}{\@tsm{#1}}{\@tsm{#1}}}}}
\def\@eatprd\prd{\prd@parens}
\def\sm@noparens#1{\mathchoice{\@dsm@noparens{#1}}{\@tsm{#1}}{\@tsm{#1}}{\@tsm{#1}}}
\def\lsm#1{\@ifnextchar\bgroup{\@lsm{#1}\lsm}{\@@lsm{#1}}}
\def\@lsm#1{\mathchoice{{\textstyle\sum}}{\sum}{\sum}{\sum}({\textstyle #1})\;}
\def\@@lsm#1{\mathchoice{{\textstyle\sum}}{\sum}{\sum}{\sum}({\textstyle #1}),\ }
\def\tsm#1{\@tsm{#1}\@ifnextchar\bgroup{\tsm}{}}
\def\@tsm#1{\mathchoice{{\textstyle\sum_{(#1)}}}{\sum_{(#1)}}{\sum_{(#1)}}{\sum_{(#1)}}}
\def\dsm#1{\@dsm{#1}\@ifnextchar\bgroup{\dsm}{}}
\def\@dsm#1{\sum_{(#1)}\,}
\def\@dsm@noparens#1{\sum_{#1}\,}
\def{\mathsf{W}}{{\mathsf{W}}}
\def\wtype#1{\@ifnextchar\bgroup%
{\mathchoice{\@twtype{#1}}{\@twtype{#1}}{\@twtype{#1}}{\@twtype{#1}}\wtype}%
{\mathchoice{\@twtype{#1}}{\@twtype{#1}}{\@twtype{#1}}{\@twtype{#1}}}}
\def\lwtype#1{\@ifnextchar\bgroup{\@lwtype{#1}\lwtype}{\@@lwtype{#1}}}
\def\@lwtype#1{\mathchoice{{\textstyle\mathsf{W}}}{\mathsf{W}}{\mathsf{W}}{\mathsf{W}}({\textstyle #1})\;}
\def\@@lwtype#1{\mathchoice{{\textstyle\mathsf{W}}}{\mathsf{W}}{\mathsf{W}}{\mathsf{W}}({\textstyle #1}),\ }
\def\twtype#1{\@twtype{#1}\@ifnextchar\bgroup{\twtype}{}}
\def\@twtype#1{\mathchoice{{\textstyle\mathsf{W}_{(#1)}}}{\mathsf{W}_{(#1)}}{\mathsf{W}_{(#1)}}{\mathsf{W}_{(#1)}}}
\def\dwtype#1{\@dwtype{#1}\@ifnextchar\bgroup{\dwtype}{}}
\def\@dwtype#1{\mathsf{W}_{(#1)}\,}
\newcommand{{\mathsf{sup}}}{{\mathsf{sup}}}
\newcommand{\ensuremath\suppsym\xspace}{\ensuremath{\mathsf{sup}}\xspace}
\def\wtypeh#1{\@ifnextchar\bgroup%
{\mathchoice{\@lwtypeh{#1}}{\@twtypeh{#1}}{\@twtypeh{#1}}{\@twtypeh{#1}}\wtypeh}%
{\mathchoice{\@@lwtypeh{#1}}{\@twtypeh{#1}}{\@twtypeh{#1}}{\@twtypeh{#1}}}}
\def\lwtypeh#1{\@ifnextchar\bgroup{\@lwtypeh{#1}\lwtypeh}{\@@lwtypeh{#1}}}
\def\@lwtypeh#1{\mathchoice{{\textstyle\mathsf{W}^h}}{\mathsf{W}^h}{\mathsf{W}^h}{\mathsf{W}^h}({\textstyle #1})\;}
\def\@@lwtypeh#1{\mathchoice{{\textstyle\mathsf{W}^h}}{\mathsf{W}^h}{\mathsf{W}^h}{\mathsf{W}^h}({\textstyle #1}),\ }
\def\twtypeh#1{\@twtypeh{#1}\@ifnextchar\bgroup{\twtypeh}{}}
\def\@twtypeh#1{\mathchoice{{\textstyle\mathsf{W}^h_{(#1)}}}{\mathsf{W}^h_{(#1)}}{\mathsf{W}^h_{(#1)}}{\mathsf{W}^h_{(#1)}}}
\def\dwtypeh#1{\@dwtypeh{#1}\@ifnextchar\bgroup{\dwtypeh}{}}
\def\@dwtypeh#1{\mathsf{W}^h_{(#1)}\,}
\makeatother
\let\setof\Set
\newcommand{\tup}[2]{(#1,#2)}
\newcommand{\proj}[1]{\ensuremath{\mathsf{pr}_{#1}}\xspace}
\newcommand{\ensuremath{\proj1}\xspace}{\ensuremath{\proj1}\xspace}
\newcommand{\ensuremath{\proj2}\xspace}{\ensuremath{\proj2}\xspace}
\newcommand{\ac}{\ensuremath{\mathsf{ac}}\xspace}
\newcommand{\un}{\ensuremath{\mathsf{upun}}\xspace}
\newcommand{\rec}[1]{\mathsf{rec}_{#1}}
\newcommand{\ind}[1]{\mathsf{ind}_{#1}}
\newcommand{\indid}[1]{\ind{=_{#1}}}
\newcommand{\indidb}[1]{\ind{=_{#1}}'}
\newcommand{\ensuremath{\mathsf{uppt}}\xspace}{\ensuremath{\mathsf{uppt}}\xspace}
\newcommand{\ensuremath{\mathsf{pair}^{\mathord{=}}}\xspace}{\ensuremath{\mathsf{pair}^{\mathord{=}}}\xspace}
\newcommand{\projpath}[1]{\ensuremath{\apfunc{\proj{#1}}}\xspace}
\newcommand{\pairr}[1]{{\mathopen{}(#1)\mathclose{}}}
\newcommand{\Pairr}[1]{{\mathopen{}\left(#1\right)\mathclose{}}}
\newcommand{\im}{\ensuremath{\mathsf{im}}}
\newcommand{\mathbin{{\ct}_{\ell}}}{\mathbin{{\ct}_{\ell}}}
\newcommand{\mathbin{{\ct}_{r}}}{\mathbin{{\ct}_{r}}}
\newcommand{\star}{\star}
\newcommand{\ensuremath{\ocircle}}{\ensuremath{\ocircle}}
\let\reflect\ensuremath{\ocircle}
\newcommand{\ensuremath{\type_\modal}}{\ensuremath{\ensuremath{\mathsf{Type}}\xspace_\ensuremath{\ocircle}}}
\newcommand{\ensuremath{\eta}}{\ensuremath{\eta}}
\let\project\ensuremath{\eta}
\newcommand{\mathsf{ext}}{\mathsf{ext}}
\renewcommand{\P}{\ensuremath{\ensuremath{\mathsf{Type}}\xspace_{P}}\xspace}
\newcommand{{=}}{{=}}
\newcommand{\idtype}[3][]{\ensuremath{\mathsf{Id}_{#1}(#2,#3)}\xspace}
\newcommand{\idtypevar}[1]{\ensuremath{\mathsf{Id}_{#1}}\xspace}
\newcommand{\coloneqq}{\coloneqq}
\newcommand{\dpath}[4]{#3 =^{#1}_{#2} #4}
\newcommand{\ct}{%
\mathchoice{\mathbin{\raisebox{0.5ex}{$\displaystyle\centerdot$}}}%
{\mathbin{\raisebox{0.5ex}{$\centerdot$}}}%
{\mathbin{\raisebox{0.25ex}{$\scriptstyle\,\centerdot\,$}}}%
{\mathbin{\raisebox{0.1ex}{$\scriptscriptstyle\,\centerdot\,$}}}
}
\newcommand{\opp}[1]{\mathord{{#1}^{-1}}}
\let\rev\opp
\newcommand{\trans}[2]{\ensuremath{{#1}_{*}\mathopen{}\left({#2}\right)\mathclose{}}\xspace}
\let\Trans\trans
\newcommand{\transf}[1]{\ensuremath{{#1}_{*}}\xspace}
\newcommand{\transfib}[3]{\ensuremath{\mathsf{transport}^{#1}(#2,#3)\xspace}}
\newcommand{\Transfib}[3]{\ensuremath{\mathsf{transport}^{#1}\Big(#2,\, #3\Big)\xspace}}
\newcommand{\transfibf}[1]{\ensuremath{\mathsf{transport}^{#1}\xspace}}
\newcommand{\transtwo}[2]{\ensuremath{\mathsf{transport}^2\mathopen{}\left({#1},{#2}\right)\mathclose{}}\xspace}
\newcommand{\transconst}[3]{\ensuremath{\mathsf{transportconst}}^{#1}_{#2}(#3)\xspace}
\newcommand{\ensuremath{\mathsf{transportconst}}\xspace}{\ensuremath{\mathsf{transportconst}}\xspace}
\newcommand{\mapfunc}[1]{\ensuremath{\mathsf{ap}_{#1}}\xspace}
\newcommand{\map}[2]{\ensuremath{{#1}\mathopen{}\left({#2}\right)\mathclose{}}\xspace}
\let\Ap\map
\newcommand{\mapdepfunc}[1]{\ensuremath{\mathsf{apd}_{#1}}\xspace}
\newcommand{\mapdep}[2]{\ensuremath{\mapdepfunc{#1}\mathopen{}\left(#2\right)\mathclose{}}\xspace}
\let\apfunc\mapfunc
\let\ap\map
\let\apdfunc\mapdepfunc
\let\apd\mapdep
\newcommand{\aptwofunc}[1]{\ensuremath{\mathsf{ap}^2_{#1}}\xspace}
\newcommand{\aptwo}[2]{\ensuremath{\aptwofunc{#1}\mathopen{}\left({#2}\right)\mathclose{}}\xspace}
\newcommand{\apdtwofunc}[1]{\ensuremath{\mathsf{apd}^2_{#1}}\xspace}
\newcommand{\apdtwo}[2]{\ensuremath{\apdtwofunc{#1}\mathopen{}\left(#2\right)\mathclose{}}\xspace}
\newcommand{\idfunc}[1][]{\ensuremath{\mathsf{id}_{#1}}\xspace}
\newcommand{\sim}{\sim}
\newcommand{\bisim}{\sim}
\newcommand{\eqr}{\sim}
\newcommand{\eqv}[2]{\ensuremath{#1 \simeq #2}\xspace}
\newcommand{\eqvspaced}[2]{\ensuremath{#1 \;\simeq\; #2}\xspace}
\newcommand{\eqvsym}{\simeq}
\newcommand{\texteqv}[2]{\ensuremath{\mathsf{Equiv}(#1,#2)}\xspace}
\newcommand{\ensuremath{\mathsf{isequiv}}}{\ensuremath{\mathsf{isequiv}}}
\newcommand{\ensuremath{\mathsf{qinv}}}{\ensuremath{\mathsf{qinv}}}
\newcommand{\ensuremath{\mathsf{ishae}}}{\ensuremath{\mathsf{ishae}}}
\newcommand{\ensuremath{\mathsf{linv}}}{\ensuremath{\mathsf{linv}}}
\newcommand{\ensuremath{\mathsf{rinv}}}{\ensuremath{\mathsf{rinv}}}
\newcommand{\ensuremath{\mathsf{biinv}}}{\ensuremath{\mathsf{biinv}}}
\newcommand{\lcoh}[3]{\mathsf{lcoh}_{#1}(#2,#3)}
\newcommand{\rcoh}[3]{\mathsf{rcoh}_{#1}(#2,#3)}
\newcommand{\hfib}[2]{{\mathsf{fib}}_{#1}(#2)}
\newcommand{\total}[1]{\ensuremath{\mathsf{total}(#1)}}
\newcommand{\ensuremath{\mathcal{U}}\xspace}{\ensuremath{\mathcal{U}}\xspace}
\let\ensuremath{\mathbb{U}}\xspace\ensuremath{\mathcal{U}}\xspace
\let\ensuremath{\mathsf{Type}}\xspace\ensuremath{\mathcal{U}}\xspace
\newcommand{\typele}[1]{\ensuremath{{#1}\text-\mathsf{Type}}\xspace}
\newcommand{\typeleU}[1]{\ensuremath{{#1}\text-\mathsf{Type}_\ensuremath{\mathcal{U}}\xspace}\xspace}
\newcommand{\typelep}[1]{\ensuremath{{(#1)}\text-\mathsf{Type}}\xspace}
\newcommand{\typelepU}[1]{\ensuremath{{(#1)}\text-\mathsf{Type}_\ensuremath{\mathcal{U}}\xspace}\xspace}
\let\ntype\typele
\let\ntypeU\typeleU
\let\ntypep\typelep
\let\ntypepU\typelepU
\newcommand{\ensuremath{\mathsf{Set}}\xspace}{\ensuremath{\mathsf{Set}}\xspace}
\newcommand{\ensuremath{\mathsf{Set}_\UU}\xspace}{\ensuremath{\mathsf{Set}_\ensuremath{\mathcal{U}}\xspace}\xspace}
\newcommand{\ensuremath{\mathsf{Prop}}\xspace}{\ensuremath{\mathsf{Prop}}\xspace}
\newcommand{\ensuremath{\mathsf{Prop}_\UU}\xspace}{\ensuremath{\mathsf{Prop}_\ensuremath{\mathcal{U}}\xspace}\xspace}
\newcommand{\pointed}[1]{\ensuremath{#1_\bullet}}
\newcommand{\ensuremath{\mathsf{Card}}\xspace}{\ensuremath{\mathsf{Card}}\xspace}
\newcommand{\ensuremath{\mathsf{Ord}}\xspace}{\ensuremath{\mathsf{Ord}}\xspace}
\newcommand{\ordsl}[2]{{#1}_{/#2}}
\newcommand{\ua}{\ensuremath{\mathsf{ua}}\xspace}
\newcommand{\ensuremath{\mathsf{idtoeqv}}\xspace}{\ensuremath{\mathsf{idtoeqv}}\xspace}
\newcommand{\univalence}{\ensuremath{\mathsf{univalence}}\xspace}
\newcommand{\ensuremath{\mathsf{isContr}}}{\ensuremath{\mathsf{isContr}}}
\newcommand{\contr}{\ensuremath{\mathsf{contr}}}
\newcommand{\ensuremath{\mathsf{isSet}}}{\ensuremath{\mathsf{isSet}}}
\newcommand{\ensuremath{\mathsf{isProp}}}{\ensuremath{\mathsf{isProp}}}
\let\hfiber\hfib
\newcommand{\trunc}[2]{\mathopen{}\left\Vert #2\right\Vert_{#1}\mathclose{}}
\newcommand{\ttrunc}[2]{\bigl\Vert #2\bigr\Vert_{#1}}
\newcommand{\Trunc}[2]{\Bigl\Vert #2\Bigr\Vert_{#1}}
\newcommand{\truncf}[1]{\Vert \mathord{\hspace{1pt}\text{--}\hspace{1pt}} \Vert_{#1}}
\newcommand{\tproj}[3][]{\mathopen{}\left|#3\right|_{#2}^{#1}\mathclose{}}
\newcommand{\tprojf}[2][]{|\mathord{\hspace{1pt}\text{--}\hspace{1pt}}|_{#2}^{#1}}
\def\trunc0{\trunc0}
\newcommand{\brck}[1]{\trunc{}{#1}}
\newcommand{\bbrck}[1]{\ttrunc{}{#1}}
\newcommand{\Brck}[1]{\Trunc{}{#1}}
\newcommand{\bproj}[1]{\tproj{}{#1}}
\newcommand{\tprojf{}}{\tprojf{}}
\newcommand{\Parens}[1]{\Bigl(#1\Bigr)}
\let\extendsmb\mathsf{ext}
\newcommand{\extend}[1]{\extendsmb(#1)}
\newcommand{\ensuremath{\mathbf{0}}\xspace}{\ensuremath{\mathbf{0}}\xspace}
\newcommand{\ensuremath{\mathbf{1}}\xspace}{\ensuremath{\mathbf{1}}\xspace}
\newcommand{\ensuremath{\star}\xspace}{\ensuremath{\star}\xspace}
\newcommand{\ensuremath{\mathbf{2}}\xspace}{\ensuremath{\mathbf{2}}\xspace}
\newcommand{{1_{\bool}}}{{1_{\ensuremath{\mathbf{2}}\xspace}}}
\newcommand{{0_{\bool}}}{{0_{\ensuremath{\mathbf{2}}\xspace}}}
\newcommand{{\mathsf{inl}}}{{\mathsf{inl}}}
\newcommand{{\mathsf{inr}}}{{\mathsf{inr}}}
\newcommand{\ensuremath\inlsym\xspace}{\ensuremath{\mathsf{inl}}\xspace}
\newcommand{\ensuremath\inrsym\xspace}{\ensuremath{\mathsf{inr}}\xspace}
\newcommand{\ensuremath{\mathsf{seg}}\xspace}{\ensuremath{\mathsf{seg}}\xspace}
\newcommand{\freegroup}[1]{F(#1)}
\newcommand{\freegroupx}[1]{F'(#1)}
\newcommand{\mathsf{glue}}{\mathsf{glue}}
\newcommand{\mathbb{S}}{\mathbb{S}}
\newcommand{\ensuremath{\mathsf{base}}\xspace}{\ensuremath{\mathsf{base}}\xspace}
\newcommand{\ensuremath{\mathsf{loop}}\xspace}{\ensuremath{\mathsf{loop}}\xspace}
\newcommand{\ensuremath{\mathsf{surf}}\xspace}{\ensuremath{\mathsf{surf}}\xspace}
\newcommand{\Sigma}{\Sigma}
\newcommand{\mathsf{N}}{\mathsf{N}}
\newcommand{\mathsf{S}}{\mathsf{S}}
\newcommand{\mathsf{merid}}{\mathsf{merid}}
\newcommand{\mathord{\hspace{1pt}\text{--}\hspace{1pt}}}{\mathord{\hspace{1pt}\text{--}\hspace{1pt}}}
\newcommand{\mathord{\hspace{1pt}\underline{\hspace{1ex}}\hspace{1pt}}}{\mathord{\hspace{1pt}\underline{\hspace{1ex}}\hspace{1pt}}}
\newcommand{\ensuremath{\mathbb{P}}\xspace}{\ensuremath{\mathbb{P}}\xspace}
\newcommand{\ensuremath{\mathcal{S}et}\xspace}{\ensuremath{\mathcal{S}et}\xspace}
\newcommand{\ensuremath{{\mathcal{C}at}}\xspace}{\ensuremath{{\mathcal{C}at}}\xspace}
\newcommand{\ensuremath{\mathcal{R}el}\xspace}{\ensuremath{\mathcal{R}el}\xspace}
\newcommand{\ensuremath{\mathcal{H}ilb}\xspace}{\ensuremath{\mathcal{H}ilb}\xspace}
\newcommand{\ensuremath{\mathcal{T}\!ype}\xspace}{\ensuremath{\mathcal{T}\!ype}\xspace}
\newcommand{\ensuremath{\mathsf{idtoiso}}\xspace}{\ensuremath{\mathsf{idtoiso}}\xspace}
\newcommand{\ensuremath{\mathsf{isotoid}}\xspace}{\ensuremath{\mathsf{isotoid}}\xspace}
\newcommand{^{\mathrm{op}}}{^{\mathrm{op}}}
\newcommand{\ensuremath{\mathbf{y}}\xspace}{\ensuremath{\mathbf{y}}\xspace}
\newcommand{\dgr}[1]{{#1}^{\dagger}}
\newcommand{\mathrel{\cong^\dagger}}{\mathrel{\cong^\dagger}}
\newcommand{\cteqv}[2]{\ensuremath{#1 \simeq #2}\xspace}
\newcommand{\cteqvsym}{\simeq}
\newcommand{\textbf{N}}{\ensuremath{\mathbb{N}}\xspace}
\let\nat\textbf{N}
\newcommand{\natp}{\ensuremath{\nat'}\xspace}
\newcommand{\zerop}{\ensuremath{0'}\xspace}
\newcommand{\mathsf{succ}}{\mathsf{succ}}
\newcommand{\sucp}{\ensuremath{\mathsf{succ}'}\xspace}
\newcommand{\mathsf{add}}{\mathsf{add}}
\newcommand{\mathsf{ack}}{\mathsf{ack}}
\newcommand{\mathsf{iter}}{\mathsf{iter}}
\newcommand{\mathsf{assoc}}{\mathsf{assoc}}
\newcommand{\ensuremath{\mathsf{double}}}{\ensuremath{\mathsf{double}}}
\newcommand{\dblp}{\ensuremath{\ensuremath{\mathsf{double}}'}\xspace}
\newcommand{\lst}[1]{\mathsf{List}(#1)}
\newcommand{\mathsf{nil}}{\mathsf{nil}}
\newcommand{\mathsf{cons}}{\mathsf{cons}}
\newcommand{\vect}[2]{\ensuremath{\mathsf{Vec}_{#1}(#2)}\xspace}
\newcommand{\ensuremath{\mathbb{Z}}\xspace}{\ensuremath{\mathbb{Z}}\xspace}
\newcommand{\mathsf{succ}}{\mathsf{succ}}
\newcommand{\mathsf{pred}}{\mathsf{pred}}
\newcommand{\ensuremath{\mathbb{Q}}\xspace}{\ensuremath{\mathbb{Q}}\xspace}
\newcommand{\mathsf{funext}}{\mathsf{funext}}
\newcommand{\mathsf{happly}}{\mathsf{happly}}
\newcommand{\com}[3]{\mathsf{swap}_{#1,#2}(#3)}
\newcommand{\ensuremath{\mathsf{code}}\xspace}{\ensuremath{\mathsf{code}}\xspace}
\newcommand{\ensuremath{\mathsf{encode}}\xspace}{\ensuremath{\mathsf{encode}}\xspace}
\newcommand{\ensuremath{\mathsf{decode}}\xspace}{\ensuremath{\mathsf{decode}}\xspace}
\newcommand{\function}[4]{\left\{\begin{array}{rcl}#1 &
\longrightarrow & #2 \\ #3 & \longmapsto & #4 \end{array}\right.}
\newcommand{\cone}[2]{\mathsf{cone}_{#1}(#2)}
\newcommand{\cocone}[2]{\mathsf{cocone}_{#1}(#2)}
\newcommand{\composecocone}[2]{#1\circ#2}
\newcommand{\composecone}[2]{#2\circ#1}
\newcommand{\mathscr{D}}{\mathscr{D}}
\newcommand{\mathsf{Map}}{\mathsf{Map}}
\newcommand{\ensuremath{I}\xspace}{\ensuremath{I}\xspace}
\newcommand{\ensuremath{0_{\interval}}\xspace}{\ensuremath{0_{\ensuremath{I}\xspace}}\xspace}
\newcommand{\ensuremath{1_{\interval}}\xspace}{\ensuremath{1_{\ensuremath{I}\xspace}}\xspace}
\newcommand{\ensuremath{\twoheadrightarrow}}{\ensuremath{\twoheadrightarrow}}
\newcommand{\ensuremath{\rightarrowtail}}{\ensuremath{\rightarrowtail}}
\newcommand{\ensuremath{\mathrel{\widetilde{\in}}}}{\ensuremath{\mathrel{\widetilde{\in}}}}
\newcommand{\ensuremath{\mathsf{SemigroupStr}}}{\ensuremath{\mathsf{SemigroupStr}}}
\newcommand{\semigroupstr}[1]{\ensuremath{\mathsf{SemigroupStr}}(#1)}
\newcommand{\semigroup}[0]{\ensuremath{\mathsf{Semigroup}}}
\newcommand{\ensuremath{\cdot}}{\ensuremath{\cdot}}
\newcommand{\vcentcolon\vcentcolon=}{\vcentcolon\vcentcolon=}
\newcommand{\downarrow}{\downarrow}
\newcommand{\ensuremath{\mathsf{ctx}}}{\ensuremath{\mathsf{ctx}}}
\newcommand{\wfctx}[1]{#1\ \ensuremath{\mathsf{ctx}}}
\newcommand{\oftp}[3]{#1 \vdash #2 : #3}
\newcommand{\jdeqtp}[4]{#1 \vdash #2 \equiv #3 : #4}
\newcommand{\judg}[2]{#1 \vdash #2}
\newcommand{\tmtp}[2]{#1 \mathord{:} #2}
\newcommand{\textsc{form}}{\textsc{form}}
\newcommand{\textsc{intro}}{\textsc{intro}}
\newcommand{\textsc{elim}}{\textsc{elim}}
\newcommand{\textsc{comp}}{\textsc{comp}}
\newcommand{\textsc{uniq}}{\textsc{uniq}}
\newcommand{\mathsf{Wkg}}{\mathsf{Wkg}}
\newcommand{\mathsf{Vble}}{\mathsf{Vble}}
\newcommand{\mathsf{Exch}}{\mathsf{Exch}}
\newcommand{\mathsf{Subst}}{\mathsf{Subst}}
\newcommand{\mathsf{c}}{\mathsf{c}}
\newcommand{\mathsf{p}}{\mathsf{p}}
\newcommand{\widetilde{\mathsf{c}}}{\widetilde{\mathsf{c}}}
\newcommand{\widetilde{\mathsf{p}}}{\widetilde{\mathsf{p}}}
\newcommand{\ensuremath{\widetilde{W}}\xspace}{\ensuremath{\widetilde{W}}\xspace}
\newcommand{\istype}[1]{\mathsf{is}\mbox{-}{#1}\mbox{-}\mathsf{type}}
\newcommand{\ensuremath{(n+1)}}{\ensuremath{(n+1)}}
\newcommand{\ensuremath{(n-1)}}{\ensuremath{(n-1)}}
\newcommand{\kbar}{\overline{k}}
\newcommand{\ensuremath{\mathbf{N^w}}\xspace}{\ensuremath{\mathbf{N^w}}\xspace}
\newcommand{\ensuremath{0^\mathbf{w}}\xspace}{\ensuremath{0^\mathbf{w}}\xspace}
\newcommand{\ensuremath{\mathbf{s^w}}\xspace}{\ensuremath{\mathbf{s^w}}\xspace}
\newcommand{\nat\mathsf{Alg}}{\nat\mathsf{Alg}}
\newcommand{\nat\mathsf{Hom}}{\nat\mathsf{Hom}}
\newcommand{\mathsf{isHinit}_{\mathsf{W}}}{\mathsf{isHinit}_{\mathsf{W}}}
\newcommand{\mathsf{isHinit}_\nat}{\mathsf{isHinit}_\nat}
\newcommand{\mathsf{W}}{\mathsf{W}}
\newcommand{\w\mathsf{Alg}}{\mathsf{W}\mathsf{Alg}}
\newcommand{\w\mathsf{Hom}}{\mathsf{W}\mathsf{Hom}}
\newcommand{\RC}{\ensuremath{\mathbb{R}_\mathsf{c}}\xspace}
\newcommand{\RD}{\ensuremath{\mathbb{R}_\mathsf{d}}\xspace}
\newcommand{\R}{\ensuremath{\mathbb{R}}\xspace}
\newcommand{\barRD}{\ensuremath{\bar{\mathbb{R}}_\mathsf{d}}\xspace}
\newcommand{\close}[1]{\sim_{#1}}
\newcommand{\mathord\sim}{\mathord\sim}
\newcommand{\rclim}{\mathsf{lim}}
\newcommand{\rcrat}{\mathsf{rat}}
\newcommand{\rceq}{\mathsf{eq}_{\RC}}
\newcommand{\CAP}{\mathcal{C}}
\newcommand{\Q_{+}}{\ensuremath{\mathbb{Q}}\xspace_{+}}
\newcommand{\apart}{\mathrel{\#}}
\newcommand{\dcut}{\mathsf{isCut}}
\newcommand{\cover}{\triangleleft}
\newcommand{\intfam}[3]{(#2, \lam{#1} #3)}
\newcommand{\frown}{\frown}
\newcommand{\smile}{\smile}
\newcommand{\diamondsuit\approx}{\diamondsuit\approx}
\newcommand{\diamondsuit}{\diamondsuit}
\newcommand{\heartsuit\approx}{\heartsuit\approx}
\newcommand{\heartsuit}{\heartsuit}
\newcommand{\tap}[1]{\bullet\approx_{#1}\triangle}
\newcommand{\triangle}{\triangle}
\newcommand{\tapb}[1]{\bullet\approx_{#1}\square}
\newcommand{\square}{\square}
\newcommand{\ensuremath{\mathsf{No}}\xspace}{\ensuremath{\mathsf{No}}\xspace}
\newcommand{\surr}[2]{\{\,#1\,\big|\,#2\,\}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\noeq}{\mathsf{eq}_{\ensuremath{\mathsf{No}}\xspace}}
\newcommand{\trianglelefteqslant}{\trianglelefteqslant}
\newcommand{\vartriangleleft}{\vartriangleleft}
\newcommand{\sqsubseteq}{\sqsubseteq}
\newcommand{\sqsubset}{\sqsubset}
\newcommand{\diamondsuit\preceq}{\diamondsuit\preceq}
\newcommand{\diamondsuit\prec}{\diamondsuit\prec}
\newcommand{\diamondsuit}{\diamondsuit}
\newcommand{\heartsuit\preceq}{\heartsuit\preceq}
\newcommand{\heartsuit\prec}{\heartsuit\prec}
\newcommand{\heartsuit}{\heartsuit}
\newcommand{\triangle\preceq}{\triangle\preceq}
\newcommand{\triangle\prec}{\triangle\prec}
\newcommand{\triangle}{\triangle}
\newcommand{\square\preceq}{\square\preceq}
\newcommand{\square\prec}{\square\prec}
\newcommand{\square}{\square}
\newcommand{\vset}{\mathsf{set}}
\def\tproj0{\tproj0}
\newcommand{\inj}{\ensuremath{\mathsf{inj}}}
\newcommand{\acc}{\ensuremath{\mathsf{acc}}}
\newcommand{\mathsf{atMostOne}}{\mathsf{atMostOne}}
\newcommand{\power}[1]{\mathcal{P}(#1)}
\newcommand{\powerp}[1]{\mathcal{P}_+(#1)}
\makeatletter
\def\defthm#1#2#3{%
\newaliascnt{#1}{thm}
\newtheorem{#1}[#1]{#2}
\aliascntresetthe{#1}
\crefname{#1}{#2}{#3}}
\newtheorem{thm}{Theorem}[section]
\crefname{thm}{Theorem}{Theorems}
\defthm{propn}{Proposition}{Propositions}
\defthm{cor}{Corollary}{Corollaries}
\defthm{lem}{Lemma}{Lemmas}
\defthm{ax}{Axiom}{Axioms}
\defthm{conj}{Conjecture}{Conjectures}
\theoremstyle{definition}
\defthm{defn}{Definition}{Definitions}
\theoremstyle{remark}
\defthm{rmk}{Remark}{Remarks}
\defthm{ex}{Example}{Examples}
\defthm{exs}{Examples}{Examples}
\defthm{notes}{Notes}{Notes}
\crefformat{section}{\S#2#1#3}
\Crefformat{section}{Section~#2#1#3}
\crefrangeformat{section}{\S\S#3#1#4--#5#2#6}
\Crefrangeformat{section}{Sections~#3#1#4--#5#2#6}
\crefmultiformat{section}{\S\S#2#1#3}{ and~#2#1#3}{, #2#1#3}{ and~#2#1#3}
\Crefmultiformat{section}{Sections~#2#1#3}{ and~#2#1#3}{, #2#1#3}{ and~#2#1#3}
\crefrangemultiformat{section}{\S\S#3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}{, #3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}
\Crefrangemultiformat{section}{Sections~#3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}{, #3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}
\crefformat{appendix}{Appendix~#2#1#3}
\Crefformat{appendix}{Appendix~#2#1#3}
\crefrangeformat{appendix}{Appendices~#3#1#4--#5#2#6}
\Crefrangeformat{appendix}{Appendices~#3#1#4--#5#2#6}
\crefmultiformat{appendix}{Appendices~#2#1#3}{ and~#2#1#3}{, #2#1#3}{ and~#2#1#3}
\Crefmultiformat{appendix}{Appendices~#2#1#3}{ and~#2#1#3}{, #2#1#3}{ and~#2#1#3}
\crefrangemultiformat{appendix}{Appendices~#3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}{, #3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}
\Crefrangemultiformat{appendix}{Appendices~#3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}{, #3#1#4--#5#2#6}{ and~#3#1#4--#5#2#6}
\crefname{part}{Part}{Parts}
\setcounter{secnumdepth}{5}
\crefname{figure}{Figure}{Figures}
\let\autoref\cref
\let\c@equation\c@thm
\numberwithin{equation}{section}
\renewcommand{\theenumi}{(\roman{enumi})}
\renewcommand{\labelenumi}{\theenumi}
\def\noteson{%
\gdef\note##1{\mbox{}\marginpar{\color{blue}\textasteriskcentered\ ##1}}}
\gdef\notesoff{\gdef\note##1{\null}}
\noteson
\newcommand{\textsc{Coq}\xspace}{\textsc{Coq}\xspace}
\newcommand{\textsc{Agda}\xspace}{\textsc{Agda}\xspace}
\newcommand{\textsc{NuPRL}\xspace}{\textsc{NuPRL}\xspace}
\newcommand{\footstyle}[1]{{\hyperpage{#1}}n}
\newcommand{\defstyle}[1]{\textbf{\hyperpage{#1}}}
\newcommand{\indexdef}[1]{\index{#1|defstyle}}
\newcommand{\indexfoot}[1]{\index{#1|footstyle}}
\newcommand{\indexsee}[2]{\index{#1|see{#2}}}
\newcommand{Zermelo--Fraenkel}{Zermelo--Fraenkel}
\newcommand{Constructive \ZF{} Set Theory}{Constructive Zermelo--Fraenkel{} Set Theory}
\newcommand{\LEM}[1]{\ensuremath{\mathsf{LEM}_{#1}}\xspace}
\newcommand{\choice}[1]{\ensuremath{\mathsf{AC}_{#1}}\xspace}
\newcommand{\mentalpause}{\medskip}
\newcounter{symindex}
\newcommand{\symlabel}[1]{\refstepcounter{symindex}\label{#1}}
|
\section{Introduction}
\label{sec:introduction}
\input{introduction.tex}
\paragraph{Preliminaries.}
\label{sec:prel}
\input{prel.tex}
\section{The Upper Bound}
\label{sec:upper-bound}
\input{upper-bound.tex}
\section{The Lower Bound}
\label{sec:lower-bound}
\input{lower-bound.tex}
\section{Conclusions}
\label{sec:conclusions}
\input{conclusions.tex}
\bibliographystyle{splncs03}
|
\section{Introduction}
Monitoring the ocean and coastal areas has traditionally been accomplished
using either moored buoys or fixed networks of bulky, partially submerged
platforms. The cost of developing, producing, deploying, and maintaining such
systems has severely limited the range and precision of the sensed data
obtained from monitoring large-scale waterbodies. To overcome these
limitations, the community has started focusing on developing simpler,
smaller, motorized, autonomous surface
vehicles~\cite{manley2008unmanned,2009AGUSMOS22A..08O,bayat2017environmental,ziccarelli2016novel,ferreira2009autonomous}. A
large group of such vehicles operating as a collective can be deployed,
retrieved, and re-deployed (partially or completely) at considerably lower
cost and operational complexity than monolithic structures. Following the
design paradigm of swarm robotics, these collectives perform complex tasks
cooperatively in a scalable fashion, with a behavior that is both robust to
failures and flexible so as to operate in dynamic
environments~\cite{bouffanais15:_desig_contr_swarm_dynam,chamanbaz17:_swarm_enabl_techn_multi_robot_system}.
Moreover, a swarm of mobile surface vehicles forms a dynamic sensor network
suitable for high temporal sampling of waterbodies.
The pressing need for small, low-cost and rapidly deployable autonomous
vehicles has been acknowledged in multiple recent
reports~\cite{matos_man_2016,nishida_development_2015,picomultipurpose,vesecky_prototype_2007}.
However, using such large distributed systems for ocean monitoring presents as many possibilities as it does challenges. On the one hand, the system modularity and the individual cost of production primes them for iterative design and fast prototyping, in which field tests provide insights enabling a constant improvement toward future models. On the other hand, this iterative process is likely to produce an heterogeneous system, in which the agents have different motor, sensing, communication, or processing capabilities. This heterogeneous character must be properly handled at the system-level design in order to benefit from the range of agent capabilities.
We have previously reported on the design, construction, and testing of a
homogeneous swarm of 50 (identical) autonomous buoys performing adaptive
deployment for applications in environmental
monitoring~\cite{Zoss2018}. Here, we investigate and study the nontrivial process of
partially upgrading this swarm robotics system by replacing a small fraction
of the original swarming units (simply denoted \bobold{}) with upgraded robotic
platforms (\bobnew{}), based on a fully
redesigned model featuring highly improved motor and processing capabilities
among other things.
\begin{figure}[htb]
\centering
\includegraphics[width=.9\linewidth]{figs/rain}
\caption{Swarm of buoys being deployed in the target area undergoing
monitoring. Note that some units are still being deployed from the shore
where the picture is taken.}
\label{fig:photo}
\end{figure}
We have combined a set of 18 \bobold{} buoys with 4 upgraded \bobnew{} buoys---hence constituting a heterogeneous swarm---and performed dynamic area coverage experiments in fully unstructured environments without the support of any external infrastructure when it comes to the system's operations (see Fig.~\ref{fig:photo}).
We measured the collective responsiveness---a proxy to assess the flexible character of a swarm~\cite{mateo17:_effec_correl_swarm_collec_respon}---by changing the target area at different frequencies in order to investigate the minimum response time at which the heterogeneous swarm is able to adapt to changes in the monitored environment.
For this series of experiments on dynamic area coverage, the only practical differences between the two platforms are the speed of the units, their inertia, and their hull design.
Specifically, design \bobnew{} is 80\% times faster in open water and 2.3 times lighter than \bobold{}, thereby making it considerably more apt to quickly correct its position to adapt to changes in the target area undergoing monitoring.
The field experiments performed reveal that the partial upgrade of the system is able to increase the average response of the system, but not necessarily the uniformity of the deployment.
\section{Robotic Platforms}\label{sec:fancy_buoy}
In~\cite{Zoss2018}, we presented a study of dynamic ocean monitoring using a
homogeneous swarm of 50 buoys. Such decentralized and cooperative systems primarily owe their outstanding
effectiveness to the large number of agents put together: a greater number of agents allows for vaster waterbody coverage and/or refined multi-point sensing.
As technology inexorably moves forward at an increasingly faster pace, there is a compelling opportunity to expand and upgrade at least some of the agents of the collective.
These new units with improved capabilities have to be able to fit and operate within the technical framework
of the original system.
The original buoy platform---design \bobold{}---is thoroughly described
in~\cite{Zoss2018},
and we refer the reader to this report for full technical details.
In this section, we focus on introducing an improved design (\bobnew{}) and its main differences with respect to \bobold{}. This new design (see Fig \ref{fig:explodedview}), re-designed mainly to facilitate operation and maintenance, can be used
to replace some units in the homogeneous swarm, or expand the original homogeneous system.
\begin{figure}
\includegraphics[width=0.9\linewidth]{figs/whole_exploded}.
\caption{Exploded view of the upgraded buoy design \bobnew{} displaying its
modular design and main components.
}
\label{fig:explodedview}
\end{figure}
The platform has a modular design consisting of (i) a cylindrical hull, (ii) a
top lid hosting the electronics on one side and antennas on the other, and
(iii) a detachable bottom section that can be customized to host a number of
environmental sensors, as shown in Fig.~\ref{fig:explodedview}.
The main hull for \bobnew{} is designed to provide the same omnidirectional vector
propulsion apparatus as in \bobold{}, while encapsulating the three thrusters
within the cylindrical shape and keeping them as close as possible to the
center of gravity of the platform. This design choice stands in clear contrast
with \bobold{} that has protruding pairs of propellers, see
Fig.~\ref{fig:buoy_comparison}. On the one hand, this new design minimizes the
wobbling when the thrusters are active and provide a
certain level of dynamic stability to the surface vehicle.
On the other hand, the new hull has a larger packing density, which
facilitates stackability and transportation of large numbers of platforms to
remote field areas (see Fig. 1 in~\cite{Zoss2018}).
The mechanical design of \bobnew{} and the distribution of components
inside the main hull has been designed and tested to not compromise
the hydrostatics of \bobold{},
thereby retaining the critical self-righting feature of the \bobold{} design.
The presence of three lateral aluminum pads allows for heat exchange between
the water and the air inside the hull---a fan ensures adequate air circulation to
homogenize the heat generated
mainly by the 3 electronic speed controllers (ESC in Fig.~\ref{fig:explodedview}) and
the single-board computer Odroid C2. These aluminum pads can themselves host small sensors or light-emitting devices
orientated 45 degrees below the water surface.
The top lid hosts the electronics and related hardware as a modular design
itself. An electronic rack built around the main PCB and battery is screwed on
the inside of this lid, thereby forming one single piece that can easily be
replaced or disassembled for rapid inspection. Following the buoy \bobold{}
design, the outer part of the top lid hosts connectivity hardware such as
communication and GPS antennas, and a charging port. The top lid of \bobnew{}
includes an LCD screen reporting a range of data and indicators,
which greatly facilitates the basic inspection and troubleshooting of the
system's digital status. Lastly, a mechanical handle has been incorporated
into \bobnew{}'s top lid design to aid both manual carrying and latching a
stacked platform on top of it.
The bottom part is designed to securely host the battery, and
can be modified to accommodate any sensor.
The latter can be easily and
conveniently interfaced with the electronic stack through a mini-PCI custom
PCB that can be slotted inside the main PCB. This configuration was specifically
designed to support
further heterogeneity of the system at the sensing level, with various units
possibly carrying different sensors.
Particular attention has been placed on identifying a simple way of assembling parts and components.
An example is the thruster attachment, in which a combination of a laser-cut gaskets and screws allows for servicing all three thrusters in less than 5 minutes---eight times faster than for the \bobold{} design.
In addition, the structure of the new electronics rack is composed of laser-cut acrylic sheets and designed using screw-less joineries to facilitate manufacture, assembly and reduce potential leaks.
Once assembled, platform \bobnew{} measures $260\times260\times245$~mm and weights 3.2~kg.
For comparison, the previous design measures $350\times350\times280$~mm and weights 7.4~kg.
The reduction in dimensions is primarily achieved by moving from a spherical
form to a cylindrical one, which allows for a compact embedding of the
thrusters. The reduction in weight comes from using plastic material for the
hull as opposed to aluminum alloy in \bobold{}. Besides weight reduction, the
usage of 3D printed nylon plastic material inherently prevents corrosion and
safely withstands seawater conditions, while reducing the amount of biofouling
accumulating on the platform during operations as compared to metal.
Due to the low number of \bobnew{} prototypes, the plastic hulls have been
3D-printed, making the cost comparable with the aluminum hull of
\bobold{}. However, when mass producing the units using plastic injection
inside a mold, the cost per unit is expected to be significantly lowered.
\subsection*{Performance comparison: buoy \bobold{} vs. \bobnew{}}
Fig.~\ref{fig:buoy_comparison} shows a visual comparison between buoy
\bobold{} and one of the prototypes of the improved design \bobnew{}. By
incorporating the thrusters into the main hull, buoy \bobnew{} is much more
compact and increases the propellers' life---the propellers are protected as
compared to \bobold{} in which the protruding pods and open propellers can
easily be damaged during transportation and/or operation.
\begin{figure}
\centering
\includegraphics[width=.90\linewidth]{figs/buoy_comparison}
\caption{The two robotic platforms: the new buoy design \bobnew{} (left) side by side with one of the previous buoy model \bobnew{} (right).
}
\label{fig:buoy_comparison}
\end{figure}
A series of performance tests were carried out under different environmental
conditions to offer a detailed comparison of the two platforms. The tests were
run over a distance of $100$~m in the direction of the prevailing the wind and
against it. The platforms were initially positioned next to one another so as
to obtain results under the practically identical conditions. The three
BlueRobotics T100 thrusters allow the buoy \bobnew{} to move about $80$\%
faster than \bobold{} in nominal test operations, although the thrusters are
capable of performing at much higher intensity.
This self-imposed limitation on the thrusters' intensity is purely related to the need to develop a more complex vectorial propulsion system than \bobold{}'s.
However, since the cooperative control strategy remains the same for \bobold{} and \bobnew{} when performing the heterogeneous
swarming tests, it was decided to not use the full power of the
thrusters of \bobnew{}. When testing the two platforms, side by side, in
adverse environmental conditions---strong winds, waves and heavy rains, the
performance of \bobnew{} was practically unaffected. On the contrary, buoy
\bobold{} drastically underperformed, with instances at which it was fully
unable to reach its target goal.
It is worth noting that buoy \bobnew{} is granted a battery 1.5 bigger than
the one in \bobold{} to compensate the need for higher power by the thrusters,
and the extra electronics energy consumption due to fan and LCD.
\section{Dynamic Area Coverage}
\subsection{Collective Operations}
Our multi-robot system can perform a range of collective behaviors
achieved by means of a cooperative control strategy supported by distributed
communications as described in full details in~\cite{Zoss2018}. For instance,
this distributed robotic system can perform a number of elementary swarming
behaviors, including flocking, navigation, and area coverage. The
effectiveness of the distributed communication setup has been verified and
analyzed in~\cite{Zoss2018}.
This large-scale networked array of mobile sensing units is designed to
monitor and characterize waterbodies, which may vary depending on the
application.
An example is the deployment of the swarm of buoys in a harbor to
assist in marine operations by monitoring key environmental and flow
parameters. More interestingly, the area to monitor might not be specified
externally or in advance, but instead be defined dynamically by the collective
of agents itself. By local processing of the sensed data, the agents may
determine the shape in which to self-deploy in order to track a particular
temperature profile, oil spill, or a range of biological markers.
\subsection{Cooperative Control for Dynamic Area Coverage}
As discussed in the previous section, the distributed robotic systems are designed to monitor features of interest in waterbodies (e.g. oil spills,
algal blooms, etc.). To do so, we require the agents to spread as uniformly as
possible across a given area that dynamically evolves over time.
It is paramount for this spreading to happen in a timely and responsive
way, as the shape of the area of interest is in general time-changing with
arbitrary dynamics. For this reason, we choose to define the behavior of the
agents in a purely Markovian fashion with agent's dynamical rules that
determine the buoys' movements based solely on the instantaneous state,
i.e. the cooperative control algorithm for agent $i$ can be cast as
\begin{equation}
\frac{d\vec{r}_i}{dt} = \vec{F}\left(t, \vec{r}_i(t), \{\vec{r}_j(t)\}_{j \sim i}\right) .
\label{eq:general-control}
\end{equation}
where $\vec{r}_i$ is the position of a given agent $i$, and $j\sim i$ is the set
of agents in the neighborhood of $i$---its ``neighbors'' according to a
specified interaction distance, be it metric, topological or else~\cite{shang14:_consen}.
The area of interest is typically determined either by external sources
(e.g. a human operator) or by the sensing capabilities of the agents
(e.g. collective tracking of a temperature gradient, or chemical concentration,
etc.). For the sake of generality, we assume that the area to monitor can be
described mathematically by
\begin{equation}
A(\vec{r}) < 0 ,
\label{eq:area}
\end{equation}
where $A$ is a signed distance function (or at least a function that increases
monotonically outside the region).
Given~\eqref{eq:area}, the cooperative control rule is defined as
\begin{equation}
\frac{d\vec{r}_i}{dt} = v_{0i} \frac{\vec{T}}{\max(1, \|\vec{T}\|)},
\end{equation}
where $v_{0i}$ is the maximum speed of an agent $i$ and $\vec{T}$ is the
``area coverage target'' defined by
\begin{equation}
\vec{T} = \frac{1}{1+\exp{(-\beta A(\vec{r}_i))}}\frac{-\vec{\nabla}A}{\|\vec{\nabla}A\|} - \sum_{j \sim i} \frac{a_R^d}{r_{ij}^d} \frac{\vec{r}_{ij}}{r_{ij}} \, .
\label{eq:geofencing}
\end{equation}
The first term in~\eqref{eq:geofencing}, proportional to $-\vec{\nabla}A$, attracts the agents towards the interior of the area and is scaled in such a way that its norm goes from being practically zero outside the area to cover ($A>0$) to being unity inside it ($A<0$).
The parameter $\beta$ controls how abruptly the transition between zero and
one is, i.e. the steepness of the exponential decrease.
The second term is an inter-agent repulsion term that causes the agents to spread inside the area.
The type of repulsion is controlled by two parameters, namely the repulsion strength $a_R$ and the power of repulsion $d$.
If the power $d$ is large ($d \gtrsim 4$), the repulsion strength is
approximately equal to the nearest-neighbor distance in equilibrium configurations.
Following numerical optimization run over simulations of the behavior \eqref{eq:geofencing}, we set the free parameters to
\begin{align}
d = 6, && a_R = 0.38 \sqrt{S / N}, && \beta = 40 / S,
\end{align}
where $S$ is the total surface area to cover and $N$ the number of agents.
\subsection{Target Surface to Monitor}
As a testbed for dynamic area exploration, we consider the following surface
\begin{equation}
A_{\alpha,\hat{e}}(\vec{r}) = r^2 - R_{\alpha,\hat{e}}^2(\hat{r}) ,
\label{eq:peanut}
\end{equation}
where
\begin{equation}
R_{\alpha,\hat{e}}(\hat{r}) = \frac{R_0}{2}\frac{2 - \alpha + 3\alpha (\hat{r}\cdot\hat{e})^2}{\sqrt{1 + \alpha/2 + 11\alpha^2/32}} .
\label{eq:radius}
\end{equation}
By changing $\alpha$, the shape of this region goes from a disk of radius $R_0$ ($\alpha=0$), to a two-lobbed area along the principal axis $\hat{e}$ ($\alpha=1$), to a ``dumbbell-shaped'' region with a nodal point at the origin ($\alpha=2$).
The normalization factor in~\eqref{eq:radius} is introduced so that the total surface of the area is kept constant at $S=\pi R_0^2$ for any $\alpha \in [0,2]$.
The gradient of this surface is
\begin{equation}
\nabla A_{\alpha,\hat{e}}(\vec{r}) = 2 \vec{r} - 6\alpha R_0 \frac{R_{\alpha,\hat{e}}(\hat{r})}{r}(\hat{r}\cdot\hat{e})\hat{e}_\bot \, ,
\end{equation}
where $\hat{e}_\bot = \hat{e} - (\hat{r}\cdot\hat{e})\hat{r}$ is the orthogonal projection of the principal axis $\hat{e}$ on the position $\vec{r}$.
The attractive geofencing force field corresponding to this target surface is presented in Fig.~\ref{fig:forcefield}.
The contours depict the lines of constant $A_{\alpha, \hat{e}}$ and the arrows
show the vector field of the first term of~\eqref{eq:geofencing} for three values of $\alpha$.
\begin{figure}
\includegraphics{figs/forcefield}
\caption{Effect of the geofencing term for the area given by~\eqref{eq:peanut}.
The arrows show the vector field of the first term in~\eqref{eq:geofencing},
and the color gradient maps its norm.
The contour lines mark isolines $A_{\alpha,\hat{e}}$.
}
\label{fig:forcefield}
\end{figure}
\section{Theoretical analysis}
\label{sec:theory}
We can study the performance of the cooperative rule~\eqref{eq:geofencing} in
\emph{dynamic} area coverage by imposing a cyclic temporal evolution of the shape of the monitoring area at different frequencies.
To quantify the response of the system, we consider two metrics for the
dynamic coverage performance that were previously used when testing with a
homogeneous swarm~\cite{Zoss2018}.
The first one is the ``tessellation performance'' $P_T$, defined as the inverse of the relative size of the largest cell assigned to any agent after segmenting the target area with a Voronoi tessellation, or
\begin{equation}
P_T = \frac{A}{A_{LVC} N},
\end{equation}
where $A$ is the total area of the target surface, $A_{LVC}$ is the area of the largest Voronoi cell (dark cell in the inset of the top panel in Fig.~\ref{fig:homo_ideal}), and $N$ is the number of agents.
The second metric is the so-called ``coverage performance'' $P_C$, defined as the percentage
of the target area covered by the agents (assuming each agent covers a disk of a certain radius $R_s$ around its position), or
\begin{equation}
P_C = \frac{\bigcup_{i=1}^{N} {c_i} \cap A}{A},
\end{equation}
where $c_i$ is a disk of radius $R_s$ centered at the position of agent $i$
(see inset at the bottom of Fig.~\ref{fig:homo_ideal}).
The tessellation performance is meant to capture the expected accuracy of the
least accurate mobile sensing unit, since it only involves the largest Voronoi cell.
The coverage performance, on the other hand, is a measure of the average
quality of the deployment, i.e. a good proxy for system-level performance.
Both metrics take only positive values lower than unity, and where the unity
corresponds to an ideal coverage.
\subsection{Ideal Homogeneous Swarm}
We first consider a swarm of $N=20$ agents following the dynamics given by
Eq.~\eqref{eq:geofencing} with a certain speed $v_0$ identical for all the
agents. We then study the responsiveness---a proxy to assess the flexible
character of a swarm~\cite{mateo17:_effec_correl_swarm_collec_respon}---of
the system when covering the surface~\eqref{eq:area} changing at a frequency
$\omega$ such that $\alpha(t) = 1 - \cos\omega t$.
The particular values of $R_0$ and $v_0$ are arbitrary, as the behavior of the system only depends on the normalized frequency $\bar{\omega} = \omega R_0 / v_0$.
The collective response of a system following~\eqref{eq:geofencing} is shown in Fig.~\ref{fig:homo_ideal}.
The obtained response, measured either with the tessellation or coverage metrics, can be fit to the form
\begin{equation}
P(\bar{\omega}) = \frac{P_0\bar{\omega}_c^\lambda + P_\infty\bar{\omega}^\lambda}{\bar{\omega}_c^\lambda+\bar{\omega}^\lambda},
\label{eq:ideal_performance}
\end{equation}
where $P_0$ and $P_\infty$ are the limit performances for $\omega\rightarrow0$ and $\omega\rightarrow\infty$ respectively, $\bar{\omega}_c$ is the ``cutoff frequency'' of the system, and the exponent $\lambda$ measures how steep the transition from $P_0$ to $P_\infty$ is.
The values for these parameters obtained by fitting \eqref{eq:ideal_performance} to the results in Fig.~\ref{fig:homo_ideal} are given in Table~\ref{tab:params}.
\begin{figure}[h!]
\centering
\includegraphics{figs/homo_performance}
\caption{Frequency response of a swarm of $N=20$ agents in dynamic area coverage according to its Tessellation performance $P_T$ (top) and its Coverage performance $P_C$(bottom).
The different markers correspond to calculations with different agent speeds $v_0$.
The dashed line corresponds to the ideal performance of (\ref{eq:ideal_performance}).
}\label{fig:homo_ideal}
\end{figure}
\begin{table}[htbp]
\caption{Parameters of the ideal performance~\eqref{eq:ideal_performance} obtained by fitting the theoretical response of an ideal homogeneous system, measured both as $P_T$ and $P_C$.}
\begin{center}
\begin{tabular}{rll}
& Tessellation ($P_T$) & Coverage ($P_C$)\\
\hline
$P_0$ & $0.758$ & $0.700$ \\
$P_\infty$ & $0.530$ & $0.526$ \\
$\bar{\omega}_c$ & $1.03$ & $300.44$ \\
$\lambda$ & $1.37$ & $0.48$ \\
\end{tabular}
\label{tab:params}
\end{center}
\end{table}
The cutoff frequency for tessellation performance is $\omega_c \simeq v_0 / R_0$.
This is revealing that the capacity of the system to maintain a uniform configuration goes down as the boundaries of the target area move too fast for a single agent to follow them.
In contrast, the cutoff frequency for the coverage performance is two orders
of magnitude larger, $\omega_c \simeq 300 v_0 / R_0$, thus showing that the
collective maintains the capacity to cover an area well beyond the individual
agents' limitations, an expected and sought feature of swarming systems.
\subsection{Ideal Heterogeneous Swarm}
Next, we consider the case where the agents are not identical, and a fraction $\rho_F$ of agents move at speed $v_F = 2v_0$ while the rest move at $v_0$.
The performance of the system could in principle have a complicated dependency on $\rho_F$, $v_0$, and $v_F$.
However, we observe that as long as the two speeds differ in approximately less than one order of magnitude, the performance only depends on these parameters through the mean speed of the collective,
\begin{equation}
\langle v \rangle = (1-\rho_F)v_0 + \rho_F v_F .
\end{equation}
The results are identical to the homogeneous case if one makes the substitution $v_0 \rightarrow \langle v \rangle$, see Fig.~\ref{fig:hetero_ideal}.
Therefore, a heterogeneous swarm with a fraction $\rho_F$ of agents moving twice as fast as the rest will have a cutoff frequency of the form
\begin{equation}
\omega_c(\rho_F) = (1+\rho_F) \omega_c(\rho_F=0) \, .
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics{figs/hetero_performance}
\caption{Frequency response of a heterogeneous swarm of $N=20$ agents where a fraction $\rho_F$ of them move twice as fast as the rest, measured by its Tessellation performance (top) and Coverage performance (bottom).
The different markers correspond to calculations with different fractions of fast agents.
Note that the frequency is normalized using the average speed, which depends on $\rho_F$.}\label{fig:hetero_ideal}
\end{figure}
\section{Experimental Results}
We have performed a series of field tests of dynamic area coverage using
$N=22$ buoys deployed in an uncontrolled environment with no supporting
infrastructure (Bedok Reservoir in Singapore).
For each test, the collective is tasked with covering the dynamic area defined by~\eqref{eq:area} with an $\alpha(t)$ oscillating at a certain frequency $\omega$.
Given the parameters of the test ($R_0 = 25$~m and $v_0 \simeq 0.5$~m/s), the system is expected to have a collective cutoff frequency of approximately $\omega_c = 0.02~s^{-1}$.
This means that, in general, one expects the swarm of buoys to be able to respond to changes in coverage at a time scale of approximately 1 minute when covering areas of the order of $1000~m^2$.
The evolution of the Coverage performance $P_C$, and Tessellation performance $P_T$ for one of these experiments is presented in Fig.~\ref{fig:experiment_evolution}.
\begin{figure}
\includegraphics[width=\linewidth]{figs/experiment_evolution}
\caption{Evolution of the two metrics for collective response $P_T$ and $P_C$ during a field experiment. The top panels show the distribution of the 18 \bobold{} buoys (black dots) and 4 \bobnew{} ones (red triangles), along with the post-processed tessellation (top) and coverage (bottom).
The frequency of the target area oscillations is $\omega = 0.02$~s$^{-1}$, close to the theoretically predicted cutoff frequency of the collective.}
\label{fig:experiment_evolution}
\end{figure}
As the core of this study is about heterogeneous swarming, we have also
carried out a series of tests to study the effect of replacing a small portion
($\rho_F=4/22$) of the $\bobold{}$ buoys by the faster and improved
model---design \bobnew{} presented in Sec.~\ref{sec:fancy_buoy}.
The theoretical prediction (see Fig.~\ref{fig:hetero_ideal}) using an
idealized model is that this heterogeneous system would display a similar
performance to the homogeneous case but with the frequencies shifted by
$\omega \rightarrow (1+\rho_F)\omega = 1.18 \omega$.
The experimental results for a total of 7 field tests are presented in Fig.~\ref{fig:experiment}.
The overall performance of the homogeneous system, measured either by the
tessellation or coverage metrics, is about $25\%$ lower than the ideal case.
This is a consequence of {\it (i)} the real dynamics and controllability of
the buoys, {\it (ii)} the precision and accuracy of the GPS-based
localization, and {\it (iii)} the finite communication rate between buoys as
part of the distributed communication network.
All these factors contribute to some extent, to the fact that the spatial
distribution of the buoys is not as uniform as the cooperative control algorithm allows for.
\begin{figure}
\includegraphics{figs/experimental_response}
\caption{Experimental frequency response of a swarm of buoys tasked with dynamic area coverage.
The homogeneous system is composed of 22 \bobold{} buoys while the heterogeneous case has 18 \bobold{} and 4 \bobnew{}.}
\label{fig:experiment}
\end{figure}
Introducing heterogeneity in the real swarming system yields interesting
results regarding its collective response (see
Fig.~\ref{fig:experiment}).
Instead of slightly improving both performance
metrics as in the ideal case (see Sec.~\ref{sec:theory}), what we observe
is that a small proportion ($\rho_F=4/22\simeq 18 \%$) of faster buoys
dramatically improves the coverage performance of the system while having a
negligible effect on the tessellation performance.
As discussed before, the tessellation metric measures the performance of the
system at the location where the deployment is the {\it least} uniform, and it
is therefore primarily sensitive to the worst-performing section of the system.
The coverage metric, in contrast, takes into account the deployment of all
buoys and thus measures the system-level performance of this swarming system.
These experimental observations reveal how the heterogeneity affects the operations of the system.
Since the agents form a regular lattice in the target area and have a relatively small mobility within that lattice, the faster buoys (\bobnew{}) can only use their higher speed to improve the deployment in the vicinity of their location in the lattice.
This improvement always translates into an improved average performance, and thus an improved coverage.
However, if the worst-performing agent happens to be far from the \bobnew{} buoys---e.g. in the snapshots of Fig.~\ref{fig:experiment_evolution}, these cannot affect its performance, and thus the tessellation performance remains virtually the same as in the heterogeneous case.
\section{Discussion}
The efficient collective operation of heterogeneous swarms has been identified as one of the key challenges in robotics~\cite{Yangeaar7650}.
In general, such a swarm can be composed of a large number of different
classes of agents where some may be specialized in specific tasks such as
sensing, communications, or (sensed) data processing.
However, with complex systems yielding emergent behaviors, it is not obvious
that such individual enhancements of a fraction of the agents automatically translate into an increased collective performance.
In this work, we study the effect of increasing the motor capabilities of
18\% of the agents on the responsiveness of the group when performing a dynamic area monitoring in real-world and unstructured water environments.
Experimentally, we observe that the timescale at which the system is able to respond is in good agreement with the theoretical prediction, confirming that this distributed smart sensor array of buoys is capable of responding to changing environments on the order of the minute, which is well beyond what is needed to track morphological changes in oil spills, algal blooms, or other surface contaminants.
We also present empirical evidence that the partial upgrade of the system
improves the responsiveness of the system when using a cooperative control
algorithm designed for homogeneous systems and that does not explicitly take
into account the different motor capabilities of the heterogeneous set of agents.
We measure the performance of the system with two metrics, the tessellation $P_T$ and the coverage $P_C$ performances.
Since $P_T$ only takes into account the agent with the largest area to cover, it is an individual measure of the ``weakest link'' in the system.
As such, it is indicative of the robustness of the collective.
$P_C$, on the other hand, measures an average collective performance that is less sensitive to any individual agent's behavior.
While the partial upgrade makes the system more capable of dynamically covering a target area on average ($P_C$), it does not improve the coverage of the least-covered regions ($P_T$).
In order to magnify the effect of the new and improved agents, the system
would need to operate taking into account the specific properties of the
different agents. For instance, a further improvement in the collective
responsiveness of this heterogeneous swarm could be achieved by developing a new cooperative control strategy
that positions the faster agents in the most critical areas, or that
implements an optimization procedure constrained by the motor capabilities of
the different agents.
Such a procedure, however, could prove difficult to maintain as it is not robust to hardware design changes: as more kinds of agents are added during the life-cycle of the collective, the control algorithm should grow in complexity to accommodate these changes.
\section*{Acknowledgments}
This work was supported by Grants from the SUTD-MIT International Design
Center (IDC) and the Singapore Ministry of Education (MOE-Tier 1 Grant \#T1MOE1701).
We are grateful for the assistance of B. Patel, H. Shekhar, S. Jain, and P. Rastogi in preparing and performing the field tests described in this work.
\input{hetero.bbl}
\end{document}
|
\section{Introduction}
Because the most massive stars end their lives as black holes (BHs), stellar mass BHs should exist ubiquitously: the stellar mass function suggests that about $10^8$ BHs exist in the Galaxy \citep[e.g.,][]{1994ApJ...423..659B}, and the nearest ones are expected to be within $\mathcal{O}(10)\,\mathrm{pc}$ \citep{1983bhwd.book.....S, 2003ApJ...596..437C}. Nevertheless, only a part of them have been probed via X-ray/radio emissions from interacting binaries with stellar companions (X-ray binaries) or from pulsars, which are presumably observable only in a fraction of the systems' lifetime and/or the parameter space of such binaries.
The nearest known BH in an X-ray binary is about $1\,\mathrm{kpc}$ away \citep{2010ApJ...710.1127C, 2018arXiv180411349G}, and we likely miss many nearby BHs.
A larger population of stellar mass BHs can be probed if we have a means to search for the quiescent systems with stellar companions on wider, \textit{detached} orbits.
They do not only help completing the census of nearby compact objects, but the visible companions allow for reliable measurements of BH mass and kinematics of the system in the Galaxy, which are both direct probes of the mass loss and kick during the supernova (SN) explosion \citep[e.g.,][]{2017hsn..book.1499C}. If they are in tight orbits, we may also learn how the outcomes of binary interactions depend on the systems' property. In this aspect, elemental abundance of the stellar companion also helps to probe the signature of mass exchange. The information will be useful to understand the formation of compact objects and those in close binaries, such as X-ray binaries \citep[e.g.,][]{2006ARA&A..44...49R} and merging BH binaries
\citep[e.g.,][]{2016PhRvL.116f1102A, 2016PhRvX...6d1015A}.
Detached BH companions of normal stars can be searched using the same techniques to identify unresolved binaries. The spectroscopic (i.e., radial velocity) search has been considered since 1960s \citep[e.g.,][]{1966SvA....10..251G}, and has recently identified massive, yet dark companions to stars both in a cluster \citep{2018MNRAS.475L..15G} and in the field \citep{2018arXiv180602751T}, whose minimum masses imply that they are BHs or massive neutron stars (NSs). The potential of \textit{Gaia}\ astrometry has also been discussed extensively \citep{2017MNRAS.470.2611M, 2017ApJ...850L..13B, 2018ApJ...861...21Y, 2018MNRAS.481..930Y}, and hundreds or thousands of BHs may be found by the end of its five-year mission. The typical targets will be $\sim10\,M_\odot$ BHs in au-scale binaries.
This paper focuses on all-sky photometry as another means to search for stars with BH/NS companions: we especially consider the potential of the {\it Transiting Exoplanet Survey Satellite} \citep[\textit{TESS},][]{2014SPIE.9143E..20R} to identify and characterize such binaries on tight orbits ($\lesssim0.3\,\mathrm{au}$). To achieve its main science goal to find transiting exoplanets around nearby stars,
\textit{TESS}\ will provide photometric light curves with sub-precent precision and at least 27-day long, for $>10^7$ stars in the almost entire sky \citep{2015ApJ...809...77S}.
The number could be sufficiently large to find such rare binaries with compact objects, as shown below. We discuss two methods that have successfully identified white dwarf (WD) companions to normal stars in the photometric data from the \textit{Kepler}\ mission \citep{2009Sci...325..709B}: (i) periodic brightening due to in-eclipse microlensing known as ``self-lensing" \citep{2014Sci...344..275K, 2018AJ....155..144K, 2019arXiv190707656M}, and (ii) phase-curve modulation due to ellipsoidal variations and Doppler beaming \citep[e.g.,][]{2007ApJ...670.1326Z, 2010A&A...521L..59M, 2015ApJ...815...26F}.
Those BH--star binaries, if detected with \textit{TESS}, would necessarily be nearby systems amenable to various follow-up observations,
with their short-period and repeating signals being ideal for detailed characterization. In particular, the self-lensing BHs, if detected, provide unambiguous evidence for their compact nature, which is otherwise difficult to confirm.
They are also complementary to the BHs detectable with \textit{Gaia}: they have shorter orbital periods and will be a more sensitive probe of the post-interaction systems that likely followed similar formation paths to X-ray binaries or merging BHs.
In the following, we estimate how many \textit{TESS}\ stars will be searchable for BH/NS companions with given masses and orbital periods (Section \ref{sec:searchable}). This information will then be combined with simple population models of BH--star binaries to estimate the number of detectable BHs around those searchable stars (Section \ref{sec:yields}). The estimated yields are checked against the known population of X-ray binaries in Section \ref{sec:xb}.
We summarize and discuss future prospects in Section \ref{sec:summary}.
\section{The Expected Signals}\label{sec:signals}
The photometric signal from detached BH--star binaries consists of the following (up to) three periodic components:
\begin{enumerate}
\item ellipsoidal variations (EVs): tidal force due to BH's gravity changes the geometric shape of the star as well as brightness distribution on the stellar surface,
\item Doppler beaming: relativistic aberration of light, time dilation, and Doppler shift caused by the star's orbital motion produces the light variation proportional to its radial velocity toward the observer, and
\item self-lensing: when the BH eclipses the star, the BH acts as a lens to gravitationally magnify the star.
\end{enumerate}
The former two produce variability in phase with the orbital motion \citep[``phase-curve" modulation; see, e.g.,][]{2017PASP..129g2001S}, while the self-lensing causes pulse-like brightening only during the eclipse. Thus the self-lensing signal has a timescale shorter than the phase-curve variation by a factor of $R_\star/a$, where $R_\star$ is the star's radius and $a$ is the orbital semi-major axis, and their detectabilities can be discussed separately.
The amplitudes of all three signals are determined once the binary period $P$, BH mass $M_\bullet$, star's mass $M_\star$ and radius $R_\star$, and orbital inclination $i$ are specified (see Section \ref{ssec:signals_formula} for quantitative details). Thus the periodicity of the detected signal, combined with the prior knowledge about the star's mass and radius, gives a handle on the companion's mass to select candidate BHs and NSs. For the phase-curve signals, this practically reduces to searching for companions more massive than the visible ``primary" star: the absence of light from the apparently more massive companion suggests its compactness, which will need to be confirmed with follow-up spectroscopy. For the self-lensing signal, on the other hand, the compactness of the companion is indisputable.\footnote{Even the null detection of spectral features of a massive companion can sometimes be ambiguous. The stellar luminosity is not always a monotonic function of mass, and/or the kinematically detected companion may be a close pair of two stars whose total luminosity is smaller than expected for a single star with the same total mass.} Especially, if the pulses with periods $\lesssim10\,\mathrm{days}$ are detected, the companion is most likely a BH or NS, because WDs on short-period orbits do not usually exhibit self-lensing pulses because of their larger physical radii (e.g., \citealt{2003ApJ...584.1042S}, figure 8 of \citealt{2018AJ....155..144K}). The self-lensing systems are also the best targets to measure BH masses, because of the known orbital inclination and clear physical relation between the pulse height and companion mass (see Section \ref{ssec:signals_formula} below).
Figure \ref{fig:amp_period} plots the amplitudes of each component against the orbital period, for a stellar companion with $M_\star=1\,M_\odot$ and $R_\star=1\,R_\odot$. The EV signal (proportional to $a^{-3}$) is strongest at $P\lesssim1\,\mathrm{day}$, while at longer periods the beaming or self-lensing signal dominates. These features are shown in the model light curves for the edge-on case in Figure \ref{fig:lc_period}. When the orbit is not nearly edge-on, the central pulse due to self-lensing vanishes.
\begin{figure*}
\epsscale{0.85}
\plotone{signal_period.pdf}
\caption{Amplitudes of the EV, beaming, and self-lensing signals versus orbital period for different BH masses. The stellar companion is assumed to be a Sun-like star. \label{fig:amp_period}
}
\end{figure*}
\begin{figure*}
\epsscale{0.545}
\plotone{lc_p0_5_m10_0_cosi0_00.pdf}
\plotone{lc_p3_0_m10_0_cosi0_00.pdf}
\plotone{lc_p10_0_m10_0_cosi0_00.pdf}
\caption{Model light curves ($30$-minute cadence) of detached BH--star systems with different orbital periods: $0.5\,\mathrm{days}$ (\textit{top left}), $3\,\mathrm{days}$ (\textit{top right}), and $10\,\mathrm{days}$ (\textit{bottom}). The primary is a Sun-like star, the BH mass is $10\,M_\odot$, and the orbit is assumed to be circular and edge-on. When the orbital inclination $i$ is far from $\pi/2$, the ``self-lensing" pulse at time $0$ vanishes.
\label{fig:lc_period}
}
\end{figure*}
When only the phase-curve variation is detected, a measurement of the spectroscopic orbit is most likely required to exclude other astrophysical sources of variability including pulsations or star spots.
While the visible star's location on the Hertzsprung--Russell diagram as available from the \textit{Gaia}\ data \citep{2018arXiv180409382G} will be helpful to check if the star can be such a pulsator that mimics the detected modulation, star-spot modulations are likely more difficult to distinguish from the phase-curve signal. Because spotted stars are much more common than the stars with BH/NS companions, they could cause too many false positives that make the search based on phase-curve signals to be impractical. Even in the actual BH/NS systems, the star-spot modulations can severely interfere with the phase-curve signals: if the stellar rotation is tidally synchronized with the orbit of a close BH/NS companion, the spot modulations have the same periods as the phase-curve signals, and their activities can even be enhanced due to rapid stellar rotation associated with short orbital periods.
We will examine these issues in Section \ref{sec:searchable}, and find that sufficiently strong phase-curve signals could still be distinguished from spot modulations and serve as a reasonable means to select promising candidates, although false positive rates could be as high as $\sim50\%$.
If both the EV and beaming signals are detected, the consistency of the relative phase and amplitude of the signals, as well as their coherence, will also make them easier to distinguish from stellar activities. Although we assume circular binary orbits throughout this paper, BH--star systems on highly eccentric orbits, if exist, are also good targets because the resulting ``heartbeat" light curve \citep{2012ApJ...753...86T} is easier to distinguish from other astrophysical sources of variability. The shape of the light curve may even yield the orbital inclination to break the degeneracy in the spectroscopic mass \citep{1995ApJ...449..294K}. In our population models in Section \ref{sec:yields}, tides do not necessarily circularize the orbit for $P\gtrsim$ a few days, if the orbit is eccentric after the formation of a BH.
\subsection{Quantitative Descriptions of the Signals}\label{ssec:signals_formula}
\begin{description}
\item[Ellipsoidal variation] Tidal force due to BH's gravity changes the geometric shape of the star as well as brightness distribution on the stellar surface via gravity darkening \citep{1924MNRAS..84..665V}. This causes the light modulation with an amplitude of
\begin{equation}
\label{eq:ev}
s_{\rm ev}=\alpha_{\rm ev}\,{M_\bullet \sin i \over M_\star} \left(R_\star \over a\right)^3 \sin i
=1.89\times10^{-2}\alpha_{\rm ev}\sin^2i\,\left({P \over 1\,\mathrm{day}}\right)^{-2}
\left({\rho_\star \over 1\,\mathrm{g\,cm^{-3}}}\right)^{-1}\left(1\over 1+M_\star/M_\bullet\right).
\end{equation}
Here
\begin{equation}
\alpha_{\rm ev}=0.15\,{(15+u)(1+g) \over 3-u},
\end{equation}
where $g$ is the gravity-darkening coefficient and $u$ is the linear limb-darkening coefficient \citep{1993ApJ...419..344M}.
For the circular orbit, the modulation is symmetric with respect to our line of sight, and the signal has the period half the orbital one. In this paper we neglect the dependence on the eccentricity and set $\alpha_{\rm ev}=1$ just for simplicity, since our purpose is not to give a precise prediction for the expected yield.
Figure \ref{fig:amp_period} shows that, at a fixed orbital period, the amplitude of the EV signal saturates with increasing companion masses. This is because $a^3 \sim M_\bullet$ when $M_\bullet\ggM_\star$ and this dependence cancels the companion mass dependence of the tidal force (Eqn. \ref{eq:ev}). Hence it will be difficult to precisely weigh the most massive companions with the EV amplitude alone.
\item[Doppler beaming] Relativistic aberration of light, time dilation, and Doppler shift caused by the primary's radial acceleration produces the photometric variation with the amplitude of \citep{2003ApJ...588L.117L}
\begin{equation}
\label{eq:beam}
s_{\rm beam}=\alpha_{\rm beam}\,4\,{K_\star \over c}
=2.8\times10^{-3}\alpha_{\rm beam}\sin i\left({P \over 1\,\mathrm{day}}\right)^{-1/3}
\left( M_\bullet+M_\star \over M_\odot \right)^{-2/3}\left(M_\bullet \over M_\odot\right),
\end{equation}
where $\alpha_{\rm beam}$ is given by integrating the following wavelength-dependent factor
\begin{equation}
\alpha_{\rm beam,\nu} ={1\over4}\left(3-{{\mathrm{d}\log F_\nu}\over{\mathrm{d}\log \nu}}\right)
\end{equation}
over the photometric bandpass, and $F_{\nu}$ is the spectrum of the star. Again we set $\alpha_{\rm beam}=1$, which is reasonable for a Sun-like star \citep[e.g.,][]{2017PASP..129g2001S}.
This signal is anti-phased with the radial velocity variation of the star and has a different orbital-phase dependence from the ellipsoidal variation (see Figure \ref{fig:lc_period}).
Thus, if both the EV and beaming signals are detected, the consistency of their phases and amplitudes will be useful to confirm the binary origin. An independent measurement of the spectroscopic orbit also helps the interpretation of the light curve because the shape and amplitude of the beaming component are essentially fixed.
\item[Self lensing] If the system is eclipsing, the BH passing in front of the stellar disk acts as a lens to gravitationally magnify the background star \citep[e.g.,][]{1995ApJ...446..541G, 2003ApJ...584.1042S, 2011MNRAS.410..912R}. The light curve exhibits periodic pulses with the height \citep{1969ApJ...156.1013T, 1973A&A....26..215M, 2003ApJ...594..449A}:
\begin{equation}
\label{eq:sl}
s_{\rm sl}=2\left(R_{\rm E} \over R_\star\right)^2
=7.15\times10^{-5}\left(R_\star \over R_\odot\right)^{-2}
\left({P \over 1\,\mathrm{day}}\right)^{2/3}\left({M_\bullet \over M_\odot}\right)\left({M_\bullet+M_\star \over M_\odot}\right)^{1/3},
\end{equation}
where
\begin{equation}
R_{\rm E}=\sqrt{{4GM_\bullet a \over c^2}
=4.27\times10^{-2}R_\odot \left({P \over 1\,\mathrm{yr}}\right)^{1/3}\left({M_\bullet \over M_\odot}\right)^{1/2}\left({M_\bullet+M_\star \over M_\odot}\right)^{1/6}
\end{equation}
is the Einstein radius. The duration of the signal is
\begin{equation}
\label{eq:duration}
\tau_{\rm sl}={R_\star P \over \pi a}\cdot {\pi \over 4}
=1.8\,\mathrm{hr}\times {\pi \over 4}\left(P\over1\,\mathrm{day}\right)^{1/3}\left({M_\bullet+M_\star}\overM_\odot\right)^{-1/3}\left(R_\star\overR_\odot\right)
\end{equation}
when averaged over the impact parameter, and the signal vanishes if $\cos i>R_\star/a$ for a circular orbit. The pulses have the same period as the orbital period.
Since stellar light is also blocked by the eclipsing compact object, the pulses are not observed unless $\sqrt{2}R_{\rm E}$, which increases with the binary separation, is larger than the physical size of the object. This is why all the known self-lensing binaries containing WDs \citep{2014Sci...344..275K, 2018AJ....155..144K, 2019arXiv190707656M} have orbital periods ranging from months to years, longer than typical eclipsing systems. On the other hand, the physical radius has negligible effects for BHs and NSs, and so they always show pulses regardless of the orbital period. Depending on the detected period the WD case can be excluded by sheer presence of pulses.
If the orbital period is measured from multiple pulses, the BH mass is derived from the pulse height and the stellar radius. For the latter, the parallax information from \textit{Gaia}\ has already allowed for $10\%$-level measurements for tens of millions of stars \citep{2018A&A...616A...8A}. If combined with the spectroscopic effective temperature, the precision better than five percent can be achieved \citep{2018ApJ...866...99B}; this translates into $10\%$-precision for the BH mass, provided that the pulse height is sufficiently well constrained.
\end{description}
\section{The Number of Searchable Stars}\label{sec:searchable}
Here we estimate how many stars that will be observed by \textit{TESS}\
will enable the detection of the above photometric signals due to BH companions, if exist.
We focus on the cases where (i) the self-lensing pulse is detectable, and (ii) either the EV or beaming signals is detectable.
We choose the detection threshold so that the signal can be recovered in the presence of realistic (i.e., both correlated and uncorrelated) noise, and that the expected false positive (FP) rate is reasonably low, as will be discussed in the following subsections. In particular, we focus on the effect of enhanced stellar noise associated with rapid stellar rotation: given the tight orbits and large BH/NS masses of our interest, the rotation of our target stars can well be synchronized with the binary orbit, and stars with such short rotation periods $\lesssim10\,\mathrm{days}$ are known to exhibit enhanced activity. We evaluate this effect using actual light curves of $\approx34,000$ spotted stars from the prime \textit{Kepler}\ mission \citep{2014ApJS..211...24M}.
We also exclude the systems where the star fills its Roche lobe (to focus on non-interacting systems), as well as the systems where gravitational wave (GW) emission causes orbital decay on a timescale significantly shorter than the star's lifetime.
The number of searchable stars counted this way are shown in Figure \ref{fig:ns} for $5\times5$ log-uniform bins covering the BH/NS mass $1$--$100\,M_\odot$ and period $0.3$--$30\,\mathrm{days}$; the details of the counting are described below. The numbers are computed for the values at the bin centers, and slant dashed lines show the corresponding values of the semi-major axis assuming that the total mass is dominated by the BH. These numbers take into account the fraction of systems with suitable orbital inclinations: the self-lensing signal can be observed only if the binary orbit is nearly edge-on, and the phase-curve signals gradually become weaker as the system becomes closer to face-on (Eqns.~\ref{eq:ev} and \ref{eq:beam}). This is partly why fewer stars are searchable via self-lensing. We find that self-lensing and phase-curve signals can be used to search effectively $\sim10^5$ and $\sim10^6$ stars, respectively, for stellar companions with periods up to $\sim 10\,\mathrm{days}$. While the EV and beaming signals are mostly sensitive to the shortest-period companions, the detectability of the self-lensing signal has a weaker dependence on the orbital period, because the signal becomes stronger with increasing orbital periods (Eqn.~\ref{eq:sl}).
\begin{figure*}[ht!]
\epsscale{1.1}
\plottwo{min2_ns_lens.png}{min2_ns_phase.png}
\caption{Number of stars in the \textit{TESS}\ input catalog for which BH companions with given masses and periods are searchable with self-lensing ({\it left}) and phase-curve modulations ({\it right}). Here we count effective numbers of stars, considering the fraction of systems with suitable orbital inclinations assuming that the orbits are isotropic. Here the companion mass is extended down to $1M_\odot$ to check the searchability of NSs as well. \label{fig:ns}
}
\end{figure*}
\subsection{The Stellar Sample and Noise Property}\label{ssec:white_noise}
The main purpose of the \textit{TESS}\ mission is to find transiting planets around nearby stars. \textit{TESS}\ employs four cameras each with a field of view (FOV) of $24^{\circ}\times24^{\circ}$, which point each sector of the sky for $27.4\,\mathrm{days}$ (two spacecraft orbits). Over its two-year mission, 26 sectors with partial overlaps will be observed to tile the almost entire sky excluding the region near the ecliptic; the region near the ecliptic poles will be most densely covered and continuously observed for one year. \textit{TESS}\ will provide images stacked at two-minute cadence for a few $10^5$ targets pre-selected for planet search and asteroseismology, as well as 30-minute cadence full-frame images for all sources.
To enable the selection of optimal targets for the planet search, a catalog of luminous sources on the sky, \textit{TESS}\ Input Catalog \citep[TIC,][]{2017arXiv170600495S}, has been created.
We adopt the magnitude in the \textit{TESS}\ bandpass, stellar mass $M_\star$, stellar radius $R_\star$, and stellar effective temperature $T_\star$ in the TIC version 6 \citep{2017arXiv170600495S}. We focus on the objects classified as stars, on ecliptic latitude $|l|>6^\circ$ (i.e., outside the gap in the \textit{TESS}\ FOV), and with both mass and radius estimated in the catalog. The last selection limits the targets from $\sim400$ million to $\sim20$ million low-mass stars (mostly $<2.5\,M_\odot$) with \textit{TESS}\ magnitudes brighter than $15$. We do not consider this to be a serious limitation because the quality of the \textit{TESS}\ photometry will be limited for fainter stars in any case.
We use the fitting formula in \citet[][ver.~20170531]{2017arXiv170600495S}
to evaluate the white component of
the expected photometric noise as a function of the \textit{TESS}\ magnitude presented in \citet{2015ApJ...809...77S}.
We adopt the observing duration $T=27.4\,\mathrm{days}$ (i.e., minimum observing duration) for all the stars to give a conservative estimate.
We always assume $30\mathchar`-\mathrm{min}$ cadence observations to deal with all the stars that will be in the full-frame images.
\subsection{Searchability for Self-lensing Signals}\label{ssec:searchable_sl}
If the stellar rotation is synchronized with the orbit, the star may show enhanced spot activity with a similar period to that of the self-lensing pulses. Since the duration of self-lensing pulses is much shorter than the orbital period (and hence rotation period of the spin-synchronized star), the rotational modulation can basically be filtered out without affecting the self-lensing signal. However, this removal may not be perfect, and also stellar variabilities may enhance correlated noise on a timescale similar to the duration of self-lensing events. Thus, estimates considering the uncorrelated noise alone (as described in Section \ref{ssec:white_noise}) most likely overestimate the detectability of the self-lensing signal.
To evaluate the detectability in the presence of such effects, we perform injection-and-recovery simulations using the actual light curves of spotted stars from the prime \textit{Kepler}\ mission. Because \textit{Kepler}\ has a $\sim$10x larger aperture than \textit{TESS}, the \textit{Kepler}\ light curves include much smaller photon noise and can be considered as a realistic template of astrophysical variability of the stars. Therefore we can simulate the light curves of spotted stars observed with \textit{TESS}\ by adding white noise component due to its smaller aperture (as described in Section \ref{ssec:white_noise}) to the actual \textit{Kepler}\ light curves.\footnote{This may overestimate the noise for the brightest stars observed by \textit{TESS}, but those stars constitute only a small fraction and would not affect the detectability of such rare systems as BH/NS binaries.} Here we assume that the stellar rotation is always synchronized with the orbit of the self-lensing systems, and
inject self-lensing signals with the same period as the rotation period of each \textit{Kepler}\ star.
We take random segments of $27.4\,\mathrm{day}$ long from the \textit{Kepler}\ light curves of rotating stars and detrend them with median filters with the length of $0.25\times(P/30)^{1/3}\,\mathrm{days}$, which is typically a few times longer than self-lensing durations and much shorter than rotation period. Then we run the box least-squares (BLS) algorithm\footnote{We used the python implementation {\tt bls.py} by Daniel Foreman-Mackey: {\url https://github.com/dfm/bls.py}.} \citep{2002A&A...391..369K} on these light curves, after randomly injecting Gaussian noise with log-uniform dispersions spanning from 100 to 10,000 ppm (to cover the expected noise range from \textit{TESS}) and flipping the light curves around their median values (to evaluate any possible source of positive excursion). We then set the threshold value for the peak signal-to-noise of the BLS spectrum, so that no detection is found from $10^6$ trials to achieve false-positive rate less than $10^{-6}$. Given that $\sim10\%$ Sun-like stars exhibit spot modulation detectable with \textit{Kepler}\ \citep{2014ApJS..211...24M}, this yields the FP rate of $\sim10^{-7}$, which corresponds to $\mathcal{O}(1)$ FP in our sample of $20$M stars. Considering the strong dependence of the spot activity on the rotation period, we choose different thresholds for each period bin, and the resulting thresholds turn out to be higher for shorter-period systems. We also count the BLS peaks as detections only if the detected period is close to the measured rotation period of the stars.\footnote{Here we implicitly assume that the rotation period of the star showing spot modulations can be measured well: this assumption may not hold for longer-period binaries in the 27.4-day data, but the detectability turns out to be low for such systems in any case.}
Then we perform the simulations injecting both Gaussian noise and self-lensing pulses of various amplitudes (100--$10^5$ppm; see Figure \ref{fig:amp_period}) and phases, and check what fraction of the injected signals are recovered using the thresholds as defined above. We fit the resulting recovery rate as a function of the signal-to-noise of the phase-folded signal $\sqrt{T/P}(s_{\rm sl}/\sigma_{\rm 30min})$ using the gamma cumulative distribution function \citep{2017AJ....154..109F} scaled by the asymptotic value of the recovery rate at high signal-to-noise ratios: the values were $\approx100\%$ for $P=0.8$--$4.8\,\mathrm{days}$, $\approx90\%$ for $P=0.3$--$0.8\,\mathrm{days}$ and $P=4.8$--$11.9\,\mathrm{days}$, and $\approx20\%$ for $P=11.9$--$30\,\mathrm{days}$. The resulting function $C_{\rm rec}(\sqrt{T/P}\,(s_{\rm sl}/\sigma_{\rm 30min}))$ is used to evaluate the expected number of searchable stars.
For a star with mass $M_\star^j$ and radius $R_\star^j$, the effective searchability for a BH/NS companion with period $P$ and mass $M_\bullet$ on randomly oriented orbits is evaluated as:
\begin{equation}
\label{eq:neff_sl}
N_{\rm eff}(P, M_\bullet, M_\star^j, R_\star^j)={R_\star \over a}\,C_{\rm rec}\left(\sqrt{n}\,{s_{\rm sl}\over \sigma_{\rm 30min}}\right),
\end{equation}
where $R_\star/a$ corresponds to the eclipse probability for a circular orbit, and $n$ is the number of pulses in the data. The number $n$ can be either $n_0\equiv[T/P]$ or $n_0+1=[T/P]+1$ depending on the orbital phase, where $[x]$ is the greatest integer that does not exceed $x$. The corresponding probabilities are $p_{n_0}=[T/P]+1-T/P$ and $p_{n_0+1}=T/P-[T/P]$, respectively. We count the number of searchable stars for both $n_0$ and $n_0+1$, and average them with the weights $p_{n_0}$ and $p_{n_0+1}$. We set $p_{n_0=1}=0$ to count only the pulses that are observed at least twice, although such longer-period systems turned out to be difficult to search in any case.
\subsection{Searchability for Phase-curve Signals}
The interference with stellar activity is a more serious issue for the phase-curve signal, because the star-spot modulation includes signals with both periods $P/2$ and $P$, if the rotation is synchronized with the orbit. This means that simple periodicity search will introduce a large number of FPs from spot modulations, because there exist more spotted stars than the BH/NS systems. Here we consider a search focusing on EV/beaming signals that are likely too large to be produced by a spotted star with a given rotation period and mass, and examine if the search could provide reasonable number of candidates without being completely swamped by false positives.
To set the threshold, we compute Fourier transform of the \textit{Kepler}\ light curves of all the spotted stars in \citet{2014ApJS..211...24M} and measure the amplitudes of the peaks corresponding to the rotation period $s_P$ and its first harmonic $s_{P/2}$. We then divide the sample into the period bins as we used for orbital periods, as well as temperature bins ($<4000\,\mathrm{K}$, 4000--5000$\,\mathrm{K}$, 5000--6000$\,\mathrm{K}$, $>6000\,\mathrm{K}$), and take the maximum values of $s_P$ and $s_{P/2}$ in each bin. We set the threshold for the EV and beaming signals, $s_{\rm ev,th}$ and $s_{\rm beam,th}$, so that the Fourier amplitudes from the EV and beaming signals are larger than $s_{P/2}$ and $s_P$, respectively. This allows us to pick up phase-curve signals that are likely too strong to be caused by spot modulations, taking into account the dependence of their amplitudes on the stellar rotation period and effective temperature.
These thresholds are much larger than the noise level and statistical false positives are negligible. Thus, assuming that Sun-like stars observed by \textit{TESS}\ have similar properties as the \textit{Kepler}\ sample, the FP rate due to spotted stars would roughly be less than one in $10^5$. The rate corresponds to some $100$ FPs in our 20M sample stars.\footnote{
Since \textit{Kepler}\ has observed only $10^5$ stars, we cannot estimate the threshold that guarantees a lower FP rate with this method.
}
Although this number of possible FPs is not small, if a similar (or larger) number of BH/NS candidates can be detected with this choice of thresholds, the search would still provide meaningful information to select promising candidates, especially considering that further vetting (e.g., presence or absence of any RV variations) would readily be possible for many stars using the archival spectroscopic data \citep[e.g.,][]{2018arXiv180602751T, 2019ApJ...872L..20G}. We will see that this can be the case in Section \ref{ssec:yields_results}.
Following Eqns.~\ref{eq:ev} and \ref{eq:beam}, the above detection criteria, $s_{\rm ev}>s_{\rm ev,th}$ and $s_{\rm beam}>s_{\rm beam,th}$ translate into
\begin{equation}
1-\cos^2 i > {s_{\rm ev,th} \over s_{\mathrm{ev},\cos i=0}} \equiv \beta_\mathrm{ev},
\quad
\sqrt{1-\cos^2 i } > {s_{\rm beam,th} \over s_{\mathrm{beam},\cos i=0}} \equiv \beta_\mathrm{beam},
\end{equation}
where the subscript ${\cos i=0}$ refers to the amplitude for a system with $\cos i=0$.
Thus, averaging over random orbital inclination (i.e., uniform $\cos i$), the effective searchability of each star is given by $H(1-\beta_\mathrm{ev})\sqrt{1-\beta_\mathrm{ev}}$ for the EV signal and $H(1-\beta_\mathrm{beam})\sqrt{1-\beta_\mathrm{beam}^2}$ for the beaming signal, where $H$ is the Heaviside step function ($H(x)=1$ for $x>0$ and $0$ otherwise). Considering that either of the signal, if above the threshold, would be sufficient to flag the system to be anomalous, we take the larger of the two values to evaluate the searchability of each star:
\begin{equation}
\label{eq:neff_phase}
N_{\rm eff}(P, M_\bullet, M_\star^j, R_\star^j, T_\star^j)=\mathrm{max}\left[ H(1-\beta_{\rm ev})\sqrt{1-\beta_{\rm ev}},\ H(1-\beta_{\rm beam})\sqrt{1-\beta_{\rm beam}^2} \right].
\end{equation}
\subsection{Counting Searchable Stars}
\begin{comment}
Given the signal amplitude $s$ for the set of system parameters $\theta=(P, M_\bullet, M_\star, R_\star, i)$, the above conditions \ref{eq:limit_sl} and \ref{eq:limit_phase} yield the limiting noise level for the detection, $\sigma_{\rm lim}$:
\begin{equation}
\label{eq:sigmalim}
\sigma_{\rm 30min}<\sigma_{\rm lim}(s(\theta)) \propto s(\theta).
\end{equation}
Since the inclination $i$ is random, we separate the dependence on $i$ and rewrite the above condition as (cf. Eqns.~\ref{eq:ev}, \ref{eq:beam}):
\begin{equation}
\label{eq:beta}
\beta(P, M_\bullet, M_\star, R_\star)\equiv\left({\sigma_{\rm 30min} \over \sigma_{\rm lim}|_{\cos i=0}}\right)
< \left(\sigma_{\rm lim}\over {\sigma_{\rm lim}|_{\cos i=0}}\right)
=\begin{cases}
1-\cos^2 i \quad &(\mathrm{EV})\\
\sqrt{1-\cos^2 i} \quad &(\mathrm{beaming})\\
H\left(1-|{a\overR_\star} \cos i|\right) \quad &(\sl)
\end{cases},
\end{equation}
where $\sigma_{\rm lim}|_{\cos i=0}$ is $\sigma_{\rm lim}$ for $\cos i=0$ and $H$ is the Heaviside step function ($H(x)=1$ for $x>0$ and $0$ otherwise).
Now $\beta$ is independent from $i$. If $\beta>1$, no value of $i$ satisfies the above condition so the star is not searchable; if $\beta<1$, on the other hand, a fraction of $i$ satisfies the searchability condition. Assuming isotropic orbital orientations, $p(\cos i)=1/2$, we can thus define the effective searchability of each star by:
\begin{equation}
N_{\rm eff}(P, M_\bullet, M_\star, R_\star
=H(1-\beta)\int_{\beta<{\sigma_{\rm lim}\over {\sigma_{\rm lim}|_{\cos i=0}}}} p(\cos i) \mathrm{d}\cos i
=H(1-\beta)
\begin{cases}
\sqrt{1-\beta} \quad &(\mathrm{EV})\\
\sqrt{1-\beta^2} \quad &(\mathrm{beaming})\\
{R_\star / a} \quad &(\sl)
\end{cases},
\end{equation}
which is a number between $0$ and $1$.
When we discuss the detectability of phase-curve signals, we can use the smaller of $\sqrt{1-\beta}$ and $\sqrt{1-\beta^2}$ in the last equality, with $\beta$ computed for each signal separately using Eqn.~\ref{eq:beta}.
\end{comment}
Since we focus on detached systems, we exclude the stars filling their Roche lobes. We compute the effective Roche-lobe radius for the star by \citep{1983ApJ...268..368E}
\begin{equation}
\label{eq:roche}
R_{\rm L}(q,a)={0.49q^{2/3} \over {0.6q^{2/3}+\ln(1+q^{1/3})}}\,a
\equiv \tilde R_{\rm L}(q)\,a,
\end{equation}
where $q=M_\star/M_\bullet$. We require that $R_\star<R_{\rm L}$, or $a>R_\star/\tilde R_{\rm L}$.
We also exclude the systems where the GW emission causes rapid orbital decay compared to the main-sequence lifetime of the star. We compute the decay time as $t_{\rm GW}=3.3\times10^8\,\mathrm{Gyr}\times{(a/\mathrm{AU})^4 \over M_\bulletM_\star(M_\bullet+M_\star)/M_\odot^3}$ \citep{1964PhRv..136.1224P}
and the main-sequence lifetime as $t_{\rm MS}=10\,\mathrm{Gyr}(M_\star/M_\odot)^{-2.5}$, and omit the systems with $t_{\rm GW}<t_{\rm MS}/2$.
Taking all these into account, the number of searchable stars for a set of $(P, M_\bullet)$ is computed by:
\begin{align}
N_{\rm searchable}(P, M_\bullet)= \sum_j N_{\rm eff}(P, M_\bullet, M_\star^j, R_\star^j, T_\star^j)\cdot H(a^j-R_\star^j/\tilde R^j_{\rm L})\cdot H(t_{\rm GW}^j-t_{\rm MS}^j/2),
\end{align}
where the index $j$ runs over all the stars and $N_{\rm eff}$ is computed for each of the self-lensing signal (Eqn.~\ref{eq:neff_sl}) and phase-curve signal (Eqn.~\ref{eq:neff_phase}). The results are shown in Figure \ref{fig:ns} for the self-lensing signal (left) and for the phase-curve signal (right).
\section{Expected Number of Detectable Black Holes}\label{sec:yields}
Here we estimate how many detectable BH companions actually exist around the above searchable stars. To do so we multiply the number of searchable stars by the intrinsic occurrence of BH companions as a function of the BH/stellar mass and binary orbital period. More specifically, given the probability density that a star with mass $M_\star$ has a BH companion with period $P$ and mass $M_\bullet$, $p(P, M_\bullet|M_\star)$, the number of detectable companions per unit $P$ and $M_\bullet$, $n_{\rm det}(P, M_\bullet)$, is given by:
\begin{equation}
n_{\rm det}(P, M_\bullet)\Delta P\DeltaM_\bullet
=\sum_j N_{\rm eff}(P, M_\bullet, M_\star^j, R_\star^j, T_\star^j) \cdot p(P, M_\bullet|M_\star^j)
\Delta P\DeltaM_\bullet,
\end{equation}
\begin{comment}
or
\begin{equation}
N_{\mathrm{bin}\,[M_{\bullet i}, P_j]}
= \int_{\mathrm{bin}\,[M_{\bullet i}, P_j]} n(M_\bullet, P)\mathrm{d}M_\bullet\mathrm{d}P
\approx
\sum_k n_{\rm eff}(M_{\bullet i}, P_j, M_\star^k, R_\star^k) \cdot p(M_{\bullet i}, P_j|M_\star^k)\,
\Delta M_\bullet \Delta P,
\end{equation}
\end{comment}
where the index $j$ runs over the searchable stars in each $\Delta P\DeltaM_\bullet $ bin.
The occurrence of short-period BH--star binaries $p(P, M_\bullet|M_\star)$ is quite uncertain both theoretically and observationally. So we adopt two simple models to provide crude estimates in the two extreme cases, with and without binary evolution. These estimates are thus not meant to be a precise prediction, but are rather to provide an order-of-magnitude sense of feasibility.
\subsection{Population Models of BH--Star Binaries}\label{ssec:yields_bhpops}
\subsubsection{Estimate Based on Field Binaries}\label{sssec:yields_bhpops_field}
Here we construct the population of BH--star binaries assuming that their properties follow those of field binaries \citep[cf.][]{2017MNRAS.470.2611M}.
We pick up ``BHs" from a power-law mass function, $I_{\rm BH}(M_\bullet)\propto M_\bullet^{-2.3}$, with a minimum mass of $5\,M_{\odot}$.
The mass function is normalized such that all the stars with $\geq 20M_{\odot}$ end up as BHs.
This gives the occurrence of BHs as:
\begin{equation}
{\d N_{\rm BH} \over \d M_\bullet}=H(M_\bullet-5\,M_\odot)\,I_{\rm BH}(M_\bullet).
\end{equation}
Then we assign companion stars to these BHs, based on the occurrence of massive star binaries in the field (e.g., \citealt{2012Sci...337..444S}):
\begin{equation}
f_{\rm companion}(q, P)\,\d q\,\d P= C\,{q^0\over P}\,\d q\,\d P, \label{eq:comp}
\end{equation}
where $q\leq1$ is the binary mass ratio, $P$ is the orbital period, and $C$ is the normalization constant that fixes the binary fraction (see below). Thus the occurrence of BHs with stellar companions is:
\begin{equation}
{\d N_\mathrm{BH\mathchar`-star} \over {\dM_\bullet\,\dM_\star\,\d P}}
={1\overM_\bullet}{\d N_\mathrm{BH\mathchar`-star} \over {\dM_\bullet\,\d q\,\d P}}
={1\overM_\bullet}\,{\d N_{\rm BH} \over \d M_\bullet}\,f_{\rm companion}\left({M_\star\overM_\bullet}, P\right)
={C\overM_\bullet}\,{\d N_{\rm BH} \over \d M_\bullet}\,{1\over P}.
\end{equation}
Here we assume circular orbits for simplicity.
This yields
\begin{equation}
\label{eq:bhocc}
p_{\rm field}(P, M_\bullet| M_\star)
={{\d N_\mathrm{BH\mathchar`-star}/{\dM_\bullet\,\dM_\star\,\d P}}\over \d N_{\rm star}/\dM_\star}
={C\overM_\bullet}\,{H(M_\bullet-5\,M_\odot)\,I_{\rm BH}(M_\bullet) \over I(M_\star)}\,{1\over P},
\end{equation}
where $\d N_{\rm star}/\dM_\star=I(M_\star)$ is the initial stellar mass function from \cite{2001MNRAS.322..231K}.
We choose $C$ so that the binary fraction integrated over $P=0.1\,\mathrm{days}$ to $P=10^{3.5}\,\mathrm{days}$ is $0.5$, following \citet{2012Sci...337..444S}. Strictly speaking, it is not clear whether the normalization remains the same for binaries containing BHs. The choice is for an optimistic scenario in which formation of a BH does not lead to any loss of binary companions.
\subsubsection{Estimate Based on a Simple Model of the Common-envelope Evolution}\label{sssec:yields_bhpops_ce}
The above estimate does not take into account any interactions in the binary. If they go through the common-envelope (CE) phase, the low-mass companion may merge with the BH progenitor. Even if it survives, the binary orbit should dramatically shrink during the process. Here we examine the outcome of this CE evolution. Unlike in \citet{2018ApJ...861...21Y}, we do not consider the case where the mass transfer occurs stably, since our focus is on systems with initially large mass ratios.
We first follow the same procedure as in Section \ref{sssec:yields_bhpops_field} to sample the population of binaries consisting of a BH progenitor with mass $M_{\bullet\mathrm{i}}$ and its stellar companion with mass $M_\star$: $p(P_{\rm i}, M_{\bullet\mathrm{i}}|M_\star)$.
Then we use the fitting formulae provided by \cite{2000MNRAS.315..543H} to compute the core mass $M_{\bullet\mathrm{i},\mathrm{core}}$ of the BH progenitor as well as its maximum radius $R_{\rm max}$ (typically $\sim 1000$--$3000R_{\odot}$) achieved during its evolution.
We then assume that all the sampled binaries with initial semi-major axes $a_{\rm i}=\left[P_{\rm i}^2G(M_{\bullet\mathrm{i}}+M_\star)/4\pi^2\right]^{2/3}$ smaller than $R_{\rm max}$ goes through the CE phase. The companion survives if the orbital energy is sufficiently large to completely strip the envelope of the BH progenitor; we assume only the stripped core of the BH progenitor is left when this happens.
The semi-major axis after this process, $a_{\rm f}$, is computed by \citep{1984ApJ...277..355W}:
\begin{equation}
\alpha \left( {GM_{\bullet\mathrm{i},\mathrm{core}}M_\star \over 2a_{\rm f}} - {GM_{\bullet\mathrm{i}}M_\star \over 2a_{\rm i}}\right)
= {G(M_{\bullet\mathrm{i}}-M_{\bullet\mathrm{i},\mathrm{core}})M_{\bullet\mathrm{i}} \over \lambda R_{\rm Roche,i}},
\end{equation}
where $R_{\rm Roche,i}=R_{\rm L}(M_{\bullet\mathrm{i}}/M_\star, a_{\rm i})$ with $R_{\rm L}$ defined in Eq.~\ref{eq:roche} and we adopt $\alpha \lambda = 1$ \citep{2002ApJ...572..407B}.
\begin{comment}
This solves to
\begin{equation}
a_{\rm f}=a_{\rm i}\left(M_{\bullet\mathrm{i},\mathrm{core}} \over M_{\bullet\mathrm{i}}\right)
\left(1+{2\over\alpha\lambda}\,{M_{\bullet\mathrm{i}}-M_{\bullet\mathrm{i},\mathrm{core}} \over M_\star}\,{a_{\rm i}\over R_{\rm L}(M_{\bullet\mathrm{i}}/M_\star, a_{\rm i})}\right)^{-1}
\sim R_{\rm L}(M_{\bullet\mathrm{i}}/M_\star, a_{\rm i}){M_\star \over M_{\bullet\mathrm{i}}-M_{\bullet\mathrm{i},\mathrm{core}}}{M_{\bullet\mathrm{i},\mathrm{core}} \over M_\bullet}
\end{equation}
for $a_{\rm i}\gg R_{\rm L}$.
or equivalently
\begin{equation}
a_{\rm f}=a_{\rm i}\left(M_{\bullet\mathrm{i},\mathrm{core}} \over M_{\bullet\mathrm{i}}\right)
\left(1+{2\over\alpha\lambda}\,{M_{\bullet\mathrm{i}}-M_{\bullet\mathrm{i},\mathrm{core}} \over M_\star}\,{1\over R_{\rm L}(M_\star/M_{\bullet\mathrm{i}})}\right)^{-1},
\end{equation}
where
\begin{equation}
a_{\rm i}=\left[P_{\rm i}^2G(M_{\bullet\mathrm{i}}+M_\star) \over 4\pi^2\right]^{2/3}, \quad
a_{\rm Roche,i}=a_{\rm i} R_{\rm L}(q_{\rm i}).
\end{equation}
In terms of the binary period $P \propto a^{3/2} M^{-1/2}$:
\begin{equation}
P_{\rm f}=P_{\rm i} \left(M_{\bullet\mathrm{i},\mathrm{core}}+M_\star \over M_{\bullet\mathrm{i}}+M_\star\right)^{-1/2}
\left(M_{\bullet\mathrm{i},\mathrm{core}} \over M_{\bullet\mathrm{i}}\right)^{3/2}
\left(1+{2\over\alpha\lambda}\,{M_{\bullet\mathrm{i}}-M_{\bullet\mathrm{i},\mathrm{core}} \over M_\star}\,{1\over R_{\rm L}(M_\star/M_{\bullet\mathrm{i}})}\right)^{-3/2}.
\end{equation}
\end{comment}
Finally, we assume that the remnant core of the BH progenitor loses mass via stellar wind and SN explosion before it becomes a BH.
For the wind mass loss during the Wolf--Rayet star phase we follow \citet{2017A&A...607L...8V}.
At the core collapse, the remnant BH mass and SN ejecta mass are determined by the formula in \cite{2002ApJ...572..407B}. Both mass loss processes change the semi-major axis.
The above procedures define a transformation between $(P_{\rm i}, M_{\bullet\mathrm{i}})$ and $(P, M_\bullet)$.
The final distribution after the evolution is then computed by:
\begin{equation}
p_{\rm CE}(P, M_\bullet|M_\star
= \left| {\partial (P_{\rm i}, M_{\bullet\mathrm{i}}) \over \partial(P, M_\bullet)}\right|\,p_{\rm field}(P_{\rm i}, M_{\bullet\mathrm{i}}| M_\star).
\end{equation}
\subsection{Results}\label{ssec:yields_results}
Figure \ref{fig:bhpops} shows the BH occurrence rates for the field binary model and the CE model with $\alpha\lambda=1$ described in Section \ref{ssec:yields_bhpops}. The values are averages of $p(P, M_\bullet|M_\star)\DeltaM_\bullet\Delta P$ over $M_\star$ of the stars searchable with self-lensing in each bin; the differences are mostly within a factor of a few when the stars searchable with phase-curve signals are used instead.
Figures \ref{fig:yield} and \ref{fig:yield_ce} show the estimated numbers of detectable BHs based on the occurrences in Figure \ref{fig:bhpops} and the number of searchable stars in Figure \ref{fig:ns}.
Both field binary and CE models predict some 10 BH companions detectable with self-lensing, and some 100 with phase-curve variations. This is the consequence that
the BH occurrence in the relevant region of the parameter space is about $10^{-4}$ (Figure \ref{fig:bhpops}),
while the effective numbers of stars searchable with self-lensing and phase-curve signals are $\sim10^5$ and $\sim10^6$, respectively.
Curiously, the estimated occurrence appears to be compatible with the discovery of a BH/NS system by \citet{2018arXiv180602751T} among $\gtrsim10^5$ stars from Apache Point Observatory Galactic Evolution Experiment \citep[APOGEE,][]{2017AJ....154...94M}, although the period we focus on is much shorter.
While the resulting occurrences are similar between the field binary and CE models, the detectable binaries from the CE models were initially in much wider orbits ($P=10^2$ to $10^4\,\mathrm{days}$) and surrendered their orbital energies to survive the CE evolution.
\begin{figure*}[ht!]
\epsscale{1.1}
\plottwo{bhprob.png}{bhprob_al1.png}
\caption{Occurrence rates of BH companions as a function of the BH mass and orbital period from ({\it left}) field-binary model and ({\it right}) CE model with $\alpha\lambda=1$. The values are computed for stellar companions searchable with self-lensing.\label{fig:bhpops}
}
\end{figure*}
\begin{figure*}[ht!]
\epsscale{1.1}
\plottwo{min2_nbh_bhprob_lens.png}{min2_nbh_bhprob_phase.png}
\caption{Expected numbers of detectable BH companions for the field binary model, using ({\it left}) self-lensing and ({\it right}) phase-curve modulation. \label{fig:yield}
}
\end{figure*}
\begin{figure*}[ht!]
\epsscale{1.1}
\plottwo{min2_nbh_al1_lens.png}{min2_nbh_al1_phase.png}
\caption{Same as Figure \ref{fig:yield}, but for the CE model with $\alpha\lambda=1$. \label{fig:yield_ce}
}
\end{figure*}
These calculations suggest that the number of searchable stars with \textit{TESS}\ may indeed be sufficient to actually detect BH/NS companions. This also indicates that even an upper limit on the occurrence rate from the null detection, if quantified, will be able to provide a meaningful observational constraint on the poorly-understood process of the CE evolution.
We also note that the searchability increases both in terms of the number and period range, if individual stars are observed for a longer duration in the overlapping regions of the observing sectors or during a potential extended mission \citep{2017arXiv170508891B}.
In these population models, the typical mass of the stars with detectable BH companions is $1$--$2\,M_\star$, mainly by construction of the sample. The brightest star with a BH companion is predicted to have a $V$-band magnitude of $\sim 10$ for the self-lensing population, and will be even brighter for the phase-curve sample. These magnitudes correspond to the distance of a few tens to a few hundreds of parsecs. If actually found, these systems will therefore be among the nearest known BHs.
\subsection{Application to the \textit{Kepler}\ Sample}
For a sanity check, we applied the same procedure to the \textit{Kepler}\ sample, using stellar properties and the combined differential photometric precision \citep{2010ApJ...713L..79K} of the stars in the \textit{Kepler}\ input catalog \citep{2017ApJS..229...30M}. Adopting the same detection thresholds, we found that no detection is expected with either self-lensing or phase-curve signals, consistently with the null detection of such systems so far. The conclusion for the self-lensing BH systems is insensitive to the adopted threshold because the null detection is expected simply due to their low occurrence rate, as pointed out by \citet{2003ApJ...594..449A}. The interpretation of the null detection of phase-curve signals is more subtle. Our population models predict that there would be some 10 such systems on close-in orbits that would have been detectable in the absence of stellar variabilities: thus either they have been swamped in (or confused with) the stellar variabilities and await confirmation with Doppler observations, or our model overestimates the number of BH systems on close-in orbits. In the latter case, the detection with \textit{TESS}\ still remains to be plausible even if the actual rate is 10 times smaller than our estimate; if the overestimation factor is 100 or larger, the detection with the \textit{TESS}\ data may not be feasible.
\vspace{0.5cm}
\section{Connections to X-ray Binaries}\label{sec:xb}
Given the lack of observational knowledge on the population of detached BH--star binaries, in Section \ref{sec:yields} we adopted simple models for the population to estimate the expected yields. On the other hand, we do have observational constraints on {\it interacting} systems, which are observed as X-ray binaries containing BHs (BHXBs). These BHXBs may be considered as the shorter-period end of the spectrum of detached systems,
which can be extrapolated to give a rough estimate on the occurrence of tight but detached systems as discussed in this paper. Here we try this more empirical approach as an independent check of the feasibility. This exercise also reveals the potential of \textit{TESS}\ to characterize X-ray binaries via optical light curves, as has routinely been performed with the ground-based photometry \citep[e.g.,][]{1978ARA&A..16..241B}.
To do this, we count the number of BHXBs where the self-lensing and phase-curve signals induced on the stellar companion would be detectable with \textit{TESS}, assuming that the system is in a quiescent phase and the system luminosity is not dominated by the accretion disk (i.e., we artificially ``turn off" interactions in these systems). We also take into account the incompleteness of the known BHXB population; those with weak X-ray emissions are likely more relevant for our estimate, but they are also more easily missed in the search for X-ray binaries.
\begin{comment}
While we have little observational knowledge on the population of detached BH--star binaries, they may be seen as analogs of low-mass X-ray binaries (XBs): if their orbits would have been slightly wider or the stellar companions would have been less evolved, or if they are in a quiescent phase, those XB systems would look similar to detached BH--star binaries. Here we estimate the number of XBs containing BHs (BHXBs) where the self-lensing pulse and phase-curve variation induced on the stellar companion would be detectable with \textit{TESS}, if the system is in a quiescent phase and the system luminosity is not dominated by the accretion disk (regardless of whether these can be true or not). Such an extrapolation from the known BHXB population provides another estimate for the number of detectable BHs in detached binaries. This exercise also reveals the potential of \textit{TESS}\ to characterize XBs themselves via optical light curves, as has routinely been performed with the ground-based photometry \citep[e.g.,][]{1978ARA&A..16..241B}.
\end{comment}
\subsection{The Sample of X-ray Binaries with Confirmed BHs}
Our sample consists of $18$ X-ray binaries with dynamically confirmed BH companions in Table 1 of \citet{2006ARA&A..44...49R}, with two systems in the Large Magellanic Cloud being excluded. Distances to the systems are adopted from Table 4.1 of \citet{2006csxs.book..157M} except for GS 1354-64 \citep{2004ApJ...613L.133C}\footnote{\citet{2018arXiv180411349G} derived a much smaller distance than adopted in this work using the parallax from Data Release 2 of \textit{Gaia}. The source of potential discrepancy is unclear, but here we simply adopt the previously estimated (larger) distance to give a conservative estimate.} and XTE J1650-500 \citep{2004ApJ...616..376O}.
\subsection{Maximum Searchable Distance $d_{\rm max}$}
We adopt $P=0.8\,\mathrm{days}$ and $M_\bullet=7\,M_\odot$ as the representative values of the above BHXB sample, and assume a Sun-like star with $M_\star=1\,M_\odot$ and $R_\star=1\,R_\odot$ for a typical companion. For these parameters, the amplitudes of the self-lensing and phase-curve signals are $\approx 9\times 10^{-4}$ and $\approx 6\times 10^{-3}$, respectively, where the detectability of the latter is limited by the beaming signal (cf. Figure \ref{fig:amp_period}). Assuming that these signals are detectable when the signal-to-noise ratio of the phase-folded signal is larger than 10 for the white noise model in Section \ref{ssec:white_noise},
these amplitudes are roughly translated into the limiting magnitudes of $V\approx 11$ and $V\approx 15$, respectively, or the maximum searchable distances of $d_{\rm max}\approx 0.25\,\mathrm{kpc}$ and $d_{\rm max}\approx 1.3\,\mathrm{kpc}$ for the assumed Sun-like companion. These are shown with vertical bands in Figure \ref{fig:yield_bhxb}. Note that $d_{\rm max}$ for self-lensing is rather sensitive to the property of the stellar companion because the signal scales as $R_\star^{-2}$ (Eqn. \ref{eq:sl}). It varies from $\approx0.1\,\mathrm{kpc}$ to $\approx0.4\,\mathrm{kpc}$ for A to K dwarf companions. The dependence is much smaller for the beaming signal (see Eqn.~\ref{eq:beam}).
\subsection{Cumulative Distribution of Distances to BHXBs}
Given the cumulative distance distribution of BHXBs, $N(d)$, the value of $N(d_{\rm max})$ roughly corresponds to the number of such systems whose self-lensing (if present) and phase-curve signals would be detectable with \textit{TESS}, assuming no other source of optical light variations. We estimate $N(d)$ as $N(d)=fN_{\rm confirmed}(d)$, where $N_{\rm confirmed}(d)$ is the cumulative distance distribution of the above confirmed BHXB sample and $f$ is a correction factor for the incompleteness of the X-ray binary search. \citet{2018MNRAS.474...69A} evaluated the detectability of X-ray and optical signals as well as X-ray outbursts for the above BHXB sample, and concluded that the completeness of the detection is $\approx 1/30$ on average. We show $N(d)$ with $f=30$ motivated by this result as a thin histogram in Figure \ref{fig:yield_bhxb}, along with the filled histogram for $N_{\rm confirmed}(d)$.
For $d\lesssim1\,\mathrm{kpc}$ at which no BHXB has been detected, we extrapolate the distribution assuming $N(d)\propto d^2$, i.e., they are dominated by the disk population and their space density is roughly constant. Although the validity of these assumptions is uncertain, this extrapolation only matters the estimate regarding the self-lensing population.
\subsection{Results}
Figure \ref{fig:yield_bhxb} shows that the phase-curve signal will be detectable among several tens of BHXB systems for $f=30$. While we have focused only on BHXBs, the number of potentially accessible systems doubles if we also consider X-ray binaries with confirmed NSs or candidate BHs \citep{2006csxs.book..157M, 2017hsn..book.1499C}. Moreover, there are a few hundreds of sources whose detailed properties are still unclear \citep{2006A&A...455.1165L, 2007A&A...469..807L}. Thus we expect that \textit{TESS}\ photometry will be useful for characterizing at least some X-ray binaries with phase-curve variations. For the self-lensing signal, on the other hand, $N(d_{\rm max})$ is $\mathcal{O}(1)$ or smaller. This number needs to be down-weighted by the eclipse probability, which is typically $\mathcal{O}(0.1)$, to estimate the number of actual detections. Thus the detection of self-lensing signals in X-ray binary like systems is not promising for $f=30$.
If we set $f=1000$ (dotted line in Figure \ref{fig:yield_bhxb}), we expect a few hundred systems showing detectable phase-curve signals and a few with self-lensing signals (after taking into account the eclipse probability). These numbers are roughly consistent with the estimates in the corresponding ($P$, $M_\bullet$) range from our population models in Section \ref{sec:yields} (Figures \ref{fig:yield} or \ref{fig:yield_ce}). This suggests that our population model is roughly consistent with the observed BHXB population if there exist $\approx 30$ times more detached BH--star systems than X-ray systems. The value appears to be compatible with the result of a population synthesis study \citep{2002ARep...46..667T} in the order of magnitude sense.
\begin{figure*}[ht!]
\epsscale{0.8}
\plotone{yield_bhxb.pdf}
\caption{Estimates for the cumulative distribution of distance $d$ to X-ray binaries with BHs, $N(d)$. Maximum searchable distances $d_{\rm max}$ for the self-lensing and phase-curve signals are also shown for $M_\bullet=7\,M_\star$, $P=0.8\,\mathrm{days}$, and $M_\star=1\,M_\odot$ (vertical bands). The value of $N(d_{\rm max})$ corresponds to the number of systems whose optical signals would be detectable with \textit{TESS}\ assuming no other source of optical signal. The gray histogram in the bottom right is for the observed BHXB sample, and thin histogram is the distribution inflated by a factor of $f$ to take into account the incompleteness of the X-ray binary detection. The dashed and dotted lines show the extrapolation of this distribution for $f=30$ and $f=1000$, respectively, assuming $N(d)\propto d^2$.\label{fig:yield_bhxb}
}
\end{figure*}
\vspace{0.5cm}
\section{Summary and Outlook}\label{sec:summary}
We estimated that the self-lensing and phase-curve signals induced by BH companions on tight ($\lesssim0.3\,\mathrm{au}$) but detached orbits will be strong enough to be detectable in the \textit{TESS}\ light curves of effectively $\sim 10^5$ and $\sim 10^6$ stars, respectively (Figure \ref{fig:ns}), taking into account inclination dependence of the signals. The detectability of these signals were evaluated using the stellar properties in the \textit{TESS}\ input catalog \citep{2017arXiv170600495S}, \textit{TESS}\ noise model in \citet{2015ApJ...809...77S}, and the light curves of spotted stars from the prime \textit{Kepler}\ mission \citep{2014ApJS..211...24M} to gauge the impact of stellar activities.
If combined with simple models for the population of detached BH--star binaries (Figure \ref{fig:bhpops}), these ``searchable" stars are expected to host $\sim 10$ and $\sim100$ detectable BH companions (Figures \ref{fig:yield} and \ref{fig:yield_ce}), although we cannot exclude the possibility that the latter may be associated with a comparable number of false-positives due to stellar activities. The most promising targets turned out to be BHs with masses of a few $10\,M_\odot$ and orbital periods of a few days. This large population of BHs, if identified, will reveal the BH mass function down to $\sim1\,M_\odot$ range, semi-major axis/eccentricity distribution of their orbits, positions and velocities of the systems in the Galaxy, and chemical compositions of the companion stars. These constraints will be valuable to probe mass ejection and natal kick during the BH formation, as well as the binary evolution process that may result in the observed close-in orbits. Since non-zero detections are expected from our models, even the null detection, if quantified, will provide critical information on the models of interacting binaries containing BHs. Although we have focused on systems with BHs, the \textit{TESS}\ light curves are also sensitive to NS companions down to $\sim1\,M_\odot$ (Figure \ref{fig:ns}). We may also detect phase-curve signals from optical counterparts of X-ray binaries in a quiescent phase, which potentially allow for better characterization of both known and unknown X-ray binaries.
Suppose that candidate BH companions are identified from the \textit{TESS}\ photometry, what needs to be done next? The self-lensing signal, if identified, will provide the least ambiguous targets that are also best suited for further characterization: the precise orbital inclination and period from the light curve allow for the precise mass determination with radial velocity measurements. The stellar radius estimate using the \textit{Gaia}\ parallax, if available, even allows for mass determination with light curves alone; because the pulse height constrains $M_\bullet/R_\star^2$, the BH mass can be determined at least to the precision of $20\%$, or even better if spectroscopic effective temperature of the companion is available.
The candidates identified with the phase-curve signals would require further vetting with follow-up spectroscopy to confirm their ``SB1" nature and to measure spectroscopic orbits. While the candidates identified from either method will most likely be bright enough for follow-up spectroscopy, the archival data from large spectroscopic surveys, such as APOGEE \citep{2018ApJS..235...42A}, RAVE \citep{2017AJ....153...75K}, LAMOST \citep{2012RAA....12.1197C}, and GALAH \citep{2018MNRAS.478.4513B}, will also play an essential role to complement the \textit{TESS}\ photometry, both in terms of target vetting and dynamical/chemical characterization. Indeed, those archival data alone might even provide sufficient information to confirm or reject some of the candidates. Eventually, \textit{Gaia}\ will also provide further information based on astrometric orbits and/or multi-epoch radial velocity measurements.
There are also other classes of objects that can be searched with similar methods. This includes BH/NS companions of WDs, which were not considered in this paper because the eclipse probability is small, timescales of the detectable signals may be too short for the $30$-minute cadence photometry, and most of them are likely too faint for \textit{TESS}. Nevertheless, they may still be good targets for all-sky photometric surveys from the ground \citep[cf.][]{2002A&A...394..489B}. If the \textit{TESS}\ mission is extended, and/or for stars in the overlapping regions of observing sectors, BH/NS companions on longer-period orbits around evolved stars, as identified in \citet{2018arXiv180602751T} and \citet{2019ApJ...872L..20G}, will also be within reach.
\acknowledgments
We thank Josh Winn, Adrian Price-Whelan, and Hajime Kawahara for helpful conversations. We thank Almog Yalinewich and Tsevi Mazeh for comments on the manuscript.
This paper includes data collected by the \textit{Kepler}\ mission. Funding for the \textit{Kepler}\ mission is provided by the NASA Science Mission directorate.
Work by K.M. was performed under contract with the California Institute of Technology (Caltech)/Jet Propulsion Laboratory (JPL) funded by NASA through the Sagan Fellowship Program executed by the NASA Exoplanet Science Institute.
\bibliographystyle{aasjournal}
|
\section{Introduction}
Machine learning (ML) is one of the most important concepts in computer science which has many applications in the real world such as face recognition\cite{0-1}, image processing\cite{0-2}, criminal recognition\cite{0-3}, medical images\cite{0-4}, computer vision\cite{2-9}, data mining\cite{2-8}, etc.
In the literature of ML, each sample is considered to be a vector. This means that in well-known ML algorithms like logistic regression classifier,
SVM \cite{0-5}, LDA\cite{0-6}, PCA\cite{0-7}, SVD\cite{0-8} and others,
the input data in other formats like matrix or tensor should be folded to the vector format.
This folding causes two major problems. At the first, by converting a matrix or a tensor to a long (wide) vector, the number of free variables of any learning model will be increased sharply, which can make overfitting in the model.
Also, by vectorizing the spatial relationships of features for some data like images and videos are not
considered. In other words, each datum treated individually\cite{0-9}. For example, a grayscale image represented by $m\times n$ matrix in this approach will be reshaped to a vector with the size $nm$. Therefore, not only we have many free variables but also the local spatial relations among pixels of images are not considered.
After a while in order to tackle mentioned drawbacks of vector-based methods, another approach
is known multi-linear(tensor-based) learning has been proposed which consider the data in their original multidimensional format\cite{1-1}\cite{1-2}. In this approach,
it is not necessary to reshape the data with multidimensional format anymore, and so the spatial relationships in data will be preserved\cite{0-9}. Furthermore, in contrast with vector-based methods, these multilinear methods have much fewer variables which can reduce the computational complexity and also the probability of overfitting\cite{1-0}.
In this approach different methods like Generalized low rank approximation(GLRAM) \cite{1-8}, multilinear PCA(MPCA)\cite{1-3},
multilinear LDA(MLDA)\cite{1-4}, support tensor machine(STM)\cite{1-5}, have been investigated as tensor counterparts of SVD, PCA, LDA, SVM, respectively\cite{1-6}\cite{2-3}.
Despite the mentioned appropriate properties, these methods have some problems, either.
The main problem of tensor-based methods is their limited search space which is only a subset of vector-based one. So, the probability of finding an optimal answer in such a small search space is less than the large feasible region in traditional methods by far.
The problem of dimensionality reduction (DR) is an essential tool for removing noise\cite{3-3}, reducing redundancy\cite{3-4} and so extracting appropriate new features\cite{3-5}.
Singular value decomposition is one of the main matrix decomposition methods that could be used as a DR method and is related to the well-known PCA method. Since in this method each data is considered as a vector, so the drawbacks of the vector-based methods in dealing with data like matrices and tensors exist for SVD, too. Generalized low-rank approximation of matrices (GLRAM) is an extension of SVD method for data samples in the matrix format, which by one pair of left and right projectors, transfers the data into a smaller subspace without folding the data into the vector format\cite{1-8}. ALso, in some data with small number of sample this methods works like SVD method even with smaller space.
In recent years some variants of this method have been proposed. For example, since at each iteration of GLRAM two SVD should be computed, this increases the time complexity. The authors in\cite{2-7} show that instead of SVD, its approximation by Lanczos could be used which improves its speed.
Although GLRAM preserve the spatial relationships of features and has less complexity than SVD, its search space is smaller than SVD\cite{1-8}.
In this paper, to overcome this drawback of GLRAM method, we proposed a method that by applying k-pair of left and right projectors to data,
while maintaining good properties of the GLRAM method has a larger search space. This new method will be named Multiple-Paris of GLRAM (GLRAM).
Expanding th search space of multilinear method at the first is done for STM method by
Hou et al, in their paper\cite{1-7}. They proposed a multiple-rank multilinear SVM for classification, that expands the search space of STM in order to gain a more accurate answer same as SVM.
Theoretically, we show that by this multi-pair projections the search space of the obtained method is increased. So the quality of approximations in this method will be better than GLRAM. Experiments show the quality of the proposed method.
In machine learning, there is a trade-off between the number of variables and occurrence of the overfitting. Although by increasing the number of parameters the search space of the model increases but at the same time this increase the occurrence of over-fitting.
In experiments, we found that, although the search space of our proposed method becomes a bit
larger than GLRAM, its quality always becomes better than GLRAM. Also, despite its low search space in comparison with SVD, almost gives better or equal results in comparison with SVD. This could be interpreted by the overfitting phenomenon. It's clear that for SVD with larger search space, the possibility of occurrence of the overfitting is more than
our proposed method, especially for data with the larger "$\#$feature-$\#$sample" ratio.
Also, it should be mentioned that the same idea could be applied to other tensor-based dimension reduction methods like MPCA and MLDA, easily.
The rest of this paper is organized as follows. In section 2, we analyze SVD and GLRAM methods and the relationship between them.
In section 3, we present our proposed method.
Next, the experimental results will be discussed in section 4. And finally, the conclusion stated in section 5.
\section{Related works}
In real applications, data usually contains some noisy and redundant features which affect the quality of the learning process, especially for high dimensional data. Dimensionality Reduction (DR) is a process that by transforming data into a lower dimension, tries to eliminate noise and redundancy in data\cite{1-6}\cite{1-9}.
Therefore, the occurrence of the curse of dimensionality and other undesired properties of high-dimensional spaces will be reduced, which has an important role in many applications\cite{2-0}.
In the last decades, a large number of DR techniques with different viewpoints like PCA, SVD, Fisher LDA and so on, have been investigated. Similar to other ML methods the input of the mentioned DR methods should be in vector format and so data types like images and videos (Matrix or Tensor) should be represented as a vector. This folding of high-order data to vector not only has a high complexity but also can cause losing some important spatial relations of the features in the data.
In recent years some multilinear versions of the mentioned DR methods have been proposed which are able to work with high-order data like matrices and tensors directly without reshaping them to the vector format. For example, MPCA, GLRAM, and Multilinear LDA are the multilinear versions of PCA, SVD, and LDA, respectively.
The main advantage of the multilinear methods could be summarized as follows:
\begin{itemize}
\item They maintain the structure and so the spatial relations in the data.
\item The parameters of the multilinear methods are less than the vector-based ones and so the computational complexity becomes less than linear methods. Also, for multilinear methods, the probability of occurrence of overfitting will be decreased
\end{itemize}
However, the multilinear methods are not convex typically, and also their search space is much smaller than linear ones.
In this paper, we expand the search space of GLRAM, which can cause to gain better results. It should be mentioned that this approach could be applied to other multilinear DR methods, too. In the following we review some linear and multilinear DR methods.
\subsection{Linear DR methods based on low-rank approximation}
PCA and SVD are the main DR methods that are related to each other. Let $X=[x_1,..., x_N]\in \mathbb{R}^{n\times N}$ be a centralized data set. The principal component analysis (PCA) project the data from $\mathbb{R}^n$ to $\mathbb{R}^d, (d\ll n)$, by orthogonal transformation $W\in \mathbb{R}^{n\times d}$ such that the variance of the projected data $Y=W^TX$ maximized.
It is easy to show that this can be formulated as follows:
\begin{align}\label{3-0}
&\max_W \ trace(W^TXX^TW)\\ \nonumber
&s.t. \ \ W^TW=I.
\end{align}
The $k$ eigenvectors corresponding to the $k$ large eigenvalues of $XX^T$ are the columns of the solution $W$ in Eq.\ref{3-0}.
Hence, if $X=U\Sigma V^T$ be the Singular Value Decomposition (SVD) of $X$,
the first $d$ left singular vectors $U_k=[u_1,..., u_d]$ are the $d$ first eigenvectors of $XX^T$ and so $W=U_d$ \cite{2-1}. Therefore $Y_d=U_d^TX$ becomes the projection of $X$. By SVD, it is clear that the projected data $Y_d$ becomes
\begin{equation}
Y_d=U_d^TX= \Sigma_d V_d^T,
\end{equation}
where $\Sigma_d= diag(\sigma_1, \sigma_2,...,\sigma_d)$.\\
In addition, this projection could be interpreted with another viewpoint. Since $X_k=U_d\Sigma_d V_{d}^{T}=WY_d$, is the best rank-k approximation of $X$, so we found that $Y_d$ is a reduced form of original data $X$ such that have the gives the smallest construction error and, the PCA equals to the following problem
\begin{align}
&\min_{Y_d, W} \Vert X-WY_d\Vert_F^2\\ \nonumber
&s.t. \ \ W^TW=I.
\end{align}
Therefore, the PCA and SVD dimension reduction also could be rewritten as follows:
\begin{align}\label{3-1}
&\min_{W,Y} \sum _{i=1}^N \Vert x_i- Wy_i\Vert_F^2\\ \nonumber
&s.t. \ \ W^TW=I.
\end{align}
It should be mentioned that for general data matrix $X$, where is not centralized, the SVD on $\bar{X}=[x_1-\mu, x_2-\mu, ..., x_N-\mu]$ where $\mu$ is the mean of the data is equal to applying PCA on $X$ .
\subsection{Generalized low-rank approximations}
Nowadays by increasing the usage of matrix datasets, the SVD (PCA) could not be used directly on data. Now, If we have a dataset like $\left\lbrace A_1,..., A_N\right\rbrace$ where $A_i \in \mathbb{R}^{n_1\times n_2}$, to apply SVD (PCA), every matrix $A_i$ should be fold to a vector as follows:
\begin{equation}
a_i= Vec(A_i)=\left[ a_{1}^{i^T},a_{2}^{i^T},...,a_{n_2}^{i^T}\right] ^T,
\end{equation}
where $a_j^i$ is the $j$-th column of matrix $A_i$.
It is obvious that this folding maybe destroys some spatial relations\cite{0-9}. Figure.\ref{fig.1} shows this phenomenon.
\begin{figure}[h]
\centering
\includegraphics[scale=.25]{ax.jpg}
\caption{{\footnotesize{ Vectorizing a grayscale image
}}}\label{fig.1}
\end{figure}
Recently, an extension of the DR based on a low-rank approximation to matrix data named Generalized Low-rank Approximation of Matrices(GLRAM) is investigated. In dimension reduction on $A_i$,
GLRAM by unknown orthogonal transformation matrices $L\in \mathbb{R}^{n_1\times k_1}$ and $R\in \mathbb{R}^{n_2\times k_2}$ looks for reduced data $D_i\in \mathbb{R}^{k_1\times k_2}$ where its reconstruction $LD_iR^T$ be the best low-rank approximation of $A_i$. Mathematically this can be modeled as follows
\begin{align} \label{1-3}
&\min_{\substack{L\in \mathbb{R}^{n_1\times k_1} :\ L^TL=I_{k_1}\\ R\in \mathbb{R}^{n_2\times k_2} :\ R^TR=I_{k_2}\\ D_i\in \mathbb{R}^{k_1\times k_2} : i=1,2,...,n}} \ \sum_{i=1}^N \Vert A_i-LD_iR^T\Vert _F^2.
\end{align}
Jieping Ye in his article\cite{1-8} showed that the optimal values of $L$ and $R$ should be the solution of the following maximization problem
\begin{align}\label{2-1}
&\max_{\substack{L\in \mathbb{R}^{n_1\times k_1} :\ L^TL=I_{k_1}\\ R\in \mathbb{R}^{n_2\times k_2} :\ R^TR=I_{k_2}}} \ \sum_{i=1}^N \Vert L^TA_iR\Vert _F^2,
\end{align}
and the optimal value of $D_i$ is $D_i=L^TA_iR$.
So instead of solving Eq.\ref{1-3}, tried to solve Eq.\ref{2-1}.
Also, to solve Eq.\ref{2-1} an alternating schema is used and at each step, this equation is substituted with the following two subproblems according to $R$ and $L$
\begin{align}\label{3-5}
&\max_R \ trace(R^TM_RR)\nonumber \\
&s.t\ \ R^TR=I,
\end{align}
and
\begin{align}\label{3-6}
&\max_L \ trace(L^TM_LL)\nonumber \\
&s.t\ \ L^TL=I,
\end{align}
where $M_R=\sum_{i=1}^{N} A_i^TLL^TA_i$ and $M_L=\sum_{i=1}^{N} A_iRR^TA_i^T$.
The optimal value of $L\in \mathbb{R}^{n_1 \times k_1}$ in Eq.\ref{3-5} and $R\in \mathbb{R}^{n_2\times k_2}$ in Eq.\ref{3-6} are the first $k_1$ and $k_2$ eigenvectors of $M_L$ and $M_R$ matrices, respectively\cite{1-8}.
The process of solving the GLRAM problem has been shown in Algorithm.\ref{alg:1.1} in detail.
\begin{algorithm}[!h]
\caption{GLRAM}\label{alg:1.1}
\begin{algorithmic}[1]
\REQUIRE Input: Matrices $\lbrace A_i\rbrace _{i=1}^N$ \\
\ENSURE Output: matrices L,R, and $\lbrace D_i\rbrace _{i=1}^N$
\STATE Obtain initial $L_0$ and set $\textit{i}\leftarrow 1$
\STATE While not convergent
\STATE \ \ \ \ Form the matrix $M_R=\sum_{j=1}^{n} A_j^TL_{i-1}L_{i-1}^TA_j$
\STATE \ \ \ \ Compute the $k_2$ eigenvectors $\{\phi_j^R\}, j=1,\ldots,k_2$ of $M_R$ \\
\ \ \ \ corresponding to the largest $k_2$ eigenvalues
\STATE \ \ \ \ $R_i \leftarrow[\phi _1^R,...,\phi _{k_2}^R]$
\STATE \ \ \ \ Form the matrix $M_L=\sum_{j=1}^{n} A_jR_iR_i^TA_j^T$
\STATE \ \ \ \ Compute the $k_1$ eigenvectors $\{\phi _j^L\},{j=1, \ldots,k_1}$ of $M_L$ \\
\ \ \ \ corresponding to the largest $k_1$ eigenvalues
\STATE \ \ \ \ $L_i \leftarrow[\phi _1^L,...,\phi _{k_1}^L]$
\STATE \ \ \ \ $\textit{i} \leftarrow \textit{i}+1$
\STATE EndWhile
\STATE $L\leftarrow L_{i-1}$
\STATE $R\leftarrow R_{i-1}$
\STATE For j from 1 to n
\STATE \ \ \ \ $D_j \leftarrow L^TA_jR$
\STATE EndFor
\end{algorithmic}
\end{algorithm}
Due to heavy computation for obtaining each eigenvalue, particularly in a deal with large data, the time and computational complexity will be increased which cause dire problems. Then, a variant of this method has been proposed named Bilinear Lanczos components (BLC)\cite{2-7} by using Lanczos method, operate faster than getting eigenvalues exactly.
\section{Low rank Approximation of matrices based on multiple-pairs of left and right transformations on matrix Samples}
In this section, we investigate the benefits and drawbacks of GLRAM method over SVD and based on this investigation propose a method that by preserving the benefits of GLRAM method try to cover its drawbacks. To start we analyze the relation between GLRAM and SVD.
Consider the data becomes $\lbrace A_1,...,A_N\rbrace$, where each $A_i\in \mathbb{R}^{n_1\times n_2}$. For reduction of each sample to data $D_i\in \mathbb{R}^{k_1\times k_2}$, the objective function of GLRAM method is Eq.\ref{1-3} with variables $L\in \mathbb{R}^{n_1\times k_1},R \in \mathbb{R}^{n_2\times k_2}$. But SVD works on vectorization $a_i=vec(A_i)\in \mathbb{R}^n$ of each sample. Here, the objective function of SVD (best low-rank approximation) on these data becomes
\begin{align} \label{3-3}
&\min_W \ \sum_{i=1}^N \Vert a_i-Wy_i\Vert _F^2, \ \ \ \ \ \ n=n_1n_2, \ d=k_1k_2, \\
&W\in \mathbb{R}^{n\times d}, \ W^TW=I.\nonumber\\
&y_i\in \mathbb{R}^{d},\nonumber
\end{align}
By using the properties of Kronecker product\cite{2-4} and
vectorization, it is easy to show that GLRAM model in Eq.\ref{1-3}, is equal to the following form
\begin{align} \label{3-4}
&\min_{L,R,D_{i}} \ \sum_{i=1}^N \Vert a_i-(L\otimes R)d_i\Vert _F^2.\\
&L\in \mathbb{R}^{n_1\times k_1} :\ L^TL=I_{k_1}\nonumber\\
&R\in \mathbb{R}^{n_2\times k_2} :\ R^TR=I_{k_2}.\nonumber
\end{align}
Now, by comparison between the Eq.\ref{3-3} and Eq.\ref{3-4}, The following benefits of GLRAM over SVD could be understood:
\begin{itemize}
\item GLRAM works directly on data with their own format without folding them into vectors.
\item In GLRAM we are free to choose the amount reduction in each arbitrary mode,which is not meaningful in SVD.
\item GLRAM has $n_1k_1+n_2k_2$ parameters that should be estimated, while there are $n_1n_2k_1k_2$ ones for SVD. So the complexity and so the possibility of overfitting in GLRAM is much less than SVD.
\end{itemize}
Although, fewer parameters in GLRAM has the mentioned benefits, we show that this causes GLRAM has the smaller search space over the SVD.
To show this fact mathematically, GLRAM objective function has could be rewritten as follows
\begin{align}\label{3-7}
\sum_{i=1}^{N}\Vert A_i - LD_iR^T\Vert_F^2
=\sum_{i=1}^{N}\Vert a_i-(L\otimes R)d_i\Vert_2^2.
\end{align}
So,
\begin{align}\label{4-3}
\Phi =\left\lbrace V \vert V=L\otimes R, L\in \mathbb{R}^{n_1\times k_1} \ and\ R\in \mathbb{R}^{n_2\times k_2}\ are\ orthogonal \right\rbrace,
\end{align}
denotes the search space of GLRAM. But the search space of SVD method is
\begin{align}\label{4-33}
\Psi =\left\lbrace U \vert U \in \mathbb{R}^{n\times d} \ is \ an\ orthogonal\ matrix\,\ n=n_1n_2, d=k_1k_2\right\rbrace .
\end{align}
Therefore, it is clear that $\Phi \subset \Psi$ and the search space of GLRAM is a subset of the search space of the SVD method.
We can see the summary of comparisons between vector-based methods and tensor-based ones in Table.\ref{tab.1}.
\begin{small}
\begin{table}[h]
\centering
\caption{\footnotesize{
Comparison between vector-based and tensor-based methods
}}\label{tab.1}
\begin{tabular}{| c || c | c |}
\hline
Methods & Vector-Based & Tensor-Based \\
\hline
\hline
Search space & Large & Small \\
\hline
Complexity & High & Low \\
\hline
Spatial Relationship & Ignored & Considered\\
\hline
\end{tabular}
\end{table}
\end{small}
\subsection{Proposed method by multi-pair of projections}
In this section, we try to extend the search domain of GLRAM in order to improve its quality without losing its aforesaid advantages.
To design our proposed method, we should have new insight to search region $\Psi$ of SVD method.
\begin{lemma}
Consider $W\in \mathbb{R}^{n\times k}$ be a solution of SVD model in Eq.\ref{3-3} applied on vectorization $\left\{a_i=vec(A_i)\right\}$ of data samples $\{A_i\in \mathbb{R}^{n_1\times n_2}\}, i=1,\ldots,N$ and $n=n_1n_2$ to reduce them to vectors $\{y_i\in \mathbb{R}^{d}\}$, where $d=k_1k_2$. Depended to the data there exists an integer number $l \leq \min\{n_1k_1,n_2k_2\}$ such that, $W$ could be rewritten as the following form
\[
W=\sum_{j=1}^{l}L_j \otimes R_j
\]
where $L_j\in \mathbb{R}^{n_1\times k_1}, R_j \in \mathbb{R}^{n_2\times k_2} $ for $ j=1,\ldots, l$ .
\end{lemma}
\begin{proof}
When SVD model Eq.\ref{3-3} is applied on data samples $\{A_i\in \mathbb{R}^{n_1\times n_2}\}$, according to Eq.\ref{4-33}, the solution will be lie on the feasible set $\Psi$ and each feasible solution will be an orthogonal matrix $W \in \mathbb{R}^{n\times d}$, where $n=n_1n_2$ and $d=k_1k_2$. Now we design the following partitioning on
matrix $W$
\begin{align}
&W=
\left(
\begin{array}{c|c|c}
W _{11} &\cdots & W _{1k_1}\\
\hline
\vdots & \ddots & \vdots \\
\hline
W _{n_11} &\cdots & W _{n_1k_1}
\end{array} \right),
\end{align}
where $W$ contains $n_1k_1$ numbers of block matrices $W_{ij}\in \mathbb{R}^{n_2\times k_2}, i=1,..,n_1, \ j=1,...,k_1$. Based on this partitioning we define the follwing reshaping
\cite{3-2
\begin{equation}\label{4-2}
\tilde{W}=[Vec(W _{1,1}),..., Vec(W _{n_1,1}),..., Vec(W _{1,k_1}),..., Vec(W _{n_1,k_1})]^T\in \mathbb{R}^{n_1k_1\times n_2k_2}.
\end{equation}
If $rank(\tilde{W})=l$ where $l \leq \min \{n_1k_1,n_2k_2\}$,
the SVD decomposition of $\tilde{W}$ will be
\begin{equation}\label{3-8}
\tilde{W}=\sum_{i=1}^l \sigma_i u_iv _i^T.
\end{equation}
By defining $\bar{u}_i=\sqrt{\sigma_i}u _i \in \mathbb{R}^{n_1k_1}$ and $\bar{v}_i=\sqrt{\sigma_i}v_i \in \mathbb{R}^{n_2k_2}$ the Eq.\ref{3-8} becomes
\begin{equation}
\tilde{W}=\sum_{i=1}^l \bar{u}_i\bar{v}_i^T.
\end{equation}
By the properties of Kronecker product and definition of reshaped matrix $\tilde{W}$, it is easy to show that
\begin{equation}\label{3-9}
W=\sum_{j=1}^l L_j\otimes R_j,
\end{equation}
where $Vec(L_j)=\bar{u}_j$ and $Vec(R_j)=\bar{v}_j$\cite{2-4, 3-2}.
\end{proof}
This shows that every projection matrix $W \in \Psi$ has a form like Eq.\ref{3-9} and so the projection matrix of GLRAM belongs to $\Phi$ as in Eq.\ref{4-3} is a special case of Eq.\ref{3-9} when $l=1$.\\
By this relation, if we set a $1<k<l$, using the projection matrix like
\begin{equation}\label{4-0}
W =\sum_{j=1}^k L_j\otimes R_j,
\end{equation}
enables us to use the benefits of GLRAM and SVD at the same time. This means that by the mentioned $W$ in Eq.\ref{4-0} as a projection matrix in GLRAM model, we obtained the following model which will be named Multiple-pairs of GLRAM (MPGLRAM).
\begin{align}\label{4-1}
&\min_{\{L_j, R_j\}_{j=1}^{k},
\{ D_i\}_{i=1}^{N}} \sum_{i=1}^N \Vert a_i-\sum_{j=1}^k (L_j\otimes R_j)d_i\Vert_F^2\\ \nonumber
&=\sum_{i=1}^N\Vert A_i-\sum_{j=1}^k L_jD_iR_j^T\Vert_F^2.
\end{align}
Here $d_i=Vec(D_i)$ and $D_i\in \mathbb{R}^{k_1\times k_2}$ which similar to GLRAM works on matrix data with their own format and at the same time its search space is larger than GLRAM method.
Here,
\[\overline{\Phi}_k=\left\lbrace \sum_{j=1}^k L_j\otimes R_j \vert L_j \in \mathbb{R}^{n_1\times k_1}, R_j\in \mathbb{R}^{n_2\times k_2} \right\rbrace \]
denotes the search space of MPGLRAM model in Eq.\ref{4-1} and from Eq.\ref{4-3} we can conclude that $\Phi \subset \overline{\Phi}_k$.
In the following we list the appropriate properties of the proposed model:
\begin{itemize}
\item The search space of the proposed method is larger than GLRAM method.
\item From Eq.\ref{4-1} it is clear that our proposed method is applied to the data with their own format without folding to vectors.
\item This proposed method has $(n_1k_1+n_2k_2)k$ parameters that should be estimated. Since we consider $k$ as a small number, the complexity of this method is not much higher than GLRAM, and still, the probability of occurrence of overfitting is less for this method in comparison with SVD or PCA.
\end{itemize}
\subsection{Solving the proposed model}
In the proposed MPGLRAM model, we deal with the following minimization problem:
\begin{align}\label{1-4}
&\min_{\substack{L_j\in \mathbb{R}^{n_1\times k_1} :j=1,2,...,k\\ R_j\in \mathbb{R}^{n_2\times k_2} j=1,2,...,k \\ D_i\in \mathbb{R}^{k_1\times k_2} i=1,2,...N,}} \ \sum_{i=1}^N \Vert A_i-\sum_{j=1}^k L_jD_iR_j^T\Vert _F^2.
\end{align}
To solve Eq.\ref{1-4} like GLRAM, we use a coordinate descent\cite{3-1} approach. So at each step of the algorithm, we have some subproblems that are solved only according to one variable.
So, after $p$ steps let
$L_j^{(p)}$, $R_j^{(p)}$ and ${D_i^{(p)}}$ are the estimations of projections and data matrices.
At the first in this step we consider the matrices $L_j^{(p)}$ and $R_j^{(p)}$ be known from last step and try to update reduced data $\{D_i\}, i=1,\ldots,N$. So, this leads to the following subproblem
\begin{align}\label{1-5}
\lbrace D_i^{(p+1)}\rbrace_{i=1}^N&=& \arg\min _{\{D_i\}_{i=1}^{N}} \sum_{i=1}^N\Vert A_i-\sum_{j=1}^k L_j^{(p)}D_iR_j^{(p)^T}\Vert_F^2\nonumber\\
& =&\arg\min _{\{d_i\}_{i=1}^{N}} \sum_{i=1}^{N}\Vert a_i-(\sum _{j=1}^k R_j^{(p)} \otimes L_j^{(p)})d_i\Vert_2^2,
\end{align}
where $d_i={\sf vec}(D_i), \quad a_i={\sf vec}(A_i) $.
If we set $B^{(p)}=\sum_{j=1}^{k}\left( R_j^{(p)}\otimes L_j^{(p)}\right)$, this problem can be reformulated as the following least squares problem \cite{2-5}.
\begin{align}\label{1-6}
&\min _{\{d_i\}_{i=1}^{N}}\sum_{i=1}^N \Vert a_i-B^{(p)}d_i\Vert _2^2
=\min_{D} \Vert A-B^{(p)}D\Vert _F^2,,
\end{align}
where $A=[a_1,...,a_N]\in \mathbb{R}^{n_1n_2\times N}$ and $D=[d_1,...,d_N]\in \mathbb{R}^{k_1k_2\times N}$.
This is a well-known least square problem and could be solved easily by direct and iterative matrix computation techniques.
After solving the mentioned problem we should find $L_j$, $R_j$ parameters successively by coordinate descent approach for j=1,...,k.\cite{3-1}
So if we assume
$\lbrace L_j, R_j\rbrace_{ j=1,...,j'-1,j'+1,...,k}$ and $\{D_i\}_{i=1}^{N}$ are known, we should estimate $L_{j'}$ and $R_{j'}$ in the next step. By these assumption equation Eq.\ref{1-4}, according to Eq.\ref{1-5} and Eq.\ref{1-6} leads to
\begin{align}\label{1-7}
&\min_{\substack{L_{j'}\in \mathbb{R}^{n_1\times k_1} \\ R_{j'}\in \mathbb{R}^{n_2\times k_2} }} \ \sum_{i=1}^N \| A_i-\sum\limits_{\scriptstyle j = 1\hfill\atop
\scriptstyle j \ne j'\hfill}^k L_jD_iR_j^T - L_{j'}D_iR_{j'}^T \| _F^2.
\end{align}
By replacing $\bar{A}_i= A_i-\sum\limits_{\scriptstyle j = 1\hfill\atop
\scriptstyle j \ne j'\hfill}^k L_jD_iR_j^T$ this equation becomes
\begin{equation} \label{1-8}
\min_{\substack{L_{j'}\in \mathbb{R}^{n_1\times k_1} \\ R_{j'}\in \mathbb{R}^{n_2\times k_2} }} \ \sum_{i=1}^N \Vert \bar{A_i}- L_{j'}D_iR_{j'}^T \Vert _F^2.\\
\end{equation}
For solving Eq.\ref{1-8} we use an alternating schema. At first, we fixed $L_{j'}$
and solve the problem according to $R_{j'}$.
By replacing $M_i=L_{j'}D_i$ in Eq.\ref{1-8} and regarding to properties of Trace function of matrices we have \cite{2-6}
\begin{align}
&\Vert \overline{A_i}- M_iR_{j'}^T \Vert _F^2=
tr\left((\overline{A_i}- M_iR_{j'}^T)^T(\overline{A_i}- M_iR_{j'}^T)\right)\nonumber\\
&=tr(\overline{A_i}^T\overline{A_i}-2\overline{A_i}^TM_iR_{j'}^T+R_{j'}M_i^TM_iR_{j'}^T).
\end{align}
By removing the constants in this term the Eq.\ref{1-8} leads to the following problem
\begin{align}
&\min_{R_{j'}} \ \sum_{i=1}^Ntr(-2\overline{A_i}^TM_iR_{j'}^T+R_{j'}M_i^TM_iR_{j'}^T)\nonumber\\
&=\min _{R_{j'}} \ -2\sum_{i=1}^Ntr(\overline{A_i}^TM_iR_{j'}^T)+\sum_{i=1}^Ntr(R_{j'}M_i^TM_iR_{j'}^T) \nonumber\\
&=\min _{R_{j'}} \ -2tr((\sum_{i=1}^N\overline{A_i}^TM_i)R_{j'}^T)+tr(R_{j'}(\sum_{i=1}^NM_i^TM_i)R_{j'}^T),
\end{align}
By defining $N_R=\sum_{i=1}^{N}\bar{A}_i^TM_i$ and $B_R=\sum_{i=1}^{N}M_i^TM_i$,
this optimization problem becomes
\begin{align} \label{1-9}
&\min _{R_{j'}}\ -2tr(N_RR_{j'}^T)+tr(R_{j'}B_RR_{j'}^T),
&R_{j'}\in \mathbb{R}^{n_2\times k_2}.
\end{align}
This problem is quadratic convex and so its derivative according to $R_{j'}$ in the optimal point should be zero. Therefore
by setting the derivative of the objective function equal to zero, we have
\begin{equation}
-2N_R+2R_{j'}B_R=0,
\end{equation}
and consequently, $R_j'$ will be
\begin{equation}
R_{j'}=N_RB_R^{-1}.
\end{equation}
Also, the same process could be used to find $L_{j'}$. Here by known $R_{j'}$ and setting $M_i=R_{j'}D_i^T$ Eq.\ref{1-8} according to $L_{j'}$ becomes
\begin{align}\nonumber
\min _{L_{j'}} \ \Vert \overline{A_i}^T- M_iL_{j'}^T \Vert _F^2=&tr\left((\overline{A_i}^T- M_iL_{j'}^T)^T(\overline{A_i}^T- M_iL_{j'}^T)\right)\\ \nonumber
=&\min _{L_{j'}} tr( \overline{A_i}\overline{A_i}^T-2\overline{A_i}M_iL_{j'}^T +L_{j'}M_i^TM_iL_{j'}^T)\\
=&\min _{L_{j'}} \
-2\sum_{i=1 }^N tr((\sum_{i=1}^N\overline{A_i}M_i)L_{j'}^T)+tr(L_{j'}(\sum_{i=1}^NM_i^TM_i)L_{j'}^T).
\end{align}
By replacing $N_L=\sum_{i=1}^{N} \bar{A}_iM_i$ and $B_L=\sum_{i=1}^{N}M_i^TM_i$, we have
\begin{equation}
\min _{L_{j'}}\ -2tr(N_LL_{j'}^T)+tr(L_{j'}B_LL_{j'}^T).
\end{equation}
which its solution is
\begin{equation}
L_{j'}=N_LB_L^{-1}.
\end{equation}
Since we should obtain all $k$ variables, we do previous stages $k$ times to determine $L_j$ and $R_j$, for all amount of $j=1,...,k$.
Eventually, at each time, $k-1$ parameters will be assumed to be fixed except one parameter that should be estimated.
And for the next parameter, the updated form of the previous ones will be used. Besides, we can repeat this alternative process more than once.
The details of the proposed coordinate descent process can be seen in Algorithm.\ref{alg:1.0}. It is easy to show that this proposed method based on coordinate descent approach for our proposed MPGLRAM model is a descent algorithm and at each step the objective function decrease.
\begin{algorithm}[h]
\caption{MPGLRAM} \label{alg:1.0}
\begin{algorithmic}[1]
\REQUIRE Input: matrices $\lbrace A_i\rbrace _{i=1}^N$,$k$, $iter$ \\
\ENSURE Output: matrices $\lbrace D_i\rbrace _{i=1}^N$
\STATE Initialize $\lbrace L_j,R_j\rbrace _{j=1}^k$
\STATE Construct $B=\sum_{j=1}^k (R_j\otimes L_j)$
\STATE Update $D$ by solving Eq.\ref{1-6}
\STATE for i from 1 to \textit{iter}
\STATE \ \ \ \ For each $j', j'=1,...,k$ Update ${L_j'}$ and ${R_j'}$ by ?? and ??.
\STATE \ \ Update $D$ by solving Eq.\ref{1-6}
\STATE EndFor
\end{algorithmic}
\end{algorithm}
\begin{lemma}
The proposed coordinate descent algorithm for MPGLRAM is a decreasing process.
\end{lemma}
\begin{proof}
The objective function of MPGLRAM method is
\begin{equation}
\min_{\{L_{j},R{j}\}_{j=1}^{k},\{D_{i}\}_{i=1}^{N}} f(\{D_i\}_{i=1}^N, \{L_j,R_j,\}_{j=1}^k),
\end{equation}
where
\[
f\left(\{D_i\}_{i=1}^N, \{L_j,R_j\}_{j=1}^k\right)=\sum_{i=1}^N \Vert A_i-\sum_{j=1}^k L_jD_iR_j^T\Vert _F^2.
\]
For simplicity, we set $Z_1=\{D_1,...,D_N\}, Z_{2l}=L_l,$ and $Z_{2l+1}=R_l$ for $l=1,...,k$. So, we have
\begin{equation}\label{n1}
g(Z_1,...,Z_{2k+1})=f(\{D_i\}_{i=1}^N, \{L_j,R_j,\}_{j=1}^k),
\end{equation}
and to show the decreasing property of the proposed algorithm, it is enough to indicate
\begin{equation}
g(Z_1^{(p+1)},...,Z_{2k+1}^{(p+1)})\leq g(Z_1^{(p)},...,Z_{2k+1}^{(p)})
\end{equation}
where $Z_i^{(p+1)}$,$Z_i^{(p)}$ for $i=1,...,2k+1$ denote the approximations of solution of Eq.\ref{n1} at $(p+1)$-th and $p$-th steps, respectively. Moreover, it is obvious that at $(p+1)$-th step of the proposed algorithm, coordinate descent process is applied $(2k+1)$ times, and at each step, one of the variables is updated. As a result, we can maintain that
\begin{align}\nonumber
g(Z_1^{(p+1)},...,Z_{2k+1}^{(p+1)})=&\min_{Z_{2k+1}} g(Z_1^{(p+1)},...,Z_{2k}^{(p+1)},Z_{2k+1})\\ \nonumber
&\leq g(Z_1^{(p+1)},...,Z_{2k}^{(p+1)},Z_{2k+1}^{(p)})\\ \nonumber
&\vdots \\ \nonumber
&\leq \min_{Z_l} g(Z_1^{(p+1)},...,Z_{l-1}^{(p+1)},Z_l,Z_{l+1}^{(p)},Z_{2k+1}^{(p)})\\ \nonumber
&\leq g(Z_1^{(p+1)},...,Z_{l-1}^{(p+1)},Z_l^{(p)},Z_{2k+1}^{(p)})\\\nonumber
&\vdots\\ \nonumber
&\leq g(Z_1^{(p)},...,Z_{2k+1}^{(p)}),
\end{align}
which finish the proof.
\end{proof}
This proves the decreasing behavior of the proposed algorithm to solve MPGLRAM.
The number of parameters that should be estimated in MPGLRAM method by k-pair of projections is $k(m+n)$, which is much less than $mn$ parameters in SVD. But still
close to the $(m+n)$ parameters of GLRAM method for small $k$.
Therefore, by using k pairs of projectors we expand the search space merely to find the optimal answer and protect it from tending to overfitting due to many parameters like SVD.
\section{Experimental Results}
In this section to show the quality of the proposed MPGLRAM method, we present some experiments on well-known data sets
ORL
\footnote{\text{http://www.cad.zju.edu.cn/home/dengcai/Data/FaceData.html}}
, Yale
\footnote{\text{http://web.mit.edu/emeyers/www/face\underline{ }databases.html}}
, YaleB
\footnote{\text{http://vision.ucsd.edu/~leekc/ExtYaleDatabase/ExtYaleB.html}}
,and PIE
\footnote{\text{http://featureselection.asu.edu/old/datasets/pixraw10P.mat}}.
The details of the data are listed in Table.\ref{tab.2}.
\begin{small}
\begin{table}[h]
\centering
\caption{\footnotesize{
Characters of different data sets.
}}\label{tab.2}
\begin{tabular}{ c | c | c | c }
\hline
Data & Size & Scale & Class number \\
\hline
ORL & 400 & $32\times 32$ & 40\\
\hline
Yale & 165 & $32\times 32$ & 11 \\
\hline
YaleB & 2414 & $32\times 32$ & 38\\
\hline
PIE & 210 & $44\times 44$ & 10\\
\hline
\end{tabular}
\end{table}
\end{small}
Also Fig.\ref{fig.2} shows some samples of YaleB data set.
\begin{figure}
\begin{center}
\includegraphics[scale=0.20]{YaleB.jpg}
\caption{{\footnotesize{10 samples of 4 individuals with YaleB database}}}\label{fig.2}
\end{center}
\end{figure}
Here we compare the MPGLRAM with GLRAM and SVD based based on the quality of reconstruction error and the accuracy of classification on the projected data.
\subsection{Comparison based reconstruction error}
At the first type of evaluation, we applied MPGLRAM, GLRAM and SVD methods on the mentioned data sets to project them to different smaller dimensions and used the quality of reconstructions by these projected data as evaluation of the quality of these DR methods. The quality of the reconstruction is evaluated via Root Mean Square Reconstruction Error (RMSRE) measure.
For different values of $d$ we reduced each data matrix (sample) by GLRAM and MPGLRAM
to matrices in $\mathbb{R}^{d\times d}$. Also to compare these methods with SVD we have to project the data to a vector in $\mathbb{R}^{d^2}$ by SVD method. The RMSRE of all the mentioned approximations for different values of $d$ are presented in Figure.\ref{fig.3}.
Here we applied MPGLRAM with different $k=2,\ldots,5$
\begin{figure}[!h]
\begin{center}
\centering
\includegraphics[scale=0.053]{error4dataset4-4.jpg}
\caption{{\footnotesize{ RMSRE of MPGLRAM , GLRAM and SVD versus different values of $a$ applied on a) ORL, b) Yale, c) YaleB, and d) PIE data sets.}}}\label{fig.3}
\end{center}
\end{figure}
As we can see in Figure.\ref{fig.3}, the results of the proposed MPGLRAM method by all $k$ and for all datasets and different values of $d$ are better than GLRAM. Also by increasing the value of $k$, the RMSRE of MPGLRAM is improved for all data and become near the SVD method. This confirms the effect of increasing the search space in the quality of the reconstruction of the methods.
Also, here the SVD method gives the best results. This phenomenon was predictable due to the large size of its search space.
But as we mentioned before, this large space causes a large amount of parameters that could be led to over-fitting that will be investigated in the following.
\subsection{Evaluation based on classification accuracy on Projected data}
In this section, we compare the classification accuracy on the projected data by the proposed MPGLRAM, GLRAM, and SVD methods. In our experiments, we applied \textit{$K$-fold cross-validation} measure with different amount of $K$, $K=2,5,10$.
For classification, we used discriminant analysis classifier, which in Matlab its command is ${\sf fitcdiscr}$\cite{bish}. In experiments, we found that this classifier works better than KNN and SVM for our data sets.
From Table.\ref{tab.2}, its clear that the ratio of the number of samples over the number of features for ORL, Yale, PIE, and YaleB are $0.39, 0.16, 0.11$ and $2.36$, respectively.
So, we could understand that the possibility of overfitting for YaleB is much less than other data sets. This means that we expect that for this data set the SVD method could work better than others, but for other datasets, due to the possibility of overfitting we could expect that multilinear methods could work better than SVD.
In the following by the experiments we will investigate this issue.
We start with YaleB dataset. Here we applied SVD, GLRAM, and MPGLRAM on this data to project each sample to a vector with dimension $d^2$ for SVD and matrices with dimensions $d\times d$ for GLRAM and MPGLRAM, for different values $d=5,6,7,8,9$.
Also for MPGLRAM we report only the best results of the best rank-$k$ between 2,3,4,5.
The obtained accuracies of all mentioned methods could be seen in Table.\ref{tab.5}.
\begin{small}
\begin{table}[H]
\centering
\caption{\footnotesize{
Comparison the percentage of accuracy in SVD, GLRAM, MPGLRAM on YaleB.
}}\label{tab.5}
\scalebox{0.7}[0.7]{
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{k-fold} & \textbf{d} & \textbf{SVD} & \textbf{GLRAM} & \textbf{MPGLRAM}\\ \hline
\hline
\multirow{5}{*}{\textbf{2}} & 5 & 70.34 & 62.88 & 65.53 \\ \cline{2-5}
& 6 & 74.40 & 69.76 & 74.52 \\ \cline{2-5}
& 7 & 77.84 & 74.11 & 78.21 \\ \cline{2-5}
& 8 & 80.74 & 78.67 & 81.35 \\ \cline{2-5}
& 9 & 83.97 & 81.23 & 82.80 \\ \hline
\hline
\multirow{5}{*}{\textbf{5}} & 5 & 71.41 & 63.63 & 66.78 \\ \cline{2-5}
& 6 & 75.23 & 71.00 & 76.18 \\ \cline{2-5}
& 7 & 79.33 & 75.23 & 79.08 \\ \cline{2-5}
& 8 & 82.44 & 80.07 & 82.27 \\ \cline{2-5}
& 9 & 84.18 & 82.52 & 84.05 \\ \hline
\hline
\multirow{5}{*}{\textbf{10}} & 5 & 72.04 & 64.21 & 67.61 \\ \cline{2-5}
& 6 & 75.23 & 71.33 & 76.59 \\ \cline{2-5}
& 7 & 79.78 & 75.60 & 79.45 \\ \cline{2-5}
& 8 & 82.64 & 80.41 & 82.60 \\ \cline{2-5}
& 9 & 85.05 & 82.56 & 84.42 \\ \hline
\end{tabular}}
\end{table}
\end{small}
From this table, it is clear that the MPGLRAM method works better than GLRAM for all dimensions and for all K-fold experiments. Also, although the parameters of $MPGLRAM$
is less than SVD, but almost these methods give a similar performance. This shows the power of the proposed method because in this situation the possibility of overfitting for SVM was less than other datasets and due to large search space we expected that SVD works better than MPGLRAM method, but the results do not show this.
As the second data set, we consider the Yale dataset. Table.\ref{tab.4} shows the obtained results for this dataset.
\begin{small}
\begin{table}[H]
\centering
\caption{\footnotesize{
Comparison the percentage of accuracy in SVD, GLRAM, MPGLRAM on Yale.
}}\label{tab.4}
\scalebox{0.7}[0.7]{
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{k-fold} & \textbf{d} & \textbf{SVD} & \textbf{GLRAM} & \textbf{MPGLRAM}\\ \hline
\hline
\multirow{3}{*}{\textbf{2}} & 5 & 75.15 & 73.33 & 82.42
\\ \cline{2-5}
& 6 & 75.15 & 75.76 & 82.42 \\ \cline{2-5}
& 7 & 69.69 & 69.09 & 78.18 \\ \cline{2-5}
& 8 & 58.78 & 50.30 & 58.18 \\ \cline{2-5}
& 9 & 63.03 & 56.36 & 63.03 \\ \hline
\hline
\multirow{3}{*}{\textbf{5}} & 5 & 81.82 & 83.03 & 84.85
\\ \cline{2-5}
& 6 & 83.64 & 82.42 & 87.88 \\ \cline{2-5}
& 7 & 83.64 & 81.82 & 89.09 \\ \cline{2-5}
& 8 & 83.64 & 84.85 & 89.70 \\ \cline{2-5}
& 9 & 83.03 & 81.21 & 85.45 \\ \hline
\hline
\multirow{3}{*}{\textbf{10}} & 5 & 79.39 & 83.64 & 86.06
\\ \cline{2-5}
& 6 & 85.45 & 84.85 & 87.88 \\ \cline{2-5}
& 7 & 84.85 & 86.67 & 90.30 \\ \cline{2-5}
& 8 & 87.27 & 87.27 & 91.52 \\ \cline{2-5}
& 9 & 88.48 & 87.27 & 89.09 \\ \hline
\end{tabular}}
\end{table}
\end{small}
This table shows that our proposed method not only works better than GLRAM, but even its performance is also better than the SVD method. Here, our proposed method achieves its best result in
$d=5$ with accuracy $82.42$, while this for SVD is $75.15$ with $d=5$ and for GLRAM is $75.76$ with $d=6$ in 2-fold. For 5-fold the best results of MPGLRAM, SVD, and GLRAM methods are $89.70, 83.64$ and $84.85$, receptively. Here we see that the performance of the proposed method at least $5\%$ larger than its nearest competitor, i.e., GLRAM.
For 10-fold we see that the best results of MPGLRAM, SVD, and GLRAM are $91.52, 88.48$ and $87.27$ which are obtained for dimensions $8,9,8$. Here we see that our proposed method with small $d=6$ gives better accuracy in comparisons with $GLRAM$ and $SVD$ with larger dimensions $d=7,8,9$. By these explanations, we could conclude that our proposed method works better than other methods in Yale dataset.
As a third test, we consider the results on ORL dataset. The results can be found in Table.\ref{tab.3}
\begin{small}
\begin{table}[H]
\centering
\caption{\footnotesize{
Comparison the percentage of accuracy in SVD, GLRAM, MPGLRAM on ORL.
}}\label{tab.3}
\scalebox{0.7}[0.7]{
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{k-fold} & \textbf{d} & \textbf{SVD} & \textbf{GLRAM} & \textbf{MPGLRAM}\\ \hline
\hline
\multirow{5}{*}{\textbf{2}} & 5 & 96.25 & 96.25 & 96.25 \\ \cline{2-5}
& 6 & 95.50 & 96.50 & 98.00 \\ \cline{2-5}
& 7 & 96.25 & 97.00 & 98.00 \\ \cline{2-5}
& 8 & 96.75 & 98.25 & 98.25 \\ \cline{2-5}
& 9 & 95.00 & 97.00 & 97.75 \\ \hline
\hline
\multirow{5}{*}{\textbf{5}} & 5 & 96.75 & 97.00 & 97.75 \\ \cline{2-5}
& 6 & 97.75 & 98.25 & 99.25 \\ \cline{2-5}
& 7 & 97.75 & 98.75 & 99.50 \\ \cline{2-5}
& 8 & 98.25 & 99.50 & 99.50 \\ \cline{2-5}
& 9 & 98.75 & 99.00 & 99.50 \\ \hline
\hline
\multirow{5}{*}{\textbf{10}} & 5 & 96.50 & 97.00 & 98.50 \\ \cline{2-5}
& 6 & 98.00 & 97.75 & 99.25 \\ \cline{2-5}
& 7 & 99.00 & 99.00 & 99.25 \\ \cline{2-5}
& 8 & 98.50 & 99.25 & 99.75 \\ \cline{2-5}
& 9 & 98.75 & 99.25 & 99.75 \\ \hline
\end{tabular}}
\end{table}
\end{small}
In our experiments, as we stated in Section.3, we expect that MPGLRAM gains results better than GLRAM and in some cases, even better than SVD. Eventually, not only our results are better than GLRAM, but also there are a sizable number of cases that we reach more accurate classification than SVD. Since SVD has more parameter in comparison with our method so
in overall one could see that the proposed method has better performance in comparison with SVD. Also in all situations is better than or equal to GLRAM method.
As the last experiment, we report the result for PIE dataset in Table.\ref{tab.6}.
\begin{small}
\begin{table}[H]
\centering
\caption{\footnotesize{
Comparison the percentage of accuracy in SVD, GLRAM, MPGLRAM on PIE. }}\label{tab.6}
\scalebox{0.7}[0.7]{
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{k-fold} & \textbf{d} & \textbf{SVD} & \textbf{GLRAM} & \textbf{MPGLRAM}\\ \hline
\hline
\multirow{3}{*}{\textbf{2}} & 5 & 100 & 97.14 & 100
\\ \cline{2-5}
& 6 & 98.57 & 99.04 & 100 \\ \cline{2-5}
& 7 & 99.04 & 96.67 & 100 \\ \cline{2-5}
& 8 & 98.09 & 96.67 & 99.52 \\ \cline{2-5}
& 9 & 92.85 & 95.23 & 99.04 \\ \hline
\hline
\multirow{3}{*}{\textbf{5}} & 5 & 100 & 98.57 & 100
\\ \cline{2-5}
& 6 & 99.04 & 99.04 & 100 \\ \cline{2-5}
& 7 & 99.52 & 99.04 & 100 \\ \cline{2-5}
& 8 & 99.52 & 98.57 & 100 \\ \cline{2-5}
& 9 & 99.04 & 99.52 & 100 \\ \hline
\hline
\multirow{3}{*}{\textbf{10}} & 5 & 100 & 99.04 & 100
\\ \cline{2-5}
& 6 & 99.52 & 99.04 & 100 \\ \cline{2-5}
& 7 & 99.52 & 99.04 & 100 \\ \cline{2-5}
& 8 & 99.52 & 99.04 & 100 \\ \cline{2-5}
& 9 & 99.04 & 99.04 & 100 \\ \hline
\end{tabular}}
\end{table}
\end{small}
From this table, we can see that our method works better than others.
As we see from Table.\ref{tab.6}, due to the fact that the figures for classification accuracy for almost all of these situations are just near 100,
we cannot perceive the effects of these 3 different methods very well. As a result of this, in order to show that our proposed method yields better results in comparison with the GLRAM and the SVD, we have done our experiments by another classifier, $\textit{K Nearest Neighbors}$ on PIE dataset.
In Figure.\ref{fig.5}, we demonstrate the result of 1-NN, 2-NN, and 3-NN classifiers with 2,5, and 10-fold cross-validation on PIE dataset respect to different values of $d=[2,...,16]$.
\begin{figure}[H]
\begin{center}
\begin{center}
\includegraphics[scale=0.55]{plottt.jpg}
\end{center}
\caption{{\footnotesize{ Comparison between SVD with rank-$d^2$, GLRAM, and MPGLRAM with $(d,d)$ ranks on PIE. a) 2-fold with 1-NN, b) 2-fold with 2-NN, c)2-fold with 3-NN, d) 5-fold with 1-NN, e) 5-fold with 2-NN f) 5-fold with 3-NN, g) 10-fold with 1-NN, h) 10-fold with 2-NN, and i) 10-fold with 3-NN.}}}\label{fig.5}
\end{center}
\end{figure}
From this table, the quality of the proposed MPGLRAM over other methods cloud be found clearly.
\subsection{$k$-Parameter}
In our proposed methods, we expanded the GLRAM search space by using k-pair projections. So, at first glance, this seems to play a vital role to achieve the best accuracy. While, as we have seen from the experimental results, by increasing the value of $k$ the RMSRE will be decreased as well, but in classification, the best accuracy occurred in different values of $k$. Also, even for small values of $k$ MPGLRAM gives results better than other methods. To show this issue we report the accuracy of the MPGLRAM for different datasets and different values of $d$ according to $k=2,3,4,5$ and 2-fold, 5-fold, and 10-fold cross-validation in Figure.\ref{fig.6}, Figure.\ref{fig.7}, and Figure.\ref{fig.8} respectively.
\begin{figure}[h]
\centering
\includegraphics[scale=.55]{plot4mpk2-fold.jpg}
\caption{{\footnotesize{ The accuracy of MPGLRAM regarding to different values of $k=2,3,4,5$ on a) ORL, b) Yale, c) YaleB, and d) PIE.
}}}\label{fig.6}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[scale=.55]{plot4mpk5-fold.jpg}
\caption{{\footnotesize{ The accuracy of MPGLRAM regarding to different values of $k=2,3,4,5$ on a) ORL, b) Yale, c) YaleB, and d) PIE.
}}}\label{fig.7}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[scale=.55]{plot4mpk10-fold.jpg}
\caption{{\footnotesize{ The accuracy of MPGLRAM regarding to different values of $k=2,3,4,5$ on a) ORL, b) Yale, c) YaleB, and d) PIE.
}}}\label{fig.8}
\end{figure}
Here we see in different situations the best results obtained for small values of $k$.
The value of $k$ can change from 1 to the minimum amount of the size of data. For example for a $m\times n $ matrix it can be $1\leq k \leq \min \lbrace m,n\rbrace$. When $k=1$ our method behaves like GLRAM, except its orthogonality constraints, so it has a small search space $m+n$. While when $k=\min \lbrace m,n\rbrace$, the search space is equal to the vectorized form of the matrix. So, in MPGLRAM we use $k$ to make a balance between these two methods. Therefore, in MPGLRAM, the search space will be $k(m+n)$ which can be larger than GLRAM and smaller than vectorized dimension reduction method, SVD.\\
In our experiments, we use $k=2,3,4,5$ to show that the proposed method works better than GLRAM especially when the dimension reduced to a lower value. An appropriate value for $k$ could be obtained by cross-validation approach.
\section{Conclusion}
In this paper, we proposed a novel method using the advantages of both SVD and GLRAM simultaneously to find a more accurate answer rather than GLRAM with lower complexity than SVD. This is done by k-pair of transformation in GLRAM method to enlarge its search space. By this method by few numbers of parameters which is impotant in the reduction of the possibility of over-fitting, we able to find results better than GLRAM and even SVD.
The reported experimental results confirm the quality of the proposed method. Here we found that our method
at the same time have the benefits of SVD and GLRAM methods and in fact, gives a trade-off between the size of search space(free parameters) and occurrence of overfitting and so almost gives better results in comparison with SVD and GLRAM.
Also, this approach could be used on other multilinear methods like multilinear LDA.
\section*{References}
|
\section{Introduction}
Let $F$ be a graph. We say that a hypergraph $\mathcal{H}$ is a \emph{Berge}-$F$ if there is a bijection
$f : E(F) \rightarrow E(\mathcal{H})$ such that for each $e \in E(F)$ we have $e \subset f(e)$.
In other words, $\mathcal{H}$ is a Berge-$F$ if we can embed a distinct graph edge into each hyperedge of $\mathcal{H}$ to create a copy of the graph $F$ on the vertex set of $\mathcal{H}$.
Viewed another way, given a graph $F$ we can construct a Berge-$F$ by replacing each edge of $F$ with a hyperedge that contains it.
Observe that for a fixed $F$ there are many hypergraphs that are a Berge-$F$. For simplicity, we use the term ``Berge-$F$'' to refer to this collection of hypergraphs. This definition was introduced by Gerbner and Palmer \cite{gp1} to generalize the established concepts of ``Berge path'' and ``Berge cycle'' to general graphs.
\vspace{2mm}
For a fixed graph $F$, if a hypergraph $\mathcal{H}$ has no subhypergraph isomorphic to any Berge-$F$ we say that $\mathcal{H}$ is \emph{Berge-$F$-free}. We denote the maximum number of hyperedges in an $n$-vertex $r$-uniform Berge-$F$-free hypergraph by
\[
\ex_r(n,\textrm{Berge-}F).
\]
In this paper we prove two general lemmas and use them to prove several new results and give new short proofs of previously-known bounds on $\ex_r(n,\textrm{Berge-}F)$ for various graphs $F$. In most of these cases, these proofs lead to improved theorems.
\vspace{2mm}
Results of Gy\H ori, Katona, and Lemons \cite{GyKaLe} and Davoodi, Gy\H ori, Methuku and Tompkins \cite{DavoodiGMT} establish an analogue of the Erd\H os-Gallai theorem for Berge paths.
Gy\H{o}ri and Lemons \cite{Gyori_Lemons} proved that the maximum number of hypereges in an $n$-vertex $r$-uniform $\textup{Berge-}C_{2k}$-free hypergraph (for $r \ge 3$) is $O(n^{1+1/k})$. This matches the order of magnitude of the bound found in the graph case (see the even cycle theorem of Bondy and Simonovits \cite{BS1974}). They also prove the unexpected result that the maximum number of hyperedges in an $n$-vertex $r$-uniform $\textup{Berge-}C_{2k+1}$-free hypergraph (for $r \ge 3$) is also $O(n^{1+1/k})$ which is significantly different from the graph case. Very recently, the problem of avoiding all Berge cycles of length at least $k$ has been investigated in a series of papers \cite{fkl,EGMNSTLongBerge,KostochkaLuo}.
For general results on the maximum size of a Berge-$F$-free hypergraph for an arbitrary graph $F$ see Gerbner and Palmer \cite{gp1} and Gr\'osz, Methuku and Tompkins \cite{GMTthreshold}.
\vspace{2mm}
A notion related to Berge-Tur\'an problems is the subgraph-counting problem. Following Alon and Shikhelman \cite{AlonS}, let us denote the maximum number of copies of a graph $H$ in an $n$-vertex $F$-free graph by
\[
\ex(n,H,F).
\]
For an overview of bounds on $\ex(n,H,F)$ see \cite{AlonS, gp2}.
A simple observation (that will be helpful later) by Gerbner and Palmer \cite{gp2} connects these two areas,
\begin{prop}[Gerbner, Palmer \cite{gp2}]\label{Bergecontainment}
For any graph $F$ we have \[\ex(n,K_r,F)\le \ex_r(n,\textup{Berge-}F) \le \ex(n,K_r,F)+\ex(n,F).\]
\end{prop}
Note that this proposition yields asymptotics for many graphs $F$ as $\ex(n,F)$ gives at most a quadratic gap between the upper and lower bounds.
\vspace{2mm}
Another connection appears in a paper of Palmer, Tait, Timmons and Wagner \cite{pttw} which follows from the combination of results from Mubayi and Verstra\"ete \cite{MuVe} and Alon and Shikhelman \cite{AlonS}. In particular, when $F$ is a graph with chromatic number $\chi(F) > r$, then
\[
\ex(n,K_r,F) \sim \ex_r(n,\textrm{Berge-}F) \sim \binom{k-1}{r}\left(\frac{n}{k-1}\right)^r.
\]
In the next section we prove two main lemmas that will be used in the later sections to derive new results and improve several existing theorems.
In particular, we give bounds on $\ex_r(n,\textup{Berge-}F)$ when $F$ is a path (reproving a result of Gy\H{o}ri, Katona and Lemons), a cycle (extending a result of F\"uredi and \"Ozkahya), a theta graph (improving a result of He and Tait), or a $K_{2,t}$ (extending a result of Gerbner, Methuku and Vizer).
We also establish new bounds when $F$ is a clique (which implies extensions of results by Maherani and Shahsiah and by Gy\'arf\'as) and when $F$ is a general tree.
\section{General lemmas for Berge-\texorpdfstring{$F$}{F}-free hypergraphs}
Throughout this section we will deal with graphs that are $2$-edge-colored with colors red and blue.
We will refer to a graph with such an edge-coloring as a {\it red-blue graph}.
In a red-blue graph $G$ we denote the graph spanned by the red edges by $G_{\textrm{red}}$ and the graph spanned by the blue edges by $G_{\textrm{blue}}$. Let $\mathcal{N}(H,G)$ denote the number of copies of the graph $H$ in the graph $G$. Finally, for a positive integer $r$ and a red-blue graph $G$, put
\[
g_r(G)=e(G_{\textrm{red}}) + \mathcal{N}(K_r,G_{\textrm{blue}}).
\]
For several graphs $F$, our theorems in following sections imply that the lower bound in Proposition~\ref{Bergecontainment} is sharp. However, one can easily see that the lower bound is not always sharp. For example, if $r>|V(F)|$, an $F$-free graph cannot contain a clique of size $r$, but a Berge-$F$-free hypergraph can contain hyperedges of size $r$.
The following lemma improves the upper bound in Proposition~\ref{Bergecontainment}.
Note that an essentially equivalent statement (with different proof) has recently been independently discovered by F\"uredi, Kostochka and Luo (\cite{fkl}, Lemma 4.3).
\begin{lemma}\label{main}
Let $\mathcal{H}$ be an $r$-uniform Berge-$F$-free hypergraph. Then we can construct a $F$-free red-blue graph $G$
such that
\[|\mathcal{H}| \leq g_r(G) = e(G_{\textrm{red}}) + \mathcal{N}(K_r,G_{\textrm{blue}}).\]
\end{lemma}
\begin{proof}
Let $\mathcal{H}$ be an $n$-vertex $r$-uniform Berge-$F$-free hypergraph with $\ex_r(n,\textup{Berge-}F)$ hyperedges. Define an auxiliary bipartite graph $X$ with classes $A$ and $B$ as follows. The class $A$ is the set of hyperedges of $\mathcal{H}$ and the class $B$ is the set of all pairs of vertices contained in some hyperedge of $\mathcal{H}$.
A vertex $b \in B$ is joined to $a \in A$ if the pair of vertices corresponding to $b$ is contained in the hyperedge corresponding to $a$.
A matching in $X$ induces a set of vertices in $B$; these vertices correspond to a set of pairs on the vertex set of $\mathcal{H}$, i.e., they form a graph on the vertex set of $\mathcal{H}$.
Let $M$ be a maximum-size matching in $X$ and let $G$ be the graph corresponding to the endpoints of $M$ in $B$. Note that $G$ is $F$-free as each edge of $G$ is associated with a hyperedge in $\mathcal{H}$ that contains it.
If $M$ saturates $A$, then $G$ satisfies the statement of the theorem if we color all edges red.
Now consider the case when $M$ does not saturate $A$. An {\em alternating path} in $X$ is a path that alternates between edges in $M$ and edges not in $M$ (beginning with an edge of $M$).
Let $A_1 \subset A$ and $B_1 \subset B$ be the vertices of $X$ that are not in $M$. As $M$ is maximum, there are no edges between $A_1$ and $B_1$.
Let $A_2 \subset A$ be vertices of $M$ in $A$ that are connected by an alternating path to a vertex in $B_1$. Let $B_2 \subset B$ be vertices matched to $A_2$ by $M$.
Suppose there is an edge $ab$ where $a \in A \setminus A_2$ and $b \in B_2$. By definition, there is an alternating path from $a$ to a vertex in $A_1$. Adding the edge $ab$ to this alternating path gives an alternating path with both start and end edges not in $M$, i.e., $M$ would not be maximal. Therefore, every edge incident to $B_2$ is incident to $A_2$.
Similarly, let $B_3 \subset B$ be vertices of $M$ in $B$ that are connected by an alternating path to a vertex in $A_1$. Let $A_3 \subset A$ be the vertices matched to $B_3$ by $M$.
For any edge $ab$ of $M$, if there is an alternating path from $a$ to a vertex in $B_1$, then there is no alternating path from $b$ to a vertex in $A_1$ otherwise $M$ can be increased. Therefore, $A_2$ and $A_3$ are disjoint as are $B_2$ and $B_3$.
Finally, let $A_4$ and $B_4$ be the remaining vertices in $A$ and $B$, respectively.
Thus $G$ is spanned by the pairs represented by vertices in $B_2 \cup B_3 \cup B_4$.
Now color red the edges of $G$ that are represented by vertices of $B_2$ and color blue the edges of $G$ represented by vertices of $B_3 \cup B_4$.
The number of hyperedges in $\mathcal{H}$ is
\[
|\mathcal{H}| = |A_1|+|A_2|+|A_3|+|A_4| = |B_2|+|A_1|+|A_3|+|A_4| = e(G_{\textup{red}}) + |A_1|+|A_3|+|A_4|.
\]
The vertices in $A_1 \cup A_3 \cup A_4$ are only adjacent to vertices in $B_3 \cup B_4$. Thus
\[|A_1|+|A_3|+|A_4| \leq \mathcal{N}(K_r,G_{\textrm{blue}}).\]
Thus $|\mathcal{H}| \leq e(G_{\textrm{red}}) + \mathcal{N}(K_r,G_{\textrm{blue}})$.
\end{proof}
Observe
that Lemma~\ref{main} implies the following corollary
\begin{cor}\label{blue-redtobergeF}
\[
\ex_r(n,\textup{Berge-}F) \le \max \{ g_r(G): G \textrm{ is an $n$-vertex $F$-free red-blue graph}\}.
\]
\end{cor}
Now we prove a general lemma that will be used throughout the later sections.
\begin{lemma}\label{genenew}
Let $F$ be a graph and let $F'$ be a graph resulting from the deletion of a vertex from $F$. Let $c = c(n)$ be such that $\ex(n, K_{r-1},F')\le c n$ for every $n$. Then
$$\ex_r(n,\textup{Berge-}F)\le \max\left\{\frac{2c}{r}, 1 \right\} \ex(n,F).$$
\end{lemma}
\begin{proof}
Let $G$ be an $n$-vertex $F$-free red-blue graph.
Let $m$ be the number of blue edges in $G$. Thus, the number of red edges in $G$ is at most $\ex(n,F)-m$. Now we give an upper bound on the number of $r$-cliques in $G_{\textup{blue}}$. Let $d(v)$ be the degree of $v$ in $G_{\textup{blue}}$.
Obviously the neighborhood of every vertex in $G_{\textup{blue}}$ is $F'$-free. An $F'$-free graph on $d(v)$ vertices contains at most $$\ex(d(v),K_{r-1},F')\le cd(v)$$ copies of $K_{r-1}$. Thus $v$ is contained in at most $c d(v)$ copies of $K_r$ in $G_{\textup{blue}}$.
If we sum, for each vertex, the number of copies of $K_r$ containing a vertex, then each $K_r$ is counted $r$ times.
On the other hand as $\sum_{v\in V(G_{\textup{blue}})} d(v)=2m$, we have $\sum_{v\in V(G_{\textup{blue}})} cd(v)=2cm$.
This gives that the number of $r$-cliques in $G_{\textup{blue}}$ is at most $2cm/r.$
Thus we obtain
$$
g_r(G) \le (\ex(n,F)-m)+ \frac{2c}{r}m
\le \max\left\{1, \frac{2c}{r} \right\} (\ex(n,F) - m +m) = \max\left\{1, \frac{2c}{r} \right\} \ex(n,F).
$$
Now, using Corollary~\ref{blue-redtobergeF}, the proof is complete.
\end{proof}
Note that Lemma \ref{genenew} gives an improvement (with a simpler proof) to a theorem of Gerbner, Methuku and Vizer (\cite{gmv}, Theorem 12) in the case when $c$ is a constant. We will only apply the lemma in this case. Our proof uses a special case of Claim 26 from \cite{gmv} about the number of $r$-cliques in $F'$-free graphs with $n$ vertices and $m$ edges. Another upper bound on this number is given in \cite{gmv}, which implies a better upper bound on $\ex_r(n,\textup{Berge-}F)$ in case $c$ is not a constant, i.e., in case when $F$ cannot be made acyclic by deleting a vertex.
\section{Berge trees}
Erd\H os-S\'os conjectured \cite{ErSo} that the extremal number for trees is the same as that of paths, i.e.,
\begin{conjecture}[Erd\H os-S\'os conjecture]
Let $T$ be a tree on $k+1$ vertices. Then
\[
\ex(n,T) \leq \frac{k-1}{2}n.
\]
\end{conjecture}
A proof of the conjecture for large trees was announced by Ajtai, Koml\'os, Simonovits and Szemer\'edi. The conjecture is known to hold for various classes of trees. In particular, the conjecture holds for paths by the Erd\H os-Gallai theorem \cite{Er-Ga} and more generally for spiders by a result of Fan, Hong and Liu \cite{faholi}. Recall that a \textit{spider} is a tree with at most one vertex of degree greater than $2$.
Gy\H{o}ri, Katona and Lemons \cite{GyKaLe} generalized the Erd\H os-Gallai theorem to Berge-paths. More precisely, they determined $\ex_r(n,\textup{Berge-}P_k)$ for both the range $k>r+1$ and the range $k \le r$, where $P_k$ denotes a path of length $k$ (i.e., a path on $k+1$ vertices).
\begin{thm}[Gy\H{o}ri, Katona, Lemons \cite{GyKaLe}]
\label{GKL}
If $k>r+1>3$, then
\begin{displaymath}
\ex_r(n,\textup{Berge-}P_k) \le \frac{n}{k} \binom{k}{r}.
\end{displaymath}
If $r \ge k>2$, then
\begin{displaymath}
\ex_r(n,\textup{Berge-}P_k) \le \frac{n(k-1)}{r+1}.
\end{displaymath}
\end{thm}
For the case $k=r+1$, Gy\H{o}ri, Katona and Lemons conjectured that the upper bound should have the same form as the $k>r+1$ case. This was settled by Davoodi, Gy\H{o}ri, Methuku and Tompkins \cite{DavoodiGMT} who showed that if
$k = r+1>2$, then
\begin{displaymath}
\ex_r(n,\textup{Berge-}P_k) \le \frac{n}{k} \binom{k}{r} = n.
\end{displaymath}
We generalize the above theorem to every tree and prove a sharp result under the assumption that Erd\H os-S\'os conjecture holds.
\begin{thm}\label{trees-thm}
Let us suppose that the Erd\H os-S\'os conjecture holds for all trees.
Let $T$ be a tree on $k+1$ vertices. If $k > r+1>3$, then
\[
\ex_r(n,\textup{Berge-}T) \leq \frac{n}{k}\binom{k}{r}.
\]
Moreover, if $k$ divides $n$, this bound is sharp.
If $k \leq r+1$, then
\[
\ex_r(n,\textup{Berge-}T) \leq \frac{k-1}{2}n.
\]
\end{thm}
\begin{proof}
Let us fix $k-r$ and proceed by
induction on $r$. Put $r=3$ and let us remove a leaf from $T$ to get a tree $T'$ on $k$ vertices.
By the Erd\H os-S\'os conjecture we have $\ex(n,K_2,T') = \ex(n,T') \leq \frac{k-2}{2}n$.
Thus, $c=\frac{k-2}{2}$ in the statement of Lemma~\ref{genenew}. First let us consider the case $k > r+1$. Then we have $\max\left\{\frac{2c}{r}, 1 \right\} = \max\left\{\frac{k-2}{r}, 1 \right\} = \frac{k-2}{r}$. So Lemma~\ref{genenew} gives
\[
\ex_3(n,\textup{Berge-}T) \leq \frac{k-2}{3}\ex(n,T) \leq \frac{k-2}{3} \cdot \frac{k-1}{2}n = \frac{n}{k} \binom{k}{3}
\]
proving the base case.
Now assume that
\[
\ex_{r-1}(n,\textup{Berge-}T') \leq \frac{n}{k-1}\binom{k-1}{r-1}.
\]
By Proposition~\ref{Bergecontainment} this implies
\[
\ex(n,K_{r-1},T') \leq \frac{n}{k-1}\binom{k-1}{r-1}.
\]
Similar to the base case this gives $c=\frac{1}{k-1}\binom{k-1}{r-1}$ in the statement of Lemma~\ref{genenew}. Moreover, we have $$\frac{2c}{r} = \frac{2}{r(k-1)}\binom{k-1}{r-1} = \frac{2}{k(k-1)}\binom{k}{r} = \frac{\binom{k}{r}}{\binom{k}{2}}.$$
Since $k > r+1$, we have $2 \le r \le k-2$, which means $\binom{k}{r} \ge \binom{k}{2}$, so $\max\left\{\frac{2c}{r}, 1 \right\} = \frac{2c}{r}$. So we may apply Lemma~\ref{genenew} again to get
\[
\ex_r(n,\textup{Berge-}T) \leq \frac{2}{r(k-1)}\binom{k-1}{r-1}\ex(n,T) \leq
\frac{2}{r(k-1)}\binom{k-1}{r-1} \frac{k-1}{2}n = \frac{n}{k} \binom{k}{r}.
\]
When $k$ divides $n$, the result is sharp by considering a hypergraph obtained by partitioning the vertex set into sets of size $k$ and taking all possible subsets of size $r$ in each of these sets.
Let us continue with the case $k\le r+1$. The proof is similar to the previous case; we proceed by fixing $k-r$ and by applying induction on $r$. However, now $\max\left\{\frac{2c}{r}, 1 \right\} = 1$. Thus we obtain $\ex_3(n,\textup{Berge-}T) \leq \ex(n,T) \le \frac{k-1}{2}n$ proving the base case. Similarly in the induction step we get $c = \frac{k-2}{2}$, so $\max\left\{\frac{2c}{r}, 1 \right\} = 1$ again. Thus we obtain $\ex_r(n,\textup{Berge-}T) \leq \ex(n,T) \le \frac{k-1}{2}n$ (by Lemma~\ref{genenew} and the Erd\H os-S\'os conjecture), finishing the proof.
\end{proof}
Note that the above proof uses only the fact that the Erd\H os-S\'os conjecture holds for $T$ and its subtrees. In particular, this gives a new proof of Theorem \ref{GKL} in the case $k>r+1>3$ and a sharp result for spiders (as Erd\H os-S\'os conjecture holds for paths and spiders).
By Proposition~\ref{Bergecontainment} we have $\ex(n,K_r,T) \leq \ex(n,\textup{Berge-}T)$ so Theorem~\ref{trees-thm} gives the following corollary for the subgraph-counting problem.
\begin{cor}
Let us suppose that the Erd\H os-S\'os conjecture holds for all trees.
Let $T$ be a tree on $k+1$ vertices. If $k>r+1$, then
\[
\ex(n,K_r,T) = (1+o(1))\frac{n}{k} \binom{k}{r}.
\]
\end{cor}
When $k > r+1$, then Theorem~\ref{trees-thm} gives a sharp bound. When $k \le r$ we give the following result without assuming that the Erd\H os-S\'os conjecture is true. Our proof of the theorem below is inspired by some ideas in \cite{EGMNSTLongBerge}.
\begin{thm}\label{delt}
Let $T$ be a tree on $k+1$ vertices with maximum degree $\Delta(T)$. If $k \le r$, then we have $\ex_r(n,\textup{Berge-}T) \leq (\Delta(T)-1)n$.
\end{thm}
\begin{proof}
We proceed by induction on $n$. The base case $n \le r+1$ is easy to check. Suppose the statement of the theorem is true for all values less than $n$ and let us show it is true for $n$.
Let $\mathcal H$ be a $\textup{Berge-}T$-free hypergraph on $n$ vertices. We may assume that:
\begin{equation}
\label{eq:Multifoldhalls}
\text{Any set $S \subseteq V(\mathcal H)$ is incident to at least $(\Delta(T)-1)|S|$ hyperedges of $\mathcal H$.}
\end{equation}
Indeed, otherwise we may delete the vertices of $S$ (and the hyperedges incident to them) from $\mathcal H$ to obtain a new hypergraph $\mathcal H'$ with $n-|S|$ vertices. Note that we deleted less than $(\Delta(T)-1)|S|$ hyperedges. By the induction hypothesis, $\mathcal H'$ has at most $(\Delta(T)-1)(n-|S|)$ hyperedges. This implies that $\mathcal H$ has less than $(\Delta(T)-1)|S| + (\Delta(T)-1)(n-|S|) \le (\Delta(T)-1) n$ hyperedges, and we are done.
Now consider an auxiliary bipartite graph $A$ with parts $A_1$ and $A_2$, where
$A_1 = V(\mathcal H)$ and $A_2 = E(\mathcal H)$ and $v \in A_1$ is adjacent to $e \in A_2$ in $A$ if the hyperedge $e$ is incident to the vertex $v$ in $\mathcal H$. We will use the following claim. For a vertex $a \in A_1$, let $N(a) \subseteq A_2$ denote the set of vertices adjacent to $a$ in $A$, and for a set $S \subseteq A_1$, let $N(S) = \cup_{a \in S} N(a)$.
\begin{clm}
\label{private_hyperedges}
For every vertex $a \in A_1$, there exists a set of $\Delta(T)-1$ vertices $S_a \subseteq N(a)$, such that if $a \not = a'$ then $S_a \cap S_{a'} = \emptyset$.
\end{clm}
\begin{proof}\renewcommand{\qedsymbol}{$\blacksquare$}
Note that \eqref{eq:Multifoldhalls} implies $|N(S)| \geq (\Delta(T)-1)|S|$ for any set $S \subseteq A_1$. Let us replace each vertex $a \in A_1$ with $\Delta(T)-1$ new vertices $a_1, a_2, \ldots, a_{\Delta(T)-1}$ so that each $a_i$ has the same neighborhood as $a$. Let $A'$ be the resulting bipartite graph with parts $A'_1$ and $A_2$ (note $|A'_1| = (\Delta(T)-1)|A_1|$). Then it is easy to see that for each $S \subseteq A'_1$, $|N(S)| \ge |S|$, so Hall's condition holds. Thus we can find a perfect matching $M$ in $A'$ that matches all of the vertices in $A'_1$.
Now consider an arbitrary vertex $a \in A_1$ of $A$. The corresponding vertices in $A'$ are $a_1, a_2, \ldots, a_{\Delta(T)-1}$, and let $a_i b_i \in M$ ($1 \le i \le\Delta(T)-1$) be the edges of $A'$ that match these vertices. Let $S_a = \{b_1, b_2, \ldots, b_{\Delta(T)-1}\}$. Then it is easy to see that if $a \not = a'$ then $S_a \cap S_{a'} = \emptyset$, proving the claim.
\end{proof}
Claim \ref{private_hyperedges} shows that we can assign a set $S_v$ of $\Delta(T)-1$ ``private" hyperedges to each vertex $v \in V(\mathcal H)$ such that for any two distinct vertices $v, v'$, the sets $S_v$ and $S_{v'}$ are disjoint.
Since any tree contains a leaf vertex, we can order the vertices of $T$ as $v_1,v_2,\ldots,v_{k+1}$ such that $v_1$ is a leaf and every $v_i$ ($i\ge 2$) is adjacent to exactly one of the vertices $v_1,v_2, \ldots, v_{i-1}$, say $v_{i'}$. This vertex $v_{i'}$ will be called \emph{backward neighbor} of $v_i$.
Let $v_1$ be represented by an arbitrary vertex $u_1$ of $\mathcal H$, and consider a hyperedge $h \in S_{u_1}$ (note that $h$ is a hyperedge of $\mathcal H$ containing $u_1$). Let $v_2$ be represented by a vertex $u_2 \in h$ different from $u_1$.
Let the subtree of $T$ spanned by $v_1, \ldots, v_{i-1}$ be denoted by $T_{i-1}$. Now suppose we have already found a Berge copy of $T_{i-1}$ in $\mathcal H$ (for some $i \ge 3$), where the vertex $v_l$ of $T_{i-1}$ is represented by the vertex $u_l \in V(\mathcal H)$ in this Berge copy (for each $1 \le l \le i-1$). Moreover, suppose the Berge copy has the additional property that for each $j$ with $1 \le j \le i-1$, if $v_{j'}$ is the backward neighbor of $v_j$, then the edge $v_{j'}v_j$ is represented by a hyperedge in $S_{u_{j'}}$.
We now wish to find a vertex $u_i$ in $\mathcal H$ to represent $v_i$ and obtain a Berge copy of the subtree $T_i$ spanned by $v_1, \ldots ,v_i$ such that if $v_{i'}$ is the backward neighbor of $v_i$ then the edge $v_{i'}v_i$ is represented by a hyperedge in $S_{u_{i'}}$. To this end, let $v_{i''}$ be the backward neighbor of $v_{i'}$. Note that $v_{i'}$ has at most $\Delta(T)-1$ neighbors among the vertices of $T_{i-1}$ (recall that $v_{i'}v_i$ is an edge of $T$). Since the hyperedges in $S_{u_{i'}}$ were only used to represent edges incident to $v_{i'}$ in $T$, and we assumed the edge $v_{i''}v_{i'}$ was represented by a hyperedge in $S_{u_{i''}}$, we obtain that at most $\Delta(T)-2$ hyperedges of $S_{u_{i'}}$ have been used to represent the edges of $T_{i-1}$. Hence (as $|S_{u_{i'}}| = \Delta(T)-1$) at least one hyperedge, say $h$, of $S_{u_{i'}}$ has not been used to represent any of the edges of $T_{i-1}$, so we can use it to represent the edge $v_{i'}v_i$ provided $h$ contains a vertex $u_i$ not in $T_{i-1}$ (which can be used to represent $v_i$) -- this is the case if $|h| > |V(T_{i-1})|$; this inequality holds whenever $r > k$ because $|h| = r$ and $|V(T_{i-1})| \le k$.
It only remains to deal with the case $r = k$; in this case the inequality $|h| > |V(T_{i-1})|$ does not hold only when $|V(T_{i-1})| \ge k$, so when $i = k+1$ (i.e., when we want to embed the last vertex $v_{k+1}$ of the tree). Let $v_{i}$ be the backward neighbor of $v_{k+1}$. Then, as we argued before, we can find a hyperedge $h$ in $S_{v_i}$ that has not been used to represent any edges of $T_k$. This hyperedge $h$ cannot contain any vertex $u_{k+1} \not \in V(T_k)$, because otherwise we can use $u_{k+1}$ to represent $v_{k+1}$ and we have found a Berge-$T$ in $\mathcal H$, a contradiction. So $h = V(T_k)$. Now consider the backward neighbor $v_{i'}$ of $v_i$; then we know the edge $v_{i'}v_i$ was represented by a hyperedge $h' \in S_{u_{i'}}$. Note that $h \not = h'$ since $S_{v_{i'}} \cap S_{v_{i}} = \emptyset$. So $h'$ must contain a vertex $u \not \in V(T_k)$. Moreover both $h$ and $h'$ contain the edge $v_{i'}v_i$, so we redefine the edge $v_{i'}v_i$ to be represented by $h$ and we use $h'$ to represent the edge $v_iv_{k+1}$, where the vertex $u \in V(\mathcal H)$ represents the vertex $v_{k+1}$. This gives us the desired Berge copy of $T$ in $\mathcal H$; a contradiction.
\end{proof}
Now we prove and upper bound on $\ex_r(n,\textup{Berge-}T)$ for every uniformity $r$ and every tree $T$ without the need for Erd\H os-S\'os conjecture.
\begin{proposition}
\label{without_using_ErdosSos}
Let $T$ be a tree on $k+1$ vertices. If $k>r$, then
\[
\ex_r(n,\textup{Berge-}T) \leq \frac{2(r-1)}{k}\binom{k}{r}n.
\]
If $k\le r$, then
\[
\ex_r(n,\textup{Berge-}T)\le (k-1)n.
\]
\end{proposition}
\begin{proof} Let us remove a leaf from $T$ to get a tree $T'$ on $k$ vertices.
We will show that $\ex(n,K_{r-1},T')\le \binom{k-2}{r-2}n$ by induction on $n$. Assume the statement is true for $n-1$ and prove it for $n$. Let $G$ be a $T'$-free graph on $n$ vertices. First we claim that there is a vertex $v$ of degree at most $k-2$. Indeed, otherwise we can embed $T'$ greedily into $G$. Thus the number of copies of $K_{r-1}$ containing $v$ is at most $\binom{k-2}{r-2}$. We delete $v$ (and the edges containing it) to obtain a graph $G'$ on $n-1$ vertices. By induction, $G'$ contains at most $\binom{k-2}{r-2} (n-1)$ copies of $K_{r-1}$. Thus the number of copies of $K_{r-1}$ in $G$ is at most $\binom{k-2}{r-2} (n-1) + \binom{k-2}{r-2} = \binom{k-2}{r-2} n$, as desired.
So we can choose $c = \binom{k-2}{r-2}$ in Lemma \ref{genenew} to obtain that
\[\ex_r(n,\textup{Berge-}T) \leq \max\left\{\frac{2}{r} \binom{k-2}{r-2}, 1 \right\} \ex(n,T).\]
Suppose $k > r$. Then $\binom{k-2}{r-2}\ge r-1$, which implies $\frac{2}{r} \binom{k-2}{r-2} \ge \frac{2}{r} (r-1) \ge 1$, where the last inequality holds because $r \ge 2$. Therefore, $\max\left\{\frac{2}{r} \binom{k-2}{r-2}, 1 \right\} \ex(n,T) = \frac{2}{r} \binom{k-2}{r-2} \ex(n,T)$. It is well known (and easy to see) that $\ex(n,T) \le (k-1)n$. Thus,
$$\ex_r(n,\textup{Berge-}T) \leq \frac{2}{r} \binom{k-2}{r-2} (k-1)n = \frac{2(r-1)}{k}\binom{k}{r}n.$$
On the other hand, if $k \le r$, then $\max\left\{\frac{2}{r} \binom{k-2}{r-2}, 1 \right\} = 1$. So by Lemma \ref{genenew}, we have $\ex_r(n,\textup{Berge-}T) \le \ex(n,T) \le (k-1)n$.
\end{proof}
Let us finish this section by considering stars. Let $S_k$ denote the star with $k$ edges.
\begin{thm}
If $k > r+1$, then
\[
\ex_r(n,\textup{Berge-}S_k)\le \frac{n}{k}\binom{k}{r}.
\]
Moreover, this bound is sharp whenever $k$ divides $n$.
If $k\le r+1$, then
\[
\ex_r(n,\textup{Berge-}S_k)\le \left \lfloor \frac{n(k-1)}{r} \right \rfloor.
\]
Moreover, this bound is sharp whenever $n$ is large enough.
\end{thm}
\begin{proof}
First let us consider the case $k > r+1$. As it is known that Erd\H os-S\'os conjecture holds for stars, Theorem \ref{trees-thm} gives the desired (sharp) bound.
Now consider the case when $k\le r+1$. Every vertex in a Berge-$S_k$-free graph has degree at most $k-1$. Indeed, assume $v$ is contained in the hyperedges $e_1,\dots, e_k$. Let us consider the auxiliary bipartite graph where part $A$ consists of the $(r-1)$-sets $e_1\setminus \{v\},\dots, e_k\setminus \{v\}$, and part $B$ consists of the vertices contained in these sets. We connect a vertex in $A$ to the vertices in $B$ that are contained in the corresponding $(r-1)$-set. It is easy to see that a matching covering $A$ would give us a Berge-$S_k$. If there is no such matching, then by Hall's condition there is a subset $A'$ of $A$ with $|A'|>|N(A')|$, where $N(A')$ denotes the set of neighbors of $A'$ in $B$. As every vertex in $A$ is connected to $r-1$ vertices, we have $|A'|>|N(A')|\ge r-1 \ge 1$. As two different $(r-1)$-sets together contain at least $r$ vertices, we obtain $|A'|>|N(A')|\ge r$, thus $|A'| \ge r+1 \ge k$, so $|A'| = r+1 = k$. However, $r+1$ different $(r-1)$-sets together contain at least $r+1$ vertices, a contradiction.
This gives the upper bound $\ex_r(n,\textup{Berge-}S_k)\le \lfloor n(k-1)/r\rfloor$.
It is easy to see that for large enough $n$, there exist $r$-uniform hypergraphs on $n$ vertices such that less than $r$ vertices have degree $k-2$ and the remaining vertices have degree $k-1$. This gives the desired lower bound.
\end{proof}
\section{Berge-\texorpdfstring{$K_{2,t}$}{K2t}}
Gerbner, Methuku and Vizer \cite{gmv} showed that if $t \geq 7$, then
\[
\ex_3(n,\textup{Berge-}K_{2,t})=(1+o(1))\frac{1}{6}(t-1)^{3/2}n^{3/2}.
\]
They also gave bounds for higher uniformities. Using Lemma~\ref{genenew}, we show that the same result holds for $t = 4,5,6$ as well, and also improve their bounds for higher uniformities as follows.
\begin{thm} If $t \geq r+1$, then $$\ex_r(n,\textup{Berge-}K_{2,t})\le(1+o(1)) \frac{\sqrt{(t-1)}\binom{t}{r-1}}{r t} n^{3/2}.$$
If $t\le r$, then $$\ex_r(n,\textup{Berge-}K_{2,t})\le (1+o(1)) \frac{\sqrt{t-1}}{2} n^{3/2}.$$
In particular, if $r=3$ and $t\ge 4$, then
\[
\ex_3(n,\textup{Berge-}K_{2,t})=(1+o(1))\frac{1}{6}(t-1)^{3/2}n^{3/2}.
\]
\end{thm}
\begin{proof}
We apply Lemma~\ref{genenew} with $F = K_{2,t}$, and $F' = K_{1,t}$.
In a $K_{1,t}$-free graph, since the degree of any vertex is at most $t-1$, there are at most $\binom{t-1}{r-2}$ cliques of size $r-1$ containing any vertex. Therefore, we get the following.
$$\ex(n, K_{r-1}, K_{1,t}) \le \frac{n}{r-1}\binom{t-1}{r-2}= \frac{n}{t} \binom{t}{r-1}.$$
\noindent
Thus $c = \frac{1}{t}\binom{t}{r-1}$ in Lemma~\ref{genenew}. We have $\max\{2c/r,1\}=2c/r$ if and only if $t\ge r+1$. Thus Lemma~\ref{genenew} gives
\[
\ex_r(n,\textup{Berge-}K_{2,t})\le \frac{2c}{r}\ex(n,K_{2,t}) = \frac{2}{r t} \binom{t}{r-1}\ex(n,K_{2,t})
\]
when $t \geq r+1$, and $\ex_r(n,\textup{Berge-}K_{2,t})\le \ex(n,K_{2,t})$ if $t\le r$.
Now using a result of F\"uredi \cite{F1996} which states $\ex(n,K_{2,t}) \le (1+o(1)) \frac{\sqrt{t-1}}{2} n^{3/2} $, the proof is complete.
\end{proof}
\section{Berge-\texorpdfstring{$C_{2k}$}{C2k}}
Gerbner, Methuku and Vizer \cite{gmv} improved earlier bounds due to F\"uredi and \"Ozkahya \cite{FO2017} by showing that if $k \ge 5$, then
\[\ex_3(n,\textup{Berge-}C_{2k})\le \frac{2k-3}{3} \ex(n,C_{2k}).\]
Now using Lemma~\ref{genenew}, we show that the same statement holds for $k\ge 3$, that is, it holds for $C_6$ and $C_8$ as well.
\begin{thm}
\label{2kcycle}
If $k\ge 3$, then
\[
\ex_3(n,\textup{Berge-}C_{2k})\le \frac{2k-3}{3} \ex(n,C_{2k}).
\]
\end{thm}
\begin{proof}
We apply Lemma~\ref{genenew} with $F = C_{2k}$, and $F' = P_{2k-2}$ (a path of length $2k-2$). The Erd\H{o}s-Gallai theorem implies that $\ex(n, K_2, P_{2k-2}) \le \frac{2k-3}{2}n$. Then $c = \frac{2k-3}{2}$ in the statement of Lemma~\ref{genenew}. Moreover, $2c/3 \ge 1$ whenever $k \ge 3$. Thus Lemma~\ref{genenew} gives that $$\ex_3(n,\textup{Berge-}C_{2k})\le \frac{2c}{3} \ex(n,C_{2k}) = \frac{(2k-3)}{3} \ex(n,C_{2k})$$
whenever $k \geq 3.$
\end{proof}
Note that for larger $r$, Jiang and Ma \cite{JM2016} proved $\ex_r(n,\textup{Berge-}C_{2k}) \le O_r(k^{r-2}) \ex(n,C_{2k}).$ In \cite{gmv} Gerbner, Methuku and Vizer gave a different proof of this result with an improved constant factor. With a similar calculation to that in the proof of Theorem \ref{2kcycle}, we can again reprove this result with an improved constant factor using Lemma~\ref{genenew} but for a larger range of $k$.
\section{Berge theta graphs}
A theta graph $\Theta_{k,t}$ is the graph of $t$ internally-disjoint paths of length $k$ between
a fixed pair of vertices. When $t=2$ the theta graph $\Theta_{k,2}$ is exactly the even cycle $C_{2k}$. An upper-bound of $C_{k,t} n^{1+1/k}$ (for some constant $C_{k,t}$ depending only on $k$ and $t$) on the extremal number of $\Theta_{k,t}$ is given by Faudree and Simonovits \cite{FS}. A lower bound is given by Conlon \cite{Conlon}. Recently, He and Tait \cite{HT} generalized the upper bound to the Berge setting.
\begin{thm}[He, Tait \cite{HT}]\label{hetait}
\[
\ex_r(n,\textup{Berge-}\Theta_{k,t}) \le M_{k,t,r,2} \cdot \ex(n,\Theta_{k,t}) = O(n^{1+1/k}).
\]
where $M_{k,i,r,m} = \sum_{j=1}^{k+1} \binom{mk(i-1)+jm-m}{r-m} + k + 1$.
\end{thm}
They also showed that for fixed $r$ and any $k \ge 2$, there exists $t$ such that the above upper bound is sharp in the order of magnitude.
\vspace{2mm}
Now we improve the constant factor in Theorem \ref{hetait}.
\begin{thm}\label{improvedtheta}
\[ \ex_r(\textup{Berge-}\Theta_{k,t}) \leq \begin{cases}
\frac{2}{r(r-1)} \binom{(k-1)t-1}{r-2} \ex(n,\Theta_{k,t}) & \textup{if } (k-1)t > r \\[1em]
\frac{(k-1)t-1}{(k-1)t}\ex(n,\Theta_{k,t}) & \textup{if } (k-1)t = r \\[1em]
\frac{2(t-1)}{r}\ex(n,\Theta_{k,t}) & \textup{if } (k-1)t < r
\end{cases}
\]
\end{thm}
\begin{proof}
First suppose $(k-1)t > r$. Observe that we can remove a vertex from a theta graph $\Theta_{k,t}$ to get a spider $T$ on $(k-1)t+1$ vertices. Since the Erd\H{o}s-S\'os conjecture is known to hold for spiders, Theorem~\ref{trees-thm} combined with Lemma~\ref{genenew} implies that $\ex(n,K_{r-1},T) \le \frac{1}{(k-1)t} \binom{(k-1)t}{r-1}n$. Therefore, by Lemma~\ref{genenew} we have
\begin{align*}
\ex_r(n,\textup{Berge-}\Theta_{k,t}) & \le \frac{2}{r(k-1)t} \binom{(k-1)t}{r-1}\ex(n,\Theta_{k,t})
= \frac{2}{r(r-1)} \binom{(k-1)t-1}{r-2}\ex(n,\Theta_{k,t}).
\end{align*}
Suppose now $(k-1)t = r$. Similarly combining Theorem~\ref{trees-thm} with Lemma~\ref{genenew}, we obtain $\ex_r(n,\textup{Berge-}\Theta_{k,t})\le \frac{(k-1)t-1}{(k-1)t}\ex(n,\Theta_{k,t})$.
Finally suppose $(k-1)t < r$. Then combining Theorem \ref{delt} with Lemma~\ref{genenew}, we obtain $\ex_r(n,\textup{Berge-}\Theta_{k,t})\le \frac{2(t-1)}{r}\ex(n,\Theta_{k,t})$.
\end{proof}
Note that an upper bound of $O(n^{1+1/k})$ in Theorem \ref{hetait} also follows from a result of Gerbner, Methuku and Vizer \cite{gmv} which states that if $F$ contains a vertex such that deleting it makes $F$ acyclic, then $\ex_r(n,\textup{Berge-}F)=O(\ex(n,F))$. Now we will reprove this result from \cite{gmv} using Lemma~\ref{genenew} to give an improved constant factor.
\begin{thm} Let $F$ be a graph on $k$ vertices and $v$ be one of its vertices such that deleting $v$ from $F$ we obtain a forest $F'$.
\[ \ex_r(n,\textup{Berge-}F) \leq \begin{cases}
\frac{4(r-2)}{(r-1)r}\binom{k-3}{r-2}\ex(n,F) & \textup{if } k>r+1 \\[1em]
\frac{2(k-3)}{r}\ex(n,F) & \textup{if } \frac{r}{2}+3<k\le r+1 \\[1em]
\ex(n,F) & \textup{if } k\le \frac{r}{2}+3
\end{cases}
\]
\end{thm}
\begin{proof} If $k>r+1$, then Proposition \ref{without_using_ErdosSos} implies that $\ex(n,K_{r-1},F')\le\ex_{r-1}(n,\textup{Berge-}F')\le \frac{2(r-2)}{k-2}\binom{k-2}{r-1}n$. Thus if $k>r+1$, we use Lemma~\ref{genenew} with $c = \frac{2(r-2)}{k-2}\binom{k-2}{r-1}$. If $k>r+1$, then it is easy to see that $\max\{\frac{2c}{r},1\}=\frac{2c}{r} = \frac{4(r-2)}{(k-2)r}\binom{k-2}{r-1} = \frac{4(r-2)}{(r-1)r}\binom{k-3}{r-2}$, so Lemma~\ref{genenew} gives the desired bound.
If $k\le r+1$, then Proposition \ref{without_using_ErdosSos} implies that $\ex(n,K_{r-1},F')\le\ex_{r-1}(n,\textup{Berge-}F')\le (k-3)n$, thus we can use Lemma~\ref{genenew} with $c = k-3$. If $\frac{r}{2}+3<k\le r+1$, then $\max\{\frac{2c}{r},1\}=\frac{2c}{r}$, while if $k\le \frac{r}{2}+3$, then $\max\{\frac{2c}{r},1\}=1$. In both cases Lemma~\ref{genenew} gives the desired bound.
\end{proof}
\section{Berge-\texorpdfstring{$K_r$}{Kr}}
In this section, for brevity, we use the term {\it $r$-graph} to refer to an $r$-uniform hypergraph.
Let $T_r(n,k)$ be the complete $k$-partite $n$-vertex $r$-graph where all the parts have size $\lfloor n/k\rfloor$ or $\lceil n/k \rceil$. Erd\H os \cite{erd} showed that the Tur\'an graph $T_2(n,k-1)$ maximizes not only the number of edges among $n$-vertex $K_k$-free graphs, but also the number of $K_r$'s for any $r<k$.
\begin{thm}[Erd\H os, \cite{erd}]\label{erd} For any $k, r$ and $n$,
\[
\ex(n,K_r,K_k)=\mathcal{N}(T_2(n,k-1),K_r).
\]
\end{thm}
Observe that if $2<r<k$, then $\ex(n,K_r,K_k)$ is at least cubic, so Proposition \ref{Bergecontainment} gives asymptotically tight bounds on $\ex_r(n,\textup{Berge-}K_k)$. In this paper we are interested in exact results for every $n$. For $r < k$, let us define the threshold $n_0=n_0(k,r)$ to be the smallest possible integer such that $T_r(n,k-1)$ is the largest Berge-$K_k$-free $r$-graph for every $n\ge n_0$. We will see that $n_0$ exists.
\vspace{2mm}
The expansion $F^{+r}$ of a graph $F$ is an $r$-uniform hypergraph obtained by adding $r-2$ distinct new vertices to each edge of $F$. Pikhurko \cite{pik}, improving an asymptotic result of Mubayi \cite{mub}, showed that for $r<k$ there is $n_1=n_1(k,r)$ such that the largest $K_k^{+r}$-free hypergraph is $T_r(n,k-1)$ for $n\ge n_1$. Observe that the expansion of $F$ is one specific Berge copy of $F$. Thus Pikhurko's result shows $T_r(n,k-1)$ is the largest Berge-$K_k$-free $r$-graph if $n\ge n_1$ provided $r<k$, so $n_0$ exists and is at most $n_1$. However, the value of $n_1$ which follows from Pikhurko's proof is quite large.
\vspace{2mm}
Our goal is to give better bounds on the threshold $n_0(k,r)$.
In this direction, research has been carried out for $3$-graphs: Maherani and Shahsiah \cite{masha} showed that for $k\ge 13$, we have $n_0(k,3)=0$, i.e., for every $n$, $T_3(n,k-1)$ contains the largest number of hyperedges among all Berge-$K_k$-free $3$-graphs, provided $k\ge 13$. Gy\'arf\'as \cite{gyarfas} proved that $n_0(4,3)=6$. The situation is different if $n < n_0(4,3) = 6$. In this case Gy\'arf\'as showed $\ex_3(5,\textup{Berge-}K_4)=5$; moreover, any $3$-uniform hypergraph on $5$ vertices with $5$ hyperedges shows that this bound is sharp (because we need at least $6$ hyperedges to form a $\textup{Berge-}K_4$). For $n \le 4$, trivially $\ex_3(n, \textup{Berge-}K_4) = \binom{n}{3}$ because a complete $3$-graph on at most $4$ vertices is $\textup{Berge-}K_4$-free.
\vspace{2mm}
Our theorem below implies most of the results for $3$-graphs mentioned above and provides new bounds for all uniformities $r$.
\begin{thm}\label{compl} For any $n$, $k$ and $r$, we have $$\ex_r(n, \textup{Berge-}K_k) \le \max\{\ex(n, K_k), \ex(n, K_r, K_k)\}.$$
\end{thm}
Our proof is based on a careful adaptation of Zykov's symmetrization method (see \cite{zyk}) to red-blue graphs.
\begin{proof}[Proof of Theorem \ref{compl}]
We will show that if $G$ is an $n$-vertex $K_k$-free graph $G$, then $g_r(G)$ is maximized when $G$ is the Tur\'an graph $T_2(n,k-1)$ with all edges of the same color.
This implies $g_r(G)$ is at most $\max\{\ex(n, K_k), \ex(n, K_r, K_k)\},$ so applying Lemma \ref{main} completes the proof.
Let $\cG$ be the family of $n$-vertex $K_k$-free red-blue graphs that maximize $g_r$.
Let $v_1,v_2,\dots, v_n$ be the vertex set of each of these graphs.
Then let $\cG'$ be the subfamily of red-blue graphs in $\cG$ with the maximum number of edges.
Let $\cG''$ be the subfamily of graphs in $\cG'$ with the maximum number of red edges. Let $d_{\textup{red}}(v)$ denote the number of red edges incident to a vertex $v$.
Let $\cG_1$ be the subfamily of red-blue graphs in $\cG''$ which maximize $d_{\textup{red}}(v_1)$. We recursively define further subfamilies. For $2\le i\le n$, let $\cG_i$ be the subfamily of red-blue graphs in $\cG_{i-1}$ which maximize $d_{\textup{red}}(v_{i})$.
\begin{clm}\label{symm}
Any red-blue graph $G\in \cG_n$ is a complete multipartite graph such that for any pair of classes $A,B$ all edges between $A$ and $B$ are of the same color.
\end{clm}
\begin{proof}[Proof of Claim \ref{symm}]\renewcommand{\qedsymbol}{$\blacksquare$}
Let $G$ be an arbitrary red-blue graph in $\cG_n$.
For two non-adjacent vertices $u$ and $v$ of $G$ let us create a new graph $G'$ by
deleting all edges incident to $u$ and adding new edges that join $u$ to the neighbors of $v$. Moreover, we color each new edge $uw$ with the same color as the edge $vw$.
We call this procedure {\it symmetrization} and say that we \textit{symmetrize $u$ to $v$}.
We claim that $G'$ is $K_k$-free. Indeed, a copy of $K_k$ in $G'$ must contain a new edge and therefore must contain the vertex $u$. As $u$ and $v$ are non-adjacent, the $K_k$ does not include $v$. However, since $u$ and $v$ have the same neighborhood in $G'$ this means that there is a copy of $K_k$ containing $v$ in $G$; a contradiction. This implies that $G$ remains $K_k$-free under symmetrization.
For a vertex $v$, let $d^*(v)$ denote the number of red edges incident to $v$ plus the number of blue $r$-cliques of $G$ containing $v$. When we symmetrize $u$ to $v$, the number of red edges plus the number of blue $r$-cliques decreases by $d^*(u)$ and then increases by $d^*(v)$. Since $G\in \cG$, we have that the value of $g_r(G)$ is maximal. Thus, $d^*(u)\ge d^*(v)$.
As we could also symmetrize $v$ to $u$, we must have $d^*(u)=d^*(v)$.
Similarly, as $G\in \cG'$ we must have that $d(u)=d(v)$ as otherwise we can symmetrize $u$ to $v$ (or $v$ to $u$) to get a graph with more edges. This would imply that $G \not \in \cG'$; a contradiction. A similar argument combined with the fact that $G \in \cG''$ implies that $d_{\textup{red}}(u)=d_{\textup{red}}(v)$ .
Now we show that $G$ is a complete multipartite graph. Assume not, then it is easy to see that we have three vertices $x,y,z$ such that $y$ and $z$ are adjacent, but $x$ is adjacent to neither $y$ nor $z$. By the previous paragraph, we have $d^*(y)=d^*(x)=d^*(z)$ and $d(y)=d(x)=d(z)$. Now we symmetrize $y$ to $x$ to obtain $G'$ and then $z$ to $x$ to obtain $G''$. Note that in the first symmetrization step $g_r$ and $d^*(y)$ do not change, while $d^*(z)$ does not increase (it might decrease if the edge $yz$ is red or contained in a blue $K_r$). This implies $g_r(G)=g_r(G')\le g_r(G'')$, which gives $g_r(G)=g_r(G')= g_r(G'')$ as $G\in \cG$. Similarly, in the first symmetrization step the total number of edges and $d(y)$ does not change, but this time $d(z)$ decreases by one. Thus in the second symmetrization step the total number of edges increases, a contradiction to the assumption that $G$ is in $\mathcal G'$.
Thus we obtained that $G$ is a complete multipartite graph, so any two non-adjacent vertices $u$ and $v$ have the same neighborhood. Let us define
\[X(u,v)=\{x\in V: ux \text{ and $vx$ are of different colors}\}.\]
Assume $X(u,v)$ is non-empty and let $i$ be the smallest index with $v_i\in X(u,v)$. Without loss of generality $v_i$ is connected to $u$ by a red edge. Then we symmetrize $v$ to $u$. By the above observations, $g_r$, the total number of edges and the total number of red edges does not change. Also $d_{\textup{red}}(u)$ and $d_{\textup{red}}(v)$ do not change, and $d_{\textup{red}}(x)$ does not change for every $x\not\in X(u,v)$. In particular, $d_{\textup{red}}(v_j)$ does not change for $j<i$, showing $G'\in \cG_{i-1}$. But $d_{\textup{red}}(v_i)$ increases, contradicting our choice of $G$.
This finishes the proof of the claim. Indeed, assume $xy$ is red and $x'y'$ is blue such that $x$ and $x'$ are in the same class and $y$ and $y'$ are together in a different class. If the edge $xy'$ is blue, then $X(x,x')$ is non-empty, while if $xy'$ is red, then $X(y,y')$ is non-empty.
\end{proof}
Observe that if $u$ and $v$ are in the same part $A$ of a graph $G\in \cG_n$, then $d^*(u)=d^*(v)$. Let this value be denoted by $d^*(A)$. Similarly we have $d_{\textup{red}}(u) = d_{\textup{red}}(v)$, so denote this value by $d_{\textup{red}}(A)$.
Note that a red-blue graph has an \textit{underlying} uncolored graph with the same vertex and edge set.
Let $G_0$ be an arbitrary red-blue graph in $\cG_n$. Note that $G_0$ is a complete multipartite graph with classes $A_1,\dots,A_j$ (note that $j \le k-1$ as $G_0$ is $K_k$-free). There may be several red-blue graphs in $\cG''$ with the same underlying graph $G_0$; let $\cH$ denote the family of such red-blue graphs.
By Claim~\ref{symm} for any red-blue graph in $\cH$ all edges between a pair of classes have the same color.
Let $\cH_1$ denote the subfamily of those graphs in $\cH$ which maximize $d_{\textup{red}}(A_1)$. We recursively define further subfamilies. For $2\le i\le j$, let $\cH_i$ be the subfamily of graphs in $\cH_{i-1}$ which maximize $d_{\textup{red}}(A_{i})$.
\begin{clm}\label{equi}
In any graph $G\in \cH_{j}$, being connected by red edges is an equivalence relation.
\end{clm}
\begin{proof}[Proof of Claim~\ref{equi}]\renewcommand{\qedsymbol}{$\blacksquare$}
Let us consider two parts $A$ and $B$ of $G$ that are connected by by red edges. We define another \textit{symmetrization step} as follows: For every part $C$ (distinct from $A$ and $B$), we change the color of the edges between $A$ and $C$ to the color of the edges between $B$ and $C$; we refer to this symmetrization step by saying that we \textit{symmetrize $A$ to $B$}.
Note that the underlying graph $G_0$ does not change. In this way, the number of red edges plus blue $r$-cliques decreases by $|A|d^*(A)-|A||B|$ and then increases by $|A|d^*(B)-|A||B|$. This implies that $d^*(A)\ge d^*(B)$ as $G\in \cG$. As we can symmetrize $B$ to $A$ we obtain that $d^*(A) = d^*(B)$. Similarly, $d_{\textup{red}}(A)=d_{\textup{red}}(B)$ because if $d_{\textup{red}}(A)<d_{\textup{red}}(B)$ then symmetrizing $A$ to $B$ would increase the number of red edges while $g_r$ and the number of edges does not change, contradicting our assumption that $G\in \cG''$.
Now we show that being connected by red edges is an equivalence relation. Assume for a contradiction that there are two parts $A$ and $B$ connected by red edges and we have at least one other part connected by red edges to one of them and by blue edges to the other, and let $A_i$ be such a part with the smallest index $i$. Without loss of generality $A_i$ is connected to $A$ by red edges and to $B$ by blue edges, then we symmetrize $B$ to $A$. The resulting graph $G'$ is in $\cH$ as the underlying graph $G_0$ and $g_r$ do not change, and for any two classes of $G'$ the edges between them are of the same color. Also $d_{\textup{red}}(A)$ and $d_{\textup{red}}(B)$ do not change and $d_{\textup{red}}(A_{j})$ does not change for $j<i$. This shows $G'\in \cH_{i-1}$, but $d_{\textup{red}}(A_i)$ increases, showing $G$ cannot be in $\cH_i$; a contradiction.
\end{proof}
Thus we found a red-blue graph $G$ that is complete multipartite, for any two of its classes the edges between them are of the same color, being connected by red edges in $G$ is an equivalence relation, and $G$ maximizes $g_r$ among $K_k$-free red-blue graphs. We will show that all the edges of $G$ are of the same color.
If there are no red edges in $G$, we are done. Let $A$ and $B$ be classes connected by red edges. Let us assume first that the vertices of $A$ are not in a blue $r$-clique. Then we can change all the edges incident to $A$ to red. If there was any change, $g_r$ increases, which would be a contradiction. Thus all the edges incident to $A$ are red in $G$, but then all the edges in $G$ are red by Claim \ref{equi} and we are done.
Hence there is a blue $K_r$ intersecting $A$. Thus there are at least $r-1$ classes $B_1,\dots,B_{r-1}$ in $G$ that are connected to $A$ and each other by blue edges. Note that $B\neq B_i$ for any $i$ as B is connected to $A$ by red edges. Then for any $i$, $B_i$ is connected to $B$ by blue edges, by applying Claim \ref{equi}. Let us now change all the edges between $A$ and $B$ to blue, and let $G'$ be the resulting graph. We claim that this way we delete $|A||B|$ red edges and add at least $(r-1)|A||B|$ blue $r$-cliques, thus $g_r$ increases, a contradiction. To prove this claim, let us pick one vertex from $r-2$ parts among the $B_i$'s, one vertex from $A$ and one vertex from $B$. This way we obtain a new blue $r$-clique. There are at least $r-1$ ways to pick $r-2$ $B_i$'s and one vertex from each of them. There are $|A||B|$ ways to pick the remaining two vertices from $A$ and $B$.
We obtained that $G$ is monochromatic, thus we have $g_r(G)\le \max\{\ex(n,K_k),\ex(n,K_r,K_k)\}$. Since $G$ maximizes $g_r$, the proof is complete.
\end{proof}
The following corollary of Theorem \ref{compl} determines $\ex_r(n, \textup{Berge-}K_k)$ exactly for every $n$ for any $k>r+2$.
\begin{corollary}
\label{nothing}
Let $r \ge 2$. If $k>r+2$, then $T_r(n,k-1)$ has the maximum number of hyperedges among all Berge-$K_k$-free $r$-graphs for every $n$.
\end{corollary}
\begin{proof} If $n<k$, then the statement is trivial. For $n\ge k$, we are going to show that the Tur\'an graph $T_2(n,k-1)$ contains more copies of $K_r$ than edges. This statement together with Theorem \ref{compl} implies our corollary. To prove this statement we use induction on $n$. Consider the base case $n=k$. In this case observe that the Tur\'an graph $T_2(k,k-1)$ contains exactly two vertices in one part and exactly one vertex in each of the other parts. The number of edges is $\binom{k-2}{2}+2(k-2)$ and the number of copies of $K_r$ is $\binom{k-2}{r}+2\binom{k-2}{r-1}$. It is easy to see the latter is at least the former if $k > r+2$.
Let us assume the statement holds for $n$, and the Tur\'an graph $T_2(n,k-1)$ has parts $A_1,\dots,A_{k-1}$. We add one more vertex $v$ to, say, part $A_{k-1}$, to obtain the Tur\'an graph $T_2(n+1,k-1)$. Let $G$ denote the subgraph induced by the other parts $A_1,\dots,A_{k-2}$. The number of edges added (by adding $v$) is the number of vertices of $G$, say $n'$, while the number of $r$-cliques added is the number of $(r-1)$-cliques in $G$. Observe that $G$ is the Tur\'an graph $T_2(n',k-2)$. Thus it is enough to prove that the number of $(r-1)$-cliques in the Tur\'an graph $T_2(n',k-2)$ is at least $n'$. Once again, we can prove this statement by induction on the number of vertices. Note that $n'\ge k-2>r$. For the base case $n' = k-2$ the Tur\'an graph $T_2(n',k-2)$ is a complete graph, so it has $\binom{k-2}{r-1} $ copies of $K_{r-1}$, and it is easy to see that $\binom{k-2}{r-1} \ge k-2$ as $k>r+2$.
For the induction step, if we add any vertex, the number of $(r-1)$-cliques increases by at least one, finishing the proof.
\end{proof}
Note that if $k\le r$, then the Tur\'an hypergraph is empty. If $k=r+1$ or $k=r+2$ and $r\ge 3$, then $\ex_r(n, \textup{Berge-}K_k)$ is not given by the Tur\'an hypergraph $T_r(n,k-1)$ for small $n$, for example when $n=k$. Indeed, if $n=k=r+1$, then the Tur\'an hypergraph $T_r(n,k-1)$ contains two hyperedges, while even the complete $r$-uniform hypergraph on $n=r+1$ vertices does not contain a Berge-$K_k$ and it has $r+1>2$ hyperedges. If $n=k=r+2$, then the Tur\'an hypergraph $T_r(n,k-1)$ contains $2r+1$ hyperedges, while any $r$-uniform hypergraph on $n$ vertices with $\binom{k}{2}-1>2r+1$ hyperedges does not contain a Berge-$K_k$.
\vspace{2mm}
However, it is not hard to compute the upper bound that Theorem \ref{compl} gives on the thresholds $n_0(r+1,r)$ and $n_0(r+2,r)$ for any fixed $r$. One can easily see that the upper bound we obtain this way on $n_0(r+1,r)$ is $r+2\log r + O(1)$, while the upper bound on $n_0(2+1,r)$ is $r+\log r + O(1)$ as $r$ increases.
\vspace{2mm}
Let us finish this section by considering the $3$-uniform case. Theorem \ref{compl} and Corollary \ref{nothing} imply the following bounds for $3$-graphs: $n_0(4,3)\le 9$, $n_0(5,3)\le 7$, and for $k\ge 6$ we have $n_0(k,3)=0$.
For $\textup{Berge-}K_5$, by a simple (but tedious) case-analysis, one can show that $n_0(5,3) = 6$. The situation is different when $n < n_0(5,3) = 6$: Firstly, $\ex_3(5, \textup{Berge-}K_5) = 9$ since any $3$-uniform hypergraph on $5$ vertices with $9$ hyperedges is Berge-$K_5$-free. Now if $n < 5$, $\ex_3(n, \textup{Berge-}K_5) = \binom{n}{3}$ because a complete $3$-graph on fewer than $5$ vertices is obviously $\textup{Berge-}K_5$-free. Combining these results with the results of Gy\'arfas \cite{gyarfas} (concerning $\textup{Berge-}K_4$) mentioned earlier, we have the exact value of $\ex_3(n, \textup{Berge-}K_k)$ for all $n$ and $k>3$, as summarized below:
\begin{corollary} Let $n\ge 1$ and $k\ge 4$ be integers. Then
\[ \ex_3(n,\textup{Berge-}K_k) = \begin{cases}
|T_3(n,k-1)| & \textup{if } k\ge 6 \textup{ or } k=5, \, n\ge 6 \textup{ or } k=4,\, n\ge 6 \\[1em]
\binom{n}{3} & \textup{if } k=5, n\le 4 \textup{ or } k=4, n\le 4 \\[1em]
5 & \textup{if } k=4,\,n=5 \\[1em]
9 & \textup{if } k=5,\,n=5.
\end{cases}
\]
\end{corollary}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.